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

[2O1-GS-8-03] Imitation of Human Behavior by a Robot Using CycleGAN

〇Masaya Nakano¹, Masatoshi Nagano¹, Tomoaki Nakamura¹ (1. The University of Electro-Communications)

Keywords:Imitation learning, Deep learning, CycleGAN, Convolutional Neural Network, Robotics

There have been studies on imitation learning, where the robot learns behaviors from human movements using motion capture or deep learning. However, the drawback of these methods is that they require the transformation of human motion into robot joint angles or paired data of human and robot images. To overcome these problems, in this paper we propose an imitation learning method using CycleGAN and CNN. CycleGAN learns the visual correspondence between a human and a robot, and CNN learns the correspondence between the image of the robot and the joint angles. Using CycleGAN, the visual correspondence can be learned from unpaired datasets obtained by randomly moving human and robot arms, which significantly reduces the cost of data collection. Experimental results show that a robot can imitate meaningful human behaviors using the proposed method.