Repository containing the code base for training the multi-agent coordination policy used in the paper "A Framework for Real-World Multi-Robot Systems Running Decentralized GNN-Based Policies".

Supplementary video material:

If you use any part of this code in your research, please cite our paper:

@inproceedings{blumenkamp2022decentralizedgnn,
  title={A Framework for Real-World Multi-Robot Systems Running Decentralized GNN-Based Policies},
  author={Blumenkamp, Jan and Morad, Steven and Gielis, Jennifer and Li, Qingbiao and Prorok, Amanda},
  booktitle={IEEE International Conference on Robotics and Automation},
  year={2022},
  organization={IEEE}
}

Clone with git clone --recursive to include submodules. Run ./build.sh to set up a docker container for training. Training can be performed without docker by installing all requirements in the host system according to the Dockerfile.

Start training by running

./run.sh python3 src/train.py

The training will stop automatically after 5000 training iterations.

Optionally, evaluate the policy performance by running

./run.sh python3 src/evaluate.py results/MultiPPO_simple_xxxx/checkpoint_yyyyyy

where the path to the model checkpoint has to be adapted accordingly.

For ROS2 inference (refer this ROS2 project)

./run.sh python3 src/export.py results/MultiPPO_simple_xxxx/checkpoint_yyyyyy

The exported torchscript model will be saved in the same checkpoint folder and can then manually be copied to the correct location in the ROS2 project.