Skip to content
Lex Fridman PodcastLex Fridman Podcast

Chris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA | Lex Fridman Podcast #28

Lex Fridman and Chris Urmson on chris Urmson on Safely Scaling Real-World Self-Driving Car Technology.

Lex FridmanhostChris Urmsonguest
Jul 22, 201944mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:00 – 3:04

    DARPA Grand Challenge mindset: believing the “impossible” is doable

    1. LF

      The following is a conversation with Chris Urmson. He was a CTO of the Google self-driving car team, a key engineer and leader behind the Carnegie Mellon University autonomous vehicle entries in the DARPA Grand Challenges, and the winner of the DARPA Urban Challenge. Today, he's the CEO of Aurora Innovation, an autonomous vehicle software company he started with Sterling Anderson, who was the former director of Tesla Autopilot, and Drew Bagnell, Uber's former autonomy and perception lead. Chris is one of the top roboticists and autonomous vehicle experts in the world, and a longtime voice of reason in a space that is shrouded in both mystery and hype. He both acknowledges the incredible challenges involved in solving the problem of autonomous driving and is working hard to solve it. This is the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter at Lex Fridman, spelled F-R-I-D-M-A-N. And now, here's my conversation with Chris Urmson. You were part of both the DARPA Grand Challenge and the DARPA Urban Challenge teams at CMU with, uh, Red Whittaker. What technical or philosophical things have you learned from these races?

    2. CU

      I think the, the high order bit was that it could be done. I think that was the thing that was incredible about the, first, the, the grand challenges.

    3. LF

      Mm-hmm.

    4. CU

      That I remember, you know, I was a grad student at Carnegie Mellon and there we, it was kind of this dichotomy of it seemed really hard, so that would be cool and interesting. But, you know, at the time, we were the only robotics institute around. And so, you know, if we went into it and fell on our faces, that would, that would be-

    5. LF

      Okay.

    6. CU

      ... embarrassing. Uh, so I think, you know, just having the will to go do it, to try to do this thing that at the time was marked as, you know, darn near impossible. And, and then after a couple of tries, be able to actually make it happen, I think that was, you know, that was really exciting.

    7. LF

      But at which point did you believe it was possible? Did you from the very beginning? Did you personally? 'Cause you were one of the lead engineer, you actually had to do a lot of the work.

    8. CU

      Yeah, I was the technical director there and did a lot of the work (laughs) , uh, along with a bunch of other really good people. Did I believe it could be done? Yeah, of course. Right? Like, why would you go do something you thought was impossible, completely impossible? Uh, we thought it was going to be hard. We didn't know how we were gonna be able to do it. We didn't know if we'd be able to do it the first time. (laughs) Turns out we couldn't. That, yeah, I guess you have to. I th- I think there's a certain benefit to naivete, right? That if you don't know how hard something really is, you, you try different things and, you know, it gives you an opportunity that others who are, you know, wiser maybe don't, don't have.

  2. 3:04 – 5:12

    What made early autonomy hard: end-to-end engineering and unclear requirements

    1. LF

      Well, what were the biggest pain points? Mechanical, sensors, hardware, software, algorithms for mapping, localization, uh, just general perception control? What, like hardware/software, first of all.

    2. CU

      Well, I, I, I think that's the joy of this field, is that it's all hard.

    3. LF

      (laughs) Yeah.

    4. CU

      Um, and that you have to be good at, at, at each part of it. So for the first, for the urban challenges, uh, if I look back at it from today, uh, it should be easy today. That, you know, it was a static world, there weren't other actors moving through it, th- is what that means. Uh, it was out in the desert, so you get really good GPS. You know, so that, that when, and, you know, we could map it roughly. And so in retrospect now, it's, you know, it's, it's within the realm of things we could do. Back then, just actually getting the vehicle and the, you know, there was a bunch of engineering work to get the vehicle so that we could control it and drive it.

    5. LF

      Mm-hmm.

    6. CU

      That's, you know, that's still a pain today, but it was even more so back then. Uh, and then the uncertainty of exactly what they wanted us to do was, was part of the challenge as well.

    7. LF

      Right, you didn't actually know the track heading in. You, you knew approximately, but you know, didn't actually know the route, the route that was gonna be taken.

    8. CU

      That's right. We didn't know the route. We didn't even r- really, the way the rules had been described, you had to kind of guess. So i- if you think back to that challenge, the idea was to, uh, that the, the government would give us, uh, DARPA would give us, uh, a set of waypoints and kind of the width, um, that you had to stay within between the line that went between, you know, each of those waypoints. And so the, the most devious thing they could have done is set, you know, a kilometer-wide corridor across, you know, a field of scrub brush, um, and rocks and said, you know, "Go figure it out." Uh, fortunately, it really, it turned into basically driving along a set of trails, which, you know, is much more relevant to, to the application they were looking for. But no, it was, it was a hell of a thing back in the day.

  3. 5:12 – 6:16

    Leadership lessons from Red Whittaker: empowering people to grow

    1. LF

      So, uh, the legend, Red, was, uh, kind of leading that effort-

    2. CU

      Yeah.

    3. LF

      ... uh, in terms of just broadly speaking. So you're a leader now. What have you learned from Red about leadership?

    4. CU

      I, I think there's a couple things. One is, you know, go and try those really hard things. That, that's where there is a, an incredible opportunity. Uh, I think the other big one though is to see people for who they can be, not who they are. It, it's one of the things that I actually, one of the deepest lessons I, I learned from Red was that he would look at, um, you know, undergraduates or graduate students and empower them, uh, to be leaders-

    5. LF

      Mm-hmm.

    6. CU

      ... to, to, you know, have responsibility, to do great things, that I think...... another person might look at them and think, "Oh, well, that's just, you know, an undergraduate student. What, what could they know?" And so I think that, that, you know, kind of trust but verify, have confidence in what people can become, I think is, is a really powerful thing.

  4. 6:16 – 9:23

    Technical evolution: HD maps, multi-beam LiDAR, and Bayesian estimation

    1. LF

      So through that, let's just, like, fast-forward through the history. Can you maybe talk through the technical evolution of autonomous vehicle systems from the first two Grand Challenges to the Urban Challenge to today?

    2. CU

      Sure.

    3. LF

      Are there major shifts in your mind or is it the same kind of technology, just made more robust?

    4. CU

      I think there's been some big, big steps. So the, for the Grand Challenge, the real technology that unlocked that was HD mapping.

    5. LF

      Mm-hmm.

    6. CU

      Prior to that, a lot of the off-road robotics work had been done without any real prior model of what the vehicle was going to encounter. And so that innovation, that... The fact that we could get, you know, decimeter resolution models was really a, a big deal. And, and that allowed us to, to kind of bound the complexity of the driving problem the vehicle had and allowed it to operate at speed, because we could assume things about the environment that it was going to encounter. So that was a, that was one of the... That was the big step there. For the Urban Challenge, you know, one of the big technological innovations there was the multi-beam LIDAR, and being able to generate, um, high-resolution, you know, mid-to-long range 3D models of the world and use that for, you know, f- for understanding the world around the vehicle. And that was really a, you know, kind of a game-changing technology. In parallel with that, we saw a bunch of other technologies that had been kind of converging, half their, their day in the sun. So, uh, Bayesian estimation-

    7. LF

      Mm-hmm.

    8. CU

      ... uh, had been... You know, SLAM had been a big field i- in robotics. You know, you, you would go to a conference, you know, a couple of years before that and, you know, every paper would effectively have SLAM somewhere in it.

    9. LF

      Mm-hmm.

    10. CU

      And so seeing that, you know, the, that... those Bayesian estimation techniques, you know, play out on a very visible stage, you know, I thought that was, that was pretty exciting to see.

    11. LF

      And most of SLAM was done based on LIDAR at that time?

    12. CU

      Uh, well, yeah.

    13. LF

      So-

    14. CU

      And in fact, we weren't really doing SLAM per se, be- you know, it... You know, in real time, because we had a model ahead of time, we had a roadmap. But we were doing localization.

    15. LF

      Mm-hmm.

    16. CU

      And we're using, you know, the LIDAR or the cameras, depending on, you know, who exactly was doing it to, to localize to a model of the world. And I thought that was, uh, that was a big step from, kind of naively trusting GPSINS before that. And, and again, like, lots of work had been going on in this field. Uh, certainly this was not doing anything particularly innovative i- in SLAM, uh, or in localization, but it was seeing that technology necessary in a real application on a big stage I thought was, was very cool.

  5. 9:23 – 10:44

    Maps and localization reality check: datums, coordinate frames, and centimeter accuracy

    1. LF

      So for the Urban Challenge, there was already maps constructed offline-

    2. CU

      Yes.

    3. LF

      ... in general? Okay. And, uh, did people do that indivi- did individual teams do it individually, uh, so they had their own difference of different approaches there? Or did everybody kind of share that information, at least intuitively?

    4. CU

      So, so the... DARPA gave all the teams a, a model of the world, a, you know, a, a map. And then, you know, one of the things that we had to figure out back then was, uh... And it's still one of these things that trips people up today, is actually the coordinate system. Uh, so you get a latitude, longitude, and, you know, to so many decimal places, you don't really care about kind of the ellipsoid of the Earth that's being used. Uh, but when you want to get to 10 centimeter or centimeter resolution, uh, you care whether the, the coordinate system is, you know, NADS83 or WGS84 or... You know, these are different ways to describe both the, the kind of non-sphericalness of the Earth but also, um, uh, kind of the act- i- i- in, I think, I can't remember which one, the tectonic shifts that are happening and how to transform, you know, the, the global datum as a function of that. So yeah, getting a map and then actually matching it to reality to a centimeter resolution, that was kind of interesting and fun back then.

  6. 10:44 – 11:56

    Urban Challenge perception: tracking, prediction, and interactive behavior

    1. LF

      So how much work was the perception doing there? Uh, so how, uh, how, how much were you relying on localization based on maps without using perception to register to the, the maps and how... I, I guess the question is, how advanced was perception at that point?

    2. CU

      Yeah. It, it's certainly behind where we are today, right? We're, we're more than a decade since the Gr- uh, the Urban Challenge. But the, the core of it was there, that we were tracking vehicles, we had to do that at a 100-plus meter range because we had to merge with other traffic. We were using, uh, you know, Bayesian, again, Bayesian estimates for, for state of these vehicles. Um, we had to deal with a bunch of the problems that you, you think of today, of predicting what that... where that vehicle's going to be a few seconds into the future. We had to deal with the fact that, uh, there were multiple hypotheses for that, because a vehicle at an intersection might be going right or it might be going straight or it might be making a left turn. And we had to deal with the challenge of the fact that our behavior was going to impact the behavior of that other oper- operator.

    3. LF

      Mm-hmm.

    4. CU

      And, you know, we did a lot of that in relative ni- relatively naive ways, but it, it kind of worked.

    5. LF

      You still, still had to have some kind of solution.

    6. CU

      Yeah. Yeah.

  7. 11:56 – 14:01

    From controlled demos to reality: unpredictability, new actors, and huge scale

    1. LF

      That was... And so where does that... 10 years later, where does that take us today from that, uh, artificial city construction to real cities to the urban environment?

    2. CU

      Yeah. I think the, the biggest thing is that the, you know, the-... the actors are truly unpredictable, that, uh, most of the time, you know, the drivers on the road, the other road users are out there behaving well, but every once in a while, they're not. The variety of, of other vehicles is, you know, y- you have all of them. And-

    3. LF

      In terms of behavior or in terms of perception? Or both?

    4. CU

      Both. Uh, that we have, you know, back then, we didn't have to deal with cyclists, we didn't have to deal with pedestrians, didn't have to deal with traffic lights.

    5. LF

      Mm-hmm.

    6. CU

      You know, the scale over which that you have to operate is now, you know, is much larger than, you know, the air base that, that we were thinking about back then.

    7. LF

      So what, uh... Easy question, what do you think is the hardest part about driving?

    8. CU

      Easy question.

    9. LF

      Yeah.

    10. CU

      Um-

    11. LF

      No, I'm, I'm, I'm joking. I, I'm, I'm sure no- nothing really jumps out at you as one thing, but, uh, in, in the jump from the urban challenge to the real world, is there something that's a particular y- you foresee as very serious difficult challenge?

    12. CU

      I think the, the most fundamental difference is that we're doing it for real, uh, uh, that in that environment, it was both a limited complexity environment, because certain actors weren't there, because, you know, the roads were maintained, there were barriers keeping people separate from, from robots at the time, and it only had to work for 60 miles-

    13. LF

      Right.

    14. CU

      ... which, looking at it from, you know, 2006, it had to work for 60 miles.

    15. LF

      Yeah.

    16. CU

      Right? Um, looking at it from now, um, you know, we, we want things that will go and drive for, you know, half a, half a million miles, and, you know, it's just a, it's a different game.

  8. 14:01 – 19:09

    Sensor fusion debate: why LiDAR, cameras, and radar all matter (and cost tradeoffs)

    1. LF

      So how important... You said Lidar came into the game early on, and it's really the primary driver of autonomous vehicles today as a sensor, so how important is the role of Lidar in the sensor suite in the near term?

    2. CU

      So I think it's, I think it's essential. You know, I believe, and- but I also believe that cameras are essential and I believe that radar is, is essential. I think that you, you really need to use the composition of data from, from these different sensors if you want the thing to, to really be robust.

    3. LF

      The question I wanna ask, let's see if we can untangle it, is, uh, what are your thoughts on the, uh, Elon Musk provocative statement that Lidar is a crutch, that, uh, is a kind of, um, I guess, growing pains-

    4. CU

      Yeah.

    5. LF

      ... and that much of the perception task can be done with cameras?

    6. CU

      So I think it is undeniable that people walk around without, you know, lasers in their foreheads, uh, and they can get into vehicles and drive them, and, and so there's an existence proof that you can drive using, you know, passive vision, no doubt, can't argue with that.

    7. LF

      In terms of sensors, yeah, so there's proof that-

    8. CU

      Yeah, in terms of sensors, right? So like, there's, there's an example that, you know, we all go do it, uh, many of us every day. In terms of, uh, Lidar being a crutch, sure.

    9. LF

      (laughs) .

    10. CU

      But, but, you know, in the same way that, uh, you know, the combustion engine was a crutch on the path to an electric vehicle, in the same way that, you know, any technology ultimately gets replaced by some superior technology in the future. Uh, and really, what, the way that I look at this is that the way we get around on the ground, the way that we use transportation is broken, um-

    11. LF

      Hmm.

    12. CU

      ... and that we have, you know, this, this, you know, what was... I think the number I saw this morning, 37,000 Americans killed, uh, last year on our roads, and that's just not acceptable. And so te- any technology that we can bring to bear that accelerates the, this techno- you know, self-driving technology coming to market and saving lives is technology we should be using. And it feels just arbitrary to say, "Well, you know, I'm, I'm not okay with using lasers because that's whatever, but I am okay with using an eight megapixel camera or a 16 megapixel camera." You know, like, it's just, these are just bits of technology and we should be taking the best technology from the tool bin that allows us to go and, you know, and solve a problem.

    13. LF

      The question I often talk to, uh, well, obviously you do as well, to, uh, the sort of automotive companies, and, you know, if, if there's one word that comes up more often than anything, it's cost and, and, uh-

    14. CU

      Yeah.

    15. LF

      ... tr- trying to drive cost down. So while it's, it's true that it's, um, it's a tragic number, the 37,000, the, the question is what... (laughs) And I'm not the one asking this question, 'cause I hate this question, but-

    16. CU

      Yeah.

    17. LF

      ... we, we ha- we want to find the cheapest sensor suite that, uh, that creates a safe vehicle.

    18. CU

      Yep.

    19. LF

      So in that, uh, uncomfortable trade-off, do you foresee Lidar, uh, coming down in cost in the future or do you see a day where level four autonomy is possible without Lidar?

    20. CU

      I, I see both of those, but it's really a matter of time. And I, and I think really maybe the- I, I would talk to the question you asked about, you know, the cheapest sensor.

    21. LF

      Mm-hmm.

    22. CU

      I don't think that's actually what you want. What you want is a sensor suite that is economically viable, and then after that, everything is about margin and driving cost out of the system. Uh, what you also want is a sensor suite that works.

    23. LF

      Right.

    24. CU

      And so it's great to tell a story about, um, how, you know, how it'd be better to have a self-driving system with a $50 sensor instead of a, you know, a $500 sensor. But if the $500 sensor makes it work and the $50 sensor doesn't work-... you know, who cares? So long as you, you can actually, uh, you know, have an economic oppor- you know, there's an economic opportunity there. And the economic opportunity is important because that's how you actually have a, a sustainable business, and, and that's how you can actually see this come to scale and, and, and be out in the world. And so when I look at LiDAR, I see a technology that has no underlying fundamentally, you know, expense to it, fundamental expense to it. It's, it's going to be more expensive than, uh, an imager because, you know, CMOS processes are, or, you know, fab processes are, are dramatically more scalable than mechanical processes, but we still should be able to drive cost down substantially on that side. Uh, and then I also do think that, uh, with the right business model, you can absorb more, you know, certainly more cost on the bill of materials.

  9. 19:09 – 24:25

    Level 2/3 autonomy and human factors: over-trust, marketing, and divergent tech paths

    1. LF

      Yeah, if the sensor suite works, extra value is provided, thereby you don't need to drive cost down to zero. It's the basic economics. You've talked about your intuition that level two autonomy is problematic because of the human factor, uh, of vigilance, decrement, complacency, over-trust and so on, just us being human.

    2. CU

      Yeah.

    3. LF

      We over-trust the system, we start doing even more, so partaking in the secondary activities like smartphone and so on. Have your views evolved on this point in either direction? Can you, can you speak to it?

    4. CU

      So... And I want to be really careful because sometimes this gets twisted in a way that's, that, that, that I certainly didn't intend. So active safety systems are a really important technology that we should be pursuing and integrating into vehicles. And there's an opportunity in the near term to reduce accidents, reduce fatalities and that's, uh, and we should be, we should be pushing on that. Level two systems are systems where the vehicle is controlling two axes, so, you know, breaking and s- breaking and throttle/steering. And I think there are variants of level two systems that are supporting the driver that absolutely, like, we should, we should encourage to be out there. Um, where I think there's a real challenge is in the, the human factors part around this and the misconception from the public around the capability set that that enables and the, and the trust that they should have in it. And that is where I, you know, I, I kind of, I, I, I'm actually incrementally more, you know, concerned around level three systems and, you know, how exactly a level two system is marketed and delivered, uh, and, you know, how people... how much effort people have put into those human factors. So I still believe, uh, uh, several things around this. One is people will over-trust the technology. Uh, we've seen over the last few weeks, you know, a spade of people sleeping in their Tesla. You know, I watched an episode last night of, um, Trevor Noah-

    5. LF

      Mm-hmm.

    6. CU

      ... talking about this. And, you know, him, you know, this is a smart guy who's... has a lot of resources at his disposal, describing a Tesla as a self-driving car, and that, why shouldn't people be sleeping in their Tesla? It's like, well, because it's not a self-driving car and it is not intended to be, and, you know, these people will almost certainly, you know, die at some point or, or hurt other people. And so we, we need to really be thoughtful about how that technology is described and, and brought to market. I also think that because of the economic issue, you know, econom- economic challenges we were just talking about, that that technology path will alt- the- these level two driver assistance systems, that technology path will diverge from the technology path that we need to be on to actually deliver truly self-driving vehicles, ones where you can get in it and sleep and have the, uh, equivalent or better safety than, you know, a, a human driver behind the wheel. Um, because the... again, the economics are very different in those two worlds. And so that leads to, you know, divergent technology.

    7. LF

      So you, you just don't see the economics of gradually increasing from level two and doing so quickly enough to where it doesn't cause safety, uh, critical safety concerns? You, you believe that, that it needs to diverge at this point, uh, into different-

    8. CU

      Yeah.

    9. LF

      ... ba- basically different routes of-

    10. CU

      And, and really that comes back to what are those L2 and L1 systems doing? And, and they are driver assistance functions where the, the, the people that are marketing that responsibly are being very clear and putting human factors in place, such that the driver is actually responsible for the vehicle and that the technology is there to support the driver. And the safety cases that are, are built around those are dependent on that driver attention, uh, and attentiveness. Uh, and at that point, you, you can kind of give up, to some degree... For economic reasons, you can give up on, say, false negatives. Uh, and so... And the, and the way to think about this is for a forward collision mitigation braking system, if it... half the times the driver missed a vehicle in front of it, uh, it hit the brakes and brought the vehicle to a stop, that would be an incredible, incredible advance in, in safety on our roads, right? That would be equivalent to seat belts. Uh, but it would mean that if that vehicle wasn't being monitored, it would hit one out of two cars. And so economically, that's a perfectly good solution for a driver assistance system. What you should do at that point, if you can get it to work 50% of the time, is drive the cost out of that so you can get it on as many vehicles as possible. But driving the cost out of it doesn't drive up-... performance on the false negative case. And so you'll continue to not have a technology that could, you know, really be available for, for a self-driven vehicle.

  10. 24:25 – 27:30

    Why even perfect communication won’t stop over-trust: experience beats statistics

    1. LF

      So, clearly, the communication, and this probably applies to all four vehicles as well, the, uh, marketing and the communication of what the technology is actually capable of, how hard it is, how easy it is, all that kind of stuff-

    2. CU

      Yeah.

    3. LF

      ... is highly problematic. Uh, so but, uh, say everybody in the world was perfectly communicated and were made to be completely aware of every single technology out there, what they, uh, what it's able to do. What's your intuition? And now we're maybe getting into philosophical ground. Is it possible to have a level two vehicle where we don't over-trust it?

    4. CU

      I don't think so.

    5. LF

      Right.

    6. CU

      If people truly understood the risks, uh-

    7. LF

      They wouldn't (overlapped) -

    8. CU

      ... and, and internalized it-

    9. LF

      Okay.

    10. CU

      ... then, then sure, you could do that safely. But that, that's a world that doesn't exist.

    11. LF

      Right.

    12. CU

      That people are going to, they're gonna, you know, uh, uh, uh, if the facts are put in front of them, they're gonna then combine that with their experience. And, you know, let's say they're, they're using an L2 system and they go up and down the 101 every day, and they do that for a month.

    13. LF

      Mm-hmm.

    14. CU

      And it just worked every day for a month. Um, like, that's pretty compelling. At that point, you know, just even if you know the statistics, you're like, "Well, I don't know, maybe there's something a little funny about those. Maybe they're, you know, driving in difficult places." Like, I've seen it with my own eyes, it works.

    15. LF

      Right.

    16. CU

      And the problem is that that sample size that they have, so it's 30 miles up and down, so 60 miles times 30 days, so 60, 180,000, 800 miles. That's, that's a drop in the bucket compared to the one, you know, what, 85 million miles between fatalities. And so they don't really have a true estimate based on their personal experience of, of the real risks. But they're going to trust it anyway, because it's hard not to. It worked for a month.

    17. LF

      Right.

    18. CU

      What's, what's going to change?

    19. LF

      So even if you start at perfect understanding of the system, your own experience will make it drift. I mean, that's a big concern-

    20. CU

      Yeah.

    21. LF

      ... over a year, over two years even. It doesn't have to be months.

    22. CU

      And, and I think that as this technology moves from what I would say is kind of the more technology-savvy ownership group to, you know, the mass market, you may be able to have some of those folks who are really familiar with the technology, they may be able to internalize it better and, uh, and, you know, your kind of immunization against this kind of false risk assessment might last longer. But as folks who are, who, who aren't as savvy about that, uh, you know, read the material and they compare that to their personal experience, I think there that, you know, it's, it's going to... It's gonna move more quickly.

    23. LF

      So your work, the program that you've created at Google and now at Aurora, um, is focused more on the second path-

    24. CU

      Mm-hmm.

    25. LF

      ... of creating full autonomy. So it's such a fascinating, uh, I think it's one of the most interesting AI problems of the century, right? It's, uh, I just talked to a lot of people, just regular people, I don't know, my mom-

    26. CU

      Yeah.

  11. 27:30 – 32:15

    Proving safety: evidence, process, simulation, testing, regulators, and better metrics

    1. LF

      ... about autonomous vehicles, and, uh, we, you begin to grapple with ideas of giving your life control over to a machine. It's philosophically interesting. It's practically interesting. So let's talk about safety. How do you think we demonstrate? You've spoken about metrics in the past.

    2. CU

      Yeah.

    3. LF

      How do you think we demonstrate to the world that an autonomous vehicle, an A-, an Aurora system is safe?

    4. CU

      Th- this is one where it's difficult because there isn't a soundbite answer.

    5. LF

      Right.

    6. CU

      That we have to show a combination of, uh, uh, work that was done diligently, uh, and thoughtfully, and this is where something like a, a functional safety process is part of that, is like, "Here's, here's the way we did the work. That means that we were very thorough. So, you know, if you believe that we, what we said about this is the way we did it, then you can have some confidence that we were thorough in, in, in the engineering work we put into the system." Uh, and then on top of that, the, you know, to kind of demonstrate that we weren't just thorough, we were actually good at what we did, there'll be a kind of a collection of evidence, uh, in terms of demonstrating that the capabilities work the way we thought they did, you know, statistically and, and to whatever degree we can, we can demonstrate that. Both in some combination of simulation, some combination of, of unit testing and decomposition testing. And then some part of it will be on-road data. And, and I think the, the way we're, we'll ultimately convey this to the public is there'll be clearly some conversation with the public about it, but we'll, you know, kind of invoke the, the kind of the trusted nodes, and that we'll spend more time being able to go into more depth with folks like, like NHTSA-

    7. LF

      Mm-hmm.

    8. CU

      ... uh, and other federal and state regulatory bodies. And kind of given that they are operating in the public interest and they're trusted, that if we can, you know, show enough work to them that they're convinced, then, you know, I think we're in a, in a pretty good place.

    9. LF

      That means you work with people that are e- essentially experts at safety to try to discuss and show. Do you think, um... The answer's probably no, but just in case, do you think there exists a metric? So currently, people have been using number of disengagements.

    10. CU

      Yeah.

    11. LF

      And it quickly turns into a marketing scheme to, to sort of, you, um, alter the experiments you run to adjust... I think you've spoken that you don't like n- uh-

    12. CU

      Don't love it. No, in fact, I, I was on the record telling DMV that I thought this was not a great metric.

    13. LF

      Yeah. Do you think it's possible to create a metric, a number, that, um, that could demonstrate safety outside of fatalities?

    14. CU

      So, so I, I do. And, and I think that it, it won't be just one number.

    15. LF

      Mm-hmm.

    16. CU

      So as we are internally grappling with this, and, and at some point we'll be, we'll be able to talk pub- more publicly about it, is how do we think about human performance in, in different tasks? Say, detecting traffic lights or, um, safely making a left turn across traffic, and what do we think the failure rates are for those different capabilities for people? And then demonstrating to ourselves and then ultimately, uh, folks in regulatory role and, and then ultimately the public, that we have confidence that our system will work better than that. Uh, and so these, these individual metrics will kind of tell an- a com- a compelling story, ultimately. I do think, at the end of the day, what we care about i- in terms of safety is, uh, lives saved, uh, and injuries reduced, and then, and then ultimately, you know, kind of casualty dollars that people aren't having to pay to, to get their car fixed. And I do think that you can... You know, w- i- in aviation they look at a, a kind of an event pyramid, where, you know, a c- a crash is at the top of that and that's the worst event, obviously, and then there's injuries and, you know, near miss events and whatnot, and, and, you know, violation of operating procedures. And, and you kind of build a, a statistical model of, of, uh, the relevance of the, the low severity things to the high severity things, and I think that's something we will be able to look at as well. Uh, because, you know, an event per 85 million miles is a- you know, statistically a, a difficult thing, even at the scale of the US, um, to, to, to kind of compare directly.

    17. LF

      And that event, the fatality that's connected to an autonomous vehicle, is significantly, at least currently, magnified-

    18. CU

      Yeah.

    19. LF

      ...in, uh, th- the amount of, um, attention it gets. So that speaks to public perception.

    20. CU

      Yeah.

  12. 32:15 – 34:38

    Winning public trust: let people ride in it until it becomes mundane

    1. LF

      I think the most popular topic about autonomous vehicles, in the public, is, um, the trolley problem formulation, right?

    2. CU

      Sure.

    3. LF

      Which, uh, has... Let's not get into that too much, but, uh, is misguided b- in many ways. But it speaks to the fact that people are grappling with this idea of giving control over to a machine. So how do you win the, the hearts and minds of the people that, uh, autonomy is something that could be a part of their lives?

    4. CU

      I think you let them experience it.

    5. LF

      Oh.

    6. CU

      Right? I, I think it's... I think, I think it's right. I think people should be skeptical. Uh, I think people should, um, ask questions. I think they should doubt. Because this is something new and different. They haven't touched it yet, and I think that's perfectly reasonable. And... But at the same time, it's clear there's an opportunity to make the roads safer. It's clear that we can improve access to mobility. It's clear that we can reduce the cost of mobility. And that once people try that and are... You know, understand that it's safe, and, and are able to use it in their daily lives, I think it's one of these things that will, will just be obvious. And, and I've seen this practically in... You know, in demonstrations that I've, you know, given, where I've had people come in and, you know, they're very skeptical and they, they get in the vehicle. And, you know, my favorite one is taking somebody out on the freeway, and we're on the 101 driving at 65 miles an hour, and after 10 minutes they, they kind of turn and ask, "Is that all it does?" (laughs) And you're like-

    7. LF

      Yeah.

    8. CU

      ..."It's a self-driving car. I'm not sure exactly what you thought it would do."

    9. LF

      Yeah, it just drives. Yeah.

    10. CU

      Right? Um, but they, you know, they, they... It becomes mundane-

    11. LF

      Right.

    12. CU

      ...which is, which is exactly what you want a technology like this t- to be, right? We don't really... When I turn the light switch on in here, I don't think about the complexity of, you know, the-

    13. LF

      Yeah.

    14. CU

      ... those electrons, you know, being pushed down a wire from wherever it was and being generated some- Like, it's just... It's like I just get annoyed if it doesn't work, right? And, and what I value is the fact that I can do other things in this space. I can, you know, see my colleagues, I can read stuff on a paper, I can, you know, uh, not be afraid of the dark.

    15. LF

      Perf-

    16. CU

      And, and I think that's what we want this technology to be like, is it's, it's in the background and people get to have those, those life experiences and, and do so safely.

  13. 34:38 – 38:24

    Deployment and scaling: driverless “zero-to-one,” urban first, and timeline expectations

    1. LF

      So putting this technology in the hands of people speaks to s- c- scale of deployment, right? So what do you think... The, uh, the dreaded question about the future, because nobody can predict the future.

    2. CU

      Yeah.

    3. LF

      But just maybe, uh, speak poetically about when do you think we'll see a large-scale deployment of autonomous vehicles? 10,000, th- those kinds of numbers.

    4. CU

      U- We'll see that within 10 years. I'm, I'm pretty confid-... I... We, um...

    5. LF

      What's an impressive scale? W- what moment... Uh, so you've, you've done the DARPA Challenge with this one vehicle.

    6. CU

      Yeah.

    7. LF

      At which moment does it become, "Wow, this is serious," scale?

    8. CU

      So, so I think the moment it gets serious is when we really do have a driverless vehicle operating on public roads, and that we can do that kind of continuously.

    9. LF

      Without a safety driver.

    10. CU

      Without a safety driver in the vehicle. I think at that moment, we've, we've kind of crossed the zero to one threshold. And then it is about, uh, how do we continue to scale that? How do we build the right business models? How do we build the right customer experience around it so that it is actually, you know, a useful product out in the world?

    11. LF

      Right.

    12. CU

      And I think that is really... A- at that point, it moves from a, you know, what is this kind of mixed science/engineering project into engineering and commercialization, and really starting to deliver on the value that we all see here.... uh, and, you know, and actually making that real in the world.

    13. LF

      What do you think that deployment looks like? Where, where do we first see the inkling of, uh, no safety driver, one or two cars here and there? Is it on the highway? Is it in specific routes in the urban environment?

    14. CU

      I, I think it's going to be urban, suburban type environments. You know, with, with Aurora, when we, we thought about how to tackle this, I- it was kind of en vogue to think about trucking as opposed to urban driving. And, and, you know, the... Again, the human intuition around this is that freeways are easier-

    15. LF

      Mm-hmm.

    16. CU

      ... to drive on, because everybody's kind of going in the same direction and, you know, the lanes are a little wider, et cetera. And I think that that intuition is pretty good, except we don't really care about most of the time, we, we care about all of the time.

    17. LF

      Hmm.

    18. CU

      Uh, and when you're driving on a freeway with a truck, say, at 70, 70 miles an hour, uh, and you got 70,000 pound load with you, that's just an incredible amount of kinetic energy. Uh, and so when that goes wrong, it goes really wrong.

    19. LF

      Yeah.

    20. CU

      And that, um, those, those challenges that you see occur more rarely, so you don't get to learn as o- as quickly. Uh, and they're, you know, incrementally more difficult than urban driving, but they're not easier than urban driving. And so I think this happens in moderate speed, urban environments because there, you know, if, if two vehicles crash at 25 miles per hour, it's, it's not good, but probably everybody walks away. And those, those events where there's the possibility for that occurring happen frequently.

    21. LF

      Mm-hmm.

    22. CU

      So we get to learn more rapidly. We get to do that with lower risk, uh, for everyone. And then we can deliver value to people that, that need to get from one place to another. And then once we've got that solved, then the, kind of the freeway driving part of this just falls out.

    23. LF

      Mm-hmm.

    24. CU

      But we are able to learn more safely and more quickly in the urban environment.

    25. LF

      So 10 years and then scale 20, 30 year. I mean, who knows if, if a sufficiently compelling experience is created, it could be faster and slower.

    26. CU

      Yeah.

  14. 38:24 – 42:17

    What would accelerate everything: perfect short-horizon forecasting and protecting vulnerable users

    1. LF

      Do you think there could be breakthroughs and what kind of breakthroughs might there be that completely change that timeline? Again, not only am I asking you-

    2. CU

      So-

    3. LF

      ... to predict the future-

    4. CU

      Oh, yeah.

    5. LF

      ... I'm asking you to predict breakthroughs that haven't happened yet.

    6. CU

      So, so what's the... E- I think another way to ask that was, would be if I could wave a magic wand, what part of the system would I make work today to accelerate it as quick as possible, quickly as possible?

    7. LF

      Right. Don't say infrastructure, please don't say infrastructure.

    8. CU

      No, it's definitely not infrastructure.

    9. LF

      Okay. (laughs)

    10. CU

      Uh, it's really that ca- that perception forecasting capability. So if, if, if tomorrow you could give me a perfect model of what's happened, what is happening and what will happen for the next five seconds around a vehicle on the roadway-

    11. LF

      Mm-hmm.

    12. CU

      ... that would accelerate things pretty dramatically.

    13. LF

      Are you... In terms of staying up at night, are you mostly bothered by cars, pedestrians or cyclists?

    14. CU

      So I, I worry most about the vulnerable road users, about the combination of cyclists and cars, right? Just be... Or cyclists and pedestrians, because, you know, they're not in armor. You know, with the, the cars, they're bigger, they've got protection for the people. And so the ultimate risk is, is lower there. Whereas, uh, a pedestrian or cyclist, they're out in the road and, you know, they, they don't have any protection. And so, you know, we need to pay extra attention to that.

    15. LF

      Do you think about a very difficult technical challenge, uh, of the fact that pedestrians... If you try to protect pedestrians by being careful and slow, uh, they'll take advantage of that? So the game theoretic dance.

    16. CU

      Yeah.

    17. LF

      Uh, does that worry you of how, from a technical perspective, how we solve that? 'Cause as humans, the way we solve that is kind of nudge our way through the pedestrians, which doesn't feel, from a technical perspective as a, uh, appropriate algorithm. Uh, but do, do you think about how we solve that problem?

    18. CU

      Yeah, I think, I think there's, there's... I think, I think that was actually... There's two different concepts there. So one is, am I worried that because these vehicles are self-driving, people will kind of step in the road and take advantage of them?

    19. LF

      Yeah.

    20. CU

      And I've heard this and I don't really believe it, because if I'm driving down the road and somebody steps in front of me, I'm going to stop, right? Like, uh, e- even if I'm annoyed, I'm not going to just drive through a person stood in the road.

    21. LF

      Right.

    22. CU

      And so I think today people can take advantage of this, and, and you do see some people do it. Uh, I guess there's an incremental risk because maybe they have lower confidence that I'm going to see them than they might have for an automated vehicle. And so maybe that shifts it a little bit. But I think people don't want to get hit by cars.

    23. LF

      (laughs) Yeah.

    24. CU

      Uh, and so I, I think that I'm not that worried about people walking onto the 101 and, you know, creating chaos-

    25. LF

      Right.

    26. CU

      ... more than they would today. Regarding kind of the nudging through a big stream of pedestrians leaving a concert or something, uh, I think that is further down the technology pipeline.

    27. LF

      Right.

    28. CU

      I, I think that you're right, that's tricky. I don't think it's necessarily... I think the algorithm people use for this is pretty simple.

    29. LF

      Yeah.

    30. CU

      Uh, right? It's kind of just move forward slowly and if somebody's really close then stop.

  15. 42:17 – 44:32

    Aurora’s strategy: focus, talent, culture, and infrastructure over demos

    1. LF

      That's a, that's a, yeah, that's an experience problem, not an algorithm problem.Who's the main competitor to Aurora today? And how do you out-compete them in the long run?

    2. CU

      (laughs) So we, we really focus a lot on what we're doing here. I think that, you know, I've said this a few times, that this is a huge difficult problem, and it's great that a bunch of companies are tackling it, because I think it's so important for society that somebody gets there. So we, you know, we're, we don't spend a whole lot of time, like, thinking tactically about who's out there and, and how do we beat that, that, that person individually. Um, what are we trying to do to, to go faster ultimately?

    3. LF

      Mm-hmm.

    4. CU

      Uh, well part of it is, the leadership team we have has got pretty tremendous experience. Um, and so we kind of understand the landscape and understand where the cul-de-sacs are, to some degree, and, you know, we try and avoid those. I think there's a part of it's just this great team we've built. Uh, people... This is a technology and a company that people believe in the mission of, and so it allows us to attract just awesome people to go work. We've got a culture, I think, that people appreciate, that allows them to focus, allows them to really spend time solving problems. Uh, and I think that keeps them energized. Uh, and then we've invested hard, uh, uh, or invested heavily in the infrastructure and architectures that we think will ultimately accelerate us. So because of the folks we were able to bring in early on, because of the, the, the great investors we have, you know, w- we don't spend all of our time doing demos, a- and kind of leaping from one demo to the next. We've been given the freedom to invest in infrastructure to do machine learning, infrastructure to pull data from our on-road testing, infrastructure to use that to accelerate engineering. And I think that, that early investment and continuing investment in those kind of tools will ultimately allow us to accelerate and, and, and do something pretty incredible.

    5. LF

      Chris, beautifully put. It's a good place to end. Thank you so much for talking today.

    6. CU

      Oh, thank you very much. Really enjoyed it.

Episode duration: 44:47

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode Tj6NOfdfa4o

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

iOSAndroidClaudeChrome