[1Q4-GS-11] Robot and real worlds: Machine learning

[1Q4-GS-11-01] Residual Reinforcement Learning Models Selection Considering Initial State for Item Alignment Task

〇Yusuke Kato^1,2, Tomoaki Nakamura³, Takayuki Nagai^3,4, Natsuki Yamanobe¹, Nagata Kazuyuki¹, Jun Ozawa^1,2 (1. Advanced Industrial Science and Technology, 2. Panasonic Corp., 3. The University of Electro-Communications, 4. Osaka University)

Keywords:AI, Robotics, Reinforcement Learning

In this research, the robot learned skillful behaviors performed by humans to perform the product alignment task in retail stores. Humans can perform more optimal actions by using different strategies for the same task in different initial environments. Therefore, we proposed a system in which the robot can autonomously select a strategy according to the initial state. We created multiple alignment behavior models obtained by simulation-based reinforcement learning and a selector to use them properly. As a result of performing the alignment task using our system, the alignment was more accurate than when only one model was used. In addition, using the model learned on the simulation, we confirmed that the alignment behavior was possible in the real environment.