autoware_lanelet2_to_clipgt

Converts an Autoware Lanelet2 .osm map into the parquet bundles that simulation and world-model tooling reads:

• clipgt — the ClipGT HDMap layout consumed by alpasim, through trajdata.dataset_specific.mads.mads_utils.populate_vector_map.
• cosmos_transfer2_5 — an NVIDIA cosmos-transfer2.5 world-scenario scene, in the AV2-style {clip_id}.<element>.parquet layout.

Lanelet2 comes from simple-lanelet2, a drop-in reimplementation published as a plain wheel — so this installs from an index with no ROS, no Boost and no C++ toolchain.

pip install autoware-lanelet2-to-clipgt

Converting a map

The CLI is hydra-driven. A map needs an origin, which is where local (x, y, z) = (0, 0, 0) lands: output coordinates are metres in that local Cartesian frame, right-handed and Z-up.

# the bundled odaiba origin (MGRS 54SUE + offset)
python -m autoware_lanelet2_to_clipgt \
    map=odaiba input_map_path=map.osm output_dir=out/

# a cosmos-transfer2.5 scene instead
python -m autoware_lanelet2_to_clipgt \
    target=cosmos_transfer2_5 map=odaiba input_map_path=map.osm output_dir=out/

# an origin given inline
python -m autoware_lanelet2_to_clipgt \
    input_map_path=map.osm output_dir=out/ \
    map.mgrs_grid=54SUE map.offset.x=92008.5 map.offset.y=45335.1

An origin is written one of three ways in a map config under conf/map/ (conf/map/example.yaml has all three):

mgrs_grid: 54SUE815501            # an MGRS code, on its own

mgrs_grid: 54SUE                  # a grid square plus metres within it
offset: { x: 81655.73, y: 50137.43, z: 42.49998 }

lat_lon: { latitude: 35.6895, longitude: 139.6917, altitude: 42.5 }

tileset_json=path/to/tileset.json aligns the output to the scene frame of a 3D Tiles tileset instead: points are projected to ECEF and transformed by the inverse of the tileset's root transform, then re-expressed as ENU at the origin. That is what lets the map share a frame with a trajectory embedded in the same USDZ.

Looking at the result

python -m autoware_lanelet2_to_clipgt.visualize out/                    # clipgt
python -m autoware_lanelet2_to_clipgt.cosmos_transfer2_5.visualize out/ # cosmos

Both draw a top-down preview from the parquet files. The clipgt one also accepts a .usdz, which it extracts to a temp directory first.

As a library

import lanelet2
from autoware_lanelet2_to_clipgt import convert, origin_from_map_config

stats = convert("map.osm", "out/", lanelet2.io.Origin(35.6895, 139.6917, 0.0))
print(stats)

origin_from_map_config takes the same three shapes as the YAML above, from an OmegaConf node or a plain dict.

Development

uv sync
uv run pytest                  # most of the suite needs a fixture map (see below)
uv run python tools/smoke_test.py

The tests want a Lanelet2 map at tests/data/odaiba.osm, which is not version-controlled; without it the cases that need one skip themselves. The smoke test needs no fixture — it builds a small map with Lanelet2's own API, converts it through the CLI and checks the bundles, which is also what the release workflow runs against a freshly installed wheel.

Releases are cut by labelling a pull request release:major, release:minor or release:patch; merging it bumps the version, tags it and publishes to PyPI. See .github/workflows/release.yml.

License

Apache License 2.0.