wuji-mjlab installs through pixi, which resolves the full CUDA and PyTorch stack from a lockfile.
This project is pixi-only. conda + pip install -e . is not tested and not supported.
Requirements
- Linux x86_64
- NVIDIA GPU with CUDA 12.8 (Blackwell sm_120 / RTX 50-series supported)
- pixi ≥ 0.66 (the version CI uses)
For sim-to-real deployment, you also need Wuji Hand hardware and a camera rig. See Hardware Setup for the full bill of materials.
Install
This produces a default environment for training and evaluation. For the sim-to-real bridge, add the optional deploy environment:
Verify
This lists the registered tasks and confirms the mjlab and tyro stack imports cleanly.
Slow Dependency Downloads
If pixi install stalls while pulling the PyTorch wheel, add mirror entries to the existing [pypi-options] block in pixi.toml. Keep index-strategy and the [pypi-options.dependency-overrides] block, and point the indexes at a mirror:
This is a local edit. Don’t commit it — the CI lockfile check fails if pixi.toml and pixi.lock drift.
Development Setup
To contribute, install the pre-commit hooks after cloning:
Every git commit then runs ruff, codespell, and the YAML, TOML, and large-file checks locally, before CI sees the change. Run the full tree manually with pixi run pre-commit run --all-files.
Don’t pip install packages into the pixi environment. Pip dependencies aren’t tracked by pixi.toml or pixi.lock and disappear on the next resolve. Add dependencies by editing pixi.toml, then run pixi install.
