DoppelTest for Autoware

Autoware implementation of DoppelTest for multi-vehicle scenario generation and execution.

Paper: https://dl.acm.org/doi/10.1109/ICSE48619.2023.00216

Repo Layout

• autoware_controller/: receiver API, sender, rosbag control, violation analysis
• autoware_launch/: Docker image and container launch helpers
• autoware_map/: Lanelet2 map directories used by Autoware experiments
• scenario_runner/: scenario generation and experiment orchestration
• start.sh: starts containers and receivers for the default multi-vehicle workflow

Requirements

• Linux
• Docker
• rocker
• NVIDIA GPU, drivers, and X11 support for the provided launch scripts

Python and rocker Setup

Install uv first. Official docs: https://docs.astral.sh/uv/getting-started/installation/

curl -LsSf https://astral.sh/uv/install.sh | sh

From the repo root, sync the project environment:

uv sync

This repo already vendors rocker under autoware_launch/rocker/. uv sync installs it from the local path, so do not clone it separately.

If you want to run on maps other than the default BorregasAve, install ROS 2 first using the official Humble installation guide:

• https://docs.ros.org/en/humble/Installation.html

Then source ROS 2 and install Lanelet2:

source /opt/ros/humble/setup.bash
sudo apt install ros-humble-lanelet2

lanelet2 is not managed by uv here. It comes from the ROS environment and is used for reliable Lanelet2 routing/regulatory parsing and by the violation analyzer.

Default Workflow

1. Build the multi-container image

bash autoware_launch/build/build_docker.sh --multi_container

2. Start containers and receivers

bash start.sh

This starts or reuses autoware_1 to autoware_5, assigns ROS_DOMAIN_ID=1..5, launches autoware_controller/receiver.py in the background, waits for /health, and prints the receiver URLs.

3. Run one GA iteration with 5 vehicles

uv run --script scenario_runner/test_main.py --num-vehicles 5 --generations 1

Meaning:

• --num-vehicles 5: each generated scenario uses 5 active Autoware vehicles
• --generations 1: run one GA generation

After experiments, remove the Autoware containers:

bash autoware_launch/scripts/kill_all.sh

By default, test_main.py auto-discovers running containers named autoware_1, autoware_2, ... through Docker and uses their current container IPs.

Default map:

• autoware_map/BorregasAve/lanelet2_map.osm

To run on sample-map-planning, add --map:

uv run --script scenario_runner/test_main.py \
  --num-vehicles 5 \
  --generations 1 \
  --map autoware_map/sample-map-planning/lanelet2_map.osm

test_main.py

Main entrypoint:

uv run --script scenario_runner/test_main.py

It orchestrates:

• Autoware startup and recovery
• sender restart with peer receiver URLs
• localization and route initialization
• autonomous mode switching
• pedestrian and traffic signal publishing
• rosbag control
• violation, decision, and fitness collection

Common examples:

Fixed vehicle count:

uv run --script scenario_runner/test_main.py \
  --num-vehicles 3 \
  --generations 10

Mixed-size scenarios:

uv run --script scenario_runner/test_main.py \
  --min-vehicles 2 \
  --max-vehicles 5 \
  --duration-hours 1

Explicit receiver URLs:

uv run --script scenario_runner/test_main.py \
  --url http://<vehicle1-ip>:5002 \
  --url http://<vehicle2-ip>:5002 \
  --url http://<vehicle3-ip>:5002 \
  --num-vehicles 3

Useful flags:

• --log-dir: output directory for scenario JSON files and GA logs
• --restart-wait: wait time after Autoware restart, default 60
• --max-recovery-retries: per-scenario recovery attempts before marking failure
• --conflict-only: generate only conflict scenarios
• --no-conflict-only: disable conflict-only generation

Receiver API

The receiver listens on 0.0.0.0:5002 inside each container.

Health check:

curl http://127.0.0.1:5002/health

Start Autoware:

curl -X POST http://127.0.0.1:5002/autoware/start \
  -H 'Content-Type: application/json' \
  -d '{}'

Start sender:

curl -X POST http://127.0.0.1:5002/sender/start \
  -H 'Content-Type: application/json' \
  -d '{}'

Start sender with explicit peer URLs:

curl -X POST http://127.0.0.1:5002/sender/start \
  -H 'Content-Type: application/json' \
  -d '{
    "receiver_urls": [
      "http://<peer1-ip>:5002/perception",
      "http://<peer2-ip>:5002/perception"
    ]
  }'

Runtime Parameters

Main parameters still relevant in the current code:

• ROS_DOMAIN_ID: per-container ROS 2 isolation; assigned automatically by start.sh, dev_start.sh, and dev_into.sh
• CONTAINER_COUNT: number of containers started by start.sh, default 5
• AUTO_START_RECEIVER: set 0 to make start.sh skip receiver startup
• AUTOWARE_RECEIVER_URLS: explicit receiver base URLs for test_main.py
• AUTOWARE_RECEIVER_CONTAINER_PREFIX: Docker auto-discovery prefix for test_main.py, default autoware_
• AUTOWARE_RECEIVER_PORT: receiver port for Docker auto-discovery, default 5002
• RECEIVER_LOG_ROOT: override receiver log directory

Removed:

• RECEIVER_INSTANCE

Outputs

test_main.py writes to the selected --log-dir:

• test_main.log
• Generation_XXXXX_Scenario_XXXXX.json
• GA_selection_gen_XXXXX.json

Receiver logs default to:

container_<ROS_DOMAIN_ID>/log/

Manual Control

Start one container:

bash autoware_launch/scripts/dev_start.sh --use_multi_container --container_name autoware_1

Enter a container shell:

bash autoware_launch/scripts/dev_into.sh --container_name autoware_1

Enter and start the receiver:

bash autoware_launch/scripts/dev_into.sh --container_name autoware_1 --start-receiver

For custom container names without numeric suffixes:

bash autoware_launch/scripts/dev_start.sh --use_multi_container --container_name my_av --ros-domain-id 7
bash autoware_launch/scripts/dev_into.sh --container_name my_av --ros-domain-id 7 --start-receiver

Common Mistakes

• Passing /perception URLs to test_main.py instead of receiver base URLs
• Running test_main.py from an environment that cannot access Docker when relying on auto-discovery
• Using a --map path that exists on the host but not inside the Autoware containers

Citation

If you use this repository in research, cite the DoppelTest paper and document any local changes to Autoware, maps, or analysis logic.