M. Kanoh, S. Kato, and H. Itoh (Japan)
Reinforcement Learning, Actor-Critic, Robotics
In this paper, we propose a learning method for generating a motion of humanoid robots. This method is an actor critic method which has two critics: one evaluates the state of perceptions, and the other evaluates the state of joints. This method can treat high dimensional state space rather than existing actor-critic methods, because state, which is obtained in environment, are independently evaluated by the two critics. In order to verify the effectiveness of our method, we made an experiment for learning the motion, getting up from a chair, of a robot walking with two legs. As a result, our method obtained a controller for the mo tion. Additionally, we compared the performance of our method with an existing actor-critic method which learns by using the state of joints. As a result, our method con verged quicker than the existing method, and the number of successes of the learning with our method was more than the existing method.
Important Links:
Go Back
