Super-LIO: A Robust and Efficient LiDAR-Inertial Odometry System with a Compact Mapping Strategy

TL;DR

Super-LIO tackles real-time LiDAR-Inertial Odometry on resource-constrained platforms by introducing OctVox, a density-regularized octo-voxel map with eight subvoxel representatives per voxel, and HKNN, a heuristic-guided KNN search that exploits the subvoxel layout for efficient, high-quality correspondences. Integrated into a tightly coupled IESKF pipeline, the system delivers competitive accuracy with markedly lower runtime and CPU usage across X86 and ARM platforms, validated on diverse public and private datasets. Key contributions include the compact OctVox map, the HKNN search strategy, and a fully open-source implementation that demonstrates robustness in high-speed, long-duration, sparse-sensing, and narrow-scene scenarios. The work enables practical, real-time LIO for aerial and ground robots on embedded hardware, expanding the applicability of LiDAR-inertial fusion in resource-limited environments.

Abstract

LiDAR-Inertial Odometry (LIO) is a foundational technique for autonomous systems, yet its deployment on resource-constrained platforms remains challenging due to computational and memory limitations. We propose Super-LIO, a robust LIO system that demands both high performance and accuracy, ideal for applications such as aerial robots and mobile autonomous systems. At the core of Super-LIO is a compact octo-voxel-based map structure, termed OctVox, that limits each voxel to eight fused subvoxels, enabling strict point density control and incremental denoising during map updates. This design enables a simple yet efficient and accurate map structure, which can be easily integrated into existing LIO frameworks. Additionally, Super-LIO designs a heuristic-guided KNN strategy (HKNN) that accelerates the correspondence search by leveraging spatial locality, further reducing runtime overhead. We evaluated the proposed system using four publicly available datasets and several self-collected datasets, totaling more than 30 sequences. Extensive testing on both X86 and ARM platforms confirms that Super-LIO offers superior efficiency and robustness, while maintaining competitive accuracy. Super-LIO processes each frame approximately 73% faster than SOTA, while consuming less CPU resources. The system is fully open-source and plug-and-play compatible with a wide range of LiDAR sensors and platforms. The implementation is available at: https://github.com/Liansheng-Wang/Super-LIO.git

Figures (10)

• Figure 1: Performance comparison of different LIO methods on X86 and ARM platforms. The left plot shows the average processing time per LiDAR frame, annotated relative to Super-LIO (1x) on each platform, while the right plot shows runtime CPU utilization. Note that the X86 experiments were conducted with 5x playback speed, which leads to higher CPU usage compared to ARM.
• Figure 2: Overview of Super-LIO. An IESKF fuses IMU and LiDAR data: IMU measurements propagate the state at the IMU rate; at each LiDAR frame, points are de-skewed using the IMU state, center-based downsampled, and, via a heuristic-guided KNN (HKNN) in the OctVox Map, nearest neighbors are retrieved to construct point-to-plane residuals. These residuals feed the IESKF observation update. The map stores up to eight subvoxel representatives per voxel and is updated online to ensure efficient and accurate correspondences.
• Figure 3: Illustration of the HKNN candidate selection process based on subvoxel proximity. A query subvoxel (blue) is surrounded by neighboring subvoxels at increasing distances, which are color-coded by the minimum Euclidean distance between their nearest vertices. These candidates are pre-grouped by geometric proximity; the groups are traversed sequentially in order of increasing distance until $K$ neighbors are found and no subsequent group can yield a closer neighbor, enabling efficient Top-$K$ retrieval without per-query sorting.
• Figure 4: Experimental platforms used for dataset collection. (Left) A handheld mapping unit equipped with a Livox-MID360 LiDAR, RGB camera, and an NVIDIA Orin NX embedded processor. (Right) A 220 mm quadrotor UAV carrying the same LiDAR and embedded processor, used for autonomous flight experiments.
• Figure 5: Point cloud maps produced by Super-LIO on ten self-collected evaluation sequences. The sequences span diverse environments, including forests, parks, underground garages, and office buildings. “h” denotes data collected with a handheld device, and “u” denotes data collected on a UAV platform. These varied scenarios and motion conditions are used to evaluate computational efficiency and robustness.