Prologue — The Year the Word "Autonomous" Split
In 2026, the term "autonomous driving" finally fractured. Where 2018-2020 treated "autonomous" as synonymous with robotaxis, the word now splits into at least four streams.
- Robotaxi — Waymo, Zoox, Pony.ai, WeRide, AutoX, Baidu Apollo Go.
- Consumer ADAS (driver assistance) — Tesla FSD v13, Mobileye SuperVision, Mercedes Drive Pilot, Ford BlueCruise, GM Super Cruise, BMW Highway Assistant.
- Autonomous trucking — Aurora Driver, Kodiak Robotics, Plus.ai, Gatik, Einride, Outrider.
- Last-mile delivery robots — Nuro, Starship Technologies, Serve Robotics, Coco Robotics, Kiwibot.
Each stream uses a different technology stack, regulator, and business model. One event reshaped the field — in December 2024, GM officially ended the Cruise robotaxi business. Just over a year after the October 2023 SF pedestrian-drag incident, the entire venture vanished.
That event set the tone for every other AV company in 2026.
- Cumulative valuation is not protection — Cruise had taken in over USD 10 billion.
- A single event can kill a company — nine years, hundreds of thousands of miles, one accident.
- Public acceptance beats technology — SF residents' anger was the deciding factor.
Around the same time, Waymo took a different path. LA public launch in August 2024, expansion plans for Miami and Atlanta in November 2024, partnership with Uber. The gap between the two companies was not technology — it was safety culture and incident response.
Chapter 1 · The Five Levels and What 2026 Actually Looks Like
SAE J3016 defines driving automation in six levels (0-5).
- Level 0 — no automation.
- Level 1 — single-feature assist (cruise control etc.).
- Level 2 — partial automation (Tesla Autopilot, parts of Mercedes). Driver remains responsible.
- Level 3 — conditional automation (Mercedes Drive Pilot 60 mph in CA/NV, BMW). System is responsible until takeover request.
- Level 4 — full automation in limited domains (Waymo, Zoox, Pony.ai, Aurora Driver).
- Level 5 — full automation everywhere (does not exist in 2026).
The May 2026 reality.
- Level 2 ADAS — over 70 percent of new cars. Tesla, Hyundai HDA2, BMW, Mercedes.
- Level 3 — Mercedes EQS and S-Class in select states. Honda Sensing Elite in Japan. BMW 7 Series.
- Level 4 robotaxi — Waymo in five cities, Zoox in pilot, Pony.ai, WeRide, Baidu, AutoX.
- Level 4 trucking — Aurora Driver in Texas, Kodiak, Plus.ai.
- Level 5 — does not exist. Unlikely before 2030.
Chapter 2 · Waymo — the Longest-Running Robotaxi
Waymo (waymo.com) is an Alphabet subsidiary. It started as Google X self-driving in 2009 and was spun off in 2016.
- Waymo One — commercial robotaxi service, hailed via app.
- Waymo Via — autonomous trucking (partially wound down in 2023).
- 2024 rides — about 4 million cumulative, over 150,000 weekly.
- Service area — Phoenix (since 2018), SF (2023), LA (2024), Austin (Uber partnership 2024), Miami and Atlanta expanding.
Waymo's stack is conservative.
- Hardware — in-house LiDAR (one 5th-gen plus four short-range), 29 cameras, 6 radars. Vehicles are Jaguar I-PACE and Hyundai IONIQ 5.
- HD maps — pre-built cm-grade maps for every service area, updated on change detection.
- Models — modular architecture (perception, prediction, planning, control as separate modules). Some modules are neural, but not the whole stack.
- Remote assistance — Fleet Response operators guide cars through edge cases.
Waymo's LA and Austin expansion is hailable via the Uber app. Hyundai IONIQ 5 vehicles enter the Miami fleet in 2025.
Chapter 3 · The Death of GM Cruise — End of a USD 10 Billion Bet
Cruise (getcruise.com) was founded by Kyle Vogt in 2013, acquired by GM in 2016.
- SF driverless commercial operation — granted August 2023 by California PUC.
- Trigger event — October 2, 2023, near SF Market and Embarcadero. A pedestrian struck by another vehicle was then dragged about 6 meters by a Cruise vehicle.
- DMV permit suspension — October 24, 2023, California DMV suspended Cruise's autonomous operation permit.
- CEO resignation — Kyle Vogt resigned November 2023.
- Business shutdown — December 2024, GM ended the robotaxi business and cancelled the Origin shuttle.
What remained.
- People — many Cruise engineers moved to Tesla, Waymo, Aurora.
- Vehicles — roughly 800 Origin shuttle prototypes sit in warehouses.
- Capital loss — GM took a USD 2.8 billion non-operating loss in Q4 2024.
This event shaped the incident-response playbook for every other robotaxi company in 2026.
Chapter 4 · Zoox — the Bidirectional Pod Bet
Zoox (zoox.com) was founded in 2014 by Jesse Levinson and Tim Kentley-Klay. Amazon acquired it in 2020 for roughly USD 1.2 billion.
- Bidirectional design — no driver's seat. Four passengers face each other.
- In-house manufacturing — a Fremont, California factory. Not a Toyota or Hyundai chassis but a clean-sheet design.
- Commercial pilots — employee shuttle in Las Vegas from August 2024, SF pilot from November 2024. Public launch targeted in 2025.
What sets Zoox apart.
- 75 mph — same as Waymo.
- GeoFenced — limited service area, similar to Waymo.
- Bidirectional driving — turns in tight alleys without reversing.
The 2025 challenge — Zoox is close to formal commercial launch. With Waymo already in five cities, differentiation is the question. Amazon's deep capital reduces short-term revenue pressure.
Chapter 5 · Pony.ai · WeRide · AutoX — China's Robotaxi Trio
The three companies leading robotaxi in China.
- Pony.ai (pony.ai) — founded 2016 by James Peng and Tiancheng Lou (both ex-Baidu USA). IPO'd on Nasdaq on November 13, 2024 (ticker PONY). Robotaxis in Beijing, Guangzhou, Shanghai, Shenzhen. Limited California operations.
- WeRide (weride.ai) — founded 2017 by Tony Han (ex-Baidu USA). IPO'd on Nasdaq on October 25, 2024 (ticker WRD). Guangzhou, Shenzhen, Beijing. Expanded to Dubai (UAE).
- AutoX (autox.ai) — founded 2016 by Jianxiong Xiao. Shenzhen, Shanghai, Guangzhou. Backed by Alibaba.
- Baidu Apollo Go (apollo.auto) — Baidu's robotaxi brand. In 2024 operates in roughly 10 cities including Wuhan, Beijing, Shenzhen, Guangzhou. Wuhan's driverless coverage exceeded about 3,000 km² at one point.
The shared strengths of the four Chinese robotaxi operators.
- National infrastructure — autonomous pilot zones by city, V2X (vehicle-infrastructure communication) priority deployment.
- Labor costs — safety driver wages roughly half those in the US, reducing the pressure to remove the driver.
- Data volume — Wuhan alone has periods where daily mileage exceeds the sum of all US robotaxi operations.
The weakness — US market entry is constrained. In 2024 the US categorized Chinese vehicle software as a national security risk, limiting data collection and operations.
Chapter 6 · Tesla FSD v13 — the End-to-End Neural Bet
Tesla's Full Self-Driving is the most debated AV system of 2026.
- FSD v12 (January 2024) — replaced modular perception, planning, and control with a single end-to-end neural network. Elon Musk claimed 300,000 lines of C++ were removed.
- FSD v13 (January 2025) — stronger vision-only perception, more simulation-based training, improved traffic-light handling.
- Hardware — HW3 (since 2019), HW4 (since 2023), eight cameras. Radar removed in 2021. No LiDAR.
- Pricing — USD 8,000 lump sum or USD 99 per month in 2025 (US).
The points of debate.
- The name itself — "Full Self-Driving" is under NHTSA and FTC scrutiny. The actual classification is Level 2.
- Camera-only versus LiDAR — Waymo and Zoox use LiDAR plus cameras. Tesla went cameras-plus-radar, then cameras-only with on-board AI-4 inference.
- Verification data — Tesla publishes per-mile crash rates in its own report, but NHTSA and IIHS analyses need to adjust for population differences (price tier, driver demographics).
2024-2025 NHTSA Autopilot and FSD investigations.
- April 2024 — NHTSA closed the Autopilot investigation. After reviewing 956 incidents, NHTSA ordered a voluntary recall handled by OTA software.
- October 2024 — NHTSA opened a separate FSD investigation, including four low-visibility collisions.
Tesla's promise — Robotaxi (Cybercab) reveal June 2025, customer OS integration ongoing in 2026, expansion of Austin and Phoenix driverless pilots.
Chapter 7 · Wayve — the British End-to-End Challenger
Wayve (wayve.ai) was founded in London in 2017 by Cambridge alums Alex Kendall and Amar Shah.
- End-to-end neural network — similar to Tesla, but trained without HD maps.
- GAIA-2 model (2024) — a video-generation foundation model for self-driving. Used to generate simulation and synthetic data.
- Capital — USD 1.05 billion Series C in May 2024 with NVIDIA, Microsoft, and SoftBank participating.
- OEM partnerships — UK Tesco and Asda last-mile pilots. Limited pilots with Uber.
Wayve's differentiator is "self-driving without maps." It does not pre-build HD maps. The neural network drives directly from real-time camera and radar input. In theory the cost to enter a new city is low.
The 2026 challenge — Wayve has no commercial driverless robotaxi yet. Pilots and OEM licensing are the main revenue lines. The company is expanding into the AV simulation market with GAIA-2, competing against NVIDIA Cosmos.
Chapter 8 · Mobileye — Partner to Intel, Volkswagen, and Audi
Mobileye (mobileye.com) was founded in Jerusalem in 1999 by Amnon Shashua. Intel acquired it in 2017 for USD 15.3 billion, and it was re-listed on Nasdaq in 2022.
- EyeQ chips — Mobileye's in-house SoC. New-car ADAS market share estimated above 70 percent.
- SuperVision — camera-based Level 2+ ADAS. Shipped in Zeekr 001 (Geely).
- Chauffeur — cameras plus LiDAR plus radar. A Level 4 urban autonomy solution.
- Drive — Mobileye's robotaxi platform. Volkswagen ID. Buzz pilot in Hamburg.
- REM (Road Experience Management) — anonymous customer vehicles upload road data, producing automatic cm-grade HD maps.
Mobileye's business model is licensing to other OEMs. Where Waymo and Cruise ran their own services, BMW, VW, Audi, Ford, and Honda integrate Mobileye into their own vehicles.
The 2024 challenge — Mobileye lowered revenue guidance in January 2024 and the stock dropped. Competition from China's in-house ADAS (Huawei, XPeng) intensified. 2025 revenue recovered but margin pressure persists.
Chapter 9 · NVIDIA DRIVE Thor — Autonomy on Silicon
NVIDIA (nvidia.com) has become a key pillar of AV infrastructure in 2026.
- DRIVE Thor (announced 2024 GTC, shipping 2025) — 2,000 TOPS AI inference. From Cosmos-based model training to inference.
- DRIVE Hyperion 9 — integrated platform with cameras, LiDAR, and radar.
- NVIDIA Cosmos (January 2025) — a video-generation foundation model for AV and robotics training.
- NVIDIA Omniverse — the standard for AV simulation.
- Customers — Mercedes-Benz, Volvo EX90, JLR, Hyundai, NIO, Li Auto, Lotus, XPeng.
NVIDIA's stance — "supply chips, middleware, and simulation so OEMs can build their own AV stacks."
This sets up competition with Mobileye (which emphasizes a turnkey solution) and Qualcomm Snapdragon Ride (which provides chips plus partial software).
Chapter 10 · Mercedes Drive Pilot — the First Real Level 3 Pilot
Mercedes Drive Pilot is the only formally certified Level 3 system in the US in 2026.
- Speed limit — up to 60 mph (95 km/h).
- Operating conditions — highway, clear lane markings, lead vehicle present, good weather.
- Certification — California DMV (2023), Nevada (2023), DC pilot. Germany under review to extend from 60 mph to 95 mph.
- Vehicles — Mercedes EQS, S-Class.
Why Drive Pilot matters — "the vehicle is responsible when conditions are met" carries real meaning. The driver can watch a film, read email, or play games. Until a takeover request (a 10-second countdown), liability rests with Mercedes. No other OEM accepts this level of liability so explicitly in the US.
BMW Highway Assistant at 70 mph entered a 2024 US pilot. Honda Sensing Elite ships in the Honda Legend with Level 3 Japan certification (2021), limited to 100 units in Japan.
Chapter 11 · Ford BlueCruise vs GM Super Cruise vs Tesla FSD
The big three US ADAS, compared.
- Ford BlueCruise (F-150 Lightning, Mustang Mach-E) — 130,000 miles of pre-mapped roads, mostly highways. Hands-free. Eye-monitoring camera required.
- GM Super Cruise (Cadillac, Chevrolet) — over 200,000 miles pre-mapped. Auto lane changes. Trailer-tow mode.
- Tesla Autopilot and FSD — all roads (highway plus surface streets). No pre-mapping (full vision). Eye monitoring weaker than GM and Ford.
All three are Level 2, but the user experience differs.
- Ford and GM — hands-free but eyes-on. Limited to pre-mapped highways.
- Tesla — hands always on the wheel, but usable on surface streets. FSD v12 and later handle nearly every scenario.
In 2024, IIHS rated partial automation with GM Super Cruise above Ford BlueCruise above Tesla on driver-monitoring strength. Tesla scored low on eye monitoring.
Chapter 12 · Volvo EX90 and the LiDAR Mainstreaming Trend
The Volvo EX90 (US launch 2024) was the first volume car to ship with Luminar Iris LiDAR as standard.
- Luminar Iris — 1550 nm wavelength, 250 m detection range.
- NVIDIA DRIVE Orin plus Xavier compute.
- Ride Pilot — Level 3 autonomous option, California pilot 2025.
Volvo's decision matters. While Tesla went camera-only, Volvo, Polestar, Mercedes, and BMW chose to integrate LiDAR.
Other LiDAR-equipped vehicles in 2026.
- Polestar 3 — Luminar.
- Mercedes EQS Drive Pilot — Valeo Scala.
- BMW iX5 Hydrogen Highway Assist — Innoviz.
- NIO ET7, XPeng G9 — RoboSense, Hesai.
- Lotus Eletre — Hesai AT128.
LiDAR cost fell from roughly USD 75,000 in 2020 to USD 500-1,500 in 2026, low enough for volume vehicles.
Chapter 13 · Autonomous Trucking — Aurora Driver Goes Driverless in Texas
Autonomous trucking has a sharper business case than robotaxi. Truck driver shortage plus long straight roads equals a strong fit for automation.
- Aurora Innovation (aurora.tech) — founded 2017 by Chris Urmson (ex-Waymo CTO), Sterling Anderson (ex-Tesla Autopilot), Drew Bagnell (ex-Uber autonomy). In December 2024 started driverless commercial freight runs over roughly 250 miles between Dallas and Houston, using PACCAR Kenworth and Volvo VNL trucks.
- Kodiak Robotics (kodiak.ai) — founded 2018. Atlanta-Dallas freight. Military programs (US Army autonomous vehicles).
- Plus.ai (plus.ai) — founded 2016 in China and the US. Trucks plus Iveco and Daimler partnerships.
- Gatik (gatik.ai) — short-haul B2B delivery (Walmart, Sam's Club, KFC). Texas and Arkansas.
- Outrider (outrider.ai) — yard-only autonomous trucks. Inside logistics centers. UPS, Georgia-Pacific.
- Einride (einride.tech) — Sweden. Cabless electric autonomous truck called the Pod. US and UAE pilots.
Gone.
- TuSimple — effectively bankrupt by 2024. California operations halted. Chinese business spun out.
- Embark — closed in 2023.
Aurora Driver's milestone — December 2024 to April 2025 saw roughly 200 driverless truck runs with no reported crashes. The company plans further lane expansion through 2026.
Chapter 14 · Nuro · Starship · Serve Robotics — the Delivery Robot Split
Nuro (nuro.ai) was founded in 2016 by Jiajun Zhu and Dave Ferguson (both ex-Waymo). The R1, R2, and R3 autonomous delivery vehicles.
- 2024 business pivot — reduced its own vehicle business and shifted to AV technology licensing.
- Domino's, Walmart, CVS pilots — Houston and Phoenix.
Starship Technologies (starship.xyz) was founded in 2014 by Skype co-founder Janus Friis. Six-wheel sidewalk delivery robots. Roughly 100 US college campuses and Milton Keynes in the UK.
Serve Robotics (serverobotics.com) was spun out from Uber in 2017. Uber Eats food deliveries in LA, SF, and Dallas. IPO'd in 2024. Built on NVIDIA Jetson.
Coco Robotics (cocodelivery.com) was founded in LA in 2020. Hybrid human-remote-plus-AI delivery robots.
Kiwibot (kiwibot.com) was founded in 2017 in Colombia and the US. College campus delivery.
The 2026 reality for delivery robots.
- Cost per mile — competing with human couriers per delivery mile. Highly sensitive to city minimum wage.
- Infrastructure constraints — sidewalk width, wheelchair access. SF and New York are tightening sidewalk-robot rules.
- Weather — rain and snow are hard. Northern US cities pause operations December through February.
Chapter 15 · Korean Autonomy — 42dot · Hyundai Mobis · Kakao Mobility
Korea has the potential of an AV power. Yet commercial robotaxi remains in pilot in 2026.
- Hyundai Motor Group — operates Motional (Boston). 100 percent ownership after acquiring Aptiv's stake in 2023. Pilots in Las Vegas with Uber and in LA.
- 42dot (42dot.ai) — founded 2019 by Song Chang-hyun (former Naver CTO). Acquired by Hyundai Motor Group in 2022. Self-driving shuttle pilots in Gangnam and Cheonggyecheon Seoul. ROBORIDE service.
- Hyundai Mobis Light — announced 2024. Hyundai Mobis's in-house autonomous SDV platform on NVIDIA DRIVE.
- Mando IONIQ Lab — Mando-Hyundai joint program for ADAS components and LiDAR integration.
- Stradvision (stradvision.com) — based in Pohang. ADAS perception SW, licensed to EU and US OEMs.
- RideFlux (rideflux.com) — founded 2018. Self-driving shuttle pilot in Jeju. KAIST roots.
- Kakao Mobility — works with Carrot Insurance on AV data. Pilot for AV hailing in the KakaoT app.
Seoul robotaxi pilots.
- 2024 — pilot in parts of Gangnam-gu and Seocho-gu. MOLIT pilot zones.
- 2025-2026 — expanded to Gangbuk and Cheonggye. Driverless operation still rides with a safety driver on board.
Two challenges Korea faces. First, public acceptance — a conservative approach learned from the Cruise incident. Second, business model — how robotaxis coexist with existing taxis and public transit in dense cities.
Chapter 16 · Japanese Autonomy — Tier IV · Honda · Toyota Woven
Japan occupies a distinctive position. An aging population creates a sharp driver shortage, and the government actively supports AV.
- Tier IV (tier4.jp) — founded 2015 by Prof. Shinpei Kato (University of Tokyo). The home of Autoware, the open-source AV operating system. The core of Japan's AV standard.
- Toyota Woven by Toyota — launched as Woven Planet in 2018, consolidated into Woven by Toyota in 2024. Arene OS, Toyota Research Institute (TRI).
- Honda plus Cruise — partners since 2018. The Cruise Origin Japan launch was planned. Cooperation paused when Cruise shut down in 2024.
- Honda Sensing Elite — Level 3 certified for the Legend in Japan in 2021. Limited to 100 units sold.
- Sony Honda Mobility (Afeela) — joint venture announced 2022. 2026 launch target. Built on Qualcomm Snapdragon Ride.
- ZMP (zmp.co.jp) — founded 1999. Autonomous shuttles and logistics robots.
- Boldly (boldly.com.co.jp) — Japan SoftBank and Mobileye joint venture. Autonomous shuttle pilots.
- Nidec ASI · Mitsubishi Autonomous — urban autonomy pilots.
Japan Level 4 legalization — April 2023 Road Traffic Act amendment. Driverless operation permitted under specific (GeoFenced) conditions. Eiheiji district in Fukui prefecture became the first case.
Japan's strengths.
- Aging plus rural population decline — clear demand for autonomous shuttles.
- National standard — industry alignment around Autoware.
- OEM cooperation — Toyota, Honda, and Nissan move in parallel.
The weakness — no single dominant champion like Waymo or Tesla in the US. The field is fragmented.
Chapter 17 · The Four Layers of the AV AI Stack
AV software splits into four broad layers.
- Perception — sense the surroundings with cameras, LiDAR, and radar. CNN- and Transformer-based models.
- Prediction — anticipate the next moves of other vehicles and pedestrians. Time-series models.
- Planning — decide the path and behavior for the next 5-10 seconds. Behavior planning plus motion planning.
- Control — convert plans into steering, throttle, and brake commands.
A traditional stack (Waymo, early Cruise) keeps the four layers as separate modules. Each module can be tested and validated independently.
An end-to-end stack (Tesla FSD v12, Wayve) merges all four layers into one neural network.
- Pros — no information loss between modules. Optimized directly against training data.
- Cons — hard to explain decisions. Safety validation lacks an established standard.
The 2026 trend — modular stacks replace some modules with neural nets (Waymo), and end-to-end stacks add separate safety monitors (Tesla). The field is converging on hybrids.
Chapter 18 · The End of HD Maps?
Early AV development depended on HD maps (centimeter-accurate digital maps).
- HD map providers — TomTom, HERE, Mobileye REM (Road Experience Management), DeepMap (acquired by NVIDIA), Mappers.ai.
- Data — lane markings, traffic-light locations, stop lines, road curvature.
- Updates — periodic vehicle re-surveys, patches when changes (construction, new traffic lights) appear.
End-to-end models (Tesla, Wayve) claim to operate without HD maps. They infer lanes and signals from camera input in real time.
The 2026 reality.
- Waymo — HD maps plus neural networks. Priority on perception accuracy.
- Tesla — no HD maps. Reports of perception limits in edge cases (construction zones, rainy nights).
- Mobileye REM — Mobileye-equipped customer vehicles anonymously upload data to auto-generate and update HD maps. A cost-versus-accuracy balance.
The future of HD maps is moving from "required" to "supporting." Companies that want to lower the cost of entering a new city reduce map dependency.
Chapter 19 · Simulation — NVIDIA Cosmos · Wayve GAIA-2 · Omniverse
Real-world driving miles are not enough for training. Simulation fills the gap.
- NVIDIA Omniverse Drive Sim — physics-based simulation, synthesizing city environments.
- NVIDIA Cosmos (January 2025) — a generative video foundation model. Synthesizes "scenario videos."
- Wayve GAIA-2 (2024) — AV scenario video generation. A neural simulator.
- Waymo SimulationCity — Waymo's in-house simulation.
- Tesla Simulator — in-house.
- CARLA (carla.org) — open-source AV simulator. The academic standard.
- AirSim (Microsoft) — deprecated in 2024.
What simulation buys.
- Rare scenarios — child jaywalking, a truck dropping cargo. Hard to encounter in the real world. Simulation rehearses them thousands of times.
- Cost — real driving costs tens of dollars per mile. Simulation costs cents.
- Regression testing — automatic verification that new software does not reintroduce old bugs.
The limit — the sim-to-real gap. A model that performs well in simulation can behave differently on real roads. Reducing that gap (domain adaptation) is core research.
Chapter 20 · Sensor Industry — LiDAR, Cameras, Radar Supply Chains
AVs are a major demand source for the sensor industry.
- LiDAR — Luminar (Volvo), Innoviz (BMW), Hesai (Chinese OEMs and NIO), Velodyne (merged into Ouster), Aeva (FMCW), AEye (long range), Valeo Scala (Mercedes).
- Cameras — Sony IMX series leads AV market share. OmniVision, Onsemi, STMicroelectronics.
- Radar — Bosch, Continental, Aptiv (Motional parent), ZF, Veoneer (acquired by Magna).
- Ultrasonic — short-range for parking. Bosch and Valeo lead.
Changes in the LiDAR industry across 2024-2025.
- Ouster and Velodyne merger (2023).
- Quanergy bankruptcy (2023).
- Hesai US market restrictions — proposed September 2024 US action against Chinese AV components.
- Innoviz revenue miss — 2024 guidance cut, sharp stock drop.
LiDAR is consolidating. Where 30 or so companies existed five years ago, fewer than 10 meaningful volume suppliers remain in 2026.
Chapter 21 · V2X — Thirty Years of Vehicle-to-Infrastructure Communication
V2X (Vehicle-to-Everything) has been a promised technology since the 1990s.
- DSRC (5.9 GHz) — early US and Japan standard.
- C-V2X (cellular) — US standard after a 2020 FCC decision.
- V2I (vehicle-to-infrastructure) — traffic lights broadcast remaining time to cars.
- V2V (vehicle-to-vehicle) — cars share position and speed.
The 2026 reality.
- China — Wuhan, Shenzhen and others are building V2X city infrastructure. Traffic lights broadcast to vehicles.
- United States — V2X mandate failed to pass. Some OEMs (Audi) deploy V2X on their own.
- Europe — eCall (emergency call) mandate completed. V2X is voluntary.
- Korea — KT and SKT build V2X networks. Pilot deployment in AV pilot zones.
V2X potential — at unsignalized intersections, vehicles can communicate to avoid collisions. In fog or at night it complements cameras and LiDAR.
The limit — value emerges only when both vehicles and infrastructure adopt simultaneously. Without government mandate, it is hard, as 1980s ABS mandates illustrate.
Chapter 22 · Regulation — NHTSA · California DMV · EU AI Act · UNECE
United States.
- NHTSA (National Highway Traffic Safety Administration) — federal vehicle safety regulator. Collects crash data and orders recalls.
- NHTSA Standing General Order (since 2021) — requires AV crashes to be reported within 30 days. Monthly data release.
- California DMV — the strictest AV pilot regulator. Pilot permits, paid-operation permits, and driverless permits are tiered.
- California PUC — issues paid robotaxi operating permits (similar to taxis).
- State variation — Texas and Arizona are friendly. California and New York are strict.
European Union.
- UNECE WP.29 — international vehicle safety standards. R157 legalized Level 3 ALKS.
- EU AI Act (2024) — classifies AV as a "high-risk AI system."
- Euro NCAP — assesses crash-avoidance and ADAS features.
Korea.
- MOLIT AV pilot zones — Gangnam, Sejong, Busan, Jeju.
- AV safety standards — insurance and logging requirements for test vehicles.
Japan.
- MLIT Council on AV commercialization — Level 4 legalized in April 2023.
- Road Traffic Act amendment — driverless operation in specific zones.
The regulatory trend — countries are tightening crash reporting requirements, forcing AV companies to balance transparency with trade secrets.
Chapter 23 · AV Incidents and Mandatory Reporting
NHTSA's Standing General Order has required AV crash reporting since 2021.
- Public data site — NHTSA.gov SGO dataset.
- Reportable crashes — injury, vehicle towed, airbag deployment.
- Cumulative through 2024 — Tesla over 1,000, Waymo over 200, Cruise over 80 (pre-October 2023).
Major incidents.
- Cruise pedestrian drag (October 2023) — a pedestrian struck by another vehicle ended up under a Cruise vehicle. The Cruise vehicle stopped, then moved about 6 m further (attempting to pull over). California DMV revoked the operating permit.
- Tesla Autopilot fatalities — 2016 truck-side missed perception (Joshua Brown), repeated stopped-emergency-vehicle collisions. After NHTSA reviewed 956 incidents, OTA recall in April 2024.
- Waymo bicycle collision (January 2024 SF) — a cyclist obscured behind a truck appeared suddenly. Minor injuries.
- Pony.ai truck crash (China) — an injury crash during a Guangzhou truck pilot in 2024. Operations briefly paused.
The value of transparency — when crash data is public, citizens can compare directly. Waymo publishes per-mile crash rates. IIHS and RAND analyze independently. That turns "is Waymo safer than human drivers" into a data-driven debate.
Chapter 24 · Insurance and Liability — Tesla Insurance · Progressive AV · Level 4 Liability Shift
Level 2 (driver responsibility) and Level 4 (system responsibility) are different animals for insurers.
- Level 2 — Tesla FSD crashes remain the driver's responsibility. Personal insurance applies.
- Level 3 — Mercedes Drive Pilot crashes during activation are Mercedes's responsibility.
- Level 4 — Waymo and Zoox bear operator liability. Passengers do not.
How insurers responded.
- Tesla Insurance — Tesla sells insurance directly, with prices driven by real-time driving scores. Operates in Texas and California. Loss ratios have run higher than the general market, creating operational pressure.
- Progressive AV — Progressive pilot product for AVs.
- Allstate Drivewise — UBI (usage-based insurance) plus ADAS data.
Legal infrastructure for liability transfer is still being built in 2026. Fifty US states apply different rules. The EU is pushing for a single framework. Korea and Japan rely on government guidelines.
Chapter 25 · Ethical Algorithms — Is the Trolley Problem Fake?
AV ethics conversations start with the Trolley Problem. "Five lives versus one — who do you save?"
The industry consensus is that the Trolley Problem is not a real AV problem.
- Time — real crashes are decided in under a second. There is no time for moral comparison.
- System design — rules that discriminate by race, age, or sex are illegal.
- Reality — decisions reduce to "can the collision be avoided, and what is the safest way to stop?"
The real ethical issues are elsewhere.
- Crash data sharing — should one company's crash learnings be available to competitors?
- Algorithmic transparency — can decisions be explained after the fact? End-to-end models make this hard.
- Employment effect — about 3 million truck drivers and 1 million taxi drivers in the US.
- Accessibility — visually and hearing-impaired riders gain mobility. Yet sidewalk delivery robots can block wheelchair paths.
The MIT Moral Machine experiment surveyed 40 million responses across 200 countries to measure "society-level moral weights." Result — East Asia leans toward protecting the many, the West leans toward protecting the young. How a global OEM should set a standard is the question.
Chapter 26 · Five-Year Outlook — Scenarios for 2026 to 2031
Five plausible scenarios.
- Scenario A — Waymo dominance — Waymo expands to 50 US cities. Tesla FSD remains a Level 2 ADAS. Robotaxi becomes synonymous with Waymo.
- Scenario B — Tesla catches up — FSD v15 certified driverless. Robotaxi (Cybercab) ships at scale. One million robotaxis on the road.
- Scenario C — China goes global — Pony.ai, WeRide, and Apollo Go reach Southeast Asia, the Middle East, and Latin America. US and EU entry blocked.
- Scenario D — Trucks first — Autonomous trucks fully cover key interstate lanes. Robotaxis remain pilots.
- Scenario E — No unified standard — city-by-city and country-by-country systems coexist. A single global standard does not appear by 2031.
The most likely outcome is D plus partial A. Autonomous trucking expands quickly thanks to a clean ROI. Robotaxi growth is led by Waymo, with city-by-city speed limited by public acceptance.
Chapter 27 · Self-Check — Buying and Using an AV
Items consumers should verify before using or buying AV features.
- Confirm the level precisely — do not trust the marketing name "self-driving." Most are Level 2.
- Is there eye monitoring? — GM Super Cruise and Ford BlueCruise enforce it. Tesla FSD is weaker.
- What roads is it allowed on? — highway only, or surface streets included?
- What is the takeover response time? — Mercedes counts 10 seconds. Without hands on the wheel by then, the car comes to a safe stop.
- Who is liable in a crash? — Level 2 is the driver. Level 3 Mercedes is the system. Read the contract.
- Consent for data collection — AVs collect video, location, and driving patterns. Read the terms.
- OTA update policy — software changes after sale. New features added or capabilities restricted.
- Backup plan — service shops and towing in case of system failure.
When using a robotaxi.
- Operating hours — does it run at night or in bad weather?
- Service area — check the GeoFence. The suburbs may be excluded.
- Pickup and dropoff — only at safe locations.
- Emergency contact — in-car emergency button, call center.
- Accessibility — wheelchair and visually-impaired assistance.
Epilogue — Autonomous Driving Arrives Step by Step
A one-line summary of 2026 AV.
"The technology is possible. Public acceptance and safety validation set the pace."
Cruise showed how a single incident can destroy a company. Waymo showed how publishing cumulative miles and crash rates builds trust.
AV will not arrive all at once. It arrives city by city, lane by lane, feature by feature. Autonomous trucks between Dallas and Houston. Zoox shuttles in Las Vegas. Tier IV shuttles in Tokyo Odaiba. 42dot shuttles in Seoul Gangnam.
Two stances matter for consumers. First, do not be fooled by marketing names. "Full Self-Driving" can be Level 2. Second, confirm who carries the liability. Who pays after a crash is the most practical question.
AV will not deliver every promise. Yet even partial delivery reshapes human society. Truck and taxi job change, less city parking, restored mobility for the elderly. The transformation has already started in 2026 and will be in full force in the 2030s.
References
- Waymo Safety Report 2024 — waymo.com/safety
- NHTSA Standing General Order Crash Data — nhtsa.gov/laws-regulations/standing-general-order
- California DMV Autonomous Vehicle Permits — dmv.ca.gov/portal/vehicle-industry-services/autonomous-vehicles
- SAE J3016 Levels of Driving Automation — sae.org/standards/content/j3016_202104
- IIHS Partial Automation Ratings 2024 — iihs.org/topics/advanced-driver-assistance
- Tesla AI Day 2022, 2023 keynote materials
- Wayve GAIA-2 paper — wayve.ai/thinking
- NVIDIA Cosmos announcement Jan 2025 — nvidia.com/en-us/ai-data-science/cosmos
- Mobileye REM whitepaper — mobileye.com/technology
- Mercedes Drive Pilot Level 3 announcements — mercedes-benz.com
- GM Cruise final wind-down announcement Dec 2024 — gm.com
- Aurora Driver commercial launch Dec 2024 — aurora.tech
- Pony.ai S-1 SEC filing Nov 2024 — sec.gov
- WeRide F-1 SEC filing Oct 2024 — sec.gov
- Baidu Apollo Go quarterly report 2024 — baidu.com
- Tier IV Autoware Foundation — autoware.org
- Korea MOLIT AV pilot zone briefings — molit.go.kr
- Japan METI AV regulatory framework — meti.go.jp
- UNECE WP.29 R157 ALKS regulation — unece.org/transport/vehicle-regulations
- EU AI Act high-risk AI systems — artificialintelligenceact.eu
- MIT Moral Machine Experiment Nature 2018 — moralmachine.mit.edu
- RAND Driving to Safety report — rand.org
- Federal Automated Vehicles Policy 4.0 — transportation.gov