AI tools directory / Video

NVIDIA MotionBricks

Generate real-time character and robot motion between poses.

Checked against nvlabs.github.io on Nov 16, 2026

Visit nvlabs.github.io

What NVIDIA MotionBricks does

MotionBricks is a real-time motion generation model from NVIDIA research, published as a SIGGRAPH 2026 paper. You give it a character's last few poses and up to four target poses, and it generates the movement in between. Chaining these generations produces continuous walking, running, keyframe animation or object interactions. One runtime drives both a human skeleton (SOMA) and the Unitree G1 humanoid robot. It is a developer release with Python code and pretrained weights, not an app.

Best for

  • Game and animation developers prototyping runtime character motion
  • Robotics researchers planning motion for the Unitree G1 humanoid
  • Teams with an NVIDIA GPU who are comfortable running Python demos
  • Researchers studying keyframe in-betweening and motion generation

Key features

Motion between poses

Each call fills in about 0.6 to 1.9 seconds of 30 Hz motion between four recent poses and one to four target poses. The model picks the transition length itself.

Smart Primitives

Describe an action, such as vaulting or sitting, as target poses placed around an object, then reuse it in new scenes without retraining.

Human and robot models

Pretrained models cover the SOMA human body model and the Unitree G1 robot, and the same runtime code drives both.

Demos and editors

The repo includes an Inbetweener demo, a Scene Preview where a character walks, runs, ducks and sits through a gym scene, plus a Smart Primitive Editor and a Scene Editor.

ONNX and TensorRT runtime

ONNX models are bundled, the SafeTensors checkpoints are on Hugging Face, and an exporter lets you build ONNX files for other runtimes.

NVIDIA MotionBricks pricing

NVIDIA MotionBricks has a free plan. The code, ONNX models and SafeTensors checkpoints are free to download and run on your own hardware. Licenses differ by part. Code is Apache 2.0. The SOMA and G1 weights use the NVIDIA Open Model Agreement, which NVIDIA says allows commercial use and redistribution of modified models under its terms. The bundled sample motion clips use a separate NVIDIA evaluation license limited to testing NVIDIA technologies. There is no hosted service or paid plan. Prices checked on the official site on Nov 16, 2026.

Open release$0Code under Apache 2.0; model weights under the NVIDIA Open Model Agreement

Source: NVIDIA MotionBricks pricing page

Pros and cons

What it does well

  • Weights and code are both public, and the model license permits commercial use under NVIDIA's terms.
  • You can run the demos right away because ready ONNX models ship in the repo.
  • The same model approach works for a game character and a humanoid robot.

Know before you sign up

  • The TensorRT runtime needs an NVIDIA GPU; the demos were tested on Windows and Linux x86-64 with an RTX 4090.
  • It takes poses as input, not text prompts or scene geometry, and other skeletons need retargeting to the SOMA or G1 layout.
  • Output is kinematic motion, so a physical robot still needs a separate tracking controller and safety checks.

Alternatives to NVIDIA MotionBricks

Questions about NVIDIA MotionBricks

What is NVIDIA MotionBricks?

NVIDIA MotionBricks is an AI tool from nvlabs.github.io, listed in our Video category. Generate real-time character and robot motion between poses.

Is NVIDIA MotionBricks free?

Yes, NVIDIA MotionBricks has a free plan.

Where can I use NVIDIA MotionBricks?

On Windows and Linux.

What are some alternatives to NVIDIA MotionBricks?

Tools that do a similar job include Viggle, DomoAI, Animate Artwork and Kaiber.