NVIDIA MotionBricks adds SOMA tools, but the Unreal demo is not included
NVIDIA has highlighted a MotionBricks update that gives developers SOMA human-motion models, downloadable checkpoints and a Newton reference implementation. The practical news is the usable package, not a newly invented animation method. Its October 27 announcement also draws an important boundary: the Unreal Engine integration shown in the research demonstration is not included.
The 0.2.0 changelog is dated October 23. It records the move from a Unitree G1 preview into a standalone repository, a SOMA model and rebuilt Python tools. The original paper dates to April 27, 2026. The October update packages that earlier research for developers.
What developers can actually download
The repository supplies a Python runtime using Warp, Newton and TensorRT. Its TensorRT demos require an NVIDIA GPU and were tested on Windows and Linux x86-64 with an RTX 4090. Choose the artifact for your job rather than downloading every checkpoint.
| Artifact | What it is for |
|---|---|
| Bundled ONNX models | Running the SOMA and G1 demos |
| SafeTensors checkpoints | Exporting models or building another runtime |
| Inbetweener and Scene Preview | Trying pose transitions and interactive scenes |
| Primitive and scene editors | Authoring target poses and interactions |
MotionBricks connects recent poses to requested target poses. Smart Primitives organize those targets into reusable interactions without retraining. Output motion is represented at 30 Hz. That describes the motion timeline, not how many game frames a GPU renders.
What the current speed measurements include
The released-model inference documentation reports these GPU times for one request, or N = 1. They use TensorRT 10.8.0.43 on an RTX 4090 under Windows.
| Model | FP32 | Mixed FP16 |
|---|---|---|
| SOMA | 2.50 ms | 1.28 ms |
| Unitree G1 | 2.67 ms | 1.63 ms |
Inputs were already on the GPU, with CUDA graph replay. Input preparation, output decoding and endpoint correction are excluded. Each result is the fastest of six plan builds. These timings are not an end-to-end motion-quality comparison, and build-to-build variation matters.
The older paper reported 15,000 generated motion frames per second and 2 ms inference latency on an RTX 5090. Those research figures are neither displayed game FPS nor the same benchmark as the RTX 4090 table above. A generated motion sequence and a rendered frame are different outputs.
The missing pieces matter as much as the speed
The checkpoint package contains weights and inference metadata, not a training pipeline or the complete training dataset. It also does not supply a physical robot controller. Hardware deployment needs a separate tracking controller, robot-limit checks and safety validation.
The SOMA model card says a complete end-to-end benchmark of these checkpoints is not included. It also says the model does not receive scene geometry. Your application remains responsible for obstacles and target placement, even when a demonstration looks scene-aware.
Licenses are split. Code uses Apache 2.0, weights use the NVIDIA Open Model Agreement, and covered motion samples use a separate license for internal evaluation and testing of NVIDIA technologies. Downloadable weights do not make every bundled asset interchangeable or unrestricted.
Our SuperMimic coverage examines the separate problem of making a humanoid track reference motion. Generating a desired movement and controlling a body to reproduce it are different jobs.
A useful first evaluation
Our suggested test, not a test we performed, is to build one short interaction and change its target placement. Record the complete application delay, then inspect foot contact, obstacle clearance and transitions. Compare precision settings on that same interaction before choosing one for speed. Fast model inference is useful only if the resulting movement meets your application’s needs.
