More On Robotaxis

Cybercab

Liam Denning reports on Tesla's quarterly results flop in Someone Call Tesla a Robotaxi, or SpaceX, Quick. He notes that:

The curated list of investor questions that gets teed up ahead of these affairs had indicated some brewing discomfort about the big issue: Tesla’s lack of progress on its long-promised mass rollout of robotaxis. Musk duly tried to finesse this by pointing out that the “constraint” is safety, with Tesla trying to scale as fast as possible “while trying to ensure that we do not harm anyone.”

It is good that Tesla is "trying to ensure that we do not harm anyone" but they need to try much harder. In last February's Tesla's Not-A-Robotaxi Service I quoted Fred Lambert:

By the company’s own numbers, its “Robotaxi” fleet crashes nearly 4 times more often than a normal driver, and every single one of those miles had a safety monitor who could hit the kill switch. That is not a rounding error or an early-program hiccup. It is a fundamental performance gap.

But I went on to point out that Lambert was making the wrong comparison:

However badly, Tesla is trying to operate a taxi service. So it is misleading to compare the crash rate with "normal drivers". The correct comparison is with taxi drivers. The New York Times reported that:

In a city where almost everyone has a story about zigzagging through traffic in a hair-raising, white-knuckled cab ride, a new traffic safety study may come as a surprise: It finds that taxis are pretty safe.

So are livery cars, according to the study, which is based on state motor vehicle records of accidents and injuries across the city. It concludes that taxi and livery-cab drivers have crash rates one-third lower than drivers of other vehicles.

A law firm has a persuasive list of reasons why this is so. So Tesla's "robotaxi" is actually 6 times less safe than a taxi.

In any rational jurisdiction, six times worse than the competition even with a safety driver would get the regulators to force Tesla back to the drawing board.

The New York Times article was the best I could find at the time but it was from 2006. We now have much better and more recent data. Follow me below the fold for one of the reasons why Tesla's numbers are bad, very probably even worse than six, and even some criticims of Waymo's numbers.

On the earnings call:

Tesla even got its head of AI software, Ashok Elluswamy, to tout another safety statistic, saying that Tesla’s robotaxis have thus far driven 380,000 miles unsupervised with “zero notable incidents.” This, in his telling, validates the company’s cameras-only approach to sensors, without the need for lidar, radar and other safety technologies recommended by “so-called experts.” Besides Tesla getting to define what “notable” means, however, 380,000 miles is, in robotaxi terms, like just getting to the end of your driveway. It equates to less than 0.2% of the “rider-only,” or unsupervised, miles that Waymo LLC had already racked up through March.

Source

Fred Lambert reported on how one "so-called expert", Waymo co-CEO Dmitri Dolgov, responded in Waymo CEO explains why Tesla’s camera-only self-driving falls short:

Dolgov made the comments in a talk at Y Combinator’s Startup School, walking through the lessons Waymo has learned building its driver over close to two decades. He put the sensor question on the table plainly: “there’s been a long-standing debate about what kind of sensors do you actually need for autonomous driving.”

His answer draws the line that camera-only advocates tend to skip right past. “Humans of course can drive with just eyes, so there’s that proof of existence,” he said. “If the goal were to just approximately match human performance or to build an assist product, that’s a very reasonable way to go.”

Then the catch. If you’re targeting full autonomy and strongly superhuman performance, he said, “you find that weak sensing just leads to a safety curve that flattens out way too early.”

Ojai, Dllu, CC BY-SA 4.0

Dolgov explained Waymo's multi-modal approach:

Waymo uses three sensing types, and Dolgov spent real time on why. Cameras give you high resolution and color, but they’re passive, and they degrade in darkness and glare. Lidar directly measures the 3D structure of the world. Radar punches through fog, rain, and snow, and reads velocity directly with Doppler. Lidar and radar are active sensors, so they see just as well in pitch darkness or straight into a blinding sunset.

This isn’t redundancy for its own sake. “These different sensing modalities, they’re not backups to each other,” Dolgov said. Each one runs its own encoder, and the data fuses into a single view of the world that he says is “vastly superior to what you get with any one sensor.”

He showed cases where multiple modes were essential:

A dust storm in Phoenix, where the camera sees almost nothing while lidar cleanly picks out a pedestrian at the roadside. People hopping a concrete barrier onto a road at night. Kids chasing dogs across a pitch-black street with no headlights or streetlights on them. In each one, the camera feed is close to useless and the lidar view is clean.

Arguing, as Tesla did, that lidar is too expensive might have been true five years ago, which is when Tesla started dropping radar and ultrasonic sensors. But it was in effect a bet against the sensor version of Moore's Law. Lidar is now cheap enough that on most Chinese EVs it comes standard, which gives the lie to:

Elon Musk has called lidar “a fool’s errand” and said anyone relying on it is “doomed.”

Lambert sums it up:

So far the results line up with the exact ceiling Dolgov described. Tesla’s own robotaxi data shows a crash rate about three times worse than human drivers, even with a human safety monitor in the front seat. Musk has claimed FSD is safer than humans, release after release, and the data hasn’t backed him up yet, even after being heavily massaged by Tesla.

Waymo is running the numbers Dolgov says camera-only can’t reach. Its latest safety report, based on 220 million rider-only miles, shows 94% fewer serious-injury crashes than human drivers would cause over the same distance, roughly 17 times better. Waymo is now serving around half a million paid trips a week across 15 US cities and is scaling toward a 1-million-weekly-rides target, even as Musk insists Waymo “never had a chance” against Tesla.

Tesla recently confirmed it has accumulated 380,000 driverless miles over the past year since launching its robotaxi. Waymo’s driverless service does that in a day.

Because Tesla is led by a notorious bullshitter, and because unlike the rest of the autonomous vehicle industry it hides the details of the "notable incidents" it reports to NHTSA, any numbers it claims should be treated with extreme skepticism. EV Expert applies a bit too much in Tesla’s Robotaxi Crash Rate 9 Times Worse Than Human Drivers, NHTSA Data Reveals:

New data submitted to the National Highway Traffic Safety Administration reveals that Tesla’s Robotaxi fleet crashes approximately 9 times more frequently than human-driven vehicles, even with safety monitors on board. Between July and November 2025, Tesla’s self-driving Model Y vehicles operating in Austin, Texas, were involved in nine reported crashes across roughly 500,000 miles of operation—translating to one incident every 55,000 miles, compared to one crash every 200,000 miles for human drivers.
...
While human drivers average one police-reported crash every 500,000 miles, accounting for unreported incidents suggests a more realistic rate of one crash every 200,000 miles. Tesla’s 55,000-mile-per-crash rate represents a 3.6x to 9x higher incident frequency, depending on which baseline is used.

What makes this data particularly damning is that each Tesla Robotaxi had a safety monitor in the front passenger seat during this period, ready to intervene at any moment. Despite this human backup system, the vehicles still underperformed compared to solo human drivers.

Three of the 9 crashes Tesla reported were likely the kind that would go unreported, so the 3.6x number is probably low, but the headline number is definitely high.

But a lesser level of skepticism also needs to be applied to Waymo's claims, as David Zipper explained last January in We Still Don’t Know if Robotaxis Are Safer Than Human Drivers:

Although Waymo has shared more data than its rivals, some researchers say they aren’t ready to judge its safety performance. [Prof. Henry] Liu told me that he “doesn’t have enough information” to know whether the company’s vehicles are currently safer than humans, particularly because the company maintains control over the information it chooses to share. “We have seen many reports from autonomous vehicle developers, and it looks like the numbers are very good and promising,” said Liu, who has extensively studied self-driving technology. “But I haven’t seen any unbiased, transparent analysis on autonomous vehicle safety. We don’t have the raw data,” he said. He added that he would like to see “designed experiments where we can control the situation,” taking into account variables like weather and road conditions.

Liu pointed out two more problems with comparing robotaxis to the average human driver:

So far, Waymo vehicles have mostly driven on urban streets with a speed limit of 35 miles per hour or less — much slower than the highways and arterials that human drivers regularly traverse. “It’s not really fair to compare that with human driving,” Liu said, suggesting that the 127 million miles Waymo has logged doesn’t necessarily reflect conditions in which people typically drive: “It’s not apples to apples.”

He also hypothesized that robotaxi crashes could be more deadly than equivalent collisions involving human-driven vehicles because fewer robotaxi occupants will use seat belts, mistakenly assuming that they will be fine without one. (Previous studies have found that ridehail passengers are significantly less likely to wear seat belts than those inside personally driven vehicles.)

Again, I need to stress that the relevant comparison is not with the "average human driver" but with a professional taxi driver. Carnegie-Mellon's Prof. Phil Koopman adds yet another problem:

In California, around half of Waymo’s total mileage involves deadheading, with no passengers inside the vehicle. “I would expect robotaxis to have only a third as much harm to occupants as regular taxis,” said Koopman. “They spend half their time empty and half the time with one person instead of half the time with one person and half the time with two people.”

We will return to the issue of deadheading. In the meantime there is the issue of statistics:

Waymo’s current safety record is not anything to celebrate, according to Matthew Raifman, a senior researcher at the University of California, Berkeley’s Safe Transportation Research and Education Center. With around 127 million self-driven miles under Waymo’s belt, the company has been involved in two fatal crashes (although not directly responsible for either). A simple calculation dividing total miles by total fatalities — which Raifman called “the most accurate measure of safety that we have” — would give the company a higher per mile death rate than average American drivers, who drive around 80 million miles for every fatality.

But Waymo doesn’t evaluate its data that way. Instead, it lumps fatalities together with serious injuries, a category of crashes where data is more likely to be missing — and which, unlike road deaths, is not included in the federal government’s quarterly road safety updates.

How much data would be needed for a comparison of fatality rates that would be statistically valid?

Koopman estimated that clear comparisons between humans and robotaxis could be drawn once Waymo had around 2 billion miles under its belt.

“The jury is still out on Waymo fatality rates, and will be for quite some time,” Koopman wrote in a recent blog post.

Waymo’s Victor acknowledged that “there is not yet sufficient mileage to make statistical conclusions about fatal crashes alone,” adding that “as we accumulate more mileage, it will become possible to make statistically significant conclusions on other subsets of data, including fatal crashes as its own category.”

Even if robotaxis were statistically less likely to cause fatalities on a per-mile basis this would not necessarily mean fewer deaths:

The reason is that self-driving cars are poised to increase driving — and thus create more opportunities for collisions. As Koopman put it: “Civilized countries look at car deaths per capita instead of car deaths per mile.” Reducing deaths per capita presents a much higher hurdle for AVs to clear.

In San Francisco, many residents appear willing to pay the premium that Waymo charges over human-driven ridehail. That is unsurprising, given the comfort, privacy and general pleasantness of self-driven trips. Since robotaxis offer a more appealing experience for many riders, allowing them to do other tasks and play their favorite music en route, it is a reasonable bet that people will take more and longer car rides if and when they become widely available at a more affordable price, much like the process through which added highway lanes induce new driving.

Robotaxis will also spend many miles deadheading, posing at least some risk to other road users even as they drive around empty (and potentially causing gridlock that slows everyone else on the street).
...
All this extra motion brings greater risks. If AVs are 20% safer than human drivers and expand total driving by 30%, total crashes would increase by 4%. (The effect would be higher still if people shift to AVs from transit, which is exceptionally safe.)

If the goal is to reduce traffic fatalities, autonomous vehicles have a long way to go:

This is the standard that self-driven vehicles must meet to produce an overall reduction in crashes. Simply outperforming human-driven cars is not enough; they must do so to such an extent that they compensate for any added miles driven. But current AV safety analyses seldom take such network effects into account.

Despite Musk's claims, Tesla is not even close to meeting the goal, and likely cannot get anywhere close without multiple sensors.

Reagin von Lehe's The Safety Case for Robotaxis Doesn’t Really Hold Up in New York City, Report Finds states:

That’s because in a city like New York, robotaxis would actually be replacing yellow cab, Uber, and Lyft drivers. These professional for-hire drivers are already about 14 times safer than the city’s average driver, according to a July 20 report from the Sam Schwartz Transportation Research Program at Hunter College, published with advocacy group Open Plans.
...
The study found Waymo’s national rate of crashes causing serious injury or death was about 1.5 times higher per mile than New York City’s for-hire fleet, and more than three times higher in the Bay Area specifically.

But the cited report now states:

As of July 23, 2026, we have temporarily removed our safety analysis from this report, as well as relevant references to the analysis throughout the report (denoted in-text by a reference to this section). We were made aware of data discrepancies between City agency data for New York City for-hire vehicle crashes with serious injuries or deaths, as well as City agency data definitions related to such crashes. We are working to resolve these discrepancies and ensure our findings properly represent the data and the driving records of New York City for-hire vehicle drivers and Waymo vehicles. Other than this section, the remainder of our report remains unchanged.

While the exact comparison with the average human driver is unclear,a claim that robotaxis are safer than professional for-hire drivers is not credible:

While the autonomous vehicle company Waymo has recently boasted a lower crash rate than human drivers, citing a new analysis by the Insurance Institute for Highway Safety, researchers suggest New York City needs its own data before autonomous vehicles arrive at scale.

“There’s a mantra going around that autonomous vehicles are more than 90% safer than human driven cars,” said Sam Schwartz, transportation engineer, former NYC traffic commissioner, and founder of the program that bears his name, in a press release. “There is no solid evidence as of yet, to make such a claim compared to the for-hire vehicles already serving this city. We find that NYC taxi and Uber/Lyft drivers may in fact be safer per mile.”

Apart from safety, there are two other major issues in New York. First, deadheading:

On congestion, the study calculates a “vehicle miles traveled multiplier,” or total miles driven per mile actually carrying a passenger. Waymo’s is 1.71 in California, compared with 1.49 for New York’s existing high-volume for-hire fleet. That’s a roughly 15% efficiency gap driven by “deadheading,” or empty miles spent waiting for or driving to the next fare.

In cities like London and New York where you can hail a passing taxi with a simple wave of the hand, deadhead miles will be significantly less than for rides that must be hailed via an app. As anyone who has driven in these cities will appreciate, a second issue will also cause additional deadhead miles:

In California, Waymos can park during about 58% of that idle time. But the study says that’s unrealistic in a city with New York’s curb-space scarcity. Modeled at just 25% park time instead of 58%, the report estimates roughly 62,000 additional vehicle miles traveled per day.

After coexisting with them for years in the Bay Area and riding them, I am certainly prepared to believe that, in their service areas, Waymos are safer than most human drivers. I'm skeptical that this would be true without the caveat "in their service areas". I have definitely seen them behave in ways that, while safe, were unlike normal human behavior and which blocked other vehicles leading to frustration. And, I repeat, the safety standard for robotaxis is not being as safe as the average human driver. It is being significantly safer than the average professional driver. That is who robotaxis are intended to replace.

I consider it likely that Waymo will eventually reach that goal, though whether their externalities in cities like New York and London could be mitigated is less likely. I consider it extremely unlikely that Tessla's camera-only technology will ever be as safe as a professional driver.