They Invested Their Own $1.5M For 6 Years in a Robot

In this interview, Navneet Dalal and Mehul Nariyawala reveal how a ruined $2,000 rug sparked the idea for Matic, why they invested $1.5 million of their own money and spent six years building 200 prototypes, what Paul Graham told them repeatedly at Y Combinator, and why they believe market risk, not technology, is what kills hardware startups.

Navneet Dalal (co-founder and CEO of Matic): I'm Navneet Dalal. I'm a co-founder and CEO of Matic.

Mehul Nariyawala (co-founder of Matic, focused on product): Hi, I'm Mehul. I'm a co-founder of Matic and I focus on the product. We are a consumer robotics company. Our first product is a floor cleaning robot. It has taken us a long time, six years, to make it happen.

Navneet: We at Matic have raised $30 million and our investors are John and Patrick Collison, Matt Rogers, Jack Dorsey, and Naval Ravikant.

Navneet: Why do a hardware startup? Why not go do software? One of the things we realized back in 2017 when we had left Google and we were thinking of what we want to work on, the honest way of saying it is that we wanted to do life's work and we wanted to work on the one key aspect which is missing in society for the next 20, 30 years.

Mehul: I think it boils down to solving your own problem. I had gotten a golden retriever. We have a joke that my golden retriever sheds twice a year, six months each. I got a robot vacuum and I actually ended up buying Dyson's version of the robot. Turns out Dyson's version was really bad as well. The suction part of it is amazing, but the vision part of it was really bad where it couldn't find its dock.

It got onto one of the nice rugs and because it has a high suction, it got stuck on it. When we came home and picked it up, the entire patch of the rug was gone. So here's a thousand dollar device that ruined our $2,000 rug and I couldn't return it. So we knew there was a problem here. No one was solving it.

Back then, 2016, 2017, self-driving was the space to go in. Imagine trying to build self-driving cars without Google Maps or GPS. The answer is no matter how smart the car is, if it doesn't know where the road is going or where it's located, it's kind of useless. In the same exact way, if a robot is supposed to clean, it has to know whether it's on the right side of the table or left side of the table. And that didn't exist. So we thought there was a huge opportunity.

Floor cleaning was funny because this is the only type of robot accepted in our home even today. Yet the category is growing 25% year over year. So the problem is so intense that we are willing to tolerate an inferior product. We asked ourselves this question: if we can build a level 5 fully autonomous robot for indoor spaces, what does that mean? Well, if a level 5 self-driving car drives like a human, then level 5 indoor robots must behave like a human and clean like a human.

So why don't these robots constantly patrol over the home, look for dirty spots, and if they find one, just clean it? If they're truly intelligent, then they should understand that this is a rug, this is a hard surface, that there is a wine stain, and change their efficacy method. We as humans know that on rugs with frills, if we take a vacuum, it will get stuck. We don't do that. But if it's a hard surface with a wine stain, we need mopping. The robot should be able to adjust the efficacy method similarly.

But then most importantly, we realized that today's robots don't have a concept of memory. If you start a robot, it will clean for 15 minutes. If you pick it up, put it back on the dock and start again, it cleans the same stupid area for another 15 minutes. Little things like that where we questioned everything: why is the brush roll designed that way? Why is mopping designed that way? Why are there small wheels and not big wheels? That was the journey. You question everything, you come back and then you build a system, put it back together from a very first principles level.

It was a proof of concept. Can we do this with a camera? Just like Tesla. The reason cameras alone were a really important piece of the puzzle is that for every single sensor you add in hardware, you have to assume there are three software engineers on the flip side making sense of that sensor. More sensors you have, the bigger the software team. More sensors you have, the more calibration. More sensors, the more failure points. Each sensor's complexity rises exponentially.

Our point of view was that nature has given us two RGB cameras and algorithms for a reason. We don't have tons of sensors, we only have five, and we primarily rely on vision. So there has to be a way to get it done. This is where we thought, let's absorb everything into the software instead of hardware. That's a much more scalable platform over the long term. So that's how we started.

Navneet: We didn't have 3D printers back then and we built a hardwood shell in some sense. We took an existing Makita vacuum cleaner and put its motor in, then put the cleaning head and attached the stereo camera. We sent Wi-Fi across, the Wi-Fi sends the images to a laptop, does all the algorithms on the laptop, and then sends the results back. That was the first prototype within three months.

After that we migrated it to this tall black robot and we built two of these. We called them Batman and Robin. We thought there would be a hose on the top that you can also clean around. You can ask for the robot to come to you and you can clean around it and put it back. We later realized that if we are doing a very good job of the floor cleaning, maybe people can just push the dirt on the ground and let it do it, and it will allow us to build a much simpler product.

Mehul: We put in about a million and a half dollars during the bootstrap stage of our own money. But luckily when you start a company, sometimes luck helps you. For us one of those trends was that 3D printers were getting really ubiquitous. If we had to go and use someone else's machine to build parts, that would have been hard. But because of 3D printers we were able to shrink that down. Overall we've built about 200 plus prototypes and all those are not going to be part of the production unit. You can say all of them along the way got destroyed in some sense.

Navneet: We have been working together since 2006 when I was at Like.com. Mehul also joined Like.com and we were doing computer vision for a Google Photos-like product, but back in 2006. Like.com was the first computer vision startup.

Mehul: Navneet did his PhD in computer vision. I didn't know anything about it. For me it was very serendipitous that I got to Like.com, fell in love with everything computer vision. I was doing most of my work on product marketing, business side of things.

The funny story is that as we were recruiting, I would go and tell other computer vision experts that my co-founder and CEO is Navneet Dalal and he has this PhD thesis in histogram of oriented gradients, and they would say, whoa, your co-founder is Dr. Dalal. And I would always say, wait, who's Dr. Dalal? For us it was just Navneet. Navneet never shared that his PhD thesis was really popular and he was very well known in the computer vision community. That was a good revelation. That's how we started working together and that's when we connected.

Navneet: Flutter was a very geeky thing. The idea was you use the front-facing cameras in the laptop and you can control it when watching YouTube, Netflix, or listening to a song on Spotify by just gesturing play and pause. When we built the technology, a lot of people believed it couldn't be done because we were using a single small laptop camera. There are not that many computes in 2011 laptops, and how can you even run it in a live product? We very soon realized that we are solving a living room problem: this is a product which should sit in your set-top box.

Mehul: This was also a time when Microsoft Kinect was very popular and Nintendo Wii was very popular. Part of the thesis was whenever people do gestures, they turn your finger or hand into a mouse. That's a point-and-click metaphor. That's not intuitive. The whole idea was if you want to mute a TV, why can't you just shush it? That was the insight.

When we launched this app to allow users to control iTunes, Spotify, YouTube, Netflix using Mac, it was number one app in 73 different countries worldwide on the Mac App Store for about three months. We had altogether 77 million gestures performed. Technology-wise it was really great.

But we were also part of Y Combinator with Flutter. The entire time we were there, Paul Graham was still working at Y Combinator actively. He would just come to us and say, I know you guys, you guys are a technology looking for a problem to solve. Have you found one yet? He would repeat that over and over again. Initially it was very hard feedback to take.

But over time we realized that no one wakes up in the morning and says today I'm going to buy gestures. That as a product isn't there. It's a really nice feature, it's a really nice technology, but it's not necessarily a product. That's where it being part of Google or Apple or Samsung or some other platform made more sense. Luckily for us, Apple and Google were both interested in acquisition and we ended up getting acquired by Google.

I feel like we were honestly very naive the first time. The acquisition definitely allowed us to feel like we can build anything. The technology part of it we got right. But we took a year, at least that first year, just analyzing what we got wrong. One of them was this realization that no one wakes up in the morning and says today I'm going to buy this, which is we weren't solving a real problem. It was really cool technology, but it wasn't a problem.

Number two, as Navneet mentioned, hardware: we only solved the problem one way. Because we were using cameras inside your MacBook, we didn't have access to frame rates per second, we didn't have access to auto exposure or autofocus. That means you're building an algorithm for an eye that is wavering and that didn't make sense. We felt like we built a half solution. If you really want to solve a problem, we have to do hardware as well as software. A product can't be limited to just software or just hardware.

That understanding was the reason why we ended up at Nest to learn how to do hardware. It was a very deliberate move on both of our part. That was really the mindset: solve a problem, build a product, and make sure there is a clear-cut revenue model on day one so that when you start a company again, these are all problems that you're not thinking about after the fact.

A hardware startup is definitely not for the weak-hearted. It takes time. But if you are passionate about the problem and if you want to solve it, absolutely. The best way to say this is if you look at the top technology companies today, from SpaceX to Tesla to Nvidia, these are all trillion dollar companies and these are all hardware, and they're actually very hard to beat.

If you look at the NASDAQ or stock market and look at the top technology valued company, I think out of the top five, four are hardware startups. So yes, hardware startups are hard, but if you can build something amazing, you can really build an extremely valuable company.

Navneet: I would also add that you've got to be a lot more patient and ask the three whys in a row: why is this thing done in that specific manner? Is this really the fundamental reason why it needs to happen that way? Another key piece in hardware, at least our perspective, is because startups are inherently risky, you want to reduce your risk as quickly as you can. What part is the most risky part and can you find a solution to that most risky part as soon as you can?

At least in the Valley, most of us think that building technology is the risky part. That is indeed risky. But at least as a team or founders, you control it. We tend to think the most risky part is the market risk. Is there a market? Are there customers who are really interested in buying the product at the scale you need to sell it to be a profitable business? Because without profitability your business will die and along with it all the dreams and aspirations of the future.

Mehul: If you have a problem and if you can solve it just using software, then absolutely just solve it using software. There is no reason to do hardware. The only reason to do hardware is if it's a problem you can't solve without it. If you're trying to build self-driving cars, you're going to have to deal with some sort of hardware. You can't just say I'm going to build software and not worry about sensors. That's really what it comes down to: what is the problem we want to solve?

Navneet: When we started working on Matic, we very distinctly recall saying, Mehul, we need to pause on thinking through this thing. We need to think through what's the future right after we build this thing. What's next? That's when it dawned on us that what we are really building is autonomous solutions. Our mission is to enable people to save time and energy through truly autonomous robots. We see there are so many things we can automate by working on robotics, machine learning, and computer vision.