speaker-0: Hello, welcome to the Robotics Stack Podcast. Today we're joined by Matthew Moore. He's the CPO of Menlo Research, the company behind Asimov, the open source robot. We're gonna talk about the robot itself, of course, the development, Matthew's journey, and what the future holds for Menlo and Robotics. This is episode 19. My name is Oscar. Hey, my name is Leo. Matthew, thank you very much for being here with us today. Thanks for having me. First we are going to get into your background and your journey to joining Mandalore Research and building Azimov of course. But first I would like to ask you a quick question that will kind of be the underlining of our entire episode. Why do you think an open source humanoid is what the future of robotics need?
speaker-1: All right. I'm gonna preface that with I think open robotics is what the future needs. It's not just a humanoid, but let's talk about a humanoid. We all want robots in our house. Everybody has that one little step in their house, if not a full s set of stairs. And so eventually you're gonna need legs. Today, today, you know, battery performance and things like that probably need to improve a lot before We start seeing these in our house, but I think open is really what we need to raise all the tides, if you will. It's just such a hard, hard space to build in. Companies need maybe hundreds of millions or billion dollars. And I don't think that's right for humanity. We want to lower that so that everybody can get it.
speaker-0: I can only agree. Maybe you can tell us a little bit about your background, what you've done before, and how did this lead you to work with Azimorph and Menu Research today?
speaker-1: Absolutely. So I by trade am a software person. I started to get into machine learning, sort of classical machine learning for some of the startups I did. And in twenty fourteen, I thought, you know, Alex Net is coming out. Computer vision is really starting to be able to do things that are really, really impressive. What applications do I care about that computer vision is really good at? Well, it turned out everything I wanted in the home had to do with manipulation. And so, okay, that meant robotics, right? If I cared about computer vision, that meant robotics. And so the first part of my journey was working with a mechatronics person that I found to build a toilet cleaning robot. I didn't have kids at the time. And that was like the worst thing in my day to day at my house that I could imagine. Doing that I wanted a robot to do. And so went down the path, got kind of distracted by data labeling. And you know, having to label data, particularly computer vision to recognize toilets and dirty toilets and all that stuff. So I went down a one ball in synthetic data. But in twenty nineteen I had a kid and he had food allergies. And so my wife who works and I work, we both have to start cooking like all the meals all the time. And it was just so tiring. And I thought, hey, I think I can bring the stuff I learned about robotics into actually cooking meals like just four different meals a week. That's it, just the staple. And so we start to build this sort of like a 3D printer type robot. It had an arm on a gantry, it had burners, it could tend, you know, it could extrude pasta, boil the pasta, boil down sauce. So this is like 2022 now, 2023. And I as a startup founder really want to focus on the application. What recipes can I bring to the table that are important for parents? Can I partner with grocery like delivery services to package these things up? But here's what happened. Anybody building robots hopefully can relate to this. The first thing that happened was that our open source motor controller decided to close source the firmware because they were getting copied by, you know, others, and that was their main way to make money. And so now all of a sudden we couldn't change the firmware to operate a little bit better with Ross. And sorry, there's a little bit of a long diatribe, but it really helps elucidate why I think Menlo is so important. Another thing that happened was that we needed to change platforms from an Intel-based chip to then a we wanted to go to a Jetson or Nano. All of a sudden, the way the CAN drivers worked were completely different. And then the other thing we kept doing was we are on ROS2, you know, on this little side project. And every subsequent update of ROS2. completely changed how some of the action servers worked and how the path planning worked. And so inevitably every time we made this little change here or there all across the stack, we'd have to like completely reinvent the wheel all the way at the bottom, the bottom of the bottom of the stack. And all I wanted to do was tend a stove and control how hot it got, basically. I end up getting introduced to Daniel and Nicole, the founders of Menlo, and they have this vision of this open robot where everything from even the motor controller to the firmware to the RPU which we call the robotics processing unit is open and the higher level path planning stack is all open. And boy, I was thinking if I had that to start this startup all over again, I could do it in, you know, less than a year instead of a few. So I apologize. That was a really long way to say that. This is just super important for my next journey as well, whenever that happens to be. And I'm excited to be a part of helping build this.
speaker-0: I think it's a perfect introduction and it completely makes sense between your story and the leak to Asimov. And it's interesting you mentioned cooking because yeah, I've seen some content about interest on cooking robots. There is activity in this space.
speaker-1: I have a lot of thoughts on cooking robots.
speaker-0: Why ⁓ would you be dissatisfied with the current cooking robot, all of these advanced cooking aids that can do pretty much anything? So
speaker-1: So, yeah, for a parent working, like especially two parents working, you have this desire to get as fresh and as raw ingredients as possible. You want to take literally flour and eggs and milk, right? And not start with like pre-dried pasta, for example, where you're not exactly sure in the factories where they're making that stuff, what went into it. I'm really a huge fan of like pascha and some of these other cooking robots that are making their first steps, but they're very limited in terms of what. they can cook because they don't have an arm that tends that you have to preload these pre-chopped ingredients and you as the human are doing like the lowest of low level labor and then it's doing the fun part to me, which is over the heat. I I want something in reverse where I'd like I don't have to do the prep and then I just throw it on the fire.
speaker-0: I mean cooking is probably one of the most advanced tasks. You must have gained a lot of experience, maybe also through a bit of frustration, but ultimately it guided you to the right place. And now you are at Menlo. What was like the initial impression when you got there and you were looking at all of this and like what these people are doing? And you were right in
speaker-1: Yeah. You know, my initial impression, especially being in the US, my initial impression was wow, they are shipping robots. I think that's such an undervalued, underappreciated thing. I went to Actuate. Actuate is a a big industry event in robotics. A lot of folks are there. This was in August of 2026. The only humanoid robots, by and large, were from China. And I think that says something, right? There were like one robot. From the US and it was the stretch robot, which I love them, but it's one form factor. You know, there's a whole bunch of form factors we should be seeing. And we really need to sort of lift up the rest of the world. China has such a head start in it. You know, I was like, Wow, Menlo's actually shipping robots to people. This is such a leg up. Let's take that and get further ahead and lift the whole ecosystem.
speaker-0: And how because we both have been following Menlo for quite a bit. And maybe you got some insights into how did this development happen? Because they started with the legs. I mean we've seen it with other companies as well, usually starting with the legs for locomotion, which was Asimov Zero, then Asimov One, which we obviously can get into, and maybe even Asimov two, you know, you're kinda teasing that. But maybe you can take us a little bit through this like development.
speaker-1: Yeah. I'm even gonna rewind further back because I think it's important part of the journey. So what happened is that Menlo existed three years ago, but didn't start in robotics until like just under a year ago. Three years ago, the team was working on essentially brains for robotics. You know, they trained their own LLMs, VLAs, now world models. And what happened is that about a year ago, they're like, We want to control robots with this. Let's go out and buy somebody's robot. And they look for robots they could buy and basically only could buy a Chinese robot. Nothing wrong with that, but it's not open, right? As soon as you go and buy Unitry or anything like that, if you want access to the raw sensor readings, well, you gotta like upgrade to the SEK. If you want any warranty coverage, well, you gotta pay for the warranty coverage. And by the time they bought it, it was in the six figures. And first of all, not everybody can afford six figures. There's that. But second, These things break like you're kind of abusing them when you're running these policies, right? And so they start to run their policies on this Chinese robot and it broke. And so what do you do? Well, ideally you could fix it yourself. Sorry, you avoid the warranty if you do that. Okay. So now your backup choice is to send it to Unitree or one of the other providers. And it is two months of downtime if you do that because you send it and you get one back. It is not like A hundred thousand dollar car, for example, where you get a loaner. It is you send it back, you're down. And we talked to other researchers from other universities and they had the same situation. Like their research would literally stop for three months and they would just be out and down unless they had some other solution. And so they started down this journey of after they bought it and were down, what if we could make our own and actually have it be repairable? That was kind of the premise. Like right to repair, these things are gonna break. Let's make it so easy and so well documented. They basically started this journey because they wanted to essentially use their own product, which I also think is a great way to build. One thing I didn't touch on in my introduction is I started, I think, like four or five companies now. The most efficient way to build a company is to scratch your own itch to solve your own problems. And that's exactly what what Menlo did. So I was even more attracted because of that. So
speaker-0: Did they just have the robotics expertise? I mean, when it comes to the hardware, I've seen the other product which they have, which was this kind of open LLM pretty much, right? Yes. I mean there's some overlap, but did they have some internal robotics knowledge already or did they hire people or how did that come?
speaker-1: Yeah, great question. So I think on the what we call the brain side or the intelligence side, pure software side, right? We definitely had expertise. Where we lacked expertise was on the hardware side. So they were able to find people. Menlo's really well situated in Singapore. It's sort of this melting pot of a lot of talent. So they were able to find some really key people that had experience from Tesla, had experience from Amazon Robotics and Designing a humanoid actually is like just the design is maybe half the problem. The other half of the problem that is harder to appreciate without being in it is the supply chain that goes into it. Finding basically actuators that are readily available that are easy to sim is actually just a really big issue in and of itself. There's also the batteries and all this stuff, right? Like if they aren't kind of made for humanoids already, you're gonna have a heck of a time actually producing this at scale and then. You need to throw in a bunch of money just to like spin up manufacturing. So I think a big part of the key insights that they had actually was that they needed to be really up to speed on the supply chain as much as the design. But let me double click a little bit on actuators. One thing that I think you also don't appreciate necessarily on a humanoid, as a humanoid or anything with a lot of actuators, you want zero shot sim. You want to train everything in simulation, but then be able to turn around and deploy that on the robot and have the robot act. Pretty much as similarly as possible as in simulation. And that actually goes all the way back down to how the actuator performs. If, for example, You have an actuator that performs nonlinearly. That gets very hard to send because it can't just be like this simple joint that you can model. You know, you have to like go in and actually create all the physics essentially around it and how it performs. So even the selection of the actuators is like it seems like it's all a hardware decision, electromechanical decision. It is an AI decision. Like the AI team has to have some feedback. into what hardware you're gonna choose, which is kind of unprecedented, right? It's not like that in the car industry as much from what I understand. So we're very careful and because we had this sort of brain first motion in the hardware design, we knew what to aim for and could kind of like intersect all these different competing priorities. I see.
speaker-0: Yeah, maybe just to get everybody involved from our audience, what is the product right now? Asimov one just started shipping the units. Really nice moment. Maybe give us a little rundown of what is the product itself and what is the current state of development pretty much. I have
speaker-1: Absolutely. So the current state is it's called Asimov One. It is a humanoid robot. I mean in the US we'd say it's about seventy pounds. It has a hip you know payload of about thirty pounds, which is pretty beefy. And ⁓ boy, I'll have to look back at exactly how tall it is. I think it's about five feet tall or something along those lines. It ships in a kit, actually. And you get all the actuators, all the 3D printed plastic, all the 3D printed metal. the C and C metal, and we bring you through really smoothly how to assemble it. I think today, today, it's a very research lab and corporate research oriented product. I mean, or if you just want to be walking around with this thing, you can do that too. I call it like entertainment and corporate and academic research oriented product. But we are shipping. We're shipping worldwide, which is great. I mean we have so many different countries that have ordered this and I think that's what really makes my heart happy is that it's not just like the US or it's not just China that is getting to play in this game now. We're really making it easy for folks all over the world to like be able to experiment with legged locomotion, local manipulation, and so on and so forth. And then just to give you a little bit of a sneak peek on Asimov One, it doesn't actually ship with gripers quite yet. We're working very hard on getting the grippers out so that you can really start to do things. Actual tasks. That's why I say today, today is more of a research oriented platform. I think once it gets grippers, you'll be able to do a lot more with it.
speaker-0: I see. I have so many questions. So first of all about act traders. I had no idea about what you said that actuators are hard to simulate. I mean that's very insightful because I would have thought that with ⁓ you know computer assisted design and so on, you can recreate the pretty much the entire Act Trador into solid works or equivalent pieces of software. So I would have imagined that it was easy to simulate them. We had a guest earlier in the podcast whose name is Bryce Yancioto and he's a young American engineer. designing his own new Actuator, which he calls the Orbit Actuator.
speaker-1: I could go off on a whole tangent, sorry, about US made actuators, which I think is a great thing for the US audience.
speaker-0: So it's basically three actuators stacked in the limb and then you get the three degree of motion over like half a sphere. So the device can act like as a shoulder or hip joint, basically. That's great. Yeah, kind of. Do you think this will solve kind of the simulation problem or not, or make it even harder? What's your take on that?
speaker-1: Yeah, it's more like a ball joint. So for us it depended a lot on how the gears behaved. If it was a pretty simple gear mechanism like planetaries for the classic example that you see, it became very easy to simulate. Obviously, there are some downside of planetary gears, but the fact that you can simulate it so easily, you know, you're you need GPUs to like train this stuff already with simple planetary gears. Like just imagine if you start getting nonlinearity across you know, twenty-seven actuators, like how much that blows up the space in in training. So I would have to look at it and our team would more accurately about whether they they are very like linear style performance on the actuators. Harmonics are really where things start to go a little bit nonlinear, for example, and that ends up being really much more computationally expensive to simulate.
speaker-0: I see. And so talking about American icers, are you working with American manufacturer? Are you manufacturing your own iThriters yourself? Are you using external supplier? How much can you tell us about this?
speaker-1: Yeah. I would need to double check if we can talk about which supplier we use for actuators. I'm happy to like put it in the show notes if we can, but we know we need to, especially once we start pre-assembling these. Right now it's a kit, right? But what people want is a pre-assembled robot. I mean, fewer people are excited to like screw in all the nuts and bolts. So we are actively seeking out partners that can help us assemble these in the US. And then that's even like one step removed from getting actuators made in the US, right? So like even just getting all those parts together and and assembling it in the US is like one thing we are working actively on to go one step deeper. To get actuators made in the US at a medium volume is actually not something that exists very well in the US. So we're actively also looking at that. We're very excited to do that, particularly me. I'm sort of the token American right now on the on the team and I'm opening up the US office and so we're creating a US company and all that stuff. So very excited to to do that. Yeah. And looking for partners. So if anybody's working on it, please reach out to me. We're actively seeking you out. What we?
speaker-0: I actually talked recently with Randolph Briggs, who is building Steelboard, and his vision is to build an open source humanoid a hundred percent made in the US. And he mentioned that part of his vision is actually making acters end to end on US soil with exclusively US sourced components.
speaker-1: So if you like really go into the wormhole, today today the hardest part is getting the magnets made in the US. You can wind the coils, you know, you can manufacture the gears, you can make the PCBs for the encod and get the encoders. But yeah, those magnets like are not made here. And so I hope we can spin that up or they're also alternative designs that that don't use the same kind of magnets. So we can also explore that.
speaker-0: Yes, it's definitely something I've encountered before, the famous neodium magnet bottleneck that always ends up in China.
speaker-1: And like really good for China for like figuring that out, right? Like I think we need to appreciate the amazing stuff that they've done.
speaker-0: Maybe just quickly, is that also because you said you are now pretty much creating a new company in the United States? A response to this trade ban situation?
speaker-1: Yeah, yeah, absolutely. So it's to be determined, just to be open and honest, that whether we are a Singapore owned and US subsidiary or vice versa. It's not like we've fully decided on that. To do anything in the US, you need a US company. So we will always have a Singapore operation and a US operation. Manufacturing specifically in the US, I think is a big part, to be honest, because with some of the trade bans that are coming through. I'm from here, so I want to see our country be able to manufacture things again, especially at a medium scale, right? But yeah, we have doubled down on our focus on manufacturing specifically because of some of these bans, particularly in advanced robotic systems. But being the US person, let me actually talk about why Singapore is so interesting. Singapore, you know, it's in the US it's actually very hard to get external talent here. I'll say that. It is very hard. Singapore is this wonderful melting pot of a lot of talent. And their government also has a lot of grants and basically a lot of funding for bringing in such talent and spinning up companies. You know, we think of the US as being a great place to start up companies. Singapore, for startups particularly, for advanced robotics and physical AI, particularly, it's a great place to start as well because you can get a lot of the salaries like partially covered as a startup when you don't have a lot of funding to begin with. And so we're just getting this immense talent that we see as a competitive advantage for us that if you know we really only operate in the US, we'd be competing with anthropic and open AI. And by and large, guess who's going to win that war? ⁓ David or Goliath. You know, I mean it's a different situation for us being a, you know, today today a pre seed startup. That will change over time as we mature, as we raise different rounds of funding and as we build more stuff in the open compared to some of the other of our competitors. I think it gets more and more attractive as a potential employee to like work with an open system. And so that's one of our competitive advantages that will help us going forward and maybe we'll change the equation of whether we're a US parent or a Singapore parent down the line. Makes sense.
speaker-0: Back to Asimov as a product. What is the compute on it? And then what is your training approach right now? Because I just saw you were publishing your locomotion policy just recently. Do you train mostly in simulation? What is your approach?
speaker-1: Yeah, let's start with locomotion training. I mean, it's basically all sim. You know, actually, we should be in Isaac Lab relatively soon as one of the built-in robots. I hope I'm speaking correctly. We're waiting on NVIDIA on a couple things. We also have our own training platform. Part of the great thing about Menlo that we don't always talk about is we do have a data center of our own ⁓ that we own. We do have, you know, this open source VLA and LM. training platform on top of that data center. And increasingly we're going to be moving towards a harness for like basically a gentic robotic control. And so having that full stack lets us play, you know, optimize the product at all levels to like do the end goal, which is accomplish some tasks. But to come back to training, you train it on these NVIDIA chips essentially, and then zero shot that into real. So far it's working well. We have These two levels actually, it was interesting for me to learn. Most people try to make everything work on a Jetson or something like that. Like everything from canvas, like really low-level messages all the way up to the like high level step. We took a slightly different approach, which has worked out well for us, which is the very low level, is on a eight-core ARM chip, actually. So that is like what I would call maybe the lizard brain. It's CPU only, you know, it makes sure that if you're Pulling on the robot, it can stabilize and things like that. It's sort of the core locomotion policy. Then as a developer, a user on the platform, you get to decide what the higher order brain on it is to be. It ships with a Raspberry Pi just to keep costs down. You can stream video to it. Obviously, if you want to do anything really intelligent with it, you're gonna be going to the cloud and using some cloud GPU, or you're gonna be finding a Jetson of some level, depending on your application to put on it. But Today, today it's Raspberry Pi five and we're looking at what a Jetson might look like on it as well as an upgrade. But it's open and you can mod it all yourself, right? Like all the software and everything's open, so
speaker-0: Yeah. You mentioned locomotion policies, VLA. Are these all open? Do you ship them with the robots? Do you put as much resources into researching those as you do to build the robots? What are your priorities right now? Yes.
speaker-1: We are opening up everything that we have. And you know, we're cleaning up the GitHub repos, I think, literally as we speak, just to like make things much more clear. Maybe even by the time you see this episode, that'll all be online and available. So out of the box, you're gonna be able to control it from your like mobile phone on a mobile web app. You're gonna be able to obviously do the same thing on a computer. You're gonna be able to train the locomotion policies yourself. put it, you know, flash it onto the robot. So we're we're a small team. I think we're about thirty people right now. And we're spread very evenly. Portion of those are even data center folks. A portion of those are on our agency harness for the robot. A portion of those are on locomotion and robotic control. A portion of those are on electromechanical design and then even a portion of those are on manufacturing and supply chain and delivery. You know, if you divide 30 by the number of teams I just mentioned. It's not very many people per team. I think I listed all the teams. That's pretty transparent ⁓ view into it and everybody's divided fairly evenly at the moment. But we have a decent number of generalists that also can go back and forth.
speaker-0: I really like the design of Ozimov. It's really bad.
speaker-1: Yes. Thanks. I think again as a small company, it was important for us and continues to be important for us to have even if it's like a strictly mechanical person, make sure that they can wear a few hats simultaneously just because we are short on resources. And so they have to be a little bit of an industrial designer, right? They have to think that way. They have to understand a bit about even through to the AI, at least appreciate some of that stuff. I'd say one of the main strengths of our team is that You know, it's not just a hyper specialization. There's a little bit of bleed back and forth and I think we care about how it looks, certainly. So, you know, we'll bring in even the part time resources here and there at the appropriate times because we do care.
speaker-0: Yeah. So my next question is about how was the company created and funded? Because from what we have been able to read online, the company has been bootstrapped. It also comes from Singapore, which is not the cheapest place in the world by far. Thirty people is not nothing. You have also a data center. So I'm curious, I have a feeling that maybe the data center came first and ⁓ the company was built on top of that. So I'm curious where does the funding come from and how do you think about funding for the
speaker-2: Yeah
speaker-1: So Daniel and Nicole, our founders, are also our benefactors. They have seen this company through everything and very much believe in the company. Again, another thing I look for when I'm joining a company is how committed are the founders. And they are the most committed out of anybody, right? It's their livelihood and everything. And nothing has been cheap about this journey. I'll just say that. It's been a research company. So the things that they made. at the very, very beginning didn't necessarily look like robot brain, right? There are some other iterations really, really early on. It's always been AI related, but really early on, might have been turned into a creative company, more creative AI side of things. So yeah, data center came because of that. And we started just getting all of these assets, which you can really only connect the dots looking backwards, as one of my favorite people has said.
speaker-0: About the data center, can you give us a little bit, a few details about the scale? Like is it GPU, is it CPU, is it NVIDIA, is it not?
speaker-1: It's all NVIDIA for sure. And you know, I'd have to look at exactly which chips it is. I don't happen to know that offhand. But I think the other thing we can bring to the table, particularly in our joint research with academia, is that you know, it's hard for academia oftentimes to get access to spare GPU cycles. And so part of what we're doing is is bringing some of that to the table. And that makes it really excited. We can work with institutions, you know, with names like Stanford and other big ones because of that and we'll take advantage of that ability and asset that we have and try and push robotics forward. Nice.
speaker-0: Yeah. What I ⁓ perceive when I see you, even the little community talks that you have, really nicely designed, the videos that you do are really well made. The docs I've seen they've been updated. It is done with a lot of care. And I think like that is my impression. There's like a really strong vision that is pushing through that. Since you mentioned the two founders, I truly believe that they're committed with all they got there. What is also interesting to me is ⁓ maybe back to Who are the type of customers right now and also the applications that you see? Of course, I mean if you look at the general robotics market, it's mostly research, but then also how many people assemble the robot right now and how many people buy the pre assembled one, for example, maybe some numbers there.
speaker-1: So I like to bucket our customers into three main groups. It's a little bit actually mirror of what you saw Unitries mention about their humanoids and their filings. About a third of it is to what I would actually classify as like individuals and entertainment. People who want to have a robot, bring it around, just live in the future, right? Like if you want to invent the future, the best way to do that is to live in it. People who can afford to, you know, spend twenty thousand dollars, which is not nothing, to do that. Two thirds of it actually is split pretty evenly between academic research and corporate research. But even a part of that, like I'd say maybe a third of that is actually product development within a corporate research arm. So it's not just corporate research for the sake of corporate research, like how do we create models that we could then, you know, put on hugging face or something like that. A third of it is actually how do we get and deploy robots? And the reason that that is so interesting to me is because I think particularly in the last couple of months, like legged robots have gotten kind of a bad rap in terms of their usefulness. And in some ways I agree, like, yeah, having two legs is kind of a weakness in our current state of battery life and things like that. But mobile two arms with perception, very important across the board. So some of our customers are very interested in like half of Asimov, the upper half, but we're modular. I think that's one of the other main things I want to get across. Like our future of a company, it's really open modular pieces, both on the hardware side and the firmware side, all the way up to to the brain side that can lift all boats, right? Like raise the tide and lift all boats and make it so you don't need a billion dollars to start a robotics company. We're gonna help with the actuator. We're gonna help with a motor controller and we're gonna help with a robotic processing unit. We're gonna help with training. and have modules around that. We're gonna help with data collection, not as a service, but like stuff you need to do data collection and so on and so forth. It's really exciting to me that even some of the deployments need legs, some of the deployments don't need legs, but I think we'll probably be very oriented towards bimanual manipulation on some kind of mobile platform. I think that's really where our strengths are.
speaker-0: Interesting. And speaking of data collection, you also have the Manlo store, right? And we've seen it also with other companies, Unitree. There are a lot of these data collectionslash skill stores. What kind of role does it play right now?
speaker-1: Yeah. You know, we're still in the early days of forming that. I'm gonna give you more my take on it. Right. And this has not been formalized, so I don't want you guys to take it as sort of Menlo's final direction. But having been a robotics builder and attempted entrepreneur in the past, I think what is very missing in the world of building robotics is actually just shedding light on like the ecosystem globally of robotics. Like Who is a good technician in the Bay Area? Who's a good technician in Australia? Right. Just like where are all these service providers? Who is a good CNC, you know, mill? Which CNC mill can anodyze? You know, who can do 3D printed metals? Just even finding that is hard. Other parts of the supply chain, like who makes a mobile base where the payload is at this? Like, like who can make fabrics to cover? The robot who could make clothing like costumes for the robot. There's so many things across the board that are so hard to discover. I think like even a Yahoo style directory, like old school Yahoo style directory, Craigslist style directory, might be more helpful than any skill store at day zero. But yeah, again, that's to be determined.
speaker-0: Okay. Speaking of data, do you purchase data externally? Do you make your own data acquisition? Do you teleoperate your own robots to get robot data? Do you also use UMI data? I mean what's your take on all this?
speaker-1: Yeah. We are in early days. So I think we are moving towards a world where we can start to collect our own data. But working with models, we're working a lot with fine-tuning other models. That's basically it. So I think that might answer the question. Like we're not purchasing a lot of data or anything like that, but that data has been used in training the models we fine-tune and build off of.
speaker-0: Garlet. Back to the robot itself. You mentioned the robots will soon have grippers. I'm assuming parallel grippers. Are you also looking at hands for the future or is it too much resources to ask for for now? ⁓
speaker-1: Absolutely. So I I think the baseline is parallel grippers, right? So we're definitely doing that. That's gonna be the first thing that we ship. I think if you look at the ecosystem, there are so many hands and like open hands and stuff. I think there's a question for us of like how much value are we adding if we design our own versus just make it compatible? You know, ⁓ part of the problem actually we've seen is that all these different hands have different power requirements and stuff like that that make it harder to interface. with every single robot out there. So maybe what we provide is an open interface like maybe a PCB and some connectors or something like that to interface a whole bunch of hands to our system. So C B D it's actually a very active topic for us right now. But yeah, I don't think we fully decided there are so many talented ⁓ shops working on hands with three fingers, five fingers, you know, and everything in between with different payloads and run times and all that. So Maybe we don't have to reinvent the way.
speaker-0: Yeah, definitely. Yeah, we discussed ⁓ a potential universal interface for hands. Definitely. ⁓ I think it will come up in the future. I mean you have a point. When you see what Sharpa is able to demonstrate with its hands, like the amount of dexterity they can achieve and how fast and precise they are, you sure wonder, yeah, I can just plug these hands on my robots and you will get the best of those worlds, basically. And
speaker-1: Hands are so like the price differentiation between the cheapest hand and the most expensive one is just so wide. It feels like it's gonna be hard to choose the perfect one for any application. You know, it's really gonna be potentially application specific.
speaker-0: Regards to the software that you're shipping. I was reading about Asimov OS before. I don't know if that's still like the main thing because so much is changing and you're early. But what is like the plan for this? Because I see this videos of you guys constantly in your office basically trying new things out. Do you want to push the capabilities that you ship with to the maximum, pretty much?
speaker-1: Yes, absolutely. I think we again in the spirit of lifting all boats by raising the tide, I think we want to ship open solutions people can build on top of. And so the safer we can make it by default, the more capable we can make it by default, the more easy we can make it to fine tune by default, I think the better. I'll have to get back to you on whether it's still called Asabov OS. I think we may be rebranding it a little bit. But yes, certainly all the components. that we've been talking about before, where you have a control, you have a way to manage it and see what's coming out of the robot and send commands to it, tele operate it. Like all those things are still at play.
speaker-0: And is already Asimov two in development because I saw that is being teased already on the website.
speaker-1: Yeah, absolutely. So I think the reason we want to tease Asmov Two and we can argue whether that was the best decision is we want to collect feedback and be very open that we're collecting feedback for the future so that we can make something that is more designed for manufacturing and designed for assembly and actually designed for deployment. So I'll give you an example of one thing that is hard today with Asimov One. It has something like four or five different actuator models on it. If you are deploying Asimov one in the field and you need five different actuators in stock, that's kind of a pain. What if you only had two different actuators you needed to stock at any given site that you added Asimov? Decisions like that will make the product developer, product deployer's life a lot easier. And so we're working towards decisions like this where it's much more And deployment oriented as opposed to purely research where, you know, we can send you an actuator in a few days and that's okay. But if this is doing a real job on a real site, like one or two days you've maybe lost a million dollars as a company and that's not okay. So we're paying a lot of attention to to things like that. Yeah. Do
speaker-0: People actually assemble it quite a bit, or would you say the majority of people that are buying it right now is truly people that buy pre-assembled?
speaker-1: So we're still working through the backlog actually of of sending it to people who have ordered the kit that they have to assemble. So nobody has gotten an assembled one. But I anticipate the number of people that would want an assembled one is like ten times the number of people or maybe a hundred times the number of people who want to build one themselves. ⁓ it's a great learning experience. That being said, if you just want to work on the software, maybe it's not the learning experience you want.
speaker-0: I mean if you want the true learning experience you probably even get the bill of materials and source it completely yourself. Maybe more general looking onto the market. How do you see we already touched a little bit on this, but how do you see the robotics market evolving right now? Also with major labs coming into ⁓ the play, software, hardware, maybe a bold prediction that you have for the future.
speaker-1: A bold prediction. A bold prediction is that while open AI and anthropic will play a big role in the models, there will still be room for big folks like Cursor and all the others, lovable or what have you, the equivalents, to play in application development and certainly play in the silicon side of things and play in the electromechanical side of things. I think that. Maybe the other bold prediction that is kind of against the grain is that there is still a major role for mechanical engineering and mechanical designs in the end effector tooling. Like I do not actually think that five fingers is the end all be all for what we need in the factory scenario, in construction, in farming. Right. Like there's a reason that humans invented tools to help us do things like. Knit faster and stuff is because these five fingers, they're great for general stuff, but to be the most optimal at speed, they're not the greatest. So even like to take it back to food handling ingredients, like the best thing to peel a carrot is not fingers, right? Like you're gonna peel a carrot much faster with an actual tool, maybe that's moved to the right place with. locomotion and and arms, but not necessarily manipulation. So I think again, it's gonna be forever this blend of like electromechanical custom with generalized models. I I don't think that'll change if you're truly trying to hit like 99.9999% at really high throughput. It's just like these fingers don't work for a lot of things. Maybe that's the boldest thing I could say against the where a lot of funding has gone in the last two years.
speaker-0: No, super interesting. It makes total sense. Yeah, follow-up questions a little bit on that is I mean, it's two questions, but I I'm gonna start with the first part. What company you think is doing something right and will get much bigger than it currently is in the space?
speaker-1: ⁓ my gosh. The easy one is NVIDIA. If you think LLMs use a lot of data, just wait until you see world models. So yeah, that's the easy, quickest answer off the cuff. But you know, the interesting thing if I can look at sort of a front of me, if you look at skill day, like look really closely at the videos. It kind of goes back to something I just talked on and see if they're using the same end effector. I don't think that they are. And there's something really interesting about that. And then you know I'm gonna answer a slightly different question, but it's like there's still this question of like what use cases are is gonna be sort of this what we call internally a golden use case that's just like breaks everything open and makes it so that a general robot, you know, you buy for the golden use case, but you get all these other ones kind of for free because you're paying a lot of money for the hardware for the golden use case and and then have the hardware around. I think that's still TVD, but in the US. There are certain shortages we talked about that maybe could use a lot of help from robots. And so I think that's what excites me the most are those types of applications. Like, how do we help fix our supply chain and manufacturing in all the areas that were deficient? And the US isn't that different from Europe. I feel like Europe has to solve these problems too. They have to become a lot more self-sufficient in a lot of ways. So I would be looking for companies solving stuff in in these areas.
speaker-0: Questions. Are you actively hiring? Are you looking for a partnership at Menro Research? What can you tell us about that?
speaker-1: Absolutely. First, we are hiring a lot all across the stack in product development, engineering, research. A lot is gonna be in Singapore. So if you're in Singapore willing to go to Singapore, please definitely reach out. And then actually in the rest of the world, we're very much focused on what we're gonna call research evangelists or research, you know, developers, sort of like developer relations, if you will. It's a really cool role. It's someone that will actually go and meet with research labs. to see if there's any opportunity to collaborate on Asimov. So particularly if you're doing anything like that, please reach out. We're staffing that up a lot as well.
speaker-0: This call to all engineers interested in getting into open source robotics. About scaling production. How many robots are you aiming to ship this year and in three years from now? Do you have a number in mind or is it confidential or how is it going?
speaker-1: My guess is this year, and again, I don't have like the perfect numbers. My guess is this year it's somewhere close to a hundred. Hopefully that's not surprising anybody. Obviously as many as we can going forward. I think we are preparing for success, but also operating conservatively when we have to. I think particularly when we have these evangelists out with research labs and figuring out partnerships, we'll be in a great position to show what our product can do and Hopefully provide them to a lot of people.
speaker-0: I think since it's open source at the end of the day, the users of course will also show what is possible. ⁓ they will pretty much push out these ideas on the robot. To everybody watching at home right now, if you want to get an open source robot, then maybe you should check out Azimov of course. We are really excited for the future. Thank you so much, Matt, for joining us today. This was the Robotics Tech Podcast episode nineteen. My name is Oscar. Hey, my name is Leo.
speaker-1: All right. I'm gonna preface that with I think open robotics is what the future needs. It's not just a humanoid, but let's talk about a humanoid. We all want robots in our house. Everybody has that one little step in their house, if not a full s set of stairs. And so eventually you're gonna need legs. Today, today, you know, battery performance and things like that probably need to improve a lot before We start seeing these in our house, but I think open is really what we need to raise all the tides, if you will. It's just such a hard, hard space to build in. Companies need maybe hundreds of millions or billion dollars. And I don't think that's right for humanity. We want to lower that so that everybody can get it.
speaker-0: I can only agree. Maybe you can tell us a little bit about your background, what you've done before, and how did this lead you to work with Azimorph and Menu Research today?
speaker-1: Absolutely. So I by trade am a software person. I started to get into machine learning, sort of classical machine learning for some of the startups I did. And in twenty fourteen, I thought, you know, Alex Net is coming out. Computer vision is really starting to be able to do things that are really, really impressive. What applications do I care about that computer vision is really good at? Well, it turned out everything I wanted in the home had to do with manipulation. And so, okay, that meant robotics, right? If I cared about computer vision, that meant robotics. And so the first part of my journey was working with a mechatronics person that I found to build a toilet cleaning robot. I didn't have kids at the time. And that was like the worst thing in my day to day at my house that I could imagine. Doing that I wanted a robot to do. And so went down the path, got kind of distracted by data labeling. And you know, having to label data, particularly computer vision to recognize toilets and dirty toilets and all that stuff. So I went down a one ball in synthetic data. But in twenty nineteen I had a kid and he had food allergies. And so my wife who works and I work, we both have to start cooking like all the meals all the time. And it was just so tiring. And I thought, hey, I think I can bring the stuff I learned about robotics into actually cooking meals like just four different meals a week. That's it, just the staple. And so we start to build this sort of like a 3D printer type robot. It had an arm on a gantry, it had burners, it could tend, you know, it could extrude pasta, boil the pasta, boil down sauce. So this is like 2022 now, 2023. And I as a startup founder really want to focus on the application. What recipes can I bring to the table that are important for parents? Can I partner with grocery like delivery services to package these things up? But here's what happened. Anybody building robots hopefully can relate to this. The first thing that happened was that our open source motor controller decided to close source the firmware because they were getting copied by, you know, others, and that was their main way to make money. And so now all of a sudden we couldn't change the firmware to operate a little bit better with Ross. And sorry, there's a little bit of a long diatribe, but it really helps elucidate why I think Menlo is so important. Another thing that happened was that we needed to change platforms from an Intel-based chip to then a we wanted to go to a Jetson or Nano. All of a sudden, the way the CAN drivers worked were completely different. And then the other thing we kept doing was we are on ROS2, you know, on this little side project. And every subsequent update of ROS2. completely changed how some of the action servers worked and how the path planning worked. And so inevitably every time we made this little change here or there all across the stack, we'd have to like completely reinvent the wheel all the way at the bottom, the bottom of the bottom of the stack. And all I wanted to do was tend a stove and control how hot it got, basically. I end up getting introduced to Daniel and Nicole, the founders of Menlo, and they have this vision of this open robot where everything from even the motor controller to the firmware to the RPU which we call the robotics processing unit is open and the higher level path planning stack is all open. And boy, I was thinking if I had that to start this startup all over again, I could do it in, you know, less than a year instead of a few. So I apologize. That was a really long way to say that. This is just super important for my next journey as well, whenever that happens to be. And I'm excited to be a part of helping build this.
speaker-0: I think it's a perfect introduction and it completely makes sense between your story and the leak to Asimov. And it's interesting you mentioned cooking because yeah, I've seen some content about interest on cooking robots. There is activity in this space.
speaker-1: I have a lot of thoughts on cooking robots.
speaker-0: Why ⁓ would you be dissatisfied with the current cooking robot, all of these advanced cooking aids that can do pretty much anything? So
speaker-1: So, yeah, for a parent working, like especially two parents working, you have this desire to get as fresh and as raw ingredients as possible. You want to take literally flour and eggs and milk, right? And not start with like pre-dried pasta, for example, where you're not exactly sure in the factories where they're making that stuff, what went into it. I'm really a huge fan of like pascha and some of these other cooking robots that are making their first steps, but they're very limited in terms of what. they can cook because they don't have an arm that tends that you have to preload these pre-chopped ingredients and you as the human are doing like the lowest of low level labor and then it's doing the fun part to me, which is over the heat. I I want something in reverse where I'd like I don't have to do the prep and then I just throw it on the fire.
speaker-0: I mean cooking is probably one of the most advanced tasks. You must have gained a lot of experience, maybe also through a bit of frustration, but ultimately it guided you to the right place. And now you are at Menlo. What was like the initial impression when you got there and you were looking at all of this and like what these people are doing? And you were right in
speaker-1: Yeah. You know, my initial impression, especially being in the US, my initial impression was wow, they are shipping robots. I think that's such an undervalued, underappreciated thing. I went to Actuate. Actuate is a a big industry event in robotics. A lot of folks are there. This was in August of 2026. The only humanoid robots, by and large, were from China. And I think that says something, right? There were like one robot. From the US and it was the stretch robot, which I love them, but it's one form factor. You know, there's a whole bunch of form factors we should be seeing. And we really need to sort of lift up the rest of the world. China has such a head start in it. You know, I was like, Wow, Menlo's actually shipping robots to people. This is such a leg up. Let's take that and get further ahead and lift the whole ecosystem.
speaker-0: And how because we both have been following Menlo for quite a bit. And maybe you got some insights into how did this development happen? Because they started with the legs. I mean we've seen it with other companies as well, usually starting with the legs for locomotion, which was Asimov Zero, then Asimov One, which we obviously can get into, and maybe even Asimov two, you know, you're kinda teasing that. But maybe you can take us a little bit through this like development.
speaker-1: Yeah. I'm even gonna rewind further back because I think it's important part of the journey. So what happened is that Menlo existed three years ago, but didn't start in robotics until like just under a year ago. Three years ago, the team was working on essentially brains for robotics. You know, they trained their own LLMs, VLAs, now world models. And what happened is that about a year ago, they're like, We want to control robots with this. Let's go out and buy somebody's robot. And they look for robots they could buy and basically only could buy a Chinese robot. Nothing wrong with that, but it's not open, right? As soon as you go and buy Unitry or anything like that, if you want access to the raw sensor readings, well, you gotta like upgrade to the SEK. If you want any warranty coverage, well, you gotta pay for the warranty coverage. And by the time they bought it, it was in the six figures. And first of all, not everybody can afford six figures. There's that. But second, These things break like you're kind of abusing them when you're running these policies, right? And so they start to run their policies on this Chinese robot and it broke. And so what do you do? Well, ideally you could fix it yourself. Sorry, you avoid the warranty if you do that. Okay. So now your backup choice is to send it to Unitree or one of the other providers. And it is two months of downtime if you do that because you send it and you get one back. It is not like A hundred thousand dollar car, for example, where you get a loaner. It is you send it back, you're down. And we talked to other researchers from other universities and they had the same situation. Like their research would literally stop for three months and they would just be out and down unless they had some other solution. And so they started down this journey of after they bought it and were down, what if we could make our own and actually have it be repairable? That was kind of the premise. Like right to repair, these things are gonna break. Let's make it so easy and so well documented. They basically started this journey because they wanted to essentially use their own product, which I also think is a great way to build. One thing I didn't touch on in my introduction is I started, I think, like four or five companies now. The most efficient way to build a company is to scratch your own itch to solve your own problems. And that's exactly what what Menlo did. So I was even more attracted because of that. So
speaker-0: Did they just have the robotics expertise? I mean, when it comes to the hardware, I've seen the other product which they have, which was this kind of open LLM pretty much, right? Yes. I mean there's some overlap, but did they have some internal robotics knowledge already or did they hire people or how did that come?
speaker-1: Yeah, great question. So I think on the what we call the brain side or the intelligence side, pure software side, right? We definitely had expertise. Where we lacked expertise was on the hardware side. So they were able to find people. Menlo's really well situated in Singapore. It's sort of this melting pot of a lot of talent. So they were able to find some really key people that had experience from Tesla, had experience from Amazon Robotics and Designing a humanoid actually is like just the design is maybe half the problem. The other half of the problem that is harder to appreciate without being in it is the supply chain that goes into it. Finding basically actuators that are readily available that are easy to sim is actually just a really big issue in and of itself. There's also the batteries and all this stuff, right? Like if they aren't kind of made for humanoids already, you're gonna have a heck of a time actually producing this at scale and then. You need to throw in a bunch of money just to like spin up manufacturing. So I think a big part of the key insights that they had actually was that they needed to be really up to speed on the supply chain as much as the design. But let me double click a little bit on actuators. One thing that I think you also don't appreciate necessarily on a humanoid, as a humanoid or anything with a lot of actuators, you want zero shot sim. You want to train everything in simulation, but then be able to turn around and deploy that on the robot and have the robot act. Pretty much as similarly as possible as in simulation. And that actually goes all the way back down to how the actuator performs. If, for example, You have an actuator that performs nonlinearly. That gets very hard to send because it can't just be like this simple joint that you can model. You know, you have to like go in and actually create all the physics essentially around it and how it performs. So even the selection of the actuators is like it seems like it's all a hardware decision, electromechanical decision. It is an AI decision. Like the AI team has to have some feedback. into what hardware you're gonna choose, which is kind of unprecedented, right? It's not like that in the car industry as much from what I understand. So we're very careful and because we had this sort of brain first motion in the hardware design, we knew what to aim for and could kind of like intersect all these different competing priorities. I see.
speaker-0: Yeah, maybe just to get everybody involved from our audience, what is the product right now? Asimov one just started shipping the units. Really nice moment. Maybe give us a little rundown of what is the product itself and what is the current state of development pretty much. I have
speaker-1: Absolutely. So the current state is it's called Asimov One. It is a humanoid robot. I mean in the US we'd say it's about seventy pounds. It has a hip you know payload of about thirty pounds, which is pretty beefy. And ⁓ boy, I'll have to look back at exactly how tall it is. I think it's about five feet tall or something along those lines. It ships in a kit, actually. And you get all the actuators, all the 3D printed plastic, all the 3D printed metal. the C and C metal, and we bring you through really smoothly how to assemble it. I think today, today, it's a very research lab and corporate research oriented product. I mean, or if you just want to be walking around with this thing, you can do that too. I call it like entertainment and corporate and academic research oriented product. But we are shipping. We're shipping worldwide, which is great. I mean we have so many different countries that have ordered this and I think that's what really makes my heart happy is that it's not just like the US or it's not just China that is getting to play in this game now. We're really making it easy for folks all over the world to like be able to experiment with legged locomotion, local manipulation, and so on and so forth. And then just to give you a little bit of a sneak peek on Asimov One, it doesn't actually ship with gripers quite yet. We're working very hard on getting the grippers out so that you can really start to do things. Actual tasks. That's why I say today, today is more of a research oriented platform. I think once it gets grippers, you'll be able to do a lot more with it.
speaker-0: I see. I have so many questions. So first of all about act traders. I had no idea about what you said that actuators are hard to simulate. I mean that's very insightful because I would have thought that with ⁓ you know computer assisted design and so on, you can recreate the pretty much the entire Act Trador into solid works or equivalent pieces of software. So I would have imagined that it was easy to simulate them. We had a guest earlier in the podcast whose name is Bryce Yancioto and he's a young American engineer. designing his own new Actuator, which he calls the Orbit Actuator.
speaker-1: I could go off on a whole tangent, sorry, about US made actuators, which I think is a great thing for the US audience.
speaker-0: So it's basically three actuators stacked in the limb and then you get the three degree of motion over like half a sphere. So the device can act like as a shoulder or hip joint, basically. That's great. Yeah, kind of. Do you think this will solve kind of the simulation problem or not, or make it even harder? What's your take on that?
speaker-1: Yeah, it's more like a ball joint. So for us it depended a lot on how the gears behaved. If it was a pretty simple gear mechanism like planetaries for the classic example that you see, it became very easy to simulate. Obviously, there are some downside of planetary gears, but the fact that you can simulate it so easily, you know, you're you need GPUs to like train this stuff already with simple planetary gears. Like just imagine if you start getting nonlinearity across you know, twenty-seven actuators, like how much that blows up the space in in training. So I would have to look at it and our team would more accurately about whether they they are very like linear style performance on the actuators. Harmonics are really where things start to go a little bit nonlinear, for example, and that ends up being really much more computationally expensive to simulate.
speaker-0: I see. And so talking about American icers, are you working with American manufacturer? Are you manufacturing your own iThriters yourself? Are you using external supplier? How much can you tell us about this?
speaker-1: Yeah. I would need to double check if we can talk about which supplier we use for actuators. I'm happy to like put it in the show notes if we can, but we know we need to, especially once we start pre-assembling these. Right now it's a kit, right? But what people want is a pre-assembled robot. I mean, fewer people are excited to like screw in all the nuts and bolts. So we are actively seeking out partners that can help us assemble these in the US. And then that's even like one step removed from getting actuators made in the US, right? So like even just getting all those parts together and and assembling it in the US is like one thing we are working actively on to go one step deeper. To get actuators made in the US at a medium volume is actually not something that exists very well in the US. So we're actively also looking at that. We're very excited to do that, particularly me. I'm sort of the token American right now on the on the team and I'm opening up the US office and so we're creating a US company and all that stuff. So very excited to to do that. Yeah. And looking for partners. So if anybody's working on it, please reach out to me. We're actively seeking you out. What we?
speaker-0: I actually talked recently with Randolph Briggs, who is building Steelboard, and his vision is to build an open source humanoid a hundred percent made in the US. And he mentioned that part of his vision is actually making acters end to end on US soil with exclusively US sourced components.
speaker-1: So if you like really go into the wormhole, today today the hardest part is getting the magnets made in the US. You can wind the coils, you know, you can manufacture the gears, you can make the PCBs for the encod and get the encoders. But yeah, those magnets like are not made here. And so I hope we can spin that up or they're also alternative designs that that don't use the same kind of magnets. So we can also explore that.
speaker-0: Yes, it's definitely something I've encountered before, the famous neodium magnet bottleneck that always ends up in China.
speaker-1: And like really good for China for like figuring that out, right? Like I think we need to appreciate the amazing stuff that they've done.
speaker-0: Maybe just quickly, is that also because you said you are now pretty much creating a new company in the United States? A response to this trade ban situation?
speaker-1: Yeah, yeah, absolutely. So it's to be determined, just to be open and honest, that whether we are a Singapore owned and US subsidiary or vice versa. It's not like we've fully decided on that. To do anything in the US, you need a US company. So we will always have a Singapore operation and a US operation. Manufacturing specifically in the US, I think is a big part, to be honest, because with some of the trade bans that are coming through. I'm from here, so I want to see our country be able to manufacture things again, especially at a medium scale, right? But yeah, we have doubled down on our focus on manufacturing specifically because of some of these bans, particularly in advanced robotic systems. But being the US person, let me actually talk about why Singapore is so interesting. Singapore, you know, it's in the US it's actually very hard to get external talent here. I'll say that. It is very hard. Singapore is this wonderful melting pot of a lot of talent. And their government also has a lot of grants and basically a lot of funding for bringing in such talent and spinning up companies. You know, we think of the US as being a great place to start up companies. Singapore, for startups particularly, for advanced robotics and physical AI, particularly, it's a great place to start as well because you can get a lot of the salaries like partially covered as a startup when you don't have a lot of funding to begin with. And so we're just getting this immense talent that we see as a competitive advantage for us that if you know we really only operate in the US, we'd be competing with anthropic and open AI. And by and large, guess who's going to win that war? ⁓ David or Goliath. You know, I mean it's a different situation for us being a, you know, today today a pre seed startup. That will change over time as we mature, as we raise different rounds of funding and as we build more stuff in the open compared to some of the other of our competitors. I think it gets more and more attractive as a potential employee to like work with an open system. And so that's one of our competitive advantages that will help us going forward and maybe we'll change the equation of whether we're a US parent or a Singapore parent down the line. Makes sense.
speaker-0: Back to Asimov as a product. What is the compute on it? And then what is your training approach right now? Because I just saw you were publishing your locomotion policy just recently. Do you train mostly in simulation? What is your approach?
speaker-1: Yeah, let's start with locomotion training. I mean, it's basically all sim. You know, actually, we should be in Isaac Lab relatively soon as one of the built-in robots. I hope I'm speaking correctly. We're waiting on NVIDIA on a couple things. We also have our own training platform. Part of the great thing about Menlo that we don't always talk about is we do have a data center of our own ⁓ that we own. We do have, you know, this open source VLA and LM. training platform on top of that data center. And increasingly we're going to be moving towards a harness for like basically a gentic robotic control. And so having that full stack lets us play, you know, optimize the product at all levels to like do the end goal, which is accomplish some tasks. But to come back to training, you train it on these NVIDIA chips essentially, and then zero shot that into real. So far it's working well. We have These two levels actually, it was interesting for me to learn. Most people try to make everything work on a Jetson or something like that. Like everything from canvas, like really low-level messages all the way up to the like high level step. We took a slightly different approach, which has worked out well for us, which is the very low level, is on a eight-core ARM chip, actually. So that is like what I would call maybe the lizard brain. It's CPU only, you know, it makes sure that if you're Pulling on the robot, it can stabilize and things like that. It's sort of the core locomotion policy. Then as a developer, a user on the platform, you get to decide what the higher order brain on it is to be. It ships with a Raspberry Pi just to keep costs down. You can stream video to it. Obviously, if you want to do anything really intelligent with it, you're gonna be going to the cloud and using some cloud GPU, or you're gonna be finding a Jetson of some level, depending on your application to put on it. But Today, today it's Raspberry Pi five and we're looking at what a Jetson might look like on it as well as an upgrade. But it's open and you can mod it all yourself, right? Like all the software and everything's open, so
speaker-0: Yeah. You mentioned locomotion policies, VLA. Are these all open? Do you ship them with the robots? Do you put as much resources into researching those as you do to build the robots? What are your priorities right now? Yes.
speaker-1: We are opening up everything that we have. And you know, we're cleaning up the GitHub repos, I think, literally as we speak, just to like make things much more clear. Maybe even by the time you see this episode, that'll all be online and available. So out of the box, you're gonna be able to control it from your like mobile phone on a mobile web app. You're gonna be able to obviously do the same thing on a computer. You're gonna be able to train the locomotion policies yourself. put it, you know, flash it onto the robot. So we're we're a small team. I think we're about thirty people right now. And we're spread very evenly. Portion of those are even data center folks. A portion of those are on our agency harness for the robot. A portion of those are on locomotion and robotic control. A portion of those are on electromechanical design and then even a portion of those are on manufacturing and supply chain and delivery. You know, if you divide 30 by the number of teams I just mentioned. It's not very many people per team. I think I listed all the teams. That's pretty transparent ⁓ view into it and everybody's divided fairly evenly at the moment. But we have a decent number of generalists that also can go back and forth.
speaker-0: I really like the design of Ozimov. It's really bad.
speaker-1: Yes. Thanks. I think again as a small company, it was important for us and continues to be important for us to have even if it's like a strictly mechanical person, make sure that they can wear a few hats simultaneously just because we are short on resources. And so they have to be a little bit of an industrial designer, right? They have to think that way. They have to understand a bit about even through to the AI, at least appreciate some of that stuff. I'd say one of the main strengths of our team is that You know, it's not just a hyper specialization. There's a little bit of bleed back and forth and I think we care about how it looks, certainly. So, you know, we'll bring in even the part time resources here and there at the appropriate times because we do care.
speaker-0: Yeah. So my next question is about how was the company created and funded? Because from what we have been able to read online, the company has been bootstrapped. It also comes from Singapore, which is not the cheapest place in the world by far. Thirty people is not nothing. You have also a data center. So I'm curious, I have a feeling that maybe the data center came first and ⁓ the company was built on top of that. So I'm curious where does the funding come from and how do you think about funding for the
speaker-2: Yeah
speaker-1: So Daniel and Nicole, our founders, are also our benefactors. They have seen this company through everything and very much believe in the company. Again, another thing I look for when I'm joining a company is how committed are the founders. And they are the most committed out of anybody, right? It's their livelihood and everything. And nothing has been cheap about this journey. I'll just say that. It's been a research company. So the things that they made. at the very, very beginning didn't necessarily look like robot brain, right? There are some other iterations really, really early on. It's always been AI related, but really early on, might have been turned into a creative company, more creative AI side of things. So yeah, data center came because of that. And we started just getting all of these assets, which you can really only connect the dots looking backwards, as one of my favorite people has said.
speaker-0: About the data center, can you give us a little bit, a few details about the scale? Like is it GPU, is it CPU, is it NVIDIA, is it not?
speaker-1: It's all NVIDIA for sure. And you know, I'd have to look at exactly which chips it is. I don't happen to know that offhand. But I think the other thing we can bring to the table, particularly in our joint research with academia, is that you know, it's hard for academia oftentimes to get access to spare GPU cycles. And so part of what we're doing is is bringing some of that to the table. And that makes it really excited. We can work with institutions, you know, with names like Stanford and other big ones because of that and we'll take advantage of that ability and asset that we have and try and push robotics forward. Nice.
speaker-0: Yeah. What I ⁓ perceive when I see you, even the little community talks that you have, really nicely designed, the videos that you do are really well made. The docs I've seen they've been updated. It is done with a lot of care. And I think like that is my impression. There's like a really strong vision that is pushing through that. Since you mentioned the two founders, I truly believe that they're committed with all they got there. What is also interesting to me is ⁓ maybe back to Who are the type of customers right now and also the applications that you see? Of course, I mean if you look at the general robotics market, it's mostly research, but then also how many people assemble the robot right now and how many people buy the pre assembled one, for example, maybe some numbers there.
speaker-1: So I like to bucket our customers into three main groups. It's a little bit actually mirror of what you saw Unitries mention about their humanoids and their filings. About a third of it is to what I would actually classify as like individuals and entertainment. People who want to have a robot, bring it around, just live in the future, right? Like if you want to invent the future, the best way to do that is to live in it. People who can afford to, you know, spend twenty thousand dollars, which is not nothing, to do that. Two thirds of it actually is split pretty evenly between academic research and corporate research. But even a part of that, like I'd say maybe a third of that is actually product development within a corporate research arm. So it's not just corporate research for the sake of corporate research, like how do we create models that we could then, you know, put on hugging face or something like that. A third of it is actually how do we get and deploy robots? And the reason that that is so interesting to me is because I think particularly in the last couple of months, like legged robots have gotten kind of a bad rap in terms of their usefulness. And in some ways I agree, like, yeah, having two legs is kind of a weakness in our current state of battery life and things like that. But mobile two arms with perception, very important across the board. So some of our customers are very interested in like half of Asimov, the upper half, but we're modular. I think that's one of the other main things I want to get across. Like our future of a company, it's really open modular pieces, both on the hardware side and the firmware side, all the way up to to the brain side that can lift all boats, right? Like raise the tide and lift all boats and make it so you don't need a billion dollars to start a robotics company. We're gonna help with the actuator. We're gonna help with a motor controller and we're gonna help with a robotic processing unit. We're gonna help with training. and have modules around that. We're gonna help with data collection, not as a service, but like stuff you need to do data collection and so on and so forth. It's really exciting to me that even some of the deployments need legs, some of the deployments don't need legs, but I think we'll probably be very oriented towards bimanual manipulation on some kind of mobile platform. I think that's really where our strengths are.
speaker-0: Interesting. And speaking of data collection, you also have the Manlo store, right? And we've seen it also with other companies, Unitree. There are a lot of these data collectionslash skill stores. What kind of role does it play right now?
speaker-1: Yeah. You know, we're still in the early days of forming that. I'm gonna give you more my take on it. Right. And this has not been formalized, so I don't want you guys to take it as sort of Menlo's final direction. But having been a robotics builder and attempted entrepreneur in the past, I think what is very missing in the world of building robotics is actually just shedding light on like the ecosystem globally of robotics. Like Who is a good technician in the Bay Area? Who's a good technician in Australia? Right. Just like where are all these service providers? Who is a good CNC, you know, mill? Which CNC mill can anodyze? You know, who can do 3D printed metals? Just even finding that is hard. Other parts of the supply chain, like who makes a mobile base where the payload is at this? Like, like who can make fabrics to cover? The robot who could make clothing like costumes for the robot. There's so many things across the board that are so hard to discover. I think like even a Yahoo style directory, like old school Yahoo style directory, Craigslist style directory, might be more helpful than any skill store at day zero. But yeah, again, that's to be determined.
speaker-0: Okay. Speaking of data, do you purchase data externally? Do you make your own data acquisition? Do you teleoperate your own robots to get robot data? Do you also use UMI data? I mean what's your take on all this?
speaker-1: Yeah. We are in early days. So I think we are moving towards a world where we can start to collect our own data. But working with models, we're working a lot with fine-tuning other models. That's basically it. So I think that might answer the question. Like we're not purchasing a lot of data or anything like that, but that data has been used in training the models we fine-tune and build off of.
speaker-0: Garlet. Back to the robot itself. You mentioned the robots will soon have grippers. I'm assuming parallel grippers. Are you also looking at hands for the future or is it too much resources to ask for for now? ⁓
speaker-1: Absolutely. So I I think the baseline is parallel grippers, right? So we're definitely doing that. That's gonna be the first thing that we ship. I think if you look at the ecosystem, there are so many hands and like open hands and stuff. I think there's a question for us of like how much value are we adding if we design our own versus just make it compatible? You know, ⁓ part of the problem actually we've seen is that all these different hands have different power requirements and stuff like that that make it harder to interface. with every single robot out there. So maybe what we provide is an open interface like maybe a PCB and some connectors or something like that to interface a whole bunch of hands to our system. So C B D it's actually a very active topic for us right now. But yeah, I don't think we fully decided there are so many talented ⁓ shops working on hands with three fingers, five fingers, you know, and everything in between with different payloads and run times and all that. So Maybe we don't have to reinvent the way.
speaker-0: Yeah, definitely. Yeah, we discussed ⁓ a potential universal interface for hands. Definitely. ⁓ I think it will come up in the future. I mean you have a point. When you see what Sharpa is able to demonstrate with its hands, like the amount of dexterity they can achieve and how fast and precise they are, you sure wonder, yeah, I can just plug these hands on my robots and you will get the best of those worlds, basically. And
speaker-1: Hands are so like the price differentiation between the cheapest hand and the most expensive one is just so wide. It feels like it's gonna be hard to choose the perfect one for any application. You know, it's really gonna be potentially application specific.
speaker-0: Regards to the software that you're shipping. I was reading about Asimov OS before. I don't know if that's still like the main thing because so much is changing and you're early. But what is like the plan for this? Because I see this videos of you guys constantly in your office basically trying new things out. Do you want to push the capabilities that you ship with to the maximum, pretty much?
speaker-1: Yes, absolutely. I think we again in the spirit of lifting all boats by raising the tide, I think we want to ship open solutions people can build on top of. And so the safer we can make it by default, the more capable we can make it by default, the more easy we can make it to fine tune by default, I think the better. I'll have to get back to you on whether it's still called Asabov OS. I think we may be rebranding it a little bit. But yes, certainly all the components. that we've been talking about before, where you have a control, you have a way to manage it and see what's coming out of the robot and send commands to it, tele operate it. Like all those things are still at play.
speaker-0: And is already Asimov two in development because I saw that is being teased already on the website.
speaker-1: Yeah, absolutely. So I think the reason we want to tease Asmov Two and we can argue whether that was the best decision is we want to collect feedback and be very open that we're collecting feedback for the future so that we can make something that is more designed for manufacturing and designed for assembly and actually designed for deployment. So I'll give you an example of one thing that is hard today with Asimov One. It has something like four or five different actuator models on it. If you are deploying Asimov one in the field and you need five different actuators in stock, that's kind of a pain. What if you only had two different actuators you needed to stock at any given site that you added Asimov? Decisions like that will make the product developer, product deployer's life a lot easier. And so we're working towards decisions like this where it's much more And deployment oriented as opposed to purely research where, you know, we can send you an actuator in a few days and that's okay. But if this is doing a real job on a real site, like one or two days you've maybe lost a million dollars as a company and that's not okay. So we're paying a lot of attention to to things like that. Yeah. Do
speaker-0: People actually assemble it quite a bit, or would you say the majority of people that are buying it right now is truly people that buy pre-assembled?
speaker-1: So we're still working through the backlog actually of of sending it to people who have ordered the kit that they have to assemble. So nobody has gotten an assembled one. But I anticipate the number of people that would want an assembled one is like ten times the number of people or maybe a hundred times the number of people who want to build one themselves. ⁓ it's a great learning experience. That being said, if you just want to work on the software, maybe it's not the learning experience you want.
speaker-0: I mean if you want the true learning experience you probably even get the bill of materials and source it completely yourself. Maybe more general looking onto the market. How do you see we already touched a little bit on this, but how do you see the robotics market evolving right now? Also with major labs coming into ⁓ the play, software, hardware, maybe a bold prediction that you have for the future.
speaker-1: A bold prediction. A bold prediction is that while open AI and anthropic will play a big role in the models, there will still be room for big folks like Cursor and all the others, lovable or what have you, the equivalents, to play in application development and certainly play in the silicon side of things and play in the electromechanical side of things. I think that. Maybe the other bold prediction that is kind of against the grain is that there is still a major role for mechanical engineering and mechanical designs in the end effector tooling. Like I do not actually think that five fingers is the end all be all for what we need in the factory scenario, in construction, in farming. Right. Like there's a reason that humans invented tools to help us do things like. Knit faster and stuff is because these five fingers, they're great for general stuff, but to be the most optimal at speed, they're not the greatest. So even like to take it back to food handling ingredients, like the best thing to peel a carrot is not fingers, right? Like you're gonna peel a carrot much faster with an actual tool, maybe that's moved to the right place with. locomotion and and arms, but not necessarily manipulation. So I think again, it's gonna be forever this blend of like electromechanical custom with generalized models. I I don't think that'll change if you're truly trying to hit like 99.9999% at really high throughput. It's just like these fingers don't work for a lot of things. Maybe that's the boldest thing I could say against the where a lot of funding has gone in the last two years.
speaker-0: No, super interesting. It makes total sense. Yeah, follow-up questions a little bit on that is I mean, it's two questions, but I I'm gonna start with the first part. What company you think is doing something right and will get much bigger than it currently is in the space?
speaker-1: ⁓ my gosh. The easy one is NVIDIA. If you think LLMs use a lot of data, just wait until you see world models. So yeah, that's the easy, quickest answer off the cuff. But you know, the interesting thing if I can look at sort of a front of me, if you look at skill day, like look really closely at the videos. It kind of goes back to something I just talked on and see if they're using the same end effector. I don't think that they are. And there's something really interesting about that. And then you know I'm gonna answer a slightly different question, but it's like there's still this question of like what use cases are is gonna be sort of this what we call internally a golden use case that's just like breaks everything open and makes it so that a general robot, you know, you buy for the golden use case, but you get all these other ones kind of for free because you're paying a lot of money for the hardware for the golden use case and and then have the hardware around. I think that's still TVD, but in the US. There are certain shortages we talked about that maybe could use a lot of help from robots. And so I think that's what excites me the most are those types of applications. Like, how do we help fix our supply chain and manufacturing in all the areas that were deficient? And the US isn't that different from Europe. I feel like Europe has to solve these problems too. They have to become a lot more self-sufficient in a lot of ways. So I would be looking for companies solving stuff in in these areas.
speaker-0: Questions. Are you actively hiring? Are you looking for a partnership at Menro Research? What can you tell us about that?
speaker-1: Absolutely. First, we are hiring a lot all across the stack in product development, engineering, research. A lot is gonna be in Singapore. So if you're in Singapore willing to go to Singapore, please definitely reach out. And then actually in the rest of the world, we're very much focused on what we're gonna call research evangelists or research, you know, developers, sort of like developer relations, if you will. It's a really cool role. It's someone that will actually go and meet with research labs. to see if there's any opportunity to collaborate on Asimov. So particularly if you're doing anything like that, please reach out. We're staffing that up a lot as well.
speaker-0: This call to all engineers interested in getting into open source robotics. About scaling production. How many robots are you aiming to ship this year and in three years from now? Do you have a number in mind or is it confidential or how is it going?
speaker-1: My guess is this year, and again, I don't have like the perfect numbers. My guess is this year it's somewhere close to a hundred. Hopefully that's not surprising anybody. Obviously as many as we can going forward. I think we are preparing for success, but also operating conservatively when we have to. I think particularly when we have these evangelists out with research labs and figuring out partnerships, we'll be in a great position to show what our product can do and Hopefully provide them to a lot of people.
speaker-0: I think since it's open source at the end of the day, the users of course will also show what is possible. ⁓ they will pretty much push out these ideas on the robot. To everybody watching at home right now, if you want to get an open source robot, then maybe you should check out Azimov of course. We are really excited for the future. Thank you so much, Matt, for joining us today. This was the Robotics Tech Podcast episode nineteen. My name is Oscar. Hey, my name is Leo.
