Skip to content

CARLA simulator support#

CARLA is an open-source simulator for autonomous driving research. Autoware connects to CARLA through the autoware_carla_interface package of Autoware Universe. This page shows what Autoware 2.0.0 supports, how the support grew, and what is still missing.

The autoware_carla_interface bridge sends CARLA sensor data, vehicle status, and the clock to Autoware, and sends the control commands, the turn indicator and hazard light commands, and the initial pose of Autoware back to CARLA.Autowaree2e_simulator.launch.xmlsimulator_type:=carla, sensor_model:=carla_sensor_kitBridgeautoware_carla_interface.launch.xmlautoware_carla_interfaceSpawns the ego vehiclePublishes sensors and statusTicks the simulation clockautoware_raw_vehicle_cmd_converterCalibrated for the CARLA Priusmulti_camera_combinerJoins six cameras for RVizactuationcommandcameraimages/sensing/*/vehicle/status/*/clock/control/command/turn_indicators_cmd/control/command/hazard_lights_cmd/initialpose/control/command/control_cmdCARLA servercarlasim/carla:0.9.16RPC port 2000Sensor dataand ego stateVehicle controlspectator_followChase camera, optional
The bridge sends CARLA sensor data, vehicle status, and the clock to Autoware, and sends the control commands of Autoware back to CARLA.

Status in Autoware 2.0.0#

PartWhereWhat it does
Bridgeautoware_carla_interface in autoware_universe 0.53.1Connects to the CARLA server, spawns the ego vehicle, publishes the sensor data and the vehicle status, and sends the control commands to CARLA. It runs autoware_raw_vehicle_cmd_converter with a calibration for the vehicle.toyota.prius blueprint.
Sensor kitcarla_sensor_kit_launch and carla_sensor_kit_description in autoware_launch 0.53.0Six cameras for 360-degree coverage, a top LiDAR, a GNSS receiver, and an IMU.
Launch filee2e_simulator.launch.xml in autoware_launch 0.53.0simulator_type:=carla includes the bridge. launch_simulator_interface:=false leaves it out, so that it can run in another process or container.
Docker demodocker/examples/demos/carla in autoware 2.0.0Runs CARLA 0.9.16 and Autoware in four Docker Compose services: carla-simulator, carla-interface, spectator-follow, and autoware. The Autoware services use the universe-cuda-jazzy image.
MapsThe demo_artifacts Ansible role in autoware 2.0.0Downloads the carla-kashiwanoha map for the Docker demo, and the Town10HD_Opt map for CARLA 0.10. The Docker demo of Autoware 2.0.0 still loads Town01, which you download yourself.
End-to-end modelautoware_tensorrt_vad in autoware_universe 0.53.1Runs the VAD model, which is trained on CARLA simulation data. See Autoware Architecture 2.0.

The README of autoware_carla_interface lists Ubuntu 22.04, ROS 2 Humble, and CARLA 0.9.15 as the supported environment. The Docker demo of Autoware 2.0.0 runs ROS 2 Jazzy with CARLA 0.9.16.

Run Autoware with CARLA#

Two ways are available:

  • Docker demo: Follow the README of the CARLA demo. You need an NVIDIA GPU with the NVIDIA Container Toolkit, Docker Compose v2, a CARLA Lanelet2 map such as Town01 in ~/autoware_data/maps/Town01, and the perception models in ~/autoware_data/ml_models. Then run these commands from the root of the autoware repository:

    xhost +local:docker
cd docker/examples/demos/carla
HOST_UID=$(id -u) HOST_GID=$(id -g) docker compose up

  • Source workspace: Install CARLA, the CARLA Python API, and the maps as the README of autoware_carla_interface explains. Start the CARLA server, then launch Autoware:

    ros2 launch autoware_launch e2e_simulator.launch.xml \
  map_path:=$HOME/autoware_data/maps/Town01 \
  vehicle_model:=sample_vehicle \
  sensor_model:=carla_sensor_kit \
  simulator_type:=carla

In RViz, set the initial pose with Init by GNSS, set a goal, wait for the route, and engage.

The CARLA simulator page also lists community projects that connect Autoware and CARLA.

Timeline#

  1. 2024 May: e2e_simulator.launch.xml gets an argument to launch the CARLA interface (autoware_launch#924).
  2. 2024 July: The autoware_carla_interface package is added to Autoware Universe (autoware_universe#6859).
  3. 2025 October: The CARLA sensor kit packages are added to autoware_launch (autoware_launch#1655).
  4. 2026 May: The launch_simulator_interface argument makes the bridge optional in e2e_simulator.launch.xml (autoware_launch#1836).
  5. 2026 August: The Docker Compose demo for CARLA is added (autoware#7264).
  6. 2026 September: The demo_artifacts role downloads the Kashiwanoha map (autoware#7301) and the CARLA 0.10 Town10HD_Opt map (autoware#7308).
  7. 2026 October: Autoware 2.0.0 is released with all of the above.

Next steps#

  • CARLA 0.10: CARLA 0.10 uses Unreal Engine 5. Autoware 2.0.0 does not run it yet. PyPI has no CARLA Python client after 0.9.16. The pedal map and the vehicle list of the bridge are for CARLA 0.9, and the default ego vehicle vehicle.toyota.prius does not exist in CARLA 0.10. The 0.10 maps of demo_artifacts wait for this support. For details, see Maps for CARLA 0.10.
  • Traffic lights: The CARLA Lanelet2 maps do not have traffic light elements for Autoware. Add them to the map, as the README of the bridge explains. Then you can set traffic_light.publish:=true, so that the bridge publishes the traffic light states of CARLA instead of camera recognition. The bridge matches each CARLA light to a traffic light element of the map, for example by position with traffic_light.map_path.
  • Closed-loop evaluation: autoware_carla_scenario is a scenario test framework for Autoware on CARLA. The design discussion CARLA × Autoware integration architecture for closed-loop evaluation proposes how the CARLA tools fit together. See also Assessment of safety and benchmarks.