Reinforcement Learning of Giant Swing Motions for a Life‐Sized Dual‐Arm Robot
The combination of robotics and reinforcement learning (RL) has enabled a wide range of dynamic motions. However, demonstrations of arm‐centered aerial movements that humans can perform remain extremely rare. This is primarily because aerial motions are inherently unstable compared to ground movements, making simulation, reward design, and Sim‐to‐Real transfer particularly challenging. Previous studies have been limited to experiments on small‐scale robots or in simulation, and the aerial motion capabilities of robots remain far inferior to those of humans. In this study, we propose a method to enable a life‐sized dual‐arm humanoid robot to achieve stable giant swing motions on a horizontal bar using RL. The robot features a durable metal structure and hands designed to securely grasp the bar. On the control side, we divide the target motion into distinct phases with different dynamic characteristics, each governed by a separate policy. The policy training was conducted in two stages: the first stage for motion acquisition and the second stage for motion refinement using a mechanical energy‐based reward function. In experiments, the robot successfully achieved more than ten consecutive and stable giant swing motions on the real hardware. These results demonstrate the necessity and effectiveness of the proposed method and are expected to contribute to advancing the development of aerial dynamic motion control and to expanding the behavioral space of robots.
Authors
- Kento Kawaharazuka (ORCID: https://orcid.org/0000-0002-7464-7187)
- Kei Okada (ORCID: https://orcid.org/0000-0001-6606-6692)
- K. Yoneda (ORCID: https://orcid.org/0009-0005-2538-8763)
- Ayumu Iwata (ORCID: https://orcid.org/0009-0000-2915-1652)
Institutions
- The University of Tokyo (JP)
Publication Details
- Journal
- Advanced Robotics Research
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1002/adrr.202600008
- Primary Topic
- Reinforcement Learning in Robotics
- Type
- article
- Field-Weighted Citation Impact
- 0.00