Experimental evaluation of a new robot learning controller

Abstract

The result of the implementation of a learning control algorithm on an IBM 7545 robotic manipulator are presented. The implementation of the joint space algorithm on the IBM manipulator confirms and exceeds the performance estimated by previous simulation work. Feedback was obtained from optical joint position encoders, and velocity was estimated by simple numerical differentiation in software. No acceleration feedback of any kind was used, and no dynamic parameters, dynamic equations of motion, or kinematic equations were needed. The performance of the algorithm was compared to that of a simple PD feedback system (common in many current industrial robots) and an adaptive algorithm which performs at least as well as a computed torque controller. The learning algorithm outperformed both of these controllers by a large margin, and exhibited convergence within approximately three cycles.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>