PHUMA

A large-scale humanoid reference-motion dataset, filtered and retargeted under physical constraints.

PHUMA is a humanoid-robot motion dataset released in October 2025 by Jaegul Choo's group at KAIST (the Korea Advanced Institute of Science and Technology). Motion extracted from internet video and then retargeted (such as Humanoid-X) often contains physically impossible artifacts like floating, ground penetration, and foot skating. PHUMA first applies physics-aware filtering, removing clips with excessive jitter, an unstable center of mass, or feet that never touch the ground; it then uses PhySINK (physically constrained inverse-kinematics retargeting) to stay close to the original motion while satisfying joint limits, keeping feet planted on the ground, and preventing the supporting foot from sliding. The data merges motion capture such as AMASS with several video sources, totaling 76,010 clips and about 73 hours, released in versions for both the Unitree G1 and the H1-2. Motion-tracking policies trained on it achieve higher success rates than ones trained on AMASS or Humanoid-X, and transfer zero-shot to a real G1.

ExampleA general-purpose motion-tracking policy trained on PHUMA's G1 version lets a real Unitree G1 reproduce walking, turning, and other motions performed by a person in video.

Also called

Physically Reliable Humanoid Locomotion Dataset

Related

Foot Skating / Floating / Penetration · Humanoid-X · AMASS (Archive of Motion Capture as Surface Shapes) · Motion Retargeting · Motion Tracking · Unitree G1

Sources

PHUMA: Physically Reliable Humanoid Locomotion Dataset (arXiv 2510.26236)
PHUMA 论文 HTML 全文 (Chinese)

As of

2026-06