Control of multi-legged robot using reinforcement learning with body image-realization of generalization capability by utilizing properties of the body-.

In this paper, we describe the development of a 6-legged robot that is designed to move on rubble. We investigate the effectiveness of applying reinforcement learning that utilizes body image to a real robot. We conduct a robot learning process in the simulated simple world and apply the policy thus acquired in the real complex world. The results show that the robot body abstracts the necessary information and applies it in the real environment, and thus effective locomotion is realized. [ABSTRACT FROM PUBLISHER]