Unofficial community guide · SIGGRAPH 2026

MotionBricks: NVIDIA's real-time motion model, explained end to end

MotionBricks (also written Motion Bricks) generates humanoid animation and robot motion live instead of playing back authored clips. This guide walks you from a working G1 demo to UE5 integration — every number linked to NVIDIA's own sources.

Install MotionBricks on your CUDA box Free code (Apache 2.0) · no account, no license key to read this site

15,000FPS
real-time generation throughput (official: 15000 FPS)
2ms
latency per official project page
350,000+clips
motion clips modeled by a single neural backbone

Many community posts repeat a “5,000 FPS” figure. It is wrong: the official project page and README both state 15,000 FPS / 15000 FPS. Cite the paper, not the hype thread.

What it actually does

MotionBricks animation teaser: a stylized character moving through varied terrain
Animation teaser — official GIF, hosted by NVIDIA. Source
MotionBricks robotics teaser: a Unitree G1 humanoid executing whole-body motion
Robotics teaser on the Unitree G1 — official GIF, hosted by NVIDIA. Source

Two ideas carry the whole design. A modular latent backbone models a 350,000-plus clip corpus in one model, so adding skills does not mean adding systems. Smart primitives — locomotion, objects, interaction — give you one command surface for navigation and object contact, as described in the paper.

Start here

Setup guide

Clone with Git LFS, pull the ~2.2 GB checkpoints, avoid the 1 KB pointer-file trap, run the demo. The page most people need.

Interactive G1 demo

All 15 keyboard controls (WASD plus 11 style keys), the camera tricks, and what the crawling-mode limitation means.

UE5 integration

How the official 2:40 uncut demo was authored, and what is public versus still unreleased for engine work.

Unity status

The honest answer on Unity: no official plugin today — here are the realistic bridge paths and their costs.

How it works

VQVAE tokenizer, pose and root models, what zero-shot means here, and where the training data comes from.

Style library

Zombie, skipping, strafing, crouch-strafing, injured walk — every released style with its keybind and source GIF.

Where it sits in NVIDIA's humanoid stack

MotionBricks is the motion-generation layer of the GR00T Whole-Body Control effort, next to GEAR-SONIC (robotic whole-body control) and Kimodo (offline generation). See the stack map for how the pieces differ.

GR00T WBC overview

The full stack: Decoupled WBC, GEAR-SONIC, MotionBricks — and the release timeline.

SONIC model selector

Default vs Low-latency vs v1.1 — which checkpoint for which job.

VLA workflow

Collect teleop data, fine-tune Isaac-GR00T N1.7, deploy with SONIC.

Teleoperation

VR, keyboard, gamepad, live camera — and the ZMQ streaming layer.

Motion representation

The 418-dim feature vector, Global vs Local subsets, decoded.

Troubleshooting

The master error table for the whole GR00T/SONIC stack.

Frequently asked questions

What is MotionBricks?
MotionBricks is NVIDIA's real-time generative motion framework, presented at SIGGRAPH 2026 (ACM Transactions on Graphics). It generates humanoid animation and robot control motion on the fly — 15,000 FPS of synthesis at 2 ms latency, trained on over 350,000 motion clips with a single neural backbone — instead of playing back pre-authored animation clips.
MotionBricks vs Motion Bricks — same thing?
Yes. The project is officially written MotionBricks (one word, camel case). Community posts also write motion bricks, motion-bricks, motionbrick, and occasionally nvidia motion brix — all refer to the same NVIDIA Labs release covered here.
Is this the official MotionBricks website?
No. This is an unofficial community guide. The official sources are the project page at nvlabs.github.io/motionbricks and the GR00T-WholeBodyControl GitHub repository; every claim here links to one of them.
Does MotionBricks need an animation blueprint or state machine?
No. The core claim of the paper is replacing hand-authored transition graphs with smart primitives: you command velocity, heading and style, and the single model composes the motion. The official demo states no foot-locking, no blending, no collision detection and no hand-authored transitions.
Can I use it for free in a commercial project?
The source code is Apache 2.0. Pretrained weights are under the NVIDIA Open Model License, which permits commercial use with attribution subject to its trustworthy-AI terms. This is an unofficial guide — verify the license texts in the official repository before shipping.