Three Learning Architectures to Improverobot

In this paper, learning control schemes for robot manipulators are tested and compared. The controller consists of a xed-gain feedback controller and an adaptable feedforward controller. The feedforward is calculated with a model of the inverse dynamic of the plant. This model is acquired during repeated motions using the feedback as error signal. Three diierent learning methods are used: simple look-up table, neural net CMAC and parameter identiication of a rigid body model. The learning approach is simple and fast. Neither precise modeling nor parameter estimation is required. The control performances of the presented methods are compared with respect to path-tracking errors and to generalization properties. Proofs of system stability during adaption and convergence of the learning are also given.