Abstract
We propose a novel method for learning dynamic multi-DOF motion with fewer trials and less pre-specified knowledge about the tasks. It is a combination of a motion representation called Virtual Goal Switching Pattern, and a search method consisting of initial random exploration followed by hill-climbing. The Virtual Goal Switching Pattern consists of a controller, some phases and goal-postures for each phase. Both of the intervals of the phases and the goal-posture parameters are searched. This method was applied to 4 learning tasks: a targeted throwing motion with a 3-link arm, a jumping motion, a Roll-and-Rise motion, and a throwing motion with a bendable arm. In these experiments, dynamic multi-DOF motions were acquired quickly (less than 1000 trials). Several interesting related phenomena are also observed.
