MotionBricks Offers Real-Time Character and Robot Animation via Modular Latent Models

MotionBricks is an open-source project featuring a modular latent generative model designed to animate characters and robots in real time. Based on a SIGGRAPH 2026 paper, the repository provides pretrained models, a Python runtime, interactive demos, and authoring editors. Users can verify the announcement directly through the official GitHub repository.

Core Architecture and Operation

The system operates by filling in motion between recent poses and target poses at 30 Hz. Three learned modules handle specific tasks:

• Root Module: Predicts the duration and path of the root.
• Pose Module: Employs a masked token transformer to predict motion tokens along the path.
• Decoder: Utilizes a VQ-VAE with codebook heads to convert tokens into continuous motion.

Longer animations are created by chaining generations together, where the runtime uses the last four recorded poses as context for the next sequence.

Included Tools and Components

The repository supplies pretrained models for the SOMA body model and the Unitree G1 robot as ONNX files, with checkpoints hosted on Hugging Face. Key components include:

• Demos: An Inbetweener demo for pose-to-pose generation and a Scene Preview demo showcasing continuous locomotion and interactions.
• Editors: Smart Primitive Editor and Scene Editor for authoring actions and environments.
• Runtime Requirements: TensorRT integration requiring an NVIDIA GPU. Testing was conducted on Windows and Linux (x86-64) systems equipped with an NVIDIA GeForce RTX 4090.

Practical Implications and Licensing

MotionBricks aims to assist developers and researchers working on humanoid motion, animation, and physical AI. Component licenses vary: the code is licensed under Apache 2.0, documentation under CC-BY-4.0, models under the NVIDIA Open Model Agreement, and sample motions under the NVIDIA Sample Data Evaluation License. The repository currently does not accept external contributions.