Distinguishing Imitation Error from Intrinsic Motion Learning Difficulty

Physics-based motion imitation often gets judged by end results like MPJPE, but that can hide the real failure mode. This paper argues that a policy can miss a motion either because the controller is weak or because the motion is intrinsically difficult to learn. TVS, the proposed metric, estimates that difficulty by measuring how much torque variation is needed to correct small pose perturbations.

The practical takeaway is that high imitation error is not always a training problem. The authors report that high-TV motions create flatter reward landscapes and vanishing policy gradients, which helps explain why some motions stay hard even with strong methods. They tested the idea against UHC and PHC+, and say TVS correlates strongly with imitation error while also helping attribute whether the issue is policy deficiency or motion-inherent difficulty.

For developers, the interesting part is the diagnostic angle. TVS is used to define Maximum Imitable Difficulty (a way to gauge policy capability), Difficulty-Stratified Joint Error for more granular profiling, and Flawed Motion Detection to flag suspicious mocap segments. That makes it relevant not just for humanoid RL researchers, but also for teams cleaning motion libraries or trying to understand why certain clips never converge.

The broader context is that motion learning pipelines often mix data quality issues, controller limits, and reward design problems into one number. A metric like TVS could help separate those concerns earlier, which is useful if you are building character controllers, evaluating mocap sets, or...