Initial experiments on reinforcement learning control of cooperative manipulations

摘要

The paper deals with instance-based reinforcement learning control of autonomous robots. A classifier system, defined in the continuous state and action spaces, is outlined. Based on the sensory state space analysis, we define a learning strategy and fix the structure of the action rules. The classifier system features a nonconservative bucket brigade algorithm and a fast reproduction mechanism. The system developed is then applied to learning cooperative behavior by two robots coupled via a common object, with each robot controlled by its own classifier. The feasibility of this scheme is tested under experiment with two Lynxmotion robots, and a motion pattern of cooperative behavior (lifting up an object) is evolved using the two interacting classifier systems.