← Articles
Read in another language
Technology

Robotaxis in 2026: Where Self-Driving Actually Stands

Jayden

Analyzes global supply chains, industrial policy, and technology issues.

Published

Key points

  • Waymo has passed 500,000 paid driverless rides a week and has logged more than 220 million rider-only miles across five US metros — Phoenix, San Francisco, Los Angeles, Austin and Atlanta.
  • Every commercial robotaxi today is SAE Level 4: no human is needed inside a defined, geofenced Operational Design Domain. Level 5 — driving anywhere a human could — does not exist as a product.
  • Three separate safety datasets point the same way, but they rest on different mileage bases and different producers: a company hub (220.6M miles), a peer-reviewed journal paper (56.7M miles) and a reinsurer's claims analysis (25.3M miles).
  • The honest framing is operator-reported and independently reviewed, not independently generated — and the figures cover geofenced surface streets, excluding freeways.
  • Failure is already on the record: Cruise lost its California permits in 2023 and GM halted its funding after more than $10 billion, while Alphabet's Other Bets posted a $2.1 billion quarterly operating loss even as weekly rides passed half a million.

Sometime in early 2026, a milestone slipped by with surprisingly little fanfare: Waymo, the autonomous-driving unit of Alphabet, passed 500,000 paid rides a week in cars with no one in the driver's seat [source: CNBC, 2026]. Its vehicles have now logged more than 220 million rider-only miles across five U.S. metro areas — Phoenix, San Francisco, Los Angeles, Austin, and Atlanta — and the company has said its first international city will be London [source: Waymo, 2026][source: Waymo, 2025]. After two decades of demos and delays, the robotaxi is no longer a promise; in a handful of cities it is a service you can hail. But "no one in the driver's seat" and "self-driving everywhere" are very different claims, and the gap between them is where this story actually lives — in the safety data, the economics, and the rules that decide where a driverless car is allowed to go.

What this article covers

  • Why robotaxis suddenly feel everywhere in 2026
  • What "self-driving" does and does not mean today
  • The safety evidence — and how to read it without over-reading it
  • What happens when a driverless car gets it wrong
  • The business problem: lots of rides, no profits yet
  • The regulatory patchwork that gates expansion
  • What to watch next

Why robotaxis are suddenly everywhere

For years, autonomous driving was a technology you heard about but never saw. That has changed fastest in a few specific places. Waymo now operates fully driverless, paid ride-hailing across five U.S. metros and crossed roughly half a million rides a week in early 2026 [source: CNBC, 2026][source: Waymo, 2026]. It has announced London as its first city outside the United States, using Jaguar Land Rover vehicles and beginning with trained safety operators on board [source: Waymo, 2025]. Tesla, taking a very different technical route, launched a small robotaxi pilot in Austin in mid-2025 [source: CNBC, 2025].

Three things converged to make this moment feel sudden. The vehicles finally accumulated enough real-world mileage to publish safety data rather than promises. The rides became genuinely paid, genuinely driverless products rather than employee demos. And the operators began expanding city by city instead of testing endlessly in one. None of that means driverless cars are now general-purpose. It means a narrow version of the technology has quietly become a real business in a small number of well-mapped cities — which is a more modest and more interesting fact than "the self-driving car has arrived."

What "self-driving" actually means today

The single most useful thing to understand is that "self-driving" is not one capability but a ladder of them, and today's robotaxis sit on a specific rung.

The SAE ladder

Engineers classify automation using the SAE J3016 standard, which runs from Level 0 (no automation) to Level 5 (a vehicle that can drive anywhere a human could, in any condition). The driver-assistance features in ordinary cars today — lane-keeping, adaptive cruise — are Level 2: the human is still driving and legally responsible. Commercial robotaxis are Level 4: genuinely no human driver is needed, but only within a defined Operational Design Domain (ODD) — a specific geography, set of road types, speeds, and weather conditions the system is validated for. Level 5, a car with no geographic or condition limits, does not exist as a product. Every driverless service running in 2026 is a geofenced Level 4 system, which is why it launches one mapped city at a time rather than everywhere at once.

Two philosophies: sensors versus cameras

Within Level 4, the leading operators disagree about how to get there. Waymo's vehicles combine lidar, radar, and cameras with detailed high-definition maps and a remote-assistance team, an approach built for reliability inside a mapped domain [source: Waymo, 2026]. Tesla is pursuing a camera-only system without lidar, betting that a more general vision-based approach can eventually scale beyond pre-mapped areas; its Austin pilot began with a small fleet of Model Y cars and a company safety monitor seated in front [source: CNBC, 2025]. These are genuinely different bets about cost, scalability, and safety, and it is too early — and would be a claim beyond the evidence — to declare either the winner.

The safety evidence — what the data shows

The strongest case for robotaxis is a safety case, and the numbers are striking. Across its 220-plus million rider-only miles, Waymo reports large reductions versus human-driver benchmarks for the same areas: 94% fewer crashes involving a serious injury or worse, 93% fewer pedestrian-injury crashes, 84% fewer cyclist-injury crashes, and 96% fewer injury-causing crashes at intersections [source: Waymo, 2026]. These figures come from crash data reported to regulators under a federal standing order, not from a company press release alone.

Two independent-leaning checks point the same direction. A peer-reviewed study in the journal Traffic Injury Prevention, analyzing 56.7 million driverless miles, found 92% fewer pedestrian-injury crashes, 82% fewer cyclist-injury crashes, and 85% fewer serious-injury-or-worse crashes compared with human benchmarks [source: Traffic Injury Prevention, 2025]. Separately, the reinsurer Swiss Re compared 25.3 million Waymo miles against a baseline drawn from more than 500,000 insurance claims and found 88% fewer property-damage claims and 92% fewer bodily-injury claims than the human-driver population [source: Swiss Re, 2024]. When a peer-reviewed journal and an insurer's actuarial data move in the same direction as the operator's own numbers, the signal is worth taking seriously.

Reading the safety numbers carefully

It is precisely because these figures are impressive that they deserve careful reading. Three cautions matter, and none of them is a debunking.

The mileage problem

A few hundred million miles sounds enormous, but road fatalities are statistically rare — roughly one per hundred million miles driven by humans. In an influential 2016 analysis, RAND researchers calculated that autonomous vehicles would need to drive hundreds of millions to hundreds of billions of miles to demonstrate, with statistical confidence, that they are safer than humans specifically on deaths and serious injuries — far more than any fleet has driven, which is why they argued the industry cannot rely on test miles alone to prove safety [source: RAND, 2016]. Today's injury-crash reductions are real and measured, but they are early evidence, not a closed case on the rarest and most serious outcomes.

Who is being compared to whom

The comparison also is not apples to apples. Waymo's miles are concentrated on geofenced surface streets and largely exclude high-speed freeways, and the benchmarks are adjusted for the areas it serves [source: Waymo, 2026]. That is the right way to build a fair comparison, but it still means the numbers describe how the system performs inside its Operational Design Domain, not across all the conditions a human drives in. And most of the largest datasets are produced or funded by the operator; peer review and third-party actuarial data strengthen them but do not make them fully independent. The honest summary is "operator-reported, independently reviewed, and pointing strongly in one direction" — not "settled." That is a distinction between a measured result and a proven verdict, and it is worth preserving.

When it goes wrong

The other half of a balanced picture is what failure looks like, because it has already happened. In October 2023, a Cruise robotaxi in San Francisco struck a pedestrian who had first been hit by a human-driven car, then dragged her about 20 feet; within weeks the California DMV suspended Cruise's driverless permits, citing an unreasonable risk to public safety and faulting the company's account of the incident [source: CNBC, 2023]. It was a vivid demonstration that a permit can be revoked almost overnight.

Tesla's rollout has drawn its own scrutiny. After videos showed its Austin robotaxis making apparent traffic-rule errors, U.S. regulators at NHTSA opened a review of the system [source: CNBC, 2025]. By 2026, Tesla robotaxis had been involved in more than a dozen crashes in Austin — property damage rather than serious injuries in that tally, but enough to keep federal investigators engaged [source: CBS News, 2026]. These incidents do not erase the aggregate safety gains, but they are a reminder that the technology's worst moments are exactly the ones that shape public trust and regulatory patience.

The business problem: rides don't yet mean profits

A robotaxi that is safe and popular is still not, yet, a profitable one. Alphabet reports Waymo within its "Other Bets" segment, which posted a $2.1 billion operating loss in the first quarter of 2026 — a loss that widened from a year earlier even as ride volume climbed past half a million a week [source: CNBC, 2026]. Growth and profitability are not the same thing, and here they are still pointing in opposite directions.

The reason is capital intensity. Sensor-laden vehicles, high-definition mapping, remote-assistance staff, depots, and cleaning all cost money before a single fare scales. The clearest illustration of the stakes came from General Motors, which disclosed it had invested more than $10 billion in its Cruise unit before halting funding for robotaxi development at the end of 2024 and folding the technology into its driver-assistance work [source: CNBC, 2024]. The open commercial question for 2026 is not whether people will ride — they clearly will — but whether ride growth can outrun the cost of the fleet fast enough to reach profitability before investors' patience runs out.

The rules of the road: a regulatory patchwork

Technology is only half of what gates expansion; permission is the other half, and it is fragmented. In the United States there is no single national robotaxi license. Individual states and agencies grant and can withdraw permits case by case — as California's rapid suspension of Cruise showed [source: CNBC, 2023]. The federal regulator, NHTSA, largely works through defect investigations and crash-reporting requirements rather than approving autonomy before it hits the road [source: CNBC, 2025]. The result is a patchwork in which a service can be legal in one metro and prohibited in the next.

Other countries are writing their rules now, and that timing shapes the map as much as the engineering does. The United Kingdom is moving to permit driverless services around 2026, and Waymo's planned London launch is proceeding under that emerging framework [source: Waymo, 2025]. Whether robotaxis spread quickly or slowly over the next few years will depend not only on whether the cars can drive, but on how many jurisdictions decide to let them — and how fast they pull the permit when something goes wrong.

What to watch

The realistic picture in 2026 is neither the hype nor the backlash. Driverless ride-hailing is a genuine, paying service in a small set of well-mapped cities, its measured safety record so far is strongly favorable, and it is still a geofenced, unprofitable, tightly regulated technology rather than a car that drives itself anywhere. Watch three things. First, whether the safety advantage holds as fleets expand into messier conditions and longer mileage, and whether more of the evidence becomes genuinely independent rather than operator-reported. Second, whether unit economics improve enough — cheaper vehicles, leaner operations — to turn ride growth into profit rather than a widening loss. Third, how regulators respond to the next serious incident, because the pace of the whole industry is set as much in permit offices as in engineering labs. Self-driving has arrived, but only inside carefully drawn lines. The interesting question for the next few years is how far those lines can be pushed out — and how safely.

Charts

Waymo's own safety hub: crash reductions vs. human-driver benchmarks

Waymo's own safety hub: crash reductions vs. human-driver benchmarksInjury-causing intersection crashes 96% fewer crashes, Serious-injury-or-worse crashes 94% fewer crashes, Pedestrian-injury crashes 93% fewer crashes, Cyclist-injury crashes 84% fewer crashes, Airbag-deployment crashes 82% fewer crashes96% fewer crashesInjury-causing intersection crashes94% fewer crashesSerious-injury-or-worse crashes93% fewer crashesPedestrian-injury crashes84% fewer crashesCyclist-injury crashes82% fewer crashesAirbag-deployment crashes
Company-reported, drawn from data Waymo files under NHTSA's Standing General Order, across 220.6 million rider-only miles through March 2026. Benchmarks are human-driver crash rates matched to the same service areas — geofenced surface streets, excluding freeways.Waymo Safety Impact (2026) (opens in a new tab)

Peer-reviewed study: crash reductions over a smaller mileage base

Peer-reviewed study: crash reductions over a smaller mileage baseInjury-causing intersection crashes 96% fewer crashes, Pedestrian-injury crashes 92% fewer crashes, Serious-injury-or-worse crashes 85% fewer crashes, Cyclist-injury crashes 82% fewer crashes, Motorcyclist-injury crashes 82% fewer crashes96% fewer crashesInjury-causing intersection crashes92% fewer crashesPedestrian-injury crashes85% fewer crashesSerious-injury-or-worse crashes82% fewer crashesCyclist-injury crashes82% fewer crashesMotorcyclist-injury crashes
Kusano et al., Traffic Injury Prevention (2025), covering 56.7 million rider-only miles — the same direction as the company hub at roughly a quarter of the mileage. The authors are Waymo employees; the journal review is independent, the data collection is not.Kusano et al., Traffic Injury Prevention (2025) (opens in a new tab)

A reinsurer's view: insurance claims, not crash counts

A reinsurer's view: insurance claims, not crash countsBodily-injury claims vs. all drivers 92% fewer claims, Bodily-injury claims vs. newer ADAS vehicles 90% fewer claims, Property-damage claims vs. all drivers 88% fewer claims, Property-damage claims vs. newer ADAS vehicles 86% fewer claims92% fewer claimsBodily-injury claims vs. all drivers90% fewer claimsBodily-injury claims vs. newer ADAS vehicles88% fewer claimsProperty-damage claims vs. all drivers86% fewer claimsProperty-damage claims vs. newer ADAS vehicles
Swiss Re (2024) compared 25.3 million Waymo miles against more than 500,000 claims drawn from over 200 billion baseline miles. Claims are an actuarial measure — money paid out — not the same unit as the police-reportable crashes in the two charts above.Swiss Re study, via Waymo (2024) (opens in a new tab)

How much driving sits behind each safety claim

How much driving sits behind each safety claimWaymo Safety Impact hub (through March 2026) 220.6million rider-only miles, Peer-reviewed journal paper (2025) 56.7million rider-only miles, Swiss Re claims analysis (2024) 25.3million rider-only miles220.6million rider-only milesWaymo Safety Impact hub (through March 2026)56.7million rider-only milesPeer-reviewed journal paper (2025)25.3million rider-only milesSwiss Re claims analysis (2024)
Three different publications over three different mileage bases, so this chart carries no single source link: the totals come from Waymo's Safety Impact hub, from Kusano et al. in Traffic Injury Prevention, and from Swiss Re's claims study respectively. RAND's 2016 analysis argued that proving a fatality-rate advantage would take hundreds of millions to hundreds of billions of miles — which is why injury crashes, not deaths, are what these datasets can actually measure.

Timeline

  1. RAND researchers (Kalra and Paddock) calculate that autonomous vehicles would have to drive hundreds of millions to hundreds of billions of miles to show statistically that they are safer than humans on deaths and injuries — a bar test-driving alone cannot clear.

    RAND, Driving to Safety (opens in a new tab)
  2. The California DMV suspends Cruise's driverless permits after an October 2 crash in which a Cruise vehicle dragged a pedestrian roughly 20 feet, citing unreasonable risk and misrepresentation of the footage.

    CNBC (opens in a new tab)
  3. Swiss Re publishes its claims analysis of 25.3 million Waymo miles, reporting 88% fewer property-damage and 92% fewer bodily-injury claims than the overall driver population.

    Swiss Re study, via Waymo (opens in a new tab)
  4. General Motors halts funding of Cruise's robotaxi development after investing more than $10 billion, folding the technology into driver-assist and citing capital intensity and a crowded market.

    CNBC (opens in a new tab)
  5. A peer-reviewed paper in Traffic Injury Prevention reports crash reductions over 56.7 million rider-only miles, moving the safety case from a company page into a reviewed journal.

    Kusano et al., Traffic Injury Prevention (opens in a new tab)
  6. Tesla launches its Austin robotaxi pilot with roughly 10 to 20 Model Y vehicles, a camera-only stack and a safety monitor in the front passenger seat.

    CNBC (opens in a new tab)
  7. NHTSA opens a preliminary evaluation into Tesla's driver-automation systems after reported traffic-rule violations in Austin — an after-the-fact defect inquiry, not a pre-market approval process.

    CNBC (opens in a new tab)
  8. Waymo announces London as its first city outside the United States, using Jaguar Land Rover I-PACE vehicles with safety operators onboard at first, working with UK authorities ahead of rules expected around 2026.

    Waymo, "Hello, London" (opens in a new tab)
  9. Tesla robotaxis are reported to have been involved in more than a dozen crashes in the months after the Austin launch — property damage, with no serious injuries in that tally.

    CBS News (opens in a new tab)
  10. Alphabet's first-quarter results show Waymo past 500,000 fully autonomous paid rides a week, while the Other Bets segment that contains it posts a $2.1 billion operating loss, widened from about $1.2 billion a year earlier.

    CNBC (opens in a new tab)
  11. Waymo's Safety Impact hub reports 220.6 million rider-only miles across five metros, with 94% fewer serious-injury-or-worse crashes than area-matched human benchmarks.

    Waymo Safety Impact (opens in a new tab)

Analysis

Level 4 does not mean "anywhere" — it means "inside this line"

On the SAE J3016 scale, Level 2 is driver assistance: lane keeping and adaptive cruise with a human legally responsible. Level 4 removes the need for a human inside a defined Operational Design Domain — a mapped, geofenced service area with rules about weather, roads and hours. Level 5 would drive anywhere a person could, and no company sells it. Every driverless ride available today is Level 4, which is why the story is always told city by city rather than country by country.

Read the denominator before you read the percentage

The reduction figures are not measured against all human driving. They are measured against human crash rates in the same service areas, on geofenced surface streets that exclude freeways. That is an honest comparison for the roads the cars actually use, and a limit rather than a debunk: it does not tell you how the same system performs at highway speed or outside the mapped zone.

Operator-reported, independently reviewed — not independently generated

The largest safety datasets in this field are produced or funded by the operator, because the operator is the only party with the miles. Peer review in Traffic Injury Prevention and an actuarial analysis by a reinsurer both add scrutiny that a company blog post cannot supply. Neither one changes who collected the data. That layering is the accurate description of the evidence, and it is stronger than most technology claims ever get without being settled.

The mileage problem means deaths are still unproven either way

RAND's 2016 argument still holds: because human road deaths run at roughly one per hundred million miles driven, demonstrating a fatality-rate advantage with statistical confidence would take hundreds of millions to hundreds of billions of miles. Injury crashes are frequent enough to measure at today's scale; fatalities are not. Anyone claiming robotaxis are proven to save lives is running ahead of the arithmetic — in either direction.

Rides are growing faster than the business is

Half a million paid driverless rides a week is a genuine consumer milestone. It sits inside an Alphabet segment that lost $2.1 billion in a single quarter, widened from about $1.2 billion a year earlier, with no company-level robotaxi profitability disclosed. Cruise is the cautionary version of the same equation: more than $10 billion spent before GM stopped funding it. The open question for 2026 is not whether people will ride, but whether ride growth can outrun fleet cost.

A permit can disappear overnight

There is no single federal robotaxi framework in the United States. States license the service — the California DMV and CPUC issue permits and can suspend them case by case — while NHTSA works after the fact through defect investigations and Standing General Order crash reporting rather than pre-market approval. Cruise's 2023 suspension showed how fast a service can be switched off. In the UK, driverless services are planned to be allowed around 2026, which is the window Waymo's London plan is aimed at.

Comparison

Three safety datasets, three evidence tiers — the numbers agree, the provenance does not
DatasetMileage baseWho produced itEvidence tierHeadline figure
Waymo Safety Impact hub (2026)220.6 million rider-only milesWaymo, from its own NHTSA Standing General Order filingsOperator-reported94% fewer serious-injury-or-worse crashes
Traffic Injury Prevention (2025)56.7 million rider-only milesWaymo employees, reviewed by an academic journalOperator-generated, independently reviewed85% fewer serious-injury-or-worse crashes
Swiss Re claims study (2024)25.3 million Waymo milesA reinsurer, using its own claims recordsThird-party actuarial analysis of operator miles92% fewer bodily-injury claims
Two bets on how a car should see the road
WaymoTesla
SensingLidar, radar and cameras togetherCameras only ("Tesla Vision"), no lidar
MapsHigh-definition prior maps of the service areaCamera-driven approach without HD prior maps
Human in the loopRemote assistance; no one in the vehicle on rider-only tripsSafety monitor in the front passenger seat at the Austin launch
Deployment shapeFive US metros, plus London announced for 2026Austin pilot launched with roughly 10 to 20 Model Y vehicles
Public safety recordCompany hub, peer-reviewed paper and reinsurer studyNHTSA preliminary evaluation opened; more than a dozen crashes reported
What each stakeholder is actually measuring
ActorThe question they askThe number they watch
RiderWill it show up and get me there?Rides per week — now past 500,000
RegulatorIs this an unreasonable risk on my roads?Crash reports filed under the Standing General Order
InsurerWhat does this cost to underwrite?Claims frequency per mile, not crash counts
InvestorWhen does the segment stop burning cash?Other Bets operating loss — $2.1 billion in a quarter

Process

  1. Map and define the ODD

    Build high-definition maps of a specific service area and set the boundary — which roads, which weather, which hours. Everything after this is scoped to that line.

  2. Test with a safety operator onboard

    The pattern Waymo describes for London: run the route with a trained operator in the vehicle before anyone rides alone.

  3. Get a state or national permit

    In the US the licence comes from state bodies such as the California DMV and CPUC, not from a federal approval. In the UK, the plan depends on rules expected around 2026.

  4. Go rider-only and charge for it

    The step that turns a demonstration into a business: no one in the driver's seat, paying passengers in the back.

  5. Report every crash under the Standing General Order

    NHTSA's reporting duty is what makes the datasets in this article exist at all — the safety numbers are built from filings, not from voluntary disclosure.

  6. Expand — or lose the permit

    Cruise's 2023 suspension shows the branch that is not expansion: a single incident and a regulator's finding can end the service in a state overnight.

Sources

  1. Waymo — Waymo Safety Impact (accessed 2026-07).View source (opens in a new tab)
  2. Waymo — Hello, London: your Waymo ride is arriving (2025-10).View source (opens in a new tab)
  3. Kusano et al. — Comparison of Waymo Rider-Only Crash Data to Human Benchmarks at 56.7 Million Miles, Traffic Injury Prevention (2025).View source (opens in a new tab)
  4. Swiss Re / Waymo — New Swiss Re study: Waymo cuts liability claims (2024-12).View source (opens in a new tab)
  5. RAND Corporation (Kalra & Paddock) — Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability? (2016).View source (opens in a new tab)
  6. CNBC — Tesla robotaxi incidents caught on camera in Austin draw NHTSA concern (2025-06-23).View source (opens in a new tab)
  7. CBS News — Tesla robotaxis involved in more than a dozen crashes since Austin launch (2026).View source (opens in a new tab)
  8. CNBC — California DMV suspends Cruise's driverless permits (2023-10-24).View source (opens in a new tab)
  9. CNBC — GM halts funding of Cruise robotaxi development (2024-12-10).View source (opens in a new tab)
  10. CNBC — Alphabet (GOOGL) Q1 2026 earnings (2026-04-29).View source (opens in a new tab)

Tags

  • #robotaxi
  • #self-driving-cars
  • #autonomous-vehicles
  • #waymo
  • #av-safety
  • #tesla