MODELLING OF LEARNING PROCESS FOR THE MANIPULATOR MOTION BASED ON THE ADAPTIVE PRODUCTION SYSTEM
Hirohisa Hirukawa, Shinzo Kitamura · Biomechanisms · 1986
Recent results of cognitive science suggest that learning is understood from three viewpoints: knowledge acquisition, representation and utilization. This paper studies, based on such an understanding, the learning of a robot manipulator with seven degrees of freedom for avoiding obstacles in the working space. The kinematics of the manipulator was represented by Denavit-Hartenberg's notation. For the knowledge acquisition to control, the minimization strategy of a potential function was employed. The potential function was defined as a sum of a quadratic form of state error vector and the penalty function; the former stood for the problem without an obstacle and the latter for a barrier yielded by existing obstacles which are mathematically defined as primitives like cuboids, cylinders and so on. In our studies, the learning process for manipulator motion is considered as a sequential change of knowledge structure. For implementation of such a scheme, an adaptive production system was used. The knowledge was represented in the form of production rules and could be modified by adaptation as a generalization or specialization. These rules were stored in a discrimination net for fast computer retrieval. The manipulator was implemented as a geometric simulator on a micro-computer and the production system by Lisp 1.9 on a large-scale computer. Simulation results were presented with computer graphics.