2026-06-18 — views

Physical AI Extreme Weather 2026 — Waymo LIDAR Fog Performance vs Tesla FSD Camera Snow: The AV All-Weather Benchmark

Waymo handles SF fog with 1550nm LIDAR. Tesla FSD uses snow-belt training data. No AV system is validated for driverless heavy snow or ice as of mid-2026.

Article 207 in the Physical AI Benchmark Series — The All-Weather Deployment Challenge

Weather is the most underappreciated constraint on the commercialization of autonomous vehicles. Every AV sensor modality — LIDAR, camera, radar — has a fundamentally different relationship with precipitation, fog, and temperature extremes. The geographic deployment pattern of commercial AV services in 2026 is not random: it directly reflects where today’s sensor stacks perform reliably enough to operate without a human backup driver. The Sunbelt dominance of AV commercial deployments (Phoenix, Austin, Los Angeles, San Francisco, Atlanta) and the complete absence of commercial driverless service in Chicago, Boston, Detroit, and Minneapolis is a technical fact about sensor physics as much as it is a business or regulatory choice.

This article benchmarks Waymo’s LIDAR-dominant sensor fusion stack against Tesla’s camera-dominant FSD system across the full spectrum of adverse weather conditions. The analysis draws on physics principles (wavelength-dependent scattering, millimeter-wave radar penetration), publicly disclosed operational design domain (ODD) boundaries, and the training-data breadth advantage that Tesla’s 6 million-plus vehicle fleet provides in snow-belt geographies. All figures marked (est.) are estimates based on public disclosures and physics principles where official data has not been published.

Section 1 — The Weather Gap in Physical AI Deployment

The geographic deployment pattern of commercial AV services in 2026 is not accidental. Waymo’s operating cities — San Francisco, Phoenix, Los Angeles, Austin, Atlanta — share a critical characteristic: they are either desert (Phoenix: annual average rainfall est. 8 inches (est.)), mild-weather coastal (SF: regular fog but mild temperatures, rare ice), or warm urban (Austin, LA, Atlanta). The notable absence is stark: no major commercial driverless AV service operates in Chicago (snow, extreme cold), Boston (snow, freeze-thaw cycles), Detroit (snow, winter storms), Minneapolis (extreme cold, heavy snow), or New York City (complex urban environment plus occasional snow).

This geographic gap is both a technical limitation and a business prioritization — but the technical limitation came first. AV companies expanded to Sunbelt cities because their technology performs more reliably there, and because regulatory environments in Texas and Arizona have been more permissive. The snow-belt absence is not primarily a regulatory failure; it is a direct reflection of where current sensor stacks reach their performance limits.

The three primary AV sensor modalities have fundamentally different weather performance profiles:

LIDAR (light-based, laser pulses): Weather-sensitive. Laser pulses at 905nm or 1550nm wavelengths scatter when they encounter water droplets in rain and fog, and snowflakes in precipitation events. Heavy rain degrades LIDAR effective range substantially. Dense fog reduces LIDAR effective range from an estimated 200-plus meters in clear conditions to an estimated 50–70 meters in dense coastal fog (est.). LIDAR sensor housings can become physically blocked by snow accumulation on the vehicle exterior. Best performance: clear weather. Worst performance: dense fog and heavy snowfall combined.

Camera (visible light and near-IR): Weather-sensitive. Water droplets on camera lenses reduce image clarity in rain. Reduced contrast in fog makes object boundaries harder to distinguish. Snow-covered roads remove lane markings entirely — both painted lines and road boundaries disappear under white accumulation. Night plus rain plus fog is the worst-case combination for any camera-based system, as reduced ambient light compounds reduced contrast. Best performance: clear daylight. Worst performance: nighttime with heavy rain and fog combined.

Radar (millimeter-wave, 77 GHz automotive standard): Weather-robust. Millimeter-wave radar signals penetrate rain, fog, and snow with minimal attenuation. Radar effective range and velocity accuracy are largely unaffected by precipitation. The critical limitation is not weather performance but spatial resolution: radar provides range and velocity data but cannot reconstruct the shape of objects the way LIDAR point clouds or camera images can. Radar detects that something is there and how fast it is moving; it cannot determine whether that something is a pedestrian, a trash can, or a parked vehicle.

The sensor stack architecture determines overall weather robustness. Waymo’s triple-redundancy approach (LIDAR plus camera plus radar) is structurally more weather-robust than Tesla’s camera-dominant approach (camera primary, radar supplemental) because radar maintains full performance when LIDAR and camera both degrade. However, Tesla’s training data breadth — estimated tens of millions of miles per week across all US weather geographies including snow-belt states (est.) — partially compensates for camera physics limitations through learned pattern recognition in adverse conditions.

Temperature extremes add a further dimension. Phoenix reaches estimated 115°F-plus in summer (est.), stressing EV battery thermal management and sensor electronics. Minneapolis and Chicago experience estimated minus 20°F winters (est.), reducing EV battery range by 30–40% (est.) and affecting tire friction and sensor calibration. These thermal extremes are additional factors in AV geographic deployment decisions beyond precipitation alone.

Section 2 — Waymo Weather Performance: LIDAR in Fog, Rain, and Mild Adverse Conditions

Waymo’s commercial driverless service operates across San Francisco, Los Angeles, Phoenix, and Austin — four cities that collectively expose Waymo’s sensor stack to urban fog, moderate rain, and severe desert thunderstorms, but not to heavy snow or sustained ice conditions. The table below characterizes Waymo’s sensor stack behavior across the weather conditions it currently encounters commercially.

Weather condition	Waymo sensor stack behavior	Operational response	Current deployment status
Urban fog (e.g., SF Karl the Fog)	LIDAR range reduced in dense fog; Waymo’s 1550nm LIDAR wavelength scatters less in fog than shorter-wavelength 905nm systems; Waymo has developed fog-filtering algorithms that distinguish fog-return signals from object returns using temporal consistency; cameras maintain partial visibility in moderate fog	Waymo operates commercially in SF where fog is regular; fog-specific perception stack tuning; SF operational success demonstrates that urban fog is within Waymo’s commercially validated ODD	Commercial driverless service operating in SF; SF fog conditions handled within ODD; dense coastal fog events may trigger conservative driving behavior (slower speeds, increased following distance)
Light to moderate rain	LIDAR point clouds become noisier in rain; raindrop returns appear as false-positive objects; Waymo’s software filters raindrop returns using temporal consistency (raindrops do not appear in the same 3D location in consecutive scans); cameras experience lens-water degradation and reduced contrast; radar maintains full performance	Waymo operates through rain in SF and Austin; rain-specific signal processing algorithms active; conservative speed reduction and increased following distance in rain	Commercial operations continue in light to moderate rain in all operating cities; heavy rain may trigger service speed reduction or geofence contraction
Heavy rain and thunderstorms	Heavy rain significantly increases LIDAR point cloud noise; torrential rain can overwhelm Waymo’s rain-filtering algorithms; cameras are severely degraded in heavy rain; radar remains functional but with limited object classification ability	Waymo’s operational policy: reduce speed, increase following distance, pull over if conditions exceed ODD; Austin and Phoenix occasionally experience severe thunderstorms (Phoenix summer monsoon season est. June–September)	Phoenix monsoon season creates the most challenging rain conditions in Waymo’s operating cities; Waymo has published that it operates through light-to-moderate monsoon rain but maintains conservative protocols for severe monsoon conditions
Dense fog with visibility below est. 50m (est.)	Dense fog is the most challenging condition for LIDAR-based systems; at very short LIDAR effective range (est. 50m), highway speeds require stopping distances longer than the LIDAR can detect ahead; cameras are equally degraded; radar maintains object detection but cannot provide the spatial resolution needed for complex urban maneuvering	Waymo reduces speed significantly below 25 mph in dense low-visibility conditions; in extreme dense fog, Waymo vehicles may pull over and request remote operator assistance	Dense fog exceeding Waymo’s ODD triggers service interruption in affected areas; more common in SF coastal fog events than in Phoenix or Austin
Ice and snow (currently outside Waymo ODD)	Waymo does NOT currently operate in ice and snow conditions commercially; Waymo’s ODD explicitly excludes heavy snow and icy road conditions; Waymo’s sensor stack faces compounded challenges in snow: LIDAR housing accumulation, point cloud noise from falling snow, loss of road surface reflectivity markings	Waymo’s operational boundary: snow and ice are explicitly outside current ODD; this is a deliberate operational constraint defining where the system has been validated, not a safety failure in normal operating conditions	Waymo’s geographic limitation to Sunbelt cities is directly linked to snow and ice ODD exclusion; expansion to Northeastern cities requires multi-year snow and ice ODD research and validation

The key finding from Waymo’s SF deployment is that urban fog at commercially acceptable density levels is within Waymo’s 1550nm LIDAR-based system’s validated operating domain. The sensor fusion advantage — LIDAR plus camera plus radar — provides redundancy that allows the system to continue operating through fog conditions that would render a camera-only system significantly less reliable.

Section 3 — Tesla FSD Weather Performance: Camera Training Breadth vs Physics Limits

Tesla’s camera-dominant approach to adverse weather relies on a fundamentally different strategy from Waymo’s sensor fusion: rather than maintaining sensor redundancy across modalities, Tesla trains its neural network on the broadest possible distribution of real-world weather conditions. With an estimated 6 million-plus vehicle fleet generating an estimated tens of millions of miles per week (est.) across all US geographies including Minnesota, Michigan, Colorado, and New England, Tesla has accumulated more snow and ice driving training data than any competitor. The table below characterizes Tesla FSD behavior across adverse weather conditions.

Weather condition	Tesla FSD behavior	Training data advantage	Current deployment status
Fog and mist	Tesla cameras are affected by fog like any optical system; however, Tesla’s neural network trained on fog conditions from an estimated 6 billion-plus miles of diverse weather data (est.) has learned fog-specific visual patterns including reduced-contrast edge detection and fog-depth estimation from visual cues	Tesla’s training data breadth includes fog driving from Pacific coast routes (CA, OR, WA), Great Lakes fog events, and Atlantic coast conditions; this diversity of fog conditions in training makes Tesla FSD more fog-trained than any LIDAR-based system with limited fog-geography deployment	FSD deployed across all US states including high-fog areas; supervised FSD operates through fog conditions nationwide; driverless Austin service operates in mild-fog conditions
Rain (light to heavy)	Cameras get water on lenses in rain; Tesla FSD uses multiple camera views and temporal consistency to filter rain noise; heavy rain significantly reduces camera effective range; Tesla’s neural network trained on heavy rain from Florida, Texas, Gulf states, and Pacific Northwest conditions	Tesla’s fleet presence in Florida, Texas, and Pacific Northwest provides heavy-rain training data at a scale unavailable to Waymo; Tesla FSD neural network has likely been trained on more heavy-rain driving miles than any other AV system (est.)	Tesla FSD supervised operates through moderate rain nationwide; heavy rain triggers conservative behaviors including speed reduction, increased following distance, and manual takeover alerts in extreme conditions
Light snow and slush	Light snow creates snowflake-noise in camera images similar to rain; neural network distinguishes snowflakes from objects through temporal consistency (snowflakes are in motion, objects are stationary or predictably mobile); lane markings partially visible in light snow; road boundaries may be inferred from elevation change and texture even without visible painted lines	Tesla’s fleet operates in Minnesota, Michigan, Colorado, and New England; Tesla FSD has accumulated substantial light-snow training miles from these markets; light-snow driving is within Tesla FSD’s training distribution	FSD supervised is used by owners in light snow conditions; Tesla Autopilot highway mode functions in light snow with speed reduction
Heavy snow and snow-covered roads	Heavy snow is the hardest condition for camera-based systems: camera visibility can drop below an estimated 20 meters (est.); lane markings are completely hidden under snow accumulation; road boundaries disappear; cameras cannot reliably detect road edges when covered by snow; radar detects other vehicles but cannot locate road boundaries	Tesla has NOT claimed FSD capability in heavy snow; the physics limitation of cameras in whiteout conditions is not fully compensated by training data breadth; this is a fundamental sensor physics limit, not only a training gap	FSD does NOT operate reliably in heavy snow; Tesla owners in snow-belt states report FSD disengagements in heavy snow conditions; Tesla recommends human supervision in heavy snow; no driverless heavy-snow validation by Tesla
Ice and black ice	Black ice is invisible to cameras (appears as normal wet road surface) and to LIDAR (similar reflectivity to wet asphalt); only road temperature sensors, GPS-correlated weather data, and prior knowledge of ice formation conditions can anticipate black ice; radar cannot detect ice on the road surface	No AV system currently detects black ice reliably from on-board sensors alone; Tesla’s approach uses correlated alerts based on temperature plus precipitation plus location data to warn of potential ice conditions; the black ice detection gap is shared across all AV systems	No AV system — Waymo, Tesla, or any competitor — has been validated for driverless operation on black ice; this is a universal limitation of current Physical AI sensor capabilities

The most important finding from the Tesla weather analysis is the distinction between two types of limitation: training gaps (which can be closed by exposing the system to more diverse real-world data) and physics limits (which cannot be closed by more training data alone). Tesla’s camera-dominant system faces genuine physics limits in heavy snow and dense fog that no amount of training data can fully overcome — because cameras cannot detect what they cannot see. The radar reinstated in some Tesla vehicles (2023–2024 “Tesla Radar”) is an acknowledgment of this physics limit.

Section 4 — Sensor Modality Weather Performance Matrix

The table below provides a structured comparison of all three primary AV sensor modalities across the full spectrum of adverse weather conditions. Star ratings are qualitative assessments based on physics principles and publicly available performance research; they are not official benchmarks.

Sensor type	Clear weather	Light rain	Heavy rain	Light fog	Dense fog	Light snow	Heavy snow	Ice
LIDAR (spinning, 1550nm)	★★★★★ Excellent range and resolution	★★★★ Point cloud noise manageable with filtering	★★★ Significant noise; filtering required	★★★★ Good; 1550nm scatters less in fog than 905nm	★★ Reduced range est. 50–70m (est.)	★★★ Housing accumulation risk; falling-snow noise	★★ Noise plus housing accumulation; degraded	★★★ Black ice invisible; snow on road reduces reflectivity markings
Camera (visible light, 8-plus cameras)	★★★★★ Excellent visual recognition	★★★ Lens water plus contrast reduction	★★ Significant contrast loss; reduced effective range	★★★ Moderate contrast loss in light fog	★★ Significant visibility loss in dense fog	★★★ Snowflake noise manageable with temporal consistency	★ Lane markings hidden; road boundary invisible	★★ Black ice invisible visually
Radar (77 GHz automotive)	★★★★ Good object detection; limited shape resolution	★★★★★ Excellent; radar unaffected by rain	★★★★★ Excellent; radar unaffected by heavy rain	★★★★★ Excellent; radar unaffected by fog	★★★★★ Excellent; radar unaffected by dense fog	★★★★★ Excellent; radar unaffected by snow	★★★★★ Excellent for other-vehicle detection; road boundary not directly measurable	★★★★★ Other-vehicle detection maintained; road ice not directly detectable
Sensor fusion (LIDAR plus camera plus radar — Waymo)	★★★★★ Best-in-class with all sensors active	★★★★ Rain affects LIDAR and camera; radar compensates	★★★★ Heavy rain degrades LIDAR and camera; radar maintains vehicle detection	★★★★ Fog degrades LIDAR and camera; radar compensates	★★★ Dense fog: LIDAR range est. 50–70m (est.); radar compensates partially	★★★ Snow degrades LIDAR and camera; radar compensates for vehicle detection	★★ Heavy snow: LIDAR housing risk plus point cloud noise; radar compensates for vehicles but not road boundary	★★ Black ice and road surface: all sensors limited
Camera plus radar (Tesla FSD)	★★★★★ Excellent camera; radar supplemental	★★★★ Camera degraded; radar compensates for vehicle detection	★★★ Heavy rain severely degrades camera; radar compensates for vehicles	★★★ Moderate fog: camera plus radar; no LIDAR for additional depth	★★ Dense fog: camera severely degraded; radar for vehicle detection only	★★★ Light snow manageable with training-data pattern recognition	★ Heavy snow: camera cannot see lane markings or road edge; radar detects vehicles only	★ Black ice: camera cannot detect; radar cannot detect; GPS weather correlation only

The radar column reveals the core insight: radar is the most weather-robust sensor modality across every adverse condition category. The fact that Waymo includes radar in its triple-redundancy stack and that Tesla has reinstated radar in newer vehicles confirms that both leading AV developers recognize radar’s irreplaceable role in adverse-weather scenarios. A purely camera-based AV system with no radar would have significant weather vulnerabilities across the full spectrum.

Section 5 — All-Weather Deployment Benchmark Scorecard

Weather dimension	Waymo (LIDAR plus camera plus radar)	Tesla FSD (camera plus radar)	Aurora (LIDAR plus camera plus radar, trucking)	2028 outlook
Urban fog	Good: SF commercial service in regular fog; 1550nm LIDAR fog-resistant; SF operational success validates fog capability within ODD	Moderate: camera training breadth compensates for physics limits; radar backup; not deployed as driverless in high-fog commercial service	Good: highway AV in light fog manageable; heavy fog triggers speed reduction; Aurora’s I-45 TX corridor not a major fog market	Waymo’s 1550nm LIDAR fog advantage grows as solid-state LIDAR improves; Tesla’s training breadth continues expanding
Rain (light to moderate)	Good: operational in all operating cities through rain; sensor fusion with radar provides weather robustness	Good: camera plus radar; training on diverse rain conditions from FL, TX, and PNW fleet; supervised FSD through moderate rain nationwide	Good: sensor fusion on highway AV in rain manageable with reduced speeds and increased following distance	Both Waymo (LIDAR plus radar) and Tesla (camera plus radar) handle moderate rain adequately; heavy rain remains a challenge for both
Heavy rain and thunderstorms	Moderate: LIDAR degraded in torrential rain; radar compensates; conservative speed reduction; Austin and Phoenix monsoon season managed	Moderate: camera severely degraded in torrential rain; radar compensates for vehicle detection; conservative speed reduction	Moderate: similar approach to Waymo; conservative speed reduction; Aurora avoids operating in severe thunderstorms	Weather-robust sensor fusion remains a structural competitive advantage for Waymo and Aurora vs Tesla’s camera-heavy approach in extreme rain
Snow (light)	Not commercially deployed in snow; LIDAR handles light snow with housing protection; outside current ODD	Moderate: FSD deployed in snow-belt states; trained on substantial light-snow data; supervised FSD used by owners in light snow	Not commercially deployed in snow conditions on current TX corridor	Waymo’s snow ODD expansion needed for Northeastern market entry; Tesla has most operational experience with light snow deployment
Heavy snow and ice	Not deployed: outside Waymo’s ODD; multi-year research and validation required for heavy-snow commercial service	Not driverless: FSD disengages in heavy snow; owner-supervised only; fundamental camera physics limit	Not deployed: Aurora’s current commercial corridor (TX) avoids snow; heavy snow is outside current ODD	Heavy snow and ice remains the hardest AV challenge for ALL systems regardless of sensor modality; no validated driverless snow operation by any major AV company as of mid-2026
Overall verdict	The all-weather deployment benchmark reveals a fundamental asymmetry in Physical AI’s 2026 deployment reality: ALL major AV systems avoid heavy snow and ice as driverless operating conditions. Waymo’s sensor fusion (LIDAR plus camera plus radar) provides better adverse-weather robustness in fog and moderate rain than Tesla’s camera-dominant approach — but both systems share the same unresolved limitation in heavy snow and ice. The Sunbelt-first deployment pattern is not just a business or regulatory choice — it directly reflects where today’s AV technology performs reliably enough for commercial driverless service. The estimated 47% of the US population living in snow-affected climates (Midwest, Northeast, Mountain West) (est.) is currently outside the operational design domain of all commercial driverless AV services. This is the largest single constraint on the Total Addressable Market of robotaxi services as of mid-2026.

The path to all-weather AV deployment runs through two parallel research streams: (1) improving sensor robustness — solid-state LIDAR with better weather sealing, radar angular resolution improvements, camera computational dehazing — and (2) expanding ODD validation to include snow and ice road conditions through systematic testing in Northern climates. Neither stream has a clear completion date as of mid-2026. The companies that solve heavy-snow driverless operation first will unlock access to the most populous and economically significant US metro areas currently excluded from the AV addressable market.

Sources: Waymo safety and ODD documentation (waymo.com/safety); LIDAR weather performance research, IEEE Transactions on Intelligent Transportation Systems (ieeexplore.ieee.org); Tesla FSD adverse conditions documentation (tesla.com/support/autopilot); NHTSA AV weather performance standards (nhtsa.gov/vehicle-safety/automated-vehicles). All figures marked (est.) are estimates based on physics principles, public disclosures, and third-party research; they have not been independently verified and may differ from official company data.