Learning-based automatic generation of collision avoidance algorithms for multiple autonomous mobile robots

Abstract

Using a model based on the omni-directional robots developed at the Institute of Physical and Chemical Research (RIKEN), we discuss the possibility of automatically generating a collision avoidance algorithm for autonomous mobile robots. To this end, we show that an effective collision avoidance algorithm for two robots can be generated by a very simple learning algorithm that simulates a naive human trial-and-error learning process, using only the robots' sensor outputs and a suitable reward function, where the exact form of the reward function is also learned autonomously by the robots. We also discuss how a robot can use its "experience" gained in a simple environment to adjust itself to a more complex environment, by automatically generating a collision avoidance algorithm for a three-robot situation utilizing a reduced state space resulting from the learning process for the case of two robots. The results of computer simulation and the experiments conducted at RIKEN using physical robots demonstrate the effectiveness of the collision avoidance algorithms generated and our learning-based approach.