acceptodds
Withdrawn from ICLR 2027

MimicX: Policy-in-the-Loop Supervision Refinement for Physics-Executable Humanoid Skills from Motion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

Human videos provide rich motion targets for humanoid learning, yet visually plausible references can still produce persistent failures under physics-based execution. These failures reveal where training supervision should change. We present MimicX, a policy-in-the-loop framework that uses execution feedback to refine video-driven humanoid motion tracking. Starting from reconstructed and retargeted motion, MimicX localizes difficult transitions and affected body regions, then jointly adapts tracking objectives and the reset curriculum for policy continuation. Repeated rollout verification selects execution-priority improvements subject to tracking guards. Across four core video tasks, MimicX consistently improves tracking accuracy and Robust Execution Horizon relative to the Fixed Reference baseline. Task-averaged results show a 25.7% reduction in body-tracking error and a 255.6% increase in execution horizon. Additional video, motion-reference, and collision-scene studies evaluate the method beyond the core tasks, while MimicX-HLoop accelerates feedback through heterogeneous execution. Overall, MimicX turns policy failure into actionable supervision for deciding what to refine and which refinement to retain.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Motion Tracking, Learning from Human Videos, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SkvAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes MimicX, a framework for learning humanoid motions from videos. It uses execution failures to adjust tracking rewards and training start states, then trains candidate policies and selects updates through repeated evaluation.

Critical strengths
  1. The paper addresses a valuable problem: motion references reconstructed from human videos can be difficult for humanoid robots to execute.
  2. The comparisons show improvements on several tasks.
Critical weaknesses
  1. Poor writing. The method and experimental comparisons are not explained clearly enough to assess the contribution. The terminology, such as “practice,” “consolidation,” “continuation,” and “protect”, and mathematical notation are unconventional in the field and not clearly explained. The method description is fragmented and confusing, making it difficult to understand how the components work together and what the core mechanism is.
  2. Limited methodological novelty. The method combines heuristic reward adjustments, failure-focused sampling, standard PPO training, and rule-based selection among three candidates. The contribution appears incremental, and the paper does not clearly establish what new algorithmic insight this combination provides.
  3. The core comparisons do not separate the benefit of the proposed method from that of additional training. In the core four-task study, MimicX trains three candidates for 250 PPO updates each, while each baseline receives only 250 updates. It is therefore unclear how much of the reported improvement comes from the proposed refinement mechanism rather than the larger training budget.
  4. No real-robot deployment. All experiments are conducted in simulation, leaving it unclear whether the improvements transfer to the real world.
Questions for authors

See weaknesses.

Reviewer_AdDUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MimicX, a framework to improve humanoid-motion tracking performance on data reconstructed from video. It is built upon BeyondMimic and uses policy rollouts to find challenging motion sgments. It then adjusts two things, tracking rewards and reset curricula. It will launch multiple training candidates and select among them. The proposed method is evaluated on Unitree G1 in simulation testing on a few sports videos.

Critical strengths

This paper tries to automating motion tracking refinement through failure diagnosis, adjusting the reward and curriculum, which could reduce manul tunning efforts.

Critical weaknesses

Limited novelty. The method mainly combines existing techniques with manually designed heuristic rules.

Poor scalability. Each motion clip requires a separately trained policy, with no evidence of generalization.

Unfair comparisons. MimicX uses 3× the training budget, and some tasks use different references compared to baselines. And it seems simple curriculum learning achieves much of the improvement, while the full method offers inconsistent benefits.

No hardware experiments.

Questions for authors

Why does the paper repeatedly show G1 rendered in original video scenes instead of actual simulation results? I don't see rendering G1 reference back to the video is meaningful. The unusual anonymous-author/link formatting and supplementary materials also suggest heavy AI-assisted preparation; can the authors clarify the AI usage and compliance with ICLR submission guidelines?

How can this approach scale to thousands of motion clips if each requires separate training?

Why are Dance and Kung Fu counted as improvements when all refinement proposals were rejected?

Reviewer_oT2MOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an interesting work on video-driven humanoid motion tracking. It localizes difficult transitions and affected body regions, then jointly adapts tracking objectives and the reset curriculum for policy continuation, thereby enabling the relatively precise extraction of robot motion from human videos.

Critical strengths

My rating of this paper is primarily given to the problem itself: motion retargeting from human video is quite important, and the authors did provide a method to lower the tracking error (according to Tab. 2, 0.055m Local body error, which is quite acceptable). The proposed HLoop algorithm should be very useful and valuable to the community.

Critical weaknesses

Despite the advantages mentioned above, I still found some weaknesses and have some questions regarding the experimental settings, which I will combine here.

  1. I think the comparison might not be fair enough. For instance, for Dance and Kung Fu, MimicX uses repaired references, but Tennis and Football share the original references. I feel quite confused, cause in this way, the comparison on Dance and Kung Fu actually involves two variables: one is the reference itself, the other is the policy learning method. Could the authors refine or justify their comparison?
  2. For the comparison in Fig 6, I just wonder how the authors directly compare the MjLab reproduction of the listed methods. For example, beyondmimic is a single-policy tracking method, while Sonic, although a general-purpose motion tracker, has reportedly performed poorly in handling high-dynamic motion. Therefore, I think Appendix F should also list the training or fine-tuning methods for these methods.
  3. A minor weakness is that this paper does not currently provide the deployment results. It should be fine since the authors seem to have already provided the sim2sim results in MuJoCo, but a real-world deployment would still increase the article's impact.
Questions for authors

see weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.