AI transcript
0:00:06 [MUSIC]
0:00:07 Pushkin.
0:00:11 [BLANK_AUDIO]
0:00:14 As a general matter, I’m a fan of technological progress.
0:00:19 But I’ll admit that humanoid robots kind of creep me out for
0:00:23 obvious, normy, uncanny valley type reasons.
0:00:29 And yet, there is an exchange that you’ll hear near the end of today’s show
0:00:33 that’s the most compelling argument I’ve ever heard for humanoid robots.
0:00:36 And it’s not just an intellectual argument.
0:00:39 It’s an emotional argument, if that’s a phrase.
0:00:44 It’s really a very human argument for humanoid robots.
0:00:50 [MUSIC]
0:00:52 I’m Jacob Goldstein, and this is What’s Your Problem,
0:00:56 the show where I talk to people who are trying to make technological progress.
0:00:59 My guest today is Jeff Cardenas.
0:01:02 He’s the co-founder and CEO of Aptronic.
0:01:04 Jeff’s problem is this.
0:01:10 Can you make a safe, reliable humanoid robot for less than $50,000?
0:01:15 We started our conversation talking about the DARPA Robotics Challenge.
0:01:19 DARPA, of course, is the government agency that helped to create the internet and
0:01:23 that gave a big push to early self-driving cars, among other things.
0:01:27 And Jeff says, the agency’s Robotics Challenge, which happened a decade ago,
0:01:32 happened in 2015, played a key role in launching a bunch of the companies
0:01:34 that are now working on humanoid robots.
0:01:41 >> The DARPA Robotics Challenge was a challenge that was created in the wake
0:01:43 of the Fukushima disaster.
0:01:48 Fukushima happened and there was a meltdown in the nuclear reactors.
0:01:51 And it was unsafe for people to go in.
0:01:56 And essentially what happened was they needed a robot to go in to sort of
0:01:59 prevent a broader catastrophe.
0:02:02 And as they went out to the robotics community,
0:02:09 the idea was calling all roboticists, we need a robot to go in and to help out here.
0:02:13 And there was no robots that could do all the different tasks that were required
0:02:15 to actually get to the meltdown reactors.
0:02:19 So you had to go down steps, you had to open doors, you had to do a whole range of
0:02:24 things and in the wake of that, basically what DARPA said was,
0:02:28 certainly there’s got to be the technology to enable us to have much more
0:02:32 versatile robots for natural disaster relief.
0:02:34 So this never happens again.
0:02:38 So out of that, DARPA created something called the DARPA Robotics Challenge.
0:02:42 And so there’s a variety of teams around the world that were put together
0:02:47 to build these general purpose robots.
0:02:51 And the team that we came out of was the NASA Johnson Space Center team
0:02:54 working on a robot called Valkyrie.
0:02:59 >> So I want to go back to this moment when the DARPA Challenge ends,
0:03:05 there’s this big final contest and what we have from it is not some incredible
0:03:09 breakthrough, but a blooper reel of robots, what are they doing?
0:03:12 Falling down stairs or driving cars into walls or something?
0:03:19 >> Pretty much, I mean, the blooper reels make it seem worse than it was.
0:03:24 But yeah, we had, basically the realization was the technology’s not there.
0:03:29 It’s going to take time to continue to mature till it can get to the point where
0:03:31 it’s actually commercially viable.
0:03:34 >> And so it’s interesting, it’s super interesting, right,
0:03:39 that this moment is not the beginning of some humanoid robot winter,
0:03:43 but rather the beginning of this humanoid robot industry.
0:03:46 How does that work, how do people, whatever, start companies and
0:03:49 get money out of this seemingly disappointing moment?
0:03:55 >> Well, I think it actually was a winter, when we started and
0:03:59 the company was incorporated in 2015 and we started in 2016.
0:04:04 And for the most part, nobody wanted to talk about humanoids and
0:04:06 nobody was really paying attention to it.
0:04:10 There was a handful of folks that I sort of think of as the true believers
0:04:15 that were really working on this problem and thought, we don’t care how long
0:04:18 this is going to take, we’re just going to keep working on this no matter what.
0:04:24 And but for the most part, the entire robotics industry was very anti-humanoids
0:04:26 coming out of the DARPA Robotics Challenge.
0:04:30 And in fact, there was many people that were saying they’ll never be viable.
0:04:32 Like, why would you ever use a humanoid robot?
0:04:36 They’re too complicated, they’re way too expensive, and
0:04:38 you’ll always use a simpler robot.
0:04:41 So actually, most of the people that we met when we decided to start
0:04:46 Eptronic were doubters, and we’re saying, humanoids will never make sense.
0:04:50 We’ll use these special purpose robots, maybe in 50 years,
0:04:52 humanoids will make sense, but not for a long time.
0:04:58 >> I mean, special purpose robots is a pretty compelling case, right?
0:05:03 Like, whatever, we all have dishwashing robots and
0:05:05 clothes cleaning robots in our houses.
0:05:09 And wheels seem way easier than legs for lots of things.
0:05:13 And obviously, there have been robot arms for, what, I don’t know,
0:05:15 70 years now or something.
0:05:18 So robots, in a way, are all around us.
0:05:21 Why would you build a machine that looks like a dude when that’s wildly hard, right?
0:05:26 >> Yeah, I mean, I was naive coming out of graduate school.
0:05:28 And to me, it seemed obvious.
0:05:32 And the way I used to think about it was you could either have thousands of robots
0:05:35 that do one thing, or you could build one robot that could do thousands of
0:05:37 different things.
0:05:40 And when I would talk about this with Nick, my co-founder,
0:05:45 Nick would say, look, you can either invest all of this engineering and
0:05:47 each of these sort of narrow solutions.
0:05:52 Or yes, a humanoid robot, a viable humanoid robot, could take you years.
0:05:53 It could take you a decade.
0:05:57 But once you invest all this time in this single platform,
0:06:01 then you can reap the benefits of that across, you can spread.
0:06:06 The research of that across many different applications.
0:06:11 >> I mean, is there not a middle case where there’s some core kind of
0:06:16 functionality that you develop that works across many different types of robots?
0:06:20 Is that a less straw manny version of the non-humanoid robot kind of argument?
0:06:23 >> I think there could be, but I came into robotics and
0:06:27 basically just saw a lot of challenges with the business models.
0:06:30 So you build this special purpose robot.
0:06:33 You custom program the robot in the industrial space.
0:06:39 You can spend six times the price of the robot on just systems integration.
0:06:41 And the robot just does one thing.
0:06:44 So this idea that you could have a much more versatile robot, to me,
0:06:45 seemed obvious.
0:06:47 Like if robotics is going to scale,
0:06:50 we have to have much more versatile robots than we’ve had in the past.
0:06:55 So if you sort of think of that as the premise is we need more versatile robots.
0:06:57 Then the question is, well, how do you get there?
0:06:59 And what does versatility mean?
0:07:04 And that’s where it led me to the humanoid making a lot of sense.
0:07:07 Because if you have to modify the environment for
0:07:12 every new task that the robot can do, you run into the same problem that we had
0:07:14 in sort of classical robotics.
0:07:18 But if the robot can retrofit into the environment such that you don’t have to
0:07:23 change or modify the environment for every new task that the robot can do,
0:07:27 then it seemed to me that this would maybe be the key unlock for
0:07:29 robotics to actually scale to the masses.
0:07:32 The demand would be infinite if you had a thing that was the size and
0:07:35 shape of a person with arms and legs.
0:07:39 Like scale would be off the charts and presumably that’s what drives costs down.
0:07:42 And that’s like the good flywheel, right?
0:07:43 Yeah, exactly.
0:07:47 So okay, so you had this big idea about humanoid robots and you started a company.
0:07:50 But at the moment you started the humanoid robot company,
0:07:54 the prevailing sentiment was like deeply skeptical.
0:07:54 What happened?
0:07:55 What did you do?
0:07:57 Well, a handful of us kept working on it.
0:08:00 So I didn’t know any better.
0:08:02 Sometimes it’s better that you don’t know any better.
0:08:07 I thought humanoids were really cool and I thought that it just seemed,
0:08:12 it made sense to me that how are we gonna get to millions of robots that are
0:08:16 working and with and around humans in all these environments?
0:08:19 And to me, this seemed like the only way that that was gonna happen.
0:08:24 And the way I looked at it was even if we failed, this was a worthy pursuit.
0:08:27 And I would be proud that I tried to do it.
0:08:31 And so the way that we did it was we bootstrapped the company.
0:08:35 There was no investors that were willing to invest in humanoid robots at
0:08:39 the time that we got started, especially for hardware,
0:08:42 which we can talk about that as we move forward.
0:08:45 And so we bootstrapped the company.
0:08:49 And we basically got paid to build robots for a lot of different folks.
0:08:54 And for the first five years of the company, we just built the company on revenue.
0:08:58 We would get project after project and somehow never died.
0:09:01 >> Like what kind of jobs were you taking at that time?
0:09:02 What’s an example?
0:09:04 >> Well, our first contract was with NASA.
0:09:07 So we had a contract with NASA to build Valkyrie II,
0:09:10 to take the lessons learned from the DARPA challenge and
0:09:12 build the next iteration of Valkyrie.
0:09:16 We were really kind of pioneering new ways of building these systems.
0:09:20 So US Special Forces ended up coming to us about a year in and
0:09:23 said, hey, we wanna do Ironman suits.
0:09:27 And our view was this was kind of a humanoid robot that you wear.
0:09:28 We worked in automotive.
0:09:30 We helped build humanoid robots for
0:09:37 a couple major companies that are still working on these things today.
0:09:40 And we would help sort of pioneer new ways of building their platforms.
0:09:44 So we’ve done 15 unique robots since we got started.
0:09:46 And we’re now on a ninth iteration of humanoid.
0:09:49 And I’ve only raised money in the last couple of years.
0:09:57 >> So where did the idea to build a robot for $50,000 come from?
0:10:03 >> The idea of where to build a robot for $50,000 was what will it take for
0:10:06 these robots to be economic and reach mass market?
0:10:13 So when we got started, sort of my view was what will
0:10:16 a truly viable commercial humanoid look like?
0:10:20 And what would the bomb cost need to be for this to make sense?
0:10:24 And if you sort of just do that bottoms up about $50,000 for
0:10:27 the robot, you’re in the money for mass market.
0:10:32 You can still do some tasks in a very economic way at even 100,000 or
0:10:36 150,000, but 50,000 was the goal.
0:10:39 This has now been blown by by some of the new entrepreneurs that are coming out
0:10:42 that are talking about sub $20,000.
0:10:48 But it never made sense to me that robots were as expensive as they were at the time.
0:10:51 If you look at a humanoid compared to a car,
0:10:55 there’s about 4% the raw material by weight.
0:10:59 So in one of our robots, there’s about $300 of raw aluminum,
0:11:02 which is the base metal of the system.
0:11:09 And so it never made sense to me that these robots would need to be any more
0:11:11 than $50,000 as you could reach scale.
0:11:14 And as you could start to think about new ways of building them and
0:11:16 similar ways that we build other machines.
0:11:21 >> So you decide you want to build a $50,000 robot.
0:11:24 Like, what do you actually do to do that?
0:11:31 How do you go from having an idea of building a $50,000 robot to having a $50,000 robot?
0:11:34 >> Well, you iterate until you solve any problem.
0:11:37 So what we would do is basically we would get a project or
0:11:40 a contract to build a robot.
0:11:47 And we would put a lot of different ideas into those designs.
0:11:49 And in early days, it was all about performance.
0:11:53 How can you get the robot to just do these tasks, to stand, and
0:11:55 have a battery life that’s long enough.
0:12:01 And then as we kept evolving, we started to really focus on cost in addition,
0:12:05 and scalability, and assemblability, and robustness.
0:12:09 And the key building block to drive cost and performance is the actuator.
0:12:12 So I mentioned we’ve done nine iterations of humanoids, but
0:12:15 we’ve done 60 iterations of electric actuators.
0:12:19 >> Actuators are basically the thing that makes the robot move, right?
0:12:20 >> Yeah, the muscles of a robot.
0:12:23 >> So where are you now?
0:12:24 Tell me about Apollo.
0:12:27 >> Yeah, we’re now at an exciting point.
0:12:30 We have about 170 employees at Aptronic.
0:12:33 We are piloting these robots right now.
0:12:38 So I think the entire industry is still in the pilot stage overall.
0:12:40 There’s some commercial orders that are happening, but
0:12:42 still early days for humanoids.
0:12:47 We’re working with a handful of really great partners, folks like Mercedes and
0:12:49 GXO.
0:12:54 And we’re getting the robots out into the real world, and it’s pretty big ways.
0:12:57 So we’ll have more announcements over this year.
0:12:59 We have a big partnership with Google DeepMind,
0:13:02 which is something that I always dreamed of coming out of graduate school.
0:13:06 We had a lot of respect for the folks at Google, and
0:13:10 they have a whole history and legacy in the humanoid space as well.
0:13:15 And basically right now, we’re getting these robots out into the world and
0:13:22 gearing up for real commercialization, which we expect to happen in 2026.
0:13:24 >> What’s the robot look like?
0:13:27 >> The robot kind of looks like a superhero, maybe.
0:13:31 This has been kind of the idea that we’ve had from the beginning.
0:13:33 It’s got two eyes and a face.
0:13:39 It’s five foot eight, weighs 160 pounds, has four hour swappable batteries.
0:13:44 Yeah, it’s got a screen on its chest and a face, and that’s about it.
0:13:51 >> Two arms, two legs, what’s it have in the way of hands?
0:13:53 >> So it has hands, it has five-fingered hands.
0:13:57 There’s these debates that I think of as false debates in the humanoid space.
0:14:01 So a lot of people, when they sort of knock humanoids and
0:14:03 the viability of humanoids, it usually has to do with,
0:14:06 do they need legs and do they need hands?
0:14:08 And the answer to that question for me is no, they don’t.
0:14:10 It’s a robot, and robots are modular.
0:14:14 So we can put Apollo on any mobility base, we can put it on wheels,
0:14:20 we can put it on tracks, we could stationary mount the upper torso of Apollo.
0:14:23 And the same thing’s true for the hands or the grippers.
0:14:27 We can use parallel grippers, or we can use five-fingered hands.
0:14:29 >> Hands are like a whole thing, right?
0:14:33 Like hands are, is it partly because they’re so hard?
0:14:35 Like what’s going on with robots and hands?
0:14:37 >> It turns out hands are a whole thing.
0:14:41 This is another one of those things that, it’s almost better that you don’t
0:14:45 understand the complexity before you get into it or else you might not
0:14:47 have done it in the first place.
0:14:53 98% of all tasks that humans do are done with our hands.
0:14:59 So there are narrow things that humanoids can do without more dexterity.
0:15:03 But it’s very limited relative to the whole sort of,
0:15:05 all the different types of tasks that humans do.
0:15:08 Most of the things we do involve our hands.
0:15:12 And certainly in the industrial space, most of the work is done with hands.
0:15:17 So solving the end effector or the hand problem is a big deal.
0:15:19 There’s a lot of different debates about what you need and
0:15:22 how you get something that can actually perform industrial work.
0:15:27 We’ve chosen the five-finger hand route and
0:15:32 we’re working across the space to really make some big advancements there overall.
0:15:37 >> It’s part of the trade off like I could build whatever two,
0:15:38 what do you call them, prongs?
0:15:41 Like if you had two fingers basically, like a claw?
0:15:43 Like you could do a lot of things with a claw.
0:15:47 Presumably it would be way easier, but you couldn’t do everything.
0:15:49 Is it kind of like what are you optimizing for and
0:15:52 sort of how much payoff now versus how much payoff later?
0:15:54 >> Yeah, I think that’s exactly right.
0:16:00 It’s versatility compared to robustness and
0:16:04 cost basically, how much complexity do you want to have on the system?
0:16:09 And for these robots to be really viable in the long run, especially in the industrial space,
0:16:13 they got to be able to operate two shifts a day minimum,
0:16:16 really 22 hours a day, seven days a week.
0:16:18 But solving that problem in a hand, so
0:16:23 just getting the performance of the hand first, but then the robustness for
0:16:26 them to do that type of work is the next piece and
0:16:28 that’s a trade off of performance and complexity and cost.
0:16:31 >> Cuz like it gets delicate, right?
0:16:37 Presumably the fingers, so to speak, would be fragile, right?
0:16:37 >> Yeah, they can break.
0:16:39 >> Easy to break, yeah.
0:16:41 >> Yeah, yeah, and you got to maintain it and
0:16:44 you got to support those systems and fix them out in the field.
0:16:46 And so what’s the trade off there?
0:16:50 And that’s a whole trade space that we’ve been working on over a long time.
0:16:53 So we’ve been talking about hardware.
0:16:57 Let’s talk about the software side, what’s happening with that?
0:17:00 >> A lot’s happening on that side.
0:17:06 I think we’re really an exciting point for robotics overall.
0:17:09 Think of the AI as really the last piece of the puzzle.
0:17:16 So we’ve had the ability to build complex robots for a relatively long time.
0:17:20 We’re just now really figuring out how to take the lessons from automotive and
0:17:25 consumer electronics and build much more economic systems.
0:17:28 And we’ve had some advancements in things like motors and batteries and
0:17:32 compute and sensors that have all sort of built up to this moment.
0:17:37 But the final piece of the puzzle was the AI and the intelligence.
0:17:39 And essentially the way to think about it, and
0:17:43 I think Jensen does a great job of explaining this, but the advance.
0:17:45 >> Jensen Wong from NVIDIA.
0:17:46 >> Yeah, that’s right.
0:17:51 I feel like Jensen is not quite the Elon level of one name, household name.
0:17:53 But sorry, go on, he’s getting there.
0:17:55 >> Yeah, he should be.
0:17:58 >> He should be, he should be.
0:18:02 >> Basically the advancements in generative AI turn out to apply very well
0:18:06 to robotics and particularly to humanoid robotics.
0:18:08 So you can basically map human movement and
0:18:13 trajectories from humans doing things and build big data sets and
0:18:17 use that to train robots to do similar tasks in similar environments.
0:18:22 And these transformer architectures that we’re using in generative AI
0:18:24 actually apply very well to robotics.
0:18:28 And so this has been a big sort of breakthrough moment for robotics.
0:18:31 And so I think as an industry as a whole,
0:18:35 everybody’s really excited right now because we’re reaching new heights and
0:18:40 we’re able to do things that we dreamed about doing with robots only even a few
0:18:44 years ago are now possible and we’re seeing a really rapid advancement in
0:18:45 performance overall.
0:18:49 >> What’s an example of a thing that you could only dream of a few years ago
0:18:51 that robots can do now?
0:18:53 >> I think it’s more dexterity and versatility.
0:18:55 So just the range of things that you can do.
0:19:00 So the challenge for robotics was that each new task,
0:19:04 even if you build something like a humanoid robot, and this is true for us.
0:19:09 And say you build an application to pick boxes off of a palette and
0:19:12 place those boxes onto a conveyor.
0:19:15 Well, you hand build that application and
0:19:18 maybe takes you 18 months to sort of wring that out and
0:19:21 get it to a certain amount of robustness.
0:19:25 Well, now you want to do the reverse of that and pick off of a conveyor and
0:19:27 palletize something.
0:19:31 That will take you the same amount of time that it took you to build the initial
0:19:32 application.
0:19:35 >> You have to basically have to write a whole other piece of software.
0:19:38 You have to start from scratch almost, yeah.
0:19:39 >> Yeah, exactly.
0:19:45 And so basically what is happening now is that we now have these much more
0:19:50 sort of general models that where you can collect a lot of data at the top layer.
0:19:54 And so each new task that you want to perform actually takes less and
0:19:57 less incremental amount of work.
0:20:00 So what is opening up now is more dexterous applications.
0:20:04 [MUSIC]
0:20:05 >> Still to come on the show,
0:20:08 how Jeff’s grandparents inspired his work on robots.
0:20:17 [MUSIC]
0:20:22 So you were talking about using the transformer model that has been the
0:20:27 breakthrough that has driven large language models in training robots,
0:20:29 essentially.
0:20:33 I mean, of course, a key sort of serendipitous thing that happened with
0:20:39 language models was there is this crazy large data set of words and
0:20:40 pictures, which is the internet.
0:20:45 And there’s not an analogous data set for the physical world, right?
0:20:45 >> Yeah.
0:20:49 >> It seems like that is, is that the rate limiting step?
0:20:54 Is that the big problem in sort of AI for robots?
0:20:57 >> Yeah, I mean, there’s a lot of work that’s still happening at the research
0:21:02 level for how can you pull that kind of data from videos.
0:21:06 So you can think of big data sets of humans doing things that can be really
0:21:09 interesting to train robots in the future.
0:21:10 >> Interesting.
0:21:12 >> And that will come into play over time.
0:21:17 But yeah, it’s the chicken or the egg problem and data is one of the key
0:21:21 things that we need to enable the next wave of breakthroughs.
0:21:26 And this is kind of the race is can you get robots out into the real world,
0:21:32 into the field and collecting data at very high, in very high volumes.
0:21:35 Whoever does that will have better models.
0:21:37 This is the data flywheel.
0:21:40 So this is kind of the race that’s on right now where you hear a lot of
0:21:45 other humanoid CEOs talking about getting a lot of robots out into the world.
0:21:48 Largely, those are going to be under tele-operation, collecting data and
0:21:52 then training and building these models of the future.
0:21:57 >> So it’s like whoever gets there first will win just because that’ll be the
0:21:58 accelerant.
0:22:02 Like once you have robots out in the world and you’re collecting data,
0:22:05 then you’re immediately getting ahead of whoever has fewer robots out in the world
0:22:07 because they’re collecting less data.
0:22:07 >> Yeah.
0:22:11 >> So tell me about tele-operation.
0:22:16 >> Tele-operation is basically just remotely controlling the robot.
0:22:18 So you’re taking over the robot.
0:22:22 You can see through the robot’s eyes with the VR headset and
0:22:27 then you’re controlling the robot’s arms and hands to do a particular task.
0:22:29 It’s like a video game and you’re controlling a robot.
0:22:33 It’s a simple idea.
0:22:35 There’s a couple of reasons it’s important.
0:22:38 The first thing is that it tells you what the robot’s physically capable of doing.
0:22:44 So if I’m completely controlling the robot and I can’t do a task under tele-operation,
0:22:46 then that means the robot’s not physically capable of doing it.
0:22:51 So it’d be very difficult for an AI control system to control the robot to do that.
0:22:56 So this is how we understand the physical capabilities of the robot.
0:23:01 As these new models have come along, the simple idea is that if you can tele-operate
0:23:06 the robot to do a task, then you should be able to automate that task on the other end.
0:23:10 So if you can collect enough data under tele-operation,
0:23:16 then you can automate it by running it through these similar architectures that we talked about.
0:23:22 >> So it’s the basic idea that you use remote control to drive the robot to do a thing,
0:23:26 whatever, a thousand times, some number of times.
0:23:28 In doing that, you’re training the robot.
0:23:31 You’re training the software, training the AI.
0:23:32 >> Yeah, that’s exactly right.
0:23:35 >> What’s an example of a thing that you’ve done that way,
0:23:39 and how many times did you have to remote control it before the robot could do it?
0:23:41 >> So each is picking is a good example,
0:23:44 or you’re taking objects and you’re putting them into a box.
0:23:47 To do that in a simple context,
0:23:52 thousands of demonstrations is what you need.
0:23:55 We think of this as generally hours.
0:24:00 So how many hours of data collection have we done?
0:24:04 And thousands of iterations can get you to,
0:24:06 let’s say 80 percent of human rate.
0:24:10 If you want to get to 95 percent or better of human rate,
0:24:12 then you need more and more data.
0:24:14 But it’s in the thousands, it’s not millions.
0:24:18 >> Yeah, thousands makes it seem totally tractable.
0:24:22 >> Yeah, I was actually surprised by how well these models work,
0:24:26 and actually how little data they need to get relatively good performance.
0:24:30 You’re seeing a lot of demonstrations of this out there today.
0:24:33 >> Presumably, that’ll get better and better.
0:24:37 As the software side of AI gets better and better,
0:24:39 it’ll learn faster essentially.
0:24:40 The other obvious thing,
0:24:43 but I’m just going to say it is once you have done it once,
0:24:45 then it works for every robot.
0:24:48 Then you can make a million robots and they all know how to do the thing.
0:24:49 >> Yeah, that’s exactly right.
0:24:52 One of the interesting things about these models is actually
0:24:56 the diversity of data is almost more important than task-specific data.
0:24:59 So we want to go wide across a range of tasks.
0:25:00 >> Interesting.
0:25:03 >> Then you’re basically building all these skills into the robot,
0:25:07 and then it becomes better at doing any one particular task.
0:25:09 >> It really is like learning.
0:25:12 It really is human-esque.
0:25:13 >> Yeah, that’s right.
0:25:17 >> So I know you’re in a few pilot projects with Mercedes and with,
0:25:20 what is it, GXO, Big Logistics Company.
0:25:23 When do you want to start selling robots for real?
0:25:24 When do you think that might happen?
0:25:26 >> 2026.
0:25:27 >> Okay.
0:25:27 >> Yeah.
0:25:30 >> Suddenly that’s next year, almost now.
0:25:34 >> Years of long time and these are dog years.
0:25:38 It’s a long time in this space.
0:25:43 >> 2026 could be almost two years for now.
0:25:47 Who are you going to sell robots to and how much are you going to charge?
0:25:48 What are they going to do?
0:25:52 >> So initially in manufacturing and logistics,
0:25:54 so folks like Mercedes and GXO,
0:25:59 these are the initial customers of these systems.
0:26:02 We are not announcing pricing yet,
0:26:08 but you can think of it as take what it costs to do these tasks today,
0:26:12 and with some discount to what it costs to do these tasks today.
0:26:15 We have a RAS model that we use.
0:26:18 >> Robot as a service?
0:26:20 >> Yeah, robot as a service model,
0:26:25 where you’re paying the robot basically by the hour
0:26:27 effectively to do a particular task,
0:26:31 and that’s at a discount to what it costs to do that task today.
0:26:34 >> How far are you from the $50,000 robot?
0:26:39 >> We’re not there yet, so not very far.
0:26:41 So we have the architecture to be able to do this.
0:26:45 So getting the cost down on these robots is a two-step process.
0:26:49 So first step is new architectures.
0:26:52 So if you still require this very high precision in the system,
0:26:56 and you’re using bespoke components that are only used for robotics,
0:26:59 these robots will still be expensive.
0:27:01 The challenge of humanoid robots is they have
0:27:03 a lot more motors than traditional robots.
0:27:06 So a traditional robot has six or seven motors,
0:27:10 a humanoid robot has 30 to 40 plus.
0:27:14 >> Okay, so that means it’s expensive or you got to figure out how to
0:27:16 get cheaper actuators.
0:27:17 >> Yeah, so we’re there. So for us,
0:27:23 that was a $500 actuator and we have a $500 actuator now today.
0:27:25 So once you solve that problem,
0:27:27 and once you solve the architecture problem,
0:27:29 now it’s about scale and manufacturing.
0:27:31 So a lot of where we spend,
0:27:35 a lot of where the cost is driven at low volumes,
0:27:37 is in just the structures of the robot,
0:27:38 where we’re seeing and seeing,
0:27:42 we’re milling at a big blocks of metal,
0:27:45 parts and very small quantities.
0:27:50 But there’s other techniques that are much more cost-effective,
0:27:53 like casting or stamping,
0:27:56 and these will allow these robots to be much cheaper.
0:27:59 As I mentioned, look at automotive and look at the scale of automotive.
0:28:01 There’s 4 percent the raw material by
0:28:05 weight in a humanoid robot as compared to a car.
0:28:09 So once you solve the architecture problem such that you can
0:28:13 build a lot of these systems and they’re simpler to make,
0:28:15 then the next piece is just applying
0:28:18 mass manufacturing approaches to
0:28:20 this to make them a lot cheaper as you scale.
0:28:26 >> Well, I mean, that’s a hard leap to make, right?
0:28:27 Like, what do you do?
0:28:30 You get a ton of capital and just build
0:28:33 a factory and hope there’s demand on the other end.
0:28:36 Like, how do you go from this bespoke expensive thing to
0:28:39 a mass produced much less expensive thing?
0:28:40 >> Well, it’s a gradient.
0:28:42 So like I said, step one is you have
0:28:45 new approaches that allow you to make them
0:28:48 cheaper just inherently on a unit-to-unit basis.
0:28:51 So the early humanoids were like millions of dollars,
0:28:53 and now we’re in the hundreds of
0:28:55 thousands of dollars range for building one.
0:28:58 >> So you just got to get one more order of magnitude out of it.
0:29:02 >> Yes. We’ve already dropped the price by an order of magnitude.
0:29:04 Then now as we build more,
0:29:06 even as you add a zero,
0:29:08 as you go from 10 to 100,
0:29:10 the price drops pretty dramatically.
0:29:14 So you don’t need the volume that you might think.
0:29:15 My view is we can get to
0:29:20 the sub $50,000 price point in the thousands of unit quantity.
0:29:23 So without hundreds of thousands or millions of these.
0:29:25 >> So one big buyer,
0:29:28 one big car company or logistics company
0:29:31 might place an order of thousands of units, right?
0:29:34 >> Yeah. You made a comment that you said,
0:29:35 and you hope that there’s demand.
0:29:37 One of the things I think it’s important to
0:29:41 note is the demand for these robots is enormous.
0:29:45 We have demand for hundreds of thousands of units already
0:29:48 today with the customers that we’re working with.
0:29:51 So the demand is enormous.
0:29:53 So we’re ramping up.
0:29:55 We’ve got to get the robustness and
0:29:58 the safety of the system and really bring out the design,
0:30:02 and these are really credible,
0:30:04 thoughtful people that are coming from
0:30:06 other industries that are now joining us,
0:30:08 that now see that we’ve crossed
0:30:12 this threshold of technical viability,
0:30:16 and now taking lessons from how you scale and
0:30:17 manufacture other things and bringing that
0:30:21 into the robotic space and the humanoid space overall.
0:30:24 >> So in a year or at least next year,
0:30:27 you want to be selling robots for real.
0:30:31 Where do you want to be in five, say, years?
0:30:34 >> My view is that where this evolves is it’s
0:30:36 going to start in logistics and manufacturing,
0:30:38 and then as we solve safety as an industry,
0:30:40 I’m really interested in health care,
0:30:43 and particularly in elder care over time.
0:30:49 So how can these robots be used to improve the way we live and work?
0:30:53 That was really the lens that I came into this on.
0:30:55 So I think over the next five years,
0:30:59 you’ll start to see the early stages of the next three years.
0:31:04 You’ll start to see early applications for robots entering the home.
0:31:06 There’s some folks that are really working hard on this.
0:31:11 I think we’re about three years out from that being really viable,
0:31:13 but I hope people prove me wrong.
0:31:15 I hope it’s faster than that.
0:31:16 >> Three years seems fast.
0:31:21 What’s the first use case,
0:31:23 first job you imagine a robot doing for
0:31:26 real in somebody’s house in three years?
0:31:28 >> Well, everybody wants laundry.
0:31:30 If everybody I talked to says,
0:31:32 “When is this thing going to do my laundry?”
0:31:33 I want that as well.
0:31:36 >> There’s literally already a machine to do your laundry.
0:31:39 All you have to do is put it in one machine and then put it in another.
0:31:41 The remaining work is trivial.
0:31:42 >> Yeah. I mean, look,
0:31:45 I’m not the person to talk about the home.
0:31:47 I think we’re still a ways out,
0:31:52 but there’s humanoid companies like 1X that are really focused on the home,
0:31:55 and we’ve got a lot of respect for what they’re doing over there,
0:31:57 and so I hope they do it.
0:31:59 I know that they’re working hard on it,
0:32:05 and I want a robot for the home as well.
0:32:09 A lot of the things that are happening with these models that I talked about,
0:32:11 these more generic models,
0:32:12 the things that we’re learning in
0:32:16 the industrial base can apply to the home over time as well.
0:32:18 >> In terms of the AI models, sure.
0:32:22 I mean, the AI models are basically teaching a robot how to deal with the physical world.
0:32:22 >> That’s right.
0:32:24 >> How to move around, how to pick things up,
0:32:26 how to put things down.
0:32:31 >> Yeah. The task in the home is tough because even a robot does your dishes,
0:32:34 your laundry, cleans, and cooks for you.
0:32:38 How much are you willing to pay for that on a yearly basis?
0:32:41 >> I’m imagining the first household tasks.
0:32:45 I would have thought you would have said people who are quadriplegic.
0:32:47 There are a lot of people who have
0:32:48 various kinds of mobility problems,
0:32:50 who can’t do very basic things around
0:32:52 the house where essentially a robot could do it for them.
0:32:54 I would think that would be the first use case.
0:32:56 >> I think that’s a great use case,
0:33:00 and for me, that’s in the realm of what I say is elder care,
0:33:04 which is like the assistive robots that help you with just base tasks.
0:33:07 Like my granddad, one granddad went to a home,
0:33:09 the other granddad had in-home care,
0:33:11 and the one that had in-home care,
0:33:12 it was very simple things.
0:33:16 Remind you to take your medication and bring the medication over,
0:33:17 get you a glass of water,
0:33:21 help you to get up and out of bed to go to the bathroom,
0:33:24 just help you stabilize to go to the bathroom.
0:33:28 That’s not something that we’re largely paying
0:33:32 attention to industrial applications right now,
0:33:34 but that is the dream long-term,
0:33:36 so I’ll be excited to see how it shakes out.
0:33:40 >> Yeah. Rationally, what you were saying is it makes sense.
0:33:42 I understand most people would rather stay at home.
0:33:44 I understand that in-home care,
0:33:48 it’s impossibly expensive for most people.
0:33:52 At the same time, my emotional response to a robot taking care of,
0:33:55 say, my parents, it makes me feel sad.
0:33:57 I recognize that that’s perhaps irrational,
0:33:59 but that is at some level my emotional response,
0:34:02 but the happy thing is I should take care of them,
0:34:05 but that’s hard and it’s probably not going to happen
0:34:07 and it’s for its own set of reasons, right?
0:34:10 It’s more than we bargained for in this conversation.
0:34:12 I don’t know. There is something though.
0:34:16 A humanoid robot starts to get to some weird places in that way, right?
0:34:18 >> Yeah. I’ve thought a lot about this,
0:34:21 and I think it’s a great place to go to.
0:34:22 I’m happy to talk about it.
0:34:26 I think what we want is we want humans taking care of other humans.
0:34:28 That’s what we want, right?
0:34:30 But we don’t have that today,
0:34:33 where look at the way that we age.
0:34:36 For me, I was very close to both of my grand-dads.
0:34:42 They both lived into their 90s and outlived my grandmothers oddly enough.
0:34:46 So I watched them age through their lens,
0:34:48 and that was a big driver of doing this,
0:34:52 and these are people that both of them were war heroes.
0:34:53 They contributed to society.
0:34:55 They did all these amazing things,
0:34:57 and then at the end of their life,
0:35:00 they felt like they were a burden to their family.
0:35:04 They had this feeling like they never had to rely on anyone for anything,
0:35:09 and now they’re completely reliant on people for everything.
0:35:15 What I saw them do as they age was they lost their dignity.
0:35:21 For me, this idea that you could have a machine that carries your secrets,
0:35:25 that is your machine that does things for you,
0:35:30 allows you to keep your dignity such that then you as a human that’s aging,
0:35:31 you’re fresher.
0:35:34 You don’t have to rely on your son or your daughter or
0:35:39 your spouse to get you a glass of water or to do things for you.
0:35:44 You still have your own agency and your own autonomy through a machine,
0:35:48 and that then helps your family to be much fresher because they don’t have
0:35:51 the burden of having to do all these things to support you,
0:35:53 where then they can be fresher.
0:35:58 My hope is that this is not designed to replace what humans do for each other.
0:36:01 This is designed to augment and enhance that.
0:36:05 Remember just like my granddad as he was getting older,
0:36:11 I was working and I was busy and I would try to go over as many days as I could,
0:36:14 but it was always really tough and I didn’t want to be alone.
0:36:17 It was this whole battle that I think everybody goes through,
0:36:19 and my hope for the future,
0:36:22 I actually think it’s a much more optimistic version is that hopefully my
0:36:28 parents have a robot and that robot is basically programmed for their happiness.
0:36:33 It’s designed to remind them when they’re down of their favorite song and play it.
0:36:39 Remind them that of the movies that they like watching or whatever it might be.
0:36:42 I think that’s more optimistic.
0:36:46 I think that’s exciting and that makes me hopeful about the future.
0:36:49 I think that’s the worst part of the human experience is the way that we age.
0:36:52 I think that these robots and AI,
0:36:57 embodied AI and AI in general can hopefully allow us to take better care of each other.
0:37:02 I don’t think it is creepy.
0:37:05 I think it can actually be pretty beautiful if properly done.
0:37:08 That’s what you asked at the beginning,
0:37:12 how is Eptronic different and what are we focused on?
0:37:15 For me, I say human-centered robotics.
0:37:21 What that means is that we want this to be an optimistic future for humanity.
0:37:29 We are tool makers and we want to build tools for humans to enable us to live in better ways.
0:37:32 I think that if we really focus on that,
0:37:36 I think that there can be really amazing ways of doing this.
0:37:41 I think elder care is a great example of how this can be used in that way.
0:37:47 We’ll be back in a minute with the Lightning Round.
0:38:00 Now we are back as promised with the Lightning Round.
0:38:05 What’s the biggest difference between Austin today and Austin 10 years ago?
0:38:09 Oh man, it’s changed quite a bit.
0:38:15 One of the things that made Austin really a great place to live and work
0:38:18 was just how small it was and how accessible everybody was.
0:38:22 We used to have these house parties and somebody would bring a violin,
0:38:26 and someone would bring a sitar and these world instruments,
0:38:31 and you’d have just all sorts of eclectic creative people doing really interesting things.
0:38:40 I think one of the things that I am worried about is that was what made Austin special.
0:38:42 The things that make you special,
0:38:49 people want to commercialize and they want to take this and they want to scale it.
0:38:52 It’s almost special because it’s not commercialized,
0:38:54 it’s just this raw organic thing.
0:38:58 How does that as more tech and more money comes into Austin,
0:39:03 how does what made Austin great,
0:39:06 how does that continue to evolve?
0:39:09 So I think though I welcome it,
0:39:14 I’d rather be in the place where everybody’s coming and everyone wants to build the future.
0:39:18 So I’m not one of those that is resisting the changes.
0:39:22 I think it’s really exciting and I think more people with
0:39:27 new ideas about the future and the world and a free place to do it.
0:39:33 There’s a real ability here in Texas and in Austin to do what you want,
0:39:39 and there’s a real culture around the freedom to do the things that you want to do.
0:39:44 So it’s a unique place where all that’s coming to the creativity,
0:39:49 and the capitalism, and all that’s all coming together.
0:39:51 Is Austin still weird?
0:39:53 Still, there’s pockets of weird.
0:39:56 Yeah, certainly there’s still weird Austin is still there.
0:39:59 It’s all growing up, but yeah, certainly.
0:40:06 What’s your favorite humanoid robot in fiction, in books, in movies?
0:40:08 C-3PO for sure.
0:40:10 Okay, you were ready with that one.
0:40:12 You had that one on deck.
0:40:15 Yeah, I want to make C-3PO, the human helper, right?
0:40:22 What’s one thing that you’ve learned about the human body from building robots?
0:40:28 Oh man, at a high level, what I’ve learned is how amazing the human body really is.
0:40:37 I think there’s this fear from humans that as we continue down this pursuit of replicating
0:40:42 humans and building machines that can do what humans do, that that diminishes what it means
0:40:44 to be humans.
0:40:48 But what it’s actually done for me and most of the people working on this is it just makes
0:40:52 you appreciate even more how amazing humans are.
0:40:55 So the hand is something that you think a lot about.
0:41:01 You just do all these things and you don’t appreciate how incredible your hands are.
0:41:10 And when you start to try to build a hand for a robot, you just appreciate all the limitations.
0:41:14 How we walk, how we move, the fact that we can-
0:41:15 It’s so hard, right?
0:41:20 All the things we do, just like pick up an egg or open a door, like that’s wildly difficult.
0:41:21 It’s amazing.
0:41:24 Or you eat that egg and it powers you for a day.
0:41:30 It powers this neural network, your brain, billions of parameters, right?
0:41:33 I mean, the humans are amazing.
0:41:38 And I think as we continue to learn more about what it means to be human, what does it mean
0:41:43 to be conscious, all of these kind of big ideas, I think we’ll only grow to appreciate
0:41:46 what we actually have here.
0:41:49 Last one, tell me about your grandfather.
0:41:53 Oh, man.
0:42:00 So two grandfathers, one Gilberto Cardenas, the other one George Smith.
0:42:04 Both of them were great.
0:42:09 My granddad, Gilberto Cardenas, came from Puerto Rico when he was 17.
0:42:13 He joined the army and fought in the Korean War.
0:42:14 He spoke five languages.
0:42:22 He was self-educated and he had the American dream and dreamed of what he could do.
0:42:23 He was in the army.
0:42:27 He was actually a field medic, but became a hospital administrator.
0:42:35 And he’s a big sort of driving force in our family and I watched him age and watched all
0:42:36 the things he went through.
0:42:38 He actually fell and lost his vision.
0:42:44 So his brain was still intact in his body largely, but he couldn’t see.
0:42:49 And so he had to have around the clock care when he was in his 90s and in the home.
0:42:55 And he wasn’t wealthy by any stretch, but he had done okay and it saved his money and
0:43:00 it was $17,000 a month for in-home care.
0:43:07 And it was like this revolving door of people that would rather be doing anything else than
0:43:10 sitting in a room with my granddad and taking care of him.
0:43:18 And so for me, I admired my granddad so much and just seeing sort of that as the end of
0:43:25 his life sitting in a room, counting the days down, I just thought there’s got to be a better
0:43:26 way than this.
0:43:30 And that was a big driver for me doing all this.
0:43:34 My other granddad actually ended up getting George Smith.
0:43:40 He had to go to a home, he had colon cancer and his brain still functioned.
0:43:45 He had a great sense of humor and he lost control of his bowels.
0:43:51 And so you can imagine how humiliating that is as you age to be fully aware of what’s
0:43:56 going on, never rely on anybody, but not be able to control your bowels.
0:44:03 And so he had to get multiple showers a day and every time I would go see him, it was
0:44:05 a humiliating experience.
0:44:10 And so these are the things that are just my story, but everybody has their own story
0:44:17 of taking care of an aging parent or grandparent and just what that looks like.
0:44:22 And my hope is that as humans, as tool makers, I think we can do better than that.
0:44:28 And I think that these machines we can create will allow us to take better care of each
0:44:29 other.
0:44:36 My parents are already naming their robots, they can’t wait to get them, they’re almost
0:44:44 70 now and they’ve got to have these ready for whenever our time is there so that we
0:44:53 can age more gracefully than our parents did.
0:44:57 Jeff Cardenas is the co-founder and CEO of Aptronic.
0:45:03 Today’s show was produced by Gabriel Hunter-Chang, it was edited by Lydia Jean-Cott and engineered
0:45:05 by Sarah Brugger.
0:45:09 You can email us at problem@pushkin.fm.
0:45:12 I’m Jacob Goldstein and we’ll be back next week with another episode of What’s Your Problem?
0:45:20 [Music]
0:45:22 you
0:45:32 [BLANK_AUDIO]
0:00:07 Pushkin.
0:00:11 [BLANK_AUDIO]
0:00:14 As a general matter, I’m a fan of technological progress.
0:00:19 But I’ll admit that humanoid robots kind of creep me out for
0:00:23 obvious, normy, uncanny valley type reasons.
0:00:29 And yet, there is an exchange that you’ll hear near the end of today’s show
0:00:33 that’s the most compelling argument I’ve ever heard for humanoid robots.
0:00:36 And it’s not just an intellectual argument.
0:00:39 It’s an emotional argument, if that’s a phrase.
0:00:44 It’s really a very human argument for humanoid robots.
0:00:50 [MUSIC]
0:00:52 I’m Jacob Goldstein, and this is What’s Your Problem,
0:00:56 the show where I talk to people who are trying to make technological progress.
0:00:59 My guest today is Jeff Cardenas.
0:01:02 He’s the co-founder and CEO of Aptronic.
0:01:04 Jeff’s problem is this.
0:01:10 Can you make a safe, reliable humanoid robot for less than $50,000?
0:01:15 We started our conversation talking about the DARPA Robotics Challenge.
0:01:19 DARPA, of course, is the government agency that helped to create the internet and
0:01:23 that gave a big push to early self-driving cars, among other things.
0:01:27 And Jeff says, the agency’s Robotics Challenge, which happened a decade ago,
0:01:32 happened in 2015, played a key role in launching a bunch of the companies
0:01:34 that are now working on humanoid robots.
0:01:41 >> The DARPA Robotics Challenge was a challenge that was created in the wake
0:01:43 of the Fukushima disaster.
0:01:48 Fukushima happened and there was a meltdown in the nuclear reactors.
0:01:51 And it was unsafe for people to go in.
0:01:56 And essentially what happened was they needed a robot to go in to sort of
0:01:59 prevent a broader catastrophe.
0:02:02 And as they went out to the robotics community,
0:02:09 the idea was calling all roboticists, we need a robot to go in and to help out here.
0:02:13 And there was no robots that could do all the different tasks that were required
0:02:15 to actually get to the meltdown reactors.
0:02:19 So you had to go down steps, you had to open doors, you had to do a whole range of
0:02:24 things and in the wake of that, basically what DARPA said was,
0:02:28 certainly there’s got to be the technology to enable us to have much more
0:02:32 versatile robots for natural disaster relief.
0:02:34 So this never happens again.
0:02:38 So out of that, DARPA created something called the DARPA Robotics Challenge.
0:02:42 And so there’s a variety of teams around the world that were put together
0:02:47 to build these general purpose robots.
0:02:51 And the team that we came out of was the NASA Johnson Space Center team
0:02:54 working on a robot called Valkyrie.
0:02:59 >> So I want to go back to this moment when the DARPA Challenge ends,
0:03:05 there’s this big final contest and what we have from it is not some incredible
0:03:09 breakthrough, but a blooper reel of robots, what are they doing?
0:03:12 Falling down stairs or driving cars into walls or something?
0:03:19 >> Pretty much, I mean, the blooper reels make it seem worse than it was.
0:03:24 But yeah, we had, basically the realization was the technology’s not there.
0:03:29 It’s going to take time to continue to mature till it can get to the point where
0:03:31 it’s actually commercially viable.
0:03:34 >> And so it’s interesting, it’s super interesting, right,
0:03:39 that this moment is not the beginning of some humanoid robot winter,
0:03:43 but rather the beginning of this humanoid robot industry.
0:03:46 How does that work, how do people, whatever, start companies and
0:03:49 get money out of this seemingly disappointing moment?
0:03:55 >> Well, I think it actually was a winter, when we started and
0:03:59 the company was incorporated in 2015 and we started in 2016.
0:04:04 And for the most part, nobody wanted to talk about humanoids and
0:04:06 nobody was really paying attention to it.
0:04:10 There was a handful of folks that I sort of think of as the true believers
0:04:15 that were really working on this problem and thought, we don’t care how long
0:04:18 this is going to take, we’re just going to keep working on this no matter what.
0:04:24 And but for the most part, the entire robotics industry was very anti-humanoids
0:04:26 coming out of the DARPA Robotics Challenge.
0:04:30 And in fact, there was many people that were saying they’ll never be viable.
0:04:32 Like, why would you ever use a humanoid robot?
0:04:36 They’re too complicated, they’re way too expensive, and
0:04:38 you’ll always use a simpler robot.
0:04:41 So actually, most of the people that we met when we decided to start
0:04:46 Eptronic were doubters, and we’re saying, humanoids will never make sense.
0:04:50 We’ll use these special purpose robots, maybe in 50 years,
0:04:52 humanoids will make sense, but not for a long time.
0:04:58 >> I mean, special purpose robots is a pretty compelling case, right?
0:05:03 Like, whatever, we all have dishwashing robots and
0:05:05 clothes cleaning robots in our houses.
0:05:09 And wheels seem way easier than legs for lots of things.
0:05:13 And obviously, there have been robot arms for, what, I don’t know,
0:05:15 70 years now or something.
0:05:18 So robots, in a way, are all around us.
0:05:21 Why would you build a machine that looks like a dude when that’s wildly hard, right?
0:05:26 >> Yeah, I mean, I was naive coming out of graduate school.
0:05:28 And to me, it seemed obvious.
0:05:32 And the way I used to think about it was you could either have thousands of robots
0:05:35 that do one thing, or you could build one robot that could do thousands of
0:05:37 different things.
0:05:40 And when I would talk about this with Nick, my co-founder,
0:05:45 Nick would say, look, you can either invest all of this engineering and
0:05:47 each of these sort of narrow solutions.
0:05:52 Or yes, a humanoid robot, a viable humanoid robot, could take you years.
0:05:53 It could take you a decade.
0:05:57 But once you invest all this time in this single platform,
0:06:01 then you can reap the benefits of that across, you can spread.
0:06:06 The research of that across many different applications.
0:06:11 >> I mean, is there not a middle case where there’s some core kind of
0:06:16 functionality that you develop that works across many different types of robots?
0:06:20 Is that a less straw manny version of the non-humanoid robot kind of argument?
0:06:23 >> I think there could be, but I came into robotics and
0:06:27 basically just saw a lot of challenges with the business models.
0:06:30 So you build this special purpose robot.
0:06:33 You custom program the robot in the industrial space.
0:06:39 You can spend six times the price of the robot on just systems integration.
0:06:41 And the robot just does one thing.
0:06:44 So this idea that you could have a much more versatile robot, to me,
0:06:45 seemed obvious.
0:06:47 Like if robotics is going to scale,
0:06:50 we have to have much more versatile robots than we’ve had in the past.
0:06:55 So if you sort of think of that as the premise is we need more versatile robots.
0:06:57 Then the question is, well, how do you get there?
0:06:59 And what does versatility mean?
0:07:04 And that’s where it led me to the humanoid making a lot of sense.
0:07:07 Because if you have to modify the environment for
0:07:12 every new task that the robot can do, you run into the same problem that we had
0:07:14 in sort of classical robotics.
0:07:18 But if the robot can retrofit into the environment such that you don’t have to
0:07:23 change or modify the environment for every new task that the robot can do,
0:07:27 then it seemed to me that this would maybe be the key unlock for
0:07:29 robotics to actually scale to the masses.
0:07:32 The demand would be infinite if you had a thing that was the size and
0:07:35 shape of a person with arms and legs.
0:07:39 Like scale would be off the charts and presumably that’s what drives costs down.
0:07:42 And that’s like the good flywheel, right?
0:07:43 Yeah, exactly.
0:07:47 So okay, so you had this big idea about humanoid robots and you started a company.
0:07:50 But at the moment you started the humanoid robot company,
0:07:54 the prevailing sentiment was like deeply skeptical.
0:07:54 What happened?
0:07:55 What did you do?
0:07:57 Well, a handful of us kept working on it.
0:08:00 So I didn’t know any better.
0:08:02 Sometimes it’s better that you don’t know any better.
0:08:07 I thought humanoids were really cool and I thought that it just seemed,
0:08:12 it made sense to me that how are we gonna get to millions of robots that are
0:08:16 working and with and around humans in all these environments?
0:08:19 And to me, this seemed like the only way that that was gonna happen.
0:08:24 And the way I looked at it was even if we failed, this was a worthy pursuit.
0:08:27 And I would be proud that I tried to do it.
0:08:31 And so the way that we did it was we bootstrapped the company.
0:08:35 There was no investors that were willing to invest in humanoid robots at
0:08:39 the time that we got started, especially for hardware,
0:08:42 which we can talk about that as we move forward.
0:08:45 And so we bootstrapped the company.
0:08:49 And we basically got paid to build robots for a lot of different folks.
0:08:54 And for the first five years of the company, we just built the company on revenue.
0:08:58 We would get project after project and somehow never died.
0:09:01 >> Like what kind of jobs were you taking at that time?
0:09:02 What’s an example?
0:09:04 >> Well, our first contract was with NASA.
0:09:07 So we had a contract with NASA to build Valkyrie II,
0:09:10 to take the lessons learned from the DARPA challenge and
0:09:12 build the next iteration of Valkyrie.
0:09:16 We were really kind of pioneering new ways of building these systems.
0:09:20 So US Special Forces ended up coming to us about a year in and
0:09:23 said, hey, we wanna do Ironman suits.
0:09:27 And our view was this was kind of a humanoid robot that you wear.
0:09:28 We worked in automotive.
0:09:30 We helped build humanoid robots for
0:09:37 a couple major companies that are still working on these things today.
0:09:40 And we would help sort of pioneer new ways of building their platforms.
0:09:44 So we’ve done 15 unique robots since we got started.
0:09:46 And we’re now on a ninth iteration of humanoid.
0:09:49 And I’ve only raised money in the last couple of years.
0:09:57 >> So where did the idea to build a robot for $50,000 come from?
0:10:03 >> The idea of where to build a robot for $50,000 was what will it take for
0:10:06 these robots to be economic and reach mass market?
0:10:13 So when we got started, sort of my view was what will
0:10:16 a truly viable commercial humanoid look like?
0:10:20 And what would the bomb cost need to be for this to make sense?
0:10:24 And if you sort of just do that bottoms up about $50,000 for
0:10:27 the robot, you’re in the money for mass market.
0:10:32 You can still do some tasks in a very economic way at even 100,000 or
0:10:36 150,000, but 50,000 was the goal.
0:10:39 This has now been blown by by some of the new entrepreneurs that are coming out
0:10:42 that are talking about sub $20,000.
0:10:48 But it never made sense to me that robots were as expensive as they were at the time.
0:10:51 If you look at a humanoid compared to a car,
0:10:55 there’s about 4% the raw material by weight.
0:10:59 So in one of our robots, there’s about $300 of raw aluminum,
0:11:02 which is the base metal of the system.
0:11:09 And so it never made sense to me that these robots would need to be any more
0:11:11 than $50,000 as you could reach scale.
0:11:14 And as you could start to think about new ways of building them and
0:11:16 similar ways that we build other machines.
0:11:21 >> So you decide you want to build a $50,000 robot.
0:11:24 Like, what do you actually do to do that?
0:11:31 How do you go from having an idea of building a $50,000 robot to having a $50,000 robot?
0:11:34 >> Well, you iterate until you solve any problem.
0:11:37 So what we would do is basically we would get a project or
0:11:40 a contract to build a robot.
0:11:47 And we would put a lot of different ideas into those designs.
0:11:49 And in early days, it was all about performance.
0:11:53 How can you get the robot to just do these tasks, to stand, and
0:11:55 have a battery life that’s long enough.
0:12:01 And then as we kept evolving, we started to really focus on cost in addition,
0:12:05 and scalability, and assemblability, and robustness.
0:12:09 And the key building block to drive cost and performance is the actuator.
0:12:12 So I mentioned we’ve done nine iterations of humanoids, but
0:12:15 we’ve done 60 iterations of electric actuators.
0:12:19 >> Actuators are basically the thing that makes the robot move, right?
0:12:20 >> Yeah, the muscles of a robot.
0:12:23 >> So where are you now?
0:12:24 Tell me about Apollo.
0:12:27 >> Yeah, we’re now at an exciting point.
0:12:30 We have about 170 employees at Aptronic.
0:12:33 We are piloting these robots right now.
0:12:38 So I think the entire industry is still in the pilot stage overall.
0:12:40 There’s some commercial orders that are happening, but
0:12:42 still early days for humanoids.
0:12:47 We’re working with a handful of really great partners, folks like Mercedes and
0:12:49 GXO.
0:12:54 And we’re getting the robots out into the real world, and it’s pretty big ways.
0:12:57 So we’ll have more announcements over this year.
0:12:59 We have a big partnership with Google DeepMind,
0:13:02 which is something that I always dreamed of coming out of graduate school.
0:13:06 We had a lot of respect for the folks at Google, and
0:13:10 they have a whole history and legacy in the humanoid space as well.
0:13:15 And basically right now, we’re getting these robots out into the world and
0:13:22 gearing up for real commercialization, which we expect to happen in 2026.
0:13:24 >> What’s the robot look like?
0:13:27 >> The robot kind of looks like a superhero, maybe.
0:13:31 This has been kind of the idea that we’ve had from the beginning.
0:13:33 It’s got two eyes and a face.
0:13:39 It’s five foot eight, weighs 160 pounds, has four hour swappable batteries.
0:13:44 Yeah, it’s got a screen on its chest and a face, and that’s about it.
0:13:51 >> Two arms, two legs, what’s it have in the way of hands?
0:13:53 >> So it has hands, it has five-fingered hands.
0:13:57 There’s these debates that I think of as false debates in the humanoid space.
0:14:01 So a lot of people, when they sort of knock humanoids and
0:14:03 the viability of humanoids, it usually has to do with,
0:14:06 do they need legs and do they need hands?
0:14:08 And the answer to that question for me is no, they don’t.
0:14:10 It’s a robot, and robots are modular.
0:14:14 So we can put Apollo on any mobility base, we can put it on wheels,
0:14:20 we can put it on tracks, we could stationary mount the upper torso of Apollo.
0:14:23 And the same thing’s true for the hands or the grippers.
0:14:27 We can use parallel grippers, or we can use five-fingered hands.
0:14:29 >> Hands are like a whole thing, right?
0:14:33 Like hands are, is it partly because they’re so hard?
0:14:35 Like what’s going on with robots and hands?
0:14:37 >> It turns out hands are a whole thing.
0:14:41 This is another one of those things that, it’s almost better that you don’t
0:14:45 understand the complexity before you get into it or else you might not
0:14:47 have done it in the first place.
0:14:53 98% of all tasks that humans do are done with our hands.
0:14:59 So there are narrow things that humanoids can do without more dexterity.
0:15:03 But it’s very limited relative to the whole sort of,
0:15:05 all the different types of tasks that humans do.
0:15:08 Most of the things we do involve our hands.
0:15:12 And certainly in the industrial space, most of the work is done with hands.
0:15:17 So solving the end effector or the hand problem is a big deal.
0:15:19 There’s a lot of different debates about what you need and
0:15:22 how you get something that can actually perform industrial work.
0:15:27 We’ve chosen the five-finger hand route and
0:15:32 we’re working across the space to really make some big advancements there overall.
0:15:37 >> It’s part of the trade off like I could build whatever two,
0:15:38 what do you call them, prongs?
0:15:41 Like if you had two fingers basically, like a claw?
0:15:43 Like you could do a lot of things with a claw.
0:15:47 Presumably it would be way easier, but you couldn’t do everything.
0:15:49 Is it kind of like what are you optimizing for and
0:15:52 sort of how much payoff now versus how much payoff later?
0:15:54 >> Yeah, I think that’s exactly right.
0:16:00 It’s versatility compared to robustness and
0:16:04 cost basically, how much complexity do you want to have on the system?
0:16:09 And for these robots to be really viable in the long run, especially in the industrial space,
0:16:13 they got to be able to operate two shifts a day minimum,
0:16:16 really 22 hours a day, seven days a week.
0:16:18 But solving that problem in a hand, so
0:16:23 just getting the performance of the hand first, but then the robustness for
0:16:26 them to do that type of work is the next piece and
0:16:28 that’s a trade off of performance and complexity and cost.
0:16:31 >> Cuz like it gets delicate, right?
0:16:37 Presumably the fingers, so to speak, would be fragile, right?
0:16:37 >> Yeah, they can break.
0:16:39 >> Easy to break, yeah.
0:16:41 >> Yeah, yeah, and you got to maintain it and
0:16:44 you got to support those systems and fix them out in the field.
0:16:46 And so what’s the trade off there?
0:16:50 And that’s a whole trade space that we’ve been working on over a long time.
0:16:53 So we’ve been talking about hardware.
0:16:57 Let’s talk about the software side, what’s happening with that?
0:17:00 >> A lot’s happening on that side.
0:17:06 I think we’re really an exciting point for robotics overall.
0:17:09 Think of the AI as really the last piece of the puzzle.
0:17:16 So we’ve had the ability to build complex robots for a relatively long time.
0:17:20 We’re just now really figuring out how to take the lessons from automotive and
0:17:25 consumer electronics and build much more economic systems.
0:17:28 And we’ve had some advancements in things like motors and batteries and
0:17:32 compute and sensors that have all sort of built up to this moment.
0:17:37 But the final piece of the puzzle was the AI and the intelligence.
0:17:39 And essentially the way to think about it, and
0:17:43 I think Jensen does a great job of explaining this, but the advance.
0:17:45 >> Jensen Wong from NVIDIA.
0:17:46 >> Yeah, that’s right.
0:17:51 I feel like Jensen is not quite the Elon level of one name, household name.
0:17:53 But sorry, go on, he’s getting there.
0:17:55 >> Yeah, he should be.
0:17:58 >> He should be, he should be.
0:18:02 >> Basically the advancements in generative AI turn out to apply very well
0:18:06 to robotics and particularly to humanoid robotics.
0:18:08 So you can basically map human movement and
0:18:13 trajectories from humans doing things and build big data sets and
0:18:17 use that to train robots to do similar tasks in similar environments.
0:18:22 And these transformer architectures that we’re using in generative AI
0:18:24 actually apply very well to robotics.
0:18:28 And so this has been a big sort of breakthrough moment for robotics.
0:18:31 And so I think as an industry as a whole,
0:18:35 everybody’s really excited right now because we’re reaching new heights and
0:18:40 we’re able to do things that we dreamed about doing with robots only even a few
0:18:44 years ago are now possible and we’re seeing a really rapid advancement in
0:18:45 performance overall.
0:18:49 >> What’s an example of a thing that you could only dream of a few years ago
0:18:51 that robots can do now?
0:18:53 >> I think it’s more dexterity and versatility.
0:18:55 So just the range of things that you can do.
0:19:00 So the challenge for robotics was that each new task,
0:19:04 even if you build something like a humanoid robot, and this is true for us.
0:19:09 And say you build an application to pick boxes off of a palette and
0:19:12 place those boxes onto a conveyor.
0:19:15 Well, you hand build that application and
0:19:18 maybe takes you 18 months to sort of wring that out and
0:19:21 get it to a certain amount of robustness.
0:19:25 Well, now you want to do the reverse of that and pick off of a conveyor and
0:19:27 palletize something.
0:19:31 That will take you the same amount of time that it took you to build the initial
0:19:32 application.
0:19:35 >> You have to basically have to write a whole other piece of software.
0:19:38 You have to start from scratch almost, yeah.
0:19:39 >> Yeah, exactly.
0:19:45 And so basically what is happening now is that we now have these much more
0:19:50 sort of general models that where you can collect a lot of data at the top layer.
0:19:54 And so each new task that you want to perform actually takes less and
0:19:57 less incremental amount of work.
0:20:00 So what is opening up now is more dexterous applications.
0:20:04 [MUSIC]
0:20:05 >> Still to come on the show,
0:20:08 how Jeff’s grandparents inspired his work on robots.
0:20:17 [MUSIC]
0:20:22 So you were talking about using the transformer model that has been the
0:20:27 breakthrough that has driven large language models in training robots,
0:20:29 essentially.
0:20:33 I mean, of course, a key sort of serendipitous thing that happened with
0:20:39 language models was there is this crazy large data set of words and
0:20:40 pictures, which is the internet.
0:20:45 And there’s not an analogous data set for the physical world, right?
0:20:45 >> Yeah.
0:20:49 >> It seems like that is, is that the rate limiting step?
0:20:54 Is that the big problem in sort of AI for robots?
0:20:57 >> Yeah, I mean, there’s a lot of work that’s still happening at the research
0:21:02 level for how can you pull that kind of data from videos.
0:21:06 So you can think of big data sets of humans doing things that can be really
0:21:09 interesting to train robots in the future.
0:21:10 >> Interesting.
0:21:12 >> And that will come into play over time.
0:21:17 But yeah, it’s the chicken or the egg problem and data is one of the key
0:21:21 things that we need to enable the next wave of breakthroughs.
0:21:26 And this is kind of the race is can you get robots out into the real world,
0:21:32 into the field and collecting data at very high, in very high volumes.
0:21:35 Whoever does that will have better models.
0:21:37 This is the data flywheel.
0:21:40 So this is kind of the race that’s on right now where you hear a lot of
0:21:45 other humanoid CEOs talking about getting a lot of robots out into the world.
0:21:48 Largely, those are going to be under tele-operation, collecting data and
0:21:52 then training and building these models of the future.
0:21:57 >> So it’s like whoever gets there first will win just because that’ll be the
0:21:58 accelerant.
0:22:02 Like once you have robots out in the world and you’re collecting data,
0:22:05 then you’re immediately getting ahead of whoever has fewer robots out in the world
0:22:07 because they’re collecting less data.
0:22:07 >> Yeah.
0:22:11 >> So tell me about tele-operation.
0:22:16 >> Tele-operation is basically just remotely controlling the robot.
0:22:18 So you’re taking over the robot.
0:22:22 You can see through the robot’s eyes with the VR headset and
0:22:27 then you’re controlling the robot’s arms and hands to do a particular task.
0:22:29 It’s like a video game and you’re controlling a robot.
0:22:33 It’s a simple idea.
0:22:35 There’s a couple of reasons it’s important.
0:22:38 The first thing is that it tells you what the robot’s physically capable of doing.
0:22:44 So if I’m completely controlling the robot and I can’t do a task under tele-operation,
0:22:46 then that means the robot’s not physically capable of doing it.
0:22:51 So it’d be very difficult for an AI control system to control the robot to do that.
0:22:56 So this is how we understand the physical capabilities of the robot.
0:23:01 As these new models have come along, the simple idea is that if you can tele-operate
0:23:06 the robot to do a task, then you should be able to automate that task on the other end.
0:23:10 So if you can collect enough data under tele-operation,
0:23:16 then you can automate it by running it through these similar architectures that we talked about.
0:23:22 >> So it’s the basic idea that you use remote control to drive the robot to do a thing,
0:23:26 whatever, a thousand times, some number of times.
0:23:28 In doing that, you’re training the robot.
0:23:31 You’re training the software, training the AI.
0:23:32 >> Yeah, that’s exactly right.
0:23:35 >> What’s an example of a thing that you’ve done that way,
0:23:39 and how many times did you have to remote control it before the robot could do it?
0:23:41 >> So each is picking is a good example,
0:23:44 or you’re taking objects and you’re putting them into a box.
0:23:47 To do that in a simple context,
0:23:52 thousands of demonstrations is what you need.
0:23:55 We think of this as generally hours.
0:24:00 So how many hours of data collection have we done?
0:24:04 And thousands of iterations can get you to,
0:24:06 let’s say 80 percent of human rate.
0:24:10 If you want to get to 95 percent or better of human rate,
0:24:12 then you need more and more data.
0:24:14 But it’s in the thousands, it’s not millions.
0:24:18 >> Yeah, thousands makes it seem totally tractable.
0:24:22 >> Yeah, I was actually surprised by how well these models work,
0:24:26 and actually how little data they need to get relatively good performance.
0:24:30 You’re seeing a lot of demonstrations of this out there today.
0:24:33 >> Presumably, that’ll get better and better.
0:24:37 As the software side of AI gets better and better,
0:24:39 it’ll learn faster essentially.
0:24:40 The other obvious thing,
0:24:43 but I’m just going to say it is once you have done it once,
0:24:45 then it works for every robot.
0:24:48 Then you can make a million robots and they all know how to do the thing.
0:24:49 >> Yeah, that’s exactly right.
0:24:52 One of the interesting things about these models is actually
0:24:56 the diversity of data is almost more important than task-specific data.
0:24:59 So we want to go wide across a range of tasks.
0:25:00 >> Interesting.
0:25:03 >> Then you’re basically building all these skills into the robot,
0:25:07 and then it becomes better at doing any one particular task.
0:25:09 >> It really is like learning.
0:25:12 It really is human-esque.
0:25:13 >> Yeah, that’s right.
0:25:17 >> So I know you’re in a few pilot projects with Mercedes and with,
0:25:20 what is it, GXO, Big Logistics Company.
0:25:23 When do you want to start selling robots for real?
0:25:24 When do you think that might happen?
0:25:26 >> 2026.
0:25:27 >> Okay.
0:25:27 >> Yeah.
0:25:30 >> Suddenly that’s next year, almost now.
0:25:34 >> Years of long time and these are dog years.
0:25:38 It’s a long time in this space.
0:25:43 >> 2026 could be almost two years for now.
0:25:47 Who are you going to sell robots to and how much are you going to charge?
0:25:48 What are they going to do?
0:25:52 >> So initially in manufacturing and logistics,
0:25:54 so folks like Mercedes and GXO,
0:25:59 these are the initial customers of these systems.
0:26:02 We are not announcing pricing yet,
0:26:08 but you can think of it as take what it costs to do these tasks today,
0:26:12 and with some discount to what it costs to do these tasks today.
0:26:15 We have a RAS model that we use.
0:26:18 >> Robot as a service?
0:26:20 >> Yeah, robot as a service model,
0:26:25 where you’re paying the robot basically by the hour
0:26:27 effectively to do a particular task,
0:26:31 and that’s at a discount to what it costs to do that task today.
0:26:34 >> How far are you from the $50,000 robot?
0:26:39 >> We’re not there yet, so not very far.
0:26:41 So we have the architecture to be able to do this.
0:26:45 So getting the cost down on these robots is a two-step process.
0:26:49 So first step is new architectures.
0:26:52 So if you still require this very high precision in the system,
0:26:56 and you’re using bespoke components that are only used for robotics,
0:26:59 these robots will still be expensive.
0:27:01 The challenge of humanoid robots is they have
0:27:03 a lot more motors than traditional robots.
0:27:06 So a traditional robot has six or seven motors,
0:27:10 a humanoid robot has 30 to 40 plus.
0:27:14 >> Okay, so that means it’s expensive or you got to figure out how to
0:27:16 get cheaper actuators.
0:27:17 >> Yeah, so we’re there. So for us,
0:27:23 that was a $500 actuator and we have a $500 actuator now today.
0:27:25 So once you solve that problem,
0:27:27 and once you solve the architecture problem,
0:27:29 now it’s about scale and manufacturing.
0:27:31 So a lot of where we spend,
0:27:35 a lot of where the cost is driven at low volumes,
0:27:37 is in just the structures of the robot,
0:27:38 where we’re seeing and seeing,
0:27:42 we’re milling at a big blocks of metal,
0:27:45 parts and very small quantities.
0:27:50 But there’s other techniques that are much more cost-effective,
0:27:53 like casting or stamping,
0:27:56 and these will allow these robots to be much cheaper.
0:27:59 As I mentioned, look at automotive and look at the scale of automotive.
0:28:01 There’s 4 percent the raw material by
0:28:05 weight in a humanoid robot as compared to a car.
0:28:09 So once you solve the architecture problem such that you can
0:28:13 build a lot of these systems and they’re simpler to make,
0:28:15 then the next piece is just applying
0:28:18 mass manufacturing approaches to
0:28:20 this to make them a lot cheaper as you scale.
0:28:26 >> Well, I mean, that’s a hard leap to make, right?
0:28:27 Like, what do you do?
0:28:30 You get a ton of capital and just build
0:28:33 a factory and hope there’s demand on the other end.
0:28:36 Like, how do you go from this bespoke expensive thing to
0:28:39 a mass produced much less expensive thing?
0:28:40 >> Well, it’s a gradient.
0:28:42 So like I said, step one is you have
0:28:45 new approaches that allow you to make them
0:28:48 cheaper just inherently on a unit-to-unit basis.
0:28:51 So the early humanoids were like millions of dollars,
0:28:53 and now we’re in the hundreds of
0:28:55 thousands of dollars range for building one.
0:28:58 >> So you just got to get one more order of magnitude out of it.
0:29:02 >> Yes. We’ve already dropped the price by an order of magnitude.
0:29:04 Then now as we build more,
0:29:06 even as you add a zero,
0:29:08 as you go from 10 to 100,
0:29:10 the price drops pretty dramatically.
0:29:14 So you don’t need the volume that you might think.
0:29:15 My view is we can get to
0:29:20 the sub $50,000 price point in the thousands of unit quantity.
0:29:23 So without hundreds of thousands or millions of these.
0:29:25 >> So one big buyer,
0:29:28 one big car company or logistics company
0:29:31 might place an order of thousands of units, right?
0:29:34 >> Yeah. You made a comment that you said,
0:29:35 and you hope that there’s demand.
0:29:37 One of the things I think it’s important to
0:29:41 note is the demand for these robots is enormous.
0:29:45 We have demand for hundreds of thousands of units already
0:29:48 today with the customers that we’re working with.
0:29:51 So the demand is enormous.
0:29:53 So we’re ramping up.
0:29:55 We’ve got to get the robustness and
0:29:58 the safety of the system and really bring out the design,
0:30:02 and these are really credible,
0:30:04 thoughtful people that are coming from
0:30:06 other industries that are now joining us,
0:30:08 that now see that we’ve crossed
0:30:12 this threshold of technical viability,
0:30:16 and now taking lessons from how you scale and
0:30:17 manufacture other things and bringing that
0:30:21 into the robotic space and the humanoid space overall.
0:30:24 >> So in a year or at least next year,
0:30:27 you want to be selling robots for real.
0:30:31 Where do you want to be in five, say, years?
0:30:34 >> My view is that where this evolves is it’s
0:30:36 going to start in logistics and manufacturing,
0:30:38 and then as we solve safety as an industry,
0:30:40 I’m really interested in health care,
0:30:43 and particularly in elder care over time.
0:30:49 So how can these robots be used to improve the way we live and work?
0:30:53 That was really the lens that I came into this on.
0:30:55 So I think over the next five years,
0:30:59 you’ll start to see the early stages of the next three years.
0:31:04 You’ll start to see early applications for robots entering the home.
0:31:06 There’s some folks that are really working hard on this.
0:31:11 I think we’re about three years out from that being really viable,
0:31:13 but I hope people prove me wrong.
0:31:15 I hope it’s faster than that.
0:31:16 >> Three years seems fast.
0:31:21 What’s the first use case,
0:31:23 first job you imagine a robot doing for
0:31:26 real in somebody’s house in three years?
0:31:28 >> Well, everybody wants laundry.
0:31:30 If everybody I talked to says,
0:31:32 “When is this thing going to do my laundry?”
0:31:33 I want that as well.
0:31:36 >> There’s literally already a machine to do your laundry.
0:31:39 All you have to do is put it in one machine and then put it in another.
0:31:41 The remaining work is trivial.
0:31:42 >> Yeah. I mean, look,
0:31:45 I’m not the person to talk about the home.
0:31:47 I think we’re still a ways out,
0:31:52 but there’s humanoid companies like 1X that are really focused on the home,
0:31:55 and we’ve got a lot of respect for what they’re doing over there,
0:31:57 and so I hope they do it.
0:31:59 I know that they’re working hard on it,
0:32:05 and I want a robot for the home as well.
0:32:09 A lot of the things that are happening with these models that I talked about,
0:32:11 these more generic models,
0:32:12 the things that we’re learning in
0:32:16 the industrial base can apply to the home over time as well.
0:32:18 >> In terms of the AI models, sure.
0:32:22 I mean, the AI models are basically teaching a robot how to deal with the physical world.
0:32:22 >> That’s right.
0:32:24 >> How to move around, how to pick things up,
0:32:26 how to put things down.
0:32:31 >> Yeah. The task in the home is tough because even a robot does your dishes,
0:32:34 your laundry, cleans, and cooks for you.
0:32:38 How much are you willing to pay for that on a yearly basis?
0:32:41 >> I’m imagining the first household tasks.
0:32:45 I would have thought you would have said people who are quadriplegic.
0:32:47 There are a lot of people who have
0:32:48 various kinds of mobility problems,
0:32:50 who can’t do very basic things around
0:32:52 the house where essentially a robot could do it for them.
0:32:54 I would think that would be the first use case.
0:32:56 >> I think that’s a great use case,
0:33:00 and for me, that’s in the realm of what I say is elder care,
0:33:04 which is like the assistive robots that help you with just base tasks.
0:33:07 Like my granddad, one granddad went to a home,
0:33:09 the other granddad had in-home care,
0:33:11 and the one that had in-home care,
0:33:12 it was very simple things.
0:33:16 Remind you to take your medication and bring the medication over,
0:33:17 get you a glass of water,
0:33:21 help you to get up and out of bed to go to the bathroom,
0:33:24 just help you stabilize to go to the bathroom.
0:33:28 That’s not something that we’re largely paying
0:33:32 attention to industrial applications right now,
0:33:34 but that is the dream long-term,
0:33:36 so I’ll be excited to see how it shakes out.
0:33:40 >> Yeah. Rationally, what you were saying is it makes sense.
0:33:42 I understand most people would rather stay at home.
0:33:44 I understand that in-home care,
0:33:48 it’s impossibly expensive for most people.
0:33:52 At the same time, my emotional response to a robot taking care of,
0:33:55 say, my parents, it makes me feel sad.
0:33:57 I recognize that that’s perhaps irrational,
0:33:59 but that is at some level my emotional response,
0:34:02 but the happy thing is I should take care of them,
0:34:05 but that’s hard and it’s probably not going to happen
0:34:07 and it’s for its own set of reasons, right?
0:34:10 It’s more than we bargained for in this conversation.
0:34:12 I don’t know. There is something though.
0:34:16 A humanoid robot starts to get to some weird places in that way, right?
0:34:18 >> Yeah. I’ve thought a lot about this,
0:34:21 and I think it’s a great place to go to.
0:34:22 I’m happy to talk about it.
0:34:26 I think what we want is we want humans taking care of other humans.
0:34:28 That’s what we want, right?
0:34:30 But we don’t have that today,
0:34:33 where look at the way that we age.
0:34:36 For me, I was very close to both of my grand-dads.
0:34:42 They both lived into their 90s and outlived my grandmothers oddly enough.
0:34:46 So I watched them age through their lens,
0:34:48 and that was a big driver of doing this,
0:34:52 and these are people that both of them were war heroes.
0:34:53 They contributed to society.
0:34:55 They did all these amazing things,
0:34:57 and then at the end of their life,
0:35:00 they felt like they were a burden to their family.
0:35:04 They had this feeling like they never had to rely on anyone for anything,
0:35:09 and now they’re completely reliant on people for everything.
0:35:15 What I saw them do as they age was they lost their dignity.
0:35:21 For me, this idea that you could have a machine that carries your secrets,
0:35:25 that is your machine that does things for you,
0:35:30 allows you to keep your dignity such that then you as a human that’s aging,
0:35:31 you’re fresher.
0:35:34 You don’t have to rely on your son or your daughter or
0:35:39 your spouse to get you a glass of water or to do things for you.
0:35:44 You still have your own agency and your own autonomy through a machine,
0:35:48 and that then helps your family to be much fresher because they don’t have
0:35:51 the burden of having to do all these things to support you,
0:35:53 where then they can be fresher.
0:35:58 My hope is that this is not designed to replace what humans do for each other.
0:36:01 This is designed to augment and enhance that.
0:36:05 Remember just like my granddad as he was getting older,
0:36:11 I was working and I was busy and I would try to go over as many days as I could,
0:36:14 but it was always really tough and I didn’t want to be alone.
0:36:17 It was this whole battle that I think everybody goes through,
0:36:19 and my hope for the future,
0:36:22 I actually think it’s a much more optimistic version is that hopefully my
0:36:28 parents have a robot and that robot is basically programmed for their happiness.
0:36:33 It’s designed to remind them when they’re down of their favorite song and play it.
0:36:39 Remind them that of the movies that they like watching or whatever it might be.
0:36:42 I think that’s more optimistic.
0:36:46 I think that’s exciting and that makes me hopeful about the future.
0:36:49 I think that’s the worst part of the human experience is the way that we age.
0:36:52 I think that these robots and AI,
0:36:57 embodied AI and AI in general can hopefully allow us to take better care of each other.
0:37:02 I don’t think it is creepy.
0:37:05 I think it can actually be pretty beautiful if properly done.
0:37:08 That’s what you asked at the beginning,
0:37:12 how is Eptronic different and what are we focused on?
0:37:15 For me, I say human-centered robotics.
0:37:21 What that means is that we want this to be an optimistic future for humanity.
0:37:29 We are tool makers and we want to build tools for humans to enable us to live in better ways.
0:37:32 I think that if we really focus on that,
0:37:36 I think that there can be really amazing ways of doing this.
0:37:41 I think elder care is a great example of how this can be used in that way.
0:37:47 We’ll be back in a minute with the Lightning Round.
0:38:00 Now we are back as promised with the Lightning Round.
0:38:05 What’s the biggest difference between Austin today and Austin 10 years ago?
0:38:09 Oh man, it’s changed quite a bit.
0:38:15 One of the things that made Austin really a great place to live and work
0:38:18 was just how small it was and how accessible everybody was.
0:38:22 We used to have these house parties and somebody would bring a violin,
0:38:26 and someone would bring a sitar and these world instruments,
0:38:31 and you’d have just all sorts of eclectic creative people doing really interesting things.
0:38:40 I think one of the things that I am worried about is that was what made Austin special.
0:38:42 The things that make you special,
0:38:49 people want to commercialize and they want to take this and they want to scale it.
0:38:52 It’s almost special because it’s not commercialized,
0:38:54 it’s just this raw organic thing.
0:38:58 How does that as more tech and more money comes into Austin,
0:39:03 how does what made Austin great,
0:39:06 how does that continue to evolve?
0:39:09 So I think though I welcome it,
0:39:14 I’d rather be in the place where everybody’s coming and everyone wants to build the future.
0:39:18 So I’m not one of those that is resisting the changes.
0:39:22 I think it’s really exciting and I think more people with
0:39:27 new ideas about the future and the world and a free place to do it.
0:39:33 There’s a real ability here in Texas and in Austin to do what you want,
0:39:39 and there’s a real culture around the freedom to do the things that you want to do.
0:39:44 So it’s a unique place where all that’s coming to the creativity,
0:39:49 and the capitalism, and all that’s all coming together.
0:39:51 Is Austin still weird?
0:39:53 Still, there’s pockets of weird.
0:39:56 Yeah, certainly there’s still weird Austin is still there.
0:39:59 It’s all growing up, but yeah, certainly.
0:40:06 What’s your favorite humanoid robot in fiction, in books, in movies?
0:40:08 C-3PO for sure.
0:40:10 Okay, you were ready with that one.
0:40:12 You had that one on deck.
0:40:15 Yeah, I want to make C-3PO, the human helper, right?
0:40:22 What’s one thing that you’ve learned about the human body from building robots?
0:40:28 Oh man, at a high level, what I’ve learned is how amazing the human body really is.
0:40:37 I think there’s this fear from humans that as we continue down this pursuit of replicating
0:40:42 humans and building machines that can do what humans do, that that diminishes what it means
0:40:44 to be humans.
0:40:48 But what it’s actually done for me and most of the people working on this is it just makes
0:40:52 you appreciate even more how amazing humans are.
0:40:55 So the hand is something that you think a lot about.
0:41:01 You just do all these things and you don’t appreciate how incredible your hands are.
0:41:10 And when you start to try to build a hand for a robot, you just appreciate all the limitations.
0:41:14 How we walk, how we move, the fact that we can-
0:41:15 It’s so hard, right?
0:41:20 All the things we do, just like pick up an egg or open a door, like that’s wildly difficult.
0:41:21 It’s amazing.
0:41:24 Or you eat that egg and it powers you for a day.
0:41:30 It powers this neural network, your brain, billions of parameters, right?
0:41:33 I mean, the humans are amazing.
0:41:38 And I think as we continue to learn more about what it means to be human, what does it mean
0:41:43 to be conscious, all of these kind of big ideas, I think we’ll only grow to appreciate
0:41:46 what we actually have here.
0:41:49 Last one, tell me about your grandfather.
0:41:53 Oh, man.
0:42:00 So two grandfathers, one Gilberto Cardenas, the other one George Smith.
0:42:04 Both of them were great.
0:42:09 My granddad, Gilberto Cardenas, came from Puerto Rico when he was 17.
0:42:13 He joined the army and fought in the Korean War.
0:42:14 He spoke five languages.
0:42:22 He was self-educated and he had the American dream and dreamed of what he could do.
0:42:23 He was in the army.
0:42:27 He was actually a field medic, but became a hospital administrator.
0:42:35 And he’s a big sort of driving force in our family and I watched him age and watched all
0:42:36 the things he went through.
0:42:38 He actually fell and lost his vision.
0:42:44 So his brain was still intact in his body largely, but he couldn’t see.
0:42:49 And so he had to have around the clock care when he was in his 90s and in the home.
0:42:55 And he wasn’t wealthy by any stretch, but he had done okay and it saved his money and
0:43:00 it was $17,000 a month for in-home care.
0:43:07 And it was like this revolving door of people that would rather be doing anything else than
0:43:10 sitting in a room with my granddad and taking care of him.
0:43:18 And so for me, I admired my granddad so much and just seeing sort of that as the end of
0:43:25 his life sitting in a room, counting the days down, I just thought there’s got to be a better
0:43:26 way than this.
0:43:30 And that was a big driver for me doing all this.
0:43:34 My other granddad actually ended up getting George Smith.
0:43:40 He had to go to a home, he had colon cancer and his brain still functioned.
0:43:45 He had a great sense of humor and he lost control of his bowels.
0:43:51 And so you can imagine how humiliating that is as you age to be fully aware of what’s
0:43:56 going on, never rely on anybody, but not be able to control your bowels.
0:44:03 And so he had to get multiple showers a day and every time I would go see him, it was
0:44:05 a humiliating experience.
0:44:10 And so these are the things that are just my story, but everybody has their own story
0:44:17 of taking care of an aging parent or grandparent and just what that looks like.
0:44:22 And my hope is that as humans, as tool makers, I think we can do better than that.
0:44:28 And I think that these machines we can create will allow us to take better care of each
0:44:29 other.
0:44:36 My parents are already naming their robots, they can’t wait to get them, they’re almost
0:44:44 70 now and they’ve got to have these ready for whenever our time is there so that we
0:44:53 can age more gracefully than our parents did.
0:44:57 Jeff Cardenas is the co-founder and CEO of Aptronic.
0:45:03 Today’s show was produced by Gabriel Hunter-Chang, it was edited by Lydia Jean-Cott and engineered
0:45:05 by Sarah Brugger.
0:45:09 You can email us at problem@pushkin.fm.
0:45:12 I’m Jacob Goldstein and we’ll be back next week with another episode of What’s Your Problem?
0:45:20 [Music]
0:45:22 you
0:45:32 [BLANK_AUDIO]
Jeff Cardenas is the co-founder and CEO of Apptronik. Jeff’s problem is this: Can you make a safe, reliable humanoid robot – for less than $50,000?
In the short term, Apptronik’s robots will work in factories. But Jeff’s long-term goal – based on the experience of his own grandparents – is to build robots that can help care for the elderly.
See omnystudio.com/listener for privacy information.