Presentation Information
[2G5-OS-47b-02]Motion Learning from Human Videos using Image-to-Image Translation
〇Masatoshi Nagano1,2 (1. Kyoto University, 2. The University of Electro-Communications)
Keywords:
Imitation learning,Style Transfer
Robots coexisting with humans are required to possess the capability to flexibly learn and generate motions according to their environments and targets. Since humans can imitate the actions of others after observing them only once, the ability to learn through watching is considered essential for robots to achieve adaptive motion generation. Conventional imitation learning methods, however, face significant challenges regarding data collection costs, often requiring large-scale training datasets and specialized equipment. To overcome these issues, this paper proposes a method that learns the domain translation between humans and robots using easily collectible, task-agnostic data and generates robot motions directly from the translated visual information. Experimental results show that the proposed method enables a robot to imitate motions using only one human demonstration.