A neural network based control strategy for flexible-joint manipulators

Abstract
A scheme is proposed for robust control of manipulators with flexible joints, using neural networks. A multilayer backpropagation neural network is designed and trained to compute the inverse dynamics of a flexible-joint manipulator. This network is implemented in the feedforward path. The main advantage of this scheme is that it does not require any knowledge about the system dynamics and nonlinear characteristics, and therefore it treats the manipulator as a black box. It is shown that the manipulator must be observable to ensure convergence of the neural net training procedure, and some suggestions for selecting manipulator outputs so as to make it observable are proposed. Simulation results for a single-link flexible-joint manipulator exemplify the performance of the resulting open- and closed-loop control systems.

This publication has 14 references indexed in Scilit: