ForgetMimic: Motion Unlearning for Reinforcement Learning Humanoid Control
Xukun Luan, Zhongxiang Lei, Chen Gong, Shaowei Li, Yuanguo Bi, Jinyan Liu
cs.RO · cs.CR · cs.LG
Quick verdict
First motion-level unlearning for humanoid RL via anti-reward fine-tuning; promising selective forgetting, but internal number inconsistencies weaken claims.
Summary
The paper introduces ForgetMimic, presented as the first motion-level unlearning method for reinforcement-learning-based humanoid control. Given a PPO policy trained on 12 motions for Unitree G1 and H2 robots, the method continues fine-tuning with an anti-reward on target motions — a penalty proportional to tracking quality, so better tracking costs more — while the rewards of remaining motions stay unchanged. The authors further show that two standard training mechanisms, adaptive Reference State Initialization (RSI) and Assistive Wrench (AW), actively counteract unlearning (failing target motions get sampled more and physically assisted) and must be disabled. On G1 with Fight and FightAndSports1 as targets, success rates drop to 0% with episode lengths of ~4-6 steps, while non-target tracking stays near 0.83-0.85 (degradation ratio down to 0.628); adapted baselines from Ye et al. and Gong et al. (and random reward) forget the target but collapse non-target skills (success rates 6-23%). A sensitivity study finds a sharp threshold at lambda = W = 8.0. Real-robot evidence on G1 and H2 is qualitative only. Main limitations: internal inconsistencies between Section 5.2.1 text, Tables 1-2 and Appendix A.3 numbers; single-run training without seed variance; and unlearning shown to be behavioral suppression, with forgotten motions recoverable in a few thousand iterations.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- ForgetMimic is the first motion-level unlearning method designed for RL-driven physical humanoid control.
- Anti-reward fine-tuning — penalizing target motions proportionally to tracking quality — selectively degrades target motions while preserving the rest.
- Adaptive Reference State Initialization and Assistive Wrench actively counteract unlearning and must be fully disabled during it.
- On Unitree G1/H2 across 12 motions, target success rates drop to 0% (episode lengths ~4-6 steps) while non-target tracking stays near 0.83-0.85, with the degradation ratio down to 0.628 at lambda=16.
- Forgotten motions are substantially harder to re-acquire: recovery costs about 8x the unlearning optimization (Appendix A.4), offered as indirect evidence of effective forgetting.
Strengths
Novel problem with a clean, testable formalization
First to target motion-level unlearning on physical humanoids; the selective goal is stated precisely (Eq. 2-3) with an explicit preservation bound epsilon. The anti-reward design (Eq. 5-8) predicts tracking is neutralized exactly at lambda = W; W = 8.0 is derivable from the Table 3 weights (six 1.0 terms, two 0.5, joint pos 1.0, joint vel 0), and Table 2 indeed shows a sharp transition at lambda = 8.0.
Mechanism-level diagnosis of why unlearning fails
Section 4.2 identifies adaptive RSI as a positive feedback loop that over-samples now-failing target motions (Eq. 14) and AW as physical assistance that cancels the penalty (Eq. 17). Table 1 isolates each: ARFT+RSI leaves target tracking at 0.7330 vs 0.6653 for the full method (lambda=8); ARFT+AW halves non-target SR to 43%. The Config A/B dynamics in A.3/Fig. 7 (entropy 22.0→31.6 vs 12.9→11.9; AW gain rising 0.007→0.026) substantiate the story.
Strong selectivity against adapted baselines
Honest scope with complementary forgetting evidence
Weaknesses
Internal numerical inconsistencies across text, figures and tables
Real-world evaluation is qualitative only
Suppression, not deletion — mismatch with the GDPR framing
No training-seed variance or statistical testing
H2 reporting gaps and an unacknowledged non-target regression
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ForgetMimic (arXiv 2609.28378) addresses motion-level unlearning for RL-trained humanoid policies: given a PPO policy trained on N motions, degrade a target subset of K motions while preserving the remaining N-K. The paper claims: (1) the first motion-level unlearning method for physical humanoid control; (2) an anti-reward fine-tuning (ARFT) mechanism plus identification of Reference State Initialization (RSI) and Assistive Wrench (AW) as mechanisms that actively counteract unlearning; (3) validation on Unitree G1 and H2 across 12 motions. The motivating framing leans on safety (removing combat motions such as Fight) and on GDPR right-to-be-forgotten — a framing examined under Limitations.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ForgetMimic and how does it remove a motion from a humanoid robot policy?
ForgetMimic is an unlearning method for RL-trained humanoid control policies. It continues PPO fine-tuning while adding a negative 'anti-reward' on the motions to be forgotten, proportional to how well the robot currently tracks them — so good tracking is penalized — while rewards for all other motions stay unchanged. It also disables two standard training aids (adaptive Reference State Initialization and Assistive Wrench) that would otherwise re-teach or physically assist the motions being removed.
Does ForgetMimic truly delete the motion data from the model (e.g., for GDPR compliance)?
No deletion guarantee is provided. The authors state (Appendix A.8) that it is approximate unlearning: it suppresses execution of the target motion rather than provably removing the underlying information. Their own relearning experiment shows the forgotten motion can be recovered to about 95% of baseline performance with a few thousand further training iterations (roughly 8x the unlearning cost), so the GDPR framing should be read as motivation, not a compliance claim.
Did they test ForgetMimic on real robots, and did it work?
Yes, on Unitree G1 and H2 hardware after simulation in MuJoCo/mjlab. Real-world results are presented qualitatively (success/failure visualizations of motions like Fight and Dance): target motions fail as intended while other motions still execute. No quantitative real-robot metrics are given. The authors recommend testing in a controlled environment, and Flip was not deployed on hardware because of damage risk.
How much cheaper is ForgetMimic than retraining from scratch?
Retraining without the target motions took 100,000 iterations (about 40 GPU-hours on an RTX 4090). The main comparison table (Table 1) uses 1,000 fine-tuning iterations — roughly 100x cheaper — but the lambda sensitivity study needed 30,000 iterations to converge, about a third of the retraining cost. So the savings depend strongly on the desired unlearning strength and convergence criteria.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures were extracted only as embedded text/axis labels; Fig. 2 and Fig. 4 numeric values are partially garbled, so figure-vs-table comparisons carry extraction risk (the primary inconsistency evidence is text-vs-table and stands independently).
- Full text including appendices appears present, but real-robot imagery (Figs. 5, 9, 10) is unverifiable from text.
- The code repository (github.com/Zili1000/ForgetMimic) is stated in the author footnote; its availability and contents were not verified.
- Real-robot protocol details (number of hardware trials, metrics) are absent from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.