AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Adapt Human Motion for Robots with NVIDIA SOMA Retargeter

Collected Nov 13, 2026

NVIDIA has published an open-source tool called SOMA Retargeter that converts human motion data into joint trajectories a specific humanoid robot can execute. The developer blog post describes a v0.2 release that ships reference configurations and retargeted example motions for five humanoid platforms: Unitree G1, Unitree H2, Booster T1, AgiBot X2 Ultra, and AgiBot A3T3. Sample motions come from the BONES-SEED motion-capture dataset, supplied by Bones.

The core problem the tool addresses is embodiment mismatch. A human demonstration encodes coordination and timing, but a robot has different limb proportions, joint layouts and range of motion, so the same joint angles cannot simply be replayed. SOMA Retargeter sits on top of the shared human-motion representation in NVIDIA SOMA-X, so heterogeneous sources — mocap sessions, online datasets, video reconstruction or generative models — are first expressed on a common SOMA base skeleton and then adapted per robot.

The pipeline takes a Biovision Hierarchy (BVH) animation on that SOMA skeleton and runs three stages. First it scales the motion targets to the robot's proportions. Then it solves inverse kinematics (IK) to find joint poses that satisfy the scaled targets — the solver is Newton, built on NVIDIA Warp. Finally it applies configured post-processing corrections. Output is a CSV file recording the robot's root position, orientation and joint angles for every frame, which downstream learning or control systems consume as the desired reference motion. Many retargeted clips together form a motion library for training.

Motion can come from three broad channels. For recorded data there is BONES-SEED. For video, GEM-X reconstructs whole-body 3D human motion from a single video using the SOMA body model, and its demo workflow wires video reconstruction directly into G1 retargeting. For generation, Kimodo produces motion sequences from text and pose constraints, while ARDY adds interactive steering while motion unfolds; both offer models that emit motion on the SOMA skeleton. This suggests the retargeter is meant as a normalization layer, not a standalone motion source — its value grows with the range of upstream inputs that can be expressed on SOMA.

Getting started is interactive. After installing the package, a developer launches the BVH-to-CSV converter with a default configuration, which opens a viewer with the Unitree G1 selected. Loading a bundled BVH file shows the human animation; a Retarget action scales targets to G1, solves the IK objectives per frame and applies post-processing, displaying the robot animation alongside the human source. Positioning gizmos let the two be separated for inspection. Playback is the first sanity check: a recognizable forward walk and continuous motion. The robot's poses will differ from the human's by design. Results save to a new CSV, which can be loaded next to the supplied G1 reference for the same motion; the blog notes exact values can vary with configuration and dependency versions.

The Robot Configurator handles new embodiments or refinement of existing ones. It loads a robot description and lets the user define how human motion maps onto it: aligning human and robot reference poses, adjusting scale, and mapping source skeleton joints to robot body parts. The saved configuration is reusable across clips. The post uses a one-handed wave as an illustration: an unaligned reference pose leaves Booster T1 holding both arms out, while aligning the robot's reference pose to the SOMA human reference produces a proper single-arm wave with the other arm at rest, with identical input motion and IK settings.

Because a robot often cannot satisfy every target simultaneously, IK weights decide which body parts take precedence. An experimental IK weight optimizer tunes those weights on a small set of motions representative of the intended tasks, testing adjustments by retargeting and checking how closely and smoothly the robot tracks the targets, then saving the chosen weights into a new configuration.

Foot sliding is a separate correction. A moving reference foot can conflict with an intended planted contact and degrade a downstream tracking policy. Foot-plant correction adjusts foot motion during detected contact intervals, and contact timing and correction settings are tunable before the motion enters training. Once a configuration is tuned on representative clips, headless batch retargeting applies it across a larger collection without the viewer.

The retargeted trajectories then feed policy training. A motion-tracking policy learns to convert the reference plus the robot's current state into control actions. The SONIC workflow includes a SOMA CSV converter for G1 that turns trajectories into SONIC's motion-library format; ProtoMotions is cited as another physics-based motion-tracking framework. The blog stresses that input representation and policy must match the target robot. Simulation is the evaluation stage for balance, ground contact, actuator limits and tracking accuracy. A separate experiment adapted Bones human motion for a robot called MicroDuck and trained a per-clip controller in physics simulation, shown side by side with the human performance and the retargeted reference.

There is also a generative loop. MotionBricks offers a SOMA human model, and its G1 model generates continuous whole-body motion directly for G1 between changing target poses in real time. That G1 model was trained in part on BONES-SEED motions adapted to G1 with SOMA Retargeter, and the MotionBricks paper and project video demonstrate the approach on a physical Unitree G1, with G1 MotionBricks supplying targets and SONIC controlling execution.

In practice, the interesting engineering is the split between scaling, IK and post-processing, and the exposure of weight and contact tuning as first-class settings rather than hardcoded heuristics. A likely trade-off is that quality depends heavily on how well a configuration is tuned and on how closely a given robot's kinematics resemble the five supported platforms; teams with unusual mechanisms will do more work in the Robot Configurator.

Why it matters: Teams training humanoid locomotion and manipulation policies typically struggle to turn human demonstrations into usable references, and a shared SOMA skeleton plus a batch pipeline lowers that cost. Developers get a repeatable path from mocap, video or generated motion to robot-specific CSVs that plug into SONIC or ProtoMotions. Inference: expect retargeting quality, contact handling and weight tuning to become routine parts of humanoid data pipelines rather than one-off scripts.

Read at NVIDIA Developer Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Human motion offers roboticists a great starting point for teaching humanoid robots how to move. A human demonstration captures the coordination and timing...