LiDAR-Based Autonomous UAVs

• LiDAR-based autonomous UAVs are advanced unmanned aircraft that use diverse LiDAR sensors for precise navigation, localization, and mapping in complex operational settings.
• They employ various sensor types—including spinning, solid-state, and 2D LiDAR—to optimize field of view, range, and cost-efficiency in low-light or GNSS-denied contexts.
• Their autonomy pipelines fuse techniques like scan-matching, semantic segmentation, and reinforcement learning to enhance target tracking, obstacle avoidance, and mission task planning.

LiDAR-based autonomous unmanned aerial vehicles are UAV systems in which LiDAR is a primary sensing modality for perception, state estimation, mapping, planning, control, or task execution, especially in GNSS-denied environments and other settings where illumination is poor, terrain is cluttered, or scene geometry is operationally critical (Ren et al., 12 Sep 2025). Across recent work, the term encompasses spinning multi-beam LiDARs, solid-state LiDARs, LiDAR-as-camera sensors that expose fixed-resolution panoramic images in addition to point clouds, lightweight 2D scanners, and hybrid LiDAR–camera–IMU stacks. The resulting autonomy pipelines range from scan-matching localization and LiDAR-inertial odometry to semantic fusion, dynamic obstacle prediction, model-predictive control, inspection-task scheduling, and safe-site selection for landing or docking (ik et al., 2023, Wu et al., 24 Apr 2025).

1. Sensor modalities and airborne system architectures

The LiDAR sensing layer in autonomous UAVs spans multiple hardware classes. A recent survey distinguishes mechanical spinning LiDAR, solid-state LiDAR, MEMS-based and flash LiDAR, and FMCW LiDAR, emphasizing the trade space between field of view, density, range, SWaP, cost, and sampling uniformity (Ren et al., 12 Sep 2025). In the UAV literature summarized here, spinning sensors are represented by Ouster OS1-64 and OS0-128 units, while solid-state sensing is represented by Livox Mid-360 and Livox Avia; parsimonious platforms also employ 2D scanners such as Scanse Sweep v1, SLAMTEC RPLIDAR A3, and RPLidar A1, as well as paired tilted 2D LiDARs for tunnel navigation (Catalano et al., 2023, Sier et al., 2023, ik et al., 2023, Tardioli et al., 2024).

Configuration	Representative properties	Representative use
Spinning 3D LiDAR	OS1-64: 360° × 45°, 20 Hz, up to 120 m	Multi-LiDAR UAV tracking (Catalano et al., 2023)
LiDAR-as-camera	OS0-128: 360° × 90°, 10 Hz, 1024×128 panoramic images	YOLOv5-based UAV tracking (Sier et al., 2023)
Solid-state 3D LiDAR	Mid-360: 360° × 59°, 30 Hz or 100 Hz depending on system, non-repetitive scan pattern	Dynamic avoidance, forest flight, inspection (Wu et al., 24 Apr 2025, Liu et al., 2024)
Solid-state 3D LiDAR	Avia: roughly circular 70.4° × 77.2°, 100 Hz, up to 450 m	High-performing subset in UAV tracking benchmarks (Catalano et al., 2023)
2D LiDAR	360° single-plane scans at 5–10 Hz, or two ±45° tilted units at 60 Hz	Dark indoor navigation, tunnel flight, low-cost autonomy (ik et al., 2023, Tardioli et al., 2024, Wu et al., 2023)

Mounting geometry is not incidental. In the tunnel-navigation system, two identical 2D LiDARs are rigidly mounted with the top sensor tilted upward by +45° about the UAV’s $y$-axis and the bottom sensor tilted downward by –45°; this configuration breaks the sign-ambiguity that would arise if both planes were vertical (Tardioli et al., 2024). In LiDAR-as-camera tracking, the OS0-128 provides both point clouds and panoramic images encoding depth, signal strength, or ambient near-infrared, enabling a single-sensor detection-and-tracking stack (Sier et al., 2023). In dense-vegetation slope inspection and dynamic-obstacle avoidance, the Mid-360 is rigidly fixed on the UAV and integrated with IMU and onboard compute, typically an Intel NUC class computer running FAST-LIO2, ROG-Map, and MPC-based planning modules (Wu et al., 24 Apr 2025, Liu et al., 2024).

Calibration and synchronization recur as structural requirements. In the multi-LiDAR tracking dataset, extrinsics were found by manually measuring planar or corner features, integrating ten scans from each solid-state LiDAR to densify the scene, and aligning to the OS1-64 reference frame via Generalized ICP; camera–LiDAR extrinsics were derived from depth cloud alignment via ICP plus optional checkerboard procedures (Catalano et al., 2023). Systems using LiDAR-inertial pipelines rely on IMU-assisted deskewing, on-board PTP or estimator timestamps, and known rigid transforms between LiDAR, IMU, and body frames (Catalano et al., 2023, Liu et al., 2024).

2. Localization, odometry, and mapping

LiDAR-based UAV autonomy is anchored in state estimation. One canonical formulation models the LiDAR measurement as a range–bearing observation of a world point from a sensor pose $X \in \mathrm{SE}(3)$, with the union of returns back-projected into a point cloud; the same survey places feature-based LOAM, tightly coupled LiDAR-inertial odometry, ICP variants, EKF-SLAM, and particle-filter methods within the standard estimation taxonomy (Ren et al., 12 Sep 2025). In practice, recent UAV systems overwhelmingly adopt LiDAR–IMU coupling.

FAST-LIO2 is the dominant front end in several fielded systems. In dynamic obstacle avoidance, raw LiDAR and IMU are sent to FAST-LIO2 at 100 Hz to produce the point cloud in body frame before dynamic segmentation, map construction, and control (Wu et al., 24 Apr 2025). In dense-vegetation slope inspection, FAST-LIO2 delivers 200 Hz pose and 30 Hz registered point clouds, with the state vector $x = [p, v, q]^T$ and an iterated error-state EKF implementation (Liu et al., 2024). In autonomous inspection, FAST-LIO2 outputs use IMU preintegration at 100 Hz and LiDAR scan registration at 30 Hz, minimizing a point-to-plane objective

$$E_{\text{LIO}}(\xi) = \sum_{i,j} \bigl\|\,n_i^{\top}\bigl(T(\xi)\,p_j - \hat p_i\bigr)\bigr\|^2$$

with respect to the 6-DoF pose update $\xi$ (Liu et al., 29 Mar 2025).

Where 3D LiDAR is unavailable or undesirable, scan-matching localization with lightweight 2D sensors remains viable. The dark-indoor system uses a single-plane 2D scanner, sequential matching for relative displacement, global matching against an octree-based map, and Kalman fusion with IMU, rangefinder, barometer, and compass to recover the full 6-DoF state (ik et al., 2023). Its scan-matching core solves for the 2D rigid transform

$$\xi = \begin{bmatrix}\omega & T_x & T_y\end{bmatrix}$$

by minimizing an ICP-style cost, then improves robustness through dual correspondences, FRMSD outlier rejection, and distance weighting (ik et al., 2023). In underground tunnels, a closely related architecture combines a horizontal 2D RPLidar A1, a vertical TFmini altitude sensor, and IMU, and compares Hector SLAM, Gmapping, and Cartographer. The reported offline 2D translation RMSE shows Hector SLAM at 9.4 cm for linear nominal motion and 24.7 cm for circular fast motion, while onboard 3D UAV flights stayed below the 20 cm target accuracy across circular and “8-shape” trajectories (Chow et al., 2019).

Mapping layers built on top of odometry differ by mission. Occupancy-grid and log-odds updates dominate reactive navigation and inspection. ROG-Map with static/dynamic layers is used in dynamic obstacle avoidance, while vegetation-oriented systems add unknown-grid inflation, infinite points ray casting, and incremental frontier updates to improve conservative avoidance in under-canopy scenes (Wu et al., 24 Apr 2025, Liu et al., 2024). Semantic mapping changes the representation itself: in large-scale under-canopy flight, trunks are encoded as cylinders $c=(q,u,r)$ and ground as planes $\pi=(n,d)$, reducing a 1 km × 1 km forest with roughly 11,000 trunks to $\lesssim 2$ MB, compared with $\sim 1.25$ GB for a 0.1 m voxel ESDF over the same area (Liu et al., 2021).

3. Perception, target tracking, and semantic understanding

A major branch of LiDAR-based UAV research concerns target-centric perception rather than only ego-motion. In GNSS-denied UAV tracking, the multi-LiDAR multi-UAV dataset provides indoor and outdoor sequences with a spinning LiDAR, two solid-state LiDARs, and an RGB-D camera, spanning micro-aerial vehicles through standard commercial UAV platforms and including outdoor trajectories flown within $X \in \mathrm{SE}(3)$0 m of building façades (Catalano et al., 2023). The paper’s baseline study is notable less for a best-case result than for its negative conclusion: current methods have shortcomings and are unable to track UAVs consistently across different scenarios (Catalano et al., 2023).

That conclusion is supported by the benchmark structure. Dynamic Multi-Frequency Scan Integration, particle-filter or clustering+PF methods, and LiDAR-as-camera YOLO-style detection are compared across sensors and UAVs; on a subset of trajectories, the Avia-based EKF achieves 0.04–0.18 m RMSE with 100% success rate, the Avia-based KF achieves 0.03–0.13 m RMSE with approximately 80% success, while Mid-360 methods mostly diverged or achieved at most 20% success, with RMSE up to 0.10 m when they converged (Catalano et al., 2023). A plausible implication is that scan-pattern dependence is not a secondary implementation issue but a first-order algorithmic constraint.

The LiDAR-as-camera paradigm makes that dependence explicit. Using an Ouster OS0-128, a UAV can be detected in a 1024 × 128 reflectivity panorama via YOLOv5 and then localized by projecting the detected ROI back into the 3D point cloud, removing the ground plane, clustering with DBSCAN, and smoothing the observed centroid through a constant-velocity Kalman filter with state

$X \in \mathrm{SE}(3)$1

In three indoor sequences with maximum UAV–LiDAR distance up to 8 m, the fused approach reports mean APE 0.061 m and RMSE 0.067 m, compared with 0.104 m and 0.142 m for point-cloud-only tracking and 0.078 m and 0.088 m for image-only tracking; on a Jetson Nano, the fused pipeline runs at 6.04 Hz (Sier et al., 2023).

Semantic perception extends LiDAR autonomy beyond geometry. A multi-modal UAV system fuses LiDAR semantic segmentation, RGB semantic segmentation, and RGB/thermal object detection in a late-fusion architecture that augments 3D point clouds, image masks, and a sparse voxel semantic map (Bultmann et al., 2021). LiDAR scans are projected into 64 × 512 range images for SalsaNext, while RGB segmentation uses DeepLab v3+ / MobileNet v3 and detections use MobileDets; onboard output is approximately 9 Hz for fused semantic images and point clouds (Bultmann et al., 2021). Quantitatively, raw LiDAR segmentation without normalization adaptation yields mIoU = 17.4%, normalization adaptation raises this to 23.8%, fused semantic point clouds reach 28.3% over the full 360° LiDAR FoV, and within the camera FoV the fused mIoU is 48.9%, against 54.2% for pure image overlay (Bultmann et al., 2021). This indicates both the value and the boundary condition of fusion: it can recover small dynamic objects missed by image-only segmentation, yet camera FoV still limits the fused result.

Dynamic-scene perception adds explicit motion reasoning. In cluttered and dynamic flight, M-detector classifies each point as static or dynamic through voxelized multi-frame features, then clusters dynamic points via DBSCAN and associates them over time with a Hungarian assignment cost

$X \in \mathrm{SE}(3)$2

Per-object states are propagated under constant-velocity or free-fall models before entering the planner (Wu et al., 24 Apr 2025). At the semantic-mapping end of the spectrum, the forest semantic SLAM system detects tree trunks with a trellis graph and cylinder fitting, ground planes through coarse circular-grid binning, and uses these landmarks to constrain pose and compensate drift in real time (Liu et al., 2021).

4. Planning, control, and decision-making

LiDAR-based autonomy couples perception to action through planners and controllers that are typically kinodynamic and corridor-constrained. In tunnels, a proof-of-concept system derives yaw relative to the tunnel axis from two tilted 2D LiDARs by projecting compensated points to the UAV $X \in \mathrm{SE}(3)$3–$X \in \mathrm{SE}(3)$4 plane, rasterizing them into a 60 × 80 × 3 tensor, and regressing yaw with a compact CNN trained for approximately 400 epochs on approximately 120,000 samples (Tardioli et al., 2024). After yaw compensation, the tunnel cross-section is reduced to a 2D polygon and the safest lateral point is computed as the center of the maximal inscribed circle: $X \in \mathrm{SE}(3)$5 Closed-loop control then uses $X \in \mathrm{SE}(3)$6 for yaw and commands $X \in \mathrm{SE}(3)$7, $X \in \mathrm{SE}(3)$8, with forward speed adapted from the ratio between a safety-circle radius and the estimated inscribed radius (Tardioli et al., 2024).

More general obstacle avoidance systems rely on integrated planning and control. DynIPC formulates dynamic obstacle avoidance as a quadratic program over a finite horizon, with state trajectory $X \in \mathrm{SE}(3)$9, control sequence $x = [p, v, q]^T$0, kinodynamic bounds, Safe Flight Corridor constraints, and a cost that combines reference tracking, control effort, control smoothness, terminal penalties, and soft repulsion from predicted obstacles (Wu et al., 24 Apr 2025). On an i7 NUC, the QP is solved in approximately 3–5 ms, contributing to a total loop of approximately 11.5 ms at a 100 Hz back end (Wu et al., 24 Apr 2025). Closely related IPC stacks appear in dense-vegetation slope inspection and autonomous cluttered-environment inspection, where A* supplies a reference path, SFC generation yields convex free-space constraints, and MPC plus differential flatness produces body-rate and thrust references (Liu et al., 2024, Liu et al., 29 Mar 2025).

Task planning can sit above local control. In the dual-phase inspection system, a human-in-the-loop phase supports untrained pilots with autonomous obstacle avoidance while they create a map and record inspection poses $x = [p, v, q]^T$1; the autonomous phase then orders the points via a Traveling Salesman Problem over A*-derived pairwise distances and executes the route (Liu et al., 29 Mar 2025). In wind-turbine inspection, LiDAR model matching localizes the turbine by fitting a pre-generated thin-cylinder model to accumulated point clouds, after which a waypoint sequence is generated at offset distances and heights, with a collision-avoidance condition $x = [p, v, q]^T$2 and camera triggering delegated to a gimballed laser range finder (Sikora et al., 2023).

Reinforcement learning provides a different decision layer in parsimonious systems. RELAX constructs a 2D occupancy map with Hector-SLAM, plans a static route with RRT, and then uses a Dueling Double DQN online re-planner whose state concatenates goal-relative geometry and eight sector-wise LiDAR minima, with eight discrete planar motion actions (Wu et al., 2023). In the reported dynamic-obstacle benchmark, RELAX(D3QN) reaches a 90% success rate with online re-planning time per step in $x = [p, v, q]^T$3 s, compared with 98% for a FAST-LIO-based 3D-LiDAR baseline and 100% for a Deep PANTHER RGB-D baseline (Wu et al., 2023). This suggests that learning-based online correction can narrow, but not erase, the capability gap between 2D-LiDAR-only and richer perception stacks.

5. Operational domains and validated applications

The application envelope of LiDAR-based autonomous UAVs is unusually broad, but it is not homogeneous. Tunnel and mine navigation emphasize the absence of GNSS, sparse visual texture, and limited geometric diversity. In the tilted-2D-LiDAR tunnel system, a smooth “double-S” tunnel with radius = 2 m permits $x = [p, v, q]^T$4 rad up to 4 m/s, while ten laps in an 8-shaped 1 m-radius tunnel at 3 m/s remain within $x = [p, v, q]^T$5 m cross-track error; in a rough “triple-S” tunnel of 260 m length, average forward speed adapts to approximately 2.4 m/s, with a minimum of 1 m/s and no collisions (Tardioli et al., 2024). In dark indoor environments, scan-matching localization works in total darkness and across churches, warehouses, forests, and tunnels, with Chlumin church experiments reducing global ATE from 1.0642 m for raw ICP to 0.0979 m after the full pipeline of filtering, interpolation, IMRP, weighting, and FRMSD (ik et al., 2023).

Under-canopy and dense-vegetation flight impose different constraints: thin obstacles, false unknown cells, repetitive trunks, and limited clearances. In the large-scale semantic-SLAM forest system, a 1.1 km manual loop reduces total drift from 10.10 m for VIO-only to 3.99 m for SLOAM + VIO, and the same framework supports guided coverage over a 0.77 km² forest with a 10 km path and 10,897 cylinder landmarks (Liu et al., 2021). A later lightweight under-canopy platform using LTA-OM SLAM and IPC reports, after optimization, success rates of 12/15 in a medium-density forest and 15/15 in a dense forest at a target flight velocity of 1 m/s, and 12/15 and 5/15 respectively at 2 m/s; 60 m missions in the optimized system average 78.7 s in medium-density and 86.5 s in difficult-density forest at 1 m/s (Karhunen et al., 16 Dec 2025). In a separate dynamic-avoidance system, real-world forest trials in a 30 m corridor achieve 100% success against tennis balls, pendulums, moving cars, and pedestrians, with observed maximum UAV speed 4.55 m/s and relative speed up to 8.75 m/s (Wu et al., 24 Apr 2025).

Inspection tasks add mission-specific geometry. For slope inspection in dense vegetation, a Mid-360-equipped quadrotor with FAST-LIO2, log-odds occupancy mapping, unknown-cell inflation, and 100 Hz IPC was deployed in six field environments, including five flexible debris-resisting barriers and one landslide site; trajectory lengths ranged from 79 to 426 m and flight times from 3:43 to 8:28 min, with reported localization drift below 0.1 m RMS over 100 m trajectories and 100% obstacle-avoidance success in narrow-area and dynamic tests (Liu et al., 2024). In the dual-phase autonomous inspection system, autonomous operations reduce trajectory length by 14–33% and flight time by 10–57% across slope, landslide, agriculture, factory, and forestry scenarios, relative to the human-in-the-loop phase (Liu et al., 29 Mar 2025). Wind-turbine inspection extends LiDAR autonomy to rotating infrastructure: at Pometeno Brdo wind farm, ICP matching between point clouds and a turbine model yields scores 0.262, 0.251, and 0.278 across three runs, with score $x = [p, v, q]^T$6 interpreted as very good alignment and successful autonomous image acquisition without collisions (Sikora et al., 2023).

Landing is another task domain in which LiDAR changes the autonomy problem from navigation to terrain assessment. A camera–LiDAR fusion system accumulates non-repetitive LiDAR returns over adaptive time, completes a dense depth map, evaluates it via SSIM-based self-assessment, derives slope and roughness, and extracts the largest inscribed safe landing circle (Chen et al., 2020). On 20 test flights in known and unknown terrains, the UAV consistently found a safe 2 m landing circle, with no crashes and average selection latency of approximately 2 s, including confirmation (Chen et al., 2020).

6. Limitations, recurrent misconceptions, and open research problems

A recurrent misconception is that LiDAR alone guarantees robustness in GNSS-denied flight. The evidence is more specific. The multi-LiDAR tracking benchmark shows that current methods are unable to track UAVs consistently across different scenarios, and that algorithms tuned for dense solid-state Avia data do not transfer cleanly to sparse Mid-360 data, while spinning-LiDAR methods were not fully evaluated because the baseline code was not tailored to them (Catalano et al., 2023). In dynamic inspection, current static-map approaches explicitly identify dynamic obstacles such as people and vehicles as a limitation and propose dynamic-object filtering and reactive local planning as future work (Liu et al., 29 Mar 2025).

A second misconception is that 2D LiDAR is categorically insufficient for autonomy. The record is mixed rather than dismissive. The dark-indoor system achieves sub-decimeter global ATE in a church dataset and fully autonomous SE(3) stabilization using a lightweight 2D scanner, but it also states that 2D LiDAR provides no direct altitude observability and that scan rate and point count limit high-speed maneuvers (ik et al., 2023). RELAX reaches 90% success in dynamic navigation with a single 2D LiDAR, yet remains below 3D-LiDAR and RGB-D baselines in the same benchmark (Wu et al., 2023). The tunnel system demonstrates that two tilted 2D LiDARs can be sufficient for straight and curved tunnel navigation, but only under a geometry-specific pipeline that exploits tunnel cross-sections and yaw-relative image regression (Tardioli et al., 2024).

A third misconception is that wider FoV or higher point rate alone resolves perception difficulty. The tracking literature argues otherwise: limited FoV adaptability, non-repetitive scan gaps, time skew in multi-modal fusion, and the very small number of returns from small, fast UAVs all remain failure modes (Catalano et al., 2023). Likewise, semantic fusion improves some classes while preserving others’ blind spots; persons behind the UAV, outside camera FoV, remain unlabeled or misclassified due to LiDAR-only model limitations (Bultmann et al., 2021).

Open problems are correspondingly consistent across papers. Several works call for scan-pattern-agnostic algorithms, adaptive FoV control, learned sensor fusion from raw LiDAR “images” and RGB, real-time calibration updates, integration of aerodynamic models or flight-control telemetry, more conservative loop-closure gating in under-canopy SLAM, and standardized field protocols with at least $x = [p, v, q]^T$7 flights per condition for reproducible forest evaluation (Catalano et al., 2023, Karhunen et al., 16 Dec 2025). The survey broadens this agenda to SWaP-constrained sensors, online spatio-temporal calibration, hybrid LiDAR+RGBD+Radar fusion, multi-agent cooperative SLAM, and learning-based planners that can exploit LiDAR geometry without discarding hard safety structure (Ren et al., 12 Sep 2025). Taken together, the literature suggests that LiDAR-based autonomous UAVs are best understood not as a single architecture, but as a family of tightly coupled sensing, estimation, and decision systems whose reliability depends on how well sensor physics, motion models, map representations, and task geometry are matched.