Human Motion Retargeting Pipeline for Humanoid Robots
End-to-end pipeline from IMU motion capture to RL training on the Unitree G1 via NVIDIA Isaac Lab.
The problem: Training humanoid robots to perform dexterous tasks requires large amounts of demonstration data. Collecting it directly on real hardware is slow and expensive. Recording human motion and retargeting it to robot morphologies is the scalable alternative — but technically non-trivial: human and robot joint structures differ fundamentally, and naive retargeting produces physically implausible trajectories.
What I’m building: An end-to-end pipeline from human motion capture to robot learning:
- Motion capture: IMU-based bodysuit + VR headset
- Trajectory format conversion and cleaning for different robot architectures
- Retargeting to Unitree G1 humanoid
- RL policy training in NVIDIA Isaac Lab
Status: Active.
-
1. Animation export (SMPL) -
2. Retargeting -
3. Policy Deployment