EP #541 - Stefan Weirich: Building AI-powered Robots

Are the really smart people are building a multipurpose robot company? 10 years ago you would build a use case specific robot company, you would optimize that one use case, but now AI is helping us to cut out the integration effort, so essentially the way that we build these models and these robots is very much multipurpose, so you could solve a ton of different challenges with them. Welcome to the Swisspreneur Show, a podcast about start-up stories and learnings from experienced entrepreneurs. If you enjoy listening to our show, make sure to follow the podcast and leave a rating. And now, here's your host, Madly. So, today I'm sitting here with Stefan, who is the co-founder and CEO of MIMIC, where you build foundation AI models for robotic manipulation. We will talk about what that means in a little bit, but first of all, welcome, Stefan, happy to have you. Yeah, thank you for having me, it's a super nice place. We're excited to talk about everything that comes with MIMIC, but first we want to learn a little bit more about you. So you've studied in Germany, in China, in Switzerland, so you've really been moving around, and in that time you've also been moving from engineering and design across everything a little bit. So where did your interest and robotics begin? Take us back. I really don't know. I think I've always been a bit of a nerd as a kid, so I think I was like 10 years old or so when I was doing these Lego mind-storms robot competitions. So we actually compete about building and programming robots to do some random stuff, right? And then, obviously, as a kid, you go through different phases of what exactly it was one of them. But I think I was always just excited by tech and everything physically that moves as even more so. So I think. So yeah, as a good German writer, I went to study mechanical engineering as he usually worked. And for some reason, I had a stupid idea of doing two master degrees in parallel, which doesn't make much sense, because no one pays even more if you have two master degrees. But I felt like it was a good idea at the time. And it had to be the two best universities in the country, so one of which was the technically nice of Munich, and the other one was a nacho, not WTH. So yeah, I was living in Munich at the time, I started out and I had to move to Munich. And in Munich, I got dragged into the robotics industry much more in the start of ecosystem as well. So as a student already, I was welcome with a few early-stage companies, including Moshe's and Robco, back then, this was like free guys or for guys in their university lab. But it was like some super interesting landings. And then I also went to work for a few years at a big corporate Siemens Healthiniers, also in the R&U robotics space. So all this sort of has been around just some longer timelines, because that's a medical device industry. It's always a bit tricky. Yeah. That sounds really cool, it's always been sort of entangled into your path, more or less. And then Munich officially launched in 2024. When did the first idea for that start? So I originally came to ETH, just looking for a really cool research project. And I obviously, as a hardware person, I started out more from a hardware topic, so we started to work in a robotic manipulators and robotic hands and so on. And this was like a super cool research project, right? But the research robots we would build, right, they would look really good in publications, they would basically follow part right after you got the figures done, right? So even with a lot of super glue, like it was always a bit of a mess, right? And so just building something that looks good on paper was for me, it's actually not good enough, right? If someone has a bit more impact, right, than whatever was possible in academia as a vehicle. And I knew it had to be done faster than big corporate as the alternative that I had explored before, right? And luckily I was fortunate enough to bump into the right people, the right time to actually do this. And how did you, how did you meet them? So how did basically the co-founding team meet itself? I think it grew together quite naturally. So we all work in within the same lab more or less at some point, right? So Elvis, our now CTA, I was doing his PhD on imitation landing at DTA to I Center. And he was somewhat remotely affiliated with with robotics lab I was at with having a second advisor there or something. And he walked in there with a lot of very crazy ideas from the really early days of alternative models for image and language. Back when it would work really badly, right? So before GPT came out, right? And basically he was contradicting everything that conventional ETH roboticists would tell it. So this was obviously super exciting and we started connecting over that. We started sort of bumping into each other every now and then working on grants together. He started to train models on robots that I built and so on. And then we started thinking a bit about sort of how we can build a team around us so we brought in and Stefan and Ben to build a properly well-rounded team that really covers everything from AI research to robotics software to hardware engineering. And that's also how we continue to grow the team now of the time, right? And then what exactly is it that you do now? So what's the product behind mimic? That's a very good question. So basically we train AI models for robotic manipulation, right? So the key goal is that you want to build a model that can predict an action in the real world, right? Similar to how you want to have LLMs predict you with a text token or something. And if we're able to do that, right? We can have these robots be much more autonomous than they are today, right? So so robot would be able to react to changing positions and orientations of objects to handle disturbances if you bump into it or you take a part away from it and to even like self-correct of some behavior, right? So if it sees it's done a mistake, right? It's going to readjust and make sure everything is all right. And this is sort of fundamentally different than how you program robots before where you were thinking and ex-wise that coordinates that you had to hard code and trajectories to get from one to another that were very deterministic, right? And now this is basically the opposite of doing it, right? Can you give us an example of like where you would see that in, I don't know, some kind of industry or where in real life? I think applications, you have a ton of them, right? And there are also different companies working on different applications. There's a lot of hype around humanoids and in the home, I don't know if you've seen an announcement this week. And that could be an interesting use, guys. We personally think, so if our sweet spot is manufacturing an assembly, essentially where you have to assemble different components, right? And then you have to deal with them all being somewhat unsorted, right? Okay. Or you have to deal with different variations of a component. And they might all be like cluttered, arriving and have been from a supplier that you can't control, right? And there might be like an intricate assembly step of how you have to put things together, where you might have to deal with soft bodies, right, that are actually really hard to simulate or to program with robots. And then on the other side, the other thing that we'll look at is sort of everything packaging, sorting, where you move objects in and out of boxes, where you have to handle like a thousand different objects and neatly arrange them with packaging material, opening and closing boxes and articulated objects, or basically objects that you have to interact with, right? All of these things is something that's conventionally really, really challenging for robots to do, right? And it's exactly sort of where these models become very, very useful. So it could be in an Amazon warehouse. Potentially. It might not be the first thing I do, but potentially. And what would you say, like your background in mechanical engineering, what would you say, like was the most valuable that you can now apply? Yeah, I think it was actually very useful. I mean, in the end, I'm a little bit of my own customer, so I've seen plenty of manufacturing operations and processes, right, add companies that are, you know, our customers or appears of our customers that they already know that they're familiar with, right? And understanding just sort of how these guys operate, how to communicate with them, and sort of what constraints of these departments and corporates are, and the challenges of dealing with, right? And how we can best support them. I think that's actually super helpful. Yeah. But why is it actually so hard to apply, like, AI to physical manipulation? You set up like heart coding. Can you explain in easy understandable words why it is so hard to apply this? Well, fundamentally, there are different methods of robot learning, right? So like actually learning a behavior, rather than programming it, rather than telling robot to x, y, and z, right? And conventionally, it is always a complex problem, right? Because you have to deal with a lot of parameters. It's in the physical world. Things can go wrong. Things can be different. And you fodder when you try to simulate them and so on. So it's never as easy as anything that you would run on a computer because you have a physical machine. It's a fast challenge. And then in robot learning, there are two distinct areas. And the main sort of pathways of going about it is either you use reinforcement learning, which is basically, you build a simulation of the world, right? Not the real world. You just sort of simulate what it could look like, what your robot could look like. And then you have thousands of robots that in simulation on your computer, right? Try out different variations of doing things at random. And you give them rewards when they come close to solving a task that you want to solve, right? So eventually they become better and better at this. And this is what a lot of ETH robotics got big on, right? This is what any botics got big on, for example, it's really good if you want to optimize for one task, right? So you have one really nice reward function that you optimize it for. And then you come out with a policy or a model that is basically really good at solving this task because it just learned and has optimized this rigorously, right? There is a bit of a challenge when you try to get into the real world, right? Because simulation is never exactly what the real world is like. But for sort of locomotion, right, making a robot walk is generally a really good approach and it was proven to be quite successful. But in manipulation, it's very different because in manipulation, you have to interact with a thousand different objects. And they might have to require you to rethink how you reward your model entirely different for each of these. Right? So you have to shape a different reward function. You might run into other problems and you seem to real transfer when you're dealing with soft bodies or liquids or anything else, right? And usually you don't have when you just try to make things walk, right? So what came up just over the last couple of years really is what's called imitation learning. And that's pretty much the opposite. So you're not randomly generating something new, right? But you look at something that's already there, right? And you try to learn from an existing demonstration. So in imitation learning, you would look at demonstrations of someone forming a task as an example, right, with with the robot, for instance, remote controlling it, right? 100 times or so, a few hundred times. And then eventually you can train a model to imitate that behavior, right? And it will be able to sort of because you have the baseline of data, right, to quite quickly arrive at something similar that is able to solve the problem. And the key problem with imitation learning is that you don't have any data to start with, right? You can train on real world data, which makes it much closer to reality and to execution on the robot and whatever you do in simulation, but it is always bottlenecked by the availability of data. So if you build an LLM or a VLM or something, right, the good thing is that you go online and you have all the training data that you need available for free, right? And in robotics, basically, you have to start from zero. And that's the key problem. That's the problem. All the humanoid companies are struggling with all the defensive robot learning startups and SF are struggling with. And a lot of people are trying to brute force this problem by just hiring, I don't know, 100 people to remote control robots and tell operation to just collect data for you to train a model on. But in our opinion, it's not scale, right? So Mimic is really about solving the state of problem and finding a smarter way to our ground. And where are you at today? Have you solved it? We're on the way, right? Yeah. It's the journey. So, Atlantic, we do three things in its core, right? So we build our own frontier imitation learning AI models, right? So that's basically the model that you need to train, right? And then we also build a very unique data collection pipeline. So we basically tap into human data at scale, right, to train robots, right? And to do that, we use human video data and we use wearables, sort of basically censorize gloves that humans can wear to collect data. And we also use tell operations or real robot data or something like this, right? And this allows us to sort of massively scale the amount of data that we can get in, right? To train these models. And in the end, the last sort of final component is then that we want to match the execution of the robot to how we collect the data, right? And that's why we are building these robotic hands, so we have our own proprietary robotic hand design that's really designed to just match our data collection methods when it comes to integration of sensing, cameras, and so on, so that we can best control the entire stack, right? And then we can combine us with off-to-shelf robotic arms, right? Anything that's faster shipped to a customer, we can deploy real quick, that gives us execution speed advantage over a trying to build an entire humanoid body. And yeah, with that, we can already, today, train models, right, morally piloting us with fast customers that solve tasks, right? And there is still obviously a step to like generalizable physical intelligence, right? That everybody's still working on, and it's going to take us a bit of time to see how do scaling laws play out in real time as we get better and better at this? Can you take us through the whole like thought or strategy process behind it, because I imagine like it probably started with like the vision, or this is kind of what we would like to achieve. But then how did you from there get to, okay, let's create these, these gloves, for example, that we can collect data, et cetera, and then we will create that. How did this, how did this evolve? It came a bit incremental, right? So we started by sort of just taking a basic existing model that was out there, replicating that, and training it in tell operation, right? So by someone remote controlling the robot. And then already the easiest thing to do in tell operation is if you do it very close to how a human can do it, right? So you can wear AR glasses, for example, right, that track your finger positions. So you can have a very natural user interface when it comes to you remote controlling that robot. And that's never sort of the faster. And with that, we quickly ran into the same bottleneck that a lot of people ran into that we didn't have enough resources to just have a lot of people collect a lot of data on a lot of robots, right? So we somehow came up with this idea of just using a workaround, right? And using data that's already out there, basically, about humans who do their job every day. Okay. And what do you, what do you personally think is the advantage that you have in comparison to maybe others that are working on something similar? That's a good question. I think in general, the space is still very much at its infancy, right? So no one's really solved a problem yet. No one's really, like no one is deploying imitation lining to industry at scale, right? And a lot of people are at a similar stage, right? So I think we have a very good advantage in two things, essentially, right? So first of all, in how we collect the data, right? And sort of how we scale this as faster than what competitors could do. And then secondly, I think we're very much focused on sort of deploying as simply as possible. So we can go to the customer as fast as possible. So we use off-the-shelf robotic arms, right, already certified and existing products that our customers are also familiar with, right? And then we can scale faster, right? And I think crucially also one of the learnings that we made is that the customers don't really care about a robot doing everything at once, right? So a lot of people are braiding these generalist models, as they call it, which is basically an AR model, supposed to do everything at once, right? So it's supposed to fold your laundry and assembly your cars, right? And currently they're medium and best at all of these tasks, right? And it's going to take them a long time before they ever get to reliability level that would be relevant in the industry. So taking another shortcut to how we can deploy faster is really just making sure we specialize our models for each car's task that we get from a customer, right? So that we collect task-specific fine-tuning data, we fine-tune our models from that specific application. So we get really good reliability at that one thing, right? Because that's what our customer needs for that application, right? So these type of shortcuts allow us to deploy faster, are really valuable. Because then the more robots you have out in the field, the more data you can collect from the point robots, right? And then the faster you will get a data advantage that no one can get up with. And do you focus at the moment on one specific application or do you really see what the customer ones and then adopt accordingly? I think today, if you are building a robot company, all the really smart people are building a multi-purpose robot company. And that's fundamentally different than how you would have done it 10 years ago, right? 10 years ago, you would build a use case specific robot company, you would optimize that one use case, and you would find another 100 customers you can copy paste that use case to because your integration method was too much work to find another use case to build, right? But now AI is helping us to cut out the integration effort. So essentially the way that we build these models and these robots is very much multi-purpose. So you could solve a ton of different challenges with them. But from a go-to-market perspective, right, when you're building a company and you're building a distribution channel, obviously it is very helpful if you can focus on one thing, right? So you don't have to talk to 100 different customers that all operate very differently at the same time. So for us, we're really focused on manufacturing and logistics, that's just two main subgroups, right? And that's because we see here that there's a lot of readiness for adoption, right? There's a good understanding, especially in a manufacturing space for what are the conventional ABB and KUKA robots can do and what they cannot do. There's a good understanding for the potential of technology, there is a good business case behind most of the use cases that we're looking at that there is a high margin for them, and a lot of operational costs involved that they're interested in reducing, right? And similarly in logistics as well, right? So these things that you have to take into account, but even if we just go to one customer, we usually still always find like 10 different use cases that we could work on, right? And then it's a bit practical, right? So you try to find something that is simple enough that you're able to execute it rather quickly, right? And show a first proof point. And at the same time, it just supposed to be a little bit too hard that is for a conventional robot to solve, right? So that's usually the sweet spot where how we enter one specific customer. So we know that where specifically can be quite cost-intensive. Where are you currently in your front raising journey and what's your plan to make it scalable? Yeah, I think that's a very good point, right? So obviously deep tech has different capital requirements than your conventional B2B SaaS startup. We just closed a larger seed round with some of the best European deep tech investors out there. And I think by now there's a good understanding for how this works, right? So I think there is an understanding for the potential scale of the technology and also for the advantages of it, right? So robots are extremely sticky, for example. If you build a B2B SaaS company and someone else does it slightly better, right, you're going to be out of market in a second, right? You're going to lose your customer. If you have a robot in production, that's thing is running, no one is going to touch it. That's the reason why a lot of factories still operate on ABB robots from the 90s, right? So there's a lot of defensibility in solving hard problems. And I think there's by now also decent understanding for sort of the capital requirements for scaling these. And we've seen a couple of other startups, right? Show that there's a clear path to do that successfully, right? So any botics and very to just round the corner here. And there's like, decently well established supply chains around most of the components and the manufacturing methods that we need. So yes, obviously it's a different game to play, but I think we're in a very good path to playing it. Which way do you think that maybe the ETH network, but also other startup supporting networks that are here present, helped you in your journey as a startup, maybe personally also as founders, but also in your fundraising journey? Yeah, I think proper deep tech startups always take time and they take resources, right? So university based research allows you to do risk a little bit, right? So to build a little bit of the foundations, right, you can leverage grant funding to build the first demo and so on, there is a ton of really good talent, right? Some of the smartest people I've met, right? And it's a good starting point to build a team, right? And get the idea somewhat ready for investors, right? Much more than it would be required in other fields. But at some point, university bureaucracy is also not the ideal setting, right? So that's when a startup is actually much more efficient. And I think what's been very beautiful for me personally is sort of to see how the ecosystem at ETH and Zurig especially evolved over the last couple of years that I've been here. So there's a lot of these programs, right? Like there was when I arrived already, like a decent amount of older generation startups, right? That you could learn from. Yeah. But initially the ecosystem was much more research focused. And Munich was much more entrepreneurial as such. And in addition to sort of oldest startups to land from, there was also a set of these institutional programs like Innocious Grans or Venture Cake or something that were quite helpful, right? But sort of the actual entrepreneurial community completely took off for over the last three years here. It's crazy. Like they all have like proper San Francisco vibes now, right? And so there's a ton of networking events, AI centers been in a lot of great work, right? ETH from the club, ETH robotics club is just starting out really strong, right? Talent cake and all these programs haven't really been around before and there's also like an entirely new group of founders at similar stages than we are that are tremendously helpful to sort of be in touch with and learn from as you're solving the same challenges. What would you say has been the most helpful so far? Probably the ETH AI center in this network and the Innocious Grans are obviously a major asset. Mm-hmm. Is there anything that you're missing in the ecosystem? The Xeroxata ecosystem is still relatively small, right? So we've seen a lot more of these stage investors now focus on Xerox more and more, but it's still nowhere compared to like a San Francisco London Paris or Berlin, right? So I think there's still a lot of work that could be done to when it comes to investor readiness, readiness of corporate to try out new things. This is something that we see a lot in another European country is with big corporates having dedicated starter programs and venture clienting and so on that is rather uncommon to do in Switzerland locally. So also our customers are usually spread out across Europe. And then there is not yet, but I'm seeing this sort of established itself now, a clear pathway of sort of usual accelerates and incubator programs that in other cities are much more established. But that's an interesting point also with the corporates. And you also just mentioned Europe, which brings me to my next question is like, where do you focus on geographically? So you have most of your clients in Europe, you said, is this like, is that your goal? Do you want to like how fast can you be, you know, global? What's your plan? Very good question. I think as a startup, if you want to compete globally, you also have to play a globally, right? So you, in the end, you have to go fundraise in the US. You have to go manufacturing China or in Asia and you have to be there for your customers or whatever you're all right. I think naturally we were quite embedded in sort of the industrial ecosystem in Europe. So a lot of our first customers come from this ecosystem, but I think we also have a lot of strong leads in the US and we're having first plans of opening up operations there next year. So that's going to definitely going to be one of the first moves to make. And yeah, I don't know. And when you talk to potential customers, what do you usually, like, what do you say is your USP? I think that the USP is really, it sounds a bit vague, but it's really autonomy, right? It's really the ability to solve tasks that are too complex for the robots, right? That are previously considered somewhat impossible to automate and handling these reliably without a customer having to worry about it, right? In the end, it's saving our potential cost, right? It's sort of handling unstructured environments, variety of components, changing positions and all of these things, right? That make these imitation learning models so much more valuable. And when you look into the future and you think where you'll be in, I don't know, let's say, three to five years, what do you see? Hmm. A little question. I think I want to, in my daily life, go around and buy and use products where I know that they have been built or distributed with mimic robots. I think that would be a really cool thing to do so that you can actually see your own impact that you had with what you're building. And obviously, we're planning to continue to grow the company. I think there is still a big opportunity space to become a true hyperscaler in this physically eye space and we're planning to be the European one. How many people are you now? We're on 25 people, I demand. And what's standing between that vision and the status quo at the moment? The solution speed is always a critical component for startups, I would say. So there's definitely one of the topics we're concerned with, especially in robotics. It comes a lot with physical and logistical challenges, right, just getting your robots to customers and making sure they work is much more complex and just downloading a piece of software. And that's not an easy thing to do as a startup that fast has to build up this infrastructure. And the problem is right now, really that we have more customers and projects coming in than we are trying to keep up with scaling the team fast enough, right. So it's always going to be a resource and execution game to play in robotics. And yeah, we're doing our best to stay on top of it. And how do you see your technology evolving? Do you have any image of the boundaries that you kind of want to push? Yeah, I think that the lines in between different methods like reinforcement landing and imitation landing will blur more and more. And new robot platforms right in the form factor is going to come to the market and be more commoditized that will make these models much more efficient and useful for customers as well. And I think in the next years, we'll be able to sort of watch in real time how these data scaling effects play out. Because right now, these models are the worst they will ever be, right. This is always what they say about jetwpd, but it's even more true for robots because we're very much at the infancy of it, right. So seeing proper generalization to different tasks and real industry deployments at scale, right. And there's going to be so much very interesting things are going to make our lives a lot easier with the technology for sure. And before we finish up, I always like to do a round of rapid fire questions to finish up the interview. You all just ask a question and you say the first scenario comes into your mind. So Stefan, what's the most unexpected thing that you've learned from studying how humans move? The human is over-engineered. It doesn't need all the things that it has evolved to have necessarily. So there are things that you can reduce, right, in its complexity. That's a really good answer. If you were in building robots, what kind of product would you love to design? I think I would probably jump on the next super challenging tech problem and dive into that. Not sure what it is. It's not robots, but we'll get to it when it's time when robots are solved. What's your favourite prototyping tool, physical or digital? We use a lot of feeding printing, and the bamboo printers are usually quite nice for quick prototypes. So, yeah, these are pretty good. What does creativity look like to you? I think usually in engineering, you try to break down a problem into different subproblems and you're trying to explore the solution space for each of them. So you're trying to come up with what are different ways of solving this. Usually there's something that's very straightforward, but it might not be the best solution, right? I think if you have a little bit of just thinking outside of the box, creativity, that allows you to tap into solutions that others might not think of, that you get from different areas and different spaces in life, that help you to solve that problem, I think that is what you could refer to as creativity in engineering, but essentially all of it is more or less creative work, right? Apart from the fact that you have to execute it afterwards to you to build it. I love it. And last question, coffee in the lab or brainstorming walked by the lake? I don't drink coffee. Tea in the lab? I'm mostly in the lab, yes. Or illiterate in the lab. Yes. All right. Well, thank you so much. Thank you for sharing all these insights. It was super interesting. I learned a lot. And we're very excited to see what happens with Mimic and where you've been in the next couple of years. Yeah. Thank you for having me. It was a pleasure. We hope you enjoyed today's episode. If you did, you can support us by rating our show on Apple Podcast. This way, we can reach an ever-growing number of aspiring entrepreneurs. [MUSIC]