Contributing¶

Contributing to blendmpc¶

Thanks for considering a contribution! The project is young and small — this is a great time to shape it.

New blend patterns from the MPC×RL literature (src/blendmpc/blends/). A blend must: compose against MPCPolicy only (no solver-specific code), cite the paper(s) it implements in its module docstring, and ship with a unit test, a closed-loop behavioral test, and an example script.
New solver backends (src/blendmpc/solvers/). A backend implements MPCPolicy.solve() and must pass the existing blend test suite unchanged.
New env/model pairs (src/blendmpc/envs/). The OCP model should match the environment's dynamics exactly where possible, with a test proving it (see test_pendulum_model_matches_gym_dynamics).
Bug reports with a minimal script — behavioral bugs in control code are subtle, numbers beat descriptions.

git clone https://github.com/assawayut/blendmpc && cd blendmpc
pip install -e ".[crocoddyl,test]"
pre-commit install
pytest

Style is enforced by ruff via pre-commit; CI must be green.
Keep the core thin: solver features belong in solvers, RL algorithms belong in RL libraries. When in doubt, open an issue before writing code.
Every public function gets a docstring; blends document their math and cite their sources.