Abstract
We describe a distributed approach to controlling autonomous arm robots. The robots need to acquire cooperative behaviors in order to smoothly lift an object. Each arm robot has its own reinforcement learning unit for decision-making. In investigating this task, we are primarily interested in the question of how to design a reinforcement learning control system for a multi-agent system. An applied reinforcement learning algorithm uses Bayesian discrimination method to segment continuous state and action spaces simultaneously, thereby generating of a set of effective rules. The proposed approach is examined empirically with two real arm robots. The basic dynamics of the reinforcement learning process are also analyzed.
