Metatron Browser Demo

The Metatron demo runs a G1 policy and MuJoCo simulation in the browser. Its tasks include steering, location, path following, motion tracking, scene interaction and interactive IK.

Open the demo. No Python installation is needed. The first visit downloads the unified scene/recovery policy from Hugging Face and the remaining assets from the site. No account or token is needed. The manifest pins an immutable revision and SHA256; verified model bytes are cached when browser storage permits. Speed depends on browser and device; WebGPU is used where supported, with a WASM path for other devices. Mobile performance is not guaranteed by desktop emulation.

Distribution

The repository carries the compiled static bundle, not the application’s TypeScript project or source maps. GitHub Pages verifies and extracts that bundle beside the Sphinx documentation; it does not compile the application. Executable JavaScript, WASM and model assets must still be downloaded by the browser and remain inspectable.

Local preview instructions are in demos/metatron/README.md. The bundle checksum, file inventory and release status are recorded in demos/metatron/bundle.json.

Usage And Notices

The demo is a simulated research interface, not a hardware controller or safety system. Task completion, precise IK and recovery are not guaranteed for arbitrary goals or fallen states.

Code, third-party runtime libraries, robot assets, policy weights and motion assets retain their respective licenses and notices in the bundle. Demo motions are not licensed for reuse through the code or model license; consult the in-demo motion notice and the original rights holders.