Direct-Adaptive Neurocontrol of Robots with Unknown Nonlinearities and Velocity Feedback

Abstract

A neural network (NN) adaptive tracking controller for rigid revolute robots is presented that requires position measurements only. The controller is synthesized using a computed torque like control part of which a modified version of the nonlinear part of Lagrangian dynamics is learnt online by a neural estimator that needs no offline training phase. Therefore, the implementation of the control algorithm needs a reasonable knowledge of the inertia matrix alone. The combined neurocontroller-linear observer scheme can guarantee the uniform ultimate bounds (UUB) of the tracking errors and the observer errors under fairly general conditions on the controller-observer gains.