AutowareFoundation/carla-ue5-maps

CARLA UE5 Autoware maps

Autoware maps for the CARLA towns as they exist in CARLA 0.10, the release that re-authored the towns in Unreal Engine 5. One directory per CARLA world.

The point cloud maps published for Autoware alongside the CARLA towns were recorded on CARLA 0.9. The 0.10 towns are new geometry, so those maps no longer describe the world a vehicle drives in. The point clouds here were recorded from CARLA 0.10 itself.

Contents

autoware_maps/
└── Town10HD_Opt/
    ├── pointcloud_map.pcd
    ├── lanelet2_map.osm
    └── map_projector_info.yaml

File	Read by	What it is
pointcloud_map.pcd	Autoware	The town as a point cloud, recorded from CARLA 0.10
lanelet2_map.osm	Autoware	The road network in lanelet2 form
map_projector_info.yaml	Autoware	The origin of the map

The directory name is the CARLA world name, so it is also the value to pass as carla_map:=.

Worlds

Town10HD_Opt

CARLA's Town 10: an inner-city map, described upstream as "a mixture of features from inner-city environments with glossy skyscrapers and industrial buildings next to waterfront promenades, apartment blocks, hotels, public buildings and tree-lined boulevards". The road network is a ring around a grid of signalised intersections, which makes it short routes with a lot of turns rather than long straights.

lanelet2_map.osm	pointcloud_map.pcd
168 lanelets, 279 linestrings, 9,214 points. Drawn by simple_lanelet2, which also renders it in the browser — click it to open the file in this dataset in that viewer.	1,851,689 points over 368 x 357 x 31 m, voxel-downsampled at 0.2 m. Seen from above, coloured by height — the road surface is the dark ground, and what stands beside it is the buildings, the street trees along the boulevards and the roundabout at the top right.

The two are the same town in the two forms Autoware reads it in, drawn to the same frame. What the point cloud shows and the lanelet2 map does not is the reason the point cloud is here at all: this is Unreal Engine 5 geometry, not the 0.9 geometry the published map was recorded on.

Why the point cloud map is recorded again

On CARLA 0.10 with the published 0.9-era Town10HD point cloud, a run engages, drives 40-60 m and then emergency-stops:

ndt_scan_matcher: Score is below the threshold. Score: 1.99097, Threshold: 2
-> /autoware/localization/scan_matching_status WARN
-> system.mrm_handler: EMERGENCY_STOP is operated.

The match is not wrong. It is as good as that map allows: a correct match against 0.9 geometry scores only 2.0-2.3 on the 0.10 town, which sits on the threshold and dips under it while driving. The knock-on effects point elsewhere — a backward path from behavior planning, then planning trajectory is too far from ego in longitudinal direction from the planning validator — so the failure does not look like a map problem.

Against the point cloud in this dataset, NDT on Town10HD scores about 3.1 and a 110 m route completes with routing state ARRIVED and no emergency stop.

The lanelet2 map needs no such rebuild. The published one matches CARLA 0.10 lane centres within 0.5 m.

Download

hf download AutowareFoundation/carla-ue5-maps \
  --repo-type dataset \
  --include "autoware_maps/Town10HD_Opt/*" \
  --local-dir ~/autoware_data/maps

That writes ~/autoware_data/maps/autoware_maps/Town10HD_Opt. Drop --include to get every world.

Use

Pass the world directory as map_path, and the same name as carla_map:

ros2 launch autoware_launch e2e_simulator.launch.xml \
  map_path:=$HOME/autoware_data/maps/autoware_maps/Town10HD_Opt \
  vehicle_model:=sample_vehicle \
  sensor_model:=carla_sensor_kit \
  simulator_type:=carla \
  carla_map:=Town10HD_Opt \
  vehicle_type:=vehicle.lincoln.mkz \
  use_light_weight_sensor_mapping:=True

The autoware_carla_interface README covers the CARLA install, the build and the rest of the launch command.

The last two arguments are what CARLA 0.10 asks for on top of that. It ships a different vehicle catalogue, and the bridge's default vehicle.toyota.prius is no longer in it. And the default sensor mapping spawns six 1600x900 cameras, which the UE5 build renders far slower than 0.9 did; one front camera keeps the simulation about three times faster and is enough for the rule-based stack.

One more thing to expect: 0.10 vehicles do not pull away on the throttle the 0.9 pedal map creeps at. A run that engages and then sits still is a throttle floor problem, not a planning one.

Map origin

projector_type: Local

Autoware and CARLA share an origin here, so map_origin_x and map_origin_y stay at 0. The only difference between the two frames is that CARLA's y axis points the other way. Anything that converts between them negates y, and the recording below already did.

Lane changes need tagged lane boundaries

The published lanelet2 maps carry geometry only: their boundary ways have no type. Lanelet2's traffic rules read the shared boundary to decide whether a vehicle may cross it, and an untagged boundary is one it may not. The routing graph then holds adjacentLeft / adjacentRight relations but no left / right ones, Autoware's route handler finds no lane change target, and the ego drives to the end of its own lane whatever the route says.

The map here is the published one, untouched, so it has this property. Tagging the dividers between same-direction lanes as line_thin / dashed is what makes crossing legal. Tag only the lanes a scenario drives: these towns are built from thousands of very short lanelets, and with lane changes legal everywhere the behavior planner has been seen to spin without ever publishing a path.

Provenance

lanelet2_map.osm is a verbatim copy of maps/vector_maps/lanelet2/Town10HD.osm in carla-simulator/autoware-contents, with SHA-256 cdd0a28664bcf9b5bd8867584a6fa6a02fa0f75a7c3dbb901ca216128ef86a17. map_projector_info.yaml states the projector that file is drawn against.

pointcloud_map.pcd was recorded from a running CARLA 0.10.0 server. A LiDAR with no vehicle attached is placed at every lane sample of the world, 8 m apart and 2.0 m above the road, with an 80 m range. The returns are accumulated in the Autoware map frame and voxel-downsampled at 0.2 m. The recorder is docker/devcontainer/carla/build-pointcloud-map.py in autowarefoundation/autoware.

License status

This dataset is labelled CC-BY-4.0. That is a reading, not something the source states.

Both files describe CARLA's own town assets. The CARLA project states that "CARLA specific assets are distributed under the CC-BY License", so this dataset carries that license on. The statement is made at the project level, in carla-simulator/carla. The autoware-contents repository that lanelet2_map.osm is copied from carries no license notice of its own: it holds no LICENSE file, and its README does not mention one.

Autoware Foundation releases its work under Apache-2.0. We intend to open an issue with CARLA to ask whether these map files may be redistributed under Apache-2.0, and to ask for a license notice on autoware-contents either way. This section will be updated with the answer, and the label changed if the answer calls for it.

Please credit CARLA when you use these maps.