TRAINING BACKPROPAGATION AND CMAC NEURAL NETWORKS FOR CONTROL OF A SCARA ROBOT

The dynamic control of a robotic manipulator is accomplished by the computation and application of actuating torques required for the manipulator to follow desired trajectories. A considerable amount of work has been reported in the literature concerning the application of classical model-based and adaptive control techniques to the above problem. However, many of the available schemes suffer from the fact that they require an accurate model of the robot dynamics, including nonlinearities, which may be difficult to obtain beforehand. In order to address the problem of adaptive control in unknown environments, it is possible to utilize artificial neural networks to learn the characteristics of the system rather than having to prespecify an explicit system model. In this paper, two artificial-neural-network-based strategies are implemented for the accurate trajectory tracking by a SCARA-type IBM 7540 robot. The performance of a backpropagation-based neural controller is compared with that of one based on a scheme similar to Albus' Cerebellar Model Articulation Controller (CMAC)1 [Albus J. S. Trans. ASME J. Dynamic Syst. Measur. Control, pp. 220-227 (1975)].