Making a Submissionยถ


This document will show you how to submit a test submission in 10 minutes.

๐Ÿš‚ All aboard!ยถ

In the next ten minutes, you will submit your first agent to the NeurIPS 2020 Flatland Challenge and will see your name on the leaderboard.


If you have any problem along the way, take a look at the troubleshooting tips at the end of this page. If things still donโ€™t go your way, donโ€™t hesitate to ask about it in the forum.

๐Ÿ“ฆ Setupยถ

Starter Kitยถ

Start by cloning the starter kit:

$ git clone
$ cd neurips2020-flatland-starter-kit

The starter kit comes with a sample agent which performs random actions. We will see how it works in more details in the last section. For now weโ€™ll just submit it as is to see how the process works.

Create the conda environmentยถ

The start kit uses the conda package manager. Install it if it is not setup on your machine.

You can now run the following:

$ conda env create -f environment.yml # creates the flatland-rl environment
$ conda activate flatland-rl # activates it


Flatland is tested with Python 3.6 and 3.7 on modern versions of macOS and Linux. We are unable to support Windows at this time. WSL is known to work but you may encounter problems with graphical rendering. Your contribution is welcome if you can help with this!

โœ… Test your local setupยถ

We will now run the agent locally to check that it works as expected.

Letโ€™s download the test environments. Head to the challenge resources and download the provided test set. Untar them in ./scratch/test-envs.

Your directory structure should be as follow:

โ””โ”€โ”€ test-envs
    โ”œโ”€โ”€ metadata.csv
    โ”œโ”€โ”€ Test_0
    โ”‚ย ย  โ”œโ”€โ”€ Level_0.pkl
    โ”‚ย ย  โ””โ”€โ”€ Level_1.pkl
    โ”œโ”€โ”€ Test_1
    โ”‚ย ย  โ”œโ”€โ”€ Level_0.pkl
    โ”‚ย ย  โ””โ”€โ”€ Level_1.pkl
    โ”œโ”€โ”€ Test_2
    โ”‚ย ย  โ”œโ”€โ”€ Level_0.pkl
    โ”‚ย ย  โ””โ”€โ”€ Level_1.pkl
    โ”œโ”€โ”€ Test_3
    โ”‚ย ย  โ”œโ”€โ”€ Level_0.pkl
    โ”‚ย ย  โ””โ”€โ”€ Level_1.pkl
    โ”œโ”€โ”€ ...

We will now replicate the setup used on AIcrowd on your local machine to ensure that your submission will be evaluated without problem when you submit it. This involves three components: your agent, the evaluator service, and a Redis server to let them communicate.


The communication between your agent and the evaluator is done through a Redis server. You should ensure that a redis server is running on localhost. Follow these instructions to set it up.

You can check that things are ready by running:

$ redis-cli ping

Evaluator serviceยถ

Letโ€™s start the evaluator service. You should use different terminals for the evaluator and the agent as they will need to run at the same time.

$ flatland-evaluator --tests ./scratch/test-envs/

The agentยถ

You can now start the agent:

$ export AICROWD_TESTS_FOLDER=./scratch/test-envs/
$ python

The agent should now start interacting with the evaluator, and you should see the results coming in:

$ flatland-evaluator --tests ./scratch/test-envs/
['Test_12/Level_0.pkl', 'Test_1/Level_1.pkl', 'Test_10/Level_1.pkl', 'Test_11/Level_0.pkl', 'Test_8/Level_0.pkl', 'Test_4/Level_1.pkl', 'Test_2/Level_1.pkl', 'Test_5/Level_0.pkl', 'Test_7/Level_0.pkl', 'Test_6/Level_1.pkl', 'Test_3/Level_1.pkl', 'Test_7/Level_1.pkl', 'Test_9/Level_1.pkl', 'Test_10/Level_0.pkl', 'Test_6/Level_0.pkl', 'Test_2/Level_0.pkl', 'Test_11/Level_1.pkl', 'Test_0/Level_1.pkl', 'Test_0/Level_0.pkl', 'Test_9/Level_0.pkl', 'Test_4/Level_0.pkl', 'Test_13/Level_0.pkl', 'Test_12/Level_1.pkl', 'Test_8/Level_1.pkl', 'Test_3/Level_0.pkl', 'Test_5/Level_1.pkl', 'Test_13/Level_1.pkl', 'Test_1/Level_0.pkl']
Listening at :  flatland-rl::FLATLAND_RL_SERVICE_ID::commands
Evaluating : Test_12/Level_0.pkl
Evaluating : Test_1/Level_1.pkl
Evaluating : Test_10/Level_1.pkl
Evaluating : Test_11/Level_0.pkl
Evaluating : Test_8/Level_0.pkl
Evaluating : Test_4/Level_1.pkl

You donโ€™t need to let the evaluation run until the end, since right now it is just using a random agent. The goal is simply to check that everything works as expected.

When you will start implementing your own agents, this will allow you to check that your solutions are fully working.

๐Ÿ—‚๏ธ Code structureยถ

There are two files that need to be present in your repository for the evaluation to work as intended: aicrowd.json to indicate which challenge you are taking part in, and which serves as the entrypoint of your solution.


Each repository must have an aicrowd.json file with the following content:

  "challenge_id": "neurips-2020-flatland-challenge",
  "grader_id": "neurips-2020-flatland-challenge",
  "tags": ["RL"],
  "debug": true

This is used to map your submission to the proper challenge. The starter kit repository includes a sample aicrowd.json file with the correct values.

You need to ensure that you set the proper tags with each submission. The tags indicate the methods you use in that submission, and need to be at least one of:

  • "RL" if you used reinforcement learning,

  • "OR" if you used operations research,

  • "other" if you used another method.


Different prizes are available depending on the method you use! Therefore it is important to fill this tag correctly. Winning solutions will be verified by the organizers to ensure the method was properly declared.

If you set debug to true, then the evaluation will run on a smaller set of 28 environments, and the logs from your submitted code (if it fails) will be made available to you to help you debug. These test submissions will appear at the bottom of the leaderboard (score of -1.0).


By default debug is set to true, so when you are ready to make a competitive submission, make sure to set debug to false!


The starter kit repository includes a sample file that you donโ€™t need to change. The default file calls the file, which is where you would usually implement your solution.

๐Ÿ“ค Submitting!ยถ

To submit to the challenge, you will use git tags. You are allowed to submit up to 5 submissions per day.

Create your repositoryยถ

Head to to create your private repository. You can use any name you want for it.

Push a tagยถ

You first need to add an SSH key to your GitLab account by following these instructions (replacing references of with If you do not have SSH Keys, you will first need to generate a pair.

You can then create a submission by pushing a tag to your repository.

First add a git remote pointing to your newly created repository:

$ # change the line below to use your AIcrowd username and repository name:
$ git remote add aicrowd<YOUR_AICROWD_USER_NAME>/<YOUR_REPO_NAME>.git
$ git push aicrowd master

Finally submit your solution by creating a tag for your submission and pushing it:

$ git tag submission-v0.1 # needs a new tag name for each submission!
$ git push aicrowd master
$ git push aicrowd submission-v0.1

Submission tags

Any tag push where the tag name begins with โ€œsubmission-โ€ to your private repository is considered as a submission! You are allowed up to 5 submissions per day.

Note that if the content of your repository does not change, then pushing a new tag will not trigger a new evaluation.

You should now be able to see the details of your submission at:<YOUR_AICROWD_USER_NAME>/<YOUR_REPO_NAME>/issues

You should start seeing something like this at the address above:

submission issue

Be patient, the evaluation will take some time! ๐Ÿ•™

๐Ÿš‰ Next stopsยถ

Take a look at the agent provided in the starter kit. It simply takes random actions for each agent at every timestep:

def my_controller(obs, number_of_agents):
    _action = {}
    for _idx in range(number_of_agents):
        _action[_idx] = np.random.randint(0, 5)
    return _action

Surely you can do better! ๐Ÿ’ช

Head over to the reinforcement learning in Flatland introduction to get started with simple RL methods such as Double DQN.

To go further, explore the research baselines which use RLlib to train using advanced algorithms such as Ape-X, PPO or imitation learning methods such as DQfD.

๐Ÿ› Troubleshootingยถ

โ€œenv_client.step() called before env_client.env_create() callโ€ยถ

This occurs if a previous local evaluation was interrupted. The client communicates with the evaluator service through Redis, and sometimes data in Redis can be left in an intermediate state that prevents a new evaluation from starting.


The commands that follow will delete all the data in your Redis database. Donโ€™t run them if you use this Redis database for other purposes!

To solve this problem, run the following commands:

$ redis-cli> FLUSHALL

If you often interrupt submissions, you can systematically cleanup the Redis database before starting the evaluator:

redis-cli -c "flushall"; flatland-evaluator --tests ./scratch/test-envs/

โ€œunknown locale: UTF-8โ€ยถ

This happens on macOS. Append this to your ~/.bash_profile:

$ export LC_ALL=en_US.UTF-8
$ export LANG=en_US.UTF-8

And then run:

$ source ~/.bash_profile

More details

โ€œactivate is not a conda commandโ€ยถ

This error can have various causes. Most commonly, this means that your conda installation is either too old, or misconfigured in some way. The easiest fix is to update conda to the latest version and re-install it if it keeps failing.