In this AMA segment, Mehul Nariyawala recounts his journey from leaving Nest/Google in 2017 to founding a home‑robotics venture. He describes early disappointment with the Dyson 360 i vacuum—its vision system failed to locate its dock and it damaged a rug—highlighting the inadequacy of existing robot‑vacuum perception. Observing that indoor robotics had been built “upside‑down,” he argues that, unlike self‑driving cars that rely on GPS and maps, home robots need a human‑like perception system: two RGB cameras and on‑device processing. He explains that a vision‑only, on‑device approach addresses three core challenges: privacy (no cloud streaming of home images), low latency (critical for safety), and the need for dynamic, self‑updating maps in highly variable indoor spaces. These insights shaped the team’s decision to redesign the core localization and mapping stack from scratch in 2020, aiming for Level‑5 autonomous robots that can understand and act in homes as humans do.
Good. To start, we have a question from Ken Lee. It's, what's the most surprising obstacle you've encountered in making robots truly autonomous in unpredicted home environments?
That's a great question. So I'll actually take a step back and start from the beginning and give everyone a background on what we've been— why we started and what we had to go do it and some of the decisions we made. But really, in 2017, we left Nest and Google. Because we had been at Nest, we had been thinking about homes. And it sort of just peaked— we realized that, hey, there were 200+ self-driving car startups in the world at that point in time. There were equal amount of industrial robotics companies. But when it came to home, all we had was Roombas or Disc Robots. And initially, I had this inkling that Dyson eventually will come into this space and do a really, really great job. with the robot vacuums. So I was actually one of the first ones to order Dyson 360i robotic vacuum back in 2016, even before we left Nest, and used it and realized that— and Dyson has an amazing suction power and battery and vacuum components in itself. They're phenomenal mechanical design and mechanical company. But when it came to algorithm, that vacuum was probably the worst vacuum I used, robot vacuum I used, and simply because 9 out of 10 times, it wouldn't wouldn't even be able to find its own dock. So that's when we realized that computer vision approach that Dyson was taking was potentially the right one, but they were really, really lacking behind. The second thing was also that it got onto— Dyson has an amazing suction power, and it got onto one of our really, really nice rugs. And because of the suction, it got stuck on it. And while it was stuck, it actually kept the vacuum on for 45 minutes, and we weren't at home. So when I came back home, look, found the robot and picked it up, I realized that the entire patch of the rug was gone. So that's when I realized that actually, even if it doesn't— it didn't even know that it was stuck on the rug. It didn't have any understanding, and it kept going on. It would really ruin rugs itself. And second part was that maybe the suction that Dyson talks about isn't needed. So those were some of the things that were back of our mind. And it piqued our curiosity that, look, why isn't anyone tackling this home chores? home robotics problems? Why aren't there any robots there? So Navneet and I spent about 2 months just digging into robotics itself. Prior to that, we had no background in robotics. So we just spent a couple of months doing it. And very quickly, we came to this conclusion. And this was our conclusion that the entire field of indoor robotics, and specifically robots in homes, is built a bit upside down. It was like putting a cart before the horse. And The reason I say that is because even though self-driving cars requires all these sensors and vision and amazing algorithms, there are 2 things that self-driving cars have, which is Google Maps and GPS. So at least they know where they are located onto the road and which road and where road will be going. On the flip side, when it came to indoor robots, the analogy we realized is that in 2002, Roomba did a phenomenal job when there was no AI or technology available, did a phenomenal job of building a robot. which had sort of this blindfold on it, would just bounce around and clean our home. That was a really great way to start robotics. But then along the way, more sensors got added, but those sensors weren't really converging into intelligence. A robot didn't really understand which part of the home they were, whether they were in the kitchen or dining area, right side of the living room or the left side— I mean, right side of the couch or the left side of the couch. So we just realized that, look, if you're going to build fully autonomous indoor robots, then we should give it the same perception system that we as humans have, which is 2 RGB cameras and a bunch of algorithms. And this is because— and this is the insight that Navneet and I have been operating on— that indoor world, and specifically homes, are built by humans for humans to take advantage of, fit our perception system. They're optimized for the vision-based system. So if we're going to have Level 5 robots, and definition of Level 5 being that if Level 5 cars drive like humans, then Level 5 robots must behave, navigate, and manipulate objects, and do tasks on our behalves like humans. So if that's what they're supposed to do, then they need to have the same perception system as us. So that's where the decision came that, hey, we're going to actually do the whole thing on the device using— sorry, using cameras only, and see if we can solve the problem by giving a similar perception. perception system. But along the way, we realized that if you're going inside homes just with cameras, then privacy is an issue. Latency is an issue because you don't want a robot that says, oops, I fell down the stairs, or I fell on your pet or kid because Wi-Fi is weak. So latency is an issue. And then the third thing we realized that the indoor space is actually far more dynamic than the outdoor space. So you can't just have a static map like Google Maps. You need dynamic maps that can map itself. So your robot that maps itself in an environment. And that's where we decided that we're going to do the whole thing on the device so that robot can always update itself and observe the changes in the environment, delta in the environment, self-correct, as well as preserve the privacy and be really, really low latency. So that was the thought process. That's how we started thinking about it.
Yeah, no, for sure.
And what was the obstacle? Well, everything about robotics. And now I like to say that if hardware is hard, you know, there is a cliché in Silicon Valley that hardware is 10x harder than software. And I'd like to say that now that robotics is probably 10x harder than the hardware itself. So it's a much harder thing because we had to do everything from ME and mechanical design to platform to OS-level optimization all the way to algorithms. So just putting everything together in a way where robot just works and everything behaves in an expected fashion is quite a challenge in itself. And then the second part was certain assumptions we made along the way. So one of the assumptions was that, hey, SLAM, which stands for Simultaneous Location and Mapping, which is the precise localization algorithm, theoretically SLAM was solved in mid-'80s. There were a lot of papers published on it, and there was this assumption that SLAM is a solved problem, and there are tons of open-source libraries available as well. So we started with the assumption that we'll be able to use it. But when we started digging into it, we realized that none of the open-source algorithms had the required accuracy. The best analogy for that is that, you know, touch interfaces were available prior to iPhone, prior to smartphones. But back in those days when we used stylus, those touch interfaces weren't really that great. Or if you've used ATM or early TVs behind the seat in airlines, you know that you had to jab your finger with all the power you can muster to make their touch interfaces work until iPhone came out and made them really, really amazing. In a similar way, SLAM systems were just about 70% to 80% accurate, and we actually decided in 2020, I want to say fall of 2020, that we're going to rewrite everything from scratch. And that was one of the obstacles that we didn't anticipate having to solve, but we did.
