ttktjmt/mjswan

What it solves

It simplifies the process of sharing and demonstrating Reinforcement Learning (RL) simulations. Instead of requiring users to install complex physics engines or run heavy servers, it allows these simulations to run entirely within a web browser as static sites.

How it works

The framework integrates three core technologies to move the simulation to the client side: MuJoCo WASM for physics simulation, ONNX Runtime for executing AI policies, and Three.js for 3D rendering. This architecture ensures that all computation happens on the user's device, removing the need for a backend server.

Who it’s for

It is designed for RL researchers and developers who want to create interactive, portable demos of their robotic policies that can be easily hosted on platforms like GitHub Pages or embedded in Google Colab notebooks.

Highlights

• Client-side execution: Runs simulations and policy control in real-time without a server.
• WebXR support: Includes a native viewer for VR and AR with tracked hand interactions.
• Cross-platform: Compatible with both desktop and mobile browsers.
• mjlab-native: Quickly converts tasks from the mjlab library into web demos.
• Easy deployment: Can be exported as a static site for effortless distribution.