2026-06-18 — views

Physical AI Public Safety 2026 — Waymo Police Stop Protocol vs Tesla NHTSA Emergency Vehicle Recall: Edge-Case AV Benchmark

Waymo built a driverless police stop protocol. Tesla had a 2021 NHTSA recall for Autopilot failing to detect stationary emergency vehicles on the roadway.

Article 189 in the Physical AI Benchmark Series — Emergency Response and Edge-Case AV Performance

AV systems can achieve 99.9% performance on standard driving scenarios. It is the remaining 0.1% — emergency vehicle response, police traffic stops, manual traffic direction, and construction zone navigation — that defines public safety outcomes and determines regulatory approval. This article benchmarks Waymo and Tesla FSD across the highest-stakes edge cases, with documented real-world incidents, regulatory records, and a public safety scorecard.

Section 1 — Why Emergency and Edge-Case Response Defines AV Public Trust

The AV industry refers to difficult and uncommon driving scenarios as the “long tail” of edge cases. Standard driving — open-road lane-keeping, traffic-light compliance, freeway following — is a solved problem for modern AV systems. The long tail is what separates systems that are safe enough for full public deployment from systems that work most of the time.

Emergency vehicle response is one of the highest-stakes categories in this long tail. A vehicle that fails to yield to an ambulance or fire truck creates a potentially life-threatening obstruction. A vehicle that stops in the wrong location and blocks a fire station exit can delay emergency response by minutes in scenarios where seconds matter. These are not edge cases in the sense of being rare and inconsequential — they are edge cases that can directly determine outcomes in life-threatening situations.

Regulatory bodies have recognized this. The California Public Utilities Commission (CPUC) and the National Highway Traffic Safety Administration (NHTSA) have both issued specific guidance and opened investigations on AV emergency vehicle response. California requires AV operators with commercial permits to document and report any interaction with emergency vehicles to CPUC. NHTSA has used its Standing General Order reporting requirements to build a public dataset of AV-involved incidents, with emergency vehicle interactions specifically flagged.

Documented real-world incidents establish why this benchmark category matters:

Waymo vehicles in San Francisco have been stopped by police and found themselves in a novel legal situation: no driver is present to interact with an officer, answer questions, produce a license and registration, or receive a citation. In documented incidents, officers approached Waymo vehicles and found empty driver seats. The interactions were resolved via remote operator contact through the vehicle’s in-vehicle speaker and screen — but the incidents exposed a gap that neither Waymo nor law enforcement had fully anticipated before deployment at scale.

A Waymo vehicle in San Francisco was reported to have stopped in a location that blocked a fire truck’s path. This type of incident — pulling over in response to an emergency signal but stopping in a tactically wrong location — is distinct from failing to detect an emergency vehicle. The vehicle correctly identified the emergency stimulus and took action, but the action was wrong in context. It reflects the difficulty of training an AV not just to respond to an emergency vehicle but to respond correctly across the full range of spatial situations in which a pull-over maneuver might be executed.

Tesla Autopilot failures to respond to stationary emergency vehicles with lights active led NHTSA to open an investigation that resulted in Safety Recall #21V-857 in 2021. The recall addressed a pattern in which Autopilot did not adequately respond to police cars, fire trucks, and ambulances stopped on roadways with emergency lights active. The recall was addressed via an over-the-air software update — but the underlying issue (camera-only detection of stationary emergency vehicles in complex roadway lighting conditions) remained a known challenge.

Tesla FSD has also had documented struggles with construction zones where lane markings are covered or altered by temporary barriers. NHTSA opened a separate investigation into FSD construction zone performance in 2023. Construction zones present a particularly difficult challenge for AV systems because they alter the road geometry that HD maps or camera-based scene understanding assume will be consistent.

The public trust dimension: A single high-profile incident — an ambulance blocked for an extra five minutes by a confused AV, a police confrontation with an empty vehicle that goes viral — can set the AV industry back years in terms of public acceptance and regulatory policy. Emergency and edge-case response is where AV systems most visibly succeed or fail in front of the public, lawmakers, and regulators.

Section 2 — Waymo’s Emergency and Edge-Case Response Architecture

Edge case category	Waymo approach	Real-world performance	Regulatory status
Emergency vehicle detection (lights/sirens)	LIDAR plus camera plus audio detects emergency vehicle lights (flashing strobes visible to LIDAR and camera) and sirens (microphone array detects audio); system trained to pull over to right and stop when emergency vehicle approaching with lights and sirens active	Functional in tested scenarios; LIDAR makes flashing light detection more reliable than camera-only in bright daylight; siren detection adds audio redundancy	CPUC requires AV operators to document emergency vehicle response; Waymo has published incident reports to CPUC
Police traffic stop protocol	When stopped by police, Waymo vehicle activates a stopped-by-police mode: vehicle pulls to right and stops; remote operator is alerted and can interact with officer via in-vehicle speaker and screen; Waymo has trained 911 call centers in operating cities (SF, Phoenix, LA, Austin) on how to interact with a driverless vehicle	Real incidents: Waymo vehicles in SF pulled over; officer initially confused by absence of driver; interaction via remote operator resolved the situation; Waymo posted a guide for law enforcement	Novel legal territory: who receives the citation? Waymo LLC (the AV operator) has acknowledged it would receive moving violations on behalf of its vehicles; CA DMV has specific rules for AV traffic enforcement
Emergency vehicle blocking and obstruction	Waymo has had incidents where vehicles stopped in locations that interfered with emergency vehicle access; Waymo updated its routing algorithms to avoid stopping in locations (intersections, fire station exits, hospital approaches) that commonly require emergency vehicle clearance	Improvement trajectory: earlier incidents led to specific software updates; Waymo has published commitment to CPUC on obstruction prevention	CPUC ordered Waymo to submit incident reports for any emergency vehicle interaction; ongoing regulatory monitoring
Unprotected left turns	Waymo vehicles execute unprotected left turns (turning left with oncoming traffic, no traffic light green arrow); early Waymo versions were overly conservative and would not turn left until the gap was very large; current version handles unprotected lefts with human-like gap acceptance	Significant improvement from early versions; Waymo has described unprotected left turns as one of the hardest AV maneuvers; current performance is functionally human-like in familiar environments	No specific regulatory requirement; part of general safety performance evaluation
Construction zone navigation	Construction zones alter road geometry from what HD maps contain — a significant challenge for HD-map-reliant systems; Waymo addresses this via map update crews that re-map construction zones before vehicle deployment, and temporary construction zone detection by sensors that detect temporary barriers and adjusted lanes	Real limitation: fast-moving or unexpected construction can outpace map updates; Waymo performs better in construction zones that are stable enough to be remapped	No specific regulatory metric; CPUC reviews construction zone-related incidents reported by AV operators
Manual traffic direction by police and flaggers	Police or construction flaggers using manual hand signals to direct traffic override normal traffic signals; one of the hardest AV edge cases because it requires interpreting human gestures in context	Known AV limitation: recognizing human traffic direction gestures is hard because they are not standardized; Waymo vehicles have struggled with this scenario; current approach may include remote operator takeover when manual direction is detected	CPUC incident reporting applies; no specific regulatory standard for manual direction compliance

Waymo’s multi-sensor architecture provides meaningful redundancy for emergency vehicle detection. LIDAR detects the flashing light strobes of emergency vehicles with high reliability even in bright daylight conditions where camera-based detection can be unreliable. The addition of microphone arrays for siren detection provides a second independent detection channel. This redundancy is architecturally significant: if the camera fails to detect an emergency vehicle due to lighting conditions, LIDAR and audio can still trigger the pull-over response.

The police stop protocol represents Waymo’s most operationally novel edge-case challenge. No prior legal or operational framework existed for a traffic stop of a fully driverless vehicle. Waymo has worked with CA DMV, operating city governments, and local law enforcement to develop protocols — a process that took years of operation to mature. The framework that now exists (remote operator interaction via in-vehicle speaker, Waymo LLC receipt of moving violations, law enforcement training guides) is a genuine institutional innovation that required sustained multi-stakeholder engagement.

Section 3 — Tesla FSD Emergency and Edge-Case Response

Edge case category	Tesla FSD/Autopilot approach	Real-world performance	Regulatory and safety record
Emergency vehicle detection	Camera-based detection of emergency vehicle lights; audio detection via microphones; trained to pull right and stop when emergency vehicle approaching	2021 NHTSA investigation: multiple crashes involving Tesla Autopilot that did not respond appropriately to stationary emergency vehicles with lights active (police cars, fire trucks, ambulances stopped on roadway); resulted in over-the-air update requiring Autopilot to monitor for stationary emergency vehicles	Significant safety record concern: NHTSA Safety Recall #21V-857 covered Autopilot failure to respond to emergency vehicles; the recall was addressed with an OTA update
Police traffic stop	Supervised FSD: human driver present and can interact with officer normally; no driverless police stop scenario currently — Tesla Robotaxi and Cybercab would face this challenge	For current FSD (supervised): human driver handles police interaction normally; future Cybercab challenge: same legal uncertainty as Waymo driverless stops	Future challenge: as Cybercab enters driverless service, Tesla will face the same “who is the driver” question that Waymo has already navigated
Emergency vehicle blocking	FSD training includes not blocking emergency vehicle access; however, FSD has had documented incidents of pulling over in suboptimal locations	Training-based approach; FSD relies on camera and learned behavior rather than HD map updates to identify appropriate pull-over locations	No specific NHTSA emergency vehicle blocking investigation beyond the 2021 stationary vehicle recall
Unprotected left turns	FSD handles unprotected left turns; camera-only system assesses oncoming traffic gap via visual estimation; challenging in complex urban environments	Functional; some reported FSD hesitancy at unprotected lefts in complex urban environments where occlusion limits sightlines	No specific regulatory record on unprotected left turns
Construction zone navigation	FSD has documented struggles with construction zones: altered lane markings, temporary barriers, covered lane lines; NHTSA opened a separate investigation into FSD construction zone performance in 2023	2023 NHTSA investigation into FSD and construction zones: multiple incidents of FSD not correctly navigating reduced-speed or altered-geometry construction zones; resulted in additional FSD updates	NHTSA investigation into FSD construction zone performance; addressed with FSD updates; ongoing monitoring
Manual traffic direction	Known FSD challenge; camera system must interpret human gesture in context of traffic flow; FSD training includes some manual direction scenarios	Documented FSD incidents with police manual traffic direction (Waymo and Tesla both have struggled; human police direction is a recognized AV industry challenge)	Not specifically recalled or investigated; recognized as a hard AV problem industry-wide

The 2021 NHTSA recall (Safety Recall #21V-857) is the most significant item in Tesla’s emergency response safety record. The recall addressed a pattern of Tesla Autopilot failures to detect and respond appropriately to stationary emergency vehicles — police cars, fire trucks, and ambulances stopped on roadways with emergency lights active. This is a camera-specific vulnerability: stationary emergency vehicles in complex road environments, often in bright daylight or with competing light sources, created detection failures that LIDAR-based systems handle more robustly because LIDAR is not affected by ambient light conditions in the same way as cameras.

The 2023 NHTSA construction zone investigation is the second major regulatory action in this category. Construction zones with altered or covered lane markings present a fundamental challenge for systems that rely on camera-based lane detection: the visual cues the system was trained on are absent or misleading. Tesla’s end-to-end neural network approach has improved construction zone navigation through large-scale training data, but the NHTSA investigation documented that failures were still occurring at a rate that warranted regulatory action.

Tesla’s Cybercab driverless vehicle will face the police stop protocol challenge that Waymo has spent years navigating. As Cybercab enters commercial driverless service, Tesla will need to establish the same legal and operational framework — relationship with CA DMV, law enforcement training, remote operator interaction protocols — that Waymo built incrementally through operation in San Francisco beginning in 2022.

Section 4 — The Regulatory and Legal Framework: Who Is Responsible?

Regulatory dimension	Current status	Waymo position	Tesla position
Who receives a traffic citation in a driverless AV?	Novel legal question; no uniform US standard; California has DMV guidance that AV operators receive citations for their fully driverless vehicles	Waymo LLC has acknowledged it would receive moving violations for fully driverless Waymo One vehicles; has worked with CA DMV on this framework	Tesla Robotaxi (supervised): human driver receives citation; future Cybercab (driverless): Tesla or the operator would receive the citation under the same CA framework
NHTSA reporting requirements	Tesla has been under NHTSA Standing General Order: must report all crashes involving Autopilot/FSD/driver assist within 24 hours; Waymo under similar but different reporting obligations	Waymo’s CPUC reporting obligations: detailed incident reports to CPUC for any collision, emergency vehicle interaction, or disengagement	Tesla: NHTSA SGO reporting has produced a large public dataset of Autopilot and FSD-involved crashes; visible through NHTSA data releases
CPUC incident reporting (California-specific)	CPUC requires AV TNC permit holders to file incident reports for: collisions, injuries, emergency vehicle interactions, disengagements	Waymo is subject to CPUC incident reporting; Waymo’s reports are publicly available and show the learning trajectory	Tesla’s Robotaxi as a CA TNC would be subject to the same CPUC reporting if operating driverless
Federal AV safety framework	No comprehensive federal AV safety law passed as of mid-2026; NHTSA operates via existing safety standards plus voluntary guidance plus SGO reporting	Waymo has advocated for a federal AV framework that provides consistent national standards rather than a state-by-state patchwork	Tesla has also advocated for a federal framework; concerns about state-level restrictions on AV deployment have been raised publicly
Liability in AV crash	When a driverless AV is at fault: product liability (AV company) rather than driver negligence; fundamentally shifts the tort framework	Waymo has public statements on liability acceptance for fully driverless vehicles; Alphabet’s deep pockets provide financial backing	Tesla’s supervised FSD: liability question depends on whether the driver was engaged and whether FSD was active; FSD’s limited liability waiver in the user agreement has been tested in litigation

The police stop legal question is the sharpest regulatory edge in driverless AV deployment. When a human driver is stopped by police, a centuries-old legal framework applies: the driver identifies themselves, produces documentation, and receives or contests a citation. A driverless vehicle stops this framework cold. Who is the driver? Who can be cited? Who has authority to consent to a vehicle search?

California has developed working answers to these questions through the CPUC and DMV rulemaking process. The AV operator (Waymo LLC, or in future a Tesla entity operating Cybercab) receives moving violations. The remote operator can interact with law enforcement via in-vehicle communication systems. The vehicle itself cannot consent to a search — that question remains legally unresolved and will likely require state or federal legislation.

The NHTSA SGO reporting system has created an unprecedented level of public data on AV incidents that did not exist before. Every Autopilot and FSD crash is reported within 24 hours. This data has been the primary driver of NHTSA’s regulatory actions against Tesla — the 2021 emergency vehicle recall and the 2023 construction zone investigation were both driven by patterns identified in SGO data. Waymo’s CPUC reporting in California creates a similar accountability structure at the state level, with quarterly public disclosure of incidents, disengagements, and miles driven.

Section 5 — Public Safety Benchmark Scorecard

Public safety dimension	Waymo	Tesla FSD/Robotaxi	Cybercab (est.)	Edge
Emergency vehicle detection	Good — LIDAR plus camera plus audio provides multi-sensor redundancy; stationary and moving emergency vehicle detection	Moderate — camera plus audio; 2021 NHTSA recall for stationary emergency vehicle detection failures	Similar to FSD architecture	Waymo — multi-sensor redundancy; LIDAR specifically addresses the lighting-condition failure mode of camera-only systems
Police stop protocol (driverless)	Developed and documented; law enforcement training underway in SF, Phoenix, LA, Austin; remote operator interaction via in-vehicle speaker/screen; Waymo LLC accepts moving violations	Not applicable for current supervised fleet; future challenge for Cybercab	Will face the same legal and operational challenge that Waymo navigated 2022-2026	Waymo — already built and operationalized the framework through years of driverless operation
Emergency vehicle obstruction	Earlier incidents addressed with software updates targeting routing near fire station exits, hospital approaches, intersections; CPUC oversight drives improvement	Functionally addressed in training; no specific obstruction investigation	Similar challenge	Roughly equal — both improving; Waymo has more documented learning trajectory due to CPUC reporting
Construction zone navigation	Challenged by fast-moving construction that outpaces HD map updates; addressed by proactive remapping crews and sensor-based barrier detection	NHTSA investigation in 2023; addressed with FSD updates; camera-only detection of covered lane markings is a structural vulnerability	Similar to FSD camera-based approach	Roughly equal — industry-wide challenge with different failure modes; Waymo’s map-update approach vs. Tesla’s learned-behavior approach each have distinct limitations
Manual traffic direction	Known limitation; remote operator takeover is the current mitigation when manual direction is detected	Known limitation; camera-based human gesture recognition is hard to generalize across the full range of officer and flagger behaviors	Similar	Roughly equal — industry-wide unsolved problem; neither company has demonstrated reliable manual direction compliance
Regulatory track record	CPUC incident reporting provides public accountability and a documented improvement trajectory from 2021 to 2026; no major federal safety recalls	NHTSA SGO crash reporting; 2021 emergency vehicle recall; 2023 construction zone investigation; two major regulatory actions in this category	No track record yet — starting from a clean slate but also lacking the regulatory framework Waymo has already built	Waymo — more transparent public accountability through CPUC; no major safety recalls equivalent to the 2021 NHTSA emergency vehicle recall
Overall verdict	Emergency and edge-case response is the AV dimension where incidents become national news and trigger regulatory action. Waymo’s multi-sensor architecture (LIDAR plus radar plus audio) provides meaningful redundancy for emergency vehicle detection that camera-only cannot match. Tesla has the more concerning safety record in this dimension: the 2021 NHTSA Autopilot recall for failure to detect stationary emergency vehicles was a significant safety event, and camera-only systems remain more structurally vulnerable to the specific lighting conditions that caused it. Both companies struggle with manual traffic direction — an industry-wide unsolved problem. Waymo’s driverless police stop framework (law enforcement training, remote operator protocol, CA DMV coordination) gives it a head start on the legal and operational challenges that Tesla Cybercab will face as it enters driverless commercial service.

The bottom line on public safety edge-case performance: Waymo and Tesla are not equal in this benchmark category, and the inequality is not minor. Tesla’s 2021 NHTSA recall and 2023 construction zone investigation represent documented, federally-acknowledged failures to handle two of the most critical edge-case categories — emergency vehicle detection and construction zone navigation. Waymo’s incidents (police stop confusion, fire truck obstruction) are also real, but they represent early-deployment operational challenges that were addressed iteratively through CPUC-mandated reporting and software updates, and they did not produce a federal safety recall.

The public trust implication is significant. A federal safety recall is a nationally visible event that shapes public and regulatory perception of AV safety for years. Waymo’s regulatory relationship is characterized by accountability through CPUC reporting and incident-driven improvement. Tesla’s regulatory relationship in this dimension has been characterized by NHTSA-initiated investigations and recalls — a fundamentally different and more adversarial dynamic.

As Cybercab enters driverless commercial service, Tesla will face the police stop protocol challenge, the citation framework question, and the need for law enforcement training programs that Waymo has already built. The question is whether Tesla can navigate that institutional-building process with the same deliberateness that Waymo brought to it — or whether it will encounter incidents during driverless deployment that generate the same kind of federal regulatory response that has already occurred twice in the supervised FSD era.

Section 6 — About This Series

This is article 189 in the Physical AI Benchmark Series. Previous articles in this series have covered the ramp index, the humanoid race, unit economics, global competition, HD mapping, fleet operations, software and OTA, insurance and liability, consumer demand, partnerships, competitive moats, Cybercab versus Model Y, safety data, Waymo Gen 6, Optimus manufacturing, scorecard snapshots, the 2030 forecast scenarios, the investor framework, city expansion, state approval maps, AV weather and climate constraints, the talent war, regulatory calendar, robotaxi fare pricing, the data flywheel comparison, humanoid deployment, supply chain analysis, consumer adoption demand, Waymo valuation and IPO, and many further benchmark dimensions.

This article adds the emergency and edge-case response dimension: how Waymo and Tesla handle the highest-stakes scenarios — emergency vehicle detection, police traffic stops, construction zone navigation, manual traffic direction — that define public safety and regulatory trust in AV systems. The edge-case performance gap between multi-sensor and camera-only architectures is most visible here, and the regulatory record in this category will shape the trajectory of AV public trust for the remainder of the decade.