Comparison of Adaptive Neural Network Controllers of a Non-Linear Robotic Manipulator

This paper presents several neural network based control strategies for the trajectory control of robot manipulators. The neural networks learn the inverse dynamics of a robotic manipulator without any a priori knowledge of the manipulator inertial parameters or equation of dynamics. Compared are; a delta rule type that does not learn on line, the HSA which is similar but has a small stack of previous input output pairs that are used to train the network on-line, and the CMAC type that also learns on-line. Training strategies and difficulties with on-line training are discussed. Simulation of a two degree of freedom serial link manipulator allows comparison of the effectiveness of the algorithms. Results show various levels of performance.