Physical AI, Play, and the Path Beyond Humanoids with Brad Porter of Cobot

Transcript

Brad Porter (00:00)

You think about just like washing dishes, right? And sure, we can kind of rote train things, but understanding the texture of a dry sponge versus a wet sponge, understanding how to flip open the lid of a dish detergent, because each one's a little bit different, it is physically problem solving in the real world. I think we need to imbue robotics with that ability to kind of learn things through trial and error. As we unlock that, We won't be playing around in simulation. The robots will just be trying new things in the world like you or I do.

Brian Heater (00:37)

Hey, everyone. Welcome to Automated. My name is Brian Heater. I am the managing editor at a three very excited to bring you this week's conversation with Cobot founder and CEO Brad Porter. Prior to Cobot, Brad spent 13 years at Amazon where he was the VP of robotics and helped lead the rollout of the world's largest fleet of industrial mobile robots.

Brad was also recently a fireside speaker at our humanoid robot forum in Seattle, where he provided a very entertaining counterpoint to the ongoing humanoid hype cycle while discussing the company's proxy robot. It's always a good time speaking with Brad and this podcast certainly isn't an exception. Don't forget to like, subscribe and tell a friend all of that good stuff. Enjoy. So as we were programming our humanoid summit agenda. Somebody suggested that we have a panel or presentation that was kind of a reality check as far as the subject of humanoids. Two people immediately came to mind. The first one was recent podcast guest, Rod Brooks, and the other one is sitting across from me right now. How does that feel that you're like the one or two person that people think of when they think about skepticism around this form factor?

Brad Porter (02:06)

You don't like go out to be a skeptic. Like there was kind of a lot of enthusiasm and excitement about humanoids. felt like it was work putting out some content to explain what I had seen, you know, thinking deeply about humanoids at Amazon, you know, a little bit ahead of the curve. And then I thought about this a lot in 2018. I think agility had showed some really impressive bipedalism. That technology looked like it was starting to be viable. The humanoid grand challenge had just taken place.

And so we thought about it a lot. And then we came to the other side of that exercise realizing that for the use cases where a humanoid would be helpful at Amazon, it wasn't actually the best form factor. And so that led me to realize there was really what we needed in commercial environments was quite a bit different. And then in the home space, I never quite figured out exactly how a humanoid helps us either.

? and so then, you know, I put out content on this and then, you know, some, I'm, I'm like branded the kind of skeptic cause they're, know, on any, on any interesting story, people are always looking for, for two sides and to create this kind of polarity. And, and so, yes, I guess I've ended up anchoring the, ? the maybe more pragmatic, ? what, what are commercial environments really need? ? and, and pour in a bit of a dose of cold water on the humanoid hype.

Brian Heater (03:31)

When you're saying that for any given thing that the robot would be doing, ? like in a warehouse specifically, they're maybe not the best suited for that. Are you talking specifically about a robot with two legs when you talk about humanoids?

Brad Porter (03:46)

As we looked at humanoids and as I continue to look at robotics, there were a few things that stood out as really critical in commercial environments. One is you want to be safety rated and safety real. Because most of these process paths that would be, that aren't going to be automated in some other way, benefit from being able to work in and around humans. And if you're trying to work in and around humans, you don't want any risk of falling on a human, injuring a human. And when you have a balancing system, that's very tough to guarantee, right? Because any glitch between sensing and electrical and actuation can cause an instability that causes that robot to fall, right? And so that was the first thing was, and then you ask yourself, do you really need bipedalism?

And then, you know, in commercial environments, people are moving things around with the assistive of wheels of carts. So you're not carrying things up and down stairs that, you know, that might exist in apartment buildings in New York City, trying to move in and out, but that doesn't exist in a commercial warehouse, right? Nobody's really carrying things up and down stairs. There's not a real need for bipedal. Yes, there are ramps and lips and things like that, but you can get over those with wheels, with maybe some degree of suspension -- active or passive suspension. you don't need bipedalism. You don't need quadrupeds either, right? Like good wheels, the right size wheels, the right size suspension will help you in whatever you need to do. You really want swappable batteries? Nobody in humanoid is really solving. Like if you have to sit on a charger, lose, you know, look, we as humans, we sleep seven, eight hours a day. We lose a third of our productivity every day to sleep.

Right? Why would you do that in an industrial environment if you could avoid

Brian Heater (05:39)

productivity or you have to buy it like three times as many robots.

Brad Porter (05:43)

And you need a whole bunch of space for the robots to park as they charge. I'm even more bullish that like sitting on a charger is, or maybe more opinionated that sitting on a charger is a completely horrible use of a robot. Your whole goal of a robot is to be operating in production continuously and taking on meaningful work. so sitting on a charger doesn't make any sense. ?

You really want your sensors not to move around and not to move up and down. So the problem is if you're a balancing system, your sensing is always moving around. And so you have to then compensate with your AI algorithms for the fact that your sensors are never stable. And so your calibration is a lot harder. They're just a whole bunch of reading the cost, the amount of... And then I think the biggest problem is they're not strong enough, right? You can't get the torque density in one of these like...to be able to what humans can lift. And look, we use linear actuation in our muscles and look, we're very, very finely tuned for it. You know, our robot, Proxy, which is a, you know, is more of the form factor that obviously espoused four wheels on the ground, the latest generation can lift 200 pounds, right? Now it does that because it has a screwdriver in the spine, right? Which is a much better way to get torque, right?

Now, if you or I had a screwdriver in our spine, like we could lift a lot too, right? And so you actually want robots that are more capable than humans, right? You want to move 1500 pound carts around that helps to have like four electric wheelchair motors in the base, right? That can move 1500 pounds without any trouble. You or I moving 1500 pounds, we can do it, but we're using a lot of shoulder, a lot of quads to try to get that moving.

? And humanoids just don't quite have that strength. And so where I became kind of disillusioned was you didn't need bipedalism, you needed safety rating, you wanted battery swap, you wanted something really, really strong, you wanted the sensing to be stable, and you were just never gonna get that if you constrained yourself to the human film factor.

Brian Heater (07:55)

Yeah, I'm glad that people are really having a safety conversation right now because I think maybe something a lot of people, or at least a lot of people outside of the industry don't realize is that I think to this day, the vast majority of the AMRs at Amazon are still in a caged environment.

Brad Porter (08:13)

Proteus is not an occasion. It's a beautiful piece. Yeah, it's the green autonomous one with the kind of blinky eyes that chirps at you. That program actually predated me a little bit when I was in and then was something we invested in heavily at Amazon and they got into production after I had left. a super impressive robot. I love that robot.

Brian Heater (08:16)

That's the autonomous version.

Brad Porter (08:41)

It has to perceive the world from the ground, right? And trying to perceive, mean, imagine you or I trying to understand what's going on in a warehouse around us if we had to lay on the floor and look up from six inches. Like, it's very hard. So.

Brian Heater (08:55)

I'm guessing they would build warehouses very differently if we were that close to the ground.

Brad Porter (09:00)

Yeah, I think so. ?

Brian Heater (09:02)

Interesting. To hear you say that at that early stage at Amazon that you were having this conversation around humanoids. Like obviously right now as things are bubbling up, there's all these reports coming out that Amazon is more actively looking at humanoids. We know that they had to pilot with Agility a couple of years ago. What did those, what shape did those conversations take that far back when discussing humanoids?

Brad Porter (09:25)

Look, Amazon, when I was there and continues, ? is in a position and has a most imperative to understand all of the latest technology and how it can be used to optimize Amazon's operations. From a shareholder standpoint, that's what shareholders would expect. From a customer standpoint, we continue to want our stuff faster, lower cost, right? so, it's like,

Brian Heater (09:56)

It's like a balancing act, right? It's like being as forward thinking as possible, but not actually implementing any of that technology until it's ready to work full time.

Brad Porter (10:05)

Well, I think that that's always the question, right? Always the question is, is this ready now? Is it something we can help accelerate? ? Is it something we can help accelerate by we? I'm speaking as if I'm still at Amazon. left Amazon for recording, but you know, the question was, could we help accelerate it? Should we develop some of that cable? So for drones, for example, Amazon decided that the best way to accelerate that was to invest in building them themselves, right? In the humanoid space, Amazon's made big investments, right? We didn't have a humanoid in-house investment, whether they do now or not is there's speculation, but there's no confirmation that I know. ? And so there they decided to invest in the partner ecosystem. look at Amazon, like our view was we just needed more robotic capability.

Right? The more capability out there, the better. And then we as Amazon could choose which technologies, partners we felt had the most potential and could go deeply. And some of them were more speculative and some of them were less. And so we really liked the agility team. And so I'm not surprised that that relationship continued after I left. And I think it expanded after I left. And as far as I know, it's continuing. ?

But we also, as I said, knew it was speculative. I think from my perspective, I started to become disillusioned that it was the right solution for a lot of these use cases that Amazon had, but that hospitals have, that manufacturing has. These use cases that are really, know, human scale material movement, moving boxes, totes, and carts around the world, I thought there was a much better form factor.

Brian Heater (11:57)

I don't know if you're having these conversations right now, but at the points in Kobot's progress when you were having conversations with VCs, especially the last couple of years, has there been push for you to make Proxy more humanoid because that's the hot thing that everybody's seeing and shoving money into right now?

Brad Porter (12:21)

Does the presence of like multiple solutions, it does two things, right? One, it both drives interest in the space, which is exciting, right? I mean, we raised our series B about the same time figure raised its big round, right? And that was a little bit just because there were a lot of investor attention in that space.

Brian Heater (12:40)

rising tide, yeah.

Brad Porter (12:42)

Investors can only have so much attention and bandwidth, right? And so they can only be focusing on so many segments at once. And so that created a little bit of a crucible where everyone was focusing on robotics for a bit. Now, the interesting thing is what some of the top investors in the world will tell you, I'm not going to name them because maybe I'm putting words in their mouth, but what's so important that is that you, to make a really successful investment, you need to invest in something that's contrarian.

And right. Right? So think Airbnb, think Uber, right? These things were contrarian. Not everyone believed in them. Not everyone bought into the idea that we were going to sleep on each other's couches, right? And yet they were right. And so that's a little bit the argument behind behind cobot and praxis. Like, OK, Brad's a little bit the contrarian here, but is he right? Now, the problem is you have multiple perspectives out there, you have these things that are really exciting. And so that also creates like some cognitive dissonance, right? And some level of like confused minds don't always don't always invest either. Right. And so, you know, I think the the landscape is still shaping out. What I would say is in the past year, it certainly feels like. You know, what we've seen is a proliferation of humanoid companies. So the idea that what some of the folks maybe like figure did where they put together a humanoid very, quickly seemed like shockingly impressive. Now we see like teams of five in a garage coming up with a humanoid. It's not, it was an impressive technical feat. It's impressive when the teams of five do it too. ? But you start to realize, wait, it wasn't maybe that hard. The other thing that happens when investors get interested in the space and then they're they're maybe confused about it for a bit is they study it really, really hard, right? And they kind of learn the whole space and they consume everything. They read everything they possibly can. They talk to everyone they possibly can. And then they get a much more refined thesis as to what might work and what might not. And I think you, you see now that there's a little more caution in thinking about humanoid investments now. And, but I think also, you know, a little bit of a dose back to reality of like, how hard robotics is, right? And I think, you know, investors have always been really excited about robotics, but robotics always goes through these kind of like ups and downs where people realize, oof, it's really, really hard, right? And I think it takes a fair bit of capital intensity and the deployment cycles are a little bit longer. And so I think we're in a very good position, ? but it is tough in robotics right now for a lot of folks.

Brian Heater (15:34)

Is there a sense in which proxy is the more cautious bet versus a lot of these legged robots?

Brad Porter (15:42)

Well, I think in some ways it's the more ambitious bet, right? Like it's easier to kind of follow the trend line and say, you know what? I can build that, but I can build it with a better shoulder actuator or something like that, right? I think what we've done in trying to bring in ? a conceptual model that people haven't seen before, there's a kind of product adage that you should

If you're bringing out something new, make it look a lot like something that already exists, right? Or if you're bringing out something that already exists, make it look new. What we've done with proxy is kind of interesting in that we didn't make it look like anything that exists. With a humanoid, you make it look like something we know, humans, and you can kind of project onto it what its capabilities are.

Brian Heater (16:30)

You mean more that it doesn't look like something that exists in nature versus it doesn't look like existing robots?

Brad Porter (16:36)

Maybe the folks at Mayo Clinic will disagree with me on this, but I think when we first showed them Proxy, they were interested in our consulting services, our flywheel services, to help them think about how to bring robotics into the Mayo ecosystem. But when we first showed them Proxy, I looked around the room and I didn't feel like that many people fully got it. It didn't look like anything that they had seen before. It didn't look like a little two-wheel diff drive. It didn't look like, a lot of like Locust, Fetch, those robots look similar. lot of AMRs are low to the ground, like Amazon's robots.

Brian Heater (17:12)

Let me interject right there. So when speaking of proxy specifically, like this is going to sound silly, but what's to get? Like what weren't they getting?

Brad Porter (17:22)

The use of the swerve drive, the use of four independently steerable wheels, the fact that this robot can do extremely precise positioning, like very, very precise positioning, means that it can like come up to the edge of a table or go through a narrow passageway very, very precisely, right? Or if a human's walking alongside, it can just like slide three inches to the right, like you or I would. And so it makes the robot far more trustworthy in these environments. If you have a two-wheel diff drive,

? In order for you to like move sideways slightly, you've got to make this like spline like curve But it makes it look Unnatural makes it look kind of drunken. It makes it look like it's wobbling. That's to correct itself It's kind of swerving back and forth in fact, you know I've walked into these places that have you know fetch and locus and they the operators describe it as like yeah That one's drunk today, right and it's drunk because like it's calibration slightly off and it's just like constantly repositioning so it's just moving a little bit and better.

Brian Heater (18:22)

They're better the robot drunk than the forklift driver.

Brad Porter (18:24)

That's right, but frankly, no one trusts things that look drunk ? in an operational environment. And so they kind of get used to it, but they don't fully trust it. And Proxy doesn't have that problem. Even if there's like an inch or two of calibration drift, it just kind of moves slightly to side like you or I would, but stays on path, right? So that's one thing they didn't fully get. The fact that it can just move existing carts.

It's funny because, you know, all these other ones that were like in conversations about, you attach to some of the cart? Do we need to modify the cart? Do we need to bring your cart in? the robot itself is a cart, but what if the cart, what if the robot's not used much? Cause the carts don't move all the time. Like just that like realization that if you have a bunch of carts, like 90 % of them are probably idle, right? And so if you're putting all the sensing and autonomy on, on an idle cart, that's a huge amount of costs. Where if you can move the existing carts. So again, these little things of like, Hey, the positioning is really precise. The sensing is really, really stable and secure because we see the world from six feet in the air. We can move your existing cart so it's okay that 90 % of your carts might not be moving at any one time. We're not adding to the cost of that. And people are like, yeah, swappable battery sounds great. Until you've actually tried to run robotics at scale and you realize how much of a pain in the ass charging and charging stands are, right? And so that's kind of...benefit we had from Amazon was we knew precise positioning mattered. We knew where you saw the world from mattered. We knew that carts were often idle and you need to be able to drop them off and leave them somewhere. We knew that sitting on a charger was terrible. Like you think about sitting on a charger, one of the things people don't always realize is that, you know, circuit breakers are 15 amps usually, right? Well, you know, your charger, you have four or five robots charging and you're exceeding the 15 amps.

Right. And so where are you, you now have to like upgrade the electrical in the system in order to get enough amperage going to that part of the building to charge 20 or 30 robots. It's and, so like, once we went into Mayo and put the robots in their specimen processing lab and people saw in the week, these weren't the first robots they had tried in there. People saw how well it worked. Then the light bulbs started to go off. Then it was like, this just works so much better, right? And sometimes you just have to see it to realize that. But to see, this works so much better. And now, like, I think pretty much every week, another department at Mayo is saying, hey, can we think about getting these robots for this use case? So the light bulb has kind of gone off, like, across the whole campus. And now it's a little bit on us to, how quickly can we get robots ? going for them in all of these process paths? But the point is, when you're bringing something new to market that people haven't seen before, ? it's not always immediately obvious why it's so much better. I think about the first iPhone didn't look quite like a phone, right? It looked very, very different. And it took a little while for sales to really ramp for people to realize why that experience was so much better. ? But once they do, then the adoption just, and that's what we're excited about with Cobot.

The worry you have with robots is when you put them in the field, people start to say, okay, well, this doesn't work as well here as I hoped. And look, we had things we were excited about at Amazon that didn't work as well as we hoped. And that's not great for the startup, right? Because they put a lot of time and capital into investing in that technology. And if it doesn't work as well as you'd hoped, then it can be kind of a dead end for a hardware company. So we're really excited that like Proxy has been working--frankly better than we and the customers hoped at taking on these tasks. And actually the folks in the lab at Mayo started to work with us on some ways to use it that we hadn't expected. And that's always great. When you start to realize you do have a platform, right? You do have something that can be used in lots of different ways. that's what we discovered and learned with Amazon's Hercules kind of KivaRoe. We could turn them into drive sorters. We could modify that platform in really interesting ways. And so I think you're going to see that with proxy too. You're going to see that the manipulators we use, some of the ways that we interact with boxes, totes, carts, items, trays is going to evolve in really interesting ways. And we're to be able to do very, very cool things.

Brian Heater (23:02)

It's interesting because it seems like there's two, there's almost two opposing forces here. It's obvious to me why, whether a VC or, you know, even especially this example of the Mayo Clinic, somebody that really has to be pragmatic and practical about deploying these things because, you you can't shut the Mayo Clinic down to have new robots.

There also seems to be this counterintuitive thing that's happening there as well, which is in a certain sense, it might be more difficult to sell small changes in the sense that like, they want this like really radical thing that doesn't look like anything that exists. But obviously, you know, there are a lot of ways in which proxy is more like an AMR than a legged humanoid. There's a difficulty selling like ? the big radical change, but you know, if it's

Brad Porter (23:51)

Sure, okay.

Brian Heater (23:56)

trying to convince them that no, actually, like a wheeled robot is the right solution. I'm suspecting could be difficult, especially around kind of the, you know, all of the conversations being had around humanoids right now.

Brad Porter (24:07)

For better or worse, the humanoids aren't working in these production environments. it's not, know, as much as this, you know, created some investor excitement, there was, there has been some customer excitement about humanoids as well, but that customer excitement dissipates very quickly when people realize the technology is not available and not ready and doesn't work. And I mean, again, this is why I get branded a skeptic, because I sound harsh when I say that.

my God, saying, know, Tesla said they're going to, you know, deploy 10,000 next year. Okay. Well, great. When that happens, that'll happen. But like, that's not what anyone on the inside is saying is going to happen. Right. And the deployments that we do have are fenced off, right. They're not collaborative. So they're not working in spaces like, like hospitals and you know, the economics are still prohibitively expensive. So, you know, it's. Maybe Tesla has deep enough pockets to go build a thousand of, you know, these quarter million plus dollar robots, but even Tesla seems to be maybe worried about cash right now. Cause yeah, as I said, this is why I ended up getting branded as a skeptic. I think for us, the good news is like our robot works. It's, running, know, see nowadays shifts at Marist gets running every day at Mayo. Like it's doing the work it's doing, you know, pretty much 100 % of the eligible work every day. so, you know, so maybe there's a little bit of a, a little bit of a like trying to help people understand how the robot can help them. But then when we do, then we can deliver the robots and put them out and we're, you know, we're going to launch a couple of new customers this, this year, you know, talking to, to our existing customers about, you know, much bigger expansions by the end of the year and next year. And so I both think we have a better solution and I think, you know, we're well positioned to start to deploy now and that's just gonna give us a head start. I did talk to an innovation leader, big manufacturing company that, you know, seriously tested two humanoids in the past year and they said their conclusion was the tech is at least five years away. ? And that was their assessment.

And I had that same, you we were looking in 2018 and I was like, this feels like it's eight to 10 years away. And I was probably optimistic.

Brian Heater (26:43)

obviously, it's not just skepticism, but it's also setting realistic expectations. When you're talking about Elon Musk, know, realistic expectations maybe aren't the first thing that comes to mind. He's, you know, he's had a history of over promising some some things in the past, but there's also a sense in which the kind of over-promising is built into at least the way that robots are being sold. You and I, a couple of years ago when I was at TechCrunch, we collaborated on a story about fudging robotic demos. And I think that humans, we've seen a similar trend play out in LLMs where humans are sort of predisposed to think that these systems are smarter and more capable than they actually are.

Brad Porter (27:28)

I think there's incredible amount of pressure to oversell your capability, right? Because...

You know, one of the rules that I've learned with operators, right, who are running warehouse logistics manufacturing is if a technology is more than a year away, then come back to me in a year. It makes perfect sense from an operator standpoint. If it's more than a year away, I've got things I got to solve right now. I've got plenty of problems that are, you know, on the urgent side. And if it's more than a year away, I'll just park this until next year. You kind of have to bring them solutions that can get in the field within a year, right? Otherwise they're not that interested, but you need to show some interest to get investors interested and investors don't want to hear that it's more than a year away either. So there's this constant pressure to basically promise us things right around the corner. Now the game people play is to then, know, look, Elon's been promising self-driving cars or Teslas are right around the corner every year. And it's because that is what people want to hear. And I think what he's learned ? is that the penalty for under delivering is maybe not as high as the reward for over promising. To Elon's credit, if you can over promise and over deliver just a few times, right? Then you become credible. And I think Elon has done that enough, right? mean, catching a rocket on chopsticks is freaking insane, right? It is absolutely insane.

No one would have said that was possible in the timeline he did it. Again, he's promised a lot of things that haven't come through, but he's promised some things where they've over-delivered. And as a result, he's credible, right? I mean, there's a low Earth orbit network of satellites that gives you high speed, relatively low latency internet everywhere right now. Like, it's insane. It's great. And he did that reasonably fast. Again, Elon is a...isn't 100 % and over delivered, but he's over delivered enough that he's credible. So I think that's the tricky thing in robotics is you've got a bit over promise and you've got to over deliver. I remember a funny story with Maris, because I knew I had to promise within a year and Rez Aglomeli, who's an amazing leader in technology innovation, he moved over to the venture capital world, but he was excited about our technology very early and was willing to sign on very early and had it. He understood where it was going to be useful right away. And I didn't have a robot. I had sketches. I'm not even sure when I pitched him, I had a mechanical engineer yet. He came to visit like nine months later and I still didn't have a robot. He'd worked with a lot of startups. understood founders over promising, but I think he gave me a little bit more credibility and respect because I came out of Amazon and Big Tech and you know, maybe I'm not.

Brian Heater (30:23)

You

Brad Porter (30:33)

like these founders who constantly over-promise things and under-deliver it. I've since talked to him about it. I said, when you came that June, you thought I was completely bullshitting you, didn't you? He's like, yeah, no, I was not happy. ? And what we had decided to do was we realized that we couldn't incrementally build our way to the product. That we had to go into the design tools and just design the whole robot and build the first ones up from scratch. It was a huge bet.

Right? was a huge, it was kind of like the Nvidia bet where they just decided to do everything in simulator because they didn't have enough money to do it. Otherwise, I knew we couldn't meet the Marist timelines if we tried to progressively build. we'd stop building and we started just designing everything and then ordering parts. And sure enough, like a year later, maybe a year and a month, we might've been a little, little short. We built the first full one full up. ? I showed a rez, the robot moving around and doing stuff. then within a couple more months, we had it in his facility and he was shocked. again, this is the like, you earn some credibility because we did deliver. I knew I had to promise that we were going to get it out there and I knew I had to find a way to deliver it. And if you can do that and you can over deliver, then I think you earn respect, you earn credibility and you earn that with investors as well. And so that's our approach.

Brian Heater (31:56)

Yeah, I think there's another element of this too that maybe doesn't get discussed enough. think Boston Dynamics played a role in this early on. Obviously, lot of their successes, they were an early adopter of YouTube and that's what got their name out there. But now we're at this point.

And I don't think cobot is in quite the same position, but all humanoid robot makers are where they feel like they have to sell to non-roboticist normal people watching YouTube as much as they do to customers or investors.

Brad Porter (32:28)

We have an amazing marketing advisor. He used to be the chief market officer for T-Mobile. And he, he's observed that with new technology, it does sometimes pay to play almost a more consumer style marketing campaign because you're trying to create awareness. You're trying to create momentum. You're trying to, you know, appeal to a more general audience. It is true that I think you're trying to humanoid comes through and in fairness like we might not be doing it quite to the same degree but you look at the industrial design of proxy you look at the videos we put out they're closer to kind of emotional content like star wars and those type of robots that are more mass appeal than they are to like the industrial amr like our robot would look odd in the amr catalog of you know industrial magazine that comes out but we do think that robots are going to go mainstream we're going to see them working in airports and hospitals in these places and customer adoption, mass market adoption, consumer appeal are going to matter. So yes, I would say we're all adopting a bit of a more consumer playbook. Well, at the same time, if you look at, you know, the conferences we're going to and the events and where our sales teams are, they're at these industrial forums, but all of us are trying to appeal to the public kind of enthusiasm for robotics.

Brian Heater (33:58)

Yeah, which I think is a perfectly fine thing if that means that these videos are going to be better lit or there's higher production values. But I also worry, and I think you share this worry, that also incentivizes faking in demos.

Brad Porter (34:14)

I think people have gotten better, right? As the public has gotten more sophisticated about consuming the videos, some of the games people play go away. This idea that we project human capabilities onto things with the human form factor. If you think robots from movies, Transformers and Terminator, all of them are stronger than humans. So we naturally think that a robot that looks like a human is stronger than a human. We don't realize that it's like 60 % the strength of a human.

Actually, it complicated things a little bit that Atlas, Boston Dynamics' first humanoid, is pneumatically driven. And pneumatic actuation is almost 10x the strength of electric motors. Atlas was a robot. Atlas could almost do a backflip off its ankle motors. So Atlas also helps set this impression that humanoids are incredibly strong. And in practice, they're not. But you can't get that from a video, really.

Brian Heater (35:11)

Sure.

Brad Porter (35:12)

But if you pay attention, if you know this and you go look at the video, like you notice that all the totes are empty, right? That the robots are moving around or that, you know, that part that it moved from the automotive station to one other piece appears to be either carbon fiber or aluminum, not heavy steel, right? And so the consumers of these videos have to become smarter in order not to be duped by that. But it's almost impossible not to want to project keep the capability onto these things.

Brian Heater (35:44)

So there's that element of things. There's the expectations when it comes to what the hardware can do, like as far as what the robot can lift. But, and I'm going to give you a perfect opportunity right here to plug your newsletter, ? people similarly project these ideas onto what AI is capable of doing and how close we actually are to something that any reasonable person would define as AGI.

Brad Porter (36:10)

The advances in artificial intelligence have been really, really quite incredible, right? The fact that if you'd asked me five years ago whether we would have a robot that could clear the dishes from a restaurant table or could understand semantically the concept of clearing the dishes from the table first, I would have thought that we would have been able to clear the table before we semantically understood what it meant to clear a table. Now, LM has clearly semantically understood what it means to to clear a table and can provide a high level robotic plan to lower level actuation to go do that. We still don't have all the actuation to pick up a plate and not have the fork fall off. But we do have sophisticated AI that understands these semantic concepts and can transform them into action plans. So the field of AI is evolving very, very quickly. think what we know about AI is it's going to keep improving. It's going to improve a bit in step functions.

? It tends to improve as we get the right techniques against a very large data set and we kind of collapse it down into a model. I think the challenge we have in physical AI is we don't have the data sets. And I also think that the way that we interact with things, right? I give this example of like, how do you actually open this thing, right? It's not obvious how it opens. ? And then when you pull it out, if you just grip it, it doesn't come out.

You learn by just playing that you put your thumb back here and then you transfer it, right? And we don't even realize we did that, right? It's not even a conscious thing. We just kind of play with our hands and our hands are just very, they have their own level of intelligence to them. And I think it's a slightly different type of intelligence. I've been calling it kind of the equivalent of large language models where we're doing structured reasoning. I think we play with our hands, the way we reason.

We don't call it reason. We call it play. play with that loose doorknob or we play with that stuck valve or whatever, right? And ? we call it play actually because it's rewarding. Like trying to figure something out with your hands is kind of fun. There's something to reinforcement learning. I think there's something to the self reward, the kind of knowledge of what it feels like to get that valve unstuck that is going to be the key to having dexterous manipulation that can play in the world. so that's what we're working on on the AI front. I have written a piece on this recently on kind of physical AI and why we talk about stochastic parrots, the idea that LLMs are just parroting back things they've heard before. I think of it as stochastic puppets in robotics, the idea that these things are just being reanimated with the animations that they were instructed before. I don't think.

I think stochastic parenting has actually worked incredibly well. ? I am not as convinced that stochastic puppets are going to be how we get to dexterous manipulation at scale. And then, you know, thank you for mentioning the newsletter. I have just put out a newsletter. I've been doing these Sunday videos just to share my thoughts on the space. They've become incredibly popular. I get a lot of comments on them. I get a lot of people who...feel like they understand me and how I'm thinking about better. did it a little bit, not to just be perceived as like curmudgeonly skeptic. Like there's a lot more nuance and depth to how I think about this stuff. so trying to put longer form comment content out there. ? but now, I'm trying to fan it out in a, in a newsletter, like almost 4,500 subscriptions in the first week. So that that's awesome. And then we just took the weekly content and put it out as a podcast this week as well. And so we'll keep doing that. So I'm trying to just get the content out there in part because there's no value in being a skeptic in such an exciting new space. There is a value to nuance and pragmatism and reality of what really works. And so in as much as I was getting soundbited down into a skeptic, I'm trying to get out there with a little longer form and kind of give some nuance.

Brian Heater (40:21)

I'm glad that you brought the AirPods out for the playing example, because as I was reading the newsletter on stochastic puppetry, and you threw out the word playing, my mind's again, I'm creative writing major over here, ? UC Santa Cruz graduate over here, ? my mind automatically goes to like a robot solving a Rubik's Cube or playing soccer, or these sorts of pastimes that we would ? generally define as playing. Obviously, like those do have value. There's a reason why roboticists and researchers are operating that space. But it sounds like you've got a maybe squishier or at least like broader definition of what constitutes playing.

Brad Porter (41:04)

do think of puzzles here. I'll bring out another one that's exciting. are apparently blacksmiths used to make these as like, they're called tavern puzzles. And so how do you get the ring off? And it's really, really hard and complicated. It took me like, it took me about an hour every night for four nights to like figure out how to get the ring off of this. And you start to like learn to reason in 3D and you develop a new way of.

Anyhow, that's a very sophisticated version. This to me is like almost the Turing test of manipulation, right? Like if you can figure out how to solve a tavern puzzle with a robot, that'll be incredibly impressive. What I think though about like soccer, soccer is a great example, like learning to dribble a ball, bounce it from your knees to your feet. There is a rewarding thing like when you succeed in dribbling the ball 20 times, right? Like without dropping it. But there's this repetitive thing. What we do is we coach, we demonstrate, right? So teach a kid to swing a baseball bat. You're going to demonstrate it a few times. Then they're going to try it and they are not going to do it very well necessarily the first couple of times. You're going to give them a little bit of coaching because you can't do it for them, right? You can't tell it. You can't. It's very hard to puppet a kid, right? We demonstrate for them and then we coach them a bit. Hey, keep your elbow up. Keep the bat in tighter.

you know, pull your elbow forward, like we give them some, some verbal instruction and then we let them practice. I think there are two pieces where we're not sophisticated today. I think the coaching piece, I think we can do a lot more in the coaching space in how do you give small amount of guidance and correction instruction to a robot without having to re demonstrate it. But then I think the play part is the place that we have to make self play interesting and rewarding.

And I hesitate to share more there, because I think we're on to some things that we're quite excited about. And so I don't really want to give too many hints about how we think that works. But there's something to how we play that we need to imbue into robots.

Brian Heater (43:10)

So feel free to dodge or completely disregard this question. as I was reading this stochastic puppetry piece, it seemed as though you were saying that plague graduates into practice. But as we're talking right now, it seems like you're saying that practice is part of play. So are these actually two distinct ideas?

Brad Porter (43:34)

I think there is a continuum there between practice and play. Like to some extent practice, and we use those words semantically differently, practice means kind of repetitiveness of like learning a basic motion over and over again. And I think of play as like, there's a bit of a discovery element to it, right? There's a bit of a figuring things out. ? And so, but when you're practicing for the first time, when you're early on, you're also discovering, right? And so, I think it's that discovery piece that's the exciting piece is the rewarding piece. ? And so I think there's a link between practice and play. ? But I think of like playing with a doorknob that's a little stuck, right? Like we're not gonna, you can practice opening a doorknob over and over again, but you have to be able to come up to a new doorknob that's a little bit stuck and wiggle it and figure it out and like, then figure out, I've got to push the handle in slightly and then twist it and then it opens. You don't want to have to use the whole demonstrate, coach, practice, framework. You want to get to the point where there's a, there's an understanding of what it takes to open the door, what, what opening the door feels like, what it looks like, what it feels like, what that experience is. And then the you want the creativity and flexibility in your dexterity in order to solve that. We run into these little puzzles every day. The refrigerator didn't open the way we expected it to. We pulled something out of the refrigerator and something else started to fall. And we very, quickly solve these little physical puzzles all the time without that same level of cognitive self-awareness. We're not like, talking to ourselves in our head. There's no large language model like in our head, maybe for some people there are. For me, I'm not talking to myself as I try to pull something out of a refrigerator. I'm not narrating that activity. Where when I'm trying to solve a work-related problem or a customer problem, I'm often talking to myself in my head. So physical, I think, is just a different layer, a different way. And I think there's something really intrinsic to this idea of... the kind of fun element of learning to do something new with your hands.

Brian Heater (46:01)

I know that Cobot, like everybody else in the free world has a partnership with Nvidia and is using some of that technology, but something that really again jumped out at me in this piece was, you said something along the lines of, simulation is going to be helpful, but the vast majority of the training is actually going to happen in person in the physical world. And that surprised me a bit.

Brad Porter (46:27)

You run into so many different scenarios that are novel, right? I mean, you think about just like washing dishes, right? And sure, we can kind of rote train things, but understanding the texture of a dry sponge versus a wet sponge, understanding how to flip open the lid of ? a dish detergent because each one's a little bit different. to some extent, we spend all this time gathering a lot of the world's knowledge and then we can, you know, push it back out as speech or text, right? ? In that form, think we're a lot like, I actually do think we're a lot like stochastic parrots, right? But when I think about what you're doing every day, it's not... It's not directly leveraging some past knowledge. is physically problem solving in this space of, and people say, hey, we'll have this 3D world and we'll create these, you know, 3D world models. And then we'll be able to reason about, but that's not how we work either, right? Yes, we roughly project the world into 3D, but you know, you can put us underwater.

It's very different physics and dynamics and we can still figure out how to deal with a, you know, take up a neutral buoyancy ball underwater and like playing with that is very different than playing with a ball above the air. Right. And you know, when the first time you play with that underwater, you're not, you don't have any prior knowledge of how to do it. And so you just play with it and you just figure out what its dynamics are. And then you figure out that, you know, this is kind of cool. And like, a little torpedo rocket in water will go quite a bit further with just a little bit of, and like, you just discover these things. I think we need to imbue robotics with that ability to kind of learn things through trial and error in the real world. And I think as we unlock that, we will be playing around in simulation nearly as much. The robots will just be trying new things in the world like you or I do.

Brian Heater (48:43)

Brad, always a pleasure. Thank you so much for joining us.

Brad Porter (48:46)

Yeah, thank you, Brian. This was great.

Brian Heater (48:49)

Thanks so much to Brad and everyone at Cobot who helped set up that conversation. Check out Brad's top of mind newsletter and video podcast over on LinkedIn. And you can find out more about what Cobot is up to over at their site. It's co.bot. As easy as possible to find that. Don't forget to like and subscribe to the automated newsletter.

The podcast board game to sequel to the motion picture. You can find all of that at automated.fm also fairly easy. It's not code out bot, but you know, it's pretty straightforward and we will catch you next week with another episode of automated.