Exploring the possibility of mechanical energy as a multipurpose evaluation function for learning of whole body dynamic motions

To optimize dynamic motions of a humanoid is difficult and any analytical method to optimize these motions has not been established. Motion learning is an effective way to improve motions which are difficult to be optimized analytically and there are many studies on it. However, in the past studies, evaluation functions of motion were dependent on motions and made by human for each motion. To solve the problem, we propose to use mechanical energy as an evaluation function in learning of whole body dynamic motions of a humanoid. In this paper, we tried to optimize forward rolling motion and smashing roof tiles motion as examples of whole body dynamic motions using mechanical energy as an evaluation function. By verifying effectiveness of the learning result, we indicates the possibility that mechanical energy could become a multipurpose evaluation function in learning of whole body dynamic motions of a humanoid.