Marc Raibert: The hockey stick was a way of showing how good the robot was performing. We were just showing off the balancing capability of the robot, and we weren't being mean in any sense. You know, the robot doesn't have any feelings or ability to feel pain or anything, and people just, especially with humanoid robots, people just assume that it's like a person.
And in almost every respect, it's not like a person.
Brian Heater: Hello, and welcome to the inaugural episode of Automated, I'm Brian Heater, the managing editor at A three. And we're gonna be hanging out for about 45 minutes or so every week with some of the top names in robotics, ai, and automation. This is something my producer Jan and I have wanted to put together for a, a very long time.
Now, it's a show that dives into the cultural, historical, and sociological implications [00:01:00] of these technologies that are taking an increasingly central role in our lives. When we were in the planning stages, I put together a short list of the folks that we absolutely had to have on the program, and I'm excited to tell you that we've already recorded episodes with most of them.
Mark Rayer was at the top of that list. Of course, uh, he and I have spoken. A dozen or so times over the years and I learned something new from him every time. You'll also find that he isn't afraid to give me and my journalistic instincts a bit of crap during our conversation. It's always great and often a bit humbling catching up with Mark.
So it was the Biodynamics AI Institute and now it's RAI. How did you settle on that name?
Marc Raibert: We call it the Ray Institute. Robotics and AI Institute, how do we decide on it? Oh, you know, we did the usual thing where we had, uh, made [00:02:00] lists and circulated among all the people and came up with a bunch of different ideas and sorted it out to this one.
I think it was important that the name had robotics in it, in addition to ai. 'cause it's really. How the two of them fit together that we're all about.
Brian Heater: Yeah, I remember, uh, like in the very early days, you know, after Rob was announced, the CEO and you would announce you were doing this. I, I think he might have been bouncing a few things off of me 'cause you were kind of going back and forth and it did initially have Boston Dynamics in the name. Was that, was that a point of, I don't know, confusion? I think
Marc Raibert: that caused confusion 'cause everybody thought we were, didn't know if we were just Boston Dynamics and. You know, people know generally that I'm was at Boston Dynamics for a long time, and the founder, same owners, but a separate company.
So I thought it was better if we had our own identity. Was, was retirement ever a serious consideration for you? Uh, you know, during CODI thought about it because during COVID, uh, I'd already, uh, [00:03:00] turned over the helm to Rob and, uh. You know, I kind of rekindled some hobbies because I was at home and had time to do that.
Uh, so I thought about it, but then and now, uh, it still feels like those hobbies would be fun for a while, but not anything like working on robotics switches, you know, my lifelong love.
Brian Heater: Yeah, you seem to be in a really good position, and I think you're probably in a similar position by, uh, the end of your time at Boston, dynamics of the company where you get to fly around, you get to talk to people, you get to sort of be the face and the voice of the institute.
Marc Raibert: Yeah. But the real fun is having your hands on the robots or working on the designs and stuff. I mean, you're right. Last year I spent a lot of time flying around and being the face. Uh, this year I'm trying to do less of that and. Focus more on, uh, on the work we're doing. I think that's where the real fun is.
I mean, I'm, you know, I'm a [00:04:00] gadget guy. A builder, uh, so that's what, that's what robotics really is for me. Like sort of the, the ultimate in, uh, in, uh, building things.
Brian Heater: I was gonna ask, what, what are Mark Bert's hobbies outside of I guess robotics?
Marc Raibert: Uh, lately I've, uh, uh. Been working on reviving uh, guitar amplifiers from the fifties and sixties and even, uh, I just got one from the forties and, uh, I kind of, you know, got the electronics working again and, uh, building what looks like a period, correct cabinet for it.
'cause it came in something else.
Brian Heater: The, the, the Guitar amp thing is really interesting to me. 'cause I think probably those are, are two bands. But working with that, you know, half Century or like 60, 70-year-old technology, is there anything that you can still learn that you feel like is, continues to be [00:05:00] applicable for these bleeding edge robots are building?
Marc Raibert: It's funny, we had a, uh, a breakfast to welcome new people at the company. Recently. And uh, one of the things we do is talk about not work, but our outside activities. And I said the same thing to them that I said to you that I was working on these old tube amps, and several of people said, what's a tube?
It surprised me a little bit. I don't think there's anything about tube amps. It's important for robotics particularly, but except for the culture of both a. Computing and AI and uh, mathematics culture combined with a build hardware and hook the two up culture. And I think it's really important to do those two and have them work, interact with each other.
Brian Heater: Well, one of the things that, that really interested me the first time I came and visited you over there in the, uh, Akamai building in, in Boston was, and, and maybe this, maybe this goes back to the decision to remove Boston Dynamics from the [00:06:00] beginning of the name, is that the company is relatively or maybe entirely, uh, hardware agnostic, right?
I mean, you were buying a lot of spot robots, but you've also been buying different humanoids to test out.
Marc Raibert: We have about 50 robots here. Some of them are arms mounted to the table, like, uh, like an industrial arm, which we bought from, um, you know, universal robot. We have a couple of unit tree humanoids here.
Uh, both the big one and the little one. We're also have a lot of spot robots from Boston Dynamics. I think we have maybe 20 of them here. We have, uh, some antibiotics, robots, you know, we, we don't have a direct relationship with them other than being customer, but, uh, Marco Hutter was one of their founders, is our, uh, the director of our Zurich office.
Uh, and he's working, you know, halftime with us and halftime with, uh, ETH, where he's been for a professor for a while. Uh, so yeah, we're, we're hardware agno agnostic and we build our own [00:07:00] hardware. We have, you know, a variety of, uh, hardware activities. You've probably seen the, the jumping bicycle, the, you know, the park parkour bicycle.
Uh, we have, uh, arms and hands and torsos that we've been building and we're, you know, we're building more stuff.
Brian Heater: Yeah, you've been pretty open about this too, that, you know, you were, I don't wanna speak for you, but may, maybe to a certain extent you were, uh, hesitant to productize because you feel like productization can kind of, in a way, get in the way of innovation.
Marc Raibert: Hesitance too weak. We are not making any products. Uh, I think that we're really trying to work on the future, uh, the next generation after the current generation. You know, we wanna make robots. Really smarter, more like eat or interact with more like people. You know, people know so much already that the task you're giving them is just kind of fits into the world, uh, that they know about.
They have understand situational [00:08:00] understanding and that's really important I think, for getting robots to be the next level beyond, uh, where they are now. You know, anybody who uses a robot today finds out pretty quickly what the limitations are. And, you know, that doesn't mean they're not useful, but it's a lot of work, uh, to, uh, make them do the things that, that they need, that you want them to do.
So we're trying to, you know, go past that. I think when you work on a product, there's lots of demands that you don't need to be working on if you're working on the future. Uh, underlying technology like the reliability of the robot. You know, Boston Dynamics spends a lot of resources on making their, you know, the spot robot out in the thousands of hours of, uh, uh, time between failures or between interventions.
And, uh, you know, we want to concentrate on getting the new thing just starting to work.
Brian Heater: Yeah, I, I suspect maybe one of the biggest hurdles there. Um, and I, and I'm curious, you know, what your thoughts are as far as the, uh, [00:09:00] electric Atlas humanoid, but is pricing, right? I mean, fossil Dynamics has been able to build these incredible robots for, for decades now, but in terms of actually being able to sell a humanoid robot, you kind of have to.
Rethink it from scratch.
Marc Raibert: Well, it's true at the institute, we're not worrying about how much the hardware costs to build. Yeah. You know, that's, uh, that's another specialized skill requiring a whole system of supply chain and, uh, uh, understanding what volume you're gonna make the thing at. Uh, and we're, we're not doing any of that.
We're making prototypes that can be very expensive and that lets you, uh, you know, explore more quickly. Uh, than if you're trying to design a product.
Brian Heater: What was that initial conversation with Hyundai like when you were trying to get this together? Uh, what was the pitch like, we're gonna spend all this money and not make any products?
Marc Raibert: I think, uh, you know, I, I wrote up a, a short document, uh, [00:10:00] that they reviewed and then we had a conversation and I think, uh, they believe that the, uh, investing in the future is, uh, an important thing for them to do. It wasn't a hard ...
Brian Heater: sell. Yeah, it, it's interesting too. I mean, one of the things that I've noticed being more involved in robotics, in the robotics community over the past several years is there is.
There's a sense of openness that you get in that world. As somebody who's covered like Apple for 20 years, there's a sense of openness, you know, working with different institutes. Obviously you mentioned ETH and you come from MIT. But there, there is a lot of sort of sharing of information and, and cross collaboration that happens.
Marc Raibert: Well one of our goals here is to take advantage of that, uh, in both directions. You know, I like to say that we're a halfway between a corporate lab and an academic lab. Academic labs are as open as you can find anywhere where, you know, they wanna share everything. Corporate labs usually don't want to share that much.
Um, and we're trying to be in the middle [00:11:00] somewhere. So we are, people write papers, we open source some amount of work, uh, and you know, we're trying to take advantage of that community, help build the community. You know, we're funding about a dozen university labs. Uh, and we are actually, uh, at least in the most recent negotiations, asking them to open source everything they do, which both makes it easier for us to access it and also the, the community to access it.
So I think openness is a good thing. Uh, and, you know, we're trying to participate in that, but we're still not as open as a, you know, a university. I'm most interested in making the technical progress and really solving some of the hard problems. Uh, I think some people are just inspired by the sharing the, the goodness of, you know, feeling about sharing stuff.
Less so for me, but, you know, whatever it takes.
Brian Heater: Yeah. I mean, it, it, it's undeniably an exciting time right now.
Brian Heater: you, you've got, you know, you, you've got a, obviously a lot of excitement around, um, AI and LLMs, but also all of this hardware coming out at the same time. Does this moment feel different for you in any meaningful way?
Marc Raibert: Now I've been in robotics for an actual 50 years. And to be fair, you know, there've been ups and downs over that whole period. I mean, there was a time in the eighties when DARPA was, uh, involved when all of a sudden, you know, AI was the hot thing and there was lots of funding. And then, you know, a year later, the someone at the office.
At DARPA would change and all of a sudden it's, you know, uh, down. So there have been some up and downs, but I don't think there's ever been an up as high as we are now. Both in terms of the availability of technology. Uh, I think everybody's spoiled now, but over my time, you know, computers have gone from [00:13:00] a big thing in a room that's expensive to are in our pockets with enough compute to do a lot of the tasks in a robot.
Uh, I think that there's been a lot of progress in educating people on building hardware and the availability of hardware and, uh, and now AI is the latest, uh, you know, cherry on the cake. Uh, you know, I'm in the middle somewhere. I think that. The AI revolution that's happening right now with LLMs and foundation models is really good and important, but it's not like it's just automatically overnight solving everything for robotics.
I think, uh, you know, physical, physically embodied robots have some extra demands that, uh, that we're working on, and we'll make progress there too.
Brian Heater: Yeah, I mean, I, I assume to a certain extent, part of your job, and maybe this is why. You release like bloopers for example, is, is like tempering expectations as far as like what robots can actually do right now?
Marc Raibert: I think [00:14:00] expectations are incredibly high and people just, especially with humanoid robots, people just assume that it's like a person and in almost every respect, it's not like a person, you know, it doesn't understand like a person. It doesn't have, they don't generally have the common sense of a person.
Uh. You know, you could come into my office and sit down in the chair. You probably have and sit down the chair without any instructions, without any maps. Robots still aren't quite there. They, you know, they're getting better in that dimension. Uh, you know, I could, you could do on the job training, uh, to do almost anything a human can do.
Uh, and you know, robots don't have that. You know, I actually think that the whole word humanoid is kind of a misnomer. People have kind of treated a humanoid. If it's got two arms and two legs and a head like thing, they call it a humanoid. But the things that make [00:15:00] humans human and some robots more like a human are the.
A combination of the intellectual capabilities, the, the things that the intellectual capabilities unlock in the physical capabilities. You know, our dexterity ability to use our hands to do complicated things is partly a mechanical design thing, but partly an intellectual capability. And you could have that capability in all sorts of structures.
You know, two arms, two legs, and a head, or. One arm, six legs or six arms in, you know, one leg. So. I'm a little disappointed that everybody is so glommed on to this idea that two arms to legs and a head makes a humanoid robot and that's the thing we should be working on. 'cause it's like a person.
Brian Heater: Yeah. I, I, as of the recording of this, a few weeks ago, I was in London for [00:16:00] the, uh, humanoid Robotics Summit and there was a speaker from DeepMind and he said something that really jumped out at me, which is.
At Google, we believe that general purpose robots will come in all shapes and sizes. In terms of, um, let, let's go way back in terms of, as you said, there have been a lot of ebbs and flows and ups and downs around excitement and breakthroughs in robotics. When the leg lab was founded at MITA few years back, what was the climate like at the time?
Marc Raibert: Uh, so first of all, the leg labs started when I was back at CMU, uh, in 1980. 1981. Uh, when, when that happened, uh, you know, robots robotics wasn't anything like the popular mainstream area that it is today. Uh, there were people [00:17:00] building robots. Uh, Engel Berger had his industrial robot. Uh, the Unimate. When I moved to MIT in, you know, in the later eighties, there was, uh, some, a couple of arms available that we did research on.
I think there were probably only maybe five US universities doing robotics at the time, whereas now there's many more than that. Uh, in the eighties, Japan was seen as the center of robotics. That, that there were many groups there working on it. I I don't really remember specifically by the years. I know that, um, in the early two thousands we started proposing to do at, at Boston Dynamics, we started to propose, uh, doing the big dog project, which was really a big influence on us, uh, in particular on Boston Dynamics in particular.
'cause we hadn't been doing much in the way of physical robots [00:18:00] up till that point. Uh, we, we'd been working with Sony for a few years. On their IBO and, uh, Cheerio. Uh, but that was it.
Brian Heater: My, my understanding of the Big Doc project. Obviously, again, you were heavily working with DARPA at the time, um, that it was basically kind of a, a pack robot, a a a, a transportation, uh, robot, but ultimately.
In, in the days when you weren't working with electric Motors, that it was like entirely too loud to be in and around the battlefield.
Marc Raibert: So it is a funny thing, uh, you know, you're sort of referring to all the news about why the program ended, but the fact is that the program actually didn't end, uh, when the media thought it did.
It is true that Google acquired us. Uh, but we still had funds and we were still working on the DARPA project, uh, into our, uh, period at [00:19:00] Google. Uh, and even though some people had reported that it was too loud, and so we, and for some missions it might have been too loud, there were plenty of people still interested in the, in the loud robots.
So, uh, I guess I'm trying to set the story today, even though I appreciate it. Uh, the minute you reported that it was too loud and the program got canceled, it never got canceled. We just, uh, uh, several years later decided that, uh, uh, it was in our interest to, uh, work on more commercial things and we, uh, actually voluntarily, uh, ended the, the contract.
Brian Heater: I, I, I asked partly because again, I was, uh, watching some of the talks that you were giving last year and you said that spot was not built for any one task. Now Spot obviously is a descendant of, of the big dog robot. Um. Building and, and you mentioned engelberg, you mentioned these arms. You mentioned, you know, these industrial robots that are very specifically built for a single task [00:20:00] in a sense, working on a, I don't wanna say general purpose robot, but a, I guess a potentially general purpose robot.
Is there a way in which you're also kind, you're almost kind of working backwards as far as figuring out what it can do and, and why it's useful?
Marc Raibert: Yeah, I mean, our thinking was, uh, that we'd been around for a long time. Uh, we had ma you know, the technology that we were building kept maturing, but we were still doing research.
And then we thought, well, one way to do research is to get this thing out there and see what people can use it for. And so we specifically designed spot to be a platform where you could attach physical devices to it, like sensors that we hadn't made. You could also attach software to it through using it, through controlling it through the API.
And so we just started selling them to whoever wanted them. We mostly lots of universities, innovation labs, some people who had applications in mind. [00:21:00] I think Boston Dynamics has matured beyond that kind of. Putting it out there for any use and now has focused on more vertical things and provided the kinds of support that the robot needs in order to be used in those, uh, specific areas.
You know, service, integration services, some other kinds of software that make it possible to have a fleet of robots in a facility. And that's really, uh, helping Boston Dynamics with promoting spot and eventually will help with the other robots that they're developing.
Brian Heater: uh, talking, again, getting back to this idea of, you know, ebbing and flowing of, I guess I.
Uh, we to cynically call them hype cycles. Uh, you know that that moment when, when Google purchased Boston Dynamics, obviously a big change, you know, longstanding, um, research facility, uh, that must have felt like one of those potentially really exciting moments. They [00:22:00] bought about a dozen companies and they were really trying to do something really big.
Did you feel at the time, like you were on the cusp of something?
Marc Raibert: I think what, um, yeah, I thought it, it was a great thing. I thought that, uh, uh, that Google had a lot of resources and, uh, you know, this field needs resources in order to make progress. They had a lot of brains and a lot of, uh, you know, computing in the brain, in the AI area, and, uh, that seemed like a great thing.
Brian Heater: Were they pushing you at the time to productize, or did that really come entirely under Hyundai? Uh, more, more under SoftBank. SoftBank, okay. Soft. So in the middle,
Marc Raibert: Google and then Hyundai more recently.
Brian Heater: So they, they were happy to have you be a, a strictly research facility at the time.
Marc Raibert: I mean, we kept doing what we had always done, you know, with, with me in charge.
We're a research organization. I think it's always been that way. Uh, I haven't, I've never been that [00:23:00] much. Concerned about commercialization.
Brian Heater: Yeah. One, one thing that is becoming really apparent to me over the last few years, again, I've been a journalist for 20 years, so maybe when we cover things, it's a little bit more, uh, black and white.
And, and there aren't the nuances that you see within a company. But even in situations like that, Google buying a bunch of robotics companies, even if it doesn't. Excuse me. Even if it doesn't end up exactly how they had planned, it's, it's not never all for Naugh and a lot of there, there's still a lot of people there, right?
There's still a lot of people from several years ago, those acquisitions and they're working on Deep Mind now, like it there, there's still forward progress happening and there's still positives that come out of an acquisition like that.
Marc Raibert: So you once had a headline. That's, no, this is relevant to your point.
Yeah. You had a headline when you were at TechCrunch, I think. Yeah. That said finally after 30 years. Yeah. Uh, they have a [00:24:00] product. And I looked at that and thought, you know, first of all, we weren't even trying to build robots for the, for a lot of that time, but beyond that, I thought you had a very narrow view of the world 'cause you were saying, well, if you don't make a product, you're not doing anything. And my view of the world is that, uh, learning about what robots can be and how to make them that, and developing the technology is just a much a, a contribution as, uh, making a product. In fact, when you make a product, you al almost seal yourself off from, uh, from the opportunity to keep advancing what the technology can do.
Brian Heater: Yeah, my, my, my outlook has obviously evolved, you know, over the last several years, but, uh, a lot of it, and I, and I think maybe this might be like a broader thing with a lot of the journalism happening around robotics. We don't have to go too deeply into this, but obviously something I think a lot about is that it's people coming over from the consumer side, like myself who are coming over from the Apple side of the world.
You're coming over [00:25:00] from research, which is like an entirely different beast.
Marc Raibert: Yeah. I'm in this because I love. Both building the things and, and creating things that haven't existed before. You know, one of the things I brag about is I love that I'm an engineer, not a scientist. I, I think people think that scientists are here and engineers are here, but I think of it this way.
I think that. Scientists, really their job is just to describe what already exists and explain it. Whereas we engineers get to do all that and also create things that never existed before. So it's like, it's like an art form. If you, uh, if you're successful,
Brian Heater: yeah. I mean, maybe this isn't entirely clear with, again, you know, like 40 years of hindsight, but, um, again, going, again, going way back to, to the leg lab, as you were saying, there were industrial arms and that was about it at the time.
Was that really what we would now call a moonshot?
Marc Raibert: You know, for me, the key event that started. Things for me was seeing a robot that had six legs. I was interested in animals and I saw this six legged robot that was moving so slowly. You know, it has six legs because then it couldn't tip over 'cause it could move three at a time and keep a tripod of support.
It went really slow because if it was going too fast, it still might tip over. Uh, if it tried to stop or start quickly, you know? Um, and I just looked at that and said, wow, there couldn't be more difference between that and how an animal works where you run, you know? Or a person works where you fly through the air sometimes.
And then you have these springy legs and you bounce off them. And that was just a chance for, you know, to try something different and, and, uh, you know, may possibly make something cool. And, you know, I'm very pleased with having done that. And I think some of those ideas still dominate, you know, what we're doing today.
You know, don't, don't do slow [00:27:00] static. Uh, you know, I would say the same thing about robot manipulation, where we're still stuck in the, here I'll, I'll do a, a demo. Um, you know, here, here's an object and a lot of the robotics world is figuring out how to grasp the stable object by looking at the geometry and looking where it is.
Well, geez, people don't handle objects that way. Right. We can handle 'em and manipulate them and, and toss them, and, uh. And I think dexterity, you know, the ability to use our hands to do useful work comes from that dynamic interaction much more than static grasping. So I think there's a chance for another revolution in robotics doing that.
And that's, you know, what we're trying to do here.
Brian Heater: Maybe what you're getting at is really, um, the way that robots are trained, right? I mean, we, we've got a, we've got a couple of paradigms that we're, uh, reinforcement learning. Um, it, how much is AI going to just [00:28:00] completely blow that away and completely transform the way that robots learn how to do tasks?
Marc Raibert: I mean, it partly depends on what you're calling ai. Uh, but I, I think. We're really making rapid progress. Now, I don't know if you've seen the reinforcement learning results. We've done working with Boston Dynamics on Ellis, but we've been doing it on, uh, on other stuff as well. And you know, we've gone from doing work where we do many trials on the hardware.
Measuring the behavior and the reaction and the results. And then, you know, making decisions about how to advance to things. Where we do all that testing in simulation. Uh, and then do in some cases one shot transfers, where the first time we do a complicated maneuver on the robot, uh, it works pretty well.
Maybe we might go back in and, you know. Look at some of the performance and, and tweak things [00:29:00] that is, have a human go and look at it. But it's amazing that the simulator simulator can capture enough of the behavior. Now, the simulators aren't good enough in some domains yet, uh, to do everything that way, but I think we'll get there.
Brian Heater: I don't think this is maybe entirely accurate, but I've, I've heard humanoid specifically by some people compared to self-driving cars. From the standpoint of, and this is the case across. Many are all technologies that it's the last like 0.5% or 0.001%. That's the hardest. It's, it's the edge cases. You can build what looks to be a very capable robot, but once you actually get out in the world, like things are gonna happen.
Marc Raibert: Unfortunately, it's not just a few percent. If you look at a lot of the research that's going on that's called successful, it's getting into the. Uh, seventies, eighties percent or 90% in a few cases. And you know, that's nowhere for a product. [00:30:00] So you need much higher levels of performance than that. But you're right.
The further you go on that curve, the harder it is, the more, the more work it is to figure out how to resolve those things.
Brian Heater: Yeah, getting, getting back to transparency, and I, and I think this is really important and I think it's something that, you know, Boston Dynamics has done more and more over the years, that certainly you're doing as the research institute is showing the work that goes into a, a system.
Not gonna name any names, but I think we've seen examples of current companies where they can produce a really flashy video and, you know, um, more than anyone else, how much work goes into that, obviously. Videos, viral videos were a big part of, you know, getting the word out there for, for Boston Dynamics.
What did you get, did you get a lot of pushback as far as the things that you were showing the world and what these systems were actually capable of? Uh, sometimes,
Marc Raibert: um, the one that's most memorable for me, or the, or [00:31:00] an example of the complexity of the topic you're talking about. Uh, when we had big dog climbing this hiking trail at, uh.
At the, it was the Guadalcanal Trail at the, uh, Marine Corps training base, which is where we did testing. Uh, we were honestly trying to show that the robot could get up the hill, uh, and that that was the achievement. And we had a person driving the robot up the hill. So while the perception was done by the person, the person wasn't in the shot.
And we were trying to show that the robot had the capability getting up the hill, which at the time we thought was a, a very strong achievement. And I think it was still, but some people thought we were cheating. It turned out, you know, maybe years later people would criticize us for cheating because we hadn't shown the person in the scene who was driving.
So I have some sympathy for the case. You were talking, the case you're talking about, you [00:32:00] know, there's a lot of things to show. Uh, about progress. You could show that the hardware is capable of the behavior. You could show that the brain, it's much harder to show that the brain is able to, by itself, control the robot.
But, you know, people talk about full autonomy, meaning that there's no person anywhere in the picture, but there's always, you know, you, you, Brian, don't have full autonomy. You know, you have a family who, uh, probably constrains what you do. At one point you had parents, there's the laws that you obey. Uh, so the idea of full autonomy is really a, a screwed up notion in my mind.
And, uh, so I'm being def maybe defensive that, uh, when you're showing a video of something, you know, you have to pick your target and try and show whatever it is. Obviously, you don't wanna. Uh, cherry pick one successful result, [00:33:00] uh, in a sea of failures, uh, and try and claim, uh, that it works all the time. Uh, but I also think that showing what's possible, even if it's not reliably possible, is can be a useful thing to do too.
We could go on for hours about the nuances of, uh, showing one's work in video. Obviously Boston Dynamics was very successful at that and. Sort of put us on the map. We were lucky we cut YouTube in its early days when it wasn't so congested, uh, with other stuff. Uh, and you know, we kind of just fell into it, uh, by accident.
It wasn't a big strategy for us. You asked about bloopers before. Um, you know, bloopers in my mind are partly to show what's realistic, but the thing I found is if people have seen your bloopers. They're much more impressed with your successes 'cause they can see how hard it is. Uh, the very [00:34:00] first time I showed the one-legged hopping robot, it was boring to just see it balancing and hopping.
It looked easy. But the first time you saw a failure where it could fall over, uh, almost instantaneously if something went wrong, then you, then you appreciated what the good behavior was. So that's another motivation for showing it. And also. It's just fun. It makes the whole field fun to, uh, see the challenge you're up against and see all the different ways things can go wrong.
I mean, it's incredible.
Brian Heater: Yeah. Maybe something that doesn't get talked about enough. Uh, and, and obviously YouTube plays a, a big role in this, is that. A lot of times when these videos go, go out, like the audience is everybody, right? So you've gotta appeal to the, the viral people on, on the gadget blogs and now TikTok and everything else, and at the same time show what the robot is capable of doing.
And that's, it's impossible to, to produce a short video that hits all of [00:35:00] those points.
Marc Raibert: Yeah. I mean, you, you know, you have an audience, you, there's an audience you care about. There's other audiences. Uh, sometimes it's hard to figure out what you wanna really try and accomplish.
Brian Heater: Yeah. And, and maybe another, uh, dimension of this too, in, in talking about bringing into productization is the pressure that comes with venture capital.
You know, when one, once they, once they enter the picture, um, you kind of really have to focus on one or two things it seems like, and, and I get the sense, and I wonder if you. Feel this way too, that there, there was so much excitement about humanoids over the past few years that we've really got caught in one very specific model that for a while there, um, you know, we were being told that that wheels were a thing of the past, but I know that you're fairly agnostic when it comes to things like that.
Even as the founder of the Leg lab,
Marc Raibert: it's funny that you're talking in the past tense as though the intense. Uh, [00:36:00] humanoid isn't, isn't ongoing, which, which I think it is. Um, you know, you said, you know, so I have never been, uh, an investee of a, of venture capitalists in my whole career, so I don't really know firsthand what the pressures are.
There's, there's lots of purposes for showing your work. There's recruiting talent. There's showing that you're making progress. There's kind of the personal pride of, uh, of showing that you, uh, you know, you move the needle ahead. I think that's a big one. Uh, for me, I like, uh, I like it when we come up with a new thing.
And, you know, I used to say that when I was a professor, I would count how many people read my papers, you know, through citations. Now I count YouTube hits and likes and dislikes. Those are the, the coin of the realm. Uh, sometimes.
Brian Heater: I mean, how does that change the math for you though, as far as what the audience is and, [00:37:00] and what the feedback is and, and you know, whether you're getting likes or dislikes.
Does that actually have an impact on the work that you're doing?
Marc Raibert: The only thing we tangibly did is, uh, adjust how much pushing on the robot we do
Brian Heater: you mean literally, like literally the hockey stick and literally like kicking and, yeah, for a while there.
Marc Raibert: Uh, you know, I didn't care if people said, oh, you're being mean to the robot, but some of the people, other people I work with cared, uh, some, sometimes our owners cared, so we would, uh, adjust the style that we showed, uh, the behavior of the robot.
You know, to be clear, the hockey stick was a way of showing how good the robot was performing. We were, we were like proud parents showing off the balancing capability of the robot, and we weren't being mean in any sense. You know, the robot doesn't have any feelings or, uh, ability to feel pain or anything, right?
We were just proud parents showing off what the [00:38:00] robot could do. Uh, but the fact that some people in the public reacted how they did, uh, made us adjust. How much of that or what style of, uh. Uh, we did those kinds of, uh, tests with
Brian Heater: Yeah, I, I, I think that, that, that is a big risk that you run when nature and biology are a huge influence.
As you said earlier, that the more lifelike a robot is, the more you know, your reptile brain will react negatively to anything like that. Right.
Marc Raibert: You know, it was perceived as abuse. So I guess if you're, if it's perceived as abuse by some, then, then maybe you wanna a, adjust what you're doing. That's, I guess, what we've done.
Brian Heater: As we're kind of, uh, coming up on time and, and wrapping up here, I'm, I'm just really curious from your standpoint, whether it's what you're doing at the institute or what's happening outside, like what are you especially excited about in robotics and ai?
Marc Raibert: You know, I think we've come a long way. In the time [00:39:00] I've been doing robotics, but we still have a long way to go and I think we have the resources and the talent and the, and the interest both from within and from, you know, sponsors to keep pushing on that.
So, you know, that's the fundamental In terms of specific technical things, you know, I've been really gassed by, uh, what the reinforcement learning progress, how that's influenced controls. I'm really excited about the challenge of. Making dexterity happen. I think it, it's been a little bit of a failure in robotics that, that they're not that good with their hands.
You know, they get around pretty well now. Uh, they can see things and perceive, at least in some cases, you know, you, you can find all the cats in your pictures on your phone, and I think we can use some of that, even though, uh, a moving robot in the world has additional perception requirements. Uh, but despite those advances, and of course computers have gotten, uh, massively better [00:40:00] despite all those things, we're still pretty limited in what we can do with, uh, with our hands.
But, you know, we're devoted to making progress on that, and I think that's gonna have a big impact on the productivity of robots. Even though I'm not gonna be making the product that uses it, I think it'll impact the productivity and that's gonna fuel, uh, you know, more work and more progress and more. Uh, interest from, from all over.
Brian Heater: I know again that you're not, obviously, as we've said many times, you're not really a product guy, but as somebody who has lived through these again, hype cycles over the years, um, how confident are you that like five to 10 years from now, it'll be meaningfully different and that the, that there will actually be, you know, robots out in the world in a meaningful way?
Marc Raibert: I'll point at three layers. Uh, there's. Moving stuff around, like in warehouses. I think that's happening and you'll continue to see lots of valuable work, [00:41:00] pr, you know, money making work going on there. I think the factory is next, and even though. I'll call them old fashioned. It's not fair. You know, fix the arms, already do a lot of things in factories.
I think having robots that are traveling around and doing more complex human-like tasks is the next frontier. And I think the home is the, is gonna be the toughest frontier. You know? The home suffers from three simultaneous requirements. Safety among very tender people cost. Among normal family budgets or household budgets and performance, doing enough useful stuff to justify, uh, the cost.
And so that's a, getting all three of those in one package is gonna take a while. I think factories are more organized. You can have more rules. You can keep the people and the robots separated enough or at least educated about what their interactions [00:42:00] are to make them safe enough. I think there's enough opportunity for return on investment, so I think that's the next frontier.
And I think warehouses is, you know, obviously still a lot to do, but is heavily underway.
Brian Heater: It is really funny. I, I've heard a lot of people in talking about, I guess what they used to call elder care and what we now call age tech is the reason why a lot of folks are invested in it is because when they hit that point in their life, they're gonna want these robots to actually be there to help them.
Um, I, I think you're gonna be with us for a long time, but do you feel like while you're still around that, that. You know that you will have a robot in the home to help you do the things you need to do?
Marc Raibert: Well, I absolutely want to have my own car, self-driving car when I, you know, don't have the skills to drive myself.
Fortunately, I still can. Um, I did have an elderly aunt, uh, my, my mother's generation was, my mother's sister lived with me, uh, through her last years, and we had, [00:43:00] you know, hired human people, humans to take care of her. And she hated that there were strangers there, there was a level of intimacy with strangers that she didn't like.
Um, and I think I saw the opportunity that robots would've been a better solution, even though some people think, oh, that's cold, to have a robot taking care of you. But I'm absolutely certain she would've preferred that to the strangers. She, I'm sure she would've preferred it for me, take to me taking care of her, which I didn't, you know, do in an intimate way, even though she was in my home.
Um. So, yeah, I think, uh, that day is coming and I, I hope to be, uh, to take advantage of it.
Brian Heater: Great. Well, mark, always a pleasure. Thank you so much.
Marc Raibert: Okay, thanks Brian. Good to talk to you.
Brian Heater: There you go. Episode one of Automated is in the books. Thank you so much to Mark Rayer for taking the time to speak with us.
You can learn more about the Robotics and AI Institute over [00:44:00] at RAI dash INS t.com. They're doing some really fascinating work over there, and as always. With projects that Mark is involved in, highly recommend subscribing to their YouTube channel. Um, thank you for listening. You could tell everyone that you got in on the ground floor of our little podcast, and this is actually one of two automated properties that we are launching this week.
The other is my new newsletter of the same name. You can subscribe to that
[email protected] for news interviews, job listings, and more talk soon.
