Low-angle shot of humanoid robot legs climbing concrete stairs
AI-generated image

NVIDIA

NVIDIA's open source SOMA Retargeter v0.2 maps human motion onto five humanoid robots

NVIDIA has published a walkthrough of SOMA Retargeter, an open source tool that adapts recorded or generated human motion to a specific humanoid robot's proportions and joints. In a post on its developer blog, the company says the v0.2 release includes reference configurations and retargeted examples for five robots: Unitree G1, Unitree H2, Booster T1, AgiBot X2 Ultra and AgiBot A3T3.

The problem it addresses is simple to state: a robot cannot copy a person's joint angles directly because its limbs, joints and range of motion differ. NVIDIA says the tool reads a BVH animation on its shared SOMA skeleton, scales the motion targets to the robot, solves inverse kinematics using Newton, built on NVIDIA Warp, and applies corrections such as foot-plant adjustment.

The output is a CSV file with the robot's root position, orientation and joint angles for each frame. NVIDIA says those trajectories can be converted into motion-library formats for policy training workflows such as SONIC and ProtoMotions, then checked in simulation for balance, contact and tracking.

Input motion can come from motion capture, from video reconstruction with GEM-X, or from generative models such as Kimodo and ARDY, according to the blog. The release ships sample clips supplied by Bones from its BONES-SEED dataset, and the code is on GitHub.

More on: Unitree · NVIDIA · China

← All news