ABSTRACT

LiDAR is widely regarded as indispensable for autonomous vehicles and safety-critical robotics. Yet a candid examination of current technologies reveals that neither the dominant pulsed Time-of-Flight (ToF) approach nor the widely promoted Frequency-Modulated Continuous Wave (FMCW) architecture is ready — as currently conceived — for mass deployment in life-critical applications. This essay systematically reviews the performance requirements that safety-critical LiDAR must meet, dissects the fundamental error sources and failure modes of both ToF and FMCW architectures, evaluates the limitations of sensor fusion with cameras and radar, and confronts the unresolved challenges of beam steering, coherent chirp generation, and uncertainty quantification. The conclusion is not pessimism but urgency: the field requires a new technical paradigm — spanning signaling physics, illumination strategy, detection architecture, and data processing — before LiDAR can be trusted with human lives at scale.

1. What Safety-Critical LiDAR Actually Requires

Before critiquing existing architectures, we must be precise about what the application demands. The phrase 'safety-critical' is not rhetorical decoration — it carries specific, quantifiable engineering obligations. An autonomous vehicle operating at highway speed is a system in which sensor failure can produce irreversible harm in under one second. This reality defines a set of non-negotiable performance characteristics that any candidate LiDAR technology must satisfy simultaneously and unconditionally.

1.1 Range and Spatial Resolution

A highway-capable autonomous vehicle requires reliable detection of pedestrians and objects at a minimum range of 200 metres, with sufficient resolution to distinguish a child from a traffic cone at that distance. This demands angular resolution on the order of 0.1° or better, and range resolution of better than 5 centimetres. These are not aspirational targets — they are derived from stopping distances, reaction times, and the physical size of vulnerable road users. Neither requirement is trivially met under real-world conditions.

1.2 Absolute and Relative Accuracy

Range accuracy must be expressed as absolute error bounds — not percentages. A meaningful specification requires a root mean square error (RMSE) of less than 3–5 centimetres under defined operating conditions, with statistical confidence bounds explicitly stated. Relative accuracy within a single scan — the internal consistency of the point cloud — must be even tighter, since downstream perception algorithms depend on the geometry of object surfaces. A system that is accurate on average but noisy shot-to-shot produces point clouds that are geometrically unreliable at the object level.

1.3 Update Rate and Latency

At 100 km/h, a vehicle travels approximately 28 metres per second. Sensor update rates below 10 Hz introduce positioning uncertainty that cascades into perception and planning errors. Real systems must deliver dense, complete point clouds at 20 Hz or better, with bounded and deterministic latency from photon return to processed output. Jitter in the processing pipeline — not just in the timing electronics — is a safety hazard.

1.4 Environmental Robustness

Safety-critical LiDAR must operate across the full envelope of conditions the vehicle will encounter: heavy rain, dense fog, driving snow, direct solar illumination, extreme temperature excursions from -40°C to +85°C, and high-humidity environments. Critically, the system must not merely degrade gracefully — it must accurately report its own degradation so that the vehicle's safety architecture can respond appropriately. A sensor that fails silently is more dangerous than one that fails loudly.

1.5 Functional Safety and Fault Detectability

ISO 26262 Automotive Safety Integrity Level D (ASIL-D) is the highest functional safety classification in automotive engineering. Achieving it requires rigorous fault detection, diagnostic coverage exceeding 99%, and clearly defined safe states for every failure mode. This is not a software problem alone — it demands that the hardware signal chain be designed from the outset with testability, redundancy, and failure mode analysis embedded at every level. SOTIF (ISO 21448, Safety Of The Intended Functionality) adds a further requirement: the system must not cause harm even when operating exactly as designed, covering the vast space of edge cases where the design itself is insufficient.

1.6 Longevity and Calibration Stability

Automotive-grade components must survive a minimum service life of ten years or 150,000 miles under conditions of vibration, thermal cycling, moisture ingress, and UV exposure. For a LiDAR system, this requirement extends to the stability of its internal calibration. Any drift in the geometric relationship between emitter and detector, or in the timing characteristics of the signal chain, translates directly into growing range error over time — error that may be invisible to the operator until it causes a failure.

2. The Fundamental Error Sources of Pulsed Time-of-Flight

Pulsed ToF LiDAR has been the dominant commercial architecture for two decades. It operates on a conceptually elegant principle: fire a laser pulse, measure the time until the echo returns, and compute distance as half the round-trip travel time multiplied by the speed of light. In practice, this simplicity conceals a cascade of physical error sources that become increasingly difficult to manage as performance requirements tighten.

2.1 Timing Jitter: The Fundamental Limit

Because distance in ToF is computed as d = (c × Δt) / 2, any uncertainty in Δt — the measured round-trip time — maps directly into range error. The speed of light is approximately 3 × 10⁸ m/s, which means that 1 nanosecond of timing uncertainty corresponds to 15 centimetres of range error. Achieving centimetre-level range accuracy therefore requires timing precision in the sub-100 picosecond regime — sustained, across temperature, and over the component lifetime.

Timing jitter arises from multiple sources in the signal chain. The laser driver circuit introduces jitter at the moment of pulse emission — the pulse does not fire at the exact commanded instant. The detector — typically an avalanche photodiode (APD) or a single-photon avalanche diode (SPAD) — introduces further jitter at the moment of photon detection, as the avalanche process is inherently stochastic. The time-to-digital converter (TDC) that digitises the time interval has its own clock jitter. These contributions are statistically independent and add in quadrature, meaning all sources must be controlled simultaneously.

A single nanosecond of uncontrolled timing jitter produces 15 cm of range error. Centimetre-level accuracy demands sub-100 ps precision — maintained over temperature, aging, and varying signal levels.

2.2 Range Walk: The Signal-Strength Dependency

In threshold-based ToF systems, the detector triggers when the received optical power exceeds a fixed threshold. Because the amplitude of the return pulse varies with target reflectivity and range, the threshold crossing point shifts relative to the true pulse peak — a phenomenon called range walk. A highly reflective target at short range triggers the detector earlier than a dark target at the same range, introducing a systematic range offset that depends on conditions the system cannot directly control. Range walk corrections can be applied, but they require accurate knowledge of the return pulse amplitude, which is itself noisy and condition-dependent. At the margins — dark targets, long range, adverse weather — the correction uncertainty grows and becomes a dominant error source.

2.3 Multi-Return Ambiguity and Ghost Points

When a laser pulse encounters a semi-transparent surface — rain droplets, fog, glass — it can produce multiple returns: one from the obstruction and one from the surface behind it. Distinguishing valid multi-returns from noise returns requires sophisticated pulse processing, and errors in this classification produce either ghost points (false objects) or missed returns (missed objects). In dense precipitation, the density of interference returns can overwhelm the signal from the actual scene, producing a point cloud that appears plausible but is geometrically wrong. The system typically cannot self-diagnose this condition from the point cloud alone.

2.4 Solar Background and Crosstalk

Direct sunlight at the laser wavelength (typically 905 nm or 1550 nm) introduces a shot-noise floor that degrades the signal-to-noise ratio at long range and reduces the maximum reliable detection distance. In high-ambient-light conditions, performance can degrade by 20–40% relative to nighttime specifications. Additionally, in future high-density deployment scenarios — urban intersections with dozens of LiDAR-equipped vehicles — pulse interference between systems becomes statistically significant. Although FMCW partially addresses this, pulsed ToF has essentially no architectural immunity.

2.5 Calibration Drift and Aging

The timing characteristics of laser drivers, APDs, and TDCs change with temperature and age. A system calibrated at manufacture may have accumulated significant systematic range error after several years of thermal cycling. Without active, real-time calibration monitoring — which most deployed systems lack — this drift is invisible. It manifests not as a sudden failure but as a gradual, silent degradation of accuracy that falls outside the sensor's self-reported confidence bounds.

3. FMCW LiDAR: A Promising Architecture with Unresolved Foundations

Frequency-Modulated Continuous Wave LiDAR has attracted enormous research and investment interest, and for good reasons. By replacing the timing measurement of ToF with a frequency measurement, FMCW appears to offer a path around the timing jitter problem. It provides simultaneous velocity measurement via the Doppler effect, offers inherent immunity to pulsed interference from other LiDAR systems, and is compatible with silicon photonics integration. However, FMCW does not eliminate fundamental uncertainty — it relocates it, from the timing domain to the frequency domain, where equally formidable unsolved problems await.

3.1 The Chirp Linearity Problem

FMCW LiDAR measures range by mixing the outgoing chirped laser signal with the delayed return and measuring the beat frequency. The beat frequency is proportional to target range only if the frequency chirp is perfectly linear over time. In practice, generating a highly linear optical frequency chirp across a bandwidth of several gigahertz at an optical carrier frequency of approximately 200 THz (1550 nm wavelength) is one of the most demanding challenges in applied photonics.

The relationship between laser drive current and output frequency is nonlinear and hysteretic. Temperature shifts the lasing mode. The cavity response has memory. Any deviation from perfect linearity broadens the peak in the range FFT, reducing range resolution and introducing systematic range bias that varies with target distance, scene content, and thermal state. Pre-distortion of the drive waveform can compensate for known nonlinearities, but the nonlinearity itself drifts with temperature and aging, making static pre-distortion insufficient for automotive-grade deployment.

FMCW does not eliminate fundamental uncertainty — it relocates it from the timing domain to the frequency domain. A perfectly linear chirp at 200 THz carrier is not a solved engineering problem at automotive scale.

3.2 Laser Coherence Length Requirements

FMCW requires that the laser coherence length exceed twice the maximum measurement range. For a 200-metre range requirement, the laser must maintain temporal coherence over a path length of at least 400 metres. This corresponds to an instantaneous laser linewidth of less than approximately 750 kHz — a specification that pushes into the regime of ultra-narrow linewidth lasers. Such lasers exist in laboratory and telecommunications settings, but their cost, sensitivity to environmental perturbation, and susceptibility to mode hops make them extremely challenging to deploy at automotive scale with the robustness required for ASIL-D compliance.

3.3 Real-Time Linearisation: An Open Research Problem

The most credible engineering response to chirp nonlinearity is real-time feedback linearisation using an auxiliary interferometer with a known reference path. By measuring the actual chirp trajectory in real time and correcting it, the system can achieve much better linearity than static pre-distortion allows. However, this approach introduces significant additional complexity: a reference interferometer, a high-bandwidth feedback loop operating at optical frequencies, and a control system that must remain stable across the full temperature and vibration envelope of an automotive deployment. The latency of the feedback loop creates a window during which the chirp is uncorrected, and rapid temperature transients can exceed the loop's tracking bandwidth. This is an active area of research, not a solved engineering problem ready for mass production.

3.4 Mode Hops and Frequency Discontinuities

Semiconductor lasers are susceptible to mode hops — discrete, instantaneous jumps in emission frequency caused by changes in cavity conditions. A mode hop during a chirp sweep corrupts the entire range measurement for that interval, producing either a false target or a missed detection. Mode hops are difficult to predict, difficult to detect in real time with low latency, and their frequency increases with temperature extremes and aging. For a system that must operate reliably across -40°C to +85°C over ten years, mode hops represent a failure mode with no robust mitigation in current semiconductor laser architectures.

3.5 Silicon Photonics Integration: Promise and Reality

Much of the excitement around FMCW is coupled to the prospect of integrating the entire optical front end onto a silicon photonics chip, enabling low cost and high-volume manufacturing. Silicon photonics has achieved remarkable results in optical communications. However, the requirements of LiDAR impose constraints that communication photonics does not face: high output power, 1550 nm or 905 nm emission, beam shaping for a free-space application, and environmental robustness far beyond data centre conditions. The integration of laser sources with the linewidth requirements of FMCW LiDAR onto silicon remains a fundamental materials challenge. Silicon does not lase. Hybrid integration of III-V gain materials with silicon waveguides introduces coupling losses and thermal management challenges that scale poorly with the temperature range and longevity requirements of automotive deployment.

4. The Beam Steering Problem: An Independent and Unsolved Dimension

A critical conceptual error pervades much public discussion of LiDAR technology: the conflation of the measurement modality (ToF vs. FMCW) with the beam steering mechanism (spinning vs. solid-state). These are orthogonal dimensions of system design. FMCW says nothing about how the beam covers the scene. Eliminating mechanical spinning does not imply FMCW. Yet both transitions are necessary for automotive-grade reliability, and each carries its own set of unsolved problems.

4.1 Why Mechanical Steering Must Be Eliminated

Rotating mechanical assemblies — whether full sensor rotation as in early spinning-sensor designs, or rotating mirror systems — introduce multiple categories of failure risk. Bearing wear is a time-dependent failure mode that is difficult to detect before it affects performance. Mechanical resonances change with temperature and age, introducing scan pattern distortions that corrupt the geometric accuracy of the point cloud. Electrical connections through rotating joints — slip rings or flex cables — are sites of intermittent failure that are notoriously difficult to diagnose. From an ISO 26262 perspective, each mechanical element requires a diagnostic strategy, a failure mode analysis, and a defined safe state, multiplying the complexity of the safety case.

Beyond reliability, mechanical steering limits the scan rate to the physical rotation speed, imposes constraints on scan pattern flexibility, and makes the system large and expensive. These limitations alone would motivate solid-state alternatives even without the reliability argument.

4.2 MEMS Mirrors: Small Moving Parts Are Still Moving Parts

Micro-electromechanical systems (MEMS) mirrors replace the macro-scale mechanical assembly with a microscale silicon mirror that is electrostatically or electromagnetically actuated. This eliminates gross mechanical wear but introduces its own failure modes. MEMS mirrors have resonant frequencies that change with temperature, altering the scan pattern in ways that must be continuously corrected. The mirror surface can degrade under UV and particulate exposure. The drive electronics must maintain precise amplitude and phase control to keep the scan pattern stable — a control loop with its own failure modes. MEMS devices are also susceptible to shock and vibration in ways that macro-scale mechanics are not: a brief mechanical impulse can drive the mirror outside its designed oscillation regime and corrupt multiple scan frames before recovery.

4.3 Optical Phased Arrays: The Calibration Nightmare

Optical phased arrays (OPAs) steer the beam by controlling the phase of optical emission from an array of antenna elements, with no moving parts. In principle, this offers unlimited reconfigurability, extremely fast beam switching, and full solid-state reliability. In practice, OPAs for LiDAR face challenges that have prevented deployment at automotive scale despite years of intensive research. Each element in the array must be phase-controlled with sub-wavelength precision — errors of a few nanometres translate into beam steering errors and sidelobe growth. At 1550 nm, thermal expansion coefficients of silicon create phase errors of tens of wavelengths across the automotive temperature range, requiring continuous active phase correction of every element. The control system for a large OPA is itself a complex, power-hungry, potentially failure-prone subsystem. Sidelobes — secondary beam peaks arising from array periodicity — create false targets at predictable angular offsets, which must be distinguished from real objects in the perception pipeline.

Eliminating mechanical steering and adopting FMCW are two separate transitions, each with its own unsolved engineering challenges. A system can be FMCW with a spinning mirror, or solid-state with pulsed ToF. The combination of both is the goal — but neither solves the other's problems.

4.4 The Case for Staring or Flash Architectures

If the objective is a sensor with no beam steering mechanism — eliminating bearings, MEMS resonances, and OPA phase control loops simultaneously — the architectural direction leads toward staring or flash LiDAR designs. Flash LiDAR illuminates the entire scene at once and uses a two-dimensional detector array to capture the full point cloud in a single acquisition event. This eliminates all beam steering complexity and reduces the signal chain to a fixed, deterministic optical path that can be fully characterised and monitored. However, flash LiDAR faces severe challenges in range performance, eye safety at the required illumination power, and the cost and noise performance of large-format avalanche detector arrays. These challenges are formidable but have a clearer engineering path than the phase control problems of OPAs or the coherence problems of FMCW. They deserve far more research investment than they currently receive.

5. The Limits of Sensor Fusion: Cameras and Radar

Proponents of current LiDAR architectures frequently respond to concerns about LiDAR limitations by invoking sensor fusion: the combination of LiDAR with cameras and radar to produce a perception system that is more robust than any single sensor. This argument is valid in principle and important in practice. However, it is too often used to defer rather than solve the fundamental problems with LiDAR. The limitations of cameras and radar are not complementary to LiDAR's limitations in the clean, additive way that this argument implies.

5.1 Cameras: Rich in Information, Fragile in Condition

Cameras provide dense spectral and texture information — lane markings, traffic signs, facial expressions of pedestrians — that LiDAR fundamentally cannot. Machine learning perception systems trained on camera images have achieved remarkable performance in controlled conditions. However, cameras share several failure modes with human vision and add some of their own. Direct solar glare can saturate an image sensor in milliseconds. Rain on the lens or condensation on a cold morning degrades image quality in ways that are difficult to detect algorithmically. Dynamic range limitations mean that a camera adapting to a bright scene misses detail in shadows, and vice versa. Most critically, cameras provide no direct range measurement — depth must be inferred from stereo disparity, optical flow, or learned priors, each of which introduces uncertainty that is difficult to bound precisely.

5.2 Radar: Reliable in Weather, Poor in Resolution

Radar operates at centimetre wavelengths and is largely immune to precipitation and fog — conditions that severely degrade both LiDAR and cameras. It provides direct velocity measurement via Doppler shift, which is highly valuable for tracking. However, current automotive radar has angular resolution an order of magnitude worse than LiDAR, making it unable to resolve the geometry of complex scenes. It detects the presence of objects reliably but cannot distinguish their shape, orientation, or the fine structure needed for classification and trajectory prediction. Radar also suffers from multipath reflections in dense urban environments, producing ghost targets and missed detections that are geometrically plausible and therefore difficult to reject.

5.3 The Fusion Problem: When Sensors Disagree

The fundamental unsolved problem in sensor fusion for autonomous driving is not combining sensors when they agree — it is deciding what to believe when they disagree, under conditions where the disagreement itself may not be detectable. Consider a scenario in which heavy rain has degraded the LiDAR point cloud to the point where it shows a clear road ahead, while the camera sees a pedestrian through a wet lens at reduced confidence. The fusion algorithm must adjudicate between a high-confidence but wrong LiDAR reading and a low-confidence but correct camera detection. No current production system handles this class of problem reliably, because no current system can accurately estimate the confidence of its own outputs under arbitrary degradation conditions.

Sensor fusion does not make a broken sensor safe — it makes the safety case more complex. Robust fusion requires that each sensor can accurately quantify its own uncertainty under degraded conditions. This capability does not exist in current deployed systems.

5.4 Correlated Failure Modes

Sensor fusion provides maximum safety benefit when the sensors have independent failure modes. However, several critical failure conditions degrade multiple sensors simultaneously. Dense fog scatters both LiDAR pulses and camera images. Heavy precipitation reduces LiDAR range and compromises camera imaging at the same time. A direct solar glare event may simultaneously saturate the camera and reduce LiDAR signal-to-noise ratio. In precisely the conditions where the vehicle most needs reliable sensing — poor visibility, adverse weather — the redundancy benefit of multi-modal fusion is reduced. The sensors are not independent; they share a physical environment whose failure modes are correlated.

6. The Meta-Problem: Sensors That Do Not Know What They Do Not Know

Underlying all of the specific technical challenges described above is a single meta-problem that may be the most important unsolved issue in LiDAR for safety-critical applications: current LiDAR systems do not provide accurate, real-time estimates of the uncertainty associated with each measurement they produce.

A LiDAR point cloud is typically delivered as a set of (X, Y, Z) coordinates, perhaps accompanied by intensity values. There is no per-point confidence estimate, no indication of whether the return was a strong echo from a nearby retroreflector or a weak, noisy signal from a dark surface at maximum range, no quantification of how much timing jitter or chirp nonlinearity affected that particular measurement. The downstream perception system — and ultimately the vehicle's safety system — must treat all points as equally trustworthy, because the sensor provides no basis for differentiation.

This is not a minor omission. In aviation — a field that has spent decades developing frameworks for safety-critical system design — a navigation system that does not report its own integrity is not certifiable, regardless of its average accuracy. The GPS system provides not just position but a Protection Level: a bound on position error that is valid with a specified probability. No equivalent concept exists in production automotive LiDAR.

The most dangerous failure mode is not a sensor that stops working — it is a sensor that continues to produce plausible but wrong outputs without any indication that something has changed. Current LiDAR systems have no architectural mechanism to prevent this.

7. The Case for a New Paradigm

The analysis above is not a counsel of despair. It is a diagnosis. Each of the failure modes and limitations described has a physical origin, and physical origins imply physical solutions. However, those solutions cannot be found by incrementally improving existing architectures. The field needs a deliberate reframing — a new technical paradigm that addresses the root causes rather than their symptoms.

7.1 Signalling: From Single-Mode to Information-Rich Illumination

Current LiDAR systems extract a single piece of information per pulse or per chirp sweep: range, and in the FMCW case, velocity. A new paradigm should design the illumination waveform to carry more information — about the target's reflectivity, about the atmospheric path, about the signal quality of the return — simultaneously with the range measurement. This suggests multi-dimensional coding of the illumination: not a simple pulse or a simple linear chirp, but a waveform designed for both ranging and integrity monitoring. The illumination design and the detection design must be co-optimised from the outset, rather than treating the laser as a dumb emitter and placing all the intelligence in the detector.

7.2 Detection: From Intensity to Coherent State Estimation

Current detection architectures measure when light arrives (ToF) or what frequency shift it carries (FMCW), but discard most of the information contained in the electromagnetic field of the return. A coherent detection architecture that measures both the amplitude and phase of the return field — not just integrated power — provides access to information about the scattering process, the target's surface structure, and the propagation path that intensity-only detection permanently discards. This richer measurement is the basis for per-measurement uncertainty quantification: the variance of the amplitude and phase estimates directly encodes the confidence in the derived range.

7.3 Illumination: Rethinking Coverage Without Steering

The beam steering problem suggests that the field should seriously reconsider whether sequential, single-point illumination is the right architecture for safety-critical LiDAR. An illumination strategy that simultaneously covers the full scene — through structured light, compressive sensing, or spatially coded flood illumination — eliminates the entire class of steering-related failures. The cost is illumination power budget and detector array complexity. But these are engineering challenges with known solutions in adjacent fields, unlike the fundamental physics problems of OPA phase control and FMCW chirp linearity. A new paradigm should treat scene-wide simultaneous acquisition as the baseline and steering as the exception, rather than the reverse.

7.4 Processing: From Point Clouds to Uncertainty Fields

The output of a LiDAR system should not be a point cloud — a list of positions. It should be an uncertainty field: a probabilistic representation of the scene in which every estimated position is accompanied by a covariance matrix encoding the sensor's confidence in that estimate, conditioned on the actual signal quality of that specific return. This requires a fundamental change in the processing pipeline, from threshold-based detection and fixed-formula range computation to Bayesian state estimation that propagates measurement uncertainty from photon arrival statistics through to 3D position estimates. The computational cost of this approach is non-trivial, but the safety benefit is transformative: for the first time, the fusion algorithm and the vehicle safety system would have a principled basis for deciding how much to trust each measurement.

7.5 Integrity Monitoring: Learning from Aviation

Aviation's Receiver Autonomous Integrity Monitoring (RAIM) concept should be a model for LiDAR system design. RAIM continuously monitors the internal consistency of GPS measurements and alerts the navigation system when the available data is insufficient to guarantee a position estimate within required accuracy bounds. An analogous system for LiDAR — continuously monitoring the internal consistency of the point cloud, the stability of calibration parameters, the signal quality statistics, and the agreement between redundant measurements — would provide the real-time integrity assessment that safety-critical applications require. This is not a post-processing step; it must be an integral part of the sensor architecture.

7.6 Materials and Manufacturing: Designing for the Automotive Lifetime

A new paradigm must also confront the materials science of longevity. Laser diodes degrade. Detector bias voltages drift. Optical surfaces accumulate contamination. A LiDAR system designed for ten-year automotive service must either use materials and processes with demonstrated stability over that period under automotive conditions, or include active compensation mechanisms that track and correct for aging-related changes in calibration. This requires bringing accelerated life testing, failure mode analysis, and calibration stability measurement into the design process from day one — not as an afterthought at the end of development.

8. Conclusion: The Honest Reckoning

The autonomous vehicle industry has, for a decade, operated under an implicit assumption: that LiDAR is essentially a solved problem and that the remaining challenges are matters of cost reduction and manufacturing scale. This assumption is wrong, and its persistence is a safety risk.

Pulsed ToF LiDAR suffers from fundamental timing jitter constraints, range walk, silent degradation under adverse conditions, and the absence of any real-time integrity monitoring. FMCW LiDAR replaces these problems with equally formidable challenges in chirp linearity, laser coherence, mode stability, and feedback linearisation — none of which are solved at automotive scale and operating conditions. Both architectures rely on beam steering mechanisms — mechanical, MEMS, or OPA — each of which introduces failure modes that are inconsistent with ASIL-D functional safety requirements. Sensor fusion with cameras and radar is essential but insufficient, because the sensors share correlated failure modes in precisely the worst-case conditions, and because no current system can accurately quantify its own measurement uncertainty under degradation.

The path forward is not a faster chirp or a smaller MEMS mirror. It is a fundamental rethinking of LiDAR architecture — from illumination strategy through detection physics to processing and integrity monitoring — guided by the uncompromising requirements of safety-critical operation and informed by the lessons of mature safety-critical fields like aviation. The technology to do this exists, in fragments, across photonics, signal processing, statistical estimation, and system engineering. What is needed is the intellectual honesty to acknowledge that current approaches are insufficient, and the engineering ambition to build something genuinely new.

Until that paradigm exists — validated not just in the laboratory but over automotive lifetimes and across the full envelope of real-world conditions — LiDAR-based autonomous systems should be deployed with the humility of engineers who know what they do not yet know, not with the confidence of marketers who have confused a promising prototype with a solved problem.

— End of Essay —