Signal HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark

Summary

Humanoid motion tracking plays a central role in teleoperation and whole-body imitation. This paper argues that the evaluation methods widely used so far often diverge from the errors people actually notice when watching motion videos. Standard kinematic error metrics only average per-frame pose differences and fail to capture the physical problems people are most sensitive to, such as unstable support or mistimed foot contacts like skating and incorrect touch-downs. The authors also point out that the test datasets commonly used in this field tend to be small and narrow in scope. To address these gaps they propose a new benchmark called HumanTracker. HumanTracker is intended to evaluate humanoid motion tracking systems comprehensively in a way that better matches human perception.