Path Searching of Robot Manipulator Using Reinforcement Learning-Reduction of Searched Configuration Space Using SOM and Multistage Learning-

Abstract

Reinforcement learning is applicable to a robot manipulator required to search for a path adaptable to an unknown environment. Searching for an optimal path in configuration space (C-space), i.e., joint angle space, however, takes much convergence time and memory resources. We propose two ways to overcome this problem. One is restructuring C-space by using Self-Organizing Maps (SOM). Another is doing reinforcement learning at multistage, stage 1 of which searches a path in C-space without considering obstacles, so does stage 2 with considering them near path 1, reducing searched space and convergence time. We propose further reducing searched space by adjusting the path in stage 2 to that in stage 1 through dynamic programming (DP) matching.