2026-06-18 — views
Physical AI Weather and Edge Cases — Tesla FSD vs Waymo in Rain, Snow, Fog, and Construction Zones
Waymo excludes snow from commercial ODD; Tesla FSD camera is vulnerable to sun glare. Radar saves both in rain and fog. Snow is Waymo expansion bottleneck.
Article 153 in the Physical AI Benchmark Series — Physical AI Weather and Edge Cases: How Tesla FSD and Waymo Handle Rain, Snow, Fog, Construction Zones, and the Long Tail of Unusual Scenarios
Weather and edge cases represent the most technically demanding dimension of autonomous vehicle development — and a primary reason why neither Tesla nor Waymo has achieved full national-scale driverless commercial deployment. The physics of sensors in adverse conditions, the operational design domain (ODD) restrictions each company imposes, and the long tail of unusual scenarios that training data cannot fully anticipate collectively define the practical ceiling of today’s AV technology. This article is Article 153 in the Physical AI Benchmark Series. It benchmarks sensor physics in adverse conditions, maps ODD restrictions by company and weather type, catalogs documented edge case failures, and delivers a structured scorecard comparing the two leading Physical AI architectures across the full weather and edge-case dimension.
All figures labeled “(est.)” are derived from public disclosures, industry research, documented incident reports, and analyst estimates rather than independently verified primary data.
Section 1 — Sensor Physics: How Weather Affects Cameras, Lidar, and Radar
The fundamental challenge of AV weather performance is sensor physics. Each sensor type interacts differently with precipitation, atmospheric water, and ambient light. Understanding these physics is prerequisite to evaluating why Tesla chose camera-plus-radar and why Waymo maintains camera-plus-lidar-plus-radar despite the cost and complexity of triple-fusion.
| Sensor type | Rain effect | Snow effect | Fog effect | Bright sun / glare | Night |
|---|---|---|---|---|---|
| Camera (visible light) | Moderate degradation — water on lens reduces clarity; heavy rain reduces visibility similar to human vision | Heavy degradation — snow on lens blanks camera; falling snow creates visual noise | Moderate to severe — fog scatters light, reduces contrast and range | Severe — direct sun glare can blind camera temporarily; NHTSA investigated Tesla FSD in sun glare conditions | Moderate — relies on streetlights and headlights; no active illumination |
| Lidar (laser pulses) | Moderate — raindrops scatter laser returns; heavy rain increases false positives | Severe — snowflakes return strong laser pulses mimicking solid objects; accumulation on sensor window blocks beam | Severe — water droplets scatter 905nm and 1550nm laser returns; reduces effective range significantly | Minimal effect — lidar is active illumination, unaffected by visible-light glare | Excellent — active illumination independent of ambient light |
| Radar (millimeter wave) | Excellent — radar penetrates rain with minimal degradation | Good — radar penetrates snow; some accumulation on radome can affect signal | Excellent — radar penetrates fog with minimal degradation | Excellent — unaffected by visible light | Excellent — active illumination independent of ambient light |
| Camera plus radar fusion (Tesla HW4) | Camera degrades in heavy rain; radar maintains detection — fusion preserves safety margin | Camera degrades severely in snow; radar maintains detection but no fallback lidar | Camera degrades in fog; radar maintains — fusion handles significantly better than camera-only | Camera vulnerable to glare; radar unaffected — NHTSA investigated Tesla FSD sun glare; OTA improvement deployed | Camera needs headlights and streetlights; radar provides excellent active detection |
| Camera plus lidar plus radar fusion (Waymo) | Camera degrades; lidar moderate degradation; radar excellent — triple redundancy provides resilience | Camera degrades severely; lidar degrades severely in falling snow; radar good — dual vulnerability in heavy snow | Camera degrades; lidar degrades severely; radar excellent — radar is the critical all-weather sensor | Camera vulnerable to glare; lidar unaffected by visible light; radar unaffected — strong structural advantage in glare and sun | Camera needs ambient light; lidar excellent active illumination; radar excellent — two independent active sensors |
| Key insight | Radar is the all-weather sensor; lidar fails in heavy precipitation; camera fails in glare and darkness. Tesla’s camera-plus-radar is robust in rain, fog, and night via radar. Waymo’s triple-fusion has redundancy but lidar vulnerability in snow is a genuine operational gap. |
The Radar Paradox
The central irony of the sensor physics comparison is that radar — the oldest and cheapest of the three sensor types — is the most weather-robust. Both Tesla’s architecture (camera-plus-radar) and Waymo’s architecture (camera-plus-lidar-plus-radar) depend on radar as the fallback all-weather sensor. Lidar, despite its superior 3D point-cloud resolution in clear conditions, is the weakest sensor in precipitation. Camera, despite its ability to read text and interpret complex visual scenes, is the weakest sensor in glare and darkness. This creates a counterintuitive result: Tesla’s simpler and cheaper sensor architecture may be more weather-robust than Waymo’s more expensive triple-fusion stack, because Tesla’s radar is not masked by a dominant lidar that fails in snow.
Section 2 — Operational Design Domain (ODD) Restrictions by Company
An ODD defines the conditions under which an AV system is certified to operate without human intervention. ODD restrictions are as important as sensor performance — a system can have superior sensor fusion and still fail commercially if its ODD excludes major geographic markets.
| Condition | Waymo commercial ODD | Tesla FSD supervised ODD | Tesla Robotaxi Austin ODD (est.) | Notes |
|---|---|---|---|---|
| Rain (light to moderate) | Operational with reduced speed (est.) | Operational | Operational (est.) | Both handle light to moderate rain; radar maintains detection in both architectures |
| Rain (heavy) | Operational with restrictions; heavy rain may trigger safe stop (est.) | Operational but disengagement rate rises (est.) | Restrictions (est.) | Heavy rain degrades both systems; both companies reduce capability or trigger safe stop |
| Snow (any) | NOT operational — Waymo does not operate commercially in snow; all 4 commercial markets (SF, Phoenix, LA, Austin) are low-snow geographies; explicit ODD restriction | Supervised FSD: driver can engage in snow with driver bearing responsibility; intervention rate higher (est.) | Not in Austin ODD (Austin receives minimal snowfall annually) | Waymo’s explicit snow exclusion is a fundamental geographic expansion barrier |
| Fog (light to moderate) | Operational | Operational | Operational (est.) | Both handle light to moderate fog |
| Fog (dense) | Reduced speed; may trigger safe stop in near-zero-visibility fog (est.) | Driver intervention rate rises in dense fog | Restrictions (est.) | Dense fog challenges both architectures |
| Night | Fully operational — lidar active illumination and radar both independent of ambient light | Fully operational — camera with headlights plus radar | Fully operational (est.) | Waymo has structural night advantage with dual active-illumination sensors |
| Sun glare | Fully operational — lidar and radar unaffected by visible-light glare | NHTSA investigated sun glare issue; OTA update deployed; improved but camera-based glare risk persists | Operational with residual glare challenge (est.) | Waymo structural advantage in direct sun glare via lidar and radar |
| Construction zones | Requires pre-mapped construction zone update; unmapped construction triggers vehicle stop and remote assistance request (est.) | Driver must monitor; intervention rate significantly higher in construction zones | Active challenge (est.) | Both systems struggle with novel construction layouts not in training data or HD map |
| Unmarked or faded road markings | Challenging — HD map provides lane context but faded paint reduces lane detection confidence | Very challenging — FSD relies heavily on visible lane markings; faded paint is a documented failure mode | Active challenge (est.) | Camera-based lane detection is structurally vulnerable to degraded road markings |
| Temporary traffic control (flaggers) | Safe stop plus remote assistance; AV cannot interpret human flagger gestures reliably (est.) | Documented FSD failure mode; driver intervention required when human flaggers are present | Active challenge (est.) | Human gesture interpretation for traffic control remains an unsolved AV problem across all companies |
The ODD as a Commercial Constraint
Waymo’s ODD restrictions are the most consequential commercial constraint in Physical AI today. By explicitly excluding snow from its commercial operating domain, Waymo cannot expand to Chicago, Boston, New York, Denver, Minneapolis, or any major US snow market without either solving the lidar-in-snow problem or accepting a fundamentally inferior product in those markets. The 10 largest US metro areas by population include multiple snow markets that Waymo currently cannot serve commercially. Tesla’s camera-plus-radar architecture allows supervised FSD operation in snow markets, giving Tesla a broader geographic operating envelope at the cost of requiring human oversight.
Section 3 — Tesla FSD Documented Edge Case Challenges
Tesla FSD’s end-to-end neural network approach — which has evolved from modular rule-based systems through v12 and v13 — has improved qualitatively across many edge cases. However, documented failures from NHTSA investigations, driver reports, and published disengagement data reveal persistent challenges in specific scenarios.
| Edge case | FSD behavior | Status | Notes |
|---|---|---|---|
| Sun glare (direct camera blinding) | FSD temporarily loses visual input during direct sun glare events; NHTSA investigation conducted in 2024 | OTA update deployed; residual risk in extreme glare conditions | Camera-only visual input has no alternative sensor to fall back on when camera is blinded by direct sun |
| Phantom braking | FSD applies unexpected hard braking for objects that are not real hazards; NHTSA investigation 2023 | OTA update significantly improved frequency; occasional reports persist | Phantom braking caused documented real-world rear-end crashes; classified as serious safety concern by NHTSA |
| Construction zones (novel layouts) | FSD misidentifies lane boundaries in construction setups not matching training data patterns; driver intervention required | Active improvement area; disengagement rate in construction zones remains elevated (est.) | Novel construction configurations by definition cannot be pre-trained — represents a fundamental long-tail challenge |
| Emergency vehicles (stationary) | FSD initially failed to recognize and yield to stationary emergency vehicles; NHTSA 2021 investigation | OTA update improved performance; NHTSA continues monitoring | Emergency vehicles in unusual positions (stationary, at angles, with complex light patterns) remain challenging |
| Roundabouts | FSD struggles with multi-lane roundabout navigation; driver often takes manual control | Active improvement area; v14-plus showing progress | Roundabouts have jurisdiction-varying right-of-way rules; multi-lane navigation requires complex gap acceptance |
| Low-speed urban complexity | FSD performs well at highway speeds; reduced performance in complex urban low-speed scenarios including parking lots and ambiguous pedestrian crossings | Improving significantly with v12 and v13 end-to-end neural net approach vs prior rule-based system | End-to-end model meaningfully improved low-speed complex scenarios that rule-based systems could not handle |
| Left turns across traffic | Historically challenging; drivers frequently disengage at complex unprotected left turns | Improved in v13 and v14; disengagement rate still elevated vs straight driving and right turns (est.) | Unprotected left turns require gap acceptance decisions under time pressure with incomplete information |
| Cyclists and scooters | Classification challenges with unusual two-wheeled vehicles; prediction errors in urban cycling environments | Active improvement; neural net approach handles novel vehicle types better than rule-based predecessor | Less common vehicle types have proportionally less training data; bicycles and scooters are edge cases within training distributions |
The Long-Tail Problem for Neural Networks
Tesla’s shift to end-to-end neural networks (v12 onward) is a direct response to the long-tail problem. Rule-based systems require explicit programming for every scenario — which fails catastrophically when encountering scenarios the engineers did not anticipate. Neural networks generalize from training data, potentially handling novel scenarios better. However, neural networks still fail on out-of-distribution inputs — scenarios that appear infrequently in training data. The long tail of AV edge cases (unusual construction, rare vehicle types, extreme weather, unusual pedestrian behavior) by definition appears rarely in training data. The engineering challenge is not solved by choosing neural networks over rules — it is reduced but not eliminated.
Section 4 — Waymo Edge Case Handling
Waymo’s approach to edge cases relies on three complementary mechanisms: HD map context, sensor redundancy (triple-fusion), and remote operator assistance for situations the autonomous system cannot resolve. This produces a different failure mode profile than Tesla’s — Waymo tends to fail by stopping safely and requesting human remote help, while Tesla’s failure mode is more likely to require driver physical intervention.
| Edge case | Waymo behavior | Status | Notes |
|---|---|---|---|
| Construction zones (pre-mapped) | Handles well — HD map updated with known construction layout; vehicle navigates with high confidence in mapped zones | Strong performance in known, pre-mapped construction | Mapping team updates construction zones as detected; turnaround time estimated at 1 to 7 days per zone |
| Construction zones (unmapped) | Vehicle decelerates and stops safely; remote operator contacted; may re-route or wait for operator guidance | Known operational gap — novel construction requires remote assistance; cannot self-resolve | Dependency on HD map currency creates vulnerability for sudden road changes between map updates |
| Emergency vehicles | Waymo has published performance data showing reliable yielding to emergency vehicles (active lights and siren) | Strong — specific training on emergency vehicle response; lidar and radar detect emergency vehicles regardless of light conditions | Waymo’s documented emergency vehicle response is one of its more publicly evidenced safety capabilities |
| Pedestrian behavior (unusual) | Jaywalkers, unusual crossing patterns, crowded sidewalks — Waymo trained heavily on San Francisco urban pedestrian behavior | Strong in trained markets; performance in new geographic markets with different pedestrian culture may differ | San Francisco’s dense, varied urban pedestrian environment is arguably the world’s best AV training ground for pedestrian edge cases |
| Cyclists | Strong — urban cycling is common in Waymo training markets; HD map encodes bike lane locations and cycling infrastructure | Strong in mapped markets with cycling infrastructure | Bike lane HD map context augments sensor-based cyclist detection; camera-plus-lidar provides better cyclist classification than camera alone |
| Night and low-light | Strong — lidar active illumination plus radar both operate independently of ambient light; camera quality less critical | Structural advantage — two active-illumination sensors independent of ambient light | Waymo’s dual active-illumination architecture provides a genuine structural night advantage over camera-dependent systems |
| Rain (heavy) | Reduced performance — lidar point cloud degrades significantly in heavy rain; safe stop may trigger | Known gap; partially mitigated by operating in low-precipitation markets (SF, Phoenix, LA, Austin) | Geographic ODD selection partially mitigates heavy rain vulnerability but limits addressable market |
| Snow | Does not operate commercially in snow — explicit ODD restriction across all four current commercial markets | Active capability gap; geographic expansion to snow-intensive markets blocked without solving lidar snow problem | Expanding to snow markets (Chicago, Boston, NYC, Denver) requires either a snow-robust sensor architecture or accepting supervised-only operation |
Remote Assistance as a Safety Architecture Choice
Waymo’s remote operator model is a deliberate architectural choice that reflects a different safety philosophy than Tesla. When Waymo’s autonomous system encounters a scenario it cannot resolve with high confidence, it stops safely and routes the situation to a remote human operator who can provide guidance (re-route, proceed at reduced speed, wait for clearing). This produces a measurably safer failure mode — safe stop with human oversight — at the cost of operational scalability. Remote operators represent a per-ride labor cost that scales with fleet size. As Waymo’s fleet grows from tens of vehicles to thousands, the remote operator ratio is a key operational leverage point. Tesla’s approach does not have remote operators — it relies on the supervised driver to intervene when FSD reaches its confidence limit.
Section 5 — Weather and Edge Case Benchmark Scorecard
| Condition | Tesla FSD advantage | Waymo advantage | Edge | Notes |
|---|---|---|---|---|
| Rain (light to moderate) | Camera plus radar handles well | Camera plus lidar plus radar handles well | Even | Both operational in light to moderate rain; radar is the robust sensor in both architectures |
| Snow | Camera plus radar — radar robust in snow; camera degrades but driver supervises | Does not operate commercially in snow | Tesla (can attempt with driver; Waymo excluded entirely) | Tesla operates in snow markets with driver oversight; Waymo is excluded from commercial snow-market operation |
| Fog | Camera degrades; radar maintains; reduced total capability | Camera degrades; lidar degrades in dense fog; radar maintains | Even (radar saves both) | Both architectures are limited in dense fog; radar is the critical all-weather sensor for both |
| Sun glare | Camera-based; NHTSA investigation conducted; OTA improved but residual camera glare risk | Lidar and radar unaffected by visible-light glare; no structural glare issue | Waymo | Waymo has a genuine structural advantage in sun glare via non-camera active sensors |
| Night | Camera (requires ambient light from headlights and streetlights) plus radar (active) | Camera plus lidar (active illumination) plus radar (active) | Waymo (dual active illumination) | Waymo’s two active-illumination sensors provide a structural night advantage |
| Construction zones | Higher disengagement rate; driver intervention typically required | Handles pre-mapped zones well; stops for unmapped zones; remote operator assists | Even (different limitations) | Tesla relies on driver; Waymo relies on map currency and remote operators |
| Pedestrians (complex urban) | End-to-end neural net improving in complex pedestrian scenarios | Strong in trained urban markets with extensive pedestrian training data | Slight Waymo (more driverless urban miles trained) | Waymo has accumulated more driverless urban miles in complex pedestrian environments |
| Geographic expansion | Can operate with supervised FSD in snow and ice markets; broader geographic ODD | Restricted to low-snow geographies for commercial driverless operation | Tesla (broader geographic ODD under supervision) | Tesla’s camera-plus-radar allows supervised FSD in all weather; Waymo cannot commercially operate driverless in snow |
| Overall verdict | Tesla’s camera-plus-radar architecture is weather-robust in rain, fog, and night via radar; structurally vulnerable in direct sun glare and relies on driver oversight for snow markets. Waymo’s triple-fusion is superior in glare and night; structurally limited in heavy precipitation due to lidar and commercially excluded from snow markets. The paradox: Waymo’s multi-sensor redundancy handles more conditions better within its operational domain, but Tesla’s camera-plus-radar covers more geographic conditions (including snow markets with driver supervision). For national-scale commercial driverless deployment, neither architecture fully solves weather today — but solving snow is Waymo’s larger commercial expansion bottleneck. |
The Snow Bottleneck as a Strategic Constraint
The single most consequential weather finding in this benchmark is Waymo’s snow exclusion. Snow is not an edge case in the US market — it is a dominant weather condition for six or more months per year across the Midwest, Northeast, and Mountain West. Chicago (population 2.7M metro area), Boston, New York, Denver, Minneapolis, Detroit, and dozens of other major markets experience significant snowfall annually. Waymo’s commercial strategy of operating exclusively in low-snow Sun Belt cities is a rational response to the lidar-in-snow limitation — but it means that Waymo’s addressable US commercial robotaxi market is structurally constrained to the approximately 30% of the US population living in low-snow geographies. Tesla’s camera-plus-radar architecture does not have this constraint under supervised FSD, giving Tesla a broader theoretical deployment envelope even before the question of full driverless capability is resolved.
Note: All figures labeled “(est.)” are derived from public disclosures, industry research, analyst estimates, NHTSA investigation reports, and reported data as of mid-2026. This article does not constitute safety certification or regulatory assessment.
Sources
- Waymo operational design domain — Waymo safety report ↗
- NHTSA Tesla FSD sun glare investigation — NHTSA ↗
- Tesla FSD phantom braking NHTSA investigation — NHTSA ↗
- Lidar performance in adverse weather — IEEE Intelligent Transportation Systems ↗
- Tesla quarterly vehicle safety report — Tesla ↗