PFM-HR: Pose Flow Matching for Humanoid Robots

This research presents PFM-HR, a reusable flow matching prior designed for humanoid robots that is trained directly on large-scale unordered pose data. The method introduces the Pose Geometry Score (PGS) to quantify how joint coordinate changes during rollouts align with the local geometry of pose variation captured by the prior. By using PGS to modulate tracking rewards while keeping the prior frozen, the approach guides policy exploration toward structured pose changes. Experiments indicate that PFM-HR improves both single motion and general motion tracking, particularly for highly dynamic motions.