Episode Transcript

Note: this transcript has not been edited.

Stefan Seltz-Axmacher (00:04):

Hey everybody, I am Stefan Seltz-Axmacher and this is Automate It, a Polymath Robotics podcast on automation whether that's robots we've built before or crazy combos of technologies and settings that someone might automate. I of course, am Stefan Seltz-Axmacher CEO and Co-founder of Polymath Robotics and with me is...

Ilia Baranov (00:24):

And I'm Ilia Baranov. I'm co-founder and CTO here at Polymath and this is the first time that we're doing this.

Stefan Seltz-Axmacher (00:29):

Which means it's probably going to be the best.

Ilia Baranov (00:31):

Absolutely. Absolutely. So let's see how it goes.

Stefan Seltz-Axmacher (00:36):

To start off, we're going to play a little game and this is a game inspired by basically how Ilia and I talked to all of my friends and his friends and everyone's friends and your friends who think about building robots, had built robots before. Basically we're going to take a tech assigned at random from a stack of cards and we're going to take, what are we going to take Ilia?

Ilia Baranov (00:55):

And we're going to take settings. We got little physical cards here and we'll pick them out and see what we got. So let's start it off picking randomly here. We got home use.

Stefan Seltz-Axmacher (01:06):

And on my tech side, I actually just drew a setting card, the tech side we have not 3D printed. So basically we have to come up with a automated company to use robots somehow in homes without 3D printing.

Ilia Baranov (01:24):

Well, everybody knows when 3D printers first came out, they were just going to revolutionize everything. Everybody's going to print everything at home. And I kind of fell for that belief early on. I definitely have a 3D printer at home. I use it quite a lot.

Stefan Seltz-Axmacher (01:37):

I would say that is true. Ilia is actually the only person I know who has ever given me something that they made with their 3D printer. [inaudible 00:01:44]. That's not only did that come from his general Canadian generosity, but frankly I think there's an underlying need of him as a 3D printer owner to find uses for the 3D printer, which can then coalesce with his generosity and desire to build weird things and then hands Fibonacci years or whatever.

Ilia Baranov (02:06):

How else do I justify buying all the rolls of 3D print material to my wife? I have to some use for it, but in the home use. So what have I seen? If we can't use 3D printing I think we're probably off to a better start.

Stefan Seltz-Axmacher (02:19):

I think this is going to be more fun if we do a similar take on where most people would use 3D printing in home use, and I think that's obviously 3D printed homes and how do we build a robot that builds a home without 3D printing? So there's some obvious ideas that now rational, sane ideas might be the type of things that we do for work, which is say automate construction equipment that could build a 3D printer, but that's not really fun. So how about just a humanoid robot to build a house?

Ilia Baranov (02:51):

Yeah, because if you're going to lab bricks, you make a house. Definitely, the best way is first build a humanoid that works just as well as a human and put down bricks and that's- Absolutely.

Stefan Seltz-Axmacher (03:03):

Yeah, absolutely.

Ilia Baranov (03:04):

No, there's been a few good startups in this space doing brick laying, actually depositing bricks, putting down mortar, that kind of thing actually I think is much more promising than three printing the whole house. Have you ever seen the finishings on these three printed house?

Stefan Seltz-Axmacher (03:16):

Yeah, I mean it looks a bit like you're living in a cave.

Ilia Baranov (03:20):

Yeah, cave.

Stefan Seltz-Axmacher (03:22):

I've been to Cappadocia, which is an underground city like dug out of sandstone and that's about the vibe that most 3D printed houses seem to be going for. It's like the Flinstones need the Jetsons and no one's really happy. Just like when the Flinstones met the Jetsons.

Ilia Baranov (03:37):

Techno cavemanism, that's something [inaudible 00:03:42] wasn't going to say. But yeah, definitely deploying bricks and trying to build a house so that what would be the main problem.

Stefan Seltz-Axmacher (03:48):

So I think the business model for something here is in any of the trades, there's a shortage of workers. Basically, I mean the hype on this is always, "Hey, people stopped going for trades because everyone went to college, blah, blah, blah, blah, blah, and now I can't find the master mason." So maybe what we're looking at is a robot, probably a shredded robot that can take bricks and lay them out according to a plan, and that therefore replaces someone who's maybe making $45 an hour as a master mason. But I'm not really sure how you would constantly feed a supply of bricks into something like this while still keeping it mobile enough to be an interesting robot.

Ilia Baranov (04:31):

Yeah, I mean the scale of the thing is going to be pretty big.

Stefan Seltz-Axmacher (04:35):

Which is not going to be hard if it's another big heavy thing you need to move from house site to house site.

Ilia Baranov (04:39):

Yeah, yeah. It's a thing you got to unload and take time. But at the same time your pallet of bricks is pretty damn heavy anyway. So if you could make it roughly the same scale as that, then it could just sit and just grab bricks from a pile itself being within the dimensions and weight class of a pile frame.

Stefan Seltz-Axmacher (04:57):

I mean, I think also though problem is, [inaudible 00:04:59] from the friends I have who have done stuff in construction and especially home construction, there's not really a great digital model of even where the- [inaudible 00:05:07]. You have a CAD drawing, an architecture drawing, and then the GC does whatever the GC does and somehow the other side you get a house. So I think the first step is you have to take this kind of shaky drawing, turn it into an actual digital plan of saying bricks should go roughly here, stacked whenever 10 feet worth of bricks and layers is, and then you also have to move back and forth.

(05:35):

I imagine also our rolling robot is going to have trouble as you get beyond the first four feet of bricks. So probably the value add here is you do the really crappy basic brick work the first four feet or so and you have a master mason in air quotes, or maybe just someone who comes and adds the next four feet of bricks.

Ilia Baranov (05:58):

Or you could just have a really tall arm, a really long-

Stefan Seltz-Axmacher (06:01):

Because robotic arms are so easy.

Ilia Baranov (06:03):

Well yeah, absolutely, absolutely. No, the same stuff that they use for pouring cement at long distances, those really long boomers, same kind of idea.

Stefan Seltz-Axmacher (06:13):

[inaudible 00:06:13] exercise on their own and definitely don't have a person the whole day like a stick or a rod's in control.

Ilia Baranov (06:17):

Yeah, exactly. Exactly. Well, so maybe we were thinking of homes not the right way, might be thinking of houses, but this kind of stuff, a lot of the construction [inaudible 00:06:28] has come into apartment buildings. We were feeding the same plan over and over. You know exactly we are putting stuff 'cause everything's been planned out in huge detail.

Stefan Seltz-Axmacher (06:35):

In fact, one of the guys on our team, Navine actually had a Robotics company where he was framing out commercial real estate and essentially the bat was there is kind of a better plan, let's put some metal rods at 90 degree angles so that people be come in and install a sheetrock or whenever afterwards. But the problem we get is still the plans. The problem again is when you have a 25-story building, it's really hard to localize, especially if the broader environment's constantly changing. You start to look a lot like a $800,000 robot to replace someone who makes $150,000.

Ilia Baranov (07:13):

So you know how you solve this. You need VR for everybody on site.

Stefan Seltz-Axmacher (07:16):

I mean, yeah, the metaverse is obviously the solution.

Ilia Baranov (07:21):

That's the solution. Why do you need a house if you could just pretend you own a house?

Stefan Seltz-Axmacher (07:24):

But I think our setting is in home.

Ilia Baranov (07:26):

Yeah.

Stefan Seltz-Axmacher (07:26):

It's not even construction.

Ilia Baranov (07:28):

Home use, maybe we've drifted away from home.

Stefan Seltz-Axmacher (07:30):

Yes. This sounds a lot like a robot that you worked on the Amazon Astro robot.

Ilia Baranov (07:34):

Well, yeah. I mean there's definitely some early 3D printed parts in there.

Stefan Seltz-Axmacher (07:38):

Yeah, well no, it just more mean a robot that's driving around that's not using 3D print tech.

Ilia Baranov (07:42):

Yeah, I mean for sure anything you're going to deploy, any kind of scale isn't really going to use any kind of 3D printed parts. The machine there was really kind of built as an interactive... Some people have described it as Alexa on Wheels in a Right Interactive-

Stefan Seltz-Axmacher (07:58):

[inaudible 00:07:58] have described it as Alexa on Wheels.

Ilia Baranov (08:01):

I mean everybody does it. It's quite funny because at the time actually Jeff Bezos put something on Twitter where his kids had taped an Alexa to a Roomba, and this is halfway through the project. Our team was like, "Yeah, basically this." Five year olds have figured out what we're doing better than we figured out what we're doing.

Stefan Seltz-Axmacher (08:17):

I think the problem of home robots is it's such a solution in search of a problem. If you want someone to grab you a drink from a fridge, you have children. That's why we procreate to say, "Hey, go get me a soda."

Ilia Baranov (08:31):

That is definitely the next skill I'm teaching my two year old. Why teach her anything else? Yeah. The [inaudible 00:08:39] get me a drink from the fridge kind of thing is the ultimate dream of ever to want assist. It's a weird goal.

Stefan Seltz-Axmacher (08:47):

I mean, if we're going to do home use, not 3D printing, it could be exactly what we've talked about in the office before of taking your old clear path jackal and figuring out how to get it to bring us drinks from the fridge downstairs.

Ilia Baranov (09:00):

Exactly. Yep. Yeah, you know what? Instead of that, how about this? How not 3D printing, right?

Stefan Seltz-Axmacher (09:06):

Yep.

Ilia Baranov (09:06):

So 3D printers use a gantry to get around different styles of gantry, but they're moving the hot end around that way. Why not just make the entire interior of your room cable driven gantry so that the robots can... Great. So you just have wires everywhere. You only need three wires. Come on. Not that big of a deal. You walk in, there's just a spider wave of wires in your room and this end effector pin popped to any position in your space really freaking fast. Like wire end effectors can travel 100 miles an hour. So you walk in, its zooop, right to the fridge, whips it open, there you go. If you're going to commit to robots in the home-

Stefan Seltz-Axmacher (09:44):

Your roommates changing a light bulb [inaudible 00:09:46].

Ilia Baranov (09:45):

Yeah. Why would you not do it this way?

Stefan Seltz-Axmacher (09:51):

That's what great about this, can use ML to tag, say the most thousand we interacted with household devices and have constant cameras everywhere looking downward in your house to grab anything that you want from any corner of your house at any time.

Ilia Baranov (10:05):

So you're saying you take Amazon Go, make it less useful and way more intrusive in your house.

Stefan Seltz-Axmacher (10:13):

Yes. But here's a thing that I often do. Now, I take showers on a regular basis as I'm proud to admit.

Ilia Baranov (10:19):

I believe you.

Stefan Seltz-Axmacher (10:20):

Yeah. When I walk to the shower, I'm in my boxers, take my boxers off, I'll take a shower, leave with a towel. Now sometimes I forget my boxers in the bathroom. Now that is unpopular. That is an unpopular thing that I do with the other members of my household, namely my wife. With this Gantry system, we could say, "Go find underwear, grab it, put it in [inaudible 00:10:42]."

Ilia Baranov (10:42):

I have a better system, not a better system, better use case. So speaking of Amazon, the whole idea of the Amazon Go stores is that they use a bunch of camera systems and they figure out what anybody's grabbing. There's no rough id, it's all vision based. You have the same thing to your point, you have cameras and it tracks what you have in your house and not only does it get you the thing, it knows what you need to buy next. And after the initial phase, after you've gone in people's houses, it starts to accidentally knock over fragile so that it can't see the five more of them. You only need to do is 2% of the time and your profit margins are awesome.

Stefan Seltz-Axmacher (11:18):

[inaudible 00:11:16]. There's actually been a bunch of SaaS companies who wanted to solve this tracking your personal inventory type.

Ilia Baranov (11:23):

Oh yeah, absolutely.

Stefan Seltz-Axmacher (11:24):

So like, "Hey, do you have extra AA batteries?" If you have cameras everywhere in your house, you'll always know if you have the extra AA batteries and you can grab them with a simple push of the hat.

Ilia Baranov (11:37):

And the best part of this whole thing is if you just have a glue gun sitting somewhere, this gantry can grab that glue gun and there you have a giant 3D printer.

Stefan Seltz-Axmacher (11:47):

So thinking about this as a business, basically what we're doing is we're making it so that at the touch of your fingertips, so you can never need to get off the couch and you can get anything in your home as well as no, every item in your home. Now how will we price this out? So you have the cables, you have the nest of cables, the roof, the ceiling-

Ilia Baranov (12:07):

High tension steel cables.

Stefan Seltz-Axmacher (12:09):

Yep. There's probably going to be some complicated bracing for stairs.

Ilia Baranov (12:13):

Yeah.

Stefan Seltz-Axmacher (12:13):

And then you need, I mean honestly don't need that good a camera. It's just like a cell phone in every corner of your house. That's like facing downwards, constantly taking your information to the cloud.

Ilia Baranov (12:24):

Exactly.

Stefan Seltz-Axmacher (12:25):

So I mean if the actual gantry crane itself, maybe that's 50 grand, 25, 50 grand.

Ilia Baranov (12:34):

I mean depend, how fast do you want this thing to move? Do you really want it to whip around at top speed?

Stefan Seltz-Axmacher (12:38):

So my apartment is 1400 square feet, so it needs to travel, say 1500 square feet within a minute.

Ilia Baranov (12:46):

That's slow.

Stefan Seltz-Axmacher (12:47):

What's that in miles per hour, freedom units?

Ilia Baranov (12:49):

I mean it will take you less than a minute to walk across your apartment. So to slower than walking speed. So you're talking like a mile an hour maybe.

Stefan Seltz-Axmacher (13:00):

That's not so bad. All right. So maybe actually, this gantry crane can move at two miles an hour. So probably it's only like $5000, $10,000 or 1500 bucks. Well, has to be reliable, it has to pick things up. I think it should probably be able to pick up things that way up to 25 pounds.

Ilia Baranov (13:16):

I think we're moving away from my dream of this thing whizzing around so quickly that it's a health hazard.

Stefan Seltz-Axmacher (13:20):

Okay, well I'm just trying to productize. So it moves to two miles an hour we need in my apartment, let's call it, I don't know, 50 cell phones worth of computing cameras.

Ilia Baranov (13:30):

Yep. That's about right.

Stefan Seltz-Axmacher (13:32):

So I think realistically we need a good name, something that both encompasses, it's always watching you and it can constantly grab you. Maybe Panopto Grab.

Ilia Baranov (13:43):

Big Brother Octopus.

Stefan Seltz-Axmacher (13:44):

Big Brother Octopus, or more Panopto Grab as it's known in some circles. And we think that we can deliver the solution to you in a 1500 square foot house for the low price of $150,000 and you can sign up for our Kickstarter today and be able to grab anything or anyone in your home.

Ilia Baranov (14:04):

Not including the insurance required. So let's move into some personal history time. I've heard you've had some previous experience with the Autonomous vehicles just a little bit.

Stefan Seltz-Axmacher (14:16):

Yep.

Ilia Baranov (14:17):

And of course when you started your previous company, Starsky, you knew exactly the direction to going.

Stefan Seltz-Axmacher (14:23):

Yep.

Ilia Baranov (14:23):

All right, well, let's tell the real story. So how did that start off?

Stefan Seltz-Axmacher (14:26):

Of course. Yeah. So as background, Polymath is my second Robotics company. My first one was company called Starsky Robotics. We at peak were raised $26 million. We're about 100 people. We were a self-driving trucking company. So we did everything from operated trucking business all the way through install sensors on trucks and we completed, what to my knowledge, is the first unmanned trip on a public highway in the world. And if that is surprising to you, it's because in general the narrative around autonomy is, "Oh, it's basically here and blah, blah, blah, blah, blah." When in reality very few people have ever driven a vehicle on a public road without person today.

Ilia Baranov (15:08):

So you started off knowing that autonomy was difficult then?

Stefan Seltz-Axmacher (15:11):

No. No, I did not. Basically, I was a SaaS sales guy who had worked around some startups and have done something more interesting than just being a SaaS sales guy. But I went on a road trip with a Craigslist roommate who had turned our desk in our living room into Robotics lab and we had a five hour drive with nothing much to talk about. So we started talking about what is next Robotics project should be? And I was like, "Oh, hey, what if we made trucks into drones? What if we made trucks remote control? If Northrop Grumman can have guys in Reno fly drones in Afghanistan, how hard could it be to have people in Reno drive trucks in Arkansas?" That seems like it must be easy and it's valuable because of driver shortages.

Ilia Baranov (15:58):

So how hard can it be as one of those famous Robotics lines that just get suddenly people in trouble. On another note though, of roommates building labs in the home, I guess Robotics lab is at least harmful or maybe most harmful version of that, depending your perspective.

Stefan Seltz-Axmacher (16:12):

We destroyed that IKEA then.

Ilia Baranov (16:16):

At least it's just the desk.

Stefan Seltz-Axmacher (16:18):

That's true.

Ilia Baranov (16:18):

Yeah. So how hard could it be? Well, what happened?

Stefan Seltz-Axmacher (16:21):

It turns out quite and not even for the fun reasons. When people hear that initial idea, they think, "Yeah, well sure what about latency and how are you going to solve that?" And that sounds like it should be the gotcha question, but actually just making a thing remote control is pretty hard.

Ilia Baranov (16:42):

In what sense?

Stefan Seltz-Axmacher (16:43):

So I don't know if you know this about self-driving trucks. You need a truck to make one self-driving. You need a self motors on it. You need someone who drives this truck that happens to not be in public road. None of those are things that I had when I served retail this company. When we first had money, we had 150 grand, a truck cost 150 grand. So it was like, are we going to have no money and just buy a truck? That was hard for a while because I had this really crappy 15 year old stick shift car every time we wanted to test even just our mechatronic device, it was like, let's get around, install some actuators on it in a non-destructive way, find a sleazy parking lot that's mostly used by drug dealers and love lost teenagers and try to drive in circles without hitting anything too hard. Every part of it was actually really hard even before the fun stuff to talk about latency and regulation and insurance and whatever.

Ilia Baranov (17:44):

Yeah. And you mentioned before that there's a problem with standardization on these vehicles. How did that show itself in your roadmap?

Stefan Seltz-Axmacher (17:51):

Yeah, so Ilia, do you know for example, the angle that the accelerator is, let's call it 2017 Freightliner Cascadia just between front, the angle of the accelerator is at and the distance of that accelerator from the floor?

Ilia Baranov (18:10):

I do not.

Stefan Seltz-Axmacher (18:11):

Yep, neither the Freightliner.

Ilia Baranov (18:15):

I feel like they should.

Stefan Seltz-Axmacher (18:16):

I would think, but in reality they make something like 10 million Prius' a year, which works out to something like a Prius is built every five to 15 seconds. They make 50,000 Freightliner Cascadias a year, which is a lot less frequent than once every 15 seconds and as a result, it's just like this Lego block jumbotron of crap just jammed together. And essentially trucks are made at such lower volumes and as a result they're all way more different than cars are.

(18:45):

And in the early days that was troubling with things where exactly do you put the actuator for the accelerator? And in later days it was hard because we would literally have a subject trucking company that we could take trucks out of and we couldn't guarantee that two trucks we took out of it that we leased for the purpose of this would actually be very similar for tuning, say our longitudinal controller, tuning so that when we push the accelerator a little bit so that we could go from 62 to 65 miles an hour, it wouldn't instead put us to 70 miles an hour, which might happen if it wasn't tuned through the right way.

Ilia Baranov (19:22):

Well, but Stefan, I'm an engineering genius. I'm just going to machine learn this problem.

Stefan Seltz-Axmacher (19:26):

Yeah, I mean that would be great. I still don't really understand why you can't. In terms of controls, one would think machine learning could just be this magical solution that could change it all and I've definitely argued at two autonomy companies that I work at, one of which I'm currently at and I've been told [inaudible 00:19:46] again that no, you should just hire controls engineers. And do you know what controls engineering is listener? My guess is unless you have a Robotics degree, you have no idea what I'm talking about. This is a made up discipline. Controls engineering that's not real. There's mechanical, there's software, there's civil and sure, maybe there's some stuff with electronics. Controls, not a real discipline.

Ilia Baranov (20:10):

I mean basically Stefan has cracked the secret of controls engineering where actually controls engineers just connect remotely and just twiddle the wheels manually and that's how all the controls are over. We just ship a little man in a box with every system [inaudible 00:20:26].

Stefan Seltz-Axmacher (20:27):

Ilia, maybe you can explain to the listener what controls engineering is and what it actually, why is it so hard?

Ilia Baranov (20:33):

Well, I'm going to steal one of your talking points because I think it's one of the better ones I've heard. So you get into a shower and you turn on the hot water and it gets too hot so you make it colder, it gets too cold, you make it hotter, and you're basically trying to control the temperature. So you do this kind of wiggle back and forth. Controls engineering is that, but doing that via software and doing it as efficiently as possible.

(20:53):

The ideal case is your system, your plant so well, you have such a good model that you can predict that if you just set the knob in your shower to this value, you're going to get exactly the right temperature, which I keep hoping I'm going to do at home, but I never speak... But it's basically that it's taking an input, having a model of your plant or your system and controlling the output of it. So it does what you want. For robots, it's steering, it's throttle. Those are the main things you're trying to control. But there's a bunch of secondary things. How quickly do you want to accelerate? How quickly do you want your acceleration to change hospital?

Stefan Seltz-Axmacher (21:29):

Oh, why is that so important how good you accelerate?

Ilia Baranov (21:29):

Well, slamming on the brakes is basically deceleration and doing that randomly without controllers would be pretty bad.

Stefan Seltz-Axmacher (21:37):

It's a uncomfortable for rider.

Ilia Baranov (21:40):

Yeah.

Stefan Seltz-Axmacher (21:40):

Yep.

Ilia Baranov (21:41):

Yeah. It's the change of velocity is called acceleration. The change of acceleration's called jerk and didn't it feel like a jerk when you're sitting in one of these and would make slams on the brake with no warning.

Stefan Seltz-Axmacher (21:53):

I would also call out for the aspiring autonomous vehicle founder, controls engineers is somewhat hunting for [inaudible 00:22:02] or I don't know, Florida Jaguars or Albatrosses or Dodos or some animal that is so rarely seen that we could argue as mythological.

(22:11):

Basically nobody studies to be a controls engineer anymore. 90% of the controls engineers I've ever met were from some country where they thought, "Oh cool. I'll have a great job if I learn how to run the power plant." And controls engineering is a great skillset for running a power plant because hey, if you start popping up too much power, you have a meltdown. If you have too little people can't watch Netflix and both of those are problematic.

Ilia Baranov (22:34):

Different kind of meltdown.

Stefan Seltz-Axmacher (22:36):

Very different kind of meltdown, but nonetheless severe. So as a early Robotics and autonomy founder, I found myself needing to hire for this role in this skillset that frankly I didn't really understand and couldn't really find anyone who did because before the autonomy boom of 2016, 2017, there was like 15 controls engineers in the Bay Area. Maybe that number's 150, but I think I once had a recruiter show me a report in 2020 that said there was in the realm of 1500 controls engineers in a market that literally has 50,000 software engineers.

Ilia Baranov (23:13):

Yeah, that sounds about right. And a bigger scope what makes Robotics really, really complex is that you need those specialists in controls and you also need the specialists in machine vision and you also need the specialists in kind of software quality and deployment and those things to actually make product.

Stefan Seltz-Axmacher (23:28):

But can they just use machine learning for the controls?

Ilia Baranov (23:30):

Yeah, I will see it one day.

Stefan Seltz-Axmacher (23:32):

So why isn't that true?

Ilia Baranov (23:35):

Using machine learning for controls is kind of, it's using an atomic weapon to kill a fly-

Stefan Seltz-Axmacher (23:42):

But you kill the fly.

Ilia Baranov (23:43):

Yeah. You also take out several city blocks or [inaudible 00:23:47] to yourself.

Stefan Seltz-Axmacher (23:47):

I didn't like anyone who lived in that city.

Ilia Baranov (23:49):

Yeah. It's overkill in the extreme in that you're running a very black box system to what hopefully by definition, maybe one of the things about controls is you really want predictability. You want to be able to predict across your range of inputs, here's how the plant is going to work, here's your output's going to work. Especially for mission critical, safety critical stuff our plants healthcare system-

Stefan Seltz-Axmacher (24:15):

Steering on the highway and-

Ilia Baranov (24:16):

Steering on the highway. Exactly. And a lot of what controls will do is they'll do this kind of linear analysis where they'll say, "Given my entire range of inputs, what did my output look like?" And they'll test the extremes. Extreme minimum, extreme maximum, the middle. In machine learning, it's very difficult and there doesn't seem to be a good understanding right now of proving that given a set of inputs, your machine learning algorithm will do the same output every time.

Stefan Seltz-Axmacher (24:41):

And couldn't you just say have a unit test that said your steering angle output is now 45 degrees and we're driving on the highway in fifth at 65 miles an hour. Don't do that.

Ilia Baranov (24:49):

Absolutely. And you will prove that for that set of inputs, your outputs are predictable. But what about if your outputs are slightly different? The problem with machine learning is that there's no good understanding of what the neural network is doing at any given time. And so it's very hard to fully prove across all possible sets.

Stefan Seltz-Axmacher (25:09):

Ilia, how big is a neural network? So that those of us who didn't study Robotics might have an understanding of what's going on.

Ilia Baranov (25:16):

How big is a neural network?

Stefan Seltz-Axmacher (25:17):

Yeah, can you just draw it out on a whiteboard?

Ilia Baranov (25:19):

Oh yeah. Status. It scales with the complexity of your inputs. So the most common ones for image recognition and those kind of things, you're talking about few 1000 input nodes, a few 10s of 1000s of hidden layer nodes and a few dozen output nodes that kind of size. There's more modern techniques that kind of abstract away from particular node sizes, but you're still talking about 10s of 1000s of variables for anything that's not completely straightforward.

(25:51):

And so maybe you could scale down your control problem to the point where your machine learning system is trying to predict one or two variables, but at that point a classical controller is going to perform as well, if not better. You haven't done anything. Speed control via machine learning seems to be overkill whereas machine learning for control, a very complex system seems very risky because you don't know what's going to happen. So I don't really like at below end or at the high end, I don't see a good reason for this approach. I'll probably be, this'll one of those moments where like 64k is good enough for everyone's forever. I'll probably be proven wrong, hopefully probably be proven wrong in the next few decades, but I haven't seen a good approach for it yet.

Stefan Seltz-Axmacher (26:34):

So in terms of other things that are hard, so your Robotics company. So a thing that was persistently hard early on is, I don't know if you know I many truck drivers or trucking executives.

Ilia Baranov (26:47):

Unfortunately not.

Stefan Seltz-Axmacher (26:48):

Oh, that's bummer. They turn out many people to not have a great understanding of what is easy or hard in Robotics. So I found myself off in this place where I talked to the trucking company, they're like, "Yep, this is fantastic. Give us all the automated trucks you can make." Fantastic. Do you guys know how to use them?" No, not really. Okay, cool. Do you guys, for example, have all of your information about where you should drive on a road by road basis? Does that exist somewhere? No, not really. When we rate things, we say drive from Atlanta to New York. Okay, well that I'm just putting things into Google Maps and that's pretty hard.

(27:27):

And then they say things like, "Well yeah, when you do this route you'll pick up this load in San Antonio and you'll then hop on the highway, drive for a while, get to Dallas and then this last part's really fun and you're going to take this exit that isn't really an exit, it's not really marked. It's about three quarters of the way between this exit and this exit and you're going to follow that road for about a mile turn left at the church drive on a dirt road up a hill and that road sometimes it washes out. So it moves around a little bit. And then once you get up there, you're going to go into this dusty area where you have to pull up and then open up a lever on the bottom of your trailer. You guys can automate opening of that lever, right?" And we'd be asked stuff like that that I would then go to the engineering team and be like, "Hey guys, can we do this?" How hard Ilia could that possibly be?

Ilia Baranov (28:14):

No, not [inaudible 00:28:15]. I mean to solve the trucking problem at that point you have to invent artificial general intelligence. In the hierarchy of needs of artificial general intelligence, driving a truck is low down that scale.

Stefan Seltz-Axmacher (28:32):

Even things like for example, that opening up, that moving that lever on the bottom of the trailer, you jump out the... There was another version of that question I'd be asked actually had someone on my team who was like, "Yeah, we can swap out trailers all the time, just have the engineers throw together a little arm that unplugs the pigtails, unplugs the air brake lines and we're good to go." Why is that so hard?

Ilia Baranov (28:56):

Humans are incredibly dextrous.

Stefan Seltz-Axmacher (28:59):

I thought we were lazy and worthless.

Ilia Baranov (29:01):

We don't admire this enough about ourselves, but our human hands and our perception of body are second to none in the animal kingdom much less then a machine kingdom. Even things like picking up a ball and manipulating it with your fingertips or throwing it up and down and catching it is cutting edge, if not impossible in the Robotics world and trying to do arbitrary manipulation. So not only do you have to move this handle on this truck model, but it could be any handle on any truck model. It could be painted different colors, it could be stained with oil, it could be jammed at a certain point because it seems to be a little bit sticky.

Stefan Seltz-Axmacher (29:34):

But I can figure that out.

Ilia Baranov (29:36):

Yeah, humans are great at it. You have this great processing node sitting on your shoulder and you have really, really accurate manipulation. If you look, there's products out there, there's a shadow hand that tries to reproduce human manipulation costs. I mean, I'm going to lie about the exact number, but multiple 10s of 1000s of dollars and it's a fraction of as good as a human hand.

(30:01):

So you're talking to get something even close, you're talking about 100s of 1000s of dollars, at which point you've blown your cost budget out of the water for a guy who has to come over and flip a lever. So that's one option. The other option, of course-

Stefan Seltz-Axmacher (30:14):

What I ended up finding and what our product decisions were at Starsky was, all right, so these trucking companies, they don't really know how to use robots because frankly, a self-driving truck is a more complicated product than an airplane and trucking companies on average, not always, but on average, are less technologically sophisticated than airlines. So what we ended up deciding to do was essentially we be the trucking company so that the need of the Robotics could be less. So we could just say, "These are always going to be our trailers. Hopefully we don't have to add anything to the trailers, but we can, if we really need to. We're only going to drive to these special whitelisted routes that we already know we're really, really good."

(30:59):

And the problem is that business really sucks for investors because not only do we have all the capital requirements of a trucking company, we have the margin of maybe good trucking company and then it actually almost sped up the problem of what is safety and then how do you make this problem safe?

Ilia Baranov (31:18):

So as an investor though, why do I care about those things? Why would I care about that?

Stefan Seltz-Axmacher (31:22):

So I used to really not understand it and at the end of Starsky I really dug into, "Hey, I could make, with a total of a $100 or $200 million investment, we could be $100 billion a year company that makes a 50% margin. And that's so much better than needing 10 times the money and making $1 billion a year within 90% margin, I thought. Another thought that I had that scared the bejesus out of me is that if I ever mentally broke, if the stress of dealing with angry truck drivers and weird state legislators and engineers who got their feelings hurt and engineers who said stuff that offended those other three groups, if I ever broke it felt like the company would shut down.

(32:06):

And I realized that's not how SaaS works and that's why investors really like SaaS because basically half the people in San Francisco can run a billion dollar SaaS company. It's a very easy skillset. You show up, you talk to some people sometimes, they think you're not a jerk and you can say some stuff about how x, y, z widget for B2B SaaS is going to change the world and democratize electronic signatures. It's a way easier thing than, "Hey, I'm going to single-handedly operate a trucking business that employs 10s of 1000s of people while building cutting edge technology. No one else in the world get news or cares about.

Ilia Baranov (32:45):

Yeah, I mean that sounds like a harder problem.

Stefan Seltz-Axmacher (32:47):

Yeah, it's definitely.

Ilia Baranov (32:48):

Yeah. And especially for engineers who get into engineering for the people skills. That's the main reason we go into that. So tell me about safety in Starsky.

Stefan Seltz-Axmacher (32:58):

So an interesting thing at Starsky was essentially we could plausibly take the person out of the vehicle very quickly. The first time that we did something that looked like that was, I want to say September or October of 2017, we drove 60 miles with a loaded trailer where the human who was sitting behind the wheel didn't need to do anything from beginning to end of the trip.

Ilia Baranov (33:23):

Nice.

Stefan Seltz-Axmacher (33:23):

If that person had not been in the truck, we would not have killed anybody. We would've driven. That could have been a driverless truck. It is then kind of terrifying to realize there's no clear playbook for taking the person out and that was weird.

Ilia Baranov (33:43):

What about ISO 26262 there [inaudible 00:33:46]?

Stefan Seltz-Axmacher (33:46):

It's hell of a standard, but you're on lower level than that. When you're in school, you get a rubric at the beginning of the semester that says, "Here's what the tests are going to be. They're worth this much towards your grade. Here's what the homework is going to be, it's going to be worth this much towards your grade. Here's how you get a good grade in the class." And then I would properly ignore all that. And you would think when it came to taking a human out of an 80,000 pound vehicle, driving 55 miles an hour two feet away from your mother, you would think that'd be a similar rubric and you would think the smart people would have a rubric. But it basically turned out that that didn't exist. And I'd like go and I'd talk to safety people and I realized that most safety people are working off of a rubric made for the particular industry that they work in, be it automotive or aviation and that if you can't fit that spreadsheet, they're like, "I don't know, might not be safe or I don't know, outside my paying grade."

(34:46):

Basically we were operating where there wasn't a rubric, where there wasn't a scorecard outside of like, don't kill anyone please and that's kind of terrifying. That's like can I intellectually terrify. Realizing not only you're the adult in the room, but maybe you're the adult in the warp.

Ilia Baranov (35:03):

Kind of makes me think of the original one, cars got on the road. There was no seatbelt laws, no even driver's license early on. There's barely a windshield. You felt like you could drive and you would start to drive and that was it.

Stefan Seltz-Axmacher (35:16):

As the world was great then.

Ilia Baranov (35:17):

Yeah. Plenty of safety on their system.

Stefan Seltz-Axmacher (35:20):

It actually reminds me more of my favorite story, which is how the aviation industry got regulated, which is the Wright Brothers pretended to be the first airplane in 1906, the Red Baron shot downs of people in World War I. And then suddenly people were doing air mail in whatever the '20s, but all these bozos would go up in the air and play ping pong on the wings of an airplane at 5000 feet and do stunts and jump from one plane to another just to show how cool they were.

(35:48):

And what ended up happening is the nascent serious aviation industry, literally Hunter Douglas or Boeing, whatever his first name was like Jeff Boeing, Jim Boeing was lobbied together to create the FAN to stop bozos from doing the equivalent of falling asleep behind the wheel of their Tesla. This is how aviation regulation happened because they realized that if this was an unsaved industry, because that idiotic barnstormers, they would never become billionaires and that didn't fly with them.

Ilia Baranov (36:19):

So my comment on that is if you're going to send your Wright brothers hate mail, please send it to Stefan, right? I want no part of this.

Stefan Seltz-Axmacher (36:29):

I think I just started a fight with the whole state of North Carolina.

Ilia Baranov (36:32):

I am completely neutral in that state.

Stefan Seltz-Axmacher (36:36):

As a Yankee, it wouldn't be the first time. So essentially we kind of got scared by this and I realized talking to other people in industry that others were similarly scared people. There was pole teams that would be able to hand wave away unmanned deployment by saying, "Oh, the machine learning and the [inaudible 00:36:58], and probably it's a state senator's fault." In our case, you couldn't say any of that and you just had to be like, all right, here's how we're going to make this drive from the [inaudible 00:37:08]. And the problem with that outside the fact that there's no real agreed upon standard, there's no real agreed upon rubric is like there's this combination of what is morally acceptable to deploy? What is engineering smart to deploy? How do you manage the PR of deploying something? And those all kind of sound like safety, but are all kind of different. So what we ended up deciding is like, all right, we want to be really safe. We want to hold ourselves to a standard that maybe no one else could even understand.

(37:39):

On the engineering side, the most nasty and terrible thing about safety is that if you're serious about safety, you're going to hurt some numerator over some denominator of people and that numerator is greater than zero. But if you think your system can never hurt anybody, you're not doing engineering, you're living in a fantasy fairyland, and that's cute and you can go back to kindergarten. But if you actually want to deploy a safety critical system, you have to say, "All right, one time out of some number of times we're going to kill someone's child or mother or father or uncle or best friend or soulmate," one time out of some number.

Ilia Baranov (38:20):

Hopefully billions.

Stefan Seltz-Axmacher (38:21):

Hopefully billions, but maybe it's 10s of 1000s. Maybe it's 1000s in certain applications. So what we decided from an engineering perspective, because it's the most measured number, is that our goal would be greater than whatever the human number is because a lot of people study what the human number is. There's a lot of data, there's a lot of justification of the human number. Is this the human number or is that? Cool we have a data set, we'll say greater then. If the human number is one out of a million miles, we'll say one out of a million one miles or greater because we can incredibly measure around it. And then from a PR perspective, I learned that that's a whole other thing altogether. Fun lies that people would love to say is like, "Oh, well the first time one of these that hurt somebody, the whole industry's going to shut down. Your company is going to fail," which isn't really true. Like safety PR ends up being its whole own weird thing that somehow actually incentivizes hurting more people than less.

(39:15):

I had a conversation with, maybe I shouldn't say who exactly, but some people who ran safety for a very prominent app that's on your phone and the first time people on that app accidentally killed somebody in accident, it was a whole big deal. The CEO flew out, they paid for the funeral, they gave the family a couple $100,000 dollars is pumping juice for three months. The second time it was one month worth of news. The third time it was a week worth of news and now that company in accidents that are justifiable kills somebody every single day and you don't notice it and you still use that app. In reality that actually think about safety PR is the best number of people to hurt is zero, the second best number is greener than 350 and that's terrifying. That completely conflicts with that moral and ethical idea of let's build a thing that is safe and makes the world better as opposed to worse, which greater than 350 is worse.

Ilia Baranov (40:08):

But Stefan I have low disengagement per mile on my autonomous vehicle.

Stefan Seltz-Axmacher (40:13):

Oh my God! So I lived through this and this was the worst. For a while working on Starsky, the premier metric of measuring an autonomous vehicle was disengagement per mile. Basically, if you are driving your self-driving car from Best Buy to Target, how many times in that trip does the human safety driver need a turn off the autonomous system to make sure that doesn't get into an accident and you, the listener who has not read 18 blog post about the subject could be forgiven for thinking that makes sense. That makes a lot of sense because obviously the lower disengagements per mile, the better the thing.

(40:49):

Now it's really, really hard to measure other people's robots. Really, I mean like I-

Ilia Baranov (40:55):

It's hard to measure your own robots.

Stefan Seltz-Axmacher (40:56):

Yes. But other people's robots, especially if you were say a poli-sci major who now works in the state government of California, it's really hard to measure somebody else's robots, especially when the industry doesn't have standards about safety, let alone what is a disengagement per mile because it doesn't count as a mile if the system's fully on, but goes into shadow mode if there's a failure. Is a mile on the highway the same as a mile in the city?

(41:22):

All these things take a number into a number with a couple of paragraphs attached, which doesn't work. But my favorite of all of this was that at Starsky summer of, I want to say 2017, we had a disengagements per mile up on the dashboard and it started going up and it went from 1215 to 75 miles per disengagement inside of a month. Like, holy shit guys, we figured out the machine learning, it had solved. We had built an autonomous truck, we're going to be billionaires next week. This is awesome. So my co-founder flew out, spent a week in the truck with a test driver and realized why the miles were going up so much because essentially our test driver knew that we cared about this metric, knew that we were telling her great job, the test you just did, I don't know if we know this, there's only two disengagement in your 150 mile trip.

(42:25):

And what she was doing was basically, as long as there was no one around the truck that she was driving, she would never turn that system off. So you just veered through a lane. Let's keep it rolling boys. Oh, there's someone coming up. Maybe I'll just like nudge the steering wheel to the right because there's no torque disengagement sensor so that we don't bump into them if something bad happens and boom we have great disengagement for mile. I'd heard of top tier robotaxi companies do stuff like if the vehicle is driving and the safety driver gets uncomfortable, they push a button where the vehicle keeps on driving in shadow mode, the safety driver does all the driving, then it's no longer in a sensitive position. They turn off shadow mode and they get credit for all of shadow mode when the vehicle didn't get to an accident as against disengagement per mile, which like sure, but it's just a bad metric.

(43:20):

A better metric frankly, is trips in a row without a person needed. And that's essentially what we used at Starsky for when we took the person out of the vehicle. We did something like 150 trips in a row without needing a disengagement from beginning to end and that gave us a reasonable amount of statistical certainty that on the 151st or the 152nd trip, we could go from end to end without needing a person and we'd be safe.

Ilia Baranov (43:47):

Yeah. It's interesting because it seems like the industry has centered on this disengagement for Mile and Google and those folks we'll talk about that quite a bit. Do you think that's kind of more of a policy problem or is it more of a PR problem or how did we get here?

Stefan Seltz-Axmacher (44:02):

Yeah, I mean I think the problem is essentially that it's a nice simple number that can mask a lot of complicated stuff.

Ilia Baranov (44:10):

People love simple numbers.

Stefan Seltz-Axmacher (44:11):

Yeah. When you think about what a metric really means is it's a number to capture a broader sense of how things are going without needing much more context. So disengagement per mile ends up looking like a really great metric and people have locked into, and frankly, we all thought we'd have self-driving cars by now. So as a result, the legislators, the regulators, they locked in a bunch of things that they thought would be relevant in 2015 into state law and that's not because they're dumb, that's because they believed the hype that everyone else will leave. They saw 15 articles a week coming out of reputable newspapers saying, "Look how close autonomous vehicles are." They thought, "Hey, I guess autonomous vehicles are close. Maybe we should address this issue. It's probably going to matter."

Ilia Baranov (45:02):

We should install a new segment next to your about bets on how far away level five autonomy is for road.

Stefan Seltz-Axmacher (45:11):

There's a guy named Steven Slotover, Dr. Steven Slotover, I hope that's his name, who's been a researcher in [inaudible 00:45:21] autonomous vehicles for 40 or 50 years. And I had a coffee meeting with him once a couple years ago, and at the end of it I said like, "Hey man, this must be really cool for you. You've been working on this field your entire career. Must be pretty exciting to see it finally happened." And he scoffed me off in a very grumpy old man type of way and basically said, "The thing I learned about this industry is the more you learn about it, the further off you think it is," and given that we're saying this in a week where a billion dollar company just wound down, I think that's absolutely the case, which is why now we're not working on on-road autonomous vehicles.

Ilia Baranov (45:59):

Yeah. Argo AI shut down and the dream of on-road autonomy seems to be getting a little bit further. What are we working on, Stefan?

Stefan Seltz-Axmacher (46:07):

Yep. So essentially what we're working on, a weird thing that happened while working on Starsky is we would constantly have big weird companies reach out to us and say, "Hey, you drive a truck on a highway, could you automate our turf cutter that we sell to sod farms all across the world?" And we'd say, "Not really, because in Robotics you have to build every single part of the stack. You're not just building the autonomy, you're also building the hardware layer. You're also building the business logic that tells that autonomy what to do and in our case at Starsky, you're also the trucking company." So pivoting all of that to be sod cutting robots or yard trucks or hauling shale whatever, would actually be essentially a whole pivot. What we're doing at Polymath is we're building a generalized autonomy layer for industrial vehicles that a technical team can apply to any vehicle large enough that it should come to a stop when it gets into trouble.

(47:09):

And the reason for that is that kind of essentially mobility unlocks Robotics in much the same way as operating systems unlocks computers. Before there were common operating systems like Windows or even early macOS to make anything on a computer, you needed an army of IBM people to design some awful data centers thing with punch cards, build all the operating system for it and build some awful primitive application to multiply numbers so then you can figure out where boats were and stupid stuff like that nobody cares about. With operating systems, you made it so that people could just build the program and that program could fit to tell you where a boat was, or it could tell you how much money you lost this quarter with your boat navigation software company and it could go on to do stuff like give you internet access and then the whole world that we know have happened. In the same way, when people try to automate industrial vehicles, all of them spend the first two to four years of their life getting able to drive from point A to point B without hitting anything.

(48:16):

We're building that so that the teams who care about specific applications can focus on the specific apps for those applications, whether that app is till this field, whether that app is push this dirt in this quarry, whether that app is pull this trailer around my farm. And that's why listener, you should immediately go to Polymath Robotics and sign up to build all of your stuff on top of it so that whether you're as stupid as me or as smart as Ilia when you build your thing, it can actually do something neat sometimes soon.

Ilia Baranov (48:50):

That's straight out of the 1920s radio book.

Stefan Seltz-Axmacher (48:54):

So Ilia, thank you so much for interviewing me today.

Ilia Baranov (48:56):

Inquisition mode.

Stefan Seltz-Axmacher (48:57):

For the next episode, the turntables will turn as some say, and I'll be peppering Ilia with questions about robots that he's built in the past and hopefully sounds smarter than I might've sounded today. But in general, this is Automate It. We're going to just talk about building robots. I've talked not necessarily about which algorithm where, but what are the weird, hard product problems that you find? What are the weird problems when you have to pitch the thing that you can actually build versus the thing that investors care about? What are the weird problems that you think will be really, really easy? How hard can it possibly be and end up being the part of the product you fight against for the next five years?

Ilia Baranov (49:36):

The answer of Robotics is always, it's harder than you think.

Stefan Seltz-Axmacher (49:38):

Always.

Ilia Baranov (49:40):

In every situation.

Stefan Seltz-Axmacher (49:41):

Regardless how many PhDs or how many of your team went to [inaudible 00:49:45]. It's always going to be harder.

Ilia Baranov (49:47):

Or Waterloo.

Stefan Seltz-Axmacher (49:47):

Yeah, you could find us at polymathrobotics.com. I'm sure we'll have some subpage that says our podcast, but for now, you'll just have to search a little bit harder, which I believe in your ability to do. So next week I'll be interviewing Ilia about his experiences at companies like ClearPath and Amazon's Astro program. In the future, we'll have some friends on who are hopefully just as a cross and open as we are, and if not, we'll kick them off and no matter what, we'll certainly be running through more of these cards because the next one that I just drew is a walking robot. Ilia, what's the setting that you... Walking robots for delivery is what we have next. So will it be Spot? Will it be [inaudible 00:50:28]? Will it be Ameca Warrior? We'll find out soon. Please join us again. Feel free to subscribe. I promise I won't send you too many crappy emails.

Ilia Baranov (50:37):

And we look forward to your Wright Brothers hate mail. Talk to you soon.

‍