How To Design The BEST Robot Vacuum Cleaner In The World (Matic Robots) - w/ Anshuman

So far we've talked about big wheels, big bag, big ground clearance, brush roll that doesn't get stuck in here and then you also have your self cleaning model right about the compute happening within the device itself. How does that happen? Are there operating systems for this now? Is there like a vacuum OS or how does that work? There's a lot of people looking to hire good hardware engineer. I know because I'm one of them. I'm Nam Pandi. This is the ready set do podcast and in this episode featured not expert is Anuman Kumar.

featured not expert is Anuman Kumar. Anaman is the head of hardware at Matic Robots, whose $1,100 smartome robot has been rated by Wired magazine as the greatest robot vacuum they've ever reviewed. In our discussion today, Anjuman takes us through some of the seemingly unintuitive design decisions that he has helped make during his time at Matic, starting out as a design engineer 7 years ago. Now, if you're interested in what truly goes into designing worldclass hardware products, you might be surprised by just how often safe design decisions can be a trap.

safe design decisions can be a trap. When you map your space, it will automatically label all these different things. If it sees a wire in the scene, it will label it as a wire. You can semantically understand where, let's say, there's like a dog poop. Do you make a prototype and then it goes to the factory for mass production? like how does that entire thing work? In line with our theme of learning from somebody that's just two steps ahead of us instead of an expert, my goal with this episode is to spotlight the fascinating world of hardware design and encourage anyone considering a career in this incredible domain to take the first step which is by building something.

which is by building something. Debugging a hardware system is a lot more slo than debugging code. It's a physical sport. You can't be sitting in a chair all day. Subscribe on YouTube and any of your favorite podcast apps for weekly episodes featuring not experts and daily clips from those episodes on Instagram and YouTube. And now without any further ado, here's Anuman. Welcome to the Ready Set Do podcast where we learn from journeys of not experts who are just two steps ahead of us. Anan welcome. Thank you. Um, so I want to start off with the wired mention here. So first of all, congrats on that.

here. So first of all, congrats on that. How did that come to be? And I the only reason I bring that up is because that's how I came across Matic, right? So just want to start off on, you know, like a lighter note here and yeah, just go through does this stuff matter to you? And so when something like this comes in, does it move the needle for a company like Matic or not really? Is it just like empty press as they say? Sure.

just like empty press as they say? Sure. Um I mean I think it doesn't. It doesn't. for us. Um, it it's a big deal uh because it's a first uh step in in sort of like getting external uh reviewers to look at our product from a holistic standpoint and then and then make a assessment of it. Um right you know to answer your previous question uh we obviously haven't been working necessarily just to get press reviews.

necessarily just to get press reviews. Um, of course, yeah, for us, we looked at this space and I've I've been working at Madic for 6 years, six and a half. Um, and this entire time, we've been obsessing over what the customer experience is going to be. Um, we've tried every single vacuum robot that's out there, spent a lot of time with it in our homes, in the lab, uh, and we have questioned every single design decision that potentially, um, you know, could be done differently and and in a lot of places we've ended up in in uh, you know, with a design which is it looks very different from what's out there, right? So, a clear example would be how tall our robot is. M and so that

be how tall our robot is. M and so that is something I'd love to discuss if if we you know have time for that cameras for our entire sort of like external sensing stack. Uh we have these big wheels. We have a particular kind of brush roll. We use bags instead of bins. Uh we also have this mopping system which takes up water rather than just spreading a dirty rag around. So there's all these different design decisions that we've taken over time and you know for the last six years it's just like okay this is what we believe is best for the customer because we put ourselves in the shoes of the customer. Uh we try to have like a deep empathy and and just take the design decision we believe is best and go with it. So when when you're

best and go with it. So when when you're doing that for a long time people just look at your design choices and they're just like oh this is weird. Why are you doing this? And even today, obviously, you know, a lot of people look at Maddox's design and and maybe at a first sight, it appears to be an odd design, but I know that everybody who's tried product actually in their home, you know, a large large majority of them are like, "Oh my god, this is amazing." And for somebody like you know uh you know the the the particular reviewer from uh Wired who looked at our robot and for them to come out and and sort of and and the article they did a great job too.

the article they did a great job too. They went very ind depth analyzed it from 20 different angles and to come out and comment on all of that was very very satisfying. U and and thankfully a lot of it came out looking positive. So it definitely helps. Yeah. I mean I would say 10 on 10 is as positive as it gets. Right. So I mean for us I think internally uh this is a little bit of a insider thing but like uh when 10 on 10 came out a lot of us are like but wait we know we can do so much better like we have so many things on this product that u we're looking forward to improving. Um

u we're looking forward to improving. Um but at the same time we know it's it's it's a good product already. It's way better than everything that's out there. Uh so it does feel great to have that you know external sort of like you know nod of approval. Absolutely. And yeah, one of my goals here is definitely to, you know, deep dive into the, as you said, counterintuitive sometimes design decisions that make MATX so different and so unique. But before we kind of go there, I do want to, you know, not yeah, not really like take a step back, but I want to focus on your journey at Matic.

want to focus on your journey at Matic. Like you said, you've been there six years, you've seen so much. So do you mind just um providing like a contrast of the arena like this specific you know hardware maybe hardware is too broad but yeah like the robot vacuum/cleaner arena when you first joined versus right now and I'm really just looking for you know a snapshot of what was it like at the time what were the technological limitations that have since gone away um and like you know just really how far have we come just from a um software and hardware point of sure um So I was I was at Tesla before this. I really wasn't thinking that much about robotic vacuums per se. Uh but I was thinking about

per se. Uh but I was thinking about robots and uh it was an intentional decision to uh step away from what was a great role at Tesla. Like I was having a great time, great company, building amazing products. Uh but it was an intentional decision to go work on consumer electronics and particularly focus on um indoor robotics. But then eventually I came across Madic and I once I heard of the vision and the actual problem we were trying to solve at Madic I was hooked and sorry just to really sorry what was this exact problem I curious to know that yeah yeah yeah I mean um I don't know if you ever saw the Jetsons growing up no I did not sorry

Jetsons growing up no I did not sorry particular uh uh cartoon program that used to air I'll pull it up Yeah. Yeah. Uh, so Jetsons and and if you look up Rosie the robot, uh, it'll be this cartoon that'll show up and, uh, Okay. So, I don't remember the exact year it was set in, but maybe 2015 or something like that. And, uh, Rosie the robot is this is this robot that goes around and does everything in their house. And and some of it is pretty comical because the robot will go and pick up the uh dad from the bed and brush their teeth and then like you know put them in a tube and they're automatically suctioned off into their office and they just show up right obviously that's a caricature

right obviously that's a caricature that's an extreme thing. Um but there's always been this idea that robots are going to be amazing and robots are going to do a lot for us. But the learning that we found was if you talk to people uh about their experience with these robots, robots, universally universally not big fans of the product experience, right? If you can look at there's a there's a popular metric among product managers, NPS, you might have heard of it. It's called net promoter score.

it. It's called net promoter score. Correct. The net promoter score is simply a metric of how likely is a person to recommend a product to somebody else. Right? And uh from what we've seen, it's pretty bad for for uh iRoot products. So it's it's minus one, right? Oh wow. Uh if you look at female users only, it's minus 30. Oh yeah. And this is not just iRoot story. This is this is every other competitor that's out there. And they're not much better.

out there. And they're not much better. Right. And um so we come in and we look at this like, okay, this this product is almost universally hated. Nobody really loves their robot vacuum. Yeah. Right. Except except for the like maybe the 5% power users who would make anything work because they're technically savvy and and they they almost find some joy in tinkering with this thing and so they're able to make it work. It's definitely not working for like the super busy or the or maybe even people who don't want to tinker with their products as much.

to tinker with their products as much. They just want it to work. It's not working for those people. Yeah. And those are the target audience, right? That's the whole point. That's why it exists and it's not even working for them. Yeah. And and it shows up in when you look at like market adoption like you know it's 10% users have a robot vacuum if you look at all households. Gotcha. So so looking at it we our early insight was hardware is probably not the problem and the the large problem is on the perception and the planning side.

the perception and the planning side. And what do I mean by that? Okay. Yeah. Exactly. Perception means how well does the robot understand the world around it. Mhm. And planning would be once you understand the world around it, can you take actions in that world? So why is perception bad? Well, perception is bad because robots mostly didn't use cameras. A Roomba when it came out, it used a bump sensor. So it would go bump into a wall, turn around, then keep going until it bumped into another wall.

going until it bumped into another wall. Correct. Yeah. Then eventually people added like, you know, cliff sensors. People added like uh what is called a single pixel LAR. It's essentally a laser. Gives you depth. I I think of it as a walking stick. uh for for somebody you know uh who doesn't have eyesight and like a blind person and they're just like trying to figure out like what the depth of thing is. I mean you can you can get around but you're not going to be able to play an intense sport for example right because you stretch the analogy a little bit. So and then eventually people started adding cameras. However, when you add cameras, uh that's only the start because vision even in humans is part hardware, part software. Uh if you look at the

part software. Uh if you look at the human anatomy, uh in the brain, a large part of the human brain is dedicated to processing the images your eye capture. That's right. The visual cortex, right? And people can add a camera. If they don't have the right software, they are not doing much with the images that are coming in. Interesting. So, is the intent to add a camera just to kind of map out the uh environment or is does it go deeper than that? It's a good starting point. It's a great starting point, right? So, what we do is we have five cameras on the robot. Mhm. And there's um two cameras facing forward, two cameras facing backwards, and one kit facing up. And interesting the the two cameras thing I have to explain a

two cameras thing I have to explain a little bit. This is how human vision also works, right? So, why two? Well, when you look at an image from two vantage points, Mhm. you now have enough data to calculate the depth of any pixel in the scene. I see. So, you can do the thing a lot of us do it as kids. It's like if you close one eye and you look at your finger and then you close another eye, you will see the finger shift. Yeah. Exactly. Right. And that is because what your brain is naturally doing is it's looking at those two images and first thing it's doing is combining it into one image. Correct. So you perceive the world as one image. You don't constantly keep seeing two video streams in your in your mind. Um but

streams in your in your mind. Um but then second it's it's also giving you a understanding of how far every single point is in your field of view. And so Matic does a very similar thing. It has two cameras which are very akin to two human eyes. And based on that we are able to figure out what is the depth of every single point in the scene. So that's the first thing we do. We do we build what is called a depth map. So quick question on that. So does that entail that when say it sees a table maybe this table has legs that are wider than most most tables, right? So is it just is the intent to map out that distance and then figure out whether it can go through or if if it should just

can go through or if if it should just avoid it all together. Is that kind of the right thinking on this? Yeah. No, you're you're you're almost there. So it's initially it's just depth which is how far is every single thing from where the robot is standing. Okay. I see. I see. Doesn't even know this is a table. It doesn't even know this is a wall. It just knows there's something there. And just an object. Yeah. Yeah. It gets you to what is called an occupancy map.

to what is called an occupancy map. Totally. You could also look at it as a traversability map which is like I can go here because there's nothing here and I can't go here because there's a table here. It doesn't know it's a table. It just knows there's something. Okay. So that's the first layer. Yep. Right. The the interesting thing about it is uh it's 3D. It's not just so in a single pixel LAR, which is what most robots have. Mhm. It's one plane. It's one laser that spins, right? So you only have a very sparse amount of data. A robot is a 3D object. So we have 3D data all around it, right? Then what we do is the second layer of it is what we call

the second layer of it is what we call SLAM, simultaneous localization and mapping. So the way that works is it identifies specific points in the environment around the robot. This is how humans do it as well. The underlying logic is the same, right? Like imagine walking into a new space. You will identify and you you don't do it while thinking about it. This is all subconscious, right? You know, so so we work through uh the 3D and then we also work through slam. There's a third layer to it. Okay? And and that's where things start to get very very interesting. We call it the semantics layer. Lulu semantics layer is when you can label things and start to apply logic on it.

things and start to apply logic on it. Uh so how do I mean? So currently this feature already exists in matic when you go and map your space. Mhm. Uh it will know what is the kitchen. It knows what is the bedroom, it knows what is the bathroom, so on and so forth. So when you when you map your space uh it will automatically label all these different things. That's a high level semantics.

things. That's a high level semantics. load semantics. If it sees a wire in the scene, it will label it as a wire. I see. And where that is when you So we we're still talking just about perception, right? We haven't gotten the planning stage yet, but when you go do your path planning, you know where the wires exist in the scene. So you can avoid all the wires because wires are big hazards for robots and get tangled up. Right. Right. you you can semantically understand where uh let's say there's like you know a dog poop and so that you don't go over it so you don't spread it everywhere. It's a big deal because if you do that you know

deal because if you do that you know you're a done deal. The robot has to be deeply cleaned or or discarded. Yeah. And on top of all the damage it might do in a person's home right. Uh and we do we use this very extensively. We understand what is carpet, what is hardwood. We understand uh in the future we're going to very quickly start understanding, hey, this is a cat, this is a dog, uh this is a person, you know, and maybe we even start understanding this is user one, this is user two, user one has these preferences, so let's clean it this way, user two hasn't this preference, so let's clean it that way.

preference, so let's clean it that way. And and this ties it back to why do we use cameras? Because cameras is the only sensor where you have all this depth of information. So you can do all these things. Imagine just having a single pixel light up. You don't even know what's going on. You're blind essentially. So you can't do all these things, right? Um so that's that's the whole perception side of things. Uh and planning of course builds on the perception, right? And and so that's how the robot moves around and takes actions. So real quick just to uh you know just a question that came to mind.

know just a question that came to mind. What part of this is stuff that other robot vacuum cleaners don't have? So, I think every robot has to have some some level of perception and some level of planning. Otherwise, it's it's a it's a machine that stays in a corner, doesn't do anything. So, when you look at the map that Madic builds, it is a you know how those panorama photos work.

you know how those panorama photos work. If you on your phone, you take a pan, right? That's what we do to your floor. We stitch together all the images and we provide a photorealistic uh almost like you know floor. Yeah. As a 3D model, it renders in your app. Why that's helpful is because when you look at it, you as a human being, you know what your kitchen looks like. So when you look at it in your phone, you can identify, oh, this is where my kitchen is on the map. So if you need to send your robot to the kitchen, it's very easy. Is I what other robots build is a just like a bunch of boxes. It looks

just like a bunch of boxes. It looks like a blueprint. M and you can Google this and you'll find like blueprint looking maps and it's very difficult. It's all an abstract shape. It's like which one's the kitchen, which one's the bedroom, which one's So that's one thing. The other thing is they don't have the semantics layer. Mhm. They don't have So they can't tell dog poop from you know just water or something like people are starting to do it. Where it gets very tricky is the data is not processed on the robot for other other companies. companies. So they see take the data from your home. It might be a camera, right? The the data gets uploaded through the robot through your Wi-Fi to a server that's sitting somewhere else. That's where the inference is running and then the inference is sent back to the robot.

inference is sent back to the robot. Interesting. There's there's a few problems with this. Yep. The first biggest problem is privacy. If if you are sending that data, if anybody, you know, works their way into that pipeline, they can get access to the images from your home. Mhm. The second problem is latency. It's just generally going to be slower running on device, right? And the last problem just from like you know a a business perspective is is you're going to have to maintain those servers and that's also problematic in the long term because you will struggle to make the product accessible for a lot of people. How do you bring the cost down?

people. How do you bring the cost down? It's it's just a generally ineffective architecture. Um and so what we do is we do all of this compute on the robot. So it's 100% private. none of the videos and images are ever leaving your device unless you give us explicit permission uh for a debugging scenario or something like that. Um and you know latency is very low because all of it is done on the device. We're not sending your data over Wi-Fi. So we don't we don't need to wait for that. Uh and then we're also we don't have any server costs. So so we don't need to keep charging you uh just so that you can run your robot for

so that you can run your robot for example. So, so that's that's where the difference comes in. And then everything else kind of has to be experienced, right? Like it's like people can say like, "Oh, I'm a better driver than you." It's like, "Okay, I can believe you or I will actually need to sit in a car when you drive the car, right?" Like there many people who drive cars, but they're not all built the same. That's kind of how this is, right? You need to sit in the car when they're driving the car, and then you figure out like, "Oh, this this person is actually a really terrible driver. They always break late," right? So, so when you use a Matic as compared to other robots, you will notice that, hey, it always, you know, turns away from me before it would

know, turns away from me before it would bump into me. It doesn't matter if I step right in front of it. It doesn't matter if it's dark, if it's like light. Um, it it's fast. It's able to cover all of the area. It's able to like, you know, make sure that wherever I ask it to go, it goes exactly there. So, it's all these like small little things uh that add up over time. And in my experience, I'm trying to use other products, it gets pretty annoying if it doesn't understand what I'm asking it to do. Yeah, for sure. And it takes just one bad incident, right? One bad incident and trust is broken pretty much forever. And there's Yeah, that was our finding. Exactly. That was our finding when we talked to a lot of people. Uh people mentioned that, you know, they

people mentioned that, you know, they used it for a while and then the robot ended up in a cupboard and they never used it again. So that's our biggest goal. How do we create a product that people just want to keep using cuz it'll add so much value. Now quick follow up on what you said about the compute happening within the device itself. How does that happen? Like you know I I understand like there's some sort of chip involved probably some processing that's going on and since you are in hardware so I'm you know obviously hoping you'll be able to share but are there operating systems for this now? Is there like a vacuum OS or how does that work? I mean we we we basically run all

work? I mean we we we basically run all our algorithms on Nvidia uh chips. Um and and that's you know a very very like um well debated well discussed design decision uh because it really informs what you can and cannot do. Um and that's one of the big reasons that this product is even possible which is compute in general has become a lot more accessible in the last 10 years and so the amount of compute power uh that is available for example in these GPUs for us to uh put on a robot like this and it doesn't require a lot of power it doesn't require you know too too much cooling the form factor is small all of those things add up you know if you went 10 10 years ago these kind of computers

10 10 years ago these kind of computers were not even available Yeah, exactly. And so we are able to use Nvidia hardware and that's how we we run all of our init. I see. And you mentioned that there can be limitations based on what you use. So can you just give us an example of what one of the that those limitations is that in that is enabled now because we use Nvidia atmatic. Yeah.

now because we use Nvidia atmatic. Yeah. I mean things like memory for example like how much memory do you have? uh you know all of this like um you need to uh even if the data is not going anywhere on a server you need to build a map for a user and save it on the device the device needs to remember what the memory was right or what the the maps look like and so it'll be navigating in those maps and uh all of this is high definition 3D model so that's one thing um also just every single thing involves some form of AI so you remember how I was talking to

AI so you remember how I was talking to you about um the depth detection part of it, right? Yeah. Look at the left scene and the right scene. Mhm. The way that works is you identify the same pixel in the left scene and the right scene and then you figure out what is the delta of their position in the left and the right scene. Oh, right. Okay. But in order to identify what exact pixel matches what exact pixel in the two scenes, that is a neural network that needs to be run.

neural network that needs to be run. Neural network is in my understanding a bunch of matrix multiplication. Uh there's a bunch of weights that you need to have and very comput intensive. It's it's just a lot of calculations the the sort of hardware needs to do and that's just one NN that we run we run a lot of this. So every single semantics thing like identifying whether this is wire or this is not wire identifying whether this is dog poop or not dog poop this carpet is this hardwood every single thing is a neural network. That's where the AI component comes into it. Right.

the AI component comes into it. Right. Right. And all of the those calculations are very very compute intensive. What happens is if you don't have enough uh computes, right? What happens is your frames per second gets limited. So you can only process let's say one frame per second. If you have a very uh or if you don't have a powerful enough computer and what that shows up as is your robot is not as snippy. It doesn't respond as quickly. It will run into you. uh it might even be slow to take decisions, right? Versus if you have more compute um then you can you can take those decisions faster. Of course, a whole element of this you have to optimize the NN as well. Correct. Yeah. And by that do you mean having it remember things or

do you mean having it remember things or just keep making it better and better? Well, uh the way it is Yeah. It's more like making it better, right? So it's like you can take 10 matrix multiplications to get to the same result or you could do it in five. Five. Yeah. So the pipe thing would be more optimal. So you're always trying to build your software in a way where it can process more uh you know decisions without taking as much compute. So that's that's you know more more uh it's just faster and yeah I see. And then how do you get to like how do you talk to the hardware I guess? So I understand that you can write code that represents your neural networks and all of that but

your neural networks and all of that but if you know like the the gap that exists in my understanding right now is the way to do that in my head is to you know go to switch on my computer it has a CPU there's like a operating system right that lets me talk to the CPU but there is an operating system here is there like how do you get software and hardware to talk to each other yeah and know there's different layers of operating system this is like um what we've been doing for the last six years I I think we started with with and then uh Ross' robotics operating system that's a more like general framework that a lot of people use in the industry

that a lot of people use in the industry quickly found that it has so much limitations that it doesn't work for us uh and then transitioning to rust based programming and we've been using rust for all of the uh sort of work and we do on uh on the on the software side but then there's also like kernel level optimizations that we do we go deep into like how the computer itself like you processing information. And so a lot of this is just like you have to go and own every part of the stack and understand where the bottleneck is and make it better. Often times the chips that we buy from suppliers even they don't have a full understanding of the bugs that we're facing. So we need to sit down

we're facing. So we need to sit down with them and understand like hey your chip is m is behaving what's going on and then we figure out some issues together. So in the process you know you end up making the thing better for everyone. Wow. I'm just just taking a minute here cuz I'm just kind of blown away by you know just how much stuff goes into that little it's not that little but it's still kind of a small right but it's the we only just talked about the software component of MBR I you know touched the hardware which is a whole beast in itself. Yeah. And which is the perfect segue to I Yeah. I really just want to kind of go down your list of the as you said unintuitive or non-intuitive. I don't know what's the right way to say. Counterint. There you

right way to say. Counterint. There you go. Counterintuitive design choices that have been made. Like um yeah, I saw obviously some of them just on the website itself. It just looks unlike any robot cleaner that I've ever seen or I should start saying droid cleaner maybe that I've ever seen. So yeah, what were some of these differences and why were they made? I'm very curious to understand because yeah it's such an interesting case study to learn about these things right. Yeah. Yeah. No that's that's a great question. So I think I'll start from the very top. Um we talked to a lot of people right and uh especially coming from a product design background like I I wanted to make sure that we understood what the problem was. And what we learned on the hardware side was the robots out there

hardware side was the robots out there just get stuck a lot. And uh you know they they get stuck on carpets, they get stuck on wires, they will get stuck on tiny ledges. Uh you know even if there's like a piece of clothing on the floor, the robot might ride on top of it so that the wheels are not making contact with the floor anymore and then you're stuck. So that was a key insight which is like people don't want to rescue their robot. Sounds obvious, but like you know that's a big deal. So from there you get to like okay maybe we should just make the wheels pretty big.

should just make the wheels pretty big. Uh think think SUV versus sedan, right? Just make big wheels so that it doesn't get stuck. Add add a big ground clearance. Correct. Right. So it doesn't it just has that uh space for things to pass underneath it without it getting stuck. Especially if you're cleaning on big carpets and things like that. Mhm. So that was the first decision. Second, we we started looking into it and we realized that the bins in all these robots are so tiny. Uh and and we also kind of just, you know, from our standpoint, we're looking at the people who need this the most. And some of the people who need it the most are people who have like, you know, a little bit

who have like, you know, a little bit bigger homes, people who just don't time to keep cleaning all the time. Yep. Um or even if you don't have a big home, even if you have a tiny home, like you might have a lot of mess on your floor. And there's multiple reasons. Uh you might have a pet, right? People love their dogs, people love their cats, but some of them shed. Yeah. Close to bed is like Yeah. that creates a lot of lot of uh stuff on the floor. Uh kids can also be messy, right? Um so the whole purpose of bringing in a robot so it keeps cleaning, you don't have to clean, but

cleaning, you don't have to clean, but if you make the bin the size of my fist, let's say, uh you know, in most homes that fills up in 3 days and so you have to still go and empty the bin. So we knew we wanted to make like a bigger bin. Okay. Right. Okay. So, if you want to make a bigger bin, then um you know, you can try to fit that. But then the other way to do it is have a big dock, which is what most companies do out there. You have like a big dock, and then the robot goes in, and then the robot already has a vacuum on it, so it

robot already has a vacuum on it, so it sucks up all this dirt. Mhm. And then you have another vacuum. This one's bigger. Okay. And you suck the dirt from the robot into the big bag on the vacuum. I see. Exactly. Sounds great in theory. sounds lovely, but it actually doesn't sound that lovely when it's in your home. Why? Uh you can ask anybody. It sounds like a rocket tippy hop in your living room. Nobody wants that, especially if you have young kids, right? Uh I never had any kids. Uh but that's what I hear from the parents uh that I've talked to. Right. Um so, and also we just thought it's just kind of ugly to put like a massive dock. Yeah, I agree. Yeah, that is an idiot. So, so we

agree. Yeah, that is an idiot. So, so we just thought we'll solve the problem on the robot itself. So, we thought of building a bigger bag, right? So, for bigger bag, you need more space. You know, you'll notice like all of this is kind of like going towards taller robot. So, go back to that. Right. Right. Right. We also started looking at okay, what is actually on people's floors? So, a lot of the problem is just hair. Mhm.

a lot of the problem is just hair. Mhm. Like you know there's yes there's like small food, there's like small and stuff like but most of it is people just walking around. Humans also shed hair and dogs al shed hair and long hair and small hair and all sorts of hair and um in in another thing is any kind of cleaner that uses a rotating thing the hair just gets stuck on it. Correct.

hair just gets stuck on it. Correct. Yeah. Hurry heart side. You get it? Yeah. So we decided to solve that problem by having a sweeping system which is very good at handling hair. Uh so we built a brush roll that hair doesn't get jammed on. So we we decided to not do any bristles. So there's there's a counterintuitive decision. There you go. Yeah. I don't think I've ever seen any of these products that don't do bristles. Like even our just sweep sweep swe swe swe swe swe sweepers have bristles too, right? Everything has bristles. Yeah.

right? Everything has bristles. Yeah. And and there's a good reason for it which is it it cleans deeper. Correct. Correct. it it will go into the contours of the floor. It makes a lot of sense if you're ready to do all that maintenance, right? But when you're talking about a robot, I don't want to do maintenance. That's why I bought the robot. So, you know, it's it's it makes sense to bias a little bit. Uh I I'm okay to give up some cleaning efficacy to get more autonomy. That makes sense. Absolutely.

autonomy. That makes sense. Absolutely. Absolutely. That's what the product is. The product isn't going to be as good as uh somebody you hire to clean your home, but that's okay. It's it's a robot. It's super cheap as compared to hiring a human and it makes sense the cleaning efficacy is not exactly as good but as long as it doesn't also need my time to rescue it. Correct. So that's Yep. So so we did that thing and so beyond that we then looked at okay even if you sweep people's homes we found that there's still like a tiny layer of dust that stays back. So, so you can do this experiment if you take a broom, right? You know, take out all the

broom, right? You know, take out all the time in your in your day and like broom for three hours if you will, right? But then once you're done brooming to your heart's content, take a paper, you know, wet it just a little bit with some water and then and then, you know, try to clean a patch on the floor. You'll find that the paper still has some brown lur on it. Yes. And another way of saying this is like you know we and so that's the mechanical aspect of it right like you can keep brooming it doesn't fix it.

you can keep brooming it doesn't fix it. We can also like keep vacuuming it and it doesn't fix it. So we knew that no matter how much we broom, no matter how much we vacuum, some dirt is going to be left behind on the floor. Okay. And a lot of people like to walk bare feet. Uh you know and and it's not nice to feel like crunchy things or dirt on your on your feet. Right. Right. It also doesn't look as nice. It's not like shiny clean like you know if you if you ever have cleaners come in and sweep and mop it it gives you a level of shine that basically you know you can only get by

basically you know you can only get by mopping. Exactly. Yeah. I feel like mopping is the final you know layer that the must have not even analyzed to have. Yeah. Right. And so then we look at like all these other robotic cleaner products and I was just so surprised. I remember every robot, every single robot, all they did was they had a flat pad and then you drip a little bit of water on top and then it drags the flat pad around. That's the one my brother has in his house. And it it just boggles my mind which is like this is a complete like scam almost, right? It's like the Yes. You're putting water down. Sure. The thing is moving around. Mhm. It'll get saturated

moving around. Mhm. It'll get saturated so quickly like within a room. Yeah. And then if a user says, "Okay, I'm going to get up and either remove the pad and replace it with a new one or wash the pad and put it back." Yeah. So then you're doing some real mopping. But like who in their right mind realistically is doing that? Nobody. Yeah. Nobody does that. And so then it completely defeats the purpose of having a mopping robot.

the purpose of having a mopping robot. Correct. So for us it was a very strong design decision which like we're going to not only just put the clean water down, we're going to suck up the dirty water and and that was another counterintuitive thing which is like nobody's doing that six years ago. And so there's no pad then. Is is that what you're saying? Or does the pad water get sucked up? It does. So it's it's a roller. Okay. I see. I see. So on the roller we put clean water from the top.

roller we put clean water from the top. It goes and agitates the dirt. It picks it up, dissolves it in the water. And then on the back side, we have what we call a ringer. Yes. A tooth that digs into the mop and it creates a bead of dirty water. Interesting. And then we apply suction to it and we're able to suck up the dirty water. Wow. So, so, so far we've talked about big wheels, big bag, big ground clearance, you know, uh brush roll that doesn't get stuck in here. And then you also have your mopet roller cleaner. Yeah. Self cleaning mop roll, right? All of this then collects the debris in it has to collect it somewhere on the robot. And we also don't want to suck it

robot. And we also don't want to suck it up into a dock, right? Remember? Um. Yep. So now if you can like, you know, just think about it. You're collecting all this hair and maybe even food crumbs and dirt and then you're also mixing water in it. Uh that mixture if if you have a plastic bin that the user needs to remove it and throw it away, it's not going to work. The reason is it's just going to get so dirty and so Yeah. It probably starts smelling right. Exactly.

probably starts smelling right. Exactly. Yeah. Yeah. It's that I've seen people ship products like that where it's it's like a mix floor cleaner and then they ship it to the plastic bin and I I can't for the life of me figure out like what the design thinking is there because that's very difficult to clean and nobody cleans it. Yeah. Cuz nobody wants to and you don't even want to put that in your dishwasher and such, right? Cuz then it just dishwasher ruins the whole thing, you know? I wouldn't know where to clean it like oh yeah my kitchen sink do I clean it in my bathroom sink do I clean in my bathtub like all of those

clean in my bathtub like all of those places I don't want that stuff there 100% so it's it's just an impossible user experience so what we did is we decided we're going to make it use and throw so that's where the this idea just doesn't exist which is we put bags on a robotic floor cleaner bags on a handheld floor uh robotic or non-rootic a vacuum cleaner exists correct but we decided to put bags on a robotic floor cleaner that just never exists Right? And then going back to our software requirements, we needed to do cameras. Now, if you think about it, there's a reason humans have eyes in their heads and not their feet.

eyes in their heads and not their feet. And the reason is the higher the vantage point is, the better you can see things. Correct. Right. U you know that's why that's why like you know uh yeah it just makes sense to have eyes high because you have the farthest nicest crispiest vision. Yeah. Exactly. So from that perspective also we wanted to put cameras high. So all of this like big bag, big wheels, big water bin uh and and being cameras cameras needing to be high, it just led us to this idea of a tall robot, right? And also we just like the idea, the industrial design of it.

the idea, the industrial design of it. We just we just like that it looks different than everything else that's out there. True. Um but then there are real downsides. It won't get under every single bed. It won't get under every single couch. And that's a real cost. And to some people, it's really important that under the bed and under the couch gets cleaned every day. To me personally, it's not. Yeah. I'm trying to think. I mean, I I don't know if there's many couches that have enough space for vacuum cleaners to get into in the first place. At least one of the nicer ones, right? Again, I'm not referring to the $50 ones you can get off of, you know, Amazon, but but no, that's a fair point, though. Yeah, I agree. It can be a limitation for some people. Yeah, I see. Yeah, for sure. And

people. Yeah, I see. Yeah, for sure. And I think people and that's why like all the Roombas look like these circular pend pend and and we knew this going in which is like we're not going to get inside uh all of it, right? But from our perspective, one, those areas don't get that dirty, right? Cuz you know, your dog is not shedding under that tiny couch where there going obviously.

couch where there going obviously. Second, even if it gets dirty, um it's not as visible as the air. Yeah. Yeah. Yeah. That's the second thing. If you clean it once a month, it's going to stay mostly clean. I was just going to say that you could even clean it once a month and leave all the debris out right there and then have the Matic pick it up, right? And so this is where like this is a hard trade-off. It's that people are used to shallow robots and everything we've talked about so far is leading us to a tall robot. just makes so much sense for all those reasons.

so much sense for all those reasons. Now, we can we can say like, "Oh, no, screw all of that. We're not going to go in this new direction. We're not going to create a new product. We're just going to stick to what people are used to because that's what people want." Turns out like it doesn't matter that much. And, you know, we we decided to give up that whole thing to get all these other benefits. Mhm. Uh what we did do is we we made the cleaning head relatively shallow. So, so if you look at the front of it still only has a 3-in height. So there's some cases where it actually does become quite important to to clean under uh a overhang and Oh, I

to clean under uh a overhang and Oh, I see the the front part you're talking about right over there. Okay. So yeah, that's a little lower right like and correct the most important thing where it does make sense to clean under some kind of overhang would be to kicks. Are you familiar with the what tokicks are? Uhhuh. Yeah. Mhm. Yeah. So it's just like the area in a kitchen where you know your your feet go if you're if you're working on a counter. Yes. So the reason that's important is one that it's very visible and second if you're cooking all the food scraps end up there. Right. Exactly. So we didn't want to ignore that. We we did want to cover that and that's why the cleaning head is designed like that. So it can still clean all your toics

clean all your toics and it won't go under the shallowest of beds. It won't go under the shallowest of couches. And we took a stance. We said it's okay. We're not going to delineign for that because there's all these other benefits. Exactly. And you can never have it all. Right. That's I just feel like that's just generally true in life where there is no pleasing everybody. You just have to pick your lane a little bit. Yeah. Yeah. We like to say there's no perfect product.

to say there's no perfect product. There's only a simple product. There you go. That that's so much better than what I just said, but yeah. Yeah. I don't worry. It's not something I came up with. So, okay. Uh and then and then I I would also like touch upon like the noise aspect of it. Um this is it shouldn't be counterintuitive but turns out it is. Uh for many floor cleaning products uh what manufacturers will do is try to make it as loud almost in a certain way or to at least to a certain exa extent because noise is considered an analog of power.

noise is considered an analog of power. Power. Right. Right. Right. Right. This is definitely true for handheld things. True. Yeah, I agree. So, we were like, "No, no, no. We're going to go in the opposite direction and we're going to make it as quiet as possible." So much so that we actually put a dedicated muffler inside the robot. Really? Wow. And it's it's kind of like a car. So, we have like a full almost like a quarter of the back of the robot dedicated to the specially custom designed full assembly where what all we're trying to do is we're going to channel in the air that comes from the vacuum motor. Mhm.

that comes from the vacuum motor. Mhm. And we want to guide the air outside the robot without letting the noise energy out. I see. Make I just described a muffler in a way. Yep. Um, and so that's exactly what it does. And it it brings the noise drastically down. Wow. And just any loss of perceived power of well, sorry, any loss of uh perceived power Yeah. is is not a problem like it you haven't seen that effect, you know?

you haven't seen that effect, you know? No, it refuses or really But you can compensate for it, right? So So you you run the airflow higher so that the net air flow ends up where you want. I see. Gotcha. Right. So overall you you end up with the same air flow that you wanted. You have to expend a little more energy because your vacuum motor has to run higher but then the sound is also lower because of the muffler that we attached.

because of the muffler that we attached. Makes sense. Yeah, that makes total sense. I'm just Yeah, I'm just kind of you know struck by just how much iteration all of this must have needed. So you know naturally I'm inclined to question how so do you guys you know as a team at MATIC come up with like a plan right? Everybody has to be aligned and then do you make a prototype and then it goes to the factory for mass production like how does that entire thing work?

like how does that entire thing work? Could you like please dumb it down for us? Yeah, for sure. Um, it's it's very iterative for sure. Um, we have leaned heavily on 3D printing and just like having a culture of like allowing people to make uh decisions, trying them out and bailing fast and learning quick and then trying again. Uh, so a lot of it is in the process. So it's a bit of a balance. You you want to write down all your requirements. You want to question every single thing before you go into it. And then what we do is, you know, we'll we'll pick we'll create a lot of concepts. We'll create like five concepts, sometimes 10 concepts. Then we'll debate. We'll go into a room and just like talk for three hours and

just like talk for three hours and figure out what are the pros and cons of everything. And then ideally come up with one or two concepts that we really like. And then uh you know whosoever the DRRi the directly responsible individual is they will go and flesh out the entire concept. They will do the CAD work. They will do the prototyping. They will run experiments. They will compare and contrast. They will ask customers how they feel about it. They will run surveys. All of this takes months if not years depending on how complex the task is. Correct. And then once the prototypes are working really well, we got good customer feedback, we've addressed all of the reliability concerns, we've looked at cost, we've looked at integration with the rest of

looked at integration with the rest of the robot, we've looked at software complexity, we've looked at supply chain complexity, then we eventually start pulling the trigger on like let's get suppliers involved uh to the point that they start cutting tools, for example, for injection molding or something like that. And then injection molding takes its sweet time. It can take like two to three months to get parts back. Then parts come back, parts might have quality issues and then also your testing becomes more and more representative as you're going. So you do more testing, you find issues at all stages of this process. Correct. The the deeper you get in the process, the harder it is to fix anything. It takes more time, more effort. Uh but you just you just you know keep working on it.

you just you know keep working on it. And at which at which step in this process that you just went over do you decide what materials to use? because uh I think typically early on in 3D printing we we we know that we're going to have a different material for injection molding for example and I'm I'm talking a little bit more about plastic parts right now. Yeah. Uh for metal your prototypes can be pretty much the same material as as your final parts are going to be. Uh cuz it's very easy to get good prototype metal parts cuz that's subtractive manufacturing. Right.

that's subtractive manufacturing. Right. So you can right someone machine your parts. uh in for plastic parts you need to do like additive manufacturing and so initially you do 3D printing what we try to do is we match the material properties to the likely material we're going to pick in injection molding so we have an idea of what we want but we'll try to match the properties and then that's how we test but then there is a round of review before the tool is kicked off then we look at like hey what are the operating conditions uh temperature humidity what kind of chemicals Does this if any material need to be in contact? What is the stress that it's going to need to handle? What kind of durability reliability

kind of durability reliability requirements are there? And all of those things will guide us in one direction or the other. And then we will start with one material. Gotcha. What is a day in the life for you currently at the position that you are at at MATIC? Yeah, I think the last few weeks have been a hell of a roller coaster ride ever since the uh Wired article came out. Um, there you go. You know, we're just uh flooded by a lot and a lot of like orders.

by a lot and a lot of like orders. Mario, well, congrats, man. No, thank you. Yeah, it's it's been it's been great to see that. Um, so I am, you know, constantly trying to figure out ways to scale our production. Uh, so training more people on the production side because we build the robots in the US. Uh, a lot of it is done in Mountain View, California. Uh so we are looking for good people to join our production team and and you know uh grow that side of the organization. Um and then also just like you know constantly looking for uh people to add to the design team, people to add to the supply chain. So that's that's I want to say like that's 40% of what I'm doing right now. I see.

40% of what I'm doing right now. I see. Um there's also like constant like design improvements we're looking at. These are some um you know forward facing designs. Yeah. Yeah. Yeah. So there's a lot of design review work that I do uh you know sitting in meetings and and sort of trying to debate pros and cons of different design uh ideas. I would say that's another like um maybe like 40%.

that's another like um maybe like 40%. And then the last uh 20% is just like you know fires that happen every now and then. Yeah. Some some critical quality issue comes up and you need to talk to the supplier uh or or like you know uh you might have like a field failure where uh Oh boy. Yeah. That's sounding bad. Have you ever had to do like a mass call back or something like that? Has it ever gotten that bad yet? Uh the the good thing is we don't we're not building a safety critical product. So true. And you know, if you're building cars and there's a safety critical issue, you have to call back. Uh for us,

issue, you have to call back. Uh for us, I think we um we would only do a call back where if it was a safety critical issue. Uh and we haven't had anything like that. I mean, it was possible, but uh we haven't had anything like that. that I think we do a lot of rigorous analysis before putting anything in the field. Uh we have like you know let's say like it silly things happen all the time like somebody might like actually push the robot off the stair even if the robot itself wasn't going off the stair and then the robot falls off the stair and then you have to go repair it. Uh and you know once in a while um some

and you know once in a while um some some things can actually do break down like yeah there might be a quality issue or something like that that slips through the cracks. So, so there's there's that aspect as well. Incredible, man. Um, last question here before I let you go. One of the things that I get a lot from especially from um some of you know my mentees and just students that are currently in their masters courses or are considering doing that. Um there's this almost constant tug of war, right? That if we go into hardware, you know, like the hardware side of the house, just using it as an umbrella term, there are no jobs. it's so difficult to get employment blah blah.

difficult to get employment blah blah. So what is something that you having been there done that would like to say to these people to you know try and well not try and convince them but to kind of tell them just as advice that it's not all broken there are opportunities here as well and maybe you could if you could even touch on just some of like the most crucial skills or tools that really go a long way in this field that would be super helpful. Yeah, I think I will take a strong stance against there are no hardware jobs. I think there's plenty of hardware jobs. There's a there's a lot of people looking to hire good hardware engineer. I know because I'm one of

engineer. I know because I'm one of them. Absolutely. Yeah. Problem is the barrier to entry on hardware jobs is higher. Um I see. And okay, there's there's reasons for that. It's it's relatively uh you know more approachable to pick up a computer and start doing some online course and learn a new language, right? Um it's it you know you could just be sitting by yourself. Correct. U you know you you spend six hours every day for a month, you're great at Python by the end of it. Yes.

great at Python by the end of it. Yes. Right. You could even learn Rust, you could learn C++, you could learn Java, whatever. Um and and you could at least like start taking interviews at the end of Right. Mhm. And however, if you're going to get into hardware, you need real machines, access to real sort of like, you know, if if you're going to go do mechanical design, you need access to 3D printers, you need access to uh, you know, some amount of budget so you can buy off-the-shelf components. So just like even if you're an electrical engineer if most people who come out of college have not necessarily fabbed a PCB PCB and hardware it's this kind of interesting thing a lot of the skills are very difficult to learn in college so when we graduate from college most people I would say 80% people

people I would say 80% people undersshoot the amount of real hands-on experience that is required for you to be high interesting And because people have been, you know, classically just, you know, you go through high school and you think if I do good in my classes, that's great. Yeah. That's all it takes. It's good to do good in classes. But as a practicing engineer, I need to know, can you do the job? And the only way to show me that you can do the job, especially if you're applying for a design engineer role or any kind of hardware role, you have to show me what have you built. Like physically, you mean typically? Yeah, physically. What have you built? And for me personally going through college, I don't know how,

going through college, I don't know how, maybe some very very, you know, wise seniors told me this early on. I I'd have to think back how I learned this, but I knew that I need to build stuff, okay, before I go apply for jobs because if I haven't built anything, what is my claim that, you know, magically I'll come to the job and I'll know how to build a robot. Like that just doesn't work like that, right? So doing things like FSAE, Baja, doing your Hyperloop competitions, uh even like building things at home like any projects like I love when I interview a candidate and they've fixed like uh a 3D printer or a router and I know that they know what it

router and I know that they know what it takes to actually make a physical system work. work. Debugging a hardware system is a lot more grind is a lot more slo than debugging. debugging. I I can't even imagine. Yeah. Like where would I even start? It's a you to get up. You can't be sitting in a chair all day. You need to go get the parts. You need to go get on a phone call. You need to go visit the factory. It's a real full body sport. Like you know it's not not for the faint of heart. Absolutely.

not for the faint of heart. Absolutely. And so and and also you you just it's a team sport too, right? Cuz the complexity gets so high so fast. You need to know what the electrical engineer is working on, what you're working on, uh what your GSM is working on, so on and so forth. So it takes like five different skills to be a successful hardware engineer. And most people maybe have like two, even those two are not that well honed in. So there's a lot of jobs. I think people just need to sort of realize what it really takes to be high. Yeah. Well, thanks so much for taking the time today. I I feel like I've just learned so much about this, you know, whole portal of universe that exists, but I just never, you know, care

exists, but I just never, you know, care to learn that much about and wish you all the very best for, you know, Matic. And I'm, you know, really uh looking forward to seeing all of the amazing things that you guys are continue to do here and of course being an advocate for this wonderful, incredible product that you have. Yeah, thank you so much. This was a great time. Thanks for inviting me. That brings us to the end of that episode with Anuman Kumar. I was just blown away by how much I learned about this given I'm not really from the hardware field. If you would like to support me, the easiest way to do that is by subscribing on YouTube and leaving me up to a fivestar rating on Spotify or any of your favorite podcast apps.

any of your favorite podcast apps. Something that goes a really long way for me is if you share by word of mouth or if you just share these episode links with your friends and family and tell them about how you found your new favorite podcast. Catch you all in the next one. New episodes every Wednesday.