NVIDIA animates characters and robots like Legos

MotionBricks is an open-source real-time motion generation model from NVIDIA that makes animating digital characters and humanoid robots feel like snapping Lego bricks together.

Instead of manually tweaking transition frames, you feed the model your character's last four poses and up to four target poses. MotionBricks automatically generates smooth, natural motion to bridge the gap. Complex actions like vaulting an obstacle or sitting in a chair turn into modular building blocks that connect instantly, without any retraining.

How it works

Traditional character animation relies heavily on rigid state machines and labor-intensive inbetweening. MotionBricks splits that burden across three learned modules working in tandem.

First, a root module predicts the duration of the transition and the spatial path of the character's base. Next, a masked token transformer plans the motion tokens along that trajectory. Finally, a VQ-VAE decoder unpacks those tokens into fluid, frame-by-frame 3D movement.

Each step generates between 6 and 16 motion tokens, translating to roughly 0.6 to 1.9 seconds of movement at 30 Hz. The runtime simply loops the playback, records the latest four frames, and feeds them as context into the next cycle. That allows characters to transition smoothly from walking and running into crouching or styled modes like a zombie shuffle.

Performance is the real standout here. The generation pipeline can reach speeds up to 15,000 frames per second with latency as low as 2 milliseconds, making it fully capable of driving real-time game engines or live robotics controllers.

What is in the repo and how to run it

NVIDIA provided full code and pretrained checkpoints for two distinct skeletons: the SOMA human body model and the Unitree G1 humanoid robot. Both models weigh around 800MB and share the exact same runtime pipeline.

The GitHub repository includes four ready-to-run tools: an Inbetweener demo to test point-to-point transitions, a Scene Preview showing interactive traversal, and two dedicated editors for authoring custom action primitives and environment scenes.

To try it out, you will need an NVIDIA GPU running Windows or Linux x86-64. The authors tested the setup on a GeForce RTX 4090. Using uv and Git LFS, you can clone the repository, sync dependencies, and launch the built-in ONNX demos right away. The first run compiles a local TensorRT plan to cache for fast future startups.

The codebase is released under Apache 2.0, while the pretrained model weights on Hugging Face fall under the NVIDIA Open Model Agreement, which permits commercial use under specific terms.

Why it matters

For technical artists and game developers, MotionBricks eliminates the headache of creating brittle blend trees for every possible interaction. Setting up an environmental action is as simple as placing a few target poses around an object.

For robotics researchers, running the exact same model architecture on human rigs and the Unitree G1 demonstrates a practical path toward universal whole-body control. It acts as an ultra-fast local planner that translates high-level targets into valid kinematic trajectories.

We think NVIDIA made the right trade-off here. By focusing on fast modular inbetweening rather than a heavy end-to-end black box, MotionBricks delivers something nimble enough to integrate directly into existing production pipelines today.