[2O1-GS-8] Robot and real worlds

[2O1-GS-8-02] Reinforcement learning for robust control in dynamic environment

〇Kentaro Sakai¹, Sachiyo Arai¹ (1. Chiba University)

Keywords:Reinforcement Learning, Robustness, Sim-to-Real, Domain Randomization

Research using reinforcement learning to control robots and automobiles has been conducted in recent years as the performance of reinforcement learning has improved with the introduction of deep learning. This research uses the simulator for learning instead of a real-world system due to time, cost, and safety restrictions. However, due to the gap caused by the inability of the simulator to reproduce the real world perfectly, it is difficult to make the models learned on the simulator work in the real world. Research to address this gap between the simulator and the real world can be divided into two categories. One is to minimize the gap between simulation and reality by improving the simulator. The other is to learn robust policy against the gap in advance when learning with the simulator. In this study, we examined the robustness of the second approach, domain randomization, to perform reinforcement learning tasks in multiple environments with different parameters. We also compared the performance of the policy obtained by domain randomization with that of model predictive control, a method that is already in operation in the real world against changes in the environment.