Nvidia's MotionBricks animates game characters and a humanoid robot from one model trained on 350,000 clips

One neural model, 350,000 motion skills, 15,000 frames per second, no fine-tuning. Nvidia’s SIGGRAPH demo replaces a discipline’s worth of animation tooling with a single network.

What MotionBricks is

NVIDIA Research unveiled MotionBricks at SIGGRAPH 2026. It is a single universal character-animation controller trained on BONES-SEED, a dataset of 350,000 production motion-capture clips. From that training it generates any combination of locomotion, interaction and physics-driven movement in real time, at 15,000 FPS with 2 ms latency.

What it does not need is the point. No hand-crafted state machines. No manual transition graphs. No per-character fine-tuning. Those three items are a large share of what a game studio’s animation pipeline consists of, and MotionBricks is positioned as a replacement for all of them.

The demos

In the Unreal Engine 5 demonstration, a developer typed natural-language commands, “sit on the bench, stand up, pick up the sword, jump over the railing,” and MotionBricks generated all the intermediate motion live. The transitions between those actions, normally authored by hand or stitched through a transition graph, came out of the model.

The second demo is the one with wider implications. The same model drives the Unitree G1 humanoid robot through NVIDIA GR00T. Game animation and physical AI are running on one network: the motion that makes a character look natural on screen is the motion that moves a robot’s limbs.

What it displaces, and what remains open

Animation pipelines are the kind of infrastructure studios spend years building, with state machines and transition logic tuned per character. A single pretrained motion model that runs on a consumer GPU and takes text commands removes most of that work. Learned motion has been a research demo for years; the SIGGRAPH showing and the Unreal Engine 5 integration are an argument that it now ships.

The unanswered questions are the usual ones for a research unveiling: how it handles characters and rigs far from the training distribution, how much control animators retain over style, and what availability and licensing look like for studios. The humanoid demo raises a further one, which is how much of robot motion control will end up sharing infrastructure with game animation. On the evidence of this demo, Nvidia’s answer is most of it.