A Neural Network Generating Force Command for Motor Control of a Robotic Arm

Brain can be considered as a system that generates motor commands for a given perception. Understanding how movement is generated and controlled still represents unanswered questions [7]. In [3], an arm controller that learns visuo-motor associations can let emerge immediate and deferred imitation. Visual input activates some attractors in the motor space that are used to generate a speed command to control the arm. In this architecture, visual information gives the motor position to be reached. In [2], a Gaussian Mixture based arm controller can enable a robot to reproduce a sequence of gestures that was demonstrated during a first phase. In [4], a similar Gaussian Mixture model is compared with a neural network based sequence learner that uses fast learned state/action couples [5]. The comparison is done on a simple navigation task for which the neural network is designed. The Gaussian Mixture model was adapted for navigation. A left over issue is whether and how such a neural network architecture could be adapted for arm control in the motor space. These considerations about arm controlling lead to looking for a model that could work with both visuo-motor associations and motor state/action paradigms. Several issues are raised as how the attractors can be encoded, what the parameters of the motor controller are and how they should be controlled by the system. Some tests on how the attractors can be constructed have led to a model that defines independent attractors for each articulation of the robot. Biological data give us hints about answers to some of the raised questions. The basic biological motor action is to contract muscles and thus to perform force control of the position of the limbs. As we aim at studying how high level and low level aspects of motor control can be interconnected, we developed a neural network architecture that generates force commands. In this architecture, the higher level structures converge on an internal representation of the motor position that is provided to a homeostatic low level force controller. Some part of the architecture is tested on a real robot in a simple experiment as the reproduction of a temporal sequence of gestures demonstrated during passive manipulation. The influence of some of the parameters that determines the robot behavior reinforces the idea that higher level structures should have specific accesses to the low level control.