Acquiring Adaptive Gaits For Many-Legged Robots by Reinforcement Learning.

Abstract

A legged robot should be able to walk through an unknown environment and avoid unpredictable breakage. To realize this capability, the robot should learn to control its legs without being precisely informed of its internal and external environments. Since the worst condition is the situation where no environmental model is given and where only sensory input that informs the success or failure of walking is available, it is necessary to establish a learning mechanism to developing rules for walking under this situation. This paper proposes a reinforcement-learning-based method to realize this adaptive gait acquisition with minimal information. It was shown by computer simulations that the 6-and the 8-legged wall-climbing robots could successfully establish their gaits without any initial knowledge. It was also demonstrated that the simulated 8-legged robot could establish a new gait after one of its legs was broken.