So, I am super excited to be joined here by Chris Urmson, founder and CEO of Aurora. As I mentioned earlier today, Aurora has done 440,000 driverless miles on public roads, now in at least three states, Texas, New Mexico, Arizona, and growing. Did you know that the moon is a quarter million miles away? So you guys have almost been to the moon and back now.
Yeah. It's going to get faster. Like you're going to have a lot of trucks out there very soon.
Exciting. Exciting. So take us back to the founding of Aurora. You were leading Google self-driving at the time. I think right before it became Waymo.
Waymo. Yeah. Some number of months before. We were in the process of workshopping the name at the time.
Right. Right. What were the other, what got discarded? Are we allowed to know?
I can't even remember. They were memorable.
Okay. They were not as memorable as Waymo. Yeah. So tell us about the founding story of Aurora.
Yeah. So I had been at Google for I don't know seven and a half years I think and it had been incredible. Opportunity to work with great people. We got to do something that nobody thought was possible. You know it was a lot of fun. And then at some point it felt like it was time to move on, right? I had got to the point where I was less convinced than I needed to be that we were going to go and execute and deliver what I thought we could. And I did what I could to try and move the needle and ultimately said, 'Okay, I've got to get out of the way.' Other people have a vision for what's going to happen here and it's not aligned, so let me move out of the way. I spent some time figuring out what to go do next. Got to travel all over the world to meet people and think that through and ultimately came to the conclusion that there was more that I could do in self-driving that I had this pretty unique experience set. I was convinced self-driving vehicles were going to be a thing and I did not want to be talking to my grandkids and like I could have been part of that. I could have done something right and so the kind of the minimizing regret caused me to go found two amazing co-founders in Sterling and Drew to start Aurora and go build something where we thought we could build it the right way with a philosophy that from the experiences we had at Google, at Uber, and at Tesla and bring that together and do something pretty exciting.
Awesome. And so I think the most notable thing right from the beginning, Aurora was planning on doing trucks or was that post?
No, actually so we started with passenger cars because everybody was doing passenger cars. And also because like we had this idea that we should be able to build a driver that would drive anything. But turns out big trucks are kind of a pain in the butt, right? They're big. You have to have a special license to drive them. They don't store well places. They're not great for moving around cities. They're mostly useful when they're driving at high speed on the freeway. And so there's more kinetic energy and more, you know, risk associated with them. And so we started with passenger cars. The other thing that we realized is that you couldn't actually at the time we didn't think you could see far enough with robust kind of redundant technologies to be able to drive safely on the freeway. This was something where in some period of time while I was at Google we had done a bunch of experiments on freeway driving and it worked really well almost all of the time. And it turns out the almost does a lot of work there. And if you're going to have something where people are not responsible for the safety of the vehicle, it needs to work well all of the time. And we just didn't see there was a technology that could solve that. So we focused on passenger cars and then ultimately we moved to trucking.
What caused you to move to trucking then? Was there something about moving away from passenger or?
No. So over the first couple of years of building Aurora, I spent a lot of time trying to find, working with our head of corporate development, Bart Na'vi, trying to find a technology that would close that sensing gap. And we ultimately found this little company in Bozeman, Montana called Blackmore Robotics or Blackmore Analytics Sensors, Blackmore something or other and they had been developing this frequency modulated continuous wave LIDAR which was this technology that Google we had concluded just couldn't be made in a way that would be useful for automotive and I remember viscerally having a test drive with these guys and I'd looked at a lot of LIDAR point clouds and I'm driving down the road not looking out the window but looking at the point cloud I'm like, you know, and there's this set of red dots on a set of gray dots that were moving down the sidewalk behind some bushes, right? And if you just looked at the dots, you couldn't tell there's anything there. But if you could segment by the velocity of them, you're like, 'Oh, there's actually a person there.' You look out the window, sure enough, there's a woman pushing a stroller going down the sidewalk. It's like, that is magic. And so we ultimately ended up buying that company, bringing it in house. And that group has gone on to make this incredible technology we now call First Light, which allows us to see about a kilometer down the freeway. And when you combine that with long-range radar and combine it with computer vision, we now have this really robust sensing capability that allows us to drive trucks at 70 mph down the freeway. And so with that in hand and that acquisition in place, we're like, okay, let's look at the business opportunities here. And we compared the robo taxi space to the automated trucking space and it became very clear that trucking was the right place. So if you compare the size of the market, trucking in the US is about a trillion dollar industry. Ride hailing is about an 80 billion industry. If you look at the unit economics, we value driving a truck at between 40 and 80 dollars an hour. We value driving an Uber at 15 dollars an hour. And so when you think about a business where you're going to introduce it with something that's probably the most expensive it'll ever be, coming into a market where the unit economics support you being profitable quickly is valuable. And then where over time, if there's already a trillion dollar market, we don't need to be driving the cost down to expand the market. We just need to be competing with the cost in market today and do a better job. And then we can make it safer. We can double the utilization of the assets which is critical for these industries for these customers in this industry. And so we're like that feels like just a way better business to go work and it's going to be really hard. Right, it's going to be differently hard than robo taxi but we now have a unique technology. We have a unique capability with the people we have. Let's go tackle that.
So for folks not familiar, tell us about the FMCW LIDAR aside from being able to see a kilometer down the road. What's the main difference?
So you know and I apologize to physics people. I will probably get this wrong but you know most LIDAR that you'll think of are time-of-flight LIDAR. They're pulse light. So what it means they send like super bright spike of IR light out in the world, goes out in the world, bounces off the world, comes back, speed of light is constant. So you measure the time, divide it by two and you now have the distance for each point and you repeat that couple hundred thousand times a second or a few hundred thousand times a second and you build a point cloud. FMCW LIDAR works a little differently. So we send a wave of light out into the world. We interfere the reflected beam or the reflective wave with the outbound wave and we basically look at the phase difference in them and that allows us to calculate the time of flight and thus the distance. What's interesting about this is because of some physics that I probably understood at some point in my life but don't anymore. This is self-heterodyne effect which means that you get a 10 to 20-fold amplification of the signal, which means that you can measure the return, you know, two to four times further than you can with the pulse light. And you're also immune to really bright light. You're immune to sunlight. You can't get blinded by other LIDAR. It's kind of magical in all the ways. And we can measure the Doppler effect. So we get instantaneous velocity. And so all of that means that we get a sensor that is further range, more robust, more capable, and you know, really is an unlock for us.
So the decision to move into trucking actually came like after you had this LIDAR break where you met this LIDAR company and went on to acquire them.
Yeah. Yeah. We wanted to do trucking as part of what we were planning to do, but we figured it would be some point in the future when we cracked this or when we cracked this. We ultimately found this technology. We're like, okay, this technology enables us to go and take on a market where we think we'll have a unique advantage to go win and where it's a really exciting and lucrative market.
Yeah. Awesome. So, Aurora talks a lot about verifiable AI. There's a lot of different approaches out there. Certainly, we're hearing a lot more about people just doing end-to-end learning now in the self-driving world. So walk us through Aurora's kind of high level approach, what is the approach that you use in that and why?
So at Aurora from day one, you know, going on a decade ago, we believed you should learn everything you can. That data is going to be king and that if we get the right data, we can train, we can build models. And that's the right way to build a system, but we need it to be testable and we need it to be scalable. And so as we've built our system, we really think about kind of two core subsystems. We think about perceiving the world and then planning to move through the world. And you know as we all know there's lots of different things you could mean by end-to-end but if you kind of take the way that it's often marketed it's sensor comes in one side, actuation goes out the other end and there's kind of an amorphous set of weights between the two and you know you throw data at it, you get the outcome. Our concern is okay maybe you can get that to work. But what we really want to be able to say is we want to be able to validate the system, have conviction ultimately that the thing that comes out the end doesn't just seem to work but that we know that it does work. And if you think about the two sub-parts, the perceiving the world and the planning through the world part of that, each of those is an exponentially hard problem and if you kind of glom them together into one system you now have the products of exponentials that you're trying to validate. So that becomes a very complicated space to validate. In contrast, if you take those two exponential systems and you are able to create an interface between the two and validate each of the two systems, you now just have the sum of exponentials which is hard but literally exponentially easier. We can also express things and force the model to behave in ways that are guaranteed. So instead of just having a post-training step or kind of having some context that is telling the model that it should do this, we can actually construct behaviors that we want. And you know, one of the kind of comical simple examples of this is stopping at stop signs. So it turns out that people actually stop properly at stop signs about 11% of the time. And so if you just use your training data to look at that, the odds that you stop in a stop line as a self-driving vehicle is 11% unless you do something to it. And so for us, we can just express the things where it is very clear and that you actually do want this particular behavior and you want to ensure that it's always true. We'd actually express that explicitly. And that seems like a strength for us.
Certainly in an industry as regulated as ours is, being able to point to those things is actually really valuable. The stop at a stop sign example is interesting because maybe you're driving an 18-wheeler, you actually do want to stop at a stop sign, but certainly some of my experience in robo taxi in San Francisco is that if you stop at a stop sign for the correct amount of time, all of the other drivers get super confused because they're like, 'What is this guy doing? Like, why are you not going?'
Yeah. Yeah. So the stops I want like I said at least a four-way stop in San Francisco you got you need to stop momentarily right you need to get all the way actual stop.
Yeah. So coming back to trucking, I think you know maybe a common misconception is that freeway driving is easier than city streets but you've got robo taxis right you're driving around a dense environment like San Francisco but you're only going 15, 20 miles an hour if something goes wrong just slam on the brakes and you're going to come to a stop. Freeway driving gets, you know, harder. Obviously, you're going a lot faster. And then you drive a truck and we've got faster plus momentum.
Yeah. And this, like you say, this is one of the, I think, the misconceptions that's out there, you know, kind of broadly is, okay, freeway driving is easy. Urban driving is hard. And no, they're just differently hard. Like you say, in an urban environment, if you don't like something, just stop. If something breaks, kind of work your way to the edge of the road and stop. It may be inconvenient, but it's okay. If you just stop on a freeway lane where traffic is 70 mph, people are going to be completely unacceptable. And so when you think about the redundancies and robustness we have to have in the system, it's just different. If you think about things like the sensing horizon, so in an urban environment, you're moving 15 miles an hour. You probably need in most situations, you know, 150 maybe outside 200 meters of field of view or of range to be able to understand and react to cross traffic and the like. In a freeway environment, we are looking for 500 meters down the road. And so there's a direct implication to the sensing technology that you need. But then there's like these subtle things underneath where because of the angular error that every sensor has, you know, at 150 meters that lever arm is enough that you know maybe it's moving things or at 50 meters it's moving things a few centimeters out at 450, 500, 700 meters out. It's moving things multiple lanes. And the difference between understanding a thing is on the shoulder and parked and how you respond to that versus a thing is in the left travel lane and parked is so different. And so you have to actually use techniques to actually localize the perception elements at those extreme ranges which is again a subtle kind of follow-on complexity.
Imagine even the sensors are moving around a lot more maybe on a big truck.
Big trucks are disappointingly violent situation. Yeah.
Yeah. So, what is the fallback? I guess if driving around San Francisco, I can just slam on the brakes or pull over to the side. What do you do if we're going 70 miles on the freeway?
You keep driving until you can get off the freeway. So, you know, we don't get to have something just break and stop, right? You just not acceptable. Like, if the engine stops, okay, we can kind of work our way to the shoulder. If we blow out a tire, we can work our shoulder. But for our customers, being stopped on the side of the freeway is bad. Because you know it's terribly stopped in a travel lane on the side of the freeway is also not great because you know people are not always awesome drivers. You'll be shocked to learn. And so you know getting out of the way of that is really important to us.
Yeah. I've seen these YouTube videos of you know they put up the lane closed sign with like giant flashing lights and just people just plowing into them.
Yeah. It's kind of disappointing.
Don't want to stop on a freeway. So you've been through this journey of starting with passenger cars and then going to trucks. How do you think about verification and validation and how is that different from trucks versus passenger cars?
Yeah, I think at a high level conceptually the same. We just have a different set of tests I expect. You know the fact that the truck bends in the middle that's an important part of it. It's much larger. You know the kinetic energy is dramatically different from low speed urban driving. And so, you know, because we get beaten up on both the M and the V squared. And so all of that kind of contributes to the complexity of the challenge. When we validate our system today, I think we run it through something like 18 million tests that span from variety of different perception elements to motion planning to, you know, all the rest of the stuff that we do. It's one of the parts that I think is underappreciated in delivering a product to market, right? Is it's one thing to have the thing kind of work. It's another thing to be able to convince yourself and ultimately your regulators and your customers that this product's actually going to be safe out in the road.
