抄録
Learning dynamic multi-DOF motion has been a challenging task for real robots. With past learning algorithm, the number of required trials has been too large for real robots. And the use of simulation suffered from dichotomy between the simulated world and the real world. In this paper, we report on a series of learning experiments where a real robot acquires dynamic whole body action skills from scratch, solely based on real trials. We use our original learning algorithm named "Exploration of Virtual Goal Switching Pattern" which discovers semi-optimal solutions extremely quickly. In order to cope with deterioration of the search performance due to the real world noise, a new improvement to the search method is devised. A series of experiments with real robot is carried out in which the robot automatically achieves the initial posture and repeats the trials. The results show that dynamic rising actions are automatically acquired from scratch, including a variety of strategies.
