Control of multi-legged robot using reinforcement learning with body image and application to a real robot

We address the autonomous control of a 6-legged robot using reinforcement learning. In general, a robot can be made to learn adaptive behavior through reinforcement learning. However, reinforcement learning, presents a serious problem when it is applied to robots with many degrees of freedom, namely the state explosion problem. In our previous works, we proposed reinforcement learning with body image and discussed its effectiveness in applying the method to a multi-legged robot. However, this was restricted to an ideal simulated world, and we did not address its effectiveness for real robots. In this paper, we report upon an actual 6-legged robot we developed, and discuss the effectiveness of applying reinforcement learning with body image to it. We conducted the learning process in a simulated world and then applied the obtained policy to the real robot. The result was that effective locomotion was realized.