LIMMT: Less is More for Motion Tracking

TL;DR

LIMMT uses physics feasibility, diversity, and complexity to select high-quality motion data, outperforming full datasets with only 3%.

cs.RO 🔴 Advanced 2026-06-05 53 views
Yu Guan Zekun Qi Chenghuai Lin Xuchuan Chen Dairu Liu Wenyao Zhang Jilong Wang Xinqiang Yu He Wang Li Yi
motion tracking data filtering physics-based deep learning robotics

Key Findings

Methodology

The paper introduces a three-stage GQS pipeline: first, physics-based filtering using a rigid-body simulator to eliminate infeasible motions like floating or ground penetration; second, learning a semantic motion embedding via a Periodic Autoencoder (PAE) to capture behavior structure and rhythm; third, applying a weighted FPS sampling that combines motion complexity metrics—kinetic energy and acceleration—to select a diverse, dynamic subset. The approach emphasizes the importance of filtering before embedding and complexity weighting after, ensuring data quality and informativeness. This systematic process transforms noisy large datasets into compact, high-value training sets suitable for physics-based humanoid motion tracking.

Key Results

  • On AMASS, GQS-curated subsets with only 3% of data outperform full datasets, achieving success rates of 95.6%, MPJPE reduced by 15%, and outperforming random sampling by large margins. In TWIST2, MPJPE drops from 0.099 to 0.084 rad. Ablation studies confirm the critical role of physics filtering, with the full GQS framework yielding the best results. Cross-dataset tests on PHUMA show improved generalization and robustness, with success rate surpassing 92% using only 10% of data.
  • Training with curated data accelerates early optimization, providing cleaner gradients and better convergence trajectories. The method demonstrates strong transferability across different trackers and domains, reducing training costs significantly while maintaining or improving accuracy. The results validate that quality-focused data selection can be more effective than mere data volume, especially in physics-based motion learning.
  • Overall, GQS achieves a 'less is more' effect, where carefully filtered and selected motion data leads to superior tracking performance, robustness, and efficiency, setting a new standard for data utilization in humanoid motion learning.

Significance

This work challenges the conventional wisdom that more data always leads to better models, highlighting instead the importance of data quality. By systematically filtering and selecting motion data based on physical plausibility, diversity, and complexity, it significantly improves the training efficiency and robustness of humanoid motion tracking systems. The approach addresses longstanding issues of noise, artifacts, and redundancy in large-scale motion datasets, paving the way for more reliable and realistic human motion imitation in robotics and virtual avatar applications. Its plug-and-play nature ensures broad applicability across different tracking architectures, making it a valuable contribution to both academia and industry. The framework's ability to outperform models trained on full datasets with only a fraction of the data demonstrates a paradigm shift towards smarter data utilization, with implications for cost reduction and performance enhancement in real-world deployments.

Technical Contribution

The paper introduces a novel data curation framework that integrates physics-based feasibility checks, semantic behavior embeddings, and complexity-aware sampling. The physics filtering module ensures only physically plausible motions are used, preventing training on infeasible data. The harmonic motion embedding captures behavior structure and rhythm, providing a meaningful semantic space for diversity measurement. The weighted FPS sampling combines geometric coverage with motion complexity, prioritizing dynamic and challenging motions. This multi-step, hierarchical approach addresses the limitations of existing large-scale datasets, enabling more efficient and robust policy learning. It also offers a theoretical guarantee that high-quality data enhances early training trajectories, leading to better convergence and generalization.

Novelty

This is the first work to systematically incorporate physical feasibility, semantic behavior embedding, and motion complexity into a unified data filtering and selection pipeline for humanoid motion tracking. Unlike prior approaches that rely solely on dataset size or simple filtering, GQS emphasizes the quality and informativeness of data, with a carefully designed hierarchy ensuring the most relevant motions are used for training. The combination of simulator-grounded filtering, harmonic embeddings, and complexity-biased sampling constitutes a comprehensive and innovative strategy that significantly outperforms traditional methods in efficiency and accuracy.

Limitations

  • The approach depends heavily on the accuracy of the physics simulator and the motion embedding model; inaccuracies here can lead to the exclusion of useful data or inclusion of infeasible motions. Its computational cost remains high, especially for very large datasets, limiting real-time applicability. The current metrics focus mainly on physical and dynamic features, potentially neglecting semantic context or scene-specific factors. The method's effectiveness in highly complex or multi-agent scenarios requires further validation. Future work should aim to optimize computational efficiency and incorporate richer semantic information for more nuanced filtering.

Future Work

Future research could focus on integrating multi-modal data, such as visual or contextual cues, to improve filtering precision. Developing adaptive, learning-based filtering strategies could further automate and optimize the process. Extending the framework to multi-agent or non-rigid motions would broaden its applicability. Additionally, real-time filtering and data augmentation techniques could make the approach more scalable for online applications. Combining this with self-supervised learning could enable autonomous data collection and refinement, pushing the boundaries of humanoid motion understanding and control.

AI Executive Summary

Deep Dive

Plain Language Accessible to non-experts

想象你在准备一份舞蹈表演的动作库。你有很多动作片段,有些动作看起来不自然,比如漂浮在空中或穿墙。这些动作虽然被录制了,但在真实世界中根本不可能用在舞蹈里。为了让舞蹈看起来自然流畅,你会先用一个“魔法检测器”检查每个动作,剔除那些不符合物理规律的片段。然后,你会用一种“动作理解器”分析动作的节奏和结构,确保动作多样且有趣。最后,你会根据动作的动态强度,比如跳跃的高度或速度,优先选择那些更有力量、更丰富的动作。这样筛选出来的动作库,不仅符合物理规律,还能让机器人跳得更自然、更酷,就像专业舞者一样。

ELI14 Explained like you're 14

想象你在教你的朋友跳舞,但你发现有些动作根本不可能,比如在空中漂浮或者穿墙。你不会用那些奇怪的动作来教,因为它们会让机器人跳得很怪。于是,你用一个“魔法检测器”检查每个动作,剔除那些不合理的。接着,你用一个“动作分析器”理解动作的节奏和结构,确保动作多样又有趣。最后,你会优先选择那些跳得更高、更快、更有力量的动作,让机器人跳舞看起来更酷、更自然。这样筛选出来的动作库,既合理又丰富,机器人学会跳舞就会更厉害啦!

Abstract

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.

cs.RO