PaperlayerView exact revision on arXiv

Source-linked AI summary

LIMMT: Less is More for Motion Tracking

Yu Guan, Zekun Qi, Chenghuai Lin, Xuchuan Chen, Dairu Liu, Wenyao Zhang, Jilong Wang, Xinqiang Yu, He Wang, Li Yi

arXiv:2606.06953v1cs.RO

TL;DR

Physics-based humanoid motion tracking lacks a clear data-centric account of what makes motion data valuable, especially amid artifacts, redundancy, and costly dataset scaling. LIMMT addresses this gap with GQS, which evaluates feasibility, diversity, and complexity to curate compact training subsets. Training on 3% of curated data outperforms full-corpus baselines, while the authors note a scope boundary around the assumption that physically demanding skills produce more robust trackers.

  • Problem

    Large motion corpora can contain physics-violating artifacts and redundancy, while the field lacks a data-centric account of how motion quality shapes robust humanoid tracking.

  • Method

    LIMMT uses GQS to filter infeasible motions, embed feasible motions for behavioral diversity, and select a compact subset with complexity-weighted sampling.

  • Results

    3% of curated AMASS data outperforms full-corpus baselines across evaluated tracking metrics, with gains transferring across trackers and datasets.

  • Takeaways & Limitations

    Motion data value is tied to physical feasibility, behavioral diversity, and dynamic richness rather than volume alone.

  • Takeaways & Limitations

    The framework assumes that datasets biased toward physically demanding skills should yield more robust trackers than datasets dominated by static or simple motions.

Abstract

from arXiv · show

We argue that high-quality motion data can steer tracking policies toward better optimization trajectories early in training. In this work, we introduce LIMMT (Less Is More for Motion Tracking). To our knowledge, this is the first data-centric study for physics-based humanoid motion tracking. We go beyond simply removing low-quality and erroneous clips, but define motion data quality through three dimensions: physics feasibility, diversity, and complexity. We show that even training with under 3% of AMASS yields better tracking performance than training with the full dataset. We further conduct data cleaning on the estimated web-sourced mocap data. Extensive experiments and analyses validate the effectiveness of our framework.

1 Introduction

LIMMT challenges indiscriminate motion-data scaling by defining quality through physics feasibility, behavioral diversity, and action complexity. Its GQS pipeline filters, embeds, and selectively samples motion data, with curated subsets achieving strong tracking results using far less data.

  • Motivation: Physics-based tracking datasets can contain artifacts and redundancy that corrupt imitation signals, increase training costs, and undermine naive scaling.Reported artifacts include temporal jitter, foot sliding, ground penetration, and implausible contacts.
  • Quality dimensions: Motion quality comprises physics feasibility, action diversity, and action complexity, rather than merely the absence of broken clips.These dimensions respectively concern realizability, behavioral coverage, and informative dynamic supervision.
  • Method: GQS hierarchically filters infeasible motions, learns a semantic embedding for diversity, and selects a compact subset with complexity-weighted sampling.The ordering prevents broken motions from dominating diversity selection and prevents high-energy artifacts from being over-selected.
  • Results: 3% of AMASS can outperform the full dataset across evaluated metrics while using under one hour of data.The reported gains extend across multiple tracking systems, including Any2Track and TWIST-2.
  • Implication: The paper presents quality-focused curation as a cost-effective alternative to scaling motion corpora by removing harmful and redundant motions.Its contribution reframes dataset quality around feasibility, diversity, and complexity rather than volume.

2 Related work

Related work progresses from studio-constrained motion datasets toward larger, more diverse collections and increasingly generalizable humanoid tracking systems. Recent approaches use architectural, sampling, teleoperation, and large-corpus strategies to broaden motion coverage and control.

  • Generalizable tracking: Recent trackers improve generalization through mixture-of-experts, adaptive sampling, unified teacher–student policies, and universal whole-body control.GMT, UniTracker, and SONIC represent distinct architectural or training strategies for diverse motion tracking.
  • Large-scale control: Humanoid-GPT scales whole-body control with a GPT-style Transformer trained on a billion-scale motion corpus and reports zero-shot generalization to unseen motions.This work represents a large-corpus approach within the broader trend toward universal motion control.
  • Motion datasets: Early motion datasets offered high-quality studio motions but limited diversity, while video reconstruction and synthetic generation expanded activities, styles, subjects, and supervision.The passage contrasts studio constraints with newer large-scale data sources.

3 Methodology

GQS converts noisy motion corpora into compact training subsets by filtering physical infeasibility, embedding behavioral similarity, and selecting diverse, dynamically complex motions. Its sampling strategy combines latent-space coverage with complexity weighting to favor informative demonstrations.

  • Overview: GQS transforms a large motion corpus into a compact training subset through physical filtering, semantic embedding, and complexity-weighted sampling.The pipeline first removes infeasible motions, then measures diversity in a learned motion space, and finally selects the training set.
  • Stage I: Physics Filtering: Trajectories shorter than 0.5 seconds or exceeding 0.05 rad/s joint-velocity violation are discarded before soft physics scoring.Remaining trajectories are retained when their physics score satisfies Sphy ≥90.
  • Stage I: Physics Filtering: The physics score penalizes six violation modes, weighting floating, ground penetration, velocity violations, foot sliding, self-collision, and jerk according to downstream sensitivity.The framework emphasizes destructive failures such as extended floating while assigning minimal weights to rare or mild signals such as self-collision and jerk.
  • Stage II: Semantic Motion Embedding: A Periodic Autoencoder maps motions to amplitude, frequency, phase, and offset parameters, enabling reconstruction of trajectories from explicit latent dynamics.Unlike variational autoencoders, the deterministic mapping is optimized solely with reconstruction loss, preserving physical scale and temporal frequency.
  • Stage II: Semantic Motion Embedding: Global embeddings average local [A_w, F_w] descriptors to produce a compact, phase-invariant metric space for motion sampling.Amplitude and frequency capture intensity and pace, while phase and offset represent temporal alignment and pose bias.
  • Stage III: Global Weighted FPS: Global Weighted FPS initializes with the highest-complexity motion and iteratively balances latent-space distance against normalized complexity.The method preserves global exploration while favoring dynamically richer motions when candidates are geometrically comparable.

4 Experiments

Experiments show that GQS-curated subsets improve tracking accuracy, success, optimization trajectories, and cross-dataset generalization while using substantially less data. Ablations indicate that physics filtering, diversity, and complexity contribute complementary benefits.

  • Main Results: Random 3% sampling collapses Success Rate to 83.8% for Any2Track and 64.9% for TWIST2, whereas GQS 3% reaches 95.6% and 86.1%, respectively.The comparison shows that the less-is-more effect depends on curation rather than data reduction alone.
  • Main Results: GQS 10% reaches 95.9% Success Rate, while increasing the subset to 90% provides only marginal improvement despite much higher cost.Physics filtering alone improves Success Rate from 94.2% to 95.4% at full scale.
  • Main Results: TWIST2 MPJPE decreases from 0.099 to 0.084 rad, a 15% relative reduction delivered in a plug-and-play manner.Across settings, GQS provides 5%–15% relative MPJPE reductions.
  • Component Analysis: Removing physics filtering reduces Success Rate from 95% to 91.1% and worsens MPJPE to 0.121 rad, showing that infeasible artifacts harm low-data training.Without prior filtering, embedding-based sampling can select physically infeasible outliers.
  • Component Analysis: Complexity-only selection reaches 93.4% Success Rate, vanilla FPS reaches 94.6%, and full complexity-weighted GQS reaches 95.6%.The ablation supports semantic coverage as a prerequisite, with complexity providing an additional selection bias.
  • Training Dynamics: GQS-curated data yields higher reward and lower tracking error before 0.5B steps, with the advantage maintained throughout training.The result is presented as evidence of a better optimization trajectory rather than merely faster convergence.
  • Cross-Dataset Generalization: GQS achieves lower MPJPE than the full dataset across PHUMA data ratios from 10% to 90% and surpasses the in-domain performance ceiling using 30%.On zero-shot AMASS, the 10% subset reaches 92.8% Success Rate versus 91.0% for the full dataset.
  • Physical Score Analysis: Performance peaks at the 60–70% physical-score slice with 96.3% Success Rate, while the highest-scoring motions reach 94.6%.The physical score identifies toxic data but does not rank the utility of feasible motions by itself.

5 Conclusion

LIMMT presents GQS, a three-stage data-centric framework for humanoid motion tracking. Its central conclusion is that compact motion libraries perform best when they are physically feasible, behaviorally diverse, and dynamically rich.

  • Conclusion: GQS filters infeasible motions, embeds the remainder in a semantic space, and selects a compact subset through complexity-weighted sampling.The pipeline targets feasibility, diversity, and complexity as complementary dimensions of motion quality.
  • Conclusion: Training on just 3% of curated data outperforms full-corpus baselines, with gains transferring across trackers and datasets.The conclusion frames motion value around feasibility, behavioral diversity, and dynamic richness rather than volume.

A Implementation Details

Domain randomization is applied during PPO training to improve sim-to-real transfer and policy robustness.

  • Implementation Details: Training randomizes terrain, external perturbations, joint friction, link mass, and center-of-mass positions.Terrain variation includes floor friction and Perlin-noise height maps, while perturbations occur at random intervals.

A.2 Training Hyperparameters

The experiments use shared PPO hyperparameters and large-scale parallel environments to support stable and fair comparisons across curation strategies.

  • Training Hyperparameters: All policies use 32,768 parallel environments, three hidden network layers, and identical hyperparameters across data curation settings.The shared setup is intended to ensure fair comparison between strategies.

B.1 Global vs. Clustered Selection

Global Weighted FPS outperforms cluster-based selection by directly optimizing coverage and re-balancing imbalanced motion distributions. The advantage is especially clear in the low-data regime and extends to real-world deployment evidence.

  • Interpretation: Global selection naturally re-balances the distribution instead of inheriting the original dataset’s imbalance.This prioritizes information density rather than simply representing the original distribution.
  • Selection strategies: Global Selection applies WFPS across the entire embedding manifold, whereas Cluster-based Selection allocates sampling within 20 K-Means groups.Both strategies use the same physically feasible motion pool.
  • Selection strategies: Global Selection consistently outperforms Cluster-based Selection, particularly when little training data is available.The comparison is made on the same feasible-motion pool.
  • Real-world relevance: The GQS-curated policy trained on 10% data transfers to the Unitree G1 without fine-tuning across basic, expressive, and athletic motions.The deployment result is reported as qualitative real-world tracking evidence.
  • Results: At 10% data, Global WFPS achieves 95.9% SR versus 95.3% for Cluster-based Selection.The authors attribute this gap to cluster-size imbalance, where proportional sampling preserves overrepresented easy behaviors.

B.2 Real-World Deployment

GQS-curated training data supports real-world humanoid motion tracking across daily, expressive, and athletic behaviors. A policy trained on only 10% curated data transfers without fine-tuning and matches or exceeds the Full-Data policy quantitatively.

  • Deployment setting: The Unitree G1 deployment evaluates real-world tracking across daily-life, expressive, and athletic motion categories.The experiment tests sim-to-real transfer using a policy trained on GQS-curated data.
  • Qualitative results: The robot tracks walking, turning, dancing, and athletic movements with balance and coordination across the reported categories.These qualitative results cover daily-life, expressive, and athletic motions.
  • Zero-shot transfer: The 10% GQS-curated policy successfully transfers to the physical robot without fine-tuning.The authors link this result to physics filtering and complexity-biased selection.
  • Quantitative results: Across four categories, the 10% GQS-curated policy matches or surpasses Full-Data policies, averaging +7.5 SR and −15.8% MPJPE.The evaluation uses 10 trials per category on the physical robot.

B.3 Cross-Corpus Evaluation on LaFAN1

The LaFAN1 experiment tests whether GQS’s data-centric benefits extend beyond AMASS. Despite a small training pool and a substantial game-to-mocap domain gap, GQS produces a modest but consistent improvement on AMASS test data.

  • Dataset preparation: LaFAN1 is segmented into 1,532 clips of 576 frames, approximately 10% of the AMASS training pool.The segmentation matches AMASS’ average clip length.
  • Evaluation protocol: The model is trained on LaFAN1 and evaluated zero-shot on the AMASS test set.LaFAN1 is stylistically distinct from AMASS and contains long clips with multiple motion types.
  • Results: Despite the limited training pool and game-vs-mocap domain gap, GQS yields a modest yet consistent improvement.The result indicates that the data-centric advantage extends beyond the AMASS distribution.

B.4 Hyperparameter Sensitivity

The study analyzes GQS sensitivity to physics, geometric, embedding, and ratio-selection choices. The selected defaults are near-optimal and robust, while ARS estimates subset ratios from embedding diversity without ratio sweeps, though it is validated on only two corpora.

  • Sensitivity scope: GQS sensitivity analysis covers the physics threshold, foot-sliding height threshold, PAE window size, and HME embedding half-dimension.These choices are treated as partly constitutive of the curation pipeline rather than mere implementation details.
  • Physics threshold: S_phy = 90 is near-optimal, and clips below 90 are guaranteed to contain at least one severe physical violation under the calibration.The penalty weights are calibrated as w_i = 10/Boundary_i.
  • Foot-sliding threshold: A 5 cm foot-sliding threshold is grounded in Unitree G1 ankle-to-ground clearance, while performance remains essentially flat from 2 to 10 cm.The chosen 5 cm setting is slightly best.
  • Embedding settings: Four-second PAE windows and k = 8 embedding half-dimension are near-optimal, with smooth degradation outside the selected settings.The corresponding analyses use a reduced compute budget for Any2Track at 10% data.
  • Adaptive Ratio Selection: ARS estimates a subset ratio from EDR, where low EDR indicates redundancy and high EDR indicates broader diversity.EDR uses the number of PCA components explaining 95% of HME-embedding variance divided by embedding dimension D = 2k.
  • Adaptive Ratio Selection: On AMASS and PHUMA, ARS predicts 11.0% and 32.8%, closely matching empirical optima of 10% and 30%.The practical recipe adds one PCA step and avoids ratio sweeps, but validation covers only two corpora.
Loading 2606.06953v1…