What needs to happen before robots enter homes

And thank you for joining. I see attendees coming in. I'll literally give a few seconds to make sure that we capture most of you joining live and regularly live streaming, which is great. Thank you for joining again. I know that you're coming from all over the world, and no matter whether you're on the Pacific or the East Coast, which is the case of our panelists, including myself, California and Massachusetts are represented by us, but I know that the audience is coming from all over the world. My name is Mark. I'm the chief economist at MicroOne, and it's an honor to host this wonderful series that we've been running for a few months now, showcasing thought leaders, industry leaders, academics, experts, and our own communities at MicroOne. Today, I'm super happy to kind of take us into a conversation around AI robotics that for many of us might be the first time that we're addressing this. I'm joined by, by Arian Salegi, who is a colleague of mine at MicroOne. Arian is part of the robotics function that Microbione has initiated, and he will tell us more about the work that he does. By Bharat Togi, that has been working in robotics for a while, from what I learned from your journey, Bharat. And by Tauraj Parang, who is also in the space of robotics, more on the commercial side, to understand the business dimension by robotics. And so together with you gentlemen, will really try to tackle a question that many people are having. So I'm gonna start from, from you, Erin, as last time I saw you, we were having a wonderful conversation in Dubai and I remember when you joined and Robotics was announced for us. So tell us a bit more, more about MicroOne Robotics, what's the link and what brought you to where you are? And then I'm gonna ask the same question to Behrad and to Raz to learn more about your personal journeys.

Erin.

Sure. Yeah. So MicroOne has been working with AI Labs for a while now, building the human data layer, helping improve models. And so I came on in September to help launch our robotics pipelines as we believe that human data is going to be just as crucial to robotics as it is to LLMs. And so I joined as VP of Robotics Data. I had long discussions with Ali, our founder, about it for quite a while before making that decision, and we finally made the decision to go ahead and launch that. And so we've been focused on building out robotics pipelines to help with data to improve these systems.

That's great. Thanks so much, Eric, for this. There, look, when I was kind of preparing for today, I noticed that A good part of your career has been about foundation models and robotics. But so what has been your journey so far? I know that for you, robotics has been part of a larger narrative that has been within your career trajectory. Just tell us, where are you currently in your thinking with robotics and how it matches your career journey as well?

Yeah, absolutely. First of all, thank you for having me, Mark and Arianne. It's great to be here. The last almost decade and a half, I think I spent all of my career in education and robotics in different forms, robots on wheels for self-driving cars and now robots with arms and fingers for manipulation. And I think it's been an interesting journey moving from cars to now robots in home and then later robots in manufacturing as I'll talk more about, but I started my career in self-driving, autonomous driving industry, worked at Mercedes-Benz, Honda, Ford, and different capacities and their self-driving projects, and later on joined Apple's special project group. There, helped to build a robotics program, a brand new robotics program, and recently I joined GM where I lead the AI robotics work, and I'm currently the head of AI robotics engineering.

That's great. Similar to what I heard so far, Taras, you are an ergonomist, as the name, Serve Robotics, right? So you're right in the middle of it. You're more on the commercial side, but your career again is an incredible journey that I think speaks a lot about the quality of the people who are kind of sharing with us the understanding on how the industry and the fields are changing. Tell us your journey, Taras, and how how robotics became part of your life.

Yeah, sure, and pleasure to be on this panel. So, Taraj Parang, President and Chief Operating Officer at Serve Robotics. Serve is the company, for those of you that don't know, the company behind those cute autonomous robots you may have seen on sidewalks in many cities in the US, including LA, Miami, Atlanta, Chicago. And, you know, I came to Serve about 5 years ago. That was my first foray into the robotics world. And I am definitely a believer and I'm here to stay. I started my career in Silicon Valley in the late '90s. I started as a lawyer, but quickly pivoted to becoming an entrepreneur and an investor. So in 2005, I did my first startup and I've done multiple startups since then. A lot of times more on the non-technical side. I have a philosophy and economics background. So Less technical than everybody else on this panel, but definitely love technology and love entrepreneurs.

Beautiful. If it makes you feel any better, my background is in economics as well. To us, although my title is Chief Economist, but I'm Chief Economist of an AI data training company or so equally. There's always been, there's always a steep learning curve for me. It never gets eventually a plateau, right? This is a, but I, I think it's fascinating, this interdisciplinary talents they're now getting into. Into the industry. Erin, I want to start with you because you and I shared this incredible journey. The MicroOne is meaning for all of us, right? Besides knowing Ali and the vision that he has, I started as a human data scientist training data for large language models. And so I kind of understand now what is the meaning of training data. But what is the main training data when instead of a large language model is robotics? So help us understand what is data training for robotics. Some people might have never heard about this before. And yeah, I'd love to hear your insights.

Sure. And Mark, I don't know if you know this, we were actually on that, on that same project together. I was also deeply involved in the human data ops there. And so Yeah, I guess it's pretty complex and it's a wide range of data layers, but human judgment is necessary for us to reach autonomy in robotics. And it's been said a lot, but I'm going to say it again. LLMs had 20-plus years of internet text to train on. Robotics don't have that, right? Humans have not spent their lives recording what they do every day. We haven't had many years of robotic systems doing tasks within different environments. And so that data needs to be collected and in some cases created from nothing. And so these pipelines have to be started of both humans recording themselves, videos of robots doing actions and sensor data. And then, you know, you have to take that, not only create that data through human autonomy, but also you need to then have human judgment to structure that data, annotate it, mark where it succeeds and fails, identify different edge cases. And so forth.

Wow, that's fascinating. And you know, naturally, you make me think of a bridge question that I have for Barry, because I think Barry, if I'm looking at your career, you understand what is training data if we're talking about the research lab. But now that you're now in a manufacturing entity, so GM is, is, is, is an organization where there is a concept of fleet scale. Right? And you're kind of thinking about, again, like we were talking about data, AI, and I think we're talking about where the industry is going through an incredible transformation. So when you're dealing with that side of more physical production, what does AI and data mean for you when you're running your, your role at GM?

Yeah, sure. Let's, let's start first of all with what data means. I really like your question. I think that The default assumption whenever we talk about data in robotics and physical AI is we are talking in the paradigm of imitation learning, which is not always the case. Imitation learning recently have had a lot of success with foundation models, VLM, VLAs, and that is perhaps why most of us, when we talk about data, even at engineering level, I had a meeting today with my infra team and I pointed out the same thing that, hey, we are, when we talk about data, let's acknowledge we're assuming we are doing imitation learning. That is not fundamentally always true. There are many other data that's relevant for robotics and for physical AI that can come from reinforcement learning or other paradigms of learning. But that being said, I think that's what people do. And when we talk about big data, that's what people care about. In terms of scale, that's also a great question. And I recently had this chat with a group of my colleagues that When things mature, we move from science to technology to operations. At the scale that I live in, GM has 30-something thousand robots. We deploy a couple thousand every year. Or that the scale that Tuuraj's company operates, this is of an operations problem. A lot of engineering problems are either solved or there's a straightforward way to solve them. That doesn't, I'm not saying that in a minimizing way, I'm saying that the expertise is shifting from scientists to less engineering to more operations experts. And we as humanity, we have a lot of operations background. We've been flying planes and doing logistics and all these things. Those are all operations problems. Same concepts apply. So data is the same. I think collecting data, managing data, labeling data, what are failure modes, how do you do triage? The job's not done when you deploy the robot, and I think Tuuraj knows this very well. Now you have a large fleet that you have to triage and collect data back. What's failed? How do we fix that? So I'd say, I almost want to say it's an operations nightmare. But yeah, many things change at scale. You start caring about things that you never thought you will care because this will save a couple megabytes per, device and that turns into petabytes a month and that turns into hundreds of thousands of dollars a month. Wow.

Again, it's amazing because I'm thinking about what is in my back of my mind and how I was thinking about this conversation. You guys are directly, indirectly creating incredible connections and interlinkages with what I was thinking, which brings me to Turaj, right? Turaj, taking it more on The fact that you've been deploying 1,000 robots right now, working with important companies for delivery. And so in the context of the work you do, right, the fact that now we are on the verge of seeing more robotics in what could be commercial deployment, do you think it comes from the fact that we have improved the hardware? So now we really have improved an engineering process. Do you think it's more about we have improved the data? Do you find there's a combination of the two that made it possible for you to find this success? I'd love to have your thoughts on this. And then I'll have a follow-up for you, Torres, as well, because I think it's a question many people are asking. But I'm going to ask you this first.

Okay. Yeah, no, great question. And I kind of just piggyback on what Bera just said. There is definitely— with scale comes a lot of operational challenges. But also what scale sort of uncovers is that the corner cases that most of the time you don't see in a lab environment becomes real when you have a scaled fleet operating in the real world out there. Pretty much with 2,000 robots going out there, you're going to see all sorts of things that are new, and for the first time you're encountering them. And you have to have basically this ability to understand what the data is, label it correctly, annotate it correctly, and then kind of push that back into your training models so that you can get better. So with each delivery and with each mile that our robots drive, we strive to be better because of that. And that chance kind of becomes this data reinforcing flywheel that just keeps making it better. But getting to that scale deployment, I would say, is key. So that you can understand those corner cases.

Yeah. I have a follow-up on this. I think the one I was thinking, I'll ask you later on, but the one that you inspired me to ask is this one. So imagine a competitor that has the right engineering, the right hardware, but no data. How long would it take for this company to catch up? And the reason I'm asking is that how much of this is a general purpose technology access in terms of the data that people eventually can get there because it's diffused. And how much of this is really this curation of the data and the way that you are annotating the data, you're training the data, you have experts for this. So I'd love to know more about it in this, this hypothetical scenario, right? Whether data is accessible or is still a barrier to entry.

Yeah, I think it's a question of what your robots are doing. I mean, the way we kind of, in a very simplistic, my philosophy mindset way of looking at it is that our robots either learn to see and learn to understand the world around them, kind of localization and perception. That's part of what the data kind of helps our robots do. And then the other way is learning how to drive, learning how to navigate in complex crowded environments with humans, dogs, cats, goats, everything going around them. And that's literally a lot of scenes we encounter on a daily basis. The learning to see part, I would say, is actually a lot more advanced and almost becoming commoditized in the sense that it's a much more mature technology because there's a lot of data out there. You can just scrape the internet, whatever. You can build these with enough processing power and budget. You can kind of, even if you don't have the data yourself, you can kind of figure out how to get the data. Learning how to drive, to navigate the world in crowded environments, that's very unique. And there are very few people people that have access to that data. And that's kind of like, if you're a company that's kind of out there with a fleet and gathering that data, I would say you have an advantage compared to others.

Okay. Which I think is, makes me think now connected to Arian, right? Arian, one of the things that I found super cool when we started as a micro one is that we were mainly hiring people around the world to record first-person household tasks with cameras, right? And so somehow we are having access to data, which is kind of what, uh, Tuvraj equally said. But then Microwind is a company that has been thinking about global workforce. So how do you in, in your function look at the quality of the data so that you can maintain it to the purpose of training the robots? So how do you manage data quality considering that we are scaling globally, right? And that we having people all over the world doing this job. So that's an interesting part for me to understand because, uh, access to the data is one thing, but also the quality of the data is another one. And in another conversation that we're having with the CEO of, of Box together with Ali, we understood that lots of the things that was holding agents back was the quality of the evaluation. I would imagine the same applies to AI robotics. So love to have your thoughts as you're probably encountering this question as you're running the robotics function.

Yeah.

So as you mentioned, global workforce, we are now up to 5,000 folks recording data across 71 countries. And so quality is important. Diversity is just as important. If you want to generalize these robots, you do need massive scale of diversity. So with unique environments also come unique movements, ways of doing these tasks, different objects interacted with. This is all important. And, you know, that is a function of quality here. Now, when it comes to the actual quality of these folks collecting the data, it really comes down to an operations process, right? Micro One has done a great job applying this pipeline to AI labs for, for LLMs. And the reason we launched this was because it's a very similar process in the ops of it, right? We have human data managers, strategic project leads looking over all these pipelines, human data managers speaking to these folks one-on-one in group settings every single day. Making sure to give them helpful feedback, implement new guidelines, you know, reiterate based on what we're seeing on the research side of things. And so making sure that everyone involved is sort of passionate about it while also having a deep understanding of the goal here is very important. But it is, it is a large, large operations kind of subject. And so making sure to kind of keep everyone on the same page, I think, is the most important thing as far as quality. But that applies also across things that aren't just human demonstrations, right? For example, I imagine in SIRV's world of collecting data, you want to see, you know, also diverse environments, diverse settings. If you had a Serve Robot driving up and down the same exact sidewalk every single day, that would not be high-quality data. But as Tiras said, you have it all around. You have potholes, you have humans trying to mess with the robot, and navigating through those complexities is what creates high-quality data.

Yeah. I have a quick follow-up because I'm curious about this, right? Out of the different tasks that we are basically kind of delegating to people to train for robotics, which one did you find that is the hardest one to get right? And the reason I'm asking this, I remember when I was training data for large language models, and if the tasks were not finance or more quantitative and there was a bit of ambiguity, it was difficult to get to, to understanding whether the model said it right or wrong, right? There was a lot of space for interpretation. What would be in robotics training, right? In terms of the data, that task that you find out of the experience that is harder to get because of maybe the complexities, because of the, the, the landscape. I, I don't know. I'd love to hear what has been so far your experience, the curiosity.

Yeah. Robots are having a very difficult time with high dexterity tasks and much easier times, still difficult, but making much more progress on low dexterity tasks, such as like picking up an object and just placing it somewhere versus maybe sewing a thread, right? Like, or doing an extreme example, a surgery, right? So these Low dexterity tasks are more within reach for robots. I think the highly dexterous tasks that are very in the details, like those are way further out. And so we encourage folks to demonstrate data that is within reach and actually discourage certain data that we don't think is useful in the near term. Such as just anything that you can't imagine a robot doing in your home, we discourage. But all things that would make life easier when robots are fully deployed into homes, such as folding laundry, putting dishes away, those things that we've seen in those product demos already from these robotics labs, that is what we encourage most.

Okay, great. That's pretty fascinating. I wanna go back to you, Bharat, right? And, uh, there's a question that I, when I was kind of preparing for our conversation, I was really looking forward just to the answer cuz I, I discover I wouldn't know how to, um, to answer it myself, right? So if you think about what GM has been doing, there's one side of the story about manufacturing robots and there's a side of the story about autonomous vehicles and Somehow I think there is, there is a clear, I think, allocation of resources across these functions. But I was wondering, is the data stack compatible or converging at all? Or do you need to think about manufacturing robots with a certain level of data and autonomous vehicle with another one? Or do you find that they naturally converge? And this is something that I know there is a legacy conversation, depends maybe what kind of data you started to train first. But I'd like to know, considering that robotics in GM is advancing this bifurcation, what has been the, the convergence of the data stack?

Great question, Mark. Well, today I'm not representing GM here officially, so I'll give you my personal take. And as an autonomous driving industry versus humanoid manipulation industry, without alluding too much about what GM is doing, but Let's look at some of the companies that are building the infra for this. There are many companies like Fox Club or others that both self-driving companies and human aid companies and autonomous lift trucks, and all of them use the same infra. The fact is, if you go to the core principles, at the end of the day, this is multimodal data. It's camera, multi-view camera, plus some form of point cloud or depth representation, and plus some form of state of a robot. That robot can be a humanoid or it can be a serve cart or it can be an autonomous vehicle. That's kind of abstracted away from the data in terms of infra. But then what goes into their data, how do you curate it is very different. One of the large problems that I had, and it's kind of funny when I was working on self-driving is, well, Waymo, I know Cruise, all these companies have large fleets of data collection driving on the streets of Bay Area. A lot of that data was cars just sitting in parking garages waiting to be deployed on a mission or just driving on a straight line for a long time. So we went through a lot of pain trying to take this data out of our data lake that, okay, like enough of straight line driving on a highway. We don't want that anymore. And when you come to the space of humanoids and home robotics or factories, same thing as Aryan kind of pointed out that Well, we don't want the same task being done over and over and over. The information content versus the amount of data. So at the y-axis, if you have entropy, and then the x-axis, you have gigabytes or petabytes of data, this plateaus pretty quickly after some point. Think about it. If you keep driving on the same highway, whether if you do that a week or 10 million years, after some point, you're like, okay, I've seen this highway, I know it, I'm not gaining more information. Nothing is changing in the scene. So the right way of looking at data is not in gigabytes or petabytes, it's in information content. And that's not very well defined. I mean, there's not a straightforward way to measure information content in multimodal data. But I think this is something that anybody that has worked in robotics is familiar with this. It's not easy to guide your operations in a way that you collect diverse data and kind of synthetic guided environments, such as what MicroVone does, for example, is like you give instructions to people and you explicitly ask them to do the task that you want. It's predefined. So you control the distribution of your data. Whereas in nature, that's not the case. If you just put cameras, if almost magically we put chest cameras on every human in the world, And we just tell them, go live and we'll record your data. You don't have control over that. And then controlling that distribution is extremely hard because we do certain things way more often and way more frequent than other things. So I think as we move, this is the beginning of robotics, right? As we move from controlled synthetic data collection to just natural data collection, then this will become a bigger problem.

Wow, fascinating. So thanks so much for a lot. I, one of the privilege of my job is learning, right? So much from conversation like this and having the ability to turn it into a conversation somewhere else, right? Almost like pay forward. And I appreciate that. I remember when you introduced yourself about the fact of philosophy and economics, and I love that, right? And before I'm gonna ask you a question that something that was going around my mind around the fact that we're deploying basically delivery robots, right? But if somebody asks you, when is it that makes sense, a delivery robot versus a human driver on a scooter? And look, when I spent 3 to 4 months of my time in Dubai with the work I'm doing over there, and if you ever go to Dubai, you notice this incredible fleet of drivers that are basically delivering, right? And you always say, wait, why do we have so many? Why every delivery is an individual delivery? And the easy answer, the cost is low. When does it make sense, uh, to replace that specific function with a robot? And, and when you're looking at this from your point of view and people ask you maybe this question, what do you tell them? Is it a cost, uh, benefit analysis? Is it an opportunity? Is more of a, of a investment into the data you're collecting? How do you make that reasoning?

Yeah, no, there's definitely a number of variables that go into that reasoning. You know, one of our taglines is, why move a 2-pound burrito in a 2-ton car? And you can kind of see how ridiculous it is that you have a human and a car moving a burrito for someone, and it sometimes costs more for a fully loaded cost of that action is more than the cost of the food. There's environmental impact. Cars are responsible for pedestrian deaths and a lot of accidents. Those, if you talk to folks in New York City, they're not happy at all with the bikes, e-bikes and scooters zooming by and hitting people. So there's a number of things. There's of course economics. When you bring the cost of delivery down, you actually bring more business to your local neighborhood because people are ordering more from their local merchants. A lot of people aren't ordering today because the cost of delivery is so high. So there's all these sorts of economic benefits, environmental benefits, safety benefits that come with robotic delivery. Our ultimate kind of vision is that if a delivery can be done with a robot, it should be done with a robot. Now, there are cases where human delivery is appropriate, and those are more like longer distances or special use cases where, you know, the human touch is needed. But for, for like half of all US deliveries that are under 2.5 miles, robots are perfectly appropriate for that.

Fantastic.

Thanks.

It's a wonderful way of also phrasing it without an unnecessary trade-off, right, between humans and machine. It's just recognizing the context where machines are a better deployment choice, right, rather than humans. And I, I do agree with you. I think the same happens with delivery of packages where for a very small item, you have to take a big box, right? And this is why, right? I think it's rethinking a lot of the assumptions, right, that we have taken for granted. Gentlemen, I'm moving to my last question to each one of you, and then maybe we can see questions that are coming from the audience, if that's okay, because I think it's quite impressive that it feels like it's just an introductory part of the conversation, but more than 30 minutes have gone by. Just because of the quality coverage that you brought up. Aaron, my last question to you before we kind of move on to, to the question from the audience is more about the work that we're doing at Micron, right? And so I think people are becoming more sensitive to the idea of data training. There's not yet the same sensitivity about data training for robotics. If you were meeting a prospective company that is kind of asking the question, but why should I invest into this? Right. What would be the, the most natural answer you give them, especially in companies that are sensitive about AI, but they are not yet sensitive about AI robotics? So what would be your, your conversation with them?

Like, are you referring to an enterprise that wants to—

Yeah, an enterprise. Yeah.

Yeah. So Um, I think that every large enterprise, um, has a very unique process, whether that's their manufacturing process or their logistics process. It is very unique to them and their product. Now, I'm going to assume that the company is a successful company and is working at a massive scale. If that is true, then you can assume that their unique process has been successful because it's worked, right? So there's, there's lots of value in that, that they have this unique process. Not only is it unique to them, but it also works. And so if you can now capture data around that process, it becomes extremely useful to them because they want to obviously make their processes more efficient, quicker, cheaper, but they need to collect the data around that process to do so. And so there is gonna be a day where one of these robotics companies builds a model that works and brings it to one of these enterprises and say, hey, let's implement these robotic systems. I mean, it's already happening now, but you can imagine that it will probably happen across all enterprises. Robotics companies will come and say, hey, we want to help integrate these systems into your process. These robotic systems then have to be trained on that process. So if you want to move quickly, you need to start collecting that data now, whether it's, you know, just some simple, you know, camera setup or more complicated, you know, detailed, even robotics setups that are set up now to start collecting that sensor data. But you need to start now because if a robotics company comes to you in a year or two and says, hey, we want to implement these in your process, you don't want to wait 6 to 18 months to then collect that data. You're already behind the curve. And I want to make like a certain point to how important this is because let's take US companies, for example, US enterprises do their manufacturing and logistics. A lot of them do, do this outside of the US to save money and cut costs. But a lot of them have the ambition of bringing it domestically within the US, right? And so doing that with robotics is actually possible because you, you save lots of spend on, on human labor and replace it with robotics and you make it more efficient and quicker. And for some folks, the first thought may be, oh, well, it's gonna replace jobs. If you actually think about it for a few seconds, it does the opposite. It actually creates more jobs because you're starting, your starting point is zero because you're already doing it outside of the US. So you actually are starting with zero jobs in the US for that process. And then you bring these robots in-house. It's not like robots are doing it. They're running a whole warehouse by themselves. You need humans to oversee these robots. You need people to help if something goes wrong. You need people to fine-tune these robots, a human data layer. So you're cutting costs, you're making your process more efficient, and you're adding jobs. So I think it's very important to start collecting that data now because this time will come for, for these enterprises and they, they don't wanna fall behind. Yeah.

And I, what I love about what you say, Aaron, is that it's sometimes counterintuitive because people think robot automatically means displacement of labor. And I think it's our responsibility to guide them through the thinking to say, probably you want to see this from a different perspective, right? Rather than what could seems to be just an immediate trade-off, which is just great. So So my question to Barrett now, right? Barrett, you said something before I was kind of thinking and pondering about. You mentioned about synthetic data, real world. And I remember when I was doing my kind of preliminary work on learning more about you gentlemen, you have a research background that has clearly a lot of work in simulated environments. For autonomous driving and more. If somebody asks you, is there a threshold where you consider the synthetic data generally replaces real-world data good enough for what you gotta do? And is there another threshold where you say, actually, no, it just supplements it? So how do you define the proxies that the data's good enough and therefore it can definitely be in lieu of real-world data? And when do you actually think real-world data needs to supplement How do you work with this question when they come into work? Is it, is it like a taxonomy of task? Is it intuition? Is it case by case? How do you define that, that judgment, Barrett?

Great question, Mark. I can tell you've thought quite a bit about this topic because it's a pretty deep question. There are two sides to this. One is the technology side. It's difficult to build a simulator that's high fidelity, both from a visual point of view and from a physics point of view. Companies like NVIDIA and many other companies are investing billions, if not trillions, into this technology. That'll get solved at some point, maybe next year, maybe in 5 years. And then there's a first principle obstacle here that there are different terminology being used, world models, simulator, generative models. These change as we move, right? What is a world model? It's an abstraction of all the information that we have collected from our surrounding, right? So the way you train a world model is you show it the YouTube videos and the Netflix videos and books and Wikipedia, and it abstracts and compresses all that information as a model that now mimics and imitates. And the keyword here is imitation. The information that you've shown it, right? So the best this model or the simulator can do, I mean, there's a model behind, there's a world model behind the simulator, is only the oracle here is what you've shown it. It can only do as good as all the videos and movies and books that you've shown it, right? So yes, it'll be very high fidelity. It'll be very realistic physics and you can simulate many, many things. You can try many things in it. But we have to acknowledge that there's an oracle, there's an upper bound of information that just theoretically exists in this world model. Whereas our real world is not bounded. We can have new experiences and new incidents and seasons change and new things happen all the time. We are very non-sta— our universe is non-stationary as a universe scale. We're expanding and all of that. So that is, I think, the first principle thing. I don't think we care much about that part yet because we don't even have a high visual fidelity and realistic physics simulator yet. Companies have spent a lot of money in this, but that just doesn't exist. You don't have a simulator that now you can go and simulate tightening a bolt or some very dexterous manipulation task in it. If you had that, that would make my life way, way easier. So once we pass that, which I think we will probably sometime the next few years, then it becomes a— we get to first principles. That first principle is always information and power, right? Our compute is limited by power, and our source for that compute is limited by information available to us. So I think at some point we'll reach that oracle.

Great. So thanks again for this, Vera. I know this is a conversation that could take us much longer because they're equally implying a level of depth in the discovery of this as we're talking about it. I'll be shallow and move on, but I really appreciate the questions and answers that we're having. And I actually want to use the Q&A, and there's a couple that I think, Duras, that becomes almost like a perfect space for my question to you that are both what I wanted to ask you and what I think the audience equally is interested. There's one from an anonymous attendee And then another one from Sahil, but I'll try to phrase it by kind of building a hybrid. You now have data on what your company has been doing on delivery. At what point in time do you feel, and this is more about me talking to you as an entrepreneur, how do you know when is the right time for that data to move and scale to a different robotics product? You will control, uh, deliveries. We can imagine this is becoming even more. It gets to a point where though the data that you have could be transferable, or let's say there is a cross-learning function that you recognize. When do you find that? Is this like, is it really just a hunch, or do you find that we have indication in the maturity of the data that we can move to a next scaling, the next industry, the next pivot? What is your take on that, Dheeraj?

Yeah, and actually that segues nicely with what Bernd was just saying. I think he was saying like that omniscient oracle kind of simulation engine, we could reach that at some point. And I think one of the gaps right now between simulation and what happens in the real physical world, especially in places around public places, is that human intent is very hard to simulate. And also, if you can even infer intent of one person and a child, how do the rest of the crowd around that person interact in that environment given that one action? To actions and how do other people interact to that. So the more you actually have these robots out there in physical environments with crowd, in crowded spaces, you actually start learning a lot more about human intentionality and agency and all that good stuff. And that can be very much either in a hospital setting or on a sidewalk. Speaking of hospital, we actually recently acquired hospital delivery robotics company called Diligent Robotics. And that was actually the thesis behind that, that the data that we have and the data that they have can together actually create this flywheel of learning more environments and more nuances and more edge cases. So I think there is no magical answer to your question. It's more about get as much data as you can, expose yourself to as many variety of environments as possible, especially if you're not a warehouse robotics sort of company, if you're an outdoor public space robotics company, you do need the variety of inputs to be ready to go out there because reliability has to be very high for these devices in public spaces. Exactly.

Which I think is, there's a question for Ari and from an anonymous attendee. I'll try to skim it a little bit, Ari, because it's a bit lengthy. But it's related to what you were saying before. There is, of course, fear and concern about robotics being a displacement of labor. And we understand by also being a bit more of an insider, having an insider role, that actually we have to correct that. It's a distortion of what actually really happens. But there's also a question about trust in robotics, right? So people say, do we have enough trust in the current adoption of robotics for us to feel that, yeah, the power of balance could go in favor of adoption of robotics because they could help employment. So if this question comes with your team, people that they are working with you and they say, well, how do we know we can trust more adoption of robots, right? What would you tell them? How do you define that space where humans are evaluating that? Yeah, things are happening for a good reason. And when instead do you start thinking, I don't like what I see, probably we should have a guardrail about this. And this is more about how you're dealing with your team in the robotics, the people that might ask you a question or coming from prospective consumers that might ask you same question. So how do you deal with the fear factor?

Yeah, I'll, I'll answer this, um, independent of MicroOne and just on behalf of the human data layer, uh, in general. It's a, it's a valid concern because think about just the first adoption that is being marketed is households, right? Um, already you're presented with one of the largest liabilities possible of interaction with pets, valuable items, children, right? These are valid concerns. And so I think that there needs to be lots of deployment in kind of test phases to first collect the data and then from there build a reasoning judgment model on how to interact with humans. But I want to sort of build on this with, with Serve and even Diligent, right? So one thing that's important when you're applying a human data layer to raw data isn't just structuring the data for the purpose of reaching autonomy. Another reason is also how you're going to design these robots. For example, you could think in the case of Serve, there's tons of human interaction, right? Even though it's not a humanoid, it's interacting with humans every single day, walking past it, people picking up their food from the robot. There are cameras on these robots, and so from there, you can assess how humans are reacting. Are they getting scared of the robot? Are they welcoming to it? You need a human data layer to then mark how these human interactions are going, and then you can apply that to improve the design and the roadmap for the robots you're building. And then the question you asked, Serge, how can you apply that data to let's say different systems like Diligent in a hospital setting, same thing, right? Like if you use a human data layer to mark the how humans are reacting, if they're getting scared of it because of a certain approach, because of a certain speed, you can then apply that to the different robotic systems in different settings. So I think these are all valid concerns. Human interaction, is, you know, gonna be a little complicated going forward, but we do need to kind of not only collect the right data, but also structure it in a proper way to kind of assess these human interactions. Yeah.

Great. Thanks so much, very— I, I love the, this, this question. I think directly, indirectly, you all have both addressed one of the questions Antoine was asking about how do organizations continue to follow up. And although I did not read verbatim your question, but The idea is really to, to keep on engaging, right? And, and having a conversation on the data layer infrastructure that is, is, is done from a point of engagement and ownership rather than, than just a cosmetic add-on that maybe could have happened 10 years ago, but not, not necessarily anymore. So I'm gonna finish with, with you and we are gentlemen of time and, and I was wondering this. So there is a lot of purpose-built robots that do a specific function. And we have humanoids as well. But is there a logic on when a robot has to look like a humanoid and when it's actually purpose-built? I think, for us, your robots are on purpose not humanoids. And so I'd love to hear from you, what is the logic of a robot that becomes in shape, right, humanoid versus a robot that is simply driven by a task the robot will conduct? And again, it can be just your personal thoughts on this. There are Although you've got tons of experience in this.

And this is for me or Bera?

I'd love to have both of you as I kind of, we bring it over to closure. I think it's a question that serves you as our guest in this Micro Ones series. So yeah, Bera, I'm going to start with you and then maybe with you, Torres, as well. And then yeah, let's finish it.

Yeah, sure. Again, I'm not going to put my GM hat on, but I'll share what I think as a as a person involved in robotics. There is an argument behind humanoid, and a lot of very, very smart people in this planet are pursuing the humanoid story. I mean, Apptronik just raised a billion yesterday, I think. Optimus is probably one of the most talent-dense organizations in the robotics space. So I don't think it's a smart thing to just dismiss humanoid as quote unquote, doesn't make sense. But I think that the product story behind the humanoid is unclear to me as a person, and I'll explain why. The main argument that supports the humanoid narrative is our physical world is designed and built for human form factor, for 5 fingers, for leg locomotion, for 2 arms, bimanual manipulation, a torso, a head that can turn. That's how we've built the world around us. That's how we've built hospitals, factories, and such, which is a valid argument. What is being missed is that factories, for example, gets rebuilt every 3, 4 years with every new car model or a new product model that comes. We rebuilt the factory and we designed the factory for that specific car. So the next generation of factories can be designed to be robot native and be designed for robots that are not humanoid, that are maybe mobile wheeled robots. Homes is a bit harder because it has a bigger cultural dependency. It's not easy to change all human homes in the next 5 years and say, you can't have stairs anymore and can't have door handles and everything has to be. So I see the narrative in home space and industry. I'm really, and maybe somebody can enlighten me, I would love that, but I don't see the connection. And then the follow-up question is, okay, our world is designed, our homes are designed for, for humanoid form factor. Is 5 fingers enough to satisfy that? A lot of times the answer is yes. By manual, 5 fingers is sufficient to interact with our physical world. Do we need leg locomotion? A lot of times no. Um, so I, it's, I, I don't, I, I don't feel convinced about the humanoid story. But again, um, I remind myself a lot of very, very intelligent people are pursuing that story, so there must be a narrative behind it. It's just I don't know it.

Great. Torres, I think in your case it was intentional not to deploy humanoid-looking robots with Serve. Why? What was your thinking behind it?

Yeah, you know, I kind of just following the same things that Barat just said. And I would also love to hear an economist's take on this as well. So technology goes through these bundling and unbundling cycles. And, and that's a lot of times driven by the underlying economics than the state of technology. It seems right now where the state of technology is and the cost of material and levels of autonomy and sophistication of autonomy, etc., special purpose robots can be deployed commercially in a commercially viable way, whereas a humanoid that can do everything, it's very hard to bring it to a cost today that can actually be commercially sustainable. So you may need to raise billions and billions of dollars to subsidize it for the time being. So we decided not to, purposefully not to follow that path, but actually build something that has near-term commercial viability.

Yeah, great. Sure, look, I haven't been thinking about it from any client perspective. I do think that to some degree, humanoids, they create some form of emotional companionship that has been proven also by many of these chatbots, right? They have created intimacy with the way we are doing in our life. I think it's more of a human characteristic that we are trying to find affinity and we can transcend, right? I remember, and I'll finish with this, watching this documentary from BBC on a long flight where basically at one point in time they were finding fossils that were basically ritual, that were sacrificed to something transcendental, because humans could imagine something that they don't see and they can think is larger than themselves. So I think it's maybe the nature of humans capable of thinking what we can't see and touch that has inspired a lot of the common imaginary. And when there is a common imaginary, the economists say there's a dormant demand that eventually could become a demand one day. But I think I really appreciated the honesty that you all, gentlemen, have intellectually brought all of us to think about this. I have run several of the Micro One series across guests, different kind of angles of the industry. And, you know, each one of this has been an incredible experience. I must say, again, you being very personal in addressing this, and I want to be personal too. This is one of the, the one I remember with a lot of cherished memory because of the learning you have really been able to create and craft during the conversations. As you are professional working in robotics, you have not discounted your humanity in your reflection. And I love that part where we, we're not necessarily taking it for granted just because it is becoming the wave of physical AI. You have also asked questions that people should be asking and engaging. I really appreciate how much, um, more human agency, as you were mentioning to us before, is reflected in these big questions, right? We are not discounting human. We're probably rediscover the purpose of humans differently from the industrial policy was 200 years ago. Truly appreciate this. I always regret that I don't have the chance to meet you in person and then shake your hands as this finishes. But you're within the ecosystem of MicroOne friends. I hope to visit California soon and maybe there's a chance to have a coffee with you. It's been wonderful to have you on this. Aryan, Taraj, Mehrad, thank you so much. Appreciate what you have brought to all of us. I hope it's a good way for you also to, to end the week, kind of. Although we don't necessarily follow the long weekend like other people do. For us, work is more of a passion. But I wish you a great weekend moving forward. And thank you again for being part of the Micro One series. All the best, guys.

Thank you.

Thank you very much.

Thank you. Bye-bye.

Thank you.