An adaptive controller that modifies its characteristics to deal with new situations, in a timely and accurate way, would be a valuable improvement on many existing industrial plant controllers. Furthermore, if such a controller could effectively be "strapped around" the existing controller then there could be many opportunities for performance enhancement of systems already "out in the field". The new situations that this adaptive controller would need to deal with could arise from time variations in the plant's characteristics due to wear and tear, or from learning, on-line, about unforeseen new parts of the plant's operational envelope. The universal approximation abilities of neural networks, combined with permanently active on-line learning, yield powerful features that can be used to great advantage in creating adaptive controllers for such applications. Together they allow accurate control strategies to be developed for these types of plant without the need for a mathematical model. Furthermore, neuro-control algorithms can learn about these dynamical features using signals from the plant that are normally easily obtained. Multi-axis revolute-jointed robot manipulators are good examples of this class of plant. In order to illustrate some of the important advantages of these methodologies, and to inform the reader about some of the important characteristics of these adaptive structures, we review some of our recently reported experiments in applying an on-line learning neuro-control approach to joint level trajectory control of two different industrial robots. In each case, the neuro-controllers are used to enhance performance of the existing PID controllers. This paper is mainly concerned with highlighting these features via the experimental results. However, when a controller learns on-line whilst acting as part of a plant's closed-loop controller, it is crucial that a careful and rigorous approach is adopted. A strict theoretical basis that guarantees the whole system's stability is required. To set the experimental work in context therefore, we briefly review our on-line learning neuro-control method, which is used for both sets of experiments.
展开▼
