Open source

Licenses

Package	License
soarm_sdk	MIT
imu_sdk	Apache-2.0
camera_calibration	Apache-2.0
soarm_lerobot	MIT
soarm_tamp	MIT
soarm_mjlab	MIT
SO-ARM100 (vendored)	see upstream repo
This site	CC BY-SA 4.0

Workspace conventions

soarm-ws is a thin umbrella repository over independently-versioned packages, each its own GitHub remote tracked as a git submodule — no shared build system forces them into lockstep. See Architecture for why that structure is deliberate, and Workspace Installation for the submodule clone/install workflow.

Agent skills

The workspace ships a set of Claude Code skills — one per submodule, a router (soarm-start), and a dedicated calibration skill — so a session can find the right context without re-deriving the workspace layout by hand. Two conventions worth knowing if you're contributing:

• Skills are mirrored file-for-file into .claude/skills/ and .github/skills/; a change to skills/ gets copied into both mirrors in the same commit.
• Skills are kept factual about state, including what's not implemented — e.g. soarm-lerobot's documented training stubs (see Imitation Learning → Training) and soarm-mjlab's unimplemented deploy phase (see Roadmap). A skill (or a doc) that overstates readiness is treated as worse than none.

Contributing

Each package is installed and tested independently — see that package's own README.md/CHANGELOG.md as the source of truth (this site links to them throughout rather than duplicating their content). Lint and test commands follow the same pattern per package, typically:

ruff check <package>       # lint
pytest                      # tests

soarm_mjlab additionally has make lint/make test/make check wrapping the same tools.