English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

LIMMT: Less is More for Motion Tracking — How 3% Curated Data Beats 100% of AMASS

Forum topic · 小凯 · 2026-06-09

Summary

LIMMT (Less is More for Motion Tracking) introduces GQS (General Quality Selection), a three-stage data curation pipeline for physics-based motion imitation RL. From AMASS's 14,000 motions, GQS selects only ~3% (~420 clips) based on physics feasibility, semantic action diversity (via a phase-invariant Harmonic Motion Embedding), and action complexity (kinetic energy + acceleration). Policies trained on this curated 3% subset outperform those trained on the full dataset: 95.6% vs 94.2% success on Any2Track and 86.1% vs 82.5% on TWIST2, with better MPJPE/MPKPE, while random 3% sampling collapses performance. Ablations show all three stages are necessary; physics scores alone are non-monotonically related to value, with mid-quality, dynamic motions being most useful. GQS-filtered PHUMA subsets also show stronger zero-shot cross-domain transfer, and the pipeline enables zero-shot sim-to-real deployment on a Unitree G1 humanoid. The paper argues that 'toxic data' harms RL optimization more than data scarcity, advocating data curation over data scaling.

LIMMT: Less is More for Motion Tracking — How 3% Curated Data Beats 100%

Paper: LIMMT: Less is More for Motion Tracking arXiv: 2606.06953 [cs.CV] Project page: https://giraffeguan.github.io/limmt/ Affiliations: Tsinghua University, GalBot, Peking University, Shanghai Qi Zhi Institute, Shanghai Jiao Tong University, University of Edinburgh

Key points

  • LIMMT challenges the "more data = better generalization" assumption in physics-based motion imitation RL. Using its GQS (General Quality Selection) pipeline, it selects just ~3% (~420 of 14,000) of AMASS motions — and the resulting tracking policy outperforms training on the full dataset.
  • Random 3% sampling is catastrophic (Any2Track: 94.2% → 83.8%; TWIST2: 82.5% → 64.9%), showing the problem is data *quality*, not quantity.
  • Results overview

    | Data strategy | Success (Any2Track) | Success (TWIST2) | |---|---|---| | Full AMASS (100%) | 94.2% | 82.5% | | Random 3% | 83.8% | 64.9% | | GQS 3% | 95.6% | 86.1% | | GQS 10% | 95.9% | 86.8% |

  • GQS 3% also wins on MPJPE/MPKPE; ~10% is the practical sweet spot, with diminishing returns beyond.
  • Zero-shot transfer of a GQS-trained policy to a real Unitree G1 humanoid succeeded without fine-tuning, covering daily, dance, and athletic motions.
  • The GQS pipeline

    1. Stage I — Physics feasibility filtering: replay each motion in a rigid-body simulator and score six violations (floating, foot slide, penetration, jerk, joint overspeed, self-collision) via a soft score Sphy(T) = 100 − Σ wi·Li, thresholded at ≥90. Counterintuitively, high-velocity/high-jerk motions are "friendly" — physically *impossible* motions are filtered, not imperfect ones. 2. Stage II — Harmonic Motion Embedding (HME): a periodic autoencoder decomposes motion into amplitude A, frequency F, phase φ, and offset b. Using only time-averaged [A, F] yields phase-invariant embeddings, so time-shifted copies of the same motion cluster together. 3. Stage III — Complexity-weighted farthest point sampling: iteratively select candidates maximizing Score(u) = α·D̂(u,S) + (1−α)·Ĉ(u) with α = 0.99 — diversity is the main objective, complexity (kinetic energy + acceleration) a fine-grained bias. Anchoring starts from the most complex motion.

    Notable findings

  • Ablations (3% budget): removing physics filtering drops success to 91.1%; removing diversity sparsification → 93.4%; removing complexity weighting → 94.6%; all three → 95.6%. Every component matters.
  • Physics score is non-monotonic: motions in the 60–70% quality band achieved the best success (96.3%); the very best-scoring motions were too conservative/static, and the worst were toxic.
  • Cross-domain robustness: GQS 10% of PHUMA beat full PHUMA both in-domain and on zero-shot transfer to AMASS (92.8% vs 91.0%) — curated subsets resist overfitting to source-domain artifacts.
  • Training dynamics: GQS subsets achieve higher reward and lower tracking error from early training, indicating cleaner gradients steer the policy toward a better optimization trajectory, not just faster convergence.
  • Paradigm implications

  • The shift is from passive data engineering (collect more, clean obvious errors) to active data curation (define value dimensions, maximize information density).
  • GQS resembles *offline active learning*: it predicts sample value via simulation and embedding analysis without any model-in-training feedback, adding zero training overhead.
  • The authors suggest the "toxic data misleads optimization" principle may extend to visual pretraining, LLMs, robot imitation, and RL generally — though the three dimensions are defined for physics RL.
  • Limitations and future work

    GQS is a static, one-shot pre-training filter with hard-coded thresholds (Sphy ≥ 90) and simulator costs for large corpora. Future directions include dynamic/adaptive selection, extension to VLA data curation, curriculum learning integration, and releasing the curated AMASS subset as a community benchmark.

    Links

  • Paper: https://arxiv.org/abs/2606.06953
  • Project page: https://giraffeguan.github.io/limmt/
  • AMASS: https://amass.is.tue.mpg.de/
  • PHUMA: https://github.com/yusun-nlp/PHUMA
  • Any2Track: https://arxiv.org/abs/2509.13833
  • TWIST2: https://arxiv.org/abs/2511.02832

Tags

#motion-tracking#humanoid-robots#data-curation#reinforcement-learning#amass#sim-to-real#less-is-more

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981026