A Study On Applying A Reinforcement Learning For A Real Robot

Reinforcement learning is a useful method to acquire a purposive behavior with little or no a priori knowledge about the environment. In applying the reinforcement learning to a real robot, there are many di cult problems, such as long learning time and construction of its state and action space. In this paper, we show a soccer robot which learns to shoot a ball into the goal using the Q-learning, one of the reinforcement learning methods. We give the results of computer simulation and real robot experiments that show the e ectiveness of a policy transfer from the computer simulation to the real robot.