Competitive physical interaction by reinforcement learning agents using predictive models
Hiroki Noda, Satoshi Nishikawa, Ryuma Niiyama, Yasuo Kuniyoshi · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2020
Though there are a lot of researches about Physical Human-robot Interaction (pHRI) using prediction, few researches work on inducing the opponent’s action or outwitting the opponent. We made the push-hand game environment in order to focus on generating strategic actions and tried to make reinforcement learning agents to learn these actions by adding rewards which are directly proportional to the degree of inducement (induction reward) or the degree of outwitting (outwitting reward), defined in this research. As a result, we demonstrated that the induction reward decreases the agent’s predictive error and the outwitting reward increases the opponent’s predictive error, and both of them didn’t contribute to the winning percentage.