A neural network adaptive controller for robots with unknown dynamics

In this paper, a neural network adaptive controller for robot manipulators with unknown dynamics is proposed which consists of one Adaline network to identify structured system dynamics and another one to compensate for both structured and unstructured dynamic uncertainties. The former is trained off-line using a LMS type algorithm while the latter uses an on-line stable weight updating mechanism determined using Lyapunov theory. Since Adaline nets match robot regressor dynamics perfectly, the training processes of the resulting simple neural networks are computationally efficient and the proposed adaptive controller has very high potential in real-time applications. The proposed control scheme is finally illustrated through simulation and comparison studies.