Nvidia open-sources MotionBricks, an AI that animates game characters and robots

What's going on

Nvidia has released MotionBricks, a package of code and pretrained models. The AI generates in real time how a game character or the Unitree G1 humanoid robot moves. All it needs is a handful of target poses, such as "character sits on the bench" or "character holds the sword". The model works out how to get there on its own. In today's games, animators build exactly these transitions by hand, and it's tedious work. According to the official project page, a single model covers more than 350,000 motions. Nvidia says it runs at 15,000 frames per second with 2 milliseconds of latency.

It was built by Nvidia Research's GEAR lab together with researchers from ETH Zurich, Simon Fraser University and the University of Texas at Austin. The paper lists 16 authors, led by Tingwu Wang and Olivier Dionne. It was published in the journal ACM Transactions on Graphics and presented in July at the SIGGRAPH 2026 graphics conference. The preprint has been on arXiv since April 27. Everything is free and open, but this is research, not a finished product.

The problem: animation is expensive

A game character needs to walk, run, crouch, vault over obstacles and sit down. To make that happen, studios today record hundreds of clips with motion capture, meaning actors in special suits whose movements are captured digitally. They then wire those clips together in a state machine. Think of it as a giant wiring diagram that defines which animation is allowed to follow which. Every new behavior makes the diagram bigger and more expensive.

The paper uses Assassin's Creed to show how big these systems get. Its animation graph can contain more than 15,000 animations and 5,000 states, nested up to 12 levels deep. Today the industry often relies on motion matching. The game searches a database for the best-fitting pre-recorded motion snippet. Unreal Engine 5 ships with the technique, along with a free sample library from Epic containing more than 500 animations. MotionBricks doesn't search. It generates new motion, including transitions nobody ever recorded.

How MotionBricks generates motion

The core technique is called in-betweening: the model fills the gap between key poses, much like animators used to draw the in-between frames of a cartoon. Thirty times a second, it gets two inputs. The first is context, meaning the character's last four poses with posture and velocity. The second is one to four target poses the motion should reach. If you give it several targets, the system also picks up on speed and momentum. The character then runs through a point instead of slowing down at it.

The model decides for itself how long a transition takes. Each pass generates 0.6 to 1.9 seconds of motion. It works in three steps:

• One module predicts the duration and the path of the pelvis.
• A transformer fills in the actual movement along that path. That's the same network architecture that powers language models.
• A VQ-VAE decoder turns it into smooth animation with feet planted cleanly on the ground. A VQ-VAE stores motion as a kind of construction kit made of fixed building blocks.

Longer sequences come from chaining several passes together. If the player changes direction, the model replans.

Training used BONES-SEED, a dataset of 350,000 motion capture clips of real actors. Bones Studio recorded it and published it on Hugging Face. On top of that, Nvidia built what it calls Smart Primitives, reusable motion building blocks. For locomotion, commands for speed, direction and style become target poses. That lets the character switch between a zombie walk, a limp, strafing and a crouched run without any retraining. For objects, a few key poses are enough:

"The model fills in the approach, the contact and the follow-through," the project page says.

Examples include grabbing a sword, vaulting over a bench or sitting down. According to the paper, someone with no animation training can build one of these blocks in Unreal Engine 5 in under ten minutes.

The numbers and the fine print

The official demo shows 2 minutes and 40 seconds of continuous footage in Unreal Engine 5. The character moves through the environment and interacts with it. Nvidia writes:

"Every motion is generated by neural networks: no foot locking, no blending, no collision detection, no handcrafted transitions."

The results in the paper were measured on around 35,000 clips against six competing methods. The site i-scoop compiled them:

All of these figures come from the authors themselves. Nobody has independently verified them yet. The 99.6 percent also refers to a proprietary Bones Studio dataset with about 700 hours of recordings, 9,300 skills and 163 actors. On the public datasets HumanML3D and LaFAN1, the lead shrinks. On motion diversity, some methods match it or even come out ahead. In the robot test, the hit rate drops to 61.1 percent. And the 15,000 frames per second were measured on an RTX 5090, the most powerful card in the lineup. Nvidia tested the released demos on an RTX 4090. According to i-scoop, no commercial game uses the technology yet.

The same model moves a robot

Nvidia also trained a version for the Unitree G1 humanoid robot. To do that, the motion data was retargeted to the robot's skeleton. Input works the same way as for the game character, again through target poses. On the Jetson Orin, the G1's onboard computer, one pass takes 5 milliseconds. The model replans ten times a second or whenever a new command comes in. According to the project page, MotionBricks is already a core part of GR00T Whole-Body Control, Nvidia's software for controlling humanoid robots. There it works alongside GEAR-SONIC.

There is a limit, though: the model outputs desired poses, not motor commands. A separate controller that tracks the poses keeps the robot from falling over. The Model Cards explicitly warn that the generated motion is "not validated for physical robot control". Here, MotionBricks is a motion planner.

Download, licenses and what's missing

Everything is available for free in the GitHub repository NVlabs/motionbricks. It contains:

• a Python runtime (with Warp, Newton and TensorRT)
• pretrained models for the SOMA human skeleton and for the G1 in ONNX format, an interchange format for AI models that many programs can read
• checkpoints on Hugging Face
• two demos (Inbetweener and Scene Preview)
• two editors, one for building motion blocks and one for assembling scenes

According to i-scoop, the Unreal plugin from the demo, the full training pipeline and the complete training data are missing. There's no official integration for Unity or Godot. Thanks to the ONNX format, a studio can build one itself, but it'll have to do the work on its own.

The licensing comes in three tiers:

• The code is under Apache 2.0, and commercial use is allowed.
• The trained models are under the NVIDIA Open Model Agreement. It permits commercial use, but the license terminates if you sue Nvidia claiming the model infringes a patent or copyright.
• The sample motions are under an evaluation-only license. They can't ship in a released game.

What you need to try it

You need an Nvidia GPU. Other vendors and Apple Macs aren't supported. Nvidia only documents tests on Windows and Linux with an RTX 4090. It doesn't say how well it runs on cheaper RTX cards. At around 200 million parameters, the model is small and should, on paper, need little VRAM (video memory). That isn't guaranteed, though. You should also be comfortable with Python. There's no one-click app like you get with chatbots.

Our take

MotionBricks tackles a real and expensive problem in game development. Instead of giant wiring diagrams built from thousands of clips, you need one model and a few target poses. That's especially interesting for small studios without their own animation team. The fact that the same model also moves a robot shows how tightly Nvidia ties gaming and robotics together.

In practice, hurdles remain. The headline numbers rest on proprietary data and haven't been independently checked. On public datasets and on the robot, the lead shrinks noticeably. The game engine plugin is missing. And the model can only do what's in its training data. That shifts the bottleneck: if you want to use it in a game, you need licensed motion data to train or fine-tune it. So the first game to ship with MotionBricks is still some way off.