Stefan (00:03):

Live from Automate It Studios in Polymath's global, incredibly impressive, totally not an office share headquarters in San Francisco, California. I'm Stefan Seltz-Axmacher, and I'm super excited to launch this week's episode of Automate It.

Ilia (00:17):

And with you is Ilia Baranov, as well.

Stefan (00:20):

When's he coming in this side? It's not a third party that you're just talking about there?

Ilia (00:23):

Absolutely. I always talk about myself in the third person.

Stefan (00:29):

And after I'm done making fun of my co-founder.

(00:33):

This is Automate It, our weekly podcast where we talk about how exactly you build robots after all. We always start off with a fun game where we spitball a robot into existence. And then after that, today we're going to be talking a little bit about the state of the broader autonomy and robotics industry. Because if a couple blog posts said that it's dead, it almost certainly is, right?

Ilia (00:53):

Basically. If the internet says so, it must be true.

Stefan (00:56):

That's how I learned that Jennifer Hudson is an alien overlord from the reptile hollow earth. With that conspiracy theory out of the way, so this week, Ilia why don't you draw a setting card for what we should be working on.

Ilia (01:10):

Let's see. Setting card.

Stefan (01:12):

Indoor localization.

Ilia (01:14):

On campus?

Stefan (01:14):

Campuses.

Ilia (01:15):

Nice.

Stefan (01:16):

This is a great combination. So let's say in this scenario, yeah, I was a popular kid in college. I got really good at knowing the school administration, I knew all the cool kids, as I am allowed to do. We wear a lot of Abercrombie. You just figured out how to be great at indoor localization, which is easy.

Ilia (01:35):

Let's take a step back? How do we want to localize, right? So here's my proposal. A lot of people are using LiDar ...

Stefan (01:42):

Actually before we get into how we're indoor localizing, indoor localization is a bug banner for me. Just this morning, I was talking to a corporate innovation type of person who challenged me, why is indoor localization so hard in the first place, given that there's all these AMR companies that there's not that many UR autonomy companies and minds, so the AMRs must obviously be easier. So maybe as an intro for those people listening who've never tried to do indoor localization, why is it so hard to know where you are? You're just in a building.

Ilia (02:15):

Well, the first thing that some people realize, but not all, is that GPS doesn't work indoors. When you look at your location on a phone, on Android or iPhone, modern phones, they're actually faking your location by knowing what WiFi connection you're in, what towers you're connected to. They're not actually ...

Stefan (02:29):

Couldn't they just fake it good enough to tell me that I'm in this room?

Ilia (02:33):

Oh, it's faked to 30 meter accuracy, maybe. It's pretty rough. They're not going to be great, especially in rooms they've never been in. So you can't use GPS, so that's out the window. Really what you're left with is you got to look at your environment, almost as a human does, and try to figure out, have I been here before? Do I know the dimensions of the room? Do I know where I need to go to?

Stefan (02:54):

Okay. Well, can't you just do SLAM?

Ilia (02:57):

Yeah, SLAM, SLAM's ...

Stefan (02:58):

Someone told me earlier today, "Just use SLAM."

Ilia (03:01):

Just do SLAM.

Stefan (03:01):

Solved.

Ilia (03:01):

Exactly.

Stefan (03:01):

It's a great acronym.

Ilia (03:04):

The acronym stands for Simultaneous Localization and Mapping. The fun, dirty secret of the industry is nobody actually does SLAM, most companies will map first and then localize in the map. Nobody tries to both build the map and localize at the same time, because you get a lot of creeping errors, your map gets distorted.

Stefan (03:20):

What's such a big deal about a creeping error? You just find your way. I'm here and I want to go through the doorway, what's the problem that I'm going to run into? It's 10 feet away.

Ilia (03:30):

Well, the fun part is breaking down your sentence there. What does here mean?

Stefan (03:33):

From here.

Ilia (03:34):

Here. Yeah, how far are you from the wall? How far is the door? Do you even know that that's a door and not just a gap [inaudible 00:03:43]? Understanding what a door looks like. Now you got to feed an ML and all sorts of stuff, right? The classic approach is you use a 2D LiDar, you map out your space and you figure out the distance to every object around you. And then you, what's called scan matching, where you say, here's the LiDar 2D scan I have. Have I mapped this area before, and if I have, can I line up the scan I'm getting now, to the historic scan I've gotten before? The problem with trying to do that simultaneously is that as you're driving around, you're accumulating error, and so your map gets melty and distorted. There's lots of videos online you can find about error correction for SLAM. And when you get back to the same place, you have this trick of, oh, I've been here before, therefore I can correct all my errors. But that magical step ...

Stefan (04:27):

You need beacons to do that, as well.

Ilia (04:27):

Yeah and beacons can help you a little bit. Exactly. So indoor localization is one of those things that's solved 90%, but getting to a hundred is exponentially more difficult.

Stefan (04:38):

And this is a problem that definitely is easily solved when you throw large sums of money at it, because you're a FANG company and you want to build an indoor home robot.

Ilia (04:48):

My group at Amazon, there's a sister group that was working just on this problem and they spent four years getting something that mostly works.

Stefan (04:58):

And how mostly is mostly?

Ilia (04:59):

Still gets lost. I think they still have to bring it back to the dock, which is a home position that he can recenter on.

Stefan (05:07):

All right so let's ... So Mr. CTO, you've solved some flavor of indoor localization some way.

Ilia (05:14):

Let me just suggest. So for campus, you can of course use beacons, you can use LiDar, you can do all those sorts of ...

Stefan (05:19):

But campuses are large and they don't have a lot of money.

Ilia (05:21):

Exactly, so here's my alternative. If you're going to localize in dorms, why not just measure the aerosolized alcohol content? And what you do is you just build up the feedback. So what you figure out is this season, this semester, this particular dorm as a higher level of aerosolized alcohol content, so I'm going to use that as my starting pitch.

Stefan (05:45):

All right. What flavor indoor localization do you think you would have solved as a ambitious campus robotics founder? What are we working with?

Ilia (05:58):

I'm joking obviously about the aerosolized alcohol, but really what you want to do is you want to mix a bunch of sensors. So you want to mix ...

Stefan (06:05):

Maybe for the purpose of this, we're just saying you've more or less vaguely solved it, but it's still not magically accurate. You haven't solved it to centimeter level accuracy. So in terms of businesses that you could reasonably have in a campus, that are not too unfun, so definitely not campus delivery robots because too many people are doing that and it stopped being fun. What are good things that mobile robots could do within a campus? Maybe it's ... I don't want to do beer delivery, maybe, I don't know.

Ilia (06:38):

One of the things I remember back to high school is they said, "When you go to university, nobody's going to check that you go to your class." So what you do instead is you have a robot that roams the halls, detects a person, grabs them and drags them to their class.

Stefan (06:51):

I think more than that actually. When I was in school, there's some professors who do things like record their lectures, with Zoom more of them started to. But I'm sure there's a decent number of professors who do not record their lectures, who cannot set up AV equipment. But buying AV equipment for every single room in a campus is incredibly expensive. You're probably talking about quarter million, $500,000, investment. So what if we build a robot that can drive throughout the campus and go to classes that are not being recorded, to record them?

Ilia (07:27):

I feel like we'd be breaking some sort of privacy law here.

Stefan (07:31):

Shut up. What's a few dozen sensors and machine learning models pointing at kids talking about sensitive subjects?

Ilia (07:39):

That's the problem that the Astra program at Amazon had, right? The question everybody had is, "Well, if it's on Alexa wheels, isn't it cheaper just to put an Alexa in every room?" Because if you're talking about quarter million dollars, I'm making a functional robot that will work well and reliably.

Stefan (07:53):

I think that the fleet of robots, it's going to be more replaceable, it's going to be more up to date, the 80's stacks and these ...

Ilia (08:01):

And handle stairs.

Stefan (08:03):

You could solve it. So we've met a company who used treads and could take any stairs at 10 miles an hour.

Ilia (08:11):

It's built to batter down doors. I feel like that would be a little bit aggressive to smash into a [inaudible 00:08:17], you are now being recorded?

Stefan (08:19):

Definitely make an appearance. You pop it through the door, aim your camera right at the ...

Ilia (08:24):

Right at the person, yes.

Stefan (08:25):

... Zoom in and make a lot of mechanical noises.

Ilia (08:28):

And of course red targeting laser system.

Stefan (08:30):

Yeah. That's how people know if they're being watched. That delays the privacy.

Ilia (08:34):

Yeah. Increase in efficiency and teaching.

Stefan (08:37):

If we're building this system, it's a self-driving AV cart, is really what it is. So we either need to figure out how to deal with stairs or we need to be able to ... Most campuses are accessible, so we either need to work with stairs or elevators. We also need to know which classes to go to. And since the problem is overcoming tech reluctance, I think what we'd have to do is we'd have to scrape the syllabi of different classes, find out which ones have a Zoom or whatever, Hangout, Teams from like blank and use that ...

Ilia (09:12):

He is a huge fan of Teams. If you want to sponsor Stephan, ask the Microsoft Teams group.

Stefan (09:17):

Yeah. Jesus, it is the worst of all worlds. So we scrape all of that, find the ones that don't have a link. So I'm a Python master because I read three chapters of also called Automate It. I build this scraper, I can tell you which classes we need to record, and then you're good, right? Robot [inaudible 00:09:38].

Ilia (09:37):

Well then you got to localize those classes in space. So if somebody ...

Stefan (09:42):

So we have a pre-map of the campus. They have those nice little pretty maps that they hand out with the trees.

Ilia (09:48):

Those are not to scale. That's actually another fundamental problem with localizing indoors, is as-built architecture design versus what's really there, doesn't actually tend to agree. And so you got to probably add on a little ML layer to detect, oh, somebody's class is in Atrium 103. You're driving by, you look at the door number and you're trying to figure out, yes, this room is Atrium 103.

Stefan (10:12):

Well wouldn't we just have a pre-map that says these spaces are Atrium 103?

Ilia (10:16):

Yeah, a human could go through and label every single thing. For a campus ...

Stefan (10:22):

That seems easier than reading door signs on the fly. All right, so we have our map of where Atrium 103 is. We've detected that there is a class in Atrium 103 that does not have a Zoom link.

Ilia (10:34):

Well, not only that, I think we should organize this based on professors with most amount of people skipping, to get the most benefit. If it's a class where everybody's a hundred percent attendance ...

Stefan (10:45):

I actually don't know if we have to figure out how many people are skipping. You can assume a steady rate of 20 to 40% of students aren't trying to get to class. So if 20% is greater than 50, then maybe you dispatch a robot to it.

Ilia (11:02):

Yeah, yeah. Exactly, but then you got to monitor how many people are in each class.

Stefan (11:08):

Well, when you barged through the door ... Actually that might be a way to do it. What you could do is if the assumed skipping rate is 30%, you get into the classroom, you do a quick people count, and if the people count is greater than 70% attendance, then you go to the next class.

Ilia (11:25):

You could people count, you could also just have a microphone and record the amount of screaming that happens when you smash down the door.

Stefan (11:32):

Well the thing is, we all know about robots is the first time people will scream, the second time people will be shocked, and the 15th time we smash down the door, everyone's going to be like, "Oh yeah, there's the robot smashed another door. Obviously that's how robots work."

Ilia (11:44):

Exactly.

Stefan (11:46):

So we've gotten into the room, we record it. In terms of actually navigating around the building once we have indoor localization, it's basically just point and click, go to this GPS coordinate and don't hit anybody.

Ilia (12:00):

Well, except for the GPS coordinate. You have two layers of the planner. One is the global floor plan which says, I'm here and I need to get to classroom 1, 2, 3, this is the shortest, most reasonable path. Including things like entering an elevator, telling the elevator system that you want to go to a particular floor. Actually, there's a lot of hospitals in, I believe both South Korea and ...

Stefan (12:23):

Singapore.

Ilia (12:23):

... and Singapore, thank you, that have started to actually build these moving carts that talk to the elevator system. That global plan is interesting because you've got to do predictive stuff.

Stefan (12:35):

The group from Opensource Robotics is working on that. They're actually building a standard way so that different robots can ask the elevator to get to the right floor. So theoretically, if this campus is in Singapore or South Korea, we don't need to push the elevator buttons.

Ilia (12:53):

No. And you really, that's one of those classic, robots are the wrong solution. You could make an arm to go and press a button physically, which is going to cost you an amazing amount of money and really unreliable. Or you could just use the elevator control system and just WiFi tell it, and it'll be way faster.

Stefan (13:09):

I had some friends who were working on a delivery bot where their big challenge was they were in a location where the street, it never became pedestrian walking, unless the button was pushed. So they couldn't really leave Stanford's campus all that easily.

Ilia (13:23):

Yep, that's exactly it. That's exactly it.

Stefan (13:25):

So we have that tie in to the elevator system. What types of things do you think ... If we're driving through hallways that are busy and not busy, do you think we can just wait until they're not busy? Or how do you think we'd do the path planning around the busy students?

Ilia (13:40):

I think the best way to do this is you record data over time, over several days or weeks or months, and you predict. If you have to path find through the cafeteria, look back historically, and you say, "Oh yeah, around noon the cafeteria is really full. I probably shouldn't go there."

Stefan (13:56):

That probably would live on the fleet management layer of, go this path, don't go that path.

Ilia (14:01):

Exactly.

Stefan (14:01):

The central artery is almost always busy, avoid it altogether.

Ilia (14:05):

Exactly. And so again, the Opensource Robotics Foundation's work on Open-RMF is what would allow that communication to [inaudible 00:14:13] vehicles.

Stefan (14:14):

So we know the path that we should go on, but I'm still thinking about the weird stuff that happens in campus hallways. There's some trashcans that move around, there's people walking. There might be some hippie kid sitting on the ground looking at his laptop forlornly. There might be a drum circle going on. How do we ... Might be a cold campus, so all these things happen indoors.

Ilia (14:41):

There's two problems. One is making sure your robot doesn't get lost. The nice hack that I've always admired, is some robots who actually have a camera that looks at the ceiling and they map the ceiling rather than the actual floor.

Stefan (14:52):

Really?

Ilia (14:53):

Your robot could be completely surrounded by humans and still know ...

Stefan (14:58):

That's an interesting hack.

Ilia (14:59):

It's a really clever hack or ...

Stefan (15:00):

I'd be like, "Isn't ceiling tile, just ceiling tile, just ceiling tile?"

Ilia (15:04):

But there's tiny little variations, like light fixtures aren't exact in the same position.

Stefan (15:09):

I know and honestly, I don't know how many campuses would care about putting a piece of tape on the ceiling.

Ilia (15:15):

Yeah or just adding a little bit of texture. And so that's a clever hack to make sure you don't get lost even in a crowd. So not getting lost is part of it and the other part of it is just not hitting people. So that's the local planner. Once you have this Google Maps style, you're here, you need to go here, your local planner tells at any given moment, here's which direction to steer and how fast to go.

Stefan (15:37):

So what's our robot look like here? We need a rugged platform for actually moving around. We can probably do wheel diff drive, feels like the right thing.

Ilia (15:48):

Especially indoors.

Stefan (15:48):

Yeah, smooth floors.

Ilia (15:50):

It's meant to be accessible for ADA for a compliance, so smooth ...

Stefan (15:54):

It's probably a round robot so it doesn't get stuck in corners. It's a round robot, it needs a decent amount of power because it's going to ... Probably the AV cart itself is going to weigh, I don't know, 50 pounds, a hundred pounds?

Ilia (16:08):

Something like that, especially with batteries.

Stefan (16:10):

Yeah. Then we're looking ... We also might want to be able to plug ourselves in for live-streaming these things.

Ilia (16:18):

Yeah, automatically plugging in, the PR2 had a great video where it could plug itself in. But again, you need a manipulator and something that, just to use a standard outlet. If you had infrastructure, like a special dock that the thing could just drive on.

Stefan (16:29):

Well, what if you have those annoying floor outlets that they have anywhere where they have AV stuff? Couldn't you just have a little hook that pops up in the floor outlet cover, and then a very simple thing that hits it at an angle and goes in?

Ilia (16:45):

Humans always underestimate how accurate our dexterity is. Actually getting a robot to plug something into a socket, you're talking about tolerances of half a millimeter. Really accurate manipulation. Again, look up the PR2 video plugging itself in, you'll see it takes four or five attempts. Because it's actually, you need a little bit of force to jam in a plug, but you can't really tell exactly where you are and how well you're jamming it in, and some plugs are sticky and some plugs have those ...

Stefan (17:13):

But I feel like floor plugs are more consistent.

Ilia (17:18):

They might be consistent, but you still got to be crazy accurate.

Stefan (17:20):

Man, when you, what's the round part of the bottom of the plug called? Is that the neutral or the ground?

Ilia (17:26):

The ground.

Stefan (17:27):

So you can have a plug that the ground sticks out further, you target the ground into the ground part and then you prong the rest of it up as the whole thing folds into position.

Ilia (17:38):

I'm no electrician, but I think if we make the ground plug too long, the rest of it just won't plug in.

Stefan (17:43):

No, it's like spring-loaded down.

Ilia (17:45):

Oh, it's spring-loaded? So we're custom making ... You're still ... Is it solvable? Yeah. Will it do it quickly? No, it'll probably take 10 minutes of retrying and retrying.

Stefan (17:56):

No, that's 10 minutes of classes these students already were going to miss. So our robot probably can't plug itself in, so it needs batteries ...

Ilia (18:04):

It needs a dock, right? Like a Roomba has a dock, it just drives itself.

Stefan (18:07):

But they're not going to install any AV stuff, so it needs batteries to support four to eight hours of recording. The camera angle probably already needs to be about as tall as a person. And that's helpful because then it won't bump into people as it's navigating the hallways. What sensor suite do you think we're looking at?

Ilia (18:26):

You probably want a mix of LiDar ultrasound cameras. Ultrasound in particular is useful for new glitzy campuses that have glass walls. Cameras aren't really going to see them, your LiDar is not really going to see them. Even your ultrasound's going to have a real hard time, but at least it's better than nothing. There's millimeter wave radar coming online now, it still has some regulatory issues, but that could be very interesting to add.

Stefan (18:50):

But I think the thing is we're probably going to be constrained to a bomb 10 grand, 20 grand?

Ilia (18:57):

Yeah. Our bill of materials, all the parts that we put on the system, that's probably about right. You could get by with just moving slowly, with the ones that you don't have to think ...x

Stefan (19:08):

But then you're probably only recording a class a day. And then my yes would be, we probably have to record three or four classes a day.

Ilia (19:15):

Yeah, yeah. If you staggered them, depending on how far across the campus they have to go [inaudible 00:19:22]. That's going to be tricky.

Stefan (19:24):

All right. So we're probably looking at a robot with a bill of materials of $20,000, which means we could sell it for 40, $50,000 ish. And we probably need, because of all this infrastructure we're going to have to build, we'd probably we'd have to charge 10 to $20,000 per robot per year as ROS.

Ilia (19:45):

Yeah and that seems cheaper than just setting up a microphone in every room. Definitely the right solution here.

Stefan (19:53):

Especially for videos that everyone's going to watch.

Ilia (19:55):

Exactly.

Stefan (19:56):

So coming to you soon, to a campus near you, we need a name. Class Snooper? Coming to a campus near you, the Class Snooper, $40,000. It does everything at AV cart can, plus hit all the co-eds in the hallway. Alrighty, now that we've solved my bad classroom attendance, what are we going to talk about today, Ilia?

Ilia (20:19):

So we wanted to discuss, there's been a lot of thick pieces recently of an autonomy or AI winter. Is the industry over-promised?

Stefan (20:27):

And is that a winter brought about because the overlords have blocked out the sun and we are moving into ground to fight them and retake ourselves from the matrix?

Ilia (20:35):

I feel like our Venn diagram of Robotics podcast and conspiracy theory podcast is, we're really just nailing it. It's a weird conspiracy theory.

Stefan (20:46):

Listen, Jennifer Hudson is a reptIlian from the hollow earth. That's the one thing I know.

Ilia (20:51):

What do you think about this? I've seen a lot of business model issues, tech issues, social acceptance issues.

Stefan (20:59):

I think the two articles that stuck out the most to me were Anthony Levandowski quotes in a Max Chafkin article in Bloomberg Business Week and George Hotts' I'm Quitting, AI, Autonomy Is Not What It's Cracked Up To Be Post. And then I also, some schmucky guy from some company called Starsky Robotics, wrote some article a couple of years ago about how this is all a lot harder of a problem than anyone had thought.

Ilia (21:26):

Ah, it's your turn to say it.

Stefan (21:31):

There's also, I think it's TechCrunch or somebody is doing an overlay into just how bad everyone's doing. Their target a couple of weeks ago were my friends at Embark, who of course as a publicly traded company, are now worth less than their bank account has money in it. And there's a bunch of other stuff like that on the horizon and Iron Ox did a big nasty layoff. Everyone seems to be in trouble.

Ilia (22:00):

And this is the industry as a whole. The tech industry ... Let's try to split out the general tech malaise like autonomy and robotics.

Stefan (22:09):

I think what we're seeing is the other side of what we saw in 2016, 2018, where every third minute an article got posted about, look how cool this ML model just improved, pretty soon, we're never going to have jobs again. And maybe that means we'll have basic income, or maybe that means the overlords will crush us to death.

Ilia (22:31):

There's a lot of this Gartner hype cycle of inflated expectations. I think a lot of it was built because in ML, what you want do is as you're training your machine learning algorithm or generally tech in general, you can really focus down on, I want to solve this very specific, very narrow problem. That's going to give pretty decent results, but only for that very specific slice. For example, ML to detect humans, can do really, really well at detecting humans, will completely fail at detecting dogs.

Stefan (23:02):

Wait, but if it can detect humans, then it can be a human and it can pass the Chinese restaurant test.

Ilia (23:06):

Yeah or the touring test. Some of the really large language models like GPT3 and Open AI's work is starting to go in a more generalized direction, but for that you need billions of dollars and enormous, enormous compute resources to get basic ...

Stefan (23:25):

And it's still can be a little crusty and hard to find where the actual value is here.

Ilia (23:26):

Exactly. What we're seeing in my view anyway, is that a lot of companies that focused on a very specific narrow field, didn't find the economic return they needed in that particular field.

Stefan (23:39):

I think that's part of it. And I call out, both George Hotz and Anthony Levandowski are ridiculously smart people. I've met Levandowski a couple of times, I've been at an event or two with George Hots, they're both very smart people who are smart in the sense of they think for themselves and they examine the facts that they see and come to the conclusions based on the facts that they see with good inductive. Whereas a lot of the autonomy boosters have just been boosting autonomy because other people have been boosting autonomy. And that's always a bit true in a hype cycle, but I have an article that theoretically will be published by the time that this is live, if not it this will escape my Google Docs folder.

(24:24):

That essentially, I don't think the autonomy industry itself is dying or is dead, but I do think it's cocooning. I think essentially, we saw a lot of different flavors of caterpillars and weevils and this and that. Some of those got eaten by birds, some of those died, some of those are now liquefying themselves and are in a chrysalis and are in the process of becoming butterflies. Some of them will be moths. Some of the caterpillars are beautiful, some are ugly. But my overall thesis is that we're seeing some of the most exciting stuff to happen in autonomy right now. But it looks different than, here's some cool machine learning model that I programed last week.

Ilia (25:06):

Yeah, exactly. To go back a little bit in history, a lot of the early robot navigation and human interaction work was posted, if I remember correctly, an MIT early project. A professor gave it to his grad students and was like, "Yeah, solve this navigation problem in a summer. What's the problem?" That was 1970, 1980, and still this problem isn't solved. And so I think a lot of this cocooning is pulling back from the, "I can solve everything all at once," to "I can do this one very specific thing."

Stefan (25:35):

And I think more than that it's not, the valuable robots aren't necessarily groundbreaking machine learning model powered. We got really excited because Google learned how to detect a cat versus a dog. We got very excited about the implications of that. And everyone got very slippery slopey about, if they can detect a cat versus a dog, then soon they'll be able to breed new species of cat dogs that will attack us. And the logic got lost somewhere between if A then B and Z, therefore A with A2. That definitely happened. But what I think we've been seeing with our customers at Polymath is, are these companies who are saying, "What is the level of autonomy that we need so that we can solve a problem for our customer?"

Ilia (26:28):

Exactly. Do you really need to have natural human interactions if you just need to drive a truck? Do you really need to understand ...

Stefan (26:36):

Or not even that, how do you simplify the truck driving task so that it's already solved? We use this podcast to mostly talk about how to build robots, but it might be useful to talk about how to build an app. So to do a similar version of what we just did, no technology card, because that's not really how you start SaaS companies, but maybe it's more like we want to start an app for campuses, and what do they care about on campuses? Maybe they care about scheduling and syllabuses. Well, what tools exist today that I can force Ilia to integrate together so that we can scrape a list ... Or actually solve the same exact problem. We scrape that same list of classes, find out who doesn't have a Zoom link, do a tie-in with TaskRabbit to dispatch a person to show up to that class, pull out their cell phone and record it.

Ilia (27:26):

Exactly. And none of the value there would be generated by ML intelligently predicting what class to go to.

Stefan (27:35):

But more than that, you could build a billion-dollar business in that space without building anything that feels new.

Ilia (27:43):

Exactly, that's what I mean. You don't have to ... Web scraping is ancient and reasonably straightforward to do, TaskRabbit is easy to integrate with. There's nothing there that you need PhDs for.

Stefan (27:54):

Yeah and it really is a matter of, all right, let's hustle. Let's hustle and let's scrape this together. And maybe TaskRabbit isn't all that easy, so to start off with, Ilia will make this website. I will make this scraping thing, and whenever somebody clicks that they want a class recorded, I will show up and record it with my phone ...

Ilia (28:15):

With a microphone.

Stefan (28:16):

... and post it on G Drive, and we can very quickly discover that nobody cares about this product and no one is willing to pay me $50 to show up and record this on my phone.

Ilia (28:26):

Much less build the robot. So let's pull back though, what does this mean for the industry as a whole? To my mind, the lesson learned here is that there's been a lot of companies focused on, here's a particular tech I'm going to apply and jam it into a market situation. And that doesn't seem to work, but what's coming next?

Stefan (28:44):

I think what's coming next is businesses that work that happen to use a tech. A company who I like a whole lot is Kiwi Robotics. I'm an advisor of theirs, they do delivery robots. And Kiwi started because Felipe wanted to do campus delivery. He wanted to build a DoorDash like product, discovered it was really hard to hire Door Dashers and started building robots instead to solve that problem. And they were early, they were the fun travails of a Robotics company. But I think that's much closer to what the successful Robotics companies that are developing right now look like.

(29:19):

Whether it is a lawn mowing company that wants to build a great lawn mowing service and however they get the robot to drive itself is fine. Whether it is a company who wants to make it easier to thin forests to reduce fire firefighting risks. All of those ideas don't require being the 50th person to build an autonomous vehicle or integrate a LiDar into a GPU or figure out path lighting. They really require knowing what a customer cares about, knowing that that problem is best solved with a robot that's buildable today and then trying to build that robot.

Ilia (29:56):

Yeah and I think from a tech perspective, I always love, again, to use technology to solve something. But I would say that these businesses are successful despite their robots, not because of their robots. Their robots tend to be their problem center, not their solutions.

Stefan (30:12):

The businesses are successful even though robots are a pain. They're successful because humans in their particular use case, are so much more of a hassle than really problematic robots. And that's where this works, I downloaded Uber as an app in 2013 after thinking it was already too much hype and too annoying. But trying to hail a cab in San Francisco's financial district at 5:15 on a Friday and it taking an hour and literally waving at every cab that drove by, them all being full, downloading the Uber app, putting my credit card in and ordering Uber and getting picked up, before an open cab ever drove by. Uber's app at the time sucked.

Ilia (30:58):

Obviously Robotaxis will solve this problem.

Stefan (31:01):

Uber's app at the time was terrible. Basically as a company, they were able to hire a bunch of smart early engineers to tie together the Google Maps API. All their pricing was basically from a lookup table in G Sheets, and it was a bare bones crappy thing that broke all the time. But it was such a good product that users were okay with a crappy solution and that's where robots can thrive. If you are in mining, you can't hire people. You just can't get people to live seven hours northeast of Perth it's just the hardest part.

Ilia (31:39):

I've been there. It is hot and it's so hot that the stones themselves over eons have broken into pieces.

Stefan (31:46):

What's the social life there? Good place to be 22?

Ilia (31:48):

Horrendous. If you're living in company housing, there's like one company bar. Yeah, it's really bad.

Stefan (31:55):

Yeah. I think the way that this market emerges from its chrysalis or it's cocoon, to use more common word, is that companies show up where a hustler and a integration focused CTO can make a product that does something. Which is interesting because when I started Starsky, people hated the fact that we were vertical and now seemingly everyone's vertical. And now I think in this funding, in this economic environment, you got to be horizontal.

Ilia (32:27):

Yeah. It's very difficult to dig out enough value from a very narrow application case than it is to be a little bit more general. The example we keep using is that everybody and their uncle was building smartphone, or not smartphone at the time they were just cell phone applications before the iPhone. Blackberry's a big player, they're arguably the dominant player. Everybody had their own OS.

Stefan (32:50):

They were more like a telecom connected PalmPilot.

Ilia (32:55):

Or pager. Then when Apple came out and later Android, you had more of a platform where you could build, where you didn't have to build a cell phone to be able to deploy your chat app. You didn't have to own the entire text.

Stefan (33:07):

Similarly, if you wanted to build a great cell phone, you didn't have to learn how to be good at building an operating system and phones suddenly, most importantly, could have more games than just snake on it.

Ilia (33:19):

Old school snake on a Nokia. Solid.

Stefan (33:22):

That's probably my number one mobile game of all time. But it's nice now that I have Clash of Clans or whatever the cool kids play these days.

Ilia (33:30):

Old man comment.

Stefan (33:31):

Yeah, I know. Clash of Clans is probably dating in and of itself, basically I think we don't, the autonomy industry isn't dead. There's still probably going to be some of these massive raise a hundred million before they have anything companies that show up. If I said that though, none of those will be funded again. An announcement would happen the day after this podcast goes live, and everyone would point it out as when an idiot.

Ilia (33:55):

Oh, it's a fate bait. Anytime you say it's not going to happen, it's definitely going to happen.

Stefan (34:00):

But I think realistically where Waymo and Cruise are at right now, is they're focusing on how do we scale unmanned Robotaxi traps? I don't think it's easy to rationalize in giving 2 billion or 10 billion to come to a company who wants to get to that stage eight years from now. And if you are doing stuff that isn't on public roads, what the market really wants these days is, "Okay, cool. Show me your business."

(34:25):

At Starsky, we saw this whole scenario in 2019 where we basically built a truck that could drive without a person in it. We built a business where the end customers thought we were holding their freight with unmanned trucks, and I went and showed investors and they're like, "Huh, this Robotics thing seems hard and expensive. Have you considered just winding that team down and just being a software enabled trucking company?" "Oh, well, we wouldn't be able to scale truck drivers that much because of these problems." "Oh, well then what about not having a trucking company, just doing autonomy?" "Well, the robot's really hard. So we're building a system that can only really work in five years, might be able to work out a total stretch of maybe 10,000 miles." "Oh, I think we're going to pass," because the reality of being vertical was that we built both a crappy robot and we built a crappy trucking business.

Ilia (35:16):

So as an alternative, maybe what we recommend and our business philosophy is that you guys build a good business and we can focus on building the good robot.

Stefan (35:28):

Or the good autonomy part of that robot.

Ilia (35:30):

Exactly. We like to think that's the hardest part. I'm sure everybody thinks that's the hardest part.

Stefan (35:35):

We both think that's the hardest part, and know that in reality, figuring out how exactly to add value to a port or a landscaping business or an oil field or a whatever, really kind of hard.

Ilia (35:49):

That's the actual hard part. Again, succeed despite your robots, not because of your robots.

Stefan (35:55):

Well, awesome. Ilia, what are we going to talk about next week?

Ilia (35:57):

Next week we're going to talk about how to measure robots. What is Balidan? What is just going to lead you astray?

Stefan (36:03):

That's easy, just look at disengagements promote.

Ilia (36:05):

Yeah, it's a fantastic metric. We should legislate that.

Stefan (36:11):

And on that note, for those of you who are triggered, I'll leave that to you for next week.

Ilia (36:17):

Signing off. See you next time.

‍