NVIDIA/soma-retargeter

What it solves

SOMA Retargeter addresses the challenge of transferring human motion data (from BVH files) to humanoid robots. Because humans and robots have different body proportions and joint structures, motion cannot be copied directly; it must be adapted to fit the robot's physical constraints and proportions while maintaining the original movement's intent.

How it works

The tool processes SOMA-base skeleton animations through a multi-step pipeline:

1. Scaling: It adapts human-scale joints to match the robot's specific proportions.
2. Inverse Kinematics (IK): It solves IK for each frame to determine the necessary joint angles for the robot to mimic the human pose.
3. Stabilization: It stabilizes contacts (such as foot plants) and clamps joint values to ensure they stay within the robot's configured physical limits.
4. Output: The final result is exported as a CSV file containing the robot's root pose and actuated joint values, which can then be used in simulations or robot controllers.

Who it’s for

This software is designed for robotics researchers and developers working with humanoid robots who need to convert motion capture data into executable robot animations.

Highlights

• Multi-Robot Support: Includes pre-configured assets for several robots, including unitree_g1, unitree_h2, booster_t1, agibot_x2ultra, and agibot-a3t3.
• Flexible Interface: Offers both an interactive viewer for manual inspection and a headless mode for batch processing large directories of motion files.
• Optimization Tools: Includes an IK weight optimizer to refine how the robot achieves target poses.
• Integration: Part of the SOMA ecosystem, complementing tools for motion estimation and whole-body control.