This repository is for providing the basic data and code for our approach.

• FuzzScene

  • gif

    this subdirectory saves two gif pictures of simulation, the first one with blue arrow represents the predict angle of origin scene, and the second one with red angle represents the predict angle of scene after mutation, which shows on the README file

  • code

    this subdirectory includes code, basic data, trained models and some scripts for demonstrating our experiments.

    • the main folder saves the main code of our algorithms
    • the env folder saves two anaconda env used in experiments
    • the sampling_vgg folder saves the VGG-Net models for our sampling algorithm
    • the seed_pool folder saves the six basic Open Scenario scenes
    • the trained_models folder saves the four trained driving models
    • the Violated images folder saves the error images found by our algorithm, divided by four models
    • the scenario_runner folder saves the scenario_runner files which helps run the simulation

  • README.md

• Python Env

  • we use Anaconda for switching to different experimental environments, because of the difference between Carla running environment (python 3.7) and the Driving Model environment (python 3.6)
  • we named the Carla anaconda environment carla and Driving Model environment dave
  • two anaconda env have been saved in the env folder

• Carla & Scenario Runner Env

  • in order to conduct our experiment, we use Carla Simulator-0.9.13 as our autonomous driving simulator, which can be found at https://github.com/carla-simulator/carla
  • in order to build simulation scenarios, we use Scenario Runner-0.9.13 which can be found at https://github.com/carla-simulator/scenario_runner. The files has already been included in our code

Because the trained model files in our code is too large, we provide Google Drive link to obtain these files in path FuzzScene/code/sampling_vgg/ and FuzzScene/code/trained_models/

• You can simply get these files through the link below
  • https://drive.google.com/drive/folders/18LgyPcKJOIQ-_ujr9vnNLoH0hW4I9RWo?usp=sharing
• please make sure the files are in the right path consistent with the Google Drive

We have made three scripts for conducting our experiments.

First, change the anaconda path code in FuzzScene/code/main/ga_sim.sh to your anaconda path, and make sure the CARLA_0.9.13 folder and the FuzzScene folder are in the same folder

then change the current anaconda environment to dave and get into the main folder.

conda activate dave
cd FuzzScene/code/main

Then you can run the first script to set basic parameter for experiments.

python set_para.py 1 1 1

The first argument value can be '1' to '4', which represents the four autonomous driving models. The second argument value can be '0' or '1', which represents whether conduct the sampling step of the genetic algorithm or not. The last argument value can be '1' to '3', which represents the three kinds of fitness function of the genetic algorithm. The error image found by our algorithm will be moved to the Violated images folder

The second script is for the main function of our approach, which includes initialization, all steps of our algorithm and the data collection.

python fuzz_ga.py

The Third script is for moving data collected before to specified location and clear unnecessary data.

python rename.py

Data will be moved to the folder ga_output, including the file of errors found by algorithm, which will be renamed by model name, whether conduct sampling step, and the fitness function been chose, the entropy file in entropy folder, and the r_list file for details of experiments in r_list folder