Skip to content

mjlab-to-mjswan

Give it a repository that defines mjlab tasks, local or on GitHub. A coding agent ports one task, with its trained checkpoints, into a browser app.

Tip

For mjlab's own tasks with checkpoints on W&B or the Hugging Face Hub, Builder.from_mjlab does this in one line of Python.

Usage

  1. Install the mjswan plugin.

  2. Run the skill on a GitHub URL or a local path:

    /mjswan:mjlab-to-mjswan https://github.com/mujocolab/g1_spinkick_example

  3. Pick a task, and answer the agent's questions.

  4. Open the app, from the repository root:

    python -m mjswan_app.main

Requirements

  • git and Python 3.10 to 3.13. The agent installs mjswan itself. No GPU needed.
  • Checkpoints: a W&B run, a Hugging Face repository, or model_*.pt files.

Warning

If the repository pins other mujoco or mjlab versions than mjswan, the agent stops and leaves the choice to you.

What the agent does

  1. Finds the repository's tasks, and asks you to pick one.
  2. Checks, in seconds, that mjswan supports the task's actions.
  3. Converts the checkpoints to ONNX.
  4. Writes mjswan_app/ and builds it, fixing what fails.
  5. Checks that the converted code computes the same numbers as mjlab (parity check).
  6. Opens a pull request to mjswan, only if mjswan lacks something other tasks need too.
  7. Reports what was ported and what was skipped.

Watch it before you share it

The parity check compares numbers, not behavior. mjlab trains with mujoco_warp, the browser runs MuJoCo's WebAssembly build, and a policy can behave differently.

Output

mjswan_app/
├── main.py      the build script
├── terms.py     rewritten terms, only if some failed to convert
├── README.md    how to run it
└── dist/        the built app

To share dist/, see Deployment or Publishing to Cloud.