A FREQUENCY-DOMAIN APPROACH TO LEARNING CONTROL - IMPLEMENTATION FOR A ROBOT MANIPULATOR

A new frequency-domain approach to the analysis and design of learning control laws for achieving a desired repetitive behavior in a dynamical systems is presented. The scheme uses two separate filters in order to obtain rapid improvements in a specified bandwidth while cutting off possibly destabilizing dynamic effects that would bar learning convergence. In this way, the trade-off between global convergence conditions and approximate learning of trajectories is made explicit. The synthesis is presented for single-input single-output (SISO) linear systems, but the method is of general application. The proposed learning controller has been used for exact tracking of repetitive trajectories in robot manipulators. In particular, actuator inputs that enable accurate reproduction of robot joint-space trajectories are learned in few iterations without the knowledge of the robot dynamic model. Implementation aspects are discussed and experimental results are reported that show how the proposed learning scheme avoids the occurrence of unstable phenomena over iterations.