The method recovers SMPL body meshes from generated videos, retargets them onto a humanoid inside the MuJoCo physics simulator, and scores the motion along interpretable axes. These include kinematic plausibility, contact and balance consistency, and dynamic feasibility. Because each component is continuous and tied to a specific physical property, the reward can diagnose which part of the motion is wrong instead of returning only a vague visual score.
PhyMotion is useful for researchers using reinforcement learning or post-training to improve human video models. It can help align generated motion with human judgment, reduce physical artifacts, and guide models toward more believable full-body action. The project is best understood as a reward and evaluation framework for video generation, not a general consumer animation editor.
