speaker-0: Asimov could be your first true open source robot. We will talk about it on this week's episode of the Robotics Deck Podcast. But before, we will also get into Dinah's new model, which was just released. Then we go to Trossen Robotics and their entire product fleet, which is quite big. And finally, we'll also talk about margins, because you know the margins are pretty important. My name is Oscar. And let's get started.

speaker-1: And my name is Leo. Right. So, first topic of the day is Dina. So, Dinah, as you may know, a redwood city in the Bay Area startup founded in 2024 by Lyndon Gao and York Yang and Jason Ma, actually three co-founders. They raised 144 million total. Their backers include ⁓ CRV, First Round, Nvidia, and Amazon. others. It seems like they are currently trying to raise at a six hundred billion valuation. But what's interesting about them is that this week they released a new foundation model with a pretty bold claim, which is that they train it on one million hours of human video data. So we are talking about egocentric data, not robot teleoperation data, right? But this is still the largest dataset scale that we have seen in the robotic space so far. So To give a bit of background, for the past three years the binding constraint on robot foundation model was teleoperation data, right? Because the claim was that teleoperation was the only way to accurately train a robot model. But teleoperation is expensive, it is slow, and it is most of the time locked to the specific robot it was collected on. If you manipulate a pair of robotic arms, you cannot retransfer that to a humanoid that has five fingers, hands, right? Yeah. So for example, physical intelligence, they collect across ten plus different embodiments to make an embodiment agnostic model. We have covered that already. Generalist they run a collection operation that produces ten thousand hours per week worth of data. And Xiaomi collected ⁓ more than seven thousand hours in real home. So many companies are aware of this limitation and are trying to scale up data collection to address it. Dyna Robotics, ⁓ their new model is called Dyna 2 and they released Dyna 1 earlier, obviously. And it is called a world action model. World action model? Yeah. ⁓ yeah, kind of. I I'm not sure whether they are the first one to use it. I haven't looked into that. But yeah, it's kind of a world model that can also, you know, output actions. Depends where you draw the lines in terms of ⁓ what you want to do with the model. Do you want the world model to when you understand physics? And then you add

speaker-0: That's new term.

speaker-1: a small heads that you fine tune or do you want the model to be end to end? It seems like that shows the latter. So yeah, one million hours is roughly 170 years of continuous working experience. And it has zero robot data in pre-training. So pre-training is the bulk of the training of the robot, right? It's that's where the one minute hours of egocentric data comes in. Post-training you usually use robot data or a different tax specific kind of data. to actually teach the robot to do things. Something that I was not able to verify actually is that they don't really say whether it is a training from scratch or is it using an existing video generation backbone. Because you know usually we use like a larger video model that is trained to understand andor generate videos. And you kind of add another layer of training on top of it in order to adapt it to robot data. And this is called pre training, even though It's still the second layers of training, right? What's interesting is that they found out that robot performance improves smoothly and predictably with every ha added hour of human video across different orders of magnitude, including on tasks the robot has never seen. The thing is, if their predictions are correct, it kind of draws the path forward for the entire robotic industry for data acquisition and what it takes. To train a model that can perform well on certain tasks. It's a mixture of transformers built on video diffusion with separate tokenized pathways for videos, action, and proprioceptions. It means basically you have different transformers inside the model. ⁓ Each of them is called an expert, which has its own white and its own modality. So one is for video, another is for action, and they share a single action string, right? So basically the action expert inherits the general learned physics, but then they become an expert at a specific type of modality. So they become an expert at video and expert at action. But they share the same physic, the same physics understanding backboard. Okay? The video stream uses causal masking. It's what we call an attention restriction mechanism, where basically the tokens that you are currently training can only see the tokens before it in the sequence, not the one in the future. So it trains it to behave like if it were you know inferring the real world because in the real world obviously you can't see the future. The action stream, however, it can see both ⁓ the past and the future. Apparently it allows to have a richer context for the action streams. And it kind of makes sense in a way because for video you cannot predict. I mean you want the model to predict what's going to happen. You cannot know what's going to happen. But if the model has, you know, for example started an arm movement like this, it knows that in a half of a second it would be there. So it kind of makes sense for it to know. ⁓ what's going to happen in the near future. And it is trained through flow matching with the joint video and action loss. So basically flow matching it's kind of denoising from a raw bunch of noise to the target. And the target is the future video frame and the robot action together, which is why the loss, you know, burnt on the two. Just a little bit on the technical aspects of the model there. So now about the data. So you know one unhow worth of data, what's in it? That's where the bulk of the claim is. So the Free training corpus is pretty much head-mounted first-person footage of everyday manipulations like assembling, cooking, folding, and so on. They augment this data with 3D hand pose extraction. So basically from the video, you are able to estimate where the hands and fingers are in 3D. And they also estimate wrist pose from ⁓ the thumb and index aperture. And what's interesting is that they actually created four datasets out of that. One with 1000 hours. 10,000, 100,000 and 1 million. And they train different models with each dataset in order to compare the results. Talking about results, so odd robots post training performance, so just using this data, 1,000 hours yields 20% accuracy, 10,000, 28%, 100,000, 45%, and 1 million hours, 53%.

speaker-0: So it increases as the hours increase, right? Yeah.

speaker-1: Yes, it does. Not necessarily as you would expect. And again, these four numbers, they are taken out of context, you know, they don't really know what they mean. But Dyna provides real world comparison. For example, ⁓ Dyna has deployed robots at customer sites. Right now the model running these robots is called Dyna One and this one is Dyna two. So zero shot at real customer sites, the quality pass rate went from forty six percent to eighty seven percent. Which is a pretty significant increase for zero shot adaptation, right? And they also say that task success rate was ninety-nine point nine percent for both. So the only gain was in quality of the task. But what I don't like about this explanation is that quality is never defined nor explicited. It's pretty easy to say, ⁓ yeah, quality has increased, but you don't really say how me you measure quality. If it's not accurately defined, then it's not a good metric, in my opinion. But again, that's how the robotic space, you know, is doing his benchmark nowadays. One interesting task that they showed is the unscrewing of a bottle cap, you know, where you have to hold the bottle, hold the cap, turn it and open it. And this is interesting because this requires some very fine grain adaptation of force in the fingers and there is no force data in the dataset. So the robot needs to learn to adapt on its own. And they used two hoogie five finger hands. For this task and after post-training it with ten minutes of teleoperation data, they manage to get it to work. So it means they come from a video diffusion generation model, they pre-train it on one million hours, and then they just give it ten minutes of data of unscrewing a bottle, and the robot is able to open the bottle, right? Which is pretty impressive. And that brings us to the scaling law. So basically the scaling

speaker-0: Yeah.

speaker-1: Kelling law is that the amount of data do you need to determine success and they kind of manage with their, you know, four datasets with different order of magnitude to determine that. And this is pretty interesting because the correlation between their numbers and the scaling curve is around 0.9, a pretty strong correlation. And it is fitted on action prediction error. This is the metrics they are using here. And zero shot robot accuracy goes from 6.3%. to about sixteen point four percent between one thousand and one million hours. So you almost tripled the zero shot accuracy when you get one thousand X the amount of data. That's approximately what was it just just one thousand X. Only one thousand X. And ⁓ they don't make the prediction, but I did it. So if this scaling law were to hold for ten million hours

speaker-0: Just one thousand eggs.

speaker-1: Zero shot accuracy would be twenty two point six percent, right? And zero shot robot accuracy, we don't really know what's exactly needed. We don't really know what number is a good number in order to have a robot that is ready to be fine tuned in the real world. But if let's say we want to reach ninety percent, do you know how many hours we need?

speaker-0: ⁓ I don't know, multiple billions of hours. Yeah.

speaker-1: Yeah, two hundred and five billion hours. So I don't know how do you ⁓ you know record that. even using all the humans in the world, it would be, you know, a pretty crazy ⁓ amount of hours. So I I'm guessing there is ⁓ going to be, you know, a lot of research needed on ⁓ you know using lase data, standing data and so on.

speaker-0: Yeah. But of course the narrative of being able to scale. I mean, we had this in the podcast a couple of times, where you have a model and the model data combination and you really try to abstract some sort of scaling law from it. And yeah. That is of course the holy grail. So I mean I I wish them a lot of luck. In terms of architecture, it kinda reminded me a bit of Cosmos. Also Yan Le Cause architecture that you mentioned. Yeah. in the last podcast I think. So even if it's a completely new name, that's just what I'm saying is still builds on a lot of stuff that we already see in other ⁓ areas as well.

speaker-1: Yeah, yeah, definitely. I mean entire industry is going pretty much in the same direction. There is not like a completely new breakthrough approach that has nothing to do with anything else. Like everybody is building on top of every of everybody else, for sure. And a few interesting things about this model is that adding extra unlabeled videos does not really help the model predict human hands better, for example, but it does help transfer to new embodiments. So basically It means the video signal is not making the model better at modeling human, but it's making it better at transferring to other robots, right? There is something in joint video prediction about learning a representation that survives embodiment changes. That's an interesting findings they made as well. Now, if we look at the data, of course you wonder where does this data come from? ⁓ no public datasets offers more than five thousand hours and one million hours is still much more than all public data sets combined. So they don't really say where it comes from. They say it has been, you know, gathered from egocentric data acquisition device, but they don't really say where from and everything, obviously. There is another company called BuildAI. They claim they have Egocentric 1M, which is a one million egocentric data, video dataset. And they publicly released a hundredth cabin of it, right? And it's recorded with ⁓ Indian factory workers. But Dina has not claimed they are using this one. It is just the only publicly announced data set of the same scale, but Dinah has not, you know, said they are using this one. So we're not really sure. Are they using this one? Are they not? Who knows? Maybe in the future we will be.

speaker-0: Just quickly with regards to the hardware you said they're using what kind of hands? ⁓

speaker-1: ⁓ so it depends. They are using for the border caps they were using Wucci hands, but they also use grippers, they also use grippers with tools. ⁓ they are multi-embodiment basically. So on the research background of scaling law robotics, ⁓ there is like a nice CLA paper from twenty twenty five called Data Scaling Laws in Imitation Learning for Robotic Manipulation that fitted power laws over environments, objects and demonstrations. Also Xiaomi Robotics ⁓ in July they published A hundred thousand hours of embodiment free UMI trajectories. So this is robot data, not egocentric video data. It's a bit more precise. Meta VJPA2 was also trained on one minute hours of videos. Not egocentric videos, but you know, general internet videos just to learn physics. So the idea of scaling low was still very much present in the robotic space. What's new from DINA is that scaling human videos improves. Held out robot performance on a fitted curve with zero robot data in pre-training at one minute hours. Basically, using egocentric video only and no robot data at all in pre-training and still getting a state of the art result is what's new here. It means it is possible to have like a really strong backbone of pre-training without complex data, if we could find it that way, and still get good results with only fine-tuning on the expensive, you know, robot data. So yeah. Back to the, you know, opening of the bottle cap with the two five in your hands, it is a false modulation task and they did it with only ten minutes of telehealth on top of a hand three prior because basically the model is like pre trained on egocentric data, so it doesn't have you know robot hands data. And with only ten minutes of this specific dual hands five finger data for post training, they managed to get a really accurate result for bottle opening. So I think the approach of a world seems very promising. And also what's interesting is that I talked about that on Twitter already. Stanford released a paper called Fact that basically argues that a chunk of the force tactile wall, our ability of model to modelize force, is a training schedule artifact rather than a sensor problem. So basically that you can accurately estimate force and you don't necessarily need, you know, very accurate sensor from it. Because force data is pretty much a missing piece in all robotics data set right now. And it's very expensive to acquire. So if we can work around with Dyna's approach and Stanford's approach, I think will be very promising for tasks that require precise force modulation. A few caveats about this approach is that so it is a data scaling approach, 100%. And they don't talk about compute scaling, no parameter scaling, it's not peer-reviewed, no independent replication. Fourteen tasks across three different embodiments. And they also perform ten trials per task. So you kind of wonder how much of that is cherry picked, but you know, that's how the robotic world is benchmarked and evaluated right now.

speaker-0: So they tried it ten times per task.

speaker-1: Yes, exactly. Yes. Across fourteen tasks. But it's still Yeah, still a stronger report than, you know, most other robotics companies out there. So we have to give them that.

speaker-0: That was true. Yes, of course. I mean also I d I don't really blame them or anything because you have other things to do, right? If you wanna work on your model and you wanna improve it then and if it's not yet even there to be deployed, I mean did they actually deploy it where it's actually working or did they just have a trial in the real world?

speaker-1: So that's a good question. Dyna has real world robots deployed. These robots are supposed to run Dyna one and Dyna two they just announced and they compared it. So they tested Dyna two on this real world robot, but they did not claim that Dina two is shipped yet to this robot.

speaker-0: And and what do these robots do?

speaker-1: If I trust the video demonstration they do cutting, folding, this kind of task.

speaker-0: I'm really curious about what kind of arms they're exactly using and if they're actually doing them themselves. Yeah.

speaker-1: These actually use arms from a company called I2RT and the model is called YAN, YAM for yet another manipulator, and is basically a pair of research-oriented robotic arms.

speaker-0: Okay. Yeah. I mean that's really a perfect keyword because ⁓ research oriented robotic arms is pretty much what Trossen Robotics does, which is our next topic. And ⁓ yeah, I kinda wonder why they didn't go for Trossen here because it's a really established company in this area. So in general they do data capture, hardware kits and robotic arms for teleoperation. The reason I wanted to talk about them is because for one, they have been in this space for quite a long time. And two, also with the recent release of Gemini 2 Robotics, they showed in the video a couple of the arms from Trusson being used there. So even Google uses them. And also they have something called the physical residency right now. Going from August twelfth to August twenty-eighth. So if you're in San Francisco, then you can see their new products because they have a new product lineup coming up, which we will get into. But maybe first let's talk about the company because They have an interesting history and I always liked these eighty percent or whatever before you know they got very popular and everybody talked about them. So they have their headquarters in Chicago, Illinois, and their founder and CEO is Matt Trossen. And he's also referred to as the chief nerd officer sometimes. So I really like that term. And what's interesting is he's largely self taught. He's an actual art school dropout. He worked as a programmer for a little while. And he has been building this company pretty much for the last 20 years, over 20 years now. Recently, there's also another person coming in who is Anish Shah, and he's now the chief operating officer. And he has a more traditional background, McKinsey and IBM, for example. But he just joined in 2026. Now in 2004, Matt Trossen began Troston Robotics basically as an online store. So around the time he had this idea of a plug-and-play USB PC connected. sensor or motor controllers so a programmer could control physical hardware from a PC with software without any deep electronic knowledge or whatnot. He had this idea but then he discovered okay there's actually a Canadian company that is pretty much doing exactly this. They started something like this a year earlier and it's called fidgets, so physical widgets. And then he contacted them and basically convinced them to send him free product. So $1,000 worth of product. And then he pretty much launched the US version of it. So Fidgets USA as the American store online reseller. From there on the store naturally expanded. And once people bought these fidgets interface boards, they obviously noticed that they need motors, they need wheels, they need brackets, they need chassis as well. And then Trossen pretty much began stocking these components from various suppliers. There were also early versions of robotic arms like the Lynx Motion. And he turned this little shop into a broader online robotic shop focused on hobbyist and education. And then later on started developing his own hardware. So basically his own robotic kits and robotic arms. It's especially got interesting around the last three to four years, because there was a Stanford project. And it's called the Stanford Aloa project. And Aloha stands for a low cost hardware. And they basically used four of the Trossen arms for their bi-manual machine learning and imitation learning setup. And this setup went kind of viral in research. Ultimately, it led to a widespread adoption of this setup. And then from there, the researchers basically reached out to Trossen and asked him essentially to build these kits for them. because obviously they don't want to worry about it. And then Trossen started reorienting this entire business heavily towards this market and this research education market. So he pretty much rebranded the lower kids into the Trossen AI line. What's also interesting, most parts are actually manufactured in the US for this. And if we look at the scale so far, because they have been in the game for so long, they served over ten thousand customers worldwide, four hundred universities, colleges. Over sixty Fortune five hundred organizations, six US national labs, I already mentioned Google, but also Meta, Nvidia, Tesla, Toyota, and maybe before we get into their products, if you hear about this Leo, how many employees do you think they have?

speaker-1: I don't know how to say, but I would guess at least two hundred.

speaker-0: Actually only ten to thirty seven. The exact number is not there, but it's a really small team. Yeah. Okay.

speaker-1: I wonder how they managed to deliver, you know, their hardware at scale.

speaker-0: Yeah. What's also interesting is they are completely bootstrapped essentially. So they're unfunded. There's one secondary mention of a very small debt, around sixty six thousand dollars, but it's not really prominent. Couldn't really verify that. So no evidence of large institutional funding rounds. They've also had some federal contracts, so a small army research lab order, for example. Now when it comes to revenue, obviously their company's private, so they're not listed anywhere. The estimate is Five million, but I think that's not that much and I think it's probably more, especially because of the recent physical AI boom and their wide array of products. And we're gonna go into their products right now. So I encourage you Leo as well to go to their website and maybe check out their products a bit while I'm talking about them. Because it's good to see them. So their core focus is essentially these complete kits that they have, and then you can also buy the arms or grippers separately. So the first kit they have is called the Solo AI. And the Solo is a single arm portable kit for teleoperation, useful if you don't need bimanu manipulation. And it goes for around $8,000. Then they have stationary AI, and this is a fixed bimanual workstation. So this uses two of their robotic arms. The robotic arm is called Window XAI, and it uses two of them and then multiple synchronized Intel Real Sense cameras. A touch screen, this is sixteen thousand dollars. Then they have mobile AI, and this is pretty much a wield version of this setup for more dynamic environments, and that one goes for $22,000. Now, of course, you can also buy the robotics ARM themselves, the window X AI ARM. This is a standalone six degrees of freedom arm. You usually get this leader follower configuration, but they also have a customizable base version. It has quasi-direct drive servos. Hardware gravity compensation, high control frequency at around 500 Hz, force and torque sensing elements, and more or less affordable at $3,000 to $3,500 per unit. Then they have the Trossen Yumi or Trumi, which is a portable handheld gripper system. So it's a Yumi UMI universal manipulation interface for data collection, but the data collection you do with them is more centered around end effectors. So folding clothes or other deformable tasks where you don't need this completely full arm rig. And they're actually pretty affordable. So the pairs start at $2,000 and they have some GoPros attached to it as well. So they sell these kids but obviously not all of the stuff that they are selling is manufactured by them. So they kind of combine different sources sometimes as well. For example, with the cameras, they don't build the cameras themselves. And yeah, with the Trumi, for example, this can be used without the Trossen robots and it outputs structured Czar or MCAP datasets. They also have a workstation. So if you want, you can buy a PC from them, a dedicated machine learning robotics workstation, optimized for CUDA and robotic operating system. And they also have some vehicles. So they have unmanned ground vehicles that drive around themselves. They have the Bunker Series, the Scout, the Sklate, and the Omnidirectional Ranger platforms. So you can put stuff on top of them. And they have also the Trossen Data Collection SDK. So a really well integrated ecosystem. Now, as I said in the beginning, they have the residency right now in San Francisco, where they also show you their new products. First one is Rivet, and Rivet is pretty much a complete mobile robot with two arms, the Window X arms and an omnidirectional wheel platform. Then we have the Glide, which is a lightweight human-operated leader arm. We have the cockpit, a movable, height adjustable operator station that holds one or two of these glide arms. We have a VR control interface so that people can control robots using virtual reality. We have a new 7 degrees of freedom arm. So the usual Window X had six, and this one is kind of an upgrade for this. They have another upgrade for the Window X for more. Heavier industrial use cases and they also created some new clouds and remote teleoperation software. And I think that's quite interesting to operate a robot from somewhere else through the cloud. So yeah, it's a really, really big product lineup. Hopefully not too confusing because it's essentially always the same stuff. You know, it's arms or grippers or, you know, of course, software, and then these kits which put the arms into a more broader product that you can use out of the box for research, pretty much.

speaker-1: ⁓ do you have an idea of their revenue?

speaker-0: I mean I said it it's private, it's a private company, so you wouldn't find the exact revenue. I saw some estimates around five million dollars, but as I said, I think it's pretty low. ⁓ yeah, so

speaker-1: It's not that big of a company. Do you have an idea of their margin?

speaker-0: If you look at their products, I think they don't have yet figured out the entire production workflow, the manufacturing workflow. It's more of a mix and match type of situation for them because they use GoPro cameras on their rig and you know it's not really that they produce everything end to end. But yeah, maybe that's why they also brought on the new chief operating officer with a McKinsey background to really get structure into this and I think that's really his job. to create timelines for the production workflow and really see where they can take. Because I think such a company they could easily raise a lot of money. Yeah.

speaker-1: ⁓ and scale of production. And that's also why, you know, the reason I'm asking about margin is because I actually looked into ⁓ robotics companies margins and compared them to software companies' margins. And it's actually pretty interesting because if you look at growth margin, the median software growth margin is around eighty percent. A bit lower if you blend in services with that. Because usually we tend to believe that a software company like Salesforce, they make money on software, but they actually make most of their money on services, which is consulting services. You pay for a Salesforce license per seat, usually. You also pay for Salesforce consultant to be at your company to make sure Salesforce runs what you want. So this usually happens for large software product for enterprise customers, right? Not for the twenty dollar a month Chat GPT subscriptions that you probably have. And the software and ⁓ programming industry as a whole reaches almost ninety percent gross margin, which is extremely high, right? So if you look at robotics, robotics as an industry is a bit less mature if you look at the ones we are, you know, talking the most about. ⁓ industrial robots is a mature industry, on the other hand. So the range of margin goes from a highly negative margin to 80% margin as well. But the range is much wider than it is in the software industry. So if you look at gross margin, software wins, you know, and it's not even close. But now if you look at operating margin, ⁓ software on average is only thirty-four percent. While if you look at robotics companies, the most successful ones actually can reach fifty percent operating margin. And we are going to get into which ones are the best ones, you know. But The main difference, if you had to remember only one line from that, is that software high gross margin is actually spent on customer acquisition down the line. So basically you sell your software with a pretty high margin, but you are going to spend that money into making ads and marketing in order to acquire new customer. Because that's where pretty much the battle is. Everybody is overpaying for ads, overpaying for marketing. And if you don't pay for that, you're pretty much not getting customers and not getting any sales, right? Robotics has lower gross margin because you have lots of hardware, heavy equipment to deal with. You basically cannot overcome that. So if we take into account, you know, ⁓ capital immobilization and the total amount you have to spend to gain and retain customers, which industry has the highest margin? That's pretty much the question I was asking myself. And what's interesting is that in the robotics industry, ⁓ I found three companies that have over 70% gross margin. So comparable to software. One of them is Kiyans, K-E-Y-E-N-C-E. It's a Chinese company. They have more than one trillion runs in sales yearly, which is pretty considerable. You have to divide that by approximately eight to get an approximate in ⁓ in dollars, for example. And their gross margin is eighty-three percent. So similar to software, right? And they make sensors, mostly for the robotics industry. Now if you look at Konyax, a company making machine vision device. So similar to you know the sensor space of Kylians, they have a gross margin of seventy-one point five percent. Then we get to, you know, my favorite, we mentioned them several times, intuitive surgical robot surgeon that is teleoperated by natural surgeon, not autonomous. Their gross margin is seventy percent and the operating margin is forty two percent. So it's like ten points higher than the average software business. Because you know the recurring revenue from licenses and consumables is so high for them, right? And another company called Horizon Robotics has 66% gross margin and they make chips for robotics applications. Basically you can see that right now the industry is highly centralized onto industry applications of robotics and picks and shovel businesses in the robotics space are where the higher margin are. They highly depend on whether you sell your own thing or whether you sell someone else's hardware that you revamp. If you look at companies that pretty much deliver actual robots, margin is a bit lower. For example, Symbodic, they make warehouse automation devices. Gross margin is twenty two point two percent. A B B robotics, they make industrial robots, arms, and conveyors. The operational margin is twelve percent. It means that gross margin will be a bit higher, but I did not get the number for gross margin here. Fanuk, I guess the most famous robotics arm company, they have A twenty-one point one operating margin and a twelve percent market share. Even if you're not into finance, these kind of numbers is still an extremely strong number. If you have twenty percent of a market and over twenty percent operating margin, is it's a pretty strong company that you have here. Serve Robotics, which makes delivery robots, you know the little robots with wheels and with a face that runs mostly in the US, I believe. Their gross margin is minus two hundred and seventy one percent. So they have a pretty negative margin. But I I guess they also behave as a startup. So they have a growth at all cost kind of strategy here. And now we come to the fancy company that ⁓ you know, shiny object company that interests us all, which is Unitree, because all the companies I talk about, they are established industrial robotics applications company. Unitree is the first humanoid physical AI next generation company that we have. Unitree

speaker-0: Yeah.

speaker-1: is pricing its Shanghai HPO at approximately a nine billion dollar valuation. And last time I checked it was over 8,000 times oversubscribed. Just to give you an idea of how the market is excited by this space. The revenue is $135 million. Gross margin went from mid forty percent to approximately 60%. Right. So we are talking about a gross margin that is comparable to intuitive surgical or companies like that. Right. So this is ⁓ pretty interesting to have this kind of number. ⁓ Unitry is also a profitable company, their net margin is fifteen percent, which is again pretty good comparable with you know the net margin of Fanok, for example. And to give an additional bit of perspective, a public SaaS company usually runs at a loss of between eight to fourteen percent net loss because it takes like decades usually for a SaaS company to be profitable because the competition for customer acquisition is just way too strong. Most SaaS companies, even as startups, they lose money even after CVC and frequently through IPO, right? So what's interesting is that with the AI wave with ChatGPT and cloud coming in, the market has already repriced that. Approximately one trillion in enterprise software market value disappeared in one week. in early February twenty twenty six, right? Because the number of seats these companies can sell, you know, is decreasing because of AI agent automation mostly. In the same period of time, unitary gross margin went from forty to sixty percent again. So I think it's interesting to see where the market is real pricing it right now. Also because obviously robotics is harder to scale because it has interaction in the real world. But once you are established customer acquisition is cheaper. And customer retention also is cheaper. So the next question I was asking myself is is there a bubble in the robotic space? That's pretty much what everybody is asking. And if you look at funding in 2025, robotic startups they raised 15 billion and the on the over the entire year of 2025. On the first semester of 2026, they raised 18.8 billion. So we are on track to get double of 2025 when it comes to robotics funding. Largest rounds were Saronic for Autonomous vessels, neurorobotics, skilled AI that you have probably seen on Twitter recently, atronic, mine robotics, just to name a few of this year. Over, you know, the past five years, only forty two robotics companies have raised thirty plus millions. Do you know how many software companies have raised more than thirty millions? Seven hundred and forty five actually. If anything, robotics is underfunded.

speaker-0: Right. Yeah. I mean of course it's how word generally needs more capital. Yeah.

speaker-1: It's not always about hardware. So for example, if you look at maybe we are overly influenced by a few key players, you know, in the space that occupies the entire media scene. For example, Figure AI has a 39 billion dollar valuation and they don't disclose revenue. They probably don't have much, to be honest. No offense to anyone. Maybe this is another valuation or you know, physical intelligence that we also mentioned in this podcast, they are currently reportedly Trying to raise at an 11 billion valuation. And again, zero revenue and they only make software, they only make models. So maybe are this company overvalued? I don't know. It's an open question. Because for example, physical intelligence, if they manage to deliver a model that has cross-embodiment and that can, you know, deploy, that can be deployed zero shot to do pretty much any kind of task, then the upside is not even measurable. The last bit on this topic is the new in fashion term that is called RAS for robot as a service. And right now, robot as a service is thirty five to forty percent of commercial robot deployments and fifty two percent of industrial warehouse deployments. So when you see you know warehouse automation devices, usually you you pay for it, obviously, but it also comes with a license and with a yearly plan to you know to take care of it. Just like if you have an intuitive surgical surgery robot, you have to remain in touch with intuitive surgical in order to take care of it every year and also for the consumable parts. So basically that turns CapEx into recurring venue. And I think this is the robotics industry's way of providing services. You know, Salesforce engineer is provided as a consultant to the Salesforce customer and if you buy a robot from, I don't know, intuitive surgical, you have to have, you know, the person who takes care of, you know, the robots and does the maintenance as well on a daily basis. And I think that's where robot as a service kind of ⁓ comes in as a business model. So I think it's an interesting business model to follow.

speaker-0: think you mentioned it when talking about Rec, the robot fighting company. Yes. How they got really good at fixing their robots because they constantly were breaking. You can think about okay, what are the primary businesses? Of course those are the robot companies. But you can also think about okay, what are the secondary businesses? Okay, robot fixing. I can imagine in a couple of years people having different outward appearances of their robots, right? You will have Tuned robots, maybe, you know, robots that can run super fast, robots that are made to jump super high, or that look like almost like a fashion icon or something like that. Especially if we go into the robot girlfriend, robot boyfriend type of situation, which we probably cover at some point. And I'm also interested in how these skill stores evolve, right? So will it really be that you get a really complex skill? Let's say you have a company and your company does a really precise thing. Okay, something that intuitive surgical can only do right now and and suddenly there's a unitary robot who can do this kind of surgery. Of course you're willing to pay a two thousand, four thousand, ten thousand dollar retainer every single month to have this skill running. But at the same time I'm thinking, Okay, once there's a skill, maybe it's gonna leak and you can just put it on your ⁓ on your robot anyway.

speaker-1: Pretty bearish on skills and skill store and everything. I don't think this is the future proof ID at all.

speaker-0: I mean I can at least tell you one company that is not that is not bearish on it. Or is it even a company? It's an open source project and that is Azimov open source robot. So have you heard of it?

speaker-1: Yeah I have. I've seen it on Twitter.

speaker-0: Yeah, of course. It's pretty popular. It looks super cool. It's like this orange robot with a nice design. And it's open source. So it's an open source reference humanoid robot platform developed by Menlo Research. So Menlo Research is an applied research and development lab based in Singapore. I think it's the only open source platform. We saw one before with a NVIDIA ISEC Route, but the Nvidia ISEC root uses a unitree, right? And it uses a combination of different hardware and software. And it stacks it together. But what they are going for is actually a true open source platform from the ground up. And Menlo Research was founded in 2023. The founders are Nicole Chu, and she is previously a VP of fraud risk at Gojek, and she's a Stanford Alum. And then we have Daniel Ong, who is previously a CTO of Dana Cita, and he's also a Stanford Alum. They have a team of roughly 25 to 30 people. With a small dedicated Asimov hardware team. They completely bootstrap and employee owned, and they have stated that they have not raised any traditional venture funding and they distribute their equity to team members and contributors. Now Manlo's broader mission is solving the robotics data problem through an open ecosystem. So shared APIs, shared skills, and also data generation by many different developers, rather than a single closed model. And this makes sense to some degree. Before the robot, they were working on basically an open version of ChatGPT called Young. But for some reason they try to work on a robot now. I don't know what happened. And if they still do the local app, but it seems their robot is the main project right now. They have started with Asimov Zero, which is open sourced as of January. And this is the legs only version. So it's a bipedal, legs only, completely focused on locomotion. And we've seen it with generative bionics as well. With the Gene One, they had the legs only, and then later on they built on top of these legs and then created basically the entire robot. They said they built the legs in under 100 days with less than $30,000 of RD. And then they have the Asimov one. And that is kind of like the flagship, complete humanoid, open sourced since April. First units came in July 2026. Delivery starting in August. The specs are quite interesting because it's Not so tall. It's just one meter and twenty tall, weighs around 35 kilograms, 25 plus 2 degrees of freedom. But it has quite strong actuators, torque figures that are competitive, so it actually can lift meaningful loads, for example. For the sensing, it has a monocular camera, microphone, speakers. So it's designed to run a VLA, for example, or some model that needs perception. The price tag is twenty thousand dollars for the complete kit delivered. But what's interesting, you can also fully self-source it. But then they say the estimate for that is around thirty thousand dollars. And on their website, they have the bill of materials. So you can see pretty much every single component that this robot uses and you can source it yourself, and then you can build the robot yourself. And I actually requested the bill of materials and I look through it and you can basically see everything that this robot is made of, you know. every single component. So it's pretty interesting if you really want to see what are the exact parts of a humanoid. If you want to pre-order it, it's $499. And even if you are experienced, you can expect a build time of 100 plus hours. So yeah, open source comes at a price, but you also got the right to repair and you also got the right to customize so you can truly do everything you want with this robot. And it also comes with some basic locomotion and control policies. But essentially you can use it to run your own models and agents as you wish. Now there's one thing though, and that is probably a downside for you, Leo. The robot doesn't come with hands. It has no hands. So they said they probably gonna release a version at some point with hands, of course, but for now they kind of focus on the rest. So I think you can put your own hands on.

speaker-1: Does it come with grippers or yeah you are expected to provide the hands?

speaker-0: I think so, yes. Okay. And what's funny, I even saw a tweet of a guy and he said like, Why should I not get the one X? ⁓ the one X has so advanced hands and it's only twenty thousand dollars. And they just responded and said, Yes, you're right. I think you should get the one X because we are more going for the open source vision. And they didn't even mean it ironically or something. They they just said like, Okay, if you want a finished product, then you'll go with that. But if you want the open source one, we want to customize Build yourself as you wish, then you go with the Azimov.

speaker-1: Yeah, and I'm guessing the one X come with, you know, the teleoperator who can see through your home as well. So we'll have to be comfortable with that.

speaker-0: Well, only if you do like the expert training, I think. So you can't have the one X without that. But but yeah. Asimov also comes with Asimov OS, so their proprietary operating system that you can basically build on top. They have a digital Asimov. So digital twin, you can see it on the website, you can go there. It's a browser-based simulation for testing Asimov without hardware, which is I think super cool. And they have a really big vision on this. API and skills marketplace. So they want composable skills, just like we see with Unitree, or I think AGBot has it as well. And a bunch of other different companies have these skill stores. So they want to build something around this. Menlo store is kind of like the word for it. And this obviously raises the question, okay, where we draw the line between open source and non-open source. Because at the end of the day, I already see in this project there are a lot of potential parts that you can monetize, right? But that's another debate. Find their repos GitHub, so you can go there, you can see pretty much everything. They also have a really well organized doc section. So I think they're doing a good job in many different areas here. And also what's interesting is that they have a manufacturing network. So they don't really produce the robot themselves. They have a network of global partners. So yeah, open source is always a good starting point, but as we've seen it in the past, a lot of open source companies they gradually transition to a a for profit company. For the future, they're already speaking about the Asimov 2. So this is supposed to be coming soon. It's a deployment ready bi-manual robot for real work. And for their long-term goal, they want to have a state-of-the-art robot by Asimov 5. So they really want to be competing with all the major robotics companies out there. If you want to check them out, you can of course go on their GitHub, join their community. But I think it's a really interesting project.

speaker-1: And I'm thinking about hands, of course. Open source means you have to need open source hands as well. If they manage to pull that off, that would be very important.

speaker-0: Yes. Maybe we can have them on this podcast at some point because I really want to ask them what is the hand situation, have they how they're thinking about it, how they're approaching it. I mean, we discussed it before. It's like a huge chunk of the R and D goes into hands. So

speaker-1: A huge chunk of the price too. So maybe the price of the robots with hands might double, you know, literally.

speaker-0: Who knows? But I I really like their design and I also like their name Asimov. You know Asimov? Yeah. Yes, and he has those three rules.

speaker-1: I do. He has a book. three rules of robotics. Yeah. I know one like robots should not hurt human or something, but that's about all I remember.

speaker-0: Do you know them? Yeah. The second one is a robot must obey humans unless that conflicts with the first law. And the third one is a robot must protect itself unless that conflicts with the first two laws. Do you agree with that?

speaker-1: I'm not sure that's really relevant anymore. Where does that put wreck making robots fight each other, you know?

speaker-0: I mean. That's true, yeah. We leave it up to debate. If you guys have your ideas on the robotic laws, then put them in the comments, of course. Make sure you subscribe to our channel. Thank you so much for watching. This was the Robotics Tech Podcast, episode 13. My name is Oscar. Bye bye.

speaker-1: Hey, my name is Leo.