Performance improvement of robot manipulator control using an on-line neural network compensator

This paper is concerned with the application of neural networks for adaptive compensation of the structured and unstructured uncertainties of the robot manipulator. The controller consists of a model-based term and a neural network on-line adaptive compensation term. It is shown that the neural network adaptive compensation is a universal scheme which is able to cope with totally different classes of system uncertainties. Novel adaptive learning algorithms for tuning the weights of the neural network are proposed. A suitable error filtered signal for training the neural network can be easily obtained from the controller design without using any model knowledge of the robot manipulator itself. The closed-loop system with neural network adaptation on line is guaranteed to be stable in the Lyapunov sense.