NVIDIA opens MotionBricks models for real-time motion generation

NVIDIA released the MotionBricks generative model from a SIGGRAPH 2026 paper, trained on over 350,000 mocap clips and claimed to run at 15,000 FPS with ~2 ms latency: a single network generates motion for both the SOMA skeleton and the Unitree G1 robot, adding new skills without fine-tuning.

Compiled by Sergey KostenchukPublished 2026-10-27Updated 2026-10-27

2026-10-27 Research NVIDIA

🤖 NVIDIA opens MotionBricks models for real-time motion generation

The NVlabs/motionbricks repository was released alongside a SIGGRAPH 2026 paper: the model is trained on 350,000+ mocap clips, with claimed performance of 15,000 FPS and ~2 ms latency. It drives both the SOMA skeleton and the Unitree G1 robot; new skills — from walking to sword lifting — are added without fine-tuning.

🌍 MotionBricks is part of NVIDIA GR00T Whole-Body Control and removes manual animation techniques — foot locking, blending, and collision detection: a single model handles locomotion, styles, and object contact, while new tasks are added zero-shot via Smart Primitives, reducing the cost of content and motion prototypes.

👤 ONNX and checkpoints are already on Hugging Face: the Inbetweener and Scene Preview demos run on Windows and Linux with an NVIDIA GPU. Release v0.2.0 is research-grade, and the terms of the NVIDIA Open Model Agreement are not disclosed.

Source 1: https://nvlabs.github.io/motionbricks/ Source 2: https://github.com/NVlabs/motionbricks

Sources