This Week in Startups - Aurora CEO Chris Urmson on building the roadmap to fully autonomous vehicles | E1228

Episode Date: June 8, 2021

Chris Urmson the CEO of Aurora, a leading self-driving company, joins Jason to recount the history of autonomous vehicles (1:06), the roadmap for removing human drivers (16:47), what technology problems are left to solve (38:25) and more!

This week in startups is brought to you by Fundrise provides access to diversified portfolios of private real estate to all investors with their industry leading, easy-to-use platform. Sign up today at funrise.com slash twist. That's F-U-N-D, RISC.com slash twist. Our crowd helps you invest early in pre-IPO companies alongside professional VCs. If you're interested in investing, you can join Our Crowd for free at OUR, CROWD.com slash twist and BrainBase. Protecting your ideas should be simple. Built by founders, for founders,

BrainBase File is a clean and automated trademark filing platform that gives anyone the ability to protect their idea. File now for just $169 at BrainBase.com. com slash twist and at checkout use code twist. Hey everybody, hey everybody. Welcome to another episode of this weekend startups. We had a great time having Dara Cosro Shahi on the podcast from Uber recently. And we talked a little bit about how Uber had sold their self-driving unit to Aurora, a company that was founded by today's guest.

The CEO and co-founder of Aurora is with us today. His name is Chris Hermson. Welcome to the program. Thanks for having me. Glad to be here. Yeah, it's great to have you on. Now, people may or may not know this, but you are kind of a legend in self-driving. You were part of DARPA's self-driving contests. I think that was in the early aughts. When did that program start? And maybe you could explain to people what your participation was and why DARPA was doing these self-driving challenges and how absolutely pathetic and insane-looking self-driving was. Back in 2007. Yeah. So back then, self-driving wasn't even really the way we talked about it. We talked about autonomous vehicles or robot cars.

And so back in... Robot cars is great. Yeah. Or just robots. So back in 2003, DARPA had their first grand challenge. And this was to get a robot to drive from Los Angeles to Las Vegas across the desert. And I was back at Carnegie Mellon at the time, as a grad student. and DARPA's goal was to get young men and women out of harm's way, right, out of the logistics supply chain.

Because if you think way back to the Iraq War, that was where we lost the most troops was people moving goods and supplies. And so can we make that safer? And so they had this challenge where they invited teams from across the country and, in fact, from around the world to come and compete. and the idea was it would set off from, the robot would set off from a town in California, outside of L.A., drive across the desert to Las Vegas. And so Carnegie Mell and I was a grad student at the time. I was the technical director for our team,

and we built this just amazing humvee, so a bright red Humvee, where he cut the top off of it, put an electronics enclosure in there. What year was this again? This was 2003. So in 2003, almost 20 years from when we're taping this,

DARPA said, hey, we need to have supply lines during wars. You know, people die when they're driving their Humvees or trucks filled with supplies. There's got to be a better way. So they start a challenge, DARPA,

the Defense Research Project, I think. Advanced Research Projects Agency, yep. And so they go to Carnegie Mellon, an incredible school. And now is it true? Carnegie Mellon free when you go there? Is it all scholarships? No, it's not free.

I think for the graduate student program, if you get into the graduate student program, then, yeah, they have scholarships and it's supported by the research funds that we got there. So as a student there in graduate school, do you remember when you first heard about this challenge and what attracted you to it? This is like a life-changing moment for you. No, it was really exciting, right? So I at the time was down in the Atacama Desert in Chile, and we had this robot that was

called Hyperion, and it had four wheels, and it moved at about 30 centimeters a second, maybe 15. So think about it. If you had a walker, that's how fast you're moving. Great. And we were down there. This was a NASA project to explore how do you look for signs of life for the robot on other planets? Because the Atacama Desert is this beautifully incredible place. It's kind of like Mars, in that there's just.

It's desolate. And we were looking for signs of life in the form of little fungus and other things that would live on rocks very hard to detect. My advisor came down and said, we need to build a robot to drive across the desert at 50 miles an hour. And I thought, that sounds really cool. I'd love to be part of that because, you know, it just, it sounded like an incredible technological challenge. Now, at that time, the iPhone does not exist. sensors do not exist that are in the iPhone like accelerometers

LIDAR exists in some form computer visualization and computer cameras are probably limited to very low specs what were the challenges at that time pre-iPhone and then post-iPhone and that how has life changed in this because my understanding I further from a lot of people is making a couple of billion smartphones

We must be at what, 10 billion smartphones have been made over the years. The 10 billion smartphones have had a dramatic impact on your business. Yeah, we have been making fundamental advances ourselves, but we've also been benefiting

from advances in the rest of industry. So, yeah, we have, you know, I think we were pushing it at a one megapixel camera. We had LIDAR, but at the time, a LIDAR was a single beam that scanned across. So it kind of looked like,

a coffee maker. This company called Sick, a German company that made them. And so the challenges were and deep learning wasn't a thing. Machine learning was kind of coming out of the AI winter and we were using some of those ideas. But yes, so we had this Humvee. We had to drive it across the desert. We had just a single beam of laser data. And so one of the challenges, our vehicle had to, one of the challenges was how do you point that in the right place? And so our vehicle had this gyroscopically stabilized gimbal that we built from scratch to point a special laser so that it could steer down the path as it was bouncing over rocks and whatnot. If you remember the itanium, which was a big breakthrough for Intel that's now defunct, right? That was, you know, we had one of

the first itatian processors to try and get enough computation on a vehicle to do this. Today, it's almost foregone that if you're going to build a self-driving vehicle, you're going to use high-definition maps, that didn't exist back then. And so we prototyped some of the first ideas. Or it was just getting started? So Google Maps, actually Google Maps literally didn't exist. Right. But there were Nav Maps, but the idea of a really like a high definition map.

So even Google Maps today isn't high definition, isn't kind of centimeter accurate. And so we were prototyping and building those early ideas that are now very commonplace in the way that we solve the problem today. So for the audience who's listening, describe the hardware stack increase that you're using now in 2021 after suffering through one megamacamac cameras in 2003? Yeah, so today are, yeah, so back then it was one one megapixel camera. We had a couple of these coffee machine shaped LIDARs that were single plane.

We had the stabilized LIDAR. and then we had a very early concept of kind of an imaging radar. Today, and then we had, I think they were one itatium and a couple of the core one type processors. Today we're running GPUs from Nvidia plus some of the high-end multi-core processors from Intel. We run two to eight-machyxel cameras in the vehicles, but they're much higher. performance even than the one megapixel cameras that were there back in the day. We have multi-beam lasers. One of the big innovations we have at Aurora is this frequency modulated continuous wave LIDAR, our first light lightar that allows us to see dramatically further, but also not just see the shape

of the world, but see how fast the world is moving at us. We've got modern automotive radar. So it's very much night and day. And the advances on the hardware have been profound. The advances in software have been profound as well, right? When you look at the hardware advances, that's pretty obvious. Software, which has had a more dramatic impact on making self-driving a reality right now? Is it largely a software problem from this point forward and the hardware is good enough? or does the hardware need to advance another generation or two?

We're going to continue to advance the hardware. We think that's a strategic advantage for us as a company. We think it's an important technology. And we're making some, like, we have some pretty breakthrough things that we're excited about there. I think the software is probably the bigger core to push on. Studies have shown that a truly diversified portfolio needs more than the traditional mix of stocks, bonds, and mutual funds. It should also have some exposure to private real estate. Studies have shown that portfolios with an allocation to private real estate generally

delivered a better risk-adjusted return with more annual income and lower volatility over the past two decades. Why is that? Because of its consistent performance through multiple market cycles. With Fundrise, this level of powerful diversification is now available to you. Fundrise provides access to diversify portfolios of real estate to all investors with their industry leading, easy-to-use platform. you're looking to add stable cash flow via dividends or you prefer long-term growth through appreciation

Fundrise makes investing in real estate as easy as investing in stocks, bonds, or mutual funds. Fundrise's team of real estate professionals carefully vets and actively manages all of their real estate projects. And with their easy-to-use website, you can track your portfolio's performance and watch properties across the country are acquired, improved, and operated via dynamic asset updates. So here's your call to action. See for yourself how 150,000 investors have built a.

a better portfolio with private real estate. It takes just a few minutes to get started. Go to funerise.com slash twist. That's f-undr-I-S-E dot com slash twist. Funrise.com slash twist. What were in the darker challenges, I just remember seeing these videos and just being enamored with the entire concept,

but it was also kind of made fun of just how pathetic these things actually performed at the time. So what was actually, how far did they actually go? And they were crashing into each other, flipping over. I mean, it was not pretty. I think he's a generous assessment. I think like anything, it was early days.

And so that first race, our vehicle, about 10 days before qualifying for the competition, we were out testing and rolled the thing and crushed the sensors and had to rebuild it. I remember. Oh, my God, that's got to be heartbreaking. You spent a year on that thing and it cost 100 grand and you rolled it. And you rolled it. It cost a little more than 100 grand. And so, you know, like any team, we rallied and pulled it together and qualified.

And we ended up actually qualifying first and heading out in the desert on race day. And our team went the furthest. And of the 150 miles we were supposed to go, we went about seven and a half. Product of the speed and the distance we went, it was a huge step forward. We did drive through three fence posts. the poor thing ended up on a berm at the side of the road it kind of cut a corner and got high-centered

and this is actually one of the more tragic robot videos you'll see where it gets itself high-centered and it doesn't understand that so it realizes it's not moving forward and so its answer is press the gas pedal harder and so the wheels are spinning and they're just they're just grating grazing the trail

and so they start melting the tires so you get this big black cloud. And then we, of course, were not out on the course. The judges were from the Defense Department. And so they had a kill switch. So when they realized, you know, it's not getting out, they stopped it. And one of the really interesting little tidbits is Humvees, unlike your car, which will have brakes out in the wheels, actually have the brakes inboard.

And, you know, this is for a really good reason, right? If you're out in the middle of the desert and someone's shooting at you, you don't want the brakes to get shot. So it turns out when they, it turns out when they, when they hit the emergency stuff, these clamped in, but the wheels had so much spinning inertia that they snapped the half shafts. And so this poor thing has been out in the desert and trying its hardest and it gets stuck and it's almost literally on fire and then it breaks.

So it's kind of a tragic end of the day for the thing. What are you thinking and what is DARPA thinking? Was there a spirit of we're on this journey together and, hey, we got 5%? So if we just double every year, hey, you know, we'll, we'll. be, we'll be able to get there in 10 or 20 years. Or was it like, did you have this sense and pit in your stomach? Like, this is a, this is a bridge too far. We're never going to get this done.

No, it was, it was clearly, we're going to go back and do this again. And, you know, to their credit, that day, the Defense Department said, we'll be back here in a year. You know, we'll update the rules come play. And, you know, from their perspective, they, you know, they weren't funding this like a conventional research project. So the, the idea was if you get to the finish line, we'll pay. you a million dollars. If you don't, you know, thanks for coming out. So the second year,

they're like, okay, we're going to do this again. And if you get to the finish line, we'll pay you two million dollars. Uh, right, if you're the first, if you're the first of the finish line, we'll pay you two million. Um, and so that's what we did. We spent the next year, uh, working our butts off figuring out what didn't work. We built a second one of these vehicles. We're out testing in Nevada, just again, some beautiful train with these, these vehicles out there. once again we managed to roll one of them, you know, 10 days before the qualifying and got it put back together. We go through the qualifying. We've got these two Humvees.

We qualify first and third with the team from Stanford in the middle. And then comes race day. We launch it off into the desert. And it turns out this time, both of our vehicles finish, along with the team from Stanford and two others. So it was a huge step forward in terms of addressing a problem. What was the leap there that you went from nobody qualifying and going 5% to three people make it 150 miles and somebody gets too milly? Yeah, it was five of us made it the whole way. Wow.

It was time, right? So we had been, what was exciting about these challenges was there was a lot of great work happening. And they allowed us to focus all of these research advancements, you know, into a single thing and have it go operate. And so it was it was the time and energy that we put into it. over the year and the consolidation of work. And it was just, you know, it was really an incredible day for robotics. And so that was the second challenge. That year, the team from Stanford won.

Our team came second and third from Carnegie Mellon. So we were obviously a little disappointed by that. But, you know, it was nonetheless an incredible day. And then a couple of months later, the Defense Department said, that's been awesome. What we'd love to do is have another competition. But this time, instead of driving in the desert, we're going to do this on a road network. Because, you know, it's great that your vehicles can kind of stay on the road and not hit too many things. It would be really nice if they could stay on their side of the road and they could interact with other traffic.

And so this was the urban challenge, which we had the competition for in 2007. Wow. And so you get through all these journeys and here we are. It feels like self-driving is very close. I drive a Tesla every day. and I've been using autopilot and watching you get better and better and better. And obviously, you've been watching other projects get closer and closer. Yet, this idea that we would be in a car without a steering wheel, without a driver, still seems to be the elusive three to five years away, three to five years away. Sitting here today, having been through this for two decades and knowing what you know,

when do we think we'll have, you know, an Uber or a Lyft pick you up, in a major city somewhere in the country could be on a predefined route and take you from point A to point B without a safety driver, without a steering wheel. Yeah. So the steering wheel is a second thing,

but today you can go to Phoenix and the team I used to lead has some vehicles on the road. You can call one and get in and it'll drive you and there'll be no one in the front seat. and that's kind of incredible. What kind of route is that going? Is it like on a predefined route where you're absolutely certain?

And is there a safety driver like at home base listening over 5G or something? So I don't know exactly the details, but I don't think it's a, well, it's not a predefined route. That there's a collection, like there's an area that they operate within. It's not everywhere, but it's a chunk of Phoenix. Is this yours or this is Waymos? This is Waymos. This is Waymos. This is a team I used to leave. Yeah.

Yeah. So they're doing these routes. We've seen that before. And I think there's a, I think the idea is there's a safety driver somewhere. When will you, where are you at with Aurora's technology? And when do you think you all be doing point A to point B?

And where would that occur in a grid like system on a predefined route in, you know, on the 280 or something like that, or the five between L.A.? What do you think is the first taste we'll get of no driver and just lean back can sleep in the car or play chess or something. Yeah. So for Aurora, we're taking a bit of a different tactic.

So for the last few years, we've been building the foundational technology to actually build something to scale and to be able to commercialize it rather than to build a demo. And our first product will be in trucking. And so expect that to happen likely in Texas because that's one of the major freight hubs in the U.S., one of the places where trucking is needed most. And yeah, it'll be something where Shuckled depart from a terminal, get on the freeway, drive for a number of hours, get off the freeway and stop at a terminal at the other end. And that'll happen in the coming years. And we think that's the right first path for a few reasons.

So one is that it's a really hard problem. You need to see a long way. And we have the technology we think maybe uniquely that we can see far enough to make that happen. This is the LIDAR you have that I think is 1,000 feet right now. This is the first light lightar, which allows us to see, you know, several hundred meters down the road allows us to see not just the things that are there, but how fast they're moving, which we think is really powerful relative to kind of conventional LiDAR in the space. The other really exciting thing about this approach from a technical point of you, there's two other points I'd make about that. One is that the freeway network is really self-similar. And what I mean like that is if you got on a bit of freeway in Texas or you got on a bit of freeway in Arizona or you got on a bit of freeway in Minnesota, they'll basically look the same and they operate the same.

And so as you think about building a business, the ability to scale that business through operations as opposed to technical advancement is really important. In contrast, if you go to San Francisco and you go to an intersection and then you go say five blocks away, the road looks completely different. The type of actors are around changed dramatically, the structure of the intersection. And so to scale there, you have to be advancing technologically. And so as a business, it makes sense to think about things that apply the freeway network first. The other is around the way that the kind of opportunities you have to optimize and think about performance versus safety. So if I want to take a taxi from my house to the movie theater, right, I'm pretty particular about when I get there because I want to see the previews before it starts. I want to make sure I've got time to get a popcorn or whatever.

And I know that the right route is to take this road, then that road. And if I get in a taxi or an Uber and it doesn't take that route, I get really frustrated, right? Like, no, there's a much better way. We've got to go that way. Whereas if we're a truck hauling goods, the goods don't care as long as we get there within the three-hour window. And so the driver then can pick the route that is incrementally safer, incrementally easier, so long as it meets the service requirements. Right. And that ability to use our understanding that this little bit of road is safer for whatever reason means that we can start to deploy and build the product sooner. Whereas if you have to serve every customer and all of their whims, that's a super demanding and constrained environment.

Hey, it's time for another Our Crowd deal of the week. Right now you can join Our Crowd's investment in Moodyify. According to the deal memo, Mootify is radically altering the multi-billion-dollar fragrance market by digitizing scents. Mootify is the first company developed software that enables function-specific sense that, according to the deal memo, improve mental performance, eliminate the perception of bad odors, and much, much more. You can get in early on Mootify and other unique opportunities at ourcrowd.com slash twist. By the way, did you know that Our crowd investors were able to get in on some of the best IPOs of 2019 and 2020. They benefited from companies IPOing like Beyond Meat and Lemonade, and some of OurCrow's

companies have been acquired by buyers like Intel, Nike, Microsoft, Oracle, and Uber. With OurCrow, accredited investors can invest directly and easily in startups early before they IPO or get bought. Accredited investors can participate in single company deals for as little as $10,000, or one of Our Crowd's funds for a business. as little as $50,000. Again, the R-Crowd account is free. Just go to O-U-R-C-R-O-W-D.com slash twist.

So when would you feel safe driving in between two self-driving trucks in Texas? Today, next year, two years with your family, you know, in between those trucks, because people are a little nervous about this. Yeah. So today, these vehicles are on the road with highly trained operators, vehicle test operators in them. I feel very comfortable driving around them today, absolutely, no doubt. Taking out the operators, would you feel safe, or do we need another year or two?

We need some time on that, right? And that's a process where, you know, I can tell you, our technology isn't there yet today. We take safety very seriously to have confidence. We're going to have a safety case that explains why, right, and captures the work we've done. as to why we believe this is safe. And that safety case really has three pillars to it. One pillar is how is our organization operating and how are the vehicles being operated?

So are they being maintained appropriately? Are we verifying calibrations? Are we doing all of the operational stuff? Are we constraining it so it's only working in places where we know it's good? So this is operational safety and organizational safety. The second is really around what happens. happens when something breaks. And this is normally what people think about when they think about

functional safety or safety or product. So it's driving down the freeway and a screw comes loose. Does it fail catastrophically or does it actually have paths to mitigate that and behave safe? So graceful shutdown. If the LiDAR got hit by a bird tragically and, you know, you got bird guts on the LiDar, not to be graphic, something has to happen. And the very simple thing to happen is slow down, put the flashers on, and go into the shoulder. And if it's on a road, as you're saying, that's super predictable. Yep. And so that's what's called functional safety.

And that's that second pillar of safety. And then the third is, is it safe operating nominally? And kind of the... Nominal being normal. Normal, yeah. And the jargon word for this is normal. Yeah. And the jargon word is so diff, safety of the intended function.

And so this is when it's driving... Safety of the intended function. Function. SOTIF. I like that. Yeah. That's not a bad word. Is that a military acronym or?

It's a safety. It's a safety school term of art. Yeah. Got it. Action. A word I should say. And so that is when it's driving, you know, it drives down the road, not the sidewalk, right? That when it's driving normally and everything's good, it's slowing down appropriately for other traffic or it's merging safely on. And so there's a collection of data for that.

And so ultimately we have the safety case that covers all three of these pillars and explains with either this is the design decisions that were made that lead to safety. These are the tests we've run to verify those designs. This is the process of procedures we've tested and validated to show that it works. Now, you don't operate in a logical world. You operate in the emotional real world. And the real world is filled with politicians, the media. link baiting media at times who like to take maybe a safety record that is 10 times better than humans and, you know, harp on, you know, somebody using, let's say, autopilot incorrectly or, you know,

somebody stands on top of their motorcycle. Like, this is just people using the technology, not as it was intended, not the intended function. I don't know if there's an acronym for doing stupid stuff with technology. But people do stupid things all the time with all kinds of technology. they do donuts with cars, whatever. And so is they're going to be a middle phase in self-driving with these trucks in which maybe instead of trying to go fully autonomous, we say all drivers should have these safety systems as standard.

So that in, you know, we, and I know this is an optimal maybe for a business or maybe it is where we just say as a society, listen, we're going to keep the drivers, but no more drivers without LiDAR, no more drivers without computer visualization. The drivers have to have these basic things, just like anti-lock brakes, you know, seat belts, et cetera. Wouldn't that be just as good of an outcome because we keep the driver's employed? Sure, it takes another decade, but then they're not capable of running into the back of another truck because, you know, the system would not let them. Is that going to be the likely scenario? Is a decade of just, you know, upgrading all the technology until we get to self-driving. And is that a victory for you as a

business? I don't think that's how this plays out for a variety of reasons. So first, there's a lot of really good work happening in what I would call driver assistance systems, which are the kind of things you're describing, where it's, if you aren't paying enough attention and you fail to hit the brakes because you're about to crash into something, it'll start to hit the brakes for you and mitigate the collision or potentially avoid it. And I think that's a really, valuable and important technology to go, to be pushing forward because it can help in certain situations. There are limits to how far you can push that technology. So if I'm a aggressive sporty driver, I may be charging up behind another vehicle with the intent to make an aggressive

last minute move to kind of swerve past them. You know, I'm kind of a jerk, but I know what I'm doing. and if at a moment where I'm about to make that swerve, the vehicle decides there's about to be a collision, and so it hits the brakes. Well, it's using up a big chunk of the friction that I have available to make that maneuver. And so now I can't make that maneuver, and now I end up crashing into the back of that thing

because I was planning to swerp to the left, and it hits the brakes, and I can't swerve anymore. And so there's limits to how aggressive these systems can be because they don't understand the mind of the driver. And so there's a bound on the performance there. And so there's a fundamental limit to how good they can be. They can help a lot of things, but they can't help all cases. And so that's a problem.

The other is that one of the real advantages with self-driving vehicles is that the safety benefits come along for the ride, if you will. So you use a self-driving vehicle in a personal application because you can't drive anymore and you'd like to get somewhere. or it's inconvenient to park, and so it's nice to get a ride to where you want to go. And you don't use it as a customer because it's safer necessarily. Use it because of all these other advantages.

And it just turns out that it's an attentive, thoughtful, safe driver. And so you get the safety benefit. Whereas if you had to pay just for the equipment to provide the safety better, you probably wouldn't. And it would drive up the cost of the vehicle in a way that, What will these trucks cost then? I mean, that makes total sense. You have to get the value of the self-driving.

What is it going to cost? What does it cost today to outfit a truck? And then what do you project it would cost five and ten years down the road? So three numbers, maybe ballpark. Yeah, I think, so I think today, ballpark, these are hundreds of thousand dollars of equipment that you're putting on the vehicles to test. Got it. 100,000 and 200,000, yeah. Somewhere, you know, kind of in that range.

Right. And it's small volumes of prototype stuff. There'll be, like anything, it'll come down a cost curve to this gets to, you know, maybe it's 10,000. Maybe it's, you know, 15,000, 20,000, something like that, right? But, you know, and those are kind of orders of magnitude type numbers. And it turns out that those cost numbers make an incredible sense in the terms of a fleet business. So an Uber or a FedEx or a UPS or a Schneider.

Amazon, whoever. Any of those type of customers. But that's still expensive for you as a personal ownership or for personal ownership. Part of it also is that, and this is one of the longer term trends that I expect, is that for many people it won't make sense to own a car. That if you think about a car today, you drive it probably 4% of the time that you own it. Yeah, it makes no sense to own it. And particularly if you live in a city, whereas in the future,

if we bring the cost of delivering that service down to the point where it's competitive with owning your car, now you don't have to worry about parking and circling the block and seeking that out, right? You know, you can work and enjoy it or enjoy the ride instead of driving. It'll make sense for people to flip over and use these vehicles, and they'll get used dramatically more than 4% of the time of day. And so you're amateurizing the incremental cost increase over much more utilization, so it becomes dramatically more cost effective as well. Two questions about, you know, this interim period. Should certain roads or certain lanes on certain roads be designated as self-driving as part of this process as a society as we adapt this?

And then what role does speed play? Should these trucks be in their own lane going 50 miles an hour or even 40 miles an hour? I don't know what's safe or 55 and just be kind of, you know, this stream, if you will. running next to regular traffic with even cones or something or a barricade. And we have that as a starting point because if they were going 45 miles an hour in their own lane and running 20 hours a day, 10 hours a day without a driver, I think we could all feel very safe about them. Yeah, I think they have to work with the infrastructure as is today.

Okay. Our road infrastructure hasn't kept pace with the growth in utilization, meaning that our roads are busier, more congested than they ever were. And so the idea of taking one of these really valuable common resources, you know, a lane down the freeway, and dedicating it to this technology at a point where it hasn't proven itself and isn't generating real societal value, feels like not the right answer. And so as we get this technology deployed, as we start to see value, then we can have conversations about do we want to change the allocation of these shared resources. But I really think It makes sense to meet the world where it is, as opposed to try to conform the world to what we might want it to be.

Every startup needs to ensure they own their intellectual property or IP, and that starts with filing your trademarks. What is a trademark? Well, that is the name of your company, essentially. There's copyright, there's trademark. You read all about it on the internet. I have launched my investment company. I have this weekend startups. You need to trademark the names of what you do so somebody, some bad actor, doesn't come

along and say, oh, I'm going to file it. And then you've got to prove and spend a ton on legal to try to say, I use this before them and I'm already using this in commerce is going to cause all kinds. Trust me, I've been there. If you don't know where to start, look no further than brain-based file. It's a clean, simple, and automated trademark filing platform that gives anyone the ability to protect their best ideas. There is no need to spend thousands of dollars on a lawyer to file your trademark for you. Nope. Now, now you can do. everything yourself in a few easy steps. Brain-based file gives you goods and services recommendations using AI so that you can avoid back and forth office actions with the US patent and trademark office

and eliminate human era. They also offer full transparency into the USPTO's process with step-by-step notifications and real-time updates on your trademark's approval. No one likes dealing with trademarks, but BrainBase File makes it easy, elegant, simple, and affordable. Head to BrainBase.com and enter the code twist at checkout to follow your first trademark now for just $169. That's a 15% discount. Thank you to BrainBase for giving our audience a nice discount. What about speed? You know, I've always been fascinated by this kind of concept that we care about safety,

but we can very easily speed limit cars. Like when you rent a car, it's speed limited. And as a society, we have this kind of concept of freedom in America. and I guess part of that is freedom to buy a car that's capable of going 205 miles an hour, 150 miles per hour, even though there is almost no use case for that except going to the track. And why don't we have speed limits on cars, you know, some reasonable 100 miles an hour, 85 miles an hour, whatever, plus 15 of the, you know, plus 15 miles to the highest speed limit in the country? And is that something you would advocate for with self-driving is saying, hey, these are self-driving cars,

they stay in the lane and they cannot go more than, you know, the speed limit. Yeah, so I can't actually speak to the first societal decision on this. It's an interesting one about kind of American individualism and, but it's not just in the U.S., right? It's just, you know, you have the Autobot in Germany where you literally have no speed limit and you can, you know, enjoy your vehicle, assuming you're competent of driving it well. Quite an experience. If I've done it, you've done it, I take it? I haven't driven my own car there, but I have been on there.

at some reasonable speed. It's wild. I was doing 120 in a beamer, and I mean, I was getting run off the road by Porsche's doing 150, 200 miles an hour. It was bonkers. Yeah. And with that, though, comes more rigorous driver training, more rigorous adherence to the kind of pass on the left rule and don't pass the right rule because it's necessary to maintain

safe driving. So, yeah, I don't know why exactly. we don't. I guess people, you know, like people enjoy the flexibility of freedom, right? America. For self-driving vehicles, I think that one of the benefits is that they won't have ego. And that because you're able to use the time in the vehicle or you're not paying someone for the time in the vehicle, they can drive at the speed limit. And that will increase safety on roads. Because one of the things we know is that speed is a major contributor accidents.

And certainly to fatalities, right? I mean, if you look at the fatality rate, when you go above 85, 95 miles an hour, it just skyrockets. And that's one of the things I love about my Tesla is I could speed limit it and say, you know, no more going over 70 miles. I think I have mindset to 79 or 81 miles an hour or something, just in case I happen to get up there. But, you know, I'm going to live a long life. I don't want to die, you know, and have a great life. it's so dumb. So what do you think is going to happen in terms of,

there are so many people who are working on this concurrently. It's got at least like six or seven major players. Do you think everybody kind of gets there in the next five to ten years and the technology is a bit commodified? How do you see this market coming to fruition? Is it going to look like the PC market where you have a lot of options? You could buy Compact or Dell or Max. or is, you know, two players, you know, yourself and Tesla are going to get there at the same time,

or you Tesla, Waymo, are going to get there and then there's no other opportunity. How is this all going to shake out into business? And what is your strategy there? Obviously, trucks to start makes total sense with Uber freight. And, you know, they can afford to spend $100,000 on a rig, no problem. But what is the market going to look like in 10 years? Yeah. So there's a lot wrapped up in that.

Tell us exactly the few. In the future. Yes, let me look at my crystal ball here. So one is we expect consolidation to continue. I've been saying this for years that it's amazing that there were so many companies experimenting, exploring the space that's necessary in a new innovative technological space. It happens consistently. And then, you know, the teams that have the right culture, the right execution, the right people, you know, go through and end up being a few that make it. I expect there'll be most a handful of companies that deliver on this technology because it's really hard.

And you need to have the experience, you need to have the quality, you need to have the partnerships and backing to make it successful. And so we've been positioning a roar to be one of those players. Something to remember is that the scale of this space is profound, that freight in the U.S. alone is a $700 billion business. That's just in the U.S. Transportation, ground transportation in the U.S. is something like a $2 to $3 trillion space. And so there's room for, you know, incredible companies, incredible growth in this space with, you know, a handful of companies just in the U.S. and globally, it's three to four times that scale. So incredible space, incredible opportunity.

Now, what will the business model be ultimately? Yeah. Selling to, you know, somebody like selling products and hardware and services to somebody like Uber and then just adding 100,000. thousand cars to their networks or is it going to be you know as i guess Elon believes he's going to just put everything as a robotaxy and he'll have a competitor to uber i was just curious how you think the whole space kind of hashes out because it does seem to me with this many players we should have a number of people just like in the DARPA challenges reach this proverbial finish line at the same time yeah i i think that i don't think there's going to be as many as

like i don't see this becoming a commodity is is the short answer um so maybe three or four people make it. Yeah, that's kind of the scale that I think. And so the way I think this works and our business model is really to provide our platform Aurora driver as a service. And so we'll work with our partners, Pac-R, Volvo, Toyota, and others as we build them, where they'll bring the vehicle, we'll bring the driver, and those will be provided to customers through fleets, whether it's through the Uber,

network or through freight and logistics networks. And then we'll generate a revenue stream that looks like a software as a service business. So we'll get paid for every mile that the Aurora driver drives a truck for a partner. And this really comes down to kind of one of our core values of focus. So we think the thing we can do best in the world is build the driver. And I look at a business like, you know, Packard's business building trucks. This is an incredible company, right? It's been around for basically a century.

They deliver incredible products. They know they have the relationships with the customers. Why do I want to compete or replicate that? Why not work with them? And we can bring our understanding how to deliver the driver. They can bring their understanding how to build trucks, their incredible service network and dealership network. And together we can grow their business and grow our business

and serve their customers. And similarly, we look at it with people moving, that Uber's business is profoundly complicated. Why replicate that? Toyota's business, these are, again, an incredible company, 100-year-old company that knows how to make stuff well. Why don't we work together with both of them where, again, we bring the driver, Toyota brings the car, Uber has the network and serve their customers

and allow them to grow their business. How much of the business ultimately will be about the data set that you own and you update? I mean, you have Tesla's on the road that have the self-driving technology in it, whether people buy it or not, and they're just contributing day in, day out. There's million Tesla's out there probably with these, you know, collecting data. You have the way-mail cars, obviously have a much smaller footprint. Then you have every Uber out there with an app on the phone.

It's collecting some amount of data and could collect more. So how much of this will ultimately be about the dataset, the GPS, the real-world data, and then is anybody sharing this data? I always thought this would be an amazing open-source product. if you could get two or three companies saying like, hey, there's the Open Maps project, obviously, but here's our real world data. We went to this intersection. This is an intersection in the world where there's a lot of accidents and all of you pool.

Here's everything we know about this intersection. That is dicey. And you could have some sort of collaboration in the spirit of saving lives and say, listen, when the 280 and the whatever highway merge and the five, it's a disaster. this is where we have the majority of accidents, you know, in this 100-mile range, we all need to pool the data here and come up with a better solution. Yeah, so I do think data is important. I think the right kind of data matters.

So, you know, Tesla's building driver assistance technology. It's a little different than, you know, full self-driving capability that we're working towards. They, I think that, yes, we think about how do you use data effectively. We've put a major investment in simulation because it doesn't matter, it's a expensive to pull data in the real world, we can generate incredible varieties and targeted exactly the type of scenarios that matter most to advance the system and do that synthetically. And we have this, this, you know, the ability to emulate our LiDAR, our radar, our camera, and do that at a sensor realistic way that allows us to get exactly what we want superficially

and then run it super repeatably. So we think that's a messy. So that's literally like working on a like a quake engine or something, like a video game engine? It's a built from scratch engine, right? And so we've brought in folks from the computer graphics industry from film who actually understand light transport because the application is a little different. If you're building a video game engine, you want something that looks good enough, but can churn quickly, whereas we really want to actually accurately replicate the way the light moves through the world, whether that's, you know, or the energy, whether it's RF, or light, LIDAR.

So it's literally like a Pixar film where they try to get the hair right and make it perfect or, you know, the avatar or something. And then you can run a simulation and say, hey,

let's run a bicycle, let's run, and, you know, a car that's flipping over a boulder coming down the hill. You could just run simulation and for simulation and train the AI and the machine learning

to understand what's happening. Importantly validate it, right, from that. And so as we think about gathering data, there's three ways that you, three reasons to gather data. So one, or to have a fleet of vehicles in the road.

One is to understand the distribution of events that happen in the real world. So how often do you see a person or wetsuit step into the road? It turns out not very often. Scooby-diver the road. Look out. Yeah. And the reason why I bring that up is it turns out wetsuit material is not very reflective. And so they're actually a hard person to see.

So, you know, understanding that matters. So that's one reason to get data. And so there's lots of ways to understand the distribution of events that don't require your drive a million vehicles around because other people are doing that already. Then there's gathering the data you need to improve the system. And so for us, we use simulation to do that. We use very targeted on-road tests where we're like, we need to go get exactly this data for the advancements we're making. Let's go get that. And then the third is around validation verification.

So making sure that the work you've done in target. of testing offline and simulation actually closes the loop back when you go and test it against particular scenarios in the real world. And so we focus on, you know, across those applications and we've been really thoughtful based on our experience of how do we efficiently gather that distribution data? How do we make sure we're targeting get not wastefully gathering millions of miles where we're throwing, you know, 99,000 of those models out, right? How do we make sure we have the simulation tools so that we don't have.

have to hope we come across this interesting event in the real world, but we realize that's interesting. We go replicate it and then build the framework so we don't just replicate it once, but we can say, okay, given this rough configuration, generate a variety of variations around that. Maybe it's, you know, maybe in one video, it's a, you know, it's a G-Wagon, and then the next, it's a Prius, right? But, you know, automatically do that rather than having to have an artist to recreate each one. And so that powerful set of offline tools are one of those places where our experience told is, let's go invest.

That will accelerate our development and make a better, safer product to market. As we get close to wrapping up here, tell me about the false positive problem. In other words, you know, the plastic garbage bag flies across the highway and your truck thinks it's a car, a boulder, a scuba diver, and a wetsuit, et cetera, slams on the, break unnecessarily flips the car or causes an accident, unintended second order, third order effect. Where are we at? Is that a serious problem?

Because we did have, tragically in the early days of autopilot, a white truck in the sunlight, go across the road. Tesla went underneath it and killed the, tragically killed the superfan of Tesla's who was watching a Harry Potter movie and not paying attention and breaking all the rules back to not using the technology as it's intended. It's obviously not Tesla's fault there. but there is this problem of like, you know, false positives or white things, you know, that are floating in the end. So I guess there's two different types of problem there.

Yeah. So one is that example of a Tesla going under the truck and having the fatal accidents or fatal collision, that's part of the reason why we think it's so important to have different sensor modalities. If we're using LiDAR, using radar, there would have enough information. there to know that there's a truck there versus a template matching computer vision system that's just looking for the back of things. Well, you know, it wasn't that the truck was white that was the problem. It wasn't that there was a bright sky.

It was that it didn't look like the back of a truck because it was the side of a truck. When the system wasn't able to understand that and didn't have the robust sensing to deal with it appropriately. So that's why we think multi-sensors, multiple sensors are really important. the false positive problem. Is that actually an issue or do people just bring that up? This is actually going back to your question earlier about why there isn't just an evolutionary path from driver assistance systems to full self-driving systems or self-driving systems is fundamentally this. So if you are a driver assistance system, you need to only hit the brakes.

Say it's forward collision mitigation brake. When you are confident that the driver's about to have a. collision because if you're driving down the freeway and there's nothing there, you know, you hit the brakes, you frustrate the user and you do it a couple of times. They take it to the dealer. You do it a third time. They're like, take this back. I don't want this broken piece of garbage, right? And so you design the system so that it only hits the break when it's really confident that there's something there. And what that means in statistical terms is it's high precision that it only calls breaking events when there are one. And it's generally low recall, meaning that

there are a bunch of things where it would have been great if it hit the brakes but it doesn't hit the brakes there because it's not confident enough that there's actually a real event and it turns out that's okay for a driver's system

because there's a person behind the wheel paying attention and they're there to backstop it and if you happen to hit the brakes in one of those events where they're distracted for a moment maybe it only worked half the time that it should have

but that's half the accidents that could have happened avoided So that's an awesome product, even if it's low recall, even if it doesn't work, quote, unquote, all the time. In contrast with a self-driving vehicle where the driver is not paying attention, or there isn't a driver, there's only passengers, it has to work all the time. And you need to drive up, recall, and precision. And so this is kind of the heart of the problem. And to do that means more robust sensing, more robust algorithms and computation on board. And that's why there's such a disconnect between these two technologies.

And so, yes, the false positive problem is real, but the more important one is the false negative problem where you're supposed to do something and don't. What would that most likely scenario be of the false negative? I should have stopped it. I didn't. That is essentially what happened tragically with the Uber and the bicyclist crossing the street in the middle of the night, not at an intersection, and the safety driver was on their phone tragically. You have a double fail there. putting aside the driver's negligence. And a variety of other things, yeah.

Yeah. Yeah. And so a false negative is, is again, it may be that, you know, in our system we've designed carefully to avoid this, but if you have a pure classification-based system that says, you know, let's imagine you had a system that only recognizes cars, trucks, and bicycles, and pedestrians. If there's something, somebody dressed up in a chicken suit, right? They don't look like a pedestrian anymore.

They're not any of those things, so you ignore them. And that's kind of, you know, in very rough terms, and I don't have inside information, but very rough terms, that's what happened with the events. Didn't know what it was in all likelihood. And then just, right. At Aurora, we've taken an approach where we have to explain all of the data we get. So it may be that we don't understand that that's a person in a chicken suit. But we understand there's something there.

And if there's something there, we shouldn't hit it. Right. And this is kind of the common question around the technology is how do you deal with a long tail problem? All of these things you don't see very often. Catch cases, yeah. Right. And the answer is a lot of them really boil down to don't hit the stuff that's there. Right.

And you're pretty straightforward. Don't crash into things. Try to avoid things. And it's, you know, there's some subtlety, but at the heart, that's it. And so our under, you know, and again, this comes from the experience that the team has of understanding this class of problem saying, okay, we're going to explain all of the data we get. Now, we might say that's exhaust and we might be wrong about that. It might be somebody, whatever reason, looks like exhaust. But we're going to have, we're going to have done our work to explain all the different parts of the system, all the parts of the data so that we don't have these kind of weird failures.

This is super morbid, but animals on the road is a big thing. And it happens tragically all the time. People run into things. They hit a deer. They hit a bird. Have you started? And is there any best practice around that? Because swerving out of the way for, you know, I don't know, a poor raccoon could cause a lot of human life.

And we're kind of getting into trolley problems. But is the idea like if this is a small animal, just go through it. And I hate to be so graphic, rather than swerve around it and kill a hen people. I think it really is a function of the situation. What I've been told by California Highway Patrol is if there's debris in the road, if you've got a long, we've got a lot of time and you're confident you can make a link change than do so, otherwise drive through it. And if it's a pop-up situation, it's sudden, you drive through it.

And so, you know, part of... Poor squire. I mean, it's the logic, right? Yeah. Right. And part of that safety case I talked about earlier is doing that kind of analysis so that we can say this is deterministically the right thing to do, right? You know, it's the least bad thing to do, maybe.

Literally the definition of the trolley problem. Do you come into these situations a lot or is it just... No. You all want to, you know, we have an attraction to trolley car problems because they're so mentally and philosophically challenging for us. But you actually don't come into that. In practice, it doesn't come up for a much. Yeah. If, you know, think about how many times it's come up in your lifetime of driving.

It hasn't. Yeah. Right. Neither mine. Right. And, you know, I think if you took a survey of people around you at the office, if you happen to be in an office at some point in the future, right, the answer is it doesn't come up. And it's a really interesting question. about society and psychology and human nature and it's fun for that for sure.

Yeah, but it's not a reality. Listen, Chris, you've been incredibly generous with your time and you've been so candid. Thank you for that. I'm sure you're hiring PhDs and machine learning folks. If people want to come work and save human lives and create the future and work with a legend in the space, how would they be able to find out about working at your company? We'd love to. We are hiring.

We're looking for great people. I couldn't pitch it better than you did. We're at www.aure.com. And you can find more about the company. And we've got jobs list there. We'd love to hear from you. Thank you. All right.

Fantastic. If you're out there and you want to do something good for the world and not just make it 5% better or help some advertiser get 5% better, click-throughs on their ads for Facebook, go do something world-positive and challenging, as opposed to serving the privacy the industrial advertising complex.

That's me saying that, not you, Chris. But, I mean, it is tragic. We have people with PhDs who are going to try to trick people into clicking links for ads. And meanwhile, we need to get to space. We need to have self-driving cars.

We need V-tiles. There's so many more important things to work on for the love of God, don't go work for Facebook and waste your PhD people. It is one of the joys of my career, right? I've had a chance to work on something that matters. and it's something that it's important, it's fun, I work with amazing people, and you can touch it and explain it to people, right?

It's so much, you know, you can have a conversation. You can sleep at night. You can sleep at night. Everybody who's listening who's working on an ad network at Facebook, please quit and go work for Aurora. I said it, not you, Chris. We'd be more than happy. More than happy to take those refugees who are wasting away in the ad-industrial privacy-breaking complex. continued success.

And we'd love to check in with you in another two years and see how it's all going. And it's really important work. And we thank you for that. I think it's such an important thing to save these lives of people on the roads. And also just make society more efficient. And it's just wonderful what you're doing. So continue success. And we'll see you all next time on this week's startups.

Bye-bye.