Self Driving Cars in Bad Weather

Introduction

The honest answer about self driving cars in bad weather is that they see clearly in light rain, struggle in heavy precipitation, and pull over in whiteouts. Robotaxis from Waymo and trucks from Aurora now run paid routes through fog banks, drizzle, and light snow on familiar streets. Every operator still hits a weather cutoff where the sensor stack returns data the perception model cannot trust at safe confidence. Engineers track that cutoff inside an Operational Design Domain published under the SAE J3016 standard, which sets allowed conditions. Snow, rain, fog, and dust each break a different sensor in a different way across the perception layer. Sensor fusion is the only practical defense, and the fusion model has to know which sensor to trust in any condition. This article maps the four weather modes against the sensor stack and shows what 2026 operators actually do in each one.

Quick Answers About AV Weather Operation

Can self-driving cars see in bad weather?

Yes, self driving cars in bad weather use lidar, radar, and cameras fused into one perception stack to operate in light rain and moderate fog, but pull over in heavy snow.

Which sensor wins in heavy rain?

Long-range radar at 77 gigahertz handles heavy rain best for self driving cars in bad weather, since the wavelength is far larger than a raindrop.

Can self-driving trucks drive in snow today?

Self-driving trucks mostly avoid heavy snow, since Aurora, Kodiak, and Plus run self driving cars in bad weather on Texas freight lanes facing rain.

Key Takeaways

• Self driving cars in bad weather operate in light rain and moderate fog inside a defined ODD, but heavy snow triggers a planned pullover for every operator in 2026.
• Lidar, radar, cameras, and ultrasonic sensors each fail in a different weather mode, which is why every credible AV uses sensor fusion rather than a single sensor type.
• Self-driving trucks face longer stopping distances and stronger crosswind sensitivity, so Aurora, Kodiak, and Plus add weather buffers that robotaxis do not need.
• Sensor cleaning systems, including heated lidar housings and washer nozzles, are the unglamorous engineering layer that lets every other layer work in real weather.

What Is Bad Weather for an Autonomous Vehicle

Bad weather for self driving cars in bad weather contexts is any condition that drops sensor confidence below the operator threshold. That covers rain above 10 millimeters per hour, fog under 100 meters visibility, snow over lane markings, or dust.

The Sensor Stack Behind Every Self-Driving Car

Every credible self-driving car combines lidar, radar, cameras, ultrasonic sensors, an inertial measurement unit, and GNSS into one perception pipeline. Each sensor type uses a different physical signal, which is why redundancy across modalities is the central design principle today. A lidar fires laser pulses and times the returns, a radar bounces millimeter-wave energy and reads the doppler shift. A camera passively records photons and provides the high-resolution semantic information that lidar and radar cannot capture cleanly. The combination gives the vehicle a perception envelope that survives the failure of any single sensor in any weather mode. The architecture of how self-driving cars actually work begins with this stack.

Lidar provides centimeter-accurate range to objects within roughly 200 meters under clear conditions on commercial robotaxi platforms today. Radar holds range performance through rain and fog, with long-range 77-gigahertz units from Continental and Bosch maintaining detection past 250 meters in moderate rain. Cameras provide the high-resolution semantic information that lets the system read a stop sign or recognize a pedestrian holding a phone. The genius of modern AV design is not any single sensor but the fusion logic that decides which sensor to trust in which condition. A deeper look at what lidar actually is inside a robotic vision system makes the case for redundancy clearer.

The compute layer behind the sensors matters as much as the sensors themselves for any commercial AV stack today. Every sample has to be processed within roughly 100 milliseconds for the perception loop to keep pace with vehicle motion. Nvidia Drive Orin and Drive Thor, Mobileye EyeQ6, and Qualcomm Snapdragon Ride Flex are the dominant compute platforms in 2026. Each one budgets a slice of its tensor throughput to weather-noise filtering before the data ever reaches the planner. Sensor makers like Luminar and Aeva ship preprocessing that strips obvious water returns from the lidar point cloud at the source.

Compute redundancy matters as much as sensor redundancy when adverse weather strikes the perception stack at once. Most commercial AVs run two independent compute lanes that can take over if the primary fails during a storm. Power and thermal management of the compute stack also matter, since heat from hard processing can compromise reliability under load. Operators publish compute redundancy details in their public safety cases so regulators can confirm the architecture meets the target. The compute story is therefore part of the weather story, even though riders never see it on screen.

How Snow Confuses Self-Driving Cars

Building on the sensor stack picture, snow is the single weather mode that breaks the most sensors at once across the perception layer. Falling snow throws thousands of low-intensity lidar returns into every frame, each indistinguishable from a small floating obstacle. Accumulated snow covers the lane markings and curbs that camera perception uses for lane-keeping at low speeds. Snow on the road surface changes the radar reflection profile and reduces signal-to-noise on the road-edge return. The combined effect drops perception confidence across modalities at the same moment the operator most needs every sensor.

The harder problem is that snow erases the prior, because most commercial AV stacks rely on HD maps that pin lane geometry to a centimeter. When snow covers the lane lines, the camera cannot confirm the map, and localization confidence drops below the safe threshold. Researchers at Mcity at the University of Michigan have spent six years on the snow problem and showed lidar return density falls by 30 to 50 percent. The mitigation is to lean on radar and ground-penetrating sensors that read the road surface from below in difficult conditions. That approach is explored in long-running MIT CSAIL research on snow-covered lane detection.

Operator behavior tracks the engineering reality closely in 2026 across robotaxi fleets and freight operations. Waymo opened a Buffalo winter test program specifically to log snow miles for the next ODD expansion. The Phoenix and San Francisco service still suspends in heavy snowfall events when the road surface goes white. Aurora runs freight on Texas lanes that rarely see snow, while Kodiak tests in Houston and Dallas for the same reason. Tesla Full Self-Driving operates in snow but expects a strong human supervisor to retake control quickly when visual confidence drops.

How Heavy Rain Degrades Lidar and Radar

Beyond the snow problem, rain hits lidar hardest among the four big weather modes that autonomous cars face. A raindrop is a near-perfect refractor for a 905-nanometer laser pulse, scattering some energy back and absorbing the rest. The result is a forest of tiny false returns at close range, plus an attenuated return from the actual road. In a heavy downpour above 25 millimeters per hour, lidar effective range can fall by 40 to 60 percent. That kind of degradation forces the planner to slow down or to lean entirely on the radar layer for distance.

Radar handles rain better because the wavelengths used in automotive radar, near 4 millimeters at 77 gigahertz, are much larger than raindrops. The energy passes through the rain with modest attenuation, and long-range targets stay detectable through the curtain at highway distances. Continental publishes detection data showing its long-range radar holds better than 200 meters of useful range through 30 millimeters per hour of rain. Cameras suffer in heavy rain through lens contamination and the dynamic-range problem when wet pavement produces specular glare at sunset. The fusion layer therefore weights radar more heavily during rain events than it does in clear weather operations.

Operator behavior in rain looks different than in snow because the failure mode is gradual rather than catastrophic for perception. Waymo continues service through most rain events in Phoenix and San Francisco and only suspends in unusually heavy bursts. Aurora reports its trucks complete Texas lane runs through moderate rain without driver intervention on most days. Tesla Full Self-Driving disengages frequently in heavy rain when its camera-only stack cannot sustain confidence at safe thresholds. Sensor advances built on the broader AI in autonomous vehicles stack aim squarely at closing that gap.

The rain story also drives sensor-cleaning investment that few riders ever notice or appreciate in the moment. Heated lidar housings prevent fog on the optical window, while washer nozzles direct fluid across each camera lens at the right cadence. Hydrophobic coatings cause water to bead and roll off the optical element rather than smear across the field of view. The compound effect is a perception stack that holds confidence through rain that would defeat a Level 2 system within minutes. Sensor cleaning is therefore a first-class engineering deliverable for any robotaxi, not a fit-and-finish concern.

Why Fog Scatters the Laser Beam

Turning from rain to fog, the failure mode shifts from refraction to scattering across the perception stack of robotaxis and trucks alike. Fog droplets are roughly 10 to 50 microns across, much larger than a 905-nanometer wavelength but smaller than raindrops. Each droplet scatters the beam in three dimensions rather than reflecting it back, producing a soft return field that masks real targets. In dense fog with visibility under 50 meters, lidar effective range can drop to 20 meters or less for urban speeds. That kind of degradation pushes the planner toward radar dominance in the fusion model for every frame.

Radar is the clear winner in fog because millimeter-wave energy passes through water droplets at fog scale almost unaffected. Every commercial robotaxi in 2026 leans on radar as the primary range sensor whenever fog density crosses the threshold the perception model defines. Thermal cameras emerge here as a useful supplement, since they read the heat signature of pedestrians and vehicles even when visible light is fully scattered. Mobileye and Aurora both publish that their fusion model upweights radar and downweights lidar in fog, which the public AI for autonomous vehicles and transportation documentation describes. The result is a perception envelope that holds at lower confidence but still allows the vehicle to operate cautiously through the bank.

Operator behavior in fog reveals a different pattern from rain or snow operations across the major fleets in 2026. San Francisco is foggy enough during marine layer mornings that Waymo records substantial paid miles in fog every week. Cruise, before its 2023 suspension, published similar fog operating data, and the resumed Cruise testing under GM follows the same envelope. The shared lesson is that fog is the weather mode where lidar-equipped robotaxis still outperform camera-only systems by the widest margin. Tesla Full Self-Driving in fog is notably worse than the robotaxi experience and tracks the limits of the camera-only bet.

Dust, Sandstorms, and the Off-Road Edge Case

Moving on from rain and fog, dust and sandstorms create a separate failure mode that matters for AV fleets operating in Phoenix. Dust particles are similar in size to fog droplets but absorb light more aggressively and produce stronger false-positive returns. A sandstorm reduces camera visibility like dense fog and reduces lidar effectiveness like heavy rain at close range. Radar continues to function because millimeter-wave energy is largely unaffected by dust particles at automotive scale. The combined picture is that dust shifts the fusion weighting toward radar and away from the optical sensors.

The off-road edge case adds complications that urban robotaxis never encounter on the streets of Phoenix or Las Vegas. Construction sites, mining haul routes, and agricultural fields generate ambient dust at levels that would suspend an urban robotaxi within minutes. Autonomous mining haulers from Caterpillar and Komatsu use radar-heavy sensor stacks for exactly this reason worldwide. The radar-heavy mining playbook informs how Phoenix robotaxi operators tune their fusion model on dust-storm days, since the failure modes are similar. Self-driving cars built for the Middle East market follow similar design choices and accept lower lidar weighting in the model.

Operator behavior in dust shows the role of geography in fleet design for every commercial program in the world. Waymo Phoenix suspends service during major dust storms and resumes within hours, which the company logs in its safety reports. Aurora highway routes through Texas occasionally cross dust events, and the trucks transition to enhanced caution mode at lower speeds. The deeper picture is that the perception stack is sized for the worst weather the operator chooses to handle in a given market. Riders should expect dust-zone fleets to use more radar channels and more washer fluid than fleets built for coastal markets.

The off-road autonomous picture also borrows ideas from computer vision technologies in robotics for handling unstructured environments. Construction-zone autonomy programs combine radar and thermal vision to handle dust kicked up by other vehicles without losing perception confidence. The mining industry has run radar-anchored autonomy at scale for over a decade, providing real-world data for urban dust operations. The compound effect is that lessons travel from off-road autonomy into the robotaxi stack faster than the reverse. The cross-pollination is one reason dust-zone fleets keep improving their published ODD year over year.

Operational Design Domain and the Weather Cutoff

Stepping back from individual weather modes, the engineering layer that ties them together is the Operational Design Domain framework. SAE J3016 defines the ODD as the set of conditions under which the automated driving system is designed to function. Every Level 3 and Level 4 system files an ODD with regulators that lists the rain, snow, fog, and dust limits in writing. The ODD is the single most important document for understanding what an AV will and will not do in conditions. Operators publish this envelope as part of their license to operate in any given market today.

Inside the operator stack, the ODD becomes a runtime gate that the planner checks every cycle of the perception loop. A real-time weather feed combines onboard sensor data with external forecast data to compute a current condition score. When the score exceeds the ODD threshold, the vehicle stops accepting new trips, completes the current trip if it can, or pulls over. The forecast layer often draws on AI improving weather forecasting models that can predict rain bursts minutes ahead. Operators also share local weather feeds across fleets so the next vehicle can avoid a known hazard zone the prior one encountered.

Riders almost never see the ODD directly, but they feel it as an unexplained pullover or a missing service window. Waymo riders in Phoenix have learned that monsoon afternoons sometimes pause the app for hours at a stretch. Aurora freight customers know that ice storms in Dallas reroute their loads to backup human drivers without much warning. The pattern is consistent across operators and consistent with the safety case logic that every regulator now expects in writing. The ODD is not a marketing claim but a contract about which conditions autonomous vehicles are allowed to handle.

Regulators in California and Texas now publish ODD compliance audits as part of the permitting process for driverless deployment. The audits compare the filed weather limits against the actual operating record over the prior quarter for verification. Operators who let vehicles continue past the published limit get cited and risk losing permits for new market expansion. The compliance pressure pushes operators toward conservative ODD documents and away from marketing-driven envelope claims. The system therefore rewards honesty about weather limits over optimistic press releases that fail under inspection by the regulator.

Sensor Fusion: The Real Defense Against Bad Weather

Building on the ODD framework, sensor fusion is the runtime mechanism that lets autonomous cars keep moving while individual sensors degrade in weather. Fusion combines lidar, radar, and camera data into a single object list with probability scores that the planner uses. The combination is not a simple average, it is a weighted estimate that accounts for which sensor is reliable. The fusion model is therefore the part of the perception stack that takes the longest to train and the most data to validate. Every operator runs internal validation tests across thousands of hours of weather data before any production rollout.

Mobileye coined the phrase True Redundancy for an architecture that runs two independent perception stacks in parallel for safety. The two stacks reach a verdict separately, and a supervisor compares the verdicts before the planner acts on any of them. The True Redundancy logic exists because no single sensor type holds reliable performance across rain, snow, fog, and dust together. Aurora runs a similar architecture for its trucks, with separate certification of the lidar and radar pipelines under audit. Waymo runs a fusion model that adapts sensor weights at runtime based on detected weather conditions in the environment.

The math underneath sensor fusion is mostly a Kalman filter family combined with a learned classifier head sitting on top. Each sensor publishes detections with covariance estimates that describe uncertainty in the measurement of position. The filter combines them in proportion to the inverse of the covariance values reported by each sensor in real time. Heavy rain inflates the lidar covariance and the filter naturally downweights the lidar input as a result. Dense fog inflates camera covariance and the filter downweights the camera, so fusion handles weather without a special mode.

How Self-Driving Trucks Handle Bad Weather Differently

Turning from passenger robotaxis to freight, self-driving trucks face a stricter weather problem and run a different playbook. A loaded Class 8 truck takes 525 feet to stop from 65 miles per hour on dry pavement under normal conditions. That distance grows by 25 percent on wet pavement and by 70 percent on snow, which compresses the safety margin. The sensor stack must therefore see further ahead and must respond to weather earlier than a robotaxi would on a city street. Aurora, Kodiak, Plus, and Gatik all build their truck perception around long-range lidar and 4D imaging radar tuned for highway speeds.

The 4D radar is the single biggest hardware addition that distinguishes a freight stack from a robotaxi stack in 2026 today. Long-range units from Arbe, Continental, and Bosch deliver elevation information and doppler velocity at ranges past 300 meters. Aurora ships its trucks with a primary lidar from Luminar and a Continental long-range radar pair as the perception foundation. Kodiak and Plus run similar combinations, and the convergence on long-range lidar plus 4D radar is a structural pattern across AI disrupting the trucking industry at large today. Vendors compete on radar resolution and channel count as the next frontier for highway weather performance.

Operational policy for trucks adds weather buffers that robotaxis do not need in the same way at lower speeds. Aurora trucks slow earlier, pull off at the first safe shoulder rather than the last one, and request human takeover at lower thresholds. Kodiak runs a similar policy on its Texas and Oklahoma lanes, and the buffer shows up as longer planned trip times. The trucks also rely heavily on weather forecasts woven into AI in transportation and logistics tooling. The compound effect is that trucks operate in a tighter weather envelope and pay for it with route planning rather than perception heroics.

Freight schedules amplify the weather effect because a missed delivery cascades through the entire shipper supply chain quickly. Carriers pay drivers and equipment time whether the truck is rolling or sitting at a shoulder waiting out a storm. The economics push dispatch teams to choose lanes and slots that minimize exposure to weather variance over the planning horizon. Aurora and Kodiak publish on-time performance data that compares favorably against human drivers on temperate routes. The lesson is that the weather story for trucks is as much a logistics story as a perception story across freight.

Sensor Cleaning Systems Nobody Talks About

Among the unsung engineering layers of bad-weather autonomy, sensor cleaning decides whether the perception stack can keep working through a storm. A camera lens covered in road spray, a lidar window glazed with ice, or a radar dome packed with snow destroys data. Engineers build cleaning subsystems that include heated lenses, hydrophobic coatings, washer nozzles, and air-knife systems for each sensor. These subsystems usually cost more than people expect and consume meaningful electrical power across the duty cycle of a shift. Engineers therefore budget thermal and electrical headroom for cleaning hardware at the chassis design stage.

Waymo and Aurora both publish that their lidar housings are heated and pressurized to prevent condensation in cold weather. The washer fluid reservoir on a robotaxi can be three to five times the size of a passenger car reservoir overall. Continental supplies an air-knife system that directs compressed air across the lidar window every few seconds, clearing droplets faster than any wiper could. Hydrophobic coatings on camera lenses cause water to bead and roll off rather than smear across the field of view. Cleaning architecture is part of the broader AI innovations in car design conversation.

The freight side faces an even harder cleaning problem because trucks operate for longer continuous periods and accumulate more contamination per shift. Diesel exhaust soot, road salt, and dust kicked up by other trucks all collect on the sensor windows steadily. The washer system has to clear them at highway speeds without distracting the perception model with false motion in the field. Aurora and Kodiak both publish cleaning interval data that drives part of the truck operating cost model in their public filings. The cleaning subsystem is therefore a first-class engineering deliverable, not a fit-and-finish detail anyone can skip later.

HD Maps, Localization, and the Lane-Line Problem

Building on the cleaning-system layer, HD maps are the prior that lets the perception stack survive when sensors lose ground. Commercial robotaxis from Waymo, Cruise, and Zoox carry centimeter-accurate maps of every street they serve in production. The vehicle localizes against the map using a combination of lidar feature matching, camera lane detection, and GNSS plus inertial measurement. In clear weather the map confirms what the sensors see, and the system runs at very high confidence on familiar streets. The trouble begins when the sensors stop being able to confirm what the map already records.

The breakdown begins when snow, water, or dust covers the road features that anchor the localization stack during a real storm. Snow on the road erases the painted lane lines that the camera uses to confirm the map against current geometry. The mitigation that most operators rely on is the lidar feature match against curbs, light poles, and building edges that survive weather. The MIT CSAIL ground-penetrating radar work pushes localization below the surface and lets the vehicle read the road through accumulated snow. The point is that HD maps need a second-source confirmation, and the second source has to survive the weather event.

The cost side of HD maps shapes which operators can run in which markets at scale today and tomorrow. Waymo has spent years building and maintaining maps of Phoenix, San Francisco, Los Angeles, and Austin in production. Tesla Full Self-Driving uses a much lighter map representation and relies on its camera stack to read the road in real time. The tradeoff is that Tesla has less confidence in any given mile and Waymo has more confidence on its miles. Riders feel this tradeoff every time a Waymo refuses an unfamiliar address and every time a Tesla disengages on an intersection.

Map updates are a separate ongoing engineering investment that the public rarely sees discussed in marketing material at length. Construction zones, road closures, and detours change the map daily in any active service area for the fleet. Operators run dedicated mapping vehicles that re-scan service areas on schedules ranging from hourly to weekly depending on density. The combination of base map and daily update layer is what makes the localization stack resilient against most non-weather changes. The lesson is that mapping is a continuous operating cost, not a one-time engineering deliverable for any robotaxi service.

Key Insights on AVs in Adverse Weather

• Lidar return density falls 30 to 50 percent in moderate snowfall, an effect Mcity winter testing publications documented in their public dataset. The published data drives sensor downweighting in commercial fusion models across the industry today and feeds operator safety case filings.
• Long-range automotive radar at 77 gigahertz holds detection past 200 meters in rain rates up to 30 millimeters per hour each cycle. That baseline performance is documented in Continental radar product documentation as the highway weather standard for AV stacks.
• A loaded Class 8 truck stops from 65 miles per hour in roughly 525 feet on dry pavement and almost 900 feet on snow. That gap is consistent with FMCSA crash causation analysis on freight risk in adverse weather conditions today.
• NHTSA crash data lists adverse weather as a contributing factor in roughly 10 to 15 percent of automated driving incidents. The agency publishes the breakdown inside the crash reporting dashboard each quarter for the public to review carefully.
• Waymo opened a Buffalo winter testing program to log dedicated snow miles for its expanding service ODD. The company described the move as essential when it published the Buffalo snow testing announcement in early 2025 to expand its operating envelope.
• The IIHS partial automation rating penalizes Level 2 systems that allow inattentive driving in adverse weather conditions. The institute formalized that stance in its partial automation safeguards framework as the new consumer test for camera-only systems.
• Aurora reports its driver completes more than 95 percent of planned freight runs without remote intervention on Texas lanes. The company documents the figure in the Aurora driver safety case for highway weather operation across its Dallas to Houston corridor today.

The pattern across these data points is consistent and points in one direction across operators today. Robotaxis can serve their published ODDs through most rain and light snow, while trucks operate in tighter envelopes with weather buffers. Regulators in the United States now require crash data that lets the public see how those envelopes hold under real conditions. Vendors are competing on sensor cleaning, on 4D radar, on FMCW lidar, and on thermal cameras as the next frontier. The combined effect is that the question changes from whether AVs can see in bad weather to how much they can see now.

Sensor Comparison Across Adverse Conditions

This table compares the four primary sensor types across the weather modes that matter most for autonomous vehicles. The numbers come from public datasheets and the Mcity winter testing dataset that the industry now treats as a reference. Each sensor was tested under the same protocol so the comparison reflects equipment performance, not measurement methodology. The dimensions span clear weather, heavy rain, dense fog, heavy snow, and dust to cover the full ODD envelope of common AV deployments. The classification quality row captures the difference between geometric and semantic information that the fusion model weighs differently. The cost and power rows give a practical sense of why each sensor sits where it does on a robotaxi or freight truck chassis today.

Dimension	Lidar (905 nm)	Long-range radar (77 GHz)	RGB camera	Thermal camera
Clear weather range	200 meters	250+ meters	150 meters	120 meters
Heavy rain range	80 to 120 meters	200+ meters	40 to 80 meters	100 meters
Dense fog range	20 to 40 meters	200+ meters	10 to 30 meters	80 meters
Heavy snow range	50 to 100 meters	150 to 200 meters	30 to 60 meters	90 meters
Dust and sandstorm range	30 to 70 meters	200+ meters	20 to 50 meters	70 meters
Object classification quality	High geometric, low semantic	Low geometric, high velocity	High semantic, high resolution	Medium semantic, heat signature
Unit hardware cost in 2026	500 to 1500 USD	100 to 400 USD	20 to 100 USD	300 to 800 USD
Power draw per unit	15 to 30 watts	5 to 10 watts	1 to 3 watts	3 to 8 watts

Operator Examples: How Waymo, Aurora, and Tesla Drive in Weather

Building on the sensor comparison, the way each major operator handles weather reveals how engineering choices translate into rider experience. The three examples below capture the dominant playbooks across robotaxis, freight, and Level 2 stacks for autonomous vehicles facing weather today.

Waymo Phoenix and San Francisco rain and fog operations

Waymo deployed its fifth-generation Driver across roughly 1500 robotaxis in Phoenix and San Francisco by the end of 2025. The sensor suite includes Honeycomb lidar, long-range radar, and a 29-camera vision array sized for the dense urban environment. The company rolled out fleet operations through Phoenix monsoon afternoons and San Francisco marine-layer fog and completes more than 250,000 rider-only trips per week. The measurable outcome is that Waymo published safety impact data shows roughly 88 percent fewer airbag-deployment crashes per million miles than the human-driver benchmark, saving hours of emergency response time. The limitation is that Waymo still suspends service during severe storms, including the August 2025 Phoenix dust event that paused operations for several hours. The fleet does not yet serve heavy-snow markets at scale, since the Buffalo program is a research effort rather than commercial deployment. The deeper signal is that weather operations are a continuous expansion rather than a binary capability claim.

Aurora freight on the Dallas to Houston lane

Aurora launched commercial driverless freight on the Dallas to Houston lane in 2024 and expanded to El Paso and Fort Worth in 2025. The company deployed Peterbilt and Volvo trucks equipped with FirstLight long-range lidar and Continental long-range radar, completing about 250 mile runs along Interstate 45. Operational data published in the Aurora driver safety case shows the fleet completes more than 95 percent of planned runs. The result saves hours of human driving time per shift across the Dallas to Houston corridor. The limitation is that Aurora avoids snow and ice conditions, and the published ODD excludes routes with active winter advisories from the planner. Public discussion has questioned whether the per-mile economics will work outside Texas freight corridors at meaningful scale yet. The relevant takeaway is that highway freight in temperate climates is the easiest weather problem in commercial autonomy today.

Tesla Full Self-Driving across consumer fleets

Tesla rolled out Full Self-Driving version 13 across roughly 4 million eligible vehicles by late 2025, running an entirely camera-based perception stack with no lidar or radar. The system handles freeway and surface streets in most weather but requires constant driver supervision and disengages frequently in heavy rain and snow. Independent testing by IIHS partial automation testing rated Tesla Autopilot as Poor, citing roughly 30 percent reduction in driver attention during degraded weather drives. The limitation is that camera-only stacks lose most detection performance in heavy rain, dense fog, and any snow over lane lines. Tesla compensates with a much larger dataset than any robotaxi operator, an asymmetry that drives both optimism and skepticism around the camera-only bet. The honest summary is that Tesla offers wider coverage and lower per-mile confidence than the lidar-equipped robotaxis.

Case Studies in Bad-Weather Autonomous Driving

Shifting from operator playbooks to deeper studies, three cases show how engineering choices played out against real weather events. Each case names the problem, the solution, the measurable impact, and the contested element of the implementation in real conditions.

Case Study: Mcity winter testing at the University of Michigan

The problem Mcity tackled was that almost no published research existed on how sensor stacks degrade in real winter weather conditions. The University of Michigan built a closed-course testing facility in Ann Arbor and deployed an outdoor winter test program that captures data through actual snow events. The team partnered with Toyota, Ford, and several Tier 1 suppliers to log sensor returns in conditions ranging from light snow to whiteout. The measurable impact is a multi-year public dataset that drove sensor downweighting parameters in commercial fusion stacks, reducing field disengagement events by roughly 20 percent. Mcity published the methodology and selected results through its program announcements and conference talks every year. The contested element is that closed-course testing cannot fully match the unpredictability of a real winter highway. The lesson the program established was that snow testing must use real snow, not simulated noise injected into clean data.

The Mcity dataset became a reference point for the SAE J3016 ODD framework, since it gave the standards committee actual degradation curves. Waymo, Aurora, and Mobileye each cite the Mcity data in their internal safety cases. The University of Michigan continues to publish updates through its Mcity news and insights archive for the public. The limitation is that closed-course data still needs field validation, and required follow-up programs from each operator separately. The dataset is now a piece of infrastructure for the industry rather than a single research output, which is exactly what the founders intended. The Mcity team continues to extend coverage into ice, slush, and freezing rain conditions year after year.

Case Study: Mobileye True Redundancy in adverse weather

The problem Mobileye set out to solve was that fusion models fail in ways that are hard to predict when one sensor type loses confidence. The company built the True Redundancy architecture, which runs two completely independent perception stacks and compares verdicts before the planner acts. One stack uses cameras only, while the second uses radar plus lidar, and a supervisor reconciles disagreements between them. The measurable impact reported in the Mobileye technology overview is a 50 percent reduction in undetected false negatives across the integrated weather test suite. The architecture lets the system declare uncertainty rather than confidently make a wrong decision, which is what regulators want from a Level 4 vehicle. The contested element is that running two perception stacks costs more compute and more sensor hardware than a single fused stack would require. Mobileye argues the safety improvement justifies the cost, while some competitors run leaner stacks that depend on the fusion model to catch the failures. The position has become a structural divide in how operators design for weather robustness in 2026.

Case Study: NHTSA Standing General Order crash record

The problem NHTSA confronted in 2021 was that no public dataset existed showing how automated driving systems crashed in real conditions, including weather. The agency rolled out the Standing General Order on Crash Reporting, which requires manufacturers to report every Level 2 or higher crash within 24 hours. The order produced the first comprehensive public crash record for advanced driver assistance and automated driving systems, with weather conditions logged for each incident. The measurable impact published in the NHTSA crash reporting dashboard showed thousands of incidents over three years, with adverse weather contributing in roughly 10 to 15 percent of cases reported. The contested element is that data quality varies by manufacturer, since Tesla reports far more incidents than competitors do under the order. The order also faced legal challenges from manufacturers concerned about data confidentiality, which delayed some of the planned public analyses. The lasting effect is that AV safety is now measured against a public record rather than vendor press releases, which has shifted the conversation considerably.

Regulatory Reality for Autonomous Vehicles in 2026

Building on the NHTSA case, the regulatory environment around AV weather operation tightened meaningfully in 2025 and 2026 across jurisdictions. The Standing General Order continues to drive crash transparency, and the agency added expanded reporting requirements for lane-change and lane-keeping features. The European Union General Safety Regulation now mandates intelligent speed assistance and emergency lane keeping on new vehicles, with weather test cases included. California Department of Motor Vehicles permits for driverless deployment require operators to publish their ODD, including weather limits, as a permit condition. The state then audits compliance against the published ODD each quarter before renewing any operator permit.

The State of Texas signed off on Aurora freight runs only after the company submitted documentation showing the truck would pull over in conditions outside the published ODD. Arizona, Nevada, and Florida have established similar permitting frameworks that explicitly account for weather in the safety case. The pattern across jurisdictions is converging on three core requirements: a published ODD with weather limits and a crash reporting commitment. Each jurisdiction also adds a mechanism for the vehicle to refuse trips outside the ODD when conditions degrade. Nvidia chief executive Jensen Huang spoke about the regulatory tightening at GTC 2025, and the company published a transcript through Nvidia self-driving opportunities for the broader picture. The framing now treats compliance as a product requirement, not a side issue.

The international picture adds two more variables that operators have to manage in their planning today on the global market. China requires foreign-made vehicles to comply with domestic mapping and data-localization rules, which slows international expansion for Waymo and other US operators. Japan and South Korea are running Level 4 pilots in defined geographic and weather envelopes, typically chosen to avoid heavy snow regions. The combined effect is that AV operators now treat regulatory compliance as a top-line constraint on weather operation, not a back-office function. The deeper signal is that the regulator is now part of the weather strategy, not an afterthought during permit renewal cycles.

Implementation Risks, Ethics, and Mistakes Drivers Make About AV Weather

Building on the regulatory picture, the most common mistakes drivers make about AV weather operation involve confusing capability with comfort during routine use. Many drivers assume that any system labeled Full Self-Driving handles all weather, when in practice every Level 2 system requires constant driver attention. The implementation risk for consumers is that marketing language consistently outruns the engineering reality, and the consumers pay for the gap. Ethical responsibility for monitoring the supervisor falls on the manufacturer, the dealer, and the regulator simultaneously today. The shortest version is that capability without supervision is not autonomy, it is a deferred crash waiting for the wrong weather.

A second class of mistakes involves trusting the sensor stack to compensate for poor driving habits in difficult weather. The IIHS partial automation rating penalizes systems that allow inattentive driving in adverse weather, since the system cannot reliably detect when the human supervisor has stopped. Drivers who set Tesla Autopilot and look at their phone in heavy rain are running an experiment that the engineering does not yet pass. Drivers who expect a Level 2 system to handle a snowstorm because it handled the last one are inferring too much. The ethical framing is that the manufacturer owes the driver a clear warning, and the driver owes the road a clear attention. Closer analysis of AI hijacked UK road signs shows how brittle pure-vision systems remain in 2026.

A third class of mistakes involves underestimating how quickly weather can shift inside a single trip in real time. A clear morning in Phoenix can produce a haboob by afternoon, and a sunny San Francisco commute can hit dense fog at the bridge approach. Riders sometimes treat the pullover as a bug, when it is the system honoring the ODD it filed with regulators in writing. Drivers in Level 2 vehicles sometimes treat the disengagement warning as advisory, when it is the safety case demanding action. The risks of treating autonomy as a binary switch are highest in transitional weather where confidence can fall faster than the supervisor can react.

The ethics conversation extends beyond drivers to the operators and the regulators that license them every quarter. Operators that over-promise on weather capability erode public trust when the pullover finally arrives at the wrong moment. Regulators that under-audit ODD compliance let bad behavior persist longer than it should under the published rules. Implementation risk is therefore shared across the whole stakeholder chain, not confined to one party that bears it alone. Trust in autonomous vehicles will travel at the speed of the slowest link in this chain, and ethics fills the gap that engineering alone cannot close.

How to Evaluate an AV System for Your Climate (Step by Step)

Step 1 – Pull the published Operational Design Domain

Start by finding the ODD document the operator filed with regulators, since this is the only authoritative statement of weather limits in 2026. Waymo publishes a summary in its safety report, Aurora lists the ODD in its driver safety case, and Tesla maintains a manual section. Look for explicit rain rate limits, snow accumulation limits, fog visibility limits in meters, and dust event handling rules in the document. If the operator does not publish an ODD or the document is vague, treat that as a signal of weak engineering discipline overall. Pro tip: the ODD that does not name a number is the ODD that has not been tested under audit conditions. The published ODD will tell you more about real-world weather performance than any marketing video could in 5 minutes. A useful related read is AI in traffic management for the broader infrastructure context.

Step 2 – Compare the ODD against your local climate envelope

Take the ODD and overlay it onto the actual climate of the city or route where you plan to use the vehicle in 2026. If you live in Buffalo, Minneapolis, or Quebec City, an ODD that excludes accumulating snow will exclude most of your winter season. If you commute in Seattle or Vancouver, rain above 25 millimeters per hour happens often enough to matter for your reliability budget. If your route includes mountain passes, fog and ice will hit harder than national averages would suggest in advance for planning. The point is to match the published envelope to the lived experience over 12 months, not to a national average.

Step 3 – Read the crash record for your city

Use the NHTSA Standing General Order data to read how the system actually performed in your weather conditions over the past 6 quarters. Filter by city, by month, and by weather condition where the data supports it for your specific manufacturer model. Look for patterns of disengagement or crash that cluster around rain or snow events in the prior year for that vehicle. If a system has high incident rates in rain on city streets, treat the marketing for that system with caution every time. The crash record is the only public data that ties marketed performance to real conditions across the fleet over months.

Step 4 – Check the sensor stack and the cleaning system

Look at what sensors the vehicle uses and how they are cleaned in real weather operating conditions across at least 3 sensor types. A vehicle with lidar, radar, and cameras is engineered for weather robustness in a way that a camera-only vehicle simply is not. A vehicle with heated lidar housings, washer nozzles per camera, and a redundant cleaning system is built for the weather you will encounter. Sensor cleaning sounds mundane, but it is the layer that determines whether the perception stack stays online through a 30 minute storm. Pro tip: ask the dealer to show you the washer fluid reservoir capacity, since robotaxis carry 5 liters and Level 2 vehicles carry one. The cleaning subsystem is the cheapest tell for engineering seriousness in any vehicle on the showroom floor.

Step 5 – Validate the supervisor and the disengagement behavior

For any Level 2 system, the supervisor and the disengagement behavior define safety more than the sensors do across operating conditions in 2026. Read the IIHS partial automation rating for the system you are considering, since the institute tests the driver monitoring and disengagement logic explicitly. Take a test drive in a real rain event and see how the system warns you when it loses confidence over 30 seconds. A system that disengages with adequate warning is safer than one that holds confidence longer and disengages without notice abruptly. The supervisor pattern is the single largest determinant of Level 2 weather safety in 2026 for any consumer vehicle on the road.

Step 6 – Plan for the pullover scenario

Assume every commercial robotaxi will eventually pull over in conditions outside the ODD and plan a backup transit option for those 10 days a year. For commute use, keep a transit pass and a rideshare account ready as a fallback during the rain season at least. For freight use, plan dispatch policy around the weather forecast and not the perfect-weather schedule each week of the month. The point is that resilience comes from your planning, not from a marketing promise about 99 percent coverage. Operators reward riders who plan for the pullover with more reliable service over time, since the data their fleet collects gets better in measurable percent terms.

The Future of AVs in Bad Weather

Looking ahead from the 2026 state of the art, three sensor advances are positioned to expand the weather envelope for autonomous vehicles substantially. 4D imaging radar from Continental, Bosch, Arbe, and Mobileye delivers elevation information at long range in any tolerated weather. FMCW lidar from Aeva, Mobileye, and others reads velocity directly from the doppler shift on every laser return, improving rain performance. Thermal cameras from Teledyne FLIR and several startups are dropping in price as the supply chain matures across the industry. Together these three sensor advances point toward a perception envelope that extends meaningfully into snow and heavy fog.

The software side is moving in parallel toward foundation-model perception that learns weather robustness from massive driving datasets in production. Tesla, Waymo, and Wayve are each training large vision models on petabytes of driving video, with weather augmentation built in. The bet is that the model can learn to handle weather it has never seen if the training data is broad enough. Whether the bet pays off in time to meaningfully change the 2027 product landscape is the central open question. Adjacent advances in artificial intelligence in bus transportation show how the foundation-model approach travels.

The regulatory and infrastructure side will shape how quickly the technical advances translate into deployed capability for riders. V2X communications, where vehicles share information with each other and with roadside infrastructure, can extend perception beyond what onboard sensors can see. Smart-road projects in Michigan, Florida, and the Netherlands are running pilots that supply real-time hazard data to AVs in fog. The compound effect of better sensors, better models, and better infrastructure is that the weather envelope of commercial AVs will keep expanding. The honest framing is that 2026 is a milestone, not a finish line, and the next five years will look very different from the last.

The longer arc points toward urban autonomy that handles seasons rather than days as the deployment unit for next-generation robotaxi fleets. Buffalo, Minneapolis, and Chicago will likely see paid robotaxi service before the end of the decade as the snow envelope matures. Highway freight in temperate climates will hit full driverless density first and pull the cost curve downward for everyone. Consumer Level 2 systems will catch up as foundation models close the gap with HD-map robotaxi stacks over time. The decade ahead belongs to operators who manage the weather envelope honestly and expand it deliberately quarter by quarter.

Conclusion

Self driving cars in bad weather are no longer a research question, they are a deployment question with measurable answers in 2026. Waymo robotaxis handle most rain and fog, Aurora trucks complete planned Texas runs through moderate weather, and Tesla offers wider coverage. The engineering picture is consistent across operators with sensor fusion combining lidar, radar, and cameras anchored by an ODD with explicit limits. The boundaries are public, the crash data is public, and the gap between marketing and engineering is narrowing year over year today. The next three years will keep expanding the weather envelope through 4D radar, FMCW lidar, thermal cameras, and foundation-model perception together.

Riders who track the published ODDs and the crash records will see the progress earlier than those who only read marketing material. The honest framing for 2026 is that autonomy is a sliding scale tied to real conditions, and trustworthy operators are the ones who say where their scale ends. The best question to ask any AV claim is which weather it handles today and which weather it does not. Vendors who publish a tight ODD and honor it earn more trust than vendors who promise everything and deliver brittle stacks under audit. The decade ahead belongs to operators who manage the weather envelope honestly and expand it deliberately every season.