Filter-error-learning neural networks for stable trajectory tracking control of robot manipulators

Abstract

A learning control architecture using multilayered neural networks is presented and evaluated for trajectory tracking control of robot manipulators. This architecture employs a stability-guaranteeing feedback controller (SGFC) which ensures that the tracking error is ultimately bounded. With the help of the SGFC, neural networks can learn to control the robot without fear of instability. Neural networks adjust their weights using the proposed filter-error-learning (FEL) which does not require either desired neural network outputs as teaching signals or error back propagation through the plant. In FEL, teaching signals are extracted from the SGFC, and the learning rule is derived in order to minimize the tracking error. As a result, neural networks learn the inverse dynamics model of the robot. The performance of the proposed learning control architecture is illustrated through simulations with a two-link robot manipulator.