NVIDIA SOMA Retargeter Adapts Human Motion for Humanoid Robots

NVIDIA has introduced the open-source SOMA Retargeter, a tool designed to adapt human motion to the unique proportions and joint structures of humanoid robots. Built on the shared SOMA skeleton representation, the tool enables developers to configure robots, refine motion outputs, and process motion data for downstream training and control systems.

How SOMA Retargeter Works

The SOMA Retargeter processes human motion data, such as Biovision Hierarchy (BVH) animations, and converts it into robot joint trajectories. The pipeline includes scaling motion targets to match robot proportions, solving inverse kinematics (IK) using NVIDIA Warp, and applying post-processing corrections. The output is a CSV file containing the robot’s root position, orientation, and joint angles for each animation frame. This data can be used to build motion libraries for policy training workflows.

Sources of Human Motion

Developers can source human motion from various platforms, including:

• Motion-capture datasets: BONES-SEED provides human motion-capture data.
• Video reconstruction: GEM-X reconstructs 3D human motion from video using the SOMA body model.
• Generative models: Tools like Kimodo and ARDY generate human motion sequences based on text and pose constraints, outputting data compatible with the SOMA skeleton.

Adapting Motion for Different Robots

The SOMA Retargeter v0.2 release includes configurations and reference results for five humanoid robots: Unitree G1, Unitree H2, Booster T1, AgiBot X2 Ultra, and AgiBot A3T3. Developers can use these examples to explore retargeting workflows and compare their results with supplied references. For instance, a forward walk motion can be retargeted to the Unitree G1 robot, with the resulting animation displayed alongside the human source motion for inspection and refinement.

Tools for Refinement

The Robot Configurator allows developers to align human and robot reference poses, adjust scaling, and map skeletons to robot body parts. Additionally, the IK weight optimizer tunes priorities for smoother motion when robots cannot match all human targets simultaneously. These tools help refine motion outputs, ensuring accurate and realistic robot movements.

Applications and Next Steps

Retargeted trajectories can be integrated into simulation workflows, such as SONIC and ProtoMotions, to evaluate balance, contact, and tracking accuracy. Developers are encouraged to explore the SOMA Retargeter repository for installation instructions, interactive examples, and guides to configure new robots or refine existing setups.