WEBVTT - Building a Mass Market Robot 0:00:15.356 --> 0:00:23.396 Pushkin As a general matter, I'm a fan of technological progress, 0:00:23.916 --> 0:00:27.116 but I'll admit that humanoid robots kind of creep me 0:00:27.156 --> 0:00:32.156 out for you know, obvious, normy, uncanny Valley type reasons. 0:00:32.956 --> 0:00:36.316 And yet there is an exchange that you'll hear near 0:00:36.396 --> 0:00:40.116 the end of today's show. That's the most compelling argument 0:00:40.196 --> 0:00:43.916 I've ever heard for humanoid robots. And it's not just 0:00:43.956 --> 0:00:48.836 an intellectual argument, it's an emotional argument. If that's a phrase, 0:00:49.036 --> 0:00:59.676 it's really a very human argument for humanoid robots. I'm 0:00:59.756 --> 0:01:02.116 Jacob Goldstein, and this is what's your problem to show 0:01:02.156 --> 0:01:04.156 where I talk to people who are trying to make 0:01:04.236 --> 0:01:08.636 technological progress. My guest today is Jeff Cardinas. He's the 0:01:08.756 --> 0:01:12.756 k founder and CEO of Eptronic. Jeff's problem is this, 0:01:13.636 --> 0:01:17.636 can you make a safe, reliable humanoid robot for less 0:01:17.636 --> 0:01:21.636 than fifty thousand dollars. We started our conversation talking about 0:01:21.756 --> 0:01:25.996 the DARPA Robotics Challenge. DARPA, of course, is the government 0:01:26.036 --> 0:01:28.876 agency that helped to create the Internet and that gave 0:01:28.916 --> 0:01:31.876 a big push to early self driving cars, among other things. 0:01:32.316 --> 0:01:36.196 And Jeff says the agency's robotics challenge, which happened a 0:01:36.236 --> 0:01:39.516 decade ago. Happened in twenty fifteen played a key role 0:01:39.636 --> 0:01:42.156 in launching a bunch of the companies that are now 0:01:42.196 --> 0:01:43.756 working on humanoid robots. 0:01:46.116 --> 0:01:50.076 The Darker Robotics Challenge was a challenge that was created 0:01:50.116 --> 0:01:54.076 in the wake of the Fukushima disaster. Lukashima happened in 0:01:54.116 --> 0:01:56.636 the you know, there was a meltdown in the nuclear 0:01:56.636 --> 0:01:59.476 reactors and it was unsafe for people to go in, 0:02:00.236 --> 0:02:04.356 and in essentially what happened was they needed a robot 0:02:04.396 --> 0:02:07.796 to go in to sort of prevent a broader catastrophe. 0:02:08.596 --> 0:02:11.436 And as they went out to the robotics community, the 0:02:11.476 --> 0:02:15.636 idea was calling all roboticists, we need a robot to 0:02:15.676 --> 0:02:18.996 go in and to help out here. And there was 0:02:19.036 --> 0:02:21.356 no robots that could do all the different tasks that 0:02:21.396 --> 0:02:24.276 were required to actually get to the melted down reactors. 0:02:24.356 --> 0:02:25.956 So you had to go down steps, you had to 0:02:25.996 --> 0:02:28.796 open doors, had to do a whole range of things. 0:02:29.436 --> 0:02:34.396 And in the wake of that, basically what DARPA said was, certainly, 0:02:34.396 --> 0:02:36.916 there's got to be the technology to enable us to 0:02:36.956 --> 0:02:41.236 have much more versatile robots for natural disaster relief so 0:02:41.356 --> 0:02:44.996 this never happens again. So out of that, Darker created 0:02:44.996 --> 0:02:48.396 something called the Darker Robotics Challenge, and so there's a 0:02:48.476 --> 0:02:51.476 variety of teams around the world that were put together 0:02:51.756 --> 0:02:57.156 to build these general purpose robots. And the team that 0:02:57.156 --> 0:03:00.076 we came out of was the Nasagen and the Space 0:03:00.116 --> 0:03:02.956 Center team working on a robot called Valkyrie. 0:03:03.436 --> 0:03:05.636 So I want to go back to this moment when 0:03:05.716 --> 0:03:10.156 the DARPA Challenge and there's this big final test, and like, 0:03:10.556 --> 0:03:14.756 what we have from it is not some incredible breakthrough, 0:03:14.756 --> 0:03:18.636 but like a blooper reel of robots, what are they doing? 0:03:18.716 --> 0:03:21.356 Falling downstairs or driving cars into walls or. 0:03:21.276 --> 0:03:26.156 Something pretty much. I mean, the blooper reels make it 0:03:26.196 --> 0:03:29.996 seem worse than it was, but but yeah, we had 0:03:30.396 --> 0:03:33.956 we basically the realization was the technology's not there. It's 0:03:33.956 --> 0:03:37.396 going to take time to continue to mature until it 0:03:37.436 --> 0:03:40.076 can get to the point where it's actually commercially viable. 0:03:40.676 --> 0:03:43.756 And so it's interesting, it's super interesting right that this 0:03:43.836 --> 0:03:47.876 moment is not the like beginning of some humanoid robot winter, 0:03:48.236 --> 0:03:52.156 but rather the beginning of this humanoid robot industry. Like 0:03:52.196 --> 0:03:55.436 how does that work? Like how do people whatever start 0:03:55.436 --> 0:03:58.956 companiesn't get money out of this seemingly disappointing moment. 0:04:00.036 --> 0:04:02.596 Well, I think it actually was in winter, you know, 0:04:03.156 --> 0:04:05.676 when we started in you know, the company was incorporated 0:04:05.716 --> 0:04:08.796 in twenty fifteen and we started in twenty sixteen, and 0:04:09.476 --> 0:04:12.876 for the most part, nobody wanted to talk about humanoids 0:04:12.956 --> 0:04:15.876 and nobody was really paying attention to it. There was 0:04:15.916 --> 0:04:18.196 a handful of folks that I sort of think of 0:04:18.236 --> 0:04:21.876 as the true believers that were really working on this 0:04:21.996 --> 0:04:24.196 problem and thought, you know, we don't care how long 0:04:24.236 --> 0:04:25.956 this is going to take. We're just going to keep 0:04:25.956 --> 0:04:30.276 working on this no matter what. But for the most part, 0:04:30.516 --> 0:04:33.876 the entire robotics industry was very anti humanoids coming out 0:04:33.876 --> 0:04:36.396 of the Darker Robotics Challenge, and in fact, there was 0:04:36.476 --> 0:04:39.516 many people that were saying they'll never be viable, like 0:04:39.596 --> 0:04:42.436 why would you ever use a humanoid robot. They're too complicated, 0:04:42.636 --> 0:04:46.796 they're way too expensive, and you'll always use a simpler robot. 0:04:47.236 --> 0:04:49.476 So actually, most of the people that we met when 0:04:49.476 --> 0:04:52.756 we decided to start Uptronic were doubters, and we're saying, 0:04:53.356 --> 0:04:57.996 humanoids will never make sense. We'll use these special purpose robots. 0:04:58.076 --> 0:05:00.756 You know, maybe in fifty years humanoids will make sense, 0:05:00.796 --> 0:05:01.836 but not for a long time. 0:05:02.996 --> 0:05:07.396 I mean special purpose robots is a pretty compelling case, right, 0:05:07.516 --> 0:05:12.196 like whatever we all have, you know, dishwashing robots and 0:05:12.276 --> 0:05:15.636 clothes cleaning robots in our houses, and like, you know, 0:05:15.756 --> 0:05:18.956 wheels seem way easier than legs for lots of things. 0:05:18.996 --> 0:05:22.156 And obviously there have been robot arms for what I 0:05:22.196 --> 0:05:25.196 don't know, seventy years now or something like. So robots 0:05:25.236 --> 0:05:27.676 in a way are all around us. Why would you 0:05:27.676 --> 0:05:29.716 build a machine that looks like a dude when that's 0:05:29.796 --> 0:05:30.916 like wildly hard. 0:05:30.756 --> 0:05:34.636 Right, Yeah, I mean I was naive coming out of 0:05:34.676 --> 0:05:38.036 graduate school, and to me it seemed obvious. And the 0:05:38.076 --> 0:05:39.796 way I used to think about it was you could 0:05:39.796 --> 0:05:42.236 either have thousands of robots that do one thing, or 0:05:42.316 --> 0:05:44.876 you could build one robot that could do thousands of 0:05:44.916 --> 0:05:48.276 different things. And when I would talk about this with Nick, 0:05:48.396 --> 0:05:51.036 my co founder, Nick would say, look, you can either, 0:05:51.476 --> 0:05:54.996 you know, invest all this engineering in each of these 0:05:55.036 --> 0:05:58.556 sort of narrow solutions, or yes, a humanoid robot, a 0:05:58.636 --> 0:06:02.076 viable humanoid robot could take you years, it could take 0:06:02.076 --> 0:06:04.756 you a decade, But once you invest all this time 0:06:04.796 --> 0:06:08.236 in this single platform, then you can reap the benefits 0:06:08.276 --> 0:06:13.276 of that across You can spread the research of that 0:06:13.556 --> 0:06:15.356 across many different applications. 0:06:15.996 --> 0:06:18.436 I mean, is there not a middle case where there's 0:06:18.516 --> 0:06:22.836 like some core kind of functionality that you develop that 0:06:22.916 --> 0:06:25.556 works across many different types of robots. Is that a 0:06:25.636 --> 0:06:28.916 less straw many version of the non humanoid robot kind 0:06:28.956 --> 0:06:29.556 of argument? 0:06:29.916 --> 0:06:32.436 I think there could be. But I came into robotics 0:06:32.516 --> 0:06:35.196 and basically just saw a lot of challenges with the 0:06:35.236 --> 0:06:39.516 business models. So you build this special purpose robot, you 0:06:39.716 --> 0:06:42.796 custom program the robot. In the industrial space, you can 0:06:42.836 --> 0:06:46.236 spend six times the price of the robot on just 0:06:46.276 --> 0:06:49.596 systems integration, and then the robot just does one thing. 0:06:50.116 --> 0:06:51.916 So this idea that you could have a much more 0:06:52.036 --> 0:06:55.396 versatile robot, to me, seemed obvious, like, if robotics is 0:06:55.436 --> 0:06:57.756 going to scale, we have to have much more versatile 0:06:57.836 --> 0:07:00.116 robots that we've had in the past. So if you 0:07:00.196 --> 0:07:02.156 sort of think of that as the premise is we 0:07:02.236 --> 0:07:05.156 need more versatile robots, then the question is, well, how 0:07:05.156 --> 0:07:08.276 do you get there and what does versatility mean? And 0:07:09.116 --> 0:07:12.956 that's where it led me to the humanoid making a 0:07:12.956 --> 0:07:15.916 lot of sense, because if you have to modify the 0:07:16.076 --> 0:07:18.756 environment for every new task that the robot can do, 0:07:19.396 --> 0:07:21.316 you run into the same problem that we had in 0:07:21.396 --> 0:07:24.996 sort of classical robotics. But if the robot can retrofit 0:07:25.156 --> 0:07:27.876 into the environment such that you don't have to change 0:07:27.956 --> 0:07:30.756 or modify the environment for every new task that the 0:07:30.836 --> 0:07:33.876 robot can do, then it seemed to me that this 0:07:33.916 --> 0:07:37.156 would maybe be the key unlock for robotics to actually 0:07:37.196 --> 0:07:38.396 scale to the masses. 0:07:38.916 --> 0:07:40.956 The demand would be infinite. If you had a thing 0:07:40.996 --> 0:07:42.876 that was the size and shape of a person with 0:07:43.036 --> 0:07:46.556 arms and legs like, scale would be off the charts, 0:07:46.556 --> 0:07:49.316 And presumably that's what drives costs down, and that's like 0:07:49.476 --> 0:07:50.956 the good flywheel. 0:07:50.516 --> 0:07:51.956 Right, yeah, exactly. 0:07:52.436 --> 0:07:55.036 Okay, so you had this big idea about humanoid robots 0:07:55.476 --> 0:07:57.676 and you started a company. But at the moment you 0:07:57.716 --> 0:08:00.836 started the humanoid robot company, the prevailing sentiment was like, 0:08:01.116 --> 0:08:03.956 deeply skeptical, what happened? 0:08:03.996 --> 0:08:06.196 What did you do? Well? A handful of us kept 0:08:06.236 --> 0:08:08.996 working on it, so I didn't know any better. You know, 0:08:09.236 --> 0:08:11.196 sometimes it's better that you don't know any better. I 0:08:11.556 --> 0:08:16.076 thought humanoids were really cool, and I thought that it 0:08:16.196 --> 0:08:18.156 just seemed it made sense to me that, you know, 0:08:18.476 --> 0:08:20.876 how are we going to get to millions of robots 0:08:21.316 --> 0:08:24.596 that are working and you know, with and around humans 0:08:24.596 --> 0:08:26.596 and all these environments, and to me, this seemed like 0:08:26.636 --> 0:08:28.876 the only way that that was going to happen. And 0:08:28.916 --> 0:08:32.036 the way I looked at it was, even if we failed, 0:08:32.036 --> 0:08:35.076 this was a worthy pursuit and I would be proud 0:08:35.076 --> 0:08:38.196 that I tried to do it. And so the way 0:08:38.236 --> 0:08:40.836 that we did it was we bootstrapped the company. There 0:08:40.916 --> 0:08:43.876 was no investors that were willing to invest in humanoid 0:08:43.956 --> 0:08:48.596 robots at the time that we got started, especially for hardware, 0:08:48.636 --> 0:08:51.156 which we can talk about that as we move forward. 0:08:52.636 --> 0:08:55.876 And so we bootstrapped the company and we basically got 0:08:55.956 --> 0:08:58.276 paid to build robots for a lot of different folks. 0:08:58.996 --> 0:09:01.756 And for the first five years of the company, we 0:09:01.916 --> 0:09:04.836 just built the company on revenue. We would get project 0:09:04.876 --> 0:09:07.716 after project and somehow never died. 0:09:08.116 --> 0:09:10.276 Like what kind of were you taking at that time? 0:09:10.316 --> 0:09:11.756 What's an one example, Well, we. 0:09:12.116 --> 0:09:14.436 Had our first contract was with NASA, So we had 0:09:14.436 --> 0:09:17.196 a contract with NASA to build Valkyrie two, to take 0:09:17.236 --> 0:09:19.636 the lessons learned from the DARPA Challenge and build the 0:09:19.676 --> 0:09:23.436 next iteration of Valkyrie. We were really kind of pioneering 0:09:23.676 --> 0:09:27.356 new ways of building these systems. So US Special Forces 0:09:27.396 --> 0:09:30.316 ended up coming to us about a year in and said, hey, 0:09:30.316 --> 0:09:32.996 we want to do Ironman suits, and our view was 0:09:33.076 --> 0:09:35.596 this was kind of a humanoid robot that you wear. 0:09:36.196 --> 0:09:39.916 We worked in automotive. We helped build humanoid robots for 0:09:40.116 --> 0:09:44.956 a couple you know, major companies that are still working 0:09:44.996 --> 0:09:47.476 on these things today, and we would help sort of 0:09:47.476 --> 0:09:50.636 pioneer new ways of building their platforms. So we've done 0:09:50.756 --> 0:09:54.036 fifteen unique robots since we got started, and we're now 0:09:54.076 --> 0:09:56.596 in a ninth iteration of humanoid and I have only 0:09:56.676 --> 0:09:58.196 raised money in the last couple of years. 0:09:59.516 --> 0:10:04.956 So where'd the idea to build a robot for fifty 0:10:05.036 --> 0:10:07.996 thousand dollars come from? 0:10:08.396 --> 0:10:10.476 The idea of where to build a robot for fifty 0:10:10.516 --> 0:10:13.316 thousand dollars was what will it take for these robots 0:10:13.316 --> 0:10:16.556 to be economic and reach mass market? So, you know, 0:10:17.436 --> 0:10:21.316 when we got started, sort of my view was, you know, 0:10:21.796 --> 0:10:25.636 what will a truly viable commercial humanoid look like and 0:10:26.556 --> 0:10:28.316 what would the bomb costs need to be for this 0:10:28.396 --> 0:10:30.556 to make sense? And if you sort of just do that, 0:10:30.596 --> 0:10:34.636 bottoms up about fifty thousand dollars for the robot, you're 0:10:34.676 --> 0:10:37.596 in the money for mass market. You can still do 0:10:37.676 --> 0:10:40.756 some tasks in a very economic way at even one 0:10:40.796 --> 0:10:43.556 hundred thousand or one hundred and fifty thousand, but fifty 0:10:43.596 --> 0:10:46.396 thousand was the goal. This has now been blown by 0:10:46.516 --> 0:10:48.836 by some of the new entrepreneurs that are coming out 0:10:48.876 --> 0:10:51.836 that are talking about sub twenty thousand dollars. But it 0:10:52.556 --> 0:10:56.036 never made sense to me that robots were as expensive 0:10:56.076 --> 0:10:58.956 as they were at the time. If you look at 0:10:58.996 --> 0:11:02.236 a humanoid compared to a car, there's about four percent 0:11:02.276 --> 0:11:06.116 the raw material by weight. So one of our robots 0:11:06.116 --> 0:11:08.756 there's about three hundred dollars of raw aluminum, which is 0:11:08.796 --> 0:11:13.916 the the base metal of the system. And so it 0:11:14.396 --> 0:11:17.076 never made sense to me that these robots would need 0:11:17.116 --> 0:11:19.716 to be any more than fifty thousand dollars. As you 0:11:19.756 --> 0:11:21.796 could reach scale and as you can start to think 0:11:21.836 --> 0:11:24.436 about new ways of building them in similar ways that 0:11:24.476 --> 0:11:25.556 we build other machines. 0:11:26.836 --> 0:11:29.716 So you decide you want to build a fifty thousand 0:11:29.756 --> 0:11:34.956 dollars robot, Like, what do you actually do to do that? Like, 0:11:35.516 --> 0:11:37.396 how do you go from having an idea of building 0:11:37.396 --> 0:11:39.556 a fifty thousand dollars robot to having a fifty thousand 0:11:39.556 --> 0:11:40.156 dollars robot. 0:11:40.596 --> 0:11:44.156 Well, you iterate until you solve any problem. So what 0:11:44.196 --> 0:11:46.476 we would do is basically we would get a project 0:11:46.556 --> 0:11:51.836 or a contract to build a robot, and we would 0:11:52.196 --> 0:11:56.476 put a lot of different ideas into those designs. In 0:11:56.716 --> 0:11:58.836 early days, it was all about performance. How can you 0:11:58.836 --> 0:12:01.836 get the robot to just do these tasks, to stand 0:12:01.916 --> 0:12:04.876 and have a battery life that's long enough. And then 0:12:04.876 --> 0:12:07.836 as we kept evolving, we started to really focus on 0:12:09.436 --> 0:12:14.356 cost in addition and scalability and assemblability and robustness. And 0:12:14.396 --> 0:12:17.756 the key building block to drive cost and performance is 0:12:17.756 --> 0:12:21.276 the actuator. So I mentioned we've done nine iterations of humanoids, 0:12:21.276 --> 0:12:24.356 but we've done sixty iterations of electric actuators. 0:12:24.716 --> 0:12:27.676 Actuators are basically the thing that makes the robot. 0:12:27.516 --> 0:12:30.836 Move, right, Yeah, the muscles of a robot. 0:12:31.316 --> 0:12:33.716 So where are you now? Tell me about Apollo. 0:12:34.756 --> 0:12:36.996 Yeah, we're now at an exciting point. We have about 0:12:37.036 --> 0:12:40.956 one hundred and seventy employees at Electronic. We are piloting 0:12:41.396 --> 0:12:45.276 these robots right now. So I think the entire industry 0:12:45.356 --> 0:12:48.676 is still in the pilot stage overall. There's some commercial 0:12:48.756 --> 0:12:51.836 orders that are happening, but still early days for humanoids. 0:12:52.956 --> 0:12:55.876 We're working with a handful of really great partners, folks 0:12:55.876 --> 0:12:59.676 like Mercedes and GXO, and we're getting the robots out 0:12:59.676 --> 0:13:03.156 into the real world and in pretty big ways, and 0:13:03.276 --> 0:13:06.396 so we'll have more announcements over this year. We have 0:13:06.436 --> 0:13:08.916 a big partnership with Google Deep Mind, which is something 0:13:08.916 --> 0:13:11.676 that I always dreamed of coming out of graduate school. 0:13:11.716 --> 0:13:13.916 We had a lot of respect for the folks that 0:13:14.436 --> 0:13:17.076 at Google, and they have a whole history and legacy 0:13:17.676 --> 0:13:21.836 in the humanoid space as well. And basically right now 0:13:21.836 --> 0:13:25.316 we're getting these robots out into the world and gearing 0:13:25.476 --> 0:13:29.156 up for you know, real commercialization, which we expect to 0:13:29.196 --> 0:13:31.276 happen in twenty twenty six. 0:13:32.156 --> 0:13:33.276 What's the robot look like? 0:13:33.836 --> 0:13:36.876 The robot kind of looks like a superhero. Maybe this 0:13:36.916 --> 0:13:39.396 has been kind of the idea that we've had from 0:13:39.436 --> 0:13:43.076 the beginning. It's got two eyes and a face. It's 0:13:43.116 --> 0:13:46.156 five foot eight, weighs one hundred and sixty pounds, has 0:13:46.476 --> 0:13:50.116 four hour swappable batteries. Yeah, it's got a screen on 0:13:50.156 --> 0:13:53.716 its chest and a face and that's about it. 0:13:53.796 --> 0:13:57.756 Two arms, two legs. What's to have in the way 0:13:57.756 --> 0:14:00.156 of hands. 0:14:00.156 --> 0:14:03.156 So it has hands, has five fingered hands. You know, 0:14:03.516 --> 0:14:05.916 there's these debates that I think of his false debates 0:14:05.956 --> 0:14:08.516 in the humanoid space. So a lot of people when 0:14:08.516 --> 0:14:11.476 they sort of knock humanoids and the viability of humanoids. 0:14:11.476 --> 0:14:13.516 It usually has to do with do they need legs 0:14:13.556 --> 0:14:15.996 and do they need hands? And the answer to that 0:14:16.076 --> 0:14:18.476 question for me is no, they don't. It's a robot, 0:14:18.556 --> 0:14:21.716 and robots are modular. So we can put Apollo on 0:14:21.796 --> 0:14:24.156 any mobility base. We can put it on wheels, we 0:14:24.156 --> 0:14:27.436 could put it on tracks, we could stationary amount the 0:14:27.476 --> 0:14:31.076 upper torso of Apollo. And the same thing's true for 0:14:31.196 --> 0:14:33.876 the hands of the grippers. We can use parallel grippers, 0:14:34.876 --> 0:14:36.516 or we can use five fingered hands. 0:14:37.036 --> 0:14:39.796 Hands are like a whole thing, right, Like hands are? 0:14:40.436 --> 0:14:42.916 Is it partly because they're so hard? Like what's going 0:14:42.916 --> 0:14:44.036 on with robots and hands? 0:14:44.116 --> 0:14:46.356 Yeah, it turns out hands are a whole thing. This 0:14:46.436 --> 0:14:48.876 is another one of those things that you know. It's 0:14:48.956 --> 0:14:51.876 it's almost better that you don't understand the complexity before 0:14:51.876 --> 0:14:54.676 you get into it, or else you might not have 0:14:54.756 --> 0:14:58.276 done it in the first place. Ninety eight percent of 0:14:58.396 --> 0:15:02.236 all tasks that humans do are done with our hands. 0:15:02.596 --> 0:15:06.836 So there are narrow things that humanoids can do without 0:15:06.956 --> 0:15:10.876 more dexterity, but it's very limited relative to the whole 0:15:11.036 --> 0:15:13.876 sort of you know, all the different types of tasks 0:15:13.916 --> 0:15:17.476 that humans do. Most of the things we do involve 0:15:17.476 --> 0:15:19.716 our hands, and certainly in the industrial space, most of 0:15:19.716 --> 0:15:23.236 the work is done with hands, So solving the end 0:15:23.236 --> 0:15:26.356 defector or the hand problem is a big deal. There's 0:15:26.396 --> 0:15:28.916 a lot of different debates about what you need and 0:15:28.996 --> 0:15:31.756 how you get something that can actually perform industrial work. 0:15:32.876 --> 0:15:36.516 You know, we've chosen the five finger hand route and 0:15:36.556 --> 0:15:40.196 we're working across the space to really make some big 0:15:40.236 --> 0:15:41.196 advancements there. 0:15:41.516 --> 0:15:44.596 Overall, it's part of the trade off, Like I could 0:15:44.596 --> 0:15:47.436 build whatever two what do you call them prongs? Like 0:15:47.476 --> 0:15:50.516 if you had two fingers, basically like a claw, Like 0:15:50.556 --> 0:15:52.236 you could do a lot of things with a claw. 0:15:52.276 --> 0:15:56.076 Presumably it would be way easier, but you couldn't do everything. 0:15:56.156 --> 0:15:58.756 Is it kind of like what are you optimizing for? 0:15:58.836 --> 0:16:00.716 And sort of how much payoff now versus how much 0:16:00.756 --> 0:16:01.356 payoff later? 0:16:01.836 --> 0:16:05.916 Yeah, I think that's exactly right. It's you know, versatility, 0:16:06.836 --> 0:16:11.556 you know, compared to robustness and cost. Basically, how much 0:16:11.596 --> 0:16:13.596 complexity do you want to have on the system. And 0:16:14.076 --> 0:16:16.676 you know, for these robots to be really viable in 0:16:16.676 --> 0:16:19.196 the long run, especially in the industrial space, they've got 0:16:19.236 --> 0:16:22.476 to be able to operate two shifts a day minimum. 0:16:22.956 --> 0:16:25.316 Really you know, twenty two hours a day, seven days 0:16:25.356 --> 0:16:27.996 a week. But solving that problem in a hand, so 0:16:28.116 --> 0:16:30.756 just getting the performance of the hand first, but then 0:16:30.796 --> 0:16:33.796 the robustness for them to do that type of work 0:16:34.156 --> 0:16:35.996 is the next piece. And that's a trade off of 0:16:36.076 --> 0:16:38.116 performance and complexity and cost. 0:16:39.116 --> 0:16:42.436 Because like it gets delicate, right, presumably the fingers so 0:16:42.436 --> 0:16:46.796 to speak, would be fragile, right, Yeah, they can bring 0:16:46.836 --> 0:16:47.876 easy to break. 0:16:47.836 --> 0:16:50.356 Yeah, yeah, yeah, and you've got to maintain it and 0:16:50.476 --> 0:16:53.436 you've got to support those systems and fix them out 0:16:53.476 --> 0:16:55.996 in the field. And so what's the trade off there? 0:16:56.036 --> 0:16:58.396 And that's a whole trade space that we've been working 0:16:58.436 --> 0:16:59.316 on over a long time. 0:17:00.476 --> 0:17:03.556 So we've been talking about hardware. Let's talk about the 0:17:03.596 --> 0:17:06.676 software side. What's happening with that. 0:17:08.076 --> 0:17:10.796 A lot's happening on that side. I think we're really 0:17:10.796 --> 0:17:13.996 in a really you know, we're an exciting point for 0:17:14.116 --> 0:17:17.676 robotics overall. Think of the AI as really the last 0:17:17.676 --> 0:17:19.996 piece of the puzzle. So, you know, we've had the 0:17:20.036 --> 0:17:24.596 ability to build complex robots for a relatively long time. 0:17:25.116 --> 0:17:27.836 We're just now, you know, really figuring out how to 0:17:28.316 --> 0:17:31.836 take the lessons from automotive and consumer electronics and and 0:17:31.956 --> 0:17:35.356 build much more economic systems, and we've had some advancements 0:17:35.356 --> 0:17:39.036 and things like motors and batteries and compute and sensors 0:17:39.076 --> 0:17:41.196 that have all sort of built up to this moment. 0:17:42.036 --> 0:17:44.516 But the final piece of the puzzle was the AI 0:17:44.676 --> 0:17:48.356 and the intelligence and essentially the way to think about it. 0:17:48.396 --> 0:17:50.716 And I think Jensen does a great job of explaining this. 0:17:50.876 --> 0:17:55.436 But the advances Jensen Wong from Nvidio, Yeah, that's right. 0:17:55.756 --> 0:17:58.116 I feel like Jensen is not quite the Elon level 0:17:58.196 --> 0:18:01.796 of one name, household name, but sorry, go on, he's 0:18:01.836 --> 0:18:02.356 getting there. 0:18:02.556 --> 0:18:03.316 Yeah, he should be. 0:18:04.156 --> 0:18:05.236 He should be. He should be. 0:18:07.316 --> 0:18:10.996 Basically, the advancements in geni AI turn out to apply 0:18:11.236 --> 0:18:15.756 very well to robotics, and particularly to humanoid robotics, So 0:18:15.796 --> 0:18:19.596 you can basically map human movement and trajectories from humans 0:18:19.636 --> 0:18:22.956 doing things and build big data sets and use that 0:18:23.076 --> 0:18:26.596 to train robots to do similar tasks in similar environments. 0:18:26.916 --> 0:18:31.076 And these transformer architectures that we're using and generative AI 0:18:31.236 --> 0:18:34.076 actually apply very well to robotics, and so this has 0:18:34.116 --> 0:18:37.516 been a big sort of breakthrough moment for robotics. And 0:18:37.556 --> 0:18:41.156 so I think as an industry as a whole, everybody's 0:18:41.196 --> 0:18:44.596 really excited right now because we're reaching new heights and 0:18:44.636 --> 0:18:47.596 we're able to do things that we dreamed about doing 0:18:47.636 --> 0:18:50.036 with robots only you know, even a few years ago, 0:18:50.156 --> 0:18:53.276 are now possible, and we're seeing a really rapid advancement 0:18:53.316 --> 0:18:54.476 in performance overall. 0:18:54.916 --> 0:18:57.236 What's an example of a thing that you could only 0:18:57.316 --> 0:18:59.996 dream of a few years ago that robots can do now? 0:19:00.556 --> 0:19:03.436 I think it's more dexterity and versatility. So just the 0:19:03.596 --> 0:19:06.476 range of things that you can do. So the challenge 0:19:06.476 --> 0:19:10.236 for robotics was that each news even if you build 0:19:10.236 --> 0:19:12.836 something like a humanoid robot, and this is true for us, 0:19:13.156 --> 0:19:18.116 and say you build an application to pick boxes off 0:19:18.156 --> 0:19:21.716 of a palette and place those boxes onto a conveyor. Well, 0:19:21.716 --> 0:19:24.956 you hand build that application, and you know, maybe takes 0:19:24.996 --> 0:19:27.676 you eighteen months to sort of ring that out and 0:19:27.796 --> 0:19:30.476 get it to a certain amount of robustness. Well, now 0:19:30.516 --> 0:19:33.476 you want to do the reverse of that and pick 0:19:33.516 --> 0:19:36.916 off of a conveyor and palletize something that will take 0:19:36.916 --> 0:19:39.636 you the same amount of time that it took you 0:19:39.716 --> 0:19:41.636 to build the initial application. 0:19:41.516 --> 0:19:43.636 You have to write basically have to write a whole 0:19:43.676 --> 0:19:46.436 other piece of software. You have to start from scratch 0:19:46.436 --> 0:19:47.276 almost yeah. 0:19:47.156 --> 0:19:51.716 Yeah, exactly. And so basically what is happening now is 0:19:51.716 --> 0:19:55.196 that we now have these much more sort of general 0:19:55.556 --> 0:19:58.396 models that where you can collect a lot of data 0:19:58.436 --> 0:20:01.356 at the top layer, and so each new task that 0:20:01.436 --> 0:20:04.876 you want to perform actually takes less and less incremental 0:20:04.876 --> 0:20:08.076 amount of work. So what it's opening up now is 0:20:08.196 --> 0:20:11.636 more dexterous applications. 0:20:13.316 --> 0:20:16.436 Still to come. On the show how Jeff's grandparents inspired 0:20:16.476 --> 0:20:27.916 his work on robots, So you were talking about using 0:20:27.956 --> 0:20:32.836 the transformer model. That has been the you know, breakthrough 0:20:32.876 --> 0:20:37.556 that has driven large language models in training robots essentially. 0:20:38.236 --> 0:20:41.836 I mean, of course, a key sort of serendipitous thing 0:20:41.836 --> 0:20:46.156 that happened with language models was there is this crazy 0:20:46.276 --> 0:20:49.596 large data set of words and pictures which is the Internet, 0:20:49.956 --> 0:20:54.236 and there's not an analogous data set for the physical world. Right, Yeah, 0:20:54.756 --> 0:20:58.676 it seems like that is is that the rate limiting step? 0:20:58.796 --> 0:21:02.916 Is that the big problem in sort of AI for robots? 0:21:03.236 --> 0:21:05.396 Yeah, I mean there's a lot of work that's still 0:21:05.436 --> 0:21:08.396 happening at the research level for you know, how can 0:21:08.436 --> 0:21:11.876 you pull that kind of data from from videos so 0:21:11.916 --> 0:21:14.236 you can think of big data sets of humans doing 0:21:14.276 --> 0:21:17.556 things that could be really interesting to train robots in 0:21:17.596 --> 0:21:20.836 the future. Interest and that that will come into play 0:21:20.836 --> 0:21:23.636 over time. But yeah, it's the chicken or the egg problem. 0:21:23.796 --> 0:21:27.516 And data is the one of the key things that 0:21:27.556 --> 0:21:31.076 we need to enable the next wave of breakthroughs. And 0:21:31.076 --> 0:21:33.316 and this is kind of the race, is can you 0:21:33.316 --> 0:21:36.196 get robots out into the to the real world, into 0:21:36.236 --> 0:21:39.796 the field and collecting data very high you know, in 0:21:39.916 --> 0:21:43.236 very high volumes. Whoever does that, you know, will we'll 0:21:43.236 --> 0:21:46.276 have better models. And this is the data flywheel. So 0:21:46.636 --> 0:21:48.436 this is kind of the race that's on right now 0:21:48.436 --> 0:21:51.356 where you hear a lot of other humanoid CEOs talking 0:21:51.356 --> 0:21:53.516 about getting a lot of robots out into the world. 0:21:54.116 --> 0:21:57.196 Largely those are going to be under Teley operation collecting 0:21:57.276 --> 0:22:00.076