Skip to content

workflows.md

Latest commit

 

History

History
741 lines (565 loc) · 28.4 KB

workflows.md

File metadata and controls

741 lines (565 loc) · 28.4 KB
Raw

Operator Workflows

This page keeps the procedural details that do not need to stay in the top-level README.

Build Prerequisites

  • scanmatcher depends on ndt_omp_ros2
  • clone with submodules:
mkdir -p ~/ros2_ws/src
cd ~/ros2_ws/src
git clone --recursive https://github.com/rsasaki0109/lidar_slam_ros2
cd ..
bash src/lidar_slam_ros2/scripts/install_source_dependencies.sh
  • build and run the default checks:
colcon build --symlink-install --cmake-args -DCMAKE_BUILD_TYPE=Release
bash scripts/run_default_ci_checks.sh

Optional 3D-BBS support:

  • Thirdparty/3d_bbs is a small MIT-licensed vendor tree with COLCON_IGNORE.
  • graph_based_slam builds its CPU 3D-BBS sources automatically when GRAPH_BASED_SLAM_ENABLE_3D_BBS=ON and the vendor headers are present.
  • Runtime use is still off by default; enable it with use_3d_bbs_for_scan_context: true in the graph parameter YAML or with the MID360 benchmark wrapper option shown in the benchmarking docs.
  • To force-disable the optional build, pass --cmake-args -DGRAPH_BASED_SLAM_ENABLE_3D_BBS=OFF.

Main Entry Points

Goal Entrypoint
Fixed public first map lidarslam-map demo; add --viewer none for headless use or --dry-run --json for a network- and write-free plan. Add --output PLAN to retain that plan once without shell redirection.
Autoware pointcloud-map quickstart bash scripts/run_autoware_quickstart.sh
Full dogfood flow bash scripts/run_rko_lio_graph_autoware_dogfood.sh --auto-exit-secs 20
Standard NTU VIRAL benchmark bash scripts/run_rko_lio_graph_benchmark.sh
KITTI Odometry small_gicp evaluation bash scripts/run_kitti_odometry_benchmark.sh --sequence 00 --small-gicp --force-prepare
KITTI Odometry small_gicp sweep bash scripts/sweep_kitti_small_gicp.sh --dataset "$KITTI_ODOMETRY_ROOT" --sequences "00 05 07"
localization_zoo PCD/trajectory → fixed graph bag python3 scripts/pcd_sequence_to_rosbag2.py --help then bash scripts/run_offline_determinism_check.sh
MID360 cross-validation benchmark bash scripts/run_rko_lio_mid360_crossval_benchmark.sh
Offline browser 3D map preview lidarslam-map view output/my_map writes and opens a self-contained HTML preview; use --no-open on headless hosts. The lower-level export_mid360_robot_3d_map_preview.py remains available for custom sampling limits.
Recent map-session history lidarslam-map sessions validates direct child session bundles and opens a local newest-first catalog; use --status action_required, --viewer none, or read-only --json as needed.
Evidence-backed session comparison Select two cards in sessions.html, or run lidarslam-map compare output/session-a output/session-b; use --viewer none or read-only --json on headless/automated hosts.
Privacy-first maintainer report Choose Get support in sessions.html, or run lidarslam-map support output/session-a; review the fixed three-member ZIP before attaching it to a public issue.
Non-destructive 3D map cleanup Create an RKO graph map with run --editable, select unwanted boxes or accepted loops in view, then run the browser-printed edit command. Replay inputs are auto-detected; edit --help-all provides overrides for older outputs.
Multi-session map project Put the trusted anchor first: lidarslam-map merge output/day1 output/day2 --output-dir output/site_project, then inspect separately colored session paths with lidarslam-map view output/site_project.
Mixed-quality open-data GNSS smoke bash scripts/run_open_data_applanix_velodyne_gnss_smoke.sh --bag /path/to/rosbag2 --applanix-msg-dir /tmp/applanix/applanix_msgs/msg --verify-map
Mixed-quality open-data GNSS benchmark bash scripts/run_open_data_applanix_velodyne_gnss_benchmark.sh --bag /path/to/rosbag2 --applanix-msg-dir /tmp/applanix/applanix_msgs/msg --verify-map
Leo Drive classic-path suite bash scripts/run_open_data_classic_path_benchmark_suite.sh --applanix-msg-dir /tmp/applanix/applanix_msgs/msg --verify-map
Packet IMU deskew validation matrix bash scripts/run_open_data_packet_imu_deskew_validation_matrix.sh --applanix-msg-dir /tmp/applanix/applanix_msgs/msg
Dynamic-object-filter save-map benchmark bash scripts/run_dynamic_object_filter_benchmark.sh
MID360 place-recognition comparison bash scripts/run_place_recognition_benchmark.sh
Pre-tag reproducible release bundle python3 scripts/check_release_bundle_reproducibility.py /tmp/lidarslam_ros2_release_candidate.tar.gz
Release/readiness gate bash scripts/run_release_readiness_checks.sh --fail-on-profiles

Required Input Topics

Public default path: RKO-LIO + graph_based_slam

Launch:

ros2 launch lidarslam rko_lio_slam.launch.py \
  bag_path:=/path/to/rosbag2 \
  lidar_topic:=/os_cloud_node/points \
  imu_topic:=/os_cloud_node/imu

Required inputs:

  • lidar_topic: sensor_msgs/msg/PointCloud2
  • imu_topic: sensor_msgs/msg/Imu

Optional inputs:

  • /gnss/fix: sensor_msgs/msg/NavSatFix when graph_based_slam use_gnss:=true

Internal wiring in this launch:

  • RKO-LIO publishes odometry on /rko_lio/odometry
  • RKO-LIO publishes submap source clouds on /rko_lio/frame
  • graph_based_slam consumes those via odom_input and cloud_input

Not currently supported in the public path:

  • wheel odometry / vehicle speed topic fusion

GNSS note:

  • GNSS is added as translation-only pose-graph constraints in the backend
  • when covariance is present, edge weight is scaled from position_covariance
  • NavSatFix does not standardize RTK fix status, so graph_based_slam treats low horizontal covariance as RTK-like
  • default threshold: gnss_rtk_fix_max_horizontal_stddev_m = 0.3

Classic path: scanmatcher + graph_based_slam

Launch:

ros2 launch lidarslam lidarslam.launch.py \
  input_cloud:=/points_raw \
  imu_topic:=/imu

Required inputs:

  • input_cloud: sensor_msgs/msg/PointCloud2
  • TF from robot_frame_id to the LiDAR frame

Optional inputs:

  • imu_topic: sensor_msgs/msg/Imu when scanmatcher use_imu:=true
  • odom TF into odom_frame_id when scanmatcher use_odom:=true
  • /gnss/fix: sensor_msgs/msg/NavSatFix when backend use_gnss:=true

Internal wiring in this launch:

  • scanmatcher publishes lidarslam_msgs/msg/MapArray on map_array
  • graph_based_slam subscribes to map_array

Voxel-grid safety:

  • every classic scanmatcher PCL VoxelGrid call is preflighted against PCL's signed 32-bit index/layout limit;
  • a VOXEL_GRID_* warning rejects only the named stage instead of passing an unfiltered cloud downstream or terminating the node;
  • use vg_size_for_input for input_scan, registration_target, and recovery_target warnings;
  • use vg_size_for_map for initial_map and map_update warnings;
  • inspect coordinate units and outliers before increasing a leaf size. The node never changes map resolution automatically.

See the VoxelGrid refusal contract for every reason code, preserved state, and the bounded issue #69 regression.

Adapting another PointCloud2 LiDAR

Use this checklist when adapting another LiDAR that publishes sensor_msgs/msg/PointCloud2. It establishes readiness for one controlled first run; it does not validate accuracy, make a vendor part of the supported matrix, or select universal tuning values. Run the fixed public demo first so that an installation problem is not confused with a sensor-adaptation problem.

Replace every <PLACEHOLDER> below with an observed value before running a command. If a value is unknown, stop at that check instead of guessing it.

  1. Confirm the topic and message contract.

    ros2 topic list -t
ros2 topic type <POINTCLOUD_TOPIC>
ros2 topic echo --once --field header.frame_id <POINTCLOUD_TOPIC>
ros2 topic echo --once --field fields <POINTCLOUD_TOPIC>

    Expected: <POINTCLOUD_TOPIC> is listed as sensor_msgs/msg/PointCloud2, header.frame_id is non-empty, and the fields output contains FLOAT32 x, y, and z. For the RKO-LIO path, also identify a supported per-point time field named t, timestamp, time, or stamps; a header timestamp alone does not satisfy that path. If any required field is absent, fix the driver or use a conversion layer before launching SLAM.

  2. Check timestamp order and rate.

    For a rosbag2 input, run the product preflight first:

    lidarslam-map doctor /path/to/rosbag2 --json

    Review the selected topic's timestamp findings and keep sampled distinct from a full-bag proof. For a live topic, observe both a timestamp and the publication rate:

    timeout 5s ros2 topic echo --once --field header.stamp <POINTCLOUD_TOPIC>
ros2 topic hz --window 20 <POINTCLOUD_TOPIC>

    Expected: timestamps advance and the rate remains positive. Repair the publisher clock, rosbag playback clock, or timestamp conversion when they do not. Do not hide timestamp warnings by increasing a timeout.

  3. Measure the frame relationship; never invent an extrinsic.

    Use the non-empty frame observed in check 1 as <LIDAR_FRAME> and verify the directed transform to the robot base while the source is live or being played:

    ros2 run tf2_ros tf2_echo <BASE_FRAME> <LIDAR_FRAME>

    Expected: repeated transforms in the same parent-to-child direction as the configured launch. If the path is missing or the measured translation or rotation is unknown, stop and repair the broadcaster or calibration. An identity transform is valid only when it is the measured mounting relationship; guessing an extrinsic can produce a plausible but invalid map.

  4. Record the sensor period and valid range in a profile.

    The classic path uses these fields in its main_param_dir YAML:

    scan_matcher:
  ros__parameters:
    scan_period: <SECONDS_PER_SCAN>
    scan_min_range: <MIN_RANGE_M>
    scan_max_range: <MAX_RANGE_M>

    The RKO-LIO path uses min_range and max_range launch arguments or its rko_param_file; its per-point timestamps determine the scan timing. Set the values from the sensor specification or a bounded measurement, and record the source in the profile. Do not copy a value from another vendor merely because the topic type matches.

  5. Run one explicit, reviewable launch.

    For the classic path, the public remap and frame arguments are:

    ros2 launch lidarslam lidarslam.launch.py \
  input_cloud:=<POINTCLOUD_TOPIC> \
  imu_topic:=<IMU_TOPIC> \
  robot_frame_id:=<BASE_FRAME> \
  base_frame:=<BASE_FRAME> \
  lidar_frame:=<LIDAR_FRAME> \
  main_param_dir:=/path/to/custom-lidarslam.yaml \
  publish_static_tf:=false

    Set publish_static_tf:=true only when the seven static-transform values (static_tf_x, static_tf_y, static_tf_z, static_tf_qx, static_tf_qy, static_tf_qz, and static_tf_qw) come from the measured calibration. For RKO-LIO, use the corresponding public arguments and put the measured extrinsics in rko_param_file:

    ros2 launch lidarslam rko_lio_slam.launch.py \
  bag_path:=/path/to/rosbag2 \
  lidar_topic:=<POINTCLOUD_TOPIC> \
  imu_topic:=<IMU_TOPIC> \
  base_frame:=<BASE_FRAME> \
  lidar_frame:=<LIDAR_FRAME> \
  rko_param_file:=/path/to/measured-rko.yaml

    A controlled first run means that the input, frames, timestamps, period, ranges, and exact profile are recorded before mapping. It is not an accuracy or hardware-support claim. If the checklist exposes a sensor question that the existing contract cannot answer, use the sensor-support issue form with sanitized observations; do not attach raw bags, map geometry, or location-bearing logs.

KITTI / LiDAR-only evaluation path

KITTI Odometry Velodyne sequences do not include IMU. Use this path for LiDAR-only evaluation and frontend tuning, not as the public default workflow.

Download and run one sequence:

bash scripts/download_kitti_odometry.sh --velodyne
export KITTI_ODOMETRY_ROOT="$PWD/datasets/KITTI_odometry"
bash scripts/run_kitti_odometry_benchmark.sh --sequence 00 --small-gicp --force-prepare

Sweep several small_gicp parameter sets:

bash scripts/sweep_kitti_small_gicp.sh \
  --dataset "$KITTI_ODOMETRY_ROOT" \
  --sequences "00 05 07"

The KITTI wrappers write prepared rosbag2 data and benchmark artifacts under output/ by default. The raw KITTI dataset belongs under datasets/, which is local-only and ignored by Git.

For a trajectory already evaluated by localization_zoo, the converter also accepts --estimate-matrices and writes exact-timestamp /rko_lio/odometry + cloud pairs for the deterministic backend runner. Localization Zoo's csv_lidar_pose / *_evaluated_gt.txt matrices are already LiDAR-frame poses: keep the default --gt-frame lidar and do not pass --calib. For official KITTI camera-frame pose matrices, pass both --gt-frame camera and KITTI calib.txt with --calib; the explicit pair prevents accidental double conversion. The frozen sequence-00 experiment and adoption decision are in kitti00-localization-zoo-graph-loop-2026-07.md.

When multiple ROS overlays exist, pass --setup /path/to/install/setup.bash to run_offline_determinism_check.sh. The wrapper resolves the runner from that overlay and records its SHA-256, params and bag-metadata hashes, and every --param override in offline_determinism_summary.md. For a fast information-weight ablation, pass --param fixed_loop_edges_path:=/path/to/loop_edges.csv; descriptor search is skipped while the frozen constraints are replayed. Experimental Scan Context runs can reduce proposal cost deterministically with scan_context_query_stride (default 1, preserving existing behavior).

Backend only: graph_based_slam

Launch:

ros2 launch graph_based_slam graphbasedslam.launch.py

Default required input:

  • map_array: lidarslam_msgs/msg/MapArray

Optional backend aids:

  • /imu: sensor_msgs/msg/Imu when use_imu_preintegration:=true
  • /gnss/fix: sensor_msgs/msg/NavSatFix when use_gnss:=true

Alternative direct-input mode used by the RKO-LIO launch:

  • odom_input: nav_msgs/msg/Odometry
  • cloud_input: sensor_msgs/msg/PointCloud2

Useful GNSS weighting parameters:

  • gnss_topic
  • gnss_info_weight
  • gnss_use_covariance_weighting
  • gnss_covariance_min_variance_m2
  • gnss_covariance_max_variance_m2
  • gnss_rtk_fix_max_horizontal_stddev_m
  • gnss_rtk_fix_weight_scale
  • gnss_non_rtk_weight_scale

Optional save-time dynamic-object filter:

  • affects only the map written by /map_save
  • does not change live odometry, loop closure, or the published working map
  • useful when repeated passes observe parked/static structure consistently but transient objects appear only once

Parameters:

  • use_dynamic_object_filter
  • dynamic_object_filter_voxel_size
  • dynamic_object_filter_min_observations
  • dynamic_object_filter_temporal_window
  • dynamic_object_filter_max_range_from_sensor_m

Inspect a bag before enabling GNSS weighting:

python3 scripts/inspect_navsatfix_covariance.py /path/to/rosbag2 --topic /gnss/fix

For Leo Drive open-data driving bags that only expose Applanix raw GNSS status, inspect GSOF50 instead:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
python3 scripts/inspect_applanix_gsof50_quality.py /path/to/rosbag2 \
  --topic /lvx_client/gsof/ins_solution_rms_50 \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg

If the bag has GSOF49/50 but no /gnss/fix, generate a sidecar rosbag2 that publishes only sensor_msgs/msg/NavSatFix:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
python3 scripts/convert_applanix_gsof_to_navsatfix_bag.py \
  --input /path/to/rosbag2 \
  --output /tmp/applanix_navsatfix_bag \
  --gsof49-topic /lvx_client/gsof/ins_solution_49 \
  --gsof50-topic /lvx_client/gsof/ins_solution_rms_50 \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --force

If you want to test the same Applanix raw messages as sensor_msgs/msg/Imu, generate an IMU sidecar too:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
python3 scripts/convert_applanix_gsof_to_imu_bag.py \
  --input /path/to/rosbag2 \
  --output /tmp/applanix_imu_bag \
  --gsof49-topic /lvx_client/gsof/ins_solution_49 \
  --gsof50-topic /lvx_client/gsof/ins_solution_rms_50 \
  --output-topic /imu \
  --frame-id base_link \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --force

Then play the original bag together with the generated sidecar bag:

ros2 bag play /path/to/rosbag2 --clock
ros2 bag play /tmp/applanix_navsatfix_bag

For an end-to-end open-data GNSS smoke with automatic /map_save, use:

bash scripts/run_open_data_gnss_smoke.sh \
  --bag /path/to/rosbag2 \
  --verify-map

run_open_data_gnss_smoke.sh auto-detects the NavSatFix topic from --gnss-bag when provided, otherwise from --bag.

For Leo Drive driving bags that expose LiDAR as velodyne_msgs/msg/VelodyneScan and GNSS quality as Applanix GSOF49/50, use the packet-to-PointCloud2 wrapper instead:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
bash scripts/run_open_data_applanix_velodyne_gnss_smoke.sh \
  --bag demo_data/autoware_leo_drive_isuzu/driving_30_kmh_2022_06_10-15_47_42_compressed \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --verify-map

That wrapper will:

  • prefer same-bag native sensor_msgs/msg/NavSatFix / sensor_msgs/msg/Imu topics when they exist
  • otherwise generate a NavSatFix sidecar bag from GSOF49/50
  • optionally generate an Imu sidecar bag from GSOF49/50
  • extract a local TUM reference from GSOF49 with extract_applanix_gsof49_reference.py
  • build a minimal velodyne_pointcloud overlay on demand with bash scripts/prepare_velodyne_pointcloud_overlay.sh
  • convert VelodyneScan packets into sensor_msgs/msg/PointCloud2
  • run lidarslam.launch.py, call /map_save, and optionally verify the output

For rosbag2 compression_mode: FILE inputs, the smoke and benchmark wrappers stage a private playback view under the output directory. Any database that ROS 2 decompresses during playback is removed from that private view on exit, so the source bag directory is not left with a multi-gigabyte temporary file.

To benchmark the same driving_30_kmh bag as a four-way classic-path comparison, use:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
bash scripts/run_open_data_classic_path_benchmark_suite.sh \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --verify-map

That suite writes:

  • classic_path_report.md
  • classic_path_report.json
  • classic_path_report.svg

To rerun the current MID360 place-recognition comparison entrypoint, use:

bash scripts/run_place_recognition_benchmark.sh

That wrapper reruns the distance-only baseline and a Scan Context candidate, then emits:

  • place_recognition_report.md
  • place_recognition_report.json

For packet IMU deskew, the important caveat is runtime sensitivity. On the real Leo Drive all-sensors-bag1 and all-sensors-bag6 front-lidar cases, native /sensing/imu/imu_data works when the packet benchmark runs at rate=1.0. Current reference numbers are:

  • bag1_front, no_imu: APE RMSE 0.248 m
  • bag1_front, imu: APE RMSE 0.251 m
  • bag6_front, no_imu: APE RMSE 0.422 m
  • bag6_front, imu: APE RMSE 0.365 m

The benchmark wrapper defaults to rate=1.0 for all runs and deterministically prefers a /front/ packet stream when several Velodyne topics exist. To validate the same A/B automatically on the default front-lidar cases, run:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
bash scripts/run_open_data_packet_imu_deskew_validation_matrix.sh \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg

That matrix runs both no_imu and imu at rate=1.0 for determinism and writes per-case outputs plus:

  • packet_imu_deskew_validation.md
  • packet_imu_deskew_validation.json

The default acceptance criteria are:

  • no_imu path coverage >= 0.95
  • imu path coverage >= 0.95
  • imu_rmse / no_imu_rmse <= 1.10
  • imu_matched_poses / no_imu_matched_poses >= 0.80 deskew and 34.089 m with --imu-rotation-use-orientation false. That is why the public packet path still keeps --use-imu false by default.

If a bag carries NavSatFix messages whose header stamps do not track ROS time, the backend now falls back to receive time when the skew exceeds gnss_header_stamp_max_skew_sec (default 30 s). That makes all-sensors-bag6 attach GNSS edges again, but its native /gnss/fix still disagrees with the GSOF49 reference enough that it is better suited to georeferenced smoke tests than to clean GNSS cross-validation.

If you still want to test packet-based IMU deskew with a real static TF, use:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
bash scripts/run_open_data_applanix_velodyne_gnss_benchmark.sh \
  --bag demo_data/autoware_leo_drive_isuzu/driving_30_kmh_2022_06_10-15_47_42_compressed \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --use-imu true \
  --tf-bag demo_data/autoware_leo_drive_isuzu/all-sensors-bag6_compressed \
  --robot-frame-id base_link \
  --imu-frame-id base_link \
  --verify-map

That path uses:

  • convert_applanix_gsof_to_imu_bag.py
  • extract_static_transform_from_bag.py
  • PointCloud2.time-based deskew in scanmatcher
  • --imu-rotation-use-orientation false for the gyro-only rotation variant

To turn the same real open-data path into a benchmark artifact with traj_raw.tum, traj_corrected.tum, and metrics.json, use:

git clone --depth=1 https://github.com/autowarefoundation/applanix.git /tmp/applanix
bash scripts/run_open_data_applanix_velodyne_gnss_benchmark.sh \
  --bag demo_data/autoware_leo_drive_isuzu/driving_30_kmh_2022_06_10-15_47_42_compressed \
  --applanix-msg-dir /tmp/applanix/applanix_msgs/msg \
  --verify-map

Odometry and TF: two separate contracts

An nav_msgs/msg/Odometry message contains a parent frame in header.frame_id and a child frame in child_frame_id. Publishing those fields does not, by itself, guarantee that the matching transform is present in the /tf tree. Start with the read-only bag check:

lidarslam-map doctor /path/to/rosbag2

When the bag contains Odometry, doctor scans the highest-count Odometry topic and all recorded TF topics, with a 100,000-message bound per topic. It reports empty or inconsistent frame IDs, no connecting path, or a path containing only /tf_static. A multi-hop path such as odom -> base_footprint -> base_link is accepted when at least one edge comes from dynamic /tf.

When a selected PointCloud2 topic and that dynamic path both exist, doctor then makes a second bounded pass in bag record order. At each cloud record it checks whether every required dynamic edge has already appeared and whether the latest stamp observed on each edge is at least the cloud's header.stamp. It reports startup gaps and every positive future-TF gap; there is no arbitrary millisecond tolerance because the exact-stamp lookup can reject any request newer than the latest buffered transform. A clean result is necessary recorded-bag evidence, not proof of live executor scheduling, DDS delay, clock alignment, TF buffer history, or interpolation at every sensor timestamp. Use the checks below for the running system and replace every angle-bracket placeholder first.

  1. Check the Odometry message frames

    First confirm that <ODOM_TOPIC> is the intended nav_msgs/msg/Odometry topic, then sample both frame fields:

    timeout 5s ros2 topic echo --once --field header.frame_id <ODOM_TOPIC>
timeout 5s ros2 topic echo --once --field child_frame_id <ODOM_TOPIC>

    Expected: both outputs are non-empty and identify the intended <ODOM_FRAME> parent and <BASE_FRAME> child. If either is empty or unexpected, correct the Odometry publisher or launch remap before checking TF; do not invent frame names in a viewer.

  2. Check that the directed TF path exists

    ros2 run tf2_ros tf2_echo <ODOM_FRAME> <BASE_FRAME>

    Expected: repeated At time ... transforms in the same parent-to-child direction as the sampled message. If the transform is unavailable, the Odometry topic is not sufficient: enable the supported TF broadcaster or static-extrinsic configuration for the robot, then repeat this check. Do not silence TF warnings or copy a robot-specific broadcaster as a fix.

  3. Check transform freshness separately

    ros2 run tf2_ros tf2_monitor <ODOM_FRAME> <BASE_FRAME>

    Expected: the monitor reports a live publisher and bounded delay for the path. A missing path is a broadcaster/configuration problem; a large delay, future extrapolation, or stale timestamp is a timing problem. Align the clocks and message/TF timestamps or repair the actual publisher rate, then rerun doctor and the monitor. Increasing a lookup timeout alone does not repair stale data. Silencing the warning or substituting a stale transform does not repair it either.

Run RKO-LIO + graph_based_slam

The main launch entrypoint is:

ros2 launch lidarslam rko_lio_slam.launch.py \
  bag_path:=/path/to/rosbag2 \
  lidar_topic:=/os_cloud_node/points \
  imu_topic:=/os_cloud_node/imu

Useful parameter files:

  • default graph backend: graph_based_slam/param/graphbasedslam.yaml
  • default scanmatcher frontend: lidarslam/param/lidarslam.yaml
  • NTU VIRAL RKO-LIO profile: lidarslam/param/rko_lio_ntu_viral.yaml
  • MID360 tuned profile: lidarslam/param/lidarslam_mid360_rko_graph.yaml

Save Maps

Save the current map at any time with:

ros2 service call /map_save std_srvs/srv/Empty

Typical outputs:

  • map.pcd
  • pose_graph.g2o
  • pointcloud_map/pointcloud_map_metadata.yaml
  • pointcloud_map/*.pcd
  • map_projector_info.yaml

Autoware Map Output Notes

graph_based_slam always writes map_projector_info.yaml.

  • without GNSS: projector_type: Local
  • with GNSS and a stable origin: projector_type: LocalCartesian plus map_origin

To stage an existing run into an Autoware map bundle:

bash scripts/prepare_autoware_map_from_graph_slam.sh \
  output/bench_rko_lio_ntu_viral_loopgate_20260324 \
  /tmp/autoware_maps/ntu_viral_loopgate

To open the staged map through Autoware's map loaders:

bash scripts/run_autoware_pointcloud_map_viewer_docker.sh \
  /tmp/autoware_maps/ntu_viral_loopgate \
  /tmp/autoware_core \
  /tmp/autoware_map_runtime_ws

For the short supported path, use bash scripts/run_autoware_quickstart.sh instead.

Loop Closure Notes

graph_based_slam supports two loop-candidate sources:

  • distance-based revisit search
  • built-in GPL-free Scan Context place recognition

The backend validates candidates geometrically before adding a loop edge and keeps only the best local edge inside the configured dedup window.

Scan Context normally excludes the 50 most recent submaps. Experimental short sequences may override scan_context_exclude_recent, but the value must remain positive and the default stays 50. Treat a smaller window as an ablation: it can expose aliased descriptors that the geometric verifier must reject.

To regenerate the README loop-area zoom figure used for visual inspection of closing-segment duplication:

python3 scripts/generate_readme_loop_zoom_figure.py

Benchmark And Dataset Pointers

Recommended public benchmark:

bash scripts/download_ntu_viral_tnp01.sh --dry-run
bash scripts/download_ntu_viral_tnp01.sh
bash scripts/run_rko_lio_graph_benchmark.sh

The dry run performs no write or network request and reports the remaining phases, official archive identity, and conservative destination-space gate.

Current MID360 cross-validation path:

bash scripts/run_rko_lio_mid360_crossval_benchmark.sh

The public benchmark and release-report flow is documented in benchmarking.md.

Related Docs