Lex Fridman Podcast
#412 – Marc Raibert: Boston Dynamics and the Future of Robotics
Fri, 16 Feb 2024
Marc Raibert is founder and former long-time CEO of Boston Dynamics, and recently Executive Director of the newly-created Boston Dynamics AI Institute. Please support this podcast by checking out our sponsors: - HiddenLayer: https://hiddenlayer.com/lex - Babbel: https://babbel.com/lexpod and use code Lexpod to get 55% off - MasterClass: https://masterclass.com/lexpod to get 15% off - NetSuite: http://netsuite.com/lex to get free product tour - ExpressVPN: https://expressvpn.com/lexpod to get 3 months free Transcript: https://lexfridman.com/marc-raibert-transcript EPISODE LINKS: Boston Dynamics AI Institute: https://theaiinstitute.com/ Boston Dynamics YouTube: https://youtube.com/@bostondynamics Boston Dynamics X: https://x.com/BostonDynamics Boston Dynamics Instagram: https://instagram.com/bostondynamicsofficial Boston Dynamics Website: https://bostondynamics.com/ PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://spoti.fi/2nEwCF8 RSS: https://lexfridman.com/feed/podcast/ YouTube Full Episodes: https://youtube.com/lexfridman YouTube Clips: https://youtube.com/lexclips SUPPORT & CONNECT: - Check out the sponsors above, it's the best way to support this podcast - Support on Patreon: https://www.patreon.com/lexfridman - Twitter: https://twitter.com/lexfridman - Instagram: https://www.instagram.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/lexfridman - Medium: https://medium.com/@lexfridman OUTLINE: Here's the timestamps for the episode. On some podcast players you should be able to click the timestamp to jump to that time. (00:00) - Introduction (10:12) - Early robots (15:15) - Legged robots (33:55) - Boston Dynamics (37:13) - BigDog (45:20) - Hydraulic actuation (47:12) - Natural movement (52:59) - Leg Lab (59:51) - AI Institute (1:03:09) - Athletic intelligence (1:11:04) - Building a team (1:14:05) - Videos (1:21:53) - Engineering (1:25:21) - Dancing robots (1:30:08) - Hiring (1:34:00) - Optimus robot (1:42:30) - Future of robotics (1:47:24) - Advice for young people
The following is a conversation with Mark Rybert, a legendary roboticist, founder, and longtime CEO of Boston Dynamics, and recently the executive director of the newly created Boston Dynamics AI Institute that focuses on research and the cutting edge on creating future generations of robots that are far better than anything that exists today.
He has been leading the creation of incredible legged robots for over 40 years at CMU, at MIT, the legendary MIT Leg Lab, and then, of course, Boston Dynamics with amazing robots like Big Dog, Atlas, Spot, and Handle. This was a big honor and pleasure for me. And now a quick few second mention of each sponsor. Check them out in the description. It's the best way to support this podcast.
We got Hidden Layer for securing your AI and machine learning models, Babbel for learning new languages, Masterclass for learning, NetSuite for business management software, and ExpressVPN for privacy and security on the internet. Choose wisely, my friends. Also, if you want to work with our amazing team or just get in touch with me, go to lexfriedman.com slash contact.
And now onto the full ad reads. As always, no ads in the middle. I try to make these interesting, but if you skip them, please still check out the sponsors. I enjoy their stuff. Maybe you will too. This episode is brought to you by a new sponsor, an amazing sponsor called Hidden Layer. It's a platform that provides security for your machine learning models.
If you've been paying any attention, it's obvious that generative AI, machine learning, AI in general. It's going to transform basically every single industry.
If you're starting a business, if you're running a business, there's probably a lot of ways that language models or any of these generative system could be used to help automate certain things, to help empower certain things, improve certain things, all that kind of stuff, to work together with humans to improve the productivity.
Now, what comes with great, powerful new technology is security threats. There's always bad guys out there and they want to find ways in. Adversarial attacks, they want to trick you. You need experts to help you incorporate the machine learning models that you're using in a secure way. That's what hidden layer does. I think that's really interesting because the kinds of attacks...
on machine learning models is going to be really, really interesting. Because one way to do that is if you look at a model zoo repository, like Hugging Face, for example, one way is to get a model in there that looks legit, but has weirdness to it, that are hard to discover until they're actually way down the line being used at a large scale. So that's like one way to do it.
The other way is to do outside repositories and deliver the models that look legit, but again, in subtle ways, difficult to detect ways. It will look legit at first, it will look functional and correct at first, but then the kind of malware inside the machine will emerge. down the line.
So you want experts to be considering this rapidly evolving security threat that is laden within machine learning systems. Visit hiddenlayer.com slash Lex to learn more about Hidden Layer. These guys are great. They're great sponsors of this podcast. They do a really important service for the whole machine learning community. So check them out.
This episode is brought to you by Babbel, an app and website that gets you speaking in a new language within weeks. Russian, Spanish, French, German, Italian, Portuguese and more. I can keep going, but those are the ones I want to speak. German, Italian. I want to at least order stuff in Italian in a restaurant. And that's the kind of stuff you can learn with Babbel really quick.
It's like practical conversational stuff. So when you're traveling, you can use it. Portuguese, obviously, you know. I practiced a martial art called Brazilian Jiu-Jitsu. Portuguese, Brazil, I got to do it. Spanish, same thing. I'm a huge fan of soccer, a.k.a. football, and of course dream of one day interviewing some of these said soccer players and the other languages like Russian.
We're going to try to do a bunch of different translation and overdubbing very, very soon with the various podcast conversations I've been doing. I think one of the most powerful ways to bridge barriers between people is... breaking through the wall that language creates, automated translation.
And then when you actually meet them in person to speak their language or to attempt to speak their language, that's what, again, Babbel is great for. I use it, all the languages I mentioned, I've used it to learn that, even to practice my Russian. For a limited time, get 50% off a one-time payment for a lifetime Babbel subscription at Babbel.com slash LexPod.
That's 50% off at Babbel.com slash LexPod. Spelled B-A-B-B-E-L dot com slash LexPod. Rules and restrictions apply. This episode is also brought to you by Masterclass, where you can watch 180 classes from the best people in the world in their respective thing. Phil Ivey on poker. I'm a big fan of his. I'm just a big fan of poker. Aaron Franklin on barbecue and brisket.
I just ate at Terry Black's, and I've also eaten at Franklin's. Those are great barbecue joints here in Austin, Texas, and it's just delicious. I love it. There's another place I really love called JNL Barbecue. They're great. I mean, they're all like incredible in their own way. And the personalities of people that are running these places is just incredible. They all really love barbecue.
And Aaron Franklin is one of those people, but he's also really good at communicating how to make barbecue. He has great cookbooks. He has great videos. He's just good. He's a great teacher. No, he's a great celebrator of the beauty of cooking. And cooking this particular thing, which is, in this case, brisket. I mean, I highly recommend it.
It's like watching an artist describe what he loves to do. I watched it and I really enjoyed it. Get unlimited access to every Masterclass and get an additional 15% off an annual membership at Masterclass.com slash LexPod. That's Masterclass.com slash LexPod.
This episode is also brought to you by NetSuite, an all-in-one cloud business management system, ERP system as the kids call it, or the cool kids call it, or the experts call it. It's not really the cool kids, the experts. I think it stands for enterprise resource planning, but I could be wrong. Anyway, ERP, everybody knows what that is.
If you think of a company as a giant machine with different components and modules and so on, ERP is the kind of the brain of the operation that connects everything, make sure that everything speaks the common language and all that kind of stuff. That's what NetSuite does. It does it really well. manages financials, human resources, inventory and supply.
If you do that kind of thing, if you sell a kind of thing, and if you sell that kind of thing online, it also does e-commerce. It is much, much more, much, much more cool stuff. Anyway, you can even check on Reddit. A lot of people are super positive about it. So my 37,000 companies have upgraded to NetSuite by Oracle, and they're turning 25 this year. Wish them a happy birthday.
Write them letters. Send them faxes. Shout out from the rooftops. Happy birthday. It is always good to celebrate. Others celebrate companies, especially when they lasted this long. It's not easy out there for a company. So you should let NetSuite make it a tiny, tiny, tiny bit easier for you. Download NetSuite's popular KPI checklist for free at netsuite.com.
That's netsuite.com for your own KPI checklist. This episode is brought to you by ExpressVPN. I use them to protect my privacy on the internet. I've used them for many years. You know, it's the privacy thing, for sure. It's also the happiness thing. I press the button, turns on, I pick a location, instantaneously, space-time travel, solved.
I don't understand what all the hoopla is about, like, uh, Faster than light travel is not possible. You click a button. If you click it fast enough, right there, you're in a different location, a different country. It opens up all kinds of possibilities.
Just press the button, close your eyes, and imagine you're in London among all those people with the fancy accents that make you sound really smart. And that allows you to watch different content that's geographically restricted, for example. But again, I think the privacy thing is really the important thing. Everybody should be using VPN. ExpressVPN is my favorite one. It's fast.
It works on any device, any operating system. Linux is the best, the sexiest, most powerful operating system, but not many people use it. Maybe because they're not cool. If you want to be one of the cool kids, you should be using Linux. To this day, Ubuntu is my go-to Linux flavor. It used to be Gentoo. And the reason I say that is because it rhymes with go-to.
So you should go to expressvpn.com for an extra three months free. This is the Lex Friedman Podcast. To support it, please check out our sponsors in the description. And now, dear friends, here's Mark Rybert. When did you first fall in love with robotics?
Well, I was always a builder from a young age. I was lucky. My father was a frustrated engineer. And by that, I mean he wanted to be an aerospace engineer, but his mom from the old country thought that that would be like a grease monkey. And so she said no. So he became an accountant. But the result of that was our basement was always full of tools and equipment and electronics.
And from a young age, I would watch him assembling a kit, an Ico kit or something like that. I still have a couple of his old Ico kits. But it was really during graduate school when I followed a professor back from class. It was Bertolt Horn at MIT. And I was taking a an interim class. It's IAP, independent activities period. And I followed him back to his lab.
1974.
1974. So there's just this arm in pieces. Yeah. And you saw the pieces and you saw in your vision the arm when it's put back together and the possibilities that holds.
Somehow it spurred my imagination. I was in the brain and cognitive sciences department as a graduate student doing neurophysiology. I'd been an electrical engineer as an undergrad at Northeastern. The neurophysiology wasn't really working for me. It wasn't conceptual enough.
I couldn't see really how by looking at single neurons you were going to get to a place where you could understand control systems or thought or anything like that. And, you know, the AI lab was always an appealing, this was before CSAIL, right? This was in the 70s. So the AI lab was always an appealing idea.
And so when I went back to the AI lab with, you know, following him and I saw the arm, I just thought, you know, this is it.
It's so interesting, the tension between the BCS, brain cognitive science approach, to understanding intelligence, and the robotics approach to understanding intelligence.
Well, BCS is now morphed a bit, right? They have the Center for Brains, Minds, and Machines, which is trying to bridge that gap. And even when I was there, you know, David Marr was in the AI lab. David Marr had models of the brain that were appealing both to biologists but also to computer people. So he was a visitor in the AI lab at the time, and I guess he became full-time there.
So that was the first time a bridge was made between those two groups. Then the bridge kind of went away, and then there was another time in the 80s. And then recently, the last five or so years, there's been a stronger connection.
You said you were always kind of a builder. What stands out to you in memory of a thing you've built? Maybe a trivial thing that just kind of inspired you in the possibilities that this direction of work might hold.
I mean, we were just doing gadgets when we were kids. I had a friend, we were I don't know if everybody remembers, but fluorescent lights had this little aluminum cylinder. I can't even remember what it's called now. You needed a starter, I think it was. We would take those apart, fill them with match heads, put a tail on it, and make it into little rockets.
So it wasn't always about function. Well, Rocket was pretty much- I guess that is pretty functional. But yeah, I guess that is a question. How much was it about function versus just creating something cool?
I think it's still a balance between those two. There was a time though when I was, I guess I was probably already a professor or maybe late in graduate school when I thought that function was everything and that mobility, dexterity, perception, and intelligence, those are sort of the key functionalities for robotics, that that's what mattered and nothing else mattered.
And I even had kind of this platonic ideal that if you just looked at a robot and it wasn't doing anything, it would look like a pile of junk, which a lot of my robots looked like in those days. But then when it started moving, you'd get the idea that it had some kind of life or some kind of interest in its movement.
And I think we purposely even designed the machines, not worrying about the aesthetics of the structure itself. But then it turns out that the aesthetics of the thing itself add and combine with the lifelike things that the robots can do. But the heart of it is making them do things that are interesting.
So one of the things that underlies a lot of your work is that the robots you create, the systems you have created for over 40 years now, have a kind of, they're not cautious. So a lot of robots that people know about move about this world very cautiously, carefully, very afraid of the world. A lot of the robots you built, especially in the early days, were very aggressive, under-actuated.
They're hopping. They're wild, moving quickly. Is there a philosophy underlying that?
Well, let me tell you about how I got started on legs at all. When I was still a graduate student, I went to a conference. It was a biological legged locomotion conference. I think it was in Philadelphia. So it was all biomechanics people, researchers who would look at muscle and maybe neurons and things like that. They weren't so much computational people, but they were more biomechanics.
And maybe there were a thousand people there. And I went to a talk. One of the talks, all the talks were about the body of either animals or people and respiration, things like that. But one talk was by a robotics guy, and he showed a six-legged robot that walked very slowly.
It always had at least three feet on the ground, so it worked like a table or a chair with tripod stability, and it moved really slowly. And I just looked at that and said, Wow, that's wrong. That's not anything like how people and animals work. Because we bounce and fly, we have to predict what's going to happen in order to keep our balance when we're taking a running step or something like that.
We use the springiness in our legs, our muscles and our tendons and things like that. As part of the story, the energy circulates. We don't just throw it away every time. I'm not sure I understood all that when I first thought, but I definitely got inspired to say, let's try the opposite. I didn't have a clue as to how to make a hopping robot work, not balance in 3D.
In fact, when I started, it was all just about the energy of bouncing. And I was going to have a springy thing in the leg and some actuator so that you could get an energy regime going of bouncing. And the idea that balance was an important part of it didn't come until a little later. And then I made the one-legged, the pogo stick robots. Now I think that we need to do that in manipulation.
If you look at robot manipulation, we, a community, has been working on it for 50 years. We're nowhere near human levels of manipulation. I mean, we can, you know, it's come along, but I think it's all too safe.
And I think trying to break out of that safety thing of static grasping, you know, if you look at a lot of work that goes on, it's about the geometry of the part, and then you figure out how to move your hand so that you can position it with respect to that, and then you grasp it carefully, and then you move it. That's not anything like how people and animals work. We juggle in our hands.
We hog multiple objects and can sort them. Now, to be fair, being more aggressive is going to mean things aren't going to work very well for a while. It's a longer-term approach to the problem. That's just theory now. Maybe that won't pay off, but that's how I'm trying to think about it, trying to encourage our group to go at it.
Well, yeah, I mean, we'll talk about what it means to, what is the actual thing we're trying to optimize for a robot. Sometimes, especially with human-robot interaction, maybe flaws is a good thing. Perfection is not necessarily the right thing to be chasing. Just like you said, maybe being good at fumbling an object
Being good at fumbling might be the right thing to optimize versus perfect modeling of the object and perfect movement of the arm to grasp that object. Because maybe perfection is not supposed to exist in the real world.
I don't know if you know my friend Matt Mason, who is the director of the Robotics Institute at Carnegie Mellon. And we go back to graduate school together. But he analyzed... a movie of Julia Childs doing a cooking thing. And she did, I think he said something like there were 40 different ways that she handled a thing and none of them was grasping.
She would nudge, roll, flatten with her knife, things like that, and none of them was grasping.
Okay, let's go back to the early days. First of all, you've created and led the legendary Leg Lab at MIT. What was that first hopping robot?
First of all, the Leg Lab actually started at Carnegie Mellon. I was a professor there starting in 1980, about 1986. And so that's where the first hopping machines were built. I guess we got the first one working in about 1982, something like that. That was a simplified one. Then we got a three-dimensional one in 1983.
The quadruped that we built at the Leg Lab, the first version, was built in about 1984 or 5 and really only got going about 86 or so. It took years of development to get it to work.
Let's just pause here. For people who don't know, I'm talking to Mark Raber, founder of Boston Dynamics, but before that, you were a professor developing some of the most incredible robots for 15 years, and before that, of course, a grad student and all that. So you've been doing this for a really long time. So you skipped over this, but go to the first Hopping Robot.
There's videos of some of this. I mean, these are incredible robots. So you talked about the very first step was to get a thing hopping up and down. Right. And then you realized, well, balancing is a thing you should care about and it's actually a solvable problem. So can you just go through how to create that robot? Sure. What was involved in creating that robot?
Well, I'm going to start on the, not the technical side, but the, I guess we could call it the motivational side or the funding side. So before Carnegie Mellon, I was actually at JPL, at the Jet Propulsion Lab for three years.
And while I was there, I connected up with Ivan Sutherland, who is sometimes regarded as the father of computer graphics because of work he did both at MIT and then University of Utah and Evans and Sutherland.
Anyway, um, I got to know him and at one point he said, uh, he encouraged me to, uh, do some kind of project, uh, at Caltech, even though I was at JPL, you know, those are kind of related institutions. And, uh, So I thought about it, and I made up a list of three possible projects. And I purposely made the top one and the bottom one really boring sounding.
And in the middle, I put Pogo Stick Robot. When he looked at it, Ivan is a brilliant guy, brilliant engineer, and a real cultivator of people. He looked at it and knew right away what the thing that was worth doing. He had an endowed chair, so he had about $3,000 that he gave me to build the first model for
which I went to the shop and with my own hands made a first model, which didn't work and was just a beginning shot at it. Ivan and I took that to Washington. In those days, you could just walk into DARPA and walk down the hallway and see who's there. Ivan, who had been there in his previous life, We walked around and we looked in the offices. Of course, I didn't know anything.
I was basically a kid, but Ivan knew his way around. We found Craig Fields in his office. Craig later became the director of DARPA, but in those days, he was a program manager. We went in. I had a little Samsonite suitcase. We opened and it had just the skeleton of this one-legged hopping robot. We showed it to him. And you could almost see the drool going down his chin of excitement.
And he sent me $250,000. He said, okay, I want to fund this. And I was between institutions. I was just about to leave JPL, and I hadn't decided yet where I was going next. And then when I landed at CMU, he sent $250,000, which in 1980 was a lot of research money.
Did you see the possibility of where this is going? Why this is an important problem? No. The balance, I mean, it's legged. It has to do with legged locomotion. I mean, it has to do with all these problems that the human body solves when we're walking, for example.
Like, all the fundamentals are there. Yeah, I mean, I think that was the motivation to try and get more at the fundamentals of how animals work. But the idea that it would result in, you know, machines that were anything like practical... like we're making now. That wasn't anywhere in my head, no.
As an academic, I was mostly just trying to do the next thing, make some progress, impress my colleagues if I could.
And have fun. And have fun. Pogo stick robot. Pogo stick robot. So what was, on the technical side, what are some of the challenges of getting to the point where we saw, like in the video, the pogo stick robot that's actually successfully hopping and then eventually doing flips and all this kind of stuff?
Well, in the very early days, I needed some better engineering than I could do myself. And I hired Ben Brown. We each had our way of contributing to the design. And we came up with a thing that could start to work. I had some stupid ideas about how the actuation system should work. And we sorted that out.
It wasn't that hard to make it balanced once you get the physical machine to be working well enough and have enough control over the degrees of freedom. And then we very quickly, you know, we started out by having it floating on an inclined air table. And then that only gave us like six foot of travel.
So once it started working, we switched to a thing that could run around the room on another device. It's hard to explain these without you seeing them, but you probably know what I'm talking about, a planarizer. And then the next big step was to make it work in 3D, which that was really the scary part.
With these simple things, you know, people had inverted pendulums at the time for years and they could control them by driving a cart back and forth. But could you make it work in three dimensions while it's bouncing and all that? But it turned out, you know, not to be that hard to do, at least at the level of performance we achieved at the time.
So, okay, you mentioned inverted pendulum, but can you explain how a hopping stick in 3D can control, can balance itself?
Yes.
What does the actuation look like?
The simple story is that there's three things going on. There's something making it bounce. We had a system that was estimating how high the robot was off the ground. Using that, there's energy that can be in three places in a pogo stick. One is in the spring, one is in the altitude, and the other is in the velocity. And so when at the top of the hop, it's all in the height.
And so you could just measure how high you're going and thereby have an idea of a lot about the cycle, and you could decide whether to put more energy in or less. So that is one element. Then there's a part that you decide where to put the foot. And if you think when you're landing on the ground with respect to the center of mass, so if you think of a pole vaulter,
The key thing the pole vaulter has to do is get its body to the right place when the pole gets stuck. If they're too far forward, they kind of get thrown backwards. If they're too far back, they go over. And what they need to do is get it so that they go mostly up to get over the thing. And high jumpers is the same kind of thing.
So there's a calculation about where to put the foot and we did something, you know, relatively simple. And then there's a third part to keep the body at an attitude that's upright. Cause if it gets too far, you know, you could hop and just keep rotating around, but if it gets too far, then you run out of motion of the joints at the hips. So you have to do that.
And we did that by applying a torque between the legs and the body. Every time the foot's on the ground, you only can do it while the foot's on the ground in the air. You know, it, it, the physics don't work out.
How far does it have to tilt before it's too late to be able to balance itself? Or it's impossible to balance itself, correct itself?
Well, you're asking an interesting question because... In those days, we didn't actually optimize things. And they probably could have gone much further than we did and then had higher performance. And we just kind of got a sketch of a solution and worked on that.
And then in years since, some people working for us, some people working for others, people came up with all kinds of equations or algorithms for how to do a better job, be able to go faster. One of my students worked on getting things to go faster. Another one worked on... climbing over obstacles. Because when you're running, on the open ground, it's one thing.
If you're running up a stair, you have to adjust where you are. Otherwise, things don't work out right. You land your foot on the edge of the step. So there's other degrees of freedom to control if you're getting to more realistic, practical situations.
I think it's really interesting to ask about the early days because believing in yourself, believing that there's something interesting here, and then you mentioned finding somebody else, Ben Brown. What's that like, finding other people with whom you can build this crazy idea and actually make it work?
Probably the smartest thing I ever did is to find the other people. I mean, when I look at it now, I look at Boston Dynamics and all the really excellent engineering there, people who really make stuff work. I'm only the dreamer.
So when you talk about pogo stick robot or legged robots, whether it's quadrupeds or humanoid robots, did people doubt that this is possible?
Did you experience a lot of people around you kind of... I don't know if they doubted whether it was possible, but I think they thought it was a waste of time.
Oh, it's not even an interesting problem.
I think for a lot of people. I think it's been both, though. Some people, I felt like they were saying, oh, why are you wasting your time on this stupid problem? But then I've been at many things where people have told me it's been an inspiration to go out and attack these harder things. And I think it has turned out, I think legged locomotion has turned out to be a useful thing.
Did you ever have doubt about bringing Atlas to life, for example, or with Big Dog, just every step of the way, did you have doubt, like, this is too hard of a problem?
I mean, at first, I wasn't an enthusiast for the humanoids because, again, it goes back to saying, what's the functionality? And the form wasn't as important as the functionality. And also, there's an aspect to humanoid robots that's about functionality. all about the cosmetics, where there isn't really other functionality, and that kind of is off-putting for me.
As a roboticist, I think the functionality really matters. So probably that's why I avoided humanoid robots to start with. I'll tell you, after we started working on him, you could see the connection and the impact with other people, whether they're lay people or even other technical people.
There's a special thing that goes on, even though most of the humanoid robots aren't that much like a person.
But we anthropomorphize and we see the humanity. But also with Spot, you can see not the humanity, but whatever we find compelling about social interactions there in Spot as well.
I'll tell you, I go around giving talks and take Spot to a lot of them. And it's amazing. The media likes to say that they're terrifying and that people are afraid. And And YouTube commenters like to say that it's frightening.
But when you take a spot out there, now maybe it's self-selecting, but you get a crowd of people who want to take pictures, want to pose for selfies, want to operate the robot, want to pet it, want to put clothes on it. It's amazing.
yellow spot so if we uh move around history a little bit so you said i think in the early days of boston dynamics that you quietly worked on making uh a running version of eyeball yeah sony's robot dog yeah it's just an interesting uh little tidbit of history for me um What stands out to you in memory from that task? For people who don't know, that little dog robot moves slowly.
How did that become Big Dog? What was involved there? What was the dance between how do we make this cute little dog versus a thing that can actually carry a lot of payload and move fast and stuff like that?
What the connection was is that at that point, Boston Dynamics was mostly a physics-based simulation company. When I left MIT to start Boston Dynamics, there was a few years of overlap, but the concept wasn't to start a robot company. The concept was to use this dynamic simulation tool that we developed to do robotics for other things.
But working with Sony, we got back into robotics by doing the AIBO Runner. We made some tools for programming Curio, which was a humanoid this big that could do some dancing and other kinds of fun stuff. And I don't think it ever reached the market, even though they did show it. When I look back, I say that we got us back where we belonged. Yeah.
You rediscovered the soul of the company.
That's right.
And so from there, it was always about robots. Yeah. So you started Boston Dynamics in 1992. Right. What are some fond memories from the early days?
One of the robots that we built wasn't actually a robot. It was a surgical simulator, but it had force feedback. So it had all the techniques of robotics. And you look down into this... mirror it actually was. And it looked like you were looking down onto the body you were working on. Your hands were underneath the mirror, so they were where you were looking.
And you had tools in your hands that were connected up to these force feedback devices made by another MIT spin-out, Sensible Technologies. So they made the force feedback device. We attached the tools and we wrote all the software and did all the graphics. So we had 3D computer graphics.
It was in the old days when, this was in the late 90s, when you had a silicon graphics computer that was about this big. It was the heater in the office, basically. And we were doing surgical operations, anastomosis, which was stitching tubes together, tubes like blood vessels or other things in their body. And you could feel and you could see the tissues move. And it was really exciting.
And the idea was to make a trainer to teach surgeons how to do stuff. We built a scoring system because we interviewed surgeons that told us what you're supposed to do and what you're not supposed to do. You're not supposed to tear the tissue. You're not supposed to touch it in any place except for where you're trying to engage. There were a bunch of rules.
So we built this thing and took it to a trade show, a surgical trade show. And the surgeons were practically lined up. Well, we kept the score and we posted their scores like on a video game. And those guys are so competitive that they really, really love doing it. And they would come around and they see someone's score was higher there. So they would come back.
But we figured out shortly after that we thought surgeons were going to pay us to get trained on these things. And the surgeons thought we should pay them in order to, so they could teach us about the thing. And there was no money from the surgeons. And we looked at it and thought, well, maybe we could sell it to hospitals that would teach, train their surgeons.
And then we said, well, we're this, at the time we were probably a 12-person company or maybe 15 people, I don't remember. There's no way we could go after a marketing activity. You know, the company was all bootstrapped in those years. We never had investors until Google bought us, which was after 20 years. So we didn't have any resources to go after hospitals.
So at one day, Rob and I were looking at that and we said, we'd built another simulator for knee arthroscopy. And we said, this isn't going to work. And we killed it. And we moved on, and that was really a milestone in the company because we sort of understood who we were and what would work and what wouldn't, even though technically it was really a fascinating thing.
What was that meeting like? Were you just like sitting at a table? You know what?
Probably.
We're going to pivot completely. We're going to let go of this thing we put so much hard work into and then go back to the thing we came from.
It just always felt right once we did it, you know?
just look at each other and said, let's build robots. What was the first robot you built under the flag of Boston Dynamics? Big Dog?
Well, there was the AIBO Runner, but it wasn't even a whole robot. It was just legs that we, we took off the legs on AIBOs and attached the legs we'd made. We got that working and showed it to the Sony people. We worked pretty closely with Sony in those years. One of the interesting things is that it was before the internet and Zoom and anything like that.
So we had six ISDN lines installed, and we would have a telecon every week that worked at very low frame rates, something like 10 hertz. You know, English across the boundary with Japan was a challenge, trying to understand what each of us was saying and have meetings every week. for several years doing that. And it was a pleasure working with them. They were really supporters.
They seemed to like us and what we were doing. That was the real transition from us being a simulation company into being a robotics company again.
It was a quadruped? The legs were four legs or two legs?
Yeah, no, four legs, yeah.
And what did you learn from that experience of building basically a fast-moving quadruped?
Mostly we learned that something that small doesn't look very exciting when it's running. It's like it's scampering. And you had to watch a slow-mo for it to look like it was interesting. If you watch it fast, it was just like a... That's funny. One of my things was to show stuff in video, even from the very early days of the hopping machines.
And so I was always focused on how is this going to look through the viewfinder. And running AIBO didn't look so cool through the viewfinder.
So, uh, what, what came next in terms of, uh, what was a big next milestone in terms of a robot you built?
I mean, you got to say that big dog was, you know, sort of put us on the map and got our heads really pulled together. We scaled up the company. Big dog was the result of, uh, Alan Rudolph at DARPA, uh, starting a biodynotics program and he put out a, you know, a request for proposals and, uh, I think there were 42 proposals written and three got funded. One was Big Dog.
One was a climbing robot, Rise. That put things in motion. We hired Martin Buehler. He was a professor in Montreal at McGill. He was incredibly important for getting Big Dog started. Out of the lab and into the mud, which is a key step to really be willing to go out there and build it, break it, fix it, which is sort of one of our mottos at the company.
So testing in the real world. Testing. For people who don't know, Big Dog, maybe you can correct me, but it's a big quadruped, four-legged robot robot. It looks big, could probably carry a lot of weight. Not the most weight that Boston Dynamics have built, but a lot.
Well, it's the first thing that worked. So let's see, if we go back to the leg lab, we built a quadruped that could do many of the things that Big Dog did, but it had a hydraulic pump sitting in the room with hoses connected to the robot. Mm-hmm. It had a VAX computer in the next room. It needed its own room because it was this giant thing with air conditioning.
And it had this very complicated bus connected to the robot. And the robot itself just had the actuators. It had gyroscopes for sensing and some other sensors. But all the power and computing was off-board. Big Dog had all that stuff integrated on the platform. It had a gasoline engine for power, which was a very complicated thing to undertake.
It had to convert the rotation of the engine into hydraulic power, which is how we actuated
uh it so there was a lot of learning just on the uh you know building the physical robot and the system integration for that and then there was the controls uh of it so for big dog you brought it all together onto one platform right and then so you could you could take it out in the woods yeah and you did we did we spent a lot of time down at the uh marine corps base in quantico where there was a trail
called the guadalcanal trail and our uh milestone that darpa had specified was that we could go on this one particular trail that involved you know a lot of challenge and we spent a lot of time our team spent a lot of time down there those were fun days hiking with the robot what did you learn about like what it takes to balance a robot like that on a trail
on a hiking trail in the woods. Basically, forget the woods, just the real world. That's the big leap into testing in the real world.
Yeah. As challenging as the woods were, working inside of a home or in an office is really harder. Because when you're in the woods, you can actually take any path up the hill. All you have to do is avoid the obstacles. There's no such thing as damaging the woods, at least to first order. Whereas if you're in a house, you can't leave scuff marks. You can't bang into the walls.
The robots aren't very comfortable bumping into the walls, especially in the early days. So I think those were actually bigger challenges once we faced them. It was mostly getting the systems to work well enough together, the hardware systems to work, and the controls. In those days, we did have a human operator who did all the visual perception going up the Guadalcanal Trail.
There was an operator who was right there, who was very skilled at Even though the robot was balancing itself and placing its own feet, if the operator didn't do the right thing, it wouldn't go.
But years later, we went back with one of the electric, the precursor to Spot, and we had advanced the controls and everything so much that an amateur, complete amateur, could operate the robot the first time up and down and up and down, whereas it had taken us years to get there in the previous robots.
So if you fast forward, Big Dog eventually became Spot.
So Big Dog became LS3, which is the big load carrying one.
Just a quick pause, it can carry 400 pounds.
It was designed to carry 400, but we had it carrying about 1,000 pounds. Of course you did.
Just to make sure.
We had one carrying the other one. We had two of them. So we had one carrying the other one. There's a little clip of that. We should put that out somewhere. That's from like 20 years ago. Wow.
Wow. And it can go for very long distances. It can travel 20 miles. Yeah. Gasoline. Gasoline, yeah. And that adventure, okay, sorry. So LS3, then how did that leave the spot?
So, Big Dog and LS3 had engine power and hydraulic actuation. Then we made a robot that was Electric power, so there's a battery driving a motor, driving a pump, but still hydraulic actuation. Larry sort of asked us, could you make something that weighed 60 pounds that would not be so intimidating if you had it in a house where there were people?
And that was the inspiration behind the spot, pretty much as it exists today. We did a prototype the same size that was the first all-electric thing Non-hydraulic robot.
What was the conversation with Larry Page like about, so here's a guy that kind of is very product focused and can see a vision for like what the future holds. That's just interesting kind of aside. What was the brainstorm about the future of robotics with him like?
I mean, it was almost as simple as what I just said. You know, we were having a meeting. He said, yeah, geez, do you think you could make a smaller one that wouldn't be so intimidating, like a big dog, if it was in your house? And I said, yeah, we could do that. And we started and did.
Is there a lot of technical challenges to go from hydraulic to electric machines?
I had been in love with hydraulics and still love hydraulics. It's a great technology. It's too bad that somehow the world out there looks at it like it's old-fashioned or that it's icky. It's true that you do. It is very hard to keep it from having some amount of dripping from time to time. But if you look at the performance, how strong you can get in a lightweight package.
And of course, we did a huge amount of innovation. Most of hydraulic control, that is the valve that controls the flow of oil, had been designed in the 50s for airplanes. It had been made robust enough, safe enough that you could count on it so that humans could fly in airplanes. And very little innovation had happened. You know, that might not be fair to the people who make the valves.
I'm sure that they did innovate. But the basic design had stayed the same. And there was so much more you could do. And so our engineers designed valves, the ones that are in Atlas, for instance, that had new kinds of circuits. They sort of did some of the computing that could get you much more efficient use. They were much smaller and lighter so that the whole robot could be smaller and lighter.
We made a hydraulic power supply that had a bunch of components integrated in this tiny package. It's about this big, the size of a football. It weighs five kilograms and it produces five kilowatts of power. Of course, it has to have a battery operating, but it's got a motor, a pump, filters, heat exchanger to keep it cool, some valves, all in this tiny little package. Hydraulics
you know, could still have a ways to go.
One of the things that stands out about the robots Boston Dynamics have created is how beautiful the movement is, how natural the walking is and running is, even flipping is, throwing is. So maybe you can talk about what's involved in making it look natural.
Well, I think having good hardware is part of the story, and people who think you don't need to innovate hardware anymore are wrong, in my opinion. So I think one of the things, certainly in the early years for me, taking a dynamic approach where you think about what's the evolution of the motion of the thing going to be,
in the future and having a prediction of that that's used at the time that you're giving signals to it, as opposed to it all being servoing, which is servoing is sort of backward looking. It says, okay, where am I now? I'm going to try and adjust for that. But you really need to think about what's coming.
So how far ahead do you have to look in time?
It's interesting. I think that the number is only a couple of seconds for a spot. So there's a limited horizon type approach where you're recalculating, assuming what's going to happen in the next second or second and a half. And then you keep iterating. At the next, even though a tenth of a second later, you'll say, okay, let's do that again and see what's happening.
And you're looking at what the obstacles are, where the feet are going to be placed, and how to, you know, you have to coordinate a lot of things if you have obstacles and you're balancing at the same time. And it's that limited horizon type calculation that's doing a lot of that. But if you're doing something like a somersault, you're looking out a lot further, right?
If you want to stick the landing, you have to get the, you know, you have to, at the time of launch, have, you know, momentum and rotation, all those things coordinated so that a landing is within reach.
How hard is it to stick a landing? I mean, it's very much under-actuated. Once you're in the air, you don't have as much control about anything. So how hard is it to get that to work? First of all, I did flips with a hopping robot.
If you look at the first time we ever made a robot do a somersault, it was in a planar robot. It had a boom. It was restricted to the surface of a sphere. We call that planar. It could move fore and aft. It could go up and down, and it could rotate. The calculation of what you need to do to stick a landing isn't all that complicated.
You have to get time to make the rotation, so how high you jump gives you time. You look at how quickly you can rotate. And so if you get those two right, then when you land, you have the feet in the right place. And you have to get rid of all that rotational and linear momentum. But that's not too hard to figure out.
And we made, back in about 1985 or 6, I can't remember, we had a simple robot doing somersaults. To do it in 3D, really the calculation is the same. You just have to be balancing in the other degrees of freedom. If you're just doing a somersault, it's just a planar thing.
When Rob was my graduate student and we were at MIT, which is when we made a two-legged robot do a 3D somersault for the first time. There, in order to get enough rotation rate, you needed to do tucking also. You know, withdraw the legs in order to accelerate it. And he did some really fascinating work on how you stabilize more complicated maneuvers.
You remember he was a gymnast, a champion gymnast before he'd come to me. So he had the physical abilities. And he was an engineer, so he could translate some of that into the math and the algorithms that you need to do that.
He knew how humans do it. He just had to get robots to do the same.
Unfortunately, though, humans don't really know how they do it, right? We're coached. We have ways of learning, but do we really understand in a physics way what we're doing? Probably most gymnasts and athletes don't know.
So in some way, by building robots, you are in part understanding how humans do, like walking. Most of us walk without considering how we walk, really. Right. And how we make it so natural and efficient, all those kinds of things.
Atlas still doesn't walk like a person, and it still doesn't walk quite as gracefully as a person, even though it's been getting closer and closer. The running might be close to a human, but the walking is still a challenge.
That's interesting, right, that running is closer to a human. It just shows that the more aggressive and kind of, the more you leap into the unknown, the more natural it is. I mean, walking is kind of falling always, right?
And something weird about the knee, that you can kind of do this folding and unfolding, And get it to work out just, a human can get it to work out just right. There's compliances. Compliance means springiness in the design that are important to how it all works. Well, we used to have a motto at Boston Dynamics in the early days, which was, you have to run before you can walk.
And that's a good motto. Because you also had Wildcat, which was along the way towards Spot, which is a quadruped that went 19 miles an hour on flat terrain. Is that the fastest you've ever built? Oh, yeah. Might be the fastest quadruped in the world. I don't know.
For a quadruped, probably. Of course, it was probably the loudest, too. So we had this little racing go-kart engine on it, and we would get people from three buildings away sending us, complaining about how loud it was.
So at the leg lab, I believe most of the robots didn't have knees. How do you figure out what is the right number of actuators? What are the joints to have? Do you need to have, you know, we humans have knees and all kinds of interesting stuff on the feet. The toe is an important part, I guess, for humans. Or maybe it's not. I injured my toe recently and it made running very unpleasant.
So that seems to be kind of important. So how do you figure out for efficiency, for function, for aesthetics, how many joints to have, how many actuators to have?
Well, it's always a balance between wanting to get where you really want to get and what's practical to do based on your resources or what you know and all that. So, I mean, the whole idea of the pogo stick was to do simplification. Obviously, it didn't look like a human.
I think a technical scientist could appreciate that we were capturing some of the things that are important in human locomotion without it looking complicated. like it without having a knee and ankle. I'll tell you the first sketch that Ben Brown made when we were talking about building this thing was a very complicated thing with zillions of springs, lots of joints.
It looked much more like a kangaroo or an ostrich or something like that, things we were paying a lot of attention to at the time. Um, you know, so my job was to, uh, say, okay, well let's do something simpler to get started and maybe we'll get there at some point.
I just love the idea that you, you, you two were studying kangaroos and ostriches.
Oh yeah. We, we did, uh, we filmed and digitized, uh, data from horses. I did a dissection of a ostrich at one point, which has absolutely remarkable legs.
Dumb question. Do ostriches have a lot of musculature on the legs or no?
Most of it's up in the feathers, but there's a huge amount going on in the feathers, including a knee joint. The knee joints way up there. The thing that's halfway down the leg that looks like a backwards knee is actually the ankle. The thing on the ground, which looks like the foot, is actually the toes. It's an extended toe. Fascinating.
But, you know, the basic morphology is the same in all these animals.
What do you think is the most beautiful movement of an animal? Like, what animal do you think is the coolest? Land animal. Because fish is pretty cool. Like, the fish moves the water, but like, legged locomotion.
You know, the slow-mos of cheetahs running are incredible. You know, they're they, there's so much back motion and, uh, you know, grace. And of course they're moving very fast. Uh, The animals running away from the cheetah are pretty exciting.
You know, the pronghorn, which, you know, they do this all four legs at once jump called the prong to kind of confuse the, especially if there's a group of them, to confuse whoever's chasing them. So they do like a misdirection type of thing? Yep, they do a misdirection thing.
The front on views of the cheetahs running fast where the tail is whipping around to help in the turns, to help us stabilize in the turns. That's pretty exciting.
Because they spend a lot of time in the air, I guess, as they're running that fast.
But they also turn very fast.
Is that a tail thing, or do you have to have contact with the ground?
Everything in the body is probably helping turn, because they're chasing something that's trying to get away that's also zigzagging around. But I would be remiss if I didn't say humans are pretty good, too. You watch gymnasts, especially these days, they're doing just incredible stuff.
Well, especially Olympic-level gymnasts. See, but there could be cheetahs that are Olympic-level. We might be watching the average cheetah versus there could be a really special cheetah that can do like... You're right. When did the knees first come into play in you building Legged Robots?
In Big Dog. Big Dog. Yeah, Big Dog came first, and then Little Dog was later. And you know, there was a... There's a big compromise there. Human knees have multiple muscles, and you could argue that there's, I mean, it's a technical thing about negative work. When you're contracting a joint, but you're pushing out, that's negative work.
And if you don't have a place to store that, it can be very expensive to do negative work. And in Big Dog, there was no place to store negative work in the knees. But Big Dog also had pogo stick springs down below. So part of the action was to comply in a bouncing motion. Later on in Spot, we took that out. As we got further and further away from the leg lab, we had more energy-driven controls.
Is there something to be said about knees that go forward versus backward?
Sure. There's this idea called passive dynamics, which says that although you can use computers and actuators to make a motion, a mechanical system can make a motion just by itself if it gets stimulated the right way. Uh, so, uh, Tad McGeer in the, uh, I think in the mid eighties, uh, maybe it was in the late eighties starting to, started to work on that.
And he made this, uh, legged system that could walk down an inclined plane where the legs folded and unfolded and swung forward, you know, do the whole walk, walking motion where the only thing, there was no computer, uh, There were some adjustments to the mechanics so that there were dampers and springs in some places that helped the mechanical action happen.
It was essentially a mechanical computer. The interesting idea there is That it's not all about the brain dictating to the body what the body should do. The body is a participant in the motion.
So a great design for a robot has a mechanical component where the movement is efficient even without a brain. Yes. How do you design that?
I think that these days most robots aren't doing that. Most robots are... are basically using the computer to govern the motion. Now, the brain, though, is taking into account what the mechanical thing can do and how it's going to behave.
Otherwise, it would have to really forcefully move everything around all the time, which probably some solutions do, but I think you end up with a more efficient and more graceful thing if you're taking into account what the machine wants to do.
So this might be a good place to mention that you're now leading up the Boston Dynamics AI Institute, newly formed, which is focused more on designing the robots of the future. I think one of the things, maybe you can tell me the big vision for what's going on, but one of the things is this idea that hardware still matters with organic design and so on.
Maybe before that, can you zoom out and tell me what the vision is for the AI Institute?
I like to talk about intelligence having two parts, an athletic part and a cognitive part. And I think Boston Dynamics, in my view, has sort of set the standard for what athletic intelligence can be. And it has to do with all the things we've been talking about, the The mechanical design, the real-time control, the energetics, and that kind of stuff.
But obviously, people have another kind of intelligence. And animals have another kind of intelligence. We can make a plan. Our meeting started at 9.30. I looked up on Google Maps how long it took to walk over here. It was 20 minutes. So I decided, okay, I'd leave my house at 9.00. which is what I did. You know, simple intelligence, but we use that kind of stuff all the time.
It's sort of what we think of as going on in our heads. And I think that's in short supply for robots. Most robots are pretty dumb. And as a result, it takes a lot of skilled people to program them to do everything they do. And it takes a long time. And if robots are gonna satisfy our dreams, they need to be smarter.
So the AI Institute is designed to combine that physicality of the athletic side with the cognitive side. So, for instance, we're trying to make robots that can watch a human do a task, understand what it's seeing, and then do the task itself. So sort of OJT, on-the-job training for robots, as a paradigm. Now, you know, that's pretty hard...
It's sort of science fiction, but our idea is to work on a longer timeframe and work on solving those kinds of problems. I have a whole list of things that are in that vein.
Maybe we can just take many of the things you mentioned, just take it as a tangent. Okay. First of all, athletic intelligence is a super cool term. And that really is intelligence. We humans kind of take it for granted that we're so good at walking and moving about the world.
And using our hands, you know. Using your hands. The mechanics of interacting with all these things.
these two things you know you've never touched those things before right well i've touched ones like this okay look at all the things i can do right i can juggle and i'm rotating this way i can rotate it without looking i could fetch these things out of my pocket and figure out which one was which and all that kind of stuff yeah and uh i don't think we have much of a clue how all that works yet right and that's i really like putting that under the banner of athletic uh intelligence
what are the big open problems in athletic intelligence? So Boston Dynamics, with Spot, with Atlas, just have shown time and time again, like, push the limits of what we think is possible with robots. But where do we stand, actually, if we kind of zoom out? What are the big open problems on the athletic intelligence side?
I mean, one question you could ask, it isn't my question, but, you know, are they commercially viable? Will they increase productivity? And I think... You know, we're getting very close to that. I don't think we're quite there still. You know, most of the robotics companies, it's a struggle.
It's really the lack of the cognitive side that probably is the biggest barrier at the moment, even for the physically successful robots. But your question is, I mean, you can always do a thing that's more efficient, lighter, more reliable. I'd say reliability. I know that Spot, they've been working very hard on getting the the tail of the reliability curve up and they've made huge progress.
So the robots, you know, there's 1,500 of them out there now, many of them being used in practical applications day in and day out where, you know, where they have to work reliably. And, you know, it's very exciting that they've done that, but it takes a huge effort to get that kind of reliability in the robot. There's cost, too. You'd like to get the cost down.
Spots are still pretty expensive, and I don't think that they have to be, but it takes a different kind of activity to do that. Now that... I think that Boston Dynamics is owned primarily by Hyundai now, and I think that the skills of Hyundai in making cars can be brought to bear in making robots that are less expensive and more reliable and those kinds of things.
So on the cognitive side, for the Eye Institute, what's the trade-off between moonshot projects for you and maybe incremental progress?
That's a good question. I think we're using the paradigm called stepping stones to moonshots. I don't believe, that was in my original proposal for the Institute, stepping stones to moonshots. I think if you go more than a year without seeing a tangible status report of where you are, which is the stepping stone, It could be a simplification.
You don't necessarily have to solve all the problems of your target goal, even though your target goal is going to take several years. Those stepping stone results give you feedback, give motivation, because usually there's some success in there. That's the mantra we've been working on. That's pretty much how I'd say Boston Dynamics has worked, where you make progress and show it as you go.
Show it to yourself, if not to the world.
What does success look like? What are some of the milestones you're chasing?
Well, with Watch, Understand, Do, the project I mentioned before, we've broken that down into getting some progress with what is meaningfully watching something mean, breaking down an observation of a person doing something into the components, segmenting. You watch me do something. I'm going to pick up this thing and put it down here and stack this on it.
Well, it's not obvious if you just look at the raw data. what the sequence of acts are. It's really a creative, intelligent act for you to break that down into the pieces and understand them in a way so you could say, okay, what skill do I need to accomplish each of those things?
So we're working on the front end of that kind of a problem where we observe and translate the, if it may be video, it may be live, into
a description of what we think is going on and then try and map that into skills to accomplish that and we've been developing skills as well so you know we have kind of multiple stabs at the pieces of of doing that and this is usually video of humans manipulating objects with their hands kind of thing we're starting out with bicycle repair some simple bicycle repair oh no that seems complicated that seems really complicated it is but but there's some parts of it that aren't
Like putting the seat in, you know, into the, you know, you have a tube that goes inside of another tube and there's a latch. You know, that should be within range.
Is it possible to observe, to watch a video like this without having an explicit model of what a bicycle looks like?
I think it is. And I think that's the kind of thing that people don't recognize. Let me translate it to navigation. Mm-hmm. I think the basic paradigm for navigating a space is to get some kind of sensor that tells you where an obstacle is and what's open, build a map, and then go through the space.
But if we were doing on-the-job training where I was giving you a task, I wouldn't have to say anything about the room. We came in here, all we did is adjust the chair, but we didn't say anything about the room, and we could navigate it. So I think there's opportunities to build that kind of navigation skill into robots. And we're hoping to be able to do that.
So operate successfully under a lot of uncertainty.
Yeah, and lack of specification.
Lack of specification.
I mean, that's what sort of intelligence is, right? Kind of dealing with, understanding a situation even though it wasn't explained.
So how big of a role does machine learning play in all of this? Is this more and more learning-based?
You know, Since ChatGBT, which is a year ago, basically, there's a huge interest in that and a huge optimism about it. And I think that there's a lot of things that machine learning, that kind of machine learning. Now, of course, there's lots of different kinds of machine learning. I think there's a lot of interest and optimism about it.
I think the facts on the ground are that doing physical things with physical robots is a little bit different than language. The tokens sort of don't exist. Pixel values aren't like words. But I think that there's a lot that can be done there. We have... We have several people working on machine learning approaches.
I don't know if you know, but we opened an office in Zurich recently, and Marco Hutter, who's one of the real leaders in reinforcement learning for robots, is the director of that office. He's still half-time at ETH now.
the university there where he has an unbelievably fantastic lab and then he's half time leading will be leading off efforts in the Zurich office so we have a healthy learning component but there's part of me that still says if you look out in the world at what the most impressive performances are
They're still pretty much, I hate to use the word traditional, but that's what everybody's calling it, traditional controls, like model predictive control. The Atlas performances that you've seen are mostly model predictive control. They've started to do some learning stuff that's really incredible. I don't know if it's all been shown yet, but you'll see it over time.
And then Marco has done some great stuff and others.
So especially for the athletic intelligence piece, the traditional approach seems to be the one that still performs the best.