ForgetMimic: Motion Unlearning for Reinforcement Learning Humanoid Control
ForgetMimic unlearns selected humanoid motions on Unitree robots while keeping other behaviors.
Researchers propose ForgetMimic, a motion-level unlearning method for reinforcement-learning humanoid control. Given a policy trained on N motions, it degrades a chosen subset of K motions while preserving the remaining motions, and they fix two robot-control training mechanisms that cause unlearning to fail. Experiments on Unitree G1 and H2 cover 12 motions, including dance, fight, and flip. The stated motivations include removing poisoned or copyrighted motions and GDPR-style deletion.
- Motion-level unlearning for physical humanoid control policies
- Drops K target motions while retaining the other N-K
- Tested on Unitree G1 and H2 across 12 motions
- Motivated by poisoned motions, privacy, and GDPR deletion
- Products
- Unitree G1Unitree H2
- Organizations
- Unitree
From the original article
1 min read at arxiv.orgHumanoid control, leveraging human demonstrations, has achieved diverse, agile, and natural locomotion behaviors through reinforcement learning (RL). While this paradigm has yielded remarkable performance in physical humanoid control, how to eliminate specific motions from learned policies remains insufficiently explored.
By Xukun Luan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.