In this interview‑demo hosted by Ashlee Vance at Matic’s Mountain View headquarters, Mahul (co‑founder and president) and engineer Brax walk through the design, technology, and philosophy behind Matic’s new robot vacuum. The company positions the robot as the first "Robotics 2.0" product: an AI‑first, vision‑only system that processes all sensor data on‑device, eliminating cloud reliance and protecting user privacy. The robot’s sensor stack – five cameras, four microphone arrays, IR illumination and an NVIDIA GPU – feeds a greedy frontier‑exploration algorithm that builds a semantic 3‑D map of the home, enabling waypoint navigation, object avoidance, and context‑aware cleaning (vacuum, mop, and even poop detection). The demo showcases real‑world tests: continuous cleaning cycles, durability “reliability pens,” a water/mop test, and a humorous “poop test” where the robot learns to avoid simulated waste using internet‑sourced images. Mahul discusses manufacturing (hand‑assembled in the US with components sourced from Malaysia/China) and the company’s privacy‑by‑design stance: all image processing runs on the edge, Wi‑Fi communication is local or peer‑to‑peer encrypted, with no cloud servers. The conversation also touches on market context (most consumer vacuums are “dumb” disc robots), pricing ($1,100), shipment goals (7,000 units this year), and the longer‑term roadmap toward trusted home helper robots (tidying, laundry, dishwashing). The overall narrative frames Matic’s robot as a privacy‑respectful, AI‑driven alternative to conventional vacuums, promising continual software upgrades and future expansion into broader household automation.
Chapters
Some people say this is the best robot vacuum cleaner that's ever been made. It comes from a company called Matic, and it's packed full of computer vision and software smarts. It's also made in America. And this is potentially a big deal. Because yes, we have reached that stage of civilization where our home bots might be spying on us. I'm not prepared to go full conspiracy mode yet, but stick around and hear me out. Matic robots started in 2017 and began selling its product in late 2024, which is to say, holy hell, there must be a lot of engineering in this vacuum. And there is.
Took us a long time to get here because we reinvented everything from scratch, but we are here and we've shipped about 1,500 robots now. We've cleaned 30 million square feet. Inside people's homes that amounts to about 25,000 miles traveled. And it's available on our website, maticrobots.com. You can purchase it for $1,100, which is relatively speaking, much lower than some of the higher-end disc robots as well. And we are planning to ship about 7,000 units this year.
That's Mahul, the president and co-founder of Matic. He used to do research at Google and then worked at Nest before heading off on this startup adventure.
If I were to use the Peter Thiel analogy, we wanted Rosie the Robot that would clean our entire home, and all we got is this bumbling, stumbling Roomba. I personally bought robot vacuum in 2016, Dyson's first robot vacuum. It was probably the worst robot I ever used because it wouldn't find its dock 9 out of 10 times. So while I was at Nest, we started thinking about homes, and it just kind of dawned on to us that, wait a minute, there are 200+ self-driving car startups in the world. There are equal amounts of industrial robotics companies.. But when it comes to home robotics, all we have is this disc robots and nothing has changed.
Yeah. How did they build the maps? I had a couple of the early Roombas.
So early Roombas didn't build maps at all. It was literally like that Pong game. You just have a square and they just keeps bouncing around. If you bounce enough time, it'll cover it. But in a home environment, how does robot know whether it's on the right side of the couch or the left side of the couch, whether it's in front of a TV or refrigerator? This contextual awareness was missing.
There's an astonishing number of robotic vacuums on the market. And Mahul is not wrong. They're mostly quite dumb.
Oh my God. Move Roomba to a new location.
And so, Mada Set out to make something that actually knows your house and the objects inside of it.
Before coming here last night, when I was reading about these things, I hadn't checked in on the robot cleaning market in a while. Yes, you have some serious competition, at least in terms of the volume. Yeah, competitors. I mean, there were— I must have seen like 10 or 12 different stabs at this. You guys would contend that You know, are all of those limited and you have a different approach to, to all of them, or, or, you know, is the high-end bucket quite similar?
Matic, we genuinely believe, is the first Robotics 2.0 product, and that's AI first, that we are starting with algorithms and work back, working backwards, versus starting with sensors and working backwards.
I mean, the other ones all do look the same.
We genuinely believe that they're— the approach that they are taking, whatever their ceiling is of their approach, that's our baseline. That's where we start, and then it's much more possible.
Okay, you You mentioned 5 cameras on the device. I mean, tell me all the tech specs of the product.
The entire robot, the sensor stack is what we refer to as this black crown. This black crown is right here.
Okay.
And what we have is 2 cameras in the front right here. Then there are 2 cameras in the back. There is one on the top. So 5 cameras. We added a display because most of the time robots are doing something and we just wanna know what the hell they're doing. And then this is actually IR lights. So nighttime illumination.
Okay.
So it can clean at night, but IR spectrum is invisible to us. So for us, it's dark, but for a robot, it acts like a headlight. Yeah. There are also 4 microphone arrays on this, on this side right here. And then this is the NVIDIA GPU. Everything goes directly from the cameras to this GPU, gets processed, and then gets there. And that's it. That's the extent of sensors we have. There is no proximity sensor, there's no cliff sensors, there's no LiDAR, there's nothing in this robot. It's purely vision-based solution.
So all the brains and processing is up here, and the internal is storage for the stuff that gets picked up, and—
That's correct.
In the vast spaces between ambition and execution lies the architecture of tomorrow.
Brax.
To see the robot in action, we ventured into Maddox's office and test lab.
And there are just 4 of them, 2 on each side.
1, perfect. 2, great.
Just, and now pull it out, open. That's it. There we go. And then press the power button.
Hello, core memory. I like it. So you get it and it just drives off.
You turn it on, it says, hello Vance family, and it will.
Drive out.
You guys get everybody's name when they order and put it on the top.
First thing you do is you say start mapping.
It will—
yes, it's capturing all of our information and literally just building the map on the fly.
Is there like a method to what it's doing?
It's very much a greedy algorithm saying that as long as there is an open frontier, I'll keep doing it.
Once the frontier closes and I have nowhere to go, I'll go find the next biggest open frontier and then go explore that and next one.
So that's how it does it.
And so it saw, it kind of like saw the pole.
It saw the pole. It's already capturing the semantic information. So it knows that this is a carpet that it's on.
And the carpet, it's visual.
It's all visual. Now what you can do, if you just long press on this side, just long press.
Yep.
Yeah. Oh. And then say, go here.
Oh, I'm giving it like the waypoint.
Yeah.
And you can say summon. So it's like the same thing as summon or go here. And now it will just say, and it makes the entire beeline to that exact spot. Okay, so it's going that area.
You're sending it to the kitchen?
Yep, I'm sending it to kitchen.
We're about to do a water test. I had a Roomba that did do the liquid and it was terrible. So I'm curious to see how far we've come. So like right now it's vacuuming and then if you want it to clean a certain spot with the mop, you tell it to mop in this area.
In the area and in—
And then it's cool though that it switches all in the one head.
All in one head.
Automatically. Yeah.
So what it does is first is it's going to start saturating the mop and start going back.
Is it going to come backwards?
Yes.
For mopping?
Yes. Not only we want to mop, but the other thing was it moves backwards, but mop roll spins forward. So it actually creates far more friction. So yeah, as you can see, it just captures everything. And whatever gets wheels in it, it also gets used.
So that's the advantage of doing it this way.
Wow, he dodged the camera.
And then he says, "I'm done." So he says, "I'm going to go back home." He's returning to go. Yeah. So these are our reliability pens here. So right now these are only hardware things there, but we have all kinds of rugs and carpets and hard surfaces that we put in here to see how it does on various surface kinds. We also put in all kinds of dirt in here. So you will see some of the wooden chips over there. It's cleaning. And the idea is that we just let it run 24/7 and it docks, undocks, docks, undocks, sees wires, avoids them, and it just continuously keeps cleaning.
So the employees are in here with these just going all the time?
All the time, yeah. So this one is trying to test whether in that heatsink, if it turns a little too hot or if it melts any of the stuff.
So when you run the chip as high as you can, if the heatsink's good enough, if it thinks good enough.
So it shouldn't have an issue, but we're just double-checking.
I can't believe we live in a world with vacuums running neural nets. You know, that's right, that's right. On the vacuum.
That's right.
Like, the amount of computing this would have compared to the Apollo rockets is significantly higher.
Yes, yes.
This one— so this one, all these stickers come—
all the stickers, some of these robots, you get a cat and a dog, and you get a cat and a dog, and different versions that we create. And you, you get the googly eyes, you get the stickers, and you can decorate it however you want. A lot of people like to name it, so we also let them put the little label and put their names so other people can talk about it as well. Some of the software, so we had to build a lot of observability and end-to-end testing software ourselves just to see what a robot is doing.
We can visually observe, but then we can test it in a virtual environment first.
So this is sort of like our—
this is—
Mission control.
Yeah, mission control.
So this is a software which we've built, uh, and this sees every layer of the robot. And where does it start? It starts from images. These are like the 5 images the robot gets, and as the robot moves, you can see that this live updates. And this is the 3D map it builds. Yeah, so you can see that this color represents rug or carpets. This color represents hardwood floors. Like, a true robot is intelligent because of its semantic understanding. And the more you understand what a wire is, what a toy is, the more you can behave like a human. So this is where, like, the intelligence comes from.
Since the robot has software at its heart and can be ever upgraded, there are always new frontiers to explore.
I know we were gonna do the poop test. Do we do that here?
Uh, we do it over that side. Let's, let's go try it. Uh, it's still a work in progress.
Let's go test some shit. It's the poop test time. He's finding the chief poop engineer. Shitty job, but it's essential.
And we've enabled the poop algorithm on top of it. Do you know? Can we do that please? Okay.
Will it touch the poop?
Let's see.
It's got a poop gauntlet.
Does it show up?
Oh, look at that. Pinpoint poop turning.
So now it's kind of captured two, one on the front and one on the back. It's doing a pretty good job avoiding it. So there we go.
But does the pink spot mean random object or is that specific?
No, no, you'll see specifically that's exactly where the poop shows up here visually.
But like it thinks it's a poop or it thinks it's us?
No, it does.
How did you train this model? How many poops did you have to show it?
Um, luckily there is a lot of internet images.
So like some engineer here, his entire internet algorithm will never be the same.
Yeah, it will never be the same.
So the cost is not actually to clean up the poop, it's to avoid the poop.
To avoid the poop.
Cleaning up is insanely hard.
Milk spilled, it can mop it up. Or it can do cereals, but it can't do milk and cereals together.
Okay. On a more serious note, the robots are being made here at Matic's headquarters in Mountain View, where they are tested and then assembled by hand.
Can we peek at them working a little bit?
Yes, yeah, absolutely.
So everybody's just here building robots.
Everybody's here building robots. This is sort of our production area, and you can see—
Everyone's busy.
Everyone's busy. So this is the kind of subassembly now.
Yeah.
We get, I believe, from Malaysia now.
And instead of building the whole thing now, you can see the cleaning head is also actuation.
So instead of getting things, this actuation system is pre-built as well.
Yeah, cool. Part of this is about the privacy things and everything.
Yes.
But we also source components.
Yes.
From Asia.
Yeah.
How do you think that through?
So the honest answer is that I would love for our government and Trump administration to figure out how to source the raw component. At the moment, for better or worse, it is coming out of China.
Yeah.
There is just no way around it.
Most of our robot friends are, of course, being built in China. And well, people say things about bots from China, like that images of your bodies and homes are being piped over to the PRC.
I don't mean to go like full conspiracy mode, you know, Reddit, Reddit chat board, but you know, there, there are a lot of these that are made in China. I think some people, maybe they don't think about this, maybe they should, you know, that you've got this, this thing mapping your house. I have some friends that claim they have proof that these images actually get sent off, you know, and, and so is that something that you guys, um care about, that you thought about when you're making this product?
Absolutely.
Actually, that was one of the core principles. And when we started out as well, that indoor world, especially homes, are our sanctuary. And for families, it's such a great place where you can be whoever you are without worrying about it. So we started with this thesis that families shouldn't have to jeopardize that privacy just to get floors cleaned or just to get dishes washed. And so we built the whole thing private by design. So while Matic has 5 cameras, the entire processing happens on the edge device. There is no cloud compute whatsoever.
So the image processing of the house happens on the device. On the device. And then what's going to the app?
So even app to robot connection, if you're on the same Wi-Fi, happens through local Wi-Fi.
Okay.
Even there, it doesn't take a round trip. If you try to control it remotely using internet, then it just does peer-to-peer direct connection between your app and the robot using encrypted end-to-end tunnel. There is no middleman, there is no server, there is no cloud. I think it's time to sort of get— it's at least in a home space, a lot of manual or sort of a non-digital appliances that are not connected to internet. They don't take any data.
Yeah.
Refrigerators, TVs, they don't really take any data. So can we get back to that sort of a trust with customers that we're coming into your home to help you but not breach your privacy? Yeah.
So there you have it. You can walk around naked in your house with no fear. For these are robot patriots, and they shall be ever upgraded in your favor.
We don't think of it from, hey, we want to build a humanoid or taller robot. We just kind of take a step back and say, what problems users want to get solved next? Ultimately, you definitely want to get your laundry done by a robot and dishwashing done by a robot and all the chores. So the step we think about it is first we build a trusted floor cleaning robot. The next step is to say, can we build a trusted helper organizer tidying up robot that does some of the items very, very well? And then the third step is to say, can you build a trusted housekeeper robot?
And also poop.
I think this one—
We can throw like a lot of poop out.
Yeah. So now I'm going to stop it, unfortunately. This is why we need to test it because it will make mistakes and we don't want it to avoid it sometimes and not avoid it the other times.
Yeah, now it's kind of detecting the—
Honestly, it detected poop far better than I ever imagined.
So you were in the middle of toggling it on as it ate the poop.
Ah, great.
So we need to turn it off. It's on now if you want it. I'm like, oh wait, it was seeing it and it's appearing. So there we go.
