ForgetMimic: Motion Unlearning for Reinforcement Learning Humanoid Control
The paper targets a specific gap: making a humanoid policy forget selected learned motions without breaking the rest.
ForgetMimic is framed as motion-level unlearning for RL-based humanoid control. Given a policy trained on many demonstrations, it degrades performance on a chosen subset while preserving the remaining motions. The authors cite safety, poisoned or weak demonstrations, privacy, and copyright-related removal as motivations. They report tests on Unitree G1 and H2 robots across 12 motions, including Dance, Fight, and Flip. ArXiv · AI/CL/LG's note
ForgetMimic is framed as motion-level unlearning for RL-based humanoid control. Given a policy trained on many demonstrations, it degrades performance on a chosen subset while preserving the remaining motions. The authors cite safety, poisoned or weak demonstrations, privacy, and copyright-related removal as motivations. They report tests on Unitree G1 and H2 robots across 12 motions, including Dance, Fight, and Flip. ArXiv · AI/CL/LG's note
score 4