In the past five months, the MaRS Laboratory of the University of Hong Kong has successively open-sourced four sets of online SLAM frameworks for UAVs:FAST-LIVO2, Point-LIO (grid-map branch), Voxel-SLAM, Swarm-LIO2 . These four frameworks cover core application scenarios such as single-machine three-sensor fusion, high-bandwidth and high-speed maneuvering, long-term multi-level map optimization, and decentralized cluster collaboration, and almost cover the main needs of current UAV autonomous navigation.

This article will introduce the core content and technical highlights of each SLAM algorithm to help you choose an appropriate solution and get started quickly.There are also links to papers and codes at the end of the article, allowing you to access them directly with one click.

01 FAST-LIVO2

FAST-LIVO2 is an efficient localization and mapping system that integrates LiDAR (LiDAR), inertial measurement unit IMU and visual information. The core uses error state iterative Kalman filter (ESIKF) and sequential update strategy to achieve accurate state estimation. Different from traditional methods, FAST-LIVO2 avoids feature extraction and directly processes raw LiDAR and image data, improving computing efficiency and system robustness.

Image source: "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", IEEE T-RO, 2024.

Technical Highlights

Direct method fusion strategy, abandoning cumbersome feature extraction The LiDAR module directly registers the original point cloud without extracting edge or plane features. The vision module performs image alignment by minimizing photometric errors, avoiding the traditional feature point extraction process.

Unified voxel map, geometry and texture depth collaboration
The system uses a unified voxel map structure, LiDAR builds a geometric structure, and image patches are attached to the LiDAR plane to form "visual map points" to achieve dense color mapping. This strategy not only optimizes the data structure, but also ensures the consistent integration of multi-source information.

Image alignment and robustness enhancement Sparse image alignment is performed by minimizing photometric errors and the reference image patch is dynamically updated; when the LiDAR blind area or FoV is insufficient, on-demand ray projection is used to fill the blind area; exposure time is estimated in real time to adapt to severe illumination changes; FAST-LIVO2 simultaneously improves image alignment accuracy and robustness in scenes with sudden illumination changes or missing point clouds.

Image source: "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", IEEE T-RO, 2024.

Image source: "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", IEEE T-RO, 2024.

ESIKF inertia update, system convergence is faster and more stable The error state iterative Kalman filter (ESIKF) is used to sequentially update the states of LiDAR and images, solving the problem of inconsistent multi-source observation dimensions and greatly improving fusion accuracy and real-time performance.

Image source: "FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry", IEEE T-RO, 2024.

Optimize resource utilization
In order to adapt to resource-constrained embedded platforms, FAST-LIVO2 has deeply optimized the system architecture and algorithm flow. Actual measurements show that FAST-LIVO2 can run stably on low-power ARM platforms such as RK3588, Jetson Orin NX, RB5, etc., and the single-frame processing delay is controlled within 78 ms, meeting the needs of real-time navigation and mapping.

02 Swarm-LIO2

Swarm-LIO2 is the latest open-source UAV cluster LiDAR-IMU state estimation framework from the MaRS Laboratory of the University of Hong Kong in April.Decentralization + plug and play + ultra-low bandwidth + edge acceleration, in scenarios where GPS is denied, the field of view is blocked, or even LiDAR is degraded, it still provides centimeter-level local and mutual poses for 40 UAVs, which can be called the "RTK of UAV clusters."

Image source: "Swarm-LIO2: Decentralized, Efficient LiDAR-Inertial Odometry for UAV Swarms", IEEE T-RO, 2025

Technical Highlights

Fully distributed, decentralized design A distributed communication framework is built based on the Ad-Hoc network. Each UAV independently broadcasts its identity and status without the need for coordination by a central node to avoid single points of failure.

Reflective features + fast initialization of factor graph By attaching reflective strips to its body, each UAV can automatically detect new teammates during flight and pass Trajectory matching + factor graph optimization Complete automatic calibration of time delays and global external parameters to support rapid initialization of large-scale clusters.

Image source: "Swarm-LIO2: Decentralized, Efficient LiDAR-Inertial Odometry for UAV Swarms", IEEE T-RO, 2025

Efficient estimator: ESIKF + marginalization mechanism Based on ESIKF, it integrates LiDAR, IMU and mutual observation measurements to provide high-precision pose estimation. The marginalization strategy is introduced to estimate only the currently observable external parameters of teammates to avoid the increase in calculation volume with the cluster size and significantly improve the scalability.

time compensation mechanism To address the problem of sensor asynchrony between different UAVs, a sophisticated mutual-observation measurement time compensation model is designed to significantly improve mutual localization accuracy.

Fully verified

• In simulation and real environments, Swarm-LIO2 implements:Outdoor obstacle avoidance formation flight, mission through dense forest, target tracking and dynamic entry/exit, multi-UAV coordinated transportation, demonstrating that it can still maintain high robustness under extreme conditions such as communication interruption and LiDAR degradation.

Image source: "Swarm-LIO2: Decentralized, Efficient LiDAR-Inertial Odometry for UAV Swarms", IEEE T-RO, 2025

03 Voxel-SLAM

Voxel-SLAM is a complete, accurate and versatile LiDAR-Inertial SLAM system. The system modules include: initialization, odometry estimation, local mapping, closed-loop detection, and global mapping. Its core advantage is that the five major modules uniformly use the adaptive Voxel Map structure, and realize the process from dynamic initialization to multiple session closed-loop through unified modeling of four types of data associations: short-term, medium-term, long-term, and multi-graph association.End-to-end high-precision mapping。

Image source: "Voxel-SLAM: A Complete, Accurate, and Versatile LiDAR-Inertial SLAM System", arXiv:2410.08935, 2024

Technical Highlights

Unified adaptive voxel map structure：
All modules use adaptive voxel maps for data expression to achieve efficient collaboration and fast data association of each module.

Efficient and robust initialization：
Even in a non-stationary initial state, it only takes about 1 second of data to complete the high-precision initialization of the system.

Real-time local optimization of mid-term data correlation：
Through innovative LiDAR-inertial BA technology, real-time optimization status and maps are achieved, significantly enhancing the system's adaptability to rapid motion and degradation scenarios.

Multi-session loopback detection and re-localization：
Supports loopback detection between current and historical sessions, greatly improving system stability during long-term operation.

Efficient global consistency map optimization：
A hierarchical global optimization method is introduced to ensure map accuracy and consistency while ensuring efficient computing performance.

Excellent robustness and generalization ability：
Tests have shown that Voxel-SLAM has demonstrated excellent localization and mapping capabilities in narrow indoor scenes, large-scale urban outdoor environments, and UAV high-altitude operation scenarios.

Image source: "Voxel-SLAM: A Complete, Accurate, and Versatile LiDAR-Inertial SLAM System", arXiv:2410.08935, 2024

04 Point-LIO (grid-map branch)

In February 2025, HKU added the point-lio-with-grid-map branch to the original Point-LIO framework, introducing occupancy grid reconstruction and degradation detection, specifically designed for extreme maneuverability and high-bandwidth control of racing UAVs.

Image source: "Point-LIO: Robust High-Bandwidth LiDAR-Inertial Odometry", Advanced Intelligent Systems 2023

Upgrade content

• Added grid mapping and degradation detection logic.

• Sustained 4–8 kHz odometer output with no IMU saturation at 75 rad/s angular velocity.

• Get started quickly: Fully compatible with the original Point-LIO, it is recommended to use an IMU above 200 Hz and a hard-synchronized LiDAR to give full play to the advantages of high frequency.

05 How to choose + key points to get started?

Note that this article only focuses on real-time SLAM/LIO/VIO; it does not include back-end components such as multi-session map fusion (LAMM) or dense reconstruction (GS-SDF, M2Mapping).

Resource Express

FAST-LIVO2: Fast and Direct Laser-Inertial Navigation-Visual Odometer
FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry
Paper link:https://arxiv.org/pdf/2408.14035_FAST-LIVO2
Code open-source:https://github.com/hku-mars/FAST-LIVO2
Hard sync device open-source:https://github.com/xuankuzcr/LIV\_handhold
Camera and LiDAR calibration open-source:https://github.com/hku-mars/FAST-Calib

Swarm-LIO2: Decentralized UAV swarm laser-inertial navigation odometry
Swarm-LIO2: Decentralized, Efficient LiDAR-Inertial Odometry for UAV Swarms
Paper link:https://arxiv.org/abs/2409.17798
open-source code:https://github.com/hku-mars/Swarm-LIO2

Voxel-SLAM: Complete, accurate and versatile laser-inertial SLAM system
Voxel-SLAM: A Complete, Accurate, and Versatile LiDAR-Inertial SLAM System
Paper link:https://arxiv.org/abs/2410.08935
open-source code:https://github.com/hku-mars/Voxel-SLAM

Point-LIO (grid-map branch): high-bandwidth laser-inertial navigation odometry
Point-LIO: Robust High-Bandwidth LiDAR-Inertial Odometry
Paper link:https://advanced.onlinelibrary.wiley.com/doi/full/10.1002/aisy.202200459
open-source code:https://github.com/hku-mars/Point-LIO/tree/point-lio-with-grid-map