Application of a General Learning Algorithm to the Control of Robotic Manipulators

Abstract

In this paper, we discuss the use of a general learning algo rithm for the dynamic control of robot manipulators. Unlike some other learning control schemes, learning is based solely on observations of the input-output relationship of the system being controlled and is independent of control objectives. Information learned previously can be applied to new control objectives as long as similar regions of the system state space are involved. The control scheme requires no a priori knowl edge of the robot dynamics and is easy to apply to a particu lar control problem or to modify to accommodate changes in the physical system. The control scheme is computationally efficient and well suited to fixed-point implementation. The learning controller is evaluated in a series of computer simu lations involving a two-axis-articulated robot arm during simulated repetitive and nonrepetitive movements. We inves tigate the effects of varying learning algorithm parameters as well as control system performance in the presence of obser vation noise and changing manipulator payloads. The learn ing control system presented promises to provide good dy namic performance in complex situations at a reasonable cost as measured in terms of both hardware and software devel opment.