Zhanke Wang, Chengfeng Zhao, Qing Shuai, Jingzhong Lin, Heng Li, Zeyu Ling, Yuxin Wen, Jing Li, Di Kang, Chunchao Guo and Linchao Bao · Physics-based motion capture

FlowHMR

FlowHMR is a research project that recovers physically plausible 3D human motion from monocular video. The project page gives the approach as flow matching combined with physics-based GRPO. It shows qualitative comparisons (a push-up and a soccer kick) and a simulated PHC tracking view of the captured motion. Three applications are demonstrated: retargeting motions to a Unitree G1 robot with GMR and tracking them in physics simulation with SONIC; animation creation, using GPT-6 Astra for scene reconstruction, retargeting and alignment; and text-to-motion, where MiniMax H3 generates a video from a prompt and FlowHMR captures the human motion from it. The page links to a paper (arXiv 2610.03691), code, a model and a Hugging Face demo. The text gives no quantitative results, licence or release date.