LIMMT: Less is More for Motion Tracking

2026-06-05

AI Summary

This work addresses the limitations imposed by low-quality or redundant motion data in physics-driven humanoid motion tracking, where such data often hinder policy optimization. It introduces, for the first time, a data-centric framework tailored to physics-based motion tracking, evaluating and filtering motion capture datasets such as AMASS along three dimensions of data quality: physical plausibility, diversity, and complexity. Experimental results demonstrate that training on a high-quality subset comprising less than 3% of the original data not only surpasses the performance achieved with the full dataset but also significantly improves tracking accuracy while drastically reducing data volume. These findings substantiate the principle that a small, carefully curated dataset of high-quality motions can outperform large-scale raw data in physics-driven motion tracking tasks.

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.

Problem

• motion tracking
• data quality
• physics-based animation
• humanoid motion
• motion capture

Innovation

• motion tracking
• data-centric learning
• physics-based animation
• motion data quality
• humanoid control