Practical Trajectory Learning Algorithms for Robot Manipulators

Abstract

Several alternative learning control algorithms are discussed, both from an inverse dynamics and an optimization point of view. The learning laws are derived in discrete time and do not need acceleration measurements. A simple algorithm using a constant learning operator is proposed to run in addition to a simple (PD) feedback controller. Its performance is comparable to other algorithms, and it works under non-ideal conditions where the others fail. Two simulation examples on (1) learning dynamic control, and (2) learning optimal redundancy resolution, are presented.