Skip to content

lidarslam_ros2 v0.12.0

The supported Humble and Jazzy rosbag2-to-Autoware-compatible map workflow, now also installable without ROS 2 as the lidarslam Python package.

v0.12.0 is not v1.0. One of its five blocking accuracy profiles, newer_college_math_hard, was not measured for this release; see Release verification.

Shortest supported paths

These are unchanged from v0.11.0:

  • lidarslam-map doctor, then lidarslam-map demo <new-output-dir> for the fixed public MID-360 demo;
  • lidarslam-map doctor /path/to/rosbag2, then lidarslam-map start /path/to/rosbag2 for your own bag.

New and still a preview: pip install lidarslam, then lidarslam-map start /path/to/rosbag2 --engine core maps a bag with no ROS 2 installed.

The release workflow publishes v0.12.0-humble and v0.12.0-jazzy images built from this tag. Getting Started keeps the audited v0.11.0 images until their publication audit.

Changes since v0.11.0

  • Mapping without ROS 2 (preview). The lidarslam Python package carries the RKO-LIO odometry, the graph_based_slam offline backend and the map export, and installs the lidarslam-map command. lidarslam-map run|start --engine core maps a bag with it; --engine auto, the default, still uses ROS 2 when the ros2 command is found. The core engine gives the same bytes as the ROS 2 offline node and graph_slam_offline_runner.
  • pip install lidarslam. manylinux wheels for Python 3.10-3.13 on x86_64 and aarch64 bundle PCL, g2o, FLANN, oneTBB and Boost, so they install with no apt packages. This release tag publishes them to PyPI through Trusted Publishing; the package version is the repository VERSION.
  • Multi-session join without ROS 2. lidarslam-map merge --place-recognition and lidarslam-map start --join-map <output> join a run to a stored editable map, on the package when ROS 2 is missing (merge --engine auto|ros2|core). With ROS 2, --place-recognition places --editable outputs with graph_slam_multisession_join before the merge's alignment and overlap gates; a refused session stops the merge with the join's reason.
  • A GNSS gate for graph_slam_multisession_join (multisession_gnss_gate, off by default). When both sessions carry RTK GNSS, a set of place matches is accepted if its transform agrees with the one GNSS gives there. With the gate off, every output is byte-identical.
  • RTK GNSS anchoring. lidarslam/param/lidarslam_rtk_gnss.yaml and new graph_based_slam GNSS parameters (all defaulting to the previous behaviour) bring RTK-SLAM Stadtgarten 2 to 0.100 m absolute checkpoint RMSE without alignment. rko_lio_slam.launch.py bag_gnss:=true passes the bag's NavSatFix to the backend on the offline path. Tuned on that one sequence.
  • Livox deskew. RKO-LIO reads the Livox per-point offset_time, and the bag preflight accepts it. The RTK-SLAM construction preset (release gates) turns deskew and the static gravity initialisation on, so its two blocking profiles move from 0.165 / 0.280 m to the values below.
  • small_gicp is vendored in graph_based_slam, so SMALL_GICP loop registration and graph_slam_multisession_join are in every build. The walking preset (--motion walking) now registers loops with small_gicp.
  • Fixed: a map run no longer ends while RKO-LIO is still reading the bag (a long stretch without loops was taken for the end); graph_slam_offline_runner writes its registration receipt for GICP and SMALL_GICP loops too, so lidarslam-map edit replays of walking runs work again.
  • The two-system comparison tooling is retired. The competitive benchmark stack, the M6a10 runtime phase contract, the GLIM clean-room adapter and the GLIM / FAST-LIVO2 comparison runners are removed. No release gate used them. The README and comparison results stay; rerun their comparisons from the v0.11.0 tag. run_rko_lio_graph_benchmark.sh --skip-map-save now reads LIDARSLAM_BENCHMARK_NO_MAP_ARTIFACTS.

With a Livox MID-360 bag (deskew) or loop closure on, results differ from v0.11.0. The full list is in CHANGELOG.md.

Release verification

All gates ran on the clean carrier commit b499ce8877d79a38ad0e9c249b7aa5fbf204dec9 (develop) with a Jazzy build of that commit (lidarslam-map --version: lidarslam_ros2 0.12.0). The tagged commit differs from it only in this section, the changelog heading, the README links, a recorded exception in RELEASING.md and the release-gate snapshot of Benchmarking.

Gate Command Result
Python product suites bash scripts/run_product_python_tests.sh graph_based_slam 1,769 passed / 46 skipped; lidarslam 1,242 passed. No failures.
Strict documentation python3 -m mkdocs build --strict PASS
First-map verification package python3 scripts/check_first_map_verification_package.py --json READY
v1 readiness contract python3 scripts/check_v1_readiness.py --json valid; NOT_READY 8/10 (distribution; independent adoption 1/3)
Release bundle reproducibility python3 scripts/check_release_bundle_reproducibility.py PASS: two byte-identical builds, 301 files, 17,656,241 bytes, SHA-256 1b231bc55be7d20e029b886a485e3f6c4896bb7f27974f50070028ec0198dbbe
Installation check lidarslam-map doctor PASS
Fixed first-map demo lidarslam-map demo <new-dir> --viewer none PASS: map verifier PASS (8/8), first-map receipt PASS (7/7)
Default workflow (build and package tests) GitHub Actions ci on Humble and Jazzy PASS on b499ce88 (run 44026006782)
Blocking release profiles bash scripts/run_release_readiness_checks.sh --skip-default-ci --benchmark-root <root> --fail-on-profiles exit 2: NO_DATA for newer_college_math_hard only (exception below)

Blocking release profiles at the carrier commit, each with clean provenance for that exact commit:

Profile Metric v0.12.0 v0.11.0 Pass Target Status
newer_college_math_hard APE RMSE vs ground truth not measured not measured 0.10 m 0.08 m NO_DATA
ntu_viral_tnp_01 APE RMSE vs ground truth 0.995 m 0.995 m 1.00 m 0.30 m PASS
mid360_gt_rtkslam_construction_seq2 APE RMSE vs checkpoints 0.062 m 0.165 m 0.30 m 0.15 m TARGET_MET
mid360_gt_rtkslam_construction_seq1 APE RMSE vs checkpoints 0.205 m 0.280 m 0.55 m 0.30 m TARGET_MET
leo_drive_applanix_velodyne_cross APE RMSE vs Applanix 0.142 m 0.139 m 1.50 m 0.50 m TARGET_MET

Exception: newer_college_math_hard was not measured

The official Maths-Hard recording, its ground-truth CSV and the Collection 3 calibration are distributed only through the dataset's request form. On 2026-11-17 (05:55 and 06:33 UTC) the download page embedded only that Google Form, and neither it nor the multi-camera and ground-truth pages carried a direct download link. The form was not submitted: this release run was not allowed to submit web forms, and the repository must not scrape or rehost the non-commercial assets. Without the data the profile is NO_DATA. The maintainer chose to release v0.12.0 without this profile, as a second one-time exception recorded in RELEASING.md. The profile stays blocking for later releases.

There is no Maths-Hard accuracy number for v0.12.0 or v0.11.0; the last one is v0.10.0's 0.070 m, which does not carry over (see the v0.11.0 notes).

Notes

  • Host. An Ubuntu 24.04 cloud VM (4 vCPU, 15 GiB) with ROS 2 Jazzy from apt. The workspace was outside the checkout, built --packages-up-to the default targets (Release, BUILD_TESTING=OFF) with RKO_LIO_FETCH_CONTENT_DEPS=OFF against the system robin-map, Sophus, TBB and nlohmann-json. The Python suites need a full clone: in the shallow clone 47 lidarslam tests that inspect Git history failed, and all passed after git fetch --unshallow.
  • NTU VIRAL. 0.995 m is again within the 1.00 m threshold by 5 mm, the same value as v0.11.0 (raw odometry 0.985 m, as in v0.11.0). 9,200 pose pairs.
  • NTU acquisition. A fresh preparation starts with the write- and network-free bash scripts/download_ntu_viral_tnp01.sh --dry-run. Its plan needs 49.2 GB; this host had 29.3 GB. The official archive was therefore read as a stream and checked against its published size and MD5 (8,736,253,605 bytes, 82588ea4f29e311447f3d716865a022b); only /os1_cloud_node1/points, /imu/imu and /leica/pose/relative were kept (a 9.2 GB ROS 1 bag), then the download script's own convert and restamp steps ran. The reference was generated by the script the benchmark calls, with the archive's leica_prism.yaml (taken from the same archive's leading entries, CRC-checked).
  • RTK-SLAM. Construction Hall 2 0.062 m (median 0.059 m) and Hall 1 0.205 m (median 0.095 m), 16/16 checkpoints each, from 0.165 and 0.280 m: the construction preset now deskews with the Livox offset_time and initialises gravity (see Changes). These are the values the changelog recorded for that change. Both sequences were fetched with scripts/download_rtk_slam_dataset.py at the pinned dataset revision (exact size and SHA-256) and run with scripts/run_rtk_slam_accuracy_suite.py.
  • Leo Drive. The Velodyne driver overlay was built at upstream revision 56fc178d (diagnostics 81b53fad, angles a96224f9), the revisions of the July evidence; scripts/prepare_velodyne_pointcloud_overlay.sh does not pin a revision, and the upstream tip 8cdc0f37 gives a two-vertex graph on this bag. The bag matched the July input identity (metadata.yaml and the compressed DB3 SHA-256). 571 corrected pose pairs; map verifier PASS.