[3L4-GS-8] Robot and real worlds

[3L4-GS-8-03] Imitation Learning with Mid-Level Representations for Object Rearrangement

〇Makoto Sato^1,2, Ryosuke Unno^3,2, Hiroki Furuta^3,2, Tatsuya Matsushima^3,2, Ryo Okada^3,2, Pavel Savkin⁴, Genki Sano⁴, Yutaka Matsuo^3,2 (1. Saitama University, 2. Matsuo Institute, 3. The University of Tokyo, 4. Telexistence Inc.)

Keywords:Manipulation, Imitation Learning

Recently, there has been a lot of research on the use of imitation learning to enable robots to perform tasks performed by humans so far. End-to-end imitation learning that uses raw-color image has been attracting attention due to the improvement of image processing capability by deep learning. However, imitation learning using raw-color images as input has low sample efficiency and requires a large amount of expert data. In addition, when the background and the brightness of the environment are different between the environment where the expert data is collected and the environment where the learned policy is used, the policy learned using the expert data may not behave appropriately. In this study, we verified that object manipulation can be performed by imitation learning combined with a depth map, which is a mid-level representation with high generalization for background and brightness, and that learning can be performed with high sample efficiency.