Livox Avia: Solid-State LiDAR Insights

• Livox Avia is a low-cost, solid-state LiDAR characterized by a non-repetitive Risley-prism scan pattern, a concentrated forward-facing FoV, and a 240 kHz sampling rate.
• Its irregular, center-heavy sampling challenges traditional clustering, scan matching, and geometry-only registration methods, impacting runtime and accuracy in SLAM pipelines.
• In UAV tracking and localization, Livox Avia serves as a secondary, sensor-tuned modality that provides focused 3D point cloud data to complement other LiDAR systems.

Livox Avia is a low-cost solid-state LiDAR that recent work characterizes as a representative non-repetitive sensor with a forward-facing field of view, a strongly concentrated scan pattern, and point distributions that differ substantially from both repetitive mechanical LiDARs and wider-coverage Livox configurations. In the literature surveyed here, it appears both as an operational sensor in UAV tracking and localization pipelines and as a benchmark modality for studying how limited FoV, irregular sampling, and center-heavy density affect clustering, scan matching, and SLAM. The resulting picture is technically consistent: Avia can be effective in long-range sensing and in sensor-tuned localization pipelines, but its non-uniform sampling and limited instantaneous coverage make it a demanding case for geometry-only registration and for algorithms that assume structured scan lines or natural frame boundaries (Schroeder et al., 17 Jun 2026, Felix et al., 6 Jul 2025).

1. Sensor modality and reported specifications

Livox Avia is described as a solid-state LiDAR with a non-repetitive scan pattern, and one paper treats it specifically as a non-repetitive Risley-prism LiDAR with a strongly concentrated scan pattern and a forward-facing field of view. The same body of work reports a sampling rate of 240 kHz and a FoV of $70.4^\circ \times 77.2^\circ$. A comparative dataset paper reports the same FoV, lists a max range of up to 450 m, power of 8–9 W (16 peak), weight of 498 g, and the onboard IMU as BMI088. In that dataset, Avia point clouds contain the fields x, y, z, intensity, tag, line (uint8), the IMU topic is /avia/livox/imu, and the IMU rate is 200 Hz (Schroeder et al., 17 Jun 2026, Felix et al., 6 Jul 2025).

Property	Reported value	Source
Sensor type	solid-state; non-repetitive Risley-prism LiDAR	(Schroeder et al., 17 Jun 2026, Felix et al., 6 Jul 2025)
Sampling rate	240 kHz	(Schroeder et al., 17 Jun 2026)
Field of view	$70.4^\circ \times 77.2^\circ$	(Schroeder et al., 17 Jun 2026, Felix et al., 6 Jul 2025)
Max range	up to 450 m	(Felix et al., 6 Jul 2025)
Power	8–9 W (16 peak)	(Felix et al., 6 Jul 2025)
Weight	498 g	(Felix et al., 6 Jul 2025)
IMU / rate	BMI088 / 200 Hz	(Felix et al., 6 Jul 2025)

These reported specifications already encode the principal trade-off that recurs throughout the literature. Avia combines comparatively long nominal range with a narrower, forward-facing sensing volume and a non-repetitive sampling process. This suggests that its utility depends less on raw range alone than on how downstream algorithms accommodate its spatial non-uniformity.

2. Sampling geometry and algorithmic consequences

The Avia scan pattern is reported to have very high point density near the center of the FoV and much sparser coverage toward the periphery. Because it is non-repetitive, papers emphasize that there are no discrete scan lines and no natural frame boundaries. In clustering research, this breaks two assumptions commonly built into methods for repetitive mechanical LiDARs: structured scan lines and well-defined frame segmentation. One consequence is that spherical or range-grid projections can produce cells with very large local occupancy $k$, and the C-ARC paper states that insertion cost scales as $O(k)$, where $k$ is the number of points per cell. The same work defines an occupancy matrix $\mathbf{F} \in \mathbb{N}_0^{H \times W}$ and active cells $\mathcal{V} = \{ (u, v) \mid F_{u,v} > 0 \}$, precisely to quantify how non-uniform projection density affects clustering behavior (Schroeder et al., 17 Jun 2026).

The central empirical conclusion of that work is explicit: “Evaluation on the Livox Avia identifies unbounded cell occupancy as the primary limitation for sensors with strongly concentrated scan patterns.” In the reported runtime study, all experiments use raw point clouds with no preprocessing such as ground removal or downsampling, clustering retrieval runs at 20 Hz with a 50 ms per-cycle budget, and Avia uses $\theta = 80^\circ$. Under those conditions, the forest sequence stays below 50 ms at $P_{99}$ for all tested buffer sizes, while the indoor sequence exceeds 50 ms for $T_B \ge 0.3$ s. The paper attributes the indoor degradation to the interaction between center-heavy sampling and planar structures such as walls and floors, which increases local connectivity, cluster merging, and BFS retrieval cost. It therefore proposes a per-cell point cap as the most effective mitigation, but presents that only as future optimization rather than an implemented component (Schroeder et al., 17 Jun 2026).

A common misconception is that runtime on Avia is determined mainly by total point count. The cited clustering analysis argues otherwise: for Avia, runtime is driven less by global throughput than by local projection density and topological connectivity. That distinction is important for method design, because coarse global downsampling does not by itself resolve concentrated occupancy near the sensor’s central viewing cone.

3. Use in UAV tracking and pose estimation

In the CL-Det system for the CVPR 2024 UG2+ Challenge Track 5, Livox Avia is used alongside Livox Mid-360, called “LiDAR 360” in the paper. Avia’s role is explicitly secondary. The authors describe it as providing “focused 3D point cloud data for a specific timestamp” that “usually represent[s] the origin point or the drone position.” The algorithmic priority is also explicit: the method first uses LiDAR 360, falls back to Livox Avia if LiDAR 360 is unavailable, and if neither sensor is available it uses the drone’s average location derived from training datasets, given as $70.4^\circ \times 77.2^\circ$0 (Xiao et al., 2024).

The synchronization rule is intentionally simple. The pipeline first aligns Livox Avia and LiDAR 360 timestamps “to ensure temporal coherence,” and the implementation retrieves the nearest timestamp for each sequence in the test set. For Avia-specific preprocessing, the paper does not describe the environment-removal and DBSCAN steps used for LiDAR 360. Instead, it removes all points at $70.4^\circ \times 77.2^\circ$1 as noise and computes the mean of the remaining points in $70.4^\circ \times 77.2^\circ$2 as the UAV position estimate. The paper also clarifies that DBSCAN is not applied to Livox Avia point clouds in the described method. The overall method achieved Pose MSE = 120.215107, Classification accuracy = 0.322, 5th place on the final leaderboard, and 14.9 predictions/second, but it does not provide a separate quantitative ablation for Avia alone (Xiao et al., 2024).

This operational usage places Avia in a specific systems niche: not the primary structural sensor, but a targeted backup modality that can still yield a direct 3D position estimate from compact returns. A plausible implication is that the authors viewed Avia’s point clouds as sufficiently target-centered to support mean-based localization, but not as the main source for scene-scale clustering.

4. Ground-focused localization pipelines

The GroundLoc pipeline uses Livox Avia as one of its evaluated LiDAR configurations on the HeLiPR dataset and explicitly highlights Avia as a “non-conventional scanning pattern” sensor. The preprocessing stage is sensor-tuned. GroundLoc first segments each point cloud into ground and non-ground points using GroundGrid and rasterizes only the ground points into a sensor-centered BEV grid. For Avia, the reported GroundGrid parameters are $70.4^\circ \times 77.2^\circ$3, $70.4^\circ \times 77.2^\circ$4, $70.4^\circ \times 77.2^\circ$5, $70.4^\circ \times 77.2^\circ$6, $70.4^\circ \times 77.2^\circ$7, $70.4^\circ \times 77.2^\circ$8, $70.4^\circ \times 77.2^\circ$9, $k$0, $k$1, $k$2, and $k$3. The image-level normalization parameters are $k$4, $k$5, and $k$6. The BEV image $k$7 is built from three channels,

$k$8

$k$9

$O(k)$0

with slope defined as

$O(k)$1

Only ground-classified points are projected, which the paper states suppresses moving objects and stabilizes the representation in dynamic scenes (Steinke et al., 28 Oct 2025).

The localization stages after BEV construction are sensor-agnostic in principle, but the paper reports that Avia benefits strongly from learned descriptors. In single-session matching on HeLiPR KAIST04, Avia achieves 0.25 m / 0.32° / 99.70% success with SIFT and 0.30 m / 0.28° / 99.80% success with R2D2. In multi-session matching on KAIST06→evaluation, performance separates more sharply: SIFT yields 8.05 m / 35.65° / 54.15% success, whereas R2D2 yields 3.47 m / 16.02° / 81.57% success. In end-to-end localization with GroundLoc + R2D2, Avia reaches 0.57 m / 0.66° on Roundabout, 0.54 m / 0.68° on Town, and 0.57 m / 1.11° on Bridge. With SIFT, the Bridge sequence fails catastrophically at 60.00 m / 2.22°. The paper attributes the Bridge difficulty to highway driving, long bridges, repetitive structures, and higher speeds up to 90 km/h; it also notes that KISS-ICP exhibits significant pitch drift on sensors without 360° FOV, including Avia. In cross-sensor localization against an Ouster OS2-128 prior map, Avia with R2D2 obtains 0.53 m / 0.58° on Roundabout, 0.49 m / 0.63° on Town, and 0.71 m / 1.08° on Bridge, indicating that a denser map can still support robust localization for a less conventional live sensor (Steinke et al., 28 Oct 2025).

This body of evidence shows that Avia is not intrinsically incompatible with large-scale outdoor localization. Instead, successful deployment depends on sensor-specific preprocessing and descriptor choice; the paper makes clear that transfer across sensor models is harder than transfer across locations.

5. Comparative SLAM and geometry-only registration

The paper “Lidar Variability: A Novel Dataset and Comparative Study of Solid-State and Spinning Lidars” uses Livox Avia as the representative low-cost solid-state lidar with limited FoV in a dataset that also includes Livox Mid-360 and Ouster OS0-128. The sequences are IndoorOffice1, IndoorOffice2, and OutdoorRoad, all recorded as ROS rosbag data. Within this benchmark, Avia is not presented as the uniformly strongest sensor. Rather, it is the modality that exposes how non-repetitive scan pattern and limited FoV affect both SLAM and scan registration (Felix et al., 6 Jul 2025).

For SLAM, the paper benchmarks FAST-LIO2, FASTER-LIO, S-FAST-LIO, GLIM, and FAST-LIO-SAM using Absolute Pose Error (APE). Indoors, Avia generally performs worse than Mid-360 and often worse than Ouster, with larger error variances, especially under GLIM. Reported indoor APE values include 0.1436 ± 0.1390 for FAST-LIO2 and 0.1223 ± 0.0517 for FASTER-LIO on IndoorOffice1, and 0.1019 ± 0.0536 for S-FAST-LIO on IndoorOffice2; by contrast, GLIM reaches 1.1587 ± 0.6535 on IndoorOffice2. Outdoors, Avia’s best SLAM result is FASTER-LIO at 0.3013 ± 0.0818 m on OutdoorRoad, which the paper identifies as the only case where Avia beats both Mid-360 and Ouster. Its worst reported outdoor SLAM result is GLIM at 1.7026 ± 0.2774 m. The paper’s qualitative summary is therefore asymmetric: Avia can excel in long-range outdoor SLAM, but it has higher variance and weaker local consistency than Mid-360.

The registration benchmark is less favorable. The same study compares KISS-ICP (point-to-point), GenZ-ICP (hybrid/weighted), and Open3D-GICP in Scan2Map and Scan2Scan variants. The standard objectives used are

$O(k)$2

$O(k)$3

and

$O(k)$4

For Avia, the paper reports 0.1348 ± 0.1049 for KISS-ICP on IndoorOffice1 and 0.1121 ± 0.0868 for GENZ-ICP on IndoorOffice2, but Scan2Scan GICP deteriorates to 1.4341 ± 0.2804 and 1.4504 ± 0.9441 on those same scenes. Outdoors, Avia underperforms across all configurations, often with the highest error and variability, including 1.0489 ± 0.9901 and 1.0357 ± 0.8246 for Open3D-GICP Scan2Scan on OutdoorRoad cuts. The paper explicitly links this fragility to sparse vertical resolution, limited vertical coverage, and irregular scan structure (Felix et al., 6 Jul 2025).

One recurrent misconception is that Avia’s up to 450 m nominal range should predict superior SLAM or odometry. The comparative evidence rejects that simplification. In this benchmark, long range does not remove the penalties imposed by limited FoV and irregular local geometry.

6. Relation to other Livox solid-state research

Several papers do not evaluate Avia directly but are explicitly relevant to Avia-class sensing because they address the same family of non-repetitive, compact Livox LiDAR behaviors. A UAV mapping study using Livox Mid-40 describes that sensor as a robotic prism-based LiDAR with an incommensurable scanning pattern, a rosette scanning pattern, and a peak angular density at the center of the FOV, “likened to the fovea of the human retina.” It reports 100,000 points/s, 38.4° circular FoV, 2 cm range precision, 0.1° angular precision, and FoV coverage of about 20% at 0.1 s and 93% at 1 s. The authors argue that such Livox sensors are suitable for UAV mapping because UAV scenarios do not require 360° FoV, while compactness, low cost, and density accumulation over time are advantageous; they do not analyze Avia directly, but state that the discussion is broadly relevant to Avia-like platforms (Tan et al., 2020).

A LiDAR-inertial odometry paper centered on Livox Horizon reaches a similar conclusion from the algorithmic side. It states that special feature extraction is required for solid-state LiDAR because the scan pattern is irregular and non-repetitive, and proposes a time-domain patch-based feature extractor followed by a tightly coupled LiDAR-IMU backend. The paper is explicit that this logic should be conceptually applicable to Avia, but that one should retune the patching and threshold parameters rather than assume Horizon settings transfer directly (Li et al., 2020).

Likewise, the adaptive filtering paper on lightweight UAV tracking validates only on Livox Mid-360, yet repeatedly frames its contribution around sparse, noisy, nonuniform point clouds from non-repetitive scanning 3D LiDARs under strict payload and power constraints. It reports that the method is conceptually well matched to Avia-type sensing conditions, while also stating that transfer to Avia is reasonable but not experimentally demonstrated (Khosravi et al., 10 Mar 2026).

Taken together, these adjacent papers support a cautious generalization. Avia belongs to a sensor class for which compact form factor and non-repetitive scanning are often assets in UAV deployment, but those same properties require sensor-aware preprocessing, robust state estimation, and methods that do not depend on organized scan lines. That broader pattern is consistent with the Avia-specific evidence from clustering, localization, and comparative SLAM.