ReCAST

This is the repository for our paper "Revealing Safety Violations of Autonomous Driving Systems with Law-Compliant NPC Vehicles".

Dependencies

• Ubuntu 22.04 LTS
• Apollo 7.0
• Docker-CE version 19.03 and above

Usage

To reproduce the experimental results, users should follow the steps below:

Preparation

• Clone this project and create a virtual environment:
  conda create -n recast python=3.8
• Install the required packages:
  pip install -r npc_train/requirements.txt
• Download Apollo 7.0 from https://doi.org/10.5281/zenodo.7622089
• Create some necessary directories at the root directory of Apollo:
  mkdir data data/log data/bag data/core
• Start up Apollo container at the root directory of Apollo:
  ./docker/scripts/dev_start.sh -l
• Enter the container and build Apollo:
  ./docker/scripts/dev_into.sh
./apollo.sh build

Training the NPC Controller

• Set the training configuration parameters (e.g., number of surrounding ADS vehicles) in the npc_train/config.yaml file
• Train a style-specific controller on a designated road type. For example, to train an aggressive controller on a straight road:
  python npc_train/train/aggressive_straight.py

Generating the Critical Scenarios

• Set the searching configuration parameters (e.g., number of NPC vehicles) in the main_ga/config.yaml file
• Search the critical scenarios:
  python main_ga/main.py