Neural network models for the learning control of dynamical systems with application to robotics

Abstract

In this paper, two new brain-like control methodologies have been suggested. The significance of the designs rests in the novelty and simplicity of their architectures. The methodologies have several promising attributes that make them very feasible solutions to current problems in system control. First of all, the controllers mimic the functions of the cerebellum for the learning and control of voluntary movements. That is, these controllers learn the inverse-dynamics of the system by repeated trials of a given task, hence allowing the dynamic systems to track trained trajectories almost perfectly. But above that, with these controllers, systems can perform untrained trajectories quite satisfactorily. The schemes also have good adaptation capabilities which allow the controllers to respond to unexpected disturbances.

Another advantage of these methodologies is that the algorithms do not require the knowledge of the system parameters, and they are robust with respect to parameter variation and disturbances under a variety of tasks. Finally, they have parallel processing capabilities which make them fast and fault tolerant. Moreover, the parallel processing property of these architectures makes them highly suitable for the integration of a multitude of sensory information into the motion controller networks.