Aurora Innovation, Inc. (AUR) Earnings Call Transcript & Summary

Thank you for coming back for Day 3 of the Industrials Conference here in the transports track. I'm Brian Ossenbeck. I cover transports for JPMorgan here. And really excited to have Aurora Innovation come and kick off the third day here, focused heavily on freight tech. This will be one of 3 autonomous trucking companies we're going to be speaking to. So we're really excited to have the Co-Founder and CEO, Chris Urmson, here in the room with us. He's going to be starting off with a few slides. I think may be a few videos. And after that, we're going to sit down and get right into Q&A. So I think the usual applies. If you've got a question, raise your hand, and we'll get a mic over to you. But with that, let me welcome Chris and have him take us off.

Okay. Thank you, Brian. We're good? Great. Well, good morning. Thank you for joining us today. Happy to share with you a little bit about Aurora. Before I get started, here's our customary safe harbor statement. This is the first time I'm doing one of these here. So hopefully, that's long enough for you all to read that super small font, and we'll move on. So this is our company's mission. That's actually our company's mission, to deliver the benefit of self-driving technology safely, quickly and broadly. Today, I'm going to talk a little bit about the mission, the opportunity in the space and some of our technology so that you can understand a little bit about Aurora before we jump into the Q&A. When we talk about delivering the benefits of self-driving technology, we think about really 4 big things: safety; access to mobility for folks; the ability to transform our logistics network; and the ability to improve the quality of life for people who even today who could drive. Today though, the urgency in addressing this technology is even more critical. We're all witnessing the implications of the supply chain brittleness that we've had for decades that's now basically come to the fore. Not all of this is due to access to transportation and freight challenges, but a significant part of it is due to driver shortage. As we get people back on the roads, we're seeing fatalities on Americans roads go up dramatically. And so the opportunity to improve safety is real. And by the way, we're short 80,000 truck drivers, we're going to be short of 160,000 by the end of the decade. So urgency in bringing this technology to market is real. And the opportunity -- the economic opportunity is profound. Freight in the U.S. is about a $700 billion industry, globally, much, much larger, ride-hailing, local goods delivery another $135 billion opportunity today. And as we roll out this technology, we expect both of those to grow. At Aurora we're building the Aurora Driver. This is a combination of software, hardware and data services that enable a vehicle to drive through -- safely through the world. And it's really -- think of it as a platform that works across different vehicle types. We drive big tractor trailers, and we drive small light vehicles and so we can apply the technology across all of that. And it's because we've thought carefully about how to architect a common core, the essential elements of driving that are applicable across both of these to make dramatically more transferable driver than others. To make this work, we use a multimodal sensor suite, a combination of camera, laser or radar and laser data. And in-house, we're developing our proprietary LiDAR FirstLight. Let me show you why we think that's super important. It's an FMCW LiDAR that allows us to see further than others. In this video, what you're seeing is one of our vehicles on the freeway in Texas coming around a corner and as we stop here, there's construction on the corner -- on the curb. And the FirstLight LiDAR is able to see the turn to this range enough that we know that there's something there. If we roll the video forward now over the next few seconds, we'll drive and you'll see eventually on the left, a traditional LiDAR actually starts to get returns. That turns out it takes about 7 seconds, right, and it turns out in that time, the truck has traveled about 2 football fields down the road. You can understand how that makes a huge difference in your ability to react safely to other vehicles on the road. Now most of the time, you don't care, right, because most of the time, the stuff happening on the road is around you. And so if you can see 100 meters out, and you can react to that normally. In developing this technology, we don't really care about most of the time. We care about those rare events that actually lead to collision, to damage and fatalities because if you can't sell those, you don't actually have a product. And so by investing in FirstLight, we're able to address these really challenging problems that you wouldn't otherwise be able to tackle with conventional approaches. Our driving technology doesn't just work on the freeway. It has to work from our customers' depots or terminals to get onto the freeway. So here you see the Aurora Driver dealing with a relatively complicated 4-way stuff as it's moving from frontage road onto another frontage road and preparing to get back on to the freeway. And here it's navigating carefully around some light vehicle stuff on the line. And then we have obviously developed merging technology. So in this case, it's one of the more challenging situations on the bottom right, dealing with the flow of traffic moving on to the freeway, the Aurora Driver being a courteous driver, making space for those in the way a normal human would, and in the top right here, we're looking at the Aurora Driver merging on to the freeway and dealing with the challenges of traffic there. To develop the Aurora Driver, we realized early on that you're not going to be able to develop this technology just by putting cars on the road or trucks on the road and hoping you encounter interesting things. So we've built a core technology around simulation. And what you're looking at here is our photorealistic simulator where we're able to develop and test against challenging scenarios that you wouldn't normally want to -- normally encounter or want to deal with in the real world. So here, one of the Aurora trucks in simulation encounters a small child running out from between 2 parked cars. And we're able to confirm that the Aurora Driver is able to see and react to that child safely. And obviously, this is a test you don't want to have to do with, well, by chance in the real world. To deliver the Aurora Driver to market, we believe in working in partnerships. So we don't want to do everything. We want to do the thing we are best at world at and that is delivering the driving capability. And so we've built this incredible ecosystem around us. Today, we're partnered with the world's #1 ride-hailing company in Uber, the world's #1 car manufacturer in Toyota. 2 of the top 3 truck manufacturers in North America and PACCAR and Volvo. These are about 50% of the U.S. truck market. We partnered with the largest freight -- the largest carrier in the U.S. by tractor or trailer count in FedEx, forward-leading technology company in U.S. Xpress. The value we provide to these partners is for the trucking companies, we bring the ability to help their product be more applicable in the coming years as the economy comes to market. And for the folks who are using these vehicles, we're providing them access to drivers, allowing them to grow their business, allowing them to increase the revenue per truck they generate and create a huge benefit for them over time. And so that's our story. We're about 6 years old. Really excited to talk to you today and take questions and have the conversation. So thank you.

Okay. Great. Thanks for the introduction there, Chris. I'm glad that it was a simulated kid running in front of the truck. But I think that's one area I wanted to start off with when we talk about scaling and clearly, there's a lot of data and a lot of challenges. You mentioned you can't just run around hoping you'll encounter of interesting stuff that -- so simulation was what has always been really interesting to me, but I guess without the background. Maybe you can help explain how you're sure that the simulation is representative of the real world and kind of what steps you take to verify and test that because obviously that's a key component of doing all these millions, millions or billions of simulated miles.

Yes. Yes. So we do billions of miles in simulation. We're incredibly excited. It's dramatically less expensive obviously to test in simulation than it is to test with physical trucks in the real world, plus we can scale it dramatically more quickly, right? We can spin up 50,000 trucks at a moment's notice. If we went to PACCAR and said we wanted 50,000 trucks, first question would be really excited. But second, it would take years for us to go and get those. So how do we do this? So simulation is really a collection or virtual development tools are really a collection of capabilities. So there's one capability we have, which is being able to take the logs that we have generated from our vehicles operating in the real world, pull them in, convert them effectively, vectorize them to create kind of a more flexible representation and then use those. And so in that case, we know this is exactly a thing that happened in the real world, and so we can play that back and assess our performance relative to it. We have the ability to do motion simulation or motion planning simulation. So this is how does the vehicle react to other actors in the road. And so here, we can take a base scenario and create variations on it. So we can say, like, look at different traffic densities, sweep up speed of other actors and then confirm the performance of the vehicle relative to that. And there we can ground those tests against tests we do at a test track or observe performance in the physical world. And then finally, we have our perception simulation capability, which is part of what I showed in the video a moment ago. And there, we're actually trying to model the way the LiDAR, the radar, the camera, the energy that they're observing kind of bounces through the world, and we do that really accurately. And this is one of the really exciting things by having the physics and engineering team that builds the LiDAR in the same place as the folks who are building the simulation, we can kind of trade their understanding so the link budgets. We can understand deeply the refractions and reflectance properties of materials in the road. In fact, we have a lab set up. We go take bits of concrete, bits of car doors and whatever so we can study how the light interacts with them and then translate that into our model. So we get very precise modeling of the way the energy interacts in the world. And we can compare statistically the performance of our simulator and what it looks like when we put a black car a distance in the simulator versus what we get when we put a black card at distance in the real world and confer that we get -- confirm that we're getting statistically equivalent data from the 2. So that's how we ground the 2 of them together.

Okay. Well, I think one of the other interesting aspects of looking at this business at least for a couple of years now is, it seems like there's the LiDAR crowd and then the camera-centric folks, but everybody who's doing something at scale is doing something more with fusion...

Yes.

Fusing all the sensors together. So are we -- as investors trying to evaluate these things from the outside, is that kind of legacy you're one or the other, does that really make sense? Is everybody moving towards more of a fusion or I guess in your perspective...

Yes.

You have a little more emphasis on LiDAR, but it's clearly not the only thing that you rely on?

We believe deeply in sensor fusion. And it's been actually kind of fun to watch the industry in this. So I've been working in this space for 20 years and early on realized that we needed to use combination with sensors. And the intuition behind that is imagine you have to build a system that's got 3 9s of reliability. And you want to get to, say, 8 9s reliability, each additional line of reliability is 10x harder give or take, right, and we're doing hand-waving math here, but kind of go with me for a moment here. And so if you want to go from a system that's 3 9s to a system that's 8 9s, then it's, what, 10,000x harder to do that. In contrast, if you have to build -- if we take 1 unit of work to build a 3 9 system, if you build 3 3 9 systems, that's 3 units of work. And assuming that they're kind of independent, and again, this is where some of the hand-waving comes in, that's 3x as much work instead of 10,000x as much work. And so that's kind of the intuition as to why it's valuable to use different sensors and sensor fusion. They all have different failure modes that you can complement with one another. Now over the last 15 years, what we've seen is basically everyone with the exception of really one company has come to the conclusion that you actually need to use multiple sensors to do this. Even folks like Mobileye, who is very heavily invested in the camera technology over the last decade, it's gone from, well, you only need cameras and you shouldn't bother with maps and radar and LiDAR to well, maps are actually pretty useful to, hey, we're building on LiDAR and radar. So I don't think there's a whole lot of debate out there what's the right way.

Okay. Great. One of the other areas of I think differentiation for where it is taking more of the tech in-house, you mentioned having the 2 teams next to each other with LiDAR and then with the software but I guess maybe can you elaborate on why you went that route? And given what supply chains haven't really gotten any easier or leaner to figure out since then? Are there any other aspects if you look further to scaling, do you think having a hindsight of the last 18, 20 months that, yes, we might need to change things around a little bit?

So our model has really been only build what we think is absolutely essential or strategically valuable to us. And so when we started Aurora, we wanted on day 1 to build a driver that was capable of operating at low speed, high speeds and fill the whole driving task and I've been leading Google self-driving car program for many years. We talked publicly about the fact we had vehicles on the road, on the highway and then we had people using them as kind of a dog food type test. And what we found from that ultimately is that it worked really, really well, except the rare events that we couldn't deal with were the ones that -- where we had to see further than the sensor suite that was available could achieve. And so when we founded Aurora, I kind of understood that, that limitation was there and wanted to go and find a technology that would bridge that gap. And so we spent a bunch of the first 2 years or I spent a bunch of my time during the first 2 years, trying to answer the question, how are we going to be able to see farther enough to make this robust and safe to drive. And so we talked with all of the different LiDAR companies out there, and we ultimately found this one Blackmore that was doing something really strategically interesting and different. And what that was is this what's called Frequency-Modulated Continuous Wave LiDAR technology. And this is an approach that had been used in RADAR for decades moving it to operated optical frequency since radio frequency is actually quite challenging it turns out. But having come into it being very skeptical because in my previous life, we had studied this and decided it was impossible. Met these folks saw that they actually made this thing work and that this was going to be a game changer. And while there was a lot of folks that were working on variations and durations on kind of classic pulse LiDAR, the Frequency-Modulated Continuous Wave stuff would allow us to see further, allow us instantly to know how fast things were moving, which is really exciting and important when you're trying to understand the world quickly and would be much more immune to noise in the environment. So we would see through more conditions that would be a problem for conventional LiDAR. And they were the only folks that we could find that actually had cracked this and understood it. And so we thought that is a strategic asset. If we don't bring it in, someone else is going to see it, and then we're not going to have access to it. And that's going to be one of the things that will unlock this capability. So that's why we brought them in-house. In other places like simulation because of how deeply coupled that is with the sensor, and it is with the way we do our motion planning and our ability to very precisely synchronize in time that make it super repeatable. That's another place where we saw a strategic advantage. In contrast, we looked at, say, cameras. And it turns out that the folks who make cameras for these guys making for a bunch of other people, and they're really good at it. And there wasn't really a strategic advantage to be gained by trying to do something in developing our own imagers. So there -- that's an example of a place where we go out and use technology from others.

Okay. In terms of one of the other key aspects of this thing is mapping. So I know that's more in-house at Aurora as well, but is thinking of how you balance the reliance on HD maps, the ability to update this quickly, but then meeting the system, the driver in this case to be able to proceed interact with things that aren't on the map or should have been, but haven't quite been updated and all these other basically operating in the real world.

Yes. So we look at maps as a way to exploit what computers are good at, right? Computers are really good at doing math, and they're really good at accessing swaths of information in a way that people aren't. That's one of our strengths. But the kind of the intuition for this is if you think about where you feel most comfortable driving where you're safest driving, it's the places where you've driven a lot, right? You kind of understand right near my house, I understand the traffic pattern at the intersection of San Antonio and El Camino, right, and I understand what's going to happen there. And so I can anticipate reacting better in that situation than if you drop me somewhere here in New York City and asked me to drive around. And so that is a way -- maps are a way to provide the computer, provide the Aurora Driver with that understanding. Now of course, even for me, driving, sometimes something weird happens at that intersection or perhaps there is construction or someone's trimming trees or whatever and so I have to adapt to and refine my model in real time. The Aurora Driver has to be able to do that as well. But the fraction of time where the math is wrong, is so small compared to the rest of the operating time and the benefits of having it are so high that it feels like one of these just -- it's a no-brainer to deploy that technology because it takes some effort to build it upfront, but maintaining it once the fleet is operational, becomes very, very, very efficient, right. You have these perfect probes out there gathering data that you can bring back and then cycle and then over time, as you have more and more vehicles in the fleet, to be able to keep that map more and more up to date and becomes even more valuable and more useful.

In terms of like we focused on the edge cases and how to solve them and what happens if you can't and those sorts of things. So I guess in this environment, as an example, rather if you have a construction and there's like no lanes -- lane markers on the highway. How does the driver react to that now and I guess how far along do you -- are you to like solving that without a map and how you'd address that situation?

Yes. So this is one of the areas that we're actively working on is managing construction. And so you can think of different types of construction. So there's the construction that you've seen before. So most freeways, they're out there working on this thing for...

Like a dance party up somewhere.

Subway. Upstairs? Okay. That seems unexpected.

Jamie Dimon's dance party upstairs.

Yes. There we go. Okay. We don't have much of a subway network in Los Altos, so mind you. So the -- I'm sorry. I completely lost track of it, [ whichever ] dance form.

Yes, sorry, but we're talking about the construction zone.

Oh, yes, construction zone. Okay. Sorry. Yes, I'm sorry. So there's different types of construction. So there's a kind of construction that you just know about well in advance, right. If they're doing work on the I-10, they're probably going to be doing it for the next month and the month after that. And so you can kind of anticipate that. And so you can effectively bake that into the map. Then there's very short duration construction that's quite minor. And so this is -- there are debris clearing down the side of the road and really the road all the traffic behaves the same except you think of it as temporarily a lane is closed, and so you have to move around that. And then there's kind of the first time you encounter interesting construction. So for example, for whatever reason, you didn't realize they were going to close part of the I-10, and they're rerouting traffic around it. And so today, we are mostly dealing with the first 2 classes, and we're developing the technology to deal with the third. But the third one is much more rare. And honestly, it's just we haven't gotten to it yet. But these are things where you really -- again, you can exploit the fact that you drive on these areas fairly regularly. You can update them, they become very limited surprise. And the good news is that they're surprising to people. And so we put a lot of work into marketing them, right. There's big orange barrels and lots of signs and so they become relatively easy to understand for the system for the same reason, they're relatively easy to understand for people to drive through.

So we've got 3 vehicles on the screen here. I think you've got 8 different platforms that you've integrated in. So I think that's one of -- clearly, one of the differentiating components of Aurora. But there's -- one of many debates, I guess, in the industry is like whether it makes sense or not to go into both at the same time. Clearly, you are doing...

Terrible idea.

Some people would say that, but clearly, you would argue against it. And so maybe you can just give us your sense as to leaving Google and then having the chance to start from scratch and then making this choice. And then, I guess, secondarily, you've got the, I think, the beta 2.0 where you're trying -- you're moving the driver into the CNN, which would, I guess, maybe be a proof point for this working. So if you can elaborate on all that, I think that would be helpful.

Yes. So when we start again, we thought about what is going to have the biggest impact in the world and how do we build the technology to capitalize on that, have that huge social impact, but also the long-term economic impact of that. And intuitively, driving is basically the same, whether you're in a big truck or you're in a car, right. You're understanding the vehicles on the road around you, you're figuring out what are they going to do over the next little while and then you're going to move -- you're going to drive the vehicle so that you keep it safe through that. And this intuition is backed up is if you think about the training of a human driver, you can't go and become a truck driver unless you've got a driver's license to begin with. So we obviously think there's immense amount of transferability in the skill set that goes from driving car to driving a truck. And we've seen this in practice. When -- so long as you have thought upfront about the architecture, the way you represent the world, the way that you represent the planning horizon internally, this becomes actually really quite straightforward to move between the 2. If you haven't thought about it upfront and you've built a representation that's around just operating on the highway and then you try to drive in urban environments, you can't represent the way the pedestrians move. You can't represent the complexity of the geometry of the road. It just doesn't work. And so we thought really hard upfront about how do we make sure we have a system that maybe doesn't do it all on day 1, but it's architected to be able to do it all over time. And it comes to the sensor suite, the representation, the motion models, all of that good stuff. As we've gone through this experience, and we have a milestone we're working towards the end of Q1, which is to showcase this transferability, same hardware, same software, operating big trucks and light vehicles. What we're seeing is there's a massive amount of transferability as we expected, that the big differences are parametric right and what I mean by that is if you think about lane changing, for example, the concept that you find a gap that is wide enough for you to fit in. It's the same whether in a light vehicle or in a big truck. It just turns out that the gap is different, right? Big truck, obviously it's longer and so if you're merging, you have -- if you're in a big truck, you may expect other vehicles are going to yield to you a little more than you will if you're in your light passenger vehicle. And so there's parametric differences in the model, and that really takes the same algorithm. We feed it more data from a different class of vehicle and we get good outcomes from that. So we see this as a really important opportunity and it sets up our engagement in ride-hailing. So we're very clear that our first product is in trucking that we're working as far as we can towards that. But then when we go into ride-hailing, we're not going to jump into urban driving on day 1. There are companies doing stuff in that space. We're actually going to go after this very interesting part of the market, which are high-speed trips for light vehicles. And it turns out that those are critical to a lot of the high-value parts of, say, the Uber, Lyft business. And so we'll be able to enter through that partnership with Uber. We'll be able to pick up those rides that we can serve very well, things that look like starting near a freeway, get onto a freeway, get off a freeway, right, things like going from the airport to your hotel or the convention center and serve those customers. And then over time, we'll build the driver out from there. And the whole time we'll be generating really interesting revenue positive models as opposed to being stuck trying to find a way to serve all demand in a location and not really being able to get there.

So you mentioned the first product on the freight side Aurora Horizon coming out, I guess you're looking for a commercial launch in '23. But just before you can get to that part, just how are you looking to integrate in this big, fragmented ecosystem. There's different key messages. There're different actors. There's private fleets, there's for-hire. How are you approaching I guess just from a integration perspective and then just more holistically from a partnership perspective?

Yes. So our model is do the thing you do best, right, and don't get distracted doing the other stuff because there's other people that they do that best. And so we want to bring Aurora Driver to market through a driver as a service model where we will work with our partners in trucking PACCAR and Volvo. We will provide the driving capability, they'll build them a truck as a customer of them, say, FedEx, you'll go to Peterbilt and you'll say I'd like a 579 with the Aurora Driver on it. That will show up. You'll pay Peterbilt for the truck, you'll pay Aurora on an ongoing basis for the ability to drive that truck for you. And so that's kind of how we think about fitting the ecosystem. Do the thing we do, help our partners grow and expand their businesses by them doing what they do best in the world. I think that's the spirit of what you're asking.

Yes. And then just maybe as a follow on. When you look at the partnerships, a couple of slides ago, you were discussing pretty notable ones. But I think if I were to maybe try to fill that out mentally, maybe maintenance and maybe some real estate for being a transfer points. So I guess there's nothing else to publicly announce, but I guess strategically, what else do you think you need to fit in there to kind of fill up that team.

So for us right now, the core has been how do we learn about how automation is going to impact the freight business and how do we provide our partners that opportunity to learn along with us. And so we've gone through a process of building these customer partnerships, what we think is really thoughtfully. And that's been -- look at the different segments of freight and find key partners that are technology forward, that are excited about this technology to want to work with us and then go into partnership with them and develop the product together. And it seems to have resonated really well with the folks we've engaged with. And we think about that engagement really in 3 phases. So the first phase we have is one where we get -- we sit down with them, we talk about our thoughts about how this technology is going to roll out. We hear from them what are the lanes that are most interesting and valuable to them. We kind of put that together and we come back to them with this kind of what we're thinking. And we get initial indication of interest. So this is a nonbinding just like, hey, if you do this, we would like to have this many trucks on the schedule. And I want to stress nonbinding. The second phase from there will be we'll take that set of initial interest and come back with some initial indications to our partners of allocation. and say, okay, well, this is how many trucks we can offer you or drivers we can offer you on this schedule. And then ultimately, we'll get to a Phase 3, where we'll have a definitive agreement with these partners on this rollout schedule of these trucks. And what's really exciting from our perspective, and we thought this would happen, but we're excited to see it happen in practices. Those initial indications, again, nonbinding are far exceed our ability to deliver through '25. And so that's really positive news. We know we have customers. We know we have a rollout to -- for the rollout plan we're working towards. So we're excited for that.

So one of the areas actually are in the field now is with FedEx, Dallas to Houston, launched not too long ago. So maybe you can just give some high-level thoughts in terms of how that's going, what you're learning. And I think next on the map would be over to El Paso. And is that a similar sort of route you're adding different capabilities, expanding the ODD? What is that -- how does that fit in those 2 together when you get to these real-world pilot programs?

Sure. I can't talk specifically about FedEx partnership. Obviously, we respect the privacy of partners. What I can tell you is that our pilot programs have been expanding. We're now operating 5 days a week. We're operating day and night with them. We're learning a ton about how do we -- how do we do handoffs and exchanges between their folks and our folks, we're learning a lot about their demand patterns and what matters to them in terms of the capabilities of vehicles. So that's all very exciting. We've been operating on Dallas-Houston commercially. We are -- been testing on Dallas-El Paso. Dallas-El Paso is really exciting because it's a key part of the Southern Freight quarter in the U.S. Obviously, it's a nice 600-and-something mile trip, which makes it a really interesting place where human drivers just aren't excited, right. This is a tedious unpleasant journey. And so it's a really good place for automation to come in and help with the driving task. And so we look forward to having more to announce about that in the future. So the question is how long does it take to learn and new market. The driving capability is dramatically transferable. So we're not -- that isn't really the heart of it. What we do see on Dallas-El Paso is there's just a lot more construction than there is on Dallas-Houston and so that's a capability we've been developing. But once we have that capability in place, right, that looks an awful lot like a El Paso to Phoenix, for example, the same kind of driving conditions. We're not seeing anything fundamentally different about that. And that's part of the thesis and why we think trucking is a really great place to start is because a bit of freeway looks the same, right? And one part of Texas and another is in Arizona as in Minneapolis, right? And so once we crack that, we expect this business to scale operationally, whereas we look at the robotaxi businesses that are working in deep urban environments, we see that much more as a kind of a long-term technology slog to get through. And so we just see this as a much more exciting scalable business in the near term.

So when you do these programs and you have the earnings calls and all the investors asking, well, how do we measure this and what's the metric and what's the magic one and please, can we have it? And so everybody gets the same question, there's no real easy way to compare it, but at least how are you thinking of success when you're putting these expensive trucks and systems out on the road. It's obviously not risk-free because you're in a live environment. So how do you think about it internally? And then maybe as a related question, do you feel like there's proof points that we should be watching for on your milestone to '23 and beyond?

Yes. So this is one of the things that we absolutely acknowledge and recognize, this is the thing that everyone would like, okay, what's the measuring stick and how do you convey progress and where you are relative to that and we're internally trying to figure out how to do this. Obviously, it's subtle, otherwise, somebody would have done it already. We think about it as effectively how long can you operate in a commercially relevant setting safely and we're trying to figure out how do you pour into something that's useful. And we think all parts of that are critical, right? You have to be operating in a way that is commercially relevant. So if you have -- if you take 1 driver out of the truck and you kind of replace it with 4 drivers and vehicles around the truck, that's probably not commercially relevant. And so we think about how do we kind of push the technology, how do we make sure we're solving the problems that matter most and then over time tracking the rate of failure, so that we know how many hours you can get there in a commercially interesting way. So that's what we're thinking about, and we'll try and find a way to kind of express that going forward. And then on milestones, you mentioned, one, we have coming up. We'll work towards the end of Q1, which is the transferability milestone to kind of make -- really make the point that we believe for a long time about how the same hardware, software can operate in both vehicle classes. At the -- in Q3, we want to showcase a really important milestone for us, and this is the implementation of our fault management system. So most of the time, the truck is driving down the road and everything works. And again, you have to get most of the time working, but it's not good enough. You have to deal with the situations where things break. And the fault management system is the ability to understand, diagnose that -- well, first, understand that something broke, diagnose the implications of that and then take the corrective action to make sure that the vehicle and public stays safe after that. And so we think this will be a really exciting milestone. We look forward to showcasing that technology because it's one of those core elements of having a safe vehicle and unlocking true commercial operation.

Safety is obviously a huge part of this business, convincing your sales regulators getting the trust of the public. I know there's a pretty substantial safety case that's been published inside Aurora and it is out there and we kind of look at it. But how do you feel similar, I guess, from a metrics perspective, and you've got this big safety case, a bunch of check boxes, all these processes, all these documentations. But how are you going to translate something like that into the proof that you feel that you're ready and then to be able to get others to kind of come to that same conclusion without having to go through this massive case. So how do you -- I know you're going to document it and you're going to get to that point. And when you're safe enough, you'll pull the driver and be ready but that's a lot of people to convince along the way, I guess.

Yes. So the safety case is, for those who haven't seen, we have our framework published on our website. And you can think of it as a structured argument as to why you should believe the Aurora Driver is safe. And it starts with, I think, the first 3 -- think of it as a tree. And at the top level, to be safe, you have to be a proficient driver. You have to fail-safe, and then you have to be constantly learning and improving. And then there's a couple more about kind of process and culture stuff at the company. And so that kind of makes sense, right? If you want -- if you thought the driver was safe, it has to be proficient. It has to deal with when things break, and it has -- you have to be learning from those things. And so then below that is what does it mean to be proficient. And so we have a decomposition of that. And so that kind of ends up getting down to ultimately get to a leaf node that says, okay, to be proficient, it means we have to be able to merge and deal with X, Y and Z and then we'll have some kind of evidence, maybe it's simulation, maybe it's experience on the road, maybe it's some other thing, proof by analysis that says we have confidence that we can deal with merging. We have confidence we can deal with 4-way stops, whatever the situation or whatever those things are in there. And then we'll effectively roll that up. And so if you kind of agree at the top level that those are both necessary and sufficient to show this you're safe, then we kind of take each one of them and decompose it into manageable parts over time. And ultimately, you kind of take the sum of the parts and you say, okay, we're good to go. And part of the process of bringing people along is sharing this. So at this point, it's public, how we're thinking about this. We go and spend time with the regulatory bodies, help them understand what we're thinking, how we're making that progress. And then ultimately, at the end of the day, we'll have our internal team, we'll say, yes, we have now conviction that this is safe and we'll be able to put the thing on the road because this is one of the things that maybe is not well understood. In the U.S., the way the regulatory framework works is that you get to make this decision yourself, right? It's a self-certification regime. And so there isn't a regulatory barrier to launching. Now they have a lot of tools they can use to protect the public if you misbehave or misuse this capability. So there isn't like a lack of oversight, but the ability and flexibility to launch is really within our and our partners' hands.

And thinking of that trade, you have to go down to like each and every node of the kid running out from behind traffic, the red bouncing ball looks like it's more on the passenger side, but do you have to really -- I guess, how down do you have to go to really tick all the boxes that you feel or someone else might feel from all these edge cases and all the different permutations?

Yes. And that's it, right? That is kind of the art of it and the science of it is, okay, how do we have conviction? We've covered the set of things that matter. And there's a variety of engineering tools you can use to do this. There's things like FMEAs, Failure Modes and Effects Analysis. You can look at how different things break and the implications of those breaking and understand what matters and what doesn't and make sure you mitigate them. There are hazard analyses, which are kind of a top-down thing, okay, what can go wrong, what we could crash into something, how could we crash into things, well, if this happens and this happens. And by using a combination of these tools, you end up building -- you have systematically built conviction that you've covered the space and that you've got the right data, you need to be confident, yes.

So maybe we can just wrap up with the next big thing that you're targeting, at least on the freight side, which is the commercial launch in end of '23 I believe, so if you can just give us a sense as to what that entails, what success looks like and then the main I guess -- well, we can talk about the positive upside to go maybe faster, but what are some of the bigger challenges and hurdles that you and the team are working on to get to that milestone?

Yes. So we are working towards this true launch of a commercially viable Aurora Driver at the end of '23. That will mean that we have vehicles that are operating for customers on a regular basis, daily basis, pulling loads to then part of their fleet, serving their customers and doing that without people onboard, right? Today, we're doing this already with people onboard. We'll be making that transition to not have people onboard on a really interesting valuable location. Between here and there, there's just a bunch of hard work to do. So our partners have to do their work in -- our vehicle partners in preparing the vehicles so that they are ready to be automated, putting in place the redundancies, the power distributions and doing their safety analyses that are essential. And so we're working with our partners through those challenges and more. And then on the Aurora Driver side, we have to kind of complete that corpus of evidence that gives us confidence that the vehicle is safe to go. And that includes developing some of the capabilities. And things we're pushing on right now, construction is one of those, expanding capabilities there. We're pushing on debris in the road and what you do most of the time if there is debris in the road, should you just kind of ignore and drive through, but if obviously, there's some level of debris that you have to actually respond to and so those are a couple of the areas that we're working towards.

Okay. Great. Well, we are out of time, but thank you very much, Chris, for joining us today. Really appreciate it and good luck.

Thanks very much. Appreciate your time.