BRAUM: Analyzing and Protecting Autonomous Machine Software Stack

This repository contains the code for error injection and analysis for Autoware, an open-source autonomous vehicle driving software, which helps us on figuring out the robustness of each individual nodes in Autoware. Based on the analysis results, we can provide selectively protection to ensure reliability while save protection overhead.

Installation

We use CARLA as the simulator and Autoware as the AV software instance. First, the user need to set up the CARLA and Autoware.

Install Autoware with docker

sudo apt install git-lfs
git clone --recurse-submodules https://github.com/carla-simulator/carla-autoware
cd carla-autoware && ./build.sh
./run.sh

After finishing the above steps, there should be a docker container running which includes the source code of Autoware.

Install CARLA with docker

sudo apt-get install x11-xserver-utils
xhost +
docker run --rm --name carla_server --network host --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=0 -e DISPLAY=${DISPLAY} carlasim/carla:0.9.10.1 /bin/bash CarlaUE4.sh -fps=20 -opengl

After finishing the above, the CARLA simulator version 0.9.10.1 is running.

Run a scenario in Autoware

The scripts for error injection are in /injection folder. To inject errors into a running instance of Autoware, we first create a scenario of a car starting from point A to point B.

In the docker container with Autoware source code, start Autoware at a fixed point.

export ROS_HOSTNAME=localhost
export ROS_MASTER_URI=http://localhost:11311
roslaunch carla_autoware_agent carla_autoware_agent.launch town:=Town01 spawn_point:="107,59,0.5,0,0,0"

In another terminal of the docker container, set up the finish point.

rostopic pub /move_base_simple/goal geometry_msgs/PoseStamped "{'header': {'seq': 0, 'stamp': {'secs': 0, 'nsecs': 0}, 'frame_id': "world" }, 'pose': {'position': {'x': 170, 'y': 0, 'z': 0.0}, 'orientation': {'x': 0.0, 'y': 0.0, 'z': -0.00720201027314, 'w': 0.999974065188}}}" --once

Error Injection

To collect results without error injection, use the script baseline.py inside /injection folder. First copy the script into the Autoware container.

docker cp baseline.py $CONTAINER_ID:/home/autoware/

Run the baseline.py script inside the Autoware container.

python baseline.py

Use other scripts in the /injection to inject errors into different nodes. For example, use the inject_trraw.py to inject errors into the output of twist_filter node

python inject_trraw.py -signal=xl -delta=0.5 -time=5.0 -iter=false

This means to inject an error into the signal xl for an amplitude of 0.5 at the time of 5.0 seconds during the scenario running, the error injection only happens once.

Error Analysis

After collecting the results from baseline and error injection, use the script inside /analysis folder. First we create a database from the baseline results, and then analysis the results using the baseline.

python create_dataset_frombaseline.py
python traceback_twist_raw.py

Publication

This work is published in ISSRE 2022. The citation information is updated later.

License

MIT