Two arXiv papers: humanoid unlearning and faster replanning

ForgetMimic drops selected motions on Unitree G1 and H2, while Rolling-WAM reports a 4.5x replanning speedup on LIBERO, RoboTwin, and a real G1.

Two separate arXiv papers posted a day apart both involve Unitree humanoid robots but address different problems. ForgetMimic, listed under cs.CR on 23 September 2026, is a motion-level unlearning method for reinforcement-learning humanoid control that degrades a chosen subset of K motions while preserving the remaining N-K, after the authors fix two robot-control training mechanisms that otherwise cause unlearning to fail. Experiments cover 12 motions, including dance, fight, and flip, on Unitree G1 and H2; stated motivations include removing poisoned or copyrighted motions and GDPR-style deletion. Rolling-WAM, listed under cs.AI, cs.LG, and cs.CL on 24 September 2026, spreads joint video-action denoising across replanning cycles using a sliding window of chunks at staggered noise levels, fully denoising the next action while only partially refining farther-future chunks as new camera observations arrive. On LIBERO, RoboTwin, and a real Unitree G1, it matches competitive manipulation performance and reports a 4.5x steady-state replanning speedup over standard joint world action models. The sources do not disagree; they describe distinct methods that share only the Unitree G1 platform.

• ForgetMimic (arXiv cs.CR, 2026-09-23) is motion-level unlearning for reinforcement-learning humanoid control: given a policy trained on N motions, it degrades a chosen subset of K motions while retaining the other N-K.
• The ForgetMimic authors fix two robot-control training mechanisms that otherwise cause unlearning to fail.
• ForgetMimic was tested on Unitree G1 and H2 across 12 motions, including dance, fight, and flip; stated motivations include removing poisoned or copyrighted motions, privacy, and GDPR-style deletion.
• Rolling-WAM (arXiv cs.AI/cs.LG/cs.CL, 2026-09-24) spreads joint video-action denoising across replanning cycles instead of restarting the full horizon each time.
• A sliding window holds chunks at staggered noise levels: each step fully denoises the imminent action and only partially refines farther-future chunks as new camera observations arrive.
• Rolling-WAM was evaluated on LIBERO, RoboTwin, and a real Unitree G1, matching competitive manipulation performance with a reported 4.5x steady-state replanning speedup over standard joint world action models.
• The reports describe distinct methods; they overlap only in using a Unitree G1 and do not contradict each other.