NVIDIA's real-time motion model (195M parameters) runs locally with ONNX Runtime Web. Give it the robot's last four poses and up to four target poses; it generates the motion in between, about once a second, while you steer.
The model downloads once and is cached by your browser. A desktop browser with 8 GB of RAM or more is recommended.
MotionBricks (Wang et al., SIGGRAPH 2026) is NVIDIA's modular latent generative model for real-time character and robot motion. This page is an independent browser port of its Unitree G1 runtime: the official ONNX model runs in a Web Worker with ONNX Runtime Web, and the locomotion, idle and Smart Primitive logic follows the reference Python runtime.
Because a browser generation takes ~50–250 ms instead of ~3 ms with TensorRT, each request predicts where the robot will be when its motion takes over and builds the context from the poses it will have shown by then.
Model: NVIDIA Open Model Agreement. Runtime code: Apache-2.0. Sample motion clips are streamed from NVIDIA's repository under the NVIDIA Sample Data Evaluation License and are not redistributed here. Robot meshes: Unitree Robotics (BSD-3-Clause).