WEBVTT - Could Autonomous Diggers Unleash a Building Boom? 0:00:15.356 --> 0:00:15.796 Pushkin. 0:00:20.396 --> 0:00:23.996 When we started this show in twenty twenty two, the 0:00:24.076 --> 0:00:27.996 standard line about driverless cars was driverless cars have been 0:00:28.076 --> 0:00:31.796 five years away for the past fifteen years, because it 0:00:31.876 --> 0:00:35.796 seemed like they were just always around the corner, always 0:00:35.916 --> 0:00:38.396 just a few years away, but they never quite arrived. 0:00:39.276 --> 0:00:43.476 Nobody says that anymore. Today, in several cities around the country, 0:00:43.676 --> 0:00:46.236 getting a ride from a driverless car is just a 0:00:46.316 --> 0:00:49.436 normal thing people do, and it'll become normal in more 0:00:49.476 --> 0:00:52.476 and more and more cities over the next few years. 0:00:52.796 --> 0:00:56.036 Driverless cars are here now, so now we can ask 0:00:56.556 --> 0:01:05.356 what's next. I'm Jacob Goldstein and this is What's Your Problem, 0:01:05.436 --> 0:01:07.196 the show where I talk to people who are trying 0:01:07.236 --> 0:01:11.076 to make technological progress. My guest today is Boris Soffman. 0:01:11.516 --> 0:01:15.036 He's the co founder and CEO of Bedrock Robotics, a 0:01:15.076 --> 0:01:18.676 company that's figuring out how to retrofit heavy equipment to 0:01:18.716 --> 0:01:22.996 make it work autonomously. Boris's problem is this, how do 0:01:23.036 --> 0:01:25.876 you teach machines not just to drive, but to do 0:01:25.956 --> 0:01:30.196 things like grade roads and move heavy things around construction sites. 0:01:31.116 --> 0:01:34.476 Boris's company is starting with excavators, and they plan to 0:01:34.516 --> 0:01:39.196 have their first commercial excavators autonomously digging holes on construction 0:01:39.316 --> 0:01:43.676 projects next year. Later in the interview, Boris goes big. 0:01:43.716 --> 0:01:47.036 He argues that if Bedrock succeeds, the company could help 0:01:47.076 --> 0:01:51.076 push forward a broad wave of new building in America. 0:01:51.796 --> 0:01:55.396 But first we talked about his time at Weimo and 0:01:55.436 --> 0:01:58.236 how the wild evolution of the autonomous vehicles that he 0:01:58.276 --> 0:02:01.156 worked on there led him to start Bedrock. 0:02:01.836 --> 0:02:04.716 The time at Weimo was this incredible period where I 0:02:04.836 --> 0:02:07.836 was there for about five years from mid twenty nineteen 0:02:07.956 --> 0:02:10.956 till through spring twenty four, and that was this really 0:02:10.956 --> 0:02:13.836 big period where it was going through this like fifteen 0:02:13.956 --> 0:02:16.956 years of R and D, and then it finally transitioned 0:02:16.956 --> 0:02:19.876 into this hockey stick of like growth that is helping today. 0:02:19.876 --> 0:02:22.316 And so today waymos at over one hundred million miles 0:02:22.356 --> 0:02:25.036 fully driver lists. It's at five times safer than a human. 0:02:25.316 --> 0:02:27.316 It's millions of miles every single week, and so it's 0:02:27.356 --> 0:02:30.436 kind of scaling itsponentially and it's like a genuinely fantastic product. 0:02:30.836 --> 0:02:32.676 But I was there when we were like stressing over 0:02:32.716 --> 0:02:35.036 the first hundred miles, and it felt like the most 0:02:35.076 --> 0:02:38.476 incredible achievement to just go like ten fifty hundred miles 0:02:38.556 --> 0:02:40.716 to completely drop less, say now that happens at like 0:02:40.836 --> 0:02:43.156 hundreds of thousands of miles every single day. And so 0:02:44.436 --> 0:02:46.076 one of the things that really broke through it made 0:02:46.076 --> 0:02:49.276 that possible is this shift to machine learning and data 0:02:49.356 --> 0:02:51.716 driven approaches as a core of the autonomy stack. 0:02:52.116 --> 0:02:54.756 And just to be clear, like that's as opposed to 0:02:54.876 --> 0:02:58.556 a more like euristics or rule based kind of model, 0:02:58.676 --> 0:03:01.836 like the old school twentieth century eyes, like, well, just 0:03:01.876 --> 0:03:04.356 tell the car all the rules of how to drive 0:03:04.556 --> 0:03:07.556 and then it'll drive. Like that's what you're comparing machine 0:03:07.596 --> 0:03:07.956 learning to. 0:03:08.396 --> 0:03:10.916 Yeah, like drive, I'm going to like embed the cost 0:03:10.956 --> 0:03:13.436 functions and all this. So yeah, so like large scale 0:03:13.436 --> 0:03:16.276 search and heuristics and rules and you can embed them 0:03:16.276 --> 0:03:18.556 now inside those heuristics, and so you can solve almost 0:03:18.556 --> 0:03:20.876 any given problem with that sort of approach. 0:03:21.836 --> 0:03:24.276 So you can solve one problem, but you can't solve 0:03:24.316 --> 0:03:27.436 like every single possible problem that would ever arise when 0:03:27.476 --> 0:03:30.076 you're driving, which is actually what you have to solve 0:03:30.156 --> 0:03:32.036 to do full autonomy. 0:03:31.636 --> 0:03:33.996 Right right with activating whackamall, where like you fix one 0:03:34.036 --> 0:03:35.876 problem and it becomes harder and harder to kind of 0:03:35.916 --> 0:03:38.276 scale the other ones. And so and so there was 0:03:38.316 --> 0:03:41.796 this really conscious shift at WEIMA, which was kind of 0:03:41.796 --> 0:03:43.676 bold at the time because it feels obvious in hindsight, 0:03:43.716 --> 0:03:46.036 It wasn't obvious at all back then. The shift to 0:03:46.156 --> 0:03:49.796 like really embracing this as a data driven solution where 0:03:49.836 --> 0:03:52.916 you're learning from human driving and human behavior. 0:03:52.996 --> 0:03:55.916 And when you say you're learning, you mean the machine 0:03:55.996 --> 0:03:58.676 learning model. The AI basically is learning. 0:03:58.436 --> 0:04:01.036 The machine learning model, that's right. And so you're basically 0:04:01.076 --> 0:04:03.596 taking giant scales of data one hundreds of thousands of 0:04:03.596 --> 0:04:06.676 millions of miles, and you're learning the model of how 0:04:06.676 --> 0:04:09.316 do you drive and how do you interpret this? Like 0:04:09.676 --> 0:04:12.236 infinite compocity and kitchen sink of contacts, like all of 0:04:12.236 --> 0:04:14.956 this sensored data, the road structures, the things you see 0:04:14.996 --> 0:04:17.036 around you, the way people are moving. How do you 0:04:17.076 --> 0:04:20.396 go from a kind of engineered and you know, kind 0:04:20.436 --> 0:04:22.996 of trained solution into one that is like dominantly a 0:04:23.036 --> 0:04:23.676 warned solution. 0:04:24.116 --> 0:04:26.236 I mean, as you're doing that, you're sort of riding 0:04:26.316 --> 0:04:31.236 the historic machine learning AI wave, right, Like that's right. 0:04:31.316 --> 0:04:34.316 Presumably you're able to do that at that moment because 0:04:34.716 --> 0:04:38.196 of this explosion in AI, which is basically machine learning, right, 0:04:38.236 --> 0:04:38.676 you know that. 0:04:38.876 --> 0:04:40.636 And you couldn't have done that five ten years ago. 0:04:40.836 --> 0:04:42.876 That's what you're kind of coming out of at way MO, 0:04:43.036 --> 0:04:45.716 that's driving this success at way MOO. How does that 0:04:46.236 --> 0:04:47.836 set you up for what you're doing a bedrock? 0:04:48.236 --> 0:04:50.636 The most shocking thing was just how well does generalized 0:04:50.676 --> 0:04:53.196 from San Francisco to Los Angeles, Phoenix in Austin and 0:04:53.236 --> 0:04:56.516 then becomes almost like a qualification problem eventually, where you're 0:04:56.596 --> 0:04:58.476 using data to fill some gaps, but like you need 0:04:58.556 --> 0:05:00.236 less and less of it, less and less new things 0:05:00.276 --> 0:05:02.996 surprise you, like you're just your competency kind of expands. 0:05:03.396 --> 0:05:06.196 And then we unified the technology stack between cars and 0:05:06.196 --> 0:05:08.796 trucks where even jumping from a cargo a truck was 0:05:09.236 --> 0:05:11.076 needed maybe ten to fifty percent more data, but it 0:05:11.076 --> 0:05:13.876 was fundamentally like you're using data to explain why how 0:05:13.916 --> 0:05:15.196 you operate a very different potages. 0:05:15.236 --> 0:05:16.756 This is like a big truck, This is like an 0:05:16.796 --> 0:05:19.116 eighteen wheel truck. When Google was working on that, yeah, 0:05:19.116 --> 0:05:19.996 are way most like. 0:05:19.996 --> 0:05:23.676 Fifty three foot trailer like eighty thousand pounds. And so 0:05:23.756 --> 0:05:27.276 that was the big moment where we started thinking about 0:05:27.316 --> 0:05:28.996 where else can you apply this? What are the places 0:05:28.996 --> 0:05:32.396 where you have all this diversity of challenges and capabilities 0:05:32.396 --> 0:05:35.396 that benefit from this type of versatility and also have 0:05:35.476 --> 0:05:38.036 the ability to jump between platforms in this really natural way. 0:05:38.116 --> 0:05:41.036 And so we looked at a lot of spaces and 0:05:41.076 --> 0:05:44.476 really settled on automation of specialized typing machinery, and so 0:05:44.516 --> 0:05:47.796 the sipes of machines that you see in construction, like 0:05:47.876 --> 0:05:50.516 excavators and wheeloaders and bulldozers, but also frankly the sort 0:05:50.516 --> 0:05:53.396 of machines you see in all sorts of industries like 0:05:53.716 --> 0:05:57.756 agriculture and mining and lumper and garbage movement. And so 0:05:57.836 --> 0:06:00.476 you have these very diverse types of machines that are 0:06:00.516 --> 0:06:03.396 interacting with the world around them. They're fairly still moving, 0:06:03.436 --> 0:06:05.716 they're in semi controlled environments, and at the end of 0:06:05.756 --> 0:06:09.756 the day, there's astronomical scale at which they great at 0:06:09.836 --> 0:06:12.676 and a lot of the learnings that we experienced a 0:06:12.756 --> 0:06:15.636 weymle actually transfer over incredibly well. But the physics of 0:06:15.676 --> 0:06:17.756 the problem is a lot less aperspherial. It's actually like, 0:06:17.836 --> 0:06:19.476 really tackle this and get the market. 0:06:19.916 --> 0:06:26.396 Basically, big vehicles operating in semi constrained environments doing things 0:06:26.436 --> 0:06:28.796 to the world, interacting with the world in some physical 0:06:28.836 --> 0:06:30.676 way in addition to just driving across it. 0:06:30.836 --> 0:06:33.316 That's right, and these are slow moving vehicles where you're 0:06:33.316 --> 0:06:36.596 already on a closed site with people who are assumed 0:06:36.596 --> 0:06:39.236 to be knowledgeable about the world around you, and you're 0:06:39.236 --> 0:06:41.636 moving at like five off an hour, for example, You're 0:06:41.676 --> 0:06:43.556 able to slow down and stop. You're able to minimize 0:06:43.556 --> 0:06:46.516 your exposure. You always have a minimum safety condition of stopping. 0:06:46.716 --> 0:06:49.836 Your complexity is less abound interactions with others on the 0:06:49.916 --> 0:06:52.236 road and more about the interactions with the world around you. 0:06:52.756 --> 0:06:55.796 And so you can actually tackle safety not through this 0:06:55.876 --> 0:06:58.876 sort of statistical methods that drove a lot of the 0:06:58.916 --> 0:07:03.436 mileage that we collect, but through a much more direct 0:07:03.516 --> 0:07:05.676 measure of your competencies in order to just make sure 0:07:05.676 --> 0:07:08.796 that you're like you're actually capable of hurting somebody. 0:07:08.916 --> 0:07:12.316 Yeah, more like like an industrial robot al most right. 0:07:12.236 --> 0:07:15.396 That's right. Yeah, it's like a traditional Susis engineering, Yeah. 0:07:15.236 --> 0:07:17.396 Where it's like, look, even if it can't do the 0:07:17.436 --> 0:07:19.516 thing every time, that's fine. Just make sure it's not 0:07:19.556 --> 0:07:21.516 gonna like go crazy and kill somebody. 0:07:21.556 --> 0:07:22.916 That's right. You can make it to that. Like your 0:07:22.916 --> 0:07:25.676 worst case is your productivity suffers, but safety wise, you're 0:07:25.956 --> 0:07:30.076 you're you're solid. And so that's like incredibly enabling because 0:07:30.076 --> 0:07:32.196 your long tail is now no longer safety. It's the 0:07:32.276 --> 0:07:33.356 versatility what you can do. 0:07:33.556 --> 0:07:35.156 Why did you start with excavators? 0:07:36.236 --> 0:07:40.396 So excavators are the most highly utilized machine, Like they're 0:07:40.516 --> 0:07:43.596 usually the highest volume in fleet. So between twenty and 0:07:43.676 --> 0:07:45.796 twenty five percent of fleets are excavators. 0:07:45.796 --> 0:07:48.716 Fleets are just like big construction heavy equipment. 0:07:48.836 --> 0:07:51.676 Yeah, like a general contractor that alone, like a thousand 0:07:51.716 --> 0:07:54.076 machines two hundred, two hundred and fifty will probably be 0:07:54.116 --> 0:07:55.556 excavators on average. 0:07:55.596 --> 0:07:57.676 And it's basically what a kid would call it digger, right, 0:07:57.716 --> 0:08:00.476 It's like at a digger bucket an arm, and it 0:08:00.636 --> 0:08:01.636 like digs stuff up. 0:08:01.756 --> 0:08:03.596 It's like the equivalent of an arm. Like you can 0:08:03.676 --> 0:08:05.876 like dig stuff, you can demolish stuff, you can swap 0:08:05.956 --> 0:08:08.116 your tools, you can like lift pipes and put them 0:08:08.116 --> 0:08:09.396 in a holes. They can do a ton of stuff 0:08:09.396 --> 0:08:11.196 with it. It's kind of crazy. It really is a 0:08:11.196 --> 0:08:13.916 personal machine. It also makes it very complicated to work. 0:08:13.956 --> 0:08:16.476 So there's like seven degrees of freedom sometimes eight, and 0:08:16.556 --> 0:08:18.676 so it's one of the hardest machines to learn, and 0:08:18.716 --> 0:08:20.716 it takes four to five years to really become an expert. 0:08:20.796 --> 0:08:22.836 And there's a huge difference between an expert and novice, 0:08:22.876 --> 0:08:25.436 and so you kind of have this situation where it's 0:08:25.516 --> 0:08:28.876 a huge volume of work and it's really hard to learn, 0:08:28.876 --> 0:08:30.076 and so you have a really deep pull in the 0:08:30.076 --> 0:08:31.676 market for it, meaning. 0:08:31.516 --> 0:08:33.356 A lot of demand, like a lot of people want 0:08:33.956 --> 0:08:36.676 somebody who can drive one of these or a machine 0:08:36.676 --> 0:08:37.236 that could do it. 0:08:37.396 --> 0:08:39.196 Well, let me tell you about the demand. I've never 0:08:39.236 --> 0:08:42.236 seen such a divergence of supply to demand like in 0:08:42.276 --> 0:08:45.196 my career in any case, where on the demand side, 0:08:45.236 --> 0:08:48.916 you have this astronomical construction industry that's already two trillion 0:08:48.956 --> 0:08:51.436 dollars a year in the US, that's obviously very heavily 0:08:51.476 --> 0:08:54.516 building and having machinery work. And then you have this 0:08:54.636 --> 0:08:57.116 like shortage of operators that already existed, but it's going 0:08:57.116 --> 0:08:59.716 in the wrong direction where forty percent of construction workers 0:08:59.836 --> 0:09:02.356 is retiring in the next ten years. Our partners are 0:09:02.916 --> 0:09:06.996 consistently having trouble filling labor. We've met some that have 0:09:07.076 --> 0:09:08.196 one hundred percent turnover. 0:09:09.196 --> 0:09:10.556 Everybody leaves every year. 0:09:10.756 --> 0:09:12.996 It means like a third of people might be lifers, 0:09:13.036 --> 0:09:15.756 but then like two thirds transition more than once per year, 0:09:15.756 --> 0:09:19.156 and you're constantly backfilling the skill sets very a ton 0:09:19.676 --> 0:09:22.276 and one of them said that for every one person 0:09:22.396 --> 0:09:24.516 entering the workforce of the quality that they look for, 0:09:24.596 --> 0:09:27.356 their seven leaving. And so we end up having is 0:09:27.876 --> 0:09:30.796 this shortage of ability to meet this demand, and so 0:09:30.916 --> 0:09:33.836 prices go up, jobs don't get done. And what's interesting 0:09:33.916 --> 0:09:36.036 is it's not that's like isolated industry that's like just 0:09:36.156 --> 0:09:38.196 on its own kind of like having these sort of challenges. 0:09:38.556 --> 0:09:41.556 It's a horizontal it supports every industry. You can't build 0:09:41.636 --> 0:09:44.116 data centers free eye without it, you can't build houses, 0:09:44.196 --> 0:09:46.236 you can't build energy facilities. 0:09:46.036 --> 0:09:49.036 Of like a rate limiting skill or rate limiting machine. 0:09:49.276 --> 0:09:51.156 It's a whole country. That's exactly right. And then you 0:09:51.196 --> 0:09:54.596 have this need that starts with labor. But then there's safety. 0:09:54.636 --> 0:09:57.796 It's huge amounts of safety challenges. You have, you know, 0:09:57.956 --> 0:10:01.996 huge predictability challenges. So if you actually could soften out 0:10:01.996 --> 0:10:04.676 some of these constraints, we would build more, more work 0:10:04.716 --> 0:10:06.996 would get done, and it would like stimulate the whole economy. 0:10:06.996 --> 0:10:09.316 And so that's what's actually pretty exciting about this opportunity. 0:10:09.636 --> 0:10:11.996 It's not as yours some game at all good. 0:10:12.076 --> 0:10:14.836 So, like you decide to focus on excavators, you you know, 0:10:14.956 --> 0:10:16.996 you raise money for your company, Like, do you go 0:10:17.076 --> 0:10:21.076 out and buy a whatever, a million dollar excavator. You 0:10:21.076 --> 0:10:23.276 go to what is it bucket and shovel dot com 0:10:23.316 --> 0:10:24.756 and buy yourself an excavator. 0:10:25.236 --> 0:10:28.116 They're like three hundred thousand to five hundred thousand, so 0:10:28.156 --> 0:10:30.796 they're like still expensive. Yeah, they're pretty expensive. 0:10:30.876 --> 0:10:32.796 So I mean like, did you buy one like we did? 0:10:32.956 --> 0:10:33.636 Yeah? 0:10:33.716 --> 0:10:34.476 Did you drive it? 0:10:34.596 --> 0:10:36.356 Of course, like you have to. It's like it's like 0:10:36.356 --> 0:10:38.596 a ride of passage. Yeah, they're really fine. I even 0:10:38.596 --> 0:10:41.396 took a lessoners. There's a place in Vegas ALESSI drive 0:10:41.436 --> 0:10:44.156 excavators and take away and from like a trade operator. 0:10:44.156 --> 0:10:44.756 It was pretty fun. 0:10:44.796 --> 0:10:47.556 So, oh, that's genius. You get to break things. I 0:10:47.596 --> 0:10:50.276 was actually as I was walking here today to the 0:10:50.356 --> 0:10:53.876 train that there's a playground and there was an excavator 0:10:53.916 --> 0:10:57.116 just like breaking up the asphalt and then like prying 0:10:57.156 --> 0:10:59.276 it up. I was like, that looks awesome. 0:10:59.676 --> 0:11:02.156 This is the funniest thing. Anybody that gets an excavator, 0:11:02.236 --> 0:11:04.476 like your inner six year old comes out. Suddenly you 0:11:04.556 --> 0:11:07.076 have like the most mature, sophisticated person is trying to 0:11:07.076 --> 0:11:09.636 be all professional. You get an excavator and just like 0:11:09.996 --> 0:11:12.236 get like a giant glob of mud and then like 0:11:12.476 --> 0:11:14.636 bring as hig as a can and then like PLoP 0:11:14.676 --> 0:11:16.716 it down and see what happens. It's amazing, and it's like, 0:11:16.716 --> 0:11:18.836 without fail, everybody kind of reverts back to this sort 0:11:18.836 --> 0:11:20.476 of a world. Yeah, we drove and I think we 0:11:20.556 --> 0:11:23.196 bought like a half dozen excavators at this point, and 0:11:23.236 --> 0:11:25.316 then we also then use a lot of excavators from 0:11:25.316 --> 0:11:28.876 our partners who are general contractors and subcontractors. 0:11:28.956 --> 0:11:30.796 And it's just what a year or so ago that 0:11:30.836 --> 0:11:32.276 you started, like not that long in. 0:11:32.236 --> 0:11:34.436 This less than a year and a half, yeah, like 0:11:34.516 --> 0:11:36.956 last last minute, seah, it's a pretty been a good run. 0:11:37.716 --> 0:11:40.636 How much is sort of commodified of autonomy, right? How 0:11:40.676 --> 0:11:42.916 much is just like, well, we're gonna buy this, this, 0:11:42.996 --> 0:11:44.636 and this in terms of sort of hardware and we 0:11:44.676 --> 0:11:46.356 know the software, and then how much is like, oh, 0:11:46.396 --> 0:11:48.596 here's the things we have to figure out that nobody 0:11:48.596 --> 0:11:49.156 knows how to do. 0:11:49.676 --> 0:11:52.556 So it's one of the other enablers that's like way 0:11:52.556 --> 0:11:54.076 better than what we would have had to go through 0:11:54.116 --> 0:11:56.596 ten years ago. In the space, we can use a 0:11:56.636 --> 0:11:59.436 lot of the existing components on lighter on cameras, on 0:11:59.516 --> 0:12:04.116 Imus GPS, there's a lot of tailwind from automotive great 0:12:04.196 --> 0:12:05.076 cameras and compute. 0:12:05.076 --> 0:12:08.236 That's like accelerating and like presumably they're buying those things 0:12:08.276 --> 0:12:10.876 at scale you can like you've got to go get 0:12:10.916 --> 0:12:13.596 them cheap. Essentially, you're like, give me one of those, 0:12:13.596 --> 0:12:14.716 one of those, one of those. 0:12:14.516 --> 0:12:16.956 That's right, because you're gonna go to millions and millions 0:12:16.996 --> 0:12:19.516 of units as like every car against an autopilot equivalent 0:12:19.556 --> 0:12:21.596 over the next two three years. Oh, that's starting to 0:12:21.636 --> 0:12:23.316 kind of pump in a cost. So that helps. And 0:12:23.356 --> 0:12:27.436 then even the platform, we can retrofit existing machinery. 0:12:27.716 --> 0:12:29.596 When you say the platform, what do you mean in 0:12:29.596 --> 0:12:30.796 this context. 0:12:30.476 --> 0:12:33.116 Platform is like the car, the truck, the excavator, and 0:12:33.156 --> 0:12:35.316 the way in this case the excavator. Yeah, the machine, 0:12:35.356 --> 0:12:40.196 so like the excavator itself. Construction machines, particularly this latest generation, 0:12:40.676 --> 0:12:44.156 they're really well designed to where they're already effectively drive 0:12:44.196 --> 0:12:46.156 by wire, which means that every signal going through the 0:12:46.156 --> 0:12:49.276 machine is electric, and so we're able to spice into 0:12:49.356 --> 0:12:54.396 it and both read and write to these signals. 0:12:54.036 --> 0:12:58.516 And control them electric meaning like computerized basically. 0:12:58.356 --> 0:13:01.076 Meaning like you have a joystick that operates the excavatory 0:13:01.316 --> 0:13:04.076 that's not physically connected to the hydraulics. It's an electric 0:13:04.116 --> 0:13:06.756 signal that's connected to the hydraulics, and so the same 0:13:06.796 --> 0:13:09.516 for every signal and the machine, like the sense the 0:13:09.836 --> 0:13:12.716 pressure gages, and so we're able to both get the 0:13:12.796 --> 0:13:15.836 data from the machine as well as control the machine 0:13:16.356 --> 0:13:20.276 through a non invasive integration where we can upit a 0:13:20.316 --> 0:13:22.676 machine with our sensors and compute suite in like less 0:13:22.676 --> 0:13:26.196 than three hours and make it autonomy capable and it's 0:13:26.196 --> 0:13:28.716 completely reversible. We could never do that with a car 0:13:28.836 --> 0:13:31.996 or a truck because the platforms were just not designed 0:13:31.996 --> 0:13:32.316 this way. 0:13:32.476 --> 0:13:35.436 Yeah, so there's a lot that's there already. The excavators 0:13:35.476 --> 0:13:38.516 themselves are sort of easy to integrate with lots of 0:13:38.516 --> 0:13:41.356 off the shelf technology. What's not there, Like, when you're 0:13:41.356 --> 0:13:43.196 coming in, what do you have to sort of build 0:13:43.236 --> 0:13:44.516 that nobody has built before? 0:13:44.956 --> 0:13:49.476 Nobody has actually created autonomy that can solve the really 0:13:49.516 --> 0:13:51.676 nuanced problems that you need to solve in order to 0:13:51.676 --> 0:13:55.276 operate like an excavator in construction tasks. 0:13:55.596 --> 0:13:59.116 So let's talk like specifically, you buy your excavators, you 0:13:59.156 --> 0:14:02.796 got your hardware, you know, whatever, what basic AI model 0:14:02.836 --> 0:14:05.436 you're going to use, But like, what's a specific thing. 0:14:05.476 --> 0:14:07.196 You have to figure out well a. 0:14:07.116 --> 0:14:08.596 Lot of things. So first of all, it's not trivial 0:14:08.636 --> 0:14:11.956 to tap into these machines and to build a platform 0:14:12.036 --> 0:14:14.836 or kind of an integration around it where you can 0:14:15.116 --> 0:14:17.236 read them, you can control them, and so you have 0:14:17.276 --> 0:14:19.716 to basically create like a wrapper around the machine and 0:14:19.836 --> 0:14:22.276 also design in a way that scales later to new machines. Right, 0:14:22.556 --> 0:14:25.436 So that part's hard. It's a big autonomy problem. So 0:14:25.516 --> 0:14:27.516 in that respect is no different than what we tackle 0:14:27.516 --> 0:14:30.476 with a WEIMO, where you have to train massive scale models. 0:14:30.476 --> 0:14:31.836 You have to collect a huge amount of data. 0:14:31.876 --> 0:14:34.516 And in this case it's like about the digging, like 0:14:34.636 --> 0:14:37.156 it's about the digging dumb question? Is the digging the 0:14:37.156 --> 0:14:39.116 hard part? Like what happens when the thing hits the ground. 0:14:39.116 --> 0:14:41.036 There's different kinds of stuff in the ground. 0:14:41.156 --> 0:14:43.836 Like tell me in the category of what's new, suddenly 0:14:44.276 --> 0:14:47.396 a car doesn't change the environment around it. You have 0:14:47.436 --> 0:14:50.956 a really complicated tough earth and soil, and you got 0:14:50.996 --> 0:14:52.916 to figure out the physics of how you dig through it, 0:14:53.196 --> 0:14:56.276 how do you deal with clay versus top soil, how 0:14:56.316 --> 0:14:58.236 do you deal with rocks inside? Then so, now if 0:14:58.276 --> 0:15:00.796 you want to use simulation to solve parts of these problems. 0:15:00.916 --> 0:15:03.036 You have a very complicated simulation problem because you have 0:15:03.076 --> 0:15:05.356 to solve not just the sensor side, but also the 0:15:05.956 --> 0:15:09.596 manipulation side. You have to structure this problem in a 0:15:09.636 --> 0:15:13.476 way where you're learning from the data you collect in 0:15:13.556 --> 0:15:15.956 order to actually capture the nuances of how do you 0:15:15.996 --> 0:15:18.396 actually interact with the environment. You have to then control 0:15:18.476 --> 0:15:20.196 and execute it the right way. You have to think 0:15:20.236 --> 0:15:22.156 about how do you actually define the goal? 0:15:22.476 --> 0:15:24.956 So from the user point of view with an excavator, 0:15:25.116 --> 0:15:27.236 like how does it look like? I just I want 0:15:27.236 --> 0:15:29.156 to dig a hole of this size at this spot? 0:15:29.236 --> 0:15:29.996 Or like what is it? 0:15:30.276 --> 0:15:32.836 Yeah, and this is a journey like fast forward you 0:15:32.876 --> 0:15:35.396 know next year where card customer does this right, So 0:15:35.436 --> 0:15:38.116 what they would do is they would give it a 0:15:38.116 --> 0:15:40.676 model of what they want the earth to be dug 0:15:40.716 --> 0:15:42.516 towards and so that would be like a three D 0:15:42.636 --> 0:15:47.676 representation of the depth. So with length depth, here's the edges, 0:15:47.716 --> 0:15:50.516 here's the constraints. The pickups for the trucks are going 0:15:50.556 --> 0:15:52.796 to be over here and then go to work. And 0:15:52.836 --> 0:15:56.036 basically these projects, like these machines will work for many 0:15:56.036 --> 0:15:59.276 months at a time, continually just digging earth and working 0:15:59.276 --> 0:16:01.836 it towards the right foundation, and the system will respect 0:16:01.876 --> 0:16:05.796 that boundary and basically dig down to that level. It'll 0:16:05.796 --> 0:16:08.796 have precision to the edges that you've defined, it'll do 0:16:08.796 --> 0:16:10.436 it in a sequence that makes sense so that you 0:16:10.476 --> 0:16:12.596 don't trap yourself in a hole, for example. And then 0:16:12.636 --> 0:16:16.516 you specify where you're going to have dump truck pickups. 0:16:16.636 --> 0:16:19.396 Then there's projects where this happens for nine months straight 0:16:19.476 --> 0:16:21.916 or twelve months straight with like many machines, and so 0:16:22.236 --> 0:16:25.556 what you basically need to specify is what you want 0:16:25.596 --> 0:16:27.356 to dig to, and that what you want to dig 0:16:27.396 --> 0:16:31.276 to ends up basically being the foundation that you're working 0:16:31.276 --> 0:16:33.716 towards for whatever you're going to construct there. And we 0:16:33.756 --> 0:16:36.316 want that to be versatile. So sometimes you're just taking 0:16:36.316 --> 0:16:38.756 off a layer of top soil. Sometimes you're digging eight 0:16:38.756 --> 0:16:42.516 feet deep. Sometimes you're taking an existing stockpile and moving 0:16:42.556 --> 0:16:45.156 in somewhere else. So there's a lot of permutations of this. 0:16:45.276 --> 0:16:48.756 Sometimes you're taking rubble from a demolition job and loading 0:16:48.796 --> 0:16:51.356 it onto trucks, right, and so what's nice is that 0:16:51.436 --> 0:16:53.516 you start to see patterns over and over again, and 0:16:53.556 --> 0:16:54.316 this sort of work. 0:16:58.676 --> 0:17:10.116 We'll be back in just a minute. 0:17:11.956 --> 0:17:14.276 So the data problem is interesting here, right, Like you 0:17:14.316 --> 0:17:16.476 were talking about the physics, just the physics of an 0:17:16.516 --> 0:17:19.356 excavator is quite different, right, Like it's pushing down on 0:17:19.396 --> 0:17:21.676 the ground, which is pushing back on the excavator, and 0:17:21.716 --> 0:17:24.156 like as you say, the ground is changing because of 0:17:24.196 --> 0:17:28.076 its work, and there's not Well where do you get 0:17:28.076 --> 0:17:28.996 the data? 0:17:29.356 --> 0:17:31.876 Data is actually an interesting mix. We get it both 0:17:31.956 --> 0:17:35.916 on our test sites and also with our design partners. 0:17:35.956 --> 0:17:39.436 So we were working with general contractors and subcontractors. Today 0:17:39.436 --> 0:17:42.196 we have five that we're partnered with across southern states 0:17:42.196 --> 0:17:44.716 like Arizona, Texas, and so we're getting data on their size. 0:17:44.716 --> 0:17:45.996 We're getting data on our sites. 0:17:46.116 --> 0:17:48.916 It's not that much data, right, Like I mean, I 0:17:48.916 --> 0:17:52.276 guess my reference point is always like image net or 0:17:52.516 --> 0:17:54.956 you know, the Internet for large language models. It does 0:17:54.996 --> 0:17:57.996 seem like a sort of recurring problem in robotics type 0:17:57.996 --> 0:18:01.316 AI applications is sparsity of data. 0:18:01.436 --> 0:18:02.116 How do you get it? 0:18:02.596 --> 0:18:03.196 Yeah? 0:18:03.396 --> 0:18:06.196 Yeah, And so this is where there's a few kind 0:18:06.236 --> 0:18:08.676 of intersuties. So first of all, like our partners together 0:18:08.676 --> 0:18:10.596 have thousand and thousands of machines and so there's a 0:18:10.596 --> 0:18:12.916 lot of choices of which she's a partner. On the 0:18:12.956 --> 0:18:14.716 other thing that you can do is actually you can 0:18:14.756 --> 0:18:17.756 be very clever on a test site, and when you're 0:18:17.756 --> 0:18:21.956 on a real project, you're kind of getting a unbiasample. 0:18:21.956 --> 0:18:24.516 You're just getting a random distribution of the things you see, 0:18:24.596 --> 0:18:26.516 just like driving around on a road. When you're on 0:18:26.556 --> 0:18:29.796 a test site, you can actually up sample the things 0:18:29.796 --> 0:18:31.996 you actually want and you can go and you can 0:18:31.996 --> 0:18:35.236 collect ten hours of data. There's representative of five thousand 0:18:35.236 --> 0:18:38.636 hours of random data, but which is particularly useful for 0:18:38.676 --> 0:18:41.636 things like safety situations. Right, you can actually create a 0:18:41.716 --> 0:18:45.156