Hidden Autopilot Data Shows Patterns in Tesla Crashes

[00:00:00] Welcome to Tech News Briefing. It's Monday, August 5th. I'm Zoe Thomas for The Wall Street Journal. Tesla's last frontier of ultra-fast delivery in the US is reaching into the remote corners of America. We'll tell you why and what it could mean for the US Postal Service. And then...

[00:00:45] Tesla's semi-autonomous driving system relies mostly on cameras, which differs from the rest of the industry. A WSJ analysis has revealed previously unknowable patterns in crashes involving this system called Autopilot, and found Autopilot sometimes struggles to recognize obstacles or stay on the road.

[00:01:06] We'll bring you details of that investigation. But first... If you order toilet paper or baby wipes off Amazon and you live in New York City, San Francisco or other urban areas, your package could arrive in a day or less.

[00:01:25] But if you live in a remote rural area of the US, that same package could take several days to show up. Amazon wants to bridge that gap. The retail giant is expanding its one- to two-day delivery capabilities throughout the

[00:01:40] country, including small cities and towns in Arizona, Minnesota, Louisiana, and Texas. Here to tell us more is our reporter Sebastian Herrera. Sebastian, how does Amazon plan to reach these really remote places in the US?

[00:01:54] So a big part of how they're doing this is they're literally opening up delivery centers closer to remote areas of the country. And they're actually using a lot of local businesses to do the pickups through this program that's called Amazon Hub.

[00:02:10] So they're relying on people that live and work remotely to pick up the packages and deliver them themselves. A big part of this equation here is that Amazon uses its third-party seller services really to subsidize costly deliveries.

[00:02:26] And rural deliveries are some of the more costly deliveries out there. What has Amazon specifically said about this effort? They said that speed really matters to them. For Amazon, it's really like a way to stay ahead of the competition.

[00:02:41] Obviously Amazon is known for fast shipping throughout the US. And so by extending it out to more rural areas of America, it's just a way to continue that goal, especially over time as they've built out their logistics operation and they've tried to rely less on other carriers.

[00:03:00] Let's talk about other carriers for a second. I mean, why not turn to the US Postal Service or other package carriers like UPS or FedEx? Well, the US Postal Services have been raising rates. They're reducing pickup times in some areas. For years, they've had problems with profitability.

[00:03:15] And so the USPS has gone through some ups and downs. For Amazon, they don't want to have to rely on them if they don't need to. When it comes to other carriers in general, you know, Amazon famously years ago had this

[00:03:29] big disruption during the holiday season where UPS failed to deliver a lot of items to people for the holidays. And ever since that incident, Amazon sought to build its own logistics operation so that it didn't have to rely on other businesses.

[00:03:46] What has the Postal Service said about Amazon's plan? They said that they're not worried about competition and that they're focused on their customers and they're focused on how they can build a profitable system.

[00:03:58] But the reality is that Amazon does bring the USPS a lot of business historically, and they really rely on Amazon for a big part of the revenue. That was our reporter Sebastian Herrera. Coming up, longstanding concerns about Tesla's autopilot technology are showing up on the

[00:04:16] roads and putting the public at risk. We'll have that story after the break. Since 2021, Tesla has reported over 1,200 crashes related to its driver assistance system called Autopilot to federal regulators. In an investigation launched earlier this year, the National Highway Traffic Safety

[00:05:17] Administration or NHTSA tied at least 14 fatalities to the tech. But it's been hard for the public to understand the role Autopilot plays in crashes because NHTSA's reports are heavily redacted. Tesla considers information about Autopilot proprietary, and key details like the crash

[00:05:36] narrative and even the exact date are obscured in public reports. The Wall Street Journal did a comprehensive analysis of crash data, revealing previously unknowable patterns linked to Autopilot. Frank Matt, a senior video journalist at the WSJ, is here to tell us about that.

[00:05:54] Frank, as I mentioned, the public versions of NHTSA's reports on the Tesla crashes are heavily redacted. So tell us how you put this investigation together. What we did was we gathered what information we could from the data that NHTSA does make

[00:06:10] public and went directly to state crash databases, police departments, and did a bunch of records requests, gathered up all these disparate sources of data and reconstructed it. So if you can imagine a report about an individual crash in the abstract, like it would have

[00:06:25] a bunch of black boxes covering pieces of data we just took from different sources and tried to copy paste it into one complete record of what happened in an individual crash. So we did that hundreds of times and patterns start to emerge.

[00:06:39] And looking at these patterns, what kinds of crashes accounted for the most injuries and deaths? Head-on collisions with a stationary object in the middle of the road were one of the most common types of crashes and the one that accounted for the most injuries and deaths.

[00:06:56] So these are Teslas going down the road in autopilot and something's in the way. For whatever reason, the autopilot system doesn't see the object, doesn't react to it, and it just slams right into it. Has Tesla said anything about this?

[00:07:10] In general, Tesla has been very successful in defending themselves in court by putting all the responsibility on the driver. So they say that it's ultimately the driver's responsibility that they're the ones in control

[00:07:23] of the car, that even if the car is in autopilot, they need to be paying attention and ready to take over the car at any time. So if, as Tesla says, the driver is still responsible for driving the vehicle, why is

[00:07:36] it important to understand the role of the driver assistance technology in these crashes? Experts are telling us that this technology kind of creates distracted drivers. That if the technology works pretty well, that people will have an overconfidence in it and will start to check out.

[00:07:54] And this problem of overconfidence is especially acute in Tesla because of Tesla's marketing. The journal obtained video and partial data from a Tesla crash in May of 2021 that killed a man named Steven Hendrickson. He was driving his Tesla Model 3 to work.

[00:08:12] Talk us through what happened and what did the experts we shared this information with say it revealed? Steven Hendrickson is a good example of the type of crash that we discovered in the data

[00:08:23] is one of the most common types of crashes and the type of crash that's most commonly resulting in injuries and deaths in autopilot crashes. And that is a driver in autopilot, oftentimes at night driving straight into an obstacle that's clearly visible to the human eye.

[00:08:38] But for some reason autopilot does not detect. The data shows that he was repeatedly reminded to put his hands on the steering wheel, which shows us that he probably wasn't totally engaged. It says criticized that that means of driver monitoring as insufficient.

[00:08:56] What the data in that Steven Hendrickson crash didn't show us is any of the decisions the autopilot made, what it saw, what it didn't see. Did it see the overturned truck and decide it didn't know what it was and so didn't do

[00:09:10] anything or did it just fail to see it at all? None of that's clear in the data because that data is the inner workings of the autopilot software is closely guarded as a trade secret by Tesla. And that hasn't even come out by in court cases.

[00:09:24] No one's been able to get it. Let's talk a little bit about how Tesla's autopilot system works. It's different from other self-driving systems in other cars. Can you tell us how? Early on in the development of autopilot, Elon Musk made a very critical decision that was

[00:09:39] very different than any other automaker pursuing this type of technology of automated driver assistance systems. So they said they are going to have their system based primarily on cameras. So in addition to the cameras, Tesla has a few other sensors.

[00:09:56] The most important other one is that they also include radar in some models, not all models and that's meant as a backup basically for situations when vision is impaired. So if there's fog or rain or snow, other manufacturers have cameras and radar too, but other manufacturers

[00:10:14] also include LIDAR, which is laser ranging technology. So what that does is it shoots out tons of beams of lasers and creates a 3D mapping environment and it's extremely accurate at judging the distance of objects. It's also expensive.

[00:10:30] And so Tesla thought their vision system was going to get good enough that eventually they would not need this. And so why would you put in a very expensive system that you're eventually going to take out of the cars once the vision product can take it over?

[00:10:44] You told us that Tesla relies on cameras for its autopilot system. What are some of the challenges of relying on a camera based system for driver assistance or self-driving technology? So a camera based system relies on machine learning. So you have to train a computer to recognize obstacles.

[00:11:06] It's easiest to do that when the obstacles are common ones, like this is what a truck looks like when it's driving alongside you. Anyone who drives knows you encounter a lot of unusual situations on the road and the human

[00:11:19] brain doesn't need to cycle through images and see if it matches any reference that we have to decide what something is. We can just say, oh, there's something huge in the road. I should stop.

[00:11:32] The way that machine learning works is if you're trying to train an automated driving system to stop at a stop sign, you're going to show it a ton of images of stop signs so that when it encounters one, it knows what to do.

[00:11:44] If, say there's snowy weather and you would need to teach it what a stop sign looks like when it's half covered with snow or 10% covered with snow. There's just an infinite number of ways that things on the road can appear and a machine

[00:11:59] learning system and a camera based system has to figure them all out eventually. What has Tesla said about the safety of its autopilot system? Tesla says that autopilot is improving all the time. They say that they think it prevents more accidents than it causes.

[00:12:13] And they say that it's been driven safely for billions of miles. The DOJ is investigating Tesla over claims about its autopilot. What has the company said about the claims and about this case? Tesla denies that they've exaggerated autopilot's capabilities.

[00:12:29] They say that they've always made it clear that drivers need to be paying attention and that it is not a self-driving car, but a driver assist system. Tesla's critics, including NHTSA, argue that Tesla has misled customers in this regard.

[00:12:44] They point to the name autopilot as giving customers the wrong idea about the capabilities of the car. And they point to a lot of statements from Elon Musk that they say exaggerates the capabilities of autopilot. All right, that was Frank Matt, a WSJ senior video journalist.

[00:12:59] We'll link to the full video investigation in our show notes. And that's it for Tech News Briefing. Today's show was produced by Julie Chang with supervising producer Katherine Millsop. I'm Zoe Thomas for The Wall Street Journal. We'll be back this afternoon with TNB Tech Minute. Thanks for listening.