LIMMT: Less Is More for Motion Tracking

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 erroneous clips. We define motion data quality through three dimensions: physics feasibility, diversity, and complexity. We show that training with under 3% of AMASS yields better tracking performance than training with the full dataset. Extensive experiments and analyses validate the effectiveness of our framework.

We want humanoid robots to imitate human motion—walking, jumping, dancing. A common assumption is that more training data is always better: collect as much motion-capture data as possible and train the robot's controller on everything. We show this assumption is wrong. Large motion datasets contain many "low-value" clips: motions that are physically impossible for the robot, near-duplicates of common walking patterns, or motions that are too simple to teach anything new. Training on this raw mix actually hurts performance. We introduce LIMMT, which judges the value of each motion clip along three dimensions: whether the robot can physically execute it, whether it covers behavior that is not already represented in the training pool, and whether it carries enough dynamic information to drive learning. A three-stage curation pipeline then picks a small, high-quality subset from any large motion library. The surprising result: training on just 3% of the widely used AMASS dataset, after curation, beats training on the full 100%. The benefit transfers across different controller designs (Any2Track, TWIST2, BeyondMimic) and across datasets, and the curated policies also track better when deployed on a real Unitree G1 robot. This challenges the "scale-is-all-you-need" assumption common in machine learning and suggests a more economical recipe for humanoid training: before paying to collect more motions, first decide which motions are actually worth learning from.