Introduction
LiDAR sensors are the workhorses of robotics, providing precise 3D measurements. However, turning these measurements into a usable map involves a constant trade-off between accuracy, memory, and speed. Traditional methods often stack point clouds, leading to unmanageable data structures, while newer "neural implicit" methods (inspired by NeRFs) are computationally heavy and prone to "catastrophic forgetting."
Splat-LOAM addresses these challenges as the first LiDAR-only SLAM pipeline to represent the environment exclusively using 2D Gaussian primitives. By leveraging an explicit representation and a custom spherical rasterizer, it achieves state-of-the-art accuracy with a fraction of the computational cost of neural methods.
Methodology
Splat-LOAM introduces several key innovations to adapt Gaussian Splatting for the unique characteristics of LiDAR data:
- Spherical Projection Model: Unlike visual SLAM designed for pinhole cameras, Splat-LOAM uses a spherical projection to handle the $360^\circ$ panoramic field of view of LiDAR sensors.
- 2D Gaussian Primitives: The system utilizes 2D Gaussians (flat disks) rather than 3D ellipsoids. These are mathematically superior for representing thin surfaces and provide explicit surface normals essential for accurate mapping.
- Custom Differentiable Rasterizer: A tile-based rasterizer was developed to render range, normals, and opacity from these 2D Gaussians on a spherical frame.
Tracking is performed via a hybrid approach that minimizes a joint loss: a geometric residual (point-to-plane distance) and a photometric residual (comparing rendered range images to actual scans). To prevent Gaussians from growing excessively in sparse areas, a novel scaling regularization loss is applied.
Key Results
The authors evaluated Splat-LOAM across four diverse datasets: Newer College (NC), Vision Benchmark in Rome (VBR), Oxford Spires, and Mai City.
- Mapping Accuracy: On the Newer College "quad-easy" sequence, Splat-LOAM achieved an F1-score of 96.74%, outperforming neural implicit methods like $N^3$-Mapping (94.54%) and PIN-SLAM (88.05%).
- Hardware Efficiency: By using explicit Gaussians instead of heavy MLPs, the system achieves mapping speeds of 200–300 FPS.
- Minimal Footprint: Active memory usage is remarkably low, often staying under 30 MB for local maps, whereas competing neural methods require hundreds of megabytes.
Conclusion and Future Work
Splat-LOAM demonstrates that LiDAR provides a "cheat code" for Gaussian Splatting; high-quality spatial measurements allow for stable and accurate primitive initialization without the need for complex pre-processing. While the current system is sensitive to motion distortion from rapid sensor rotation, future work will focus on integrating motion compensation and loop closure to support large-scale, high-speed robotic applications.
Summary updated