Adaptive neural network control of flexible-joint robotic manipulators with friction and disturbance

Abstract

An adaptive control strategy has been developed for flexible-joint robotic manipulators in the presence of friction nonlinearities and external disturbances. As the exact inverse model is unrealizable for such systems, only an approximation can be found. The control strategy consists of a rigid linear in parameter model based feedforward controller that approximates the flexible-joint inverse model and a neural network feedback controller that compensates for parametric and modeling uncertainties such as, friction, flexibility, and disturbance. A reference model is used as a trade off strategy to alleviate joint elasticity effects. Unlike other control strategies, no a priori offline training or weights initialization is required. Results with different situations highlight the performance of the adaptive controller in compensating for structured and unstructured dynamical uncertainties, in particular nonlinear Coulomb friction terms and external disturbance. Internal stability, a potential problem with such a system, is also verified. Furthermore, the adaptive control structure stability is guaranteed by Lyapunov stability theory.