rko_slam

ROS2 LiDAR-inertial SLAM and multi-session alignment

A drive that returns to where it started, with rko_lio, and with rko_slam running on top of it

Quick start

rko_slam runs on rko_lio, my LiDAR-inertial odometry, or on your own odometry. Build it with rko_lio:

cd <ws>/src
git clone https://github.com/PRBonn/rko_lio
git clone https://github.com/PRBonn/rko_slam
cd <ws> && rosdep install --from-paths src --ignore-src -y
colcon build --packages-select rko_lio rko_slam

What it needs

• A LiDAR (sensor_msgs/PointCloud2 with per-point timestamps) and an IMU (sensor_msgs/Imu).
• The LiDAR and IMU frames linked to your base frame on TF, e.g. by your URDF.

Run

odometry_and_slam.launch.py starts rko_lio and rko_slam together and finds the LiDAR and IMU topics and your base frame by itself. Start your robot, then:

ros2 launch rko_slam odometry_and_slam.launch.py rviz:=true

From a bag, in two terminals, bag first:

ros2 bag play <bag> --clock
ros2 launch rko_slam odometry_and_slam.launch.py use_sim_time:=true rviz:=true

If your setup differs:

Your setup	Do this
Base frame named other than base_link, base_footprint or base	base_frame:=<yours>
Several LiDAR or IMU topics, e.g. the LiDAR driver's own IMU	an rko_lio config file naming yours and your base frame, below
LiDAR and IMU messages share one frame_id and no TF links it to a base frame, e.g. a Livox with its built-in IMU	an rko_lio config file with identity extrinsics, below
Topics or TF take over 10 s to appear	autodetect_timeout:=<seconds>
Something else already publishes odom <- <your base frame>, e.g. robot_localization or the bag's /tf	turn it off, or run with your own odometry, below

An rko_lio config file, passed with rko_lio_config_file:=<file>, replaces the autodetection (all keys):

lidar_topic: /livox/lidar
imu_topic: /livox/imu
# LiDAR and IMU in one frame, no TF
base_frame: livox_frame
extrinsic_lidar2base_quat_xyzw_xyz: [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0]
extrinsic_imu2base_quat_xyzw_xyz: [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0]

With your own odometry

slam.launch.py starts rko_slam alone, on any locally consistent odometry that publishes odom <- base_link on TF:

ros2 launch rko_slam slam.launch.py lidar_topic:=/points imu_topic:=/imu base_frame:=base_link deskew:=true

If your setup differs:

Your setup	Do this
Scans already deskewed	deskew:=false, the default
LiDAR and IMU in one frame, no TF to a base frame	base_frame:=<that frame>; your odometry publishes odom <- <that frame>
Odometry frame named other than odom	odom_frame:=<name>
Odometry publishes base_link <- odom	invert_map_tf:=true
No IMU	leave out imu_topic

With an IMU, gravity levelling improves the map:

A 20 km drive against ground truth, the same odometry under both runs: in height, the run without gravity levelling swings through ±40 m, the one with it stays within a few metres

Settings

• dump_results:=true writes the run to results/<run_name>_<n>/: each sub-map as it closes, then the trajectory and pose graph on shutdown.
• splitting_distance (50 m): a new sub-map starts at this straight-line distance from the last one's start, and a loop closes against a sub-map four or more back, set by no_of_sub_maps_to_skip (3, at least 1): at the defaults, a loop shorter than about 200 m never closes. For a small area, lower either, e.g. splitting_distance:=30.
• overlap_threshold (0.4): how much two sub-maps must overlap for a closure. Raise it where places look alike.

ros2 launch rko_slam slam.launch.py -s lists every rko_slam parameter. Offline processing, odometry from a TUM file, every output: documentation.

Multi-session alignment

Three sessions of the same place, recorded on different days, and the one frame they end up in

Run each session with dump_results:=true run_name:=day_1 (day_2, ...), then:

ros2 launch rko_slam align.launch.py run_dirs:="[results/day_1_0, results/day_2_0]"

It writes each session's trajectory in the joint frame and the joint pose graph to results/aligned_<n>/.

Citation

If rko_slam is useful to you, leave a star ⭐.

rko_slam is part of my PhD thesis. Until the thesis is published, please cite rko_lio, which it shares its core with:

@article{malladi2026ral,
  author      = {M.V.R. Malladi and T. Guadagnino and L. Lobefaro and C. Stachniss},
  title       = {A Robust Approach for LiDAR-Inertial Odometry Without Sensor-Specific Modeling},
  journal     = {IEEE Robotics and Automation Letters},
  year        = {2026},
  volume      = {11},
  number      = {6},
  pages       = {7420--7427},
  doi         = {10.1109/LRA.2026.3685966},
}

rko_slam builds on KISS-SLAM; its first version was a reimplementation for ROS2. Closure detection reimplements MapClosures. Please cite KISS-SLAM as well:

@INPROCEEDINGS{kiss2025iros,
  author    = {Guadagnino, Tiziano and Mersch, Benedikt and Gupta, Saurabh and Vizzo, Ignacio and Grisetti, Giorgio and Stachniss, Cyrill},
  booktitle = {2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  title     = {{KISS-SLAM: A Simple, Robust, and Accurate 3D LiDAR SLAM System With Enhanced Generalization Capabilities}},
  year      = {2025},
  pages     = {5363-5370},
  doi       = {10.1109/IROS60139.2025.11246613}
}

License

This project is free software made available under the MIT license. For details, see the LICENSE file.