International Workshop PROCEEDINGS - gichd

VIENNA UNIVERSITY OF
TECHNOLOGY - AUSTRIA
ROYAL MILITARY
ACADEMY - BELGIUM
International Workshop
Robotics and Mechanical
assistance in
Humanitarian Demining and
Similar risky interventions
Brussels-Leuven, Belgium
16-18 June 2004
PROCEEDINGS
WITH THE SUPPORT OF THE ‘FONDS NATIONAL DE LA
RECHERCHE SCIENTIFIQUE ‘ and the General Direction of
Scientific Research of the FRENCH COMMUNITY OF BELGIUM

CONTENT OF THE HUDEM’2004 PROCEEDINGS
HUDEM4-1 The Mine Ban Convention: past, present and
future
HUDEM4-2 The Mine Ban Convention and the Mine actions
technologies
HUDEM4-3 The Outdoor Robotics: challenges and
requirements, promising tools and dual-use
applications
HUDEM4-5 Robotics Systems for unstructured Outdoor
Environments
HUDEM4-6 MAXML and PARADIS: Mine Actions
software management tools
HUDEM4-7 Remote Sensing Systems for robotics and
aquatic related Humanitarian Demining and
UXO Detection
HUDEM4-8 Air/Ground Robotic Ensembles for Risky
applications
HUDEM4-9 Light, Low-cost, hybrid robotic platform for
Humanitarian demining
HUDEM4-10 Ground adaptative manipulation of GPR for
mine detection system
HUDEM4-11 A system for monitoring and controlling a
climbing and walking robot for landslide
consolidation
HUDEM4-12 Sensor Head and Scanning Manipulator for
Humanitarian Demining
HUDEM4-13 Life Driving Video System (LDV) command,
control and communicate
With the robotic device
P.T. Gscwind, M.Stuber, Innosuisse Corp,
Geneva
HUDEM4-15 Robotic systems for Humanitarian Demining:
modular and generic approach. Cooperation
under IARP and ITEP
HUDEM4-18 EUDEM-2, a useful tool for the Humanitarian
Demining Researchers/End-Users Community
HUDEM4-19 A concept of implementing technology to
encourage economic growth in de-mining
M. Lint, Ambassador
Ministry of Foreign Affairs
Belgium, President of the 4 th MSP
Prof M.Acheroy, RMA
D. Caltabiano, D.Longo, G.Muscato,
m;Prestifilippo,
G.SpampinatoUniversità degli Studi di
Catania, Chairman WG CLAWAR
Prof V.Gradetsky, M.Knyazkov,
L.Meshman, M.Rachkov , Institute for
Mechanical Problems, University of
Moscou and Ministry of the Russian
Federation for Civil Defence
Sébastien Delhay, Vinciane Lacroix ,
Mahamadou Idrissa, SIC, RMA
Dr Ch.Bostater, T.Ghir, L.Bassetti,
Florida Institute of Technology
S.Lacroix, R.Chatila, LAAS Toulouse
Prof G.Genta, Carabelli,S Polytech
Torino, Italy
Drs Hibeno.Yabushita, Kazuhiro
Kosuge, Yasuhisa Hirata, IRS Tohoku,
JPN
Dr L.Steinicke, et Al SAS, Brussels
Dr P.Gonzalez-de-Santos,E.Garcia,
J.Combano,J.Estremera, IAI-CSIC,
Madrid
P.T. Gscwind, M.Stuber, Innosuisse
Corp, Geneva
Yvan Baudoin, RMA
Mrs K.Debruyn,Prof H.Sahli, Prof
J.Cornelis, VUB Brussels
R. Dixon, Demining Systems, UK
HUDEM4-20 A concept for a humanoid demining robot. Peter Kopacek, Man-Wook Han,
Bernhard Putz, Edmund Schierer,
Markus Würzl, Vienna University of
HUDEM4-21 Towards a Semi-Autonomous Vehicle
for Landmine Neutralisation
Technologies, Austria
Kym M. Ide, Brian J. Jarvis, Bob
Beattie, Paul Munger and Leong Yen
Australian Govt, Dept of Defense, Def

HUDEM4-22 Results of Open-Air Field Trials on Chemically-
Science and Technology Organisation
B.C.Maglich, T-F Chuang, M.Y.Lee,
Specific Identification of UXO Fillers and Ch.Druey, HiEnergy Technologies, Inc
Buried AT landmine using portable Non-Pulsed Calfornia,USA, J.W.Price,G.Miller, ,
Non-directed Fast Neutron Stoichiometer L.A U.California
HUDEM4-23 Feature-Level Sensor Fusion for a Demining Svetlana Larionova, Lino Marques and
Robot
Anibal T. de Almeida, University of
Coïmbra
HUDEM4-24 European Project of Remote Detection SMART Dr Yann Yvinec, RMA
HUDEM4-25 Modelling of a Metal Detector Dr Pascal Druyts, RMA
HUDEM4-26 Extraction of Landmine Signature
I van den Bosch, S.Lambot, M.Acheroy,
from Ground Penetrating Radar Signal UCL, Belgium-RMA
HUDEM4-27 NQR-based detection: an overview K.Althoefer, et al, Kings College,
London
HUDEM4-28 The PMAR lab in Humanitarian Demining M.Zoppi, R. Molfino, E. Cepolina,
Effort
PMAR Lab, University Genova, Italy
HUDEM4-30 Robosoft’s advanced robotic solutions for
outdoor risky intervention
HUDEM4-31 Behaviour based motion control for offroad
navigation
HUDEM4-32 Planning Walking Patterns for a Biped robot
using Fuzzy Logic Controller
HUDEM4-33 Adaptative Neuro-fuzzy Control of AMRU-5, a
six-legged walking robot
HUDEM4-34 Multi-Agent-System: an efficient approach for
Sensor and Robotics Systems used in
Humanitarian Demining
HUDEM4-35 4-D GPR Image Based on FDTD Parallel
Technique
HUDEM4-36 Robotic Agents for dangerous tasks. Features
and Performances
HUDEM4-37 How to design a Haptic Telepresence System for
the Disposal of Explosive
HUDEM4-38 Detection of material and structure of mines by
acoustic analysis of mechanical drilling noise
P.Pomiers,V.Dupourqué, Robosoft,
France
M.Proetzsch, T.Luksch, Prof K.Berns,
University of Kaizerslautern
Dr A.Pajaziti, Ahmet Shala, Bujar Pira,
Technical University of Prishtina,
Kosovo
Dr JC Habumuremyi, Y.Baudoin Royal
Military Academy, P.Kool, VUB
Belgium
E.Colon, Royal Military Academy
Baikunth Nath, jing Zhang, University
of Melbourne, Australia
Stezfan Havlik, Slovak Academy of
Science, Slovakia
OrdnancesB.Petzold, M.F.Zaeh,
A.Kron,G.Schmidt,B.Deml,B.Färber,
T.U. München – Ifa University of the
Armed Forces
G.Holl, B.Schwark-werwach, C.Becker

Past, present and future of the Anti-Personnel Mine Ban Convention
HUDEM’04 - RMA - 16 th June 2004
Speech by Ambassador Jean Lint, President of the 4 th Meeting of the States Parties
Since its entry-into-force on 1 st March 1999, the Anti-Personnel Mine Ban Convention,
crafted in Vienna, launched in Brussels, negotiated and adopted in Oslo and finally signed in
Ottawa on 3 rd December 1997 has been a true success story as almost three-quarter of the States
of the world have accepted the responsibility to never use, produce or transfer anti-personnel
mines and to cooperate in addressing the devastating impact of those mines.
This success is due to the widespread recognition of this international norm and to the
spirit of cooperation between all States Parties, the International Campaign to Ban Landmines
(ICBL) and the International Committee of the Red Cross (ICRC).
This success is also due to mechanisms foreseen in the Convention, such as
• the annual Meetings of States Parties and
• the obligation to provide international cooperation and assistance for mine
clearance, stockpile destruction and victim assistance.
It is also due to those mechanisms created by the States Parties to facilitate coordination
and exchange of information:
• the Intersessional Work Program, where four Standing Committees are active on:
• the General Status and Operation of the Convention,
• Mine Clearance, Mine Risk Education and Mine Action Technologies,
• Victim Assistance and Socio-Economic Reintegration and
• Stockpile Destruction.
The next Intersessional meeting will be held in Geneva from 21 to 25/6/04 where
all participants are welcome.
• the Coordinating Committee, responsible for the organization of meetings, and
• the Implementation Support Unit, taking care of the secretariat of the Convention.
Finally, it is due to those mechanisms that have emerged on an informal basis:
• a sponsorship program for mine-affected and developing countries;
• a Contact Group designed to promote universal adherence to the Convention
(initiative of Canada);
• a Contact Group on transparent reporting by States Parties (Belgium) and
• a Contact Group on the mobilization of resources for mine action (Norway).
Thanks to that informal exchange of information and the formal provision of information
through compulsory annual transparency reporting, we now have a clear view of our progress in
the pursuit of the Convention’s core humanitarian objectives and the challenges ahead.
With respect to universalization, the international norm established by the Convention is
now consolidated as 142 States have formally accepted the Convention, including almost every
country in Europe, the Americas and Sub-Saharan Africa. Additionally, 9 signatory States are
expected to ratify the Convention in the foreseeable future.

In Europe, all countries are parties to the Convention, with the exception of Finland, Latvia
and Poland. I hope that those States will be able to join us soon. We have received encouraging
signals in this respect, as Latvia and Poland submitted article 7 reports on a voluntary basis.
We are particularly concerned by those States remaining outside of the Convention, which
still use and/or produce anti-personnel mines as well as by those that have huge stocks of antipersonnel
Mines.
We urge them to stop using and producing and to destroy their stockpiles. We need to
increase our efforts to stress that no conceivable utility of anti-personnel mines could possibly
justify the devastating human costs of these weapons. In this context, I have called repeatedly not
only on all States but also on armed non-state actors to abide by the principles of the Convention
and comply with them.
In the field of mine clearance, we know that 50 States Parties suffer from the impact of
landmines and we are working effectively to know the extent of the problem, and establish and
support national mine action programs with a view to respecting each State’s 10-year deadline to
clear mines.
As for stockpile destruction, 117 States Parties have declared they do not possess
stockpiles any longer. Together they have destroyed more than 30 million landmines. What is
important is that even States Parties with few resources took full ownership over this obligation
and that some destroyed huge numbers of mines. In addition, by taking decisive action to destroy
these weapons, the States Parties have clearly demonstrated that their armed forces can continue
to fulfill their responsibilities without anti-personnel mines. 15 States Parties still have stockpiles
to destroy before the four-year deadline prescribed by the Convention.
In the field of victim assistance, up to 40 States Parties may require assistance to meet the
care, rehabilitation, and social and economic reintegration needs of landmine survivors. Even if
that responsibility rests which each affected State Party, we also know that the countries with the
greatest numbers of mine victims are amongst the poorest of the world. In addition, the
commitment to assist survivors is not expressed in a time limit in the Convention but in the
lifetime of the victims.
Today, we need to ensure that we sustain our efforts to truly achieve our aims of a antipersonnel
landmine-free world through a more sophisticated approach to humanitarian
demining.
The Nairobi Summit for a Mine-Free World, the name given to the Convention’s first
Review Conference, which will take place in Kenya from 29 November to 3 December 2004,
will constitute a crucial opportunity to review our achievements and our remaining challenges, as
well as envision the way ahead.
I could not conclude this speech without congratulating the Royal Military Academy for
the role it has been playing for more than five years in the field of Mine Action Technologies.
It has been a pleasure for me to work with the Academy and especially with Professor
Marc Acheroy. Together, that is with politicians, diplomats, militaries, NGO’s and victims,
experts in Mine Action Technologies have an important role to play in the noble cause of putting
an end to the human suffering and casualties caused by anti-personnel mines, which are all
indiscriminate, cruel and inhumane and which destroy the lives of thousands of innocent civilian
victims each year.

Mine action technologies: building a roadmap to bring
appropriate and improved technologies into operational use
Marc Acheroy, Royal Military Academy, Brussels, Belgium
The present document wishes to reflect the outcome of three expert hearings in mine
action technologies, chaired by the author and which took place on the margins of the
Standing Committee on Mine Clearance, Mine Risk Education and Mine Action
Technologies in February 2003, May 2003 and February 2004.
1. Introduction
In 1997, at the workshop, which accompanied the signing of the Ottawa Convention,
concern was expressed at the lack of international coordination and cooperation in mine
action technology. It was noted that there were no universal standards for technology,
no common view on where resources should be directed, and inadequate dialogue and
understanding existed both within the research and development community and also
with other actors in mine action.
Even if there is still a lack of international coordination and cooperation in mine action
technologies, especially between the end-users, the donors and the R&D communities, a
lot of work has been done and some success stories can be reported. Significant progress
has been made:
- in the performance of metal detectors and handheld dual sensors, which combine
metal detectors with ground penetrating radar (GPR),
- in the development and use of mechanical devices,
- in the development of applications based on information technologies (IMSMA is a
good example),
- in manufacturing personal protective equipment and prosthetic feet,
- in the training of rodents to detect landmines,
- in the suitability and cost of personal protective equipment.
Thanks to the International Test and Evaluation Programme (ITEP), much work has been
undertaken to test and evaluate equipments, systems and methods against agreed
standards (e.g. the CEN Workshop Agreement – CWA 14747:2003 "Humanitarian Mine
Action - Test and Evaluation - Metal Detectors", published by CEN in July 2003, and the
CEN Workshop Agreement for Test and Evaluation of Demining Machines (CWA12)
approved on April 20, 2004).
Nevertheless, efforts must continue, especially to initiate and increase the coordination
and the cooperation between users, donors and technologists in order to develop and
bring to the field equipment and tools based on real needs and not assumed needs.
2. Mine action technologies: a very difficult problem

Several factors slow down real progress in the development and fielding of new
technology, with the most significant of these factors related to the fact that mine action
solutions are not simplistic and that no “silver bullet” is available. It can be said that
finding all mines in the ground without a false alarm is a challenge comparable to
sending a person to the moon but with much less money. Some of the significant
challenging factors include:
- A lack of a procurement path makes fielding a technology very difficult.
Consequently, developers can face a dead-end when research and development as
well as prototyping and test and evaluation / validation (if any) are achieved!
- Mine action solutions are not universal but rather often country / region specific
(e.g., related to specific soil type, climate, vegetation, socio-cultural environment,
level of education, etc). A system approach needs to be used.
- Mine action technologies are diverse, e.g. ITEP recognizes six different categories:
survey, detection, mechanical assistance, manual tools, personal protection and
neutralisation.
- Requirements for technologies are not easily set, nor satisfied.
- Some major advances have not been well appreciated (e.g. the very significant
improvements in metal detectors, personnel protective equipment, information
technology support tools).
- It is now clear that the market for mine action equipment is not large enough to
support bringing products to market.
- Both donors and demining organizations are naturally conservative – especially
regarding safety.
- Donors are reluctant to insist on new and more efficient technologies and deminers
often do not change successful clearance methods (even if not efficient) as long as
donors accept the status quo.
- Some of the problems of new mine action technologies are not technical (e.g.
computer staff in field offices leaving once they are trained).
These challenges emphasize the responsibilities of each of the actors of the demining
Community: the donors, the end-users and the technologists.
a. Donors responsibilities
Clearly, donors have a key role to play, especially in supporting the fielding of
appropriate technologies in order to optimize the funding of technology in the
long-term (e.g., by supporting the introduction of new technologies on the
condition that they will lead to faster operations, saving lives, and saving
money).
b. End-users responsibilities
End-users need to have a pro-active role and to be understanding and open
regarding the process of bringing appropriate technologies in the field.

c. Technologists responsibilities
Technologists need to understand the importance of bringing appropriate
technologies to the field. They should visit the field to truly understand the real
needs of end-users, avoid building technologies based on assumed needs and
work interactively with end-users. They should understand that technologies
must be affordable, adaptable and manageable in the field, and must fit in
existing procedures (IMAS). They must be aware that ergonomic and
physiologic aspects are strongly influencing the efficiency of mine clearance
activities. Technologists should increase their understanding of the fact that, in
addition to technologies related to detection, technologies related to area
reduction (to know where the mines are not), strategic planning and programme
management are also important.
3. Need for a roadmap
The Convention states that “each State Party undertakes to facilitate and shall have
the right to participate in the fullest possible exchange of equipment, material and
scientific and technological information concerning the implementation of (the)
Convention.” This implies that such an exchange is an important underpinning to
assisting States Parties in the fulfilment of their obligations. It is in the spirit of this
provision of Convention that all actors are urged to apply the recommendations in this
document. Donors need to understand that technologists need their support to
establish a sound procurement process for fielding appropriate technologies in order
to have a more cost-effective mine action. For their part, end-users need to be proactive
and be understanding and open to the process of introducing new technologies
in the field and to make use of existing tools. End-users need to understand that
appropriate technologies could save human lives and increase mine action efficiency.
Finally, technologists must accept that nothing is more important than understanding
the working environment.
The co-chairs of the ad hoc Standing Committee decided to mandate an informal
expert group, meeting on the margins of the Standing Committee and including endusers,
donors and technologists,
- To define a coherent roadmap to field as soon as possible effective mine
action technologies, taking into account real needs of end-users, priorities of
donors and mine-affected countries as well as the state of maturity of
technologies.
- To identify means to establish a sound procurement process for fielding
appropriate technologies in order to have a more cost-effective mine action.
- To investigate means to encourage and organise a close dialogue between
mine action actors
4. A possible roadmap

It is a matter of urgency to define a roadmap to secure rapid fielding of appropriate
and improved mine action technologies. This roadmap can only be defined by
achieving the following sub-objectives:
- to establish a list of prioritized operational needs (SON), based on the political
priorities and constraints;
- to establish a portfolio of priority technological projects, based on an inventory
and an analysis of existing and potential technologies as well as on the
determination of their gaps regarding real operational needs;
- to establish technology action plans in conjunction with end-users;
- to facilitate procurement processes. Thereby getting appropriate and improved
mine action technologies into operational use;
- to identify means to secure maximum support to enhance technologies, from
donors, understanding benefits of appropriate and improved technologies, from
end-users, in terms of understanding and openness, and technologists in terms of
understanding end-users’ real needs.
These main topics of this roadmap and their interconnections are sketched in Fig. 1
and described below.
Securing
maximum
support
from
Donors ,
End - Users
and
Technologists
Defining &Prioritizing Operational Needs SON
Analysing technologies &
Classifying technology needs
Technology action plans
Portfolio
Appropriate technologies
into operational use
GICHD
ITEP
Lessons
Learned
JMU
Fig. 1: A roadmap for getting appropriate and improved technologies in operational use

a. Statement of operational needs
It is expected that organizations (such as GICHD, UN and NGO’s) who are close
to end-users, continue to facilitate in defining operational needs (SON). They
should, when appropriate increase the quality of the knowledge regarding endusers’
real needs and enhance the transfer of this knowledge towards
technological organizations such as the International Test & Evaluation
Programme (ITEP).
b. Portfolio of prioritized projects: end-users’ involvement
It is expected that Organizations (such as the International Test & Evaluation
Programme (ITEP)), who are close to technologists, continue to facilitate testing
and assessing mine action technologies. They should increase the involvement
of end-users in the test & evaluation process in order to improve the confidence
of end-users in appropriate and improved mine action technologies, include cost
/ effectiveness considerations in their test and evaluation reports and share this
information with end-users, and welcome direct requests issued by end-users
and donors.
c. Portfolio of prioritized projects: importance of a cost / benefit analysis
Equipment testing should be subject to an estimated cost-benefit analysis. To
take into consideration:
- a description of the costs of the technology (including ALL the extras
like training, down-time, transport costs etc).
- a description of the potential scenarios where the technology will be
useful, including financial analysis of how much is being spent currently
in these regions on mine clearance
- a realistic delivery date for the first small batch for testing – and a
contractually agreed price.
- a detailed cost-effectiveness analysis, which shows comparison with
existing methods when appropriate
d. Technology action plans to bring appropriate technology into operational
use
Donors should consider increased support for appropriate or improved
technologies. Possible provisions could include:
- Contractual obligation for donor-supported, mine-clearance to make
available at a fair price the facilities, staff, access, etc for large-scale field
trials of appropriate or improved tools and equipment.
- Emphasis on contracts based on IMAS using the best available
technology for clearance, by insisting that demining organisations

demonstrate an objectively measurable steady improvement in efficiency,
not simply retaining existing methods.
- Funding to secure end-users participation to meetings and trials.
Donors must be willing to invest in emerging and improved technologies now,
for the best results in the near future!
The informal expert group has adopted this roadmap. To support its objectives, it is
mandatory to have a close collaboration of end-users, donors and technologists, and a
good exploitation of lessons learned by making use of the structure which exists in
JMU / MAIC.
The roadmap process must be a living interactive process, as statements of
operational needs (SON) evolve and depend on variable mechanisms and scenarios.
End-users must be involved in the whole process, from the definition of SON ’s to the
testing of equipments into operational conditions. A cost / benefit analysis is a keyelement
in the process of bringing appropriate and improved technologies into
operational use.
5. Building a test case
As, already a long time ago, technologists are promising to bring appropriate and
improved technologies into operational use and have created a structure to test and
evaluate mine action technologies (ITEP), end-users strongly suggest proving by a
“test-case” that the proposed roadmap is working and could be an effective process.
Further, it is felt that results need to be presented in NAIROBI at the end of 2004,
during the meeting of State Parties. Otherwise, the credibility of technologists will be
dramatically affected and end-users as well as donors will definitely fail to believe in
the possibility of bringing new, appropriate or improved mine action technologies
into operational use. Hence, it is particularly important to demonstrate the
effectiveness of the proposed roadmap by processing “test-cases” through it.
Therefore, dual sensors (metal detector – GPR) have been proposed as test-case: the
HSTAMIDS system from US and the system of ERA Technologies (UK) – VALLON
(GE). The objectives to achieve include:
- To present at the Nairobi meeting (beginning of December 2004) the results of
the process of bringing, according to the proposed roadmap, a dual sensor
(metal detector + ground penetrating radar (GPR)) into operational use.
- To involve GICHD in the definition of the Statements of Operational Needs
corresponding to the region selected for the operational tests.
- To involve ITEP in performing the technical testing and the comparison of
test & evaluation results against end-users’ requirements.
- To involve end-users in all the steps of the process and especially in the
organisation of operational tests in Bosnia.

End-users (Norwegian People Aid – NPA) will select a region in Bosnia (to be
confirmed) and GICHD will define the SON for this region in close collaboration with
end-users.
The process described in the roadmap will include:
- the establishment of the SON for the region of interest by GICHD in close
collaboration with end-users;
- the technical testing and evaluation as well as the comparison against SON by
ITEP (dual sensors must be included in ITEP work plan – Portfolio of
technology projects) with the participation of end-users;
- the operational testing by end-users in the selected region of interest.
6. Conclusion
Even if a lot of progresses have been booked in technological development,
especially in close-in detection, mechanical clearance and in demining campaign
management, there is still a lot of work to be done in order to convince donors and
end-users of the possible increase of effectiveness, thanks to mine action technology.
It is just the aim of the proposed roadmap to give a coherent frame, accepted by all
actors, for bringing appropriate technologies into the field.
Acknowledgement
The author wishes to thank all the participants to the informal group of experts on mine
action technologies, meeting on margins of the Standing Committee on Mine Clearance,
Mine Risk Education and Mine Action Technologies.
References
1. http://www.gichd.ch
2. http://www.itep.ws
3. http://maic.jmu.edu
4. M. Acheroy, "Mine Action Technologies: Problems and Recommendations" Journal
of Mine Action 7.3, Dec 2003.
5. M. Acheroy and I. van den Bosch, “Humanitarian Demining: sensor technology status
and signal processing aspects”, CNRS – GDR ondes, Dec 2003 Marseille, France,
(invited paper).
6. M. Acheroy, "Humanitarian demining : sensor technolology status and signal
processing aspects" , IEEE-ICRA conference 2003 (invited paper).
7. M. Acheroy, A. Sieber, "Defining in Europe a possible methodology to field faster
improved Mine Action Technologies (MAT) " Dubrovnik workshop on humanitarian
demining, Oct 2002.

The Outdoor Robotics: challenges and requirements,
promising tools and dual-use applications
Daniele Caltabiano, Domenico Longo, Giovanni Muscato
Michele Prestifilippo, Giacomo Spampinato
DIEES Università degli Studi di Catania
viale A. Doria 6, 95125 CATANIA ITALY
e-mail: gmuscato@diees.unict.it
Abstract
In the last years many new robots have been designed and built specifically for outdoor
applications. Within the Workpackage 3 of the CLAWAR Thematic Network several results of
outdoor robotic projects have been collected and analysed. In this paper a short overview of some
innovative solutions, innovative locomotion strategies and their requirements will be exposed.
Finally some projects developed at DIEES University of Catania are presented with particular
reference to the possible applications to humanitarian demining.
1. Introduction
The main aim of CLAWAR WP3 is to formulate the requirements for CLAWAR machines in
specific sectors that have already been identified as having good potential for automation [1].
Two tasks are carried out per year focusing on applications sectors that are felt to offer good
potential for exploiting the CLAWAR technology. In specifying the requirements, the concepts of
modularity and modular components that have been formulated are. In Year 2 (2003/04) one of the
tasks was concerning Natural/Outdoor applications. The activities of the task consisted in analysing
some application sectors in dangerous environments, in evaluating the capabilities of innovative
locomotion systems and in the study of methods for autonomous navigation [2].
In the following sections some dangerous applications of outdoor robots will be exposed, and then
several locomotion architectures will be discussed. Finally some projects for outdoor robotics
developed at DIEES will be briefly shown.
2. Robots for Dangerous Environment
The following sectors have been analysed in some detail and are briefly reported:
• Fire-fighting
• Exploration of volcanoes
• Demining
• Quarrying and mining
21 Fire fighting
Mobile robots are becoming an option of choice where there is a need to remove people from
danger. An example of the new commercial thrust has been to develop from existing military
technology a range of remotely operated robotic fire fighting vehicles for use in a variety of
commercial operations. The financial and human cost of fire fighting is substantial. Robotic systems
have the potential to reduce this figure quite dramatically. For natural disasters such as forest fires

and 9/11 type disasters, it may not be possible to solve these issues but there will be opportunities
where levels of assistance may be appropriate.
As a risk reducing measure, robotic machines are a valuable addition to any fire-fighting team.
Robotic fire fighters are well equipped to handle extreme temperatures and hazardous atmospheres.
As well as the ability to contain the blaze, the machines are fitted with a low cost thermal imager
and other onboard sensor options to allow them to investigate operational risk and provide statistical
monitoring by assessment.
Two of the most prominent vehicles are 'Firespy' and 'Carlos' developed by QinetiQ. As might be
expected, given the company's history, the systems developed by QinetiQ are not aimed at the fire
service's bread-and-butter deployments but are intended for use in emergency situations. The robots
have been designed for robustness and reliability in a variety of hazardous situations such as largescale
petroleum fires; aircraft engine fires; hazardous chemical leaks; and fires with the risk of
building collapse. QinetiQ's two fire-fighting robots can therefore cope with a range of applications.
A detailed analysis of the requirements of a fire fighting robot has been carried out by Ulrich
Schmucker of the Fraunhofer IFF, (Magdeburg, Germany) and is summarised in the scheme
reported in Fig. 1. [3].
Fig. 1. Fire fighting robot requirements.
22 Exploration of volcanoes
Today there are 1500 volcanoes on the Earth potentially active, 500 of them have been active
during last 100 years and about 70 are presently erupting. Ten percent of the world population live
in areas directly threatened by volcanoes, without considering the effects of eruptions on climate or
air-traffic for example. About 30.000 people have died from volcanic eruptions in the past 50 years,
and billions of euros of damage have been incurred. As a consequence it is important to study
volcanoes and develop the technologies to support the volcanologists in this process. In the last
decade alone, due to both the unpredictable timing and to the magnitude of volcanic phenomena,
several volcanologists have died surveying eruptions. So a major aim of a robotic system is to
minimise the risk for volcanologists and technicians involved in working close to volcanic vents
during eruptive phenomena. It should be noted that observations of, and measurement of the
variables relating to volcanic activity are of greatest interest during paroxismal phases of eruptions,
which unfortunately are also the time of greatest risk for humans. Among the projects for volcano
exploration we can mention the Dante II walking robot developed by Carnegie Mellon University
and NASA, the RMAX helicopter developed by YAMAHA and the recently EC funded
ROBOVOLC project [4].
2.3 Demining
There exists a big difference between the military and the civilian mine clearance. The military
demining operations accept low rates of Clearance Efficiency (CE). For these purposes it is often

sufficient to punch a path through a minefield. For the humanitarian demining purposes, on the
contrary, a high CE is required (a CE of 99.6% is required by UN). This can only be achieved
through a ‘keen carding of the terrain and an accurate scanning of the infested areas’ that implies
the use of sensitive sensors and their slow systematic displacements, according to well-defined
procedures or drill rules, on the minefields. The robots, carrying the mine-detectors, could play here
an important role. Obviously, the automatization of an application such as the detection, the
marking and removal of AntiPersonnel mines implies the use of autonomous or teleoperated mobile
robots following a predefined path, sending the recorded data to their expert-system (in charge of
processing the collected data), marking the ground when a mine is detected with a probability of
predefined level and/or, possibly removing the detected mine. This complete automatization is
actually unrealistic and may surely not be entrusted to a unique robot: the technologies allowing it
exist, but the integration of those technologies in mobile Robotic Systems moving in unpredictable
outdoor environmental conditions is not yet mature.
Prof. Y. Baudoin of the RMA Belgium has proposed an Humanitarian Demining catalogue intended
to (1) disseminate the results of research-activities related to the development of mobile roboticized
carriers of detection sensors with the support of the GICHD (Geneva International Centre for
Humanitarian Demining) , (2) assist actual and future T&E activities of ITEP (International Test
and Evaluation Programme), (3) help the research-centres focusing on this kind of solutions to
refine and continuously adapt and update generic modules that may be transferred to useful
Robotics Systems. This catalogue includes existing robotics systems but also current projects
focusing on the possible use of roboticised solutions.
A detailed classification of the requirements for a robot for humanitarian demining has also been
prepared subdivided into three levels:
- the system-level requirements
- the robot-level requirements
- the mobility-level requirements
2.4 Quarrying and mining
Quarries are an important economic source for many countries. Working in quarries is a good job
opportunity, but is very hard and dangerous. In fact, mineworkers work at open-air, subject to all
weather conditions, in a dangerous environment. Quarries change aspect every day because as
mining proceeds the ground and walls change shape. The infrastructures such as energy supply
cables and access paths result temporary and unreliable. The problem of safety in quarries is urgent
to be solved; especially for poorer countries where protection for workers is not foreseen as life
value is low, although regulation to prevent accidents exist, but for richer countries as well,
because, although protections are higher, accidents happen every day because satisfactory
technology, able to guarantee mineworker safety, doesn’t exist. Some recent project on quarrying
and mining are the ROBOCLIMBER [5], the Microdrainage, and the applications developed by
Caterpillar.
3. Next stage type locomotion system
Innovative robots for outdoor robotic applications require new locomotion systems. The analysis
performed within WP3 was concentrated on the following typologies:
• Hybrid locomotion
• Peristaltic locomotion
• Self-reconfigurable modular systems
• Aerial systems
In the following sections a brief overview of each architecture will be discussed.

3.1 Hybrid Locomotion
The design of a new mobile robot is process that entails satisfying several requirements with
respect to the environment in which the robot will be asked to work. Most classical solutions adopt
wheels as a locomotion technique. It is, however, clear that wheeled robots have some limitations in
highly unstructured environments or in difficult structured environments (like stairs). On the other
hand the current technology for walking robots is not yet so advanced as to represent a valid
alternative. Most current walking robots are very slow, have a low payload capability, are difficult
to control and, due to the high number of sensors and actuators, have low fault tolerance
capabilities.
A way to find a compromise between these two solutions is to choose a hybrid system adopting
wheels and legs together. Several examples of hybrid robots exist in the literature and belong to two
main categories.
The first category includes articulated-wheeled robots, with the wheels mounted at the end of legs.
This category comprises the WorkPartner robot [6],[7],[8] with its particular kind of locomotion
called rolking, the Roller Walker [9], where the wheels are not actuated but the robot simply skates,
the biped type leg-wheeled robot developed by Dr. Matsumoto of AIST [10][11] able to climb up
stairs, the Walk'n Roll with four legs with a wheel attached at the end of each leg [12] and the minirover
prototype developed by LRP [13], just to name a few examples. Most of these systems can
have three type of locomotion: wheel mode, where they act as a conventional wheeled mobile robot;
step mode, where the wheels are locked and they act as a simple legged system and hybrid mode
with cooperation of legs and wheels.
The second category consists of robots with wheels and legs separated, but acting always together
to locomote the system. This category includes for example the Chariot II robot [14], [15] that has
four articulated legs and two central wheels, the RoboTrac robot [16], [17] with two front legs and
two rear articulated wheels, the hybrid wheelchair proposed by Krovi and Kumar [18] and the
ALDURO, [19], [20]. This category of robot has usually only the hybrid type of locomotion but are
from a mechanical point of view more simple and consequently with a more light weight.
Legged locomotion is very effective in extreme ground conditions, but bad with high speed.
Wheeled locomotion is good in high speed but bad on natural terrain. Helsinki University of
Technology introduced a way to combine legged and wheeled locomotion to gain effective natural
terrain mobility. The mode of locomotion is called rolking (rolling-walking). In addition to effective
mobility rolking mode makes it also possible to measure the shapes and unevenness of the ground
only by touching it with the wheel. WorkPartner, as illustrated in Fig. 2, is a futuristic service robot
designed to be used mainly in urban outdoor environment. Mobility is based on a hybrid
locomotion system, which combines benefits of both legged and wheeled locomotion to provide
good terrain negotiating capability and large velocity range at the same time.

Fig.2 Workpartner robot (HUT).
The Wheeleg robot (Fig. 3) was designed and built by the University of Catania in order to
investigate the capabilities of hybrid wheeled/legged structures on rough terrain [21-27]. Possible
applications that have been envisaged are humanitarian demining and exploration of unstructured
environments like volcanoes, etc.
Fig.3. The Wheeleg robot
3.2 Peristaltic locomotion
Peristalsis has always been deeply studied because of its diffusion in nature. This mechanism is
often used when a fluid has to be pushed in contact with a natural surface, such as blood in veins
and other biological fluids, and some low-order animals use it for their locomotion.
All these animals are invertebrates, but there are differences between them according to different
anatomies and behaviours. The most studied from the point of view of locomotion are snails and
earthworms. Main applications of these robots are for diagnostic and surgery purpose, but several
prototypes have been designed and built also for disaster-relief.

3.3 Self-reconfigurable modular systems
In the last years, many efforts have been made in research in modular robotics and especially in
self-reconfigurable systems (SRS), i.e. systems which are able to change dynamically their
topology. Such systems are made of sets of mechatronic building blocks which have the capacity to
connect and disconnect together. By manipulating themselves their own building blocks, those
systems can change their topology. The advantages of modular systems are several:
• Easy and fast to deploy: the robot does not need to be designed and constructed; the modules
are already available and can be fast and easily manually assembled. The required topology
for a task is automatically generated by software and the control algorithm is downloaded to
the robot.
• Easy and fast maintain: a deficient module can be easy and fast replaced.
• High versatility and little stuff: a set of modules can be assembled in various topologies
optimised for various tasks.
• Adaptability: it is possible to perform tasks which were not originally foreseen.
• Low cost: the genericity of the modules authorizes mass-production
Furthermore, a self-reconfigurable system has even more advantages:
• Autonomy: the robot can reconfigure and repair itself
• Increased versatility: the robot can adapt its topology during a task, and not only before.
The applications of SRS are numerous: i.e. industrial manipulations, locomotion on rough and
difficult terrain, pipe inspection, space station construction, etc.
Because of their versatility and robustness self-reconfigurable systems are very interesting for space
applications. A set of robotic modules could do a very wide range of tasks, which would else need
numerous big and heavy non-modular robots, including tasks which are not originally foreseen.
Moreover a low gravity context reduces tremendously the mechanical constraints and permits much
more efficient SRS. The system can move more economically, carry heavier loads and form bigger
structures. All candidate sites in space, like orbital stations, asteroids, moons, and non-gaseous
planets, have less gravity than on Earth.
The mechatronic design of self-reconfigurable robotic modules is quite complex because it involves
a high degree of technology for energy storage or production, efficient and low energy consumption
of the connexion mechanism, distributed control system, etc.
The optimisation of module characteristics (like connectivity, mobility, geometry) which gives the
best versatility is a non-trivial problem because they are only components of a modular systems
which should be able to form an undefined number of different topologies, depending of the task
and the context, like the local shape of the terrain. Therefore the modules design must be optimised
for some main and general functions which must be defined first.
3.4 Aerial systems
Over the recent years the use of low cost Uninhibited Aerial Vehicle (UAV) for civilian
applications has evolved from imagination to actual implementation. Systems have been designed
for fire monitoring, search and rescue, agriculture and mining. In order to become successful the
cost of these systems has to be affordable to the civilian market, and although the cost/benefit ratio
is still high, there have been significant strides in reducing this, mainly in the form of platform and
sensor cost. Many new research projects are currently developed with the aim to build totally
autonomous systems, groups of cooperating UAVs and to study the cooperation between ground
vehicles and UAVs.

4. Overview of some prototype of robots for outdoor applications developed at DIEES
4.1. The Wheeleg Robot
The Wheeleg robot is a hybrid system with two pneumatically actuated front legs, each one with
three degrees of freedom, and two rear wheels independently actuated by using two distinct DC
motors. The main idea was to use rear wheels to carry most of the weight of the robot and front legs
to improve the grip on the surface and to climb over obstacles [27].
From an historical point of view there are several examples of systems with both wheels and legs.
What can be observed from all these systems is that the legs have a better adhesion with the soil and
perform most of the work traction, while the wheels support most of the weight of the structure.
The robot's dimensions are W ×L × H = 66cm × 111cm × 40cm and its total weight is 25kg.
This robot has been experimentally tested both in laboratory and in outdoor environments, including
very unstructured volcanic terrains (Fig. 4).
As a general consideration it should be remarked that the proposed robot was built as a research
prototype and is not suitable for work in real outdoor applications. Several modifications in the
mechanical structure and in the electrical connections are required to increase the reliability of the
system in real applications. In particular the kinematics of the legs is very simple and low-cost, but
it is not reliable enough for potential applications. Spherical joints should be inserted in the feet to
allow the robot to better adapt itself to different surfaces.
It is the authors' opinion that this hybrid robot does not include simply both advantages of wheeled
and legged robots, but also some of their drawbacks. For example some of the slow speed and
stability problems of legged systems are also present in the Wheeleg. Moreover some of the traction
problems typical of wheeled robots have been also experimented. However in many situations, like
in the case of surfaces with rocks, the help of the legs has demonstrated to greatly improve the
locomotion capabilities of the robot, without excessively reducing its speed. Such a situation is
probably common also to other hybrid architecture and is mainly due to the actual limitation in the
technology in designing and building more efficient mechanical legs.
The design of the control architecture was an important aspect, since a strong interaction between
the many different sensors and actuators was needed. The solution developed has shown to be
suitable for the intended purposes and has not revealed any particular problem.
In any case the experience gained from this robot was considered very useful and will be adopted
in the design of next generation of hybrid and legged systems.
Figure 4. The Wheeleg on Etna volcano

4.2. The ROBOVOLC system and the small scale prototypes
The main objective of this EC funded project was the development and demonstration of an
automatic robotic system to perform measurements in a volcanic environment. Partners of the
projects were University of Catania, University of Leeds, INGV, IPGP, ROBOSOFT and
BAESYSTEMS.
The selection of the most appropriate locomotion technique for the ROBOVOLC system was
carried out by evaluating the most promising techniques by comparing certain criteria e.g.
reliability, rough terrain performance, suitability for purpose, etc.
The selection was also based on the trial of the M6 and of the P6W platforms. The M6, shown in
Fig. 5, has an articulated chassis capable of adapting to very rough terrain [3]. However not having
a substantial central body, the payload of this system was low. Some improvement was obtained
with the use of the P6W system, a small scale prototype with a reduced mobility, but with the
capability to carry a more substantial payload (Fig.6). In particular this prototype was used to test
traction control strategies in laboratory. Following these testing, also performed on volcanic sites,
the final locomotion concept was a six wheeled rover with an adaptive chassis, with a large range
semi-active suspension and differential steering.
The final ROBOVOLC system is shown in Fig. 7 and 8. The system is composed by a
Communication system, a Power supply system, a Control system, a Navigation system, a User
interface and a Science Package.
Communication system: The telecommunications system is based on a redundancy system. This
used a standard high-power wireless LAN interface, for most of the data exchange at a high data
rate (10Mbps) and a low speed radio modem (100kbps), for essential tele-operation e.g. when the
robot is out of the line of sight of the base station. Other radio links comprises two analog video
channels and a differential correction channel for the GPS.
Fig 5. The M6 robot Fig 6. The P6W prototype platform.
Power supply: The chosen power supply solution employed lead acid batteries, however to extend
mission duration a hybrid Internal Combustion (IC) engine has been designed and manufactured.
The optimum mode of operation uses the IC generator in combination with the batteries, as a back
up and recharges system. This arrangement greatly extends mission duration and enhances the
capability of the system.
Control: The infrastructure design is based on the utilisation of general purpose embedded
computer boards, with a combination of COTS and bespoke software running on these platforms.
The various tasks are distributed over three PC104 based computers located in the rover to execute
the main tasks of manipulator control, locomotion and communication control and sensors control.

The low level motion controllers selected were RoboSoft RMPC-555. These motor controllers are
to be used for both Rover and Manipulator motion and are interfaced to the PCs through a CAN
network.
Navigation: This covers localisation and path planning by sensing by the use of multisensor data
fusion algorithms and map data. For most operations the decisions are made by a human operator
(teleoperation), but automatic obstacle avoidance has been incorporated.
User Interface: A graphical man-machine interface was designed to allow the volcanologists to
operate the robot via two joysticks and a touch screen during the missions without specialist
knowledge of the system. Specific functions have been developed to allow teach and repeat
operations, intervene preset missions to explore manually and to set additional waypoints.
Science tasks: The Science package comprises three main subsystems: the Manipulator, the Pan-tilt
turret and the gas sampling system. A 5 degree of freedom SCARA manipulator has been
specifically designed for the Robovolc system to allow the collection of samples of rocks, to place
and pick up instruments at specific locations and to collect gas samples in the proximity of
fumaroles. A three-finger gripper, has been designed with force sensors to pick up rocks with a
maximum diameter of 15 cm. The Pan/tilt turret has the following sensors: a digital video-camera
recorder, a high resolution still-image camera, an infrared camera and a Doppler radar for gas speed
measurement. The user can orient the turret and can tele-operate all the installed sensors. Gas
sampling is a specific and detailed operation needed for the Robovolc project. The main
problems for the gas sampling are: gas temperatures can rise above 600°C ; extremely corrosive
acid components are present; and contamination of the gases with the surrounding atmosphere must
be avoided at all cost. To perform such precise gas sampling a new system has been designed,
comprising of a probe for sampling the gas, a gas collection system and specific gas collection
bottles.
Several lab test and indoor and outdoor were carried out to assess the functionality of the final
system. In addition three on-site test campaign on Etna volcano were performed in September 2002,
June 2003 and August 2003, to demonstrate the functionalities of locomotion, rock and gas
sampling, terrain reconstruction, tele-operation capabilities, operational life, telemetry. In Figures 7
and 8 some pictures concerning some test campaign are reported. In the project web site more
details concerning the final tests and results are reported [4].
Fig.7. The final ROBOVOLC system tested inside the ETNA crater formed during the January 2003 eruption.

4. Conclusions
Fig.8. ROBOVOLC system collecting a lava rock sample and tested on a rocky terrain.
In this paper some of the activities carried out within the WP3 of the CLAWAR network. have been
briefly shown. In particular it resulted that many research projects are carried out within Europe to
find outdoor robotic solutions to solve different kind of problems. In many case the obtained
prototypes are valid solutions that could be efficiently also tested for humanitarian demining. It is
the authors’ opinion that even if up to now most of the robots do not appear as reliable and useful as
they should, by continuing the common research effort and by testing new solutions the distance
between the research laboratories and the real adoption of these systems in the minefields will be
rapidly reduced.
5. Acknowledgement
This work has been prepared using some part of the contributions received by many CLAWAR
members. In particular we thank H. Warren and N. Heyes (QinetiQ), Y. Baudoin (RMA), R.
Molfino (University of Genova), G. Pezzuto (Dappolonia), A. Haalme and S. Ylonen (Helsinki
University of Technology), P. Bidaud (LRP) and B. Bell (BAESYSTEMS). Their help is gratefully
recognized and acknowledged.
6. References
[1] The CLAWAR Network, http://www.clawar.net.
[2] G. Muscato, “CLAWAR 2 WP3 Application sectors Year 2 Report”, GIRT-CT-2002-05080, May 2004.
[3] U. Schmucker, Fire fighting robots requirements, notes from CLAWAR meeting Catania 09/2003 and Karlsruhe
11/2003. See also [1].
[4] The Robovolc project homepage http://www.robovolc.diees.unict.it
[5] The Roboclimber web pages http://www.t4tech.com/roboclimber
[6] A. Halme, I. Leppanen and S. Salmi , "Development of Workpartner Robot - Design of actuating and motion control
system", CLAWAR 99, Portsmouth 1999.
[7] A. Halme, I, Leppanen, S.Salmi, S, Ylonen, "Hybrid locomotion of a wheel-legged machine", CLAWAR 2000
International Conference on Climbing and Walking Robots, Madrid, Spain , 2-4 October 2000.
[8] I. Leppanen, S. Salmi and A. Halme, "Workpartner - HUT Automations new hybrid walking machine", CLAWAR
98, Brussels 1998.
[9] G. Endo and S. Hirose, "Study on Roller-Walker (Multi-mode steering Control and Self-contained Locomotion),
ICRA 2000, International Conference on Robotics & Automation, San Francisco, April 2000.
[10] O. Matsumoto, S. Kajita, M. Saigo and K. Tani, "Dynamic Trajectory Control of Passing Over Stairs by a Biped
Leg-Wheeled Robot with Nominal Reference of Static Gait", Proc. of the 11th IEEE/RSJ International Conference on
Intelligent Robot and Systems, pp. 406-412, 1998.
[11] O.Matsumoto, S.Kajita, M.Saigo and K.Tani, "Biped-type leg-wheeled robot", Advanced Robotics, Vol.13, No.3,
pp.235-236, 1999.
[12] H. Adachi, N. Koyachi, T. Arai, A. Shimuzu, Y. Nogami, "Mechanism and Control of a Leg-Wheel Hybrid Mobile
Robot", Proc. of the IEEE/RSJ International Conference on Robotics and Automation, pp. 1792-1797, 1999.

[13] F. Benamar, P. Bidaud, F. Plumet and G. Andrade, "A high mobility redundantly actuated mini-rover for self
adaptation to terrain characteristics", CLAWAR 2000 International Conference on Climbing and Walking Robots,
Madrid, Spain , 2-4 October 2000.
[14] Y.J. Dai, E. Nakano, T. Takahashi, H. Ookubo, “Motion Control of Leg-Wheel Robot for an Unexplored Outdoor
Environment”, Proceedings of the 1996 IROS Intelligent Robots and systems, Nov 4-8, 1996.
[15] E. Nakano, S. Nagasaka, “Leg-Wheel Robot: A Futuristic Mobile Platform for Forestry Industry”, Proceedings of
the 1993 IEEE/Tsukuba International Workshop on Advanced Robotics”, Tsukuba, Japan, Nov. 8-9, 1993.
[16] K. Six., A. Kecskeméthy, "Steering Properties of a Combined Wheeled and Legged Striding Excavator", In:
Proceedings of the Tenth World, Congress on the Theory of Machines and Mechanisms, Oulu, Finland, June 20-24, vol.
1, pp. 135-140. IFToMM (1999)
[17] C. Schuster, T. Krupp, A. Kecskemèthy, "A non-holonomic tyre model for non-adhesive traction and braking
manouvres at low vehicle speed", Proceedings of the NATO Advanced Study Institute on Computational Methods in
Mechanisms, Varna, Bulgaria, June 1997.
[18] V Krovi, V. Kumar, "Modeling and Control of a Hybrid Locomotion System, " In ASME Journal of Mechanical
Design, Vol. 121, No. 3, pp. 448-455, September 1999.
[19] M. Hiller, J. Müller, U. Roll, M. Schneider, D. Schröter, M. Torlo, D. Ward ,"Design and realization of the
anthropomorphically legged and wheeled Duisburg robot ALDURO." Proc. of the 10th World Congr. of theory of
Machines and Mechanisms. IFToMM, Oulu, Finnland, 1999.
[20] J. Müller, M. Hiller, “Design of an energy optimal hydraulic concept for the large-scale combined legged and
wheeled vehicle ALDURO”, Proceedings of the 10th World Congress on the Theory of Machines and Mechanisms,
Oulu, Finland, June 1999.
[21] S.Baglio, G. Muscato, G. Nunnari, N. Savalli, “Robots and sensors for HUDEM at the University of Catania”,
European Journal of Mechanical and Environmental Engineering, Vol. 44, N.2,pp.81-84, 1999.
[22] S. Guccione, G. Muscato, "Experimental outdoor and laboratory tests with the hybrid robot WHEELEG",
Proceedings of the 3rd International Conference on Field and Service Robotics, Helsinki (Finland) June 11-13, 2001
[23] S. Guccione, G. Muscato,"Test of the hybrid robot WHEELEG on a Volcanic Environment", Video Proceedings of
the 2001 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, Como, Italy July 2001, p.3.
[24] G. Lami, S. Guccione and G. Muscato, "Control strategies and architectures for the robot WHEELEG", CLAWAR
2000 International Conference on Climbing and Walking Robots, Madrid, Spain , 2-4 October 2000.
[25] G. Muscato and G. Nunnari, “Leg or Wheels ? WHEELEG a hybrid solution”, CLAWAR 99 International
conference on climbing and Walking Robots, Portsmouth, U.K. , 14-15 September 1999.
[26] M. Lacagnina, G. Muscato, R. Sinatra, "Kinematics of a hybrid robot Wheeleg", 5th IEEE International
Conference on Intelligent Engineering Systems 2001, Helsinki (Finland), 16-18 Sep 2001.
[27] S. Guccione, G. Muscato, "Control Strategies Computing architectures and Experimental Results of the Hybrid
Robot Wheeleg", IEEE Robotics and Automation Magazine, (IEEE Piscataway, U.S.A.), Vol.10, N.4, pp.33-43,
December 2003.

“Robotic System for Unstructured Outdoor Environments”
V.Gradetsky*, M.Knyazkov*, L.Meshman**, M.Rachkov***
*The Institute for Problems in Mechanics Russian Academy of Science
Tel.: (7095) 434-41-49 ; Fax: (7095) 938-20-48 ; e-mail: gradet@ipmnet.ru
**Federal State Establishment All-Russian Research Institute for Fire Protection
Ministry of the Russian Federation for Civil Defence, Emergencies and Elimination of
Consequences of natural Disasters
(FGU VNIIPO EMERCOM OF RUSSIA)
12, VNIIPO, Balashikha, Moscow Region, 143903
Fax: (095) 529-82-52
***The Moscow Industrial University
Tel.: (095) 277-28-39 ; Fax: (095) 274-63-92 ; e-mail: rachk-v@mail.msiu.ru
Abstract
The paper presents some results of a study the mobile robot’s motion in emergency
conditions to produce such operations as fire-fighting preventions, search of the mines
and demining or rescue operations after nature catastrophes and explosion results.
The mobile robot’s motion in dangerous environments is distinguished by special
design of mechanical and sensory control systems, providing the important qualities of
actions including the reliability of working operations, high level of maneuverability,
necessary capacity for cross-surface travel, possibilities for autonomous motions over
some parts of the route, availability for supervision mood of operations using manmachine
interface, cooperative control between transport system and onboard tools, the
fast exchange of measuring information and decision making ability in some arising
nonpredicted situations of environments.
The survey is delivered to emphasize some key problems of mobile robot’s motion
in unstructured environments.
1. Introduction
Operation fulfillment is very attractive for robot application when it is necessary to
work in dangerous environments instead of people. Such environments may be
characterized by high temperature, exploisive conditions, high level of radiation,
possibilities of mechanical trauma accidents. To move in such conditions, robots have
to be equipped with special protection material, and transport, sensors, detectors,
control systems, that guarantee an achievement of definitive task.
Problem of demining automation was considered in [1], where suggested that robots
may produce predicted job. Navigation and inspection methods [2, 3, 5] were evolved

to solve demining problems. Mine detection equipment developed in many countries,
for example [4, 6].
Robots for fire-fighting operations were developed up to day by some organizations
[7]. It was necessary to determine the mobile robot’s motion over the predicted
trajectory to produce reliable the operations in dangerous environments [8, 9]. The
problems of mine clearance robots were presented in Proceedings of IARP Workshop
on Robots for Humanitarian Demining and other Conferences [10–13].
The mobile robot’s motion in dangerous environments is distinguished by special
design of mechanical and sensory control systems, providing the important qualities of
actions including the reliability of working operations, high level of maneuverability,
necessary capacity for cross-surface travel, possibilities for autonomous motions over
some parts of the route, availability for supervision mood of operations using manmachine
interface, cooperative control between transport system and on-board tools, the
fast exchange of measuring information and decision making ability in some arising
nonpredicted situations of environments [10-14].
The survey is delivered to emphasize some key problems of mobile robot’s motion
in dangerous environments, such as control motion and trajectory planning to avoid the
obstacles, detection methods for automation of demining using mobile robots for
horizontal motion and wall climbing robots, fire-fighting robotic operations and some
examples of mobile robots.
The paper presents some results of a study the mobile robot’s motion in emergency
conditions to produce such operations as fire-fighting preventions, search of the mines
and demining or rescue operations.
2. Main requirements for mobile robot’s motion in unstructured environments.
Properties of dangerous environments can be characterized by undeterminateness,
dynamically changing and nonprediction.
To satisfy the demands of reliable motion, high level of maneuverability and
inspection for producing operations, robots have to be equipped with multifunctional or
fusion sensory system. The multifunctional sensory system intended for to supply the
robot control system with detection information about the route for navigation to
identify the obstacles, to looking for the objects for operations and to inspect the
technological operations.
Control system has to realize on the base of sensory information not only
programming motion, but to create special algorithm, including pattern recognition
algorithms for obstacle avoidance, algorithms for simple decision making to satisfy the
possible motion in undetermine environments. In some cases environments can be
dynamically changed, for example on the case of fire-fighting situation or moving
obstacles. Control system has to form trajectory planning with the possibility of
obstacle avoidance and realize simple decision making. Decision making for intelligent
motion can be on the base of fuzzy logic or neural networks.
In this case control system has to be rather sophisticated, and consists of navigation,
estimation and decision making blocks. The structure of such system may include local
and global feedback loops, to control robot motion not only along cruise-control curve,

ut also in near by zone of obstacles or working objects, such can be a mine. Special
blocks are responsible for the maneuverability and passibility capacity of robot for
cross-country motion. One of solution can be the combination of robot for horizontal
motion with wall climbing robot on the design of transport mechanical system.
Another solution can be multilink biped locomotion robot [7] in which the
combination of legs, pads and wheels. The combination of climbing transport system
with ordinary wheel permits to increase significantly the robot motion in dangerous
conditions. Fast reaction and small response time of control and mechanicalmechatronic
robot systems are needed for the high quality mobile robot motion.
Adaptation to changing conditions is important quality of such robots. Due to the
adaptive possibilities of the pedipulators of climbing robot to various obstacles, for
example stones, small hills, stumps of tree or artificial obstructings or barriers, robot
has to keep the working position of the demining, navigation and others sensors during
searching operations [13, 14].
The decision making production rules of control system have to satisfy the selection
of the actions, for example to overcome the obstacle or avoid it, that depends of
identification of obstacle parameters.
3. Demining mobile robots motion in unstructured environments.
The robot that has a task to move across an unknown minefield should have four sensor
blocks (Fig. 1).
Fig. 1. Sensor blocks of the robot
The internal sensors allow implementing feedback control of transport motion
and identification of robot position. They include optical sensors and accelerometers.
The navigation sensor block has an electronic compass, inertial sensors and global
triangulation system that uses position sensitive detector. Obstacle sensors are
presented by a sonar system. A metal detector, an IR detector and a chemical sensor are
applied as the mine sensors [1].
A minefield on a rough slope terrain with mines of both metal and plastic cases
was chosen as the target environment for the system to be developed. The minefield can
have stones of a height up to 150 mm. The angle of the slope can be up to 50 degrees.
The diameter of the mines can be up to 200 mm. The sensor block can have a weight up
to 30 kg.

Walking robots have high adaptability on the rough terrain and are suitable to
overcome stones and move along slope surfaces. A pedipulator (legged) robot can
achieve stable, also active, footing. When robot mistakenly detonates a landmine, the
damage will be limited to the end of the leg, although this approach requires
lightweight, inexpensive, replaceable legs. Also, by changing the broken leg, this robot
can be restored.
Mines of both metal and plastic bodies can be detected by means of a combined
sensor block that contains devices based on different principles. An analysis of the
combined sensor block information provides high reliability of mine detection.
The design of the transport module{ XE "transport module" } of the robot is
shown in Fig. 2. It consists of longitudinal pneumatic cylinders and latitudinal
pneumatic cylinders, which bodies are connected symmetrically and have 200 mm
stroke to cover maximum mine size in one stroke. Each pneumatic transport cylinders{
XE "transport cylinder" } has two pedipulators that are fixed at the ends of the piston
rods. The pedipulator consists of a lifting cylinder{ XE "lifting cylinder" } of 150 mm
stroke to overcome maximum stone obstacles and a foot with toothed contact surface to
improve robot climbing possibilities. The mine detection block{ XE "detection block" }
is connected to the front part of the robot. Linear position sensor of longitudinal motion
is placed on a body of the longitudinal cylinder. Linear position sensor of latitudinal
motion and the detection block is placed on a body of the latitudinal cylinder.
6
5
10
11
1
9
4
14
Fig. 2. Design of the transport module of the robot.
1, 2 - longitudinal pneumatic cylinders; 3, 4 - latitudinal pneumatic cylinders; 5 -
lifting cylinder; 6 – foot; 7 - metal detector; 8 - IR detector; 9 - chemical sensor; 10 -
linear position sensor of longitudinal motion; 11 - linear position sensor of latitudinal
motion; 12 – valve units; 13 - supply rotation block; 14 - electronic compass; 15 - onboard
control computer.
13
8
7
2
15
3
12

The valve units are placed at both sides of the robot providing minimum tubing
length. They are supplied from a supply rotation block connected to the main air
pressure line. The rotation block allows free rotation of the robot in relation to the
umbilical cord with the air supply pipe. An electronic compass is installed in the front
of the platform. The on-board control computer is placed in the centre of the platform.
The pneumatic cylinders can be actuated in two modes:
- the first mode is a transport mode. In this case, the longitudinal cylinder performs
motion of its pedipulators with maximum velocity by using all length of the piston rods.
During this motion the robot is connected to the ground by means of the latitudinal
cylinder pedipulators. They serve as support cylinders during motion in this direction.
The longitudinal cylinder pedipulators are lifted. After the first step the longitudinal
cylinder pedipulators must be connected to the ground and the latitudinal cylinder
pedipulators must be lifted. In such a position the sensor block can be moved for one
step towards the working zone, and so on. The robot can change a motion direction by
90°, by actuating the latitudinal cylinders as transport ones instead of the longitudinal
cylinders. The rotation of the robot can be carried out by means of simultaneous motion
of longitudinal or/and latitudinal cylinders in opposite directions during a contact of all
their feet with the motion surface;
- the second mode is a searching mode. During this mode the sensor block must carry
out searching functions and be moved along a scanning trajectory. This trajectory is
performed by means of latitudinal and longitudinal cylinders, which are actuated with a
nominal searching velocity. The value of the searching velocity is determined by the
characteristics of the sensor block.
The robot motion and mine sensing are controlled by means of a distributed control
robotic system. This system provides remote control from a safe distance in an
automatic mode or in a teleoperated mode by an operator. The distributed control
robotic system architecture is shown in Figure 3.

Location
sensor
Detection
block
Linear
sensor block
Compass
Optical
sensor
Operator
Central
computer
On -board
computer
Interface
Robot drive
system
End-effector
signals
Navigation
program
Covering
program
Data fusion
program
Fig. 3. Structure of the distributed control robotic system.
A search head of the metal detector ATMID{ XE "ATMID" } [2] is used in the
detection block of the robot. Metal detectors are based on the fact that variable
electromagnetic fields produce responses in metallic objects. The transmitting coil,
embedded in the search head of the metal detector, generates such a field. As the search
head is swept over the ground, the receiving coil in the search head detects the
variations in the electromagnetic field caused by metallic objects. The variations are
then processed to generate a signal indicating the presence of metal in the ground
beneath the search head. The metal detector is installed on the front part of the robot as
it is shown in Fig. 4.
2
5
3
7
4
1
9
6
8

Fig. 4. Position of the metal detector on the robot
1 - metal detector, 2 – robot platform, 3 – support, 4 – height adjusting unit, 5 –
pedipulator, 6 – mine, 7 – metal part of the mine, 8 – transmit field, 9 – receive field.
The detector was tested to detect a real mine but without explosives. It was the
Portuguese mine of MAPS type with the plastic case of 85 mm diameter that contains a
1 g metal part. A picture of the mine is shown in Fig. 5.
Fig. 5. Mine of MAPS type on the surface of the test field.
Figure 6 shows the influence of the ground type on the mine detection signal for
sand, soil and stones at the mine depth of 10 cm.
Frequency [Hz]
2500
2000
1500
1000
500
METAL DETECTOR - Mine at 10cm depth
Sand
Soil
Stones
0
0 20 40 60 80 100 120 140 160 180 200
Distance [mm]
Fig. 6. Influence of the ground type on the mine detection signal

Sand allows receiving the strongest signal from the mine. Stones reduce the output
signal about 5% and the compact soil leads to a reduction of about 2%. The mine in
sand was detected at the 20 mm longer distance that other grounds while scanning.
A new infrared detection sensor for detection of plastic or metallic landmines
was developed to be integrated in the robot. The mine sensing is based on an infrared
image analysis during microwave soil heating and posterior cooling. The detector
prototype contains a microwave klystron emitting 1 kW power at the frequency of 2.45
GHz and two infrared sensors sensitive in the range of 8-14 µm. Depending on the soil
dielectric properties, the emitted radiation will be absorbed, reflected or transmitted
through. Common plastic material is transmissive, metals reflect the microwaves, and
wet soil absorbs and converts the radiation to heat. Using this sensor, it is possible to
image thermal gradients in the soil surface and detect different rates of temperature
changes depending on the soil content [4].
The results shown in Fig. 7 were obtained after heating of the simulated minefield. It is
the cut of a volume of soil with a plastic landmine in the interior. For the simulation,
the initial temperature of 290°K in all materials was considered.
The simulation is composed of two phases. In the first phase, the workspace is
exposed to homogeneous electromagnetic radiation of 2.45 GHz, 1 kW in 0.1225 m 2 .
The temperature raises according to the materials properties. The initial values of the
second phase are the final values of the first phase. In this phase the material is not
exposed to radiation, so only diffusion phenomena’s occur by heat conduction.
cm
air
mine
soil
cm
K ο
ºC × 10
cm / 20 cm × 2
Fig. 7. Heat distribution. Fig. 8. IR sensor output after
scanning of the heated mine location (mine
of MAPS type)
If a mine is laid at the level of the ground and it was heated up to 40 ºC, the
output configuration of the mine location corresponds to Fig. 8 after the sensor
scanning of the ground by means of the robot.

Fig. 9. Example of the scanning Fig. 10. Multisensor demining robot on a test field.
trajectory to avoid an obstacles or a mine.
The robot starts to check an obstacles or a mine at a working distance of a
corresponding sensor at the detection point and uses this information to avoid them.
The developed multisensor robot is presented in Fig. 10 on a test field searching a
mine.
Due to the adaptive possibilities of the pedipulators to obstacles (a stone in the
picture), the robot can keep the working position of the demining sensors during
searching operations.
4. Examples of mobile robot motion in dangerous conditions
Basic conception of the control system intended for mobile robot motion in
dangerous environment make provision for navigation and estimation (Fig. 11). The
control system includes local control loops for ordinary feedback control and global
control loop for estimation and simple decision making producing. The local control
loops are working all time of the motion. Navigation subsystem produce sensory
information processing, image and pattern recognition, task and path planning and
control of drives and transdusers.
The estimation subsystem produce analysis of prescribed tasks and nonpredicted
situations as so as motion analysis including manoeuvrabilities.
Wall climbing robots with on-board manipulators (Fig. 12 and Fig. 13) are
controlling by means of this kind of control system. Appropriate path generation
algorithm (Fig. 19) is intended for obstacle avoidance in dynamically changing
environments.
Prototype of wall climbing robot for fire-fighting application was designed with cooperation
with All-Russian Research Institute for Fire Protection (Fig.14). Robot has
two arms on-board manipulator with load capacity 120 kg. Cutting and fire-fighting
material in the arm end-effectors intended for an implementation of various fire-

fighting operations or cutting holes in petroleum storage tanks under fire. Diagram of
the heat shielding and plot of temperature change are illustrated by Fig. 15 and Fig. 16.
Structure of the control system for the trajectory planning with obstacle avoidance
(Fig. 17) permits to generate necessary actions for the robot motion with the simple
decision making on the base of computer simulation. Example of the robot motion is
delivered in Fig. 18, when the robot produce obstacle avoidance with manoeuvre over
trajectory planning. The picture is reflected the computer simulation.
One of the way for mobility and reliability increasing of mobile robots is achieved
by combination of mobile robot for horizontal motion with on-board wall climbing
machine (Fig. 19). The simple manipulator of mobile robot is attached the wall
climbing machine to the wall to move this machine over the wall.
Examples of mechanical leg system (Fig. 21) and mechanical multilink system (Fig.
22) show the possibilities to overcome the angles between crossing surfaces in the
space.
Conclusion.
The results of R&D of mobile robots intended for the motion in dangerous
nonpredicted environments are delivered. The mobile robot motion in emergency
conditions is distinguished by means of special design the mechanical, sensory and
control systems, providing the high level of manoeuvrability, the motion along irregular
surfaces in the space, possibilities for autonomous motion and supervision mood of
operations to produce such working tactics operations as demining, fire-fighting and
rescue.

Fig. 12 Fig. 13
Fig. 11

Fig. 15
Fig. 16
Fig. 19
Fig. 14

Fig. 21
Fig. 20
Fig. 22
Fig. 17
Fig. 18

References
1. M. Rachkov, L. Marques, A. T. de Almeida, Automation of demining, Textbook,
University of Coimbra, Portugal, 2002, 281 p.
2. Hofmann, B., Rockstroh, M., Gradetsky, V., Rachkov, M., 1998. Acoustic
navigation and inspection methods for underwater mobile robots, In Proc. 1st Int.
Workshop on Autonomous Underwater Vehicles for Shallow Waters and Coastal
Environments, Lafayette, USA, S. IV: pp. 1-13.
3. Moita, F., Nunes, U., 2001. Multi-echo technique for feature detection and
identification using simple sonar configurations, IEEE/ASME Int. Conf. On Advanced
Intelligent Mechatronics, pp. 389-394.
4. ATMID mine detecting set, Operating manual, Schiebel Elektronische Geraete
GmbH, Austria, 2000.
5. D. Noro, N. Sousa, L. Marques and A.T. de Almeida, Active Detection of
Antipersonnel Landmines by Infrared, Annals of Electrotechnical Engineering
Technology, Portuguese Engineering Society, 1999.
6. Explosives detection equipment, ARLI SPETSTECHNIKA, Catalogue 2001,
Moscow, 2001.
7. V.Gradetsky, V.Veshnikov, S.Kalinichenko, L.Kravchuk “Mobile Robot’s
Control Motion Over Arbitrarily Oriented Surfaces in the Space”, Moscow, “Nauka”
publ., 2001, pp. 361 (in Russian).
8. Y.Baudoin “Improving the Safety and the Quality assurance with Robotics
Tools”, ISSC Conference, April 18-19, 2002.
9. Y.Baudoin, M.Acheroy “Robotics systems for Humanitarian Demining: Modular
and Generic Approach and Cooperation Under IARP/ITER/ERA Networks, IARP
Workshop on Robotics for Humanitarian Demining, HUDEM’02, Vienna, Austria,
November 3-5, 2002, pp.1-6.
10. Proceedings of VR-MECH’01 Workshop, BSME, Royal Military Academy,
Brussels, Belgium, November 22-23, 2001, pp.241.
11. Proceedings of IARP Workshop on Robots for Humanitarian Demining
HUDEM’02, Vienna, Austria, November 3-5, pp. 110.
12. A.Almeida, L.Marques, M.Rachkov, V.Gradetsky “On-Board Demining
Manipulator”, Proceedings of IARP Workshop on Robots for Humanitarian Demining,
Vienna, Austria, November 3-5, pp. 45-50.
13. Ch.Grand, F.Ben Amar, F.Plumet and Ph.Bidaud “Simulation and Control of
High Mobility Rovers for Rough Terrains Exploration”, Proceedings of IARP
Workshop on Robots for Humanitarian Demining, Vienna, Austria, November 3-5, pp.
51-56.
14. V.Gradetsky, V.Veshnikov, S.Kalinichenko, G.G.Rizzotto, F.Italia “Fuzzy
Logic Control for the Robot Motion in Dynamically Changing Environments,
Proceedings of 4 th CLAWAR 2001 Conference, September 24-26, 2001, pp. 377-386.

PARADIS: Focusing on GIS Field Tools for
Humanitarian Demining
Sébastien Delhay Vinciane Lacroix Mahamadou Idrissa
Signal And Image Center (SIC)
Royal Military Academy (RMA)
30, avenue de la Renaissance 1000 Bruxelles
sdelhay@elec.rma.ac.be Vinciane.Lacroix@elec.rma.ac.be idrissa@elec.rma.ac.be
1. Introduction
PARADIS stands for a Prototype for Assisting Rational
Activities in Demining using Images from Satellites. The
aim of this project is to improve the planning of
Humanitarian Demining campaigns using Remote
Sensing data and GIS (Geographic Information System)
techniques. In this context, a user interface has been
developed in ArcView GIS to integrate the tools needed
for the management of the campaigns. The interface is
built upon four scales (global, regional, local and
advancement) and really follows the evolution of the
campaign from global to local scale by providing the user
with scale-dedicated tools. The system uses high and very
high resolution satellite images and their interpretation,
onto which one can overlay vector data taken from the
IMSMA (Information Management System for Mine
Action) database, which is used as data repository. The
tool is also compatible with the Belgian EOD
Champassak database, aimed at gathering information
related to the collection of UXO’s on the field. The
PARADIS project is funded by the Belgian Defense; the
developed system will be used by the SEDEE-DOVO
(Belgian Armed Forces Bomb Disposal Unit) during their
campaigns in mine- and UXO-affected countries.
Several problems are inherent to the process of
supporting Humanitarian Demining campaigns with
useful data. Within the framework of a demining
campaign, contaminated areas are often very large and
represent a lot of information. This huge amount of data
has to be compiled and safely stored in a central
repository to avoid loss or corruption of data. Also, these
data need to be represented in an explicit manner, in order
for the user to work with them as effectively as possible.
Even when a system has been developed to solve this
first problem, errors of encoding may appear when the
user enters new data into that system; in order to ensure
the integrity of the data, these errors must be eliminated.
Maps are key elements for campaign management and
field work; however, as demining campaigns generally
take place in developing countries, there is often a lack of
accurate and recent maps for the zone of work.
Information systems have been developed in order to
solve those problems and are now used in countries
affected by Explosive Remnants of War (ERW). IMSMA,
the UN-standard information system for Mine Action,
addresses the first problem. It consists mainly of a
database located at the Mine Action Centre (MAC) of a
specific country or region, into which all Mine Action
related data are centralized. It also contains management
and planification tools that process the information in the
database in order to help decision makers.
EOD IS (Explosive Ordnance Disposal Information
System) focuses on the second problem – data integrity.
This system provides a GIS (Geographic Information
System) based interface to the field worker to ensure that
the data collected on the field is safely conveyed to the
IMSMA database.
This paper explains how PARADIS addresses these
problems, then focuses on the tools being developed for
the field work.

2. Overview of the system
2.1. Original ideas
PARADIS addresses the first problem by compiling
and organizing the data related to the campaign in a
geographic database (GeoDb). As PARADIS is built on a
GIS, it presents the data to the user on a map rather than
on a form. The map representation is more contextual than
the form representation, and enables the user to really
visualize the information, as it is shown on the figure 1:
GIS representation
(map)
X Y
574,659 1,608,404
574,700 1,608,450
574,602 1,608,360
574,563 1,608,306
Database representation
(form)
Fig. 1: Perimeter of a minefield
IMSMA is database-centered, thus it mainly presents
the data on forms. In order to fill the IMSMA database
using a map, the user will be able to enter the data in the
PARADIS system, then export them to IMSMA using the
maXML (Mine Action eXtended Markup Language, [2])
protocol.
Concerning the second problem, because it is
GIS-based, the PARADIS interface intrinsically enables
the user to verify that the data have been correctly entered.
For example, when defining the perimeter of a minefield,
if the user inverts the X and Y coordinates of a turning
point, the latter appears at a totally different spot than
expected. The wrong location of the turning point on the
map warns the user about the encoding error he just made,
and he then is able to correct the error.
Finally, in order to provide the user with useful maps of
the zone of interest, PARADIS uses satellite imagery (see
[1]).
2.2. Designing two dedicated interfaces
The needs of the campaign manager differ from those
of the field operator. While the manager needs tools to
help him make decisions at country and regional levels,
the field operator needs an easy-to-use, lightweight tool to
bring data to the field and collect new information while
there. The PARADIS interface as it is described in [1] has
been split into two separate interfaces in order to better fit
the needs of both the campaign manager and the field
operator.
The manager interface is located in the Mine Action
Center (MAC); it is built on a full-featured GIS. As it
needs to access the central GeoDb and may run big
processes while manipulating the data, it is based on a
desktop –or laptop- computer. It is hence referred to as the
Desktop Interface.
The field operator will work on a lightweight GIS
interface called the Field Interface. The latter is running
on a Personal Digital Assistant (PDA), which small size
and light weight make it the ideal tool for the field work.
Fig. 2: A Personal Digital Assistant
The following figure shows the procedure that a field
activity involves:

Desktop
Interface
1
3
Field
Interface
(PDA)
GeoDb Field data
MAC Field
Fig. 3: Field activity procedure
Desktop Interface (Point 1):
As the field operator needs geographic data such as maps,
satellite imagery and its interpretation to support his
activities, the first step is to export these data to the PDA.
Note that the PDA user only needs data for the zone where
he will be sent, so only a reduced set of geographic data
needs to be exported. In the Desktop Interface, the
manager reduces the set of data to the region of interest by
drawing a bounding rectangle on the map. He then exports
the enclosed data set to the PDA.
Field Interface (Point 2):
On the field, the PDA is used to collect new data
such as: location of UXOs, bridges, audio/video files of
people interviews, etc. The user may also edit the existing
set of data, which is very useful to enter information that
cannot be known a priori. In particular, the only way to
know the state of the roads is to go to the field.
Desktop Interface (Point 3):
When the field operator comes back to the office the
manager gets the data from the PDA, verifies their
integrity and puts them into the central GeoDb.
As more and more data comes into the GeoDb, the whole
region gets covered by the data brought in by the different
field operators, which enables the manager to make
decisions at regional level.
2
3. Description of the Field Interface
The process described above will be used during three
separate phases of the demining campaign: Impact Survey,
Technical Survey (Clearance) and Quality Assessment. In
the following we describe the tools that are being
developed for the Field Interface. Some tools are
dedicated to one phase and will only be used during that
specific phase, while other tools are of a more general use.
3.1. Impact Survey – Road Map tool
During this phase, surveyors are sent to the field to
collect the information that will be used to evaluate the
socio-economic impact of ERWs on the population. In
order to interview the population about the ERW situation,
they go from village to village along the roads of the
region. In this context, the Road Map tool helps managing
the itineraries the different teams of investigators will
follow.
Desktop Interface:
This tool allows a road map to be created for each team of
surveyors. In the Desktop Interface, the manager draws
the itinerary to be followed by the team; the resulting road
map is then exported to the PDA.
Field Interface:
Each segment of the road is namely given:
- a name;
- a status (not practicable, poor or good);
- seasons of practicability (if any);
- a type (tarred road, track);
- a mean speed depending on season of
practicability;
- a mined status reflecting the estimated presence
of ERWs along the road and the types of ERWs.
As one goes along the road, the status of the road may
change (e.g. from poor to good). In order for the Field
Interface to reflect this information, the road segment can
be easily split up into several segments of different status.
Conversely, two segments can get merged into one single
segment.

Fig. 4: The road color reflects the status of the road
(brown= unknown; red= not practicable; orange= poor;
green= good). The user can toggle labels to show the road
name and its mined status.
When the team goes back to the MAC, these data are
imported into the central GeoDb; after grouping the road
maps of the different teams, a regional map of roads is
obtained.
3.2. Technical Survey - Grid tool
This tool will allow the user to follow the work on a
minefield. After delineating the minefield, the user will be
able to draw a grid on the minefield. Each cell in the grid
represents the small area that a deminer works on at a time.
Then, each cell will be edited to show the ammunition
found in each cell and the status of the cell (not cleared,
ongoing, cleared). After a certain number of cells have
been edited, the system will be able to compute the time
for complete clearance of the minefield, given an
estimated future numbers of detectors on the minefield.
3.3. General tools
These tools are of a general use and can be used in
whichever phase of the demining campaign (Impact
Survey, Technical Survey or Quality Assessment). The
Layers tool makes the process of introducing new data
into the system simple and powerful; the GeoNote tool
enables the user to quickly associate different types of
data to the same location on the map.
3.3.1. Layers tool
Using this tool the user may select dedicated layers
from a list and insert them in the map (ex: airports, gas
stations, hospitals, towns, UXO’s, etc). The layers are
then loaded with the appropriate pre-defined symbology.
Possible
layers
Layers
to load
in map
Fig. 5: Layers tool form
Layer to edit
The user may also choose to edit one layer directly,
using the “Edit layer” box.
Note that the way of editing a layer changes from one
layer to another. For example, the Bridges and Minefields
layers differ in that a bridge is represented as a point in the
interface, while a minefield is represented as a polygon.
Hence, to define the bridge point the user simply taps on
the screen at the appropriate location. As the minefield is
a polygon, the system behaves differently: when the user
taps on the screen, the benchmark gets defined, then a
form opens and lets the user define the other points. This
is shown in the following figure:

Fig. 6: The form used to define the turning points of a
minefield; these points can be defined by bearing and
distance to the previous point, or by their X and Y
coordinates.
The advantage of the Layers tool is that the system
activates the appropriate editing method depending on the
layer the user has chosen to edit.
3.3.2. GeoNote tool
Sometimes it is useful to associate heterogeneous
information to a single point on the map, such as text,
voice recordings, photos, or videos; in the interface, this
point is called a GeoNote. As an example, if the PDA user
interviews a mine victim, he would create a GeoNote by
tapping the screen at the location of the interview. He
would then associate to the GeoNote the audio file of the
conversation, along with a sketch and a textual
description of where the accident happened.
4. Conclusions and future prospects
The current state of the Field Interface has been
presented; the two interfaces are still being developed
based on the ideas and methods of work of the end-users –
the deminers of the SEDEE-DOVO. Hands-on training
sessions will be put in place in order to teach the end-users
how to use the Desktop and Field interfaces. The whole
system will then be validated in situation during test
campaigns.
In particular, the validation of the Field Interface will
focus on practical details such as the possibly difficult use
of the PDA in outdoors conditions (because of the sun,
dust, etc), and the handiness of the interface for the field
operator.
5. References
1. S. Delhay, V. Lacroix, M. Idrissa, PARADIS : a
GIS Tool for the Management of Humanitarian
Demining Campaigns, In International
Conference on Requirements and Technologies
for the Detection, Removal and Neutralization of
Landmines and UXO (EUDEM2-SCOT), VUB,
Brussels, Belgium, September 2003
2. N. Klemm and R. Bailey, Sharing Mine
Action Information,
http://www.hdic.jmu.edu/conference/casualty/m
axml_files/frame.htm

Remote Sensing System for Robotics and Aquatic
Related Humanitarian Demining and UXO Detection
Dr. Charles R. Bostater Jr. 1 , Teddy Ghir, Luce Bassetti
Marine & Environmental Optics Lab and Remote Sensing Center
College of Engineering, Florida Institute of Technology
Melbourne, Florida, USA
Abstract
It is becoming more important to understand the remote sensing systems and
associated autonomous or semi-autonomous methodologies (robotic & mechatronics) that
may be utilized in freshwater and marine aquatic environments. This need comes from
several issues related not only to advances in our scientific understanding and
technological capabilities, but also from the desire to insure that the risk associated with
UXO (unexploded ordnance), related submerged mines, as well as submerged targets and
debris left from previous human activities are identified and reduced through detection
and removal. This paper will describe (a) remote sensing systems, (b) platforms (fixed
and mobile, as well as to demonstrate (c) the value of thinking in terms of scalability as
well as modularity in the design and application of new systems now being constructed
within our laboratory as well as future systems. New systems – sensing systems as well
as autonomous or semiautonomous robotic and mechatronic systems will be essential to
secure domestic preparedness for humanitarian reasons. These same systems hold
tremendous value, if thoughtfully designed for other applications which include
environmental monitoring in ambient environments.
I.1. Basic Concepts
I. Background
The engineering talent, effort and economic costs involved with the design of
remote sensing systems aboard robotic and mechatronic systems are not inconsequential.
Typically, the design of remote sensing systems first involves considerations with respect
to which part of the electromagnetic spectrum will be used. Second, one must consider
the “medium” the target is embedded in. Both of these considerations are necessary to
understand which part of the electromagnetic spectrum and associated sensor system may
be appropriate for the detection problem. In essence, most targets are thus imbedded
inside an environmental medium (water, air, soil or submerged land). In most cases, the
“ambient” environment (outdoors environment) is the location we consider when
thinking of UXO’s and mines in water or submerged land underling the water column.
1 Dr. Charles R. Bostater Jr., Associate Professor, Physical Oceanography and Environmental Sciences;
Director Marine and Environmental Optics Lab & Remote Sensing Center, College of Engineering,
Florida Institute of Technology, Melbourne, Florida, US, 32937 email: bostater@probe.ocn.fit.edu Ph:321-
258-9134, Fax: 321-600-9412. Paper submitted to Proceedings and to be presented at the HUDEM04
Conference, June 16, 2004, Brussels Belgium http://www.ihrt.tuwien.ac.at/hudem04/.

This is what we call “ambient environmental monitoring” with respect to dual uses of
sensing systems for environmental monitoring applications. Figure 1 describes the remote
sensing selection process with respect to the sensing system to be used as described
above.
Remote Sensing Systems:
Measure A Portion of the
Electromagnetic Spectrum
Targets or Objects Are Located
In an “Ambient” Environmental Medium.
The Medium & the Target both determine the Selection of:
(a) the portion of the EM Spectrum to Sense (use) &
(b) thus Remote Sensing System to apply.
Figure 1. Schematic decision tree describing the selection of the sensing system for object
detection in an “ambient” environmental medium (air, water, soil, submerged land).
I.b. Remote Sensing Platforms
Next, when designing a remote sensing or object recognition system, after
considering the above, one needs to consider the platform of choice for the detection and
localization (spatial location detection) of the object(s) or targets to be identified or
discriminated from the background medium. Generally speaking, we consider two types
of platforms, (a) fixed or (b) moving. In aquatic environments, the fixed platform could
be mounted to a (1) pier, (2) piling, (3) fixed platform (bridge, tower, or other fixed
platform away from or extending from the shore, e.g. an oil rig platform or other such
specially designed platform – e.g. breakwaters) and a (4) moored bottom, subsurface or
in-situ (at depth) or water surface moored platform (e.g. buoy). The moving platform
with a remote sensing system can be conceived of as either a (1) drifting-autonomous, (2)
semi-autonomous (tethered) or (3) autonomous-controlled platform. We commonly
would call the latter two moving platforms as a “vehicle”. Each of these moving
platforms could be classified as either: (a) surface vehicles or (b) sub-surface and/or
bottom roving vehicles. After the terminology of the CLAWAR network, we can consider
the vehicles as either the (1) walking type, (2) or crawling type both of which may
change its elevation with the land or submerged land surface in which case we may say
the vehicle is a “climbing” vehicle. The vehicles can next be classified as to either
manned or unmanned (another way to infer autonomous). Figure 2 below is a schematic
depicting the categorical types of remote sensing platforms described above with respect
to object detection and localization for objects that are spatially fixed or not moving with
the environmental medium (water, soil or submerged land).

Another important aspect with respect to remote sensing systems to keep in mind
is that systems are designed with the purpose to either (a) detect targets within a
medium or (b) to detect the characteristics of the medium itself with respect to its
properties (absorption, scattering, transmittance and or reflectance). In some cases, one
may need to consider both aspects, especially in the event that the object’s recognition is
dependant upon correction or removal of these selected processes from the remote
sensing signal in order to detect the target or object’s characteristics and location within
in the aquatic environment as depicted below in Figure 3.
Piers, Pilings, Docks, Bridges, Towers,
Rig Platforms, Breakwaters, Moored
Remote Sensing Platforms:
Fixed or Moving
Surface, Bottom, Sub-Surface
Manned or Unmanned
Drifting, Autonomous-semi (tethered)
Autonomous (controlled).
Flying, Swimming, Floating, Walking,
Crawling – Climbing or changing elevation
(roving) with terrain.
Figure 2. Schematic of remote sensing platforms for detection of objects that are spatially fixed
(not moving) within the environmental medium (water, soil or submerged land).
Detecting Properties
The Medium:
•scattering
•absorption
•trasmission
•reflectance
•homogeneous/nonhomogenius(layered)
Remote Sensing Systems:
Object:
•Detection
•Recognition
•Discrimination
Detecting Properties
The Object or Target:
•scattering
•absorption Object
•transmission Properties
•reflectance
•size - shape
Figure 3. Schematic depiction of the use of remote sensing systems as being used to detect or
characterize (a) the properties of the medium, and/or (b) the properties of the object or target. These
two considerations in the remote sensing system design and operation, including data analysis or
algorithm development and use leads to (3) the detection and recognition of the object and its
discernment or discrimination from the background medium.

I.c. Remote Sensing Systems
After considering platforms, one needs to consider the basic remote sensing
systems in terms of their detecting an object through use of the electromagnetic spectrum
they sense. For water environments one classifies the systems as UV, Visible, Near-
Infrared, Mid-Infrared, Far-Infrared, Microwave, and Acoustical systems. Each of these
systems can be either passive sensor systems or active-passive systems. Passive systems
include spectroradiometers, spectrographs, cameras (film or digital), hyperspectral
imagers, passive acoustic hydrophones and microwave radiometers. Active-passive
systems include, laser based Lidar (light induced detection and ranging) or ladar, Radar
(microwave) imaging, Sonar (acoustic) imaging, fluorescence systems which may
include Raman spectrometers and imaging systems which utilize the Raman scattering
signal. All of these systems may include the use of multiple wavelength excitation and
emission detection systems, magnetic field and acoustic or Sonar devices. Typically, all
of the active-passive systems utilize a passive detector component which measures the
medium or object’s response to an active continuous (CW) or pulsed radiation source.
Remote Sensing Systems: Measure a Part of the EM Spectrum
X-Ray, UV, Vis., NIR, Mid-IR, Far-IR, Microwave, Acoustic
Passive Systems: Spectroradiometers,
Spectrographs, Radiometers,
Hydrophones, Cameras (film, digital)
Surface, Bottom, Sub-Surface, In-Situ
Active-Passive: Lasers (CW),Lasers
(Lidar) Mulitwavelength Excitation &
Emission (Fluorescence, Raman)
Microwave (Radar), Acoustic (Sonar)
Magnetic (Field Induction)
Point, Line Imaging, 2-D Imaging, 3D-Spectral
Data Types
Figure 4. Schematic describing the different remote sensing systems and categories such as passive,
active-passive systems, spatial categories of data obtained from these systems and their data types.
II. Detection of Objects in Water
II.1 Methods of Risk Assessment & Risk Reduction
The goal of detecting objects such as mines and unexploded ordnance and related
debris using robotic systems may ultimately provide a unique method of securing a safe
environment for human activities. With the increased humanitarian need to secure a safe
environment by many nations around the world, especially safe waterways, ports and
associated coastal areas such as bridges, piers, breakwaters, etc., the detection of objects
in water is an essential aspect of national preparedness. The detection of dangerous
objects from other natural objects, such as rocks, submerged bottom types (submerged
vegetative areas, etc.), man made debris, first needs to be demonstrated before a remote
sensing system is routinely utilized in a robotic or unmanned system. The determination
of the risk associated with false identification of dangerous objects must also be

determined, estimated and then minimized. This minimization process or function is a
necessary element in the demonstration and improvements of any remote sensing system
utilized aboard a robotic autonomous or semi-autonomous platform.
Conceptually, there are 3 major methods useful to minimize this risk in false
identification of objects. The first risk assessment methodology is the use of sensor
system modeling and simulations, the second method is through system simulations and
modeling of the robotic and mechatronic systems, and the third method is through
conducting field trials where actual integrated remote sensing systems and associated
robotic and mechatronic systems are tested in an artificial environment (e.g. example a
test mine field with various targets objects or debris to be classified as to their potential
risk as depicted below.
Methods For Minimizing Risks In
Detection of Dangerous Objects:
•Remote Sensing Systems modeling & associated
detection algorithm simulations.
•Robotic-Mechatronic Systems modeling & motion
control simulations.
•Filed Trial Tests in Artificial Object Environments
Estimation Of Risk Reduction Through:
• Systems Hardware Modifcations
• Data Analysis & Detection Algorithm Modifications
• Changing Field Trial Environments & Object
Characteristics Modifications
Figure 5. Systematic process for risk assessment and risk reduction in development,
testing and utilization of possible UXO and mine detection systems through modeling,
simulations, and resulting system modifications.
II.b. Example Applications: Shallow Water Remote Sensing Systems
In the follow discussion we present design concepts for shallow water operations.
The designs have incorporated three important concepts we believe are useful
considerations and design concepts. These are (1) utilization of the “sensor payload
concept”, (2) the design of both the sensor and the robotic-mechatronic elements using
the concept of “size scalability” and (3) the use of components using “modularity” or
modular components. All three of these design elements have been used by the main
author as a guidepost in the systems constructed to date, and those currently being built.
Figure 6 shows moored or tethered shallow water remote sensing system which was built
by the author. This system is designed as a tethered shallow water system and was
designed with the concept of a “sensor payload”. This particular example contains an
internationally patented non contact scalable backscatter probe (Bostater, 2000) within

the payload sensor well which is essentially a hyperspectral remote sensing system. This
particular system was designed to measure identify and quantify chemicals in a shallow
water environment and the payload and probe hold an active electromagnetic energy
source and a passive multi-wavelength sensor. Next to the photograph is shown the side
and top perspective of the moored or fixed remote sensing platform which is scalable in
size with respect to the flotation collar as well as with respect to a sensor payload holding
area and as well as the probe.
Figure 6. Example of a coastal shallow water observation and surveillance system which utilizes a
scalable in size payload area, scalable chemical probe for multi-wavelength EM source and detector
system such as a Raman scattering system or a pulsed Laser line fluorescence based hyperspectral
sensor.
Following the design and fabrication of this system we developed the concept of
an autonomous system which can hold several payload sensor wells. Figure 7 present the
floating autonomous vehicle which is currently being built for wireless controlled active
motion and for active-passive (LIDAR) imaging, hyperspectral imaging, Raman imaging,
acoustic (SONAR) and magnetic imaging systems. This system is planned for mapping
surveys of shallow water areas around bridges, waterways, canals, channels and ports in
order to map the submerged land or bottom feature types and to identify and localize
natural and manmade objects and debris, as well as unidentified objects such as UXO
and/or mine like objects. This system is “scalable” in size, accommodates payload areas
or “sensor wells”, and is made from off-the-shelf components that are modular for
construction purposes. It is scalable in size (thrusters) to power a 3-axis motion control
system, wireless radio based fast ethernet communications and fast maneuvering control.
Figure 7. Figures showing the scalable in size, modular structural and motion control components
and scalable sensor payload or sensor well for shallow water mapping and object detection. The
vessel is currently under construction and utilizes wireless Ethernet communications, an onboard
sensor system for around bridges, canals, waterways, piers and enclosed seaport areas as well as in
lakes and coastal lagoons. The system is battery and solar powered for WAAS GPS motion control.

In addition, this robotic shallow surface water vessel is designed for easy assembly and
disassembly making it ideal for transport in a small light weight container, along with
selected, optional, scalable and modular remote sensing systems for object detection and
discrimination. The above 2 examples are next contrasted with futuristic versions of the
concepts embodied above. In Figure 8, below we depict the embedded concepts described
above to “submerged walking robotic systems”. These “walker robots” were conceived
and motivated based upon recent examination of underwater flagellate organisms found
in a European freshwater lake, Lake Balaton, Hungary. As will be noted, these robotic
concepts contain the design elements used above in addition to the additional concept of
utilizing acoustic sensors in the feet of the multi-legged robotic system, surrounding the
center payload area.
Sensor Payload
Robotic Operating System
(ROBOos)
Object Buried
In Sediments
Figure 8. Designs of submerged walking robots hold unique possibilities and advantages for
mapping and detecting objects buried in submerged land or bottom types. The center of the walking
holds one suite of sensors and the feet (which help support in soft bottom types) contain acoustic and
magnetic detectors which receive the returned signal from the central robotic EM impulse generator.
II.c. Example of Sensor Imaging System Modeling and Algorithm Simulation
In order to demonstrate the value of designing security related and homeland
preparedness systems, an example is provided of simulations for detecting targets in
different water types (clear water versus turbid water types. Analytical and Monte-Carlo
based radiative transfer models designed to quantitatively describe, simulate and predict
the electromagnetic transport properties, in an environmental medium such as water,
provide powerful tools for educating future engineers and remote sensing scientists as to
the design of remote sensing systems as well as the possible robotic systems which
contain and move the sensors. Over the last decade, the Marine Environmental
Laboratory and Remote Sensing Center at Florida Institute of Technology has developed

a suite of instruments and associated models which help describe the optimal design of
remote sensing instruments. Below are shown selective results of model simulations
regarding the detection of an object. In this case it is a line target with 80% and 20%
constant reflectance lines across the spectrum. This target could have a complex shape
and unique property characteristics in terms of its EM reflective/transmission
characteristics. In addition, the medium (in this case water) is composed of clear water
and then different chemical properties (dissolve organic matter, chlorophyll-a pigment
concentrations and suspended sediment concentrations) which make optical detection
difficult, unless the active-passive systems are employed. Figure 9 shows the simulated
line target simulated on the water bottom at a specified depth (≈2m). We also use a
spectral water surface wave model demonstrating the combined effects of wind on the
water surface and water quality influences on the target at 3 different wavelengths where
the resulting RGB image 480 nm, 530 nm and 650 nm simulated channels as shown.
Figure 9. Synthetic image sensor simulations of a hyperspectral imaging system with continuous
source illumination using a radiative transfer model (Bostater, et. al. 2000, 2004). A line target is
simulated with 80% and 20% reflectance lines. The line target (upper left) is simulated as a ≈140 cm 2
target. The water wave surface is generated using the JONSWAP water wave spectrum (upper
middle) modified for shallow water use in the Sebastian coastal inlet area (upper right) of the
Atlantic Ocean, Florida. The wind speed is 5 ms -1 from the NNE direction (30 0 ). The solar zenith
angle is set to 20 degrees off nadir (late morning) and the sensor viewing angle is near nadir. The
bottom reflectance type in the background of the target is a seagrass or vegetative bottom type. A
resulting synthetic image is presented from the clear water simulation (lower left) and the zoomed
region of the panel for the target in this area (lower middle) shows the influence of the water surface
and clear water column properties on the target. The lower right zoomed line target area shows the
influence of wind and water quality with suspended matter of ≈ 30 mgL 2 , 20 mgL 2 C as dissolved
organic matter (DOM) and chlorophyll-a pigment concentration of 20 ugL 2 .

Also, similar sensor simulation results as shown above demonstrates the above water
surface detection of an object will be influenced by the scattering and absorption effects
within the atmosphere if the sensor simulation is carried aboard an airborne sensor. The
atmosphere will decrease the contrast observed of a submerged target further than the
results shown above. Therefore detection of UXO and mine like objects and debris in
shallow waters would be best to utilize sensors flown at a low altitude (less than 3,000
meters), preferably lower, or from sensors used aboard vessels as described above. The
models developed by Bostater, et. al., 2003 and 2004 also demonstrate the effects of
subsurface imaging results from the model simulations by examining the upwelling EM
energy from just below the water surface, without the water wave effects. The lower
middle and left panels shown above result from simulations after image enhancements
have been performed upon the synthetic images in the zoomed region. Image algorithm
detection modeling may improve the results for detection, as shown in Figure 5 above.
II.d. Detection of Objects in Using Other Sensor Systems
The above sensor model results and simulations were conducted in order to demonstrate
the need for these types of sensor simulations in order to develop risk reduction
information regarding detection of objects in water. These types of modeling results are
useful not only in sensor and platform design considerations but in training applications
of future engineers, scientists, and command & control operations personnel. All of these
types of professionals involved in the detection of dangerous objects need to work
together and share a common scientific understanding for successful preparedness and
detection of dangerous objects in ambient environments. In the following discussion are
selected results of applications of sensors and robotic systems with potential use for
detection of objects in shallow water types. Figure 10 shows three different robotic
systems designed and built at Florida Tech. The upper left system shows the application
of a semi-autonomous underwater vehicle used to collect water samples in ice covered
regions. The middle panel below shows a system constructed for use as a camera
inspection system autonomous vehicle. The lower right system shows the application of s
surf zone or littoral zone robotic crawler type system. These systems are just examples of
possible future vessels that could be outfitted with more modern remote sensing systems
for object detection applications.
Figure 10. Examples of a subsurface tethered water semi-autonomous underwater vehicle used in sea
ice cover conditions and developed at Florida Tech. The middle image above shows the near surface
autonomous vehicle which was developed to move around shallow water environment applications.

The last image above on the right shows a littoral or surf zone crawling robotic system
for entering the high energy surf zone or beach area. Additional examples of imagery
obtained from remote sensing systems include sonar images as well as robotic systems
for roving in shallow waters are shown below.
Figure 11. Example of a crawling underwater robot with an active-passive acoustic (SONAR)
system. The middle image is an object identified with an acoustic system (see http://marinesonic.com)
reported by Scott and Wilcox (undated). The image on the right shows the results of an image
derived from a small vessel carrying ground penetrating radar from a shallow freshwater study
reported by Versteeg and White (undated manuscript).
Many other examples exist of remote sensing systems and resulting data for detecting
objects and bottom characteristics in shallow water environments. The above examples of
robotic systems and sensing systems are described to demonstrate that technological
advances are developing many possible systems for humanitarian applications regarding
UXO and demining applications as well as applications and related training for homeland
security and preparedness activities as well as for environmental monitoring needs.
III. Summary & Conclusions
Remote sensing systems and robotic systems utilizing advanced mechatronics
concepts hold a unique capability and thus require careful evaluation and training for
future applications and activities related to detection of mines and UXO in aquatic
environments. The above discussion summarizes a practical approach for conceptualizing
the design of such integrated waterborne sensing systems. We have presented an example
of how to model and simulate the phenomenology related to detection of targets and
objects in a shallow coastal inlet, with application of the results to sensor design, and
sensor data analysis or algorithm development. Examples of existing, under construction,
and futuristic remote sensing systems and robotic systems have been described,
We have also described those types of systems which have great potential for
future detection, removal, deterrence and preparedness activities related to securing risk
free and safe aquatic environments. The systems also have a dual use utility in
environmental monitoring in aquatic environments. All applications described above

need to always consider the sensor system requirements in terms of spatial, temporal,
spectral, radiometric and digital resolution in order to be successful.
.IV. Acknowledgements
The research and applications described in this paper have been supported in part
by the following organizations: KB Sciences; Northrop Grumman Corporation,
Melbourne, Florida; NASA; S&C Services; The Link Foundation; US Department of
State. Graduate students Teddy Ghir and Luce Bassetti are acknowledged. Teddy is
acknowledged for putting on paper the designs we discussed and Luce is acknowledged
for running the model runs requested to demonstrate the use of our radiative transfer
models in the context described in this paper.
V. References
Bostater, C., 2000, “Buoy Instrumented for Spectral Measurement of Water Quality”,
NASA Tech Briefs, Vol. 24, No. 11, p. 69.
Bostater, C., et al., 2004, Synthetic Image Generation Using an Iterative Layered
Radiative Transfer Model With Realistic Water Waves, SPIE, Vol. 5233, pp. 253-268.
Bostater, et al., 2004, Developing & Testing a Pushbroom Camera Motion Control
System: Using a Lidar-Based Streak Tube Camera for Studying the influence of Water
Waves on Underwater Light Structure Detection, SPIE Vol. 5233, pp.1-16.
Donald M. Scott, D., Wilcox, T., Side Scan Sonar Suitable for AUV Applications,
http://www.marinesonic.com/documents/SuitableAUVApplications.PDF,
Marine Sonic Technology, Ltd., 5508 George Washington Memorial Highway, P.O. Box
730, White Marsh, Virginia 23183-0730 USA, 4 pp.
Versteeg, R., White, E., Rittger, K., (undated), Ground-Penetrating Radar and Swept-
Frequency Seismic Imaging Of Shallow Water Sediments In The Hudson River
http://water.usgs.gov/ogw/bgas/publications/SAGEEP01_133/SAGEEP01_133.pdf, 11
pp.
Meyers, R., Smith, D., 1998, Development of an underwater GPR system. In: Seventh
International Conference On Ground-penetrating Radar. University of Kansas, Lawrence,
Kansas, USA: Radar Systems and Remote Sensing Laboratory, University of Kansas,
2291 Irving Hill Road, Lawrence, KS 66045-2969,USA.

Charles R. Bostater Jr. is Associate Professor in Physical Oceanography & Environmental Sciences,
College of Engineering, at the Florida Institute of Technology. He is the Director of the Marine
Environmental Optics Laboratory and the Remote Sensing Center. His research interests include remote
sensing systems, remote sensing platforms, shallow water research, coupled modeling of marine
environmental systems and sensor system modeling. Dr. Bostater is the named inventor on four
international patents concerning scalable non-contact remote sensing probes. He has won awards from
NASA for outstanding research and serving in the role of a Principal Investigator, he has managed over
four million dollars in research grants & contracts.
bostater@probe.ocn.fit.edu

Air ground robotics ensembles for risky applications
Simon Lacroix and Raja Chatila
LAAS-CNRS
7, Av. du Colonel Roche
31077 Toulouse Cedex 4 France
Simon.Lacroix@laas.fr, Raja.Chatila@laas.fr
Abstract
Numerous projects deal with the development of aerial robots, aiming at endowing them with
the functional and decisional abilities required by the autonomous execution of exploration
or monitoring missions. The cooperation of such robots with robots operating on the ground
is very relevant in various risky applications. In this paper, the development of cooperation
schemes in which the complementarity of air and ground robots is exploited to achieve
missions in risky contexts is considered. The potential advantages of air/ground robotics
ensembles are discussed. Some research directions are sketched, and preliminary results
obtained in the context of projects in which LAAS is implied are briefly presented.
1 Introduction
These last years, the robotics community has paid more and more attention to the development
of aerial robots (see e.g. [3, 14, 16, 15, 9]). As opposed to drones that executes
pre-programmed mission, aerial robots exhibits decisional autonomy: they are supposed
to achieve high level missions, such as mapping or monitoring a given area, with little human
interactions. Aerial robots rises a wide number of robotics research areas, that goes
from the study of innovative flying concepts and the associated flight control algorithms
(especially for micro drones) to high level mission planning, via real time environment mapping.
In the context of risky applications, aerial robots are very well suited for exploration
and monitoring missions: they can easily gather detailed information on the environment,
without exposing themselves or the operators to any danger.
But gathering data by aerial means might not be sufficient in most of the risky applications:
the returned information are not always complete, especially in cluttered urban
areas, some sensors can only be effective when they are close to the ground, and aerial
robots can hardly physically intervene on the environment. Actually, in all the applications
contexts where the development of exploration and intervention robotics is considered,
air/ground robotic ensembles bring forth several opportunities from an operational point

of view. Be it environment monitoring and surveillance, demining or reconnaissance, a
variety of scenarios can be envisaged. For instance, aircrafts can operate in a preliminary
phase, in order to gather informations that will later be used in both mission planning and
execution for the ground vehicles. But one can also foresee cooperative scenarios, where
aircrafts would support ground robots with communication links and global informations,
or where both kinds of machines would cooperate to fulfill a given mission.
In this paper, we discuss the interest of developing air/ground robotics ensembles in
risky application contexts, and we outline the main research issues raised by such systems.
The paper is organized as follows: the next section describes the main missions for which
robots can be employed in the context of risky applications, and describes the benefits
air/ground robotics ensembles can bring for their achievements. Section 3 presents the
main research issues to tackle in order to develop effective air/ground robotics ensembles,
and section sec—results depicts some current work and preliminary results obtained at
LAAS.
2 Robotic missions in risky applications
Whatever the actual application context is (fire monitoring, fire fighting, mine detection,
demining, . . . ), the functions that can benefit from robotised systems in risky applications
can be summarized as the following:
• Exploration or reconnaissance. This function consists in gathering and structuring
data related to a given area, in order to build a map of the environment. The term
“map” is here understood in its most general acception: a map can be a realistic 3D
model of the environment dedicated to the operators, or a simple representation that
exhibits traversable areas, or a representation in which the position of particular
elements (e.g. mines, fire alarms, chemical hazards, . . . ) is stored. Exploration
essentially relies on perception functionalities, but decisional processes are of course
also involved when it is performed autonomously.
• Monitoring or surveillance. This consists in checking that any specified event does not
occur in a given area, or in continuously assessing the evolution of given events (e.g.
monitoring active fires). This function also relies on perception functions and calls
for decisional processes. It is quite similar to the mapping function, except that time
plays an important role: monitoring an area requires that it is regularly perceived,
and monitoring a given event requires a continuous perception of the event.
• Intervention. The exploration and monitoring functions are passive, in the sense that
they do not modify the environment. On the contrary, by “intervention” we mean
all the functions that actually modify the environment, e.g fire extinction or mine
neutralization.
These three basic functions can be declined and assembled in various mission scenarios,
depending on the application context. For instance, “Search and rescue” missions calls for

an exploration phase (finding the position of victims) and an intervention phase (reaching
the victims to provide them with assistance).
If each of these generic functions can be achieved by a single robot, be it terrestrial or
aerial, it can be more efficiently achieved by an ensemble of aerial and ground robots, that
exhibits a wider spectrum of complementary capacities. Various scenarios that exhibit
a synergy between the two kinds of robots can be foreseen. For instance, for mapping
and monitoring functions, the complementarity of aerial and ground robots from the point
of view of perception is quite obvious: aerial robots can provide a global view of the
environment, while ground robots can exhibit more detailed information. Typically, an
aerial robot can build hypotheses on the presence of particular elements, that are afterwards
confirmed by a ground robot equipped with more reliable sensors.
3 Air ground robotics ensembles: research issues
3.1 Some cooperation schemes
Several cooperation schemes can be foreseen between aerial and ground robots, from simple
assistance to actual tight on-line cooperation.
• Aerial robots assist ground robots, by providing them information on the environment.
The aerial robots build beforehand a map of the environment (either a 3D
representation, a traversability map or a map that exhibits the presence of given elements),
that is used afterwards by the ground robots as an initial information. Such
schemes can be envisaged for mapping missions, as the robots operate in sequence
(i.e. the aerial robots are used in a missions preparation phase).
• Aerial robots assist ground robots, by providing them a communication link with the
operator station. This assistance scheme requires an on line operation of both kinds
of robots.
• Aerial robots assists ground robots, by estimating their localization with respect to
the environment. Ground robot localization is a key problem that is essential to
solve: aerial views of the ground robots can help to estimate their absolute position
- this scheme also requires an on line operation of both kinds of robots.
• In on line cooperation schemes, both kinds of robots cooperate on line to fulfill one of
the three functions mentioned above. For instance, the map being built by the aerial
robot is exploited to drive the ground robot, that actively completes it by perceiving
the areas not seen by the aerial robot. Such schemes are the most complex to achieve,
as it requires synchronizations and coordinations between the two kinds of robots.
To achieve any of these schemes, some robotics problems have to be tackled, either
from the functional and decisional point of view.

Figure 1: A schematic view of the “multi-source” environment information integration problem among
air and ground devices.
3.2 Air/ground cooperative functionalities
Cooperative mapping. One of the most important issue to address to foster the development
of such ensembles is the building and management of common environment
representations using data provided by all possible sources. Indeed, not only each kind of
machine (terrestrial, aerial) can benefit from the information gathered by the other, but
also, in order to plan and execute cooperative or coordinated actions, both kinds of machines
must be able to build and manipulate shared coherent environment representations.
Basically, the problem is to integrate the information acquired from the rovers and
aircrafts into a global coherent representation (figure 1). For that purpose, one must define
algorithms and representations that support the following characteristics of the data:
• Data types: the acquired data can be images, either pan-chromatic or color, from
which geometric 3D informations can be recovered, or directly 3D, as provided by a
SAR or a LIDAR for instance.
• Data resolution: the resolution of the gathered information can significantly change,
depending on whether it has been acquired by a ground or an aerial sensor.
• Uncertainties: similarly, there are several orders of magnitude of variation on data
uncertainties between ground an aerial data.

• Viewpoints changes: beside the resolution and uncertainties properties of the sensors
that can influence the detection of specific environment features, the difference of
viewpoints between ground and aerial sensors generates occlusions that considerably
change the effectively perceived area, and therefore the detectable features.
Integrating all these data is a multi-sensor fusion problem. Depending on the considered
context, the difficulties may vary: for instance, for aerial sensors, occlusions caused by
overheads are likely to occur because of the presence of vegetation or buildings that makes
some area unperceivable from air. Also, the aircraft altitude of course strongly influences
the properties of the perceived data in terms of precision and resolution.
The localization problem. But before tackling the problem of integrating ground and
aerial data into a global consistent environment representation, one must consider the
registration issue: in order to merge these informations, a way to establish the spatial
correspondence between the two kinds of data is required. This would be straightforwardly
solved if sensor positions were always precisely known, which is far from being a realistic
hypothesis. Designing algorithms that solve the registration problem is therefore a key
prerequisite to address cooperative environment modeling. Moreover, such algorithms can
help to localize the rover with respect to initial maps of the environment, as provided by
an orbiter for instance 1 .
Localization is also an essential problem to tackle, even if the two kinds of robots do
not cooperate to build a complete environment map. Besides solutions developed in single
robot contexts, the possibility to use aerial views to localize the ground robots with respect
to their environment is also a solution to investigate.
3.3 System architecture and decisional issues
Multi-robot applications calls for the development of a specific system architecture, that
embrace more concerns than a single robot architecture: in particular, designing multirobot
architectures requires to define the decision making scheme and to specify the interaction
framework among the different robots of the system - which in turn influences the
design of the individual robot architecture.
Within a multi-robot system, the “decision” encompasses several notions:
• Supervision and execution: The executive is a passive, reactive management
of tasks execution, whereas supervision is an active process that manages all the
decisional activities of the robot, whatever their extent.
• Coordination: It ensures the consistence of the activities within a group of robots.
It defines the mechanisms dedicated to avoid or solve possible resource conflicts that
may arise during operational activities. This is especially related to trajectories and
multi-robot cooperative tasks execution (e.g. simultaneous perception of the same
target by several robots).
1 A problem often referred to as the drop-off problem

• Mission planning and scheduling: These decisional activities are dedicated to
plan building, taking into account on one side models of missions and tasks, and on
the other side models of the world: robot’s perception and motion abilities, current
knowledge related to the environment, etc.
• Task allocation: This deals with the way to distribute tasks among the robots. It
requires to establish a task assignment protocol in the system, and to define some
metrics to assess the relevance of assigning given tasks to such or such robot.
These decisional components can be implemented according to different configurations:
they can be gathered within a Central Decisional Node, or be partially (or even totally)
distributed among the robots.
But one of the most important feature of the system architecture is the consideration
of the operators, that should be able to intervene anytime and at any level in the system.
Indeed, in an actual operational context the robots are always operating under human
requests - and operators should have the possibility to choose the robots autonomy level,
depending on the current execution context.
4 Current work at LAAS
We briefly present here some work that is held at LAAS in the context of air/ground
robotics ensembles.
4.1 The autonomous airship Karma
To tackle the various issues raised by the deployment of heterogeneous autonomous systems,
in the context of exploration, surveillance and intervention missions, we initiated the
development of an autonomous blimp project. The ever on-going developments in a wide
spectrum of technologies, ranging from actuator, sensors and computing devices to energy
and materials will ensure lighter than air machines a promising future. There is undoubtly
a regain of interest in this domain, as shown by the recent industrial developments on heavy
loads transportation projects, and on stratospheric telecommunication platforms. As for
small-size unmanned radio-controlled models, which size is of the order of a few tens of
cubic meters, their domain of operation is currently essentially restrained to advertising or
aerial photography. But their properties makes them a very suitable support to develop
heterogeneous air/ground robotics systems: they are easy to operate, they can safely fly
at very low altitudes (down to a few meters), and especially their dynamics is comparable
with the ground rovers dynamics, as they can hover a long time over a particular area,
while being able to fly at a few tens of kilometers per hour, still consuming little energy.
Their main and sole enemy is the wind (see [7] for a detailed and convincing review of
the pros and cons of small size airships with regards to helicopters and planes). Let’s also
note that some specific applications of unmanned blimps are more and more seriously considered
throughout the world, as shown by numerous contributions in the AIAA Lighter

Than Air conferences and European Airship Conventions for instance [1, 2]. In the context
of demining applications, the Mineseeker project is of course worth to mention [5].
Besides long-term developments related to the coordination and cooperation of heterogeneous
air/ground robots, our research work on autonomous blimps is currently twofold:
we concentrate on the navigation problem on the base of automatic control, and on environment
modeling issues using low altitude imagery. For that purpose, we have designed
an realized Karma, a 9.5 m long blimp (figure 2).
Figure 2: The autonomous airship Karma. Note the two cameras mounted on the front and rear of the
gondola.
4.2 The Comets project
Comets is a EEC funded project, that involves various academic and industrial partners
(see [12] for more information). Although it only deals with UAVs, it is worth to mention
here as the objective of COMETS is to design and implement a distributed control
system for cooperative activities using heterogeneous UAVs, in the context of forest fire
applications. Particularly, both helicopters and airships are considered in COMETS. In
order to achieve this general objective, a new control architecture has been designed and
implemented. COMETS also involves cooperative environment perception including fire
detection and monitoring, and terrain mapping.
The heterogeneity of the UAVs is twofold. On one hand, complementary platforms are
considered: helicopters have high maneuverability and hovering ability, and are therefore
suited to agile target tracking tasks and inspection and monitoring tasks that require
to maintain a position and to obtain detailed views. Having much less maneuverability,
airships can be used to provide global views or to act as communications relay - they
also offer a graceful degradation in case of failures. On the other hand, the considered
UAVs are also heterogeneous in terms of on board processing capabilities, ranging from
fully autonomous aerial systems to conventional radio controlled systems with minimal
on-board capabilities required to record and transmit information. Thus, the planning,

Figure 3: The global architecture of the Comets system.
perception and control functionalities of the UAVs can be either implemented on-board
the vehicles, or on ground stations for the low cost and light aerial vehicles without enough
on-board processing capabilities. The global architecture of the COMETS system is shown
in figure 3. It involves remotely piloted UAVs and UAVs with on-board controllers that
have the ability to perform planning and perception activities [8]. Both kinds of UAVs are
equipped with radio transmitter and receivers in order to communicate with the ground
Control Center and with other UAVs.
4.3 The Aerob project
This project is supported by the French national research institute (CNRS), in the context
of the Robea research program [13]. Aerob essentially deals with the building of environment
maps that integrates data acquired by aerial and terrestrial means: we present here
some results of terrain mapping with low altitude imagery.
Besides 3D data acquisition, the main difficulty to build a terrain model that gathers
a set of data acquired during motion is to have a precise estimation of the sensor position:
if this position is not precisely known, the built model is eventually distorted and contains
discrepancies. A huge amount of work has of course been devoted to the localization
problem in robotics. In the absence of any external absolute reference, the only way to
guaranty a sound position estimate during motions is to rely on environment features, that
are detected and localized as they are perceived by the robot. This approach is known as
“simultaneous localization and mapping”(“SLAM” - see e.g. [6, 17]): the robot position
is concurrently estimated with the position of landmarks that are detected by the robot
exteroceptive sensor.
Most of the achievements of SLAM are made in the context of indoor robots evolving
on a 2D ground, that perceive the environment with a laser range finder that is scanned
along a plane parallel to the ground. Recently, various contributions that deal with robots

evolving in the three dimensions and that use vision to detect and map landmarks have
been proposed [4].
We developed an approach that uses only stereovision to estimates both the robot
motions (prediction) and the landmark positions (observation) [11]. Landmarks are interest
points, i.e. visual features that can be matched when perceived from various positions,
and whose 3D coordinates are provided by stereovision. We use an extended Kalman filter
(EKF) as the recursive filter: the state vector of the EKF is the concatenation of the
stereo bench position (6 parameters) and the landmark’s positions (3 parameters for each
landmark). The key algorithm that allows both motion estimation between consecutive
stereovision frames (prediction) and the observation and matching of landmarks (data
association) is a robust interest point matching algorithm [10].
Figure 4: Interests points matched in two aerial images taken from different positions. All the detected
points are denoted with red crosses - the ones that have been matched are surrounded by a square. The
green squares indicates the points matches that are used to estimate the motion between the two positions,
the blue ones are the points that are memorized as landmarks
.
With this approach, it is possible to estimate the six airship position parameters on line
as it flies, with an accuracy on the translations that is as good as 0.1%. Thanks to these
precise position estimates, a digital elevation map is incrementally built by merging the
3D points provided by dense stereovision every time a stereoscopic image pair is acquired.
Figures 5 and 6 show the 5 cm resolution digital terrain maps built after a flight of
Karma over the parking of the lab at an altitude of about 30 m, with a 2.2 m stereovision
bench.
4.4 The Acrobate project
Acrobate (“Algorithmes pour la Coopération entre Robots Terrestres et Aériens”) is an
other project founded by the CNRS Robea program, initiated in late 2003, that explicitly
tackles air/ground ensembles issues. Both functional and decisional aspects related to
this context are considered. Up to now, most of the work has consisted in specifying a

Figure 5: A DTM computed with 240 stereoscopic image pairs: orthoimage and 3D view of the bottom-left
area. The map covers an area of about 3500 m 2 .
simulator that will allow to evaluate distributed mapping and decisional algorithms - no
tangible results can yet be presented here.
5 Conclusions
We sketched in this paper the potential benefits of air/ground robotics ensembles in the
context of risky applications, and presented some preliminary work achieved at LAAS.
Such systems are now becoming to be seriously considered in the robotics community, and
it appears clearly that they will mobilize more and more attention in the near future, as
they definitely bring a lot of capabilities in wide spectrum of application contexts.
Among the research issues to be tackled, it seems that the overall system architecture
and the associated decisional functionalities are the most demanding. In particular, the role
of the operators calls for the definition of a versatile, open and reconfigurable architecture,
that allows to share the decisions at any level in the system.
References
[1] AIAA. 14th Lighter-Than-Air Convention and Exhibition, Akron, OH (USA), July 2001.

Figure 6: A DTM computed with 400 images stereoscopic image pairs. The map covers an area of about
6000 m 2 .
[2] The Airship Association. 4th International Airship Convention and Exhibition, July 2002.

[3] O. Amidi, T. Kanade, and R. Miller. Vision-based autonomous helicopter research at carnegie
mellon robotics institue - 1991/1997. In Heli Japan ’98. Gifu (Japan), April 1998.
[4] A. Davison and. Real-time simultaneous localisation and mapping with a single camera. In
IEEE International Conference on Computer Vision, Nice (France), pages 1403–1410, Oct.
2003.
[5] S. Christoforato and P.K. Bishop. Mineseeker deployment to kosovo for mine survey. In 14th
AIAA Lighter-Than-Air Conference and Exhibition, Akron, Ohio (USA), July 2001.
[6] G. Dissanayake, P. M. Newman, H-F. Durrant-Whyte, S. Clark, and M. Csorba. A solution
to the simultaneous localization and map building (slam) problem. IEEE Transaction on
Robotic and Automation, 17(3):229–241, May 2001.
[7] A. Elfes, S.S. Bueno, M. Bergerman, J.G. Ramos, and S.B Varella Gomes. Project AURORA:
development of an autonomous unmanned remote monitoring robotic airship. Journal of the
Brazilian Computer Society, 4(3):70–78, April 1998.
[8] J. Gancet and S. Lacroix. Embedding heterogeneous levels of decisional autonomy in multirobot
systems. In 7th International Symposium on Distributed Autonomous Robotic Systems,
Toulouse (France), June 2004.
[9] E. Hygounenc, I-K. Jung, P. Soueres, and S. Lacroix. The autonomous blimp project at
LAAS/CNRS: achievements in flight control and terrain mapping. to appear in International
Journal of Robotics Research, 2004.
[10] I-K. Jung and S. Lacroix. A robust interest point matching algorithm. In 8th International
Conference on Computer Vision, Vancouver (Canada), July 2001.
[11] I-K. Jung and S. Lacroix. High resolution terrain mapping using low altitude aerial stereo
imagery. In International Conference on Computer Vision, Nice (France)), Oct 2003.
[12] The official Comets project site. http://www.comets-uavs.org.
[13] The CNRS Robea program. http://www.laas.fr/robea.
[14] Georgia Tech Aerial Robotics. http://controls.ae.gatech.edu/gtar.
[15] S. Saripalli, D.J. Naffin, and G.S. Sukhatme. Multi-Robot Systems: From Swarms to Intelligent
Automata, Proceedings of the First International Workshop on Multi-Robot Systems,
chapter Autonomous Flying Vehicle Research at the University of Southern California, pages
73–82. Kluwer Academic Publishers, 2002.
[16] S. Sukkarieh, E. Nettleton, J-H. Kim, M. Ridley, A. Gvktogan, and H.F. Durrant-Whyte.
The anser project. International Journal of Robotics Research, 22(7-8):505–540, 2003.
[17] S. Thrun, D. Fox, and W. Burgard. A probabilistic approach to concurrent mapping and
localization for mobile robots. Autonomous Robots, 5:253–271, 1998.

Light Hybrid Robotic Platform for Humanitarian Demining
N. Amati, B. Bona, A. Canova, S. Carabelli, M. Chiaberge, G. Genta
Laboratorio Interdipartimentale di Meccatronica (www.lim.polito.it)
Politecnico di Torino – Italy
Introduction
The combination of legs and wheels, so-called hybrid solution, seems to be a promising
solution to move in a number of environment where purely wheeled machines may be
ineffective and humanoid robots are too complex (and costly).
Hybrid and light robots should move autonomously, they should communicate with other
within a fleet and with a fixed base, they should know their exact position, they should
allow remote teleoperation of their specific equipment (modular payload).
In order to keep their cost as low as possible, the concept is to use a simple and reliable
mechanical construction based on two frames, six legs and four/six wheel, to be driven by
compact electric motors and controlled by a rather general purpose control unit based on
available digital platforms.
Walking robots for space applications are a long term project at Laboratorio
Interdipartimentale di Meccatronica (LIM) that resulted in a number of working
prototypes ([4], [5], [7]).
Digital platforms for mechatronic applications are another long term project of LIM that
resulted in a number of working applications.
A number of other basic technologies are to be combined in order to build an effective
mobile platform, namely wireless communication and positioning but these are
increasingly available as modules that can be easily integrated into the digital platform
controlling the machine.
Work-in-Progress
Vehicle architecture
Among the many configurations of moving robots suggested in the past ([1], [2], [3], [6]),
a number had both wheels and legs or appendages which rotate like wheels and are
conformed like legs (or even some hybrid devices wheel-track-leg). Very seldom they
were actually built and extensively tested and often were discarded in favour of more
zoomorphic legged configurations or standard wheels ([9]). However, in many
applications as humanitarian deminig, it could be of advantage to have wheels to travel
on level ground and legs to deal with obstacles, both from the viewpoint of the average
speed and also to increase the reliability of the machine. The leg mechanisms, which are
usually highly stressed and are critical for fatigue and wear, are required to operate only
when needed, while wheels supply mobility in easier conditions.

Actually there is a wide range of solutions which combine wheels with levers to increase
mobility. The most conventional types are vehicles (military vehicles or machines for
open air mining, construction works, etc.) in which regular wheels are suspended using
long-stroke trailing arm suspensions, which can be active or at least supplied with loadlevelling
devices. They are here considered as wheeled vehicles. The suspensions can be
made by two articulated levers, more similar to a leg than to a trailing arm. Here the
ability to walk of what is essentially a wheeled vehicle depends on the actuators and the
control system used.
On the other side there are walking machines with small wheels either attached at the end
of the legs ([10]) or under the body ([8]), which can be put on the ground by raising the
legs. Wheels at the end of the legs have the advantage of allowing the body of the vehicle
to ride high on the ground, clear from obstacles, but either require that the legs behave as
active suspensions or that some sort of elastic and damped suspension is placed in the
feet – except if no suspension at all is used, a thing possible only for very low speeds.
Wheels under the body make it easier to use a more or less conventional suspension, but
the body rides very low and the legs must be able to “get out of the way” in a somewhat
unnatural position, which in some cases, particularly when a mammalian configuration is
used, can be impossible at all.
Generally speaking, to perform well as a walking machine the device must have feet as
light as possible, so the use of large wheels at the end of the legs indicate that walking
has been considered as an auxiliary locomotion method to allow a wheeled vehicle to
extend its capabilities to very rough ground. On the contrary, very small wheels show that
the primary locomotion mode is walking. In any case, the presence of wheels not only
allows a walking machine to increase its speed on level ground and to be operated in a
simpler way, but also allows a limited performance in case of failure of some actuators or
of the walking control system.
A hybrid robot with wheels at the feet can perform in the “wheeled” or in the “legged”
modes, the first one being better suited to smooth ground and the second one to rough
surfaces, but it can also operate in a mixed mode. The legs supporting the robot have
locked wheels and behave as legs. The advancing legs are not raised from the ground, but
roll on it, maintaining a certain support and perhaps supplying a certain traction. The
stability of the vehicle is so much improved and a fully static stability can be obtained
also in the case of a quadruped machine with a high value of the duty factor, at the
expense of added complexities in the control system.
Rigid frames machines are particularly suitable for building wheel-leg hybrid vehicles.
The wheels can be located under the body, all on one of the two frames (and then a
steering system is required) or two on one frame and two (or one) on the other, so that
steering is provided by the rotation of the frames with respect to each other. The wheels
can be easily provided of a suspension system, but some of the throw of the legs is lost.
As an alternative, the wheels can be located under some of the legs. In the case of an
hexapod similar to WALKIE 6 (Figure 1, Figure 2) the wheels can be located under two
legs of one frame and two of the other one, so that no purposely designed steering system
is required.

Figure 1: Mechanical subsystem of Walkie 6.2 prototype.
Figure 2: Walkie 6.2 walking on outdoor rough terrain. Mount Etna, September 2002.

However, either the speed is very low and the legs act as a load-levelling suspension, or
some compliant connection is required between the feet and the wheels. For low speed
operation on fairly level ground the suspension system can be dispensed of even without
using the legs as an active suspension system if three wheels are used. Since there are
many possible layouts (number and location of wheels, presence of a dedicated
suspension and steering system, control strategy of the body and legs while rolling), a
vehicle of this type needs to be designed following its tasks, the type of terrain expected
and other operational constraints.
Digital platform
The objective of combining an embedded actuator control unit with enough software and
firmware power to seemless integrate communication, positioning and payload needs can
be achieved by means of a modular digital platform featuring a processing unit and a
logic device.
The processing unit is based on a combination of a general purpose processor (e.g.
ARM9), with integrated communication devices, (802.11), and a digital signal processor
(DSP) with mixed signals I/O (AD and PWM drives). The general purpose processor
should run a real-time operating system, for instance Linux based RTAI.
The presence of Field Programmable Logic Device (FPGA) allows a substantial freedom
for the later integration and management of digital devices needed by the system or its
payload, e.g. a stereoscopic vision subsystem..
The proposed digital platform is to be intended as a prototyping system that can be used
in real terrain conditions without the need to go through an full engineering phase, at least
for small series production.
The use of Open Source software as well as firmware and hardware is extensively
adopted in order to use what is already available and, at the same time, leave the project
fully open to specific modifications and adaptations.
Wireless communication and positioning
A wireless communication system is thought to be a key feature for almost any
application of mobile robotics. A solution based on a widely known standard such as
IEEE 802.11 should lead to a low cost solution that makes use of available hardware
devices and software drivers and applications.
The relatively short range, less than 1 kilometer in open field, may be conveniently
extended by means of multihop connections within a fleet of mobile robots that “keep in
touch” with a central operating center via an ad hoc network. During the search phase of
operation data streaming is used to convey detection sensor information to the central
unit to be added to the location map. On the other hand, during the actual demining phase
teleoperation may be needed; transport protocol for real-time communication, such as
Real-Time Protocol (RTP), is to be adopted.
The combination of autonomous moving capabilities and positioning on field is another
topic of current research. Global Positioning System (GPS) may be used to locate the

mobile robot on a map in order to produce a resonable reference route and leaving to its
autonomous capability to deal with local obstacles.
Vision
The stereoscopic configuration of the proposed vision system helps to find distance of
possible “visible” obstacles on the trajectory of the rover. It is composed of two CMOS
cameras that are interfaced with the digital platform to perform camera handling and
configuration on FPGA and image management, analysis and compression on the DSP.
The vision system algorithm is based on the principle that the two images acquired by the
two CMOS sensors are very similar (the distance between the two CMOS sensors will be
10cm). This similarity is used to optimize the compression (scientific stereoscopic
images) and for distance extraction of visible and near obstacles.
This approach is costly effective and simplify the whole control architecture of the
mobile robot introducing the vision system as a payload to be used also for navigation
and obstacle avoiding tasks.
Group cooperation
It is agreed that a number of complex tasks in mobile robotics is better carried out by a
team of several simple mobile platforms than by a single complex unit. Since the various
tasks involved in humanitarian demining depend heavily on the capacity of thoroughly
exploring an unknown mine-filled terrain, and autonomously react to sensor
measurements related to mine threats, a team of autonomous or semi-autonomous mobile
robots is the best choice in this case.
The number of team members may depend on the extension of the area to be searched, as
well as on the accepted level of a-posteriori probability of leaving a mine undiscovered.
A team can in principle be composed by homogeneous or non-homogeneous robots: in
the first case, each team member carries the same payload and is functionally
interchangeable with other members, while in the second case there can be different
specialized robots: for example, while some robots are devoted to mine recognition, one
robot can be in charge of localization and mapping, or take care of the medium range
radio link with the coordinator.
The hybrid platform proposed in this paper suit well the mobility aims of the demining
team, since it can perform both high speed deployment on a wide area, and low speed
tasks, such as obstacle avoidance on difficult terrain or mine identification and
neutralization.
The platform can, in principle, carry homogeneous payloads or dishomogeneous ones,
with minor modifications of its layout and power consumption.
The control system can be partitioned into two levels: a low level behaviour-based
control, assuring fast and prompt action with minimal computing power, and a high-level
layer, performing model-based tasks, such as map reconstruction and safe areas
recognition.

Conclusions
A number of expertize and research topics currently developed at the Mechatronics Lab
at Politecnico di Torino are thought to be well fitted to design and prototype a hybrid
robotic platform to be used in the technical effort toward an effective humanitarian
demining application.
The main objective of the proposed hybrid solution for light mobile robots is make them
available at reasonable cost in order to be applied in a number of demining scenarios both
for the main removal task or as a support to it.
References
[1] D.J. Todd, 1985, Walking Machines: an Introduction to Legged Robots, Kogan
Page Ltd., London.
[2] J. Peabody, H. B. Gurocak, 1998, Design of a Robot that Walks in Any Direction,
Journal of Robotic System, pp. 75-83.
[3] M. E. Roseheim, 1994, Robot Evolution: the Development of Anthrorobotic, Wiley,
New York.
[4] Amati N., Chiaberge M., Genta G., Miranda E., Reyneri L.M, 2000, WALKIE 6–A
Walking Demonstrator for Planetary Exploration, Space Forum, Vol. 5, N° 4 pp.
259-277.
[5] N. Amati, G. Genta, L.M. Reyneri, 2002, Three Rigid Frames Walking Planetary
Rovers: a New Concept, Acta Astronautica, Vol. 50, N°.12, pp. 729-736.
[6] S. M. Song, K. J. Waldron, 1989, Machines that Walk: the Adaptive Suspension
Vehicle, Cambridge-MIT.
[7] G. Genta and N. Amati, 2001, Planar motion hexapod walking machines: a new
configuration, Proceedings of the Fourth Int. Conf. on Climbing and Walking
Robots, London, pp. 619-629.
[8] D.S. Goldin, S.L. Venneri, A.K. Noor, 2000, The great out of the small, Mechanical
Engineering, Vol. 122. N°. 11, pp. 70-79.
[9] G. Genta and N. Amati, 2002, Non-Zoomorphic Versus Zoomorphic Walking
Machines and Robots: a Discussion, European Journal of Mechanical and
Environmental Engineering, Vol. 47. N°. 4, pp.223-237.
[10] A. Halme, I Leppanen, M. Montonen, S. Ylonen, 2001, Robot motion by
simultaneously wheel and leg propulsion, 4 th Int. Conference on Climbing and
Walking Robots, pp. 1013, 1020.

Ground Adaptive Manipulation of GPR for Mine Detection System
Hidenori Yabushita, Kazuhiro Kosuge and Yasuhisa Hirata
Department of Bioengineering and Robotics, Tohoku University,
Aoba-yama 01, Sendai 980-8579, Japan,
email:hidenori,kosuge,hirata@irs.mech.tohoku.ac.jp
Abstract
In this paper, we propose the mine detection system
which consists of the ground penetrating radar
(GPR) and the arm for operation of the GPR. The
GPR is expected as a most effective mine detection
sensor which can detect both metal and plastic mines.
Many researchers research the GPR for detection of
underground conditions. However they have not considered
a rough ground surface for utilizing the GPR.
In actual mine fields, a rough ground surface influences
the GPR signal. The underground information
would be buried in noise influenced by the ground surface.
The manipulation of the GPR along a ground
surface would reduce the effect of the ground surface
and make the precise information of the underground.
The proposed system has high capability for
mine detection.
Key W ords : Ground penetrating radar, Mine detection,
Robotics, Ground scanning
1 Introduction
In more than 70 countries, 120 million anti personnel
mines have been buried[1]. These mines are located
in not only battlefields but also residence areas, community
roads, gardens and so on. It is necessary to
detect land-mines on several situations. The purpose
of the study is to realize the ground adaptive manipulation
of a sensor head on several environmental
situations. The ground adaptive manipulation is a
motion control for improvement the GPR signal.
A GPR is expected as a most effective mine detection
sensor which can detect both metal and plastic
mines. However the GPR is not perfect because
the GPR signal includes unexpected signals such as
ground reflection and clutter reflection. It is very
difficult to eliminate these unexpected signals from
received signal for getting the expected signals from
mines.
Most of research topics concerning a GPR
is a signal processing to extract underground
characteristics[2][3][4]. Those studies try to solve
Figure 1: Concept model of mine detection robot.
above problems and have succeeded. However these
conventional approaches are not enough, because
they only assume a flat ground surface. If we use
the GPR on an actual sensing area, we should consider
the effect of the shape of the ground surface.
Uneven ground surface makes complex reflections. A
ground reflection disturbs to detect mines from the
GPR signal. We could reduce an effect of the ground
reflection by using signal processing, however the signal
processing also affects the underground information.
To reduce an effect of a ground reflection, a
position control of a GPR is important[5].
If we could use the position control of the GPR, we
could keep a relative distance between a ground surface
and the GPR as a constant, therefore a ground
reflection is detected as a constant intensity signal.
We could find characteristics of underground objects
from a received signal. In this paper, the position
control of a GPR which we call as the ground adaptive
manipulation is useful for mine detection.
2 Concept of proposed mine detection
robot system
Robot technologies are expected to be applied to hazardous
fields such as space, disaster sites, nuclear
power plants and so on. Minefields are also danger-

ous area, where we hope to apply robot technologies
to. By applying robot technologies to a demining
process, we could execute dangerous tasks in safety.
In this section, we introduce the concept model of our
mine detection robot. The concept model of mine
detection robot is illustrated in Fig.1. The concept
model has three characteristics. The first characteristic
is a low-pressure tire. The low-pressure tire has
a great flexibility to make a broad contact surface.
A broad contact surface decrease the ground pressure
less than the actuated force of land mines. The
concept robot with these tires realizes a demining
process on undemined fields without exploding the
mines. The driving ability on undemined area enables
multiple robots to execute mine detection in
parallel.
The second characteristic is the small-reaction manipulator
for realizing the high-speed and precise
sensor head maneuver. The high-speed sensing will
reduce a working time and improve the cost effectiveness.
The small-reaction manipulator counteracts
the reaction force of the sensor head motion by
using counter weights. This functionality is important
for the concept robot. The concept robot has
low-ground-pressure tires which are extremely flexible,
therefore the fast sensor head motion causes oscillations
on the concept robot. Oscillations of the
robot impair the accuracy of the sensor position and
oscillations also make the GPR signal complex. A
signal from a buried object is hidden in complex
noise based on oscillations of the robot. By using
the small-reaction manipulator, we could execute the
high-speed and precise detect motion without oscillations.
The small reaction manipulator has been
developed in our laboratory, and the functionality is
already validated in our research [6].
The third characteristic is the ground adaptive manipulation
of a sensor head which is mentioned in the
following sections. The purpose of the ground adaptive
manipulation is to reduce an effect of the ground
reflection. The ground reflection has a large proportion
of the received signal and a complex-ground
reflection hides the signal from a buried object. A
ground adaptive manipulation keeps a distance between
the GPR and the ground surface as a constant
in order to keep a ground reflection as a constant. It
is easy to extract a target signal from a received GPR
signal which has a constant ground reflection.
3 Sensor head maneuver along ground
surface
The ground adaptive manipulation is a sensing motion
which traces a ground surface. By keeping the
distance between the sensor head and the ground surface,
it is possible to reduce the effect of the ground
[m]
Buried object reflection
Effect of ground surface
Figure 2: Cross section diagram of mine field with
GPR
surface and to emphasize the characteristics of buried
objects. The data processing of sensing results also
enables to decrease the effect of the ground surface.
However the data processing also affects target signals,
as a result a target signal will be deformed.
On the other hand the ground adaptive manipulation
could detect underground information without
deforming.
The sensing system uses the GPR for detection of
land mines which are made with not only metal but
also plastic. However, conventional GPR systems
have a defect on its operation, because the position
control of the GPR has not considered scanning conditions.
It is difficult to extract the target signal
from the received signal, because the received signal
consists several complex factors such as target
reflection, ground-surface reflection, and clutter reflection.
The greater part of the received signal is
based on a ground-surface reflection. Especially a
reflected signal from shallow land mines is buried in
a ground-surface reflection. A research topic of a
GPR signal processing is to extract the target signal
from the complex received signal.
Fig.2 shows the cross section diagram of a flat mine
field which is measured by the GPR. From Fig.2 we
can see that the ground surface reflection has a great
intensity in the received signal. Uneven-ground surface
generates a complex signal depending on the
shape of the ground surface. If the sensor head traces
the ground surface, the ground surface reflection generates
a constant intensity. In this case, we could
eliminate the ground reflection and recognize the
target characteristics from the received signal. We
call the tracing a ground surface as a ground adaptive
manipulation. By applying the ground adaptive
manipulation for the GPR system, we could obtain
a clear underground information which emphasize
characteristics of buried objects. From next section,
we show the detail of the proposed mine detection
system.

Figure 3: Laser range finder (LMS 200) for ground
scanning
4 Mine detection system
The ground adaptive manipulation is prospective for
improving the GPR signal by reducing the ground
reflection. The proposed mine detection system consists
of the GPR, the surface capturing system, and
the manipulator. The surface capturing system obtain
the 3D map of the ground surface, and the manipulator
realizes an effective path control of GPR.
Details of these functions are discussed below.
4.1 GPR performance
The GPR is utilized as a land mine detector in the
system. In this subsection, we introduce the GPR
performance briefly. The GPR is a step-frequency
radar which uses from 0 [GHz] microwave to 2 [GHz].
The GPR measures the reflective intensity with each
step-frequency from underground objects. By using
the inverse Fourier transform, we obtain the timedomain
signal from a measured information. A stepfrequency
radar has several advantages compared
with a impulse radar. For example, a time-domain
signal which is measured by the step-frequency radar
is sharper than an impulse radar. A feeding to
the step-frequency radar is smaller than the impulse
radar.
Table 1: Performance of laser range finder
Type LMS 200
Range max 8 [m]
Resolution 10 [mm]
Statistical error ±15[mm]
Scan speed 200 [ms/scan]
Angular range 180[deg]
Angular resolution 0.5[deg]
Blind spot
Laser range finder
Measurable area
Figure 4: Example of ground scanning
Figure 5: 3D captured ground surface
4.2 Ground surface capturing system
In order to operate the GPR along a ground surface,
it is necessary to measure the shape of the ground
surface. We use the laser range finder which is called
as SICK LMS 200. The specification of the laser
range finder is shown in table 1. By using the laser
range finder, we get the distance between the laser
range finder and the ground surface with respect to
a plane. The laser range finder scans the ground
surface from two different view, because a scanning
path from one position has possibility to have blind
spots as shown in Fig.4. To avoid making blind spots
we utilize a superposition of two scanning data which
are measured from two different view.
The superposition of the measured data makes the
3D map of the ground surface. For superposition
of the data it is necessary to collect several slices
of the ground scanning data. To collect the ground
scanning data, the laser range finder is implemented
on the 2nd joint of the manipulator as shown in Fig.3.
The height of the laser range finder is 60 [cm] above
the ground surface.
To crate a 3-D map of a ground surface, we digitize
the scanning data along the x axis and the y axis. In
this study we choose 1 [cm] for the grid interval of

the 3D map. An example of the 3D ground surface
is shown in Fig.5. From Fig.5 we can see that the
map has no blind spot.
The laser range finder is connected to a PC through
RS-232C. The scanning process takes 200[ms] per 1
slice image. For scanning the ground surface which is
700[mm] width and 900[mm] length a total scanning
process takes around 30[sec] with 1[cm] interval. By
applying the 3rd order function to the 3D ground
surface map we could make a sensing path of the
GPR smoothly.
4.3 Mine detection manipulator
The proposed system is designed and developed in
order to realize the ground adaptive manipulation
for a high-performance mine detection. On unevenground
surfaces the ground-adaptive manipulation is
useful to obtain a clear sensing image.
In this research we assume a height of unevenness
of a ground surface is less than 30[cm]. Conventional
mine-detection systems could not be applied
to such situations, because those systems exclude uneven
ground conditions. The manipulator operates a
sensor head along the 3D ground surface map which
is made by the ground surface capturing system.
The developed mine detection manipulator is shown
in Fig.6. The manipulator has 3 degree of freedom.
The movable range of the x axis is 900[mm], y axis
is 700[mm], and z axis is 400[mm] as shown in Fig.7.
The speed of the each axis is the max 0.6 [m/s] for
x axis and y axis, and 0.4 [m/s] for z axis as shown
in table2. These speed of each axis enables a mine
detection widely. Conventional detection area which
is executed by human is about 3-6[m 2 ] per one day.
However the developed manipulator has ability of detection
of 1[m 2 ] with 15 minutes. If we use the proposed
robot for 8 hours, the robot can detect 32[m 2 ].
Each axes of the manipulator are driven by ACmotors
and the ball-screws. Ball-screws realize the
straight movement and precise position control. The
GPR is mounted on the end-effector of the manipulator.
Fig.8 shows a GPR system which we use in
this research. We apply the ground adaptive manipulation
to the manipulator. Fig.10 shows the result
of the motion.
5 Experiments
To evaluate the proposed system, we apply the control
system to the mine detection system. Following
subsections we explain experimental conditions and
show experimental results.
5.1 Experimental conditions
We made a sand pool illustrated in Fig.9 for testing
a proposed mine detection system. The volume of
y
x
Figure 6: Mine detection manipulator
400
900
z
700
[mm]
Figure 7: Movable range of sensor head
the sand pool is 1.0[m]×1.0[m]×0.42[m]. The sand
pool has electromagnetic wave absorbers on each wall
and on the bottom in order to prevent artificial unnecessary
reflection. By using electromagnetic wave
absorbers, the sand pool has a electro-magnetic property
as almost same as an actual field.
For evaluating the proposed system, we made two
ground surface forms as shown in Fig.11. The small
mountain and the small valley has a 10[cm] in vertical.
In this experiments we use two dummy mines as
shown in Fig.12. These dummy mines have a similar
shape and a similar electro-magnetic property as real
landmines. The size of M-14 is 55 [mm] in diameter
and 40 [mm] in height, and the size of TYPE 72 is
77 [mm] in diameter and 40 [mm] in height.
Dummy mines are buried under 10[cm] at the top of
the mountain or at the bottom of the valley. In the
experiments, we scan the detection area at 5[mm]
interval on horizontally. When the ground adaptive
maniplation was applied to the manipulator, the
Table 2: Specification of mine detection manipulator
Range x: 900 [mm]
of y: 700 [mm]
movement z: 400 [mm]
Maximum x: 0.6 [m/s]
working y: 0.6 [m/s]
speed z: 0.4 [m/s]

(a) (b) (c) (d) (e)
Control unit
PC
Antenna
Figure 8: Ground penetrating radar system
Woodboard
Electromagnetic wave absorber
Dry sand
1.0 [m]
Figure 9: Sand pool
Figure 10: Ground adaptive manipulation
0.42 [m]
height of the GPR was kept as as about 4[cm] above
from a ground surface.
5.2 Results of mine detection
In the experiments, we made two type ground surfaces
as a mountain and a valley. In each condition,
we manipulate the GPR with two type operation.
One is the constant height manipulation of the GPR,
and another is the ground adaptive manipulation.
Fig.13 shows the result of mine detection with constant
height of manipulation of the GPR on the small
mountain, and Fig.14 shows the result of mine detection
with ground adaptive manipulation. From these
¢¡¤£ ¥ ¦¤§©¨���� ¦¤���
¢¡¤£ ¥ ¦¨§�©���� ¦¨���
(a) Mountain (b) Valley
Figure 11: Experimental ground surface
(a) M14 (b) TYPE 72
Figure 12: Imitation model of landmine
results we can see that the Fig.13 includes the effect
of ground surface, however Fig.14 shows the constant
intensity of ground reflection. From the Fig.14 we
can find the feature of a landmine easily.
Fig.15 shows the result of mine detection with constant
height on the valley form, and Fig.16 shows
the result of mine detection with the ground adaptive
manipulation. From these results we can see that
the ground adaptive manipulation keeps a constant
intensity of the effect of the ground reflection. From
Fig.16 we can find the feature of a landmine easily
than Fig.15. Through these experimental results, the
proposed mine detection system is validated.
6 Conclusion
In this paper we propose the ground adaptive manipulation
which is realize the high ability of landmine
detection. By using the proposed system, we could
emphasize the features of buried objects by keeping
a constant ground reflection. The experimental results
illustrated the validity of the proposed mine
detection system.

GPR
GPR
GPR
GPR
References
(a) Sensor path (b) Without mine (c) M14 (d) TYPE72
Figure 13: Constant height manipulation on mountain surface
(a) Sensor path (b) Without mine (c) M14 (d) TYPE72
Figure 14: Ground adaptive manipulation on mountain surface
(a) Sensor path (b) Without mine (d) M14 (d) TYPE72
Figure 15: Constant height manipulation on valley surface
(a) Sensor path (b) Without mine (c) M14 (d) TYPE72
Figure 16: Ground adaptive manipulation on valley surface
[1] Jacqueline MacDonald et al.: Alternatives For
LANDMINE DETECTION. RAND, 2003.
[2] Andria can der Merwe, Inder J. Gupta: A Novel Signal
Processing Thechnique for Clutter Reduction in
GPR Measurements of Small, Shallow Land Mines
,IEEE TRANSACTION ON GEOSCIENCE AND
REMOTE SENSING, vol.38, No.6, pp.2627-2637,
2000.
[3] Nada Milisavljevic, Isabrllr Bloch: Sensor Fusion
in Anti-Personnel Mine Detection Using a Two-
Level Belief Function Model, IEEE TRANSACTION
ON SYSTEMS, MAN, AND CYBERNETICS-PART
C:APPLICATIONS AND REVIEWS, vol.33, No.2,
pp.269-283, 2003.
[4] Guangyou Fang, Motoyuki Sato: GPR detection
of landmine by wavelet transform. Proceedings of
the 6th SEGJ International Symposium, pp.449-454,
2003.
[5] Claudio Bruschini, Bertrand Gros, Frederic Guerne,
Pierre-Yves Piece, Oliver Carmona: Ground penetrating
radar and imaging metal detector for antipersonnel
mine detection, Journal of Applied Geophysics,
vol.40, pp.59-71, 1998.
[6] Hidenori Yabushita, Yasushisa Hirata, Kazuhiro Kosuge:
Small Reaction Manipulator for GPR Sensing
Head Maneuver. SSR2003 Proceedings of the First
International Symposium on Systems & Human Science,
pp.339-344,2003.

A SYSTEM FOR MONITORING AND CONTROLLING A CLIMBING AND
WALKING ROBOT FOR LANDSLIDE CONSOLIDATION
Leif Steinicke
David Dal Zot
Thierry Benoist
Space Applications Services
Leuvensesteenweg 325
B-1932 Zaventem
Belgium
www.spaceapplications.com

ROBOCLIMBER HUDEM ‘04
1. INTRODUCTION .................................................................................................................................... 3
2. CLAWAR TECHNOLOGY USED FOR SLOPE/LANDSLIDE CONSOLIDATION...................... 4
2.1 SLOPE CONSOLIDATION SCENARIO........................................................................................................ 4
2.1.1 Setting the anchor ........................................................................................................................ 5
2.1.2 Positioning ................................................................................................................................... 7
2.1.3 Drilling......................................................................................................................................... 8
2.1.4 Iterating ....................................................................................................................................... 9
3. REQUIREMENTS ON ROBOT TECHNOLOGY IN SLOPE/LANDSLIDE CONSOLIDATION
...................................................................................................................................................................... 10
3.1 REQUIREMENTS ON THE ROBOT .......................................................................................................... 10
3.1.1 Moving and positioning ............................................................................................................. 10
3.1.2 Power supply.............................................................................................................................. 10
3.1.3 On-board equipment .................................................................................................................. 11
3.1.4 Other system features................................................................................................................. 11
3.2 REQUIREMENTS ON THE ROBOT HMI.................................................................................................. 11
- High level operations : the operator specifies a depth and the HMI automatically has the robot
drill a hole of the specified depth (high level command). ................................................................... 12
- Medium level operations : the operator controls the robot on a rod by rod basis and sends
commands such as "insert rod", “remove rod” and so on.................................................................. 12
- Low level operations : likely to be directly accessed only for testing and debugging. At this
level, the operator would control the drilling subsystem on a very low level and would send
commands such as “rotate joint x by n degrees". ............................................................................... 12
3.3 ROBOT HMI DESIGN ........................................................................................................................... 13
3.3.1 General considerations.............................................................................................................. 13
3.3.2 Tablet as a command console .................................................................................................... 13
3.3.3 HMI Operator Mode .................................................................................................................. 14
3.3.4 HMI Super-User Mode............................................................................................................... 16
4. CONCLUSION....................................................................................................................................... 16

ROBOCLIMBER HUDEM ‘04
1. Introduction
This paper describes the application of space robotics control technology applied to the
slope/landslide consolidation sector. The description is made from the point of view of
monitoring and controlling a walking/climbing machine used in the Roboclimber project.
Figure 1a.
Figure 1a shows an early prototype of the Roboclimber, used to test the concept,
operating on a near vertical rock face. The process of stabilising a risky slope is initiated
with a geological survey, after which a series of holes are made to consolidate the wall.
The objective of a tele-operated climbing robotic system for the maintenance and
consolidation of mountain slopes is to drill the deep holes for the insertion of 20-metre
long rods. This saves time and operating costs and is less dangerous. Roboclimber is
made by a consortium of European SMEs and builds upon the technology used to control
space robotics on-board satellites or spacecraft such as the International Space Station.
Figure 1b.
Figure 1b shows the robot in an early phase of development without drilling equipment.

ROBOCLIMBER HUDEM ‘04
2. CLAWAR Technology Used for Slope/Landslide
Consolidation
2.1 Slope consolidation scenario
This section describes a typical use of monitoring and control technology in the field of
slope consolidation. Figure 2 below shows a rendering of a slope to be consolidated.
Figure 2. Rendering of a slope to be consolidated
The part of the slope needing consolidation is indicated in light green.
The following describes the different stages involved in consolidation of a slope, using a
climbing and walking robot. The general approach is that the robot climbs down from an
anchoring point at the top of the slope, and then is controlled to walk to a number of
positions on the slope where it will drill rods into the slope. The typical steps are thus:
• setting the anchor on the top of the slope;

ROBOCLIMBER HUDEM ‘04
• climbing down the slope from the anchoring point
• positioning the robot on the slope at a drilling location;
• drilling a rod (or several rods) into the slope at drilling location;
• move to new drilling location and repeat drilling, until all reachable drilling locations,
from the current anchor point, have been consolidated with rods;
• moving the robot back up the slope and walk it to the next anchoring position;
• iterate the process until the entire slope is consolidated with rods.
During the whole process, at all times the monitoring and control system will keep the
operator aware of the situation through a Human-Machine Interface (HMI) providing him
with necessary information about the robot, such as its position, its speed, leg position,
and various relevant telemetry for each particular task.
2.1.1 Setting the anchor
The first step to be completed is for the robot to anchor itself to the top of the slope at the
edge of the zone to be drilled. The robot may feature a so-called tirfor which will allow it
to use two ropes attached to two anchor points to climb down the cliff.
At this stage, the operator can control the robot and move it around so that it can go by
itself to the first designated anchor point, see figure 3.

ROBOCLIMBER HUDEM ‘04
Figure 3. Positioning the robot before anchoring
The robot is positioned on the edge of the drilling zone, and is ready to anchor itself.
Individual drilling sites are shown in red. The operator can give directions to the robot at
a high level and does not have to “worry” about individual components - legs, joints - to
achieve this, though he is capable of controlling the robot at this level if deemed
necessary. When the robot reaches the designated anchor point, the operator orders the
robot to anchor itself. The robot will in effect drill a hole and insert an anchor attached to
the extremity of one of its tirfor ropes.
The operator repeats the process to move and anchor the robot to the second anchor point.
Now that the robot is properly anchored, with its two ropes attached to their
corresponding anchor point as shown in figure 4.
The robot is then ready to start its descent.

ROBOCLIMBER HUDEM ‘04
2.1.2 Positioning
Figure 4. Anchoring
With its ropes secured on the top of the cliff, the robot must now position itself above the
drilling sites. In an ideal drilling strategy, the robot will drill holes some 2 meters apart in
every direction. In the light of stability issues, it is not possible to drill more than
approximately three holes horizontally for each anchor point without making the lateral
deviation too big. Refer to figure 5 a and b.
To achieve this, the operator will send high level commands to the robot to climb down,
up, left or right. The robot will automatically use its tirfors and legs
in a joint effort to manoeuver horizontally and laterally, climb up and down, avoid
obstacles, and so on.

ROBOCLIMBER HUDEM ‘04
Fig 5.a The robot in a stable drilling position.
Fig 5.b The robot at a drilling location, at the limit of its reach envelope.
These figures illustrate the stability issue when moving the robot laterally on the slope.
2.1.3 Drilling
Before the drilling can actually commence, the robot needs to be as stable as possible.
Therefore, the operator will manually fine-tune the position of each leg to ensure proper
contact with the ground. Fine-tuning the position of a leg involves manually trying
different triplets until the telemetry is satisfactory. When
the four legs are properly positioned, the operator specifies a depth and instructs the robot
to start drilling. The robot will automatically load and use drilling rods and will

ROBOCLIMBER HUDEM ‘04
automatically recover them when the specified depth is reached. Of course, should an
anomaly occur, manual control will immediately be teturned to the operator. The operator
will be kept informed via a progress indicator displaying relevant information (drilling
speed, current depth, heuristic analysis from robot, etc.)
2.1.4 Iterating
After the drilling is complete, the robot will be moved to the next drilling site as
explained previously. If the robot has reached the bottom of the drilling zone, the
operator will order it to climb all the way back up to the top of the cliff, back at the
anchor site, using high level commands to climb up and walk the final distance to an
anchor point.
Recovering and setting up an anchor is performed using high level commands. After the
new anchor point has been chosen, the process is then iterated until all the drilling
locations have been processed.

ROBOCLIMBER HUDEM ‘04
3. Requirements on Robot Technology in
Slope/Landslide Consolidation
From the typical slope or landslide consolidation scenario described in the previous
section, a number of requirements can be derived for the robot technology applied in this
sector, and the associated monitoring and control.
3.1 Requirements on the robot
3.1.1 Moving and positioning
A tele-operated climbing robot shall be able to move vertically, laterally and horizontally
during slope consolidation.
The robot structure should be provided with an ad hoc base plate to act as a skid support
for quick vertical movement and as a protection against protruding rocks.
The robot shall be able to move laterally and vertically by a combination of the ropes by
which it is suspended from the top of the slope and its own legs.
The positioning of the robot shall be able to be provided by the robot/robe system.
Due the working conditions all on-board equipment should be well-anchored with the
robot frame and work independently from the position on the slope.
3.1.2 Power supply
A climbing robot for slope consolidation is likely to require three types of power sources:
• pneumatic: compressed air (at e.g. 12-20 bar) for a drilling unit (for the operations
of drilling and flushing debris from the drilling);
• hydraulic: oil (at e.g. 200 bar) for a drilling unit (to rotate and advance), for the
leg movements and other services;
• electric: for sensors, control system, cameras and lighting.

ROBOCLIMBER HUDEM ‘04
3.1.3 On-board equipment
A robot for slope consolidation is likely to require the following types of on-board
equipment:
• drilling rig;
• drilling rods;
• drilling bit;
• 1 camera with pan/tilt/zoom) for monitoring leg positions;
• 1 fixed camera for monitoring the drilling area;
• electronics unit for performing the on-board control.
3.1.4 Other system features
The robot shall incorporate features that actively and passively ensures that the robot will
not become disconnected from the ropes that support it from the top of the slope.
3.2 Requirements on the robot HMI.
The HMI shall allow the operator to move the robot as a whole or component by
component depending on the operational context (climbing, pre-drilling positioning,
walking to anchor points, etc.).
The HMI shall incorporate a simple to use and easily recognizable emergency stop
button.
The HMI shall allow the operator to manage and monitor the drilling, using high level
macros or low level primitives.
The HMI shall be able to operate in a harsh environment: wide temperature range,
humidity, precipitation, dust, etc.
The HMI shall be able to operate in a bright daylight environment, leading to special
needs of screens to provide high contrast, and provide features to reduce reflections of
light.
The HMI shall be ruggedized to sustain being dropped and/or treated roughly.
The HMI shall be user friendly and accessible to non computer-literate users. This
requirement translates into employing devices such as joysticks and buttons, rather than
keyboard and mouse.
The HMI should follow a layered approach allowing the operator to access high level or
low level commands depending on the situation.
HMI should support different modes, each mode being tailored and adapted to a
particular phase of the overall operation.

ROBOCLIMBER HUDEM ‘04
These modes are likely to include:
• Prestart mode
Used to specify terrain-specific data.
• Positioning mode
• Drilling mode
- Moving/Climbing/Anchoring : the operator uses the joystick to
indicate the robot where to move (high level control).
- Finetuning : for a given leg, the operator will send directives such as
"move forward 0.5 meter”. Alternatively, the Operator can adjust its
rotation, extension and elevation.
- High level operations : the operator specifies a depth and the HMI
automatically has the robot drill a hole of the specified depth (high
level command).
- Medium level operations : the operator controls the robot on a rod by
rod basis and sends commands such as "insert rod", “remove rod” and
so on.
- Low level operations : likely to be directly accessed only for testing
and debugging. At this level, the operator would control the drilling
subsystem on a very low level and would send commands such as
“rotate joint x by n degrees".

ROBOCLIMBER HUDEM ‘04
3.3 Robot HMI design
3.3.1 General considerations
The command console will be PC-based. It must be light so that it can be carried around
and easily strapped around the neck of the operator.
As the robot has to be operated outdoors, this command console has to be harsh condition
resistant (rain, dust) and the screen must have outdoor-displaying capabilities.
Finally, the console must provide remote control capabilities, to prevent the user from
being too close from slopes, and the display must be large enough for displaying camera
views.
The camera system is network based, tcp/ip cameras typically broadcast video in the form
of MJPEG, MPEG-1 or MPEG-4 streams.
3.3.2 Tablet as a command console
The hardware solution that has been chosen is the Fujitsu-Siemens tablet-pc.
Fig. 6. The Roboclimber Command
The 4121 model has an outdoor screen and a harsh environment casing that enable
manipulations even with dust or rain, and the screen is large enough to display all the
relevant information for controlling the robot.
The control station also has an embedded 802.11b wireless Ethernet card.

ROBOCLIMBER HUDEM ‘04
3.3.3 HMI Operator Mode
When running in Operator Mode, Roboclimber will be controlled using one joystick and
a set of buttons (provided by the joystick).
These buttons will be used to change the operational mode of Roboclimber.
One button is used to cycle through the different modes.
All the commands are sent via a TCP/IP connection.
Fig. 7. The control software main display
Figure 7 is an example of what the operator sees when he is controlling the robot in “high
level mode”. With a single button, he can switch to the desired sub-system he wishes to
control.

ROBOCLIMBER HUDEM ‘04
Fig. 8. The upper left leg is selected Fig. 9. The upper right leg is selected
Operating the robot is done by selecting the wanted move and setting the value or
increment to reach.
Figure 8 and 9 show how the user fine-tunes the rotational angle of the upper left leg, by
setting the angle to 42 degrees.
An arrow indicates the direction of the leg after having confirmed the command and realtime
display is provided during the move.
It is possible to send emergency stop signals during a move.
Also, for security reason, the server running on the robot pings the client regularly and it
stops immediately any move if the client does not answer (for instance, if the console
runs out of battery power or if it breaks down).
Fig. 10. Sending a command to rotate a joint to a specified angle value

ROBOCLIMBER HUDEM ‘04
3.3.4 HMI Super-User Mode
Super-User mode provides access to low-level functionalities for testing or for resolving
contingencies.
Fig 11. The low level console
Fig 12. A dialog displaying messages sent by
the robot
Figure 11 shows the low level console that is used to send commands to the root. The
user can select a command from a pull down menu and modify the parameters prior to
sending it.
The figure 12 shows the messages that are sent by the robot in response to the command.
“ACKNOWLEDGE $70 2” states that the command “INITIALIZE” has been properly
received and queued for processing.
4. Conclusion
Stabilising risky slopes is a dangerous job. Today it is done manually by workers who
have to build and climb tall scaffolding and who remain exposed to siliceous dust and
loud penetrating noise.
Maybe the most important aspect of the Roboclimber is that it will make risky jobs safer.
Thanks to automated interventions and the use of remote control, accidents related to
operating on high scaffolding can be eliminated; additionally, any sudden soil movement
will endanger nobody as the operators will all be working at a safe distance.

Sensor head and scanning manipulator for humanitarian de-mining
Abstract
P. Gonzalez de Santos, E. Garcia, J. Cobano and J. Estremera
Industrial Automation Institute-CSIC
Ctra. Campo Real, Km. 0,200- La Poveda
28500 Arganda del Rey, Madrid, Spain
Detecting and removing antipersonnel landmines is a significant humanitarian activity to
accomplish all over the world. More than 100 million mines have been scattered on many
countries over the last twenty years, and full de-mining would take several more decades,
assuming no more mines would be deployed in future. New technologies such as improved
sensors, efficient manipulators and mobile robots can help to clean fields up. This paper
presents both the configuration of a sensor head and a manipulator for detecting and locating
antipersonnel landmines. The sensor head can detect certain landmine types, while the
manipulator moves the sensor head over large areas. This paper describes the main features of
these two subsystems that are a part of a whole system under development.
1 Introduction
Detection and removal of antipersonnel landmines is at present a serious political, economic,
environmental and humanitarian problem. There exists a common interest in solving this
problem, and solutions are being sought in several engineering fields. The best solution,
although not the quickest, would be to apply a fully automatic system to this important task.
However, any such solution still appears to remain a long way from succeeding. First of all,
efficient sensors, detectors and positioning systems would be needed to detect, locate and, if
possible, identify different mines. Next—and this is of paramount importance—adequate
vehicles would have to be provided to carry the sensors over the infested fields. The case
posited above would require simple sensor arrays or terrain-scanning manipulators using just
one simple sensor. During de-mining operations, human operators would have to stay as far
away as possible for safety; thus, tele-operation is also and important issue in this application.
Walking robots exhibit many advantages as vehicles for humanitarian de-mining. Legged
mechanisms for such application have been under development for at least the last five years,
and some prototypes have been already tested. TITAN VIII, a four-legged robot developed for
general purposes at the Tokyo Institute of Technology, Japan (Hirose and Kato 1998), was one
of the first walking robots adapted for de-mining tasks. AMRU-2, an electropneumatic hexapod
developed by the Free University of Brussels and the Royal Military Academy, Belgium
(Baudoin et al. 1999), and RIMHO2, a four-legged robot developed at the Industrial
Automation Institute-CSIC, Spain (Gonzalez de Santos et al., 1999), are two more examples of
walking robots used as test beds for humanitarian de-mining tasks. COMET-1 was perhaps the
first legged robot developed on purpose for de-mining tasks. It is a six-legged robot developed
by a Japanese consortium, and it incorporates different sensors and location systems (Nonami et
al. 2000). The COMET team is currently engaged in developing the third version of its robot.
These four robots are based on insect configurations, but there are also different legged robot
configurations, such as sliding-frame systems, being tested as humanitarian de-mining robots
(Habumuremyi 1998, Marques 2002). Summarising, there is an increasing activity in
developing walking robots for this specific application field.
2 The DYLEMA project
The DYLEMA project is devoted to configure a semi-autonomous system for detecting and
locating antipersonnel land mines and it has been conceived around a mobile robot based on
1

legs. The overall system is thus broken down into the following subsystems, illustrated in
Figure 1.
1. Sensor head. This subsystem contains the mine detector and additional elements for
detecting the ground and objects in the way.
2. Scanning manipulator. The sensor head is basically a local sensor. That means it is
able to sense just one point. The efficiency of such a device can be improved by moving
the sensor head through a large area. A manipulator seems to be the ideal device for this
task.
3. Locator. After detecting a suspect object, the system has to mark the object’s exact
location in a database for subsequent analysis and deactivation. We considered that an
accuracy of about ±2 centimetres is adequate for locating landmines. This accuracy can
be obtained with commercial systems such as DGPS (Differential Global Positioning
Systems).
4. Mobile robot. It is a mobile platform to carry the different subsystems across infected
fields, which is of vital importance for thorough de-mining.
5. Controller. The global control system will be distributed into two main computers, the
onboard computer and the operator station. The onboard computer is in charge of
controlling and co-ordinating the manipulator and leg joints, and also communicating
with the DGPS, the detector and the operator station via radio Ethernet. The operator
station is a remote computer in charge of defining the mobile robot’s main task and
managing the potential-alarm database.
3 The sensor head
There are different sensor technologies for detecting mines. Simple sensors consist in metal
detectors; they are lightweight and easy to use. However, they only detect mines that have
metal parts, and they are inefficient with non-metallic mines (plastic mines). Other sensor types
are required in such cases, such as sensors based on ground-penetrating radar (GPR) (Gader et
al. 2000), chemical sensors (Albert et al. 1999) or artificial noses (Rouhi 1997).
An efficient detection system, it is commonly thought, should blend different technologies.
The DYLEMA project, however, is devoted to the development of mobile-robotics techniques
for landmine identification and location. The project’s scope does not include any sensor
Operator station
Radio Ethernet aerial
DGPS antenna
Sensor head
Scanner
(manipulator)
Figure 1. DYLEMA system
2
Onboard computer
Walking robot

development. Therefore, the simplest way would be to select a metal detector as the de-mining
sensor, just to help in the detection and location of potential alarms. After a suspect object is
detected, its location must be marked in the system database for further analysis and possible
deactivation.
The Schiebel AN-19/2 commercial mine-detecting set is used for the DYLEMA project’s
purposes. This detector is in service in the US Army as well as in several NATO countries. It
has been designed to detect very small metallic objects, typically mines with a very small metal
content. This detector and the full sensor head are shown in Figure 2. The sensor head consists
of a support holding the metal detector plus additional range sensors (infra-red sensors) for
detecting the ground and controlling sensor-head height and attitude. These sensors are located
in pairs defining the upper and lower limits of the band in which the sensor head works. This
array allows the controller to estimate the ground plane and thus to adapt the sensor head to
terrain irregularities (see Figures 2 and 3).
A sensor ring based on a set of flex sensors is located around the sensor head as an obstacle
detector. This sensor ring alerts the controller about the position of objects in the sensor head’s
trajectory, enabling the controller to steer around them. A flex sensor is a flexible-strip device
which electric resistance changes with its form. These sensors are located around the sensor
head as shown in Figure 4. When a flex sensor touches an object it is deformed, thus its electric
resistance changes, and by measuring the electric current through its resistance it is possible to
infer the contact. The flexibility of these devices allows a soft interaction between sensor head
and objects in its way.
4 The scanning manipulator
The DYLEMA project uses a sensor head based on a metal detector, which is a device that
senses a single point or very small areas. A scanning device is therefore needed that can sweep
the sensor across large areas. The easiest system would be a manipulator tailor-made for this
task. Such a manipulator would require three DOFs for positioning the sensor in a 3D area;
assuming that the system is scanning a non-flat area, motions in the x, y and z components
would be required. Also, the sensor head would have to be adapted to small terrain inclinations;
hence, two additional DOFs would be needed at the wrist to control detector attitude. The mine
detector has radial symmetry, so no additional DOFs would be needed for orientation control.
To sum up, a manipulator with at least five DOFs is needed to accomplish the task.
Metal detector
Infra-red sensors
a) b)
Figure 2. Sensor head: a) solid model and b) prototype
3
Metal detector

Sensor
Ground estimation plane
Figure 3. Infra-red sensors in the sensor head
The manipulator is designed to carry the sensor head, so the design is optimised to carry just
this load. First, the load is balanced to move the detector ±45º in its pitch and roll wrist axes
with the lower torque. This is accomplished by placing the detector in a configuration in which
no torque is required in the normal position (detector levelled at rest). In the manipulator, a
RRR arm configuration is good enough for this application. Mobility is adequate and, because
the links lie along a single vertical plane, there will be fewer collisions with the environment
(assuming the robot’s body is levelled). Another key design point is to mount the joint motors
at the required position to balance the loads and decrease required torques.
The motors have been integrated inside the mechanical structure to decrease the volume
swept by the manipulator structure when moving. This decreases the number of potential
crushes between the manipulator and the environment. Joints 2 and 3 are based on a spiroid
reducer, which also changes by 90 degrees the output shaft, thus allowing the integration of the
motor inside the manipulator structure (See Figure 5).
Figure 5a shows a detailed design of the scanning manipulator taking into account the
Figure 4. Flex sensors around the sensor head
4
Flex sensors
Sensor head
Sensor head
Ground profile

Elbow joint
Pitch motor
aforementioned design requirements and Figure 5b shows the prototype. Table 1 lists the
manipulator’s features. Some of these features, such as manipulator-link lengths, depend on the
robot’s dimensions (body height and leg span).
5 Conclusions
Attachment to robot’s body
Shoulder motor
Shoulder joint
Elbow motor
Roll motor
Pitch and roll joint
a) b)
Figure 5. Scanning manipulator
Detection and location of antipersonnel landmines is being performed mainly by human
operators using manual equipment. There is worldwide interest in eradicating deployed
Joint/Link Link length
Table 1. Main scanning manipulator features
(mm)
Motor power
(watt)
5
Gearing
Mass
1 60 14 246:1 1.5
2 341 72 287:1 2.1
3 341 26 489:1 1.9
4 -- 12 246:1 0.2
5 200 12 246:1 --
Manipulator 5.7
Sensor head 2.2
Total 7.9
(kg)

landmines, and solutions are coming from new, emerging engineering fields. The robotisation
of this task would provide many benefits to the human community in many countries.
Although new sensor technologies are required to detect landmines efficiently, mobile robots
must also be improved to carry that sensor in a safe manner. Legged locomotion offers some
important advantages for moving on natural terrain and it appears to be a good solution for
carrying mine sensors efficiently over infested fields.
Some preliminary work has been done to study the potential of using walking robots for demining.
The DYLEMA project is one more attempt to use walking robot for humanitarian demining
tasks. This paper addresses the development of a manipulator and a sensor head that are
part of the DYLEMA project, which is also briefly introduced.
The system has to be completed with the tools it needs for forming databases of potential
alarms and providing the operator with adequate images and graphs. The incorporation of new
sensors, detectors, and software for signature analysis will be addressed in the second step of
this project.
Acknowledgement
This work has been funded by the Spanish Ministry of Science and Technology under Grant
CICYT DPI2001-1595.
References
Albert, K.J., Myrick, M.L., Brown, S.B., Milanovich, F.P. and Walt D.R. (1999). “High-speed
fluorescence detection of explosives vapour”, SPIE, 3710, pp. 308-314.
Baudoin, Y. Acheroy, M., Piette, M., and Salmon, J.P. (1999). “Humanitarian de-mining and
robotics”, Mine Action Information Center Journal. Vol. 3, No. 2.
Gader, P.D., Nelson, B., Frigui, H., Vaillette, G. and Keller, J. (2000). “Landmine detection in
ground penetrating radar using fuzzy logic”, Signal Processing, Special Issue on Fuzzy
Logic in Signal Processing (Invited Paper), Vol. 80, No. 6, pp. 1069-1084.
Gonzalez de Santos, P., Armada, A., Estremera, J. and Jimenez, M.A. (1999). “Walking
machines for humanitarian de-mining”, European Journal of Mechanical and
Environmental Engineering, Vol. 44, No. 2, pp: 91-95.
Habumuremyi, J.C. (1998). “Rational designing of an electropneumatic robot for mine
detection”, Proceedings of the 1 st International Conference on Climbing and Walking
Robots, pp. 267-273, Brussels, Belgium.
Hirose, S. and Kato, K. (1998). “Quadruped walking robot to perform mine detection and
removal task”, Proceedings of the 1 st International Conference on Climbing and Walking
Robots, pp. 261-266, Brussels, Belgium.
Marques, L., Rachkov, M. and Almeida, A.T. (2002). “Control system of a de-mining robot”,
Proceedings of the 10 th Mediterranean Conference on Control and Automation, Lisbon,
Portugal, July 9-12.
Nonami, K., Huang, Q.J., Komizo, D., Shimoi, N. and Uchida, H. (2000). “Humanitarian mine
detection six-legged walking robot”, Proceedings of the 3rd International Conference on
Climbing and Walking Robots, pp. 861-868, Madrid, Spain.
Rouhi A.M. (1997). “Land Mines: Horrors begging for solutions”, Chemical & engineering
news, Vol. 75, No. 10, pp. 14-22.
6

Live Driving Video System (LDV) command, control and communicate
with the robotic device
P.T. Gschwind, M. Stuber, Innosuisse Corp, Murten, Switzerland
The LDV-System needed 15 years of development and now it’s time to integrate it in
applications.
The LDV system
The LDV system (Live Driving video system) makes it possible to control or manipulate a
unmanned vehicle or an object in real time, which is miles away. With this activity the socalled
simulator illness normally arises. With the LDV system there is no such problem, you
can work with it for hours and hours. Operations over large distances are possible. The LDV
technology replaces the usual remote control that uses standard screen and standard
joystick. With the LDV system the controller already believes after few seconds to be
inside the moving object or part of it, if he manipulates a robot. The advantage of this
procedure shows up on one hand in the response time (milliseconds) and on the other hand
in the improved range of manoeuvres (speed and precision). For instance, it is enormously
difficult to steer an unmanned vehicle by means of joystick and screen around a stone or a
hole on the road, particularly if you need to drive at 70 kilometres an hour. With the LDV
system however the pilot feels and steers this remote vehicle as if he is in it. All head,
hands and feet movements are replicated instantly in the vehicle. The controller/pilot
reads via the video system the instrumentation from in the remote vehicle. The driver,
installed in the following vehicle (or at a stationary location), controls the remote vehicle
by radio transmissions.
Radio transmission is the means for remote control. In the unmanned vehicle, a camera is
installed which replicates the head movements of the human pilot. Through a unique
camera system the “simulator driver” reads the instruments in the car, as well as,
manoeuvring for views in the rear-view or side-view mirrors. The camera transfers the
picture through the simulator helmet or eyeglasses of the “simulator driver” to the human
driver. After 15-30 minutes of training, the human driver controls and monitors the remote
vehicle with an acute human touch despite the absence of centrifugal forces present with
ordinary motoring.
The main aim was to develop a remote control for objects, which allows the user, to
manipulate in a natural way his object. I.e. to drive a car as you always drive a car, to
pilot a helicopter as you always do, to do a demining as you always do.
LDV transmits all your desired movements either to servo-motors respectively robots and
most important your head-movements instantly, so that you mean you are virtually at the
other place. All this is possible without that you get sick, LDV doesn’t cause “nausea”, the
so called simulator-sickness.
LDV and Demining: The Auto-Robot
Why a Auto-Robot?
Today NGO’s and peacekeeping forces stand in different crisis areas. International UN or
NATO troops can be found around the world. It is on the roads where many of the relief
and support vehicles are prey to undetected land mines. Over time and season’s heavy
rainfall, erosion and landslides uncover such traps. In unsecured areas, mines are
positioned at short notice. Yet, the time for the relief operation is limited mostly, it is not
worthwhile to accomplish special mine control in advance of such a mission. Also certain
roads are not at all driven on because of other uncertainties. This type of situation is
typical in countries like Angola where certain villages can only be supplied by air, as the
roads are dangerous.

To open the relief roads for NGO and peacekeeping forces in a simple way LARS can be
used in these crisis situations.
Principle
Our goal is the development of a system to patrol, examine, and detect the presence of
land mines on roads and bordering surfaces, thus, to enable the safe supply and support of
field (non-combat) operations. The “LDV-Auto Robot System” (LARS) is simply installed in a
regular road vehicle. This vehicle will be then steered remotely by the LARS system. A
LARS simulator installed in a following vehicle (Convoy) and can run at the same speed as
the unmanned vehicle to safely control it. The unmanned vehicle will detect a mine, using
scanning equipment and special installations.
How does it work?
The “LDV-Auto-Robot System” (LARS) can be installed simply. INNOSUISSE would determine
that the vehicle to be used is available worldwide - a model with power steering (e.g.
Toyota Pick-up) and automatic gearbox. The flexibility inherent in the installation of the
robotic device makes it suitable for a variety of vehicles. With each application proposal,
the appropriate vehicle will be considered. Installing LARS requires: the driver's seat be
removed and replaced by rails to position LARS, then the steering wheel is removed. The
appropriate robot arms are connected with the steering post and the pedals. The vehicle is
started through the LDV-simulator and the driver takes over the control. The instructions
are transferred to servomotors. There are three sets of servomotors, installed in fallback
or failsafe mode, to mitigate or avoid a malfunction. If the vehicle detonates a mine, the
camera and the robot installation are protected from the usual land mine impact by virtue
of the INNOSUISSE installation materials and techniques.
To protect the following vehicles optimally, the ahead-driving car has so-called mine
scanning and detonation installations. How these are defined, see below. They should
bring both anti-person mines as well as anti-tank mines to explosion and be simply
mounted at the vehicle.
Vehicle Requirements
“Pick-up” vehicle with a pulling and brake capacity around 3.5 ton. Power steering and
automatic gearbox is required. The scanned width is normally 3 meters (expansion
possible) and will be marked.
Vehicle attachments (interior and exterior) include:
> LDV-Auto-Robot
> additional observation camera (optional) on passenger seat for second man in the Convoy
At the vehicle:
> fork construction for wire and stick release of camouflaged mines
> magnet unit for the release of sensor-steered anti-tank mines (optional)
> “jammers” for remote control mines, control signal disruption (optional)
> trailers with “floating” disks, the area following, for the release of near-surface and on
surface pressure-activated mines
> marking unit of track secured (marking material disappears after one hour)
Advantages/disadvantages:
+ high speed possible (we run with the LDV today at maximum speed of 140 km/h)
+ simply applicable world-wide
+ trailer can be equipped with additional rubber wheels (pull down possibility), so that it
can be driven simply (lateral mounting of the axle) to the place of work.
+ also applicable as a forward observation and image transmission vehicle
+ favorable costing

+ only short time training necessary
- damage of the de-mining aids and the vehicle with mine contact
- no guarantee that all anti-tank mines are brought to explosion.
Price
The estimated price for small series are for the LDV-Auto-Robot and Simulator below US$
80’000.00. Installation for the Car are depending on used sensors.
The future
Innosuisse Corp. is working on a demining robot, which incorporates the LDV-System for
the Head movements and touch-sensitive robot-arms controlled by your own hands to do
demining manipulations.
Contact:
Innosuisse Corp.
Hans-Christian Stuber
President
Ryf 21
CH-3280 Murten
Switzerland
Phone: +41 (0)26 670 75 76
Fax: +41 (0)26 670 75 79
Mobile : +41 (0)79 250 47 44
Direct Mail : info@innosuisse.com
Webpage : www.innosuisse.com

Robotics systems for Humanitarian Demining : modular and generic approach. Cooperation
under IARP and ITEP
1. Introduction.
Yvan Baudoin
Royal Military Academy, Brussels , Belgium
Chairman IARP/WG HUDEM
Yvan.baudoin@rma.ac.be
Robotisation of humanitarian (ref 1) demining reflects the use of teleoperated, semi-autonomous or
autonomous robots with mobile platform. Robotic solutions properly sized with suitable modularized
mechanized structure and well adapted to local conditions of minefields can greatly improve the safety
of personnel as well as work efficiency, productivity and flexibility. Robotic research requires the
successful integration of a number of disparate technologies that need to have a focus to develop
- flexible mechanics and modular structure;
- mobility and behaviour based control;
- human support functionalities and interaction (HMI);
- integration of sensors (including the control of the inteferences) and data fusion;
- different aspect of autonomous or semi-autonomous navigation;
- planning, coordination, and cooperation among multi robots (MAS);
- machine intelligence;
- wireless connectivity and natural communication with human;
- virtual reality and real time interaction to support planning and logistics of robot service.
In the current situation, there is a need to correctly define the usefulness and needs of robotics solutions ,
essentially in pre- and post-mine detection (minefield delineation and quality assurance), to develop a
network of research-centers focusing on this kind of solutions , to define standardised modules of the
uused Robotics Systems . Beside the correct orientation of research activities , deduced from such
definitions, it will be necessary to develop test methods and procedures in order to assess the
performances of the 'System' in highly, cost-effective and most generic way. A (European) Network , with
teams focusing on work-packages related to the modules defined in the figure 1, will help to clarify the
role of the Robotics Systems (or Mechanical assistance) and assist the actual T&E activities of ITEP and
the current R&D to be pursued under European or National funded Projects
2. The Programmes and Networks
IARP (The International Advanced Robotics Programme) actually focus on three major topics: (1) the
robot dependability, including the reliability, the cost-effectiveness, the safety and the human-machineinteractions,
for industrial as well as for service and/or environmental applications (2) the humanitarian
demining, including the gathering of informations on the sensor systems that will be available in a next
future and improve the usefulness of robotics systems (3) the robotics systems aiming the reinforcement
of the security/safety in societal applications. Three working groups that are open to the scientific
community: www.eng.nsf.gov/roboticsorg
ITEP (Test and Evaluation Programme) allows truly independent testing of products produced under
either national or European funding. ITEP gives an international dimansion to the test and evaluation
programmes as well as to the standardization initiatives. Furthermore the European Commission issued a
mandate to the European Committee for Standardization (CEN) for Humanitarian mine action,
incorporating the request to coordinate these efforts with the International Mine action Standards (IMAS)
through close cooperation with GICHD (Geneva center) and UNMAS (UNO coordination) (ref 2)
Under ITEP the next task has been defined, that will be supported by IARP
Project No. 3.1.4
Title Robotics systems for the detection of Mines
Description Information on the current esearch-activities aiming the introduction of
automatisation techniques in Humanitarian Demining

Aim Repertory of existing experimental Robots (and projects)
Collection of satisfied requirements and potential usefulness through IARP
workshops (a.o.)
Request
Category Mechanical Assistance
Type Automatisation
Time frame End June/July 2004 (*)
Place
Lead nation BE
Partners Members of WG IARP/HUDEM (www.eng.nsf.com/roboticsorg) and Eur Networks
Clawar-2, Euron, Eudem-2..)
Point of contact Yvan.baudoin@rma.ac.be
Web site http://mecatron.rma.ac.be
http://www.mat.rma.ac.be
Comments (*) CD-ROM with requirements and descriptions of robots in progression (example,
fig 2)
Example of descriptive sheet (ref 3)
a. PICTURE
b. DESCRIPTION
SYSTEM Intelligent modular, open and upstream compatible
architecture (in revision)
Controllable by just one operator at the Command and
Control Station
Robot equipment not yet rugged for all kinds of outdoor
usage. Tracked Platform to be replaced by end
december 2004 (obsolete)
COMMAND/CONTROL (CC station or CCS) Operator pre-mission planning (motion and lateral
scanning along preprogrammed corridors according to
pre-choices of detection levels)
Direct commands
Robot monitoring from CCS
Small control unit, integration into existent manned vehicles
possible
MODE OF OPERATION Remote controlled
AUTONOMOUS MOBILITY Under visual control on off-road terrain way point
navigation,:
− speed up to 2 km/h
− velocity adaptation with respect to terrain structure
−restricted obstacle detection and obstacle avoidance
or vehicle stop
COMMUNICATION BETWEEN UGV and CCS Range: 50 m , wire-guided (radio-link optional)
PAYLOAD 3D Lateral Scanning Device 850x850 mm²
Digital Metal detector Vallon (MD), RMA- Ultra Wide band
radar (UWB) (Interface Acquisition Programme CORODE)
Digitised pictures from stereo-camera or elecetrical
switches for scanning
Localisation Color Camerat on fixed (CCS) pan & tilt
platform

Sufficiently modular for reconfigu-ration of alternate
payloads
TRIALS Indoor and outdoor (RMA- Dummy minefields BE)
Localisation Resolution: 0.5 %
Performance: systematic (quality control) scanning with
MD and UWB: 6 m²/H to 12 m²/H
Not satisfactory on gravelly terrain (platform/tracks to be
replaced)
C. SOME RESULTS (Hardware/Software)
LOCALISATION (Ref: VUB, Tracking with colour camera – P.Hong, G.De Cubber)
SENSOR-INTERFACE (Ref RMA, Development of a Control Software for the Demining CORODE,
E.Colon)

Figure 2. Descriptive sheet of the RMA HUNTER
EUDEM-2 is a European Network that may ease the exchange of informations among the members of
the Scientif Community and with the End-Users. More details will be given by Karin Debruyn during this
IARP WS HUDEM’2004
CLAWAR-2 (Climbing and Walking Robots and Associated Technologies) The consortium includes all
the stakeholders that are needed to develop and to promote a European robotics industry able to exploit
the Robotics technology; CLAWAR comprises 5 contractors (UoLeeds, QinetiQ, UNICT, CSIC and
Robosoft who reflect the different viewpoints from industry, academe and research centres) and 28
members from 12 countries. Many of the organisations have been involved in the initial CLAWAR TN
and are heavily involved in developing the area of robotics. The consortium includes a good balance of 10
Universities, 6 Research Centres, 15 from industry (SMEs and larger industrial organisations), and 2
members from newly Associated States. A number of End Users and Standards and Professional
Organisations are also involved as Associate Members and/or observers to assist in the development of
Guidelines and Best Practice solutions for robotic systems in a range of applications areas.
Internationally renowned Third State Partners (non-funded) are also involved as observers to give
added value as well as create a world network for this area of technology. Prof G.Muscato gave , during
this IARP WS HUDEM’04 some informations on the Work-package focusing on the Outdoor Robotics .
The members of the EN CLAWAR attach a major importance to the modular character of a robotics
system. If the modular approach of an application may be defined by the scheme of the figure 1, the next
figure 3 describes the modular definition of the robot self. Every component has been studied in details
and may lead to specific guidelines: those standard requirements will be defined by the end of the actual
contract (May 2005)

“Interaction Space”
Generic specification of modules
ACT1 ACT2
CHW1 SEN1 SEN2
POW1
MECH1
Actuator
SOFTWARE
Intelligent
actuator
SOFTWARE
Comms
system
Analogue
sensor
SOFTWARE
Intelligent
sensor
Figure 3. Clawar-Modular description of a robotics system
SOFTWARE
Power
Databus
Mechanics
Analogue
Digital
Environment
Intelligent
power supply Mechanics
The projects on (Robotics in) Humanitarian Demining , in the context of a European or International
cooperation or information-exchange (under IARP, ITEP, EUDEM-2, CLAWAR for example) aim at
making humanitarian demining safer, faster and more cost-effective by developing:
• a global system approach to Mine Action Technologies (MAT), while building, when possible, on
the results of the previous and ongoing national and EU framework programmes. The end-users and the
Standing Committee for Mine Risk Education, Mine Clearance and MAT, established to implement,
monitor and control the Mine Ban Treaty, request that MAT must be affordable, simple, effective and
manageable (see Roadmap defined by Prof Acheroy in the current proceedings)
• an effective methodology to assess the real performances of the developed tools in the field.
Under ITEP (ref 3) , the MACE T&E WG (Mechanical Assistance Clerance Equipment, Test and
Evaluation working Group) is chaired by Dr. Chris Weickert and Geoff Coley (CCMAT) and focus on
the mechanical mine-clearance, not on the mechanical assistance to the detection. Five objectives had
been defined, namely:
• Inventory of existing equipment and T&E reports
• Develop Best Practices
• Conduct T&E of Equipment
• Develop Standards
• Populate the Repository
The aspects Best Practices and Standards have been entrusted to the CEN WS 12 and will inspire the tests
and evaluation methodologies developed by the group.
The next table summarises the current inventory.

The next picture illustrates one of those mechanical clearers in test progression.
Vehicle status
Motion control
Sensor
deployment
Robot
positioning
sensor
VEHICLE CONTROL NETWORK
Cutter control
DATA
NETWORK
Control
transceiver
Sensor data
processing
Figure 4. Armtrak
Data acquisition
process
Data acquisition
process
Data acquisition
process
Data acquisition
process
Sensor data
transceiver
SENSOR SUITE
Sensor
Sensor 2
Sensor 3
Sensor n
COMMUNICATION LINK
Fig 1. Modular approach
Transmit /
Receive
Sensor data
fusion
Mission
management
HMI
Transmit /
Receive
Vehicle host
MISSION DATA REPOSITORY
CONTROL
STATION
BACKPLANE
Ref: 1. IARP BE report 2003 (www.eng.nsf.gov/roboticsorg)
Ref 2. Mine action Technologies: building a roadmap to bring appropriate technologies into operational
use (IARP WS HUDEM’04 June 2004)
Ref 3. ITEP MACE T&E WG, G.Coley, ITEP Meeting , London, Jan 2004 (geoff.coley@drdcrddc.gc.ca)

EUDEM2: A useful tool for the Humanitarian Demining Researchers and End-
Users Community
1. Introduction
Karin De Bruyn, Mike Barais,et Al
The EUDEM2 project found its existence as a follow-up activity of EUDEM which was
carried out in 1999. EUDEM 1999 was a small scale and limited study carried out by
the Vrije Universiteit Brussels, (VUB-ETRO), as coordinator and the Ecole
Polytechnique the Lausanne (EPFL-LAMI) on behalf of the European Commission.
The study lasted for 6 months and resulted in a web based database of contact
persons and actors active in Europe in the domain of Humanitarian Demining and a
report on the “State of the Art in the EU related to Humanitarian Demining
technology, products and practice”.
The EUDEM report 1 also provided an overview of technologies developed or under
development for humanitarian de-mining in different research centers, university
laboratories or in some commercial companies. Each technology was analyzed and
presented taking into account its cost and effectiveness. The report also gave some
insight in the personal conceptions of the actors by providing a set of 50 face-to-face
interviews 2 . These interviews explained the views of industrials developing
equipment for humanitarian de-mining, research institutes, academic researchers
from several European Universities, some end-users (represented by European
non-governmental organizations) and some donors of funding for both research and
mine action assignments, or national governments from some European Countries.
For details on the types of organizations interviewed we refer to the report 3 .
At the end of 2001 the new EUDEM2 project was launched by the European
Commission as a support measure in the 5 th framework. The new project started in
the beginning of 2002 and runs towards its end (December 2004). After a first and
very intensive effort to elaborate the web pages, a second intensive focus was
placed on visits to the European players in Humanitarian Demining and the studying
of technologies currently on the market, under development or with potential for
future developments in order to produce a catalogue of equipment currently used in
the field.
2. EUDEM2: aims
The EUDEM2 team has expanded, next to the two previous academic partners, a
new third one was added, namely, the Gdansk University of Technology (GUT-
MEED). Next to the additional partner the EUDEM2 team was reinforced by an
advisory panel that follows the developments of the team, guides and assists the
team by providing advice and information. The advisory board consists of
representatives of research, commercial companies, governments or donors, endusers
and military representatives. The advisory panel meets with the EUDEM2
team several times per year and also assists and supports the workshops organized
by the EUDEM2 team.
EUDEM2 supports ongoing research and the development of technologies that
could assist the research community, the European Commission and technology
1
Bruschini C., De Bruyn K., Sahli H. and Cornelis J. (1999) EUDEM1 – Final Report, 65p.
2
Bruschini C., De Bruyn K., Sahli H. and Cornelis J. (1999) Annexes to the EUDEM1 – Final report
3
K. De Bruyn, M. Barais, H. Sahli, C. Bruschini and Jerzy Wtorek (2003) EUDEM2: Overview and
Some Early Findings, Issue 7.2, pp. 92-95

developers by offering them an up to date overview of results and achievements in
the domain, but the focus of data provided is larger than just Europe. Bridging the
gaps between the research world, the military expertise, the actual practice in the
field and the commercial developers by providing information so that duplication of
efforts is avoided is the key to the success of the project. Next to this central and
most important aim, the EUDEM2 team wishes to bring the before mentioned
players together to create a synergy and cooperation possibilities.
3. Working Methodology of EUDEM2
3.1. EUDEM2 web pages
In order to fulfill its aims a new web based database was created. Initially this
database was furnished with updated information from the EUDEM1999 report and
database of contacts, but since March 2002 a complete new and much more
elaborated version is online at http://www.eudem.vub.ac.be. Data in this new
database were searched in the scattered set of information available on World Wide
Web, via recent key publications, and other scientific material. All data on the www
pages have been analyzed before integration into the database. The web pages are
presented in a user friendly format and are constantly updated whenever new
valuable information is available. The information is structured under 6 different
categories as shown below:
Each class of information is linked with all the others through a set of indexes
defining the relationships between the objects as described in the scheme depicted
below. The Data Base is not only used to store flat data structures and facts, but
also supports more complex relationships between them and explicitly incorporates
general knowledge about the objects described. A given Organization is for example
active in a given Technical Activity, and participating to several Projects whose
Publications can be listed. The figure shows the link classes that build up the
relationship between the classes and hence the navigational structure of the
application.
Projects
Activities,
Products
Services
Practices
Publications Events
Organizations
EUDEM2 Data Base Conceptual View

All topics are interrelated with each other in order to provide the users of the
EUDEM2 site a clear and complete overview of the available information. Next to
these easy navigation tools, a search engine to facilitate quick and easy retrieval of
sought information and a “New” section were added in order to easily spot new items
without browsing through al the pages.
3.2. Face-to-face interviews
One has to understand that face-to-face interviews, guarantee more effective
information gathering than questionnaires, and open the road for bi-directional
information exchange 4 . Coupled to a face-to-face interview is an on-site visit, with
better understanding of the real-life situation within the organisation, and the
possibility to collect (in person) extra information (brochures, presentations, data on
CD etc.) 5 .
Although the cost for interviewing face-to-face is significantly higher than for video- or
teleconferencing, it is necessary to speak to representatives of organizations in
person in most cases 6 . Setting up appointments for video or teleconferencing is as
time-consuming as for the face-to-face interviews, and are less trustworthy since
people have often unforeseen issues that require urgent dealing with. Coupled to
that, not all organizations involved are technically equipped for these kinds of
interview techniques. Most importantly, people tend to tell more details when you are
in front of them then when they are called up. Documentation is collected in an easier
and much more efficient way in person and is not forgotten or delayed.
In order to facilitate analysis of the data collected during the interview, the skeleton of
the general EUDEM2 interview consists mainly of a fixed set-up.
When specific projects – not necessarily directly related to humanitarian demining –
are discussed, the interviewer tries to identify the (i) Project aims, (ii) Maturity of the
different technologies involved, and corresponding Cost estimates, (iii) Testing
procedures, (iv) “Transferability” of the developed techniques to different aspects of
humanitarian demining, (v) Technical specifications of the equipment, performances
in certain circumstances, compatibility between different techniques, degree of
success in the field, and (vi) R&D activities and strategies, research funding,
commercial perspectives 7 .
In order to make it easier to analyze an interview, a rather rigid structure is
necessary. Also interview content needs to be approved by the interviewee prior to
publishing. This needs to be done fairly quickly (one week or 10 days at most) after
the interview. The questionnaires have been adapted according to the type of
organization interviewed and focus has been shifted from topics of relevance, but
following the same line of thinking and structure. Three different categories and
questionnaires have been drafted for:
� Industrial Companies/Equipment Manufacturers
4
Loosveldt, G. (1995) The profile of the difficult-to-interview respondent. Bulletin de Méthodologie
Sociologique, 48, pp. 68-81.
5
Loosveldt, G. (1997) Interaction characteristics of the difficult-to-interview respondent. International
Journal of Public Opinion Research, 9, pp. 386-394.
6
Groves, R. (1989) Survey Errors and Survey Cost. New York: John Wiley and Sons.
7
Sudman, S. Bradburn, N. and N. Schwarz (1996) Thinking about questions: The application of
cognitive processes to survey methodology. San Francisco, CA : Jossey-Bass.

� Research Centres/University Labs/ Governmental Agencies MOD, Foreign
Affairs, Development Aid (R&D related)
� Operators: NGO, MAC, Commercial/ Governmental Agencies MOD, Foreign
Affairs, Development Aid (Mine Action Related)
Although the set-up of the questionnaire is open at several places, a certain number
of questions can be evaluated as quantitative data, which greatly facilitates the
analysis and leads to more objective results. The open questions will be evaluated
but will be presented only as summaries.
3.3. The Technology Survey
Another main aim of EUDEM2 is to carry a broad scope and an in depth technology
survey. The technology survey activity was carried out worldwide, through literature
analysis, direct contacts, participation to international conferences and foremost a
large numbers of visits to the field and to some organisations.
The study of some of individual technologies emphasises on the status of
past/current R&D projects (overview of ongoing research), in particular the EC
financed ones, and some national projects.
A clear terminology and a proper classification is being established, distinguishing
between Research (product/system 5-10 years way), Development (

(2) Soil conductivity study: Soil parameters are often not sufficiently considered at
present in HD related scientific publications, although their knowledge is really
needed by the end user to be able to predict the detector’s performance in situ. This
is true not only for electromagnetic properties, but also for example for neutron or
vapour sensors. This could lead to combined soil maps, which could allow assessing
a priori on which fraction of contaminated land a given sensor is likely not to work
satisfactorily.
(3) Furthermore Research & Development (R&D) on humanitarian demining related
items or on technologies with potential to be transferred to humanitarian demining,
was carried out in Central/Eastern Europe.
During the second year and third year of the project the gathering of information and
analysis on the status of past and current R&D projects was carried out. Overviews
were generated for the projects in RTD that were funded by the European
Commission, and of the national projects of the Netherlands, UK and Germany.
During the EUDEM2-SCOT conference which took place in September 2003, many
interesting ideas were discussed and the findings and conclusions of the conference
have been published on the EUDEM2 web pages 12 and in the Journal of Mine Action
of JMU 13 .
It is generally known, that without help and inputs from the representatives from the
Humanitarian Demining communities, most studies are not valuable at all. Therefore
the EUDEM2 findings will be confronted with the opinions of representatives of the
end-user community, the military community, the research community and the
technological community at all times.
4. Overview of the work done up to May 2004 and ongoing
Interviews with some of the key players in Europe have been carried out, but as the
analysis of the interview results is being done at this very moment, it is too early to
communicate results.
Collaboration to other initiatives: The EUDEM2 team is keen on collaborating
with other organisations that are involved in the collection of technology related
research for Humanitarian Demining. Active collaboration to a German query on
technologies currently under development resulted in a preliminary set of structured
lists of equipment being developed at the moment. Further elaboration of these lists
will result in a well documented catalogue which is foreseen as the final result of the
EUDEM2 technology survey.
Next to this further collaboration with other information collecting initiatives was
launched and resulted in a recent joint study with the Geneva International Centre
for Humanitarian Demining (GICHD - Switzerland) on manual mine clearance and
costing issues.
Help Desk: This support activity consists mostly of replies to mail enquiries, phone
contacts and meetings. The Help Desk is up and running since start of project, and it
advertised on the EUDEM2 Website.
The Help desk provides the user with a direct reply when necessary; otherwise the
queries are redirected of to relevant persons/Web pages. During the first year of the
project the use of the help desk service was 5-10 requests per month, since January
12 EUDEM2-SCOT conference conclusions: http://www.eudem.vub.ac.be/eudem2-scot/
13 I.G. McLean (2003) Scientific Contributions to Demining Technology: Beliefs, Perceptions and
Realities, Journal of Mine Action, Issue 7.3, pp. 40-42

2003, the use has slightly gone up. The complexity of the queries varies a lot but the
satisfaction of the users is rather high. More than 50% of the users give as feedback
that they are satisfied with the replies received.
Bringing people together. One of the aims of EUDEM2 is to share information and
bridge gaps between the different players in Humanitarian Demining. The initiative
first initiative here was a joint venture between the Society of Counter Ordnance
(SCOT-USA) and EUDEM2. The International Conference on Requirements and
Technologies for the Detection, Removal and Neutralization of Landmines and
UXO was very successful. About 300 international guests from nearly 40 different
countries gathered in Brussels for a 4 days conference and more then 120
presentations were made.
5. Some findings and Conclusions
Call for Collaboration: Although the Humanitarian Demining market is very small
and shrinking 14 as stated and activities are mainly focused on research, rather than
on development a lot of activities take place in several research labs all over the
world. It seems that most of these activities have no connection with other similar
initiatives. Therefore it is necessary to join efforts and work together. This could lead
in the first place to savings of scarce funding available. Also duplication of activities
would be obsolete and results will be speeded up and in the end, most importantly,
it would also save LIVES.
Future research needs to focus more on the integration of what was done in
research for Humanitarian Demining up to now. How can what has been researched
and developed as prototypes be taken into the field? We should focus on a
distillation of the best items of all systems and merge them in new concepts or
systems that are field able soon. And it is time to move to the final stages of the
sensors development; there is no need to further research new sensor systems.
Emphasis should be placed on technologies and applications, currently available or
under development from within the Information Technology sector. Current
equipment needs to be elaborated further but should ideally adopt some IT solutions
into its system. Our daily use of high tech equipment like GPS, satellite use, palm
tops etc. should be visible in the research and development of technological
solutions to the mine problem as they could provide radical changes in
Humanitarian Demining practices today!
14 Assessment of the International Market for Humanitarian Demining Equipment and Technology.
Prepared for the Government of Canada, p30, by GPC international.

HUDEM04 Brussels
A concept of implementing technology to encourage economic growth in de-mining.
During this conference we have seen and will see some wonderful and exciting new
technologies and concepts. However, as de-mining techniques becomes more
successful and less people are being affected, donors are starting to look at other more
pressing causes to sponsor. It has been estimated that land mines were killing or
maiming perhaps 28,000 people each year, because of the hard work and success of
many organisations some of them represented here today, thankfully this is now
dropping to what some estimates at around 15,000 people each year. So the de-mining
industry is having to compete for donations and funding against this ever-increasing
competition from other causes, for example AIDS. People are becoming infected with
HIV at the rate of about 15,000 each day
Another major factor that must be considered is this, weather we use today’s or
tomorrow’s technologies it will become irrelevant if we do not change the method in
which these technologies are used. Why, because the de-mining industry can only
expand relevant to the amount of the funding that is available to it. So when we
consider the previous point of funding now being focused on other more pressing
needs, we could start to see the de-mining industry peeking out at this point and
starting to decrease in size. Effectively a premature causality of its own success, even
though the job is far from being finished. If we don’t comprehend this point and take
action to stop it from happening, it will be a terrible shame for humanity.
If the de-mining industry fails to become financially self sustainable, then it can only
be the size, relative to the funding that is available. However the industry could
accelerate it’s expansion by considering ways of grasping technology and
implementing it into developing new de-mining methods, not just new bigger better
machines, but a completely new way of thinking i.e. a method of generating it’s own
funds along with wider issues such as agricultural and community development and
most important; self sustainability.
De-Mining Systems were once nearly involved with a potential project in Asia in an
area that had been good agricultural land originally owned by small farmers.
(Unfortunately, then we did not have the technology to produce the necessary
machinery). An organisation using investor and development funding, were going to
take a vast area of the now mine ridden and overgrown land from the government on a
six year lease. We were to develop and build the MDM (Modular De-Mining)
machines that would simultaneously clear the vegetation, clear the personnel mines
and detect and map the location of the tank mines and larger ordinance, for later
removal by hand cultivate the ground and plant a crop.
The idea was to develop some of the infrastructure and harvest, export and sell these
crops on the open market for five years. This would have paid for the de-mining and
given a return to the original investors, or the same funding could carry on again
working and multiplying elsewhere. In the fifth year the families who owned the land
before that war and the landmines were laid, were to undergo an agricultural training
program, as some of these people were now second generation. In the sixth year when
they return home and take their land over again, there is a crop in their fields and an
already established market. (They win, we win, and the investors win,)
Page 1 of 7 HUDEM04 Brussels A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

It’s basically a matter of brining business principles into humanitarian works to create
all-round sustainability.
One of the reasons that this has never been done before on such a large scale is
because, using yesterday’s technologies, the commercial cost of removing landmines
was higher than any potential agricultural returns. However, by reducing this cost
through the novel Modular De-Mining system, that also has the ability to
simultaneously plant crops for food, or other modern crops, we can now to tip the
scales in favour of large-scale mine removal.
New crops could be considered such as starch for industry or crops to help the
economy such as oilseed rape to produce bio-diesel as mineral oil reserves diminish
or become to expensive for developing nations. A situation often occurs in a non oil
producing developing nations where they do not have any money to remove the
landmines because they are having to pay out to rich nations for their oil, whereas if
they removed the mines and planted bio-oil crops this money would remain
generating up in their own economy. Many countries have now shown a commitment
to remove landmines, the potential of this is huge.
There are many spin-offs to using these business principles in humanitarian works to
create economic sustainability. Another example of this is to assist with the logistical
difficulties and expense of providing the accommodation for the de-miners. One
possible way round it is, if we were to develop a low cost method of portable
accommodation and can also generate a revenue from selling it commercially, this
would then allow us to provide it at a subsidised rate for the de-mining industry. Here
is one example…
QuickSpace animation video.
Page 2 of 7 HUDEM04 Brussels ¡ A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

In trying to define the main problems in de-mining, it’s been found that in general,
most mines are laid in developing nations and most developing nations are in warm
climates. So, after the local farmers have moved off because of the mines, mother
nature then takes over and within a couple of years these often vast, mostly vital
agricultural areas can become heavily overgrown with dense vegetation. Although
mine detection systems work, vegetation removal systems work and mine removal
systems work, they often don’t work together causing a perpetual catch 22 situation
for the de-mining operations. As most mine detection systems can not be used close to
the ground because of the dense bush/vegetation we then can’t go in and clear the
bush because of the mines, then were unable to remove the mines because we could
not detect where they are. It’s not just a matter of developing new technologies but
also working methodology or how that technology is implemented into the industry.
(PowerPoint presentation)
P O Box 73, Hexham, NE47 0YT, England
T el + 44 (0)870 126 9121
www.deminingsystems.co.uk
With mulch blower and seeder
R em ote control version
Over the last seven years, we have self funded the development of a basic machine to
prove a new theory in soil grinding and simultaneous vegetation separation. The test
machine worked beyond expectations with many spin off benefits. Now that the basic
mechanical theory has been proven, it is now time to implement some of these new
technologies that are now available such as detection, GPS plotting and machine
control systems.
The Modular De-Mining System is built around a high-speed agricultural tractor. It
comprises two Ground Claimer 3000 units fitted to the multi-purpose vehicle/tractor.
This high speed tractor allows the machine to travel without the need for a transporter
and in places with no road infrastructure.
Page 3 of 7 HUDEM04 Brussels ¢ A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

In the working position the tractor is lifted up off the ground by the two Ground
Claimer 3000 units, which are mounted at the front and rear of the tractor. This
elevated position protects the driver, tractor unit and tyres from the effects of blast
detonation. Each GC3000 has two hydraulically driven traction rollers that provide
the forward motion for the whole machine. A hydraulically driven grinding drum
counter-rotating at high speed against a shear bar is located between the two traction
rollers on each GC3000. This has the effect of digging up and grinding/pulverizing
any material it encounters such as soil, stones, vegetation and anti-personnel mines.
The positions of the two GC3000 units are controlled through intelligent 3-point
linkage units. These devices allow the GC3000s to move independently, while
accurately following the contours of the ground at the same time as keeping the
tractor unit suspended on a straight and level course. The hydraulic drive control
system, working through the workload governor, ensures that power is efficiently
channelled to where it is needed most. The forward speed is automatically reduced
when difficult working conditions are encountered such as dense vegetation and/or
hard soil.
Detection equipment may be used independently, but it is hoped that a new system
could eventually be mounted and protected inside the front rollers of both GC3000s to
identify the presence of larger devices that could not be safely destroyed in the
grinding drum. The on-board computer, using information from the detection systems
would be able to determine the size statistics of an object. If the object's statistics are
greater than that of the pre-set parameter in the software then the system
automatically stops the forward movement of the machine, before the object is
disturbed. The driver is then able to look at the object's statistical outline on a monitor
in the cab and decide whether to carry on and grind it up or to leave it behind for later
removal by hand. If the object is left behind then the size statistics and location
coordinates are registered and mapped into the on-board computer using a global
positioning system.
The MDM system can be used in a variety of working modes to tackle different
problems:
¤¦¥ §©¨©� ������¥ ���
Ground Level
28 November 1999
MDM System
3 Metre Working Position
Patent De-Mining Systems UK
Slide 5
The 3-metre wide position is
used in overgrown areas. In
this working mode the front
GC3000 unit is tilted back so
that only the rear traction
roller is in contact with the
ground. With the mulch
blower fitted, this then acts as
a vegetation-clearing device.
The rear unit then follows,
pulverizing the soil, safely
destroying any anti-personnel
mines not detonated buy the
front machine.
Page 4 of 7 HUDEM04 Brussels £ A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

� � � � � �����©� � ����� � �
Ground Level
28 November 1999
���������¦�
��������� ���
���¦�����
T ra n sp ort position
MDM System
5.75 Metre ‘ Crabwise’ De-Mining Position
Patent De-Mining Systems UK
��� ��� � ������� � � ����� � �
5.75 Metre ‘Crabwise’ de-mining
position, seen from above.
Brown areas have been cleared,
green areas have not yet been
cleared.
�¦� � � � � � � ��� ��� � � � � �
Slide 6
Another operating mode is the 5.75
metre wide position. The front mulch
blower is replaced by a secondary
seeding unit and the machine is
operated in the crab steer mode.
Slide 7
This is for mine clearing in areas
with little or no vegetation, thus
allowing the greater operating width.
Slide 8
View auto running animation sequence
of slides 9 to 97
The MDM system is unique in its
ability to simultaneously:
• Travel to the place of work
independently and at a relatively
high speed.
• Once at the place of work, it is able
to protect the tractor unit and
operator by lifting them up off the
ground.
Page 5 of 7 HUDEM04 Brussels � A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

• It will be able to detect un-exploded ordinance, automatically, stopping the
machine if the ordinance detected is above a pre-set parameter, again
protecting the operator and machine.
• It is able to remove the vegetation and bush while keeping it separate from the
soil-processing operation.
• It will be able to grind up and detonate anti-personnel mines while registering
the size statistics via the detection system and location co-ordinates via the
GPS system of the larger UXOs for later removal by hand.
That was a small piece of vegetation, but the prototype is able to deal with trees up to
15 cm in diameter or more. Most personnel mines are too small to detect while
quickly on the move like this, and they are either detonated by the pressure of the
traction roller, or they are ground up often detonating inside the counter torque blast
suppression unit. The whole machine is able to leave the land in a more valuable
condition for agricultural viability and even seed, fertilise and roll (plant a crop), and
this is all carried out simultaneously with the one unit.
With this de-mining system being modular in design it is able to help tackle the wider
picture of global de-mining. For example after an area has been cleared the front and
rear machines can be taken off and any other machine with the standard agricultural
three point linkage system can be attached, such as a digger on the back and a cement
mixer on the front to help with small construction and infrastructure development
works.
We are designing a double acting forklift mast that can be attached on the back of the
tractor. This can be used to unload seed or fertiliser when crop planting or building
materials and we’re designing another unit called the QuickDrill system that the
forklift backs into it clips on and the whole unit then becomes a water well drilling
rig.
R em ote control version with mulch
blow er and seeder
Another use for the MDM system is in
humanitarian work such as crop planting.
The machine is able to re-establish the
crop cycle as part of a famine prevention
programme following wars, droughts and
floods etc. It can very efficiently plough
up, cultivate, seed fertilise and roll at a
width of 6 metres. (Slides 98 to 99
showing the remote control version)
This may seem futuristic, but we are going
to have to come up with something better
than spending more money on clearing the
land, than what the land is currently worth.
Our objective has to be, too develop and
produce machinery/systems that increase
the safety and efficiency of the landmine
removal industry while facilitating a
reduction in the operating costs and an
increase in revenues.
Page 6 of 7 HUDEM04 Brussels � A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

With it now being possible to send video and other data back via satellite, this could
have many other benefits such as in the transparency of charitably supported
humanitarian demining. An example of this would be in the hope of implementing a
method where donors who donate on a direct debit basis and having a sense of
ownership or people who are buying a share in, or sponsoring a particular machine,
could go on the internet and see their machine actually working or a video clip of part
of that days work. With the GPS data, donors could see a model of the area that their
machine has done. This system would also assist in persuading companies to take on
the sponsorship of a machine, It’s good public relations as their employees and
customers could go on the internet and actually see there machine working.
To clear landmines for humanitarian reasons is extremely worthwhile, but if we don’t
start to be more competitive for the limited funds against what many funders are
considering to be more pressing humanitarian problems such as Aids, malaria, TB etc,
then the de-mining industry will starve to death.
W ith mulch blower and seeder
(Slide 100)
We have to diversify and
invest now in these new
technologies and put into
practice ways of reducing
the cost of de-mining, while
increasing any potential
returns such as
simultaneous crop planting
for bio-diesel to eventually
tip the economic scales in
favour of de-mining.
If we can all work together to develop something such as a Modular De-Mining
system that could start gaining a financial return for the de-mining funders and
contractors, then this could pave the way for a dramatic increase within the industry.
Lets’ grasp these technologies and use them to turn back the hands of time and return
the land to its original purpose, that of supporting humanity.
Contact Info;
Roy Dixon
De-Mining Systems UK Ltd
PO Box 73
Hexham
NE47 OYT
United Kingdom
www.deminingsystems.org
E-mail: roydixon@deminingsystems.co.uk
Tel: +44 (0)870 126 9120
Fax: +44 (0)870 126 9121
Mobile +44 (0)797 124 8857
Page 7 of 7 HUDEM04 Brussels � A concept of implementing technology to encourage economic growth in de-mining.
PO Box 73, Hexham, NE47 0YT, UK +44 (0)870 126 9120 www.deminingsystems.co.uk

A concept for a humanoid demining robot.
Peter Kopacek, Man-Wook Han, Bernhard Putz, Edmund Schierer, Markus Würzl
Institute for Handling Devices and Robotics
Vienna University of Technology
Favoritenstr. 9-11, A-1040 Vienna, Austria
Tel.: +43-1-58801 31801, Fax. +43-1-58801 31899
Email: e318@ihrt.tuwien.ac.at
Abstract: Currently there are worldwide two categories of two legged humanoid robots available:
“Professional” humanoid robots developed by large companies with a huge amount of research
capacities either with the idea to assist humans in everyday working or with the background to
serve mostly for entertainment, leisure and hobby or in the future as personal robots. “Research”
humanoid robots: There a lot of such robots currently available or in the development stage e.g.
approximately worldwide more than 500 University institutes and research centres are active in this
field. The robots of this category are usually prototypes.
In this paper a concept of a new humanoid robot combining the features of both categories
mentioned before - professional for a reasonable price – for deminig applications is described and
discussed.
Keywords: Mobile robots, two legged robots, humanoid robots, demining robots.
1 Introduction
Removal of landmines – mostly by hand - is today a very inhuman, dangerous, time consuming and
expensive task. On the other hand the number of landmines is increasing worldwide dramatically.
Since several years there are several trials to use unintelligent and very heavy robots for demining.
There do not exist standard equipment requirements for the humanitarian demining. But the
following requirements for systems, that should improve the work of demining, can be summarized:
• The system (robot, vehicle or equipment) should be low-cost.
• Most of the parts and components should be commercially available in order to reduce
maintenance costs and availability problems.
• Approriate small and lightweight for transportation.
• The system should be of a robust design and capable to work in rugged environments.
• Operation time should be simple and failsafe (user-friendly operation) for all skill-levels to
reduce training problems
• Operation time should be at least 2 hours before re-fueling or re-charging and that work
should not take more than 30 minutes.
• Ability to distinguish mines from false alarms like soil clumps, rocks, bottles and tree roots.
• Ability to detect both types or in fact variety of different mine types and sizes.
• Operation in vegetated ground cover.
This list is a huge challenge for robot designer and there are some experts which are doubtful, that
the requirements can be fulfilled in the short run. But considering especially the first two points a
tool kit could offer notable advantages to accomplish these requirements.
In the last decade a new generation of mobile, intelligent and cooperative robots were developed
and introduced. One of the reasons for this development was the availability of reasonably cheap

sensors and the increasing computer power. This offers the possibility to use robots of this new
generation for humanitarian demining. Since mines are likely to be placed in many different terrains
it should be especially attached value to the locomotion system.
2 Locomotion Systems
The most spread type today is the wheel locomotion system. Typically the number of wheels
amounts to three, four or six. The extensively use of wheeled locomotion systems has many
practical reasons. The underlying mechanic is simple and easy to assemble and the ratio of
admissible total weight for payload to own weight is quite acceptable.
The problem is that wheeled systems are in rugged terrain not as efficient as in plain. As a basic
principle it can be supposed that wheeled vehicles have problems to overcome obstacles higher than
the radius of their wheels.
The advantage of using only three wheels is high manoeuvrability at the expense of balance. Fourwheeled
systems are well developed in the car industry and all-wheel drive would be obviously
desirable for off-road applications. But the best solutions seems to be the six-wheeled since this
configuration is the most suitable for off-road and mine fields are mostly not plain and additional
scattered with small obstacles like stones and branches. A six-wheeled robot nearly achieves the
same performance in off-road terrain than a robot with a chain locomotion system.
Chain locomotion systems are successful for rough, muddy and sandy terrains. The comparative big
tread allows robots with such a system to pass relatively big obstacles. Beside the heavier hardware
compared to wheeled robots the main disadvantage is the low efficiency of chains. Energy is
dissipated by friction in the chain and by relative movement between the chain and the ground.
Therefore the robot has need of a more efficient driving system which however increases the total
weight. But especially in the case of demining a robot has to fall below a weight limit to avoid
triggering the mines. On the other hand the use of a chain locomotion system enables the designer
to decrease the pressure force of the robot onto the ground by simply increasing the length of the
chains instead of reducing the weight of the robot.
Walking locomotion systems are in general the better solution for rugged terrain compared to wheel
and chain locomotion systems. But the development of these complex devices is only at the
beginning. Nevertheless many mobile robots demining projects are using this concept. Legged
vehicles can be omni-directional and can turn in place. A unique characteristic of walking vehicles
is the ability to move with discrete foot placements, thus allowing the robot to avoid stepping on
landmines or other delicate objects. Also, the legs of a walking robot cause much less damage to the
terrain than a standard tracked vehicle. Additional due to the freedom of leg placement, a legged
robot can stop walking and then choose suitable stable footing to work on a mine, even on very
uneven ground. Furthermore, its body posture can be changed while keeping the feet on the ground,
thus essentially adding another degree of freedom for performing the task. An unexpected explosion
damages a legged robot in a much lesser degree than a tracked robot if the posture is such that the
legs are stretched out of the centre of the robot’s body like the legs of a spider. If the robot does
mistakenly detonate a mine the damage is reduced to the end of the leg. In this way the body can be
protected, although this approach does require lightweight, inexpensive, replaceable legs.
Another advantage for the use of legged robots in demining actions is that the legs are may also
used as manipulators. In principal such legs are robot arms used for locomotion purposes. They
only have to be equipped with a tool and can be used for other purposes. Disadvantages of legged
robots are the relatively low movement speed and of course the complexity of the system compared
to tracked vehicles in regard to the used mechanical parts and the control of the movement.

Therefore it could take a long time until legged robot platforms suitable for real outdoor
applications are competitive to standard tracked robots.
3 Two legged – humanoid - robots
It was an old dream to have a personal robot looking like a human. Main features of a humanoid
robot are: biped walking, voice communication – speech recognition, facial communication. For
demining only the first is currently important.
The main feature of a real human is the two legged movement and the two legged way of walking.
Much research has been conducted on biped walking robots because of their greater potential
mobility. On the other side they are relatively unstable and difficult to control in terms of posture
and motion.
The stability during the walking process decreases with the number of the legs. Therefore there
were 8, 6 and 4 legged robots copied from the nature (insects, swarms, …) developed in the past.
With new technologies and new theoretical methods two legged robots can be realised responsible
for human tasks like service applications, dangerous tasks, tasks on the production level, support of
humans in everyday life and also for deminig actions.
Static balance is a simple variant but limited to speed and possibilities of movements that the
walking looks not like human. Therefore we have to use for a two legged robot the quasi dynamic
balance which is today standard for larger and efficient projects. This kind of balance requires
another method for the control of stability because in this method the robot will be sometimes in
unstable positions. Because of the limited computing capacity this method can be only realized with
very powerful and fast computer hardware.
Walking and running realized by two legged robots is only possible by means of appropriate actors
and sensors. Currently there is no real alternative to the classical electrical drives. One step in the
direction of humanoid walking concerning actors could be the pneumatic or artificial muscles.
Concerning sensors and take into account the huge developments of sensors in the last ten years
there should be no problem to use standardized sensors – one problem could be only the size and
the prize of such sensors. The problem arising here is only to install a reasonable computer
hardware to include all the signals from the different sensors online in the software.
A humanoid robot could only move efficient if the mechanical construction parameters are
optimized. First the mass distribution could minimize the complexity of the control as well as the
energy consumption. Currently the energy consumption is more or less an unsolved problem in
mobile robots. But there are some technologies available to solve this problem in the next years.
Approximately 40% of the total weight of the robot are batteries for the power consumption. The
reloading time of the currently available batteries is approximately between one and two hours
guarantee an operation time of the robot of 1.5 hours.
Open problems of the humanoid walking are currently the changing of the weight on each step by
means of a small roll movement of the pelvis and the other leg.
4 The new concept for a demining robot
We are currently working on a humanoid, two legged robot called ARCHIE. The goal is to build
up a humanoid robot, which can simulate in some situations a human being. Therefore Archie needs
a head, a torso, two arms, two hands and two legs and will have the following features:

Height 80 - 100 cm Intelligence “On board” intelligence
Weight less than 40kg Hands Hands with three fingers
Operation time minimum 2hrs Degrees
freedom
of minimum 24
Walking speed minimum 1m/s Other features Capable to cooperate with other
Price low Selling Price –
robots to form a humanoid Multi
using commercially
Agent System (MAS) or a “Robot
available
components
standard
Swarm”
Technical details – according to the currently available technologies:
Components 1 Head, 2 Legs, 2 Arms, 1 Torso
CPU + Main Microcontroller � PDA Module
� More than 32 MB SDRAM
� 400 MHz
� Windows CE or Linux
� Serial Interface
Sensors Ultrasonic, Temperature, Acceleration, Pressure, Force
Image Processing + Audio PDA module which communicates with the main processor
Control
Image Processing 2 small CMOS-camera-modules
Audio � 2 small microphones for voice recognition
� 1 loud speaker for communicating with the environment
� 1 amplifier
Clock System time of the PDA Module
Joints � Leg six DOF: two for the ankle (left and right of the feet, up
and down of the feet), one to lift off the leg, one for the knee
(to bend the leg), two for the hip (one for rotating and the other
one for putting the leg left or right)
� Arm four DOF: two for the shoulder section (one for lifting off
the arm and the other one for rotating), two for the elbow (for
bending the arm and for rotating the forearm)
� Head two DOF: one for rotating the head, one for up and down
movement of the head
� Torso two DOF :one for the shoulder section, one for the hip
section
� Hand six DOF: one fixed finger and two fingers with 3 DOF
Drives For each DOF a mini motor with a gear unit (up to 70 kg moment
of force) or a servo
Joint Control (Control of the For each component at least one microcontroller, e.g.: a Basic
Drives)
Stamp
Material of the body Aluminium, Plastics
Sensors: Proximity for measuring distances and to create primitive maps, temperature, acceleration,
pressure and force for feeling and social behaviour, two CMOS-camera-modules for stereoscopic
looking, two small microphones for stereoscopic hearing and one loud speaker to communicate
with humans in natural language.
The control system is realised by a network of processing nodes (distributed system), each
consisting of relative simple and cheap microcontrollers with the necessary interface elements.
According to the currently available technologies the main CPU is for example a PDA module, one

processor for image processing and audio control and one microcontroller for each structural
component, e.g.: a Basic Stamp from Parallax.
5 Summary
The humanoid robot described in this paper is just situated between expensive, “professional” and
cheap “amateur” robots. It should be used for demining but can also support humans in everyday
live and serve for entertainment, leisure and hobby. For working with kids and handicapped persons
social behaviour of the robot is necessary.
In addition he should also be able to work – like a human – in an industrial environment probably in
factory automation for transportation tasks like workpieces and tools between machines, production
cells and storage devices, to cooperate with a human on a common workplace, to operate in
hazardous environments like landmine detection and removing especially in rough terrain, to work
inside of nuclear facilities….The applications will be more and more in the future.
6 References
� Coiffet, Ph.(1998): New Role of Robotics in the Next Century, in: Proceedings of the 7th
Intl. Workshop on Robotics in Alpe-Adria-Danube-Region RAAD'98, June 26-28, 1998,
Smolenice Castle, Slovakia, p. 261-266.
� De Almeida, A.T., Khatib, O.(1998): Lecture Notes in Control and Information Sciences –
Autonomic Robotic Systems, Springer London, 1998.
� Han, M.-W.; Kopacek, P.(1998): Intelligent Autonomous Agents – Robot Soccer, in:
Proceedings of the 7 th Intl. Workshop on Robotics in Alpe-Adria-Danube-Region RAAD'98,
June 26-28, 1998, Smolenice Castle, Slovakia, p. 261-266.
� Han, M.W., P.Kopacek, G. Novak and A. Rojko (2000). Adaptive Velocity Control of
Soccer Mobile Robots, Proceedings of the Int. Workshop on “Robotics in Alpe-Adria-
Danube Region (RAAD ‘00)”, Maribor, Slovenia, p 41-46
� Hirochika, I.(1999): Humanoid and Human Friendly Robots, in: Proceedings 30th Intl.
Symposium on Robotics, Tokyo, Japan, Oct. 27-29, 1999.
� Kopacek, P. (2003): Humanoid Robots for demining, In Preprints of the On Site IARP
Workshop on Robots for Humanitarian Demining - HUDEM 2003 Vision or Reality, June
19 - 20, 2003 Prishtina, Kosovo
� Kopacek, P., Han, M.W., Putz, B., Schierer, E., Würzl, M. (2004):, NEW CONCEPTS FOR
HUMANOID ROBOTS Proceedings of RAAD'04, 13th International Workshop on
Robotics in Alpe-Adria-Danube Region. Brno June 2-5, 2004 (will be published)
� Wörn, H., Dillmann, R., Henrich, D.(1998): Autonome Mobile Systeme 1998, Springer
Verlag, 1998.

UNCLASSIFIED
Towards a Semi-Autonomous Vehicle
for Mine Neutralisation
Kym Ide, Brian J. Jarvis, Bob Beattie, Paul Munger and Leong Yen †
Weapons Systems Division
Systems Sciences Laboratory
† Land Operations Division
Systems Sciences Laboratory
ABSTRACT
Rapid Route and Area Mine Clearance Capability (RRAMCC) is a Capability Acquisition program conducted
by the Department of Defence in Australia. Initial capability requirements and risk mitigation research is being
undertaken within the Defence Science and Technology Organisation. This paper discusses progress of the
mine neutralisation research component of the RRAMCC system concept.
An earthmoving Bobcat is in the process of being converted to semi-autonomous operation for the purpose of
delivering a mine neutralisation charge. The vehicle navigation system consists of several components. High
quality video cameras housed in tilt adjustable pods give a 360-degree field of view. They are connected to a
multi-input frame grabber card that outputs a composite image configured by the remote operator. The
composite image is converted to MPEG-2 video and streamed with very low latency over a wireless local area
network (WLAN) to a remote ground control station (GCS).
Significant mechanical modifications have been made to the Bobcat to provide a platform for teleoperation. The
original hydraulic steering system was replaced with a system of hydraulic actuators and an electronic
proportional valve block connected to a “Hetronic” radio control unit. Transmissions to and from the radio
control handset are multiplexed onto the WLAN allowing the neutralisation vehicle to be driven at distances up
to 1 km from the ground control station.
Development and programming of a robotic arm for placement of the mine neutralisation charge has taken
place in a full-scale mock-up within a simulated environment. This facility including a reconfigurable road
surface provides a controlled setting for testing all aspects of semi-autonomous operation in parallel with
vehicle development.
There is a program of planned upgrades to enhance vehicle capability. These activities are expected to
lead to a field trial in the fourth quarter of 2004.
UNCLASSIFIED

UNCLASSIFIED
INTRODUCTION
Issues associated with designing a teleoperation system for a skid-steer vehicle have been studied
previously 1 . The Australian Army supplied DSTO with an earthmoving Bobcat model 843, which was
surplus to requirements. This vehicle is being converted for use as a mine neutralisation vehicle (MNV)
for the purpose of a concept technology demonstration (Figure 1 shows a concept drawing of the MNV).
The integration and testing of the various sub-systems added to the vehicle will be discussed in this paper.
Figure 1 Concept Drawing of Semi-Autonomous Mine Neutralisation Vehicle.
A mine detection vehicle will mark the position of a landmine buried in a road or route with a paint spot,
magnetic or radio frequency tag. The MNV will work after the detection vehicle. The MNV will be
driven by an operator sitting at a remote GCS to a position close to the marker using global positioning
system (GPS) coordinates as a guide (supplied by the mine detection vehicle). A visual aid indicating the
operating range of the manipula tor (robotic arm) mounted on the MNV is overlayed onto the camera
image at the ground control station display. The vehicle is stopped when the operator is confident that the
marker is within range of the manipulator. The manipulator autonomously places a mine neutralisation
charge on the ground with high precision. The MNV then retreats to a safe distance and the mine
neutralisation charge is remotely detonated ensuring that the buried landmine is disrupted.
1 Mechanical Conversion
MINE NEUTRALISATION VEHICLE SUB-SYSTEMS
A significant amount of work was required to prepare the Bobcat for teleoperation. Firstly, all extraneous
parts were removed, ie. bucket, arms, canopy, seat, rear jacks and engine cover. This left a platform with
plenty of capacity, in both load bearing and “real estate”, for mounting the various sub-systems. The
UNCLASSIFIED

UNCLASSIFIED
centre of gravity was maintained between the axles by addition of the robotic arm (discussed in Sub-
Section 6). A counterweight of welded steel plate and bar functioned as a substitute for the robotic arm
during radio control set-up and testing (as seen in Figure 2).
The main manual hydraulic control block was replaced with an electronic proportional valve block
allowing connection to a radio controller.
2 Radio Controller
The “Hetronic” Nova-L type radio controller (Figure 3) is a well proven commercial off-the-shelf (COTS)
product often used in the operation of cranes, gantries and other material handling equipment. The
controller is well suited to replace the Bobcat’s skid steering mechanism as the transmitter unit’s paddles
are fully programmable with respect to the corresponding movement of the spools in the valve block.
Differences between the left hand and right hand side hydraulic systems can be electronically
compensated to provide precise steering control.
The original radio controller units have a range of 200m and operate in the 400-470MHz band. The
transmitter unit was modified to allow data to be captured prior to radio transmission. Connected by cable
to the GCS PC’s serial port the data is packetised in software and transmitted over the LAN between the
GCS and MNV. Software on the MNV PC converts the TCP/IP data back to serial data, which is sent to
the modified receiver unit by a cable connected to the serial port. By multiplexing all controller and video
transmissions the range of teleoperation is limited to the maximum range of the WLAN. This reduces the
likelihood of mutual interference from having several radio frequency transmitters onboard. Only the
range of the WLAN and not their individual transmitters limit the range of each subsystem.
Figure 2 Current configuration of Bobcat (inset: original state)
UNCLASSIFIED
Radio Receiver

3 Camera Pod
Cables for connection to
valve block terminals
Radio Receiver
UNCLASSIFIED
Figure 3 Hetronic Nova-L type radio controller
The vision system for teleoperation consists of a number of components. Miniature cameras were selected
with low blemish Sony Super-HAD colour CCDs capable of 440 lines resolution. Two camera pods were
designed and manufactured for placement on the Bobcat. One will be installed facing forward and one
rearward but the design allows for placement in any configuration. Each pod is fitted with three CCD
cameras, as seen in Figure 4. The lenses have a horizontal field of view of slightly greater than 60
degrees. When aligned such that the field of view of each camera is slightly overlapping, two pods
installed back-to-back provide a 360 degree field of view. The housing provides protection against the
elements and shades the CCDs from direct sunlight.
Figure 4 Computer model of camera pod showing internals
UNCLASSIFIED
Programmable Steering Paddles
Radio Transmitter

UNCLASSIFIED
To ascertain the best position for placement of the camera pods the view from a number of positions was
determined by modelling. The modelling results showed (Figure 5) that the most effective teleoperation
of the MNV would be achieved when the cameras are pointed at or just below the horizon. The modelling
also showed that to operate the manipulator for placement of the mine neutralisation charge the cameras
needed to be pointed downwards with a view of the working area. To account for these requirements the
cameras were mounted on a motorised platform within the pod, which can be remotely adjusted to any
position between horizontal and a lookdown angle of 45 degrees.
4 Video Transmission
Figure 5 Model of camera views
Images from the cameras placed on the MNV are captured and transmitted to the GCS by a PC based
system described below. The remote PC contains a multi-input framegrabber card (manufactured by
Zandar Technologies) capable of capturing images from all six cameras simultaneously. The associated
software stitches these images together to form a “composite” view of any configuration desired by the
operator. Each operator may position, size, crop and overlap any of the six camera images according to
personnel preferences and save them to a configuration file. Different configurations may be preferred for
driving the MNV and placement of the mine neutralisation charge. The application is also capable of
displaying overlays to the composite view and background bitmaps, which can be used to provide
navigational aids or relevant information.
The composite view output from the Zandar card is input to an audio/visual (A/V) encoder card
(manufactured by Vwebcorp). This card encodes the input video to MPEG-2 format in hardware in
accordance with the bit rate and resolution specified by the operator. The associated server software then
packetises the MPEG-2 data and sends it over the LAN using the user datagram protocol (UDP). At the
GCS PC the client software (player) decodes the incoming data stream and displays it.
For communications a WLAN conforming to the IEEE 802.11b standard (manufactured by Proxim Corp)
was installed. This LAN connection is capable of transmitting data at 11Mbps in ad-hoc (or peer-to-peer)
mode between the MNV PC and the GCS PC. In practical terms this data rate will never be achieved and
our tests have shown a data rate of approximately 6Mbps can be attained. UDP has no facility for error
correction (data being simply unicast or broadcast across the network) hence a very low latency of
approximately 120 msec to the client software at the GCS has been achieved. Studies 2 have shown that a
maximum latency of 300 msec is acceptable for a teleoperated vehicle travelling at 20 km/h and this
increases to 800 msec at a speed of 8 km/h. The MNV has a top speed of 10 km/h.
UNCLASSIFIED

UNCLASSIFIED
Range of the video transmission system has been extended by addition of external antennas and
amplifiers. In tests a stable, low latency video image was received over a distance of up to 1000m.
5 Ground Control Station
An example of the video images transmitted in tests of the system is shown in (Figure 6). The image
shows the composite view from six cameras stitched together to form forward and rearward looking
panoramas. The A/V encoder card supports a maximum video resolution of 720?576 pixels at 25 frames
per second (fps). However, the frame rate of the video displayed at the GCS was approximately 15 fps.
Dropped frames were expected due to the lack of error correction present in the UDP used to transmit
video data.
The bandwidth of the 802.11b WLAN limited the maximum bit rate for the MPEG-2 stream to 3Mbps. It
is envisaged that by upgrading the WLAN to the 802.11g standard a practical limit for MPEG2 bit rate
should be 6-7Mbps, which is typical of bit rates used to encode commercia l DVD-Video. The A/V
encoder card supports a maximum video bit rate of 15Mbps. It also supports audio encoding and it is
envisaged that sound captured by a microphone will be transmitted to aid the driver’s situational
awareness.
Other features of the GCS graphical user interface (GUI) may include vehicle diagnostics and GPS data
on a moving map display providing location and orientation. While the manipulator operates an alternate
GUI displaying video provided by cameras mounted on the arm and gripper as well as information from
sensors will allow the operator to monitor and control the manipulator during placement of the mine
neutralisation charge.
Figure 6 Images from video transmitted over WLAN
UNCLASSIFIED
FORWARD
REAR

6 Robotic Arm
UNCLASSIFIED
An industrial robotic arm (UP-20, manufactured by Motoman Robotics) was selected for integration onto
the Bobcat. This arm has a maximum carrying capacity of 20 kg. The UP-20 was preferred to an “inhouse”
designed and manufactured manipulator because it offered mature technology with great
flexibility owing to its programming simplicity.
Shown below (Figure 7) is a full-scale mock-up of the MNV constructed in DSTO’s Robotic
Development Laboratory. This environment allows development and validation of control code and safe
function testing of the robotic arm. The laboratory permits parallel development of the mechanical
elements of the Bobcat prior to full integration of the manipulator.
The laboratory features a simulated road surface comprising gravel and sand, which can be reshaped to
form features such as potholes and bumps. Hence, the effect of these elements on the sensors and robotic
arm can be studied. A fully developed manipulator can be readied for installation onto the Bobcat for final
integration.
Figure 7 Full-scale neutralisation vehicle mock-up and simulated road surface
7 Manipulator Sensors
At the GCS, video and other sensor information will be displayed on a GUI developed for use during
operation of the manipulator. A camera mounted in the gripper provides video of the area under the
manipulator (see Figure 8). A marker designating the location of a buried mine will be visible on the GUI
display. The operator, by means of a mouse or other input device, clicks on the centre of the mine marker.
The manipulator then proceeds to autonomously place a mine neutralisation charge on the ground. Firstly,
it moves along a vector, keeping the marker in the centre of the video image, until the infrared range
finder senses the ground. Calculation of the exact position in three-dimensional space is sent to the XRC
(robotic arm controller) relative to the robotic arm’s base coordinate system. The manipulator proceeds to
UNCLASSIFIED

UNCLASSIFIED
scan over the ground near the marker and the range finder takes multiple measurements from which a
plane of best fit can be calculated through the surface.
The manipulator picks up a mine neutralisation charge in the gripper jaws. Using the plane of best fit a
movement step for the robotic arm can be calculated so that the mine neutralisation charge can be put
down perpendicular to the ground. This movement is sent to the XRC and the manipulator proceeds to
slowly put the mine neutralisation charge down on the marker. One or more of the gripper pressure
sensors will indicate when the bottom of the mine neutralisation charge has come into contact with the
ground. The manipulator rotates the mine neutralisation charge around the point of contact until all
pressure sensors indicate that the mine neutralisation charge is in full contact with the ground. The
operator manually fine tunes the position of the mine neutralisation charge and/or releases it. The
manipulator returns to a rest position.
Camera
Claws (retracted)
Figure 8 Gripper showing pressure sensors and internally mounted camera
CONCLUSIONS
Integration and testing of sub-systems for conversion of a Bobcat to teleoperation for mine neutralisation
is progressing well. Programming the Hetronic radio controller will compensate for mechanical and
hydraulic variations between each side of the skid steering mechanism.
The gripper sensors and manipulator require integration with subsequent programming of the robotic arm
for autonomous pick and place of the mine neutralisation charges. Once fully developed in the simulated
environment, the robotic arm will be installed along with its controller, PC and hydraulic generator set
onto the MNV.
These activities are expected to lead to a field trial in the fourth quarter of 2004.
UNCLASSIFIED
Infrared Range Finder
LEDs
Pressure Sensor

UNCLASSIFIED
REFERENCES
1. Barton, D., Connell, T., Tamblyn, T., Thorburn, J. and Wickham, T., “Advanced Landmine Proofing System –
Teleoperation Control System Design”, Proceedings of Level IV Design & Research Project Seminar Program,
Department of Mechanical Engineering, University of Adelaide, Adelaide, September 2001, pp108-113.
2. “Remote Operation of Off-Highway Vehicles” CSIRO Robotics & Automation Team, CSIRO Report No. CMIT-
B C#1 2003, Dec 2002 (Revised February 2003), CSIRO.
ACKNOWLEDGEMENTS
Mr Bob Beattie and Mr Paul Munger performed the bulk of the mechanical modifications to the Bobcat.
Design and manufacture of sub-systems was performed by Scientific and Engineering Services personnel,
particular thanks to: Mr Adriano Tranfa, Mr Frederico Lorenzin and Mr Peter Richards.
UNCLASSIFIED

Results of Open-Air Field Trials on Stoichiometric
Identification of UXO Fillers and Buried AT landmines
using Portable Non-Pulsed Non-Directed Fast-Neutron
Stoichiometer (Models 3B2 and 3AT2)
Mu Young Lee, Tsuey-Fen Chuang, Christian Druey, Bogdan C. Maglich,
HiEnergy Technologies, Inc.,
1601 Alton Parkway, Irvine, California 92606
www.hienergyinc.com
John W. Price
Department of Physics and Astronomy, University of California, Los Angeles, California
90095
George Miller
Department of Chemistry, University of California, Irvine, California 92697
Abstract
Stoichiometer TM is a remote online decipherer of the quantitative empirical chemical formulas of
unknown substances - through steel, soil and other barriers. It retrieves the formulas in the form CaNbOc,
where a, b, c are the atomic proportions of carbon, nitrogen and oxygen respectively, which are determined
to an accuracy of 5-10%. It also graphically displays on the operator’s push button command, each formula
as a point on a triangular screen while flashing a message of “This is explosive”, or “Not a known
explosive” or “Marginal, double check!” The contents of other chemical elements whether or not related to
explosives can also be displayed on screen. The top-of-the-line version of Stoichiometer is SuperSenzor; it
uses electronically directed neutrons and it is currently in the development stage.
We report here the results with the already developed simpler, slower, shorter range and less expensive
version of Stoichiometer named MiniSenzor, which uses non-directed 14 MeV neutrons. In open-air field
UXO filler identification tests, MiniSenzor Model 3B2 scored 100% in answering "yes/no" to the question
"Is it explosive?” It scored 80% in answering "What explosive is it?” The trials were conducted at the
Navy's EOD Technology Division (NAVSEA), Indian Head, MD. A separate series of trials with a similar
MiniSenzor, 3AT2, was carried out at a University of California test site in Irvine. Samples of 5 Kg of
TNT or RDX simulant AT (Anti-Tank) landmines were chemically identified and discriminated from nonexplosives
with no false alarms through 2.5 cm of wet soil.
The decision time varied from 2-20 minutes, depending on the sample size. The time is projected to drop
to 1-5 minutes with the improved non-directed model 3B3; and to 12-60 seconds with the directed-neutron
model, SuperSenzor 7AT7, which is being developed under an SBIR Phase II contract with the U.S.
Army’s RDECOM Night Vision and Electronic Sensors Directorate.
1. Introduction
The Stoichiometer analyzes the gamma rays emitted from substances that are probed with
fast neutrons. For the detection of explosives, the key elements are carbon, nitrogen and
oxygen. The Stoichiometer’s analysis results in the calculation of the relative proportions

Figure 1. A graphical representation of the three-element chemical formulae for several
substances.
of those three elements present in the substance. A simple way of depicting this threeelement
empirical formula is in a plot such as the one shown in Figure 1. The location of
the markers which are associated with various substances shown in the figure depends on
the CNO ratios for that substance. The distances from a point within the triangle to the
three sides are proportional to the elemental ratios for that substance. The operator does
not need to actually see the results depicted this way since an automatic alert message
will be generated if a threat is detected but the triangular plot is instructive in
understanding how the Stoichiometer works. Notice that explosives are generally
clustered around a certain region of the plot. Although some substances have fixed
chemical formulas such as TNT, which is a specific molecule, there are some explosive
substances which may consist of two or more other explosives (and possibly nonexplosive
additives as well as inadvertent impurities) in a mixture of inexact proportions.
For this reason, if the results of the interrogation fall within the region of explosives then
the Stoichiometer will report to the operator that a threat is present. This ability to

calculate the empirical chemical formula obviates the need for the creation and storage of
a database of gamma ray spectral signatures against which the experimentally acquired
spectra must be compared. Instead, the Stoichiometer’s analysis resides in a
parametrized space where the dimensional axes are the quantitative carbon, nitrogen and
oxygen content.
2. Key Components
MiniSenzor Models 3AT2 and 3B2 consists of 5 main components:
1) Portable (34 lb, 75 Watt) fast neutron generator & control unit.
2) Solid state gamma ray detectors (less than 25 lbs each) made of high purity
germanium crystals.
3) Detector interface, control, and digital signal processing module.
4) Shielding materials.
5) Portable computer with Windows-XP operating system.
6) Optional battery for system power.
The neutron generator produces up to 10 8 /sec of 14 MeV fast neutrons which are emitted
isotropically. Shielding materials between the gamma ray detector and neutron source
are used to protect the detector from neutron damage as well as to mitigate neutroninduced
noise signals. The fast neutrons that reach the target excite the nuclei in the
target which, in turn, decay to the lower and ground states via gamma ray emission. The
gamma ray detector receives the gamma ray signals and transmits them to a multichannel
analyzer for digitization and processing. The digitized data then are transferred
to a remote laptop computer for analysis.
At first glance, the PELAN system from SAIC appears very similar to the Minizensor in
that it also utilizes a 14 MeV neutron source and gamma ray detector for UXO detection.
PELAN’s utilizes PFTNA (Pulsed Fast Thermal Neutron Analysis) technique instead of
Minisenzor’s continuous fast neutron technique. The gamma detector used in competing
devices is BGO (Bismuth Germanate) which cannot perform stoichiometry because of its
much poorer energy resolution (7%) compared to the HPGe detector (0.2%) used in the
Minisenzor. Without stoichiometric capability, such systems must rely on pattern
recognition algorithms to determine the presence of explosives. Another consequence of
the poor resolution is that the fast-neutron-induced nitrogen gamma rays are basically
indiscernible from the background noise signals. Since nitrogen is the key element found
in explosives, a system which is incapable of detecting nitrogen cannot conclusively
detect explosives.
3. Test Results
In the photograph shown in Figure 2, a rocket warhead was placed in front of an
experimental prototype of the Minisenzor 3B2 system. This object and a total of 14 other
objects were interrogated during a blind test of the Minisenzor conducted by the EOD
Technology Division (NAVSEA) at Indian Head in Maryland in January of 2003 as part

of a UXO filler identification program. Among the targets, an explosive was identified as
Composition B (C1N1.2O1.3) with an empirically determined chemical formula of
C1N1O1.2 in a 5 minute interrogation. The result for the target of ammonium nitrate in a
cardboard box was identified in a 5 minute interrogation. A bottle of gasoline was also
identified in a 5 minute interrogation. The range of interrogation times was between 5
and 20 minutes for the various objects. Although the performance score of the
Minisenzor on an object by object basis was not disclosed to HiEnergy by Navy
personnel, it was revealed that the device had achieved overall score of correctly
identifying 80% of the filler materials.
More recent tests of the
Minisenzor conducted at
HiEnergy’s labs indicate that
under similar test conditions
a minimum quantity of 0.114
kg of TNT simulant
(chemical ratios: C=2, N=1,
O=2) can be detected in 20
minutes.
To first order, the detection
time, T, is dependent on the
mass of the target, M, the
distances from the target to
the neutron source, L1, and to
the detector, L2, as described
in the equation bellow:
2
cL1
L2
T =
M
Here c is an empirically
determined constant which is
dependent on the setup
configuration, target and
background.
Depending on the particular
configuration, there is also a
minimum mass threshold
below which a conclusive
2
Figure 2. Rocket warhead being interrogated by a
Minisenzor system.
stoichiometric interrogation is not achievable. Different masses of TNT simulant from
2.18 kg down to 0.114 kg were detected by using the MiniSenzor at HiEnergy’s labs in a
geometry similar to the one shown in Figure 2. Examples of the sizes of the interrogated
targets are as follows: A 4.8 lb (2.18Kg) target sample was placed in a cylindrical metal

canister, 17cm in diameter and 13.5 cm in height. A 1 lb (454g) target sample was placed
in a cylindrical can 9.5 cm in diameter and 13 cm in height. A 0.5 lb (227g) target
sample was placed in a cylindrical can 6.5 cm in diameter and 12 cm in height. A 0.25 lb
(114g) target sample was placed in a cylindrical can 6.5 cm in diameter and 12 cm in
height.
In order to make a conclusive determination of explosive content, the Stoichiometer must
collect a statistically sufficient number of nitrogen gamma rays in addition to those from
carbon and oxygen. Since TNT has a lesser nitrogen content compared to equal masses
of RDX, HMX and Composition B, the detection time for these other explosives with
higher nitrogen concentration is shorter. In situations where the target is not on a table,
so far, the minimum detectable mass of TNT simulant on the ground is 5 lb (2.3 kg) and
11 lb (5 kg) buried under 2.5 cm of soil. The latter results were from tests conducted on
the campus of the University of California at Irvine using the Minisenzor. The detection
of buried explosives is more difficult due to the large amount of background gamma ray
signals which originate from the soil. In order to properly compensate for the effect of
this background, a careful measurement of the soil must be made. This particular result
with Minisenzor is not representative of what might happen in other types of soil that are
different from that of the Irvine test site. Tests conducted on rainy days did indicate
however that similar results were achievable, at least for this soil with 5 kg of TNT
simulant buried 2.5 cm below the surface.
It is important to note once again, that the device relies on calculation of an empirical
chemical formula of the object that is being interrogated. The scientific soundness of this
approach was validated in the fact that prior to the testing conducted at Indian Head, the
Minisenzor had never acquired any gamma ray spectra from actual explosives yet was
still very successful in identifying the UXO fillers that were presented during the Navy’s
blind test. Up until that point, the research, development and testing had been conducted
only with various mixtures of chemical reagents that had served as simulants for real
explosives.
In addition to the stoichiometric evaluation of CNO content, there are certain other
elements that can also be detected. Stoichiometric results for some of these elements is
possible with Minisenzor whereas certain other elements are merely detectable. It is
expected that with the directional-neutron Supersenzor device that many more elements
can be evaluated stoichiometrically as compared to the Minsenzor. A partial list of other
detectable elements includes hydrogen, phosphorous, sulfur, fluorine, chlorine, calcium,
silicon, aluminum and iron. An example of the typical computer printout from the Indian
Head tests is shown in Figure 3. The interrogated object that generated those results is
shown in Figure 4.
In this case, the red circular dot on the triangle plot shows that the content of the box is
close to TNT or perhaps smokeless powder. In either case, it was identified by the
computer as being explosive. The bar graph in Figure 4 shows that the aluminum from
the box was also detected. The reason some of these other elements were displayed in
the analysis window is that their detection was required in order to properly identify some

of the non-explosive targets that we knew might be placed before us. Examples include
the detection of chlorine in order to identify bleach and the detection of sulfur to
distinguish gasoline from diesel fuel.
Figure 3. Computer display window from the Indian Head blind tests of Minisenzor. An
automatic response message of either explosive or chemical weapon (as opposed to
“inert”) is shown in the upper right. The stoichiometric ratios of CNO are prominently
displayed at the top. The numerical suffix associated with Hydrogen is a nonstoichiometric
result.
Figure 4. An ammunition box being interrogated by Minisenzor. The corresponding data
are shown in Figure 3.

4. Minisenzor vs. PELAN: quantitative analyzer vs. qualitative analyzer.
Minisenzor is a quantitative analytic tool that performs ‘stoichiometric’ chemical
analysis of unknown materials. Stoichiometry is the scientific term for deciphering the
quantitative chemical formulas.
Minisenzor measures the atomic proportions of all 3 chemical elements always
present in military explosives: Carbon, Nitrogen and Oxygen as well as that of other
elements, if present. Obtaining accurate proportions amounts to the empirical chemical
formula of the substance investigated thus, MiniSenzor performs definite chemical
identification of explosives as well as non-explosives. For example, it gets the formula of
RDX, C1N2O2 to an accuracy of 5% (Figures 1 and 2).
In contrast, PELAN (“Pulsed Neutron Elemental Analysis”) performs ‘Elemental
Analyses.’ Elemental analysis generally establishes only what chemical elements are
present but cannot determine the amounts of each of the 3 elements which are required
for explosive identification. E.g. RDX elemental analysis will tell only that this material
contains Carbon, Nitrogen and Oxygen; since dozens of common materials, too, contain
Carbon, Nitrogen and Oxygen, it requires a subsequent inspection to confirm what the
object really contains. An elemental analysis of 3 elements is not a ‘confirmation sensor’
but - provided it can detect all 3 elements - an anomaly sensor or a half way between
anomaly and confirmation sensor.
The problem is that PELAN cannot perform even a 3-element analysis.
Figure 5. HiEnergy’s core technology, STOICHIOMETER, chemically identifies explosives,
contraband, and other illicit substances through steel, online, in real time with no operator
intervention or interpretation.

Figure 6. What the operator sees on screen.
• STOICHIOMETRIC ANALYSIS = art & science of deciphering
quantitative empirical chemical formulas of unknown substance.
• ELEMENTAL ANALYSIS: deciphering qualitative presence of some
elements.
STOICHIOMETER
Continuous 14 MeV Neutrons
FAST and AUTO-THERMAL
γ Energy Resolution: 0.2%
Deciphers CaNbOc
Quantitatively Obtains a, b, c
to + 5%
Table I.
Operating Parameters of MiniSenzor 3UXO3 compared to PELAN:
5. PELAN cannot detect Nitrogen.
PELAN
Pulsed 14 MeV Neutrons
Fast and Thermal
γ Energy Resolution 6%
Measures Carbon: Oxygen
Ratio
Detecting nitrogen in the fast neutron induced gamma spectrum is a difficult task.

First, one cubic meter of air contains nearly one kilogram (925 g) of Nitrogen. Second,
the 5.11 MeV Nitrogen peak is inseparable from the 5.10 MeV Oxygen peak, (known as
the ‘second escape oxygen peak’); they are on top of one another. This is illustrated in
Figs 3 and 4 in which the gamma spectra observed by stoichiometric MiniSenzor and
SuperSenzor are compared to those observed by PELAN. It is clear that PELAN cannot
observe even the combined Oxygen plus Nitrogen 5.1 MeV peak, let alone to measure
what fraction of it comes from N and what from O. Third, to observe the first escape N
peak at 4.6 MeV, a gamma energy resolution of at least 0.5 % is required to separate it
from the broad Carbon peak are 4.4 MeV (Figs 3 and 4). Gamma energy resolution of
PELAN is 6%, i.e. 1000% poorer.
PELAN 2003
PELAN 2002
STOICHIOMETER
SuperSenzor 7UX 7
Figure 7.
PELAN CANNOT
DETECT NITROGEN
S = 1
N 10
S = 10
N 1

PELAN
2000
STOICHIOMETER
MINISENZOR
3UX3
Figure 8.
PELAN CANNOT DETECT NITROGEN
PELAN Cannot Detect Nitrogen
GAMMA ENERGY (KeV)
SIGNAL = 1
NOISE 10
SIGNAL = 2
NOISE 1
6. PELAN cannot differentiate explosive from non explosive in a realistic, nonengineered
environment.
PELAN claims to be able to detect explosives from the Carbon-to Oxygen ratio alone.
Confirmation detector for military explosives without detecting nitrogen is not feasible in
a realistic environment. Nevertheless, PELAN claims to be able to do so by measuring
the ratio of the Carbon to Oxygen. There are about 4000 known materials that have a
similar C/O ratio as the explosives.
For example, for RDX, C/N = 0.5 while for alcohol C/N = 0.42.
> TNT has C/O = 1.17, while starch C/O = 1.2.
> 362 INORGANIC COMPOUNDS; 414 ORGANIC, 1,000s MIXED
MATERIALS have C/O in the ballpark between TNT and RDX.
> Hence, C/O alone cannot tell explosives from non-explosives
> The knowledge of the percentage of NITROGEN IS CRITICAL. One can
unambiguously determine an explosive only by deciphering quantitatively the
empirical chemical formula involving at least 3 elements.
• CaNbOc
• ALL 3 must be known: a, b, c

7. PELAN cannot identify UXO filler on the ground
Any soil that contains both C and O would be declared by PELAN as “explosive”.
Hence, PELAN cannot be used for the UXO lying on the ground. . The only way to
analyze UXO is to place it at least 3 ft above the ground. Since UXO is found on the
ground or in the ground in 99% of the time, PELAN can be used only in 1% of the cases.
8. How to Make an Explosive Non-Detector into an “Explosive Detector.”
Tests can be engineered, however, so as to create an impression that an N/O ratio
detector can differentiate between explosives and non-explosives.
In his recent article ‘One in a Million’, Freemen Dyson has elaborately described
two conducts by which a methodology that is intrinsically incapable of getting correct
information can produce the ‘right answer’: inadequate control and biased sampling
[New York Review of Books, March 25, 2004, p.4.].
An example of such biased sampling would be to compare explosives, which
always contain both C and O, with a selected, heavily biased sample of the nonexplosives
that contain only O (no C) or contain only C (no O) . That is, by the exclusion
of over 4,000 common materials thatcontain both C and O. Then ANYTHING that
contains both C and O must be explosive by definition.
EXAMPLE: Use explosive TNT C/O=1 C and O C/O=1
and use only the following non-explosives:
> WATER H2O no C C/O=0
> CEMENT CaSiO3 no C C/O=0
> Plaster of Paris CaSO4 no C C/O=0
> Sand SiO2 no C C/O=0
> Bleach NaCl+ H2O no C C/O=0
> Gas/Diesel C1H1 no O C/O=∞
This is exactly what was done in the competitive test on UXO detection between PELAN
and MiniSenzor by U.S. Navy Explosive Ordnance Disposal Technology Center at Indian
Head, Maryland, in the period December 2002-January 2003. The tests were artificially
tailored as if to circumvent the shortcomings of PELAN. This is summarized in Fig. 5.
Moreover, the UXO and other containers examined in the test were placed on a table 3 ft
above ground, a scenario that has nothing to do with the real world in which the UXO is
either on the ground or in the ground.

This Was Exactly The Case During Navy UXO Detection
Tests At EODTECH DIV Indian Head Jan. ‘03
Figure 9.
Water no C
Cement no C
Plaster of Paris no C
Sand no C
Bleach no C
C
RDX
TNT
Only C-Deficient or O-Deficient Substances were
contrasted with C-Rich and O-Rich Substances, so that
PELAN can pass as an Explosive Detector:
“Tailoring The Tests by Client To Fit The Client”
N
O
4000 COMMON SUBSTANCES
CONTAINING BOTH C and O
ARE EXCLUDED
no O (Gas/Diesel)
9. U.S. Navy Explosive Ordnance Disposal Technology Center. Testing Provided the
Venue for Proving the Effectiveness of Our Technology Over the Competing
PELAN.
Prior to the competitive test runs between PELAN and MiniSenzor, Navy’s EODTECH
had given to PELAN a two-week preparation period (May 13-24, 2002) to test 164 cases
of UXO.
Although, EODTECH had given to the MiniSenzor one hour preparation time on January
9, 2003, MiniSenzor has scored 100% in the yes/no explosive determination and 80% in
the exact identification of the type of explosive and non-explosive tested. PELAN cannot
approach this.

Parameter/ Function
1. Can it decipher the chemical formulas of substances?
2. Can it definitely answer “explosive yes or no” without
secondary inspection (opening)?
3. Can it detect Nitrogen quantitatively? (Can it tell how
much Nitrogen is present?)
4. Can it detect presence of Nitrogen qualitatively? (Does it
have a chance of detecting any explosives?)
5. Range in materials? (Penetration Depth)
6. Emitter/Receiver Timing Synchronization*
7. Time required for STOICHIOMETRIC detection of car
bomb from a distance of :
A.6 inches
B. 1 ft.
C. 2 ft.
8a. Interrogation time to detect explosives at distance of 6
inches
8b. False Alarm Rate
8c. Probability of detection
(PoD) (Likelihood of
detecting explosives)
9. High precision gamma
detector (solid state):
Degree of precision
(resolution) of gamma
energy
untested
Untested
NO
+ 6%
PELAN
(Pulsed Elemental
Analysis with
Neutrons)
NO
NO
NO
NO
2-3 feet
10 %
(Necessary)
A. incapable
B. incapable
C. incapable
untested
10%
90%
YES
+ 0.2%
MiniSenzor 3C3
100%
(Not Needed)
12 sec.
1%
99.5%
YES
YES
YES
YES
2-3 feet
A. 15 sec.
B. 1 min.
C. 2 min.
30 sec.
Advantage of 3C3
over PELAN
∞
∞
∞
NONE
NONE
1000%
(Represents
Simplified
Electronics)
A. ∞
B. ∞
C. ∞
unknown
unknown
Unknown
∞
3000%
Table 2.
Performance of Stoichiometer Model MiniSenzor 3C3 (CarBomb Finder 3C3) compared to Pulsed Fast
Neutron Device “PELAN.”
∞

Summary
MiniSenzor 3B2 was able to identify a variety of UXO fillers on a table top 75 cm above
the ground. The minimum detectable TNT simulant on the table is 0.114 kg. The
minimum detectable TNT simulant is 2.3 kg on the ground and 5 kg buried under 2.5 cm
of soil. The results show that MiniSenzor 3AT2 is able to detect an AT landmine
simulant buried below 2.5 cm of soil. Tests and development with Minisenzor 3B3 and
3AT3 to detect smaller amounts of material, in shorter times and under deeper burial
depths are ongoing.
References
B.C. Maglich et al, “Demo of Chemically-Specific Non-Intrusive Detection of Cocaine Simulant by Fast
Neutron Atometry,” Proceedings of the 1999 International Technology Symposium, pp. 9-12 to 9-22, Mar
8-10, 1999, Washington D.C., Office of National Drug Control Policy of the Executive Office of the
President of the United States.
E. Rhodes, C.E. Dickerman, C.W. Peters, “Associated-Particle Sealed-Tube Neutron Probe for
Characterization of Materials,” SPIE vol 2092, SubstanceDetection Systems, pp. 288-300, Oct. 5-8, 1993,
Innsbruck, Austria.

Feature-Level Sensor Fusion for a Demining Robot
Svetlana Larionova, Lino Marques and A.T. de Almeida
Institute of Systems and Robotics,
University of Coimbra, Polo II,
3030-290 Coimbra, Portugal
{sveta, lino, adealmeida}@isr.uc.pt
May 28, 2004
Abstract
This paper presents a practical implementation of landmine sensor fusion techniques in a pneumatic demining
robot. The proposed method is based on a two-step strategy for the sensor fusion. This strategy allows separating
two different classification tasks: Object/Background and Mine/Another Object. The first step - Regions-Of-
Interest (ROIs) extraction and objects association - provides collections of ROIs extracted from several sensors.
Then the second step performs classification of the extracted objects using specific landmine features. The
ROIs extraction algorithm uses less parameters, is reliable in different environmental conditions and can be
implemented online.
The proposed strategy allows to create a database of landmine signatures which can be used for classifiers
training combining data from different experiments. Results of experimental implementation and conclusions
are presented.
1 Introduction
The problem of humanitarian demining has a well known importance nowadays. This task is very dangerous
and has very strict requirements of landmine clearance. As manual demining process provokes more victims, the
development of automatic demining devices becomes more important. ISR (Coimbra, Portugal) is developing a
demining robot able to explore autonomously a specified area searching for antipersonnel landmines situated on
it [8].
One of the main problems of such robot is the lack of landmine sensing devices able to provide enough information
to guarantee the required clearance rate with a low false alarm rate. Using the current sensing technologies, it is
considered that a combination of several different sensors and sensor fusion techniques are required to provide
acceptable results.
Sensors which are used for landmine detection have very different physical principles which do not allow to apply
a deep fusion provided by signal-level or pixel-level fusion techniques. Thus, there are basically two types of sensor
fusion which can be used for demining: feature-level and decision-level. Feature-level fusion is a deeper process
that uses particular features from each sensor to make a general decision. Both decision-level and feature-level
approaches are currently in the research by different groups and from the experiments there is not still a conclusion
about which approach should be considered the best (for a comparison see, for instance,[7]). In these conditions we
consider a feature-level fusion being more promising for demining.
2 General structure
Feature-level sensor fusion can be divided into the following steps: registration, filtering and mapping; ROIs extraction
(segmentation); objects association; features extraction; classification.
Two sets of classes are usually used in demining: Mine/Background or Mine/Clutter/Background. This strategy
provides information about landmine existence without paying attention to other existing objects (considered as
clutter and background). On the other side, the method proposed in this paper takes into account all the objects
which might improve the results of classification and sensor fusion in general.
Landmine detecting sensors do not provide conclusive information about presence or absence of mines. They
only allow to sense a heterogeneity of some physical characteristic against a background. This heterogeneity can be
caused by a landmine, by other artificial objects (e.g., clutter), by natural objects (e.g., stones) or by changes in the
environmental conditions (e.g., sun shining and humidity). In order to distinguish a landmine from other objects,

(a)
measurement point
X−direction movement
Figure 1: a - Demining robot equipped with two active IR sensors with bandwidths 6.5-14 µm and 8-14 µm and a
Shiebel Atmid metal detector, b - a path followed by the scanning device
its particular fingerprint in spatially mapped sensors readings should be identified. According to this specificity
the following sensor fusion methodology is proposed: the process is divided into two steps, the first step separates
all objects from background, then during the second step a feature-level fusion allows to separate landmines from
other objects (using classes Mine and Another Object). In other words the first step is a ROIs extraction and the
second step is a classification which uses features extracted from the ROIs. The criterion for selecting ROIs consists
in separating all possible objects from the background.
The advantages of this methodology are:
Unification of the classification process for classes Mine/Another object. Classifiers always need
experimental data for training and evaluation. But resources for the experiments (fields and mines) are usually
unavailable in most research labs. Furthermore the size of the training set required by the classifier ([9]) is much
higher than the size usually employed in the current experiments (about 20-40 samples and leave-one-out evaluation
method [5]). The methodology proposed in this paper helps to mitigate the influence of current experimental
conditions by providing a more unified way of using landmine signature data and allowing the integration of
different experiments. Databases of experimental data are already freely available through the Internet for public
usage [2, 1]. This data is used to create a database of landmines signatures stored as collections of ROIs obtained
from the ROIs extraction and objects association algorithms.
Focusing the classifier on “landmine features”. The proposed methodology allows to focus a classifier only
on the important features which distinguish a landmine from other objects. The classifier does not need to pay
attention to features which reflect the difference between background and any object (including landmines): mean
value, minimum, maximum etc.
Controlling the algorithm failures. Requirements for humanitarian demining are very strict and it is not
certain that an automatic system will achieve the required levels, being important to consider the possibility of
misclassification. Selection of ROIs provides visual information that can help to recognize failures leaving for an
operator the responsibility of the final decision about checking the identified region by other means, for instance,
manually.
robot path
regular grid
3 Separation of objects against the background (Step 1)
3.1 Registration, filtering and mapping
These tasks provide the necessary preprocessing of the sensors data for the ROI extraction algorithm. The data
is gathered by a scanning device (our demining robot) that performs a Cartesian movement similar to the one
presented in Figure 1(b). Due to difficulties in positioning our pneumatic scanning platform, the acquired data is
usually non regular. Thus the sensed data should first be mapped into a regular grid-map in order to build an
image where ROIs might be found.
In the case of the demining robot represented in Figure 1(a) this task can be simplified by considering the following
assumption: X coordinate does not change during movements in Y-direction, Y coordinate does not change during
movements in X-direction. Then a 1D, instead of 2D interpolation can be done. This algorithm does not require
significant processing time, thus a grid-map for each sensor can always be updated online. Following the same
principle, 1D filtering is used instead of 2D filtering and a grid-map is formed from the already filtered data.
(b)
Y−direction movement

(a)
interesting point
(b) (c) (d) (e)
Figure 2: ROIs extraction process: a - sample grid-map, b - encountering of interesting points (two iterative filters
with Q = 40 and Q = 0.004), c - Segmented map, d - extrema searching in the Segmented map, e - selected ROI
3.2 ROIs extraction
3.2.1 Overview
The task of ROI extraction is pointed by several works about feature-level sensor fusion. The main idea of these
algorithms is to use computer vision techniques applying for the processing of the grid-mapped sensor readings
(which can here be refered as grey-level images). Techniques used for this tasks usually are: thresholding ([5]),
Tophat filter and Hough transform ([10],[5]), mathematical morphology ([6]), texture segmentation and Gabor
transform ([3]).
The main problems of the considered approaches are:
• A region of a grid-map for processing is required before the processing can be started, in some cases this
region is used for the selection of parameters which does not allow performing the process online
• Tuning of the required parameters is made heuristically by experts and is hard to implement in automatic
systems (the result hardly depends on the selection of thresholds, kernel sizes etc.)
• Computer vision based algorithms mentioned above search for objects with certain shapes (in particular,
circles), this requires images obtained under well controlled conditions which is not always possible
3.2.2 Proposed Approach
The proposed algorithm intends to improve the problems pointed above. The main idea of the algorithm is: ROIs
can be found in places where heterogeneities in sensor readings are situated. To make the explanation of the
algorithm more clear we consider a ROI extraction from a sample grid-map presented on Figure 2(a).
The proposed approach has the following steps:
• Detecting interesting points. Heterogeneity in sensor readings can be detected by analysing changes of its
value. It is proposed to use for this purpose two iterative filters with different parameters based on the
principles and equations for the Kalman filter. Selecting different values for the process noise covariance Q
it is possible to achieve filters with different behaviours: the bigger Q, more exact and faster a filter follows
the signal (Figure 2(b)). An interesting point is detected if the difference between the slow filter and the fast
filter exceeds a threshold (which is equal to the Q of the fast filter).
• Forming a 2D segmented map. Segmented map is a grid-map which represents interesting points spatially. It
is formed as follow:
1. An initial integer current value is chosen (it can be any number, for example 100)
2. If V aluefastfilter − V alueslowfilter > threshold then current value is incremented
If V alueslowfilter − V aluefastfilter > threshold then current value is decremented
3. Current value is mapped to the segmented map according to the interesting point coordinates
This process forms a map with homogeneous segments that differ from each other by an integer value (2(c)).
• 1D maximum/minimum searching in the segmented map. An important property of the segmented map is:
the centre of the ROI along X coordinate corresponds to the local maximum/minimum on the segmented
map along X coordinate. A maximum/minimum searching is performed to detect all local extrema for each

value of Y coordinate (Figure 2(d)). The nature of the sensor determines which type of extremum should be
used: for metal detector only maxima are appropriate, while for IR sensor both maxima and minima can be
considered.
• Region growing in the segmented map. Each extremum encountered on the previous step starts a region
growing process:
1. Selecting a segment to which the extremum point belongs - start segment
2. Detecting all segments adjacent to the start segment
3. For any detected segment its HW ratio and size (see Features extraction) are tested according to the
rule: HW ratio of the ROI can not exceed 4, size can not be bigger when a certain threshold. If the
segment passes the test it is joined to the start segment
4. If there are no new segment joined to the start segment then the region growing is finished, otherwise
repeat from step 2.
• Association with the initial map. The final ROI is a rectangle which contains all the selected segments.
This rectangle is brought in correspondence with the initial grid-map. The final result of ROI extraction is
presented in Figure 2(e).
3.3 Objects association
The ROIs extraction algorithm selects for each sensor a set of regions which contain some objects. The ROIs from
different sensors which represent the same object should be combined together for the futher features extraction
and classification (by other words, each ROI should be associated with some object). A one-to-one or a one-tomany
association can be made. One-to-one association provides a certain result but can cause problems due to the
position errors in the sensor data. On the other side one-to-many association can provide uncertain results. Thus
the proposed algorithm for the objects association is as follow: ROIs are associated together if the distance between
them is less than 200 mm, one ROI can be associated with different objects, the value of the distance is included to
the features set used for the classification. This algorithm allows to burden the classifier with the possible mistakes
in the objects association caused by uncertainty.
4 Detecting landmines among all objects (Step 2)
4.1 Features extraction
Selection of the features which reflect the difference between landmines and other objects is a complex task because
the landmine detection sensors do not provide specific information about landmines, thus, statistical characteristics
are usually used as features ([11],[5],[6],[4]). The proposed methodology does not consider features which depend on
the current conditions of the experiment: mean value, minimum value, maximum value etc. Information provided
by these features is not significant on the current step and only confuse the classifier.
For each collection of ROIs obtained by the objects association algorithm a set of features is calculated. Two
types of features are used: ROI features (calculated for each ROI) and collection features (reflect the relationships
between the ROIs).
ROI features: Size S = height ∗ width; HW ratio HW = height
width
Standard deviation σ = √ µ2; Skewness γ1 = µ3
µ 2/3 ; Kurtosis β2 =
2
µ4
µ 2 2
contrast in the point (i, j).
Collection features: Distance between centers of ROIs; Correlation.
4.2 Classification
if height > width, and HW = width
height otherwise;
; Contrast C = max(Ci,j), Ci,j is a local
The lack of experimental data does not allow yet to make an investigation about which classifier provides better
results for demining. There are many types of classifiers used in the current works about feature-level and decisionlevel
sensor fusion for demining: rule-based, Bayesian, Dempster-Shafer, fuzzy probabilities, neural networks etc.
For this work a decision-tree classifier was implemented. It is simple to implement and to analyse the system behavior
in different situations. The classes used in this step are: Landmine, Another object and No object. Existence of No
object class reflects possible mistakes in the objects association.

Number of test 1 2 3 4
Training set A1 top A1 top + A5 bottom A1 top + A5 A1 top + A5 + A3 top
Evaluation set A1 bottom A1 bottom A1 bottom A1 bottom
N. of landmines 4 4 4 4
Detected landmines 3 1 2 3
False alarms 6 7 4 4
Table 1: Classification results
(a) (b) (c)
(d) (e)
Figure 3: ROIs extraction results obtained from MsMs experimental data (subplot A1): pulsed metal detector (a),
continuous metal detector (b) and IR camera (c). Failure tests (d,e). Results of applying two different thresholds
to the pulsed metal detector data (f ). ROIs extraction result obtained from demining robot IR sensor (g)
5 Algorithms evaluation
A project maintained by Joint Research Centre in Ispra provides a database (MsMs database [1]) of data collected
from different sensors over different test lanes populated with surrogate mines and clutter.
Evaluating the proposed ROI extraction algorithm can be carried out using data from the MsMs database. The
results obtained for the pulsed metal detector, continuous metal detector and IR camera are presented in Figure 3.
To show the reliability of the proposed ROI extraction algorithm two failure cases are simulated using a sample
segment of data: failure of positioning system (Figure 3(e)) and changes in environmental conditions (Figure 3(d)).
A thresholding technique was applied to the database data obtained from the pulsed metal detector to compare
it with the proposed algorithm (Figure 3(f)). It can be noticed that two objects selected by white rectangles are
never detected together (they require different thresholds), whereas the proposed algorithm detects both of them.
A simple experiment was made to show performance of the proposed ROI extraction algorithm when working
with the demining robot. The data from one IR sensor was processed; one cold and one hot object were used to
simulate the presence of possible landmines. The result is presented on Figure 3(g).
The performance of the whole system was tested using experimental data from MsMs database (subplots A1, A3
and A5). Data obtained from pulsed metal detector and IR camera were processed to analyse the results of sensor
fusion. The bottom half of the subplot A1 is used for evaluation. Other data are processed by ROIs extraction and
objects association algorithms; the created collections of ROIs are used to train a decision-tree classifier. Different
sizes of the training sets are tested. The results of the classification are presented in Table 1.
(f)
(g)

6 Conclusions
A two-step sensor fusion method for a demining robot was proposed and implemented. The developed algorithms are
included into the control software of the demining robot allowing its automated functioning. A database containing
the collections of ROIs is implemented using Mysql. Special ROIs extraction algorithm is implemented. Its main
features are:
• Automated online operation. The nature of the algorithm allows online implementation and selection of a
ROI right after an object being fully scanned by the landmine sensing devices
• Reduced number of parameters
• The criterion for separating objects from the background is general, so the algorithm does not require special
shapes for the objects (e.g., circular)
• Reliability tested with experimental and public database data
The evaluation of the algorithm over database data shown promising results (e.g., in the images of Figure 3 all
objects were correctly identified for metal detectors data, processing of the IR images still requires improvements).
Experiments carried out with a demining robot show that the algorithm is able to provide reliable result in real
conditions. The results of classification show the nature of the decision-tree classifier: during test 1 it is overtrained
because the training and evaluation data are from the same subplot, then test 2 shows that the training set has too
small size, the results become more stable as the bigger training set is used (tests 3 and 4). The passive IR data
used for the tests provided only a little of usefull information for the classifier. Whereas, it seems that GPR data
can significantly improve the classification process, their integration in the algorithms is part of future work.
Acknowledgement
This research was supported by the FCT (Portuguese Foundation for Science and Technology) under the project
DEMINE - POSI/36498/SRI/2000.
References
[1] Joint multi-sensor mine signature database (joint research centre - ispra). http://demining.jrc.it/msms/.
[2] Uxocoe database. http://www.uxocoe.brtrc.com/default.asp.
[3] G. A. Clark, J. E. Hernandez, N. K. DelGrande, R. J. Sherwood, S. Lu, P. C. Schaich, and P. F. Durbin.
Computer vision for locating buried objects. In Asilomar Conference on Signals, Systems, and Computers,
pages 1235–1239, 1991.
[4] G. A. Clark, S. K. Sengupta, M.R. Buhl, R. J. Sherwood, P. C. Schaich, N. Bull, R. J. Kane, M. J. Barth,
D. J. Fields, and M. R. Carter. Detecting buried objects by fusing dual-band infrared images. In Asilomar
Conference on Signals, Systems, and Computers, volume 1, pages 135–143, 1993.
[5] F. Cremer, W. Jong, K. Schutte, A. G. Yarovoy, and V. Kovalenko. Feature level fusion of polarimetric infrared
and gpr data for landmine detection. In Int. Conf. on Requirements and Technologies for the Detection, Removal
and Neutralization of Landmines and UXO, 2003.
[6] H. Frigui, P. Gader, and J. M. Keller. Fuzzy clustering for land mine detection. In NAFIPS, 1998.
[7] A. Gunatilaka and B. Baertlein. Comparison of pre-detection and post-detection fusion for mine detection. In
Detection and Remediation Technologies for Mines and Minelike Targets - SPIE, volume 3710, pages 1212–1223,
1999.
[8] L. Marques, M. Rachkov, and A.T. de Almeida. Mobile pneumatic robot for demining. In EEE Int. Conf. On
Robotics and Automation (ICRA 2002), pages 3508–3513, 2002.
[9] D. W. McMichael. Data fusion for vehivle-borne mine detection. In Int. Conf. on the Detection of Abandoned
Land Mines, pages 167–171, 1998.
[10] W. Messelink, K. Schutte, A. Vossepoel, F. Cremer, J. Schavemaker, and E. Breejen. Feature-based detection
of landmines in infrared images. In Detection and Remediation Technologies for Mines and Minelike Targets -
SPIE, volume 4742, pages 108–119, 2002.
[11] M. Roughan and D. W. McMichael. A comparison of methods of data fusion for land-mine detection. In Int.
Workshop on Image Analysis and Information Fusion, 1997.

Abstract
European Project of Remote Detection:
SMART in a nutshell
Yann Yvinec
Signal and Image Centre - Royal Military Academy,
Av. de la Renaissance 30, Brussels, Belgium,
yvinec@elec.rma.ac.be
This paper shortly describes the principles and ideas behind the project SMART, a European project intended
to help Mine Action Centres (MAC) or Mine Action Authorities (MAA) in their task of area reduction
by providing a GIS-based environment with specific tools to ease the interpretation work of the
operator. Using multi-spectral optical data as well as SAR data obtained during a flight campaign in
Croatia and satellite data from before the conflict, the tools will help the land-cover classification and the
detection of indicators of presence or absence of mine-suspected areas. The results of these tools will be
given to a data fusion module that will summarise all data and contextual information available to facilitate
the creation of maps of indicators from which mapss of danger can be derived.
A more detailed and technical description of SMART has been given – and some results described – in
[13]. This paper focuses on the principles of the project and shortly presents some new results.
1 Area reduction: a key process
Area reduction has been recognized as a mine action activity where reduction in time and resources could
help a lot. Long-term empirical data from CROMAC, the Croatian Mine Action Center, show that we can
estimate that around 10% to 15% of the suspected area in Croatia is actually mined. The minefield records
alone do not contain enough information for the proper allocation of limited de-mining resources to reallymined
areas. Their completeness and reliability are not high enough. Decision makers need additional
information. SMART is intended to provide some of this additional information that would help in two
ways: it can reduce the suspected area on some places and reinforce the suspicion of others. The goal of
the SMART project is to provide a GIS-based system – the SMART system – augmented with dedicated
tools and methods designed to use multispectral and radar data in order to assist the human analyst in the
interpretation of the mined scene during the area reduction process.
The usefulness of such image processing tools to help photo-interpretation has already been studied:
the possibility to process automatically a large amount of data and help a visual analysis is among their
advantages [6] [7].
The use of SMART includes a short field survey in order to collect knowledge about the site, a flight
campaign to record the data – multispectral with the Daedalus sensor and SAR with the E-SAR –, and the
use of the SMART system by an operator to detect indicators of presence or absence of mine-suspected
areas. With the help of a data fusion module based on belief functions [1][9][10][11][12] the operator will
prepare thematic maps that will synthesise all the knowledge gathered with these indicators. These maps
of indicators can be transformed into danger maps showing how dangerous an area may be according to
the location of known indicators. These maps are designed to help the area reduction process as described
in the present paper.

A
C
Abandoned Land
Fields in use without vegetation
Fields in use with vegetation
Residential areas
Roads
Pastures
Forests, hedges or shrubs
Water
Radar shadows
Figure 1: A: part of the speckle-reduced polarimetric L-band image (R:HH, G:HV, B:VV) B: "detection
image" for abandoned land (in bright). C: SAR classification results
2 Reducing the suspected area with SMART
The use of SMART can help technical surveys and reduce the clearing of non-risky or non-hazardous areas.
Experience shows that, in parts of the country that are considered to be suspected, there are areas
actually in agricultural use. Some farmers cannot wait for the official reduction or clearing of their fields
and take the responsibility to clear them by themselves or have them cleared unofficially – leading sometimes
to incomplete clearing or casualties. These behaviours lead to discrepancies between reality and
CROMAC’s records of suspected areas. By using multi-spectral and SAR data and processing them to
provide a classification of the areas, an operator of SMART can quickly have an objective point of view
of the real land cover and land use of a large, theoretically-suspected area. Once the use of SMART has
updated the MAC’s records and identified cultivated fields inside suspected areas, a short field survey –
shorter than what would have been needed without SMART – can be organised to determine if these fields
can be officially declared reduced. Figure 1 presents a result of classification performed on SAR data. See
[4] for more details including a quality assessment of this classification by confusion matrix and [13] for
results on Daedalus classification.
2
B

Figure 2: Left: a 400m x 400m area in Croatia as seen on optical data. Right: Detection of hedges (green)
and trees (red) from SAR data. Images courtesy of DLR
Using of satellite data from before the war can help determine if a field that is abandoned now was
already abandoned before the war. If this is the case the neglected state of the field cannot be attributed to
mine infection – although it does not mean that the field is not mined.
3 Suspicion reinforcement
The use of SMART can also help to detect abandoned fields in suspected areas, thus reinforcing the suspicion
about these fields. For instance before the contamination by landmines, agricultural fields and pastures
in Croatia were enclosed by hedges but rarely with trees. If some fields are abandoned and not used, their
borders change; hedges become mixed with smaller trees; bushes grow inside the field borders; low bushes
become small trees. The trees and bushes inside the field or at its borders are significant indicators that the
field is abandoned.
Detecting abandoned areas can be done by classification of multi-spectral data. By making it possible
to make the difference between trees and bushes, SAR provides also very valuable information for this
analysis. Since hedges were often used as hiding places and may therefore be mined, they are important
indicators of mine-suspected areas. See Figure 2 for an example of the use of SAR data to automatically
make the difference between trees and hedges.
If a field is abandoned now, it may be because the soil is simply not suited for agriculture. Using
satellite data from before the war can help to determine if the field was cultivated then. If it was, then it
reinforces the suspicion.
Multi-spectral data can also be used to detect locations where creating a minefield would have made
sense: river shores, forest borders, crossroads, bridges and any other places that are better located on images
than on old and obsolete maps. See also [13] for results on change detection focusing on roads and paths.
As a general rule all information gained from the use of SMART must be appraised by the operator.
4 From indicator maps to danger maps
The indicator maps show the locations of the various indicators of presence or absence of mine-suspected
areas that have been gathered. They synthetise the knowledge that has been accumulated during the area
reduction process. The danger maps show how likely it is that a location may be mined.
To create danger maps from indicator maps, danger zones are defined near indicators based on expertise
on the mine situation. For instance borders of forests on Croatia are suspect. So danger zones around forest
3

Figure 3: Discrete ’danger map’ of Glinska Poljana (preliminary) Red: Danger (buffers) Orange: Danger
(areas no longer in use) Green: No danger (residential areas, cultivated areas...) Other: No status (forests)
Source: ULB
borders are drawn. The size of these zones are defined from mine action expertise. See Figure 3 for an
example of what a danger map will look like.
5 Limitations
The general knowledge used in SMART is strongly context-dependent. It has been currently derived from
the study of three different test sites in Croatia chosen to be representative of the country. In the case of
another context a new field campaign is needed in order to derive and implement new general rules. Before
using SMART the list of indicators must be re-evaluated and adapted. For instance it has been noted that
the asumption that a cultivated field is not mined, although quite valid in Croatia, may not apply in other
countries such as South Africa or Colombia. It must also be checked if the indicators can be identified on
the data and if the new list is enough to reduce the suspected areas.
6 Conclusion and acknowledgment
Despite these expected limitations the ideas presented here make us confident that SMART has the technical
potential to be a working solution for an airborne general survey applied to area reduction.
This work is performed in the scope of the European project SMART: Space and Airborne Mined
Area Reduction Tools (IST-2000-25044). It is co-funded by the European Commission. The project is
coordinated by TRASYS (BE) and Renaissance/RMA (BE). The project partners are CROMAC (HR),
DLR (DE), ENST (FR), ixl (DE), Renaissance/RMA (BE), RST (DE), TRASYS (BE), ULB (BE) and
Zeppelin (DE). For more information see:
http://www.smart.rma.ac.be/
In addition to the author, points of contact for SMART include:
at CROMAC Milan Bajic (milan.bajic@zg.htnet.hr) as representative of the end-users
at DLR Helmut Süß(Helmut.Suess@dlr.de) and Martin Keller (Martin.Keller@dlr.de) for the data
collection, pre-processing and SAR processing
at ENST Isabelle Bloch (isabelle.bloch@enst.fr) for data fusion
4

at RMA Marc Acheroy (Marc.Acheroy@elec.rma.ac.be)for the technical management, Dirk Borghys
(Dirk.Borghys@elec.rma.ac.be) for SAR processing, and Nada Milisavljević (nada@elec.rma.ac.be)
for data fusion
at TRASYS Jacques Willekens (Jacques.Willekens@trasys.be) for project management and Olivier
Damanet (olivier.damanet@trasys.be) for the integration
at ULB Eléonore Wolff (ewolff@ulb.ac.be), Sabine Vanhuysse (svhuysse@ulb.ac.be) and Florence
Landsberg (Florence.Landsberg@ulb.ac.be) for field surveys, classification, change detection and
danger maps
References
[1] I. Bloch, “Some Aspects of Dempster-Shafer Evidence Theory for Classification of Multi-Modality
Medical Images Taking Partial Volume Effect into Account,” Pattern Recognition Letters, No 8, vol
17, pages 905-919, 1996.
[2] D. Borghys and V. Lacroix and C. M. Acheroy, “A Multi-Variate Contour Detector for High-Resolution
Polarimetric SAR Images,” Proc. ICPR 2000, Barcelona, vol 3, pages 650-655, Sep 2000.
[3] D. Borghys and V. Lacroix and C. Perneel, “Edge and Line Detection in Polarimetric SAR Images,”
Proc. ICPR 2002, Quebec, vol 2, pages 921-924, Aug 2002.
[4] D. Borghys, Y. Yvinec, C. Perneel, A. Pizurica and W. Philips, “Hierarchical Supervised Classification
of Multi-Channel SAR Images,” Proc. 3rd Int. Workshop on Pattern Recognition in Remote Sensing
(PRRS’04), Kingston-upon-Thames, Aug. 2004.
[5] S. Cloude and E. Pottier, “An Entropy Based Classification Scheme for Land Applications of Polarimetric
SAR,” IEEE-GRS, Vol.35, No.1, January 1997.
[6] P. Druyts, Y. Yvinec, M. Acheroy, “Usefulness of semi-automatic tools for airborne minefield detection,
” Clawar 98 -First International Symposium, pages 241-248, Brussels, 1998.
[7] P. Druyts, Y. Yvinec, M. Acheroy, “Image processing tools for semi-automatic minefield detection,”
ORS99, Second International Symposium on Operationalization of Remote Sensing, Enschede, Netherlands,
August 1999.
[8] V. Lacroix and M. Acheroy, “Feature extraction using the constrained gradient,” ISPRS Journal of
Photogrammetry & Remote Sensing, No 2, vol 53, pages 85-94, 1998.
[9] S. Mascle and I. Bloch and D. Vidal-Madjar, “Application of Dempster-Shafer Evidence Theory to
Unsupervised Classification in Multisource Remote Sensing,” IEEE Transactions on Geoscience and
Remote Sensing, No 4, vol 35 , pages 1018-1031, 1997.
[10] N. Milisavljevic and I. Bloch, “Fusion of Anti-Personnel Mine Detection Sensors in Terms of Belief
Functions, a Two-Level Approach”, IEEE-SMC, 2003.
[11] G. Shafer, “A Mathematical Theory of Evidence”, Princeton University Press, 1976.
[12] P. Smets, “The Combination of Evidence in the Transferable Belief Model,” IEEE-PAMI, No 5, vol
12, pages 447-458, 1990.
[13] Y. Yvinec, D. Borghys, M. Acheroy, H. Süß, M. Keller, M. Bajic, E. Wolff, S. Vanhuysse, I. Bloch,
Yong Y., and O. Damanet, “SMART: Space and Airborne Mined Area Reduction Tools - Presentation,”
EUDEM2-SCOT-2003 International Conference on Requirements and Technologies for the Detection,
Removal and Neutralization of Landmines and UXO, Brussels, Belgium, September 2003.
5

Metal Detector Modeling
Pascal Druyts
Royal Military Academy
Brussels - Belgium
Pascal.Druyts@elec.rma.ac.be
Risk assessment
• Risk assessment is crucial in HD
– permanent task
– take into account many factors:
• background information (kind of mines, etc.)
• detector capability
• operation procedure
• accidents, QA,...
Detection capability of MD
• Good knowledge mandatory for risk assessment
• Detection capability includes:
– a) Max depth at which a mine may be found
– b) Max depth at which mines are reliably found. PD
close to 100%
• Both criteria are not equivalent
– detector ranking may change with criteria (a or b)
– Deminers are mainly concerned by (b)

Evaluation of MD detection
capability
• On the field:
– put most difficult mine expected at required clearing
depth
– tune sensitivity to ensure a clear alarm
• Controlled tests
– count number of detected mines
• binary test (detection/no detection)
• practicl constraints -> limited mines in limited configurations
– More advanced measurement
• detection margin
• curve detection depth/metal content
Discussion
• Deminers claim near 100% detection
– clear alarm on calibration mine
– condition may change (soil answer, mine
oxydation, …) but if margin sufficient, mines
still detectable
• Claim seems reasonable but not formally
proven. Lack definition of the clear alarm.
Discussion
• Accidents are recorded in database
– the accident rate is low:
• gives a high lower bound on PD
• OP may also be responsible for accidents:
– area not scanned (marking)
– detector not used properly: bad calibration …
– Supports deminer claim but:
• no formal prove
• not proactive:
– OP are evolving. If bad OP, too much time before feedback from
accident records

• Quality assurance
Discussion
– not efficient to guarantee PD
• mine density very low
• area to re-check to guarantee near 100% detection
prohibitivly large
• usually same sensor used. Does only prove no more
detectable mine
– can efficiently check full coverage:
• no detectable metal may be left
• density of metal much higher than mines
Discussion
• Controlled tests:
– show unsatisfactory detection rate (much lower that
100%) in difficult soils even above 10 cm.
– Seems to show deminer claim (close 100% detection) is
false
• Seen as casting doubt on their work by deminers -> critics:
– usually not performed by experienced deminers
– not enough time to get used to all detectors
• Seen by scientist as a proof that 100% detection is:
– not reached by MD
– should not be requested for new technologies (GPR,..)
– This is a WRONG interpretation
Bad interpretation of tests
• tests are maybe not fully realistic nevertheless they
provide very useful information:
– 50% detection depth
• they DO show that in some circumstances detectors
will not find all mines to an acceptable depth
• they DO NOT show that mines are missed during real
demining:
– calibration would show no clear alarm on reference mine
– other demining technique would be used (proding, dogs, …)

Can we prove near 100% detection?
[Gasser]
• Using crude PD (detection/no detection)
– NO
– prohibitive number of cases needed
• Demining is not the only safety-critical
application
– nuclear plants, airplane, …
– reliabilty CAN be proved but NEVER done by
counting the number of occasional failures
– other approach such as detection margin needed
How to prove near 100% PD
• One measurement:
– detection margin bigger than expected
variation expected (soil, mine, …)
– difficult to reliably predict variability
– security coefficient -> very high margin needed
• Measure margin in a representative number
configurations:
– residual variability lower
– lower margin needed
Modeling to prove near 100% PD
• Model to predict detection margin as function of
relevant parameters
• mine depth and orientation
• scan path
• soil relief
• mine and soil EM properties (conductivity, electrical
permittivity)
• measurements to estimate error on model
• measurement to estimate distribution of
parameters

Modeling to prove near 100% PD
• Statistical tools to compute confidence intervals on
estimations
• monte-carlo to estimate PD
– model must be fast enough
– possible to use big number of samples ( real tests)
• Check conclusions of model on:
– test scenarios where model predict 50% detection (50%->
reasonable number of tests)
– test a reasonable sample number a bit above the 100% depth limit
(objective stat. relevance)
– test a reasonable sample number a bit below the 0% depth limit
(objective stat. relevance)
Modeling allows to
• Define objectively what is a reasonable detection
margin obtained from:
– a realistic parameter distribution obtained by
measurement
– an accurate prediction of the precise effect on detection
margin (model)
• With a reasonable number of measurements
– measurements are more informative than detection/no
detection
• Define objectively constraints on scanning path
Conclusion
• PD close to 100% is plausible in demining
• Was never formally proven
• This can be done using a model and well
chosen measurements
• The model must be
– Accurate
– Simple
• Development is ongoing and promising

Extraction of Landmine Signature
from Ground Penetrating Radar Signal
Idesbald van den Bosch
Microwaves UCL
3, Place du Levant
1348 Louvain-la-Neuve
Belgium
vandenbosch@emic.ucl.ac.be
Abstract— With its principle of detection based upon electromagnetic
(EM) wave propagation parameters inhomogeneities
within the medium under consideration, ground penetrating
radar (GPR) is poised to be a very valuable tool in the field
of humanitarian demining, especially for the detection of plastic
landmines. In this regard, however, its performances strongly
depend upon the EM properties of the medium surrounding the
target, and upon the dielectric contrast between the target and
the surrounding medium.
After a short introduction to the world landmine problem,
this article presents the extraction and modeling of buried
targets signatures from a stepped-frequency continuous wave
GPR signal. The signal is decomposed in its soil and targetin-soil
contributions. The mine signature extraction process is
detailed, and the outline of the GPR model used to simulate
these signatures is also presented. More importantly, the reasons
for mine signature extraction and simulation are discussed. This
development is completed by the validation of the numerical
model, which shows encouraging results for the path of research.
I. MINE FACTS AND FIGURES
The following figures can be found at the following
adresses: www.state.gov/t/pm/wra/ and www.
cirnetwork.org/info/fact_sheet.cfm.
A. Mines by the numbers
Landmines
• are estimated to be 45-50 million infesting at least 12
million km 2 of land
• affect more than 100 countries
• heavily affect 20 countries: Angola, Afghanistan, Croatia,
Egypt, Cambodia, . . .
• kill or maim a reported 10,000 people annually (UNICEF:
30–40% of victims are children of 15 or less)
• cost US$3–30 the unit.
B. Heavy economical impact
Landmines
• create millions of refugees
• prevent hundreds of thousands of km 2 of agricultural land
from being used
• deny thousands of km of roads for travel
Sébastien Lambot
Department of Geotechnology
Delft University of Technology
Mijnbouwstraat 120
2628 RX Delft
The Netherlands
s.lambot@citg.tudelft.nl
Marc Acheroy
Signal and Image Centre
Royal Military Academy
Av. de la Renaissance, 30
1000 Brussels
Belgium
acheroy@elec.rma.ac.be
• create food scarcities, causing malnutrition and starvation
• interrupt health care, increasing sickness and disease
• inflict long-term psychological trauma on landmine survivors
• hinder economic development
• undermine political stability.
C. Demining: a slow/costly process
Demining process
• is extremely dangerous: 1 accident/2,000 mines destroyed
• costs US$300–1,000 the unit
• is slow: ∼100,000 mines/year cleared.
So there is a real need for a faster, cheaper and more secure
demining process. It includes, but is not limited to:
• improve currently used detectors
• increase research on new technologies
The list of currently available detectors and prototypes is
presented at table I.
II. RADAR SYSTEM DESCRIPTION
We use a stepped-frequency continuous wave (SFCW) radar
becauses it possesses several advantages over the time domain
technology [1], among which:
• possible to control the signal over a very large bandwidth
(0.75–4 GHz)
• ease of characterization of the elements of the system
(cables, antenna)
TABLE I
LIST OF CURRENTLY AVAILABLE DETECTORS AND PROTOTYPES.
sensor type sensor maturity effectiveness
prodders & prodder in use high
acoustics acoustic R&D high (wet soil)
MD in use very high
EM
GPR R&D/in use high (dry soil)
MWR R&D medium
biosensor
dog
rodent
in use
in devel.
medium–high
medium–high
nuclear all prototype —
chemical all prototype —

• better signal-to-noise ratio (SNR) than for time-domain
GPRs.
The antenna is used off-ground and has the following features:
• TEM horn (high directivity)
• used in monostatic mode (simpler modeling, shorter
round-trip path).
The radar system is emulated by a vector-network analyzer
(VNA). The measured signal is S11(ω).
III. RADAR SYSTEM MODELING
The signal measured by the radar is the sum of two
contributions:
S total
11
= Ssoil
11
+ Starget
11
which can be rewritten using a transfer functions formalism
as (Fig. (b); Hm is neglected):
where
(1)
S total
11 = Hi + H (Gs + Gt) (2)
• Gt is the “signature” of the target
• Gs is the Green’s function of the soil, and can be
computed once (ε1, µ1) , . . . , (εN, µN) are known
• Hi, H are the transfer functions of the antenna.
Gt depends upon many parameters, among others the depth of
the target in the medium and the soil EM parameters. From
(2) we have that
Gt = Stotal 11 − Hi − H Gs
H
. (3)
Gt can be computed thanks to a proper soil-target system
modeling, for which we used the method of moments for
objects embedded in stratified media [2] (not developed here).
IV. MINE SIGNATURE EXTRACTION AND SIMULATION
A. Measured signature extraction process
The extraction of the mine signature is done in three steps:
1) determination the characteristics Hi, H = HtHr of the
antenna by measurements
2) extraction of the soil EM parameters
(ε1, µ1) , . . . , (εN, µN) (Fig. (a)) by inversion of
the radar model for a soil without a target
3) subtraction of the soil contribution from the total radar
signal and division of the difference by H (see (3)).
The advantage of this method w.r.t. a classical extraction by
moving average window is that once the soil EM parameters
are known, the background can be subtracted from the total
signal, for any height of the antenna, as Gs can be computed
once the soil EM parameters are known [3]–[6].
B. Example of measured signatures
Measurements for a soil at different moistures and containing
different targets have been made (Fig. 2). Fig. 2(a) shows
the time-domain S11(t) for a 4-layered propagation medium
containing an AP PMN Russian mine, without filtering out
the soil response; Fig. 2(b) shows the time-domain signature
of the mine after extraction of the soil response in the
frequency domain, which had been previously measured. The
abscissa “configuration” refers to the water content of the layer
containing the objects (1 is 0%, 9 is 25%). In configurations
1–6, the filtered signal shows much more clearly the presence
of the mine, and the time position permits to accurately
retrieve the depth of the target and correspond very well to
the theoretical value. The reasons why configurations 7–9 led
to less satisfactory results has still to be explained.
C. Signature simulation
Simulating mine signatures allows one to:
1) get an understanding of the physics of real experiments
2) set an upper limit for the performances of a GPR given
a physical environment.
The first point is demonstrated at Fig. 3, where a virtual B
scan in free space and in a layered medium are shown. The
characteristic time domain hyperbola appears twice in the right
figure, as the last layer is a PEC metal sheet. One can also
see that the multiple reflections between the top and bottom
of the cylinder are decaying faster in the layered medium, as
the layer containing the target is lossy.
The second point can be inferred from the observation that
in the simulations, no statistical imperfections such as soil
roughness or clutter for example, are taken into account. So,
if the modeling has been validated in quasi-ideal conditions,
we can use it as an upper limit as follows: if, for a given soil
and mine, the simulator gives a “not detected” answer, it is
highly probable that in real field conditions the detector will
also miss the target.
V. RADAR AND SOIL-TARGET MODELS VALIDATION
In this scope experiments involving a conducting sphere
embedded in a stratified medium have been led; the resulting
extracted signatures are compared to simulations at Fig. 4.
The PEC sphere has a 5 cm radius and the 4-layers medium
is constituted as follows: air, sand with 5% moisture and 15.4
cm thick, sand with 10% moisture and 13.5 cm thick, metal
plane.
The reason why the computed signature is smaller in
amplitude than its measured couterpart (see left figure) may
be due to:
• directive radiation pattern of the TEM antenna not accounted
for in the radar model
• remove-and-replace process necessary to introduce the
mine in the soil which introduces air in the sand and
lowers the value of its EM parameters.
The right figure shows these computed and measured signatures
of the sphere in the time-domain. The time dependence

εN , µN
εN−1, µN−1
εN−2, µN−2
ε1, µ1
t (s)
x 10
0
−9
2
4
6
8
10
Ht
S11 = Y
X Y
X
(a) (b)
Fig. 1. Left: physical modeling of the reality; right: conceptual transfer function model
1 2 3 4 5 6 7 8 9
Configuration
20
10
0
−10
−20
S 11 ( t) (−)
t (s)
x 10
0
−9
Hi
Hm
Gs
Gt
1 2 3 4 5 6 7 8 9
Configuration
Fig. 2. Left: measured raw GPR signal; right: measurement extracted mine signature.
2
4
6
8
10
Hr
10
5
0
−5
−10
H o ( t) (−)
air
soil

t (s)
x 10
0
−9
1
2
3
4
5
6
7
8
9
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
relative position [m]
t (s)
x 10
0
−9
1
2
3
4
5
6
7
8
9
−0.6 −0.4 −0.2 0 0.2 0.4 0.6
relative position [m]
Fig. 3. Left: signature simulation of PEC cylinder in free-space. Right: signature simulation of PEC cylinder in 4-layers medium. Both images are
B-scans.
450
400
350
300
250
200
150
100
50
computed
measured
1 1.5 2 2.5 3
x 10 9
0
−2
−4
−6
x 109
8
6
4
2
0
measured
computed
0 0.2 0.4 0.6 0.8 1
x 10 −8
−8
Fig. 4. Measured and computed PEC sphere signature Gt in a 4-layers medium show good agreement. Left: frequency domain; right: time domain.
is correctly predicted; only the amplitude of the first reflection
seems underestimated, probably for the reasons exposed
above. There are two reflections: the first one is due to the
sphere, while the second one is due to the lowest layer, a
PEC plane. This validation process is very important, as the
modeling will serve to establish an upper limit of detection
capabilities for a given soil and mine.
VI. SUMMARY
Landmines cause heavy problems in affected countries, and
demining is costful and slow. It is necessary to improve the
existing demining tools and methods w.r.t. speed, security,
effectiveness and cost. Research has been focused on the GPR,
in measurement information retrieval—that is, target signature
extraction—as well as in modeling.
From the measurement point of view, mine signature extraction
appears promising for GPR signal SNR enhancement
(noise = soil reflections) and is important in view of target
classification. It is based upon the knowledge of the soil EM
parameters, which we can extract from soil GPR signal.
From the modeling point of view, mine signature simulation
could be used for establishing an upper limit of
the GPR detection capabilities for a given soil and mine.
Furthermore it can be used to make “virtual” experiments,
which yield a good “feeling” of what would happen in reality.
A simulated signature simulation has been compared to its
measured counterpart and have shown good agreement.
VII. FUTURE WORK
It includes, but is not limited to:

• better modeling of the antenna radiation properties
• wider testing of the model
• study of the degradation of the signature extraction w.r.t.
soil roughness.
ACKNOWLEDGEMENT
This work has been partially funded by the Belgian Ministry
of Defence (HuDem project) and is partially funded by the
FRIA as well as by the Royal Military Academy of Brussels.
The authors are grateful to Ir. Pascal Druyts for the many
discussions involving this subject.
REFERENCES
[1] I. van den Bosch, S. Lambot, and A. Vander Vorst, “A Unified Method
for Modeling Radar and Radiometer Measurements,” in Proceedings of
the EUDEM2 SCOT 2003 conference, H. Sahli, A. M. Bottoms, and
J. Cornelis, Eds., vol. 2, VUB, Brussels, September 2003, pp. 523–528.
[2] K. A. Michalski and J. R. Mosig, “Multilayered Media Green’s Functions
in Integral Equation Formulations,” IEEE Transactions on Antennas and
Propagation, vol. 45, no. 3, pp. 508–519, March 1997.
[3] S. Lambot, E. C. Slob, I. van den Bosch, B. Stockbroeckx, B. Scheers, and
M. Vanclooster, “GPR design and modeling for identifying the shallow
subsurface dielectric properties,” in Proceedings of the 2nd International
Workshop on Advanced Ground Penetrating Radar, A. Yarovoy, Ed., TU
Delft, the Netherlands, May 2003, pp. 130–135.
[4] ——, “Estimating Soil Dielectric Properties from Monostatic GPR Signal
Inversion in the Frequency Domain,” Water Resources Research, 2003,
in Press.
[5] S. Lambot, E. C. Slob, I. van den Bosch, B. Stockbroeckx, and M. Vanclooster,
“Accurate modeling of GPR signal for an accurate characterization
of the subsurface dielectric properties,” IEEE Transactions on
Geoscience and Remote Sensing, 2003, accepted with minor revisions.
[6] S. Lambot, I. van den Bosch, and E. Slob, “Dielectric characterization
of the shallow subsurface using ground penetrating radar for supporting
humanitarian demining,” in Proceedings of the EUDEM2 SCOT 2003
conference, H. Sahli, A. M. Bottoms, and J. Cornelis, Eds., vol. 2, VUB,
Brussels, September 2003, pp. 535–541.

Detection of Landmines Using Nuclear Quadrupole Resonance (NQR): An Overview
S.D. Somasundaram, J.A.S. Smith, K. Althoefer, L.D. Seneviratne; King’s College London, Strand, London WC2R 2LS
{Samuel.Somasundaram, John.Smith, K.Althoefer, Lakmal.Seneviratne}@kcl.ac.uk
Abstract: Estimates show that there are more than 100 million landmines worldwide and at present rates it will take more
than 500 years and US$33billion to clear them. This has led to an increase in research into new techniques for demining.
Nuclear Quadrupole Resonance (NQR) is one of the few detection methods that is able to detect the explosives found
within landmines. This paper provides an overview of the theory and practical aspects of NQR and its advantages and
limitations. A discussion on recent improved hardware and signal processing techniques that have been proposed to try
and improve the sensitivity of detection is provided.
1 Introduction
1.1 The problem
Demining is an issue affecting both the military and the
general public. According to the “International
Campaign to Ban Landmines”, there are more than 100
million land mines buried in over 80 countries and more
are still being laid. They estimate that if the demining
rate remains at the current rate, it will take over 500
years and US$33 billion to clear them. This problem
has led to an increase in demining research.
1.2 Current demining technologies
The military usually only require a small path through a
minefield to be cleared. One method is one of “brute
force”; landmines are triggered using explosives and/or
heavy machinery. Such a method would not be suitable
for clearing an entire minefield. A common method
used in civilian demining is to detect the mines using
metal detectors or ground penetrating radar (GPR).
Once the mine has been found, it can be carefully
disarmed or destroyed. Unfortunately, many mines are
cased, not in metal, but in plastic or wood and may
contain only a relatively small amount of metal. This
means the sensitivity of the metal detector has to be
increased; consequently lots of other metal objects
(clutter) are detected. GPR has a similar drawback.
Every suspicious object detected has to be carefully
investigated, which usually involves probing the ground
with a “prodder”, a slow, laborious and very dangerous
task. These problems have lead to a worldwide drive
into researching new mine detection methods, which
may be used to detect mines accurately, quickly and
safely within the hostile and varying environments in
which mines may be found.
1.3 Emerging demining technologies
New mine detection methods generally fall into one of
two categories. The first category is where the method
seeks to detect the mine casing or, other anomalies in
the ground introduced by the buried mine (table 1). In
the second category, it is the actual explosive within the
mine or its vapour that is being detected (table 2).
Demining Technology Description Main Limitations
Ground penetrating radar (GPR) Images dielectric of the soil,
discontinuities such as mines can be seen.
Infra-Red/Thermal Imaging Images the temperature of the ground.
Soil above mines cools at a slower rate
than surrounding soil.
Soil is generally inhomogeneous,
therefore it is not necessarily that
easy to see mines, particularly near
the surface.
Use is limited to certain times of the
day.
Table 1 – Methods for detecting anomalies in the ground (caused by the mine) or the mine casing.
Demining Technology Description Main Limitations
Nuclear Quadrupole Resonance After excitation, detect a frequency Received signal is very weak.
(NQR)
signature specific to the explosive in the
landmine.
Thermal Neutron Activation Detects presence of nitrogen in sample by Uses ionising radiation and requires
(TNA)
irradiating with slow neutrons and then
measuring energy levels of returned
gamma rays
heavy equipment.
Mine Detecting Dogs Dogs trained to “sniff” out landmines. Easily tired.
Table 2 – Methods for detecting the explosives within landmines.

One point worth noting is that the equipment needed for
NQR detection of landmines is relatively light and
simple compared to GPR, TNA and IR detection
systems and could therefore be easily mounted on some
sort of robotic device. Researchers at King’s are
currently investigating the feasibility of integrating
various mine detection methods, including NQR, onto
robotic vehicles.
2 NQR
2.1 Theory
Nuclear quadrupole resonance is a radio frequency (RF)
technique that may be used to detect the actual
explosives found within landmines. Many explosives
used in the manufacture of landmines are nitrogen-rich
(e.g. TNT, PETN, and RDX). The 14 N nitrogen nucleus
has spin quantum number I = 1, and thus behaves as
though it has a non-spherical charge distribution and
therefore possesses an electric quadrupole moment, Q.
When such a nucleus is placed in a non-zero electric
field gradient (EFG) different quadrupole energy levels
arise as some nuclear orientations will be at lower
energy levels than others. The EFG at the nucleus is
produced by neighbouring charges, both electrons and
nuclei, and is therefore directly related to the structure
of the compound. The symmetry of the EFG, the
nuclear electric quadrupole moment and the spin
quantum number of the nucleus determine how the
energy levels are split. In NQR, the applied RF
radiation drives transitions between the quadrupole
energy levels. As 14 N has spin-quantum number I = 1,
there are three energy levels -1, 0 and +1. The EFG is a
tensor quantity and its three principal components can
be expressed in Cartesian coordinates as qxx, qyy, and qzz.
As the sum of these three components is zero it is usual
to define qzz or q as the maximum principal component;
the deviation from axial symmetry with respect to the zaxis
of the EFG is given by the asymmetry parameter η
a positive number varying between zero and unity and
defined as
qxx − qyy
η = .
(1)
qzz
The energy of interaction between the electric
quadrupole moment of the 14 N nucleus and the EFG of
the surrounding charges is given by solving the
quadrupolar Hamiltonian for a spin-1 nucleus [2]. In the
general case, three transitions are allowed (figure 1)
with frequencies:
Ey
− Ez
vx
( 0 → + 1)
=
h
3 e
2
qQ η
= ( 1+
),
4 h 3
E
( 0 1)
x − E
v
z
y → − =
h
3 e
2
qQ η
= ( 1−
),
4 h 3
Ey
− Ex
1 e
2
qQ
vz
( −1
→ + 1)
= = η,
h 2 h
(2)
where e is electronic charge; h is Planck’s constant and
e 2 qQ/h is known as the “quadrupole coupling constant”
[1]. These frequencies lie in the radio frequency region;
hence electromagnetic radiation in this range interacts
with the quadrupolar nuclei. It is worth noting that the
NQR signals are observed only in the solid state.
Figure 1 - Quadrupole energy levels in axially
symmetric (top) and non-axially symmetric (bottom)
fields.
2.2 Experimental methods
Currently, NQR experiments are conducted using
pulsed RF techniques. The sample is subjected to bursts
of RF radiation at or near the resonant frequencies and
the resulting signals monitored. Two types of signals
that are commonly monitored are free induction decays
(FIDs) and echoes.
2.2.1 Free Induction Decays (FIDs)
When an RF pulse at the resonant frequency, vQ, is
applied to the NQR sample at the correct geometry, the
magnetic component of the RF radiation couples with
the nuclear magnetic moment and alters the orientation
of the 14 N nucleus within the EFG. This effectively
excites the nucleus, as the nuclear magnetisation
precesses about B1 away from its equilibrium
orientation.

Figure 2 – Interaction of RF radiation with nuclear
magnetic moment
Once the RF pulse stops, the nucleus returns to a lower
energy level via various relaxation processes. This is
observed as a free induction decay (FID) immediately
following the end of the pulse, which consists of a
damped sinusoidal oscillation that is observed to decay
with a time constant T2 * , known as the FID or spinphase
memory decay time.
The amount by which the nuclear magnetic moment is
“tipped” away, by the radiation, from its equilibrium
position in the EFG, is given by the flip angle α (figure
2), which is defined by,
, t B γ α = (3)
1 w
where γ is the nuclear gyromagnetic ratio, B1 is the
magnitude of the oscillating RF field and tw is the time
for which the pulse is applied.
The FID intensity depends on the flip angle and is
maximum when α =119°, and not at 90° as in pulsed
NMR. This is due to the NQR sample being
polycrystalline, which means that the EFG at a given
nitrogen site can assume all possible orientations with
respect to the B1 field. It is customary to refer to a 119°
degree pulse as “90ºeff”. Equation (3) shows that the
hardware must be configured to deliver an RF pulse of
sufficient magnitude and for long enough to produce a
90ºeff pulse and get the maximum FID intensity.
One of the processes by which the nuclei lose energy is
governed by the spin-lattice relaxation time, T1, by
which the nucleus returns to its equilibrium orientation
in the EFG by losing energy to the thermal motions of
the solid. Following a pulse, a time of 5T1 is required to
produce a fully relaxed system.
2.2.2 Spin Echoes
Another set of signals that are usually monitored are
spin echoes, which are also useful as they can be used
to sustain the NQR signal for longer than the FIDs
described earlier. The long “echo trains” that are
produced can be summed to reduce the signal to noise
ratio (SNR).
There are several relaxation processes that contribute to
the decay of the FID. One of these processes is caused
by the inhomogeneous nature of the sample, which
means that the EFG at each of the 14 N sites can vary
slightly, so that their respective NQR frequencies also
vary slightly. This means that the signals from the 14 N
nuclei become out of phase with each other and
therefore lose coherence. It is, however, possible to use
certain pulse sequences to refocus the signals and bring
them back into phase. This is the basis for the formation
of echoes. An echo can be produced by applying a
second pulse a timeτ after the initial pulse, often with
its phase shifted by 90° (with respect to the first pulse).
This has the effect of refocusing the de-phased signals;
consequently, a time τ after the second pulse, the 14 N
signals are all back in phase and the peak of the echo
corresponds to this point. If we keep applying
successive pulses, then an echo train is produced, a
sequence known as pulsed spin locking (PSL). The
echo train cannot be sustained indefinitely as the nuclei
dephase; the time constant for this process following
two pulses is T2, the spin-spin relaxation time.
However, in a PSL sequence for example, an even
longer decaying echo train is produced with a time
constant T2e. Once the amplitude of the echoes has
decreased below noise, an experimental wait time of
5T1 must be allowed before the sample can be excited
again.
2.3 Landmine NQR Data
85% of landmines are manufactured using RDX and/or
TNT. Table (3) tabulates some of their NQR properties
at room temperature. Note that the last three columns of
table (3) list average values.
Compound v x (kHz) y v (kHz) T 1 (ms) T 2 (ms) Line width (kHz)
RDX 5047, 5192, 3359, 3410, 13 7 0.64
5240
3458
TNT 837, 843, 844, 714, 739, 743, 5000 - 1.2
(monoclinic
form)
848, 859, 870, 751, 769
Table 3 – NQR data

2.4 Signal to Noise Ratio (SNR)
In NQR, the SNR can be expressed by the equation:
1/
2 2 3/
2 1/
2
VCQ
Aζh
γN0vQ
μ0
SNR = (4)
1/
2
30. 08kT(
kTBF)
which needs to be multiplied by a factor of 0.436 for
polycrystalline samples. VC is the RF coil volume, Q is
the loaded Q factor, A is a constant close to unity, ζ is
is
the coil filling factor, h is Planck’s constant, γ
the
nuclear gyromagnetic ratio, N0 is the number density of
the 14 N nuclei, vQ is the NQR frequency, μ0 is the
permittivity of free space, k is Boltzmann’s constant, T
is temperature, B is the receiver bandwidth and F is the
noise factor.
The formula was derived assuming a solenoid type coil
and the assumed noise is the thermal noise of the
hardware. Two points worth noting are that the greater
the NQR frequency the greater is the SNR and the
higher the Q factor the greater is the SNR.
2.5 Limitations of current NQR detection
methods
The main problem with NQR is that the signals detected
are often weak compared to the ambient noise floor.
The main sources of noise are:
1. Thermal (Johnson) noise of the RF antenna
and sample noise.
2. Radio frequency interference (RFI).
3. Spurious responses e.g. piezoelectric and
magnetoacoustic responses from sand in the
soil.
In the laboratory, sample noise at these low
radiofrequencies is less than the random thermal noise
of the RF coil [8] which then becomes the limiting
factor, as the sensor can be shielded from RFI and
substances which produce spurious responses can be
removed from the sample. Equation (4) shows that
SNR ∝ v . The data in table (3) shows that the TNT
3/
2
Q
frequencies are less by a factor of at least four than
those of RDX, so their SNR is less at least by a factor of
about 4 1.5 =8. Since the sensitivity of detection is poor,
the SNR must be improved by repeating the experiment
and averaging the signals obtained. In section (2.2) we
mentioned that a wait time of 5T1 needs to be observed
between the last excitation pulse in one experiment and
the first excitation pulse in the next. The T1 values for
TNT and RDX, shown in table (3), indicate that this
quantity is around 13ms for RDX, whilst for TNT its
value rises to 5 seconds. For landmine detection using
NQR, we would like to be able to determine whether a
mine exists (in the area being investigated by the
antenna) in less than a minute, which means we may
only be able to collect 2 echo trains from a TNTcontaining
mine, insufficient to give an acceptable SNR
for many anti-personnel mines.
The problem of RFI is not included in equation (4) and
this further reduces the SNR in field experiments where
it is not possible to fully shield the sample under
investigation. Automobile ignition, radio station
transmissions and lightning are just a few of the
possible sources of RFI. These are commonly found in
the AM band, which is specifically a problem for the
detection of TNT, which has NQR frequencies lying
between 700 and 900kHz [1]. These problems show
that the detection of TNT presents one of the toughest
challenges for NQR based landmine detection systems.
In addition, the presence of certain substances in the
soil, e.g. quartz, can generate strong piezoelectric
signals that can completely mask the weak NQR
response [1]. Magnetic materials and shrapnel in the
terrain may also generate strong magnetoacoustic
responses.
Another problem is not being able to detect the
FID/echo immediately following the excitation pulse
due to “ringing” of the coil and associated hardware.
Once the excitation pulse has stopped, the coil
continues to ring for a time of the order of 23 time
constants before NQR signal can be detected.
Unfortunately, valuable NQR signal is lost during this
time, known as the dead time.
3 Dealing with NQR limitations
There are several approaches to dealing with the
limitations discussed, but they can be broken down into
the two following broad categories,
1. Improved antenna/hardware design
2. Improved signal processing techniques.
A discussion on recent technologies follows.
3.1 Improved antenna/hardware design
Gradiometer coils have been used to reduce the
problems of RFI. A gradiometer coil has the property
that it is a poor antenna for distant sources, but not for
nearby sources. Unfortunately, with simple
gradiometers, some signal loss occurs. This problem is
partly solved by the use of asymmetric gradiometers.
Here the dimensions of the antenna are optimised so
that the field received in a particular inspection volume
is optimised, whereas fields from outside this region are
rejected. Suits et al. [6] have devised a computational
method for optimising single and two-layer surface

coils. The general approach is to find the solution to
Maxwell’s equation, which optimises a certain
parameter. Then the current densities that produce such
a field are calculated allowing the construction of an
optimised coil.
Equation (4) shows that SNR∝ Q 1/2 . This and the advent
of high temperature superconducting (HTS) ceramics
are the main reasons for investigating the use of cooled
coils with high quality factors, which we are studying at
the present time. To produce a high Q coil the
resistance of the coil must be reduced, which also
reduces the noise term in the denominator of equation
(4), achieved by either cooling the coil or using
superconducting materials. In magnetic resonance
imaging (MRI) experiments, SNR enhancements by a
factor of 3 or more have been reported from the use of
HTS antennae cooled to 77K [8]. However, additional
obstacles must be overcome. Equation (5) relates the
quality factor to the dead time td.
( πvtd
)
Q = (5)
ln( A / A )
0 d
A0 is the initial circuit voltage immediately following
the pulse, which could be as high as 10kV, and Ad a
value less than the signal (typically < 1µV).
Thus the higher the Q factor, the greater the dead-time
required as the coil rings for longer. For high Q values
this could pose a significant problem as the NQR signal
could have decayed to negligible levels by the time the
hardware could be set to receive signals. Another
problem is caused by the fact that the bandwidth of the
system is inversely related to the Q factor. The NQR
frequency is temperature dependant, and under some
soil conditions may move outside the range of the
instrument.
The use of a circularly polarised RF magnetic field has
been investigated as a way to increase the amount of
signal detected, reduce the problems of spurious
responses and improve multiple pulse sequences [5]. In
standard NQR experiments, RF is applied along one
axis only, hence only nuclei whose EFGs are correctly
aligned with this axis will see the RF field. If circularly
polarised RF is used, say in the x-y plane, then nuclei
aligned with the x and y axes will see the RF and hence
more nuclei will be excited. Suits et al. [5] reported a
net gain in the SNR of 21%. The use of circularly
polarised RF also leads to a significantly more
homogenous RF field, which can lead to improvements
in the effectiveness of multiple pulse sequences. The
technique may also be used to distinguish interfering
signals, such as piezoelectric responses, from the NQR
signal. The NQR signal from a powdered sample will
have the same polarisation as the excitation pulse,
whereas the interfering signal will be linearly polarised
along an axis dependant upon its orientation, hence the
signals can be distinguished. The main problem with
implementation of such a technique in landmine
detection is the fact the sample must be surrounded by
at least two mutually orthogonal coils.
Three-frequency NQR has been suggested as a way of
alleviating the problem of “coil ringing” [5]. Here, the
NQR sample is excited at two out of the three NQR
frequencies and the third NQR frequency is observed.
Unfortunately, this requires three coils which are
mutually orthogonal to one another and thus poses
practical problems for implementation in a landmine
detector.
4 Signal Processing
Essentially, the aim is to decide whether a mine is
present or not, using the data collected in one minute or
less. Signal processing is a vast field and there are many
ways of tackling the problem.
As the characteristic NQR frequencies are known, one
approach is to Fourier transform the time domain signal
into the frequency domain and obtain a measure of the
energy of the signal over the different frequencies. A
threshold is then applied at the NQR frequencies to
determine whether explosive is present or not. This is
known as peak energy detection. This method is fine
when the signal to noise ratio is relatively high, and the
NQR signal is significantly higher than any spurious
responses. However, if the energy of the noise at the
NQR frequencies is of the same magnitude (or greater)
than the energy of the NQR signal then this method is
not very useful. Work has been done in the field of
spectral estimation. Classical spectral estimators based
on periodogram methods suffer when noisy truncated
data sets are used as they do not assume any knowledge
about the signal whose spectrum is being estimated.
However, parametric spectral estimators which assume
some knowledge of the signal to be estimated can be
used to exploit the different characteristics of the noise
and the NQR signal. In laboratory experiments, where
the sample can be shielded from RFI and substances
that produce spurious responses can be removed, the
observed NQR signals can be modelled as a set of
sinusoids buried in white noise. Parametric methods for
line spectra, such as the matrix pencil method and the
multiple signal classification (MUSIC) method, can
then be employed. We have shown that matrix pencil
methods can be used to improve the sensitivity of
detection [9]. L. Collins et al. showed that improved
detection performance could be achieved by using the

MUSIC method instead of periodogram based methods
[4].
One approach to reduce the RFI is by using adaptive
noise cancellation algorithms. These use two sets of
antennae. One main antenna set measures the NQR
signal and the RFI, whilst the reference antenna set
measures the RFI only. Algorithms are applied to the
reference set and are used to obtain an estimate of the
RFI. This estimate is then subtracted from the main
antenna data. L. Collins et al. used a 2-tap least mean
squares algorithm to reduce the RFI power by more
than 50dB [3].
Y. Jiang et al. [10] consider the case of array signal
processing, where both spatial and temporal
information is available. They have used maximum
likelihood (ML) and Capon methods for signal
amplitude estimation in the presence of temporally
white, but spatially correlated noise. Both estimators are
asymptotically statistically efficient for large data
lengths, but not when the SNR is high. The Capon
estimate outperforms the ML estimate when the SNR is
low. They also consider the more general case, where
the noise is both temporally and spatially correlated,
which they model as an autoregressive random process.
They present an alternating least squares method for
parameter estimation and show this to be superior to the
ML method.
The problem of piezoelectric responses generated by
quartz has been alleviated using phase cycling
techniques [1]. These rely on the fact that the phase of
an NQR signal depends on at least two of the preceding
pulses, whereas that of the piezoelectric response is
determined entirely by the phase of the preceding pulse
which generated it. Unfortunately, since the
characteristics of the piezoelectric responses change
with time, they cannot be entirely mitigated using these
techniques.
5 Conclusions
This paper presented the basic theory behind NQRbased
landmine detection and highlighted some of the
problems with this approach. Hardware and software
methods that have been proposed to improve the
sensitivity of detection have been presented.
6 Acknowledgements
We thank DSTL at Fort Halstead for supporting this
project.
7 References
1. J.A.S. Smith and M.D. Rowe, “Mine detection
by nuclear quadrupole resonance”, 1996,
Proceedings of the Eurel International
Conference on the Detection of Abandoned
Landmines. IEE Conference Publication No.
431. 62-66.
2. J.A.S. Smith, “Nuclear Quadrupole
Resonance Spectroscopy”, 1971, J. Chem.
Educ., 48, pp.39, A77, A147 and A243.
3. S. Tantum, L. Collins and L. Carin, “Signal
Processing for NQR discrimination of Buried
Landmines”, 1999, Proceedings of SPIE
Vol.3710.
4. Y. Tan, S.L. Tantum and L.M. Collins,
“Landmine Detection with Nuclear
Quadrupole Resonance”, 2002, Geoscience
and Remote Sensing Symposium, IEEE
International, Vol.3, pp1575-1578.
5. B.H. Suits, A.N. Garroway, J.B. Miller and
K.L. Sauer, “ 14 N magnetic resonance for
materials detection in the field”, 2003, Solid
State Nucl. Magn. Reson. 24, 123-136.
6. B.H. Suits and A.N. Garroway, “Optimising
surface coils and the self-shielded
gradiometer”, 2003, Journal of Applied
Physics, Vol.94, 4170-4178.
7. Pound, Phys. Rev, 1950, 79, 685.
8. L. Darasse and J.C. Ginefri, “Perspectives with
cryogenic RF probes in biomedical MRI”,
2003, Biochimie, 85, 915-937.
9. Y.Y. Lin, P. Hodgkinson, M. Ernst and A.
Pines, J. Magn. Reson., 1997, 128, 30-41.
10. Y. Jiang, P. Stoica and J. Li, “Array Signal
Processing in the Known Waveform and
Steering Vector Case”, IEEE Transactions on
Signal Processing, 2004, Vol.52, No.1, 23-35.

THE PMAR LAB IN HUMANITARIAN DEMINING EFFORT
1 Introduction
E.E. Cepolina, R.M. Molfino, M. Zoppi
PMAR Robot Design Research Group
University of Genova - Dept. of Mechanics and Machine Design
{emacepo, molfino, zoppi@dimec.unige.it}
www.dimec.unige.it/PMAR
According to current estimates, at the moment up to 90 countries throughout the world are affected by
landmines [1]. Land affected by mines is inaccessible for long time (basically until fields have been
cleared) and, besides mine risk education efforts, the number of injuries stays very high for long after
landmine placement.
In order to give their contributions in the field, researchers in engineering disciplines can direct their
efforts into two different fields: effective techniques for mine localization and clearance and
mechanical tools for victims care.
Some premises are necessary. As manual mine clearing is very slow, costly and quite dangerous for
personnel, the benefit of a possible introduction of automatic systems and devices is remarkable.
Landmines are disseminated in ex-confrontation lines that divided military factions as river banks,
abandoned industrial sites, residential areas as well as in strategically important sites or resources such
as cultivable ground. Due to this large variety of operative environments, the field is open to specific
environment oriented conventional techniques as well as to unconventional solutions, i.e., new
locomotion principles and new localization strategies.
People living in affected areas often have no choice other then entering minefields for sustaining their
families, i.e. for harvesting fruits, cultivating land, crossing the boarder for working in the near richer
country, and the majority of accidents related to landmines occur when people deliberately enter such
areas. For example, in Cambodia, in 2002, 33% of incidents happened while the victim was tampering
with the mine for extrapolating explosive for selling, the 13% of incidents happened when the victim
was farming and another 13% when he was collecting wood [2]. The need of prosthetics for upper
limbs and legs in mine affected countries is high.
The effectiveness of any technique or tool aimed at developing countries depends on its acceptability
by the local people who will use it. Therefore, taking into account the expectations and capabilities of
local end-users, from all technical, psychological, and cultural (anthropological) points of view, is a
fundamental premise to successful solutions.
This paper aims at presenting the efforts of the researchers of the PMARlab of the University of
Genoa into the development of technical solutions to landmine related problems, friendly to localusers.
2 The PMARlab profile
The PMARlab is the laboratory of Design and Measurements for Automation and Robotics of the
University of Genoa, Italy. The laboratory has large research interests. The main research areas are:
Robotics, Intelligent Automation and Measurements. The aim of the lab is to constantly support a
didactic and research environment to face the requirements of analysis and design of mechatronic
systems during their whole life-cycle, with a particular attention to the development of theoretical
methods. Moreover, the lab has specific skills in modeling and simulation, CAD and virtual
prototyping.
From 2000, the research interests of the lab has enlarged to humanitarian issues. In particular, the lab
is engaged in design of machines and systems for humanitarian demining in uneasy accessible areas
covered by thick vegetation [3-5], in development of prostheses for upper limbs [6] and in design of
rescue robots using unconventional locomotion, e.g., peristaltic [7].
The research work inside the laboratory is shared between researchers and students in order to reach
research results and, at the same time, to form the students. Each year within the courses on Robots
Mechanics and Industrial and Service Robotics a specific subject is proposed to the students and
interdisciplinary teams are composed: after a deep discussion about user needs, literature proposals
and market search, the design work starts. Each team works in parallel performing the design of the

specific module it is in charge of taking into account the pre-defined skeleton layout and needed
interfaces.
3 Design methodology
Hereafter the service robots design methodology [8-9] adopted within the PMAR laboratory is shortly
outlined. During the first part of the project, the combined knowledge, expertise and experience of the
researchers and of the local field experts and practitioners are used to know user needs [10], to derive
specifications and eco-requirements [11], and to set robotic devices performances. A comprehensive
gathering of information from scientific literature, as well as from expert humans knowledge is done
to guide in the definition of the technologies and methodologies for system design and development.
The large use of physical phenomena modelling [12], computer simulation and virtual reality testing is
then scheduled, to provide the throughout characterization of the artefact life-cycle behaviour [13];
this will allow to test, at the design phase, competing architectures (of the mechanical structure, of the
controller, of the sensorial system) and find out those improving the overall figure of merit.
Simulation, in fact, after throughout investigation of achievements and drawbacks, offers affordable
commitment, making possible to rank competing robotic solutions (Fig. 1a). The computer simulation,
moreover, offers, during the system utilization phase, an important aid about the overall work
management and tasks allotment; taking into account the environment-robot interactions
characteristics and the system’s current statics and dynamics, the simulation will be helpful to define
the operating tasks. Kinematics and dynamics models are very useful for control laws setting, mainly
if high performances are required and non-linear, model based control systems have to be
implemented. Static models give important information about low motion intrinsic stability.
USER NEEDS
mechanisms,
geometry, materials,
sensorial system
actuators
control logics
specificated
desired functions,
RAW
MATERIAL
DESIGN
Function
SYNTESIS
MANUFACT
URE
(a) (c)
Form
System
Representation
ANALYSIS
(b)
ASSEMBLY
MODELING
verification of the
current state of
the design by:
modeling
simulation
virtual mock-ups
Behaviour
CUSTOMER
SERVICE
REUSE
RECYCLE
DISPOSAL
ENVIRONMENT
Figure 1 - Design approach and methodology (a) and relationship between form, function and behaviour of a system (b).
The mathematical models are defined and implemented in suitable software modules to be interfaced
and included within the general purpose CAD/CAE packages. These modules will be useful for
motion planning and mission execution purposes. Based on the selected configuration, a reduced
performance digital prototype with basic control features is implemented in order to assess the system
performances and to guide in further development work, while reducing risks and relevant cost of
wrong physical prototypes.
Throughout all the design phases mechatronics methodology and modularity issues [14] are used to
achieve an optimal design of the e-mechanical product. As a design philosophy, mechatronics serves
as an integrating approach to engineering design [15]. It is important for the designer to be able to

simulate the behaviour of the current state of the design: as the design evolves its form, behaviour and
function should be consistent with each other (Fig. 1b). Modularity allows economical and easy set-up
and maintenance of several configuration: a modular robot is built from physically separate sub-units,
each contributing to the operation of the robot.
Digital mock-up and graphic rendering, directly connected with CAD tools, enhance concept design,
moving up life-cycle (Fig. 1c) assessments by virtual prototypes let, so, to devise the optimal layout
and the best mechanical architecture of the robotic system [16] subject to specific cost and
performance constraints. Preliminary investigation by digital mock-up is becoming the engineering
mean to range the technological appropriateness and leanness between competing solutions.
4 Robots designed for rescue and landmine localization
The most of the realized and working systems for landmine localization is designed for terrains clear
and, frequently, enough even and consistent. So far, effective solutions for difficult terrains, very
uneven, rocky and covered by thick vegetation, are lacking.
For such difficult terrains and environments traditional machines are precluded due to size or shape
and appendages such as wheels or legs may cause entrapment and failure.
The PMARlab studies robotic solutions based on biological locomotion suggested by the nature. The
aim is to single out the locomotion mechanical principle and laws to be used in the design of bio
inspired mechanisms.
The main applications of these bio-inspired robots are the mines localization in difficult terrains and
rescue. The robots are thought as mobile devices equipped with suitable sensors in order to fulfill
specific inspection tasks.
This approach led the PMARlab researchers to consider and study unconventional locomotion systems
and propose innovative robotic solutions by exploiting the recalled principles of mechatronic and
modular design. This section is devoted to the presentation of some of these solutions.
Figure 2 [4] shows some tele-operated robotic modules based on different locomotion principles.
a
b)
c)
Figure 2 – Virtual mock-ups of smart crawling robots for landmine localization in thick vegetation using REST (Remote
Explosive Scent Tracing).

The locomotion of the module of Fig. 2a is performed by means of peristaltic movements that provide
trust, while grip is provided by needles that are put inside/outside the external envelope. Figure 2b
presents a module trusted by counter rotating screws, fit for muddy and sandy grounds. The module in
Fig. 2c is trusted by four needle crawlers; it is equipped with suitable detaching wheels for foliage
entangled in the needles of the crawlers. All these modules are powered by means of an umbilical that
also includes wires for signals data. It is noteworthy that these modules have been designed in close
collaboration with students involved in robotic courses.
A further activity recently launched at the PMARlab is the simultaneous design of a modular worm-like robot
by the Pro/Intralink facility by PTC. Different student teams work on the same project: each one is charged of
the design of a specific module. The different types of modules can be combined to satisfy different mission
requirements. Typically modules are designed to perform an elementary function, such as one or more degree
of freedom motion or service: e.g. trust modules (generating the peristaltic contractions-relaxations); steering
modules (bending the snake to steer it in a desired direction); sensor modules (segments that host sensors,
eventually endowed with some degree of freedom); head modules (with a camera, eventually carrying tools
such as grippers or sampling tools).
Figure 3 – Robotic modules for steering and peristaltic trust.
Figures 3 shows the motion cycle of a worm segment composed by three 3dof modules actuated by SMA
springs. Each module can contract and relax or can bend in any direction to steer the snake. The central steel
spring gives shear and torque rigidity to the module. The dark rubber bellows on each module preserve the
inside from any dirtiness coming from the environment, such as water or ground. Moreover, these bellows
provide grip by swelling and contracting depending on the distance between the two bases of the segment
(that is a function of the length of the system of springs). It is possible to arrange plug&play connectors for
signals and power on the interfaces in order to get smart interfacing of the modules.
Figure 4 shows some preliminary physical prototypes of the module.
Figure 4 – Robotic modules for steering and peristaltic trust.
Figure 5 – Conceptual design of other new trust peristaltic modules.

Finally, Figure 5 shows the conceptual design of some innovative peristaltic modules. Small elastomeric
umbrellas are opened and closed by suitable inner mechanisms with flexible joints. These umbrellas provide
grip both by enlarging outside the external diameter of the snake and, mechanically, by biting the ground with
the edge of the umbrella. The result is a one way movement (a preferential direction of motion, from left to
right in the pictures). Once it is asked for recovering the snake, all umbrellas of the trust modules are closed
and the snake is pull from the umbilical. The actuation is again by SMA elements placed inside the modules.
Figure 6 – Simulation of the snake robot trusted by a pushing machine; this robot has been designed for minefield area
reduction by REST (Remote Explosive Scent Tracing).
Figure 6 shows the virtual mock-up of a worm robot (co-bot) [5] trusted by a pushing machine. The
idea is to exploit the bending stiffness of the worm to move it forward no matter the nature of the
ground and independently from the presence of obstacles that would make difficult any distributed
locomotion. The body of the worm does not contain mechanics for locomotion, so it presents lot of
space to host sensors. This is one of the reason why this robot can be fruitfully used for landmine area
reduction by REST (Remote Explosive Scent Tracing). The sampling filters can be rolled up inside the
segments and, once known the geometry of the robot and the number of segments that have been
inserted, it is simple to compute the location of each sample filter with respect to the pushing machine.
So doing it is possible to scan an area (by inserting the snake many times on parallel paths, collecting
each time the filters and registering their location at the moment of the sampling). Once again, the
researches and the design have been carried on involving students of robotic courses and one master
thesis.
5 Design of prostheses for third world countries
Prostheses for people with disabilities, who live in developing countries, should be designed with
requirements different from the ones for developed countries. The main characteristics of such prostheses
shall be: *inexpensive to produce, purchase and maintain (local available or easy to purchase materials and
technical skills), *easy to use, *effective and *designed in consultation with users in a way well suited to the
users' diverse social and physical environments [17]. Other important concerns are the use of human body
energy for the actuation and underactuated mechanisms to simplify the mechanics.
Figure 7 – Prototype of underactuated hand, with mechanics and actuation integrated in an elastomeric structural matrix.
Figure 7 gives an idea of the outcomes of this approach. The figure shows the underactuated hand
prototype realized at the PMARlab in 2003 [6]. Main characteristics of this prototype are a very low
cost, thanks to the shrewd use of everywhere-available material and very easy skills to realize it. The
working principle is simple and there are no constrains about the type of actuation and control, in the
first prototype body energy has been considered for actuation. This hand can fully adapt to every
object shape, but it is not fit to perform manipulation tasks. In fact, dexterity is not a leading
requirement compared to the others such as cost, locally production and time to use. The elastomeric
polymer chosen for the structure, the TST MS 939 from Locktite (probably to be imported), is soft to
the touch and its consistence reminds the one of the natural hand, this improve the acceptability from

the users. The main drawbacks are a low grasping force (but this depends on the implant of the
prosthesis and can be enhanced) and the weight that is a little bit heavy. Design of prostheses for third
world countries is now the subject of a PhD Thesis.
6 Conclusions
The PMAR lab carries on a branched research efforts in humanitarian demining and associated fields.
In particular the PMAR lab is collaborating with: HUDEM IARP work group, EUDEM2, GICHD. At
the moment we are actively involved in the study “Providing demining technology end-users need”:
assessment of end-users needs for humanitarian demining technologies through extensive collection of
data in the field.
Acknowledgement
We would like to thank very much EUDEM2 for the support they are providing within the
collaboration on the study “Providing demining technology end-users need”. We would like to thank
GICHD for constantly giving us high technical level advises. The teams of the 2004 Industrial and
Service Robotics course: F. Canonica, A. Ferrara, E. Micheli, L. Rimassa and M. Sawusch are kindly
acknowledged for the help in the development and design of part of the robotic modules.
References
[1] Landmine Monitor Report: Toward a Mine-Free World, International Campaign to Ban Landmines
(http://www.icbl.org), 2002;
[2] National Census of the Victims and Survivors of Landmines and Unexploded Ordnance in Cambodia,
Cambodian Red Cross, 2002.
[3] E.E. Cepolina, Low cost robots for landmine location in thick vegetation. On-site IARP Workshop on Robots
for Humanitarian Demining: HUDEM 2003, Prishtina, June 19, 2003
[4] E. Cepolina, M. Zoppi, Cost-effective robots for mine detection in thick vegetation. CLAWAR03, September,
17-19, Catania, Italia, 2003.
[5] E.E. Cepolina, Worm: a simple robotic solution to landmine location. Int. Conf. on Requirements and
technologies for the Detection, Removal, and Neutralization of Landmines and UXO, Brussels September 15-18
[6] E. Ullman, R. Molfino, M. Zoppi, A smart and low-cost hand-like prosthesis for amputees. Poster at 35th
Intl. Symposium on Robotics, ISR 2004, Paris, Nord Villepinte, March 23-26 2004
[7] F. Cotta, F. Icardi Non conventional strategies for peristaltic locomotion, Tesi di Laurea, University of
Genova, April 2004
[8] R.D. Schraft, M. Hagele, R. Dillmann (1997): Technologies, state-of-the-art and challenges in service robot
technology, design and applications, IARP International Advanced Robotics Programme service and personal
robots: technologies and applications, Genova, Italy, 23-24 October 1997
[9] P. Anthoine, L.E. Bruzzone, F. Cepolina, R. Molfino: Metodologia di progettazione meccatronica per
robotica di servizio, 1a Conferenza Nazionale “Sistemi Autonomi Intelligenti e Robotica Avanzata”, Frascati,
Italia, 29-31 Ottobre 2002
[10] E. E. Cepolina, C. Bruschini and K. De Bruyn Providing demining technology end-users need Survey on
minefields – Methodology & Aims, EUDEM2, March 2004
[11] S.B. Billatos, N.A. Basaly: Green technology and design for the environment, Taylor & Francis, 1997.
[12] Oelen, W., Modeling as a tool for design of mechatronic systems – design and realization of the Mobile
Autonomous Robot Twente Ph.D.- thesis, University of Twente, 1996
[13] C. Luttropp, J. Lagerstedt: Customer benefits in the context of life cycle design, Eco-Design '99: 1st Intl.
Symp. on Environmental Conscious Design and Inverse Manufacturing, Tokyo, Feb., 1999
[14] S. Farritor, S. Dubowsky, N. Rutman, J. Cole A Systems-Level Modular Design Approach to Field Robotics
Robotics, IEEE International Conference on Robotics and Automation, Minneapolis, MN, 1996, .Vol. 4, pp.
2890-5.
[15] P. Lambeck, B. Bertsche G. Lechner Creating Size Ranges Easily Using An Integrated Design Tool The
Sixth International Conference on Mechatronic Design and Modeling Cappadocia – Turkey, September 4 - 6,
2002
[16] R. Molfino, M. Zoppi. Virtual engineering techniques: application to the design of a prehensor for picking
up limp pieces. EVEN Workshop on Virtual Eng. Appl. for Design and Product Development: the Footwear
case, September, 11-12, Gallipoli, Italia, 2003.
[17] Kejlaa, G.H. Consumer concerns and the functional value of prostheses to upper limb amputees. Prosthetics
& Orthotics International 17:157-163, 1993.

Robosoft's Advanced Robotic Solutions for outdoor risky
interventions
Pierre Pomiers, Vincent Dupourqué
Robosoft, Technopole d'Izarbel, 64210 Bidart, France
Tel: +33 5 59 41 53 66
Fax: +33 5 59 41 53 79
E-mail: pierre@robosoft.fr
Abstract
Tomorrow, advanced all-terrain robotic applications have to be able to cope with various situations and
perform many different tasks in a dynamic and changing environment. Furthermore, finding concrete
applications in a wide range of user oriented industrial products, such systems, embedding several
computing units, have to cope with an increasing demand of interactivity and to support number of noncritical
pieces of hardware and software. This whole set of different capabilities needs to be performed
reliably and safely over long time periods. To this aim not only advanced programming techniques, but
also appropriate control architectures are required. For these reasons, Robosoft proposes both a set of
hardware and software components developed from our own experience in the field of automatic
transportation of people and goods, which can be easily adapted to robotic solutions for outdoor risky
interventions.
1 Overview
Most papers concerned with real-time embedded application design present experimental tests showing that
theoretical results, obtained from formal analysis, match real-time behavior of embedded system. But they
do not consider critical aspects of integrating, or interfacing, with other non-real-time compatible
processes such as complex high-level control system or user-end applications. To override the complexity
of computing such application algorithms, the control software is built using a dedicated software
environment: iCORE. iCORE relies on the SynDEx 1 data flow graph formalism (introduced by INRIA)
which objective is to provide rapid prototyping and error free implementation procedures for distributed
and heterogeneous application.
From this flexible and reliable development approach, we present how our advanced mobile
systems could be easily used and customized for implementing outdoor applications, and guaranteeing
integrity during risky interventions. Each point of the method is mainly discussed through one of the
Robosoft all-terrain products: the robuCAR TT. It is an all-terrain car-like mobile platform designed
specifically for exploring, performing both measurements and actions in hazardous environments. Last, in
order to illustrate the possible robot operating mode, we will focus on software modules covering various
needs: autonomous navigation, fleet management, remote control, specific HMI...
2 iCORE development environment
Robotics solutions, we describe here, make use of custom control architectures (composed of one Intel x86
Linux/RTAI machines and from one to 8 Motorola MPC555 based control boards 2 ) with CAN buses as
communication media. This section covers both the development and embedded targets environment.
1 Synchronized Distributed Executive
2 cb555 boards manufactured by Robosoft as part of his own control system products

2.1 iCORE: an approach based on SynDEx methodology
The application development method discussed here makes use of both SynDEx tools and Robosoft
kernels. Developed by INRIA, SynDEx V6 is a graphical interactive software (see Fig. 1) with on-line
documentation (refer to [3]), implementing the AAA 3 methodology. Here is the list of services offered by
the couple SynDEx and Robosoft proprietary kernels:
• specification of an application algorithm as a conditioned data-flow graph (or interface with the
compiler of one of the Synchronous languages ESTEREL, LUSTRE, SIGNAL through the common
format DC)
• specification of a multicomponent as a graph
• heuristic for distributing and scheduling the algorithm on the multicomponent with response time
optimization
• visualization of predicted realtime performances for the multicomponent sizing
• generation of dead-lock free executives for real-time execution on the multicomponent with optional
real-time performance measurement. These executives are built from a processor-dependent executive
kernel. SynDEx comes presently with executives kernels for various digital signal processors,
microprocessors and micro-controllers.
The distributing and scheduling heuristics as well as the predicted real-time diagram, help the user to
parallelize his algorithm and to size the hardware while satisfying real-time constraints. Moreover, as the
executives are automatically generated with SynDEx, the user is relieved from low level system
programming and from distributed debugging. This allows optimized rapid prototyping and dramatically
reduces the development cycle of distributed real-time applications.
3 Algorithm Architecture Adequation
Fig. 1: Application design example using SynDEx CAD

The SynDEx development system described above is able to generated executive binaries for various types
of target including the ones that compose the Robosoft control platform. Robosoft control architecture
typically embeds one or more MPC555 based boards and an Intel x86 Real-time Linux computer.
2.2 Robosoft cb555 control board
The Robosoft cb555 [4] board (see Fig. 2) is a stand-alone four axis controller designed for critical
industrial process handling. Including a 32-bit PowerPC architecture, it provides high computation
performance (refer to Table 1 for a detailed description of boards IO capabilities).
Fig. 2: The cb555 Robosoft control board
2.3 Robosoft emPC and wsPC computers
Table 1: Robosoft control board connectors
description
The context of real-time embedded applications programming is quite different from the classical one, user
usually meets. This notion of "real-time" is not present in normal Linuses. Such real-time dedicated
mechanisms can be added by installing an RTOS 4 on top of Linux standard kernel [1] [2]. Robosoft based
both emPC and wsPC product ranges (resp. for embedded and workstation computers) on RTAI version
which is widely used in embedded industry for prototyping and which is supported by very active
companies.
RTAI basic principle is rather simple. RTAI provides deterministic and preemptive performance
in addition to allowing the use of all standard Linux drivers, applications and functions. To this aim, RTAI
decouples the mechanisms of the real-time kernel from the mechanisms of the general purpose Linux
kernel so that each can be optimized independently and so that the real-time kernel can be kept small and
simple. In our case, the primary function of RTAI kernel is to provide real-time tasks with direct access to
the raw hardware, so that they can execute with minimal latency and maximal processing resource, when
required.
3 The robuCAR TT implementation
The robuCAR TT platform (see figure 3) is a mobile robot offering great all-terrain capabilities, able to be
driven over very hazardous landscapes (such as forests, military sites, building sites...). Consequently, this
platform perfectly fits a wide range of missions: autonomous explorations, all-terrain tele-operation,
supervised demining operations, ... Table 2 gives the main electromechanical specifications of the
robuCAR TT. The platform control is handled by an heterogeneous architecture composed with both cb555
4 Real-Time Operating System

oards and an embedded PC. Figure 3 shows how hardware units are organized: two cb555 controllers
dedicated to low-level critical loops, while the embedded PC focuses more on advanced algorithms and
interfaces with asynchronous devices such as wireless network, GPS, laser scanner... Realtime
communications sequences between cb555 boards and PC rely on a CAN bus, while higher level
communications (toward supervisors, network databases, web server, as well as any other non-realtime
devices) are realized trough classical Ethernet links or serial lines.
Fig. 3: The robuCAR TT platform
Fig. 4: robuCAR TT control architecture
Table 2: robuCAR TT electromechanical
specifications
Thanks to a very modular hardware and software structure, robuCAR TT can handled a wide set of
application. As an illustration, let us focus on figure 5. It represents the organigram of an all-terrain
supervision application. Several software levels are depicted, from the most critical 1kHz task to the fully
aperiodic web oriented processes:
• 1kHz control loops driving the four independent motors and the two independent steering jack servos
• a 100Hz loop dedicated to speed and steering profile computing
• an asynchronous level (running under Linux) handling both DGPS and LMS 5
• aperiodic supervision processes with both web page and database update (running over one or more
computers connected to a network)
Relying on iCORE environment (merging the best of SynDEx programming methodology and Robosoft
5 LMS SICK laser scanner

modular proprietary kernels), these software levels implementation lead to highly predictable and safe
application executions, as well as safe (non-blocking) interactions between software levels.
Fig. 5: robuCAR TT application structure
4 iCORE approach: a decisive step to both hardware and software modularity
Over the robuCAR TT example, we show that fairly complex and exhaustive software applications may by
implemented, covering all the needs required for all-terrain missions (supervision, exploration, ...). The
iCORE approach, we propose, appears to be very well adapted to robotics software development and,
moreover, bring a very interesting modular concept. With iCORE approach, modularity is twofold.

First, relying on SynDEx methodology, iCORE provides users with hardware modularity. This
means that once an application is written for a given architecture (composed with a set of cb555, PC and
CAN buses), it is able to run on an extension of this architecture without any modification. Hence,
extending the computing capabilities of a robotics platform is totally effortless: application is automatically
redistributed in order to handle the new architecture resources.
Secondly, iCORE approach also provides user with software modularity. As shown on figure 1, in
our context, an application is designed as a block diagram. Each block contains either a basic feature (from
iCORE kernels) or another set of blocks implementing a more complex feature. Thus, iCORE approach
make software part be easily reusable: simply by copying and pasting blocks subsets.
Fig. 6: Example of possible platform modularity
Figure 6 gives a striking illustration of possibilities offered by iCORE modularity. On the left side,
a four-wheeled platform is shown, composed with two pods. Each pod is driven by the same piece of
control software (running on its own cb555). Hence, programming control for a six-wheeled version (with
three pods) is nothing but duplicating a diagram subset and adding a new cb555 into architecture graph.
Code distribution and execution will require no other step. The same way, assuming a 6-DOF robot arm
control has been previously realized using iCORE approach, adding such an arm to the 6-wheeled platform
is nothing but merging the two application diagrams together and modifying the architecture description in
order to fit the right set of cb555 boards and PC.
Conclusion
Relying on this flexible and reliable development methodology, the iCORE approach, we present here,
bring an efficient help to users and researchers, interested in overriding all-terrain application
implementations complexity. Thanks to his own experience in the field of automatic transportation of
people and goods, Robosoft has adapted robotic solutions for outdoor risky interventions, leading to a
dedicated product range: rugged hardware components (cb555 and various types of embedded PC), as well
as a set of realtime software components (control loops, I/O primitives, laser and wire guidance, obstacle
detection, ...). As shown over the given examples, iCORE approach allow user to easily implement and
customize outdoor applications. Finally, making use of programming methodology introduced by SynDEx,
iCORE guarantee applications executions integrity during risky interventions.
References
[1] E. Bianchi, L. Dozio, P. Mantegazza.
DIAPM RTAI Dipartimento di Ingegneria Aerospaziale - Politecnico di Milano
Real Time Application Interface.
[2] RTLinux
The Realtime Linux.
[3] Thierry Grandpierre, Christophe Lavarenne, Yves Sorel.
Modèle d'exécutif distribué temps réel pour SynDEx. 1998..
[4] MOTOROLA SEMICONDUCTOR TECHNICAL DATA.
MPC555 Product Preview PowerPC TM Microcontroller, MOTOROLA INC., 1998.

Behaviour-based Motion Control for Offroad
Navigation
Martin Proetzsch † , Tobias Luksch ‡ , Karsten Berns
AG Robotersysteme, Fachbereich Informatik, TU Kaiserslautern, Germany
‡ luksch@informatik.uni-kl.de
Abstract— Many tasks examined for robotic application like
rescue missions or humanitarian demining require a robotic
vehicle to navigate in unstructured natural terrain. This paper
introduces a motion control for a four-wheeled offroad vehicle
trying to tackle the problems arising. These include rough
ground, steep slopes, wheel slippage, skidding and others that
are difficult to grasp with a physical model and often impossible
to aquire with sensoric equipment. Therefore a more reactive
approach is choosen using a behaviour-based architecture. This
way a certain generalisation in unknowned environment is
expected. The resulting behaviour network is described and
initial experiments performed in a simulations environment are
presented.
Index Terms— outdoor navigation, behaviour-based robotics,
wheeled vehicles
I. INTRODUCTION
Locomotion in rough natural environment is a basic precondition
for applications like humanitarian demining or rescue
and monitoring after natural disaster. The field of application
is normally restricted to only a few kilometers. Therefore, it
is necessary to have a UGV able to surmount middle-size
obstacles as well as to cover a distance of some kilometers
in an adequate time. In literature there are several robotic
systems like walking machines, snake-like robots, wheel and
chain driven vehicles for different types of natural terrain. As
a good compromise we examine a high flexible wheel-driven
machine for the locomotion in rough natural terrain.
A. Offroad Locomotion
When aiming at locomotion in rough, natural terrain with
a wheel-driven vehicle one has to consider several problems.
For one thing the characteristics of the environment have to
be examined:
Climate These characteristics include extreme tempertature
or humidity.
Physical features of the terrain One has to consider the
slope which can range from flat to near vertical and can
cause the vehicle to stop or even tip over; surmountable
or not surmountable obstacles like rocks, holes, trees,
builinds or even dynamical obstacles like animals or
other vehicles; the soil consistancy which can affect the
possible torque the wheels can create as well as the
vehicle’s stability.
Furthermore different classes of outdoor terrain should be
taken into consideration. These include partially structured
† m_proetz@informatik.uni-kl.de
terrain with roads or paths, relativly even open land, uneven
terrain with medium hills, difficult terrain with more obstacles
and steeper slopes, or even extreme to near vertical terrain.
This work will focus on uneven or even more difficult terrain
with low obstacle density. Other activities in this area include
[1], [2], and [3]. Efforts focusing on more structured terrain
like [4] are not subject of this paper.
Keeping these contraints in mind there arise several problems
when addressing the control side of outdoor locomotion.
These can be devided in navigation and motion control [5].
For the navigation part questions like environment modelling,
localisation including relative and absolute position estimation,
and trajectory generation have to be faced. The motion control
will have to handle the lower level control and handle difficulties
imposed by e.g. the character of the terrain or certain
kind of obstacles in a reactive manner. Motion control is the
main subject of this paper whereas the navigation problems
will not be addressed.
B. The Test Platform Ravon
The test platform used to verify the motion control is a
wheel driven outdoor vehicle by Robosoft (see figure 1). It has
a size of about 1.4 meters in width and 2.4 meters in length
and wheels of 76 cm in diameter. Each wheel has an individual
D/C motor, the front and rear steering is independent. The
robot is capable of driving up to 3 m/s and of climbing slopes
with 30% inclination.
Fig. 1. The four-wheeled outdoor vehicle Ravon
II. KINEMATIC MODEL
This section introduces an abstract single track kinematic
model to describe the essential features for the motion control.
1

This model is extended to four wheels to calculate the single
wheel velocities and to analyse the errors imposed by the
coupled steering.
A. Single Track Model
The motion of a four wheel steering vehicle can be described
using a single track model [6] as illustrated in figure 2.
In this model the front wheels of a four wheel vehicle are
merged together to one wheel in the centre line of the vehicle.
The same action is applied to the rear wheels. This model
describes the features being essential for the motion of a
four wheel steering vehicle, i.e. translational and rotational
movement.
The description of the vehicle position and motion always
refers to a point C representing the kinematic centre of the
vehicle. The yaw angle ψ describes the heading of the vehicle
body. The vehicle moves with a velocity v to the direction β
(called “side-slip angle”) which is measured relatively to the
vehicle body. At the front reference point the velocity is called
vf , at the rear point vr.
The length between C and the position of the merged front
wheels is called lf , the length to the rear point lr. Thus, the
vehicle length l = lf + lr.
When moving, the vehicle motion has a centre called
“instant rotating centre” (IRC) which, in case of parallel
steering of the front and rear wheel, can reach to infinity. The
distance from C to IRC is called rc, the distances from the
front and the rear wheel rf and rr.
In order to force the vehicle to follow a given track the
steering angles (δf and δr) relative to the vehicle body must
be set so that the lines perpendicular to the front and rear
wheel meet in the intended IRC.
There is the assumption that the steering angles of the
wheels are limited (|δf | < π
2 and |δr| < π
2 ) so that the IRC
never lies at the vehicle body line. In most practical cases
this special case need not be considered due to the constraints
imposed by the mechanical structure restricting the steering
angles.
Fig. 2. Single Track Model
Using these assumptions the side-slip angle can be calculated
as presented follows:
� �
lf · tan (δr) + lr · tan (δf )
β = arctan
(1)
lf + lr
The distances to the IRC can be calculated as follows:
rc = lf · sin � �
π
2 − δf
sin (δf − β) = lr · sin � �
π
2 + δr
(2)
sin (β − δr)
rf = lf · sin � π
2 + β�
sin (δf − β)
rr = lr · sin � π
2 − β�
sin (β − δr)
Assuming that the mechanical constraints of the vehicle do
not allow turning around the kinematic centre C, the vehicle
can be controlled by providing the vehicle velocity v and
the front and rear steering angle δf and δr. In this case the
velocities of the two virtual wheels are of interest:
vf = v · rf
rc
vr = v · rr
rc
The movement of a vehicle in an outdoor terrain consists
of the change of the pose in three dimensions having the
following elements:
• coordinates for x, y, and z
• angles for roll (ρ), pitch (φ), and yaw (ψ)
To determine the components of a movement the angle λ
is introduced which denotes the angle between the x-y-plain
and the vehicle velocity vector:
2
(3)
(4)
(5)
(6)
λ = φ · cos(β) − ρ · sin(β) (7)
Using this angle the change of the position in three dimensions
can be calculated as follows:
˙x = v · cos(λ) · cos(ψ + β · cos(ρ)) (8)
˙y = v · cos(λ) · sin(ψ + β · cos(ρ)) (9)
˙z = v · sin(λ) (10)
The change of the heading angle is the following:
� �
vf ˙ψ
· sin (δf ) − vr · sin (δr)
= arctan
lf + lr
The changes of the roll and pitch angle in three dimensions
are determined as follows:
B. Four Wheel Model
˙φ = ˙ ψ · sin(−ρ) (11)
˙ρ = ˙ ψ · sin(φ) (12)
The vehicle this work is based on is constructed so that
the front wheels and the rear wheels respectively are steered
in pairs. This is realised in a way that—like in the case of
normal car steering—the steering centre of the front wheels
always lies on the extension of the rear axis (see figure 3).
Accordingly the steering centre of the rear wheels lies on the
extension of the front axis.
This constraint determined by the mechanical realisation on
the one hand simplifies the kinematic calculation. However,
on the other hand it introduces errors in comparison with the
correct case that the wheels are steered independently so that
all wheels’ perpendicular lines meet in one point.

Fig. 3. Four Wheel Model
The radii of the wheel tracks determined by the position of
the IRC can be calculated using the law of cosines:
�
rfl = sign (rf ) · r2 f +
�
w
�2 − 2 · rf ·
2
w
2 · cos (δf )
�
rfr = sign (rf ) · r2 f +
�
w
�2 − 2 · rf ·
2
w
2 · cos (π − δf )
�
rrl = sign (rr) · r2 �
w
�2 r + − 2 · rr ·
2
w
· cos (δr)
�
2
rrr = sign (rr) · r2 �
w
�2 r + − 2 · rr ·
2
w
· cos (π − δr)
2
Using the radii calculated above the single wheel velocities
can be calculated as follows:
vfl = v · rfl
rc
vfr = v · rfr
rc
vrl = v · rrl
rc
vrr = v · rrr
III. BEHAVIOUR-BASED MOTION CONTROL
The classical control approach requires a complete physical
model of the robot and the environment, which cannot be
acquired with the given sensors and the examined terrain. This
is the case due to possible wheel slippage, skidding and other
problems implied by the vehicle-terrain interaction. Therefore
this paper introduces a behaviour-based motion control not
requiring a complete knowledge of the environment to achieve
robust locomotion. The general behaviour architecture, the developed
behaviours, their purpose and specific implementation
as well as the interaction in the resulting behaviour network
are described in this section.
rc
A. Behaviour Architecture
The behaviour architecture used for this work is based on
the one developed by Albiez and Luksch (see e.g. [7] or
[8]) which again is loosely based on Brook’s subsumption
architecture [9]. It is originally used for controlling walking
machines and is inspired by the activation patterns in the
brain and the spinal cord of animals. Individual behaviours are
structured on layers in a hierachical network, the coordination
problem is solved by using special meta signals generated by
each behaviour. Only the interaction of the behaviours and
their placement in the network will result in the desired actions
of the overall system.
Fig. 4. The layout of a single behaviour B
Each individual behaviour B = (r, a, F ) is a module of the
form outlined in figure 4. The input vector �e can be composed
of sensor information, data processed by other behaviours or
coordination signals of other behaviours. The activation or
motivation value ι ∈ [0..1] is used by higher level behaviours
to influence the behaviour’s degree of activity, the inhibitation
input i ∈ [0..1] allows lower level behaviours to repress
its activity. The output vector �u = F (�e, ι, i) is composed
of control signals for the robot or lower level behaviours
and is calculated by evaluating the transfer function F . This
function describes the behaviours functionality and can range
from a simple P-controller to complex state machines or AI
algorithms.
To solve the coordination problem each behaviour generates
two meta information signals a and r. The activity a(�e, ι, i) ∈
[0..1] states the degree of action the behaviour is producing.
The target rating r ∈ [0..1] indicates the behaviour’s own
evaluation of the current situation based on the input vector �e.
A target rating of Zero denotes that the behaviour is “content”
with the situation, a rating of One represents maximum discontentedness.
The target rating can conveniently be seperated
in an absolute and relative fraction with r = rabs · rrel. The
combination of a and r can provide all necessary information
on the state of the behaviour (see figure 5).
Behaviours can further be classified into target-based,
progress-based and activation behaviours. Target-based behaviours
try to reach a certain goal, e.g. a mobile robot driving
to a point on a map or a walking robot stabilising. Progressbased
behaviours a performing specific movements without
having a target state, e.g. an exploration strategy. Finally
activation behaviours are controlling lower level behaviours
using their activation ι. They are not directly controlling
3

Fig. 5. The state of a (target-based) behaviour can be deduced from its
activity a and its target rating r
hardware.
As already mentioned all behaviours are arranged in a
hierachical network where a flow of meta information (a, r
and ι) is controlling the coordination. For example by using
the activity of one behaviour as inhibitation input of another,
exclusion can be realised. If more than one behaviour tries
to influence a lower level behaviour, a fusion knot is inserted
where the weighting of inputs is deduced from the influencing
behaviour’s activities.
B. Motion Control
The constraints introduced in section I highly influence
the design of motion control functionality. First of all the
vehicle’s abilities, i.e. vehicle velocity, front and steering, form
the basis of this approach. Following a bottom up strategy
behaviours handling these controls are added. The next layer
consists of behaviours generating an intended vehicle velocity
and a translatory or rotatory movement. Finally high level
behaviours are introduced following a top down strategy, i.e.
defined by high level targets, and use the low level behaviours
to fulfill it.
As a main challenge of the movement in outdoor terrain is
handling the occurrence of slopes, the behaviours introduced
below react on influences concerning the pose of the robot.
These behaviours are embedded in a framework which is
introduced in the following section.
1) Basic Functionality:
a) Unit Conversion: To unify the units used in controlling
and sensing modules the values received from and sent
to the user interface are adapted, i.e. all length values are
converted to metres.
b) Drive Input Conversion: To convert different inputs
to the standard control procedure, i.e. steering angles for front
and rear axis and vehicle velocity, a group which contains
modules for each provided steering mode is implemented.
A multiplexer is introduced for selecting the currently active
control mode. These are:
• The standard control mode including front and rear
steering angle as well as the vehicle velocity. This mode
can be used for manual steering or for the behaviourbased
control.
• A steering mode receiving the reciprocal value of the
radius of the robot motion, the side-slip angle, and the
vehicle velocity.
c) Kinematic: The kinematic of a four wheel steering is
described in section II. These considerations are used to form
a module having two tasks: On the one hand it receives vehicle
motion control data, i.e. vehicle velocity and steering angles,
and calculates single wheel velocities. On the other hand data
acquired from sensors specifying single wheel velocities and
steering angles is converted to the assumed vehicle velocity
and the steering angles of all four wheels.
d) Kinematic and Dynamic Constraints: The automatic
control of a vehicle using a behaviour-based approach is independent
from its mechanical characteristics. The control output
of behaviours describes intended values ignoring kinematic
and dynamic constraints of the vehicle’s structure. Therefore
a central unit controlling the vehicle constraints is introduced
regarding the following restrictions:
• maximum vehicle velocity
• maximum vehicle acceleration and deceleration
• maximum front and rear steering angle
• maximum steering velocity
• maximum steering acceleration and deceleration
e) Wheel Velocity Calculation: Encoder values delivered
by the robot hardware are converted to velocity values used
in the control structure.
f) Odometry: For position estimation using odometry in
this module the movement due to wheel velocities and steering
angles is calculated. The following input values are used:
• the vehicle velocity v
• the velocities at the front and rear reference point of the
vehicle (vf and vr)
• the steering angles δf and δr
• the current vehicle orientation consisting of roll (ρ), pitch
(φ), and yaw (ψ)
With the calculations from section II the following output
values are generated:
• the change of the position ( ˙x, ˙y, and ˙z)
• the change of the orientations ( ˙ρ, ˙ φ, and ˙ ψ)
g) Position Determination: As position estimation must
never rely on only one measurement, information from diverse
measurement techniques are merged in this group. Both absolute
and relative position information can be used.
The functionality introduced previously provides the frame
for the behaviour-based motion control comprised by the
behaviours described in the following section. First the targetbased
behaviours beginning from low-level velocity and steering
control are presented following a bottom-up strategy.
Afterwards some supporting progress-based behaviours are
introduced. Finally a higher level activation behaviour is
introduced following a top-down strategy for reaching a point.
2) Target-Based Behaviours:
a) Velocity: The control of the movement of a vehicle
starts with setting a velocity. Behaviours suggesting a velocity
do this by providing an intended normed absolute value v ∈
[0, 1] and the desired direction vsgn (i.e. vsgn = +1 for
forward movements, vsgn = −1 for backward movements, and
4

vsgn = 0 if a behaviour wants to perform movements without
caring about the direction). These values are weighted separately
by a fusion node. The output of this fusion node is fed
into the velocity behaviour, having the following properties:
The intended velocity of the behaviour is defined as
vintended = vmax · v · sign(vsgn) (13)
where vmax is the maximal velocity to be generated, v and
vsgn are the fusioned values of input behaviours. For maximal
response the internal activity aint is set to 1. The absolute and
relative target rating rabs and rrel are calculated as follows:
rabs = abs (vintended − vcurrent)
� � �
∆v
max 0.5, 1 −
rrel =
2·∆t if ∆v > 0
1 else
where ∆v = abs(vintended − vprevious) − abs(vintended − vcurrent)
refers to the change of velocity in respect to the intended
velocity. Here vcurrent is the current vehicle velocity and vprevious
the vehicle velocity measured one time step ∆t before.
b) Steering: The low-level steering behaviour (used
both for the front and the rear steering) transforms translatory
and rotatory movement into a steering angle. The target angle
δintended is calculated the following way (for front and rear
steering):
δf,intended = δtrans · atrans + sign(v) · δrot · arot
atrans + arot
δr,intended = δtrans · atrans − sign(v) · δrot · arot
atrans + arot
The internal activity and the target rating are calculated
similar to the velocity behaviour using the current and previous
steering angle.
c) Translatory Movement to a Target Point: Tasks a robot
has to achieve are often expressed as a movement to a given
position. The target of this behaviour is to reach a target point
by a translational movement. Depending on the current pose
of the robot and the target position the intended translation
(i.e. the side-slip angle) is generated. Additionally, depending
on the distance to the target the velocity is set.
3) Orientate to a Target Point: While the previous behaviour
handled the translatory movement to a target point this
behaviour has the goal to rotate the vehicle so that it points to
the target. The rotational movement and the velocity are set
depending on the deviation between the vehicle orientation and
the target direction. The direction of the movement depends
on the distance to the target point: If the vehicle is far away
a forward movement is executed. If it is near the target a
backward movement is performed in order to minimise the
deviation. For avoiding thrashing a transition zone is used
where the previously set direction is continued.
a) Minimise Roll: The purpose of the Minimise Roll
behaviour is to reduce the absolute value of the roll angle
of the vehicle. This behaviour can be used to stabilise the
vehicle’s posture.
The behaviour receives the roll and pitch angle of the
vehicle. Depending on the roll angle a rotational movement is
performed, assuming that the near environment only changes
in small amounts.
b) Minimise Pitch: Similar to the Minimise Roll behaviour
this behaviour tries to reduce the pitch angle by setting
a rotational movement. The purpose of this behaviour is to try
and orientate the vehicle along an isoline so that a movement
is not hindered by steep slopes.
4) Progress-Based Behaviours:
a) Follow Absolutely Minimal Gradient: A problem to
be considered in outdoor terrain is the elevation that has to be
overcome. A movement along isolines has the advantage of
generating a smooth path concerning height. For this purpose
this behaviour has the goal to follow the absolutely minimal
gradient, i.e. not to change height.
The intended translation is calculated considering the min-
imal slope:
⎧
⎪⎨ 0 if φ = 0 and ρ = 0
π
τ = 2 � � if φ = 0 and ρ �= 0 (14)
⎪⎩
ρ
arctan φ else
The velocity is determined using the change in height—the
higher this value the faster the behaviour wants to move. As
the direction of the movement does not influence the goal of
this behaviour, vsgn is set to 0.
b) Descend: Especially for resolving situations where
the robot movement is stagnating due to a too high slope
a behaviour moving the vehicle downward is introduced. By
using the current information about pitch (φ) and roll (ρ) the
desired direction of the movement τ is calculated similar to
equation 14.
5) Activation Behaviours: Finally a behaviour with the goal
to reach a given target point is introduced controlling the
behaviours mentioned above by influencing their activation
ι. The algorithm used can automatically detect degrees of
progress or stagnation and adjusts the activations of the
behaviours controlled using gaussians to resolve the problems.
A. Simulation Environment
IV. EXPERIMENTS
Fig. 6. The user interface with the simulated terrain
For simulating the four wheel steering robot a user interface
(see figure 6) on the one hand provides a 2D and 3D view of
5

the robot and its environment and on the other hand gives the
user the ability to control the robot manually or by activating
behaviours.
B. Results
One of the experiments carried out compares the performance
when driving towards a target point with most
supporting behaviours activated and without. In the latter case
the vehicle will move on a straight line and possibly even stop
at a slope too steep instead of searching a more energy efficient
way. Figure 7 and 8 plot the energy used during both runs over
the time. One can observe higher peaks and a nearly double as
high energy sum at the end of the run with most behaviours
turned off. Figure 9 shows the activation, activity and tar-
14000
12000
10000
8000
6000
4000
2000
0
300000
250000
200000
150000
100000
50000
0
0 10 20 30 40 50 60 70
Fig. 7. Energy (top) and accumulated energy (bottom) required for reaching
a target point using all supporting behaviours
25000
20000
15000
10000
5000
0
600000
500000
400000
300000
200000
100000
0
0 5 10 15 20 25 30 35
Fig. 8. Energy (top) and accumulated energy (bottom) required for reaching
a target point without supporting behaviours
get rating of the behaviour generating translatory movement
towards the target point. The activation ι is reduced by the
activation behaviour in case the target rating is low enough to
increase the activation of the supporting behaviours. As soon
as the movement of the vehicle is diverging too much from the
goals of the behaviour (high target rating), its activity raises to
get the robot back on track. The activation behaviour increases
ι again to give the behaviour a higher weighting. The target
rating drops towards Zero as soon as the target point is getting
close.
ι
a
r
1
0
1
0
1
0
0 10 20 30 40 50 60
Fig. 9. Activation (ι), activity (a), and target rating (r) of the translatory
point access behaviour while reaching a target point
V. CONCLUSION AND OUTLOOK
This paper introduced a behaviour-based motion control for
a mobile outdoor robot. Its kinematic has been modelled using
a simplifying single track modell to deduces translational and
rotational movements and the odometric reckoning. It is than
extended to a four wheel modell to calculate e.g. single wheel
velocities. Based on this information a behaviour network
has been designed to allow the vehicle save locomotion
towards a target point in uneven terrain. This is achieved
by the cooperation of a set of behaviours reacting on sensor
information like the pose of the robot. Initial experiments in a
simulation environment indicate the approach’s feasibility and
show promising results.
Next steps will be further experiments under real world
conditions with the mobile robot to verify the results from
the simulation. Additional behaviours will be included to
provide higher level functionality and to incorporate sensor
information to a greater extend. Localisation and navigation
algorithms using natural landmarks will be some of the next
problems to be addressed.
REFERENCES
[1] A. Hait, T. Siméon, and M. Taïx, “Robust motion planning for rough
terrain navigation,” in IEEE International Conference on Intelligent
Robots and Systems, 1999.
[2] M. Cherif, C. L. J. Ibañez Guzmán, and T. Goh, “Motion planning for
an all-terrain autonomous vehicle,” in International Conference on Field
and Service Robotics, 1999.
[3] S. Singh, R. Simmons, T. Smith, A. Stentz, V. Verma, A. Yahja, and
K. Schwehr, “Recent progress in local and global traversability for
planetary rovers,” in Proceedings of the IEEE International Conference
on Robotics and Automation, 2000.
[4] D. Coombs, K. Murphy, A. Lacaze, and S. Legowik, “Driving autonomously
offroad up to 35 km/h,” in Proceedings of the IEEE Intelligent
Vehicles Symposium, Dearborn USA, 2000.
[5] A. Singhal, “Issues in autonomous mobile robot navigation,”
http://www.cs.rochester.edu/research/mobile/docs/areapaper.ps.gz, 1997.
[6] D. Wang and F. Qi, “Trajectory planning for a four-wheel-steering
vehicle,” 2001.
[7] J. Albiez, T. Luksch, K. Berns, and R. Dillmann, “A behaviour network
concept for controlling walking machines,” in 2nd International Symposium
on Adaptive Motion of Animals and Machines (AMAM), 2003.
[8] J. Albiez, T. Luksch, K. Berns, and R. Dillmann, “An activation-based
behavior control architecture for walking machines,” The International
Journal on Robotics Research, Sage Publications, 2003.
[9] R. Brooks, “A robust layered control system for a mobile robot,” vol.
RA-2, no. 1, pp. 14–23, Apr. 1986.
6

PLANNING WALKING PATTERNS FOR A BIPED ROBOT USING
FUZZY LOGIC CONTROLLER
Arbnor Pajaziti, Ismajl Gojani, Ahmet Shala, Bujar Pira
arbnorpajaziti@hotmail.com , ismajlgojani@hotmail.com , shalaa@un.org , b_pira@hotmail.com
Mechanical Engineering Faculty, University of Prishtina
Prishtina - KOSOVA
Abstract
As biped robots assume more important roles in applicable areas, they are expected to perform difficult
tasks like mine detecting and quickly adapt to unknown environment. Therefore, biped robots must
generate quickly the appropriate gait based on information received from visual system. In this paper a
Conventional PD Controller and Fuzzy Logic Controller for Planning Walking on flat ground of a planar
five-link biped robot is presented. Both single support and double support phases are considered. The joint
profiles have been determined based on constraint equations cast in terms of step length, step period,
maximum step height and so on. When the ground conditions and stability constraint are satisfied, it is
desirable to select a walking pattern that requires small torque and velocity of the joint actuators. Using
Computed Torque Method is given input - torque on biped robot. The locomotion control structure is based
on integration of kinematics and dynamics model of biped robot. The proposed Control scheme and Fuzzy
Logic Algorithm could be useful for building an autonomous non-destructive testing system based on biped
robot. Fuzzy Logic - Rule base, is optimized using Genetic Algorithm. The effectiveness of the method is
demonstrated by simulation example using Matlab software.
Key words: Biped Robot, Conventional Controller, Fuzzy Logic, Genetic Algorithm, Planning Walking.
1. INTRODUCTION
Researches of biped humanoid robots are currently one of the most exciting topics in the field of robotics and
are many ongoing projects [1]. From the view point of control and walking pattern generation these works can
be classified into two categories. The first group requires the precise knowledge of robot dynamics including
mass, location of center of mass and inertia of each link to prepare walking patterns. Therefore, it mainly relies
on the accuracy of the models. This group is called the Zero-Moment Point (ZMP) based approach since they
often use the ZMP for pattern generation and walking control [7], [8].
Contrary, there is the second group which uses limited of dynamics e.g. location of total center of mass, total
angular momentum, etc. Since the controller knows little about the system structure, this approach much relies
on a feedback control. This can be called as the inverted pendulum approach, since they frequently uses an
inverted pendulum model.
Further, because the environments where humanoid robots operate are unpredictable, the robot must generate
in real time the appropriate gait based on the information received from the visual system [2]. The robot must
walk with different step lengths, overcome obstacles, etc. Until now, the humanoid robot gait is for the most part
prescribed based on humanoid motion. However, measuring the angle trajectories during human walking for a
wide range of step lengths and step times is difficult and time consuming.
In order to generate a human like motion, we proposed the Control scheme and Fuzzy Logic Algorithm that
could be useful for building an autonomous non-destructive testing system based on biped robot. Therefore,
when the humanoid robot is constrained to walk with a particular velocity and step length, the inputs to FLC are
desired position and velocity. Inputs to FLC firstly are labeled depending on their value on PO – positive, ZE –
zero and NE – negative. GA generated the optimal gaits for a wide range of parameters, like step length and step
time during walking, which are used to create the FLC structure.
The paper is organized as follows. The solution of kinematics and dynamics modeling of the used mobile
robot is given in Section 2. Section 3 discusses the FLC-GA approaches, in details. The results of computer
simulations are presented and discussed in Section 4. Some concluding remarks and scope for future work are
made in Section 5.
2. KINEMATICS AND DYNAMICS MODELING
For modeling the biped with four degrees of freedom (DOFs) we have used the rotation angles of joints: θ 1 , θ 2 ,
θ 4 andθ 5 .

Fig. 1. Full gait cycle of a five-link biped walking in the sagittal plane
From Figure 1 follows:
xh = xa1
+ l1
⋅ cos( θ1) + l2
⋅ cos( θ 2 )
xh = xa2
+ l5
⋅ cos( θ5 ) + l4
⋅ cos( θ 4 )
yh = H h = ya1
+ l1
⋅ sin( θ1) + l2
⋅ sin( θ 2 ) ................................................ (1)
yh = H h = ya2
+ l5
⋅ sin( θ5 ) + l4
⋅ sin( θ 4 )
where:
y trajectories of lower limbs (foots) on x respectively y- axes.
x a ,
1
x a ,
2
y a and
1
a2
x h = xhs
+ xhD
and y h = yhs
= yhD
x hs and y hs are trajectories of hip (body) on x respectively y- axes during single support.
x hD and y are trajectories of hip (body) on x respectively y- axes during double support.
hD
Solution of equations (1) by joint angles [ ] T
v = θ1
θ 2 θ 4 θ5
, represents the inverse kinematics of biped.
The Lagrange formulations are used to derive the dynamic equations of the biped.
After the calculation of Lagrange function, the dynamical equations of the biped in Figure 1, can be expressed
in the matrix form:
D( v)
⋅ v�
� + H ( v,
v�
) = τ ............................................................ (2)
where: dimension of matrix D(v) is 4x4 and dimension of H ( v,
v�
) and τ is 4x1.
3. CONTROL DESIGN
Conventional PD controller is used both, as an ordinary feedback controller to guarantee asymptotic stability
during motion period, and as a reference model for responses of the controlled biped.
The proposed control scheme for the biped is presented in Figure 2.
Computed torque τ(t) is:
d
τ ( t ) = D(
v)[
v�
� + Kv
( e�
+ Yv
) + K P ( e + Y p )] ................................. (3)
where KP, KV PD are gains, while YV, YP is additional angle and velocity, respectively.

Fig. 2. Control Scheme for planning of walking pattern of biped using Fuzzy Logic Controller
3.1. Design of optimal FLC using GA
Based on Neural Network (NN), neurons between layers are connected in all-to-all basis, through respective
weights. Inputs to FLC are desired position and velocity. Inputs to FLC firstly are labeled depending on their
value on PO – positive, ZE – zero and NE – negative. Weights between input layer (fuzzy) and hidden layer Vij
and weights between hidden layer and output layer Wjk are calculated using back-propagation algorithm [3].
Weights depends on trajectory tracking error: e = v d – v, where v d – desired values, v – estimated values.
Hidden layer is activating function called sigmoidal (extended sigmoidal for hidden and output layers [4],
sigmoidal-linear [5]. Figure 3a shows structure of FLC, three layers (fuzzy-sigmoid-linear) after optimization by
Genetic Algorithm (GA). General rule base between inputs and hidden layer is shown in Figure 3b.
d
v
d
v�
PO
ZE
NE
PO
ZE
NE
Input layer
Fuzzy
Vij
S1
S4
S9
S1
S4
S9
Hidden layer
Sigmoid
Wjk
R11
R22
Yp
Yv
Output layer
Linear
d
v
PO
d
v�
ZE NE
PO S1 S2 S3
ZE S4 S5 S6
NE S7 S8 S9
Rj k
Wj k
W11 R11
Rj1 Rj2
W12 R12
W21 R21
W22 R22
a) b)
Fig. 3. a) Structure of FLC optimized with GA and b) general rule base between layers.
Genetic Algorithm used for optimization is same as in [5] with small adaptation. As fitness function is used
quadratic error [6]. Important GA parameters are used as per following: Population size: 80, Crossover
probability: 0.95, Mutation probability: 0.02. Number of generations: In first run was 200 generation and second
run of GA 100 generations. After run the GA the best solution was found.
Only rules 2xS1, 2xS4, 2xS9, R11 and R22 (total 8 rules) are selected as good rules. Other rules don’t indicate or
their indication is bad for minimization of respectively quadratic error.
4. SIMULATION RESULTS
In this section, a joint profile for a five-link biped walking on flat ground with both single and double support
phases are determined using the method discussed in Section 3. The values of the parameters mi, Ii, li, and di, of
the five-link biped robot are listed in Table I and the walking speed is chosen 1.06 m/s with step length
SL=0.72m, TS=0.6 s, TD=0.1s, Hm=0.05m and Sm=0 m [7].

Table I. Parameters of the bipedal robot.
__________________________________________________________________________________________
Mass Moment of Length Location of centre of
Link (kg) inertia (kgm 2 ) (m) mass to lower joint (m)
__________________________________________________________________________________________
Torso (3) 14.79 3.30 x 10 -2 0.486 0.282
Tigh (2,4) 5.28 3.30 x 10 -2 0.302 0.236
Leg (1,5) 2.23 3.30 x 10 -2 0.332 0.189
__________________________________________________________________________________________
Figure 4 shows a stick diagram of the five-link bipedal model walking on flat ground. From this diagram,
one can observe the overall motion of the biped during the single and double support phases. The dashed lines
represent the movements of the hip, upper and lower left and right limbs. The stance limb propels the upper
body while the upper body is maintained in the upright position (è3=0). The posture of the biped at the end of
each step is close to that at the beginning of each step, which indicates that the repeatability condition is
satisfied.
Figure 5 shows the motion of the lower limb joint angles during the single and double support phases for two
steps.
Fig. 4. Stick diagram of bipedal walking for Fig.5. Designed joint angles simulation time 1.4s,
two steps, first 0.7m and second 0.6m two steps, first 0.7m and second 0.6m
Figure 6 a) and b) shows the horizontal displacements and its velocity of the hip during the single and double
support phase of one step.
It can be seen that the hip displacement remain approximately at the centre of the ZMP stability region,
which ensures the greatest stability [7].
a) b)
Fig. 6. Designed: a) Trajectory and b) Velocity of the hip for first step 0.7m and second step 0.6m
Figure 7 shows the displacements in x axis of three step lengths for first, and second lower limb and hip
desired and designed trajectory, respectively. All the trajectories are smooth; i.e., all the velocities are
continuous. The hip trajectory designed with FLC-GA shows better performances than designed hip trajectory.
The quadratic errors “desired-designed FCLGA”, and “desired-designed” for three step lengths: 0.7, 0.6 and
0.5 [m], are given in Figure 8, respectively.

Legende:
trajectory, first lower limb, x axis
trajectory, second lower limb, x axis
desired hip trajectory
designed hip trajectory
hip trajectory designed with FLC-GA
Legende:
quadratic error “desired – designed FLC-GA”
quadratic error “desired – designed”
Fig. 7. Displacements in x axis, three step lengths: Fig. 8. Quadratic errors, three step lengths:
0.7, 0.6, and 0.5[m]. 0.7, 0.6, and 0.5 [m].
5. CONCLUSIONS
In this paper a FLC-GA based approach for optimization of biped gait synthesis is presented. We considered an
important task for the biped robot – walking with different step lengths. Our method can be applied to a wide
range of step lengths. The performance evaluation is carried out by simulation. Based on the simulation results,
we conclude:
• Fuzzy Logic Controller and Neural Networks contain high number of mathematical operations.
• To decrease in minimum mathematical operations in this paper is represented the successfully called
fuzzyfication of FLC and it’s optimally rule base using GA-s.
• Using GA-s we have found minimal number of rules, 8 rules from total 22 rules. Optimal FLC has decreased
error with 63.6% less number of mathematical operations.
• For FLC is necessary to have general rule base, after optimization with GA can be found optimal (minimal)
number of rules.
• From the stability point of view, it is better for the robot body not to be fixed during motion.
• Using GA for the gait optimization of the biped robot when going up-stairs, down-stairs, etc.
REFERENCES
[1] Sh. Kajita et al.: “Biped Walking Pattern Generation by using Preview Control of Zero-Moment Point”,
Proceedings of the 2003 IEEE International Conference on Robotics & Automation Taipei, Taiwan, pp.
1620-1626, September 14-19, 2003, Robotica, Volume 13, pp. 477-484, 1995.
[2] G. Capi et al.: ”Real time gait generation for autonomous humanoid robots: A case study for walking”,
Robotics and Autonomous Systems 42, pp. 107-116, 2003.
[3] S. Jung; Hisia T.C.: “A new Neural Network Control Technique for Robot Manipulators” Robotica,
Volume 13, pp. 477-484, 1995.
[4] G. Campa: “Adaptive Neural Networks” toolbox for MATLAB implementation, represented on
www.mathworks.com, last updated June 22 2003.
[5] C.R. Houck, J.Joines and M.Kay.: “Genetic Algorithm Optimization Toolbox” A genetic algorithm for
function optimization: A MATLAB implementation. Free Software Foundation, Inc., 675 Mass Ave,
Cambridge, MA 02139, USA. www.matlab-exchange.com , last updated 2003.
[6] A. Shala, R. Likaj, A. Geca, A. Pajaziti, F. Krasniqi: “Trajectory tracking by using Fuzzy Logic
Controller on Mobile Robot” International workshop, HUDEM 2003, Prishtina, Kosova, 2003.
[7] X. Mu, Q. Wu: “Synthesis of a complete sagittal gait cycle for a five-link biped robot”, Robotica, Volume
21, pp. 581-587, 2003.
[8] Q. Huang, et al.: “Planning Walking Patterns for a Biped Robot”, IEEE Transactions on Robotics and
Automation, Vol. 17, No. 3, pp. 280-289, 2001.

Abstract
Adaptive Neuro-Fuzzy Control of AMRU5, a six-legged walking robot
J-C Habumuremyi, P. Kool and Y. Baudoin
Royal Military Academy-Free University of Brussels
08 Hobbema Str, box:MRTM, 1000, Brussels, Belgium
E-mail:Jean-Claude.Habumuremyi@rma.ac.be;Pkool@vub.ac.be;
Yvan.Baudoin@rma.ac.be
Due to the complexity of walking robots which has in general a great number of degrees of freedom, cognitive
modelling controller such as Fuzzy Logic, Neural Networks…seems to be reasonable in the design of adaptive control
of such robot. Fuzzy Logic Controller is more used because it lets you describe desired system behaviour with simple
“if-then” relations. But it has a major limitation because in many applications, the designer has to derive “if-then”
rules manually by trial and error. On the other hand, Neural Networks perform function approximation of a system but
we cannot interpret the solution obtained neither check if its solution is plausible. The two approaches are
complementary. Combining them, Neural Networks will allow learning capability while Fuzzy-Logic will bring
knowledge representation (Neuro-Fuzzy). In this paper, we show an original method to design an adaptive Neuro-Fuzzy
controller which consists in five steps that are: the initial design of an ANFIS controller, the identification of the
dynamic model of the leg joints, the estimation of the parameters of the dynamic model, the calculation of an ideal
torque and the updating of the parameters of the controller and finally the design of the supervisory control.
1 Introduction
Many robots (manipulator and mobile robots), until now, are controlled using linear controllers (PID) which are
independent for each joint. It can be proven that those controllers are fairly effective. The two main reasons are [1]:
- The large reduction ratios between the actuators and the link mechanism (non-linearities and coupling terms
become less important)
- The large feedback gains in the control loops (they enlarge the domain where the complete robot dynamics is
locally equivalent to a linear model).
These controllers operate over a small range in which the dynamics of the system are considered as linear. These
controllers limit the use of such robots to slow motion applications and fixed payload. However, the normal operational
range of a robot may be large, and its payload also could change. To have a controller which works on different
operational range and take into account the change of the payload, the environment and the uncertainties (friction,
flexibility,…), necessitate a sort of on-line parameter estimation scheme in it (an adaptive controller). Most of classical
adaptive controllers are based on the well-known dynamic properties of robots which stipulate that the dynamic model
of a system is linear with respect to the dynamic parameters (mass, moment inertia, link lengths…). Even for simple
cases, it remains difficult to have the relation which expresses the linearity of the dynamic parameters in the dynamic
model. The problem become more complex for a walking robot which has in general a large number of degrees of
freedom (we have 18 just for the robot to walk) and which requires changing internal parameters depending on the
environment that it explores. Also, it seems practically difficult to build a representative model of a walking robot due
to the problem of having accurate internal parameters (distance between joints, moment inertia…) and to accurately
model some complex phenomena such as backslash, friction…In this case, cognitive modelling such as Fuzzy Control
and Neural Networks seems to be reasonable.
Fuzzy Logic Controller (FLC) is more used because it lets you describe desired system behaviour with simple « ifthen
» relations. In many applications, this gets you a simpler solution in less design time. In addition, you can use all
available engineering know-how to optimise the system performance directly. While this is certainly the beauty of fuzzy
logic, at the same time it is a major limitation. In many applications, knowledge that describes desired system behaviour
is contained in data sets. The designer has to derive the « if-then » rules from the data sets manually, which requires a
major effort with large data sets. This is often done by trial and error. Without adaptive capability, the performance of
FLCs relies on two factors: the availability of human experts, and knowledge acquisition techniques to convert human
expertise into appropriate fuzzy « if-then » rules and membership functions. These two factors substantially restrict the

application domain of FLCs. Changing shapes of membership functions can drastically influence the quality of the
FLC. Thus methods for tuning fuzzy controllers are necessary.
Artificial neural networks are highly parallel architectures consisting of simple processing elements, which
communicate through weighted connections. They are able to approximate or to solve certain tasks by learning from
examples. When data sets contain knowledge about the system to be designed, a neural net promises a solution because
it can train itself from the data sets. However, only few commercial applications of neural nets exist due to the lack of
interpretation of the solution, the prohibitive computational effort and the difficulty to select the appropriate net model.
It becomes obvious that a clever combination of the two technologies delivers the best of both. Neuro-Fuzzy [2] is a
combination of the explicit knowledge representation of the fuzzy logic with the learning power of the neural nets. In
this paper, we show the way to design an adaptive Neuro-fuzzy controller in order to track determined trajectories in
different situations. Five steps have been considered in the design of such a controller.
2 ANFIS 1 Architecture
Figure 1: Walking Robot AMRU5
There are many approaches to combine fuzzy logic and Neural Networks, known among them are:
- A Cooperative Neuro-Fuzzy system where ANN learning mechanism determines the FIS membership
functions or fuzzy rules from the training data after ANN goes to the background.
- Concurrent Neuro-Fuzzy systems where ANN assists the FIS continuously to determine the required
parameters.
- Fused Neuro-Fuzzy systems are the methods where FL and ANN share data structures and knowledge
representations. In the above methods FL and ANN are separated. In fused systems, ANN learning algorithms
are used to determine the parameters of FIS. For that, Fuzzy system is represented in a special ANN like
architecture. Some of major works in this area are NEFCON, ANFIS[3], FALCON, GARIC, FINEST and
many others.
In our application, we have used the ANFIS [3] architecture. It keeps the structure of the fuzzy controller that is
determined by the fuzzy rules as depicted in Figure 2.
At layer 1, every node is adaptive (premise parameters) with a node function which is the membership function. Node
output at layer 2 represents the firing strength of a rule. In our application, it is a product of all the incoming signals but
can be in general any T-norm operators that performs fuzzy AND. At layer 3, a normalised firing strengths is realised
by making a ratio of rule’s firing strength to the sum of all rule’s firing strengths. Nodes at layer 4 are adaptive
(consequent parameters) with node function which can be a first-order Sugeno, zero-order Sugeno, Mamdani or
Tsukamoto fuzzy model.
1 Adaptive Neuro-Fuzzy Inference Systems

Figure 2: ANFIS Architecture
3 Step 1: Design of an initial zero-order Sugeno Fuzzy Logic controller
It is important to have an initial controller which works properly in a closed range. One of the reasons is that during the
on-line learning, optimization algorithm will be used and the solution cannot converge to the good one if parameters of
the controller are set far away of the true. We have to make a fuzzy controller which work properly in a closed range
then make it adaptive to take into account uncertainties of the model. Many controller design avoid this problem by
making first a classical controller (PD usually) then add another controller (Fuzzy or Neural Network) to deal with
uncertainties. This makes the controller more complex. In our method, we design an initial Neuro-fuzzy controller
which works similarly as a classical one. After, we make it adaptive to deal with uncertainties. To find good parameters
of a fuzzy logic controller is not an easy task because they have in general a lot of parameters. If we have n input, m
triangular membership functions (2 parameters to adjust by membership function) and a zero-order Sugeno FIS is used,
n
we have to fix 2 mn + m parameters. For illustration of the method, we will use a fuzzy system (equivalent to a
discrete PID controller) with 2 triangular membership functions (N, P), 3 input (e(n), e(n-1) and e(n-2)) and a zero-order
Sugeno FIS as shown on Figure 2, but this method can be generalised. In this case, we have to fix 20 parameters. But if
we fix parameters of the membership function, we have only 8 parameters to fix. A typical rule of such a system has the
form:
If e(i) is N, e(i-1) is P and e(i-2) is N then the output equal
z 1 = p1e(
i)
+ q1e(
i −1)
+ r1e(
i − 2)
+ s1
(1)
Where { p 1 , q1,
r1
, s1}
: is the parameter set of one node. The equation (1) become z 1 = s1
(2) for a zero-order
Sugeno-Takagi FIS.
Figure 2: Zero-Order Sugeno: two triangular MF and three input
The first method to fix parameters of a controller is simple trial and error. Unfortunately, intuitive tuning procedures
can be difficult to develop in the case of Sugeno FIS because a change in one tuning constant tends to affect the
performance of others terms in the controller’s output. Also, the great number of parameters makes this method
practically impossible. The second method can be the analytical approach to the tuning problem. It involves a
mathematical model of the process. This method cannot be used because the advantage of fuzzy logic is precisely the
fact that it is used on complex processes where the establishment of a reliable model is unimaginable. The third

approach to the tuning problem is something of a compromise between purely self-teaching trial and error techniques
and the more rigorous analytical techniques. It was originally proposed by John G. Ziegler and Nathaniel B. Nichols
[5] and remains popular today because of its simplicity and its applicability to process which can be describes by a
“gain”, a “time constant” and a “dead time” (which is the case of joints of robots actuated by DC motor). Ziegler and
Nichols came up with a practical method for estimating the proportional, the integral and the derivative parameters of a
PID controller. In this paper, we show how these techniques can be applied in the design of a fuzzy controller.
3.1 How the method was developed?
Many techniques used to turn a Mandani Fuzzy Model (which has less parameter compare to Sugeno Fuzzy Model) are
intuitive. In many papers, books…they show which parameters to increase or to decrease by considering the rise time,
the overshoot and the steady state error [4]. These techniques seem more like the art than engineering and they are
difficult to apply them to Sugeno Fuzzy Model. The best solution to turn parameters of a Sugeno Fuzzy Model is by
fusing Neural Networks to Fuzzy Logic Systems. But we need data to train the system. The first solution can be to
collect them from a classical controller implemented to the real robot by giving random trajectories to the actuators. We
noticed that we cannot cover all possible operating regions, the time to read data (on encoders, actuators,…) and to
write them on a stored device is too short (the microcontroller has no much time sometime to write all the value) and
there is noise on the data. The error obtained after training is still big by using these data. Another original solution
could be the use of Ziegler-Nichols rules originally applied to PID controllers. The analogue PID controller is expressed
by the equation:
de
u( t)
= K pe( t)
+ K i ∫ e(
t)
dt + K d (3)
dt
Where e: is the difference between the set point and the process output and u the command signal. and K
are controller parameters.
Two practical methods can be used to have a first estimate of the PID controller parameters:
- the step-response method
- and the frequency response method (only this method will be considered in this paper)
K , d
p K i
3.1.1 The step-response method
This method is based on a registration of the open-loop response of the system, which is characterized by two
parameters. The parameters (a and L) are determined from a unit step response of the process, as shown in Figure 3.
When those parameters are known, the controller parameters are obtained from Table 1.
Figure 3: Step-response method Figure 4: Frequency response
method
3.1.2 The frequency-response method
The idea of this method is to determine the point where the Nyquist curve of the open-loop system intersects the
negative real axis. This is done by connecting the controller to the process and setting the parameters so that pure
proportional loop system is obtained. The gain of the controller is then increased until the closed-loop systems reaches
the stability limit. When this occurs, the gain and the period of oscillation shown on Figure 4 are determined.
T
The controller parameters are then given by the Table 2.
K u
u

Controller
Type
K p K i K d
P 1
a
PI 0.
9 0 . 3
a aL
PID 1 . 2 0. 6 0.
6aL
a aL
Table 1: Parameters obtained from
step-response method
Controller
Type
P
PI
PID
K p K i K d
0 . 5K
u
0 . 45K
0 . 6K
u
3.1.3 Method of turning an UFLC based on the frequency-response
u
0.
54K
T
u
u
u
1. 2K
u 0.
075K
u
Table 2: Parameters obtained from frequency-response
method
If the ultimate gain and the ultimate period T of the process were determined by experiment or simulation,
K u
u
equation 3 can be written as follow:
1.
2K
u
de
u( t)
= 0.
6K
ue( t)
+ e t dt K uTu
T ∫ ( ) + 0.
075 (4)
u
dt
There exist different methods to convert equation (3) into discrete form for digital implementation such as Tustin
approximations (or trapezoidal approximations), ramp invariance, rectangular approximations…When the sampling
time T is short, all these methods have nearly the same performance. We’ll use rectangular approximations. Equation
(4) becomes:
n
(5)
(5)-(6) gives
1.
2Ku
e(
n)
− e(
n − 1)
u( n)
0.
6Kue(
n)
+ e(
i)
T + 0.
075KuTu
T
T
= ∑
u i = 1
n 1 1.
2Ku
e(
n − 1)
− e(
n − 2)
u( n − 1)
= 0.
6Kue(
n − 1)
+ ∑e( i)
T + 0.
075KuTu
T
T
−
u
i = 1
(6)
∆u n)
= u(
n)
− u(
n − 1)
= K e(
n)
+ K e(
n − 1)
+ K e(
n − 2)
(7)
( 1
2
3
2
2
Where 3 K ⎛ ⎞
u Tu
+ 8TTu
+ 16T
, − 3Ku
⎛Tu − 4T
⎞ and 3 K uTu
K = ⎜
⎟
1
40 ⎜
⎟
K = ⎜ ⎟ K
2
3
⎝ TTu
⎠ 20 ⎝ T ⎠ 40 T
=
Equation (6) can now be used to build a FLC controller that we called a UFLC (Unit FLC). UFLC will be determined
by the equation:
∆uu ( n)
= K1eu
( n)
+ K 2eu
( n − 1)
+ K 3eu
( n − 2)
(8)
where e is between -1 and 1. If we define a step t (t equal 0.001 for example), we can define a set A of numbers
u
()
between -1 and 1 as follow:
A = { −1,
−1+
t,
−1+
2t,
K , 1−
2t,
1−
t,
1}
{ ( n),
e ( n −1), e ( n − 2
Then we constitute all possible set eu u
u )} with numbers which belong to the set A. From each
set, we calculate ∆ uu (n)
using equation(8). Finally, we can use the set { eu ( n),
eu
( n − 1),
eu
( n − 2),
∆uu
( n)}
to train the
Neuro-Fuzzy Controller. Using a hybrid learning paradigm (least square error algorithm for consequent parameters
which are linear and backpropagation for premise parameters), we noticed that the initial membership functions did not
change (premise parameters remain the same), only consequent parameters change. With 2 triangular membership
functions choose for our illustration, we have analytical expression shown in the Table 3.
The same procedure can be applied to the step-response method and to derive rules of a controller which depends from
the parameters a and L.
3.1.4 Use of the UFLC on a real process
In practice, error will not belong always between -1 and 1. We need some transformation to use the UFLC design on a
real process. If the minimum error of the system is a and the maximum is b (a and b was determined by the limitation of
each joint), a reduced error (n (error between -1 and 1) can be expressed as follow:
e u
)
T
uT

and
b − a b + a
( n)
= e ( n)
+ (10)
2 2
e u
e u
Rule e(n)
2 ⎛ ⎛ b + a ⎞⎞
( n)
= ⎜e(
n)
− ⎜ ⎟⎟
(9)
b − a ⎝ ⎝ 2 ⎠⎠
e(
n −1)
e(
n −
1 N N N
2 N N P
3 N P N
4 N P P
2)
∆ un (n)
− 6KuT
5T
2 2
− 3Ku
( 8T
− Tu
)
20T
T
− 3Ku
( 2T
+ Tu
)
10T
T
2
2
− 3Ku
( 8T
+ 8TuT
+ Tu
)
20T
T
5 P N N
2
3Ku
( 2T
+ Tu
)
10TuT
6 P N P
2 2
3Ku
( 4T
+ Tu
)
10TuT
7 P P N
2 2
3Ku
( 8T
− Tu
)
20TuT
8 P P P
6 KuT
5Tu
Table 3: Rules of the system used for illustration
Equation (10) in (7) gives
b − a
b + a
∆ u( n)
= ( K1e
n ( n)
+ K 2e
n ( n −1)
+ K 3e
n ( n − 2))
+ ( K1
+ K 2 + K 3 ) (11)
2
2
Equation (11) becomes after simplification:
b − a 6K
uT
b + a
∆ u(
n)
= ∆uu
( n)
+
(12)
2 5Tu
2
Figure 5 shows how UFLC is used on a real process.
Figure 5: Use of the UFLC on a real process
u
u
u
u
2

3.1.5 Application to a known function transfer
To allow comparison between a classical PID controller and a UFLC, we have applied the method to the process with a
transfer function
1
G ( s)
= 3
( s + 1)
(13)
This process has the ultimate gain K = 8 and the ultimate periodT
2π = ≈ 3.
63 . From the Table 2 and Table 3, we
u
can easily design a PID and an UFLC. Figure 6 shows the Matlab schematic used to compare the two controllers with
the step function as the input.
Figure 6: Comparison between PID controller and UFLC
The output of the two controllers and their result of the error (output of the PID controller subtract to the output of the
UFLC) are shown on Figure 7 and 8. The error is less than 0.0063. We can see how close they behave in the same way
and the small error between them is mainly due to the truncation of the numbers.
Figure 7: Output of the PID controller and the UFLC Figure 8: Error between the PID controller and the UFLC
u
3

4 Step 2: Identification of the dynamic model of the legs
ANFIS will be used to identify the dynamic model of each legs of the robot AMRU5. The dynamic model of the amru5
leg (or robot in general) is formulated as:
T
τ = A ( θ ) & θ&
+ C(
θ,
& θ ) & θ + F(
θ,
& θ ) + G(
θ ) − J F ( θ ) FRF
(14)
Where:
τ = [ τ1
, τ 2,
τ 3]
θ = [ θ1
, θ 2 , θ 3 ]
Is the vector of forces/torques
Is the vector of the position coordinates
A (θ )
Is the inertia matrix
( θ, θ ) & C Is the centrifugal and Coriolis vectors
( θ, θ ) & F Are frictions forces acting on the joints
G (θ )
Is the vector of the gravitational forces/torques
T
J F
Is a transpose of a Jacobian matrix
F Is the vector of the reaction forces that the ground exerts on the robot feet
RF
Equation 14 can be written in a compact way as follow:
τ = A θ & θ&
+ H θ,
& θ (15)
( ) ( )
With H ( , θ ) being defined as a whole without distinguishing the differences among the different terms.
θ &
H ( , θ ) contains centrifugal, Coriolis, gravitational forces, viscous friction, coulomb friction and reaction forces
terms.
We have used a 2D pantograph mechanism for each leg. This mechanism shown on Figures 9 and 10 has the
particularity to have a decoupling of the joint θ 3 from the joints θ 2 andθ 1.
Equation 15 can be split in two parts as
θ &
follows:
⎛τ
⎞ ⎛ A
⎜
⎟ = ⎜
⎝τ
2 ⎠ ⎝ A
A
A
⎞ ⎛ H 1 ( θ 1 , θ , &
2 θ , &
1 θ 2 ) ⎞
⎟ ⎜
⎟
⎟
+
⎜ H 2 ( 1 , 2 , &
1 , &
2 ) ⎟
⎠ ⎝ θ θ θ θ ⎠
( τ ) A & θ&
+ H ( θ ,θ&
)
= τ =
Α
( θ)
& θ&
+ H(
θ, θ
1 11 12 1
&
12
22
⎞⎛
& θ&
⎟
⎟⎜
⎜
⎠
&
⎝θ
2
3
= (17)
3
3
3
3
3
) (16)
Figure 9: 2D pantograph mechanism Figure 10: General structure of the AMRU5 robot leg
We will only consider the system of equation 16 to explain the ANFIS joints control. Equation 17 is particular case. To
identify parameters of the dynamic model described by equation 16, we need its equivalent discrete-time version

θ &
defined by nonlinear difference equations. To approximate and (i=1 to 2), we have used the Taylor series and we
find:
~ θi
( k + 1)
−θ
i ( k −1)
θi ( k)
=
2∆t
&
(18)
~ θi
( k + 1)
− 2θ
i ( k)
+ θi
( k −1)
θi ( k)
=
2
∆t
&
(19)
Where ∆t
is the sampling time.
Equation … becomes in discrete-time:
τ(k) = A(
θ(
k)
) θ&
& ( k)
+ H ( θ(
k)
, θ&
( k)
) (20)
The procedure to have an Offline (Online is also possible) ANFIS dynamic model of the coupled joints 1 and 2 is as
follows:
1. Collect ( θ( k + 1)
, τ(
k)
) ≡ ( θ1
(k + 1), θ2
(k + 1), τ 1(k),
τ 2 (k) ) from trajectories of the actuators on the
joints 1 and 2. These trajectories are chosen in such a way to cover all the possible movements of the two
joints.
2. Constitute from these collected data sets
3. The sets defined above will be used in the parallel identification model shown in Figure 11. Trial and error
have to be used to find the best type of membership function to use, the number of linguistic variables and the
type of Takagi-Sugeno FIS. The final RMSE (Root Mean-Square Error) of each trial gives a comparative
evaluation between different ANFIS models.
i
θ &
Figure 11: Offline ANFIS parallel Identification model
5 Step 3: Estimation of the parameters of the dynamic model
It is necessary to estimate to estimate the elements of the symmetric inertia matrix ( θ(
k)
)
matrix θ(
k)
, θ(
k)
( )
i
A and the elements of the
H & because they will be used in the strategy control. To estimate those elements, we use the
principle that the elements of the inertia matrix are only dependent on θ(k) and the element of the matrix
H ( θ(
k)
, θ&
( k)
) are dependent on θ(k) and θ(k) i.e. if we change and maintain θ and with the same
value, elements of the inertia matrix will remain the same.
& θ(k) & (k) θ(k) &
Suppose we have the same θ(k) , θ(k) for 3 different angular accelerations (
& θ& &
j (k) j = a,
b and c ), then
(k) = A θ k θ&
& k + H θ k , θ&
k (21)
τ j
j
By applying the principle stated above, we have:
( )
( ( ) ) ( ) ( ) ( )

τ
τ
1a
1a
(k) − τ
(k) − τ
1b
1c
(k) = A
(k) = A
11
11
( θ&
&
1a (k) − θ&
&
1b (k) ) + A ( θ&
&
12 2a (k) − θ&
&
2b (k) )
( θ&
& (k) − θ&
& (k) ) + A ( θ&
& (k) − θ&
& (k)
1a
1c
( θ&
&
1a (k) − θ&
&
1b (k) ) + A ( θ&
&
22 2a (k) − θ&
&
2b (k) )
( θ&
& (k) − θ&
& (k) ) + A ( θ&
& (k) − θ&
& (k)
12
2a
2c
) (22)
τ 2a (k) − τ 2b (k) = A12
τ 2a (k) − τ 2c (k) = A12
1a 1c
22 2a 2c ) (23)
The resolution of the system of equations 22 and 23 gives the estimated values of the elements of the inertia matrix.
There are two way to determine A12
(by using the system of equations 22 or the system of equations 23). These two
ways could give an idea on the precision of the estimated parameters.
6 Step 4: Calculation of the ideal torque to control the joints and updating of the parameters
of the controller
From equation 20, we obtain:
θ&
& −1
(k) = A θ k τ k
−1
− A θ k H θ k , θ&
k (24)
( )
( ( ) ) ( ) ( ( ) ) ( ) ( )
If we suppose that the matrices A and H are known exactly, we can define a control law as follows:
−1 τ(k) = A( θ (k)) [ A ( θ ( k )) H(
θ ( k ), & θ ( k )) + θ&
&
d (k) + ke]
(25)
T
where & θ&
d ( k ) ≡ [ & θ&
1d
( k ) & θ&
2d
] is the vector with the desired angular accelerations of the joints 1 and 2.
e e e&
e e&
T
is the tracking error vector where e = θ − θ and e& & θ − & θ (j=1,2). Also e& &
[
]
= 1 1 2 2
j jd j
will be equal to & θ&
jd − & θ&
j . ⎟ be such that all roots of the polynomial and
⎟
⎛ k1
k 2 0 0 ⎞
2
k = ⎜
s + k 2 s + k1
⎝ 0 0 k 3 k4
⎠
2
s + k4
s +
k
the polynomial are in the open left-half plane.
Introducing 25 in 24, we have:
limt →∞
j
3
e ( t ) = 0
e&
& 1 + k e&
2 1 + k1
= 0
(26)
e&
& + k e&
+ k = 0
We can see that , which is the main objective of the control. Since
2
4
2
3
j = jd j
j
A and H are not known
exactly, we replace them respectively by A ~ and H ~ obtained from the ANFIS model. The resulting control law is:
H ( ( k ), ( k )) θ (k) ke
~
A ( ( k ))
~
A ( (k))
~
−1 (k) = θ θ θ & θ + &
+ (27)
[ ]
τ c
d
τ c (k) obtained from an approximated model of the process is called the certainty equivalent controller [6] in the
adaptive control literature. This torque will be used to update the parameters of the initial controller designed. If τ a is
the output of the zero order Takagi-Sugeno controller, it can be expressed as:
n
∑
i=
1
a = n
∑
i=
1
ω r
τ (28)
Where n is the number of rules, ω i the product of the output of the membership functions which belong to the rule i and
ri
i
ω
the output weight of the rule i. We update only the output weights of the rules by minimizing the error
(
2
(backpropagation algorithm) E = τ −τ
) . For the lth rule with 1 ≤ l ≥ n
c
∂E
∂r
l
2
=
a
2
2E
∂E
= −2
∂r
l
i
i
ω
l
n
∑ ω i
i=
1
( τ −τ
)
c
a
(29)

The update law of the weight of lth rule is
⎛
⎞
⎜
⎟
⎜ ω l
r ( ) ⎟
l ( k + 1 ) = rl
( k ) + η 2 τ −
⎜ n c τ a (30)
⎟
⎜ ∑ ω i ⎟
⎝ i=
1
⎠
Whereη is the learning rate.
7 Step 5: Design of the supervisory control
For simplicity of writing, we will not put the terms between brackets. Equation 24 can be written as follows:
H ~
A ~
A A H
~
H A
~
A ~
A τ k
~
θ&
& −1
−1
−1
−1
−1
−1
(k) = − + − τ − − (31)
( ) ( ) ( )
⎡ 0 1 0 0 ⎤ ⎡0
0⎤
⎢
⎥
Introducing 27 in 31 and defining φ as ⎢
− k1
− k 2 0 0
⎢ ⎥
⎥ , b as ⎢
1 0
⎥ and e as previously, we obtain
⎢ 0 0 0 1 ⎥ ⎢0
0⎥
⎢
⎥ ⎢ ⎥
⎣ 0 0 − k 3 − k4
⎦ ⎣0
1⎦
[ H ] e bJ
~
A ~
A A ) ( A H
~ −1
−1
−1
−1
e&
= φe
+ b ( − τ c + − = φ + (32)
Where
H )
~
A ~
A A ) ( A H
~ −1
−1
−1
−1
J = ( − τ + −
c
1 T
We define a Lyapunov function V = e Pe
2
T
Where P is a positive definite symmetric 4X4 matrix which satisfies the Lyapunov equation φ P + Pφ
= −Q
(Q is
an arbitrary 4X4 positive definite matrix. The derivative of the Lyapunov candidate function gives:
1
V&
T T
= − e Qe + e PbJ (33)
2
In order for θ − , & θ − e&
to be bounded we require that V must be bounded that means V& ≤ 0 when V is greater
d
e d
than a large constantV
. However, from equation 33, it is difficult to design τ c such that e PbJ is less than zero. To
T
solve this problem we add another term
τ τ + τ
τ s to τ c called supervisory control term. The control torque
becomes = c s . With the new definition ofτ , we obtain:
H ~
A ~
A A ) A ( A H
~ −1
−1
−1
−1
−1
e&
= φe
+ b ( − τ − τ + − (34)
[ ]
c
Substituting 34 in 33, we have:
T T 1
1
1
1
1
[ c
c
s
H ) ]
1
~
T
T
1
1 T 1
e Qe e Pb [ A c A H ] e PbA s
2
~
A ~
A A A ( A H
~
1
V&
−
−
−
−
−
= − e Qe + e Pb τ − τ − τ + −
2
(35)
−
−
−
≤ − + & θ&
+ τ + − τ
Using the known properties of the dynamic model of robot in general which stipulate that the inertia matrix and its
n
inverse is positive definite and bounded (i.e. ∃0 < α ≤ β < ∞ , such that αI
≤ A(
θ ) ≤ βI
∀θ
∈ R ) and if
we suppose we know (or we estimate it with great values) the upper bound 1 2 δ δ δ of the matrix H , we obtain
1
~
T
T 1 1
T 1
V& ⎡
⎤
−
≤ − e Qe + e Pb & θ&
I 2 τ c I 2δ
− e PbA τ s
2
⎢ + +
α α
⎥
(36)
⎣
⎦
⎡ ~
T
1
If we define τ s = sgn( e Pb ) βI
&
⎢θ&
2 + I 2τ
c
⎣ α
1 ⎤
+ I 2δ
α
⎥ (37)
⎦
s
n
T ( [ ] )
n

sgn( x )
0
where equal 1 if x ≥ and -1 if x < 0 .
Substituting equation 37 in equation 36, we obtain
V&
≤ −
8 Illustration of the proposed NF Controller
1
2
e
T
Qe ≤ 0
The implementation of this controller on the real robot (which is not a negligible task and which can prove to be
difficult because of the interaction between the software and the hardware) was not going to give us quantitative and
qualitative measurements allowing the comparison of this controller with the range of some considered controllers and
that in various situations. It is for that reason we have recognized the need of having a software to simulate dynamically
the leg of the robot. We have used the simMechanics toolbox of the MathWorks, Inc. The Figure 12 shows all the block
diagrams of the considered controller. This diagram has the leg model of the AMRU5 (block legModel) and the initial
PD-like Fuzzy Logic Controller as subsystems. Figure 13 shows the comparison between the proposed adaptive
controller and the initial controllers (PD-like controller) when tracking a square signal. We can see on this Figure how
fast the adaptive controller reaches the set point and how small is the final error. On the simMechanics legModel, we
have an entry where we can change the external force acting of the robot (it can be the case when the robot is on a
slope, the payload is changed by adding for example devices on the robot (battery, sensors…)). Doing that (5N acting
horizontally at the foot), we can see on Figure 14 that the response of the non-adaptive controller becomes worse while
the adaptive controller adapts to the new situation.
Figure 12: The blocks of the ANFIS controller designed

Figure 13: Response of the adaptive controller (solid
line) and of the initial controller (doted line) to a square
signal
9 Conclusions
Figure 14: Response of the adaptive controller (solid line)
and of the initial controller (doted line) to a square signal and
change of the load
In this paper, we have shown the way to derive parameters of a zero-order Sugeno Fuzzy Logic Controller from the
Ziegler-Nichols method. This method makes a Fuzzy Logic Controller to behave as a classical controller (PID, PI, PD,
P). As zero-order Sugeno Fuzzy Logic is a particular case of known fuzzy reasoning methods (Mamdani, first-order
Sugeno, Tsukamoto), we can conclude that the performance comparisons between a PID and Fuzzy Logic controllers
which can be found in some paper are debatable. A Fuzzy Logic Controller includes the classical PID controller but it is
a non-linear controller and it can cope with more complex situations like variable payload. We have also developed a
method on how to find the model and the parameters of the process using the adaptive Fuzzy Inference System. We
have tested this method on a two link planar manipulator because we have a mathematical model of it. The comparison
between the outputs of the method developed and the mathematical mode show the validity of it. Our method is very
general because it is based on the properties of the equation describing the dynamic model of robots. Finally, we have
presented the way to adapt parameters of the initial controller design. The Lyapunov method has been used to prove that
the controller designed is bounded. This method is based on the model of the process obtained, on the estimation of the
lower and upper bound of the inertia matrix and also the upper bound of the matrix containing the Coriolis, the
centrifugal, the gravitational and the frictions vectors.
10 References
[1] T. Yoshikawa, “Foundations of Robotics: Analysis and Control”. Massachusetts Institute of Technology, USA, 1990
[2] D. Nauck, F. Klawonn and R. Kruse. “Combining Neural Networks and Fuzzy Controllers” FLAI’93, Linz, Austria,
Jun. 28-Jul.2, 1993
[3] J. –S. R. Jang, C. T. Sun and E. Mizutani. “Neuro-Fuzzy and Soft Computing” . Prentice-Hall (UK), 1997
[4] B. Subudhi, A. S. Morris, “Fuzzy and Neuro-Fuzzy approaches to control a flexible single-link manipulator ”
IMechE 2003, 29 May 2003
[5] J.G. Ziegler and N.B. Nichols, “Optimum settings for automatic controllers”, Trans. ASME, 64, 759, 1942
[6] S. Sastry and M. Bodson, “Adaptive control: Stability, Convergence and Robustness”, Englewood Cliffs, NJ:
Prentice-Hall, 1989.

Evaluation of CORBA communication models for the
development of a robot control framework
Eric Colon
Unmanned Ground Vehicles Centre
Royal Military Academy
Avenue de la Renaissance 30
B-1000 Brussels, Belgium
ABSTRACT
This paper reports on ongoing work in the specification and development of a modular
software control framework for mobile robots. In order to allow flexibility, a systematic
analysis of requirements has been conducted and their consequences on the framework
architecture are summarised in this paper. We show that the free CORBA implementation
library ACE_TAO fulfils the communication requirements of the framework. We present in
details the communication models defined by CORBA and compare results obtained by
applying each of them on a typical distributed application.
Keywords: software modularity, distributed control, control framework, CORBA
1 INTRODUCTION
Many researchers in robotics are confronted with the same problem: they have at their
disposal many excellent algorithms but due to the lack of standard it is almost impossible to
easily reuse those bricks into new applications. Existing programs have to be modified,
translated, ported, or simply (!) completely rewritten from scratch when changing/updating
the robotic platform. What is needed is a software framework that enables agile, flexible,
dynamic composition of resources and permits their use in a variety of styles to match present
and changing computing needs and platforms. Since a couple of years, some researchers have
begun to work in this way.
At the Unmanned Ground Vehicle Centre (UGV-C), we are dealing with command and
control applications including tele-monitoring, tele-operation (including shared, traded and
supervised control), collaboration (between users), for single and multi-robots (of the same or
different models). The present efforts are justified by the fact that most of the tools developed
by the research community deals with autonomous robots (MCA, DCA, MIRO, GeNoM,...).
Note that we are not dealing with hard real-time control like the Orocos project
(www.orocos.org). It addresses higher-level components like planning and user interaction.
In the next section, the requirements of the control framework are summarized and related to
the properties of the ACE_TAO library. Afterwards, we examine the communication models
defined by the CORBA specifications and relate them to application needs. Finally we
conclude with some guidelines about the communication model selection.

2 FRAMEWORK REQUIREMENTS AND CORBA
The requirements for the control framework have been presented in [1]. These requirements
have been inferred from a systematic analyse of typical tele-robotic applications [2]. In the
first reference we also justify the choice of CORBA and more particularly the ACE_TAO
CORBA implementation among other middleware technologies for our future developments.
We show here that the communication requirements of the framework are fully compatible
with the free CORBA implementation library ACE_TAO.
• Integration of different robotic systems: use of native libraries (C/C++): ACE_TAO is
mainly written in C++, which is the de facto language of most robot libraries.
• Concurrent control of several robots: distributed and multi-threaded processes, easy
communication and process synchronisation. This is the reason of the existence of
ACE.
• Universal GUI, flexible programming language and implementation solution: the
choice of ACE_TAO does not restrict the GUI developments. GUI can be based on
C++ toolkits or on interpreted language. Furthermore, CORBA communication is
straightforward in Java.
• Shared control between several users requires management of access and use policy as
well as coordination between control modules: this is not directly linked to the choice
of ACE_TAO. However, the network capabilities could facilitate the implementation
of such functions.
• Integration of user algorithms: requires run-time configuration capabilities, portability.
By modifying the server registration data in the naming/trading services different
servers (functionalities) can be selected at run-time.
• Flexibility:
− Distribution: This is the core function of CORBA.
− Modularity: by using Object Oriented programming and interfaces using the
Interface Definition Language, capabilities can be divided among several
components.
− Configurability: the naming/trading services could contribute to solve the
configurability issues.
− Portability: ACE_TAO is a universal library that can be used on almost all
platforms.
− Scalability: ACE_TAO provides many components suited for applications
involving many processes.
− Maintainability: ACE_TAO is based on many standard design-patterns.
CORBA is an industry standard.
• Performance and efficiency: ACE_TAO has been developed for time-critical
applications used in aviation and medicine. Many years of development and
improvement have lead to very efficient software implementation. ACE_TAO
provides many RT components (RT Event communication, RT-CORBA,...).

3 CORBA COMMUNICATION MODELS
CORBA offers different methods to implement communication and data transfer between
objects. The basic communication models provided by CORBA are synchronous two-way,
one-way and deferred synchronous 1 . To alleviate some drawbacks of these models
Asynchronous Method Invocation (AMI) has been introduced. The Event Service and the
Notification Service provide additional communication solutions. The following of this
section briefly describes all these models and discusses their benefits and drawbacks.
Synchronous two-way
In this model, a client sends a two-way request to a target object and waits for the object to
return the response. The fundamental requirement is that the server must be available to
process the client’s request.
While it is waiting, the client thread that invoked the request is blocked and cannot perform
any other processing. Thus, a single-threaded client can be completely blocked while waiting
for a response, which may be unsatisfactory for certain types of performance-constrained
applications.
The advantage of this model is that most programmers feel comfortable with it because it
conforms to the well-know method-call on local objects.
One-way
A one-way invocation is composed of only a request, with no response. One-way is used to
achieve “fire and forget” semantics while taking advantage of CORBA’s type checking,
marshalling/unmarshalling, and operation demultiplexing features. They can be problematic,
however, since application developers are responsible for ensuring end-to-end reliability.
The creators of the first version of CORBA intended ORBs (Object Request Broker) to
deliver one-way over unreliable transports and protocols such as the UDP. However, most
ORBs implement one-way over TCP, as required by the standard Internet Inter-ORB Protocol
(IIOP. This provides reliable delivery and end-to-end flow control. At the TCP level, these
features collaborate to suspend a client thread as long as TCP buffers on its associated server
are full. Thus, one ways over IIOP are not guaranteed to be non-blocking. Consequently,
using one-way may or may not have the desired effect. Furthermore, CORBA states that oneway
operations have “best-effort” semantics, which means that an ORB need not guarantee
their delivery. Thus, if you need end-to-end delivery guarantees for your one-way requests,
you cannot portably rely on one-way semantics.
Deferred synchronous
In this model, a client sends a request to a target object and then continues its own processing.
Unlike the way synchronous two-way requests are handled, the client ORB does not explicitly
block the calling thread until the response arrives. Instead, the client can later either poll to
see if the target object has returned a response, or it can perform a separate blocking call to
wait for the response. The deferred synchronous request model can only be used if the
requests are invoked using the Dynamic Invocation Interface (DII).
1 All specification documents over CORBA are available on the OMG web site: http://www.omg.org.

The DII requires programmers to write much more code than the usual method (Static
Invocation Interface or SSI). In particular, the DII-based application must build the request
incrementally and then explicitly ask the ORB to send it to the target object. In contrast, all of
the code needed to build and invoke requests with the SII is hidden from the application in the
generated stubs. The increased amount of code required to invoke an operation via the DII
yields larger programs that are hard to write and hard to maintain. Moreover, the SII is typesafe
because the C++ compiler ensures the correct arguments are passed to the static stubs.
Conversely, the DII is not type-safe. Thus, the programmer must make sure to insert the right
types into each Any or the operation invocation will not succeed.
Of course, if one can’t afford to block waiting for responses on two-way calls, he needs to
decouple the send and receive operations. Historically, this meant the programmer was stuck
using the DII. A key benefit of the CORBA Messaging specification is that it effectively
allows deferred synchronous calls using static stubs (automatically generated communication
methods hiding CORBA complexities), which alleviates much of the tedium associated with
using the DII.
CORBA Messaging
The CORBA Messaging specification introduces the Asynchronous Method Invocation
(AMI) model. As we saw in the preceding section, the standard CORBA doesn’t define a
truly asynchronous method invocation model using the SII. A common workaround for the
lack of asynchronous operations is to use separate threads for each two-way operation.
However, the complexity of threads makes it hard to develop portable, efficient, and scalable
multi-threaded distributed applications. Moreover, since support for multi-threading is
inadequately defined in the CORBA specification there is significant diversity among ORB
implementations.
Another common workaround to simulate asynchronous behaviour in CORBA is to use oneway
operations. For instance, a client can invoke a one-way operation to a target object and
pass along an object reference to itself. The target object on the server can then use this object
reference to invoke another one-way operation back on the original client. However, this
design incurs all the reliability problems with one-way operations described in previous
section. To address these issues, CORBA Messaging defines the AMI specification that
supports a polling and a callback model. Only the callback model of the CORBA Messaging
specification has been implemented in ACE_TAO.
The internal mechanism is actually based on two normal synchronous invocations in both
directions. Remarkably, adding asynchrony to the client generally does not require any
modifications to the server since the CORBA Messaging specification treats asynchronous
invocations as a client-side language mapping issue.
The Events Service
There are many situations where the standard CORBA (a)synchronous request/response
model is too restrictive. For instance, clients have to poll the server repeatedly to retrieve the
latest data values. Likewise, there is no way for the server to efficiently notify groups of
interested clients when data change.
The OMG COS Events Service provides delivery of event data from suppliers to consumers
without requiring these participants to know about each other explicitly. A Supplier is an

entity that produces events, while a Consumer is one that receives event notifications and
data. The central abstraction in the COS Events Service is the Event Channel, which plays the
role of a mediator between Consumers and Suppliers and supports decoupled communication
between objects. Events are typically represented as messages that contain optional data
fields.
Suppliers and Consumers can both play an active or a passive role. A PushSupplier object can
actively push an event to a passive PushConsumer object. Likewise, a PullSupplier object can
passively wait for a PullConsumer object to actively pull an event from it.
By combining the different possible roles for consumers and producers, we obtain the four
canonical models of component collaboration in the OMG COS Events Service architecture.
While Push type streams are preferred when the supplier/consumer work at the same pace,
Pull type streams are best suited when data processing is slower than possible data production
or when it is requested at random.
Benefits:
• Producers do not receive callback registration invocations. Therefore, it need not
maintain any persistent storage for such registration
• The event channel ensures that each event is distributed to all registered Consumers.
• The symmetry underlying the Events Service model might also be considered as a
benefit. It simplifies application development and allows Event channels to be chained
together for bridging or filtering purposes.
Drawbacks:
• A complicated consumer registration (multiple interfaces, bi-directional object
reference handshake,...),
• The lack of persistence that can lead to events and connectivity information lost,
• The lack of filtering that leads to increased system network utilisation especially when
multiple suppliers are involved.
The Notification Service
Two serious limitations of the event channel defined by the OMG Event Service are that it
supports no event filtering capability and no ability to be configured to support different
qualities of service. Thus, the choice of which consumers connected to a channel receive
which events, along with the delivery guarantee that is made to each supplier, is hard-wired
into the implementation of the channel. Most Event Service implementations deliver all
events sent to a particular channel to all consumers connected to that channel on a best-effort
basis.
A primary goal of the Notification Service is to enhance the Event Service by introducing the
concepts of filtering, and configurability according to various quality of service requirements.
Clients of the Notification Service can subscribe to specific events of interest by associating
filter objects with the proxies through which the clients communicate with event channels.
These filter objects encapsulate constraints which specify the events the consumer is
interested in receiving, enabling the channel to only deliver events to consumers which have
expressed interest in receiving them. Furthermore, the Notification Service enables each

channel, each connection, and each message to be configured to support the desired quality of
service with respect to delivery guarantee, event aging characteristics, and event
prioritisation.
The Notification Service attempts to preserve all of the semantics specified for the OMG
Event Service, allowing for interoperability between basic Event Service clients and
Notification Service clients. The Notification Service supports all of the interfaces and
functionality supported by the OMG Event Service.
The TAO implementation does not support Pull interfaces and Typed Event style
communication. Work is underway to implement the TAO Real-Time Notification Service.
This is an extension to TAO's CORBA Notification Service with Real-Time CORBA support.
Dead or unresponsive consumers and suppliers are detected and automatically disconnected
from the Notification Service.
4 EVALUATION OF CORBA COMMUNICATION MODELS
A comparative example using different communication models has been implemented. It
provides CORBA wrapping to serial communication.
Standard two-way communication model
A serial server has been developed using the TAO library. A client program reads inputs from
the console and sends corresponding commands to the serial server. This server
communicates through the serial port to another program simulating a micro-controller
controlling a robot (Synchro). It reads translation and rotation inputs and adapts these values
to the robot kinematics. The right and left speeds values preceded by the command code are
sent to the serial server. The next figure illustrates this application.
server
RS232
3 2
1
4
Ethernet
Synchro
Client
The time sequence and screen captures are shown in the next figure.
It has been verified that the serial server can process method calls and transfer data to/from
the serial port concurrently. TAO provides several concurrency models and in our tests the
default configuration has been used, namely a single-threaded, reactive model. One thread
handles requests from multiple clients via a single Reactor [3].

Client SerialServer Synchro
1
4
It is appropriate when the requests take a fixed, relatively uniform amount of time and are
largely compute bound. This is the case in this application where the response of the remote
serial client program is immediate and is determined by a 50 ms timer. The single thread
processes all connection requests and CORBA messages. Application servants need not be
concerned with synchronizing their interactions since there is only one thread active with this
model. [4]
AMI callback model
AMI helps solve the problems with waiting efficiently for long latency calls to complete.
With this technique, long-running calls don't interfere with other calls. It allows singlethreaded
applications to avoid blocking while waiting for responses.
One of the advantages of this model is that existing CORBA servers need not be changed at
all to handle AMI requests. Furthermore, the TAO IDL compiler automatically generates
methods that implement the AMI communication model. A reference to a handler object is
passed as an additional parameter to the invocation method and the response is received in
this handler object. The class handler declaration and implementation are also automatically
generated.
The serial server has been modified to generate a one second timeout. Requests are sent in a
loop and responses are arriving in the right order (counter value) with a time difference of 1
second.
The AMI client callback model requires client programmers to write more code than with the
synchronous model. In particular, client programmers must write the Handler servant, as well
as the associated client event loop code to manage the asynchronous replies.
2
3

Events and notification model
While above communication methods are only based on CORBA core implementation, events
and notification communication models rely on additional services. They allow the creation of
events or notification channels that will are used by producers and consumers.
In these models, producer and consumer are now decoupled. If we keep the same data flow,
the client plays now the role of producer pushing data and the serial server the role of
consumer. The data are pushed to the consumer by the event channel.
The data flow throughput is in this case limited by the serial communication speed that is far
slower than the TCIP one. In case of slow data production (manual input), the application
behaves like the original one. But if the data production runs faster, this could rapidly leads to
buffer overflow and lost data. If the communication is broken or the application on the other
side is slow or response times are variable, we cannot absorb the data flow.
In the previous implementation, if there is no response from the system connected to the serial
port, the caller blocks and returns after a timeout of 1sec. It means that if we use a push
model, the push period should be larger than 1 sec. Consequently, it should be better for the
serial server to pull the data (PullConsumer) and for the client to provide it on request
(PullSupplier).
In the case of a mobile robot a joystick could produce motion commands at regular intervals
(typically 50 ms) and consequently a push model is best suited for the Supplier. So we get a
Push/Pull hybrid model and the Event Channel act as a queue. Another solution is to modify
the serial server implementation by using non-blocking (serial) communication.
What happens if a pushconsumer blocks on the call of the push() invocation?
If we use a single threaded Event Channel, all communications will also be blocked. So, it
could perhaps be better to use a thread by connection or thread by client model for the ORB.
One of the advantages of the synchronous two-way communication model is to return the
state of the communication on the serial port to the client. By using an event model, we lose
this capability. It means that we need to use another method if we want to monitor the serial
communication state.
We see that using models with higher capabilities also requires more effort to program and to
maintain and also more resources from the system.
Which model for which task?
According to the task to be accomplished, different communication models can be selected.
From model properties and tests described above, we derive the following guidelines.
Configuration
Synchronous two-way is best suited for light operations. It can be used for configuration tasks
or to evaluate the availability of resources.
Control

An Event model is best suited for components which continuously produce/consume data and
for processes requiring RT capabilities.
Motion commands are generated continuously (periodic or not): mouse click, joystick,
manual commands, file, path generator, ... and are best propagated as events.
Data processing
Asynchronous Method Invocation can be used for components requiring long computation
times (planning, localisation, stereovision,...).
Visualisation
Data are continuously produced (converted by intermediate components and forwarded to the
next one) to finally arrive with the consumers. An event model (Notification) is best suited in
this case.
7 CONCLUSION
Developing modular control software requires a systematic and detailed analysis of the
applications requirements. Furthermore, adopting programming standards like CORBA and
the choice of open-source software is the only way, according to us, to reach this software
modularity.
CORBA provides different communication models that suit different users’ needs. By using
more sophisticated models, like the Events model, we can develop more flexible software. On
the other hand, it generally requires to write more code and to modify data flow models.
If using simple CORBA synchronous model is not more complicated than writing sockets
based programs, other models require assimilating new communication paradigms and
thinking differently. CORBA has certainly a steep learning curve but offers many benefits for
writing heavy distributed applications.
REFERENCES
[1] Software Modularity for Mobile Robotic Applications, Eric Colon, Hichem Sahli, Clawar
conference, September 2003, Catania, Italy.
[2] Telematics Applications in Automation and Robotics - TA2001, July 2001, Weingarten,
Germany
[3] Pattern Languages of Program Design, Jim Coplien and Douglas C. Schmidt, Addison-
Westley, 1995, ISBN 0-201-6073-4
[4] Configuring TAO's components, Douglas C. Schmidt, http://www.cs.wustl.edu/
~schmidt/ACE_wrappers/TAO/docs/configurations.html.

4-D GPR Image Based on FDTD Parallel Technique
ABSTRACT
Ground penetrating radar (GPR) has become an
established technology and provides some
valuable information to aid the detection and
identification image processes of landmine, so
there is a continuation of the research on the
advanced algorithms for the processing of GPR
data, especially for 4-D GPR image. FDTD
method is a useful method for 4-D GPR image of
landmine detection, but it demands a great
amount of memory and computer time. In this
paper, we will describe our experience with 4-D
GPR data image and will discuss some steps
done in the improvement of the speed of our
parallel FDTD algorithm.
1 Introduction
Many years ago image processing was
performed aiming to create an image of the
landmine detection. One of the recent
developments is four-dimensional (4D) image of
landmine [1]. The 4D (three-space and time)
measurement is concerned with a geophysical
measurement that is repeated in time in the same
three-dimensional (3-D) configuration and it has
the following physical connotation: two
dimensions are related to the spatial surface
coverage of the experiment; one dimension is
related to the recording time of one measurement
and later on in the image formation transferred to
the depth dimension coverage. The sampling in
the repetition dimension is larger than the
duration of the coverage. Moreover in the case of
the number of instants on the repetition axis will
be sparse due to economic reasons. The sampling
rates of the common 3D measurement are
selected such that the 3D heterogeneity is
sufficiently captured. In addition, the repetition
of interval has to be chosen so that the localized
* visiting from Computer Science and Technology
Institute, Harbin Engineering University, China
Baikunth Nath , Jing Zhang*
Department of Computer Science &Software Engineering,
University of Melbourne,
Melbourne, VIC 3010, Australia,
baikunth@unimelb.edu.au,jzhang@cs.mu.oz.au
change of 3D heterogeneity is sufficiently
captured.
Ground-penetrating radar (GPR) has been used
for several years as a non-destructive method of
detection and locating landmine. Just as seismic
reflections are generated when a seismic wave
hits a layer in the subsurface with different
material properties, GPR reflections are
generated when a pulse hits an object or layer
with different electromagnetic characteristics.
According to the different reflected signal data,
landmine image can be built.
For solving electromagnetic problems, Yee
proposed the finite-difference time-domain
(FDTD) technique in 1966. At beginning,
because of need sufficient computing resources,
there was little interest in the FDTD method [2].
However, with the advent of low cost, powerful
computers and advances to the method itself, the
FDTD technique has become a popular method
now. However, the requirement for high
performance computing in engineering is endless.
Many fields such as complex object modeling,
engineering design and automation, electronics,
aeronautics and medicine put forward challenges
to computing. The calculating is impossible for a
PC even through the performance of PC is much
higher than before. Now the parallel FDTD
algorithm research have already progressed a lot
[3,4], as the application of it, 4D GPR image
processing became fast.
2 FDTD Parallel Technique
Since mid 1990’s, the finite-difference time
domain (FDTD) method has been successfully
applied to solving problem in ground-penetrating
radar systems for landmine detection.
FDTD computation involves a time-based
leapfrog method [5, 6] where updating equations
alternate between electric field and magnetic
field calculations. In order to realize it,
Maxwell’s equations are transformed from their
vector, differential form into difference

equations.
In the normal FDTD method, the EM field
(electromagnetic field) values of every last time
step and other variables are stored in the memory
of a computer. The larger the object’s electrical
size is analyzed, the more memory the computer
needs. For example, a 32 MHz memory
computer only has the capable for the space size
about 80×80×80 grids. So, we often use a more
powerful workstation or even gigantic computer
to analyze the EM problems of a large electrosize
object. But if we only have a PC or a
workstation that have no enough memory, but
we still want to handle with an electrically largesize
object that we would meet a serious problem.
In order to solve this problem, a high efficiency
method is implement FDTD algorithm in a
parallel computer. There are some researchers
have already developed a little strategy for
FDTD parallel computing [7, 8, 9, 10], what’s
more, the research are becoming bloom.
This paper introduced a strategy for parallel
implementation of the FDTD algorithm using
COW (Cluster of Workstation) parallel
computing system that are built with LINUX
operating system and PVM parallel software.
Some computing examples were given to prove
the feasibility, correctness and high efficiency of
this strategy.
The main advantages of this method lie in three
aspects :Firstly, this method is a time domain
method that means the data in the whole
frequency band can be obtained by only once
calculation in time-domain. Secondly, this
method can easily model complex objects.
Thirdly, the necessary memory is relatively
smaller than other low frequency numerical
technique such as the Moment Method [11, 12].
The details on FDTD algorithm have already
been discussed in many references such as [13,
14]. The preconditions of the parallel FDTD
algorithm is the dividing of all computation task
and then every node of COW can compute one
part of all computation task. Firstly, the basic
principle of FDTD algorithm makes the division
of computational space possible. According to
the principle of FDTD algorithm, the
electromagnetic field value at certain position
can be decided by the value of last time step at
this position and electromagnetic field value of
this time step at nearby position.
The electromagnetic field value has no direct
relation to the values at position far from this
point. So, the whole computational space can be
divided into sections that can be commutated in
some nodes of parallel computing system. The
exchange of field values between nodes can be
executed only at interface between sections.
According to the basic cognition, the relay
computing between parallel nodes can be
executed to simulate the serial computing in a
single PC or workstation.
According to the structure of software described
above, we built a test COW system with two PC
and 10M Ethernet. The two PC are PП266/64M
and PП200/64M, which have Red-Flag Linux2.4
and PVM3.4.2. For proving the feasibility,
correctness and high efficiency of this system,
some computing problems have been executed
by our parallel FDTD computing.
COW system [15] can be constructed by two
parts: workstation and interlinkage network. It
Shows in Fig.1. This is the key point of our
parallel FDTD algorithm. The Master-Slave
programming-style has been used. The Master-
Slave program has three steps normally: Firstly,
slave programs are created by master program
and the information of task and the parameter of
computing are delivered from master program to
every slave program; Secondly, the parallel
computing are executed in every node of COW,
within every time step the synchronization and
communication between every node are kept;
Finally, the computing results are transferred to
master program from every node and the
computing are terminated. Taking two nodes
COW as an example, the computing task are
divided into two parts along Y axis in one
dimension. The segmentation is shown in Fig.2,
and this figure also shows the data needed to
transfer between nodes. Taking the iterative
computing of H x and E z components of onedimension
as an example, the difference formula
for j = m -1/2 and j = m are showed as follows.
From Fig 2, we know that we must deliver the
value of Ez(m) from section 2(No.2) to section
1(No.1) before the value of Hx(m-1/2) is
calculated, and same thing, we must deliver the
value of Hx(m-1/2) from No.1 to No.2 before the
value of Ez(m) is calculated.
The main loop proceed by computing data then
sending updated data values to processors
responsible for adjacent sub-domains.

Synchronization across all nodes is required
before this exchange can occur. Two cycles of
computation, synchronization and data exchange
are required for each time step.
Go through the FDTD processing, we can get the
processed GPR data from raw GPR data. That
means the image from processed PGR data can
get even clearer.
Internet
Router
m-1
m+1
m
* # *
#
*
No.1 section
Workstation PC
Interface
Hub
PC PC
Fig.1.Structure of Parallel
System
No.2 section
3 4D GPR Image Processing
* Ez
# Hx
Fig.2 the one-dimension segmentation and the
data need to transfer between nodes
GPR(Ground-penetrating radar) data is gotten by
the method as follows.
If a GPR pulse hits a layer or object with a
different dielectric constant, the pulse is reflected
back, picked up by the receiving antenna, and the
time and magnitude of that pulse is recorded, as
in many cases, the transmitting and receiving
antennas are the same.
Essentially, a reflection occurs when there is an
increase in the dielectric constant of materials in
the subsurface. The dielectric constant is defined
as the capacity of a material to store a charge
when an electric field is applied relative to the
same capacity in a vacuum, and can be computed
using Equation 1:
Where:
εr = The relative dielectric constant,
(1)
c = The speed of light (30 cm/nanosecond), and
v = The velocity of electromagnetic energy
passing through the material.
GPR data is divided into three types [16, 17, 18,
19], that is, A-scan, B-scan, C-scan. A-scan is a
1D representation of a single GPR profile (trace),
B-scan is a 2D representation of a series of GPR
traces, and C-scan is a 3D representation of a
series of 2D traces, they are shown in Fig.3 and
Fig.4.
After getting the processed GPR data processed
by FDTD technique, we still need to create 4D
image from the processed GPR data if we want
to get a procedure of the detection of landmine,
As for 4-D representation of GPR data, we got it
from processed C-scan PGR data and add time
pixel by using 3D doctor trial software, it’s
shown by Fig.5. The landmine type is T72 APL
and it’s buried under the sand in 5 cm. The
sample picture in Fig.5 is four 3D images
connection.
4 Conclusion
We need not concern the normal setting of
FDTD algorithm but assure the computing object
is absolute same in parallel computing and
reference serial computing. Taking a GPR data
calculation as an example, the whole computing
space includes 76×76×76 grids and is divided
into two parts equally along Y-axis. The
recorded results are shown in Fig.4 and Fig.5 and
the results obtained from normal FDTD are also
shown in the two figure. According to the
computing, the accelerating ratio is 1.6 and the
parallel efficiency is about 0.8. This means the
parallel FDTD system is efficient. So we can
know that this strategy of parallel FDTD

Algorithm is successful and the precision is
absolutely same as normal serial FDTD, so we
can say this system is feasible.
Fig.3 A-scan (Left) and B-scan (Right) Representations
of GPR Data; T72 APL in 5 cm Sand
Refrences
[1]-- “Block iterative techniques for fast 4D
reconstruction using a priori motion models in
gated cardiac SPECT,” Phys. Med. Biol., vol.43,
pp. 875–886, April 1998.
[2]Amir Fijany, Michael A.Jensen, Yahya
Rahmat-Samii, Jacob Barhen, “A massively
Parallel Computation Strategy for FDTD: Time
and Space Parallelism Applied to
Electromagnetic Problems”,IEEE Trans. on AP,
pp 1441-1449, 1995
[3] Cheng Lifeng. “Red Hat Linux 6.0 Network
and Communication”, China People’s Posts &
Telecommunication
Publishing House, 2000.
[4] Chen Jiu, Long Hao, “Red Hat Linux 7 Self-
Culture Guide”, Publishing House of National
Defence Industry, 2001.
Fig.4 C-scan (Left) Representations of GPR
Data; T72 APL in 5 cm Sand
Fig.5 4D Representations of GPR Data; T72 APL
in 5 cm Sand
[5] U Oguz, L. Gurel, “Frequency Responses of
Ground-Penetrating Radars Operating Over
Highly Lossy Grounds”, IEEE Trans. On
Geoscience and Remote Sensing, Vol. 40, No. 6,
pp 1385-1394,2002
[6] W.P. Pala, A. Taflove, M.J. Piket, R.M.
Joseph, “Parallel Finite Difference Time Domain

Calculations”, IEEE Trans. Antennas &
Propagation, Vol. 30, No. 3, pp 83-85, 1991
[7] A Taflove, S.C. Hagness, “Computational
Electrodynamics – The Finite Difference Time
Domain method”, Artech House, 1995
[8] V. Varadarajan, R. Mittra, “Finite-Difference
Time-Domain Analysis using Distributed
Computing”, IEEE Microwave and Guided
Wave Letters, Vol. 4, pp. 144-145, 1994
[9] Q. Chu, K. Chan, C. Chan, “Parallel FDTD
Analysis of Active Integrated Antenna Array”,
Antennas and Propagation Society International
Symposium, Vol. 1, pp 628-631, 2002
[10] Z.M. Liu, A.S. Mohan, T.A. Aubrey, W.R
Belcher, “Techniques for Implementation of the
FDTD Method on a CM-5 Parallel Computer”,
IEEE Antennas & Propagation Magazine, Vol.
37, No. 5, pp 64-71, 1995
[11] T.V. Pistor, “Electromagnetic Simulation
and Modeling with Applications in Lithography“,
Ph.D Dissertation, University of California
Berkeley, 2001
[12] A.D. Tinniswood, P.S. Excell, M.
Hargreaves, S. Whittle, D. Spicer, “Parallel
Computation of Large-Scale FDTD problems”,
Third International Conference on Computation
in Electromagnetics, Vol. 1, pp 7-12, 1996
[13] M. Sypniewski, J. Rudnicki, M. Celuch-
Marcysiak, “Investigation of multithread FDTD
schemes for faster analysis on multiprocessor
PCs”, IEEE Antennas & Propagation
International Symposium, Vol. 1, pp 252-255,
2000
[14] G.A. Schiavone, J. Codreanu, R.
Palaniappan, P. Wahid, “FDTD speedups
obtained in distributed computing on a Linux
Workstation Cluster”, IEEE Antennas &
Propagation International Symposium, Vol. 3, pp
1336-1339, 2000
[15] C. Guiffaut, K. Mahdjoubi, “A Parallel
FDTD Algorithm using the MPI Library”, IEEE
Antennas & Propagation Magazine, Vol. 43, No.
2, pp 94-103, 2001
[16]T.-S. Pan, D.-S. Luo, and M. A. King,
“Design of an efficient 3-D projector and
backprojector pair for SPECT,” in Proc. Fully-
3D Image Reconstruction Radiol. Nucl. Med.,
1995.
[17]Ma Jifu, “Generalized Computation of 3-
Dimensional EM Field of Resonant Cavities and
Its Applications”, Thesis for Ph. D of China
Institute of Electronics, 1998
[18]. Narayanan, M. A. King, E. J. Soares, C. L.
Byrne, P. H. Pretorius,and M. N. Wernick,
“Application of the Karhunen–Loève transform
to 4D reconstruction of cardiac gated SPECT
images,” IEEE Trans. Nucl. Sci., vol. 46, pp.
1001–1008, Aug. 1999.
[19]Manoj V.Narayanan,”Improved Image
Quality and Computation Reduction in 4-D
Reconstruction of Cardiac-Gated SPECT
Images” IEEE Trans. Nucl. Sci., VOL.
19,NO.5 ,MAY 2000.

Abstract
Paper deals with applications of robots for performing
risky tasks in environments dangerous for humans,
or, tasks where the robot should prevent human
life or big economical as well as ecological
damages. These applications fields require specific
robotic systems that exhibit specific features, some
limited level of autonomy as to mobility and task
oriented functions what directly corresponds to
availability of sensory information from the environment.
Any agent working in risky environment
except its functional performance should exhibit
three features: selfrecovery capabilities, minimal
risk assessment and maximal reliability in all actions.
Required performance is discussed on exa mples
of robotic agents for demining, searching /
cleaning dangerous environments and some surveillance
tasks. The concept of modular construction is
briefly presented.
Key words : Service – mobile robots / agents, hazardous
environment, demining, detection, selfrecovery,
1. Introduction
Dangerous tasks or works in hazardous environments
are those tasks or situations that would endanger the
safety of human beings and / or valuable equipment
used or the near by environment in which the emergency
management application takes place. One
could encounter as those environments: physical
catastrophes (earthquakes, sea or river debacles, etc.)
a fire extinguishing mission in uneven terrain, nuclear
accidents, cleaning terrain from landmines, etc.
As human safety is the highest priority, the interest is
to remove the operator from the scene of the hazardous
environment and / or either totally substitute him
by an onboard “intelligent” agent - which is expected
to provide the same or similar functionality, or, to
provide the operator by such means that would enable
him to perform the same mission safely.
Performing these tasks by a “robot” is a big challenge
for research in all domains of robotics. There
are several task oriented vehicles or mobile systems
Robotic Agents For Dangerous Tasks
Features and Performances
Štefan HAVLÍK
Institute of Informatics
Slovak Academy of Sciences, Severná 5,
974 01 Banská Bystrica, Slovakia
E-mail: havlik@savbb.sk
for inspection, for searching dangerous terrain or for
fighting with fires in forestry. But actual need of
humanitarian demining is one of the most challenging
world movement for development new demining
technologies that could be fast, safe and reliable. A
great research effort was devoted especially to development
sensory systems and signal processing techniques
for detection mines or any unexploded ordnance
[1,2,3,4].
Beside these, yet quite well established applications,
the occurrence of terrorist attacks give arise a new
field for applications of robotic technologies: solving
situations due to terrorists attacks, or due to menace
of their occurrence.
2. Analysis of needs for some dangerous tasks
General requirement is that such a task should be
performed fast and safe in order to reduce any risk
for humans as well as material on economics or environment.
Let’s analyze some tasks from the point of
view robotic research and robotic technology.
Humanitarian demining
Lot of research work has resulted in design of several
concepts demining technologies as well as development
new machines and especially detection systems.
But as it seems, despite this effort, several sophisticated
solutions and systems will not find such acceptance
in practical use as could be expected. Let us
mention some main rules that should be taken into
account before starting new development of any tool:
− Minefields are not laboratories. Robust and reliable
constructions as well as control techniques
should correspond harsh working conditions and
environment. This includes solving so called “selfrecovery
strategies” in most crucial situations that
could arise (occasional explosions, errors in systems
/ operators, lost of communication, etc.).
− The cost and availability of detection as well as
neutralization technologies is a very important factor
that could limit their mass use in conflict areas.
Automatic / robotic cleaning should be faster (as to
productivity in m 2 /hour), cheaper (as to total
cost/m 2 ) comparing to standard methods, reliable
and safe.

− Any new demining technology should be easily
accepted by users, as well as local authorities /
people. The robotic system should satisfy specific
conditions related to its local applications (people /
country, cost / salaries for deminers, infected terrains,
climatic conditions, mines, maintenance,
etc).
− There are no universal solutions. Robotic technologies
have not totally replace standard hand
searching / neutralization methods but they will be
applied beside them. Automatic ways are especially
suited for primary detection and cleaning
large areas under some homogenous conditions
(obstacles, mines, vegetation, etc.).
− The reliable detection and localization of mines
(UXO) as targets is the task of primary importance.
It can be said: “ When a mine is found and
localized about 90% of problems are solved”.
− Any new solution should minimize risks for
people as well as for the damage of relatively expensive
technology. This risk of the damage, or,
the lifetime by using new technology should be
calculated in expected comparable total cost for
demininig the unit of surface.
Fire fighting
Fire fighting represents other dangerous task from
several reasons. There are harsh conditions in the
operation space: toxic gases, high temperatures, possibility
of explosions / breaking down of any constructions
or trees, poor / no visibility, obstacles - bad
accessibility, etc. An intended robotic system should
exhibit some principal functions as follows:
- Mobility in complicated terrain with unknown obstacles.
The vehicle can move on wheels, belts, or
depending on obstacles, on legs.
- Sensory equipment. Principal requirement is positioning
the vehicle in a global / local reference system.
For measuring actual position in global world
coordinates can be used GPS devices able to ascertain
position within 1m, or less, resolution The other
problem is to navigate the vehicle according to actual
situations (obstacle avoidance, thermal navigation to
source of fire, solving recovery situations, etc.). To
do this beside visual feedback an additional thermal
imaging system is need for navigation under poor
vis ibility conditions.
- The board of the vehicle should be equipped by an
active system for fire intervention (water / foam
/sand gun, placement explosives to extinguish the
fire, removing objects / obstacles, etc.
- Control and communication should reliably work
and such an agent should exhibit a given degree of
autonomy to solve some unexpected situations.
Anti-terrorists interventions
This field represents a broad range of possible situations
that should correspond particular menaces and
tasks. It should be noted that the attack on the WTC
means that a new strategy should be adopted in order
to prevent / to solve such situations. Till that time the
terrorist actions were oriented more or less to provoke
a public opinion. Economical damages and
number of killed people, including their own lives,
were not so frequent, or, were not the main goal of
terrorist activities. Unfortunately, since that time, the
purpose and strategy of their attacks is to make
maximal damages as well as to kill maximal number
of people. This fact naturally should result in different
approach in prevention as well as protection
against such actions.
3. Some specific performance features for
agents pe rforming dangerous tasks
Performing operations in risky environment requires
some principal features to be satisfied by an intervention
agent.
Self-recovery
This is an important and specific feature directly
related to particular tasks. Its main purpose is to prevent
/ to avoid loses or self-destruction of the agent
and to finish a given action in risky environment
without serious damages. The self-recovery strategies
should start especially in unwanted situations as
follows:
- In cases of any failure of technique (communication,
engine, control system, sensory system,
etc…)
- In cases of fault decision made by the operator
- There are no / not enough information for further
action, it seems to be risky for the agent and it
could be destroyed.
The general requirement is to built agent and all its
functional parts as reliable as possible. A typical
requirement is solving communication as two independent
systems.
Let us consider an agent performing demining operation
in the minefield. Any its motion should be carefully
considered otherwise any wrong movement
could result in explosion of mine with its destruction.
The principal strategy should be based on following
rule: any motion of the vehicle could be realized in
such direction where no mines were reliably detected,
or, all mines were destructed / removed. This
practically means that the vehicle can move in such
direction where detection, or, destruction systems are
fixed.
Minimal risk
The other situation arises in case of any failure and
the vehicle cannot perform desired activities. The
problem is to remove if from the dangerous terrain
without any risk for persons. One of the simplest way
how to solve this situation is using a cable and pull it
out by the winch mechanisms. The other possibility

enables using another vehicle, which helps to remove
the first one from the minefield.
Solving any situation brings for operator / operation
system the decision problem: to decide for the next
action if any unexpected situation arose. The general
rule is: the operator decides for the next paths of the
agent in order to minimize any risk of damages for
the agent itself. This procedure represents the standard
decision algorithms [6] according to the scheme
for risk assessment in Fig. 1.
New approach /
option to reduce risk
Fig. 1. The general risk assessment and decision
procedure
It is obvious that some decisions can be represented
by relatively simple routines working over a given
set of options: Action :
if .< CONDITION: SENSOR ->. On the other hand,
other operator’s decisions require much more complex
assessment of possible risks with respect to
given criteria. Such a typical situation arises during
automatic demining operation when using mine detection
systems give not reliable information about
the presence of mines and there is only a suspicion if
“something is inside”. The operator should decide if
to continue in searching the terrain by the vehicle /
agent with detection systems, or to send the directly
the destruction (flailing) vehicle. It is clear that the
operator will decide for destruction tools for reason
of minimal risk.
Reliability
The reliability of performing a task should be considered
with respect to criteria given for a particular
operation. To compare different requirements and
criteria some features are given on examples in the
table.
Task Task description (exa mple) Description of the action Criteria
Humanitarian
demining
Nuclear power
plant actions
Anti-terrorist
action
No
Statement criteria
and risk levels
Cost? Time?
Evaluation of options
according to criteria
Choosing an approach / option
Estimate risk and
level of hazard
Is the risk
acceptable ?
Yes
Steps of demining:
- Cleaning vegetation, if
any exists
- Detection of mines and
localization
- Destruction of mines on
place or removing
Surveillance and manipulations
with nuclear materials.
Airport action: Remove an
unattended baggage as
potential / suspected explosives
or other dangerous
Robotic agents for performing particular
tasks are operated from the center.
Agents are equipped by detection
systems and tools for destruction /
neutralization. [1,2,3]
The robot hand executes manipulations
with dangerous materials. The
detection – measuring system verifies
level of radiation.
Equipment: Remotely-operated transport
platform, equipped with robot-
manipulator. Portable operator's console
with radio command communication
channel The vision system with
radio-communication channel. Ga mma-detector
includes static detection
unit (gamma locator) - SDU, dynamic
detection unit and informationprocessing
unit
The remotely-operated mobile robotic
agent will approach to the object,
takes it by an arm and inserts the object
into a container on its platform.
• Reliability of cleaning
(up to 100%)
• Cost / effectivity
• Reliability
• Minimum time
• Prevention of accidental
explosion

4. Modular concept
sives or other dangerous
materials inside. During the
action all persons – passengers
should go away
into a protected / remote
zone. The agent should
remove dangerous object
from the space as soon as
possible.
As follows from the study and analysis of particular
applications the agents for all applications exhibit
some “general purpose” parts to perform common
functions as well as some specific “task oriented”
parts. The modular concept of a robotic agent is the
based on separation of the functional parts. Such
approach of the open architecture system could
minimize cost of a general purpose system and the
whole agent, as well.
General purpose system includes the mobility system
– vehicle and on-board manipulation equipment.
The mobility system is usually a remotely controlled
vehicle moving on wheels, belts or legs. Particular
applications differ by requirements on speed,
weight, mechanical protection against environment.
As regards to control and communication systems,
including sensors / detectors related to motion and
maneuvering capabilities main parts of these systems
can be practically unified.
On board manipulation equipment that represent:
a heavy load manipulator / platform and a robot
arm.
Task oriented parts represent special tools and sensory
equipment dedicated to performing a given task.
Such a modular approach and open architecture solution
enables standardization for majority of functional
parts. This concept of modularity can be seen
on next Fig.2 with functional blocks:
− The mobile platform / vehicle (on wheels / belts)
− The heavy load manipulator (2-3 d.o.f)
− The long reach robot arm (5-6 d.o.f)
− Control and communication systems
Robotic arm
with the set of tools
Camera
4 DoF heavy load manipulator
Fig.2. A modular concept of the multi-purpose robotic
vehicle
ject into a container on its platform.
The object is then carried out into a
safe space for further recognition. The
environment is quite well defined as
to terrain, visibility and weight of
objects.
Equipment:
Mobility system, robotic arm, vision,
hand held camera, remote control
(~200m), other sensors / detectors
The pre-project proposal based on this modular concept
with some performance characteristics are given
below:
Fig.3. The vehicle with the platform
Fig.4. The long reach robot arm
The vehicle:
− speed 0-5 km/h
− remote vision (optionally IR)
− global / local positioning (GPS)
− remote control and communication ~ 5 km
− control panel
The 3-4 d.o.f. heavy load manipulator:
− max. load ~ 500kg
The long reach arm:
− 5-6 d.o.f.
− payload capacity 20-30kg
− reach 3m
− the set of primary tools (grippers)
Additional equipment attached to this base structure
gives a task oriented functionality required for the
agent.
5. Robotic agent for demining
Let us go out from the modular concept of agents.
The open architecture enables to build several modi-

fications performing specific tasks as regards to
demining operations.
The agent for searching dangerous terrain and target
localization.
This agent is a practically a general porter of several
detection systems performing mapping dangerous
terrain and localization of dangerous targets. Detection
systems are on the frontal platform fixed on
manipulator, or, on the robot arm.
Neutralization / destruction vehicle.
The neutralization vehicle is the general porter of
various destruction systems (flails, guns, laser gun,
explosives, etc.). The vehicle with similar maneuvering
capabilities should be more robust and be protected
against of the pressure waves during explosions
of mines. The principal configuration is shown
in Fig. 5.
Fig. 5. Vehicle with flailing destruction system
The realization of this vehicle with flailing destruction
tool can be seen in Fig.6.
Fig.6. The “DIANA” flailing vehicle for demining
Robotic arm with set of tools.
The long reach arm will be used for fine works as
follows: removing / deployment neutralization explosives,
probing, fine cutting the vegetation, localization
of mines using additional detectors, etc.
The set of tools consists of probes, cutters, various
grippers, additional sensors for detecting explosives,
etc. On the end of arm will be small camera what
will allow detailed views on the mine and place of its
vicinity. Control is considered to be in local world or
tool reference coordinates.
Besides flailing system on Fig.5, other principles that
activate explosion of mines can be used. There are:
directed energy system, laser gun or sniper rifle. One
of the most reliable methods for destruction mines is
placement additional explosive beside position of the
mine. This is task for robotic arm with set of tools
(grippers, sensors, probes, etc.). Input data for all
these systems are positions / coordinates of objects
recognized as mines.
The modular concept in Fig.7, then consists of common
functional blocks and special attachment that
specifies purpose of such an agent.
GPS
Mobility systems
perception
and
recognition
Supervising and
control center
minefield
robot arm
GIS (maps)
-scanning
vehicles
-tracking terrain
-fine motions
-scanning motions
-maneuvering by
of detectors
-navigation
tools (flail, cutter) and small
-obstacle avoidance
tools
-self-recovery
-target (gross motion
positionning
Neutralization Removing
tools:
obstacles
mechanical Tools:
(flail, digger,
shovel,..)
cutter,
gripper,
others laser,.etc) saw,...
Fig. 7. A modular concept and parts of robotic
vehicle for demining operations
5. Conclusion
The general concept of mobile robotic systems for
performing task in hazardous environment, especially
for demining operations was presented. Real
need and importance for development of these robotic
systems is a big challenge for researchers working
in robotics, control, sensing technologies, signal /
information processing and communication.
As briefly shown, any agent working in risky environment
except its functional performance should
exhibit minimally three features: selfrecovery capabilities,
minimal risk assessment and maximal reliability
in all actions.
A concept of modularity in construction of these
mobile agents is outlined. It is based on the fact that
for the majority of similar tasks it is possible to define
general and common functionalities that a multipurpose
mobile system could satisfy. Then, specific
features for an agent will give the task oriented tooling
and sensory equipment. This concept is shown on
examples of two vehicles for demining operations.
References
[1] Proc. EUDEM2-SCOT –2003 Int. Conf. on Requirements
and Technologies for Detection, Removal
and Neutralization of Landmines and
UXO. Sept. 15-18, Brussel, Belgium

[2] Havlik, S.: Some concepts of mine clearance
robots. Proc. Int. Conf. on Robotics and Automation.
Thai-Pei, Taiwan, Sept. 2003
[3] Proc. HUDEM’02 IARP Workshop on Robots for
Humanitarian Demining., Nov. 3-5, 2002, Vienna
[4] www.gichd.ch;
[5] Salvall, A. Avello, L. Briones, Two Compact
Robots for Remote Inspection of Hazardous Ar-
eas in Nuclear Power Plants. Proceedings International
Conference on Robotics and Automation,
1999.
[6] Kaminski, L. et al.: The GICHD Mechanical
Application in Mine Clearance Study. Proc.
EUDEM2-SCOT –2003 Int. Conf. on Requirements
and Technologies for Detection, Removal
and Neutralization of Landmines and UXO. Sept.
15-18, Brussel, Belgium, pp.335-341

How to Design a Haptic Telepresence System for the Disposal of
Explosive Ordnances
Bernd Petzold
and Michael F. Zaeh
Institute for Machine Tools and
Industrial Management (iwb)
Technische Universität München
Germany
bernd.petzold@iwb.tum.de
Alexander Kron
and Günther Schmidt
Institute of Automatic Control
Engineering (LSR)
Interactive Systems & Control Group
Technische Universität München
Germany
alexander.kron@ei.tum.de
Abstract
Barbara Deml
and Berthold Färber
Human Factors Institute (lfa)
University of the Armed Forces
Germany
barbara.deml@unibw-muenchen.de
At the Technische Universität München (TUM) an experimental telepresence system is developed enabling
the disposal of explosive ordnances (EOD) without any risk for the user. Most current EOD-systems enabling
user control of remote environments are single-armed configurations and display only visual feedback. In
contrast to this, the presented telepresence system constitutes of an intuitive bimanual interface and displays
visual as well as force feedback to the user. In order to generate guiding design principles, two operator
consoles were set up and evaluated experimentally.
1 Introduction
The threatening of blasting compositions to harm military and civil population is more present these days
than ever. To dispose explosive ordnances, most of the time direct human interaction is required inevitably.
Though robots are available for basic operations during the explosive ordnances disposal (EOD), they do not
suffice for all required actions [3]. This is mainly due to the poor interaction between the human operator
(HO) and the robot. Usually, the operator controls the robot by supervising the robot’s actions via a monitor
screen remotely (see Fig. 1). As the monitor screen does not provide any depth cues, it is very difficult to
determine the exact spatial positions of remote objects as well as of the robot end-effector. Besides, current
control panels do not provide an intuitive interface. Smooth movements of the manipulator afford a lot of
training and even expert users are not able to accomplish all the tasks needed for an EOD because of the
following reasons: First, most manipulators are designed to be single-armed configurations. As almost all
human interactions are two-handed they cannot be mapped onto a single manipulator in an optimal manner.
Second, present systems usually do not display force feedback to the HO which in consequence diminishes
a sensitive interaction in the remote environment. Especially for visually hidden operating areas this turns out
to be a major drawback. To improve current EOD-systems, a more sophisticated bimanual interaction,
experiencing haptic (force and touch) feedback, seems to be necessary [5]. Therefore, telepresence
technology is highly suggestive for developing EOD-systems [1, 3].
Figure 1: EOD with the commercially available single-armed TeleRob vehicle [7].
Telepresence is characterized by a high-fidelity human-system interface which enables the HO to feel
‘present’ in a remote environment. To guarantee familiar and intuitive manipulations, force feedback devices
play a major role for the design of such an interface. As the human hand is a very complex organ, the
devices vary extremely in size, shape, and function. Some devices track finger motions and display grasping
forces to the user (e.g. CyberGrasp, from Immersion, Corp. [2]). Other devices present contact forces or the
weight of objects (e.g. PHANToM-Series from Sensable Technologies [6]).

At the TUM two telepresence EOD-systems for the disposal of a remote bounding fragmentation mine are
developed (Section 2). In order to gain a deeper insight in how to design a haptic telepresence EOD-system,
both setups were evaluated experimentally (Section 3, 4).
2 Telepresence Setup
2.1 Teleoperator
The teleoperator consists of two manipulator arms each providing 4 degrees of freedom (DoF) (3
translational and 1 rotational) in motion [3, 4]. For the manipulation task the end-effectors are equipped with
two-jaw grippers; whereas one is arranged vertically for task execution of the dominant hand, the other one
is arranged horizontally accomplishing the non-dominant hand’s operations (see Fig. 2). The shared
workspace of both arms is cube shaped with 60 x 20 x 30 cm³. The maximum gripper opening is 9 cm. Both
arms are equipped with force/torque sensors that measure contact forces as well as the gripping forces while
picking up and holding objects.
2.2 Operator Side
Figure 2: Bimanual teleoperator system.
For the operator side two different setups are realized: setup A (see Fig. 3) resembles to the manipulator
kinematically and provides 4 DoF of motion and active force feedback by the same SCARA configuration as
the teleoperator [4]. The grippers of the telemanipulator are replaced by clamps by which the wrists of both
arms are linked to the haptic display. Thus, the operator perceives appearing contact forces at his/her wrist.
A force/torque sensor which is integrated below the clamps enables the implementation of a dual hybrid
control architecture [4] displaying teleoperator forces one-to-one. For the grasping procedure a CyberGrasp
exoskeleton [2] for each hand is applied so that the user can perceive gripping forces. Consequently, the
overall force transmission is done in a parallel manner by a combined finger-wrist display.
For setup B (see Fig. 3) two PHANToMs (Desktop and 1.5 / 6DoF) are used as input devices enhanced by
two-finger gripping masters so that the operator can perform grasping procedures. As PHANToM devices do
not provide integrated force measurement, a force-pose control architecture is implemented to display the
weight of objects as well as the occurring contact and gripping forces. In order to display hard environmental
contacts, the PHANToM devices are made compliant by local impedance control [3]. Despite of the adjusted
manipulator compliance, the achieved force capability is still less than the magnitude of the measured
contact forces. Therefore, the displayed forces are scaled appropriately and in consequence the maximum
forces are lower compared to setup A. Force transmission occurrs exclusively at the fingertips. As the
PHANToMs are desktop devices, the operator performs the teleoperation sitting in front of the system
console.
In order to ensure natural motions, users are encouraged to reconfigure their workspace during the
teleoperation: By switching a pedal the communication between display and teleoperator is disconnected
and the input devices can be repositioned; a succeeding switch reconnects the user and the teleoperation
continues. During this operating mode the telemanipulator might reach its limits of the workspace. Therefore,
manipulator motions coming close to constraints in joint space are augmented by mapping virtual arrows into
real stereo video images. In addition, an user is supported by augmented force feedback pushing him/her
back into the workspace smoothly [3]. The same kind of augmentation is applied to avoid any collision of
both manipulator arms.

Figure 3: Operator side for setup A and for setup B.
Instead of a conventional visualisation by video-streaming, a stereoscopic view is implemented in both
setups. The scene is recorded by a stereo camera setup with fixed view position and focus. On operator side
the stereo images are displayed by means of a Head Mounted Display (HMD). The three-dimensional
appearance by stereo-view ensures a reliable depth estimation of the scene.
3 Experimental Evaluation
It is obvious that both systems differ in more than one dimension. Though both setups are able to display
contact forces as well as gripping forces, they vary at least in three features: Setup A corresponds better to
the kinematical configuration of the teleoperator and distinguishes by providing higher force feedback,
whereas setup B is designed without a tight system linkage and therefore enables a more natural activation
of the limbs. As these variations could not be held constant or controlled during the experiment much
emphasis was put on a structured qualitative interview with which user requirements were explored
thoroughly. The experimental evaluation was guided by the following hypotheses:
H1 Higher force feedback will reduce the applied contact forces in the remote environment and will
decrease the distance error of the gripper when grasping objects. In consequence users who
prefer setup A will experience this system to be more transparent and sensitive for the task at
hand (amount of haptic feedback).
H2 By a tight system linkage a natural presentation of force feedback is diminished. In consequence
users who prefer system B will experience that interface to be more realistic and classify those
interactions to be more sensitive (presentation of haptic feedback).
H3 The kinematical correspondence of teleoperator and operator is less important than the amount
or the presentation of the haptic feedback. For this reason the commanded trajectories will not
differ and both setups will reveal the same degree of motion efficiency.
H4 A haptic augmentation of occurring constraints in the workspace will be perceived as useful and
will not increase the operator’s mental load.
3.1 Experimental Procedure
The disposal of a fragmentation mine of the type PROM (see Fig. 4) served as experimental scenario. The
task required the following operations: First, the mine had to be gripped by the non-dominant hand and a
retaining pin had to be picked up by the dominant hand. Next, the retaining pin had to be put on the
detonator so that the mine was covered and the detonator could be unscrewed without any risk. Finally, both
the detonator and the corpus of the mine were to be stored separately. The whole experimental procedure
requires bimanual activity: While the non-dominant hand held and stabilized the gripped object, the dominant
hand performed all sensitive procedures. According to the participants unscrewing the detonator was judged
to be the most demanding operation followed by inserting the retaining pin; correspondingly longer
completion times were recorded for these two subtasks.

Mainly, experienced EOD-experts from the German Armed Forces were recruited as participants. As the
participants were not accustomed to telepresence technology, a sufficient training period was provided. The
experiment was conducted as a within-subject design, so that every of the 20 participants performed the
defusing task with both setups; serial and positioning effects were avoided by a balanced design.
Figure 4: Assembly parts of a fragmentation mine of the type PROM served as experimental
scenario - original (left) and experimental reproduction (right).
3.2 Experimental Results
When asking the participants to decide for one setup two-thirds (67%) of the users chose setup A whereas
one third (33%) preferred to work with setup B. The subjective choice could be backed up by objective
criteria as the participants required less reconfiguration of the workspace when teleoperating with their
preferred system. As supposed, both groups accounted for their decision to be more sensitive for the task at
hand. While according to the first group this was due to the displayed gripping and contact forces, the
second group appreciated the available gripping masters and the way how input movements where
performed (see Fig. 5).
reconfiguration (%)
56
44
ranking for setup A (Median)
gripping forces
contact forces
40
way of movement
presentation of forces
gripping masters
weight of objects
60
setup A setup B
teleoperator correspondance
sensitivity (%)
89
89
11
11
setup A setup B
ranking for setup B (Median)
way of movement
gripping masters
teleoperator correspondence
gripping forces
presentation of forces
contact forces
weight of objects
setup A preferred
setup B preferred
lower medians indicate
higher importance
Figure 5: Participants reconfigured the dominant workspace less often when working with the setup that
they preferred, and that they judged to be more sensitive. While for system A the appearing
forces were regarded to be important, for system B the kind how movements were executed
was preferred.
H1: In order to evaluate the objective value of higher displayed contact and gripping forces, both systems
were compared according to the applied contact forces in the remote environment. Besides, the occurring
error between gripper distance input at the operator side and the measured distance at the end-effectors was
assessed (see Tab. 1). The assumption that contact forces which are displayed one-to-one to the user would
reduce the applied contact forces in the remote environment was supported by a t-test. Furthermore, the
percentage distance error, both for the non-dominant and the dominant hand, turned out to be significantly
lower when working with system A. As a lower distance error indicates that the gripped object was felt more
precisely, it is comprehensible when two-thirds of the participants classified system A to be more sensitive.

mean standard error t-test (sign. level)
A B A B
non-dominant contact forces [N] 6.0436 6.7716 0.1136 0.5367 1.327 (0.196)
dominant contact forces [N] 2.7388 4.2984 0.0974 0.2002 7.004 (0.001)**
non-dominant distance error [%] 34.80 84.36 3.52 3.79 9.581 (0.001)**
dominant distance error [%] 39.44 67.96 3.51 3.67 5.614 (0.001)**
Table 1: Comparison of appearing contact forces and gripping error for system A and B; highly significant
results are marked by asterisks.
H2: Whether the activation of the limbs is natural or not can only be judged in comparison with an identical
direct interaction. For this reason further 20 participants were recruited and asked to perform the defusion
task in a real physical environment. The manual task was executed twice whereby the participants interacted
once similar to setup A and once similar to setup B. In both conditions the participants were instructed to
concentrate on their haptic sensation so that they would be able to assign the appropriate percentage values
to the demanded limbs (see Tab. 2). When the participants were interviewed after the teleoperation they
were also asked to express the perceived activation by percentages. As setup A provides a less natural
stimulation, it is likely to expect that the gathered percentage contribution will differ from the manual
benchmark data set whereas there would be no difference for the percentage contribution of setup B.
Actually, a chi-square test revealed that the participants favouring setup A did not differ from the manual
benchmark data set, neither for setup A nor for setup B (I, III). In contrast to this, the participants preferring
setup B differed in both teleoperation settings from the manual benchmark data set (II, IV). Thus the
participants favouring setup A seemed to be more robust towards the perception of haptic feedback: they
were not aware of the altered presentation of haptic feedback displaying interaction forces in a parallel
manner at wrist and finger and tended to fuse sensed force feedback into realistic perceptions. In contrast,
the participants favouring setup B recognized a different haptic stimulation in setup A. Though their
recordings for setup B also differed from the manual benchmark data set, they demonstrated sensation of
the higher stimulation of the fingertips that has been especially characteristic for this setup. In consequence
the participants who preferred setup B turned out to be particularly sensitive towards the presentation of
force feedback at the fingertip performing precision grasps and for this reason it is comprehensible that they
decided for setup B.
Where did you sense benchmark (%) teleoperation (%)
contact forces in… A preferred B preferred
setup A: fingertip 46.75 46.39 38.89
phalanx 31.50 34.17 46.67
palm 8.50 8.33 4.78
wrist 13.25 11.11 9.66
chi-square-test I) 0.58 (0.90) II) 11.23 (0.01)*
setup B: fingertip 50.56 38.06 53.33
phalanx 30.22 41.67 36.67
palm 6.11 6.10 0.56
wrist 13.11 14.17 9.44
chi-square-test III) 7.51 (0.06) IV) 27.55 (0.01)*
Table 2: Comparison of a manual benchmark data set with the two teleoperation settings according to
the activation of the limbs.
H3: For gaining a robust mental model of the telepresence system it might be favourable when the operator
console corresponds to the kinematical configuration of the telemanipulator. To assess the motion efficiency
both setups were compared in terms of their telemanipulator’s trajectories (see Tab. 3). As assumed, in
average a t-test revealed no significant difference between both setups. Solely, the motor demanding
unscrewing operation profited by a similar kinematical configuration and was performed more efficiently
when working with setup A. In consequence, the hypothesis has to be specified: Whereas a kinematical
correspondence does not seem to be relevant in general, it becomes more important when the task difficulty
increases and when rotational movements are demanded. The participants themselves ranked the
kinematical correspondence to be rather unimportant (see Fig. 5).

manipulator trajectory [m]
Mean
setup A setup B
standard error
setup A setup B
t-test
1) gripping mine 0.0055 0.0229 0.0053 0.0117 1.356 (0.187)
2) gripping safety pin 0.6553 0.5364 0.0457 0.0546 -1.670 (0.103)
3) inserting saftey pin 1.2275 1.4756 0.2134 0.3335 0.627 (0.535)
4) unscrewing detonator 2.0896 5.7882 0.1804 0.5253 6.659 (0.001)**
5) storing mine 0.1333 0.1519 0.0281 0.0294 0.459 (0.649)
6) storing detonator 0.6663 0.7711 0.0621 0.0872 0.979 (0.335)
Table 3: Comparison of setup A and setup B in terms of motion efficiency for the dominant input.
H4: All participants classified the reconfiguration of the workspace to be helpful. Besides, the majority of the
users (81%) appreciated the haptic augmentation of occurring constraints of the teleoperator and solely a
minor part (19%) classified the force display to be an additional cognitive burden. Likewise, two-thirds (67%)
of the participants stated that they made use of the forces provided additionally, although it has to be
mentioned that one third (33%) drew the conclusion that they could not interpret the augmentation
intentionally. Their interpretation difficulty is also mirrored in the recorded data as these participants afforded
slightly more reconfiguration procedures compared to the others. When taking a closer look at the input
behaviour, these users also tended to apply higher contact forces with the telemanipulators at the remote
environment and for this reason seemed to be less sensitive towards haptic feedback (contingency
coefficient: 0.71). The augmented force display might also be less useful when working with setup B as here
higher contact forces were applied, too. In addition, due to the support with a lower amount of haptic
feedback by setup B in comparison to setup A, participants demonstrated less sensitiveness for the
augmented force feedback.
4 Conclusion
Two haptic telepresence systems were compared in terms of their usability for EOD whereby especially
setup A was accepted by expert users. In order to state the reasons for the acceptance more generally, the
following design principles could be derived: (A) Though being favourable it is not essentially necessary that
force feedback activates the limbs in the same way a direct manipulation does. The majority of the users was
not even aware of the different haptic presentation. (B) The telepresence system should be able to display
high contact and gripping forces. Thus, the distance error when gripping objects can be diminished and
lower contact forces will be applied when interacting with the remote environment which in consequence
prevents damages of the telepresence system. (C) A kinematical correspondence of operator console and
telemanipulator must not be overestimated and will yield only benefit when difficult tasks have to be
executed (e.g. rotations). (D) In most cases it is recommendable to augment occurring limitations of the
teleoperator’s workspace not only visually but also haptically. The design principles (A-D) are proposed for
novel designs of haptic telepresence systems employed to remote controlled EOD. Presently, at the TUM
bimanual haptic devices and telemanipulators with full movability in 6 DoF are under development enabling
more advanced teleoperated task execution in future.
5 Acknowledgments
The presented work is funded as part of the ‘SFB 453’ Collaborative Research Center ‘High-Fidelity
Telepresence and Teleaction’ of the DFG (Deutsche Forschungsgemeinschaft).
References
[1] Hirose, S.; Kato, K.: Development of Quadruped Walking Robot with the Mission of Mine Detection and Removal – Proposal
of Shape-Feedback Master-Slave Arm. In Proc. of the IEEE Int. Conf. on Robotics and Automation, pp. 1713-1718, 1998.
[2] Immersion, Corp. Available at: http://www.immersion.com [25.05.2004].
[3] Kron, A.; Schmidt,G.; Petzold, B.; Zäh, M.F.; Hinterseer, P.; Steinbach, E.: Disposal of Explosive Ordnances by Use of a
Bimanual Haptic Telepresence System. In Proc. of the IEEE Int. Conf. on Robotics and Automation, pp. 1968-1973, 2004.
[4] Kron, A.; Schmidt, G.: Bimanual Haptic Telepresence Technology Employed to Demining Operations. To appear in Proc. of
the EuroHaptics conference, June 2004.
[5] Lederman, S.; Klatzky, R.L.: Designing Haptic and Multimodal Interfaces. A Cognitive Scientist’s Perspective. In Proc. of the
Workshop on Advances in Interactive Multimodal Telepresence Systems, pp. 71-80, München, 2001.
[6] SensAble Technologies, Inc. Available at: http://www.sensable.com [25.05.2004].
[7] Telerob Gesellschaft für Fernhantierungstechnik. tEODor – telerob Explosive Ordnance Disposal. Available at
http://www.telerob.de [25.05.2004].

Detection of material and structure of mines by acoustic
analysis of mechanical drilling noise
Abstract:
G. Holl, B. Schwark-Werwach, C. Becker
Wehrwissenschaftliches Instiut für Werk-, Expolsiv- und Betriebstoffe (WIWEB)
Großes Cent
53913 Swisttal-Heimerzheim
Institut für Neue Basistechnologien (INBT)
Alte Leipziger Str. 50
D-99735 Bielen
The Wehrwissenschaftliches Instiut für Werk-, Expolsiv- und Betriebstoff (WIWEB) built in
cooperation with Institut für Neue Basistechnologien (INBT) a computer controlled drilling
and acoustic detection system DIMIDAC (Detection of Mines by Drill Acoustic). This system
can detect known samples different in material and shape during drilling process. The
vibration spectrum of sample in exchange with drilling process is measured, analysed and then
compared with known spectra. The position of the drilling-rode is also recorded and therefore
drilling rate is known.
The analysed characteristic parameters lay typically within limits depending on material and
shape of sample.
1. Introduction:
Because mine clearance operations procedure need much more time than mine laying the
number of polluting mines is still increasing world-wide. By humanitarian causes there is a
great demand for fast detection systems, which reduce the risk for an accident by a mine to
nearly zero. A good solution seems to be using automatic detection systems.
Such systems need one or better several complementary (see [1]) very sensitive sensors and
good signal processing tools to reduce the false alarm rate.
One idea to get information of a possible mine is to use a mechanical drilling device and to
analyse the vibration spectrum, a procedure remembering on handling with prodder.
Detection of vibration spectrum is a common method in process control (see [2],[3] and [4])
and is very sensitive as we know from our sense of hearing.

2. Experimental Device:
The Wehrwissenschaftliches Instiut für Werk-, Expolsiv- und Betriebstoffe (WIWEB) built in
cooperation with Institut für Neue Basistechnologien (INBT) a computer controlled drilling
and acoustic detection system DIMIDAC (Detection of Mines by Drill Acoustic) (Pic.1).
Picture 1: Photo of the experimental device
A high-speed drilling machine can be conveyed pneumatically in a guide rail low and up with
given force if moved against a solid sample. This force is measured by a weighing machine,
upon which the container for samples is placed. The signal of a position sensor fixed on
sliding carriage, so as the signals of two vibration sensors, one for low frequencies and one for
high frequencies, is recorded during measuring process.
A computer with a data acquisition and a control board controls the whole measuring and
detection procedure. Programming and signal processing can be done with aid of the software
DASYLAB, which allows you to program hardware directly within a graphical environment.
Detection procedure starts with working drilling machine at its highest position and with
rotating drilling rode in air. The data acquisition starts and a ground level signal is recorded.
Then movement down starts. If signal of vibration sensors or of weight sensor exceed a certain
level we have contact with a solid body. After 2 or 3 seconds we can analyse a stabile signal
which is characteristic for each type of sample.
If sample is detected or if a time or position limit is exceeded data acquisition stops, a message
with description of detected sample is on display of computer and drilling machine moves up
to its start position.

3. Analyse of Signals
In Picture 2 you can see a typical signal of a vibration sensor. You can recognise that sample
contact happens just before 5 seconds on time scale.
After Fast-Fourier-Transformation of time signals you get frequency spectra. In Picture 3
typical spectra of samples with different material are presented. It is obvious that they differ
significantly from each other.
V
10,0
7,5
5,0
2,5
0,0
-2,5
-5,0
-7,5
-10,0
plot-V10_D15_HF5_U_AXT-FE1.DDF
0
NF-Sensor
5 10
Picture 2: Typical signal of vibration sensor
0,30
0,25
0,20
0,15
0,10
0,05
0,00
0
2500
5000
FE B-WOOD
STONE K-WOOD
PLASTIC BRICK
7500
10000
Hz
s
0,30
0,25
0,20
0,15
0,10
0,05
0,00
Picture 3: Overview of spectra of sample consisting of different material

To get characteristic parameters for detection procedure we calculate an effective value in
special frequency intervals (see Picture 4), which are found out and characterise the samples
The effective value in frequency interval of the upper diagram in picture 4 is nearly equal for
every sample, while effective value in frequency interval of the lower diagram in picture 4
differs from sample to sample.
0,30
0,25
0,20
0,15
0,10
0,05
0,00
0,30
0,25
0,20
0,15
0,10
0,05
0,00
0
0
2500
5000
FE B-WOOD
STONE K-WOOD
PLASTIC BRICK
2500
5000
FE B-WOOD
STONE K-WOOD
PLASTIC BRICK
7500
7500
Hz
Hz
10000
10000
Picture 4: Selection of frequency intervals to get characteristic parameters for detection

Another detection parameter is the drilling rate, which can be calculated from position signal
after contact with sample.
With this procedure we can discriminate samples consisting of different material but also
samples with identical materials but different shapes e.g. a steel plate and a steel-tube.
4. Discussion and outlook
The results are encouraging, nevertheless we could not be sure that object, which sounds like a
mine, is really a mine and we will only recognise mines which spectra are known.
How can this detection system be used?
At first we have to find all objects, which can be possible mines, in order to have contact with
drilling rode. This can be done by using other systems as ground penetrating radar (GPR) or
set a grid of drilling positions, which have to be checked.
The next step is to drill and detect the found material, maybe to detect a special type of mine
by one drilling. But we can get more information if we set several drillings nearby the first
contact with an interesting object. So we win a contour of the object and the distribution of
spectra because they will change along object surface.
As a next step it is possible to drill into the possible mine and take a sample of material inside.
As a last step it is thinkable to destroy the mine or reduce the risk of detonation by taking
special mechanical measures at surface of mine.
Advantage of this detection system is its simple, robust and not too expensive set up with
components of production technology, which are industry standards.
Literature:
[1] M. Acheroy, M. Piette, Y. Baudoin, J.-P. Salmon: Belgian project on humanitarian
demining (HUDEM), Sensor design and signal processing aspects. Royal Military
Academy (RMA), 2000.
http://www.sic.rma.ac.be/~acheroy/worshp_Ispra_2000/workshp_Ispra_2000.html
[2] Strache, Wolfgang: Multisensorielle Überwachung des Stanzprozesses. Düsseldorf :
VDI- Verl., 2000
[3] Lindemann, M.: Erkennung von Verbrennungsaussetzern mit Hilfe von
Klopfsensoren Düsseldorf : VDI Verl., 2001
[4] Brandmeier, Thomas: Verschleißerkennung durch Schwingungsanalyse an der
Drehmaschine. Dissertation Universität Karlsruhe, 1992