speaker-0: Do you need a robot that folds your laundry? Then maybe Sunday Robotics is something you want to know about. We will talk about it this week on the Robotics Stack Podcast, as well as Underreal and Archer's collaboration on a new aircraft. It looks absolutely phenomenal. Then we go into physical intelligence and anthropic acquisition rumors, and finally we end up with an Italian startup bringing out a new humanoid. This is the Robotics Stack Podcast, episode 10. My name is Oscar. Hey, my name is Leo. And let's get started.

speaker-1: All right, so as you said, our first topic we are addressing Sunday robotics and their recent communication about Act Two, which is the model running the robot they are also making because they make both the humanoid robot and the model. So it is still closed source, but we are going to get into what makes it interesting. So to give a pi a bit of background Sunday Robotics is a Silicon Valley startup. They are funded in April 2024 by Tony Zao and Shang-Chi, and they raised 165 million at a 1.15 billion valuation Series B. If you are a humanoid startup right now and you have reached a billion dollar valuation, you are not out of the wood yet, but you are still pretty comfortable to make things happen. Shangxi CTO is an ex-Stanford. And he worked a lot on tear operation and data acquisition. And this is going to get important for later. Their flagship robot is named Memo and it's approximately 1 meter 70 centimeters tall. It has a four-hour battery. It is wheeled, so no legs, it cannot do stairs, but it can still, you know, drive around your home. And it has some kind of very specific hands that you don't see often rhinotics. Right now you see mostly you know pinchers, ⁓ clamps, or you see, you know, dexterous exactly, or dexterous hands, you know, that have like five fingers. Yeah. And that are obviously much more complex to you know maneuver. The choice Sunday Robotics is making is to have a three-finger hand. So it's pretty much like these three fingers are together, then you get an index and you get a thumb. So you can still pinch and you get also more flexibility around it. And they also released a glove that matches the robot's hands. And maybe we can show a picture ⁓ in the podcast as well because it's a very specific device. I think it is still a bit bulky, but they manufactured and shipped those hands beforehand in order to make data acquisition. So it is not data acquisition from teleoperation. It's data acquisition from human activity. So basically you were supposed to wear this glove to fold your laundry, for example. I'm not taking that example, you know, out of nowhere because this is The main example they demoed in their recent communication as well. They released like a communication because they did not release the model, so we cannot properly say that they released the model. A frozen model that folded laundry in unseen homes, so it means in new environments in more technical terms, with a 99.1% accuracy. Zero like fine-tuning. It's like one shot attempt, 99.1% accuracy. Okay. So What they want what they mean here is that it is the first robotics model to unify broad generalization with high reliability.

speaker-0: Did they show like a video of this particular one shot scenario? ⁓

speaker-1: ⁓ yes, but again it's ⁓ it's a corporate communication, so how do you know if it's one shot? How do you know the amount of data that is actually published? How do you know if there isn't like a bunch of failure attempts that you don't get to see? But yeah, a lot of videos online on their website ⁓ has been released. So they have a demo video and they are also if you go a bit further down the page, ⁓ you get to see many videos of the robots folding laundries. They narrow the scope of laundry folding onto nine garment types, right? In order to get some kind of repeatability and higher accuracy, probably. So which is pretty smart I think, but also makes me wonder if the model is able to generalize to some kind of garment that it has never seen. Or if you have some kind of weird, I don't know, of weird t-shirt that is between a polo shirt and a t-shirt or something that is, you know, not really categorizable for for the AI model. Would it fail or would it be able to fold it ⁓ accurately. I'm not sure. And they obviously have not tried that because that would increase the failure rate and you don't want to complicate on that. Also what they claim is that a single fine-tuning example can teach the model a new behavior. So in the case I mentioned with the new guardment, supposedly you are supposed to be able to demonstrate like how to fold a new type of clothes and the robot is supposed to learn it in one shot. Again, this has not been tested. outside of the company's ⁓ control environment and communication. So we don't really know for sure how accurate it is.

speaker-0: Was it this video where they were folding it on the bed? Yes. Okay, it's it does make sense to fold on a bed, but you know, a bed is also like a very plain surface. Yeah.

speaker-1: Yeah yeah, exactly. I I would love to see a demo where the robots can pick up the clothes from the dryer, ⁓ put them inside out, put them like I don't know, back in the right place if they are inside out and so on. But I I've seen a part ⁓ a video where the robot is actually able to get an inside out pants and put it back in the correct setting before folding it. So I I think th they have, you know, ⁓ addressed that issue as well. And also Something that I found very interesting in the video is that you know the robot, the memo robot has some kind of hat, ⁓ either a cap or a round hat, and that's where the camera is located. And it's a wide-angled camera. And you get to see demonstrations where the robot starts folding a piece of clothing and an operator comes up and just puts a t-shirt or something on top of the robot's head so it cannot see anymore. And the robot keeps folding. I was like, how does that work? How is this even possible? And it's because the robot has, you know, beyond just vision sensing, it also has sensing in its hand, in its spatial overall placement in its environment. So once it starts, you know, folding a t shirt, it knows where it is, you know, along the progress of the task and it can just keep doing it without the camera. Which makes sense because if you have a t shirt in your hand, you could probably fold it with your eyes closed as well, right? Because if you've done it sat you've done that in your life probably a thousand times. So It's something you can do out of muscle memory.

speaker-0: And the robot also doesn't forget what he has done before, so he can basically ⁓ follow the steps.

speaker-1: Yeah, yeah, and I think I think it's also comparable to most of memory. It's like you just do it out of habits, you could do it with your eyes closed. The robot still has feedback from his sensors in his hands forever. That's how it works. And Sunday also released something they called a benchmark, but it's not exactly a benchmark. It's called solve. They say it's able to get reliable performance across a decascope. At a stated adaptation cost. And they don't really release a way to perform evaluation or a benchmark, but it's more of a set of guidelines that other companies should follow. And it goes along three axes: so performance, scope, and adaptation cost. It's a bit ironic because adaptation cost is precisely what we don't get to see in their videos, right? You mentioned falling on the bed, and we don't know how many failure attempts are not shown in the data. So it's exactly what we don't get to see. And they say, okay, if you want to benchmark your robot on a specific set of tasks, you should follow these guidelines. So it's not really clear to me why they are not following their own guidelines. They don't share like a task suit, no standardized ride, no third party grader. And you know, I'm I'm more and more bullish on a third party grader for robotics better benchmark to emerge very soon because that's definitely what ⁓ I think the space needs right now. And It's pretty much themselves declaring, okay, this is our benchmark and we are evaluating ourselves on it. Yeah. It's a way to evaluate a result of a robot on a specific task, but not a really good way at the same time to compare to a robot.

speaker-0: Yeah, interesting. I've seen the video as well. I was impressed because the fluidity of the motion is really, really on point. And I have never seen a comparable example. So it makes sense that they show this particular example. Obviously with the things that you mentioned, the self-reporting, the specific conditions. Yeah, maybe you should have a third-party evaluation at some point if you want full credibility.

speaker-1: Yeah, definitely. Someone completely independent. And also, I mean, I found it a bit odd that they don't even I mean that they don't release the model. I understand. you you can argue that Tesla 1X figure they are making humanoids too and they don't they don't release their model. They have, you know, their incentive to remain closed source and it's fine. But they don't even communicate on the type of model, on the data set, on the architecture, on the model side. They just say Yeah, it's one shot, it's scalable, and that's pretty much it. Again, not everybody is supposed to be open source, but many open source models are coming up, so you should at least communicate on I guess the technicalities of your model and your data and your data set.

speaker-0: That's yes. And also it comes back to the question, okay, clamp three fingered hand versus full humanoid hand. You know, if you see one X hand that we discussed in the last episode, at what point do you need this sort of dexterity, of course? And if they came out with a glove, you know, as you said, to kind of record this motion and then apply this motion to the robot, right? That's what they do, as I understood it. To me it just back to question, okay, why not use

speaker-1: Yes, exactly. Yeah.

speaker-0: An actual model of the hand, why not always get like this abstraction of the human interface? I mean, there's probably a reason for it, you know, but still.

speaker-1: I have a take on it. I I think so. I don't want to seem like I'm ⁓ pushing this company down too much because that's not at all what I'm doing. I'm just comparing it to the current competition, which is extremely fierce right now. And I think at first their ID is pretty smart because the three finger hand is better than a simple clamp. And it's also much easier to train than the dexterous hand, right? And we are talking about two year company who is raising a robot already. So it's still pretty impressive. So I think the three finger hands is a pretty good idea for simple tasks. What it makes me wonder is that like n the One X, for example, you could imagine the One X robot to perform any task in your home, including cooking, putting groceries in the fridge or something. ⁓ you cannot really picture Sunday Robotics Memo doing that, right? Besides laundry folding, ⁓ I'm guessing the range of applications it can undertake are much more limited. than a robot with, you know, five fingers, dexterous hands. So still a pretty impressive achievement, but if the competition can deliver similar results with dexterous hands, ⁓ you are definitely ⁓ getting a setback with, you know, a three-finger hand.

speaker-0: Yeah. And even then, like I mean that's obviously criticizing on a high level here, but ninety nine point one percent is still not a hundred percent. There's still something missing there. So Yeah.

speaker-1: I mean, I don't know, if you were to ford laundry a thousand times, would you get it right nine hundred and ninety one times over a thousand? I'm not even sure.

speaker-0: I I'm just saying because if you think about it, you outsource this test at some point. I mean it's it's mostly about the teaching, right? It's not about the actual execution. Or is it about the execution, the ninety nine percent? Execution. Okay. But then it matters because if I leave this robot for three hours and I drive somewhere and then I g come back and then I realize okay, I just wasted like four hours with this robot, it failed and then, you know, it's one out of a hundred times, but it will still kinda bug me, right?

speaker-1: I don't think that's how you should see it. It's like at at the beginning of, you know, when we were designing CNN to detect image content and image semantics, very very quickly CNN were more accurate than human at, you know, describing what's in an image. Because as a human, if you have to look at a thousand images and for each single one of them ⁓ describe what's in it, like a bird, a car, a dog or whatever, you are going to make mistake just out of boredom, out of repetition. I think it's the same for large refolding. If you are going to fall like all your laundry at once or do a bunch of household calls or tasks, ⁓ I think eventually you are going to make mistakes. I think a human will probably make more mistake in their own home.

speaker-0: I mean I get what you're saying, but ⁓ to me it just is the question how do how do these one percent errors ⁓ materialize themselves? Yeah. Because if you have somebody as a human doing this error and they fold a t shirt and then place it on there and then they notice it's bad, right, they just do it again. But as we learned with robots, these errors compound. So maybe it folds the t shirt wrong way once, then it places it. It is not at this position where it's expected to be and then the whole stack of t shirts suddenly tilts down and you know, we have a mess again.

speaker-1: Yeah, yeah. You mean the ability to recover from a mistake is not demonstrated on the robot. And yes, that's true. They did not communicate on that. I've seen examples of other companies and models communicate on, you know, the ability to notice an error and correct it, but this time they did not communicate on that.

speaker-0: Yeah. But still I found it really impressive. Like their demo is super impressive.

speaker-1: Real world demonstrations are always impressive and I think that's definitely where everybody should go.

speaker-0: In in general, I just found it so amazing that in this short time frame we see so many young companies, you know, creating actual hardware, shipping stuff and ⁓ really going all in. I mean, of course they got a lot of money, but they're also delivering in that sense. And there's something else which is also not too old, a new aircraft. And this stems from a collaboration of Underreal and Archer. Because the lines are blurry between okay, what's an aircraft, what's a machine, what's a factory? Is it a robot? Yes or no? I mean, we'll see. But Andreel, ⁓ we talked about them before. Defense Company. We talked about their Arsenal OS, a digital platform for manufacturing. So we've covered them. But Archer we have not really touched yet. But it's quite an interesting company because it was founded by Adam Goldstein and Brad Adcock. And Brad Edcock is obviously a big name in The robotics space because he is the founder of Figure AI, which we all know, of course. So those two they founded Archer in 2018. And Archer is essentially an aircraft and defense company. And they already have one product, for example, which is the piloted air taxi. It's called Midnight. And this one is designed for four people, looks super nice. Four people plus a pilot, and it has around a 100-mile range. And with this product, They're currently in the process of getting certifications to really roll it out and bring it into the real world. And they have been raising quite a bit of money as well. They raised $430 million in equity to support Archer Defense, led by an ex-Lockheed and Sikorsky veteran. The company, Archer as well, has been public since 2021. So at one point, after they raised another 800. 50 million in 2025, they got close to 2 billion in liquidity. And that's ⁓ kind of what you need. That's why they also have a bunch of partnerships. For example, they have a partnership with United Airlines, which already did some pre-orders with Stellantis, Abu Dhabi Aviation as a launch customer for it, Palantir. But as I said, they also had a partnership with Undereal to jointly develop the next generation hybrid. Evitol aircraft for defense applications, targeting a potential US Department of Defense program of record here. That was 2024, and now they actually released this product, an Evitol. Just announced it at the Farnborough International Air Show. Evito stands for electric vertical takeoff and landing aircraft. So it's basically an autonomous aircraft, or how I like to call it, a flying robot. And it looks pretty badass. I don't know if you've seen it, but ⁓ it's a tilt rotor, looks a bit like a Plane helicopter mix and it comes in two variants. So one is called Halo, and this one is the commercial side version led by Archer, and the other one is called Thunder on the defense side, led by Andre, of course. So it's a really kind of an interesting partnership that they have going here. And tilt rotor in that sense means it can vertically take off and land, which is quite useful, of course, because it's runway independent from Austria locations. So it's a hybrid electric. With serious hybrid electric powertrain. What does it mean? It means it has a fuel powered generator with jet fuel, aviation fuel, and this one produces the electricity, and the electricity then drives electric motors on the rotors. And that gives extra range, endurance, and also a very low acoustic signature. If you think about the Tesla, obviously it's also not that loud, you know, you sometimes don't even hear them when they approach. It actually goes back to an earlier collaboration where Archer was supplying this. powertrain for the underreal and edge collaboration, the Omen drone. I don't know if you've seen the Omen drone, but Archer was already collaborating with Andereal basically on this one. It was also last year. What's interesting about these tilt rotors, they lift the aircraft up there horizontally and then they tilt towards a vertical position. And then they can also adjust the RPM on the fly for different speeds, altitudes. And they also use Underwheel's lattice AI. You talked about that, I remember that, for formation, flying, and mission autonomy. And it's also foldable and transportable. So you can ship it in a standard shipping container. And that means you can basically deploy this aircraft pretty much anywhere. And it just gets started fully autonomous and starts flying. It has a high payload capacity, long range. And we're talking like almost a thousand miles target range, high speed, so up to three hundred and twenty miles per hour. It's really fast. And of course, they want to get this into a low cost, higher volume manufacturing situation. Potentially, maybe they use Arsenal OS on some point for that, but they really want to make, you know, defense scalable. US in-house. As I said, there are two different variations. So this is kind of like the baseline that both of these variations build on top. But then there's some differences between of course the Halo 1, which is the Archer variant, and the Thunder One, which is the defense underware variant. So for the Halo, it's obviously focused on civilian and logistics missions, so freight, aeromedical, search and rescue tasks. cargo, passenger or even specialized payload configurations, but definitely no weapons. They also partner, for example, with Marobani Aerospace as a strategic launch partner, a Japanese trade company. Now Thunder, on the other hand, the defense variant, is a group five autonomous attack rotor craft. Pretty much the largest category in the Department of Defense classification for such aircraft. And it's built to serve as a loyal wingman for Apaches, you know these military helicopters or the J N2. And that's why it really needs this high speed. Situations where you have a piloted helicopter and then it's also surrounded by multiple autonomous aircrafts to kind of expand the range and expand the capabilities and of course also security for the piloted one. And it is modular so it has multiple possibilities for internal weapon base. So for example, hellfire or Jack ⁓ missiles. You can get rockets onto this, counter drone weapons, and you can see a pretty nice visualization of that on the website of Unreal. Obviously, this one is combat focused, so it aims to give you fire support in contested environments, for example, or do surveillance, spotting enemies, etc. So all things that are pretty much related to warfare. And they're actually targeting the first flight in 2027. So it's really getting started for this one as well. And they've already done multiple tests. And for Archer, it was a very good announcement that they have because it was just announced on this air show. And after that, the stocks of Archer surged around like 20%. So people really seem to like this one. And it also is interesting because of the form factor, right? If you look at it, I've never seen something like this before. Of course, I've seen tilt rotors, and you could say, okay, it's just a stylish tilt rotor. But I really see how all these technologies and design languages come together now and we see all these super badass looking hardware with badass backs today.

speaker-1: Yeah, I see what you mean. what I I really like everything Angri makes and they also showcase some really cool video at every product announcement, ⁓ which I find super interesting for a defense company. It's a new way to to do marketing for them in a very traditionally old school ⁓ domain. So I like that a lot about them. And the thunder yeah looks extremely impressive. I also read somewhere, I'm not sure if it was in Iran or in Ukraine. I don't want to, you know, provide ⁓ unconfirmed information. But I think recently ⁓ the US lost ⁓ a couple Apaches helicopters in a conflict. And that's also where this idea of a companion kind of drone comes from. ⁓ they were like, okay, if we have an expensive Apache with actual humans inside to pilot it and maybe there is also a gunner. So having like drones around it that can serve as basically missile bait when you are in ⁓ enemy territory was already an idea booming up for a while actually. So I think it's interesting that you get something that is also kind of like an helicopter but can also launch missile at the same time, which is just n not just protection, it can also serve as basically launching as many missiles as the main Apache has. I guess with a bit less flexibility. But let's say if you have like I don't know four thunders next to an Apache, suddenly it's as powerful as five Apaches at the same time, maybe. And I would be very curious to see if it works. But again, I'm not an expert on on defense, but we can see in Ukraine and Iran what's happening that warfare is definitely changing and the United States, which has by far the larger the largest and most financed military of the world, is adapting to that quite quickly. Because you can see that as an attacker, you pretty much have the worst role ever. Because I mean defending yourself is becoming much cheaper than it ever was with drones and, you know, cheap manufacturing. So there is an asymmetric warfare. And I think Andrew is working hard towards making it more equitable if you could frame it that way.

speaker-0: Yeah. And it's really interesting to me, ⁓ the pricing because they frame it as low cost. So we don't really have a clear price on that. Maybe it will never be published.

speaker-1: usually is eventually because ⁓ there is some kind of publicity in a defense contract in the US. So for example you know the approximate price of an Apache. I don't have it at the top of my mind, obviously, but if you look at Wikipedia as the is a page for the Apache you will find the price on it usually ⁓ because ⁓ since the US is spending a lot of public money on it, they have to read that.

speaker-0: Yeah, and if you don't get it for the underreal Thunder, then you probably get it for the Halo version because that one is yes commercial and you know, companies will potentially buy it. But it will be interesting because maybe you can afford to lose three or four of them ⁓ if you protect a human, as you see.

speaker-1: Yeah, definitely. And I think that's totally the point here that you can afford to lose this acra you don't lose the human life. But I think Andrew is also very good at streamlining manufacturing. So to make it as efficient as possible. I would not say it's going to be cheap because everything in defence is still pretty expensive. But it's going to be, I don't know, much, much less expensive than an Apache, for example. That's for sure. I'm also very curious about how autonomous is it and what kind of model are they using. Is that is this aircraft going to be able to be like autonomous like any other existing drone, or is it going to be autonomous to into the extent of which it is following a manned aircraft like an Apache? Yeah.

speaker-0: It's designed to be a ⁓ a wingman actually.

speaker-1: Yeah, okay. So it's not designed to be alone.

speaker-0: No, not not not necessarily. I mean you can deploy it and and use it alone, but I think what you just spoke about, they clearly want to target this specific problem as well.

speaker-1: That totally makes sense. So I guess it makes navigation easier. And maybe that's why they don't communicate too much on the model. Because you have probably heard recently communication about the models from both Anthropic and another company called Physical Intelligence. Have you heard of that? Yeah. So I think they're also a very interesting company. They do research on, you know, models to control robots. So what we can traditionally call the robot brain. They're also sometimes called Pi.

speaker-0: Yes. Okay.

speaker-1: Because the initial the initials are obviously PI, so they are called Pi, like the symbol pi. And a few days ago, there was a really strong rumors going on everywhere about anthropic acquiring physical intelligence. But I mean, we we should say it out loud from the beginning. This is a rumor, and this has been disproved by you know physical intelligence since. So what actually happened is that tech influencer Robert Scoball posted on X that He had information from an undisclosed investors. Pretty easy to say if you ask me, that Anthropic was in talk to acquire physical intelligence. What actually seems to have happened ⁓ is that Anthropic was in talk with physical intelligence in spring earlier this year, you know, so several months ago, and the deal was cancelled after that. So it's probably like a recycled late data point coming into the public. right now and it has been denied by basically the entire physical intelligence team. But I think it is still a pretty good ⁓ in opportunity to discuss physical intelligence because we have not mentioned them in the podcast yet. So basically physical intelligence is based in San Francisco and they describe themselves as a robot foundation model startup. And they were founded in twenty twenty four. Again, many strong Startups, robotic startups in the Bay Area founded in 2024 apparently. Looks like a sounds like a good year. And one of the founders, Sergei Levine, pitches it as Chat GPT for robots. And you can also get, you know, a podcast from Wong, who is another founder. And he was invited into a podcast with Y Combinator. And the podcast was titled The Chat GPT Moment for Robots is here. So yeah. It's really I I guess one of the few good use case of making X for Y kind of situation. You know, it's not like you make Uber for dogs, it's Chagi P for robots. And from what we've seen, it actually works, which is pretty impressive. So a little bit on their funding, because I think it's a good case against them selling. So they raised a 70 million seed in 2024. They raised a series A of 400 million. led by Jeff Bezos in late 2024. And also interesting because this is going to be important later. OpenAI was a co-investor along with Jeff Bezos in this series. And there raised a 600 million Series B in November 2025. And since March 2026, they are in talk to raise 1 billion at at 11 billion valuation. And you see These talks in March that have been going on since March 2026, they are approximately the same time as they were, you know, talks within Tropic about an acquisition. So maybe it was just, you know, an introductory call and nothing too serious because quite often when you are a company that is, you know, seriously considered for funding and funding is still an endeavor and it is still takes a lot of effort. But the company is like this, they have zero uncertainty about their ability to raise. You know, they will raise. It will happen. No matter what. The only details that are being discussed is the details on valuation on who is leading and so on. So maybe at the same time Enthropic approached them and offered them to, you know, join the company in a way. Because if even if it's an acquisition, considering, you know, the level of involvement of all the participants and the fact that Enthropic wants to develop its robotic arm as well. It will be like kind of a merger, even though Enthropic is not bigger, right? I'm also thinking OpenAI being an existing investor would make it pretty complicated because open AI and Anthropic have been you know in an ongoing battle for like a couple of years now. And Anthropic has ⁓ overtaken OpenAI in both revenue and profitability. I don't know if OpenAI would sue them if they were to acquire physical intelligence, but that would steer some trouble definitely. So a little bit about the founders. ⁓ So Horsman, CEO, is an ex Goodbrain, Sergei Levine, he is still a professor at Berkeley. Chelsea film, Brian Inster. ⁓ Lachi Groom, Adnan Esmail who is working on hardware and he's an ex-NGRIL actually. And Kwan Wong who I mentioned who you know did the YC podcast. Yeah, it's a pretty big team of founders. The way I understand it they have eight founders. ⁓ I don't know if I'm missing one or adding one, so don't quote me on that.

speaker-0: That's a big team, right?

speaker-1: But Petilar founding team, Wong says it makes sense because they all have complementary skill sets. It's the first startup for most of them that they are funding. They're most of them met at Google working on robotics and they are really interested in making this company and making the model work. They are not really interested in making a quick buck. I mean they are all heavy researchers profile. So if they were interested in making money, ⁓ they would have done something else from the from the beginning, right? So I think they are really interesting in solving robotics, you know. Obviously. No, but why is that? Maybe that's around a Sam Altman type will come up and turn the company into, you know, a profitable business, you know, because OpenAI had been a non profit doing research for like five years, maybe before you know coming up with ChatGPT and turning it into a product that everybody uses now every day. So ⁓ yeah, they also make mostly open model.

speaker-0: what they that's what they all say.

speaker-1: I mean their vision, which I find really interesting, is they want to make a model that can control any robots to do any task. And the any robot is especially interesting to me because right now most models, if they are made by a ⁓ humanoid company, they are designed to work on their own hardware and not outside of it. For example, I talked about Sunday and their AC2 model earlier. And it is designed and trained and optimized to work on their memo robot. It's not going to work on any other robot. You know, they haven't tested it and they probably will not. But if you were to take this model and put it into a figure AI, you wanna it it would probably not work at all, even beyond the, you know, hand difference.

speaker-0: That's the that's the embodiment problem, right?

speaker-1: Yeah, exactly, exactly. Cress embodiment problem. And but what I like about physical intelligence is that crass embodiment is at the very center of their vision. They say that a generous model is the only way to go because you are going to have many robot platforms, and this is already true, between, you know, American manufacturers, Chinese manufacturers, and now we even see ⁓ supply the French company starting making its own robots right now in Paris called UMA. And you are going to talk about an Italian company who makes its own humanoid as well. So I mean, just like we have many many di different mobile phone manufacturers, we have many different car manufacturers, we are going to have many different robots manufacturers with different hardware. And even within a single robot manufacturer, you cannot freeze hardware. Even on a single, I don't know, Unitry G one or whatever, ⁓ Maybe the one from this year would have a slightly different shape or slightly different eye trailer from the one from next year. So you don't want, you know, a specific model to be explicitly trained on a single hardware. You want a model that is able to generalize, right? And that's exactly what it means. It's the same if you learn to eat or whatever and use, you know, a fork and a knife. You want to be able to use different shape or fork and knife. You don't want to have, you know, the exact single pair of, you know, very specific forks and knives. It just doesn't make sense. So This is yeah, one of their thesis that I really like. And there is also another very specific aspect to their technology, which no one else has, is that their model does inference in the cloud, not embedded in the robots. And I think this is the first time I hear this right now. Everybody is doing, you know, embedded compute and ⁓ buying Nvidia jets and buy the truckload to you know store to ship on their robot. But They do okay, no, no, we do cloud inference. The model runs into a data center and ⁓ they have a very fine logic where for example the robot has like two seconds in its brain of action, you know, to complete. And every second they will redo inference and in it takes 50 milliseconds to happen to happen to the next action on top of the one already happening. So it's kind of a streaming of action, you know, that's how the robot behaves. And when you think about it, if movie streaming like Netflix. works. ⁓ action streaming for robots can work as well, right? It's not that much difference. Just a good

speaker-0: Yeah, but but ⁓ I mean I I get it, it it makes sense and I think their goal is extremely ambitious. But ⁓ I always think about the example of a robot catching a ball or playing tennis with you, right? We're talking about milliseconds of delay here. And yeah.

speaker-1: I see what mean. But I think right now it's mostly designed for household tasks and things like that. It's true that I have not seen them discuss interruption or perturbations in the robots, you know, in the robots work, for example. But maybe their model is good enough already to work with it.

speaker-0: And it will I I mean in my opinion it's it's a bet on the future because I I think it will get there, right? Internet is getting better, upload stream, download stream, everything is getting faster. ⁓ so it will just happen at some point. Yeah.

speaker-1: ⁓ but ⁓ on the other hand also means that if you have an internet a temporary internet connection issue, your robot doesn't work anymore, right? Yeah. So it's it's a trade-off, but I get where they are going with this. They have two ongoing deployments. So I mean professionally, one of them is with a company called Wave. They have they deploy robots into Laundromat to fold the calls again, and they manage to get a working prototype in two weeks using physical intelligence models. And they also work with ⁓ Ultra, which is which do packing in soft pouches, you know, in e-commerce warehouse. So it's, you know, not packing in a rigid cardboard, packing in a soft pouch, which is more difficult for robots to handle. And also interesting, both of these companies are from Y combinators. So Y combinator is everywhere again. The founders of physical intelligence really believe in is in a Can bring an explosion of robotics company. So basically, upfront costs to ⁓ you know start a robotics company have collapsed. So even if all companies are raising 100 million and reaching a 1 billion valuation, upfront costs still have collapsed compared to before. You can get a humanoid for less than 10k. Training your own model, you know, is not as expensive as it used to be. So right now, differentiator is onto understanding the workflow and into the data. Right. Basically there is going to be a lot of you know Robotics company solving real world issues and it's already happening in my opinion. Right.

speaker-0: Yeah. That's why that's why we're launching the stag robot.

speaker-1: What it what was it going to do? Like it's going to be like a a third person on the podcast?

speaker-0: Maybe, maybe, maybe we're gonna replace you Leo. So or or me. Then I can ⁓ chill on the island. Shots fired.

speaker-1: Maybe. I don't know. Will Shine would you pick for me?

speaker-0: Yeah. I don't know, Meldies, something like that.

speaker-1: Maybe. Maybe too hot for me, but maybe. Alright. So yeah. So a as a closing argument, ⁓ and I think that's also a strong argument against physical intelligence selling to anybody, is that they believe that if we solve robotics, it will add approximately ten percent to the US GDP. Which makes sense. In a way, if you think of all the tasks robots can complete and all the free time you get out of that. and that you can you can dedicate to working more on your own as well. I think they have a really strong vision and really strong confidence in their ability to do it. And I think this is a good marker of them not to sub in the future.

speaker-0: Yeah. I mean speaking of tests, there still seems to be like a obsession with ⁓ packaging and folding laundry tasks. So I really hope that we can emerge out of this at some point. Maybe you can make a little Twitter post because you always make these beautiful infographics. Maybe you can make an infographic with the most common tasks that are solved so far because ⁓ that will be really

speaker-1: That's a good idea. That's a good idea. I think laundry folding makes sense because it is still pretty accessible for robots. ⁓ but if I think of the time I spend on a daily basis doing tasks, doing like curls in my home, I think cooking would be by far the one that is the most time consuming. ⁓ Yeah, maybe.

speaker-0: Cleaning and cooking. Laundry. Yeah. I think cleaning cleaning is definitely like a good half a day per week or even yeah a full day sometimes. And that is a huge annoyance. I mean, for some people cooking is fun and I think cooking is still more complex, but I agree, like having a personal cook, that will be amazing. ⁓ and yeah, that there are some simple things. We saw this argument the other day where somebody said, Okay, at the end They're just a couple of different tasks that you have to solve. You know, you don't have to solve everything. If you just solve a couple different tasks for a home robot, then it already becomes pretty useful. And yeah, my hope is just that as we see more companies, maybe there's gonna also be some diversification on what they focus on. For example, we have another humanoid startup from Italy, which is called Generative Bionics. And they're actually a spin out of the Istituto Italiano di. Technologia. I hope I pronounced that correctly. The IIT in Genoa. They specialize in humanoid robots, so just another one. The company builds on top of this institute experience basically and the whole research from it. Which is quite interesting because they had a couple of very, very interesting research projects and products in there. For example the iCup, the Ergo Cup and the Iron Cup. So maybe we can quickly ⁓ go over these because these kind of build the foundation for the new humanoid robots. So iCup, ⁓ not iPod or anything, actually started in 2004 already. So way, way back, you know. And this was a foundational child sized humanoid robot. So around one meter tall. So already we're talking over 20 years ago. And it was open source and tailored towards researchers, of course, as a research platform. It had ⁓ advanced sensors, including Tektile Skin, which we Later see in the new robot as well. Force and torque sensing, compliant joints. And even back then they deployed dozens of units and sent them worldwide to ⁓ do research on robotics. Then they also had the Ergo Cup, which is basically an evolution of the iCup. And this one ⁓ was more for real industrial and collaborative work, and it was also taller, so one meter and fifty and stronger. So for real physical human work, lifting, carrying loads. And this one was developed with the INAIL. So an Italian workplace safety institute. And it emphasized really on erg ergonomics, safe force interaction and practical tasks, which is we talked about it before, also super important because if a robot falls down on a person, it has the human weight, it's hard, it's metal sometimes. If you get hit by it, you can really get some serious issues here. So even back then. They thought about all these things. The next stage or the next evolutionary step was then the Iron Cup. And this is a really interesting one because it was actually the world's first flying jet powered humanoid. So maybe that's more of a flying robot than what Archer and Andorreel did before. Because this one, it was built on the iCub platform, but they added jet engines to the robot. So on the arms, and they also added a jet backpack basically to a robot. And you can see a video where they really have some serious thrust there. And you can see this robot. I don't know if you've seen these videos, but there are videos of people from the military as well having these like jet engines on their arms and then they're like flying across the river and stuff like that. And they basically applied this to the robot. And the robot was flying. So, you know, other people are still talking about folding lorry. I mean, these guys were flying. So that was really ⁓ a remarkable video what I've seen. I mean, maybe it's not that useful to be honest, because you don't necessarily need a flying robot. ⁓ I would pr actually prefer to make my laundry, but still, like it's a very impressive task because you have to balance everything. Physics are involved at at a very high level, have to use heat resistant materials. So ⁓ they actually achieved a stable hover and a controlled flight with this one. And yeah, that that is kind of the background because the founders, a lot of them, have been at this institute and were part of this. Technological evolution. But now they formed this new company creating this humanoid and they want to create a scalable, customizable platform rather than a one-size-fits-all robot. So it's really kind of what we said before. The form factor is still developing. Maybe it can be a humanoid, but also a humanoid with a full humanoid hand, maybe three fingers, maybe two fingers, maybe just one hook or whatever, depending on the tasks. The company As I said, it's called Generative Bionics, and they founded it again in 2024 in Genoa, Italy. With regards to their funding, they raised 70 million, one of Europe's largest early deep tech rounds for humanoid robotics in late 2025. So yeah, it may be not comparable to the United States venture game, but still, like quite a remarkable sum. It was led by CDP Venture Capital's AI Fund with participation of AMD Ventures, NI Next, Duferco. And also Tether again. Tether, we have seen them investing into neurorobotics. And ⁓ they have some team members ⁓ who are Italian, so definitely makes sense that they go in here as well. This money is going to be used obviously for the research and development, production, scaling and hiring, because right now they already scale to around 75, between 75 and 100 employees, and many have PhDs from the institute. What is their mission? They're really mission-driven. You can see it on the websites. These are researchers at heart, similar to the physical intelligence people. It's a big team and they really want to amplify the human capabilities through collaborative robots that feel, learn, and adapt to the physical world. That's really what they prioritize. ⁓ what they also prioritize is the European supply chain resilience. So they have partnerships for actuators and designs in Europe and I think most, if not everything, of their supply chain is European. Now, founders, they have six founders in total. So also not a small team. And I think this is something interesting that seems to be more common now than maybe more years before, where we always had like founding teams between one and four people, sorry.

speaker-1: Yeah. No, I think I was gonna say robotics is a pretty complex field. You need hardware, you need software, you need infrastructure, you need many different skills that one person just cannot have. So a founding team of two is usually pretty rare for

speaker-0: Sure. Yeah. But speaking about the founders, we've got Daniele Pucci. He is the CEO and co-founder, and he's a former head of the Institute's Artificial and Mechanical Intelligence Line. And he was a key contributor to these technologies I mentioned before, the iCub, the Ergo Cup, and the Iron Cup projects. Then we have Alessio Dabue, and he's an expert in computer vision and multimodal machine learning. We've got Marco Majiali, mechatronics and iCub Tech Coordinator, experienced in turning prototypes into industrial platforms. We have Andrea Pagnin, technology transfer and innovation strategy expert. We've got David Erosta, who is more on the entrepreneurial side with a tech and telecom background. And we also got Jeffrey Lipschitz, which is also for the entrepreneurial support there. So they have this collective vision. And what is their product? Their new humanoid. It's called Gene01. And Gene01. was just unveiled as a concept first at the CES 2026 during AMD's keynote. But now they released basically their fully functional platform just a couple days ago. So three days ago at July 20. And this is their core scalable platform. So again, not a single end product, but the foundation for customizable humanoids built from concept to functional sensorized robot in just six months. So they really just built it in very short amount of time. And you can go to the YouTube channel. ⁓ a lot of robotics founders have YouTube channels where they show the progress. And if you look back, you can see, for example, videos showing the lower part of the robot, only the legs basically walking in several videos. And now it kind of all comes together. They've released a little video and also a website where you can see the robot. It has a really nice design. It's it's like a sleek red and black design and it also has glowing accents and also a little bit like the T800 as well from engine. Now they don't have too much information about the features, but what they have is full body multimodal skin and tactile sensing. So they have the distributed sensors that detect touch, proximity, force, temperature, which is obviously necessary for safe and intuitive human interaction. And you can see this in the video as well where they have where they show the robot And then a hand moves closer to the robot and the robot kind of responds without the hand even touching the robot. So it's really interesting. So their sensing technology seems to be something that maybe they have an edge on. They have obviously the intelligence distributed in the body. They have mechatronics and AI as one system. And they use reinforcement learning for adaptive locomotion. For example, for blind walking or running on very terrain. or in the darkness. You can see that on the videos or the website as well, that the robot seems to be walking in dark environments. Yeah, it looks cool of course, but it's also quite useful, right? But to be fair, the the walking doesn't yet look that smooth. I've seen smoother walks, you know, but ⁓ it's walking. So who knows? Who knows what's next? So but you can still see it's very human inspired and the design language is kind of Italian. They also have a partnership with Ital Design for manufacturability. Well, and it's modular and customizable. So they have adaptable end effectors and hardware for specific tasks. For example, I found some information about a welding variant called the Gene01 slash W. I love these naming schemes. For Fincantieri shipyards. So a ship company. And they specifically created this welding. version apparently. And it has robotic arms and hands for these welding operations and also cameras and vision systems that monitor the weld scene. So they really think that this customizability is gonna be useful and hopefully it is. They also open sourcing apart. So they already I think open sourced a digital model of the robot available to developers via PiPy, Conda and RRS. That's your keyword Leo. Yeah. Definitely. And yeah, and of course they also have hands, dexterous hands, twenty four degrees of freedom. it's not clear they're tender driven, but very likely if you look at the past ⁓ yes robots they've worked on, they have the close-up sensors, of course. And for the lower part of the body, they already apparently started batch production three months ago. And yeah, it's really just a remarkable example of another company getting started just two years ago, builds on a long, long time of research and now it all comes together and yeah, it's really exciting. So ⁓ we can probably hope for a broader market push, I would say like twenty twenty seven or

speaker-1: Yeah. And it's in Europe. So it it also goes to show that ⁓ not everything is centered in the US or in China. I would be very curious about two things though, is that you mentioned that it's mostly manufactured from European supplier. But ⁓ where do they get their magnets, actually, ball bearings? I I would be curious to learn more about that eventually because that's usually the parts that you have to source from Asia and that's the parts that the US is trying more and more to manufacture on their soil, obviously. And also I would be curious to see their performance on dexterous manipulation because that's usually where things get complicated as well.

speaker-0: Yeah. I mean with regards to the actuators, I can give you an answer. They are partnering with Synapticon. I don't know if you've heard about this company, but it's a European but I don't think so. But ⁓ they haven't released too much, you know. It's it's really fresh, it's really new, but ⁓ I just saw it, so I really wanted to bring it on the pod and ⁓ yeah maybe help them a bit get some exposure even though we're still quite a small podcast.

speaker-1: Maybe have a guest from their company on the podcast as well. Would love that too.

speaker-0: Yeah. I mean to everybody listening, if you wanna be on the show and you have something to say about robotics, maybe you're working on a startup, working in science, doing research or something like that, something interesting, you can always reach out to Leo on X or me or just write in the comments and we'll get in touch with you. You can also write us an email, you'll find it on the YouTube channel and of course like and subscribe. This was the Robotics Tech Podcast, episode number ten. My name is Oscar.

speaker-1: My name is Leo.

speaker-0: Thank you so much for watching. Thank you.