Is MotionBricks state of the art?Pending benchmarks.Announced Oct 23, 2026; no independent results yet.
NVIDIA's open real-time motion generation model for human characters and humanoid robots. Given a character's four most recent poses and up to four target poses, it generates the movement between them; chaining generations gives continuous locomotion and object interactions. Two packages: SOMA (human skeleton) and Unitree G1 (robot). About 200M parameters each, weights on Hugging Face under the NVIDIA Open Model Agreement.
The details so far
- Lab
- NVIDIA
- Announced
- Oct 23, 2026
- API
- Not yet · None announced by NVIDIA
- Open weights
- Published · Hugging Face nvidia/MotionBricks (soma/ and g1/ folders), not gated
- Parameters
- About 200M (SOMA) and about 195M (Unitree G1), each split across a pose VQ-VAE, pose predictor and root predictor
- Task
- Pose-conditioned motion generation and inbetweening at 30 Hz; each generation is 6 to 16 tokens of four frames (about 0.6 to 1.9 s)
- Variants
- MotionBricks-SOMA-RP-v1 (30-joint SOMA human skeleton, step 1,500,000) and MotionBricks-G1-RP-v1 (34-joint Unitree G1 skeleton, step 2,000,000), FP32 SafeTensors; ONNX exports in the GitHub repo
- Architecture
- Modular latent generative model: a VQ-VAE pose tokenizer and decoder, a masked-token transformer pose predictor and a transformer root predictor that also picks the duration
- Inputs
- Skeletal-motion features and pose and root constraints only; no text, image, audio, video or scene-geometry inputs
- Training data
- Bones RigPlay 1 (RP1) motion capture: 313,526 training clips from a source corpus of about 350,000 clips and 700 hours (G1 data retargeted to the robot)
- Robot use
- Generates reference motion, not a robot controller; running it on hardware needs a separate tracking controller and safety checks
- Code
- GitHub NVlabs/motionbricks (v0.2.0): TensorRT runtime on Warp and Newton, Inbetweener and Scene Preview demos, primitive and scene editors, ONNX exporter; no training pipeline
- Earlier release
- The G1 weights first shipped as a preview in NVlabs/GR00T-WholeBodyControl; the SOMA model is new in this release
- License
- NVIDIA Open Model Agreement for the weights (commercial use permitted); code Apache 2.0; bundled sample motions under the NVIDIA Sample Data Evaluation License
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 15,000 FPS and 2 ms latency, covering over 350,000 motion skills with a single neural backbone (project page).
Lab claim: applies to new downstream tasks zero-shot, with no fine-tuning or task-specific tagging, through its Smart Primitives interface.
Sources
REPOnvidia/MotionBricks on Hugging Facehuggingface.co/nvidia/MotionBricks · Oct 23, 2026REPONVlabs/motionbricks on GitHubgithub.com/NVlabs/motionbricks · Oct 23, 2026LABMotionBricks project page (NVlabs)nvlabs.github.io/motionbricks/PAPERMotionBricks paper (arXiv 2604.24833)arxiv.org/abs/2604.24833REPOMotionBricks G1 preview in NVlabs/GR00T-WholeBodyControlgithub.com/NVlabs/GR00T-WholeBodyControl/tree/main/motionbricksPRESSAlvaro Bartolome (Hugging Face) post on Xx.com/alvarobartt/status/2114720297584574829 · Oct 26, 2026PRESSHuggingPapers post on Xx.com/HuggingPapers/status/2114731481906208811 · Oct 26, 2026Tracking continues until an independent board lists it.