ALIVE-LIO

ALIVE-LIO: Degeneracy-Aware Learning of Inertial Velocity for Enhancing ESKF-Based LiDAR-Inertial Odometry

GEODE-ShieldTunnel1

GEODE-ShieldTunnel5

---

News

• 2026-10-03: Accepted to IEEE T-ITS
• 2026-08-03: Code released

Pretrained Models

We release models trained for handheld and automotive platforms, one per IMU framerate. All of them ship with this repository under onnx/model/.

Model	Framerate	Description
model_handheld	200 Hz	Trained on Livox built-in IMU data
model_vehicle100	100 Hz	Trained on the GEODE dataset
model_vehicle200	200 Hz	Trained on the FusionPortableV2 dataset
model_vehicle400	400 Hz	Trained on the UrbanNav dataset

Select a model through the localizability/model_name parameter in your config file (without the .onnx extension).

To train a model on your own sequences, you can use AirIO. **Note that our preprocessing step, which removes the LIO-estimated IMU bias, is not released.**

1. Prerequisites

• Ubuntu >= 18.04 and ROS. Follow ROS Installation. We used Ubuntu 20.04.
• PCL >= 1.8. Follow PCL Installation.
• Eigen >= 3.3.4. Follow Eigen Installation.
• Sophus. Follow Sophus Installation.
• ONNX Runtime. Download from ONNX Runtime Releases and extract it under onnx/. We used the prebuilt onnxruntime-linux-x64-gpu-1.18.0, and its directory name is hard-coded in ONNX_DIR at CMakeLists.txt:21. If you use another version, edit that line to match.
• livox_ros_driver. Follow livox_ros_driver Installation, and source it before building.

2. Build

Clone the repository and catkin_make:

    cd ~/catkin_ws/src
    git clone https://github.com/cocel-postech/alive-lio.git
    cd alive_lio
    cd ../..
    catkin_make
    source devel/setup.bash

Dataset

We release two sequences recorded at Seosan Tunnel (36.0420409, 129.3612293) in Pohang-si, Gyeongsangbuk-do, South Korea. The platform specification is available in the CoCEL Handheld Dataset repository.

Both sequences start at one end of the tunnel, travel to the other end, and come back to the starting position. They were recorded with a Livox Avia moving sideways while mostly facing the tunnel wall. Since each sequence starts and ends at the same position, the end-to-end error can be used as a drift measure. **Note that no ground-truth trajectory is provided.**

Sequence	Duration	Description	Download
handheld_light	357 s	Recorded under normal exposure	link
handheld_dark	380 s	Recorded with the aperture fully closed, leaving no visual cue except the tunnel lights	link

TBD

• Docker support

Citation

If you use this library for any academic work, please cite our original paper.

@article{kim2026alive,
  title={{ALIVE-LIO}: Degeneracy-Aware Learning of Inertial Velocity for Enhancing {ESKF}-Based {LiDAR}-Inertial Odometry},
  author={Kim, Seongjun and Lee, Daehan and Hong, Junwoo and Park, Sanghyun and Jo, Hyunyoung and Han, Soohee},
  journal={arXiv preprint arXiv:2604.02706},
  year={2026}
}

Acknowledgments

Thanks a lot for the authors of PV-LIO, VoxelMap and AirIO.

License

ALIVE-LIO is released under the GNU General Public License v2.0. It is built on PV-LIO, FAST-LIO2 and VoxelMap, which are GPL-2.0 licensed; the bundled IKFoM (BSD-3) and robin-map (MIT) keep their own licenses.

Copyright (c) 2026 Seongjun Kim, POSTECH.

Contact

If you have any questions, please contact Seongjun Kim (seongjun4527@postech.ac.kr).