MuJoCo WASM + Go2 ONNX Web

中文 | English

Overview

This project runs MuJoCo (WASM) simulations in the browser and performs online inference of a Go2 (quadruped) policy using ONNX Runtime Web. The repository also includes several test terrains (steps, slopes, tracks, etc.) to validate control and policy performance. See the live demo here

Quick Start (Local)

1. Install dependencies:

npm install

2. Build the frontend bundle:

npm run build

3. Start the local static server:

python server.py
# Visit http://localhost:8090

Note: To use ONNX Runtime Web multithreading / SharedArrayBuffer, the page must be served with COOP/COEP headers (already handled by server.py). Use a modern browser (latest Chrome/Edge/Firefox recommended).

How to Use

• After the page loads you will see the simulation area and the GUI (lil-gui).
• Choose different terrains from the Scene dropdown; the page will reload the selected scene.
• In AI Controls select a model (PPO / MOECTS) and click "Enable AI Control" to activate policy inference.
• Use keyboard WASD/QE or joystick to control the

Adding / Replacing ONNX Models or Scenes

• Add an ONNX model: put the model file in models/ (e.g. my_model.onnx), then add or replace an entry in src/main.js's modelConfigs:

modelConfigs: {
  "my_model": { url: './models/my_model.onnx', history: 3, stacking: 'frame' }
}

• Add a scene: place the scene and its assets in assets/scenes/go2/ and add the scene path to the Scene options in src/mujocoUtils.js.

Acknowledgment & Contributing

This repository references from zalo/mujoco_wasm

Contributions are welcome (issues/PRs) to add scenes, models, or improve examples.