Chapter 2. Prehension

Chapter 5 - Movement Before Contact 141
The goal of this neural model is that the Cerebrocerebellum/Pao
Red Nucleus slowly takes over the work of the motor cortex and the
Spinocerebellum/Magno Red Nucleus by being able to accurately pre-
dict the motor command using its inverse-dynamics model. If the in-
verse-dynamics model is an accurate model of the system, then the
feedback torque will be zero, thus removing movement errors.
However, in order to accomplish this, the inverse-dynamics model
must be learned, and taught by the teaching signal, which here is the
total torque, u. When the actual trajectory differs from the desired
trajectory, feedback torque occurs. This becomes an error signal.
Reducing the error by updating its internal model, the inverse-dynam-
ics controller slowly replaces the internal feedback with the feedfor-
ward model. The advantage of this approach is that the model learns
the dynamics and inverse-dynamics of the system instead of a specific
motor command for a specific movement pattern.
Feedback control, whether performed by the motor cortex on the
external feedback, or else by the dynamics model controller using in-
ternal feedback, is limited in time. With the inverse-dynamics model,
feedback torque is slowly replaced with feedforward torque, which
anticipates the torque for the desired trajectory. Thus, the system is
controlled faster. A problem with this approach is the amount of
storage needed for the internal model. In Kawato et al.’s simulation of
a neural net for a three-degree-of-freedom manipulator, 26 terms were
used in the internal model’s non-linear transformations. For a 6 degree
of freedom manipulator, 900 would be needed. For an arm having 11
degrees of freedom and 33 muscles, the number of terms is
astronomical. The authors argued that these can be learned by the
synaptic plasticity of Purkinje cells, which are large neurons in the
cerebellum, Another question relates to generalizing experiences
obtained during learning and whether what’s learned can be used for
quite different movements. Kawato et al. say that their method does
not require an accurate model or parameter estimation.
5.4 Arm and Hand Together, To Grasp
While the behavior of the arm in aiming and pointing tasks is a
first step towards understanding sensorimotor integration and motor
control, reaching to grasp an object brings the additional complexity of
posturing the hand appropriately for impending interactions with the
environment. In grasping, the hand must apply functionally effective
forces, consistent with perceived object properties for a given task.
Creating a stable grasp means taking into account active forces and