Stefan (00:03):

Hey everybody, this is Stefan Seltz-Axmacher, and I am your co-host here at Automated. And my day job, is I'm CEO and co-founder of Polymath Robotics, that's a whole other thing. With me is this character who I met on the street playing with electrical wire. Who exactly are you?

Ilia  (00:19):

I'm Ilia Baranov, CTO and co-founder of Polymath Robotics.

Stefan (00:23):

What a coincidence.

Ilia  (00:24):

In this weekly podcast, we're going to explore topics in robotics and design and business cases for them.

Stefan (00:33):

But more than that, just like in the last episode, which I'm sure you've listened to 15 times, we're going to talk about how to build robots, how to make them valuable, how to make them interesting, how to deal with all that misery that comes along with doing anything in this funny little industry. And we're going to do that in two parts. The second part, we're going to dig into what Ilia has worked on in his past lives at Clearpath Robotics, a robotics company who's a way bigger deal than you even realized. And at Amazon, a tiny little startup that happened to try robots once or twice and solved all the problems in the first go.

(01:04):

And the first part, my favorite part of it, is a game that we play where we draw two cards and build a robot around them. And the cards are essentially a technology and a setting. I like to think about this as, let's say the CTO has some technology that they really love and that they have to use, or they absolutely hate that they will never use. Like last week we did 3D printing, and the assumed CTO in that case hated 3D printing, could never use it for anything. And the setting, is of course, the business environment they have to build something in. So last week that was home use, and we had to figure out a home use robot that did not involve 3D printing, but wasn't just something to bring you a tray of drinks, because who cares about that?

(01:42):

Now, at the end of the last episode, we actually drew two cards, and in true podcast host format, immediately forgot about them and actually had to re-listen to our own podcast for the 18th time, just now, so we can remember what those cards are. And Ilia,, what are those two cards that we drew?

Ilia  (01:59):

We picked walking robots for delivery.

Stefan (02:01):

All right, so if we were going to build walking robots for delivery. Now walking robots, assuming not everyone here has worked on walking robots before, they're really easy, right?

Ilia  (02:12):

Definitely.

Stefan (02:12):

Okay.

Ilia  (02:13):

Why would you even do wheels when legs are so much simpler.

Stefan (02:17):

I have legs, I've never once used a wheel to get up a flight of stairs.

Ilia  (02:21):

And there's no animals in the animal kingdom with wheels. So definitely, definitely legs.

Stefan (02:25):

Yeah.

Ilia  (02:26):

Definitely legs. It's made all the simpler because motors are completely weak compared to muscles and any leg needs multiples of them, they're heavy and power consuming.

Stefan (02:38):

And what's great about delivery for a walking robot, is delivery is one of those use cases that has a lot of robots get built about it because everyone interacts with delivery people every 18 minutes when they have a new Amazon Prime delivery of toilet paper show up at their house. So, I feel like a walking robot for this is perfect. Because what better... That Amazon guy who brings me a package out of that Sprinter van, he has to make, I don't know, $12, $13 an hour, let's replace him with a robot.

Ilia  (03:09):

Yeah, definitely. With multiple, multiple high-end motors, that's the cheap delivery case. But there are companies that have been built in this space, there's a few that are trying that, and there's a few interesting kinematic approaches. Well, let's not talk about that. Let's talk about the crazy applications where we might see this.

Stefan (03:26):

So I guess here's an application that I think is actually rather interesting. So obviously, just a dog robot walking down the street with a bunch of packages, that's always going to lose to a Sprinter van. There's no use case or that's better than a Sprinter van, I think, that's pretty self-explanatory. So a weird statistic that I know, I can't remember the word for it 'cause I don't speak Japanese, and I don't remember the number because I don't remember numbers. But somewhere between four and 9% of young men in Japan have shut themselves into their homes and are like hermits in their bedrooms.

Ilia  (03:59):

Oh, the NEETs. You're talking about NEETs. Yeah, non-education, employment or training.

Stefan (04:04):

No, I think many of them have gone to college. I think it's just-

Ilia  (04:07):

More not currently.

Stefan (04:08):

... Oh yeah. Yeah, yeah, yeah. And I feel like they are [inaudible 00:04:12] the disturb market. People talk about urban versus suburban versus rural when it comes to package delivery, no one talks about the dense urban [inaudible 00:04:21] who lives with their parents.

Ilia  (04:22):

Hang on, hang on. We just specified delivery, it's not even package delivery. What if we're not delivering packages to people, what are for delivering people to the packages? Where the robot comes to your house, picks you up and carries you out of the house [inaudible 00:04:37]

Stefan (04:36):

I'm over here thinking that your dog robot can scale the side of the building and go directly to your window so that you don't have to interact with a person or your terrible parents.

Ilia  (04:48):

You never have to exit your room.

Stefan (04:49):

Yes. And that's a need that I can relate to. If we're be doing that, let's say it's a walking robot, essentially what you could do is, you could have sprinter van full of dog robots that each get a package put on. The dog robot walks up to the building, starts skipping [inaudible 00:05:07]

Ilia  (05:07):

So it's not only walking, it's a climbing machine.

Stefan (05:09):

But what is climbing if not walking vertically?

Ilia  (05:12):

I feel like that definition misses a few things. Your motors have to be even more powerful. So, I think our primary customer actually, you'd think it'd be the people getting deliveries, but it'd actually be the people who own the buildings who you need to compete to get the least damage to their surface of their building by these robots walking up and down [inaudible 00:05:35]

Stefan (05:36):

Your business could be one part robotics delivery starting up-

Ilia  (05:40):

And one part mafia.

Stefan (05:42):

... Or just the other part, architectural design and construction firm. If there's ever been two businesses that flow together better.

Ilia  (05:49):

Robots and construction firms?

Stefan (05:50):

Yeah.

Ilia  (05:51):

Well listen, what's another spin we could take up most? Robots walking up the side of the building. So the inevitable question is, okay, well why not just use drones then, why do we need to walk up the side of the building?

Stefan (06:01):

Well, the drones have limited carrying capacity, whereas walking-

Ilia  (06:06):

I feel like walking up a foot wall could have some limitations on [inaudible 00:06:11]

Stefan (06:11):

... Not if you're okay with breaking a few windows.

Ilia  (06:13):

Yes.

Stefan (06:14):

[inaudible 00:06:14] funnel money into our construction business, if we're going to break a few windows.

Ilia  (06:18):

It's the robotically abled broken window fellas. Yeah.

Stefan (06:21):

All right. What other walking delivery things could you do? Maybe you could be... Okay, how about this? In TV shows, there's a lot of stuff that happens in the sewers, and I feel like the sewers in grownup life are a far underutilized form of transportation. So the walking robot could come [inaudible 00:06:40] out of the sewers, basically leave an Amazon warehouse, you'll be into the sewer, show up in front of your house, go into your living room-

Ilia  (06:46):

So what you're really saying is Stephen King's IT sponsored by Amazon?

Stefan (06:51):

Yes.

Ilia  (06:52):

That is even more terrifying than the original. Also steals labor. Yeah, sewer [inaudible 00:06:59] what up?

Stefan (06:59):

I feel like sewers are actually relatively constrained to [inaudible 00:07:02]. There's only so many dimensions that sewers are looking. Manhole covers are all circular as that interview question that we've all gotten. So, you just have to be able to pop off manhole covers, crawl through the sewers, and in some... There's very little sewer traffic. If you go into the sewers, you actually might be able to beat normal delivery.

Ilia  (07:23):

I feel like what-

Stefan (07:24):

The robot could gallop through the sewers.

Ilia  (07:26):

... I feel like once there's more than one robot in the sewer, you're going to have inherent traffic jam. You [inaudible 00:07:31] sewer traffic controls.

Stefan (07:32):

Yeah. I feel like later generations could climb the walls of the sewers. So passing each other like Will Smith in Men In Black, just run up the side, like little... Yeah.

Ilia  (07:43):

If they're going to climb arbitrary walls, they can climb the sewer. That's fine.

Stefan (07:46):

Yeah.

Ilia  (07:47):

I feel like the other problem you're avoiding here though, is that manhole covers tend to be in the middle of the street. So what you're going to have to do is, you're going to have to spend a bunch of your autonomy on predicting when a car is going to run over you as you can exit and [inaudible 00:07:59]

Stefan (07:58):

I actually don't think that's a big problem. First you blow the manhole cover off, then you send a periscope up. And you only have to blow the manhole cover off once, because most of the time when they put that manhole cover back down, they're not going to bolt it down or anything.

Ilia  (08:11):

That's true.

Stefan (08:12):

So you just have to push it off, put a periscope up, watch for a no car. I think this is a suburban delivery solution anyway.

Ilia  (08:19):

Yeah. And you don't even replace the manhole cover. You partner with all these autobody shops that'll have to repair all these destroyed cars.

Stefan (08:26):

No problem. Let's price this out a little bit so we can move on to our primary segment. So, a [inaudible 00:08:33] robot right now, my understanding is about $120,000.

Ilia  (08:35):

Something in that range.

Stefan (08:36):

Yeah. So let's say, the sewers are a little bit less nice than say an industrial warehouse. So let's say a good production scale dog robot's $200,000. I think most wheel delivery bots are aiming for something like a dollar, one to $2 per mile. So probably, maybe our average delivery gets us $10. But I'm not entirely confident of sewer speed. It's definitely faster than the two miles per hour that most sidewalk robots go. So maybe you could move at five miles an hour. You could be generating, what's that, a delivery an hour, $10 an hour. The sewers are open all night. So you do 24/7 sewer delivery.

Ilia  (09:17):

It's only 2,000 hours to payback, huh?

Stefan (09:19):

Yeah, that's about what the lawyer gets.

Ilia  (09:20):

2,000 hours?

Stefan (09:21):

Yeah, 2,000 hours a year. Yeah. So here you have it, for any of you interested. We have early wait lists signup available right now at sewerdog.net, the .com was taken unfortunately. Sign up to join the wait list and find out when sewer dog will come to a local suburb near you. Sewer Dog, it'll get there, but it may not smell good. So now it's time to dig into the real interesting stuff where I finally asked my co-founder, Ilia  here, what exactly he's ever done to be qualified for this job.

Ilia  (09:53):

Definitely not rob banks.

Stefan (09:57):

So to start off with, you went to a school that I believe copies my Alma Mater, Drexel University, is what's mostly known for, named after a trivial engagement in the Napoleonic War.

Ilia  (10:10):

Isn't British history, yeah. Yeah. No, that's right. I went to University of Waterloo, and just as a note-

Stefan (10:15):

Referred to as the Drexel of the north.

Ilia  (10:17):

... Yeah, yeah. I'm not so sure about that, but I will often wear my Waterloo t-shirt, I've never seen you wear a Drexel t-shirt. I feel like the branding game is off. Exhibit A. I think the opposite is true. But what you're referring to I think is our co-op programs that both schools employ. And I think that was actually pretty key in being able to explore robotics and try it out early on. I was really lucky to join Clearpath Robotics really early days.

Stefan (10:44):

So what is early days? You were employee 700 and you got a blue name badge instead of a red one?

Ilia  (10:50):

That is something I'm sure they still aspire to, they're not quite there yet, but they did have name badges by the time we left. But no, early days was the fore-founders plus one or two people. Started off as a university. Waterloo has capstone projects. It started off as one of their projects where their whole idea was doing mine sweeping robots, hence the name Clearpath Robotic.

Stefan (11:11):

And my mine sweeping for those listeners who don't know, is a task that is mimicked from the classic Windows game.

Ilia  (11:18):

Yeah. So basically robots that solve the game for you. But it's actually, we're joking obviously, but the closer analogy is even more funny, is you can do mine sweeping two ways in robotics. You can make really expensive, complicated robots that predict stuff and measure where the mine is and go and try to dig it out or warn a human operator. The Clearpath approach and arguably the more successful approach, is just randomly runs through the field hoping to blow up your robot.

Stefan (11:43):

You do that as a human or as a robot?

Ilia  (11:48):

As a robot. For sure as a robot. The robot randomly runs through the field and it's cheap and it's replaceable and it's hopefully cheaper than the mines you're clearing. And so if you could just deploy enough robots, you clear out your field. So early days they were working on this project where they would deploy, I think four or five of these robots from the back of a vehicle to clear out a minefield, hence the logo and the name and those kind of things. By the time I joined, they had moved on to more developing stuff for academics. So development platforms-

Stefan (12:16):

Well, let's get [inaudible 00:12:16] of that history of Clearpath. What was the world of robotics like through 2012 when you tried this out? What percentage of all the problems were solved? 80%, 90%, 98%? How many things were easy?

Ilia  (12:28):

Practically everything was a pain in the rear. One of the stories I was getting into is, one of the early, early surface vehicles that we had built, we ended up having to test in Kitchener City Hall's front pool or fountain, 'cause there was no other easy place to test it. We basically carried it out-

Stefan (12:47):

Well, what was the application?

Ilia  (12:49):

... So, it was a surface robotic vehicle for automatically gathered water samples and depth readings of ponds and those sort of things.

Stefan (12:56):

How deep was the fountain?

Ilia  (12:58):

Maybe knee height, roughly.

Stefan (13:00):

It's a good thing to measure accurately.

Ilia  (13:01):

Yeah, exactly.

Stefan (13:02):

It's a terrible thing when fountains grow on you.

Ilia  (13:06):

Got a few funny looks, but I think that's probably not the weirdest thing that Kitchener experiences. It worked out reasonably well. No, everything was a pain. I think one of the big reasons why ROS ended up being such a big part of Clearpath, was that there was no standard, and ROS being the robot operating system-

Stefan (13:25):

When did ROS come out in relation to when you were at Clearpath? Were you at pre-ROS Clearpath?

Ilia  (13:31):

... Yeah.

Stefan (13:32):

What was that transition like?

Ilia  (13:33):

Wonderful. So, pre-ROS was kind of, every time we would finish a robot we'd have to call over the CTO, Brian [inaudible 00:13:41] to wave his hands and do the magic links and cantations to enliven the robot with the software, which nobody really understood and had a whole bunch of weird tool chains, and yeah, it was pretty horrible. And then the post-ROS process was, you log in, you've pulled some packages and it works. So that's quite a difference.

Stefan (13:58):

How ready were they to switch from their homegrown, "We built all of this robotics code from scratch and C++.", to like, "Oh, here's a framework that speeds things up." Was it-

Ilia  (14:12):

It was drop everything and run for the hills level already, because it was so clearly the better direction. I don't think I've seen the tool chain that the early robots were built since those days, they've completely evaporated from history. And this was even at the time when Clearpath was a small, pretty scrappy group ,and dedicating time to anything was expensive. And it was pretty clear that not using a home cook thing and actually going to a standardized platform was pretty big deal.

Stefan (14:40):

What types of projects did you work on at Clearpath?

Ilia  (14:42):

I started off in electrical, designed some of the boards for that surface vehicle. And then over time kind of went from electrical more and more into software. Did a bunch of, some controls for some arms that were related from other manufacturers. And then kind of progressed through to running a small team supporting new PR2, which is one of the big ROS robots.

Stefan (15:03):

What was the PR2?

Ilia  (15:04):

It was kind of one of the early humanoid-ish robots for ROS. What's-

Stefan (15:09):

Elon Musk just built one just like that, right?

Ilia  (15:11):

... Yeah, definitely. Definitely. I'm excited when that will be ready for public use in 30 years or so, or next month, depending who you're listen to.

Stefan (15:21):

Depending on how many of the engineers end up having to work on Twitter.

Ilia  (15:24):

Yeah. With hand printed code [inaudible 00:15:28] or whatever he's been talking about. But the PR2, what's interesting about Willow Garage and the PR2 and ROS development actually is, there's one tool package that came out of that that everybody in the robotics world is very famIlia r with, is Open CV, Open Computer Vision. A lot of the libraries for processing images to actually do something useful with them in software came out of the same project as ROS did, and they both came out of Willow Garage in a lot of ways. They were adopted and adapted and so on, but it's a fair argument to say that Willow Garage was the birthplace for both of them. And then after Willow Garage shut down and moved on to other projects, PR2 still had about 50 units worldwide and they had to be maintained and updated.

Stefan (16:11):

How great was the PR2? Was it walking around, opening doors, maybe doing lock picking to open the doors and then somersaulting through the recently opened doorways?

Ilia  (16:22):

As all of Robotics, some things that seem incredibly amazing, it could do relatively easily.

Stefan (16:27):

Like what?

Ilia  (16:29):

A good story was, one of the developers, Austin Hendrix locked himself out of the Willow Garage office one day. Came in early, nobody was there, he realized he forgot his key card. So instead of going home, he logged up to the wifi, remotely tele-operated the PR2 to open the door from inside and got into that building. So, sounds like magic, right? But pretty true-

Stefan (16:50):

Wait, you're saying he used teleop, why don't you do that with machine learning?

Ilia  (16:56):

Oh, machine learning. That's is pre-modern machine learning in a lot of ways. Open CV was a lot of classical computing, so that seems amazing but is fairly straightforward. But other stuff that spent a lot of research time, there's a lot of videos you can look up online of PR2 folding clothing or folding a towel. And they were very impressed that they got the time for folding a single towel from something like three hours down to something like an hour and a half. That's quite impressive-

Stefan (17:23):

That's way better than the two weeks it takes me from a clean laundry hamper to folded laundry.

Ilia  (17:28):

... That's why I skip it. So folding towels, you could imagine is a pretty complex task. So a lot of that was magical. A lot of the developers on it did cool stuff like, "Go to the fridge and fetch me a beer.", autonomously, play pool autonomously. Stuff that still today is quite difficult to manage, and a lot of it was-

Stefan (17:47):

Well, I saw this one video that I've been meaning to ask you about online and notably with how it relates to your relationship with your wife. I saw a video where you were dancing, not with your wife but with a robot. Now, should I inform her that you're leaving her for the robot soon? How easy was it to teach a robot to dance? And most importantly, can it do the tango while biting onto a rose?

Ilia  (18:10):

I will excuse myself by saying I was practicing for a wedding dance. I'm sorry. No, but that was actually even before I met my wife, I was at Clearpath Robotics and I was working on the, what was it called? Baxter. The Baxter robot, which is a humanoid robot, also again, without legs, just like the PR2, because legs are hard, as we talked about in our walking robot segment.

Stefan (18:31):

Funny, Meta feels the same way.

Ilia  (18:34):

Feel like the cause is there [inaudible 00:18:36] Same idea, I guess. So Baxter was kind of this approach of collaborative soft dynamic robots that can work with people in factories and small assemblies from Rethink Robotics, I think they shut down a year or two after I did that project, hopefully not related to that project. But that robot was mounted on top of a Clearpath mobile base, a ridge back framework. And for a demo basically, I wanted to make this ballroom dancing robot because the Baxter had force sensitive arms. So if you pushed on the arms, it would pull away from you. So I just transmitted that force down to the base, where you could just move the arms and it would move the whole base and the whole robot around.

Stefan (19:16):

Do you think there's much of a business to be made in robotic dance partners?

Ilia  (19:20):

Not in the next 100 years, maybe after that. But what it taught me about ROS along that line, was that that demo, again, which seems a little bit like magic-

Stefan (19:31):

Well I guess, why is that so hard. I feel like if you're pushing back on a robot, obviously me and go in the opposite direction of me.

Ilia  (19:40):

... Yeah, I think the part that's difficult is it wasn't one robot, it was a Baxter on top of a ridge back. So two completely different systems that had never worked together.

Stefan (19:49):

So was that just a matter of translating the force feedback in one into a directional command of the other?

Ilia  (19:55):

Exactly, exactly. So it ended up being relatively straightforward, where you just take the force that you measured at the wrists as a ROS message, send it to the base, small Python script that would change a force into a desired velocity. Send it to the base and there you go. So that kind of thing wouldn't have been possible without a common standard where the messages were the same, where there was an easy way to connect to it, where there's an easy way to share data. So it turned something that would've taken, before ROS, maybe a month of effort to try to write custom drivers and custom software into an afternoon.

Stefan (20:28):

That's pretty cool. So moving out a little bit. After Clearpath, you were at Amazon. And a question that, that anyone who's ever had to pitch an investor can relate to is, "Wow, your idea seems really neat. What's going to happen when Amazon or Google or one of these really smart companies, business like that Jeff Bezos, he knows what's going on. What's going to happen when they just build what you're building? Won't they do way better?" You were working at Clearpath in a 200 person company, a 100 person company today or something like that. And you went to Amazon who has more money than, I don't know, than Dragons in Tolkien stories. You went to Amazon and how come all the problems weren't immediately solved? What were you working on? Why aren't we using that robot every day?

Ilia  (21:15):

Yeah, the Astro robot is available for sale, it does exist, it's a little bit expensive for what it does. But it's an interesting platform. Amazon didn't have one robotics group, it had several. It had one in, and I'm sure I'm not going to list them all, but of the ones I knew of, there's one in the warehouses of course, doing their warehouse robots. They're ex-Kiva Systems guys. There was us doing consumer robots. There's one doing delivery robots. Speaking of walking robots for delivery, they shut down their delivery robot system or their delivery robot group just recently. There's obviously the one doing the drone delivery as well.

Stefan (21:46):

NASA of Zoox.

Ilia  (21:48):

Exactly. So lots and lots of different-

Stefan (21:50):

Why is it that they can, for example, take with the guys that Zoox are working on and apply it to sidewalk robots and then apply it to home robots?

Ilia  (21:59):

Yeah, yeah. The main challenge of Astro was it has the equivalent computing power of, let's say cell phone from three years ago. So you got to do all the complex computation of a robot on something that really doesn't have the power to do something-

Stefan (22:14):

But won't chips get cheaper and cheaper and then you get more and more ability?

Ilia  (22:18):

... For sure, for sure. Moore's Law, which is this kind of rule written that-

Stefan (22:23):

I think everyone who's listening to this has heard about-

Ilia  (22:25):

... Yeah, exactly. It's sort of stalled out in the last little while. You're getting more cores, you're getting a little bit more efficiency, there's special purpose circuitry that's improving for machine learning and those kind of things. But raw computing power hasn't actually kept up with Moore's Law. And so you could wait two or three years, you might get better memory bandwidth, you might get better machine learning, but you're not going to get raw computing improvements. So we're off that treadmill a little bit unfortunately.

Stefan (22:51):

... But I guess more broadly, for those of you who don't know, before the autonomy boom that started in like 2016, the biggest robotics acquisition ever as far as I'm aware, was Amazon's acquisition in 2006, 2008 of Kiva Systems, a warehouse robotics company for something like $700 million. So they bought the company in 2007, give or take a year. They're Amazon, they have a bunch of money, they could pour a lot of money into that system. They've scaled it. They have lots of data from it operating. They have a seasoned robotics team. How come they can't build a home robot?

Ilia  (23:30):

Yeah. Well, few things. First of all, the Kiva Systems, the entire goal of it was to have really dumb robots. So if you look at a Kiva system, and this is changing just recently, their newer robots aren't this way, but their older robots, what they would do is, they would have very few sensors, they wouldn't localize very much and they'd read these barcodes off the ground. So all the time that they're moving, they're checking their position against a grid pattern. They're really not a general purpose robot. Whereas the challenge with Astro is, you have to be able to drop it into somebody's house, make no changes to their home, and be able to find your way around it. You can't cheat with barcodes, you can't cheat with known locations, you can't cheat with pre-made maps. You have to do everything [inaudible 00:24:10].

Stefan (24:10):

And how these robots driving around these warehouses and they have some sensing, they have some compute, they have some, granted not all of it. Why can't they just build a data set that you could then throw into some machine learning guy and train it on how to localize itself.

Ilia  (24:22):

The early Kiva robots didn't even have cameras, they were just blind. They just had bump sensors as a absolute last resort of something going horribly wrong. Again, the newer ones are different. But your broader question about sharing data sets, the trick with machine learning is, you need to have enough samples of that environment that you want to work in. And a warehouse environment is very different than a home environment.

Stefan (24:47):

You haven't seen my house after Prime day.

Ilia  (24:50):

Lots of Amazon boxes, but probably not a lot of people. It's hard. And part of what we're trying to do here at Polymath at least, is to try to generate a general understanding of navigability and industrial spaces like mines or farms or those kind of things. But even that's proving very challenging, where your data set has to become so big and the machine learning understanding with space becomes so tenuous that you kind of exponentially have to grow how much data you gather to get the same quality as if you just said, "You're in a house, you're in a factory, you're on a farm, learn only the things relevant to those."

Stefan (25:26):

When you say it has to grow, what does that mean? Could you give me a sense of amount of data that you need to [inaudible 00:25:34] like that?

Ilia  (25:35):

I don't have great numbers on this because we're still working on this ourselves. But comparatively I would say you need a few terabytes per space if you were to do just a farm. So if you're talking about three or four terabytes of data for just a farm, if you then move to something like a farm and a mine and a industrial site, you're talking 300, 400 terabytes. And that's not just 300, 400 terabytes of just data, that's annotated data that's been checked for errors, that's been classified, that's built into a correct pipeline, that's relevant to your use cases. So not just the same highway or not highway, but farm over and over and over, but different farms of different situations with different weird occurrences. When you grow something exponentially like that, it makes it so much more difficult to find those black swan events that you need to train your dataset to actually become robust to real world events.

Stefan (26:28):

Is that part of why things like on road autonomy for robo-taxis is so hard, because say the Financial District is so different than from the Marina, is so different from Palo Alto, is so different from anything going on in Texas?

Ilia  (26:40):

I don't know if the districts or the main difference, because on road autonomy doesn't really care what district it [inaudible 00:26:45], it more cares about road situations.

Stefan (26:47):

Well, [inaudible 00:26:48] dense urban, tight in the Financial District. At the marina, there's a bunch of bros coming out of happy hour having their bevys. And in Texas, they're probably riding a horse.

Ilia  (27:01):

Man, Japan and Texas. I think the better example, I gave a talk recently at [inaudible 00:27:07] Forum, where I threw up this piece I found from Twitter where somebody was recording their Tesla's, "full self-driving package", where the systems thought that there was stop lights flying towards it constantly. And then they pan the camera up and you see that it's actually a truck that's carrying stop lights that are going to be installed somewhere. And in my entire life, I've never seen a truck carrying stop lights, ever. And it probably never will, because it probably happens one in a million miles. So that kind of data is almost impossible to capture even for just highway driving.

(27:44):

And so, as you expand to more spaces, you again exponentially have to grow these black swan events. Now one approach is you just say, "Okay, well I'm going to ignore that situation.", but then you're not really talking about full autonomy. If the full self-driving system had a software built in that said, "If you see a stoplight, you have to check if it's valid and if it's not valid, then assume it's broken and maybe slow down.", that would've caused an issue at that point. So as you try to build up these scenarios, you start to see them fighting each other. Applications that make sense on highways don't make sense in dense urban environments. And so, if you try to build this master model that applies to both, you might end up with these enormous data sets that's [inaudible 00:28:25]. Yeah.

Stefan (28:26):

What exactly is the Astro, what all were you working on at Amazon?

Ilia  (28:30):

Yeah. So, the pitch of the Amazon Consumer Robotics Group was, is Robotics far enough along now to actually start to become a consumer product?

Stefan (28:39):

And the answer was a hard yes?

Ilia  (28:41):

The answer was a maybe, I think. I think Jeff Bezos was a big supporter of it. And I think there's continuing support. I don't know, I'm out of the group now. But I think there's continuing support for it. But the only example we had of that was robotic vacuum cleaners. There hasn't been anything else that's really-

Stefan (28:58):

To be fair, robotic vacuum cleaners are one of the most successful types of robots in the world.

Ilia  (29:01):

... Absolutely, absolutely.

Stefan (29:02):

And I know we're sarcastic a lot here, but that is a full-throated, it's really... If you look at the robots that matter in the world, it's Predator Drones and Roombas are the two flavors of robots that have had any amount of scalable success in environments that aren't controlled by roboticists.

Ilia  (29:24):

Yeah, exactly. So, it's a good bet. And I think, I don't know the answer. I don't know the answer if we have a good market for this.

Stefan (29:32):

Why would you want a robot in your house? I'm not a privacy guy, but I know people... I'm a big Sonos guy and I know they have this whole side of the product line that doesn't have a microphone on it, because people don't want Google to listen in on when they're ordering something on Amazon.

Ilia  (29:53):

Yeah, yeah. Not to reiterate too many Amazon talking points, but the Echo kind of kicked off the whole smart speaker idea in a large way. And for the first few years, it didn't really do much. You could ask it for time, you could ask it to play music, but it didn't really have all the skills, smart room integration, all that kind of stuff. Amazon's goal with the consumer robots is kind of the same thing. You start with the starting point and you build up from there to see what actual apps and skills and those kind of things can you build on top of the platform.

Stefan (30:27):

The list price of Astros, which is available publicly online, this is nothing private, is $1,500 a pop. I feel like there's a fundamental challenge in building a platform product where the cost of entry is buying a high end TV. If you're an electronics nerd, you'll buy a $1,500 TV or an $800 TV plus an $800 Sonos soundbar, itself is a, hide that receipt from your significant other. But a little robot that might not do anything that seems like a whole different scenario.

Ilia  (31:01):

Yeah. So, you might notice Stefan is very partial to Sonos, maybe we should talk to them in the future.

Stefan (31:06):

They should sponsor us.

Ilia  (31:09):

Yeah, exactly. But yeah, you bring up a great point. I think the reason why iRobot and early Roombas, not today's Roombas, and why the Echo succeeded, was that they were cheap. They were reasonably cheap impulse buy. Amazon's put a lot of their new tech through this kind of invite only day one something programs, and that's how they're hoping to get this data back to understand what the actual use cases are so they can make those lower cost targeted applications. Right now, it's like a throw everything and the kitchen sink at it. It has way more sensing than it needs, it has arguably way more compute than it needs, even how hard it is to get it there. But a lot of the stuff that it does, navigation, localization, those kind of things could in the future be done by specialized circuits, application specific integrated circuits that just do that one thing, offloading the processing, lowering the cost.

(32:01):

And I'm sure they'll get there at some point, but they need to understand what exactly is the main target market, what's the killer app for robots. And robots in general have suffered from this, what is the killer app for this thing? What exactly do you need to do with it?

Stefan (32:16):

Yeah, maybe let's think into that. What killer apps have you seen before?

Ilia  (32:19):

For robots in general or home robots?

Stefan (32:21):

As a bigger question of an industry of a discipline that you spent your entire adult life thinking of, as a fun fact to listeners that [inaudible 00:32:31] barely know myself. Illia chose becoming a roboticist over becoming an opera singer, thus making him less irrelevant to songs done by the band Cake, who I'm also partial to. So, why is it that robots don't matter right now? Where should they matter?

Ilia  (32:48):

I think part of our theory of operation for this business is that robots are still waiting for its 1980s computer days. They're still right now treated as specialized equipment owned by big companies, warehouses, industries.

Stefan (33:05):

I know I have my IBM punch card system at home. I used to use that every day.

Ilia  (33:08):

Yeah, organize the recipes. Right?

Stefan (33:10):

Yeah.

Ilia  (33:10):

Computers made the same thing. Why would you need a personal computer? Would the wife at home organize recipes? What are you going to do with this thing? Same kind of shortsightedness and cost has impacted robotics. And I think outside of these industrial spaces, it's been tough to find that niche because the technology's isn't there. One, the hardware, batteries and actuators need to improve. But the other thing is, you need this generalized layer of self-awareness for the robot. And I don't mean self-awareness like AI, I just mean, the robot knows where it is, knows how to get somewhere, knows the extent of its own body, knows how to not hit things. And that's been missing. That's been missing.

Stefan (33:47):

But how hard can that possibly be, Illia? I've seen people build robots for 30 odd years. John Deere and Caterpillar were looking at autonomous excavators in the eighties. How possibly hard can it be to go from point to point and not hit stuff?

Ilia  (34:03):

It's still a challenge. Ask my two-year-old. People still tell you how hard it is to not hit stuff when you're going for point A to point B. But no, it is. It's still on the cutting edge to do it reliably.

Stefan (34:18):

Why is it so hard today to make a robot that can go from one point to another without hitting anything?

Ilia  (34:25):

Yeah, I think, as I mentioned, the industry hasn't standardized on a way to do things. Think ROS is a great kind of, we've talked about ROS on this podcast, is a great toolkit, but using that toolkit takes a lot of experience, takes a lot of energy, takes a lot of very capable engineers, and there isn't yet this kind of plug and play solution like computing has had. You can't just go out and buy a robot that does a thing you want like you could with laptops. And that's really what we're trying to do, starting with the industrial space, but hopefully for all moving machines in the future, to provide that kind of self-awareness layer and to provide the ability for the robot to just get from point A to point B without hitting stuff and to know where it is in space.

Stefan (35:10):

Well, that's terrific. And if you want your robot to do that without spending the first two years of your life working on it, you should reach out to us at polymathrobotics.com. Well, I am of course more than a little biased, but Ilia , thank you so much for letting me ask probing yet often silly questions about your life's work to date. Should we draw next week's cards right now? Is that going to be a new tradition? I'm being told that that is not going to be a tradition by our executive production staff who is uncredited because that's just the type of business we are. With that being said, Ilia , thank you so much for talking this time. Next week we'll probably talk some more, maybe to each other and maybe to others, we'll see. Ilia , anything you want to say as parting words to our incredibly large, massive automated audience of hundreds of millions?

Ilia  (35:56):

Keep building robots.

Stefan (35:58):

And with that statement of confidence, have a great week.

‍