This repository contains an RL environment based on open-source game Gameplay Football.
It was created by the Google Brain team for research purposes.

Useful links:

• (NEW!) GRF Tournament - take part in the Tournament and become the new GRF Champion! Starting April 2020.
• GRF Game Server - challenge other researchers!
• Run in Colab - start training in less that 2 minutes.
• Google Research Football Paper
• GoogleAI blog post
• Google Research Football on Cloud
• Mailing List - please use it for communication with us (comments / suggestions / feature ideas)

For non-public matters that you'd like to discuss directly with the GRF team, please use [email protected].

We'd like to thank Bastiaan Konings Schuiling, who authored and open-sourced the original version of this game.

Open our example Colab, that will allow you to start training your model in less than 2 minutes.

This method doesn't support game rendering on screen - if you want to see the game running, please use the method below.

sudo apt-get install git cmake build-essential libgl1-mesa-dev libsdl2-dev \
libsdl2-image-dev libsdl2-ttf-dev libsdl2-gfx-dev libboost-all-dev \
libdirectfb-dev libst-dev mesa-utils xvfb x11vnc libsdl-sge-dev python3-pip

First install brew. It should automatically install Command Line Tools. Next install required packages:

brew install git python3 cmake sdl2 sdl2_image sdl2_ttf sdl2_gfx boost

and boost-python3, that supports Python 3.7:

brew install https://raw.githubusercontent.com/Homebrew/homebrew-core/527d96f84d4632d30b281ef5b26717e0a75edcb4/Formula/boost-python3.rb

git clone https://github.com/google-research/football.git
cd football

python3 -m venv football-env
source football-env/bin/activate

pip3 install .

This command can run for a couple of minutes, as it compiles the C++ environment in the background.

python3 -m gfootball.play_game --action_set=full

Make sure to check out the keyboard mappings. To quit the game press Ctrl+C in the terminal.

• Running experiments
• Playing the game
  • Keyboard mappings
  • Play vs built-in AI
  • Play vs pre-trained agent
  • Trained checkpoints
• Environment API
• Observations
• Scenarios
• Multi-agent support
• Running in docker
• Saving replays, logs, traces

In order to run TF training, install additional dependencies:

• Update PIP, so that tensorflow 1.15 is available: python3 -m pip install --upgrade pip
• TensorFlow: pip3 install "tensorflow==1.15" or pip3 install "tensorflow-gpu==1.15", depending on whether you want CPU or GPU version;
• Sonnet: pip3 install "dm-sonnet<2.0.0";
• OpenAI Baselines: pip3 install git+https://github.com/openai/baselines.git@master.

Then:

• To run example PPO experiment on academy_empty_goal scenario, run python3 -m gfootball.examples.run_ppo2 --level=academy_empty_goal_close
• To run on academy_pass_and_shoot_with_keeper scenario, run python3 -m gfootball.examples.run_ppo2 --level=academy_pass_and_shoot_with_keeper

In order to train with nice replays being saved, run python3 -m gfootball.examples.run_ppo2 --dump_full_episodes=True --render=True

In order to reproduce PPO results from the paper, please refer to:

• gfootball/examples/repro_checkpoint_easy.sh
• gfootball/examples/repro_scoring_easy.sh

Please note that playing the game is implemented through an environment, so human-controlled players use the same interface as the agents. One important implication is that there is a single action per 100 ms reported to the environment, which might cause a lag effect when playing.

The game defines following keyboard mapping (for the keyboard player type):

• ARROW UP - run to the top.
• ARROW DOWN - run to the bottom.
• ARROW LEFT - run to the left.
• ARROW RIGHT - run to the right.
• S - short pass in the attack mode, pressure in the defense mode.
• A - high pass in the attack mode, sliding in the defense mode.
• D - shot in the the attack mode, team pressure in the defense mode.
• W - long pass in the the attack mode, goalkeeper pressure in the defense mode.
• Q - switch the active player in the defense mode.
• C - dribble in the attack mode.
• E - sprint.

Run python3 -m gfootball.play_game --action_set=full. By default, it starts the base scenario and the left player is controlled by the keyboard. Different types of players are supported (gamepad, external bots, agents...). For possible options run python3 -m gfootball.play_game -helpfull.

In particular, one can play against agent trained with run_ppo2 script with the following command (notice no action_set flag, as PPO agent uses default action set): python3 -m gfootball.play_game --players "keyboard:left_players=1;ppo2_cnn:right_players=1,checkpoint=$YOUR_PATH"

We provide trained PPO checkpoints for the following scenarios:

• 11_vs_11_easy_stochastic,
• academy_run_to_score_with_keeper.

In order to see the checkpoints playing, run python3 -m gfootball.play_game --players "ppo2_cnn:left_players=1,policy=gfootball_impala_cnn,checkpoint=$CHECKPOINT" --level=$LEVEL, where $CHECKPOINT is the path to downloaded checkpoint.

In order to train against a checkpoint, you can pass 'extra_players' argument to create_environment function. For example extra_players='ppo2_cnn:right_players=1,policy=gfootball_impala_cnn,checkpoint=$CHECKPOINT'.

Solution: set environment variables for MESA driver, like this:

MESA_GL_VERSION_OVERRIDE=3.2 MESA_GLSL_VERSION_OVERRIDE=150 python3 -m gfootball.play_game