07.02.2013 Views

Session WedAT1 Pegaso A Wednesday, October 10, 2012 ... - Lirmm

Session WedAT1 Pegaso A Wednesday, October 10, 2012 ... - Lirmm

Session WedAT1 Pegaso A Wednesday, October 10, 2012 ... - Lirmm

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

<strong>Session</strong> <strong>WedAT1</strong>0 Lince <strong>Wednesday</strong>, <strong>October</strong> <strong>10</strong>, <strong>2012</strong>, 08:30–09:30<br />

Skill Learning – Dynamics<br />

Chair Rüdiger Dillmann, KIT Karlsruhe Inst. for Tech.<br />

Co-Chair<br />

08:30–08:45 <strong>WedAT1</strong>0.1<br />

Autonomous Online Learning of Velocity<br />

Kinematics on the iCub: a Comparative Study<br />

Alain Droniou, Serena Ivaldi, Vincent Padois and Olivier Sigaud<br />

Institut des Systèmes Intelligents et de Robotique - CNRS UMR 7222,<br />

Université Pierre et Marie Curie<br />

• Incremental and autonomous online<br />

learning of velocity kinematics, from<br />

scratch and in a limited time<br />

• Visual servoing task with general target<br />

and end-effector (unknown tool)<br />

• Three contexts: reaching the target in one<br />

or two different workspaces; tracking a<br />

target moving unpredictably by a human<br />

• Comparison of three ML algorithms:<br />

LWPR, XCSF and ISSGPR<br />

• Testing generalization capabilities, velocity<br />

in learning and robustness of parameters<br />

• ISSGPR performs better in all studied<br />

criteria<br />

The learning contexts<br />

09:00–09:15 <strong>WedAT1</strong>0.3<br />

Learning Concurrent Motor Skills in Versatile<br />

Solution Spaces<br />

Christian Daniel, Gerhard Neumann and Jan Peters<br />

FG Intelligent Autonomous Systems, TU Darmstadt, Germany<br />

Max Planck Institute for Intelligent Systems, Germany<br />

• Many interesting motor skill tasks<br />

have several distinct solutions.<br />

• Representing multiple solutions<br />

ensures operability of the robot<br />

even if the environment changes<br />

and can in addition lead to faster<br />

learning.<br />

• We present a hierarchical policy<br />

search method which can<br />

simultaneously learn multiple<br />

motor skills to solve complex<br />

tasks.<br />

08:45–09:00 <strong>WedAT1</strong>0.2<br />

Online Learning of Inverse Dynamics via<br />

Gaussian Process Regression<br />

Joseph Sun de la Cruz<br />

National Instruments, USA<br />

Bill Owen Dana Kulić<br />

University of Waterloo, Canada<br />

• On-line learning of the inverse dynamics of a robot manipulator with<br />

Gaussian Process Regression<br />

• Model is trained on a sparse subset of the observed data, with incremental<br />

updates to both the model and the hyper-parameters<br />

• Investigate the impact of full or partial prior information on the convergence<br />

• Comparison to existing approaches shows improved accuracy and reduced<br />

computational requirements<br />

Computation Time Required for a Single Prediction<br />

09:15–09:30 <strong>WedAT1</strong>0.4<br />

Learning Robot Dynamics with<br />

Kinematic Bézier Maps<br />

Stefan Ulbrich, Michael Bechtel,<br />

Tamim Asfour and Rüdiger Dillmann<br />

Humanoids and Intelligence Systems Lab<br />

Institute for Anthropomatics at the Karlsruhe Institute for Technology<br />

• Novel model-based machine<br />

learning algorithm<br />

• Tailored to efficiently learn the<br />

inverse dynamics<br />

• Based on the Kinematic Bézier<br />

Maps algorithm<br />

• Exact encoding of the equations of<br />

motion<br />

• Batch and incremental (online)<br />

learning<br />

<strong>2012</strong> IEEE/RSJ International Conference on Intelligent Robots and Systems<br />

–128–<br />

Plot of the Coriolis and centripetal<br />

forces on a robot joint

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!