Skip to main content

Multi Agent Reinforcement Learning on Trains

Project description


Test RunningTest Coverage


About Flatland

Flatland is a opensource toolkit for developing and comparing Multi Agent Reinforcement Learning algorithms in little (or ridiculously large !) gridworlds.

The base environment is a two-dimensional grid in which many agents can be placed, and each agent must solve one or more navigational tasks in the grid world. More details about the environment and the problem statement can be found in the official docs.

This library was developed by SBB, AIcrowd and numerous contributors and AIcrowd research fellows from the AIcrowd community.

This library was developed specifically for the Flatland Challenge in which we strongly encourage you to take part in.

NOTE This document is best viewed in the official documentation site at Flatland-RL Docs


Installation Prerequistes

  • Install Anaconda by following the instructions here.
  • Create a new conda environment:
$ conda create python=3.6 --name flatland-rl
$ conda activate flatland-rl
  • Install the necessary dependencies
$ conda install -c conda-forge cairosvg pycairo
$ conda install -c anaconda tk  

Install Flatland

Stable Release

To install flatland, run this command in your terminal:

$ pip install flatland-rl

This is the preferred method to install flatland, as it will always install the most recent stable release.

If you don't have pip_ installed, this Python installation guide_ can guide you through the process.

.. _pip: .. _Python installation guide:

From sources

The sources for flatland can be downloaded from gitlab

You can clone the public repository:

$ git clone

Once you have a copy of the source, you can install it with:

$ python install

Test installation

Test that the installation works

$ flatland-demo

Jupyter Canvas Widget

If you work with jupyter notebook you need to install the Jupyer Canvas Widget. To install the Jupyter Canvas Widget read also

Basic Usage

Basic usage of the RailEnv environment used by the Flatland Challenge (also see Example)

from flatland.envs.observations import GlobalObsForRailEnv
# First of all we import the Flatland rail environment
from flatland.envs.rail_env import RailEnv
from flatland.envs.rail_generators import sparse_rail_generator
from flatland.envs.schedule_generators import sparse_schedule_generator
# We also include a renderer because we want to visualize what is going on in the environment
from flatland.utils.rendertools import RenderTool, AgentRenderVariant

width = 100  # With of map
height = 100  # Height of map
nr_trains = 50  # Number of trains that have an assigned task in the env
cities_in_map = 20  # Number of cities where agents can start or end
seed = 14  # Random seed
grid_distribution_of_cities = False  # Type of city distribution, if False cities are randomly placed
max_rails_between_cities = 2  # Max number of tracks allowed between cities. This is number of entry point to a city
max_rail_in_cities = 6  # Max number of parallel tracks within a city, representing a realistic trainstation

rail_generator = sparse_rail_generator(max_num_cities=cities_in_map,

# The schedule generator can make very basic schedules with a start point, end point and a speed profile for each agent.
# The speed profiles can be adjusted directly as well as shown later on. We start by introducing a statistical
# distribution of speed profiles

# Different agent types (trains) with different speeds.
speed_ration_map = {1.: 0.25,  # Fast passenger train
                    1. / 2.: 0.25,  # Fast freight train
                    1. / 3.: 0.25,  # Slow commuter train
                    1. / 4.: 0.25}  # Slow freight train

# We can now initiate the schedule generator with the given speed profiles

schedule_generator = sparse_schedule_generator(speed_ration_map)

# We can furthermore pass stochastic data to the RailEnv constructor which will allow for stochastic malfunctions
# during an episode.

stochastic_data = {'prop_malfunction': 0.3,  # Percentage of defective agents
                   'malfunction_rate': 30,  # Rate of malfunction occurence
                   'min_duration': 3,  # Minimal duration of malfunction
                   'max_duration': 20  # Max duration of malfunction

# Custom observation builder without predictor
observation_builder = GlobalObsForRailEnv()

# Custom observation builder with predictor, uncomment line below if you want to try this one
# observation_builder = TreeObsForRailEnv(max_depth=2, predictor=ShortestPathPredictorForRailEnv())

# Construct the enviornment with the given observation, generataors, predictors, and stochastic data
env = RailEnv(width=width,
              remove_agents_at_target=True  # Removes agents at the end of their journey to make space for others

# Initiate the renderer
env_renderer = RenderTool(env, gl="PILSVG",
                          screen_height=1080,  # Adjust these parameters to fit your resolution
                          screen_width=1920)  # Adjust these parameters to fit your resolution

def my_controller():
    You are supposed to write this controller
    _action = {}
    for _idx in range(NUMBER_OF_AGENTS):
        _action[_idx] = np.random.randint(0, 5)
    return _action

for step in range(100):

    _action = my_controller()
    obs, all_rewards, done, info = env.step(_action)
    print("Rewards: {}, [done={}]".format( all_rewards, done))
    env_renderer.render_env(show=True, frames=False, show_observations=False)

and ideally you should see something along the lines of


Best of Luck !!



Please follow the Contribution Guidelines for more details on how you can successfully contribute to the project. We enthusiastically look forward to your contributions.



Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for flatland-rl, version 2.1.10
Filename, size File type Python version Upload date Hashes
Filename, size flatland-rl-2.1.10.tar.gz (2.5 MB) File type Source Python version None Upload date Hashes View hashes

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page