G1 Humanoid Robot Motion Capture Clips from Real Performers

Containing 60 motion capture clips for the Unitree G1 EDU humanoid robot, captured from real performers in Austin, TX. Data is provided at multiple pipeline stages including raw mocap (BVH/FBX), retargeted joint trajectories (PKL), processed RL training data (NPZ), and trained ONNX policies.

Use Cases

• Train reinforcement learning policies using the processed RL training data (NPZ) for humanoid robot control.
• Retarget raw motion capture clips (BVH/FBX) to the G1 robot's 29 degrees of freedom for motion imitation.
• Benchmark monocular video extraction (video2robot) performance by comparing raw mocap to retargeted joint trajectories (PKL).
• Deploy trained ONNX policies for real-time control of the Unitree G1 robot using the provided model files.

Strengths

• Contains 60 distinct motion capture clips, providing a foundation for robot learning.
• Data is captured from real human performers, offering natural motion patterns.
• Includes data from multiple pipeline stages (raw mocap, retargeted trajectories, training data, policies) for end-to-end workflows.
• Specifically targets the Unitree G1 EDU robot with its 29 degrees of freedom.

Limitations

• Limited to 60 clips, which may not cover a wide diversity of motions for robust model training.
• Geographic bias as all data is captured from performers in Austin, TX, USA.
• Relies on markerless motion capture (MOVIN TRACIN) and monocular video extraction, which can introduce noise compared to marker-based systems.

Provenance

Source

exptech on Hugging Face.

Collection Method

Captured from real performers using MOVIN TRACIN markerless motion capture and video2robot monocular video extraction.

Freshness

Last updated March 15, 2026.

Geography

Austin, Texas, United States.