AutowareFoundation/map-carla-kashiwanoha

Kashiwanoha test map

A map of a real site, the Kashiwanoha Campus area in Kashiwa, Chiba. It holds the same roads twice: as a lanelet2 map for Autoware, and as an OpenDRIVE for CARLA. CARLA has no map of this area, so it needs the road network in a form that it can build a world from.

Contents

File	Read by	What it is
lanelet2_map.osm	Autoware	The road network in lanelet2 form
map_projector_info.yaml	Autoware	The origin of the map
kashiwanoha.xodr	CARLA	The same road network in OpenDRIVE form
kashiwanoha_carla_demo.mp4	A person	A recording of the demo

The first three files are one set. map_projector_info.yaml states the origin that kashiwanoha.xodr was converted against. This origin is what lets Autoware and CARLA agree on where a thing is. If you change the origin on one side only, the two sides put the same road in two places.

projector_type: LocalCartesianUTM
vertical_datum: WGS84
map_origin:
  latitude: 35.9033135426554
  longitude: 139.93338978245356
  altitude: 0.0

This directory holds no point cloud map. A ground truth localization setup does not need one. pointcloud_map_loader logs PCD load failed and does not start. Routing, engage and driving still work.

This holds only while nothing downstream reads the point cloud map. A perception stack that reads it stalls instead. Its compare-map filters wait on a service that only pointcloud_map_loader offers. No occupancy grid is published, and the behavior path planner stops at waiting for occupancy_grid_map. If you need that path, add a point cloud map to this directory.

Download

The Autoware ansible role gets this map, without the video:

ansible-playbook autoware.dev_env.install_dev_env --tags demo_artifacts --ask-become-pass

The role writes the files to ~/autoware_data/maps/carla-kashiwanoha. To get the same result by hand, run:

hf download AutowareFoundation/map-carla-kashiwanoha \
  --repo-type dataset \
  --revision 0.2.0 \
  --exclude "*.mp4" \
  --local-dir ~/autoware_data/maps/carla-kashiwanoha

Build the CARLA world

CARLA cannot load this area by name. Build the world from the OpenDRIVE before you launch Autoware:

import carla

client = carla.Client("localhost", 2000)
client.set_timeout(300.0)
with open("kashiwanoha.xodr") as f:
    client.generate_opendrive_world(
        f.read(),
        carla.OpendriveGenerationParameters(
            vertex_distance=2.0,
            max_road_length=500.0,
            wall_height=0.0,
            additional_width=0.6,
            smooth_junctions=True,
            enable_pedestrian_navigation=False,
            enable_mesh_visibility=True,
        ),
    )

CAUTION: Keep wall_height at 0. CARLA raises a boundary wall for each road. Each one of the 190 roads in this network holds a single lane. A positive height therefore puts a wall between two adjacent lanes. A vehicle that touches such a wall flies into the air.

CARLA names the generated world OpenDriveMap. Add carla_map:=OpenDriveMap to the launch command, together with a map_path that points at this directory. The world holds geometry only: roads, sidewalks and markings. It has no buildings and no vegetation.

The autoware_carla_interface README covers everything else: the CARLA install, the build, and the launch command.

Provenance

lanelet2_map.osm is a copy of planning/autoware_diffusion_planner/test_map/lanelet2_map.osm in autoware_universe, with SHA-256 4fe358f2d1e182dc76de296e5619086f8d8ad2fcab3e6d8f99f6e3d724a84e6d.

tier4/autoware_lanelet2_to_opendrive v2.62.0 converted that file into kashiwanoha.xodr. The conversion targeted CARLA, used the origin above, and used left-hand traffic. Nobody edited the network by hand: every merge, move, delete and remove operation that the converter offers was left empty.

The converter is automatic, and the quality of its output still has room for improvement.

Both map files come from a file in autoware_universe, so they carry the same Apache-2.0 license.