Physically-Grounded Humanoid Locomotion Data

PHUMA is a physically-grounded humanoid locomotion dataset created by DAVIAN-Robotics. It leverages large-scale human motion data, processed through physics-constrained retargeting to overcome physical artifacts. The dataset was last updated in November 2025.

Use Cases

• Train reinforcement learning agents for bipedal locomotion using curated motion trajectories and physics states.
• Develop humanoid motion imitation models by learning from large-scale human motion data retargeted to a physical avatar.
• Benchmark physics-constrained motion prediction algorithms against the dataset's high-quality locomotion sequences.
• Simulate and analyze humanoid gait stability using the physically-grounded joint angles and body poses.

Strengths

• Created through physics-constrained retargeting to ensure physical plausibility.
• Leverages large-scale human motion data as a foundation.

Limitations

• Specific row count, column details, and file size are unknown.
• The dataset's geographic and demographic origins are not specified, which may limit generalizability.

Provenance

Source

DAVIAN-Robotics via Hugging Face.

Collection Method

Created by curating large-scale human motion data and applying physics-constrained retargeting to a humanoid avatar.

Freshness

Last updated in November 2025.