Consideration on robotic giant-swing motion generated by reinforcement learning

  • Authors:

  • M. Hara;N. Kawabe;N. Sakai;J. Huang;Hannes Bleuler;T. Yabuta

  • Affiliations:

  • Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland;Graduate School of Information Science and Technology, The University of Tokyo, Tokyo, Japan;Dept. of Mechanical Engineering, Yokohama National University, Yokohama, Japan;Dept. of Intelligent Mechanical Engineering, School of Engineering, Kinki University, Higashi-Hiroshima, Japan;Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland;Dept. of Mechanical Engineering, Yokohama National University, Yokohama, Japan

  • Venue:
  • IROS'09 Proceedings of the 2009 IEEE/RSJ international conference on Intelligent robots and systems
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

This study attempts to make a compact humanoid robot acquire a giant-swing motion without any robotic models by using reinforcement learning; only the interaction with environment is available. Generally, it is widely said that this type of learning method is not appropriated to obtain dynamic motions because Markov property is not necessarily guaranteed during the dynamic task. However, in this study, we try to avoid this problem by embedding the dynamic information in the robotic state space; the applicability of the proposed method is considered using both the real robot and dynamic simulator. This paper, in particular, discusses how the robot with 5-DOF, in which the Q-Learning algorithm is implemented, acquires a giant-swing motion. Further, we describe the reward effects on the Q-Learning. Finally, this paper demonstrates that the application of the Q-Learning enable the robot to perform a very attractive giant-swing motion.