2026-06-18 — views

AV Adverse Weather Performance — Rain, Snow, Fog and Why Waymo Stays in Phoenix

Rain kills cameras, fog kills lidar, radar survives everything — why geography determines where driverless robotaxis can actually launch.

Article 62 in the Physical AI Benchmark Series — How Adverse Weather Explains the Map of AV Deployment

Ask why Waymo operates robotaxis in Phoenix but not Chicago, or why Tesla’s Robotaxi launched in Austin rather than Boston, and the most honest answer is not about software maturity or fleet economics. It is about weather. Adverse weather is one of the most direct explanations for why AV deployments cluster in specific geographies — and the interaction between different sensor types under rain, snow, and fog is the defining performance gap that constrains where commercial driverless services can launch.

The physics of sensor degradation in adverse weather is not speculative. Rain attenuates laser beams. Fog scatters lidar pulses. Snow accumulates on sensor housings and produces false returns. Camera optics go blind beyond a few meters in dense fog regardless of what software runs behind them. The key insight is that different sensors fail in completely different ways — and the stacking of sensors into a fusion architecture is precisely the engineering response to this fact.

Section 1 — How Each Sensor Degrades in Adverse Weather

Understanding adverse weather performance requires understanding what each sensor actually measures and how weather interferes with that measurement.

Sensor	Rain	Snow	Fog	Dust / Sand	Night
Camera (visible light)	Lens droplets blur image; rain streaks on windshield; contrast and feature detection degrade	Snow accumulates on lens; whiteout conditions; lane markings buried	Visibility drops to meters; features become indistinct beyond a short range	Dust clouds obscure the scene	Depends on headlights and streetlights; no independent light source
Lidar	Rain droplets cause false returns; heavy rain attenuates beam; ~10–20% range loss in heavy rain (est.)	Snowflakes reflect beam and produce false returns; accumulated snow on housing obstructs; wet snow is worst	Fog water droplets scatter lidar pulses severely; dense fog can cut effective range by 50–80% (est.)	Dust absorbs and scatters near-IR wavelength; moderate impact	Unaffected — uses its own light source
Radar	Nearly unaffected — radar penetrates rain well; slight attenuation only in extreme downpour	Nearly unaffected — radar penetrates snow	Nearly unaffected — radar penetrates fog	Nearly unaffected	Unaffected
Thermal IR	Unaffected — detects heat, not reflected light	Unaffected	Slightly affected in dense wet fog (water absorbs IR)	Unaffected	Strong advantage — heat signatures remain clear
GPS / HD Map	Unaffected (signal); map accuracy degrades if roads are buried under snow	Unaffected (signal)	Unaffected	Unaffected	Unaffected

The central insight from this table: Radar is the all-weather sensor. It penetrates rain, snow, and fog with minimal degradation. This is why radar remains in every multi-sensor AV stack even as lidar dominates close-range 3D perception. Lidar has the worst fog performance of the active sensors. Camera has the widest range of weather-induced failure modes. Thermal IR is underutilized but has real night and fog advantages.

The ranking of weather vulnerability, from most to least: camera first, then lidar, then radar — which is nearly invulnerable to precipitation and fog. Any sensor fusion architecture is implicitly an attempt to cover the failure modes of the weakest sensors with the strengths of the more robust ones.

Section 2 — Tesla’s Camera-Only Challenge

Tesla’s Full Self-Driving is a camera-only architecture. The company deliberately removed radar from the Model 3 and Model Y hardware suite in 2021, arguing that a neural network trained on sufficient data could match or exceed the redundancy provided by radar. This bet has significant implications for adverse weather.

Weather condition	Camera-only impact	Tesla FSD mitigation
Light rain	Manageable; wipers clear windshield; a standard real-world condition	Neural net trained on rain data; wiper activation integrated; no fundamental capability gap
Heavy rain	Significant: vision range drops, lane markings obscured, pedestrian detection degrades	Fleet data includes heavy rain; v12 and v13 improved; still harder than clear conditions (est.)
Snow (light)	Lane markings buried; road boundaries become unclear	Road boundary recognition via neural net; relies on curb and vegetation cues as proxies
Snow (heavy / blizzard)	Major challenge: whiteout, buried markings, no reference points	Tesla recommends FSD off in heavy snow; supervised mode expected; driver takeover required
Dense fog	Camera range drops to meters; the system is effectively blind beyond that	Neural net cannot create information that the optics cannot see; fundamental physics limit
Ice	Not a sensor issue directly — but road surface estimation via camera is imperfect	No direct camera-based ice detection; relies on temperature sensors and traction events
Sun glare	Camera directly blinded by low sun on the horizon	Detected via image saturation; FSD may slow or request takeover; fundamental physics constraint

Geographic implication: Tesla FSD handles normal weather in most US markets well enough for supervised driving. Driverless commercial service at the safety standard required for public robotaxi operation is a different bar. Austin, Texas — with mild winters, limited snow, and infrequent dense fog — is a favorable launch environment. The Northeast, Midwest, and Pacific Northwest are harder. Camera-only physics does not change with more data when fog removes the visual signal entirely.

The Tesla thesis is that neural networks, trained at the scale of its fleet, will develop robust enough scene understanding to compensate for reduced visual signal in adverse conditions. This may hold for a wider range of weather than critics allow. The physics limit — a camera cannot see through opaque fog or measure velocity without radar — is real but applies at the extremes. The contested question is whether those extremes occur often enough, in the markets Tesla needs to serve, to block commercial driverless operation.

Section 3 — Waymo’s Multi-Sensor All-Weather Strategy

Waymo’s sensor stack combines lidar, radar, camera, and (in some configurations) thermal imaging. This architecture was deliberately designed to provide redundancy at the sensor-physics level — meaning that when one modality degrades, others maintain operational confidence.

Weather condition	Waymo sensor fusion response
Rain	Radar primary for velocity and detection through rain; lidar handles light rain well; camera provides object classification when visible; three-sensor redundancy maintains confidence
Heavy rain	Radar holds detection regardless; lidar degrades but remains functional at close range; camera degrades substantially; fusion confidence drops — service may be suspended in severe storms (est.)
Fog	Radar penetrates fog entirely; lidar degrades severely but radar maintains object detection; camera augmented by radar; multi-sensor fusion retains situational awareness that camera-only cannot
Phoenix (sunny, hot, dry)	Optimal conditions for all sensors; dust occasional but manageable; extreme heat (45°C+) is a hardware thermal management challenge, not a sensor degradation problem
San Francisco (coastal fog, mild rain)	Radar handles the marine fog layer; lidar performs well in light conditions; camera augmented by radar in heavy fog — a direct multi-sensor advantage over camera-only architecture
Snow and ice markets	Waymo has not launched in snow-primary markets; snow accumulation on sensor housings is a real operational problem; heated housings mitigate this but add complexity (est.)
Gen 6 vehicle	Purpose-built with integrated sensor housing and improved weather resilience; heated sensor covers and improved sealing reported (est.)

The operational consequence of this architecture: Waymo can maintain confident operation in conditions — fog, light rain — where a camera-only system loses significant situational awareness. It is not invulnerable to weather. Dense snow accumulation on sensor housings is a known challenge that requires heated covers and mechanical cleaning systems. Extreme precipitation can still push fusion confidence below the threshold for autonomous operation. But the degradation curve is more gradual and the failure modes are different — they tend to be operational (can we clean the sensors?) rather than fundamental (can we see anything at all?).

Section 4 — Geographic Deployment Patterns Explained by Weather

The map of where Waymo has launched driverless commercial robotaxi service is not arbitrary. It corresponds almost exactly to markets with favorable weather profiles for its current sensor stack.

Waymo market	Climate profile	Why it works
Phoenix, AZ	Desert: 300+ sunny days per year, minimal rain, no snow, dry	Optimal for all sensors; lidar performs at maximum range; camera clear; driverless since 2018 — the earliest and most mature market
San Francisco, CA	Mediterranean: coastal fog but mild; light rain; no snow	Multi-sensor handles fog layer; radar penetrates the marine layer; driverless launched 2022
Los Angeles, CA	Mediterranean: very mild, rarely rains heavily, no snow	Excellent AV weather; expanding 2023 onward
Austin, TX	Subtropical: mild winters, occasional heavy rain, rare ice events	Good AV weather; driverless launched 2025
Atlanta, GA (announced)	Subtropical: mild winters, summer thunderstorms	Manageable weather profile; next announced expansion
NOT deployed	Chicago, IL; Boston, MA; Minneapolis, MN — all have significant snow and ice seasons	Snow accumulation on sensors, buried lane markings, ice — operational challenges not yet solved at commercial scale

The pattern is clear. Every current Waymo driverless market sits in the Sun Belt or in mild coastal climates. Every major northern US city with real winters is absent. This is not because Waymo has not tried to enter those markets. It is because the operational complexity of adverse winter weather — sensor accumulation, buried road markings, reduced fusion confidence — creates a reliability threshold that current AV systems, including Waymo’s, have not crossed at commercial scale.

The same logic applies to global markets. Why is Waymo not in London, Paris, or Toronto? Regulatory complexity is part of the answer (see Article 61 on the regulatory map). But weather is the other half. The combination of regulatory friction and weather challenge means that northern European and Canadian markets are unlikely to see commercial driverless AV operation before the technology has matured considerably beyond its current state.

Section 5 — The Weather Frontier: Who Gets There First

The all-weather AV is the trillion-dollar unlock. Most of the US population lives in climates with real winters. A driverless service that cannot operate in Chicago, Boston, or Minneapolis is excluded from a very large fraction of the addressable market. The companies and technologies racing toward all-weather competency define the next competitive wave in physical AI.

Approach	Companies	Status
Multi-sensor fusion with radar dominance	Waymo, Zoox, Cruise (suspended)	Best current all-weather performance; still avoids heavy-snow markets commercially
Heated sensor housings and mechanical cleaning	Multiple AV companies	Engineering solution for sensor accumulation; adds power consumption, mechanical complexity, and maintenance requirements
4D imaging radar	Arbe Robotics, Continental, Mobileye	Next-generation radar with much higher angular resolution; approaches lidar-level detail with radar’s weather immunity — potentially the most important sensor development for adverse weather
Solid-state lidar	Luminar, Ouster, Innoviz	More ruggedized than spinning units, lower cost, better suited to harsh environments; weather performance broadly similar to spinning lidar
Camera plus 4D radar (no lidar)	Mobileye SuperVision, some competitors	Skips lidar’s fog weakness entirely; bets on radar resolution improving to the point where it substitutes for lidar’s 3D detail
Camera-only all-weather (contested)	Tesla FSD	Requires neural nets to overcome the fundamental physics limits of cameras in fog and heavy snow; whether achievable at a commercial driverless safety standard is the open question in AV

The investor signal: The first AV operator to commercially launch driverless service in a northern US city with genuine winters — Chicago, Boston, Minneapolis — will have demonstrated a capability gap that expands the total addressable market by roughly 3x. The milestones to watch: heated sensor housing announcements, winter-city pilot permits from regulators, and 4D imaging radar partnership deals. The company that crosses that weather frontier first does not just win a new city. It unlocks the majority of the addressable market that is currently off-limits to every commercial AV operator.

The adverse weather challenge is not a temporary engineering obstacle that more lidar beams or more training data will eventually dissolve. It is a physics constraint on the sensors available today, a manufacturing challenge for the sensor housings, and a regulatory and operational challenge for deployment in climates where road conditions can change hourly. Solving it will require advances in sensor technology — particularly radar resolution — as much as it requires advances in neural network performance. The companies that understand this distinction will make better bets than those who treat it as purely a software problem.

Sources: IEEE Transactions on Intelligent Transportation Systems — ieeexplore.ieee.org; Waymo Safety Report — waymo.com/safety; Tesla FSD weather guidance — tesla.com/support; Automotive World radar analysis — automotiveworld.com. All figures marked (est.) are estimates derived from public technical publications, company materials, and industry research. They have not been independently verified and should be treated as directional. This article does not constitute investment advice.