Aurora's Chris Urmson on why self-driving trucks are finally ready (Live at HumanX)

Hello and welcome back to Equity TechCrunch's flagship podcast about the business of Startups. I'm Rebecca Belan and this is the episode where we bring on industry experts to help us explore a trend in the tech world and dive deep. At the HumanX conference in San Francisco, I had the chance to sit down with one of the top names in autonomy and today we're bringing you a conversation I had with Chris Erbsen, CEO of Self Driving Vehicle TechFirm Aurora. For those who aren't following this, self driving has been almost there for a long time. But the commercialization story may be finally changing with long haul trucking. So I had Chris walk us through what is actually shifted after a decade of technical bottlenecks? Why trucking could have a better pathway than robot axes to real revenue? And what founders and investors still tend to get wrong about this market? Chris and I actually just got off a panel and talked about a lot of this stuff we're going to talk about today. So we're all warmed up for you. Chris, thank you so much for joining me. Thanks for having me and thanks for putting up with me for another half hour. So appreciate it. Please, it's a pleasure. Okay, so give us a little bit of a rundown of where you come from because you've had such a storied career in the physical AI space and where Aurora is today. Well, I've had the privilege of working in self driving for 20-something years at this point. I started when I was at Carnegie Mellon University working on the DARPA Grand Challenges, which were these robot race across the desert. I then got a chance to work with Caterpillar on these giant dump trucks for mines. That was a lot of fun. And then I got the privilege of helping found what's now Waymo, what's the Google self-driving car program and led that for seven and a half years. Ultimately stepped away from that. It took a little break to figure out what I wanted to do next and hadn't got rid of the bug. And so ended up having a chance to start Aurora, founded that with Cerlin Anderson Drew Bagnell, great co-founders. And we've been out of for nine and a half years. And as a company, we've been really focused on our mission, which is to deliver the benefit of self-driving technology safely quickly and broadly. And in the field of trucking. So today, starting April of last year, we have trucks on the road in Texas and now in New Mexico and Arizona, hauling goods for customers in a driverless way. And then this year, we're going to go from having a handful of trucks to having hundreds of trucks. So it's a big year for us and an exciting year for the industry. We talked about this a little bit on stage, but autonomy has been here for over a decade. You're now running commercial driverless trucks. Are we at the start of scale or is this more of a polished pilot phase? I think we are very much in the begin to scale phase. It's exciting. So we've done over 250,000 miles of driver's operations. So we're getting good experience. We have customers who are learning with us what this means. These are companies like FedEx and Werner and Schneider and Hershbach, some of the biggest names in trucking and what they're seeing they like. And so for us, it's this big moment of starting to scale and going from a handful of trucks to hundreds of trucks. And then ultimately, two thousands of trucks next year and onto tens of thousands trucks. So we've kind of cracked the problem and we're now off to, okay, how do we really build a business with this? And important, how do we serve customers and make sure we're building their businesses? How does your roadmap change when you think about macroeconomic trends? Like we are in a recession, not recession. People are spending a bit less. I imagine that that creates less opportunities for shipping, but maybe I'm not sure. Does that change your roadmap? I've had the privilege of working the space that is really transformational. That the impact it will have on safety of America's roads is profound. The impact it's going to have on fuel economy and sustainability is great. And the benefit to our customers in how much they can use the trucks they have and how they can build their businesses is again huge. And so the benefits that we expect and we're seeing with these trucks already really kind of outscale the macroeconomic situation. There obviously are bottlenecks. Like we've established that, you know, we all thought autonomy was going to happen a lot faster than it did. Like what are the main bottlenecks? Is it tech? Is it regulation? Is it just public acceptance? So I think historically it has been primarily technology that this is a really hard problem, whether you're driving in a city or whether you're hauling a trailer down the road to 70 miles an hour, like these are very hard problems. And it's critical that you get it right. And that you make it safe, right? Like you just can't cut corners when it comes to safety with these kind of applications. And so I think what you see is the responsible folks are making sure they get to that point where it's safe. And that has really been what's kind of been the bottleneck up until recently. At this point for Aurora, we are very much supply constrained. So the first generation of hardware that we launched with we knew we could only make a couple of, you know, 20 of those 25 of those trucks. And so we just couldn't build our fleet. We couldn't grow customers beyond that. What's very exciting is in Q2 of this year, we're going to be launching our second generation of hardware with our new international LT trucks. And that will be able to scale up to about a thousand five hundred trucks. And then next year we'll launch our hardware that we've been building and developing with a MoVIO. And that'll take us to scales of tens of thousands of trucks a year. And so we're going to unlocking the supply side. The pilots and work we've been doing with customers are really creating the demand. And you know, we've got a lot of demand this year. We're getting demand for next year. It's really exciting to see how it's resonating and the values there for customers. And then on the regulatory side, we're seeing progress. So it feels really good. I want to talk about the regulation stuff, but also, you know, as we're talking about you scaling. I'm curious, you know, we're all we hear about is chips chips. There's so much demand for chips. Data centers are using all these chips. I imagine that self-driving trucks, you know, doing a lot of onboard inference. Are you fighting with data centers and hyper-scalers and AI companies to get access to chips? Yeah, we're not directly fighting with them for access to chips. The, you know, truck is big and has a big engine, but it still has a lot of power limitations that, you know, just aren't relevant for data centers. And so the kind of chips we use are intended for automotive and physically eye applications. And so that third generation hardware that we're launched with the MoVIO will be on the Thor SOC. And that's something that's been specially designed for this. And it's not super relevant for data centers. So no, we're not seeing a direct competition for that right now. Okay. You know, going back to regulation, you know, you're running in Texas. You can't run in California yet, right? They've got, I believe, a ban on deployment or testing of fully autonomous or autonomous trucks, right? Anything over like 10,000 pounds or tons or something. How much of your road map is dictated by regulation versus the technology? So today, you know, obviously our road map is constrained by regulation. We don't break the law, but we operate places we're allowed to. And that's the vast majority of states in the US. And we can build a heck of a business if the law just stays the way it is today. We estimate in the Sun Belt there's 50 billion vehicle miles traveled. And that's mileage that we can go and support and support our customers and grow our business in. What's exciting though is that it's not just customers who are seeing the benefit and value, but we're seeing at a regulatory level and at a policy making level really interesting this. And so we're seeing California. We expect regulations to be released for trucks on the road. And we expect that to happen in the next month or so. So that's really promising. And then we see at a federal level, real interest in there being a framework that would create a consistency of rules across the United States. I think that's an important step. It's one that we look forward to, but it's not one that today is limiting our business. Yeah, you can operate in all these other states. But California always seems like the Holy Grail. I mean, you've got the, it's got the shipping like from, you know, the ocean, right? Like you need that coast, I guess. I think another way to think about it. I think California is like the fourth largest economy in the world. Right. Right. And so for sure. And they're sunny. It's a sunny state. Right. Are you operating in rain conditions? We do operating the rain today. Yeah. Yeah. Okay. So you autonomously. Yes. Okay. Yep. How long? How long ago did you start doing that? We started in April last year operating in the daytime and good weather just between Dallas and Houston. Over the course of the last year, we unlocked kind of what the, what we call operation design domain increases, right? So new place we could operate. So we started operating at night. I think in the middle of last year, we started operating to El Paso towards the end of last year. And I think in January, we unlocked operating in the rain. So yeah. What are your customers most excited about? Like is it that you can kind of run all night and just do more, like, or do more runs essentially? It's a combination of things. So for all of these companies, safety is top of mind. Right. It is, you know, one, two, and three. I think when they, when they talked to us. And so the opportunity to have a driver that is always vigilant, that never gets distracted, that is able to look three hundred six degrees and the safety implications of that, I think are real. And that's not just good for customers. Of course, that's good for us as a society. Beyond that, the ability to utilize the truck for so yet if You're a company that runs trucks. You spend $150,000 to $200,000 for a truck, and you can only operate it half the time because we very rationally limit people to driving a truck 11 hours a day. This allows us to take that truck and use it twice as much. And so if you think again, at any business where you have a major capital expenditure, the more you can utilize it, the better it is for your business. - Yeah. - They're also excited about the sustainability benefit. So we've done studies that were-- - Do we care about sustainability today? - I certainly do. - I mean, I certainly do, but it seems like it's gone out of both. - But I think you can both think of it as I do as both a societal good that we want the world to be a cleaner place. We want it there for our kids and grandkids. But also you can look at the bottom line, right? A truck that consumes less fuel is lower cost to operate and emits less greenhouse gas and other pollutants, right? And so I think it's a win on both sides. And we expected to be 14 and 34%. And we're actually seeing that in practice in our fleet. So it's, again, another big deal, particularly with our current set of adventures and the impact that has on the diesel fuel. - Right. - It's a big win. - We're talking about safety, we're talking about regulation. I remember last year, a couple of years ago, there was a whole hullabaloo about safety cones. Do you remember this? You guys, I think, filed a complaint against-- - Warning triangles. - Yes, yes, warning triangles. So for those who don't know, so much of road safety is obviously designed around humans. So when a truck pulls over on a highway, the drivers went to get out and put out safety cones at certain distances so that cars coming down the road can see this. Now, if you have a driverless truck, obviously a human can't get out and do that, right? So that causes problems. But I imagine it's like a weird little regulatory thing that would have held you back from being able to do fully driverless deployments. What's the update on that? Because I covered that when it happened, and I want to tell you, when I was surprised that this story got so many clicks, like so many people were interested in that. - It's interesting, right? And I think what's neat about is it translates something that a lot of people don't have direct contact with, right? Or like AI and machine learning, verifiable AI. And translate something very real, right? A person having to walk down the freeway and put a cone down. - It's very loud. - It is. And it's grooving here at the Manhattan. - You did a great job paying attention to your own thoughts. Well, that's happening in the background. - It's pretty fun. You know, it's, you know, it's actually a little difficulty on the podcast. But you talk about it being this idea of dropping these triangles as a safety thing. What turns out there's no data that shows that, all right? It was a rule that was put in place in the '70s. And then I think it was in the '90s. And it's actually studied, should we require that for passenger vehicles? And they said, oh, actually, there's no reason, like there's no evidence this is actually helpful. And so our approach has been like, how do other vehicles on the road tell drivers to look out? - Yeah. - You know, what does a police car do? What does a emergency vehicle do? What does a tow truck do? A construction vehicle? They all turn on flashing lights. And we're like, hey, we should turn on flashing lights and warn people. And it felt very common sense. And it turns out at this point that the administration, the Department of Transportation has come back and said, they'll give us a waiver to study this for a few months. And then that would potentially lead to an exemption that would last for five years. And we've been excited to have a good conversation with the federal regulators around this. And they're smart people and they're listening to common sense and things are moving forward. - I'm zooming back out a little bit. So you were part of WEMO here in San Francisco is everywhere. They had, I don't like a similar, but opposite trajectory to you guys, right? Both started thinking, okay, we're gonna do highway, autonomous vehicles, autonomous trucks and robot taxis. WEMO ditched the trucking, stuck with the robot taxis. You guys have not ditched the robot taxis. It'll come back, I'm sure. But now you're focused fully on trucking. What made you made that decision? - Yeah, like you say, we started wanting, and we continue to believe we're building a driver that can drive all kinds of different things. But you have to pick somewhere. And for us, we picked trucking. We did that for two big reasons. So one is, we didn't start with trucking because we didn't think you could solve the problem. When you drive a truck, you have to look a long way down the road. And we're big believers that you use a combination of sensors to do that robustly. Laser radar and camera. And there was no laser sensor that could see far enough. - You say laser, do you mean lighter? - I mean, lighter. - You guys have your own lighter, right? - We do. In fact, this is a big part of why we can do trucking. Is that when I started the company, I'd been at what's now Waymo. And we'd come to the conclusion, you couldn't see far enough with lasers or with light R to drive safely a freeway speeds. When we started a roar, we thought, okay, how do we solve this? We spent a bunch of time trying to find a technology to do it. We found this great company in Bozeman, Montana, of all places, acquired them, brought them in. And that turned into what's now first light. Which is this special kind of light R that because the way we do the measurement can see way further than a conventional light R can. And so that technological unlock was one that said, okay, we can now go after trucking. And then as a business, we just see trucking as a better opportunity to start with. It's a trillion dollar market in the US, whereas ride hailing is a $50 billion market. And so, it's gigantic to go and work with. There's a real need on the safety front. 500,000 collisions with tractor trailers, were trucks every year and 5,000 fatalities. And the economics, the unit economics are stronger that we value a truck being driven, three times the value of a Uber being driven. And so when you bring a new technology to market, having something where you can be profitable sooner and where you have a huge market to grow into, that's just ready for it, just felt like a lot of make common sense to us. It reminds me of, I mean, and you partly reminded me of this before we started recording. You were talking about serve robotics and do the little sidewalk robots. We've had their CEO on equity before Ali. And there's some serve robots here. - Yeah. - Running around handing out little electrolyte packets, which is exactly what I-- - Oh, is that what they are? - When they were something there, I was like, "It's exactly what I wanted." - I thought they were lucky tags. I didn't look closely, okay? - Yeah, they were still very delicious. - Excellent. - One thing that Ali said was, yeah, the business is not delivering food. Like that's not where I want to end. But it's a business that makes money today and it allows me to get the data I need for real world use cases of autonomy. I'm curious, what is your end goal? Is it trucking? Is it autonomy at large? Like what's your why? - Yes, so what I'd say is if all we do is, what we're doing with trucking, it will be, the company will be incredibly successful and be very proud of what we've done. That said, what we're building is really the Vanguard of physical AI. - Yeah. - There's ability to understand the physical world and to be able to have something interact with it safely. And so that capability is something where we think we can apply it to a lot of things. And whether it is other elements of logistics, so not just big trucks, but box trucks, or whether it's in adjacencies like, personal ability in robotaxies, or whether it's in mining or agriculture, or aerial drones or humanoid robotics. In a couple of years, you'll start to see as explore where we can take this muscle that we've built and apply it at other places. And so it'll be fun to see, I'm not sure exactly which of those places it goes, but what I do know is we're gonna be very uniquely positioned, not just with the capability we've built internally, but with the scale that will come from the trucking business and the cost down that comes with that scale. And so I think it's gonna be, it's gonna be a lot of fun for the next few years. >> Yeah, well, what have you learned in your however many decades, what, how long has it been? >> Oh, good gosh. It's been 24 years, something like that. >> Right, so in your 24 years of bringing AI from the lab to the physical world, what have you learned that other startup founders can take away? >> Yeah. >> And I say that about yourself, right? 'Cause if you do another business unit, you're essentially a startup founder again. >> I will not be doing another company. This is one public company that's good for me. I'm not that ambitious, I guess. But do intend to be a business company for the next century, right? I think one of them is that there's just no shortcuts. We believe this from day one in that if we are building something that people's lives depend on, right? If you're driving a 70,000 pound thing down the freeway, you need to know that it works and that you build that trust slowly over time. And so we have to be committed to that. We need to be thinking safety first. We need to be building the process and tools that allows to do that. And then grow from there. And it's baked right into our mission, which is deliver the benefit to self-trafficking technology safely, quickly and broadly. Do it safely, move as quickly as we can, and then think about scale. I think that is really served as well. And it's an important part of how I believe we're gonna be here for a long term time doing exciting things. - Yeah, it's interesting how safety is so front and center for physical AI, less so in a lot of ways, or it feels like for not physical AI, like LLM, right? You said on stage, which I thought was interesting that, you have to focus on, that there's more safety implications for self-driving trucks than there are for LLM's. And I said, well, I don't know if I agree with that because as we've seen, there's been a lot of, - Mental health. Poor mental health outcomes that have happened as a result of LLMs and before that, social media. But why do you think, is it just because you're operating in a realm that already has rules of the road upon intended? - I think it's a little different than I. I think it's more obvious. - Right. - Right, that it's clear that there are risks when you are driving a car down the road or driving a truck down the road. The harm that can be done through an LLM is not as front and center. - It's not as like cause and effect. - And it's not as direct, right? And I certainly agree that we're observing harm and I'll have to think about how express this whole differently. But really when I think about an LLM, there at least is a person intervening between the interaction that happens on the screen and whatever physical interaction happens in the world. Right? I talk about the kind of gap that Gemini had back in the day. How do you keep the cheese on your pizza? You put glue, right? And we all look at that and say it silly and I don't expect many people would take that of face value and then glue cheese on their pizza. The equivalent and the truck is if for a moment it decides, geez, I should make a right turn in the middle of the freeway, there's no human who can intervene. And so yeah, I'll try and be more a little more careful how I talk about that distinction, but I do see one being there that's this material. Yeah. Well, do you have, I guess like when we talk about safety and for self-driving or the AI being safer than a human or a better driver than a human, like who gets to decide that and how, like is that something that you kind of self-certify or are you working closely with regulators or-- Yeah, at the end of the day society gets to decide it, right? We have elected officials, they create a regulatory environment, they create a policy environment. And that is the decision about how and where these things should be used. On a day-to-day basis, we make the decision whether we think the protocol pronoun the road is safe. And we do that using what we call a safety case. And you can think of this as an explanation for why we think it's safe. And it has five core pillars. The first is that the vehicle has to be proficient. So that means that it drives safely, right? That it behaves the way you'd expect. The second is it has to be fail safe. So it has to understand if something breaks. And then figure out how to mitigate risk associated with that and be safe. The third is that it has to be resilient. So we have to think about how it might be misused or think about how a cyber attack might impact it and make sure that we're thoughtful about that and responsive to that. The fourth is that we need to be continuously improving. So we need to learn from our experience, whether it's what we see from others out in the world or we see from our system behaving on the world or our company operating and be constantly making things better. And then the last is that we need to be trustworthy and just kind of on the face value. If I tell you the other four things and I'm not trustworthy, it doesn't mean a whole lot. But it's also about how we engage with regulators and policymakers to make sure they're informed and can make informed decisions. And so we take those five pillars and we blow that out into about 450, something like that, bits of evidence, right? These are things where we said, this is what it means, this is why. And then when we get all that checked off, we say, okay, yeah, we feel good that this is not creating unreasonable risk on the road and that, you know, frankly, I would feel comfortable with my family on the road around it and that I feel comfortable that we're not putting other people at unreasonable risk. - How much is that like that feeling of safety based on your AI approach, right? Like so there's a debate I think in the industry right now between like end to end systems and then more structured, verify approaches. I think that's what Aurora is doing. So why does that matter more for safety? And for context, I believe, you know, a company like Wabi, one of your competitors, they're doing more of an end to end approach. And yeah, so I'm curious what you think about how that applies to safety. - I think it's one part of it, right? And because as I said, there's a lot of things around safety and getting the software right is obviously an important part of it, but it's one part of it. For us, we've taken this approach of verifiable AI and what that means is that we appropriately decompose the problem so that we can understand how the system's behaving. So we can have conviction that it's actually working well and it is safe. And so, you know, one of the major ways we decompose is we take an understanding of the world. So what's moving around us and in the world around us? And how do we react to that? And by decomposing that, we can actually look at the things that matter and make sure that we're understanding the world correctly. Because if you just have an amorphous blob and a single system, I don't know why I'm making the decision I am. Is it because I made a bad decision, but I understood what was going on or is it that I didn't understand the world or is it some weird interaction between the two? Or something else altogether. And so we think breaking them apart allows you to actually better understand the system and make the testing, the verification of validation actually possible. And someone would say, okay, well, you don't need to break it apart. You can just have, you know, kind of what they call heads, right? These versions, you know, these outputs from the system that tell you those things. The challenge with that is that we're seeing in these reasoning models that the system is complicated enough that it can lie or do two things at once, right? So if you take one of these reasoning models, it'll tell you an explanation for why it did what it did. - Right. - What it's really doing is telling you an explanation for what it did what it did that it thinks you would like. - Yes. - Right? - And that may not actually be, yeah, and I don't even know, I don't want to necessarily ascribe a personality to it, whether it's DB's or not, but it's learned that that's the right way to explain this thing. And then it may be coming to that answer in a very different way. And so when you think about verifying our system, we want to say, okay, let's not have that be a risk. - Right. - That it's telling us that it sees the red car there, and so it's reacting in this way. But in practice, it's doing something else, right? And the complexity of these networks are such that it could very well be. And so for us, this decomposition allows us to understand it. And then we are allowed to, are able to explicitly express constraints and guardrails. So we don't just have to tell it, please don't. We can actually put it in a box and say, it can't. - Right. - And for a safety critical system, that seems really important to me. I think what you're starting to see with the LLMs is that as we are both moving to domains where the implications of bad actions are more serious. And to your point, we're starting to understand more of the consequences that may happen, even in what felt like innocuous domains. We're seeing architectures that look more and more like what we've been doing for a while. Because it's kind of the way to limit the machine. - Zooming out really quick, I would love to hear, I mean, you're probably so heads down on Aurora right now. But what are some other companies in the autonomy space that are exciting to you? - Yeah, there's a couple that I get a chance to check in on everyone's while. So one of them is server-bottles, and you mentioned earlier, I think it's just cool to see. Like there's this wave of everybody thought, "Oh, we'll drive on sidewalks, and that'll be way easier." And it turns out it's a technically pretty hard, you know, interacting with people is not easy. And then building a real business there has been hard as well. And so to see a company that's got these things and the business seems to be starting to work, it's kind of fun and cool. And then the other one that I think is pretty cool is this company called Bedrock Robotics. And they are doing automation for excavation and kind of at the construction site. - Yeah. - And that one, there's a couple of people that founded it I've known for a long time and work within the past and they're good people. And it's just kind of cool. I'm still, like this part of me, is still like the eight-year-old boy who's like, big truck, cool, excavator, cool, right? And so seeing something where they are, both working with a physical world and doing something that actually is useful. And like seeing the technology come to work and seeing it with good people doing it, I love that. - You know, you're talking about construction sites and it's reminding me of a company that I saw, I think we wrote about them at TechCrunch maybe last year or the year before, I can't remember. But their whole thing was we wanna help construction sites be more organized. So we're gonna put cameras on all the hard hats and with that data will help you be more organized and whatever. Are there still these, is there still a priority on getting real world data or have the advancements and simulation kind of made it so that it's a little bit obsolete to get out in the field and do that? - So our approach I can speak to is you need to do both. - Right. - Right, we have some incredible simulation capabilities. But you have to ground the simulation, meaning you have to know is this actually, is my simulation accurately reflecting the real world? - Yeah. - Otherwise, you're playing a video game. - Yeah. - And you may be really good at the video game, but that, you know, anyone who's driven a Mario Kart knows that's quite different than driving a real car. - Yeah. - And of course, the simulations are much more closer than Mario Kart, but if you haven't grounded it, you don't know how close they are. And so for us, there's a set of things that we know, our simulation is very good at and we understand the limitations and then there's other things where we go to a test track or we gather data from the real world and we use that both to improve the simulation capability, but also because today we don't trust that particular part of the simulation for that particular test. And so. It's an "and" not an "or" in my mind. - Got it. Okay, well, we're just about out of time. Thank you so much for joining. Is there anywhere that our listeners can find us or find you online? - Yeah, please, we're at Aurora.tech. And even better yet, if you actually want to see our trucks in action, you can go to youtube.com, Adderora Driver, and we live stream our driver's trucks every day for I think eight to five central times. So check it out. It's a warning, it is super boring, but we have some good background music. - Yeah, oh, look really, it's not. - Honestly, that's what we want, right? If you're catching an exciting moment in a big truck, like we just don't want that, right? We wanted to be boring and smooth and easy. - Yeah, awesome. Okay, well, you can find me, I'm on Twitter, I'm on LinkedIn, I'm on Blue Sky, I'm on Substack. And you can find equity on Twitter and Blue Sky at EquityPod, talk to you next time. Equity is hosted by TechCrunch Senior Reporters and produced by Teresa Lo Consolo with Editing by Cal. Subscribe on YouTube or wherever you get your podcasts and find out what's next at techcrunch.com/events. Thanks so much for listening and we'll talk to you next time.