Skip to content
Fault Injection for Autonomous Vehicles
Branch: release
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
Dockerfile_ppc64le Fixing numpy verson to 1.15.4 to prevent crashing Jun 15, 2019
Dockerfile_x86 fixed paper title Jun 10, 2019 Added License and Readme. Repository cleanup May 10, 2019

AVFI: Fault Injection for Autonmous Vehicles using CARLA

We have provided the Dockerfile which will build an image with all the required packages and dependencies installed. Using this image is the easiest way to run the CARLA client with the Fault Injection campaign.

The self driving agent with the fault injector will run from within the docker container. The server (world simulator with the AV) has to be started separately before running the agent from the docker container.

Make sure both the server and the client are using CARLA 0.8.1

First start the Carla World Server
Example for Town01
On Windows: .\CarlaUE4.exe /Game/Maps/Town01 -carla-server -benchmark -fps=15 -windowed -ResX=800 -ResY=600 -carla-settings=.\Example.CarlaSettings.ini

The settings file can be found here:
Ensure the the 'UseNetworking' property is set to 'true'. Take note of the 'WorldPort' number. This port number should be used when starting the agent.

To start the fault injection campaign using the client, run python ./ --host [serverhostname] -p [serverport] -c [citymap]
Example for Town01
python ./ --host [serverhostname] -p 2000 -c Town01

The server and the client should have matching port and city parameters For more information on how to run the client and the server, check the and the CARLA manual online.

UIUC FI Codebase

  1. - This file sets up the experiments that the self-driving agent will execute alongside the fault injection campaign. It also logs the metrics from the AV and Fault Injector

  2. /agents/imitation/ - Given an input and output fault model, this file is reponsible for perturbing the input measurements from the AV sensors and the output controls from the self driving agent. Takes as argument objects from classes defined in input_fault_model and output_fault_model.

  3. /agents/imitation/ - Defines the CameraFaultModel, MeasureFaultModel and CommandFaultModel classes as input fault models.

  4. /agents/imitation/ - Defines Control fault models that perturb control information from the self driving agent to the AV.

  5. /agents/imitation/ - Original IL agent code has been modified to incorporate fault inejction through the FaultInjector class defined in


If you use AVFI, please cite the following papers.

  1. Avfi: Fault injection for autonomous vehicles.
    Jha, Saurabh, Subho S. Banerjee, James Cyriac, Zbigniew T. Kalbarczyk, and Ravishankar K. Iyer. DSN 2018 [PDF]
  title={AVFI: Fault Injection for Autonomous Vehicles},
  author={Jha, Saurabh and Banerjee, Subho S and Cyriac, James and Kalbarczyk, Zbigniew T and Iyer, Ravishankar K},
  booktitle={2018 48th Annual IEEE/IFIP International Conference on Dependable Systems and Networks Workshops (DSN-W)},
  1. End-to-end Driving via Conditional Imitation Learning.
    Codevilla, Felipe and Müller, Matthias and López, Antonio and Koltun, Vladlen and Dosovitskiy, Alexey. ICRA 2018 [PDF]
  title={End-to-end Driving via Conditional Imitation Learning},
  author={Codevilla, Felipe and M{\"u}ller, Matthias and L{\'o}pez,
Antonio and Koltun, Vladlen and Dosovitskiy, Alexey},
  booktitle={International Conference on Robotics and Automation (ICRA)},
You can’t perform that action at this time.