Waymo fully driverless public service: The driver's seat goes empty

The event and the narrow claim

On October 8, 2020, Waymo chief executive John Krafcik announced that riders of the company's Waymo One service in metropolitan Phoenix could now hail a car with nobody in the driver's seat. In Waymo's announcement opening its fully driverless service to the public, the company said existing Waymo One members would get access first, that more people would be admitted through its app over the following weeks, and that in the near term every ride would be fully driverless. Riders could now bring friends and family and talk publicly about the experience, which the earlier driverless riders, bound by non-disclosure agreements, could not.

Contemporary reporting filled in the operating details. Bloomberg's account of the launch described a fleet of more than 300 minivans serving a suburban area of roughly 50 square miles, smaller than the territory of about 100 square miles used before the pandemic pause, and called Waymo the first company to offer paid trips in cars with empty front seats. The Associated Press report added that nobody would be watching remotely with the ability to take over and drive.

The narrow, documented claim is therefore specific: a commercial ride-hailing service, open to members of the public rather than to vetted testers under contract, in which no employee sat behind the wheel, inside a bounded patch of suburban Arizona. It was not the first driverless car on a public road, not the first robotaxi with paying passengers, and not a demonstration that self-driving works everywhere. It also matters to separate the announcement from availability. On October 8 the change applied to people already enrolled in Waymo One; general sign-up through the app followed in stages, and only later did Waymo add an editor's note saying anyone in the United States could download the app and ride in its service territory.

Fifteen years from a desert race to an empty seat

The engineering lineage runs back to a government prize. In 2004 no vehicle finished DARPA's desert Grand Challenge; in 2005 five did, and Stanford's Stanley won the $2 million purse, as IEEE Spectrum's history of the race recounts. The same article notes that Stanford team leader Sebastian Thrun went on to work on Google's first self-driving cars. Waymo's own timeline dates that project to 2009, when the team set itself the goal of driving ten uninterrupted 100-mile routes in Toyota Priuses.

Two later steps in that timeline frame the Phoenix launch. In 2015, Steve Mahan, a blind man who had directed the Santa Clara Valley Blind Center, rode on public streets in Austin in Firefly, a small pod with a top speed of 25 miles per hour and no human driver. In 2016 the project became Waymo, a separate company under Alphabet, and in 2017 it began an early rider program in metro Phoenix using hybrid Chrysler Pacifica minivans fitted with its fourth-generation hardware. When Waymo One launched in December 2018, it covered parts of Chandler, Tempe, Mesa and Gilbert around the clock, but Waymo-trained drivers rode along to supervise.

The removal of that supervisor was slow and incremental. Krafcik told reporters, as Bloomberg recorded, that it had taken two years to run three cars at once in fully driverless mode in Phoenix and another year to reach a hundred. A few hundred early riders had been carried without safety drivers since the summer of 2019, and Waymo's own announcement put the fully driverless share of 2020 rides at 5 to 10 percent. Before the coronavirus cut demand, the AP reported, the service gave 1,000 to 2,000 rides a week, most with a backup driver aboard. October 2020 was the point at which that small, private fraction became the whole public offering.

Why the empty seats appeared first in Arizona

Law as well as geography pointed to Phoenix. Arizona's governor Doug Ducey issued executive orders in 2015 and again in 2018 directed at automated-driving developers, signalling that the state wanted their testing. Stanford law scholar Bryant Walker Smith, in an analysis of the governor's legal authority, observed that Ducey had embraced the industry's vision and questioned how far the power to govern the sector by executive order actually reached, since the state's enumerated gubernatorial powers do not explicitly include it.

The same metro area had also produced the field's defining failure. On March 18, 2018, an Uber test vehicle driving itself in Tempe struck and killed a pedestrian who was pushing a bicycle across the road away from a crosswalk. The National Transportation Safety Board's investigation found that the automated system detected her 5.6 seconds before impact but never classified her correctly as a pedestrian or predicted her path, that the design had precluded emergency braking in favor of the human operator, and that the operator had been looking at her phone. The board faulted Uber's safety culture and its lack of measures against automation complacency, and cited insufficient oversight by the Arizona Department of Transportation. Ducey suspended Uber's testing in the state.

That history explains why the empty seat was more than a symbolic gesture. The Tempe crash showed that a person nominally responsible for a supervised prototype could fail precisely because the system usually worked. Waymo's wager was the reverse arrangement: accept full responsibility for the driving task within a defined area, and make the case for safety through engineering process and data rather than through a human fallback.

What driverless meant on October 8

Waymo describes its system, the Waymo Driver, as a Level 4 automated driving system, meaning it drives without human supervision but only within an operational design domain, the set of conditions and environments for which it has been approved. Waymo's safety methodologies paper, released three weeks after the launch, explains that each expansion of that domain requires a readiness determination, a structured review of evidence leading to a judgment that the system can operate there without unreasonable risk. In October 2020 the domain was roughly 50 square miles of suburban Phoenix.

Driverless did not mean unattended. Riders could reach a human support agent from a button in the car, a feature present since the 2018 launch. The company later described a fleet response function in which human agents answer questions the vehicle poses, for example which lane a line of cones is closing, and may suggest a path. Waymo's 2024 explanation of fleet response stresses that the vehicle weighs that advice against its own sensing and remains in control throughout, so that a stuck car may stop and wait for a roadside crew rather than be driven remotely. That account is consistent with what the AP was told in 2020: there was no remote driver.

The launch also carried the marks of its year. The AP described sanitizing supplies, a mask requirement checked by in-car cameras, and a promise that rides with and without backup drivers would cost the same. Waymo said it would bring trained operators back into some cars once barriers separating the front row from the rear cabin were installed, to add capacity and serve a larger area. Krafcik told reporters the company was worried it could not meet demand with the vehicles it had, a reminder that the binding constraint in 2020 was the size of the driverless fleet, not interest from riders.

Where the learning sits inside the Waymo Driver

For a history of artificial intelligence, the significance of the service lies in what was doing the driving. A self-driving stack of this era was modular: perception turns lidar, camera and radar returns into tracked objects; prediction forecasts what those objects will do; planning chooses a trajectory. Learned models had become central to the first two, and Waymo published some of the methods and data behind them.

In December 2019 the company released the Waymo Open Dataset, 1,150 scenes of 20 seconds each with synchronized, calibrated lidar and camera data and consistently identified 2D and 3D labels, which its authors measured, by a metric they proposed, as 15 times more diverse than the largest existing camera and lidar dataset. In May 2020 Waymo researchers described VectorNet, a hierarchical graph neural network that predicts the behavior of road users by reading map elements and trajectories as vectors rather than as rendered images, matching or beating a rendering-based baseline while using over 70 percent fewer parameters.

The most direct link to deep learning's history is ChauffeurNet, a 2018 imitation-learning study co-authored by Alex Krizhevsky, the lead author of the 2012 AlexNet paper. Its finding was cautionary: even 30 million examples of expert driving were not enough for a network that simply copied human drivers to handle complex situations, so the authors synthesized perturbed scenarios such as collisions and departures from the road and added losses that penalized those outcomes. These papers are research artifacts, not a specification of the system that carried riders in 2020, and Waymo's safety framework describes processes rather than an architecture. What they show is that by the time of the Phoenix launch, learned perception and prediction were embedded in a larger engineered system whose safety argument did not rest on any single network.

The evidence released with the launch

On October 30, 2020, Waymo published two papers in support of the launch and said it was the first time an autonomous-driving company had released a framework describing the safety of fully autonomous commercial operations. Alongside the methodology paper came Waymo's public road safety performance data, by Matthew Schwall and colleagues.

That paper covered more than 6.1 million miles of automated driving in metro Phoenix, consisting of 2019 driving with a trained operator behind the wheel plus 65,000 miles of fully driverless operation from 2019 through September 2020. It counted every collision and minor contact, and also every event in which an operator took over and a simulation of what would otherwise have happened predicted contact. The result was 47 contact events, 18 actual and 29 simulated, none expected to cause severe or life-threatening injury. Nearly all involved a rule violation or error by a human road user, including all eight of the most severe, defined by actual or expected airbag deployment.

The disclosure was unusual for an industry that had largely kept such data private, and it set a pattern Waymo would repeat. But it also exposed the imbalance at the heart of the launch. Almost all of the 6.1 million miles had a human ready to intervene. The driverless portion, the part that the public service now consisted of, was 65,000 miles, about one percent of the total.

A statistical problem no fleet could drive around

Four years earlier, RAND researchers Nidhi Kalra and Susan Paddock had shown why mileage alone could not settle the question. Their report Driving to Safety took the 2013 American rate of 1.09 deaths per 100 million miles and asked how far an automated fleet would have to travel to prove it was at least that safe. To show with 95 percent confidence that its fatality rate was no higher, a fleet would need 275 million miles without a fatal crash; 100 cars driving around the clock at 25 miles per hour would take about 12.5 years. Estimating the rate to within 20 percent would take about 8.8 billion miles, and proving a 20 percent improvement over human drivers more than 11 billion.

Against those numbers, 65,000 driverless miles said almost nothing about the rare, catastrophic outcomes that matter most. Waymo's answer, visible in its methodology paper, was to combine road data with a range of other safety methods whose outputs are reviewed together, in a set order, before engineering judgment is applied to each new operating condition. RAND itself had concluded that developers would need such alternative methods, and that regulation would have to be adaptive as evidence accumulated. The Phoenix launch was, in effect, the first time a company asked the public to accept that kind of argument in place of a human driver.

From one suburb to hundreds of thousands of rides a week

The service expanded in the stepwise fashion the readiness process implies. Waymo's timeline lists a public rider-only service starting in San Francisco in 2022 through a waitlist and testers, fully autonomous rides to Phoenix Sky Harbor airport the same year, a doubling of the Phoenix area to 180 square miles with Scottsdale in May 2023, open access in San Francisco and Los Angeles in 2024, and Austin and Atlanta in 2025.

Regulators moved alongside. On August 10, 2023, the California Public Utilities Commission let both Waymo and its rival Cruise charge fares for driverless rides across San Francisco at any hour, with one commissioner conceding that the data to judge automated vehicles against human drivers did not yet exist. Less than three months later, the California DMV suspended Cruise's driverless permits, finding its vehicles not safe for public operation and that the company had misrepresented safety information. Waymo's lead was not inevitable; its closest American rival lost its license to operate without a driver.

Waymo's own figures trace the scale that followed. In August 2024 co-chief executive Tekedra Mawakana said the company was giving more than 100,000 paid rides a week across Phoenix, San Francisco and Los Angeles, as TechCrunch reported, double its previous disclosure. In its 2025 year in review, Waymo said it had been serving more than a million fully autonomous rides a month since that spring. And in February 2026, announcing a $16 billion funding round at a $126 billion valuation, it reported more than 400,000 rides a week across six metropolitan areas, 15 million rides in 2025 and more than 20 million in total, with groundwork laid for over 20 more cities including Tokyo and London. These are company disclosures, not audited operating statistics.

The safety record and its unresolved questions

The mileage RAND said was needed has, in part, arrived. Waymo's safety impact data hub reports 271.3 million rider-only miles through June 2026, 92.1 million of them in Phoenix. Compared with human benchmarks built from police-reported crashes on the same kinds of roads, it reports about 95 percent fewer crashes with serious injury or worse, 82 percent fewer injury-causing crashes and 93 percent fewer pedestrian injury crashes. Crash data come from reports required under a federal standing order, and the hub publishes the underlying files for replication.

Peer-reviewed analyses by Waymo researchers underpin those figures. Kusano and colleagues' study at 7.1 million miles, covering data through October 2023, found 0.6 injury-reported crashes per million miles against a human benchmark of 2.80, an 80 percent reduction, while noting that the any-injury reduction in Phoenix alone was not statistically significant and describing the results as a directional indicator for use within a safety case. A follow-up at 56.7 million miles found the largest gains in vehicle-to-vehicle intersection crashes, a 96 percent reduction in injury-reported crashes, and no significant disbenefit in any of 11 crash types.

The caveats are real. The studies are written by the company whose product they assess. The comparisons cover surface streets and exclude freeways, and the headline categories group fatal crashes with serious injuries rather than isolating a fatality rate, the quantity RAND's arithmetic was about. And aggregate rates can hide specific failure modes. In late 2025, NPR reported that federal regulators had opened an investigation after a Waymo car drove around a stopped school bus in Atlanta, that the Austin school district had documented 19 similar incidents, and that Waymo would file a voluntary software recall; the company said no injuries had resulted. NPR also noted that independent analyses had supported Waymo's broader claim of fewer crashes than human drivers.

How to place the Phoenix launch in AI history

The October 2020 launch was not a conceptual breakthrough in machine learning. Its methods came from a long line of work that includes the desert-race robotics of Stanley, the deep convolutional networks that spread after 2012, and specialized learned models for perception and prediction. Its novelty was operational and institutional: by Bloomberg's account, for the first time a company let paying members of the public summon a vehicle driven by software that bore sole responsibility, moment to moment, for a safety-critical physical task among other road users.

That framing sets the right comparisons. Shakey, in the 1960s, planned and moved through a laboratory; the DARPA vehicles crossed an empty desert; the Phoenix cars drove through intersections and school zones for strangers who had paid a fare. The launch also prefigured patterns now familiar from other AI deployments: capability expanding domain by domain through internal readiness judgments, a public safety case assembled largely from the developer's own data, and oversight drawn from a patchwork of state executive orders, utility permits and federal crash reporting rather than a single framework.

Its long-run influence remains provisional. Hundreds of thousands of weekly rides in 2026 are a large number for a technology that began with three cars, but they still describe a service in a limited set of cities, with open questions about cost, about edge cases such as school buses, and about how evidence will be judged once fatality-level statistics become possible. What is already settled is narrower and still significant: on October 8, 2020, the empty driver's seat stopped being a test condition and became a product.

References

• Waymo (2020). Waymo is opening its fully driverless service to the general public in Phoenix.
• Boudway, Bloomberg News via Transport Topics (2020). Waymo Begins Fully Driverless Rides for All Arizona Customers.
• Bussewitz, Associated Press via Tech Xplore (2020). Waymo removing backup drivers from its autonomous vehicles.
• Marsh, IEEE Spectrum (2025). The Starting Line for Self-Driving Cars.
• Waymo (2026). Waymo's journey.
• Waymo (2018). Riding with Waymo One today.
• Smith, Stanford Center for Internet and Society (2018). A sad irony for Governor Ducey after Uber's fatal crash.
• National Transportation Safety Board (2019). Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian.
• Webb et al. (2020). Waymo's Safety Methodologies and Safety Readiness Determinations.
• Waymo (2024). Fleet response: Lending a helpful hand to Waymo's autonomously driven vehicles.
• Sun et al. (2019). Scalability in Perception for Autonomous Driving: Waymo Open Dataset.
• Gao et al. (2020). VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation.
• Bansal, Krizhevsky and Ogale (2018). ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst.
• Waymo (2020). Sharing our safety framework for fully autonomous operations.
• Schwall et al. (2020). Waymo Public Road Safety Performance Data.
• Kalra and Paddock, RAND Corporation (2016). Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability?.
• California Public Utilities Commission (2023). CPUC Approves Permits for Cruise and Waymo To Charge Fares for Passenger Service in San Francisco.
• California Department of Motor Vehicles (2023). DMV Statement on Cruise LLC Suspension.
• Korosec, TechCrunch (2024). Waymo is now giving 100,000 robotaxi rides a week.
• Waymo (2025). Delivering more for our riders in a year of incredible growth.
• Waymo (2026). Accelerating our global growth: Waymo raises $16 billion investment round.
• Waymo (2026). Waymo Safety Impact.
• Kusano et al. (2024). Comparison of Waymo Rider-Only Crash Data to Human Benchmarks at 7.1 Million Miles.
• Kusano et al. (2025). Comparison of Waymo Rider-Only Crash Rates by Crash Type to Human Benchmarks at 56.7 Million Miles.
• Brady, NPR (2025). Waymo will recall software after its self-driving cars passed stopped school buses.