Hijacking Robot Teams Through Adversarial Communication

This is the code for submitted paper Hijacking Robot Teams Through Adversarial Communication presented as an Oral talk at CoRL 2023.

Author: Zixuan Wu, Sean Ye, Byeolyi Han and Matthew Gombolay

The main code to train the adversarial policy is run_adv_comm_offpolicy function in main_heterogeneous.py. It will:

• Load your pre-trained agent policies (MADDPG ones can be trained with base_policy.py) in folder saved_models.

• Trains surrogate policies to mimic them and adversarial communication policies offline.

• Automatically creates a folder named models to save the trained adversarial policies as checkpoints.

The conda environment for running the code can be created using: conda env create -f environment.yml.

This work is still on-going and we will continue refining this repo - the next step includes to train a defender or mainipulate the attacked agents to anywhere we want.

This code is adapted from the Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments (MADDPG) code with the citation:

@article{lowe2017multi,
  title={Multi-agent actor-critic for mixed cooperative-competitive environments},
  author={Lowe, Ryan and Wu, Yi I and Tamar, Aviv and Harb, Jean and Pieter Abbeel, OpenAI and Mordatch, Igor},
  journal={Advances in neural information processing systems},
  volume={30},
  year={2017}
}

Please cite our paper with the following format if you find it help:

@inproceedings{wu2023hijacking,
  title={Hijacking Robot Teams Through Adversarial Communication},
  author={Wu, Zixuan and Ye, Sean Charles and Han, Byeolyi and Gombolay, Matthew},
  booktitle={7th Annual Conference on Robot Learning},
  year={2023}
}