G2VLO

[RAL 2026] G$^2$VLO: Accurate and Generic 2D Gaussian based Visual-LiDAR Odometry

G2VLO is a generic visual-LiDAR odometry framework based on 2D Gaussian Splatting (2DGS). It combines LiDAR geometry and dense visual rendering for accurate odometry, and supports both synchronized and non-strictly synchronized camera-LiDAR setups.

Highlights

• Generic 2D Gaussian based visual-LiDAR odometry
• Dense visual-LiDAR joint optimization
• Support for synchronized and asynchronous sensor inputs

Pipeline

The overall pipeline of G2VLO is shown below.

Our framework contains three main components:

1. LiDAR Subsystem
   The LiDAR subsystem estimates the LiDAR pose and constructs an adaptive voxel map with plane features for robust geometric registration.

2. Visual Subsystem
   The visual subsystem estimates camera poses by minimizing the discrepancy between rendered observations and input images. A photometric bundle adjustment step further refines keyframe poses with rendered depth and normal information.

3. Joint 2DGS Mapping Subsystem
   The joint mapping subsystem initializes 2D Gaussian primitives from LiDAR points and optimizes the local Gaussian map using RGB, depth, and normal supervision.

Together, these three modules form a unified visual-LiDAR odometry pipeline that combines accurate LiDAR geometry with dense visual tracking.

Project Status

We are currently organizing the codebase and datasets for public release.

The complete implementation, training/inference scripts, and related datasets will be released after the cleanup and documentation process is finished.

Data Acquisition Device

Our data acquisition device contains the following componets:

• An Allied Vision Alvium 1800 U-508r RGB camera.
• A Livox Avia LiDAR.
• A Raspberry Pi 4B (with a small screen), which is a low-cost embedded board based on ARM, is used to start/stop the camera and the LiDAR and store the data.
• A battery is used to provide energy for all devices.

Citation

If you find this project useful in your research, please consider citing our paper:

@article{tu2026g2vlo,
  title   = {G$^2$VLO: Accurate and Generic 2D Gaussian based Visual-LiDAR Odometry},
  author  = {Diantao Tu and Hainan Cui and Peilin Tao and Yangdong Liu and Shuhan Shen},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2026}
}