NVIDIA MotionBricks Uses One Model to Drive Game Characters and Robots

NVIDIA has unveiled MotionBricks, a real-time motion-generation AI designed to further blur the boundary between simulation and the physical world. The single model can animate characters in video games and, at the same time, drive humanoid robots in reality.

NVIDIA says the system delivers throughput of up to 15,000 frames per second with only 2 milliseconds of latency. That performance lets it generate smooth, complex motions faster than the response limits of a game engine, according to the company. The company frames the model as a unified motion engine rather than separate tools for games and robotics, targeting interactive applications where responsiveness matters.

The technology is detailed in a SIGGRAPH 2026 paper. MotionBricks drops conventional animation state machines and instead relies on a generative model trained on a dataset of more than 350,000 motion clips.

The model generates subsequent motion in real time from the previous few poses and a target pose, enabling dynamic and responsive control of characters. Its central feature is “Smart Primitives,” a set of modular motion skills that covers commands such as jumping, sitting down, or grasping an object. Developers can combine these skills like Lego bricks without retraining the model. The paper describes how the generative model reacts to control inputs rather than playing back fixed clips.