LiDAR-Based SLAM: A Review of Algorithmic Modules and LiDAR-Inertial Integration Architectures
TL;DR
This review presents an architecture-centered synthesis of LiDAR-based SLAM, covering LiDAR-only and LiDAR-inertial systems through a dual-axis, architecture-centered taxonomy, and highlights the field’s progression toward continuous-time estimation, semantic scene understanding, adaptive map representations, neural implicit mapping, and uncertainty-aware sensor fusion.
Abstract
Highlights Main Findings: The review presents a dual-axis, architecture-centered taxonomy of more than 100 representative LiDAR-based methods. The first axis distinguishes LiDAR-only and LiDAR-inertial sensing configurations, while the second distinguishes their principal registration, representation, coupling, and estimator structures. The reviewed literature shows a persistent trade-off among accuracy, computational efficiency, and robustness, where methods optimized for high precision often require greater computational resources, while lightweight real-time systems may be less robust in geometrically degraded or dynamic environments. Implications of the main findings: Future LiDAR-based SLAM systems are likely to move beyond purely geometric matching toward modular frameworks that integrate continuous-time trajectory modeling, semantic scene understanding, adaptive mapping, and uncertainty-aware sensor fusion. No single SLAM architecture is universally optimal. System selection should be guided by deployment-specific factors such as platform dynamics, environmental structure, sensor configuration, computational resources, and required accuracy. Abstract Simultaneous Localization and Mapping (SLAM) enables real-time six-degree-of-freedom (6-DOF) state estimation and spatial reconstruction in environments where global navigation satellite systems (GNSSs) are unavailable or unreliable. Light Detection and Ranging (LiDAR) is particularly suited to this task because it provides direct geometric measurements and remains largely unaffected by ambient illumination. This review presents an architecture-centered synthesis of LiDAR-based SLAM, covering LiDAR-only and LiDAR-inertial systems through a dual-axis, architecture-centered taxonomy. The first axis distinguishes sensing configuration, while the second distinguishes the principal registration, representation, coupling, and estimator structures within the corresponding branches. Core algorithmic modules, including front-end registration, state estimation, loop closure, map representation, and back-end optimization, are examined to determine how their interactions affect accuracy, computational latency, robustness, and scalability. Comparisons of representative systems are integrated within the corresponding methodological sections. Quantitative results are compared only when the same dataset subset, metric, and evaluation protocol are shared, while results obtained under non-equivalent conditions are used only as contextual evidence. The analysis identifies recurring trade-offs between low-latency recursive estimation and globally consistent graph-based optimization, as well as between sparse feature-based, direct, and learning-based representations. It also highlights the field’s progression toward continuous-time estimation, semantic scene understanding, adaptive map representations, neural implicit mapping, and uncertainty-aware sensor fusion. Finally, the review organizes future research around trustworthy estimation and recovery, environmental adaptability, persistent map maintenance, resource-bounded operation, and dependable interfaces with navigation and motion planning.