Tesla’s Biggest Robotaxi Test Just Happened

本站添加: 2026-05-25

[00:00:00]Tesla's Robo Taxi is now nearly two times safer than humans in Austin. Today I'm going to do a deep dive in what we actually know about Tesla Robo Taxi safety right now. So an analyst named Raines, he ran the numbers and estimates that as of April, Tesla Robo Taxi in Austin is now about 1.8 times safer than human drivers in Austin. Top of that, for the first time, Tesla unredacted the crash reports. So we finally have a clear view of exactly what happened in each one.

[00:00:24]17 incidents since launch on June 22nd, 2025.

[00:00:28]Seven of those 17 were counted as Tesla at fault.

[00:00:31]When you actually read them, two were just curb scrapes. Two happened while a remote teleoperator had control. That leaves three that were the car itself.

[00:00:40]I'm going to walk you through what those three were.

[00:00:42]So here's the plan. First, the safety number and how Raines built it. Then we're going to talk about those 17 incidents in detail, including the seven where Tesla was at fault.

[00:00:51]Then why we still can't cleanly compare Tesla to Waymo. Then one crash report that honestly changed how I think about these cars. And finally, a New York Times piece written by doctors that frames this whole thing in a way I wasn't expecting.

[00:01:04]So let me start with the headline because it's the thing that everybody's going to argue about. Raines, who tracks this stuff carefully, he estimates that as of April, Tesla Robo Taxi in Austin was at least about 1.8 times safer than the average human driver.

[00:01:17]Now, I don't want you to just take the number. I want you to see how we got there because the method is the whole point. There's three pieces. The first piece is the human baseline. Using NHTSA data and counting both crashes that got reported and the ones that never do, he estimates a human driver has about one accident every 249,000 miles. That's the bar. That's what a Robo Taxi has to be.

[00:01:38]Okay, then there's the mileage problem, which has always been the hard part.

[00:01:42]Tesla doesn't hand you clean monthly miles driven numbers.

[00:01:45]So Raines pulled all the mileage Tesla's reported, plus the 7-day active fleet size from Robo Taxi Tracker, and found the relationship between them is actually pretty stable. Stable enough that he can estimate the monthly Austin robot taxi mileage with under 5% error.

[00:02:00]The last piece is how you handle the crashes. Tesla has months with zero crashes in Austin. February was zero.

[00:02:06]So, single month rate bounces all over the place. To smooth that, he used a 3-month rolling crash rate.

[00:02:13]And when you plot that against the human line, April is the month it crosses. For the first time, the rolling robot taxi crash rate came in safer than humans.

[00:02:21]So, here's the claim.

[00:02:22]Says that about 1.8 times safer in Austin as of last month. Built on real reported mileage, fleet data, and the actual federal crash reports. Hold on to that because the next thing we do is open up these crash reports and see if the number holds up.

[00:02:37]So, this month is different from every month before it, and the reason is the redactions.

[00:02:41]For the longest time, the Nitsa reports were redacted. You'd see that a Tesla robot taxi was involved in an incident, and then a black bar where the description should be. Tesla finally unredacted the narratives, retroactively, all the way back to the first one. So, now we can read them all.

[00:02:56]And I read them all. Robot taxi tracker also ran an analysis on who was at fault.

[00:03:01]Their breakdown came out to about 53% where somebody else was at fault. And about 41% where Tesla was. So, roughly nine versus seven out of 17.

[00:03:12]Let me take that somebody else pile first because it's bigger and it's almost comical how consistent it is.

[00:03:16]Over and over, it's a Tesla sitting still getting hit.

[00:03:20]An SUV creeps forward at a red light and rear-ends the stopped Tesla.

[00:03:24]A truck rolls forward at a stop sign and rear-ends a Tesla. A pedicab clips the side mirror as it rides past. A guy in a motor scooter pulls in behind a stopped car, steers into the back of it, and then hops the curb and rides off down the hot sidewalk.

[00:03:39]A city bus makes a right turn and the bike rack on the front sideswipes the Tesla's side.

[00:03:45]None of those are the car driving badly.

[00:03:46]That's the car parked, basically, and the rest of the road being the rest of the road.

[00:03:51]And I'm not going to soften this because honesty is the whole reason this video is worth your time.

[00:03:57]Two of those seven were under teleoperation, meaning a remote human operator had taken over control, and a human is the one who put it into something.

[00:04:06]In one, the safety monitor asked for help because the car wasn't proceeding.

[00:04:09]The teleoperator took over, sped it up, turned it left, and drove it up over the curb into a metal fence.

[00:04:17]In the other, a teleoperator took control during navigation and ran the car into a temporary construction barricade at about 9 mph, scraping the front fender.

[00:04:27]So, those two are remote human errors, not the car's own driving.

[00:04:32]Two more were curb scrapes, low-speed stuff. Reversing into a parking space and a rear tire catches the corner of a curb, that kind of thing. Cosmetic.

[00:04:41]So, that leaves three that are the real deal. The car driving itself making contact with something it should have cleared.

[00:04:47]The first one, the car's making this unprotected left into a parking lot entrance, and it drives into a metal chain strung across the entrance. Low speed, but it didn't see the chain.

[00:04:57]The second one, it's turning onto a residential street passing a house with a dump trailer in the driveway. The trailer's gooseneck hitch is sticking out into the road, and the car's side mirror clips it.

[00:05:10]And the third, it's going straight down a tight street with a heavy tow truck parked on one side and a car on the other, and the left side mirror catches the side of the tow truck's bed.

[00:05:19]So, there's a very clear pattern here.

[00:05:21]Chains, hitches, poles, mirrors, tight clearances.

[00:05:25]There's even one where it reversed into a wooden electrical pole in a blocked alley.

[00:05:30]The car's weak spot right now is low-speed precision around weird stationary objects. The slow, tight, fiddly stuff, not freeway speed or pedestrians in crosswalks.

[00:05:40]I want to be fair about one thing across all of these. Every single one had a safety monitor present. These weren't empty cars roaming the city.

[00:05:47]There's a person there. So, when you hear robo-taxi crash, picture a 5-mph mirror scrape with a monitor in the seat, because that's what the average one actually is.

[00:05:57]EV Wire ran the numbers across all of them, and the average speed was about 5.9 mph with no major injuries.

[00:06:06]Now, the question everybody jumps to is, "Okay, how does that stack up against Waymo?" And I have to be honest, right now, cleanly, it doesn't. We can't do an apples-to-apples. Here's why. On raw incident counts, it actually looks lopsided in Tesla's favor. Tobias Göbel pulled the federal numbers by month. In January, Waymo had 100 incidents. Tesla in Austin had five. February, Waymo 73, Tesla zero. March, Waymo 50, Tesla two.

[00:06:33]First half of April, Waymo two, Tesla zero. Another tracker, Mayhoff, has Waymo's all-time total over 1,800 incidents with Tesla at 17.

[00:06:42]But, you cannot just read those numbers off and declare a winner. Now, I won't let you walk away thinking that. Because Waymo's operating in something like 40 cities, actually 10 They're getting close to 10 cities, right? Atlanta, Austin, LA, Phoenix, San Francisco, and so forth. Tesla's robo-taxi count there is Austin only. Of course, the company in 10 cities has more total incidents.

[00:07:02]It's driving vastly more miles in vastly more places.

[00:07:06]So, to actually compare them, you need two things for each: in the same place over the same time, miles driven fully unsupervised, and the number of cars actually operating as robo-taxis. And for Tesla, both of those are pretty fuzzy. The exact unsupervised mileage isn't published cleanly. The active fleet size is estimated, not disclosed.

[00:07:24]That's literally why Raines had to back into the mileage from fleet size and a stable relationship. It's a smart workaround, but it's an estimate, and he says so.

[00:07:32]There's another wrinkle that matters.

[00:07:34]The federal data files Tesla's Austin operation as ADS, actual automated driving. But, it files Tesla's Bay Area operation as ADAS, driver assistance.

[00:07:43]Waymo's filed as ADAS everywhere. So, even the category labels aren't uniform across the two companies. If you're not careful about which bucket you're pulling from, you'll compare two different things and not even know it.

[00:07:54]So, my honest position is this. The incident counts are a real encouraging signal for Tesla. But, anyone telling you that they've got a clean settled Tesla versus Waymo safety ratio is selling you more certainty than the data actually supports.

[00:08:06]The right move is to watch the trend in Austin on a per-mile basis.

[00:08:10]That trend is the one Waymo needs to show crossing the finish line.

[00:08:15]I want to slow down on one specific incident because it flips the whole frame. So, in December, a robo-taxi is driving straight, and its rear right tire rolls over a patch of uneven pavement that punctures the tire.

[00:08:27]The car detects the pressure loss.

[00:08:30]On its own, it slows down to about 18 mph, pulls over to the side of the road, and puts itself into a safe stopped state. A minimal risk condition in the official language.

[00:08:40]On the way to the curb while pulling over, the tire caught the corner of the curb, which is why this one even shows up in the crash list at all.

[00:08:47]The thing that put this incident on the crash list is the car's safety like system handling a blown tire by itself.

[00:08:55]Without a human touching anything, it sensibly decided it had a problem and parked itself out of traffic. A human driver gets a flat at speed, and half of them panic. They'll brake wrong or drift. This thing just calmly pulled over.

[00:09:07]And that's what these cars really are.

[00:09:08]They're not just cameras and a driving model. They're rolling computers wired into every system in the vehicle.

[00:09:14]They got tire pressure, battery, motors, brakes, the whole car. It can file find a slow leak and act on it before a human would even notice the steering felt off.

[00:09:23]These are the most self-aware cars on the road. They can diagnose themselves and respond.

[00:09:29]And when you fold that back into the safety story, it's not just does the car avoid crashes, it's what does the car do when something physically goes wrong?

[00:09:37]And the answer in the one real world example we have is it handled it better than most people would.

[00:09:43]The last thread is one I did not see coming. The New York Times ran a guest essay from a neurosurgeon named Dr. Jonathan Slotkin.

[00:09:50]He's not coming at this as a tech guy or a Tesla fan. He's coming at it as a doctor who spends his career putting people back together after car crashes.

[00:09:58]So he put together an open letter signed by doctors and nurses asking the government to clear a faster path for self-driving cars.

[00:10:05]And his argument is blunt. He calls it a public health imperative to quickly expand the adoption of autonomous vehicles.

[00:10:11]The data he leans on is the Waymo record from a peer-reviewed study across more than 56 million driverless miles. 91% fewer serious injury or worse crashes than human drivers on the same roads.

[00:10:22]80% fewer crashes causing any injury at all.

[00:10:25]And they point to a projection that if just 30% of cars were fully automated, you might prevent 40% of crashes because the automated ones don't cause crashes and they react better when a human nearby screws up.

[00:10:38]That's a different conversation than the one we usually have online. We argue about whether the technology is cool or whether the stock goes up. A trauma surgeon is looking at the same technology and seeing fewer people on his operating table. Fewer spinal injuries, fewer funerals.

[00:10:51]When the people who deal with the records start writing letters asking for this faster, that tells you the safety case has crossed over from fan argument to medical argument.

[00:11:00]Let me give the other side a fair hearing because this is what the audience wants, right? They want signal, not a pepper rally.

[00:11:07]The first honest objection is the sample size. 17 incidents in a few months of Austin mileage is not a lot of data. One bad month, one serious crash, and that 1.8 times safer number moves. Rain knows this, which is exactly why he used a 3-month rolling average instead of a cherry-picking a good month.

[00:11:23]But it's early. And early numbers are very noisy. Treat it as a real signal, but not a final verdict yet.

[00:11:30]Second is safety monitors. Every one of these Austin incidents had a monitor present. So, we're measuring a system that still has a human backstop in the car. The real test, fully empty at scale, is still ahead. The good news is the at-fault incidents we do have are low speed and minor. The honest caveat is that safer than humans today includes that backstop.

[00:11:51]Third, the comparison problem I already laid out. We don't have clean, disclosed, unsupervised mileage and fleet counts from Tesla. Until we do, every Tesla versus Waymo ratio has an asterisk on it.

[00:12:02]I'd rather tell you that than hand you a fake clean number.

[00:12:05]And fourth, the teleoperation piece cuts both ways. Yes, the two of the at-fault incidents were a remote human, not the car. That's good for the driving model.

[00:12:13]But, it also means there's still a human in the loop steering these things sometimes. That human can put it into a fence. That's part of the current system, too. And it's fair to count it.

[00:12:22]Now, here's the reframe.

[00:12:24]Even with all of that, look at where the bar sits. The worst genuine mistakes this car made on its own were scraping a chain, clipping a couple of mirrors, and tapping a pole at walking speed. Nobody got hurt in those.

[00:12:37]Meanwhile, the human drivers around it were the ones doing the rear-ending.

[00:12:41]That's the actual picture once you read the reports instead of the headlines.

[00:12:45]So, what do you actually do with this if you own Tesla or you're thinking about it? The thing that just changed isn't the technology.

[00:12:51]The cars were already driving.

[00:12:53]What changed is that the evidence got legible. We went from a redacted count that anyone could spin to readable narratives, plus a mileage-based safety estimate, plus a peer-reviewed medical case. The story stopped being trust me and started being here, read it yourself.

[00:13:09]For the investment case, robotaxi safety is the gate everything else waits behind. The valuation people put on Tesla's robotaxi future assumes regulators eventually that the scale without a monitor in the seat.

[00:13:22]Regulators move on evidence and on political cover. Readable safety data is evidence. Doctors writing letters is the political cover.

[00:13:31]Both of those just showed up in the same month.

[00:13:34]I'm not telling you to do anything with your money on one analyst estimate. I'm telling you the kind of proof that unlocks the next phase is exactly the kind of proof that just landed.

[00:13:42]So, the smart move is to stop watching the hype and start watching the per mile trend and the regulatory response.

[00:13:49]That's where the real money gets answered.

[00:13:54]So, here's what I'd actually keep an eye on. The first thing is the monthly Nitsa crash data. It updates roughly monthly.

[00:14:01]Watch whether Tesla's Austin incidents stay low as the miles climb. Rising miles with flat incidents is a signal that matters.

[00:14:09]Then there's the per mile rate, not the raw count.

[00:14:12]Anyone waving a total crash number at you without miles behind it is doing it wrong. Ask for the denominator.

[00:14:18]Watch for the safety monitor coming out of the seat. The day Tesla runs Austin with no monitor at scale is the day safer than humans gets its final test.

[00:14:27]That's the milestone to circle and we're almost there. We're over 50% of all cars in Austin are now unsupervised and it's moving up.

[00:14:35]Keep an eye on whether the data stays unredacted and gets more complete.

[00:14:39]Disclose unsupervised mileage and disclose fleet size would let us finally do a clean Waymo comparison. Let's push for it.

[00:14:47]And then let's read the room on the medical and regulatory side. When more doctors, hospitals, and safety agencies start citing this data, that's the political path to scaling opening up.

[00:14:57]So, step back from all of it and here's where I land.

[00:15:00]For years, robo-taxi safety was an argument you can only win or lose with conviction because nobody had the receipts.

[00:15:08]This month, the receipts showed up.

[00:15:09]Readable crash reports. A mileage-based estimate that pulls the car ahead of humans in Austin. A blown tire the car handled by itself.

[00:15:18]And a surgeon telling the New York Times this is a public health issue.

[00:15:22]It's early. The numbers will move, but the conversation just changed from can it to how fast. And that's a very different place to be standing.

[00:15:30]If you want me to keep reading the supports as they update and tell you what's really in them, no spin, please subscribe. Stick around. I'll keep tracking it month by month. Thanks.