Learning Multi-Agent Policies using GNNs

Dependencies

• python==3.8
• gym==0.11.0
• numpy==1.20.0
• matplotlib==2.2.2
• PyTorch
• Gym-Flock

Available algorithms:

• Behavior Cloning python3 train.py cfg/cloning.cfg
• DAGGER imitation learning python3 train.py cfg/dagger.cfg

To test:

• python3 test_model.py cfg/dagger.cfg

Other code:

• python3 flocking_gym_test.py provides test code for the Gym Flock environments

Citing the Project

To cite this repository in publications:

@inproceedings{tolstaya2020learning,
  title={Learning decentralized controllers for robot swarms with graph neural networks},
  author={Tolstaya, Ekaterina and Gama, Fernando and Paulos, James and Pappas, George and Kumar, Vijay and Ribeiro, Alejandro},
  booktitle={Conference on Robot Learning},
  pages={671--682},
  year={2020}
}