Model reference adaptive neurocontrol of flexible joint robots

A global stable model reference adaptive neurocontrol scheme for flexible joint robots with unknown dynamic models is proposed in this paper. No off-line pre-training of the controllers is needed. Two-layer sigmoidal neural networks are adopted to realize the neurocontrollers, with their structures and weights determined from the function approximation theory viewpoint. Based on the re-formulation of the unknown models of flexible joint robot systems, the derived model reference adaptive neurocontrol law ensures that all the signals in the adaptive neurocontrol system are uniformly bounded, and the end-effector of a controlled flexible joint robot with unknown dynamics can track any given trajectory with user-specified precision. Moreover, arbitrary joint flexibility is allowed, and no acceleration or jerk measurements are needed. The neurocontroller is also robust to the representation errors of the neural networks with finite number of neurons and bounded additive external disturbance