The Vergecast - Aurora CEO Chris Urmson on what's next for self-driving cars

Episode Date: April 23, 2019

Aurora CEO Chris Urmson stops by to discuss the future of self-driving cars with The Verge's Nilay Patel and Andrew Hawkins. They explore how the industry has evolved over the years, and how long it will take before self-driving cars are commonly used on the road. Learn more about your ad choices. Visit podcastchoices.com/adchoices

Support for the show comes from Retool. Too many companies run critical operations on duct taped spreadsheets, Slack workflows, and whatever else they could cobble together. Not because they want to, but because building internal tools means weeks of waiting on someone else's backlog. That's where Retool comes in. Build custom internal tools just by describing what you need. Prompts something like,

Build Me a Revenue Dashboard on our Salesforce data. And Retool actually builds it on your company's data, in your cloud with enterprise security built in. Go to retool.com slash Verchcast. We all need to retool how we build software. What's up, y'all. I'm Skylar Diggins, seven-time WMBA All-Star, Olympic gold medalist, and mom. And I'm Cassidy Hubbard, host and reporter for nearly 20 years, covering the biggest names and stories in sports and mom. And this is Am Mom, a community for athletes, game changers, and moms of all kinds.

dropping May 14th. Tap in with us. Hey everybody, it's the Lion on the Vergecast. On this week's interview episode, I sat down with Andy Hawkins, once again, our senior transportation reporter. We talked to Aurora CEO Chris Irmson. Aurora's a new self-driving car startup. They're really focused on the software package, on the cars himself.

But Chris is a really interesting guy. He used to be at Waymo, which is Google's self-driving division. He was at Carnegie Mellon, which is a hotbed of self-driving research talent. So we talked about the industry. why it's so crazy, why all the people are suing each other, why the talent keeps leaving from one company to another, how hard it is to develop a self-driving car. We talked about the first line of code they sat down and wrote at Aurora.

It's probably not what you think, and it's way more complicated than I expected. We also talked about whether it's going to happen. This is the first question I ask every self-driving car person that comes on a show. Is this going to happen? They all say yes, because they all run the companies. The timelines are different,

and the way they think it's going to happen, usually widely divergent. So super interesting conversation with Chris Irmson, CEO and co-founder of Aurora. All right, we're here with Chris Irmson. He's a co-founder and CEO of Aurora, which is a self-driving car company. Welcome, Chris.

Thanks for having me. I'm joined again by Andy Hawkins. Two weeks in a row, Andy. Hello, yes, I'm taking over the Vergecast. A person who understands what's going on in transportation, our senior transportation reporter, Andy Hawkins. So Chris, you're now working Aurora, which is like a self-driving car tech company.

Give us your history. Google before? Yeah, so I've been working in self-driving cars for roughly the last 15 years. So I was at Carnegie Mellon as a grad student that on the faculty and there I took part in the DARPA Grand Challenges, which were these kind of robot races the Defense Department put on. And then back in 2009, went to Google, helped found the self-driving car program there that's now called Waymo, led engineering and ultimately ran that program for a number of years. And then back in 2016, left and spent a few months trying to figure out what to do. And in that time, realized it was a chance to accelerate this technology coming to market

through pulling together a great team and then working in partnership with the rest of the transportation infrastructure and founded Aurora with Drew and Sterling. And, you know, we've been off to the races for the last two years. I want to get into self-driving cars. Sure. My interest is actually the technology. But I just want to take one step back. It seems like the market for people who know how to make cars drive themselves is

crazy. Right. And you named some of the biggest players. There was Carnegie Mellon. Half of that team went to Uber. Then he went to Google, which sued. Like, right, like, there's like, there's all these people. Yeah. Why, why is this industry as fluid, as contentious as it seems, at least from my perspective on the outside? I think there's a couple of things. One is it's a new industry. And I think anytime you have a new industry, there's a lot of dynamic, right? There's a, there's a lot of, you know, flowers blooming. and mixing and whatnot. I think the other is that the technology is really great for capturing people's imaginations.

Right? We've all, well, almost all of us, probably everybody listens to this, is ridden in a car. A long-running joke on the Verstackass is we make people pull over in their cars to tweet at us. Oh, excellent. So, yeah, so they don't even need to pull over anymore. I have vested interest in these cars driving themselves.

Yeah, totally. And I think that most of us had the experience, you know, there's times when we've really enjoyed driving a car, and then there's a lot of times that's kind of sucked. And so the fact that we get to work on something that can transform the way people get around, the way our goods get around, I think that captures people's imagination. And I think that the opportunity here, both to make the road safer, to make it more efficient, to get better access to people, and ultimately to build a heck of a business, I think is compelling. So you were at three places, obviously the university is a research-oriented place.

You're at Google, which is a very commercially oriented place, and we've got cars on the road now. Yeah. What is Aurora fit? So, Rory, you know, we're a startup. You know, our mission is to deliver the benefits of self-driving technology safely, quickly, and broadly. We think about it as building the driver. So we don't want to build the car. We have a tremendous amount of respect for those people. We don't want to go build Uber or Lyft or UPS.

We have an incredible amount of respect for how hard that is. We just want to build this capability to get the vehicle from one place to another than work with partners to have a go ahead and serve people. So you want to be one of the many, many, many suppliers of the automotive industry? Well, we don't think there'll be many, many people who can do this, right? We think actually building the driver is really hard. We imagine ultimately this, you know, what's now probably 100 companies working in the space will probably consolidate down to a handful. And we expect to be one of those. Why do you, is that a technology reason you think it's going to consolidate? Is it a capital reason? Is it? Yes, it's all of them, right? I think it's really hard. It's a very complicated

problem. It's one of the more complicated engineering problems where, you know, if not the most complicated engineering problem we're trying to solve right now. The number of people who have deep experience in it is relatively small. You know, you talked about the fluidity and kind of keep moving around. Well, the way you win is kind of pull those people together and have them work towards a common goal. And that's what we're trying to do at Aurora. And then ultimately, I think the technology, once we start to get it actually really deployed in serving people, right, you know, people talk about there being self-driving cars today, but there aren't. Yeah. They're not really out there yet. Once we start to really see commercial scale happening, there'll be

in, you know, there'll be evidence that the system works well and is serving people well. And so that'll start to kind of build a bit of a flywheel, I expect. So let me ask you, you actually said you were debating this over the weekend, but this is the question I ask every person who comes on our show and talks to what self-driving cars is, is this going to happen? Is this real? Yes. Yes, it's going to happen. On what time frame? I think you're going to see small scale deployments in the next five years. and then it's going to phase in over the next 30, 50 years. And do you think the model is the same as, you know,

I don't know, we had the CEO Ford Mobility in here. He's like, we're going to do level two, and then everyone's going to be comfortable, and we're going to skip level three because that'll make people crash their cars, and we're going to, do you think it's that stages of adaptive fruit controls get better, or are you taking the steering wheel out right away? So we're not taking the steering wheel out necessarily right away,

but no, I don't think it's a continuum, right? I think that this level two driver assistance capability, I think that's great, right? That's making people's lives a little bit better, but it's very different than self-driving capability, right? Driverless vehicles. And that's what Aurora were focused on, because we look at all the big players in the automotive space, and they know how to do driver assistance. And it's really a problem of, is the product compelling enough that the consumer wants to buy it for the price they can sell it? at. When we think about driverless vehicles or self-driving vehicles, you know, the level four and five

from the SAE standard, that's where we see a transformation. That's where you can sleep in the car. That's where the vehicle can be deployed as part of a transportation service and, you know, give you a ride, then give me a ride. And we can share the benefit of that together. And I think that's where the economics swing. And that's where we see the biggest social good for the, you know, for cities like New York and San Francisco. Will they be personally owned vehicles, do you think? Will they be part of a commercial fleet? What do you think is going to be sort of the breakdown there?

So I expect this will come to market first as part of a fleet. I think that just the economics again, that that's where it makes sense. If you could imagine, you know, let's say we're a few years down the road and the kit for the self-driving capability, I don't know, is $20,000 worth of stuff, let's imagine. that's really a premium to pay on top of the price of a car, whereas if that car is out and actually contribute in commerce, right, it's people are paying to get rides in it or people are paying that vehicle to go and deliver goods, then the economics suddenly make a ton of sense. And $20,000 sounds really cheap.

And so I'm pretty convinced that's the way it comes to market first. And the good news is I think that's the way that we get the social good, because if this really comes to market as an expensive feature that, you know, rich kids used to go send out to run errands, you know, I joke about fetching ice cream for them, right? That's just going to create more traffic and that's not really helping anyone, I don't think. It's like the zombie car phenomenon that I've heard spoken of. I hadn't heard of just empty cars. You ever heard that way? Just empty cars, in the streets, fetching ice cream and Uber eats deliveries for people. Jeez, yeah, let's, let's grim. Let's not do that. Well, so some of the recent

news around Aurora was that you guys just received a very significant investment from Amazon over half a billion dollars. What do you think Amazon's interest is in Aurora? Where are the sort of the overlap here? Presumably they told you. Well, I think you'd have to talk with the Amazon guys about what they're thinking. But let me tell you what we think about it. So we see Amazon as this, you know, this incredible company that has both world leading capability in cloud, which is an important technology for us as we're building what we're doing. And then on top of that, they're just an immense logistics company. And Aurora is in the business of building drivers. And so we think ultimately that'll be a way that we collaborate. When you say, you said earlier,

you can sleep in the car. Yeah. Is that your North Star? Is there just a picture like on the wall of your office of somebody's sleep in the front seat of a car? No, that's not really our North Star, right? Our North Star is the company's mission, which is to deliver the benefit to self-driving. technology safely, quickly, and broadly. And so we really, I think a lot of our people think about the 1.3 million people that die globally on the world's roads, right? And that is one of the leading causes of death, or if not the leading cause of death, for people under 29. That's incredible, right? This is the thing where we can bring technology to bear to address it. And then we think about the accessibility of transportation, both for folks that can't, you know, don't have the

privilege of driving like you or I do and allowing them to get around with the same kind of flexibility and freedom we have. And then we think ultimately about the ability of this technology to come to market and to serve across socioeconomic boundaries. That New York has an incredible public transportation system, right? I used to this morning to get here. You're shaking your head, but I think compared to a lot of places in America, it's pretty incredible. Yeah. Just as people who live here. Fair enough. It's a C plus. C plus. Fair enough. Well, I think that means that, you know, everywhere else in the country, it's an F-minus. So I think that this technology can actually come to market in a way where it's augmenting public

transportation, where we can actually provide the level of service, you know, a much better level of service at a much lower cost, ultimately. It's funny. I love hearing new startup founders talk about not doing the technology for the technology's sake. Yeah. When you talk about, okay, we're going to make transportation more inclusive, okay, we're going to make the car safer.

How does that actually factor into the technology decisions? Well, we look at it as some parts of it as table stakes that we're actually, the status quoted is broken. Right? Like, yeah, we've had two 737 max eight crashes. It turns out if you annualize, you know, you take the 40,000 people that die every year in America's roads, that's the equivalent of four of those planes crashing a week. So we have this fundamental problem in the transportation system that we have.

So we look at that and say, that's pretty profound. We think about the amount of value that can be created by addressing that, both society, but also when we do kind of in a very cold way look at the business. And we say, we don't have an opportunity to go, we don't think we add value to driver assistance work. Companies are doing that well already. Where we do think we add value is the insights we have and the understanding we have about self-driving capability. And that means we can go build a business in that space. And it's pretty much green field because the technology doesn't exist.

And if we can go and tap into the three trillion miles that are driven everywhere in the U.S. And make, you know, cents per mile, that's an incredible business. Yeah. So, yes, our company's mission is, honestly, you know, I've been doing this for 15 years. since before it was sexy. Right. We're bought into all the social good things. But if you look at the, you know, our ability to go raise capital, we have some very smart

folks who care about the economic outcome here and they're pretty, you know, making a pretty big bet with us. So you're not making like a LIDAR versus vision decision based on the bigger stuff. You're making your own kind of independent decisions and you're saying the other stuff is going to come along for the ride because this is just so much. better. Well, we make a LIDAR versus vision versus whatever. And actually, we don't think of a versus. We think it's a combination of them. Yeah, that was my next question. Right. Because that's what you have to do to solve the problem. You know, people, we won't be able to bring this technology to market if it isn't safer

than what we have today. And so this is one of those places where kind of the, the business interest and the social interest, I think, are really well-adined. Okay, so let's do the nerdy stuff. Sure. Let's do it. And you beat me to the punch. Oh, I'm sorry. No, that's great. That's how it's supposed to go.

Everyone else is supposed to be smarter than me. Why else are people listening to this anyways? They want the nerdy stuff. Yeah, they're here for it. So describe your technology. Like, there's a lot of competing ways to do this. Yeah. It seems, at least from this conversation so far,

you think the incremental improvement of driver assistance is not the way to go. That's handled. Not going to get there. You don't think it's going to get there. Yeah. How does your system work? So our system, you know, we talk about it as version 2.0 of the self-drive. in car technology. So we were able to bring together founders and team that have spent a lot of time

building this technology. So our VP software engineering used to lead SpaceX's software team. So these are the guys who land and launch rockets, right? That's pretty cool. One of the co-founders of Ross works at Aurora. And this is the big robot operating system, which a bunch of people use. So we've taken that experience and we've said, okay, what are the insights we have now about how to solve the problem. So one of them is that, yes, you know, deep learning is important. Machine learning generally is super important. But folks who kind of forget about all of the classic kind of algorithmic work that's been done, you know, things like common filtering and all the probabilistic state estimation work that's happened in the past, maybe you could get a machine learn system to

approximate that. But turns out if you actually understand the models, you can do it really well and you can do it much more efficiently. So at Aurora, the philosophy is, let's go bring the best of the deep learning world, let's go bring the best of kind of the engineered world together to solve the problem. And based on the experience we have, we can kind of figure out where to apply that.

One of the examples we talk about is how we deploy machine learning in our motion planning system. So machine learning is great because it can pick up on feature vectors and whatnot that our engineers maybe don't intuit. The challenge is, you don't know exactly what the bounds of its behavior going to be, particularly in situations

that you may not have encountered. And so the way we address that Aurora is we actually encode invariance, which are effectively rules that the system has to maintain as it's operating. So an easy example of this is we force the vehicle to always keep enough distance, enough time, really between it and the vehicle in front, such that if the front vehicle hits the brakes as hard as it can, we can sense that, react to it, and avoid a collision. And so that invariant is coded and wrapped around the machine learned system that is then within that set of constraints saying, okay, given all of the infinite ways we could have turned the steering wheel or press the brake pedal and a gas pedal,

what feels most natural? What kind of behaves in a way that is kind of a reasonable approximation of the way human would drive? So do you have to recognize the vehicle in front? So if you're in front of like a delivery truck versus a Ferrari, you know the Ferrari's going to stop faster? We don't, actually. We kind of make a pessimistic assumption about the, you know, about what might happen. An unmaintained Ferrari. Yeah.

Or actually, it's kind of the other, right, is that we assume that the package truck has super brakes because that's, you know, that's actually the more difficult thing, right? If it slows down more slowly, we've got more time to go break for it. I've been told from folks in this industry that prediction is a very hard, sort of hard nut to crack in all of this. Could you talk a little bit about how you guys are approaching that? Yeah. So when we think about the stack for self-driving vehicles, right, it starts with you've got the sensors, you know, observing the energy in the world. You then got the system that's estimating the state right now. And then you're trying to figure out how's the world going to evolve over the next five to ten seconds.

and then the motion planning system uses that to figure out how it moves through the world. And so we are using human driving experience to kind of build an implicit model of what the world will do over the next five to ten seconds. And then the combination, really, what's interesting about one of the things we're doing at Aurora is we're actually closing the loop around both motion planning and perception and, you know, also through that prediction. And what's kind of cool about this is the way that I've seen. seen this done in the past is machine learning was primarily applied to the perception system. And the perception team would go off and they'd kind of learn as good a model as they could. They would spit that out and it would have some kind of noise and issues with it. And then the motion planning team would say, okay, we're going to go implement a feature set

and we'll use that model. So they go churn away at it and they deal with whatever the noise characteristics are. Then the perception team says, hey, we just spent the last three months. We've just kicked out a new model. And so then it ends off to the motion planning team again, and they're like, oh, that's great. I'm glad it's so much better, but now we have to go and retune everything again.

And so they kind of get stuck, you know, being swung around by the perception folks. The way we're engineering the system at Aurora, and this is back to this kind of V2 idea, is to close the loop around both. So that as soon as the perception guys have done whatever magic they've done, you know, the motion planning scene can then basically overnight incorporate that information, that new model into their,

their way of doing business. Why is that not possible other places? Because it's not obvious how to do it. Right. So this is one of the things that drew back now, one of our co-founders brings. So he's, you know, he and I have known each other for the last 20 years. We went to grad school together at Carnegie Mellon. And he's one of the world's experts in machine learning applied to robotics, applied in particular to motion planning. And so this is kind of a big chunk of what his career has been, is how does he, how do you smartly apply machine learning to that? So that experience kind of shines through. And for me, Drew and I had never really worked together in the past.

And so, you know, I had a great bunch of people I worked with at Google, and we did some very cool stuff. And since working together with Drew, what's been great is there's a few things where, you know, he said, you know, we're going to solve this problem like this. You know, my intuition is like, no, Drew, you're crazy. That's a terrible idea. But, you know, that the – but, okay, let's give it a go. And then we do.

And it turns out he's right. Give me a natural example. Well, in particular, this kind of model for closing the loop around motion plan, it's kind of hard to get into the details of it. But it was not obvious to me that investing the energy we did up front with the right answer. My instinct is, let's go out, let's get something on the road, let's try it and iterate. And his was, no, let's actually take a moment and think here. And architect a system.

So over time, as we get more data, it will automatically improve. And, you know, I'm like, okay, that's a great idea. But how do we do it? And, you know, he came up with an approach with the team. And that's rolled out. And that's exciting to see. I'll take a quick break for an ad. We'll be right back.

Support for this show comes from Shopify. Every thriving, successful business has to start somewhere. A good place to start is a relatively simple question. What if, given the right tools, I've really put my all into this. One tool that can help grow your sprouting business to new heights is Shopify. millions of businesses around the world rely on Shopify for e-commerce. They offer a host of helpful tools you can take advantage of, from payment processing to analytics to website design.

Their design studio includes hundreds of templates to help you create the exact website you've been envisioning for your business. If you're wondering, what if I need help? Then no worries, because you're never left to fend for yourself. Shopify's award-winning customer support is available 24-7. It's time to turn those what-ifs into a thriving business with Shopify today. Sign up for your $1 per month trial today at Shopify.com slash vergecast. Go to Shopify.com slash vergecast.

That's Shopify.com slash vergecast. Support for the show comes from LinkedIn. If you're a small business owner, you know that every hire counts, but time and resources are limited. finding, connecting with, and screening the right candidates takes up valuable time you could be giving to your customers. That's where LinkedIn Hiring Pro comes in. It's built to be your hiring partner, helping you find the right candidates faster. That way you can hire with confidence without turning it into another full-time job. Hiring Pro streamlines the entire process from drafting your job to shortlisting candidates

and conducting AI-powered interviews for initial screenings. It's updated conversational interface lets you describe what you need in plain language. Nearly 60% of hirers find a candidate to interview within a week. With Hiring Pro, you spend less time searching and more time connecting with the right talent. And instead of getting buried in resumes, you get a focus shortlist that actually moves your hiring forward. Join the 2.7 million small businesses using LinkedIn to hire. Get started by posting your job for free at LinkedIn. dot com slash track terms and conditions apply okay we're back with aurora CEO Chris

Irmson let me ask you really this is going to be a dumb question I have my doubts well it's a good show that I put on okay you start a company yep you're your self-driving car engineer you've been doing it for 15 years yeah call your friend hey we've never worked together it's great being in school with you let's start this company together you call some another friend you're on a room you're in the conference room yep what is the first line of code that you write? What was the first line of code we wrote? You know what it was, actually? So I think the first place we started was building the, this sounds really boring, but it was the representation for the position of the vehicle. So a vehicle once moving through

space, it's got X, Y, Z, Roll Pitch, Yaw, and some speeds. And this is another one of these places where it's actually kind of one of these geek out math things. So, So, you know, historically you would use X, Y, Z, and, you know, a rotation matrix, but rotation matrix is, you know, it's got some problems where you kind of have these singularities and the representation and, you know, things don't behave well. So you say, okay, we're going to go use Quaternians instead because that's a much better, you know, representation. And that again, has some problems.

Can you say that? Pretorians? Sorry, quaternians. It's a four-dimensional vector that you use to represent rotation. This is what we're here for. Yeah, this is my heart explains. Sorry, okay. You want me to say it more slowly?

Yes, please. Quaternian. Quaterian. I love it. All right. And so, you know, that's kind of the hotness for a while. And it turns out that this is even cooler stuff called Lee Groups. And one of our engineers, Ethan, it turns out if you go to Ethanede.com, it is basically

the world's repository of information about Lee Groups and Lee Mathematics. And it turns out this is one of these very, really interesting ways to represent transformations through space that have none of the problems of rotation matrices or quaternions. And yet, you know, I've been doing this for a very long time and had never done it before, never used it. And so, you know, as we got started, this was like, this is the right way to represent. This is the mathematically beautiful, correct way to do it. And so that's what we started with. And from there, we built out the local state estimation. So integrating the inertial system and the GPS to figure out, the wheel encoders, figure out how the vehicle's moving.

And then the global localization. So this is where is the vehicle in the world to a really high degree of accuracy or high degree of precision, 10 centimeters, you know, tenth of a degree kind of resolution. So you build that stuff. It sounds like that's a big leap of intuition and logic to get to where you are. That was pretty cool. Yeah. And then you're like, well, we need a map. We're just going to call here maps.

and we're just going to buy a map and stick it on the map. Yeah, we thought about that. And in fact, when we founded the company, you know, Drew and I said, we've built maps in the past. We've done a lot of this. I mean, you're at Google. I mean, it's like you have access. You just think of the phone, right? Well, we, you know, even in Google, we built custom maps for the self-dramed car project.

I bet the maps team was like, dudes, come on. There was a little bit of that every once in a while. But, you know, in our defense, you know, the maps guys had to produce. maps that were being used by a billion people. And, you know, we had like six self-driving cars. Yeah. And so, you know, if you're the maps team, what are you going to build? You're going to fix the six self-driving car problem?

Are you going to fix the problem that's irking a billion people? So anyway, different products. And so we started the company and, you know, Drew and I were like, and I think Sterling as well, we are not going to build maps. There are a bunch of companies out there doing this. We will use their maps. And then as we dug into it, we came to the conclusion that that was the wrong answer. One is that the maps that we were able to get, they were really built by taking maps that were intended for people and kind of plus-plusing them, right?

Just make them, you know, higher resolution. And what we realize is that the maps that you need for a self-driving car are different, right? And you have different properties that you're looking for. You do need very high precision in these maps. And what you want is you want to be able to update them frequently because it turns out if my navigation map is a little bit out of date. You know, I kind of roll with it. If the map of the self-driving car is out of date, well, we have to be able to roll with it. But, you know, what we really want is it to be not out of date, be able to update it very quickly.

And so we came to the realization that the other map systems that were out there didn't meet the requirements for self-driving vehicles. So we said, okay, well, you know, we'll actually take this on and start to build that as well. And now we think that's actually one of the strategic advantages we'll have is a really interesting representation of the world for these vehicles. I wanted to ask you about safety. You brought up the fact that 40,000 people dying car crashes every year. And it seems like on some level, as a society, we've sort of internalized that, right? And we've sort of accepted that as sort of the price of driving a car, having access. to cars and sort of the freedom, quote unquote, that comes with driving a car. And I'm just wondering

how you sell safety to a public, especially when, you know, we have a very, we have a high tolerance for people killing other people on the roads, but we probably, I would assume, and it's only happened once, but we have a very low tolerance for robots killing people on the road. So how do what's, how do we sort of bring that sort of into parity, do you think? Yeah. So I think the, the first is that we have to help educate the folks who kind of figure out what the rules should be and help them understand the risks so that they can, you know, and I'm thinking about folks like NHTSA or the state DOTs, where their job is to kind of govern our roadways and kind of manage the risks there. And so one of the things that I've spent a fair bit of time over the last several years doing is

engaging with those folks, helping them understand what we perceive as the risks and what we perceive us the opportunities and helping prepare them for this technology so they could have an independent opinion and do the right thing as this technology starts to roll out. I think we have to help educate the public about the opportunity here. I think one of things, though, that's most compelling to me is that the way this technology is coming to come to market, the kind of the safety will come with it. So most people, when they go to buy a car, they, you know, they, you know, they, they They don't really understand the risks of operating a vehicle. They don't really rationally assess that.

And they generally don't pay for extra safety features. Because if you think about it, there's a bunch of incentives against this. So the first is that when you go to the lot, the cars that are there that you can buy are selected by the dealer. And they look at what are the different features they can put in that car and where can they effectively make the most money. And so they have a choice. They can put a $3,000 stereo system.

in the car they ordered, and that's got an 80% margin, or they put a $3,000 driver assistance system in the car, and that's got a 20% margin. And so, like, well, let's get more of the stereos, because, you know, we make more money off of that. So there's less options there. And then as the consumer, you know, you get to there and you have a choice. Well, I could buy this thing that's got this nice stereo that I use every day and that I enjoy, or I can have this thing that, you know, well, I'm a better than average driver as 80% Americans, us believe. And so this thing that I will only use on a really crappy day, which am I going to pick? Well, I'm going to pick the one that's on the lot and that has the feature that I enjoy.

And so that's kind of a long way to get to. There's a reason why these advanced safety features, driver assistance features aren't more prevalent in vehicles today. It's because the whole system is kind of stacked against them getting there. Whereas the driverless technology that we're working on the level four capability, people will use it because it is more convenient. because it's a better way for them to kind of get from one place to another. They can use that time however they want. And along with that, they'll get a system that is never drunk, that doesn't get distracted, that, you know, isn't tired,

and that can see 360 degrees around it. And so the safety will come along with those kind of the convenience benefits of the technology. Can the industry absorb another incident like the Uber incident in Arizona, do you think? Because that was clearly such a setback. And it seems like it's had an effect on the public's attitude towards self-driving cars, even though it's like it kind of boggles my mind that they still do like polls and surveys about self-driving cars when 99.9% of people have had no experience with the self-driving car at all. So you're asking people questions a lot of technology that they've never encountered in the world.

So yeah. Do you think, I mean, because it seems likely that there's going to be another incident as more cars come on the road. it starts to become, you know, sort of a more sort of hybrid system of, you know, human-powered cars and robot-powered cars. Yeah. So to answer your question, yes, I do believe we can weather that storm. I think we have to be working diligently to test safely and responsibly on the roads. I think we need to be building technology where we are holding ourselves to a high safety standard. But we also have to remember that, you know, the perfect really is the enemy of the good here.

and that if we could cut traffic deaths in half with this technology, that would be as big an impact as seatbelts. But that is still 20,000 people dying, you know, on America's roads. And what I think is really compelling here is that this is a technology where the driver will continue to get better and better over time. As we deploy the technology, it'll be better than people to begin with. As we learn more from what's out there, we can then. push improvements to that driver that span the fleet and all of the vehicles will get better.

If I have a bad event on the road, maybe as I'm out on the road after that, I think about that and I behave a little differently, but you don't learn from it and you don't learn from it, whereas the whole fleet will learn from that bad experience and get better. So I actually think about this all the time. Yeah. Because I have very nerdy thoughts. I'm sure you do too. It seems like the big increase is when there's widespread vehicle-to-vehicle communication,

when all the cars of the road are talking to each other. But right now, like, there's a jerk in a Mustang, and, like, your system has to just, like, have code that's, like, going to... I mean, I'm just thinking, literally thinking of you myself as a teenager, because I was definitely a jerk in a Mustang. Like, is that the ramp? Literally in a Mustang? Literally in a Mustang.

Excellent. One time I crashed my Mustang into a bank. This is a true story. The bank didn't have a... Intentionally? or like it really depends in where you think

attention ends and statute of limitations I think it's probably good here or I guess under steer begins is that the ramp is that the inflection point when there's like

widespread V to V or is it your computer is just going to I think our computer has to deal with it right our software has to deal with it and I think again one thing that's

compelling to me is that even one self-driving car on the road is a safety benefit because that that will be a, you know, a better driver. Like a known good.

Right. And that means it's going to create less bad events for other people around it. And as we, as we integrate more of these vehicles under the road, it'll get, you know, better and better and better. I do think ultimately, you know, V-to-V is an interesting technology, but it's not one that I think we can kind of hang our hat on and wait for. So I think about my children walk into school in the morning. They don't have transponders on them. I think about, you know. Well, they have phones, right?

Well, the younger one does not. Right? Olden one does, but, you know, I don't think you can count on him having it. Right. Yeah. Right. And then, you know, if you think about your old Mustang, it's probably still actually out there. It doesn't have a transponder on it.

The kid I sold this, you totaled. But yeah. Okay. So, a bad example, but. Yeah. Sounds like that car had a hard life. That car was not treated well by any stretch of the imagination. But that's interesting to me.

Like you're going to do all this local processing in the car effectively, and then you're going to have a cloud infrastructure that improves your fleet. But that's not going to go talk to Uber's fleet. That's not going to go talk to Waymo's fleet. Well, hopefully our fleet, you know, ends up powering Uber and Lyft and the other folks. But, you know, we'll see how that goes. There's one self-driving company in your name. But it's like you're not going to talk to, like Volvo's going to get, do something over time.

Right. Yeah, they'll work with us. It'll be great. Sure. Well, it's a bold vision. But do you worry that there's going to be resonant, like, nightmare effects of, like, competing cloud fleets? No, actually, that isn't, you know, I, no, that's not one of the, one of my top

worries, right? And the reason why is there kind of is an interoperability standard today, right? And it's the rule of roads, right? It's the, you know, it's a little loose, but the, you know, the DMV for New York State or for California, you know, describes how you're supposed to behave. That kind of bounds the behavior the vehicles can take. and I think at that point you're in, you know, you're in pretty good shape. Yeah.

It's funny that you said there is an interoperability standard. And I was expecting like a white paper reference. Yeah, no. No, no, it's actually just the DNV. It's the DMV. It's the handbook. Yeah, yeah. That's amazing.

But it works, right? You know, today there's literally millions of, millions of independent agents interoperating on, you know, the roads out there. And they do that through this kind of common standard, which is the DMV handbook. Yeah. When you think about, so like, when, Andy writes about Waymo, it's always in terms of miles driven. Why is Waymo ahead they've got the most miles?

Is that a metric that you think about as well? No. And obviously I was there and it was a really good number that we could use. You were responsible for a significant portion of those millions. Well, I was a small part of folks that there's... You're responsible for like 25 miles. There was actually a period of time where I did have the most number of miles. There's a like a leaderboard in the office?

Well, it wasn't, you know, we didn't actually have a leaderboard up in the wall, but you could cut it by, you know, who is driving. Yeah. All right. So, miles driven is not a metric. What's your metric? So, so we, you know, we look internally at a bunch of different performance criteria. So the way we think about developing is we're going to go, we've got a set of features that we need to build that, you know, things like we need to be able to deal with traffic like controlled intersections. We need to build to make left turns across traffic.

We need to build right turns. We need to deal with pedestrians crossing. J-workers. So we have a collection of these features, and then what we do is we go build a bunch of simulation tests, and we go gather a bunch of human-driven representative data driving examples in those kind of environments. And then we internally set the goal to be able to, here's the N, where N is a relatively large number of tests that the system needs to be able to pass in simulation. Once we do that, then we're excited to go see, to then compare, to then can look at, you know, how's it doing relative to the human-driven data and kind of optimizing

in that. That's kind of the way we think about feature development internally. Can you give us a sense of sort of what's the current status of your test fleet? You have cars in California. You have cars in Pittsburgh. Pennsylvania, yeah. Right? So what, what are they doing right now? Yeah, so we have eight vehicles. For us, you know, again, back to the miles, for us, it's not about maximizing miles. It's about maximizing kind of learning and data for the fleet. And so every day those vehicles are on the road. What's cool is we have three offices today. We're in San Francisco, Palo Alto, and Pittsburgh, and we have cars on the road in all three locations basically every day. And that gives us a really good diversity of experiences. So San Francisco

is a relatively modern city, but high density. You know, Palo Alto is suburban, representative of an awful lot of America's kind of suburban roadcape. And then Pittsburgh is actually a very dense, you know, older city with an older infrastructure. And so by having vehicles on the road across that, we're able to build and learn from that breadth of experiences and make sure we're not pigeonholing to, you know, super wide open spaces that never have weather, you know, for example. Like Arizona. For example. So is the idea that a car, one of your cars or, say, any self-driving car company's cars, will need to operate within a certain domain for a certain period of time

before it can become commercially operationals essentially, before it can start to do the thing that makes the company money? Or will you get to a point where you can just basically take a car and drop it into any location and have it begin sort of doing that task? So for us, we're going to need a map first, so we'd have to go do that chunk of data gathering. And then I don't know is the honest answer, right? We expect that as we get better and better, it'll look more like we can just effectively drop the vehicle into a domain.

Because I think about me as a driver, that I might spend most of my time driving in suburban Bay Area. But I can also, you know, not particularly well, go drive in Boston. I would recommend it. Yeah. I don't do that often. And so, right, I think as the technology gets more advanced, it will look more like we can just drop these vehicles in places. But I think early on there will be enough lessons and, you know, to get geeky,

you know, we'll be expanding the ODD, the operational design domain for these vehicles over time. All right. So I want to, we're a little bit over time. So thank you. But I want to kind of wrap up where we begin, which is, is this going to happen? Like, and I asked a question again in that way, but more specifically, do you think that the average consumer is going to go out and purchase a car that drives itself for real within 10 years? No, I don't think that, I think, again, that this is going to come to market in fleets and transportation services.

And I think the economics that are going to be profound. I think the benefits will be profound. I think the adoption curve for personal car ownership is further out. It just, the economics makes so much more sense in the other business model. that I think that's where the energy is going to get focused. Yeah. And how long will people have to be in those cars, do you think, in those sort of early stages monitoring the situation?

Because it seems like half of it is sort of a liability issue, and the other half is a psychological issue. Elevator operators. Yeah, right. You know, like people weren't willing to get onto an elevator without somebody inside of it until they were positive that it was something that they could safely do. So I think when I say within the next five years,

you'll see small-scale deployments, that's without other people in the vehicle, where you'll be able to hail one of these vehicles, get in it, and I'll take you where you're going, or you'll be able to, you know, order something from somebody, it'll show up with, you know, vehicle pull up to the curb, you walk up, pull your stuff out of it, and then off it goes. And I think that happens in the next five years. Wow. So where can people keep up with what Aurora is doing besides Andy's terrific reporting?

So we have a blog post up on Medium, and, you know, we're at www.orgora.org. Cool. All right. Well, Chris, thank you so much for sloven by. It was great conversation. I look forward to getting in a self-driving truck that takes me over. I'm going powered by Aurora. Or self-driving Mustang.

Literally any technology would have been a superior driver than me. It could have been two robot arms. It could be like an Apple LC2. It would have been better than me. Breaking bungee gourd. Yeah. Awesome. Thanks, Chris.

Thank you so much. All right. That was Aurora CEO, Chris Ermson. Special thanks to Andy Hawkins or senior transportation reporter for helping me out with that. I'm not going to be on the Vergecast this week on Friday. I'm not in L.A. doing some L.A. stuff. But Dieter and Paul hold it down. Probably have some special guests in the mix.

We'll be back next Tuesday with another interview episode. Next Friday with the chat show, Birchcast keep rolling. Let me know what you think. Tweeting me. I'm at Reckless. Would love to know what you think of the show and who you want me to interview. I live for that feedback. We'll see you soon.