Mehul Nariyawala on Why Home Robots Must Be Vision-First and “Delegate, Not Collaborate”

Transcript

[00:00:00] Mehul Nariyawala : Most people think Roomba is the first robot vacuum, but as you very well probably know that Electrolux was, and I think it was priced, if I'm wrong, remember correctly, somewhere around 1400 or $1,500 and it failed. And then Roomba priced it at 1 99. Uh, do you know why they priced it below $200? What Rodney told us, what Dr.

Brooks told us is that that was the price under which you don't have to ask permission of your partner or your life to purchase a gadget.

 

[00:00:33] Brian Heater: Hello folks, and welcome to another episode of Automated. I am Brian Heater, the managing editor at the Association for Advancing Automation, and I am here to talk to you about vacuums, but not just any vacuums. Robot vacuums. I have been covering MADD for a number of years now, since meeting Maul. Nera, uh, a number of CSEs ago. I could tell you from firsthand experience that the company builds, uh, honestly, like the best robot [00:01:00] vacuums on the market right now. But their ambitions have always run a bit deeper than that, which is something that we get into quite a bit during this conversation. So, uh, thanks to Mahu for the conversation. Thanks to you as always for listening. If you've been enjoying the show, please like and subscribe, check out our newsletter over at Automated FM and a quick plug, because we're gonna be coming to Boston in a couple months, more information on that soon. But what I can tell you right now is that we will be hosting a, uh, live stream over on LinkedIn along with Mass Robotics on March 25th at 2:00 PM Uh, subscribe to the automated LinkedIn for more information when that drops. And with that, please enjoy this conversation with Maddick. We talk a lot about what's coming up next in automation on this show, but if you really wanna see the future in motion, you've got to be there in person. Automate 2026 is where the world's leading innovators, builders, and dreamers come together to show what's possible. Robots, ai, machine vision, motion control, you name it, all automation under one roof. Register for [email protected] to join us in Chicago June 22nd through the 25th. We'll see you there. I believe we met during a CES, um, 

 

[00:02:20] Mehul Nariyawala : correct. Two years ago. 

 

[00:02:21] Brian Heater: Nu two years ago. Uh, two years 

 

[00:02:23] Mehul Nariyawala : ago. 

 

[00:02:24] Brian Heater: It wasn't a great CES for me because I immediately caught COVID and so the rest of the week in the hotel room. But it seems like it worked out pretty well on your end. 

 

[00:02:33] Mehul Nariyawala : It did. That was the first CES we were just starting to talk about Madic. I think we had just come out of, uh, come out of. Stealth, quote unquote, uh, about a month ago before we came to that cs, and we didn't have a booth as per, but we were just meeting a few folks and kind of starting to talk about it. So it, it turned out well, but it was also very clear, uh, based on that CS that we needed to ship. So the focus as soon as I came back was back to shipping and [00:03:00] prioritizing and finishing the product. 

 

[00:03:01] Brian Heater: What changed during that event that made it more clear that you needed to actually ship the product?

 

[00:03:07] Mehul Nariyawala : There were two things about this particular category that we are in, which is robotic vacuums as a beachhead first category. There was very clear that there were, there were years and years of. Over promise and under delivery. Um, and what that meant is that every single person that we met kind of accepted this point of view on what it could be doing, but then really wanted to see it happen and make it work in different world. And I think that's generally true for robotics and especially true for robotic vacuums as a category where first robust came out in 2002 or two, and I think Electrolux launched their robot vacuum in 2001. So for 24 years they've been just basically. Disc and, and bumping and usual stuff. So, so that was one.

The second part also, Brian, uh, uh, prior to this I was at Nest as you know, and I feel like between Nest and Medic, um, we've gone from [00:04:00] world that was default trust to default skepticism. So it's not just the, the robotic vacuum category in itself. It was also macro environment where I felt like there was just a lot of. Question mark and apprehension about willing to embrace anything. So from customers to partners, to everyone, the entire idea was, Hey, let's ship. Let us see you scale, and then we'll, we'll have a conversation. 

 

[00:04:23] Brian Heater: That skepticism is because you're moving from like, you know, effectively from Google into a startup, or is it something beyond that?

 

[00:04:30] Mehul Nariyawala : I think skepticism is one, because robotics has been hyped for quite a while and many, many, many Kickstarter campaigns get funded, but they usually never ship. Or if they're shipping, they're prohibitively expensive, so can you fulfill your promise? That was one part of the skepticism. The second part, in general. I just feel like customers are no longer as attuned to someone saying this is a great product and just buying it. They really want, uh, and I think this is sort of [00:05:00] the Amazon effect where every single review for any robot vacuum is about four. So how do you know which one is actually good one? Uh, so there is a reviews proliferation, there is other stuff. So people really wanted to see it working and have their friends talk about it and have real users talk about it. And, and we are still seeing that today. It's kind of interesting. 

 

[00:05:18] Brian Heater: I'm asking you all these questions, but obviously I was coming to that conversation with a lot of that skepticism on my end and, and lot of that skepticism was coming from the fact that like, I had a lot of meetings, you know, when I was at TechCrunch. Like I was pretty, pretty overloaded at that that point. Um, you know. Again, CES, there's hundreds of robot vacuums, so you know what's gonna set us precisely this company apart? Do we need another company coming in, doing the robot vacuum thing? I'll say you had two things working in your favor immediately. One was obviously the pedigree of, you know, having spent time at Google slash Nest that'll immediately get you on somebody's radar. And the other, um, I probably shouldn't say this out [00:06:00] loud, but I can now that I'm kind of like, not entirely on the journalist side of things, but you had a very. Insistent PR person.

 

[00:06:09] Mehul Nariyawala : That's good to know. Uh, that's good to know. That's, that's awesome. Uh, no, she's, Emily was great. Uh, she, she saw the product and I think. What changed if I were to kind of, uh, uh, uh, say what changed? And the second part of that learning there, now that you brought it up, was also that when people saw the product, it made a big difference. Yeah. So across the board, this is consistently has been theme that it isn't about me talking about the product or us kind of doing marketing or ad campaigns around it. It's really just people using the product that kind of conveys the differences. Much, much better than anything that I can say over the, over the camera.

 

[00:06:49] Brian Heater: Yeah. This is an interesting aspect of, of CES that, uh, probably, uh, most people listening, um, don't really know, but the most of, [00:07:00] and in fact, the really good meetings are not happening on. The show floor. I mean, maybe, maybe that's obvious to people, right? Because the show floor is like the, is the chaos. And a lot of the, you know, when I'm, when I'm meeting with the really interesting startups, in a lot of cases they're happening in hotel suites. In this particular instance. You're very nice. I think if you had just had, if you were in like, uh, Eureka Park, which is where all the, the startups are stationed. Um, sure. You know, I don't, I don't know that, I don't know that you would've caught my eye, but, um, I know that it's. More difficult to be in a position where you have to get yourself a one-on-one with somebody, but obviously it's much easier to make your case if you could actually just sort of sit down and show somebody and walk them through your product E.

 

[00:07:42] Mehul Nariyawala : Exactly. There are elements which are different and hard to explain, uh, also, which is when you see robot behave, it kind of gives you an understanding on what's. Possible because it hadn't been done before. I absolutely understand where everyone was coming from. Like I, and I think it just [00:08:00] reinforced my point of view that look, we need to ship and we need to prove.

And I think that was therefore internally as well. Uh, to be entirely honest, it's not just that robotics, you know, category itself has been, um. Over promise, under delivery. But for us as well, we've been, we had been working at it since 2017 and this was 2024 or 2023, uh, CES, so it's been six years. So what was taking so long? Um, and then that's another part as well that wise robotics so hard to ship and wise robotics, um, you know, uh, take such a long time and I have my more thoughts on it now that I'm past that stage. But even that was also the sense of urgency to ship and prove that we can actually make it work. Was real, real imperative for us as well.

 

[00:08:46] Brian Heater: Yeah. So this was something that I really wanted to touch on was both, um, your, your internal skepticism, um, like early investor skepticism. Obviously you landed on the robot vacuum, which again was already. This, [00:09:00] this saturated space, but you must have had your own doubts that this was something that you could even potentially make a mark in.

 

[00:09:08] Mehul Nariyawala : I think we never had a doubt about whether we can make a mark in this category or the product vision that we had. I think we always had a doubt. Like I always kept telling Neath my co-founder and CEO, that I know we can do it. I don't know when. I have zero doubt about if, but I do have doubts about when, because we were taking a completely different approach of just doing Tesla like vision only just with five cameras and building the entire foundation from scratch. And what kills you in a startup? In my humble venia. It's not that what you don't know, it's that what you don't know, that you don't know is the unknown unknowns. And there were several unknown unknowns that we had encountered and, and, uh, along the way and we've overcome those things, but that is the reason why we, it took longer. Um, I think, um, I. Recently in one of the interviews, Andre Carpathy mentioned this quite well, that in robotics as well as self-driving cars, uh, the accuracy is March of nines. 

 

[00:10:05] Brian Heater: Yeah. 

 

[00:10:06] Mehul Nariyawala : Which is 90% accuracy or 99% accuracy, or 99.99. Right. And each nine is the same amount of work. Each nine is a constant amount of work. And in robotics, what ends up happening is that at 90% demos look really, really impressive. But then getting it to product is. Three x four x the work because at 99.99, you probably have a customers who want to accept the product in the market. And this is another conundrum, uh, uh, or sort of, um, slightly less understood, uh, aspect of robotics is that one, no one wants robots.

They want solution to their problems. So robots are just means to an end. So as long as you're solving the problem, that's great. So that's number one. Uh, I'm personally of an opinion that R 2D two is probably the worst idea in the world in terms of robots. The reason is because we already have an R 2D two that's called an iPhone. Uh, Luke's Skywalker didn't have it, so he had to have a robot following him around, but we have that, so, so, 

 

[00:10:59] Brian Heater: [00:11:00] yeah. C3 PO even. Yeah, he's, he was just, he was just a walking like, uh, ear, ear pod at the end of the day. Right. Correct. Exactly. In those 

 

[00:11:07] Mehul Nariyawala : translations. 

 

[00:11:08] Brian Heater: Yeah, 

 

[00:11:09] Mehul Nariyawala : exactly. Uh, so, so in that scenario, a lot of those technologies already available and we are covered, so we don't need robots for that. So what's problem you are solving, that's number one. And number two, the problems that robots solves are always the ones that are mundane and rapidly even boring and dangerous in these problems. Customers do not want to collaborate. Today's ai, we collaborate. We get to, if he ask, uh, Chad GT to create a report for us, it does 90% of it right?

And then we are happy to take it to the final 99%, uh, and final 10%. But you don't want a robot that says, I'm gonna come into your home and clean everything except one corner, and that's the 10% you'll have to do yourself. That collaboration doesn't work in case of robots. People want to delegate, which means accuracy of bar is actually much higher than what today's AI can deliver.

 

[00:12:00] Brian Heater: I hear the 9, 9, 9 9 number a lot, and I think it makes a lot of sense obviously in the context of self-driving cars, right? Mm-hmm. Because mm-hmm that 0.0, 0, 0 0 1% an accident like that could be somebody's life, right? Or even like industrial robots, things like that. But there's a somewhat higher threshold for error with robot vacuums, right, than there are self-driving cars.

 

[00:12:25] Mehul Nariyawala : Absolutely. We had assumed that in 2021 and in 2017, even when we started out, that we do not necessarily need to be, let's say, level five autonomy, where it just works 99.999% of times in order to ship. That was the thing. At the same time as we started shipping. What we realize, Brian, is that. The trivial, the task, the less patients customers have, 

 

[00:12:47] Brian Heater: the more they expect out of it 

 

[00:12:49] Mehul Nariyawala : precisely, which is self-driving.

Car is complex for human beings. Even after years of driving, you might still make a mistake, you might still get into an accident. So it, it's [00:13:00] something that is accepted as a complex behavior, complex task. And if AI does 99% correctly, you're impressed. But it, cleaning the floor or navigating your home without bumping, you don't go to school for that. You don't learn that. So in that scenario, like we get email if the single popcorn is left behind. Sorry. Uh, I know you want to jump in there, but, but yeah. Yeah. To finish my, finish my thoughts. Like we get an email if a single popcorn is left behind, if the single cru is left behind. And this is not just about it making a mistake, it's literally saying that it didn't clean. I want my home perfectly clean. If your vacuum cannot do that, or if your robot cannot do that, it just useless. 

[00:13:35] Brian Heater: This is a a point that I've been talking to a lot of folks about, especially as it pertains to humanoids, but I think it's appropriate here as well, is that people's expectation of what is easy for a robot and what is actually easy for a robot are not necessarily the same thing.

 

[00:13:51] Mehul Nariyawala : Precisely. I think it's more than the robot is the tasks that we want robots to do, which is current ai. If it does, let's [00:14:00] say, you know, cursor gets 95% of your code right? At Claude 4.05, Opus does 99% of the job correctly. You're just mized because that's something you spend years learning. Like, you know, if I, I'm not a writer, if it does, if, if, uh, you're a writer, if, if I try to write something and it gives me 80% of what you can do, I'm pretty happy with it because that's not my skillset, right. As I said, cleaning the floors, that is not something that is considered a trivial task or, or, you know, loading a dishwasher or folding a laundry. These are, I mean, that is, those are all considered trivial tasks. They're not complex tasks, and we expect 10-year-old to do them extremely well. So if a robot cannot do it, we're like, okay. Just, it's just kind of stupid. Right. The analogy I always use is that if you have a. If you have a domestic help coming to clean your, clean your home, and if they leave one corner dirty first time, you'll give them an instruction. Maybe even second time, third time, they're definitely fired. Uh, and I think that's the level of, uh, expectations that, um, robots in [00:15:00] home has on, uh, from customer's point of view that they just want it to work. They wanna delegate, not collaborate. 

 

[00:15:06] Brian Heater: The other sticking point for myself and a lot of other people too was, was the price point as well. And that's a difficult, that's a difficult math on your end because it, it, it, it, it seems to me that the higher the price point goes, the more people subsequently actually like, expect out of the robot.

Right? And this is incredibly difficult to navigate, um, for you because one, you're a startup, you're, you're not at scale yet, but also. Time-wise, supply chain. I don't know what's going on right now with tariffs. I mean, we can, you know, go down that road. I don't, don't know if you necessarily want to, but like, this must have been an incredibly difficult time to launch a consumer hardware product and, uh, to sell it even like, you know, a hair under $2,000.

 

[00:15:54] Mehul Nariyawala : Yeah, it's, uh. That's a great point. We did quite a bit of research early on and we had few intuitions. So I'll give [00:16:00] you one example of it. So most people think Rumba as the first robot vacuum, but as you very well probably know that Electrolux was one, right? Trying. And it was surprised that I believe. Yeah. And I think it was surprised, if I'm wrong, remember correctly, somewhere around 1400 or $1,500. Mm-hmm. And it failed. And then Roomba priced it at 1 99. Uh, do you know why they priced it below $200? Did you ever get an answer from that? From Rodney Brooks. 

 

[00:16:24] Brian Heater: Joe Jones just wrote a book about this. Have you seen this?

 

[00:16:26] Mehul Nariyawala : I have not. No. No. 

 

[00:16:28] Brian Heater: Did you have the answer to that question? Why they hit that number? Specifically 

 

[00:16:31] Mehul Nariyawala : what Rodney told us? What Dr. Brooks told us is that that was the price under which you don't have to ask permission of your partner of your York life to purchase a gadget. That's right. So they were actually targeting one 50, uh, and then they couldn't make it at one 50, but uh, the idea was 1 99 and it wasn't creating a market, so they wanted to make sure that price point was that. So that was number one insight that we had from very early on. The second thing is, if you take a step back and think about it, you mentioned $2,000 as a threshold, which is a great one because we realized very early on that there is not a single. Ubiquitous consumer electronics device priced higher than $2,000 beyond $2,000.

You are in AER space. Mm. Even might give Apple with Vision Pro, uh, at 3,500 struggle, right? So only thing consumers pay that is about $2,000 is, is really cars and houses and cars have been around for, you know, a

hundred years proven utility. No one questions that they're helpful yet. A $10,000 car remains a considered purchase. It's not an impulsive purchase and there is so much value associated with it. So to ha get someone to part $10,000 upfront, it's a quite a heavy, uh, ask in my opinion. And at that point in time, I think, uh, going below that price point was critical. So that's one of the biggest reasons why we did vision only autonomy. Uh, we very early on that came to conclusion that at in Home Robotics, and I'm willing [00:18:00] to probably say this now, uh, with Fair. At least 95% confidence that in, even in an indoor robotics of any kind, vision only is the only way to make things economically viable. Mm. Otherwise, you're just not gonna have enough business model or unit economics available for you to solve the problem. So unless you can do five cameras and lower very cheap sensors, just like human eyes and brains, and make it happen. It's just not gonna pan out as a company economics or long-term strategy point of view.  

 

[00:18:30] Brian Heater: I wonder though, is that a problem of scale? You know, like down the road maybe that might be viable?

 

[00:18:36] Mehul Nariyawala : no, no. So, so we had this rule of thumb at Nest, and I have to give this credit to Tony Fidel for Nest, for teaching us, but he would always talk about this point of view that if you had a single sensor in a hardware, assume three software engineers on a flip side is a permanent cost. So more sensors you have, the bigger the team, more sensors you have, the more calibration, more sensors, more failure points, more sensors, more complex, the supply chain, higher the [00:19:00] bomb cost manufacturing, so on and forth. So complexity rises exponentially with each sensor you add in a hardware. Now that complexity, even at a scale, doesn't go. And then I sort of go back to our own analogy as a human being and, and this point of view that, you know, one of the point of behind Multisensor approach is the sensor fusion. But if you actually think about it, we don't, as humans, we don't really do sensor fusion. If we see something but not here, we're gonna trust our eyes. If we hear something and not see, we're gonna trust our ears. The only thing we do is, uh, is, is our confidence in whatever is happening is reinforced. If we see and hear both. But there is a primary sensor for primary thing. Not only that, but let's assume that we lose a sensor.

Let's say you, you know, you get a eye patch where for a day and try to walk around, there is a good chance if you try to shake a hand, you will be t twin inches away from the actual hand. So the best computer and best, uh, computer engine in the world can. Uh, handle the single sensor loss immediately. It takes a [00:20:00] few days for you to adjust and figure things out. So now, if you have 20 sensors with each sensor failure, who do you rely on? Which one do you trust? Which one is the primary sensor? So complexity, compute requirements, all that just skyrockets memory requirement. Everything skyrockets and. And not only tariff, but memory is really expensive. All of a sudden when compute buying a DIAM is a pain in the behind. So in general, all these things adds up, and that's where our belief is that, that even in a indoor environment where cost may not be an issue, unit economics may not work out. Uh, I'll give you an example. Uh, typical forklift, for example, is about $20,000. If you buy any sort of robotic forklift, they're usually around 70, 80, 90 k. Then you need a fine lines of accuracy and fine lines of reliability to make it work. And generally we hear, when we did the research, we learned from many customers that ultimately with the reliability and uptime not being there, it usually ends up same as having another human being. So unless you can bring that [00:21:00] cost down to 30 K, it's not gonna happen.

 

[00:21:02] Brian Heater: Yeah, that's, that's a good that, that's exactly what I was thinking. Uptime, I think is the key there, right? Because that's a big selling point on a lot of these robots is that they can potentially, theoretically, you know, if they got hot swappable batteries and everything else, work a 24 hour shift, which I guess theoretically maybe thematic could do as well.

 

[00:21:21] Mehul Nariyawala : Uh, that's, that's the thing. Theory and making it happen. So I'll give you a funny anecdote, but we are using this robot hand called Dobo in our office to do some calibration and camera rigs adjustment. And I believe it's a company out of Sheen that is building humanoid as well. And in that $40,000 robot hand, we have the motor burns out every three months.

It doesn't matter what we do every three months motor burns out. Couple of times we got it to fix them to come and fix it for us. Now we just know how to fix it ourselves and we just keep, uh, stock in the backend. So, so, and that's just a single actuator on a robotic hand. And what that means is that the reliability that we want to look for at a price point of 10,000, $80,000, that reliability that we have in these phones doesn't exist yet. It doesn't exist yet because really in robotics there is no scale. 

 

[00:22:10] Brian Heater: Yeah. 

 

[00:22:11] Mehul Nariyawala : The only category with scale in robotics is robotic vacuums where. iRobot and Roomba cumulative have sold about 50 million robots in their entire lifetime. Um, the next highest is Amazon robotics and million robots for their warehouse. Robotics. 

 

[00:22:25] Brian Heater: Yeah. Drones, maybe you could make an argument. 

 

[00:22:28] Mehul Nariyawala : Drones. Yes, drones are definitely getting there for sure. But in robot, so motors and actuators. So it's like, you know, rotator motors or blade motors are there, but the actuators are not there yet. 

 

[00:22:37] Brian Heater: Wait, just to back up a second. You're using, what are you using the arm for? [00:22:42] Mehul Nariyawala : We are using ARM to just calibrate, like in a production system, just to calibrate the, okay. Calibrate the robot itself. 

 

[00:22:48] Brian Heater: I assume that you're also like experimenting with manipulators too. One of the big selling points that that, that you had. There were a couple things again that jumped out. You know, one of them, I went back and read my story and I, sure, I guess I really jumped on privacy.

'cause at the time there was that whole thing about the Roomba, like there was a lady on the toilet. We all remember that MIT story that came out. Um, but, but the other thing, you know, that, that jumped out at me was this idea of like. Of, of the platform is like, how do we get our foot in the door of the home? Mm-hmm. And for you, it makes sense to start with the robot vacuum.  

 

[00:23:20] Mehul Nariyawala : Correct. There are two analogies there that, uh, that I can kind of share to give us, give you an understanding. So instead of building sort of a full blown humanoid or full blown Rosie, the robot from scratch. We always thought that it's better to grow the robot the way human child grows. So if you actually take a step back and think about it, between zero to five year lifetime of a human child, all they learn is perceive how to perceive the world. They learn to navigate, perceive not run into things, but they're not doing any tasks. So instead of in the context of floor cleaning robot, what we are teaching a robot is to perceive the world in a high fidelity 3D and under have a contextual as well as semantic understanding of it. Then between five to 10 is when human child gets a little bit better at dexterity and some sort of a task planning, and they can maybe, you know, organize their room or make their bed or, or put their shoes, uh, back in the closet or, or make sure that bookcase is properly maintained. 

 

[00:24:14] Brian Heater: Mm-hmm. 

 

[00:24:14] Mehul Nariyawala : So that's the sort of a minor task that they get good at and they fine tune their dexterity. That's the second step in what we are thinking about as a product as we move forward. And then the third one is 10 to 15 is when we really start putting together. Uh, the locomotion, the dexterity, and the long horizon planning to cook and clean and dry, and all kinds of, uh, uh, complex tasks. So that's how we think about it, that it's really the three step progressions. And in each of the way, in each of those steps, how do we make. Product useful. What problems that we can solve with the current capability. So yes, we are, uh, experimenting some with some of these things, but in our case, we almost always start with a problem and work backwards. To us, a humanoid isn't a product, it's just a platform. What it does, why do I buy, do I buy it for [00:25:00] laundry? Do I buy it for dishwasher? Do I buy it for babysitting? I don't know any of those things yet. So what problem you are solving becomes really, really the critical starting point for us, and then we work backwards based on that. 

 

[00:25:10] Brian Heater: I'm sure I've asked you this at some point, but one of the things I I would be helpful for me, for you to contextualize is what you mean when you say platform with regards to, you know, what is again, a, a consumer. Product, right? I mean is mm-hmm. Is it, you know, are, are we, are we talking about like modularity and actually like letting the consumer kind of like build out on the device itself going forward? Mm-hmm. Or are we talking about it just like sharing software? 

 

[00:25:42] Mehul Nariyawala : Uh, for us it's about building the brain of the robot.

So when we think of it as a platform, it's about can we give it a capabilities? And first one is sort of like this visual cortex with the perception. And then you say, now can we have that perception combined with dexterity in a meaningful way where it can at least even do basic tasks at the high level of accuracy?

And once you have those two things, then you add the long term. Planning and memory element of it and say, let's combine everything. And, and obviously hardware has to grow and get better at it. But if you, you know, if you looked at Madic, madic has the crown, the black crown on top of, and if you remember, that's our entire sensor stack. It's just a five cameras based system. And the idea was always, and if you look at Madic, madic is actually at a same height as the crawling human child. That's the vantage point, top-down, vantage point that it gets. And that was the priority number one. That instead of a flag ant like vantage point, can we get that top-down vantage point. And that crown was always meant to scale and grow and grow taller and just work out of the walks. So just the way our perception system gets built when we are, maybe were a one feet tall, uh, but then as we scale, it just works in the same exact way we thought. We can build a system that for a robot and a brain that just works and then you use that platform for us is not. That users enable more use cases. [00:27:00] It's that we think through that. What are the next set of use cases that we can, uh, deliver to users, and they just work with that, you know, five lines of accuracy. 

 

[00:27:10] Brian Heater: What is it now? It's about around like 1600 now. Is that right? 

 

[00:27:13] Mehul Nariyawala : No, it's a, it's 1,250. 

 

[00:27:15] Brian Heater: Oh, 1200, okay. Yeah. 1245 

 

[00:27:17] Mehul Nariyawala : to be exact, 

 

[00:27:17] Brian Heater: which is great. The price keeps going now, which is, which is nice. Yes. I guess the idea is there's this upfront cost, but you know, theoretically over time, asmatic grows and your capabilities grow on the backend. I'm able to use the robot vacuum as a literal platform to grow out the functionality. 

 

[00:27:35] Mehul Nariyawala : You are absolutely correct in the sense that robotic vacuum category itself is inundated. Uh, with this idea of every six months there is a new robot that makes the previous one obsolete. Mm-hmm. So, as a consumer, if I'm investing a thousand, $1,500, I do not want to get a hands on on a robot that in six months later, will become obsolete already. In that scenario, the vision first approach allows us to do exactly what Tesla does, which is Tesla. Even after 10 years of owning your car, you'll get software updates and your car just keeps getting better and better. So since launch, we launched in a Thanksgiving, or since we started shipping to be Exact, which is Thanksgiving of 2024, first product, I think we've shipped about 70 releases ever since. Hmm. Um, and 70 software updates. So every week. Customer gets a new software update and it almost feels like you got a almost a new robot and that capability keeps increasing there. At the same time, there is a limitation in the hardware in the sense that matic doesn't have hands. So anything that may require hands it can't do, even if it has a brain. So there will be an element of building second, third, and fourth robot where, uh, complexity of task required changes where hardware required changes, but it's not based on software need. It's purely based on hardware need. 

 

[00:28:55] Brian Heater: Okay. Yeah. So that, that was a disconnect. So when you're talking about building up, you're talking about what you're doing over there at, at hq.

You're not talking about me being able to just like stick some arms on it that I 

 

[00:29:06] Mehul Nariyawala : No, no, no, no, no. We're not, we're not doing that. I don't, uh, that s. Modularity and sticking some arms on it. That stuff works for early adopters, but I don't think main actually want to do any of that stuff. 

 

[00:29:20] Brian Heater: Uh, you know, for sure. And just, and talk about like, you know, motors burning out really quickly, that seems like a, a pretty easy way to do that. Yeah. As again, as, as we're having this conversation right now. Well, one, I noticed that you're not in Las Vegas. 

 

[00:29:33] Mehul Nariyawala : Yeah, I'm not. Um. I think, uh, consumer electronic shows is, is great, but at the moment we are, we don't have anything in, quite honestly. We are just, we just need to improve the product. We don't have anything new to share or showcase. Sure. It's a show where you wanna show the hardware capability and for us, the hardware capability is ingrained in the software in itself, and we just want to do that. And, and then the second answer is honestly. At the moment, we can't build fast enough. Our challenge isn't the demand or getting customers. Our challenge is just, just we are, we are massively supply constrained, not not demand constrained. 

 

[00:30:09] Brian Heater: It must be interesting too, just from a, um, from a, a, a publicity standpoint because, you know, the, the, the team can, you know, can, can, can only grow so much and, and you're focused on this product it sounds like, and you will be probably for the next several years before you really start thinking about making entirely different piece of hardware. Um mm-hmm. So, I mean, you're, you're gonna be, you're gonna be in vacuum land for, for quite some time. 

 

[00:30:37] Mehul Nariyawala : Yes. The analogy I'll give you is that we absolutely want to win, win the battle front in a war and security before we get to the next battle front. At the same time, um, it's not function of, uh, uh, of uh, uh, justice alone. It's a function of where the capability of robotics industry is moving as well. And I think robotics is making a faster product. There are, uh, options available and then there is also customers now that we have. Uh, more than 6,000 customers. They are telling us what they want next, what use case, what are the problems they really want to solve. So that does give us insight. So really the the answer is to. Make sure that we find a problem that is as intense as for cleaning. And, and then, uh, once we do that, we'll move fast. So, I don't know if it will be several years, but, but you know, we, we should have something, some interesting experience happening in a year.

 

[00:31:25] Brian Heater: Can you gimme a couple of ex of examples of what's been floated? 

 

[00:31:28] Mehul Nariyawala : I'll tell you what customers have asked us and, and, and these are some of the interesting ones. So one, one that we never understood, we didn't anticipate very surprising was that a lot of.

Customers came back and said, Hey, I have these parents, they live in Florida.

Because families tend to be far more spread apart now than they were 50 years ago, maybe even hundred years ago. And it's like, and, and I don't trust the cleaner that come into the home. 

 

[00:31:52] Brian Heater: Oh no. 

 

[00:31:53] Mehul Nariyawala : And I don't want my parents to use Your robot actually will control it remotely. But then because you have  cameras, can you gimme 32nd time lapse at the end of the day that says, my daddy is okay, or My mom is okay. Which was really interesting that, hey, there is this guilt. Inside the kids who are, uh, who are living far away from their parents, that we are not, I'm not able to take care of them and they want to help remotely. So how do we enable that? And that was really powerful insight. And then the same parents have their own kids. And, and we've heard this, I'm sure, uh, you heard this as well, where they just say, Hey, if you gimme. Toy cleaning robot, I'll give you thousands of dollars because I'm just fed off of picking up things, uh, after my kids. So there are, there are a lot of little insights like that, that are coming through and, and then we just kind of take a step back and say, what can we enable in a way that genuinely, even if it just saves five minutes, 10 minutes, 20 minutes a day, it just saves it, which is it?

You no longer have to worry about that chore. I had a customer who just came and pick up accessories a couple of days ago and tour our office and he says, yeah, my wife no longer vacuums over the weekend, and I'm pretty happy about that. And she's pretty happy about that. And [00:33:00] that's really powerful. And that's the point that we're trying to get to that what are the time and energy oriented tasks, sucking tasks that we are enabling, eliminating from your life.

 

[00:33:10] Brian Heater: It took seven years, right, to get to the point where you were able to launch the, the product and, and, and there's no, I don't think you can. Debate that it, it's a very well made robot vacuum. The reviews across the board have been really good. Um, I have one in the other room. It's far and away the best I've ever used, you know, and I've, I've, I've tested all the top level Roombas. Thank you. But it took you seven years to get to a place that had. To a certain extent been carved out. You know, that there, there are robot vacuums out there, that is the one place in the home that mm-hmm. Robots have captured. So in a certain sense, it strikes me that getting to that next level is going to be even more difficult.

 

[00:33:54] Mehul Nariyawala : The right way to say it is building foundation of a home. It takes a long time, but [00:34:00] then building various rooms on top of it does not take as long of a time. So building this fundamental building block of a foundation, which is, you know, let's go back to our kids analogy. To get to five years old, to grow a child to 5-year-old, it takes a lot out of parents, and parents are just giving them a lot of capability. But once they get that basic skills between five to seven to 10, the amount, the speed at which they acquire new skills and speed actually. The acquire autonomy is actually quite enormous and quite gigantic and quite fast. Uh, so in the same exact way, uh, adoption ability to do many things can accelerate much faster. So a new robot, we don't have to do anything for. The new version of the robot in terms of 3D mapping and navigation, all those things are just done, they're built out of the box. So it's just adding that one more skill on top of it versus adding the entire uh, uh, foundation. 

 

[00:34:53] Brian Heater: This is the distinction we need to make. Right? So, so there's adding rooms, right? And, and, and rooms. When we're talking about rooms, to me it's like, you know, how do we make the robot vacuum? How do we improve the robot vacuum? Right? How do we make it better at its job? How do we make perception better? You know? And, and again, as somebody who has continued to use the, your robot, it's like, oh, it, now I can tell it where the sink is, right? So when it needs precisely refill, it can go over to sink. Mm-hmm. But, but, but what I'm talking about is an adding a room. I'm talking about adding like a new wing on, but I'm talking about adding a wing on that's never. It's successfully been added onto a house before. 

 

[00:35:29] Mehul Nariyawala : Let's assume that you have a humanoid, or let's say you have a rolling robot inside your home. No matter what kind of robot it is, it's needs to make sure that it doesn't get tangled in wire. It needs to make sure that it's not gonna step over the dock pool. It needs to make sure that it's not gonna go commit suicide in an indoor swimming pool. And a lot of our customers do, uh, use a robot around indoor swimming pool to clean or get tangled up in castle. So all those things are just consistent and all those skill set. We are just learning right now. Regardless if you're building a robotic vacuum or let's say humanoid from scratch, you gotta teach all that stuff. The amount of having been about 250 different homes, the amount of variance home has is actually quite gigantic. People significantly underestimate how unstructured dynamic and crazy home environments are. So all those things we've learned is there. Now, let's say we are just picking a toy picking robot. I'm no longer teaching the robot, Hey, where are, what are the toys? I'm not teaching robot where the toy room is, or where the living room is. All I have to do is teach it, how to pick it up and put it in its place, and that's the ad that I'm adding. So that's one more thing. And in terms of just reveals or, or navigations, or hardware, or even the systems, or PCBs, all those things translates very, very well. From one product to the second product. So it's just kind of building it and then you have to be, you have to be conscious about it. So

I'll give you a great example. So I, so I heard from a friend that the amount, if you take the third row [00:37:00] seats out. The six seater thing that they did for, for, uh, model Y. The overlap of components between model Y and model three is 95%. And that's the thing. So if you can somehow say that, okay, here's the platform, here's the basic navigation systems we built, and how many components we can overlap and how do we build on top of it, and how can we scale? There is an ability to do that if you kind of do it rightly. 

 

[00:37:23] Brian Heater: You're saying, all we have to do is to, to, to do the part where we, we, we pick up the I'm by the way, as as though mobile, as though mobile manipulation isn't the most difficult problem to solve.

 

[00:37:34] Mehul Nariyawala : No, no, no. It it, it is, it is. It's, no, no, no. I don't. Okay. There are two things that we've completely banned at mad. Uh, which, which I'll tell. Which is, I always say that there is nothing easy. Um, that's a very good jazz chu code, so easy as a word is banned. I would never, ever, ever say that's is easy. Uh, I think manipulation, mobile manipulation, putting hands, all the stuff is gigantically hard.

What I'm trying to kind of [00:38:00] convey is that yes, it's hard, but we have built foundations. So I am, as an entrepreneur, optimistic that we can, we won't need another seven years to make it happen. That's all. 

 

[00:38:10] Brian Heater: This is something I've been thinking about a lot is like the, is one of the biggest problems with, uh, mobile manipulation right now is that there just isn't enough data, right?

Like that the systems haven't been trained in enough real, real world data and that there's so much variation. But what you built here is this machine who is. That just going around and constantly building and rebuilding, uh, your house over and over again. 

 

[00:38:32] Mehul Nariyawala : Correct, correct. It's a, so there are two approaches in robotics at the moment. One is obviously the robotics foundation model and trying to collect all this data and do the same exact thing that LMS have done. Uh, and, and that's an approach that is obviously being used in, uh, uh, in that's how it started in 23, 20 22, 20 23. Now, as you, I'm sure you heard that a lot of people are moving on to reinforcement learning and realizing mm-hmm. That closing that gap from 80 to 99 or 99 9  as RL is required. What we are doing has always been slightly different. We've always taken the spatial AI approach where we tell robot to go ahead and actually build the entire map from scratch on its own. So it's not relying on a data to say what this home is or this living room is. There are all basic stuff, but the fundamental point of view we've always had is that the physics of the world. Physics of the 3D and gravity remains the same. So if robot can build a very precision map and indoor environment, specifically mapping a localization system, then it becomes a critical.

So the analogy there is, I'm sure you've heard of in indoor environment.

So in self-driving cars you have these two amazing infrastructures. One is the GPS, which tells car where the car is located. And then the second one is Google Maps, which tells car where the road is going. Without those two information, car would be lost. But in an indoor environment, how does a robot know whether it's on the aisle five of a grocery store or aisle seven, or whether it's on the right side of the couch or the left side of the couch? That precise localization is the missing piece of the puzzle Now. To solve this problem, researchers invented this algorithm called slam, which is, which stands for Simultaneous Location and Mapping.

Mm-hmm. And theoretically it was worked out in mid eighties, but when we started working, and this was one of the unknown unknowns when we started working initially, we thought that we would use one of the open source algorithms that are available for Slam and it should just work. Turns out none of those open source algorithm are even remotely accurate enough to productize at best, they're 70 to 80% accurate. Uh, and people did productize it, but the analogy that I always use is that it's like a touch interface is pre-phone and post iPhone. Sure touch interface is existed pre iPhone, which you had to jab your finger with all the power you can muster for it to take as an input versus now the experience is just so smooth.

 

[00:40:50] Brian Heater: Yeah. Just like, go take a flight right now and use the monitor in the back and you'll, you'll be taken back to what they were like before iPhones. 

 

[00:40:57] Mehul Nariyawala : Exactly. Um, especially airing the [00:41:00] end if you go. What we realized is that all the slam system that were built prior to it were very, very rudimentary, and we essentially have over the last three, four years, and this is what it took a long time, completely redid of our own and reinvented the entire slam system and implemented it in a way that is. I wanna say order of magnitude more accurate than anything else out there. And that is done using just RGB cameras and facial ai, the approach that we've taken. And that gives us a huge advantage because now that world already has this 3D, it's much easier to learn, even RL on it. And robot can do RL in real time in the home environment on its own without us even needing to.

Do it so we can genuinely build self-learning robots, uh, versus, versus robots that come with the preconceived models and are not improving along the way. So that's really always have been the approach that once you have this 3D and a good example is, let's say you have a 3D, let's say it's in that 3D environment in your home. It finds a toy. Let's say it's a plush toy, uh, some sort of a bear if robot. And if I tell robot to pick it up. Even in the next frame, the the uh, uh, bear isn't moving. It already knows something went wrong. It does not need a reward function for the entire completion of task. Immediately based on that 3D understanding of the role, it has that, uh, clarity and that allows us to move things much faster. So there is an element of, uh, element of, um, approach that has always been there that can be. Help teach robot to experiment and play on its own and start learning on its own. And all that stuff can only happen if we did everything on the device. So not only from privacy perspective, but we as humans are not hive minds. We learn on our own. We figure things on our own. So we always felt that robots need to be, uh, behaved, uh, uh, built in the same way as well. And, and that's how we, that's why we did everything on the device. 

 

[00:42:52] Brian Heater: It's really clear how, you know, 20 years, 20 plus years of, uh, robot vacuum precedent has really laid the foundation for you to build a good robot vacuum. It's clear to me how um, all of the work that's being done in self-driving cars has, you know, been scaled. The prices have come down. This technology can now be used, um, on a smaller scale on these devices, even like smartphones, right? A lot of those technologies have been made miniature. You can use these on these devices. As you're looking towards future applications, as you're looking towards building out hardware, as you're looking towards what life looks like beyond vacuuming for home robots, what are those technologies right now that are finally starting to scale, that are laying those foundations for the future of home robots?

 

[00:43:42] Mehul Nariyawala : I mentioned giving this analogy early on, but, but I, I never got to it. So let me just start with this. So, the vision for how we believe robotics will grow, uh, is, is laid down in our, in, in our, in our opinion, on how smartphone industry grew. So General Magic, I'm sure you're familiar [00:44:00] with it, company that was started. Bandi Hertzfeld, bill Atkinson, two of the most distinguished engineers right from Macintosh team, and they tried building iPhone in nineteen ninety four, ninety five. Massively failed. And instead what we got was Nokia feature phones first, and then we got the PDAs and then we got the Blackberries, and then we got the iPods. And then consumers said, Hey, I don't want to carry four devices in my, in my pocket. So that's when iPhone became an obvious answer. And that was sort of the final product. 

 

[00:44:26] Brian Heater: And your old boss, if I remember correctly, was involved in General Magic? 

 

[00:44:30] Mehul Nariyawala : That's correct. That that is Tony Fidel's first company. And he taught us this, all of these things along the way. And he always said. Very clearly, consumer behavior is not about making us drastic jumping capabilities. They actually like to take incremental applications and they need to have a trust in a given application before you can move forward. So we've always believed that next five years are still going to be about purpose-built robot for a single task, and you will keep adding them inside home one at a time, and then five,  10 years from now it might turn into. It might turn into sort of Rosie the robot does, does everything. So let's assume you have three, four of those thousand dollars robots in your home doing various tasks. Then all of a sudden you, you'd, you'd say that, okay, fine, spending 5,000 or maybe $10,000 on a multipurpose robot makes sense to me. Uh, so how do you actually bridge that grab? And that's a really critical part of consumer psychology that, that we still have to think through and, and push and we. Technically don't approach the problem saying that, okay, here's the technology that's ready, let's go make it happen. We kind of said that we, as I mentioned, we just kind of said, okay, what is the next problem we wanna solve and let's push the envelope of technology in that direction now.

Mm-hmm. What has happened, uh, to give credit the credits to do, there have been some amazing work happening. On the reinforcement learning side, there has been some amazing work happening on the grippers, whether it's, uh, pencil grippers or, or robotic hands along the way. There have been some amazing work happening [00:46:00] in, uh, some of this foundation models and the needles they're pushing as well. So there are pieces of the puzzle that are starting to appear that says, Hey, we can solve some of this problem, uh, whether it's, uh, toy cleaning robot, where it picks up each individual toys. If you ask me what is the gap between accuracy of a product requirement versus the research, uh, that we are not sure yet that we have to genuinely build the robot, put in various environment and understand it and trade ourselves.

 

[00:46:27] Brian Heater: We're at about time, so you know, we can close on this and I'm really curious to get your take. You know, we started talking about iRobot for obvious reasons, you know, pioneer in the space. Um mm-hmm. It's been rough going for them, you know, kind of. Death, death by a million cuts. Um, the nature of that company, while I'm sure that Abot is gonna be around in some form, will change, you know, they're now, um, owned, owned by a, a Chinese brand, you know, per, perhaps that will really impact the, the top of the market where you're currently operating.

As you're looking at the landscape going forward and where Bock operates in, like how do you see over the next few years just. Robot vacuums in general, that world's evolving. 

 

[00:47:11] Mehul Nariyawala : Great question. So, so I think of robot vacuums in a two category, which is, one is a disc robots, which are quote unquote robotics 1.0.

And right now everyone is in that bucket. And actually the company that is going to own iRobot, also white labels, dyson's, new robot vacuums. So when Dyson is going in that same disk capability, so there is no. Quote unquote unique platform, uh, madic we think of as the first AI enabled, actually intelligent robot that is breaking that mold. And we think that world is going to move towards much more of a, of, of amatic. So we think of, you know, if our goal was, and we hope it is, that mad was the iPhone moment for the robot vacuums where it just comes and, and, and you realize that software can really make a big difference. So that's really how we think about it. And we just fundamentally believe that limitation in robotic vacuum is no longer hardware. Um, if you go to CES, I think some of the robotic vacuums vendors show you all kinds of hardwares and all kinds of capabilities. They're walking upstairs. 

 

[00:48:15] Brian Heater: Now. 

 

[00:48:15] Mehul Nariyawala : There's an arm on the. Yes, and, and, um, you know, they can heat your water 200 degrees or however, whatever you want, but it's really, those are all technologies.

The bottleneck still remains on intelligence and AI and software, and that's where we think we can deliver far more understanding. Because ideally, I'd love to live in a world where just like we ask Che, GPDA very semantic question, maybe you should have a. GP interface with your app and, and, and you just tell robot what to do and it does it for you. And really that's what we want as a customers. 

 

[00:48:47] Brian Heater: So, okay, so I lied. Lemme I'm gonna ask a much bigger, harder to, to answer a question to end off with 

 

[00:48:52] Mehul Nariyawala : please. 

 

[00:48:53] Brian Heater: These sorts of things are obviously always much easier to litigate in hindsight, but as, as you're kind of thinking about this moving forward, do you have a sense of when it could possibly be clear whether you've had an iPhone moment?

 

[00:49:08] Mehul Nariyawala : Ultimately the only way to know you've had an iPhone moment is if we keep growing the way iPhone grew. Uh, yeah. Right. Uh, at the same time, we are quite proud of the fact and then we didn't expect, we think this of a product is by no means perfect. As you know, we still have a long way to go. But you know, wired Magazine gave us first 10 for 10 review in 17 years.

I think that was after iPhone. So we are in that point of view. And then the other part that I really. Or we are starting to see and has been our goal is, is customers are really kind of talking about anytime a new robot company launches, they're gonna say, Hey, is this, is this the same asmatic? So a year or two years from now, if a robot is launching and they ask this question that is this as good asmatic, then we are probably, you're 

 

[00:49:56] Brian Heater: the standard to which other 

 

[00:49:58] Mehul Nariyawala : companies' the standard, are 

 

[00:49:58] Brian Heater: prepared.

 

[00:49:59] Mehul Nariyawala : The standard, yeah. And then customers are starting to do that. I've seen that a little bit on Twitter, uh, next, where people are saying that, Hey, I'm glad that this is the right place. It's moving and if we can keep improving Madic in such a dramatic way, and, and honestly, we've only shipped about. I don't know, 40, 50% of our minimum lovable product feature set. So there's a lot more to go over the next year, uh, that we're super excited about, just purely from software updates point of view. Uh, so if we can shift that, we think we can get to a point where hopefully people will say, Hey, this is, this is the kind of robot I want in terms of interaction. And, and it's not just about medic receiving rate reviews, you know?

Our app on App store is rated 4.9. Um, and and that is a big piece of puzzle. So it's the entire experience holistically on how APP is interacting with the robot and communicating. All that stuff needs to go along. And then if we can keep improving it and get to a point where the experience is perfect and everyone is using that experience. Uh, to, to get to the next level. That's when it gets to, and, and I use word perfect, but really I, you know, over the years I've come to realizations that perfectionism rag, it doesn't exist. Point nine nine nine 

 

[00:51:05] Brian Heater: nine nine. Yeah, 

 

[00:51:06] Mehul Nariyawala : exactly. But what does, this is simple versus complex, so our goal is to try to simplify as much as possible. That's the theme for this year for us. 

 

[00:51:14] Brian Heater: Alright, so let's meet back here in a year or two and we will. 

 

[00:51:17] Mehul Nariyawala : Sounds great. 

 

[00:51:17] Brian Heater: Recess the situation. Great. Well thank you so much for joining us. Always a pleasure speaking with you. 

 

[00:51:22] Mehul Nariyawala : Likewise, Brian, really appreciate it. Thank you for having me. 

 

[00:51:27] Brian Heater: Thanks so much to me, HUL and Maddick.

Definitely a syrup to watch. Uh, thanks to you as always for tuning in. If you've been enjoying the show, don't forget to like and subscribe and check out our newsletter. Over at Automated fm. You'll find exclusive interviews. More information about our upcoming Boston trip and a lot more. We have a lot of fantastic episodes coming up. We've got a interview with the CEO of iRobot, uh, plus one Robotics, serve Robotics, the head of Boston Dynamics Atlas, uh, some [00:52:00] folks from Physical Intelligence and ESI Robotics. We have, uh, we have episodes coming out of our ears. Excited to bring them to you. So stick around because we are going to be back just about this time next week with another episode of Automated.