Ep. 8: Disha Sarawgi and Mohammed Kabir

Flying in Warehouses, Inventory, and Fleet Management

November 30, 2021 · 56:51

In this episode, Audrow Nash speaks to Disha Sarawgi and Mohammed Kabir, who are Corvus Robotics’s Head of Special Projects, and Co-founder and CTO, respectively. Corvus makes a robotic drone for warehouse inventory management, which may greatly reduce inventory loss. They speak about flying their drone in warehouses, working with customers, safety, fleet management, how Corvus is funded, and they give advice to aspiring entrepreneurs.




The transcript is for informational purposes and is not guaranteed to be correct.

Audrow Nash (A.N.) (0:03)

This is a conversation with Disha Sarawgi and Mohammed Kabir. Disha and Kabir both work at Corvus Robotics. Disha is the head of Special Projects and an early employee at Corvus, while Kabir is a co founder and the CTO at Corvus. Corvus robotics is a Boston based startup that makes a drone to fly around warehouses and help with inventory management. Their drone, the Corvus one seems to promise to reduce inventory loss in warehouses, which currently happens to a staggering rate. As you'll find out in the interview, they speak about flying their drone in warehouses working with customers, safety, fleet management, how Corvus is funded, and they give advice to aspiring entrepreneurs. This is the Sense Think Act Podcast. I'm Audrow Nash. Thank you to our founding sponsor open robotics. And here is my conversation with Disha and Kabir.

Kabir, would you introduce yourself?

Mohammed Kabir (M.K.) (1:11)

Yeah, sure. So I'm Kabir. I'm the I'm the co founder and CTO of Corvus robotics.

A.N. (1:19)

Hi, Disha.

Disha Sarawgi (D.S.) (1:22)

Hi, I'm Disha. I work as head of Special Projects and General Projects.

I lead basically all the product lines here at Corvus.

M.K. (1:34)

And she's also one of our early employees, one of the very first

A.N. (1:38)

Gotcha. How large are you guys now?

D.S. (1:43)

We do have about 15 people across engineering and on engineering goes

A.N. (1:51)

to the shot when what number employee were you?

D.S. (1:54)

I read? Second, probably.

A.N. (1:57)

Yeah. She's the second employee. Right after you come here.

M.K. (2:04)

The co founder, the founders,

A.N. (2:06)

I see. So tell me about Corvus. Yeah, about what is the mission and everything.

M.K. (2:15)

Yeah, sure. So when we started out, sometime in 2017, or so we saw this a whole bunch of robotics companies, newly formed robotics companies, really kind of doing great things in the warehouse and industrial automation space. And generally, my background is that I come from a real robots. So I really love flying things and making drones. So when we started the company, we kind of were like looking for use cases, which we could automate kind of putting together this industrial automation space, and aerial robotics. And I think we found a pretty good niche and decided to pursue it, which is warehouse inventory management. That's kind of our singular, large focus throughout the company. So all our product lines are kind of trying to solve for that one specific thing, which is warehouse inventory. And that's where the idea for purpose one was born. What purpose one is, it's a drone, which which does, which helps warehouses do inventory. This is a huge market, warehouses lose millions and billions of dollars worth of goods from inside their own facilities every year. And this is the problem we set out to solve.

A.N. (3:37)

Why is that? Why are they losing millions and billions of dollars? Right, good, right.

M.K. (3:42)

So the way warehouses do inventory management today is whenever they send a person out with like a handheld barcode scanner, and then this person goes up and down the racks. So these are huge facilities that can be millions of square feet. So this person goes up and down the racks with this handheld barcode scanner, and basically scans barcodes on the on whatever stuff they have on the shelves. And then they basically write down or somehow tie back into their warehouse management system where this thing was. Now in variably, because there's a human involved in the loop. This introduces a lot of errors. And typically, I think the nationwide average for kind of the inventory accuracy across warehouses in the US, it's something about 40 to 60%, which is actually quite bad. So they lose half of their stuff just because they don't know where they put it. How could

A.N. (4:39)

that be that they lose 42 They only retain 46 to 60%. Yeah. So so what happens is like, why why are these red? Why such a large number of loss?

M.K. (4:52)

Right? So um, basically, the way inventory comes into warehouses, you basically unload from trucks and Then there is a person who is taking it from these truck bays and then putting it in their shelf somewhere. And this person is responsible for kind of doing an incremental scan. So they can. So basically, we can input into their warehouse management system, where this particular thing was booked. So it's basically all deltas. So it's, it's as things coming.

A.N. (5:25)

What do you mean, deltas? It's, it's every, it's the change effectively?

M.K. (5:29)

Yes, yes, it is the change. So every time there's a change, they update their warehouse management system, but invariably, because there isn't like a global view of the warehouse, per se. Errors build up over time.

A.N. (5:45)

Why? Why are they're building up, though? I mean, if I just put something there and leave it there, it should stay there. And it shouldn't matter that I only have Delta's what's going on? Are people stealing some of the packages that people move? Knowing or like, what's creating 60% of the law?

M.K. (6:06)

So I think it's a bunch of human factors at the end of the day, I wouldn't say that stealing is a big one. But I'm sure it is an actual problem. One of the main things we've seen is people just put it in the wrong place in the first place. So basically, you say that you're putting it in whatever aisle a slot B, but then you put it in aisle a slot, see the one next to it? And they just don't care about where they put it at the end of the

A.N. (6:35)

day? Because you mean the person who puts it there? Like, yeah, doesn't matter.

M.K. (6:38)

Yeah, they're processing like 1000s of orders a day. So it's really hard at even at the human level to kind of keep up that 100% accuracy.

A.N. (6:49)

I mean, it seems like getting one and two wrong if you're getting that level of like loss, which is kind of nuts. Yeah. Yeah.

M.K. (6:57)

It's a huge problem. It sounds ridiculous, but it is what it is. And that's that's what we set out to solve.

A.N. (7:05)

So then, what your so tell me a bit about how you're solving it with your Corvus one, drone.

M.K. (7:12)

Sure. So what premise one is, it's basically a full cycle autonomous solution. It's a drone, and this charging dock, which we go and install in facilities. So it sits at the end of warehouse racks. And it does scheduled flights on its own, and then goes out into the aisles, scans inventory as it's flying, it finds barcode scans them comes back home lands, and then we process this data and then send it to the customer, usually through some sort of digital data dump. And what this does is we're basically taking that person who was doing this with a barcode scanner, and we put the barcode scanner on the ground. And then we can do this repeatedly because well, it's a drone day don't get tired.

A.N. (8:03)

Yeah, so you're thinking of this. So if you were thinking of the diff, so it's kind of like in edge detection, and image or something like this. So you only get like, I don't know, you're only getting the difference, it's not a very clear way to see the whole state of the warehouse, right. And so now what you're saying is your drone flies around periodically. And so you can kind of get updates over time of where all the packages are. Exactly.

M.K. (8:30)

And if you do this frequently enough, we can basically get that global view of the warehouse basically every day, depending on the size of the facility. Okay.

D.S. (8:42)

Yeah, generally, like do I do that? Generally, inventory counts are taken by houses maybe once a year. But with the drawing, we're able to get the entire inventory counts multiple times. So God gives you a great idea of like historically, timestamps data, which you can go back to and refer to when like locate every item at the warehouse at any instant.

A.N. (9:11)

Gotcha. Very interesting. Now, how I'm thinking about this crazy loss. that puzzles me honestly, that it's so large, and I imagine like goods are twice as expensive as they would be if they didn't have this loss. Is that true? Would you think? Like, that's probably included in the cost of things? Yeah.

M.K. (9:32)

So the way the business model for warehouses work is they just write this loss off. So they account for that loss as they're moving stuff around.

A.N. (9:42)

Write it off in what way

M.K. (9:44)

as just lost.

A.N. (9:47)

Okay, but yeah, but it's

M.K. (9:49)

still in their facility somewhere. It's they just don't know where.

A.N. (9:52)

Hmm. Interesting. So that with the perspective of fraud, flying The drone around, often, you can kind of get a better sense doesn't require the human to place the package as accurately or in the specific section so you can manage it, or at least in the one that they say it's placed in, because I would imagine, it would be the reason for the error. So then from that, you can kind of see if things are missing, or like if it was here, and now it's not you at least have a small window of time to investigate where it has gone to, or where, or something like that. Yeah,

D.S. (10:36)

it's not always just miss loading of items, but sometimes, because the humans in the loop, people also forget to just scan because they're taking care of 1000s of orders a day. And it's, it's easy if you're also on the ground level, but warehouse racks go up to like 20 feet, or 20 to 30 feet. And getting up a forklift scanning, and then coming down going to the next track at the same height is not an easy feat. So human error does play a big part of this as well.

A.N. (11:14)

I see. Yeah, that's a good point that they're so high that these packages are just stored with layer on layer on layer of packages. Is that the motivation for the drone?

D.S. (11:28)

Oh, badly. Yes. It's like humans can do inventory tracking at ground level, probably at the same rate. But as the height increases, the complexity increases. And that's where the drone comes in. And we get a

A.N. (11:47)

diversity increase, because you have to have some sort of ID or something to lift the person and you need a person Manning, or thing that tries and holds the basket. I'm picturing one of those like, things that they put up power lines, or go and work on the power lines with, I don't know how,

M.K. (12:03)

yeah, it's actually quite close. So they have forklifts or bucket trucks. And basically, there's a person in a harness clip to that. And then they go up in that, and they're running the barcode scanner. And it's, it's actually quite dangerous. Yeah, it is quite terrifying. Like, we've we've been in them just to get a feel for what it's like, do this annually. And it's more or less and were

A.N. (12:27)

lifted by Yes, yes. That's awesome. I like that. You guys do that? Yeah. How was it terrifying? It's it's, I'm just imagining, like dangling trying to scan packages. That sounds incredibly difficult. Yeah. Okay. So, being a drone really helps you get to those high packages very easy, it's almost free, because you just have to pilot it. Is there other other advantages to having a drone, rather than like a robot with like a real big telescoping pole, or something like this? Right. So

M.K. (13:07)

we've, when we started, we've actually looked into all of the options. Starting from like a lower complexity, like ground robot is obviously easier to build, right? Because it's not flying. But it turns out that when you get your poll 20 feet high, it becomes really hard to make the robot not tip over, just from the weight of the ball. And then if you want to kind of conquer that you need to make your robot like the you need the size of the base, you need to put in more batteries more weight. And it's like it's a pretty bad trade off at the end of the day, if you want to do anything of that kind.

A.N. (13:43)

Okay, yeah. And then so you basically with the drone, the Corvus. One, you have the drone flying and it has effectively a barcode scanner attached to it. And it's just pointing that at the packages. Is that right? Yes, exactly. And so now it's flying around from package to package just scanning all the different visible barcodes? Exactly. Hmm. Okay. So that's very interesting. How, how far into how many systems are deployed? And like, where are you guys at in terms of deploying yours to different warehouses your system to different warehouses?

M.K. (14:27)

Right. So we have a bunch of customers, we don't publish the exact number on systems deployed. But it is significant enough that we're getting very interesting fleet insights from them.

A.N. (14:39)

So order of magnitude is like 10 100. It's on the order of 100. Yes, that's really exciting. 100 robots out there. 100. Warehouses,

M.K. (14:49)

100 vehicles. Yeah.

A.N. (14:53)

And when you say fleet, insights into fleet management, you're meaning multiple Have your drone systems in a single warehouse? And how to like urgently go about scanning them? Yep. Yeah. Gotcha. That's tell me a bit about that. The different things you're realizing at a fleet level of controlling your drones? Sure.

M.K. (15:23)

Let's see. So in general, the way we design our system, it's quite scalable. So if you want to scale to a larger facility, we can just add systems basically. And they talk to each other, the charging docks can talk to each other and kind of coordinate. And one of the things we're just starting to roll out is basically this orchestration engine, which has a global view of the entire warehouse, and then it can basically automatically tasked groans to go and scan certain parks either. Maybe it's a high traffic area, which we are kind of inferring from previous flights, that if there's a lot of change, and we can see that in the data, we can, we can cast more drones to kind of focus on that one area, and things like that.

A.N. (16:13)

That's really cool. Yeah, you should do anything to add

D.S. (16:18)

a bottle. Now, we also give kind of a high level control to the people at the warehouse as well. We have scheduling online, where, whereas managers can schedule, what time they want to draw to go and scan what location, so and then the charging of orchestration figures out which John was done there, and autonomically does all the missions and comes back and charges again.

A.N. (16:50)

So that scheduling that you mentioned, Keisha, it's like you try to avoid when people are in the area, or what kind of why would I want to schedule the flights of these drones.

D.S. (17:04)

So there are times in the warehouse where there's too much traffic in an area, and you'd rather dedicate the drawing to less traffic area, the friendships during the day, were like, probably night shifts have X number of people around, and you want to get a scan of whatever was moved during the day. And that works out better for people. So we give customers that level of control to decide what and how they want to use this product as well.

A.N. (17:38)

Gotcha. And could be are that what you mentioned for it, recognizing high traffic areas that have a lot of change and trying to do them? Or I was imagining? Just I didn't understand what about the fleet management was extremely interesting before because it was like, Oh, you just patrol certain regions. But then if a certain region has a lot more change, you want to make more passes of it, because you want to hire more temporal data. That's very interesting. Yeah. Gotcha. How so now that you have 100 robots out there or so like, order of magnitude? How is this affecting? Package loss?

M.K. (18:19)

Right, right. Right. So one of the things we've begun seeing is that we've actually been able to increase the inventory accuracy for warehouses significantly. It really depends on a per facility level, how much we can help just because facility dynamics, the workers, their, their existing practices. We don't have any specific numbers, which we want to share this point, just because it's extremely facility specific. And even like any vague ballpark number idea will be wrong. So

A.N. (18:52)

you can have a specific instance of the thing. Yeah, but I mean, maybe you have like an average, or something or like a median level of improvement. Right. So

M.K. (19:01)

like, I'd be interested in,

A.N. (19:03)

like, double it, because the huge, like 60% loss is crazy to me. And if you keep an additional 30%, that would be amazing. I just walk

M.K. (19:17)

across, right, yeah, that is actually quite close to what we almost guarantee that if your facility is using Corvus drones 95% Just kind of the baseline you can expect to achieve within months. Yeah, and this involves working really closely with every customer because again, as I said, every facility is different. Everyone uses different barcodes. There's there's a lot of moving pieces there.

A.N. (19:49)

See? So let's see, can you tell me a bit about working with these customers?

M.K. (19:57)


A.N. (19:58)

and maybe the chef you have a perspective On this, I want to try to get us all talking about the same amount.

D.S. (20:05)

Sure, so, okay, since the team is pretty small right now, like, we are the ones who actually go to the deployment as well. And it's good to have engineers on site as well. Because if we build products that people cannot use in production, that it makes no sense. So going on site interacting with customers, we get a sense of what they want, and we build accordingly. So like onside, Devon been instances where customers have told us stories, like they spent hours on a Sunday night trying to locate one package, which was lost. And that's why like, and if there was something on Sunday, that sets them back for the rest of the week, because their week starts on Sunday. And that's where our drones have been able to get them a good amount of improvement like that is we have all the data stored. And we can tell them exactly where, what, where what package was seen. And it has them significantly improved their buffers and dedicated with others still actually getting packages to avoid. So

A.N. (21:24)

that's cool. So does this mean that you have to tie the data that you guys collect into their existing package, localization infrastructure, so like figuring out where all the packages are, and the map of that I imagine it's like a grid or something like a two 3d grid where they say it's in here, or at this location on this shelf, and you have to like, locate the package, and then map it to how they actually represent the location of the package. Is that true?

D.S. (21:53)

Oh, yes. So every facility has their own warehouse management system, which is, like, where every location is tagged as, depending on the rack. And it's not.

A.N. (22:08)

So it's like a sticker is on the so say, we have three shelves or something high, that each shelf of those will have a sticker or something that says this is shelf one, this is shelf two, this is shelf three. Yeah. And so that's how you localize them.

D.S. (22:24)

Yeah. So it also is like, maybe a number as to their eyes than their faces, like the two faces on each eye. So then, the next level is the face. And then they go by the level of the rag. And then each level also has number of slots. So the a hierarchical naming of each pallet position in the warehouse, which to keep track of and the warehouse management system. And we use the same mapping, and we work with customers to integrate the data we collect into their management systems as well.

A.N. (23:03)

So it's a big key value problem effectively with the key being the label of the shelf, and then the value being whatever packages are on it kind of thing. That's when I guess so if that's the case, it might not be I basically, was it challenging to kind of work with all of the different warehouse management setups that all of the different warehouses have? Or was there a challenge there?

D.S. (23:30)

Oh, okay. Kobe, do you want it?

M.K. (23:33)

Sure, yeah. This is just a huge pain generally integrating with the consuming, but the number of different systems out in the wild. We have a couple of ways of integrating life from the very simplest version, which we can basically support for everyone is we just give them a digital data dump. It's something like a CSV or something of that kind, which they can. And everyone says, yes, yes, yes, men is in commerce,

A.N. (24:02)

it can be processed like an Excel sheet. Exactly. It's like it's essentially an Excel sheet. Yeah. And then.

M.K. (24:12)

And then we can also kind of go into slightly more advanced integrations where we have some API which they can pull for data, or we can go the other way, and we can push data to their API's. This is again, very facility specific, very customer specific, and it's essentially a custom integration for every new facility.

D.S. (24:32)

We want to give as much value to the customers as we can, so we try to integrate with whatever their existing processes are and make them make it as easy as we can.

A.N. (24:48)

See, so now, getting into your drone a little bit more. How I like just how large is it? It has the autonomous charging dock just telling you bit about a bit more about it. What sensors it has, what it how large what it looks like, everything should start.

D.S. (25:12)

Sure. Yeah, I'm good tone is about.

A.N. (25:17)

And this is a cordless one we're talking about. Correct? Great. Yeah. Yeah.

D.S. (25:22)

So it's basically two components. We have the landing pad, which is the charging dock for us. And the drone. The charging station is about like, one meter or 2.7 meters, maybe. And the drawn is probably, like 3.4. Yeah, five

M.K. (25:47)

meters in diameter.

A.N. (25:51)

Okay, so for now, just kind of, yes, landing station is four feet or five feet, like a square kind of thing. And I saw it in a video. And I don't know if it's changed, but it's like a fiducial marker, like a QR tag. And then you have the charging dock in the middle of it. So I'm assuming that you're counting that in the five foot width of the station? Yes, yeah. Okay. Yeah, go ahead.

D.S. (26:19)

So these are the mountains at a height where forklift can go through underneath it, and we're not causing any obstruction to general warehouse traffic.

A.N. (26:33)

How did you What was the lesson learned for why you are mounting on the Bub forklifts? I assume that was like, there's some story there.

D.S. (26:42)

So yeah, like, we genuinely love the landing pad and the drawn to like at the end of a warehouse racks. And there is also a high traffic area for anybody who wants to go into the aisle or come out of the aisle. And we do not want to cause any obstruction for them to take like a large radius around the racks. Because some of them are also narrow passages, which is difficult to get through. So we generally mount them at a height where the drone can also operate without obstructing the traffic underneath it.

A.N. (27:26)

Did like a get run over or where people just frustrated that it was like a long path around or

M.K. (27:33)

so we've had some close calls before before we kind of realized that we should move it up even higher. So one of the things about forklifts is that they kind of have two modes of operation. So one is when they're just traveling along the ground, and therefore service fully at ground level. And there's another thing where if you have to take packages off from a higher act, they basically extend up and then the whole thing goes up in the forks also go up with it. Now what happens is, even though forklift drivers instructed to not move around at full extension,

A.N. (28:09)

sometimes even meaning the fork all the way up the exact right.

M.K. (28:13)

Yes, that is correct. So they're never supposed to actually be moving around about the whole thing extended in the forks all the way up, because as you can imagine, that's pretty dangerous. stuff. Exactly. And but it's actually it does happen sometimes. And then well, as you can imagine, they could hit our landing pad or they could hit other stuff. Um, that is why we tried to position it at such a place which minimizes any

A.N. (28:42)

of this kind. So it's kind of you have like a shelf side and then a few feet up, you'll have your actual landing pad, which is five foot by five foot, approximately one point So would you say seven meter

D.S. (28:58)

square? Oh,

M.K. (29:01)

point 1.2 by point seven. It's yeah,

A.N. (29:04)

yeah. Okay, so some square that suspended up there, like a little bird's nest. Yeah. And that that has the charging station there. You're Corvus one drone. Does it have like guards around the propellers? Or like how is it set up? So it's a quadrotor. It has four rotors. And each of those rotors has a propeller spinning. That's how it lifts up. Do the propellers have guards around them? So that they don't like hit anything? Or is it not a problem?

M.K. (29:39)

That actually, there's a story behind that too. So probably the story around Yes, indeed. Indeed. So when we started we actually did used to have propeller guards around the propellers. But after like doing 1000s of flights, one of the things we realized was that If there is any kind of collision with anything, the propeller guards don't really help in the first place, they actually end up. So the reason they don't help is whatever material you build your propeller guards out of, they typically have some amount of flex to them. And what can happen is your propeller guard can kind of bend and hit your propellers instead. And that's kind of a no win situation for everyone. Because then the drone, there's like a catastrophic failure.

A.N. (30:33)

Just falls down. Yeah, pretty much smashes on the warehouse floor. Okay, pretty much.

M.K. (30:38)

So the way we started looking at the problem was how can we prevent any kind of a situation where there could be a collision in the first place. And kind of what we've converged through is our economy stack maintains this safety envelope around the vehicle. And we basically got rid of the the hardware prop guards, and the software makes sure that the drone is just never close enough to get into any situation of that kind.

A.N. (31:08)

Gotcha. Okay, so you found basically, you could just stay far enough away from everything so that you don't need pretty much any sort of propeller guards and the propeller guards weren't even that good, because they flex and they would just go into the propeller, and then it would crash. So they'd be like, for protecting fingers or something where I would stop it. And I would hit the propeller guard, but then the propeller would hit the propeller guard, and then it would fall. So my fingers would be okay, but the drone would still suffer the same color. I don't know falling from air. Why not just use really rigid? I guess. So you've eliminated the need for it. I just wonder about like, really rigid propeller. But maybe they're heavier, they reduce your flight time or exactly. Yeah,

M.K. (31:51)

that's, that's pretty much it. If you're going to design with propeller guards in mind, you need to kind of add a lot of extra weight to the vehicle, which might be acceptable for for certain companies, or certain people working on drones. But for us, as since we kind of designed around the problem, it just was never a thing.

A.N. (32:15)

Yeah. And then the shot. Can you speak a little bit to the are you doing something with the cable?

D.S. (32:24)

No, I was blogging?

A.N. (32:28)

Ah, yes. Yeah, these lame to have the computer turn off. Can you speak a little bit to like safety concerns of customers? Because I guess I'm getting at what I'm getting at with the propeller guards. Is I wonder if people in the factory would want something like that. But just general safety things, safety concerns of the customer? This kind of thing?

D.S. (32:50)

Oh, yeah, definitely. So safety is something we take very seriously because it's drawn with models running at like 20,000 rpm, and we do not want to collide into anything or hurt anyone or not. So we have a bunch of software and hardware fail safes for ensuring safety. We have bunch of audio visual cues that we be provide while taking off by landing wire in case of

A.N. (33:26)

me that the drone has a speaker and it will go like I'm taking off now. Or like landing now or something like this. So people have some idea of its intention. Yeah,

D.S. (33:37)

we actually do. So we have Bibles on the drawn on the landing pad, which make it known that it's going to take off now. Now we have a bunch of spotlights to, like, tell people where it's going to go, or it's flying above you, you've taken out or something like that, and a bunch of flight to tell you what it's doing now, like it's on a flight or it's charging, it's not charging stuff like that, like anything that you can as much as you can gather by just looking at the system. There will be a visual

A.N. (34:18)

indicators of its status. Or audible as well. Yes. Do you do any? So I was thinking for hardware failsafe, like if the drum does happen to malfunction. Is there any like because So you mentioned flying above people? Is there any sort of like I don't know failsafe for something if the drone does say for some reason, there's just like a segfault on the drone like worst case and it just starts to fall. Any any recourse from there?

D.S. (34:49)

Yeah, definitely. That does happen sometimes. But we have tried to resolve it by software and our low level thesis. So in case have an emergency, which can be like a mission computer, losing contact with the landing pad or something like that, or just even a segfault. We try to find a safe spot underneath the drawn like the case, we have visual information of everything around the drone making figurative Potter land, and the drone navigates itself and land safely around without external intervention.

A.N. (35:33)

Gotcha. What sensors do you have on the drone?

D.S. (35:37)

So for a sensor stack, we nearly have an array of like 12 cameras onboard?

A.N. (35:49)

Or do they have like infrared? Or are there any? Like, like low light conditions or something? Maybe I think you have depth cameras as well, probably.

D.S. (36:01)

Yeah. So they're all arranged in like, do fashion. So we have infrared cameras, and then we also get information out of them. We have a bunch of time use and modifier to give you a no show information. And it's basically a visual inertial system.

A.N. (36:23)

Gotcha. So you have a bunch of cameras are all of the cameras, infrared depth cameras?

Unknown Speaker (U.S.) (36:29)

Yeah. Yes.

A.N. (36:34)

Okay. And there's no like, this is gonna sound kind of silly, but there's no like parachute on board or something. If the book, I guess the drone doesn't like hit anything ever, or it hasn't. And so that's you don't have to worry about it, falling out people's heads, or anything. Like my track record. So it isn't very

M.K. (36:54)

much. And also another way we try to kind of deal with this problem at the root is we tell the customers to schedule it in specific off hours, when there is no traffic in the aisles. And then we also do us a lot of customer training. And we train the workers there that if you see the drone, and it's about banter and aisles, don't go in there, wait for it to come back. It's like probably going to spend like five minutes there. So there's a lot of like, kind of human interaction between us and the customer and the worker stair which

A.N. (37:27)

you get exactly the aisle, the drone is using. Okay, because Disha you mentioned that flying overhead in this kind of thing. So I was imagining, like drones, scanning people, but need, and that seems risky to me. Like I imagined my propeller flies off or something that's really hard to control for.

D.S. (37:44)

I mean, we have safety systems in place, but obviously the same sense to not be that.

A.N. (37:50)

Yep. Gotcha. So it can't, like you can't fly over someone, but it's not desirable, because it's a little bit more risky. Yeah, exactly.

M.K. (38:00)

And variably. In the wild, we see that someone wanders in or someone just didn't pay attention to what we said. So we still design for that kind of extreme case where there's someone underneath it all the time. But in reality, that's probably like, 1% of our international flights.

A.N. (38:19)

Yeah, I would imagine far, far less. Yeah, gotcha. Okay, so you have this drone, it flies around. It uses depth cameras. To get a sense of the distance it is from everything it probably. So do you have to do? Like, do you have to map the warehouse first? Or how do you say you bring this new drone to a new warehouse for a new customer? How do you set up the environment other than putting this nest at? And by nest? I mean, the launchpad elevated for the drone? How else do you do the setup? Do you do some sort of mapping? Do you fly the drone around? What kind of things?

D.S. (38:57)

So yes, you are exactly right about this. Apart from setting the nest, we also pre map the warehouse, which means we get like geometric and semantic information from the warehouse and give the drone a pre map environment. Based on that we can define reasons or slots that we want to fly in and schedule missions accordingly and fly exactly that.

A.N. (39:31)

Gotcha. And so for mapping it, you just fly the drone around and have it What was it? So I understand? Right,

M.K. (39:41)

right, right. So the exact process we cannot talk about because it's something we've developed internally. And it's actually one of the hardest parts of doing this even though it may not Yeah, actually seem so. But at the end of the day, we're taking measurements, some by hand Some using the vehicle and then we're kind of putting these together and generating kind of a unified map for the vehicles views.

A.N. (40:09)

Okay, so you will custom generate a specific a map for the warehouse. For that you go off of that. Exactly. Do you have problems? I guess the shelves if they're really tall, you don't have a problem where they're like, move a few inches or a few feet on one end, as much.

M.K. (40:28)


A.N. (40:28)

we've actually

M.K. (40:29)

seen that. Yeah, that does occur. It's ridiculous again, but it does happen. Like we have to map it again. Whatever. So no, so we actually designed into the system. Yeah. So we have a map update method, which basically deals with minor change in the environment automatically. So every time a vehicle goes out, we can kind of update the map and then share it across the fleet. So all of the vehicles know that. Okay, this part changed a little bit. So next time I go there, I know what to do.

A.N. (40:57)

I see. So you're updating a little bit with your map. So you're doing some sort of simultaneous localization and mapping? Flight? And how does the so when your drone I saw I saw on the website, that you only need internet connected to the launch pad. And so when your drone goes away, it needs to probably stay within range of the launch pad? Like why like some sort of maybe it has a router or something on the launch pad? Or does it just does it just go back to where the launch pad is, and it doesn't need that internet connection for that period of time that it is in flight?

M.K. (41:37)

Right? So the way we design the system, actually, we don't need any sort of connection back to the launch pad while in flight. So everything all computation happens on board. There is a wireless link, but we don't actually need it while flying. It's just there for there's a bunch of reasons why it's there. But we don't actually require it while flying. Also,

A.N. (42:02)

things like logging data or something or what

M.K. (42:05)

pretty much pretty much basically information about flight progress, stuff like that.

A.N. (42:10)

Oh, okay. That's cool. Yeah. And then, but you don't need that actual the controller and the getting the data and save saving the data and everything that happens on board on

M.K. (42:24)

exactly, it happens all on the ground. And that's another design decision, which was actually driven by stuff we saw. So warehouse is like, typically it's a bunch of metal racks, right? It's it's a Faraday cage. And they're terrible, terrible RF propagation environments. Yeah, so any sort of wireless is destined to fail.

A.N. (42:45)

So that was that was why you decided to put all the computation on the drone. Exactly. That's amazing. I started in research in my undergrad doing this kind of thing where it was trying to put a drone and all the computation on the drone and having it localize and maintain a swarm formation. And it was really, really difficult. This was like 10 years ago, or something. And it's impressive now to have everything and you're like unbelievable suite of cameras on your drone to it's amazing how far things have come in the last nine years.

D.S. (43:26)

We have extremely long eyes in the warehouse as was so that showboating that kind of vibe firing is almost impossible. I like like you have made a backup pointer on it. So we just had to compress everything to onboard combination.

A.N. (43:48)

Let's see. So now talking a bit more about Corvus as a company. So you got you mentioned you started in 2017. Will you tell me a little just how did you decide to start a company and then dish I'd love to hear how did you decide to be employee number two after the co founders?

M.K. (44:08)

Yeah, sure, I can start. Um, so basically. So the company was, myself, Jackie, our CEO, and then our two other co founders, Jonathan and Brian. And all of us, kind of happened to come together. It was like a matter of luck or chance almost. So as the story goes, I knew Brian and then Jackie knew Jonathan. And then Jackie found me on the internet in some random IRC chat room and I was helping people with drone problems or something of that sort. And what kind of got connected there and decided to kind of start looking into the kinds of use cases which we could all automate. Because personally, I'm really into automating anything that is automatable. Just because people should be spending their time doing more high impact things. So automation is really a big part of my life. And kind of I think this is true. Also for the other co founders, all of them really care about kind of this problem space. And we sort of came together. And it was a really great, great team. That, that we started with, basically, to Brian was our hardware guy, he's our lead of hardware today. He does all of this. So anything you see on the ground, it's it's Brian, basically, all the hardware, all of the basically every piece, then Jonathan is our chief of software, he does a ton of work, both on the autonomy stack, and also kind of the almost, it's almost harder, I would say, kind of the customer integration side, which is integrating with all these custom systems, some of which are from the 90s, stuff like that. And then Jackie is our CEO. And basically, well, yeah, he does, he does everything. And then all of us came together in 2017. It was like things, some, some some point in the middle of 2017. And I told Jackie that, okay, I know this great, good hardware guy, my friend, we could probably think of doing something. And all of us came together and decided to start purpose.

A.N. (46:44)

Awesome. And how did you pick the opportunity? Like, how did you pick this specific problem to work on? Right?

M.K. (46:53)

So Jackie's family back in China. They were one of the very first generation entrepreneurs in China. And he had actually kind of firsthand seen the problems that warehouses go through. And kind of this, like this huge thing. Like as an outsider, like, I personally had no idea about what problems warehouses and generally like industrial stuff, people there go through. And Jagga kind of introduced us to this whole new problem space where these sort of non sexy use cases, which could still benefit hugely from automation, like it's not self driving cars, but there's still a huge opportunity.

A.N. (47:35)

Yeah, gotcha. Yeah, that's awesome, that Jackie knew so much. And then the shot you joining as number two, how was that for you? Like, how did you decide to join this company? Yeah. How did you like your interest in startups at this stage and this kind of thing.

D.S. (47:54)

So it was a bit of both like, I personally always wanted to work in an early stage startup, because it gives you the opportunity to work on the entire stack, get a sense of how everything's interacting with each other, have a good amount of ownership on the code as well. And the minute you get to learn is much more than you would if you were working just in a subsystem as part of a big team. So it kind of worked out that Jackie reached out to me on LinkedIn, I think. And I was in Austin at the time, and I flew down to Boston to meet the team. And I was really intrigued by the work we were doing, and just like being a small team, like they had come far, I guess, with the hardware and the software at the time, but I saw an opportunity to contribute as well and get the system to where it is now today.

A.N. (49:07)

Where are you guys located?

D.S. (49:09)

So we are in Boston, Massachusetts, at the moment.

A.N. (49:14)

Awesome. Let's see. And then just a more general question about Corvus. How are you guys funded? So are you receiving investment or have you received like a Series A or like, where are you guys at?

M.K. (49:30)

Right? We're Priya company. We went through Y Combinator in 2018. We were like the YC. Summer summer 2018. Batch. Yeah. Cool.

A.N. (49:42)

And so you've received what are they? They they give it's like seed funding kind of thing from a seed funding. Yes. Gotcha. And then you've just been running on that and maybe pretty much are you getting revenue from your customers yet? Yeah. Proper course.

M.K. (49:59)

But Yeah, exactly, yeah. So we've been, we've actually been deploying for about one and a half years now, we've been doing pilots before that. But we've had, we were doing production deployments for quite a while now. And all of these are active revenue streams for the company.

A.N. (50:18)

Gotcha. And so are you guys bootstrapping, effectively where you're reinvesting your profits back in and you're not getting much? Gotcha. Nice. I like that model. Do you think that you'll be seeking investment as a way to kind of grow faster, or just kind of stay off that keep bootstrapping and keep your equity really large? Or

M.K. (50:39)


A.N. (50:40)

I'm going forward?

M.K. (50:41)

That's that's completely open question. We are open to outside investment. Because we do want to kind of grow exponentially, we're kind of hitting the point where we want to scale up the engineering team, the sales team, all of that really quickly. So we're, of course open to that. But everything's on the table.

A.N. (51:01)

And you have to find a good fit probably, right? Because you guys are in a powerful situation where you're not exactly burning through your money. And meeting investment, possibly under bad terms, these kinds of awesome. Now, let's see, I see that we're coming close to the time that we were saying we should end. So what I'd like to do for our last little bit, is just asked you guys general questions about like, where robotic is going? And then advice things. So the shot could you start, and I would like to know what advice you would give to like a 20 year old view?

D.S. (51:47)

Um, okay, I think that's pretty straightforward. Because yeah, I mean, okay, my point is not long ago. Yeah,

A.N. (51:57)

that to someone else added up? Yeah.

D.S. (52:01)

Well, to two younger badasses, I would say, just do what you love. And do not be afraid, like, if it's an engineering solution that you don't know the answer to keep at it, because that is what engineers are helpful, like to think of problems and solve them even when you do not think of answers as the first goal. So just in general, the problem solving aspect is engineering appears to me. And that is what I love doing. And if you're good at it, keep doing it.

A.N. (52:49)

Let's see Kabir same question. Advice for younger you.

M.K. (52:53)

Vice for younger me. Um, let's see, I think one of the big things that's kind of been a part of my engineering journey has been open source software, I feel really strongly about open source, and also the stack at corpus, which we built, it leverages a lot of open source pieces. For example, we use Ross one, we use the px for autopilot stack. And I kind of started interacting with all of these at a very young age. And I think, kind of working with a lot of experienced people in the same field. And that kind of experience can only come from being on a team, like individually, there's only so much you can do. Like you can take classes, you can study, you can look up stuff yourself. But I think being a part of an open source software team really gave me a bunch of skills, which I use everyday now. I would really, really advise people to kind of pick a project you love or biggest space you love, find a project in that space, and then see where see where it goes, you will learn a lot on the way.

A.N. (54:02)

And what are some of the most valuable parts to you of contributing to open source? So obviously, your learning skills but like what specific things were most valuable in your experience?

M.K. (54:13)

Right. So when I started, like working on drones, I think this was when I was in middle school or something. I started looking at kind of what the options were if I wanted to build my own drone, and I didn't know anything at that point. I was just like a middle school kid. And then eventually, I came across the px for autopilot, which is this open source software stack for drones. And what this lets people do is there's a certain set of supporting hardware, and you keep this open source software, put it on that and then boom, you have your own drone, mess around a little bit with the hardware, but effectively it's a it's a pretty easy solution if you're trying to build something on your own with no experience before. And eventually what what my story personally looked like was I started using px for as a user. But eventually I started contributing back to upstream. A lot of the things which I wanted to do, but the features didn't exist at the time. And eventually what this led into was, I became a part of the core kind of maintenance team, which maintains the entire project. And that that branches out into helping other people doing code reviews, just a lot of skills, which there was no way I thought I would have learned unless I kind of got into it in the first place.

A.N. (55:42)

Awesome. Now, let's see a wrapping up. Is there any links that you guys would like to share? So in the show notes, we'll put a link to Corvus is website. Anything else worth share? Sure, we

M.K. (55:55)

have this demo reel of Corvus. One, which we'd probably like to show off. I can send you the link after the talk. And it'll be that'd be pretty nice. I

A.N. (56:06)

think you guys have social media or anything you'd like to share the shot? Do you have anything?

D.S. (56:13)

Yeah, we do have garbage on the garbage website, and it was on Twitter. And we will send you all the links.

M.K. (56:22)

Yeah, we'll send your Twitter handles and LinkedIn and whatever else.

A.N. (56:27)

Okay, great. Well, it's been a pleasure speaking with both of you. It's been a pleasure.

D.S. (56:33)

Thanks for taking time.

A.N. (56:38)

Thanks for listening to this conversation with Disha and Kabir. Thank you again to our founding sponsor, open robotics, and I hope to see you next time