Collision avoidance in multi-robot systems based on multi-layered reinforcement learning

It is important for a robot to acquire adaptive behaviors for avoiding surrounding robots and obstacles in complicated environments. Although the introduction of a learning scheme is expected to be one of the solutions for this purpose, a large size of memory and a large calculation cost are required to handle useful information such as motions of robots. In this paper, we introduce the multi-layered reinforcement learning method. By dividing a learning curriculum into multiple layers, the number of expected situations can be reduced. It is shown that real robots can adaptively avoid collision with each other and to obstacles in a complicated situation. (C) 1999 Elsevier Science B.V. All right reserved.