This is a community effort to solve the 2020 NeurIPS Flatland Challenge. The flatland environment is here and its Documentation Watch the introductory video here

Fork this repository and submit a pull request or contribute to any of the other community repositories. Don't be shy :)

• https://github.com/mitchellgoffpc/flatland-training
• Look at forks of this and other repositories to find more.

• Write boilerplate to interact with the environment. It is already a gym environment, so unclear if this is necessary, maybe code to visualize a run.
• Write simple baseline algorithms for pure RL and pure hard-coded heuristics.
• Write some documentation here about how to evaluate an algorithm, maybe we define some joint format for results so we can compare.

• DeepMind ACME has a lot of RL built-in.
• RLlib
• SEED

• Different solution approaches go into subfolders in src/approaches. Feel free to contribute your own or work on someone else's (maybe coordinate via discord).
• For easiness, there is a setup.py, so we can pip install -e . and don't have to use relative paths in the approaches, but let's not put requirements globally. Instead, make a README or requirements.txt in your approach folder (or utility submodule) if you rely on specific things. Not mandatory, most people will figure out what's missing by themselves.
• Shared utilities go in src/utils. As with the requirements, let's not have a global __init__.py for the utils (but __init__.pys are completely fine in submodules).
• Any code that interacts directly with the environment (wrappers, monkeypatches, visualization, etc.) can go into submodules of src/environment.

• Presentation on Graph representation for flatland
• Flatland RLlib baslines