1HKUST(GZ) 2Noitom Robotics 3SIGS, Tsinghua University 4Google
*Equal Contribution
†Corresponding author
Motion priors improve reinforcement learning for physics-based humanoid tracking, but temporal priors require ordered motion clips, while pose priors provide limited guidance for policy-induced pose transitions. We present Pose Flow Matching for Humanoid Robots (PFM-HR), a reusable flow matching prior trained directly on large scale unordered pose data. PFM-HR introduces the Pose Geometry Score (PGS), which quantifies how joint coordinate changes during rollouts align with the local geometry of pose variation captured by the prior. Using PGS to modulate the tracking reward guides policy exploration toward structured pose changes while keeping the prior frozen across tracking tasks. Experiments demonstrate that PFM-HR improves both single-motion and general motion tracking, especially for highly dynamic motions.
Visual comparison on single-motion tracking of dynamic skills. PFM-HR masters these skills with fewer training samples and higher tracking accuracy than both baselines.
Backflip
14.3% fewer samples and 6.3% lower position error vs ADD w/ PDF-HR (ADD fails to converge).
ADD
ADD w/ PDF-HR
ADD w/ PFM-HR
Double Kong
28.8% fewer samples and 9.7% lower position error vs ADD w/ PDF-HR (ADD fails to converge).
ADD
ADD w/ PDF-HR
ADD w/ PFM-HR
General-motion tracking across episode lengths. PFM-HR reaches higher success rates with fewer training samples than both baselines.
We deploy with the BeyondMimic pipeline, adding PFM-HR only as a frozen prior during simulation training. Compared to the original BeyondMimic, BeyondMimic w/ PFM-HR cuts the simulation samples needed to reach by 24.2% (Spinkick), and 15.1% (Kick combo).
@misc{gao2026pfmhrposeflowmatching,
title={PFM-HR: Pose Flow Matching for Humanoid Robots},
author={Yukang Gao and Yi Gu and Yangchen Zhou and Xingyu Chen and Zhaorui Wang and Fanghai Zhang and Hanyang Cao and Zhengyang Shen and Ji Ma and Runhan Zhang and Lei Han and Renjing Xu},
year={2026},
eprint={2608.03227},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2608.03227},
}