NVIDIA’s SOMA Retargeter Maps Human Motion to Five Robot Models

NVIDIA’s open-source SOMA Retargeter v0.2 converts human-motion data into joint trajectories for humanoid robots, giving developers material for simulation and policy-training workflows.

Built around a shared SOMA skeleton, the tool can process motion data such as Biovision Hierarchy (BVH) animations. It scales motion targets to fit a robot’s proportions, solves joint poses through inverse kinematics using NVIDIA Warp, then applies post-processing corrections. The CSV stores the robot’s root translation, rotation and joint values frame by frame.

The release includes configurations and reference results for five models: Unitree G1, Unitree H2, Booster T1, AgiBot X2 Ultra and AgiBot A3T3. Developers can compare their results with the supplied references, including a forward-walk example for Unitree G1.

A Robot Configurator lets developers align human and robot reference poses, adjust scale and map skeletons to robot body parts. An inverse-kinematics weight optimizer adjusts priorities when a robot cannot match all motion targets at once.

Human-motion inputs can come from motion-capture datasets, video reconstruction or generative models. BONES-SEED provides motion-capture data, while GEM-X reconstructs three-dimensional human motion from video using the SOMA body model. Kimodo and ARDY generate motion sequences from text and pose constraints, with outputs compatible with the SOMA skeleton.

The resulting data can be used to build motion libraries for policy training. The trajectories can also be integrated into simulation workflows such as SONIC and ProtoMotions to evaluate balance, contact and tracking accuracy. NVIDIA’s robotics work has also included an expansion of its Isaac platform.