[4Xin2] Poster session 2

[4Xin2-19] Robustness Verification in Robot Control Using Decision Transformer

〇XIYING DONG¹, Tsuyoshi Takano¹, Hiroshi Kera², Kzuhiko Kawamoto² (1.Graduate School of Science and Engineering Chiba University, 2.Faculty of Engineering Chiba University)

Keywords:Reinforcement Learning, transformer, robustness, MUJOCO

To verify the robustness of Decision Transformer in robot control, Decision Transformer is an off-line reinforcement learning model using Transformer, which has been reported to perform as well as or better than conventional reinforcement learning. In this study, we evaluate the robustness of Decision Transformer using three robots (Half Cheetah, Hopper, and Walker2D) for failure cases that are not included in the collected data for offline reinforcement learning. The robustness evaluation simulates a situation where the actuator does not work for the three datasets (Medium-Expert, Medium, and Medium-Replay). Experimental results confirmed a trend toward lower rewards for all robots for all three datasets. This result indicates that the Decision Transformer needs to be made more robust.