Grey humanoid robot walking on a snowy sidewalk in winter, no people, logos or readable text
AI-generated image

NVIDIA

NVIDIA just open-sourced the motion engine behind its humanoid stack, and it runs a Unitree G1 in real time

Making a humanoid robot move smoothly from one pose to the next is hard, and NVIDIA has just given its answer away. "NVIDIA open-sourced MotionBricks, a real-time motion model for animated characters and humanoid robots, trained on 313,526 motion clips," the TAAFT Research account posted on X on Saturday. "The same runtime drives a human body model called SOMA and the Unitree G1 humanoid robot."

What it does

You do not animate every step. "You place the poses you want a character to reach, and the model fills in the motion between them," TAAFT wrote. The model looks at the most recent poses and the targets, decides how long the move should take, then generates the motion. Chain the outputs together and a character can walk, run, vault, duck and sit in one flow.

NVIDIA's project page describes it as "a large-scale, real-time generative framework" that models "over 350,000 motion clips with a single model" and claims "15000 FPS and 2 ms latency". TAAFT's figures (313,526 clips and as little as 1.3 ms of GPU time per generation on an RTX 4090) differ slightly from the project page, so treat the exact numbers as NVIDIA's claims rather than independent tests.

Smart primitives, no retraining

The second idea is what NVIDIA calls smart primitives: reusable building blocks, such as a ledge to vault or a chair to sit on, that "snap together like bricks". The project page says MotionBricks "applies to new downstream tasks in a zero-shot manner", with "no fine-tuning or task-specific tagging required".

Why it matters for humanoids

This is not only a games tool. NVIDIA says MotionBricks "is now a core organic component of NVIDIA's GR00T Whole-Body Control effort, powering the motion-generation layer of the stack." The GitHub release includes the runtime with pretrained SOMA and Unitree G1 models, interactive demos and editors for building primitives and scenes. TAAFT says the code is Apache 2.0 and the models use NVIDIA's Open Model Agreement, which allows commercial use.

For robot makers building on the G1, that is a free, ready-to-run motion layer. The open question is how well motion generated in simulation holds up on real hardware, on real floors, outside the demo.

More on: Unitree · NVIDIA · China

← All news