Adversarially Robust Control

This is the repo for paper ARC: Adversarially Robust Control Policies for Autonomous Vehicles. The ARC framework trains control policies in an adversarial fashion for autonomous driving. The protagonist policy aims to drive safely, whilst the adversarial policy aims to cause collision. The policies are trained end-to-end on the same loss, in a GAN-like fashion, within a RL training framework. Multiple policies are used to create a more general and robust protagonist policy.

For further details see the paper: https://arxiv.org/abs/2107.04487

Installation

Clone the repo

git clone https://github.com/sampo-kuutti/adversarially-robust-control

install requirements:

pip install -r requirements.txt

Training the policies

To run ARC training run train_arc.py.

You can control the number of adversarial agents with the --num_advs argument.

Citing the Repo

If you find the code useful in your research or wish to cite it, please use the following BibTeX entry.

  title={ARC: Adversarially Robust Control Policies for Autonomous Vehicles},
  author={Kuutti, Sampo and Fallah, Saber and Bowden, Richard},
  booktitle={2021 IEEE International Intelligent Transportation Systems Conference (ITSC)},
  pages={522--529},
  year={2021},
  organization={IEEE}
}