Robust Collision Avoidance in Multi-Robot Systems - Implementation onto Real Robots -

Abstract

In this paper, we discuss robust collision avoidance in multi-robot systems. It is important for a robot to acquire adaptive behaviors for avoiding robots and obstacles in complicated environments. As we reported previously, the reinforcement learning is useful for such kind of purposes. It was, however, found that it is difficult to implement the learning method onto real robots because the method requires large size of memory storage. In this paper, the multi-layered learning is introduced to reduce the required memory size. By dividing a learning curriculum into multiple layers, the number of expected situations can be limited and the learning process itself can be structured. It is shown that real robot is able to successfully avoid collision to other robots and obstacles in a complicated situation, based on the proposed learning procedure.