[5Yin-A] Poster 8

[5Yin-A-29] Cross-Robot Knowledge Transfer for Quadrupedal Control via In-Context Reinforcement Learning

〇Naoaki Ishii¹, Hiroshi Kera^1,2, Kazuhiko Kawamoto¹ (1. Chiba University, 2. National Institute of Informatics)

Keywords:

Reinforcement Learning,Transformer,Legged Robot

Deep reinforcement learning–based robotic control policies often depend on robot-specific dynamics, requiring retraining for each new platform. Prior work reduces this dependency by increasing environmental and dynamic diversity during training, but at the cost of substantially higher training cost and computation. This work addresses control policy adaptation to unseen robots with minimal additional training. We apply algorithm distillation, a form of in-context reinforcement learning, to learn shared locomotion representations from expert trajectories of multiple quadrupedal robots. Adaptation is then performed by fine-tuning only the decoder of the distilled model, while keeping the shared representation fixed. Simulation results show stable locomotion on previously unseen quadrupedal robots, indicating rapid cross-platform policy transfer.