Technical Summary: LIMMT (Less Is More for Motion Tracking)

Problem Statement

Physics-based humanoid motion tracking aims to convert reference motion libraries into deployable, physically grounded behaviors. While the field has seen a rapid expansion of motion capture (MoCap) corpora—from studio-quality datasets like AMASS to internet-scale video reconstructions—the assumption that "more data leads to better generalization" has not consistently held for physics-based imitation reinforcement learning (RL).

Indiscriminate dataset scaling often introduces systematic artifacts (e.g., temporal jitter, foot sliding, ground penetration, and implausible contacts) that corrupt the imitation signal. These low-quality clips can lead to brittle solutions, reward hacking, or convergence toward suboptimal attractors early in training. Furthermore, training on massive libraries is computationally expensive, increasing the costs of reference sampling, curriculum design, and long-horizon optimization. The paper posits that the current challenge is not a lack of data volume, but the presence of noise and redundancy that degrades the optimization trajectory.

Methodology: General Quality Selection (GQS)

The authors propose LIMMT (Less Is More for Motion Tracking), a data-centric framework centered on the General Quality Selection (GQS) pipeline. GQS operationalizes motion data quality through three complementary dimensions: physics feasibility, action diversity, and action complexity. The pipeline is designed as a three-stage, hierarchical process:

Stage I: Physics-Based Feasibility Filtering

Raw kinematic datasets often violate the rigid-body constraints of humanoid robots. This stage acts as a binary gate to eliminate physically impossible motions before they influence representation learning.

• Hard Constraints: Trajectories shorter than 0.5 seconds or those with joint velocity violations exceeding a safety margin (0.05 rad/s) are discarded.
• Soft Scoring: A composite feasibility score ($S_{phy}$) is computed based on weighted penalties for six violation modes:
  • Toxic Artifacts: Floating (sustained lack of support) and Foot Sliding are heavily penalized as they are catastrophic for imitation-RL.
  • Neutral/Safety: Ground Penetration and Self-Collision receive medium weights.
  • Friendly/Dynamic: High Velocity and Jerk are penalized lightly or not at all, as the authors found that filtering these "violent" motions actually degrades performance by removing valuable dynamic signals.
• Outcome: Only motions with a score $S_{phy} \ge 90$ are retained.

Stage II: Semantic Motion Embedding for Diversity

To address redundancy, the framework moves beyond Euclidean pose differences to capture structural and rhythmic similarity.

• Periodic Autoencoder (PAE): A Periodic Autoencoder is trained to decompose motion into dynamic parameters: Amplitude ($A$), Frequency ($F$), Phase Shift ($\phi$), and Offset ($b$).
• Global Embedding: Unlike standard autoencoders, the PAE uses a deterministic mapping to preserve physical scale and temporal frequency. Global embeddings are derived by averaging the amplitude and frequency vectors ($[A, F]$) across the motion sequence, creating a phase-invariant representation.
• Metric Space: Distances in this latent space reflect behavioral diversity (e.g., gait styles, athletic skills) rather than superficial pose differences.

Stage III: Diversity-Aware, Complexity-Weighted Subset Selection

The final stage selects a compact training library using Global Weighted Farthest Point Sampling (FPS).

• Complexity Metric: A complexity score ($C(x)$) is calculated based on kinetic energy and acceleration magnitude, representing the intensity and abruptness of movement.
• Hybrid Scoring: The selection algorithm maximizes a hybrid score:
  $$\text{Score}(u) = \alpha \cdot \hat{D}(u, S) + (1 - \alpha) \cdot \hat{C}(u)$$
  Where $\hat{D}$ is the normalized distance to the nearest neighbor in the embedding space (diversity) and $\hat{C}$ is the normalized complexity.
• Strategy: The process initializes with the highest-complexity motion as an anchor. In subsequent iterations, it prioritizes geometric distance (to ensure broad coverage) but introduces a bias toward dynamically richer motions when candidates are geometrically equidistant.

Key Contributions

1. Data-Centric Perspective: The paper redefines "quality" over "quantity," arguing that physics feasibility, action diversity, and complexity are the decisive factors for robust tracking, challenging the blind scaling assumption in physics-based RL.
2. The GQS Framework: A hierarchical pipeline that first eliminates physically infeasible artifacts via simulator grounding, then maximizes behavioral coverage and dynamic richness using harmonic embeddings and complexity-weighted sampling.
3. The "Less-Is-More" Paradigm: Empirical evidence showing that training on a curated subset of just 3% of the AMASS dataset consistently outperforms training on the full dataset across multiple metrics and tracking architectures.

Experimental Results

The framework was evaluated on the AMASS and PHUMA datasets using two state-of-the-art tracking policies: Any2Track and TWIST2.

• Performance vs. Data Ratio:
  • Random Subsampling: Randomly selecting 3% of the raw data caused catastrophic performance collapse (e.g., Any2Track Success Rate dropped from 94.2% to 83.8%).
  • GQS Performance: Using only 3% of the data curated by GQS, the Success Rate reached 95.6% for Any2Track and 86.1% for TWIST2, surpassing the full-dataset baselines (94.2% and 82.5%, respectively).
  • Error Reduction: GQS achieved a 5–15% relative reduction in Mean Per Joint Position Error (MPJPE) compared to full-data baselines. For TWIST2, MPJPE improved from 0.099 to 0.084 rad.
• Training Dynamics: Learning curves demonstrated that GQS-curated data provides cleaner gradients, steering the policy toward better solutions from the early stages of training (before 0.5B steps), rather than just accelerating convergence.
• Ablation Studies:
  • Removing the physics filter caused a significant drop in Success Rate (to 91.1%), confirming that embedding-based sampling without filtering favors toxic outliers.
  • Removing complexity weighting resulted in suboptimal performance (93.4% SR), proving that diversity alone is insufficient without prioritizing informative, high-energy motions.
• Generalization: On the PHUMA dataset (which is already physically grounded), GQS improved precision (lower MPJPE) and zero-shot transfer to AMASS, demonstrating that diversity and complexity selection add value even when physical feasibility is already high.

Significance and Claims

The paper claims that motion data quality dominates data quantity because it shapes the initial optimization trajectory. High-quality targets provide consistent, physically meaningful gradients that steer learning toward stable solutions early, whereas low-quality motions inject biased objectives and unstable gradients.

The authors assert that their framework offers a plug-and-play solution that improves the training signal for any existing tracker without exploiting algorithmic idiosyncrasies. By curating a compact, high-value library, LIMMT reduces training costs by an order of magnitude while simultaneously improving tracking accuracy, success rates, and robustness. The work suggests that future motion data acquisition and policy training should prioritize the curation of physically feasible, diverse, and complex motions over the indiscriminate accumulation of large-scale corpora.