MotionBricks

MotionBricks is a real-time generative motion framework from NVIDIA, published in ACM Transactions on Graphics (SIGGRAPH 2026). It uses a single modular latent generative backbone trained on over 350,000 motion clips (the BONES-SEED mocap dataset) and reports 15,000 FPS with 2 ms latency. It generates motion between target poses (in-betweening) for characters and humanoid robots.

On top of the backbone, "smart primitives" give an interface for navigation (smart locomotion with velocity, heading and style commands) and object interaction (smart objects defined by proxy keyframes, e.g. sitting, picking up a sword, jumping over a bench). The authors state that it applies to new downstream tasks zero-shot, without fine-tuning or task-specific tagging. Their demo runs in Unreal Engine 5 and shows styles such as zombie, injured and skipping, with no hand-authored transitions.

NVIDIA released code at NVlabs/motionbricks with Warp, Newton and TensorRT runtimes, pretrained SOMA (human) and Unitree G1 models (ONNX bundled, SafeTensors checkpoints on Hugging Face), an ONNX exporter, interactive demos, and a Smart Primitive Editor and Scene Editor. It is also part of NVIDIA's GR00T Whole-Body Control effort. Per creator reports, the models are about 800 MB each and are released under an NVIDIA open model agreement; the exact license text was not in the provided pages.