[Update]: check out our newest GRF agent here: TiZero: Mastering Multi-Agent Football with Curriculum Learning and Self-Play

Learning-based agent for Google Research Football

Code accompanying the paper "TiKick: Towards Playing Multi-agent Football Full Games from Single-agent Demonstrations". [arxiv][videos]. The implementation in this repositorory is heavily based on https://github.com/marlbenchmark/on-policy.

Update:

• [22.8.11]: 11 vs 11 model is released! Model can be found on Google Drive.

pip install -r requirements.txt
pip install .

(a) First, you should download the trained model from Baidu Yun or Google Drive:

• pre-trained models can be found at:
  • Baidu Yun: Click to download Password：vz3a
  • Google Drive: Click to download

(b) Then, you should put the actor.pt under ./models/{scenario_name}/.

(c) Finally, you can go to the ./scripts/football folder and execute the evaluation script as below:

cd scripts/football
./evaluate.sh

Then the replay file will be saved into ./results/{scenario_name}/replay/.

• Hyper-parameters in the evaluation script:
  • --replay_save_dir : the replay file will be saved in this directory
  • --model_dir : pre-trained model should be placed under this directory
  • --n_eval_rollout_threads : number of parallel envs for evaluating rollout
  • --eval_num : number of total evaluation times

Once you obtain a replay file, you can convert it to a .avi file and watch the game. This can be easily done via:

cd scripts/football
python3 replay2video.py --replay_file ../../results/academy_3_vs_1_with_keeper/replay/your_path.dump

The video file will finally be saved to ./results/{scenario_name}/video/

Please cite our paper if you use our codes or our weights in your own work:

@misc{huang2021tikick,
    title={TiKick: Towards Playing Multi-agent Football Full Games from Single-agent Demonstrations},
    author={Shiyu Huang and Wenze Chen and Longfei Zhang and Ziyang Li and Fengming Zhu and Deheng Ye and Ting Chen and Jun Zhu},
    year={2021},
    eprint={2110.04507},
    archivePrefix={arXiv},
    primaryClass={cs.AI}
}