FADA lets humanoid robots adapt to new dynamics in 2 minutes

Humanoid robots struggle when the dynamics of their environment change – different terrain, payload, or actuator response can cause the same control input to produce different motions. Existing solutions either rely on zero-shot generalization (which often fails) or require extensive retraining with demonstrations or rewards. A new paper from researchers introduces FADA (Few-Shot Domain Adaptation via Dynamics Alignment), a three-stage framework that solves this with minimal data. FADA first trains an oracle policy using privileged information, then distills it into a deployable Planner-Inverse Dynamics Model (Planner-IDM) via DAgger. At deployment, the planner is frozen and only the IDM is fine-tuned using roughly 2 minutes of target-domain rollouts – no optimal demonstrations or reward functions are needed.

Experiments show FADA significantly outperforms both in-context adaptation (which generalizes poorly) and end-to-end retraining baselines, achieving higher task performance under dynamics shifts. The approach was validated on real humanoid robots, enabling them to perform diverse high-precision whole-body tasks such as walking on uneven ground or carrying varying loads. By aligning the IDM's action generation with the target dynamics through simple supervised learning on paired actions and observations, FADA offers a practical, sample-efficient path to robust humanoid control. Code and hardware rollout videos are available on the project page.