Reinforcement Learning can allow robots to learn a policy, which is a rule of choosing optimal behaviors in unknown environments. While it is effective for simple tasks with small action-state space, it cannot deal with mobile manipulators due to large action-state space. In this paper, we propose a method that is able to learn a policy of a mobile manipulator using Deep Reinforcement Learning. This method consists of Q-Learning, which is representative algorithm of Reinforcement Learning, and Deep Learning. To show the effectiveness of our method, we first train the system with a policy by simulation. After the training, the system achieves to press the elevator buttons in real environments.
抄録全体を表示