Autonomous Obstacle Avoidance and Navigation Technologies for Unmanned Aerial Vehicles Based on LiDAR: A Review

1. Introduction

Unmanned aerial vehicles (UAVs) are increasingly deployed in tasks that require flight through cluttered, unstructured, and often GPS-denied environments, including forest inspection, search and rescue, precision agriculture, and indoor exploration. In these settings, the vehicle cannot rely on satellite positioning or pre-surveyed maps and must instead build an internal representation of its surroundings and react to obstacles in real time. Among the sensing modalities available for this task, LiDAR has emerged as a benchmark choice because it provides direct, high-precision, and lighting-independent range measurements over a wide field of view, in contrast to cameras, which depend on scene texture and illumination, or ultrasonic and infrared sensors, which offer limited range and resolution [1,2,3].

The central engineering challenge is that raw LiDAR point clouds are dense, high-frequency, and three-dimensional, while the onboard computers carried by small UAVs are constrained in processing power, memory, and energy budget. Translating a stream of point clouds into safe, low-latency flight commands therefore requires a carefully engineered pipeline: filtering and downsampling the raw data, segmenting obstacles from traversable space, building a memory-efficient three-dimensional map or distance field, estimating the vehicle's own pose through LiDAR-based odometry, and finally generating or reacting with a collision-free trajectory. Each stage of this pipeline has produced a substantial and rapidly evolving body of literature.

This review synthesizes recent, primarily peer-reviewed literature on LiDAR-based UAV autonomy, with emphasis on studies published over the last five years. The scope is restricted to LiDAR-centred perception, mapping, localization, and obstacle-avoidance methods developed specifically for, or directly applicable to, unmanned aerial platforms; ground-vehicle and general robotics literature is referenced only where it clarifies a mechanism that is directly relevant to aerial navigation. The remainder of the paper is organized as follows. Section 2 reviews LiDAR perception and point-cloud processing. Section 3 covers three-dimensional environment representation and LiDAR-based SLAM. Section 4 discusses multi-sensor fusion. Section 5 surveys obstacle-avoidance and trajectory-planning algorithms, including learning-based approaches. Section 6 addresses dynamic-obstacle handling. Section 7 outlines current technical challenges, and Section 8 discusses future research directions. Section 9 then draws together representative systems from throughout the review in a single comparative table before the paper concludes in Section 10.

2. LiDAR Perception and Point-Cloud Processing for UAVs

Before a point cloud can be used for mapping or planning, it must be filtered, downsampled, and segmented so that obstacles can be distinguished from free space and from the ground plane. Voxel-grid and pass-through filters remain the most widely used preprocessing tools because of their low computational cost and predictable memory footprint – a direct comparison of these and related outlier-removal filters for streaming point-cloud pipelines confirms their favourable size- and bandwidth-reduction trade-offs [4,5] – while learning-based filters are increasingly used to remove noise and reduce point density adaptively rather than uniformly [6]. Ground segmentation techniques, such as RANSAC plane fitting and Euclidean clustering, separate terrain returns from genuine obstacles so that downstream obstacle-avoidance modules are not triggered by the ground itself; lightweight obstacle representations that avoid explicit ground segmentation and three-dimensional clustering altogether have also been proposed to reduce the computational burden on small platforms [7].

A parallel trend is the direct use of raw LiDAR point clouds for downstream planning tasks rather than first converting them into a heavy intermediate mapping representation. Recent high-speed MAV navigation systems have demonstrated fully onboard trajectory planning directly from LiDAR point-cloud observations, reducing the latency associated with maintaining a separate dense mapping stack, though at the cost of more sophisticated real-time collision-checking logic [8]. Deep-learning architectures designed natively for irregular point sets, such as PointNet, PointNet++, VoxelNet, and sparse convolutional networks, are increasingly applied to LiDAR data for obstacle classification and semantic labelling because of their superior recognition performance compared with hand-engineered geometric features. The PointNet++ architecture underlying many of these pipelines was originally introduced by [9], and UAV-specific adaptations such as MLF-PointNet++ have since demonstrated accurate object classification directly from LiDAR-UAS point clouds [10].

Dynamic obstacle tracking is commonly treated as a distinct sub-problem within the perception stage. Since static structures such as buildings, trees, and terrain can be incorporated into a persistent map whereas moving agents such as other aircraft, birds, and machinery cannot, modern LiDAR pipelines increasingly perform an early separation of static and dynamic point-cloud components, allowing static observations to update the map while dynamic observations are forwarded to dedicated tracking and motion-prediction modules [11,12].

3. LiDAR-Based 3D Environment Representation, Mapping, and SLAM

3.1. Volumetric Occupancy Mapping

The most widely adopted approach for converting point clouds into a persistent three-dimensional map is volumetric occupancy mapping using octree data structures, exemplified by OctoMap, in which each voxel maintains a probabilistic occupancy estimate and the tree structure keep memory usage compact even for large environments [13]. OctoMap-style representations are memory-efficient and widely adopted in robotic navigation frameworks; however, maintaining high-frequency occupancy updates can become computationally demanding in dynamic environments and on resource-constrained onboard platforms [12,14).

Vision-based occupancy prediction has expanded the field from mapping with range sensors in mostly static robotics settings to semantic occupancy prediction in dynamic driving scenes [15,16]. These methods increasingly use multimodal fusion, differentiable volumetric rendering, temporal fusion, and occupancy flow to improve semantic consistency, geometry, and dynamic-scene handling [15,16,17].

Evaluation remains a weaker point than algorithm design. Comparative work notes that comparative evaluations often focus primarily on computational efficiency, while broader quality metrics remain less standardized [18,19]. That gap matters because parameter optimization alone improved OctoMap-based performance by up to 15% in one benchmark study [19].

Volumetric occupancy mapping is therefore a mature robotic mapping paradigm for explicit free-space reasoning, but current research is still actively improving uncertainty handling, scalability, dynamic-scene modeling, and learned semantic prediction.

3.2. Euclidean Signed Distance Fields

To support gradient-based trajectory optimization directly, several frameworks convert volumetric occupancy information into a Euclidean Signed Distance Field (ESDF), in which every voxel stores its distance to the nearest obstacle. Voxblox incrementally builds such fields on board a micro aerial vehicle and allows a planner to compute smooth, collision-free trajectories through gradient descent on the distance field rather than through discrete search [20]. Related frameworks such as FIESTA offer similar functionality with improved incremental update efficiency [21]. More recent adaptive voxel-mapping schemes improve registration efficiency and estimation robustness by allocating spatial resolution according to local point-cloud density, while hash-voxel and Gaussian-based map representations further enhance memory efficiency and online update performance in large-scale environments [22,23].

The main recurring limitation is the tradeoff between accuracy and efficiency: several systems note that previous methods accept approximation errors, limit scale or resolution, or incur high update cost when maintaining full ESDF volumes [24,25]. Newer work is therefore pushing in three directions at once: faster explicit updates on CPU or GPU, more accurate non-projective or VDB-based fields, and continuous implicit ESDFs for compact storage, collaboration, and differentiable planning [25,26,27]. Overall, Euclidean Signed Distance Fields are a mature planning representation, but current research is still actively improving incremental updates, continuity, dynamic-scene handling, and scalability [28,29].

3.3. LiDAR Odometry and SLAM

Because GPS is frequently unavailable or unreliable near obstacles, LiDAR-based simultaneous localization and mapping is a foundational enabling technology for UAV autonomy. Early systems such as LOAM separate a high-frequency, low-fidelity odometry estimate from a low-frequency, high-fidelity mapping process, but are computationally demanding for small platforms [30]. LIO-SAM tightly couples LiDAR and inertial measurements within a factor-graph formulation, improving robustness during aggressive maneuvers [31]. A significant shift in the field came with FAST-LIO, and subsequently FAST-LIO2, which directly register raw LiDAR points to an incrementally updated map through a tightly coupled iterated Kalman filter, eliminating explicit feature extraction and enabling higher accuracy with substantially lower computational load than earlier LiDAR–inertial systems; the framework has demonstrated odometry and mapping rates of up to 100 Hz on both x86 and ARM-based onboard platforms, making it well suited to aggressive aerial maneuvers [32,33]. Adaptive voxel-based spatial architectures such as Voxel Map build on this direct-registration paradigm to further reduce drift and accelerate map updates [34].

Because pure LiDAR-inertial odometry can degrade in geometrically degenerate settings such as long corridors or open fields with few distinguishable surfaces, tightly coupled LiDAR-inertial-visual systems have been developed to add complementary visual constraints. FAST-LIVO and its successor FAST-LIVO2 fuse LiDAR, inertial, and camera data through a shared error-state iterated Kalman filter, using direct, feature-extraction-free methods for both the LiDAR and visual branches; FAST-LIVO2 has been shown to sustain autonomous UAV navigation and obstacle avoidance through basement and forest environments, including scenes in which LiDAR geometry alone is degenerate [35,36]. Multi-sensor fusion reviews similarly conclude that fusion becomes especially valuable precisely when LiDAR geometry is sparse or ambiguous [37].

4. Sensor Fusion Approaches

Relying on a single sensing modality introduces structural risks: LiDAR alone degrades in feature-poor corridors and in adverse weather, cameras alone depend on scene texture and illumination, and inertial measurement units alone drift over time without external correction. Consequently, modern UAV navigation stacks loosely or tightly couple LiDAR with an inertial measurement unit (IMU) and, increasingly, with monocular or stereo cameras. LiDAR–IMU fusion improves state estimation and robustness during rapid rotations and is the basis of the FAST-LIO family of odometry systems described above. LiDAR–camera fusion adds semantic and texture information that pure geometric LiDAR data lacks, enabling object classification in addition to ranging, as demonstrated by the FAST-LIVO line of work. Where absolute positioning is available intermittently, LiDAR–GPS–IMU fusion supports long-range navigation and drift correction [1]. Recent Gaussian-splatting-based odometry frameworks extend this trend by fusing LiDAR, inertial, and visual measurements within a unified, differentiable map representation that supports both accurate localization and photorealistic reconstruction [23].

5. LiDAR-Driven Obstacle Avoidance and Trajectory-Planning Algorithms

5.1. Reactive and Geometric Methods

Reactive methods compute control commands directly from current sensor measurements without maintaining a persistent global map, making them attractive for resource-constrained UAV platforms that require low-latency onboard execution [38]. Artificial Potential Field (APF) methods model the goal as an attractive potential and obstacles as repulsive potentials, and compute the resulting gradient field to generate instantaneous motion commands [39]. Because APF relies primarily on local geometric computations, it remains computationally lightweight and suitable for real-time implementation on embedded robotic systems [40]. However, classical APF formulations are well known to suffer from local minima, oscillatory trajectories, and freezing behavior in narrow or densely cluttered environments, motivating numerous augmented and hybrid APF variants designed specifically to mitigate these shortcomings [38,39,41]. Vector Field Histogram (VFH) and its three-dimensional variants collapse LiDAR returns into local polar histograms and select an opening, or safe steering direction, from the resulting occupancy bins; these methods are efficient and operate in real time but are sensitive to sensor noise and offer no guarantee of global optimality. An improved 3D-VFH variant built on an octree-based LiDAR occupancy representation has demonstrated real-time, attitude-stable local obstacle avoidance for multi-rotor UAVs in confined, GPS-denied environments, including avoidance of dynamically appearing obstacles [42]. The Dynamic Window Approach (DWA) additionally accounts for the vehicle's kinodynamic constraints when evaluating candidate velocities, which suits agile quadrotors, though its planning horizon remains myopic and it can struggle in highly cluttered scenes. A LiDAR-driven 3D extension of DWA, in which an onboard 3D LiDAR point cloud is projected onto an incrementally built OctoMap to inform real-time velocity selection, has demonstrated robust, low-latency (about 40 ms per cycle) reactive navigation for UAVs in confined indoor environments, though it can still fall into local minima when operating without a global planner [43].

5.2. Optimization-Based Trajectory Generation

Where an ESDF or occupancy map is available, trajectory optimization methods compute smooth, dynamically feasible, collision-free paths by minimizing an objective such as control effort or snap subject to obstacle-clearance constraints. Frameworks such as Voxblox-derived ESDF planners allow gradient descent directly on the distance field [20]. To avoid the computational cost of constructing and maintaining a full ESDF, EGO-Planner instead estimates collision gradients directly from an incrementally built, ESDF-free representation, substantially reducing per-cycle planning time while retaining robustness in cluttered scenes [44]. MINCO, formulated as a minimum-control trajectory class with closed-form dependence on waypoints and time allocation, further reduces the computational cost of large-scale, dynamically feasible trajectory generation and has become a common backend for spatial-temporal joint optimization in quadrotor planners [45].

5.3. Direct Point-Cloud and High-Speed Planning

A more recent generation of systems bypasses the classical perception–mapping–planning–control pipeline in favour of planning directly on raw or lightly processed point clouds, reducing latency at the cost of increased demands on the real-time collision-checking logic. The Safety-assured High-speed Aerial Robot (SUPER) generates two trajectories in every replanning cycle – one that maximizes speed and one that guarantees safety in previously observed free space – and has demonstrated fully onboard, GPS-free flight at speeds exceeding 20 m/s while avoiding obstacles as thin as 2.5 mm, such as wires and twigs, using a lightweight three-dimensional LiDAR [8].

5.4. Learning-Based Reactive Control

Deep reinforcement learning (DRL) enables agents to learn complex avoidance behaviors by mapping sensor observations directly to control actions, bypassing explicit motion prediction [46,47]. When applied to UAVs equipped with 3D LiDAR, such reactive DRL policies allow high-speed navigation in cluttered environments without the computational overhead of dynamic trajectory forecasting [47]. To circumvent the heavy perception–mapping–planning–control pipeline entirely, researchers have increasingly turned to learning-based reactive control, training neural policies that map raw or downsampled LiDAR observations directly to flight-control commands. [48] trained a convolutional sensorimotor policy in simulation via privileged imitation learning and demonstrated zero-shot transfer to real, previously unseen environments, including dense forests, snow-covered terrain, derailed trains, and collapsed buildings, at speeds around 40 km/h using onboard sensing and computation alone. Globally guided, deep value-network planners have similarly been applied to motion planning for fixed-wing UAVs, extending learning-based motion planning methods to fixed-wing UAV platforms with different aerodynamic constraints [49].

End-to-end deep reinforcement learning architectures can successfully bypass explicit map construction and trajectory generation by directly mapping raw LiDAR scans to continuous velocity and heading commands [50]. Moreover, such actor-critic policies trained purely in simulation can demonstrate robust sim-to-real transfer and obstacle avoidance capabilities during real-world MAV flight experiments [50]. Soft Actor-Critic algorithms based on fused LiDAR and depth-camera observations achieve high sample efficiency for learning obstacle avoidance policies in dynamic environments [51]. Recurrent Neural Network units integrated into a double-critic SAC/TD3 policy trained on sparse range-sensor data improve transfer inference ability and mapless obstacle-avoidance capability for UAVs in 3D environments [52].

End-to-end learning approaches compress multi-frame historical LiDAR scans into 2D obstacle maps, enabling a single neural network to handle perception and navigation without mode switching [53]. Physics-inspired DRL frameworks model robots and obstacles as electrical charges with Coulomb forces, providing explainable motion planning with anticipatory obstacle boundary rewards from LiDAR segmentation [54]. Spiking neural networks driven directly by LiDAR range data offer neuromorphic control with superior energy efficiency and robustness to perturbations compared to conventional non-spiking paradigms [55].

Pure learning-based methods struggle in specific environments where the trained policy fails to find a valid path, necessitating integration with global path planners [46]. Sim-to-real transfer remains a challenge, though training environments with dynamic characteristics and ROS implementation improve real-world viability [46].

6. Dynamic-Obstacle Handling and SLAM in Dynamic Environments

Static mapping and localization are relatively mature, but safely handling moving obstacles – other aircraft, birds, human operators, or machinery – remains substantially harder because it requires high-frequency temporal reasoning in addition to geometric mapping. If the LiDAR's effective scanning frequency is low relative to the closing speed of an incoming obstacle, the perception system may fail to estimate the obstacle's motion vector in time to guarantee a safe response.

The literature converges on the view that LiDAR-inertial-visual SLAM improves localization accuracy and mapping robustness, which are essential for reliable obstacle avoidance in dynamic settings [36], but that SLAM alone is insufficient for fast-moving obstacles unless dynamic points are explicitly filtered [12] or the obstacle's future motion is predicted [56]. Several complementary mechanisms have been shown to strengthen performance in dynamic scenes: dynamic-point filtering removes false obstacles and map artifacts introduced by moving agents [11,12]; loop-closure and global optimization reduce drift and improve long-term positioning consistency, with intensity-based loop closure shown to outperform LeGO-LOAM- and LIO-SAM-style front ends at minimal added runtime cost [11]; tighter LiDAR-inertial fusion improves robustness and state estimation under rapid motion [57,58]; explicit prediction of moving-object trajectories allows the vehicle to evade fast or suddenly appearing obstacles [56]; and lifelong SLAM with continuous map updating and relocalization maintains navigation reliability as the scene changes over time [59]. UAV-specific field trials support these mechanisms: a VIO-LiDAR SLAM system coupled with model-predictive control adjusted flight paths in real time to avoid trees and power poles during farmland operation [60], and an indoor exploration framework combining SLAM with distance-field mapping and a parallel collision-avoidance module demonstrated robust real-time navigation despite odometry uncertainty and sparse point clouds [61]. Several caveats temper this generally positive picture. Reactive, high-frequency map updates do not by themselves substitute for explicit trajectory prediction and can still miss fast movers [12]; limited sensor field of view can reduce environmental observability and degrade mapping robustness, particularly in constrained or geometrically ambiguous environments [62]; and feature-poor or geometrically degenerate scenes remain challenging for LiDAR SLAM, motivating tighter LiDAR–inertial coupling to improve robustness [57,63]. Much of the supporting evidence is also system-level rather than derived from head-to-head UAV trials that isolate LiDAR SLAM as the single causal factor [12,60,61], which limits the strength of causal claims that can currently be drawn from the literature.

7. Current Technical Challenges

7.1. Computational Constraints on Micro-UAVs

Processing high-density, multi-echo three-dimensional point clouds is computationally intensive, yet small, tactical UAVs are bound by tight size, weight, power, and cost (SWaP-C) budgets [64]). These constraints make fully autonomous LiDAR-based flight and onboard state estimation challenging [65,66]. Traditional multi-line spinning LiDARs are often impractical for micro-UAVs due to their bulky size, heavy weight (~1 kg), low resolution, and low frame rates [67]. Running multi-layered SLAM together with real-time path planning on embedded platforms such as the NVIDIA Jetson NX or Xavier frequently risks thermal throttling or increased control-loop latency, and onboard computational limitations are consistently identified as a major bottleneck for autonomous UAV deployment; a dedicated bottleneck-analysis model for UAV onboard compute design shows that, depending on the sensing and autonomy stack, safe operating velocity can be limited by the compute, sensor, or airframe dynamics in turn, underscoring how tightly LiDAR processing load is coupled to overall flight performance, and automated co-design frameworks that jointly select autonomy algorithms and onboard accelerators under the same SWaP constraints report substantial mission-count improvements over general-purpose hardware baselines [68,69]. End-to-end reinforcement learning policies based on high-dimensional point clouds introduce substantial computational demands, motivating compressed or surrogate representations for embedded deployment [47].

7.2. Environmental Degradation

Because LiDAR relies on the time-of-flight of emitted light pulses, its performance deteriorates in degraded visual environments such as dust, heavy fog, rain, snow, and smoke, where light scatters off suspended particles. This scattering introduces geometric noise and spurious returns ("phantom obstacles") that can trigger unnecessary stopping maneuvers or trajectory replanning, and weather sensitivity remains one of the primary practical limitations of LiDAR-only navigation in outdoor autonomous operations [70,71,72,73].

7.3. Dynamic-Obstacle Tracking and Temporal Latency

As discussed in Section 6, tracking fast-moving obstacles requires temporal resolution that can exceed the effective scanning rate of the sensor, and most existing systems remain considerably more robust in static than in highly dynamic scenes [12,56].

7.4. Limited Benchmark Datasets and Evaluation Protocols

The field still lacks large-scale, standardized UAV LiDAR datasets and benchmark scenarios specifically targeting dynamic obstacle avoidance, which slows reproducibility and makes head-to-head comparison across algorithms difficult [1,74]. Related to this, mirror and other highly specular or reflective surfaces can corrupt occupancy maps and induce false obstacles; intensity-based correction methods have been shown to exceed 97% detection accuracy indoors, but such corrections are not yet standard in most pipelines [75].

8. Future Research Directions and Trends

8.1. Solid-State LiDAR Integration

The UAV navigation community is increasingly shifting from heavy, mechanically rotating LiDAR sensors toward compact solid-state LiDAR (SSL) devices, owing to their reduced weight, lower power consumption, cost-effectiveness, and enhanced mechanical robustness stemming from the absence of moving parts. However, SSL sensors typically exhibit a restricted field of view and frequently employ non-repetitive scanning patterns, which present severe challenges for conventional 360° LiDAR–SLAM pipelines and motivate the development of specialized, scanning-aware mapping, localization, and motion-planning algorithms tailored to solid-state perception. Active, environment-aware control of the scanning pattern itself can further improve odometry robustness in geometrically complex or degenerate scenes by steering sensor coverage toward informative regions in real time [76]. Consequently, achieving real-time, high-speed autonomous UAV flight using solid-state LiDAR in cluttered and dense dynamic environments remains an active and evolving research domain [8,77,78].

8.2. Kinodynamic Trajectory Optimization

Frameworks such as MINCO and EGO-Planner are trending toward tighter integration of trajectory optimization within the flight-control loop itself, so that planned trajectories natively respect the vehicle's physical and aerodynamic limits rather than being generated independently and subsequently tracked [44,45].

8.3. Multi-Sensor Fusion Beyond LiDAR-IMU-Camera

Future systems are expected to extend fusion further [79], combining LiDAR with radar, which is largely unaffected by dust and fog [80], and with event cameras, which offer very high temporal resolution for fast-moving obstacles [81,82]), with the shared goal of robust, all-weather operation. Fusing radar and LiDAR data leverages the high-resolution point clouds of LiDAR to enhance radar's limited object recognition while radar compensates for LiDAR's weather-induced failures [83]. This complementary relationship makes sensor fusion essential for autonomous systems that must operate reliably across the full spectrum of environmental [84,85].

The limited temporal sampling rate of 3D LiDAR creates critical latency gaps when tracking fast-moving obstacles, which motivates the fusion of LiDAR with event cameras. Unlike conventional frame-based cameras, event sensors transmit asynchronous brightness-change events with microsecond-level temporal resolution, avoiding motion blur and enabling continuous perception between successive LiDAR scans [86]. LiDAR provides accurate metric depth measurements, while event cameras deliver high-rate temporal information to detect dynamic targets in the intervals between laser sweeps [87]. Nevertheless, joint deployment imposes additional challenges: precise spatial-temporal sensor calibration, asynchronous data fusion, and increased onboard computational load on SWaP-constrained micro-UAVs [88]. Event-augmented LiDAR-inertial odometry frameworks demonstrate improved state-estimation stability during agile flight and enhance dynamic-obstacle-tracking capability compared to LiDAR-only perception [87,89]. Comprehensive surveys confirm that event–LiDAR fusion is one of the most promising approaches to mitigate temporal-latency limitations of LiDAR perception in highly dynamic environments [88,89].

8.4. Learning-Based Navigation and Foundation Models

Reinforcement learning, imitation learning, and general-purpose robotic foundation models are increasingly being explored as alternatives to, or complements for, hand-engineered planning pipelines in autonomous navigation and control. Recent surveys indicate that foundation models may unify perception, reasoning, planning, and low-level control within a single learned architecture, potentially enabling more adaptive behavior in open and dynamic environments [90,91]. However, the principal open challenges remain the provision of formal safety guarantees, uncertainty-aware decision making, explainability and trustworthiness of learned policies, and the establishment of auditable certification pathways for safety-critical deployment [92].

8.5. Multi-UAV Collaborative Mapping

Swarm SLAM, distributed mapping, and cooperative obstacle avoidance are emerging research directions that extend single-UAV LiDAR autonomy to teams of aerial robots capable of sharing sensing, localization, map representations, and computational resources. Recent distributed collaborative SLAM frameworks demonstrate that inter-UAV loop closures, decentralized pose-graph optimization, and bandwidth-aware map exchange can enable scalable multi-robot operation in GPS-denied environments [93,94,95]. Such collaborative mapping capabilities are particularly relevant to large-scale search-and-rescue, disaster-response, and aerial survey missions, where multiple UAVs can simultaneously explore unknown environments, fuse local maps into a consistent global representation, and coordinate obstacle avoidance and exploration strategies [96].

8.6. Simulation-to-Reality Transfer

Because many learning-based planners are trained primarily in simulation, developing high-fidelity LiDAR simulators that realistically model ray-casting noise, atmospheric degradation, and material reflectivity is expected to accelerate the safe deployment of learned policies on physical hardware, and lightweight, point-realistic simulators purpose-built for LiDAR-equipped UAVs are already being developed for this purpose [97].

9. Comparative Analysis of Representative Systems

Table 1 synthesizes representative LiDAR-based UAV systems discussed throughout this review into a single comparison, contrasting their principal contribution, main limitation, and supporting evidence. The rows follow the structure of the paper: point-cloud perception and dynamic-object handling (Section 2), three-dimensional mapping, distance-field representations, and LiDAR-based SLAM (Section 3), reactive and optimization-based obstacle-avoidance and trajectory-planning algorithms, including learning-based policies (Section 5), and the multi-sensor fusion, foundation-model, and multi-UAV directions introduced in Section 8. The comparison illustrates a general trade-off: systems that minimize computation and latency (FAST-LIO2, EGO-Planner) tend to omit back-end optimization or global map consistency, whereas systems that add robustness through multi-sensor fusion or safety-assured dual-trajectory planning (FAST-LIVO2, SUPER) incur additional sensing, fusion, or platform-integration overhead.

10. Conclusions

LiDAR has established itself as a robust, high-precision perceptual backbone for autonomous UAV navigation in GPS-denied and cluttered environments, supporting a pipeline that spans point-cloud filtering, volumetric and distance-field mapping, tightly coupled odometry, and real-time obstacle avoidance. The literature reviewed here shows a consistent trajectory: reactive geometric methods remain valuable for their low computational cost, but the field has progressively moved toward map-centric, direct point-cloud, and increasingly learning-based approaches that reduce latency and improve robustness in cluttered or high-speed flight. At the same time, the strongest-performing systems no longer treat LiDAR as a standalone solution; instead, they pair it with inertial, visual, radar, or event-based sensing, and increasingly with learning-based or foundation-model-based exploration and trajectory prediction, to overcome the well-documented limitations of LiDAR under geometric degeneracy, adverse weather, and fast dynamic obstacles. The main open challenges – onboard computational constraints on micro-UAVs, atmospheric performance degradation, reliable tracking of fast dynamic obstacles, and the lack of standardized benchmarks — define a clear research agenda. Progress on solid-state LiDAR hardware, kinodynamically aware trajectory optimization, richer multi-sensor fusion, safety-certifiable learning-based navigation, multi-UAV collaborative mapping, and high-fidelity simulation-to-reality transfer is likely to jointly shape the next generation of LiDAR-based UAV autonomy.