Non-Repeating Scanning LiDAR (NRS-LiDAR)

• Non-Repeating Scanning LiDAR (NRS-LiDAR) is a sensor technology that employs non-repetitive beam trajectories via MEMS or Risley prisms to generate continuously updated point clouds.
• Its unique scanning mechanics overcome uniformity assumptions, improving real-time perception, calibration, and multi-sensor integration for robotics and mapping.
• Cost-effective and innovative, NRS-LiDAR offers adaptive scanning that introduces both enhanced coverage and new challenges in algorithm design and extrinsic calibration.

Non-Repeating Scanning LiDAR (NRS-LiDAR) refers to a class of light detection and ranging sensors that produce temporally and spatially non-repetitive beam trajectories across their field of view (FoV). Unlike classical multi-line, mechanically rotating LiDARs that generate periodic, ring-structured point clouds, NRS-LiDARs employ mechanisms such as MEMS or Risley prisms to steer a single beam along pseudo-random or Lissajous-style patterns. This scanning paradigm disrupts traditional assumptions about scan-line uniformity and frame granularity, introducing unique challenges and advantages for perception, calibration, and real-time robotics.

1. Fundamental Scanning Mechanisms and Mathematical Formulation

NRS-LiDAR systems utilize non-repetitive actuation schemes—most commonly Risley prisms or MEMS mirrors driven at incommensurate frequencies—to achieve a continuously evolving, non-cyclic scan pattern. Let $\omega_1$, $\omega_2$ be the (irrational) angular velocities of orthogonal prism axes. The beam direction at time $t$ is typically parameterized via “rose” (Rhodonea) or Lissajous curves: $$\phi(t) = A_\phi \sin(n t), \quad \theta(t) = A_\theta \cos(m t)$$ with $m/n \approx \omega_2/\omega_1$ and amplitudes $A_\phi$, $A_\theta$ covering the device’s FoV $[\Theta_{\max}, \Phi_{\max}]$. Each emitted pulse at time $t_i$ thus lands at a unique (azimuth, elevation) not generally revisited within a practical time horizon.

The instantaneous point cloud exhibits non-uniform, centrally densified distributions, with coverage progressively filling in over longer accumulations. For a planar target of area $A$ at distance $\omega_2$0, the hit density follows

$\omega_2$1

yet, due to non-repetition, the distribution varies spatially within each frame and temporally over successive integrations. This inhomogeneity invalidates conventional "ring" indexing and scan-line accumulation used in repetitive LiDAR segmentation and clustering (Schroeder et al., 17 Jun 2026).

2. Algorithmic and System-Level Distinctions

Feature	Repetitive LiDAR	Non-Repetitive LiDAR
Scanning mechanism	Multi-laser + spinner	Single beam, dual-axis MEMS or Risley prism
FoV	360° × 40° (typ.)	$\omega_2$270–98° × $\omega_2$338–77°
Point rate	1.5–2.3M pts/s (128-line)	0.2–0.24M pts/s
Frame granularity	Well-defined revolutions	No canonical frame; continuous streaming
Instantaneous density	High, ring-structured	Low, spatially uneven, increases with time
Blind zones	Yes (between fixed lines)	None (progressive fill over FOV)
Cost	$\omega_2$410–20$\omega_2$5 higher	Typically $\omega_2$6–$\omega_2$7 k USD for Livox Avia/Mid-360

This paradigm shift demands new algorithmic strategies for feature extraction, clustering, calibration, and multi-sensor fusion. For instance, the absence of canonical frame boundaries and ring indices precludes per-scan connected-components (CC) segmentation or planar image projection strategies standard with mechanical LiDARs (Schroeder et al., 17 Jun 2026, Qi et al., 28 Oct 2025).

3. Adaptive Perception and Clustering Methods

Real-time perception using NRS-LiDAR necessitates methods robust to irregular, streaming input, fluctuating grid cell occupancy, and dynamic coverage. The C-ARC (Continuous-Adaptive Range Clustering) framework establishes the state-of-the-art for real-time instance clustering under such constraints (Schroeder et al., 17 Jun 2026):

• Persistent Dual-Graph Sliding Window: Points are collected in a fixed-length buffer, simultaneously updating a fine-grained point graph $\omega_2$8 (single-linkage local clusters) and a coarse component graph $\omega_2$9 (tracking high-level cluster connectivity).
• Adaptive Grid Initialization: Grid resolution is adjusted online using an exponential control loop optimizing for density, gap, and mean bucket multiplicity, based on live occupancy statistics; this balances fragmentation versus cell-collision trade-off.
• Lazy Deletion and Deferred Retrieval: Algorithmic decoupling of high-frequency point insertion from cluster label publication (e.g., at 20 Hz) ensures bounded latency and reduced computation.
• Empirical Performance: On Livox Mid-360 ($t$0200 kHz), C-ARC in single-threaded C++17 achieves $t$1 ms for buffer lengths up to 0.5 s. It outperforms grid-based 4-conn CC methods by avoiding "black holes" and fragmentation under sparse coverage.

Limitations include unbounded bucket occupancy for highly concentrated scan centers (e.g., Livox Avia), contributing to $t$2 per-insertion cost and cluster overgrowth (Schroeder et al., 17 Jun 2026).

4. Applications in Localization, Mapping, and Robotic Perception

NRS-LiDARs are increasingly applied in infrastructure-based vehicle localization, multi-sensor fusion SLAM, reflectance imaging, and adaptive scanning:

• Infrastructure Localization: At the infrastructure scale, Bench-RNR demonstrates that multi-session, multi-modality ground-truth datasets (dual Livox Avia, Hesai OT128, GNSS/IMU) enable direct benchmarking of NRS-LiDAR in diverse scenes. Template-based ICP registration achieves mean pose errors of $t$37 cm and $t$4 with NRS-LiDAR, matching high-end repetitive units at much lower cost (Zhao et al., 19 Sep 2025).
• Large-Scale Map Construction: GM-Livox integrates up to six NRS-LiDARs with IMU, wheel encoder, and RTK to achieve robust, real-time pose estimation and mapping via factor graph optimization with time-synchronized feature point fusion and keyframe-based sliding-window marginalization (Wang et al., 2021).
• Vision-Like Reflectance Imaging: By leveraging the cumulative quasi-random filling, dense camera-like reflectance images can be synthesized from NRS-LiDAR for loop closure and lane detection, with compensation modules rectifying angular, range, and incidence angle distortions. Densification network architectures (U-Net backbones with Adaptive Fusion Modules, Dynamic Compensation Modules) have been shown to produce $t$5\% lane detection accuracy under day/night and adverse illumination conditions at $t$620 Hz (Gao et al., 14 Aug 2025).
• Adaptive Scanning and Energy Efficiency: Predictive adaptive scanning strategies (temporal cue exploitation, query-based mask generation via Gumbel-Softmax) have achieved 65\% LiDAR energy reduction without degrading object detection mAP (≤1\% loss). This is accomplished by targeting dense pulses to predicted ROIs and sparsely covering background, validated across nuScenes and Lyft (Shoouri et al., 3 Aug 2025).

5. Performance Benchmarks and Practical Trade-Offs

Extensive simulation (InfraLiDARs’ Benchmark in CARLA) and real-world evaluations reveal nuanced advantages and limitations relative to repetitive scanning systems (Qi et al., 28 Oct 2025):

• Detection Range and Density: NRS-LiDAR’s slow decay in point density ($t$7, $t$8) at long range leads to superior detection of distant targets ($t$990 m), often yielding twice as many points on vehicles at range compared to 128-line repetitive systems.
• Average Precision (BEV-AP): Across common 3D object detectors (PointRCNN, PointPillars, PV-RCNN, DSVT), BEV-AP on highway, crossroad, and curve scenarios is comparable between high-end NRS and 128-line units, especially in long-range and wide-area monitoring.
• Cost and Coverage: NRS-LiDAR provides orders-of-magnitude cost savings (e.g., Livox Avia at $$\phi(t) = A_\phi \sin(n t), \quad \theta(t) = A_\theta \cos(m t)$$0\,USD 2\,k vs. Ruby Plus at $$\phi(t) = A_\phi \sin(n t), \quad \theta(t) = A_\theta \cos(m t)$$1\,USD 24\,k), but at the expense of limited instantaneous horizontal FoV ($$\phi(t) = A_\phi \sin(n t), \quad \theta(t) = A_\theta \cos(m t)$$277°), necessitating careful system-level placement for applications requiring 360° awareness.
• Algorithmic Robustness: Classical geometric pose-fitting degrades under instantaneous sparsity/uniformity; learning-based or model-registration techniques are more resilient to the non-uniform, progressive fill of NRS-LiDAR (Zhao et al., 19 Sep 2025, Qi et al., 28 Oct 2025).

6. Calibration and Multi-Sensor Integration

The lack of regular scan lines and highly irregular single-frame sampling present unique calibration challenges for extrinsic sensor alignment:

• Checkerboard-Based Camera–LiDAR Calibration: ACSC achieves automatic extrinsic calibration by fusing temporarily accumulated NRS-LiDAR frames, extracting planar checkerboard corners via spatial–temporal filtering and registering intensity reflectance patterns to known binary models. Optimizing reflectance-and-geometry-based costs enables sub-pixel error ($$\phi(t) = A_\phi \sin(n t), \quad \theta(t) = A_\theta \cos(m t)$$32.1 px average normalized reprojection error) exceeding prior work on solid-state devices (Cui et al., 2020).
• Synchronization for Fusion: Synchronizing asynchronous NRS-LiDARs with IMU, odometry, and RTK (e.g., via GNSS-PPS hardware pulses and unified software timestamping) is essential for consistent fused point clouds and accurate trajectory estimation in multi-device mapping systems (Wang et al., 2021).

7. Limitations and Future Directions

NRS-LiDAR systems are inherently constrained by:

• High Occupancy at FOV Centers: Dense central accumulation can cause degeneracy in cell occupancy, increasing computational cost and degrading cluster resolution.
• Limited FoV: Mechanical simplicity trades off against peripheral coverage, restricting surveillance unless multiple units are integrated.
• Absence of Canonical Frames: Streaming, pattern-agnostic clustering precludes certain algorithmic primitives (e.g., scan-line-based index lookups), requiring new paradigms for perception pipelines.

Research avenues include bucket occupancy bounding, cache-optimized data structures, semantic–geometric cluster refinement, and adaptive multi-sensor fusion frameworks. The emergence of real-time, robust, and pattern-agnostic clustering frameworks such as C-ARC opens viable paths toward deploying NRS-LiDAR on cost-sensitive, high-performance robotic and intelligent transportation platforms (Schroeder et al., 17 Jun 2026).