VCAT: Vulnerability-aware and Curiosity-driven Adversarial Training for Enhancing Autonomous Vehicle Robustness

Pre-Installation

highway-env is required

Please install highway-env from here before this repo. and ensure running successfully.

Installation

Recommended system: Ubuntu 20.04 or 22.04

Step 1: Setup conda environment

conda create -n VCAT python=3.8
conda activate VCAT

Step 2: Clone this git repo in an appropriate folder

git clone [email protected]:caixxuan/VCAT.git

Step 3: Enter the repo root folder and install the packages:

cd VCAT
pip install -r requirements.txt
pip install -e .

Step 5: Replace the scenario files (6 files) in the highway-env origin file highway_env with the files in the provided highway_env folder. (The scenario file of highway_env usually is in the Home/.local/lib/python3.8/site-packages/highway_env)

Usage

Just use run the .py file! The attack training is going on!

Note: The defend training is not included in this repo. now! We will release it upon the paper the accepted.

python PPO.py

Note: env = gym.make("inverse6-env-cx-v0", render_mode='rgb_array') in which "inverse6-env-cx-v0" can change to "intersection2-env-cx-v0" etc. to change the difficulty of the environment.

Method	Description
MonteCarlo: MC.py	The initial state of attacker within a limited area is setting randomly.
Learning-to-Collide: LC.py	The concept of GAN is utilized to generate safety-critical data.
NormalizingFlow: NF.py	The normalizing flow generator is leveraged to create natural and adversarial safety-critical data.
ReinforcementLearning: PPO.py, DDPG.py. SAC.py, TD3.py	RL-based agent is employed to play the role of attacker.
Propsed: PPO_Proposed.py	The proposed method is utilized to train the attacker.

Attack Training Results

Snapshots in highway-env

The upper row is the normal training in which crashes occur, the lower row is the attack training in which the agent is more robust.

Citation

If you find this work useful in your research, please consider citing:

@article{xuancai2024vcat,
  title={VCAT: Vulnerability-aware and Curiosity-driven Adversarial Training for Enhancing Autonomous Vehicle Robustness},
  author={Xuan Cai, Zhiyong Cui*, Xuesong Bai, Ruimin Ke, Zhenshu Ma, Haiyang Yu, Yilong Ren*},
  journal={arXiv preprint arXiv:2304.02391},
  year={2024}
}