2026-06-18 👁️ — views

AV Weather Performance — Rain, Snow, Fog and the Sensor Limits Shaping AV Expansion

How rain, snow, and fog degrade lidar, camera, and radar differently — and why sensor physics explains the Sun Belt expansion pattern of every major AV company.

Article 104 in the Physical AI Benchmark Series — AV Weather Performance: How Rain, Snow, and Fog Degrade Each Sensor Type and Why Physics — Not Strategy — Explains Why Waymo Launched in Phoenix and Why Tesla Claims a Camera Advantage in Adverse Conditions

Adverse weather is one of the most frequently cited limitations of autonomous vehicles — and one of the most misunderstood. When the subject comes up in public discourse, it typically collapses into a single category: “bad weather.” But rain, snow, and fog are three distinct physical phenomena that degrade the four primary AV sensor types — camera, lidar, radar, and ultrasonic — in very different ways, at different rates, with different fallback options available to the autonomy stack.

That sensor physics, compounded across thousands of operating days per year, is the direct explanation for the geographic expansion pattern of every major AV company. Phoenix first. Sun Belt cities second. Northern cities with heavy winter snowfall last, if at all in the near term. This is not a strategic preference. It is a constraint imposed by the state of sensor technology, the economics of operating in conditions where effective sensor range drops from 200 meters to 10–20 meters (est.), and the training data limitations of fleets that have spent most of their miles in desert cities.

This article maps AV weather performance as a benchmark dimension in the Physical AI series. It covers the physics of each sensor type in adverse conditions, the specific advantages and failure modes of Waymo’s lidar-centric and Tesla’s camera-centric approaches, the geographic expansion pattern that follows from sensor limits, and what technical problems must be solved before any AV company can operate commercially in Chicago, Boston, or Buffalo.

Section 1 — How Weather Affects Each Sensor Type

The four sensors used in commercial AV platforms interact with weather through different physical mechanisms. Understanding those mechanisms is the foundation for understanding every geographic and technical decision AV companies make.

Why lidar struggles in fog but radar does not

The physics divergence between lidar and radar comes down to wavelength. Lidar operates at near-infrared wavelengths (905 nanometers or 1550 nanometers). Fog droplets — typically 1–100 micrometers in diameter — are large enough relative to these wavelengths to scatter laser pulses effectively, drastically reducing the signal that returns to the lidar receiver. A dense fog layer can reduce lidar effective range from 200 meters to less than 20 meters (est.), well below the stopping distance needed at highway speeds.

Radar at 77GHz operates at wavelengths of approximately 4 millimeters — much larger than fog droplets. At this wavelength, fog droplets are too small to scatter the radar signal meaningfully. The radar beam passes through fog with minimal attenuation. This is why Waymo’s sensor fusion architecture treats radar as the primary range sensor in low-visibility conditions, down-weighting lidar returns when fog confidence metrics trigger.

Camera systems sit between these extremes. Fog reduces optical contrast and shortens effective range, but a camera in moderate fog still produces a usable image with degraded range performance rather than a near-total loss of range as lidar experiences in dense fog.

Why rain is a different problem for lidar than fog

Rain affects lidar differently from fog. Rain drops are larger (0.5–5 millimeters), which means they intercept and return laser pulses directly rather than scattering them diffusely. This creates two problems: the laser pulse loses energy when it hits a raindrop, reducing range to the actual target behind the raindrop, and the raindrop itself returns a false point in the point cloud. Heavy rain therefore degrades lidar range (est. 200m → 50–100m in heavy rain) and adds spurious points that the perception pipeline must filter.

For cameras, rain creates a different challenge: water on the lens. A camera with rain droplets on its lens produces a blurred, distorted image regardless of what is happening in the scene. Heated camera housings and active lens cleaning systems partially address this, but heavy rain with windshield spray from other vehicles significantly degrades camera vision systems.

Radar is almost entirely unaffected by rain at 77GHz. Rain attenuation at this frequency is measurable but minor under all but the most extreme precipitation rates.

Section 2 — Why Phoenix Was Waymo’s First Commercial Market

The geographic selection of Phoenix as Waymo’s first commercial driverless market was not primarily a regulatory or population decision. It was a sensor physics decision.

The Phoenix weather profile means that on approximately 90%+ of operating days (est.), Waymo’s lidar and camera arrays perform in their optimal range. The weather-induced sensor degradation modes that challenge the perception pipeline — fog scatter reducing lidar range, rain on camera lenses, snow covering road markings — are rarely encountered. This allows the autonomy stack to develop operational confidence on the most common scenarios before being stress-tested by adverse weather.

The sequential expansion from Phoenix to Austin, Los Angeles, and San Francisco follows the same logic. Each city adds a weather challenge: Austin adds occasional heavy rain events. Los Angeles adds coastal fog. San Francisco adds frequent fog. Each step up the challenge ladder stress-tests the perception pipeline against weather scenarios not present in Phoenix, generates new training data for adverse conditions, and validates the sensor fusion architecture’s ability to degrade gracefully.

The notable gap in this progression is that no major AV company has yet expanded to cities where snowfall is a regular winter operational condition. Atlanta — which receives occasional light snow — is in Waymo’s expansion pipeline. Chicago, Boston, and Buffalo are not on any announced near-term timeline as of mid-2026 (est.).

Section 3 — Tesla’s Weather Claim: Vision-Only in Adverse Conditions

Tesla’s position in the weather performance debate is structurally different from Waymo’s. Tesla’s Full Self-Driving system uses cameras exclusively, with no lidar and (in recent hardware generations) no radar. This creates both a specific advantage and a specific weakness in adverse weather.

The training data asymmetry

The most significant Tesla weather advantage is not sensor physics — it is training data volume. Tesla’s consumer fleet is estimated at several million vehicles (est.), most of which drive in rain, snow, and fog as part of normal daily use. Every adverse weather mile driven by a Tesla with FSD enabled generates training data with the sensor context (camera images) labeled with what the correct driving behavior was (what the human driver did). This shadow mode data pipeline generates adverse weather training samples at a scale that Waymo’s commercial driverless fleet — operating in Sun Belt cities — cannot match.

Waymo’s advantage is sensor reliability in adverse conditions: radar as a weather-robust fallback for range sensing. But training data that teaches a neural network how to handle snow-covered roads, ice, and heavy fog requires actually operating vehicles in those conditions. Waymo must send mapping and validation vehicles to northern cities before launching commercial operations, while Tesla accumulates that data automatically through its existing consumer fleet.

This training data asymmetry is one of the most underappreciated structural differences between the camera-only and lidar-centric architectures when evaluated as a path to northern city deployment.

Section 4 — The Weather Performance Benchmark by Geography

The table above makes visible the underlying logic of AV geographic expansion. The “AV weather challenge” rating correlates directly with precipitation totals, snowfall, and fog frequency. Every city currently hosting commercial driverless operations sits at “Very Low” to “Medium” on this scale. No commercial driverless service operates in any “High” or above weather-challenge market.

Miami’s case is instructive: it has no snow, but 62 inches of rain per year including intense tropical downpours. Heavy rain is a different operational challenge from persistent snow, but it still stresses camera systems and generates standing water that affects vehicle dynamics. Miami’s status as a “medium-term” market for Waymo reflects both the rain challenge and the route complexity of South Florida driving, not just weather.

Section 5 — What Needs to Be Solved for Northern City Deployment

Northern city deployment represents the hardest open problems in AV weather performance. The technical challenges go beyond sensor degradation in precipitation — they include loss of road surface information that the autonomy stack depends on for localization and path planning.

The lane marking dependency problem

The most fundamental northern city challenge is not sensor degradation in precipitation — it is the loss of the road information the autonomy stack depends on most. Under heavy snow accumulation, lane markings disappear. Road surfaces become uniform white. The visual cues that camera and lidar systems use for localization and path planning within a lane are gone.

The solution architecture for this problem requires combining HD map coordinates with GNSS positioning to know where the road is even when visual lane markings are absent, using lidar point cloud analysis to detect road edges, curbs, and guardrails as lane boundary proxies, and validating those estimates against radar returns from roadside structures. This fusion approach is technically achievable but requires significantly more development work than the visual lane following that underlies most current commercial autonomy stacks.

Waymo’s HD map infrastructure — often cited as a limitation because it requires pre-mapping before deployment — becomes an advantage in this specific scenario. A vehicle that knows from its HD map exactly where a lane is located can continue operating even when visual lane markings are obscured, using the map as a ground truth that weather cannot erase. Tesla’s camera-only, mapless approach loses this fallback in heavy snow conditions.

The irony is that the weather scenario where HD maps are most advantageous — northern city winter operations — is also the scenario where HD maps are hardest to maintain (snow removal, temporary lane configurations, and road treatment change road appearance and sometimes road layout throughout winter months).

Section 6 — Weather Performance as a Physical AI Benchmark Dimension

Weather performance belongs in any serious Physical AI benchmark framework. It is not a binary capability (can operate / cannot operate) — it is a continuous degradation curve that varies by sensor type, precipitation intensity, weather phenomenon, and available fallback sensor architecture.

No AV company publicly discloses quantitative performance benchmarks across this weather matrix. The absence of such disclosure is consistent with the observation that no major AV company has launched commercial driverless service in any market requiring “High” or “Extreme” weather performance. The expansion pattern from Phoenix outward is the de facto published benchmark: the cities in which a company operates commercially define the weather performance envelope it has validated.

The gap between current commercial operation (approximately “Medium” weather challenge maximum) and the “Very High” to “Extreme” performance required for full US geographic coverage is significant. Solving it requires advances in sensor fusion for degraded lidar conditions, HD map maintenance through winter weather, lane marking–free path planning, and northern city training data at scale — all open engineering problems as of mid-2026.

Note: All performance estimates, range degradation figures, fleet sizes, city weather statistics, and operational timelines in this article are directional estimates based on publicly available information, published research, and industry analysis as of mid-2026. Figures labeled “(est.)” should not be treated as confirmed data. This article does not constitute investment advice.

Sources

• Waymo safety and all-weather operations — Waymo ↗
• Tesla FSD all-weather capability — Tesla ↗
• Lidar performance in adverse weather — SAE International ↗
• NOAA climate data by city — NOAA ↗
• Radar vs lidar in fog and rain — IEEE ↗

Tags