#360: Doordash w/VP of Autonomy Ashu Rege

Well, I have to ask this, if everybody who deploys any delivery robot sidewalk or bike lane,

how many issues have you had with vandalism? And what was the craziest one? Because when I see

them now, I'm like, oh, that's cool. But if I were 12 years old, and I saw that,

like I'd be after it with a flamethrower and like, you wouldn't be hugging it and like putting a

cool sticker on it or something. That was a bad kid. Hello and welcome to the Atana cast.

I'm Kirsten Korosek, transportation editor at TechCrunch. And I'm Ed Niedermeyer. I'm the program

director for the ride AI summit happening April 15th at SFJAS Center. I'm also the author of

Ludacris, the Unvarnished Strike Tesla Motors, and Elon Take the Wheel, my new upcoming book

from Ben Bella Books that will be available on December 1st. Available for pre-order now.

I'm also now formally taking over Alex's job as the person with the most titles

and stuff to talk about in the intro. And I'm Alex Roy, the co-founder and general partner at

New Industry Venture Capital, founder of the Human Driving Association, and fan of excellent

interior decor, which is why I'm thrilled to introduce today's guest, who is the exception

to the rule that there's an inverse correlation between success in Silicon Valley and taste.

This man has exceptional tastes. I can see from his background, which I thought was fake,

but it's real. The VP of autonomy from DoorDash, Ashu Reggae. Welcome, Ashu.

Thank you so much. Thanks for having me. I've been pretty excited about this.

Alex, you're going to force us to go to the video podcast format like everyone else.

Well, I like leaving it up to the audience's imagination. They don't need to know exactly

what everything looks like, but I will share that once again, Alex, because he cares so much about

audio is in his Tesla and recording this. So thanks again. I'm sorry, guys. I was commuting,

and I'm sorry. I love my Tesla, and I'm going to give after this episode is over about how

much I love my car. Let's move on to our guests. Okay. Well, let's start with, we want to start,

we want to talk about autonomy in robots, but I'm just curious, what do you drive?

I can drive a Tesla. Which one? I got a Model 3 for competitive reasons at the time. I was just

curious to see what was going on. It was also the pandemic. So I was kind of, you know,

sits stuck at home. So I was like, you know what, apparently they'll just deliver the Tesla straight

to your door. How cool is that? So I said, you know, it's one of those, and I must say they make

it very easy for you to just go and impulse buy. I did that. And yeah, I've been driving a Model 3.

What year? What year is yours? That was 2021. And based on your interior decor,

do you have the white interior? I now, I wish I had at the time. So there's also availability

issues back then, if you don't recall, the pandemic. So they didn't, you couldn't truly

customize, you could customize, but it really like it'll take another three months. And I was like,

no, I want this here and now, you know, what's the point of it impulse buy if it doesn't show up

in a few short days. Yeah. And we should note that when at the time you weren't at DoorDash,

you were at Zooks. I was at Zooks. That's correct. And so, yeah, it's a space gray exterior and

very nice. The darker interior. Yes. I am very jealous of your white interior, which I think

looks. Well, actually, I have the beige. The cream. Yes. Yeah. The cream is the way to go.

Dark gray exterior. I quite like it. And this is a drive at crazy. Do you have Hardware 3 or

Hardware 4? Hardware 2, I believe that's what it is. So it is inside FSD 12. And I don't think

they're gonna... The good news is it's all obsolete now, right? He's already talking about AI6. So

it doesn't matter what you have. And yet, would you agree that even over to FSD 12,

and yet it is still better as a driver assistance system than 99% of anyone else's ADAS? Would you

agree? I wouldn't use everybody else's ADAS. But yeah, I do believe it's pretty good. I use it. I've

done long trips with it. And it's quite a convenience when it's like dark at night and

maybe you don't have quite the same energy anymore. Just having something on a freeway

especially is really, really excellent, I would say. Well, we've completed the marketing section.

Alex, Roy, Tesla. I love it. I do. Let's talk about DoorDash. Let's talk about DoorDash.

But actually, what I'm really interested in is just backing up when you're at Zooks.

What made you interested in going from Zooks, which is the mission is different than DoorDash?

What made you go to DoorDash? Yeah, that's a very good question. DoorDash,

and we started talking about the possibility back in 21, about the same time. Tony, our CEO,

and Stan, our co-founder, invited me to dinner actually because I chatted with them.

My bigger meta question was, why are you guys entering this space at this point?

There are so many companies already out there. And 21, to be fair, there are quite a few entities,

some of whom are no longer with us. And why are you doing this at all in the first place?

And why not just partner? So I was probably kind of obnoxious that I'm doing dinner.

We're asking these pointed questions about the future of what they saw. And they had a viewpoint

which actually convinced me, which is the multiple axes here. So sorry if I meander a bit,

but the first one was the business case itself, which we can get to in a moment.

But my meta question or the first question I had was, why are you doing it? Why not partner with

all the entities that are out there already? And they said we are, and we are looking at it very

closely. And they had actually, they even had a pilot with Cruz, which is public knowledge during

the pandemic. And so their observation was, and this is a sort of very key phrase within DoorDash,

three words, a product market fit. They're obsessive about that to an amazing degree. And

it's really awesome to see the product. And I was a huge power user of DoorDash. So just

to FYI, even before the pandemic, I would order ahead and use it in every which way you could

possibly because I hate shopping or going out and getting stuff. So I ended up with the point

out that nobody was building the autonomy stack. And the solution they were building

were not at all suited to delivery in the way they thought of delivery. And DoorDash has completed

over 10 billion orders historically. And so they do know delivery inside and out. And I'm speaking

of them as they, but I've been there now for five years. So I should probably use we. But at the

time, their point was, at the space at the time, if you recall, was mostly robotaxi. So you have

this sort of autonomous solutions for robotaxis, quite a few of those. And then you had delivery

was mostly sidewalk robots back then. There were a few notable exceptions. But generally speaking,

their observation was, do we really need a two ton car to drop off, you know, a two ton burrito

is a catchphrase. Or in the sidewalk robots are great in a dense urban environment. The problem,

of course, is they're limited on speed. So the distances they can go, especially for, you know,

prepared food is pretty limited. So the question was, does it exist a Goldilocks solution? And

that's where we ended up. And that's what they convinced me was the right approach. The other

factor that played large in my mind at the time was, is the autonomy problem on the robotaxi

front, you know, while exciting and clearly awesome to see what Beemo and others have done now,

is that the right way to go about building a commercially viable product that actually scales?

And in my mind at the time, it felt like the delivery problem had challenges, no question,

and it's a completely different animal in many ways. But the autonomy piece, which in 21, you

could make the case, Tony wasn't solved. And but it felt far more possible to do it incrementally.

With a two ton vehicle going on, you know, even at, you know, suburban speeds, but let alone highway

speeds, that's a non trivial safety case. And so you have a big step function to get to before you

can even consider that, oh, now we can use scale. And then there's another mountain to climb after

that, which is scale and then make it commercially viable. So maybe three mountains, really.

Whereas delivery felt like you don't have to solve everything, especially when you're

within a platform that is already doing, for example, today, we do over eight million deliveries in

the US alone every day. So you don't have to go solve every delivery. In fact, even if you solve

one eighth, you are going to get, you know, you're going to have to build tens of thousands of robots

to do that. And so, so that's what's to me. Yeah, interrupt me, please. Sorry, I told you.

No, no, I just want to, I want to feel, I want to ask if I'm crystallizing like kind of what

you're saying here, which is that we've talked with a lot of people over the years about, you know,

creating something that works in autonomy as a lot of is about, you know, just defining, right,

the the use case and really the operating domain, like often the main way we think and talk about

this, it feels like what you're saying is, is, and this is something we've also talked to those

about, like, speed and mass are what creates risk, right? And that by defining the use case

away from speed and the speed and mass of the personal mobility car,

it's much better than me. You should do a podcast.

I just want to, I just want to make sure that that like, you know, for listeners that this is,

this is how we boil that down. Exactly. It's really half mass times velocity square. That's

your, you know, kinetic energy. And can you, can you be incremental, but also use case, right?

So you can literally, you can do the math because sidewalk robots kind of is a pruing

point. You can go, you know, go along at sidewalk speeds, you can still do deliveries.

You can now say, okay, what if you went a little higher, maybe you got seven miles an hour?

Yeah, that expands your, your, your scope even more. And so as long as you have autonomy as a

fundamental part of your, of your solution, then you can sort of adjust to whatever you need to

and then say, okay, I can only do these subset of deliveries, which is fine. You can make a living

and then incrementally build on that. So yeah, mass and velocity on the one end,

but also the product itself was the use case was different, right? You can't have a robot taxi go

along at seven miles an hour. Nobody's going to take it maybe for a joy ride somewhere,

but that's about it. You cannot have a robot taxi that is careful in certain situations. And so,

you know, gives you a less comfortable experience, right? I mean, we all have friends who do the

brake tap driving, that's bad enough, right? But when you're in a fully autonomous vehicle,

that's that feeling, if it feels tentative to you, you are not going to feel confident of being

in that. So to me, that really create a complete separation of the two categories. And frankly,

I, the other thing, which is my personal observation, but also I think is, I think is

true, which is ride hailing, like how many rides would you take in a day maximum? Okay,

maybe you have a very busy day and maybe you're going hopping from place to place.

It's kept by the time you have. So there's a fundamental limit here for each person of how

much they can do. How many things can be delivered to you? Quite a few, right? It's pretty much

everything you potentially buy. So maybe it's your purchasing power, maybe a credit card company.

I think I asked this very question the last time I looked at my DoorDash bill. Yeah.

How many times, Ed? 10, 20? And as a power user, that didn't take any convincing for me. It's like,

yeah. And also the other thing, the observation, again, this is my personal observation, is we

are barely scratching the surface of deliveries. If you just do the math of number of deliveries

to the population at large, and we're talking about US alone, later on in Europe, the places where

DoorDash has been active in acquiring volt and delivery, so we are combining forces there.

I think that the delivery market is, I feel personally, more stuff should be delivered to

me, right? I feel like there are more and more categories. I don't want to go out there and

spend half an hour, 45 minutes back and on each direction somewhere stuck in traffic.

So I love delivery. No question about it. I have a question. So I grew up in New York City,

where delivery was common in the 70s and 80s. You'd call anywhere and they would deliver.

And then I moved to Paris in the 90s, where delivery was not common. And so I remember

an era where delivery was common in one very specific market and nowhere else except for

Domino's Pizza. So how much consumer behavior change has there been with the expansion and

availability of delivery, literally anywhere? I mean, is it, has it replaced 20% of regular

shopping for groceries? 50%? I don't, I think it's less in my, okay, I don't know the numbers,

to be honest. I always go, I'm the tech guy. But this is actually, it's a very interesting

question. I think the behavior has changed because of the pandemic to a large degree.

I think a lot of us got used to it. And a lot of people thought, oh, now that the pandemic is over,

people are going to go back to sort of the original patterns. And they didn't. And the

reason I think is it is just so much more convenient to especially with people who have kids, like

so many of my friends who had, they realized suddenly that like, they didn't have to take

their kids into the target, because that what a nightmare that is. And they, they could,

they could completely avoid that by either doing like the curbside pickup or, or delivery.

I'm wondering from a technical perspective, though, like, not just a business one, but

you can comment on both. Why DoorDash went with a sidewalk delivery, but as opposed to,

let's say drone delivery, because we are seeing that. And I believe DoorDash might have partnered

with at least one company, but in terms of in-house, because what we're seeing, like, I just was

looking at Zipline's numbers and, you know, they're seeing that like, once people use the

at-home delivery drone, they're like doing multiple orders, like even a day now,

and they're adding more and more stuff, there's obviously a weight limit. But why not, why not

drones? Why, why was the sidewalk robot like so compelling? So let me make a correction,

which is very important, actually. It's a, it's a dot is not a sidewalk robot. That dot actually

lives in the bike lanes if they're available. It will drive on the side of the road, and it goes

up to speeds of 20 miles an hour. So the model we had for dot was actually an e-bike or person on

an e-bike. So it, but it is also small enough, thin enough to go on a sidewalk. So it will go

on sidewalk when necessary. And so, so the, the constraints we had when we designed dot was

will this work for our merchants as well as for our customers? And for merchants, this is very

important was we don't want to, if they want to, if you're going to ask them to load the robot,

they don't want to go out there and go hunting for the robot. That's part, you know, half a block

away because it was a bigger robot that works on the, on, you know, like a bit more like a vehicle,

regular robot taxi type robot. Whereas with the sidewalk robots, the constraint is, like I mentioned,

the, the reach is very limited. It works in dense urban areas where residents and

merchants are in close proximity to each other because especially for food, your half an hour is

about your SLA. So, and if your average speed is about two miles an hour, which is about on the high

end, then you're looking at about a one mile distance. So, so this is why dot is, we think of

the Goldilocks robot, it, it zips along at a pretty good clip, but it stays to the sidewalk. So,

and this is why we actually built it. We were perfectly happy to partner with our sidewalk

robot partners. They're doing great on our platform. And we also, as you mentioned, we are working

with wing, they're working in flight tracks, there are other, other drone companies. We also think

drone is an absolutely valid modality. And that's why our answer is all of the above. And that

actually the one other thing my team built, we call it the autonomous delivery platform. And the

idea is essentially coordinates and orchestrates across all of these modes. So you can, you know,

in the future, almost certainly we'll have a situation where you'll have everything available

in a particular geography, maybe Scottsdale. And, and this is like drones, you got, you know,

dots, you got sidewalk robots. And of course, we have the backbone for our fleet is our dashers.

And so, depending on their capability, so drones are payload limited. So that becomes a fundamental

limiter, right? And so you want to do use it for like, I'm in a dense congested area,

like the traffic and maybe it's like high traffic time, right? It's rush hour. And so a drone can

just skip through all of that, right? So you want a pharmaceutical, you know, order from one of those

pharmaceutical stores, and it can show up really quickly because the payload is low and it will

fly as the crow flies. So, and here it is, right? But also it has limitations on, you know, where

can it actually do the drop off and so on. So, so, so drones has its place in and they will

be continued to be successful. And we support all of the drone companies out there that are,

you know, growing their network. Same thing with sidewalk robots, we already talked about that

at LAN. We even have a partnership with Weymo. So if you want to do longer distance deliveries,

that again, meet our SLAs, you have Weymo because, you know, and dot is there for something in between

and then you need more complex deliveries, you have dashers, you know, dot cannot go up the fifth

floor of an apartment complex. So, and hand you your delivery. So, so, so we really think it is,

take the, you know, this goes back to what I mentioned before, a product market fit,

find the proper fit and the orchestration layer we have built ADP, which doesn't get as much press,

because, you know, it's not a physical product, but it's really, we think it's going to be the

key platform for the all, all modalities in the future. And that's, and that's the big piece you're

also pretty excited about. So when you go into the bike lane, which is reminiscent of a couple

startups who I don't, I don't think they exist anymore. They disappeared. I think of a bike

lane because I'm in one often enough, you're maybe encountering other bikes. Sometimes the road

conditions aren't really great. And there is a potential of getting hit by vehicles as well.

So what other challenges technically speaking, may I not be thinking of that, that you maybe

didn't even consider until suddenly you're a testing in the bike lane. And let me say we are

huge advocates for bike infrastructure because of some of the things you already mentioned. Yeah.

And by the way, we are extremely sensitive to bicyclists in the sense of, you know, we are

sharing a space with you. And we are very careful around bicyclists and other vulnerable road users

like pedestrians. Obviously, we are driving to the side of the road. So we might interact more

like, you know, kids and other things, other pets and so on. And so we are very careful around that.

It's been programmatically designed to be like that. And again, it's that's the sort of thing

you can do when you're, you know, transporting goods and not, you know, people. So but yes,

among the things we learned along the way is the split surface. So which I'm sure as a bicyclist,

you already experienced and they're not necessarily the edges of the roads are not as well maintained

as, you know, as the main surfaces. The other one interesting thing is a lot of these lights are

induction triggered. And I don't know as a bicyclist, you encountered that person, but

they don't the lights will not trigger. So you have to kind of go into the main lane to trigger

the light. So we actually been working with a lot of the city officials and they're really excited

to get this kind of feedback. Because I think, again, you know, people want to make more sort of

habitable communities and improve the biking infrastructure. So definitely these are kinds

of challenges of living in life in the bike lane, right? But the flip side is it does give us a

sort of we are out of the main that we are not in the hair of bigger vehicles, right? We are kind

of off to the side or people who might be using the sidewalk, which you're encountering a way

more complicated environment like walking their dogs, wheel, you know, people in wheelchairs,

like people with walkers, people who don't understand like in New York City,

who are visiting New York City and do not understand how to walk in New York City

is is another like kind of weird challenge. Absolutely. And this is why we use sidewalks

as necessary, mainly for pickups and drop offs. That's why the robot the width was designed in

fact was designed to go inside a regular sized door. Maybe maybe that someday we'll do that

would go all the way inside a restaurant. But but definitely the sidewalk pieces primarily at the

pickup and then at the drop off, especially for single family homes, Dodd can go all the way

right up to their door and wait, wait, you know, do the delivery right there. So so these are sort

of again, we're not trying to solve every delivery. And that that's kind of our phenomenal principle

here is that there's going to be a mode or a different type of delivery. So one of the other

pieces, you know, we talk about defining a problem and that with this technology that that you can

solve in a workable, tractable, viable way, one of the big pieces of that, of course, is regulation.

And you know, you compared your vehicle to an ebike, which I think is really interesting, right?

So so one of the challenges we've had with bigger vehicles is is where is the right like

regulatory from a regulatory perspective, you know, there's, anyway, we don't have to unpack

all of that. What I'm interested in is by calling your vehicle an ebike or comparing it to an ebike.

And I understand, you know, you your main focus is the technical side of this rather than

necessarily the regulatory side. But this is a multi piece problem. Where does that land you?

Because, you know, ebikes are like now kind of increasingly in a kind of infamously unregulated

sort of open environment. And and I see the opportunity there. Talk to us a little bit

about what what that means. Yeah, absolutely. So, you know, not just paying lip service safety is our

absolute 100%, you know, first principle here. And so, you know, very fundamental way,

I think we can all agree, and you guys have been following the AV space forever.

In general, AVs tend to be more careful because there's no notion of distraction,

right? And you have pretty much 360 degree cameras, you have other sensors. So your

our presence in the bike lane as far as safety of, you know, we are used like bicyclists and

pedestrians is our record is impeccable. And we expect it to continue to be we should ask

of that. We should ask that of all robots, right? They need to be as close, they definitely need

to be far better than humans. I think for acceptance alone, but just as an ethical and moral reasons.

And so, yes, from a regulatory framework, we are not categorized because of our speed limits and

where we operate, we are not categorized as a vehicle. And we believe that is correct because

I'm sure you followed some of the FMVSS and other requirements that would basically put us

in a very different category, which we are not in. But yeah, I do think that, yeah, I've read

New York Times got a whole bunch of pieces recently about, you know, some of the challenges with

the e-bikes and some of the ways you can go quite a bit faster, frankly, than other traffic.

But again, from our point of view, that is not that the comparison with e-bikes was mainly

as a metaphor, not that our behavior behavior is going to be far more differential, if anything.

Okay, so but just to narrow this down, because so you're not a neighborhood electric vehicle,

right? That's a regulatory category, but that's a vehicle category. That's not what you've created.

We are in some states, we are under the personal delivery device.

Okay, so okay, so there's actually a state level. Not all state private though. So that's

one of the sort of questions of where is this market going to go? Yes. But it's not automotive

equipment, right? Like, so you're not, okay. And just out of, and again, I appreciate that you're

not, you know, regulatory stuff is not your main focus here. I want to get more into the

business model and other stuff, because that's really interesting too. But like, what divides

that line? Like, what's to prevent someone? Is it just merely a speed limit thing? It's a weight

it's primarily speed limiting, which is exactly right. Speed is the single biggest determining

factor. It's the quadratic rate, it's mv squared. So so and the size, of course, right, you can't

that's the size. Okay, you have to fit within the bike lane also. So you start off, it kind of

you end up with something that looks like dot at some level, right? Because there are width

constraints, there are speed constraints, and there are, you know, size constraints in terms of

mass, etc. So so all of those effectively make make this it's sort of this category where you are,

you know, going to be sort of almost the regulation will make you into a particular form factor.

Yeah, interesting. Okay, I want to specifically talk about some autonomy challenges. And recently,

there's been a couple of incidents, not involving dots, I believe, from two other sidewalk robot

companies, where they've just like randomly crashed into the bus. That was a fun video,

like blinks after the crash. Yeah. And it just it's one of those things where maybe now suddenly

everyone's sharing these various things and this has been happening all along. But on a technical

perspective, what is happening in these cases where, you know, suddenly we're now seeing these

videos of these sidewalk robots that have been around for a while, just like crashing into glass

from I'm just I would love your viewpoint on the technical aspects of how that happens,

and how to avoid that from happening. Yeah. And this goes back to what you said earlier,

sidewalks are pretty unpredictable in high variance environment, right? You would have

high speed entities. Well, sometimes there are bicyclists who zip along, but leaving that aside,

most of the time, it's it's rather low speed, but lots of complicated things. And so

I can't I have no idea what happened with those two robots. I didn't see those videos. I also love

that guy goes around actoring the robots is so funny, but then helps them in the end.

But I can tell you what we do in dot and this is one of the things that I think from the way get

go. So a lot of sidewalk robots originally started as an idea that you're remotely driving them,

literally actually driving them, which at relatively low speeds, you know, you could

potentially say, okay, that's reasonable, perhaps our viewpoint from the very beginning was we

wanted to go obviously, we wanted this form factor. And so and a lot of the team that we put

together here came from our AV sort of, that's where our roots were. And so we built a stack,

very forward looking stack. And you can go into that if you want. But which was essentially a

full on autonomy stack, it was as L four as anything that's out there. And so that meant a lot of

investment into sensors. And the more importantly, the deep nets and everything, the perception

stack, everything that was built was, Hey, we need to be able to detect all of these different

things. And so the question then becomes, you know, again, not knowing what happened in those

situations, was it a human error that, you know, sometimes the glass is hard to see through. Also,

personally, I think we can all concede that AI and computer vision is now far better than any human

sort of visual equity can even bring to bear. And also just the fact that it's 360 degrees,

right? So once you have a full on autonomy L four stack, we absolutely don't think something,

things like that will happen to dot for that reason, because it's been built to be fully

autonomous in that sense of it detects everything, it knows what everything is. And it is all usually

always going to be, if been in doubt, it will be careful around entities. So, so all of those,

I think when you're, when if you have indeed a human driving a robot on a sidewalk, and you have

a large number of people doing that, because now you're trying to scale, you could end up with

a pretty, pretty high variance in behavior, right? There might be some people who are very

deft and they can do all kinds of stuff, but operating a sidewalk robot, and there might

be some people who were just brought on recently and assumed maybe they were not familiar with,

you know, I think it was like a bus, bus stop, right? With shelter, bus, bus shelter, bus stop

shelter. Maybe they aren't just familiar with that because they're not from that part of that

state or that city. But isn't, isn't Lidar like a pretty key piece of that too? Right? I mean,

just, but Lidar, yes, Lidar can be good vision is pretty good now. So we are, we do use Lidar,

but we are pretty vision primary now. And, but, but it is, you know, and also you can compliment

these, like radar can also be very useful for certain things. So, so it's, and frankly, all,

all the sensors are becoming pretty low cost now. So it's not clear anymore that you have to choose

between them as opposed to creating a system, which is, you know, some things are good at,

good on Lidar, some are better at vision and so on. So, but, but yeah, we have very, very much

vision primary and vision has gotten extraordinarily good now. So, and just to tease out this, this

sidewalk challenge a little bit more, because right, you define away from the, the speed and the

mass and as we said before, it makes the safety case easier. When things are moving slower and

also, you know, it's harder to predict. I think this is one of the things I've tried to explain

to people, it's very difficult about, about like, if you're crossing the street in front of an AV,

the more sort of like confidently you walk in front of it, the more it's able to sort of like

use your, the data you're generating in real time to predict your future motion.

That challenge becomes more difficult at the slow speed, right? This is why, you know, people

thought, for example, summon on, you know, in the parking lots was going to be like the easy thing

for Tesla and it's all hard, right? Again, I mean, if you ask any of our AI team members to say,

what is the hardest sort of ODD or the part of the autonomy that's the hardest,

it will be parking lots and sidewalks. I mean, that's where it's kind of chaotic. And yes,

just going slower doesn't mean any, it doesn't make you better necessarily. It does give you

time to react and your time to collisions are lower or bigger rather. And so, but, but on the

flip side, because of those, the nature of those environments, you are going to get so many actors

that are, you can almost take Brownian motion at that point, right? And for some reason, people

even driving in parking lots in particular in malls seem to feel like no rules apply. And so,

it's just, I don't know what happens to, and I've done that myself. So I'm not pointing fingers,

where, you know, you're far more careful on the road for some reason. Whereas parking

glasses where you can get a lot of, maybe it's just vendor benders, but you can definitely run

into these situations there. So yeah, I agree with the sidewalks are not as easy as they seem.

Certainly if you want to go zip along on a sidewalk and any kind of clip, that's going to,

and they're not also necessarily maintained. That's the other part, right? It's the road surfaces

can be pretty, pretty challenging on a sidewalk. So we've talked a lot about certainly what DoorDash

is working on. But I would love to get your opinion on all the companies out there that are working

on, you know, robotaxis as a, you know, a technical perspective. I have to imagine that you've

written in certainly the Waymos and you were working at Zooks. Have you, have you tried

in Austin the Tesla robotaxis? Because the one in California doesn't count. There's no

permit. It's not autonomous. It's not. No, I haven't, I haven't been to Austin.

Well, great city though. But no, I haven't. I'm looking forward to it. I definitely will try it

if I get the opportunity. I mean, I'm a participant, but I'm also a huge fan. You know, my PhD,

if you ask me, was in robot motion planning 30, almost 30 years ago, I'm dating myself here.

And there was, there was literally nothing interesting in robot motion planning going

on at the time. So to me, it was almost like, yes, finally, we have a ride.

So nothing interesting was happening at the time he said, so why did you,

why did you pursue that? Because I loved, I loved robots. The math was cool.

I don't know. I mean, it's been here, been here in academia. You just go, you know,

it's like, you don't think of practical things like, am I going to be employed or not?

You just go. You go after what is interesting. What do you think of this shift that has happened

where, you know, going back to like the DARPA days, there was much more of a robotics focused

approach in terms of like the programming and things like that to autonomy. And now that seems

to have shifted a few times, but this idea of like these, this end to end units, what, as, you

know, an academic, do you think of that? And do you think that that's the ultimate end or will

this shift again? And companies are going to go, Oh, actually, no, no, no, we're going to go back to

either something more old school or some new, some new approach.

Yeah, I think I do think directionally, that is the right approach. The question becomes twofold

in my sort of the way we look at it. And we are by the, our stack is pretty straight of the art,

but I think the approach we have taken, we can talk about it in a bit, might be at least a way to go

about it. So the problem with just pure end to end, and it's all about definitions, by the way,

here. And it's very careful, you have to be careful on what the actual what somebody means

by end to end. So the pure end to end approach is taken mostly camera data, just taking camera

sensor information, take the raw sensor information, and outcomes, the controls for the robot,

right. And you don't even have like low level controllers necessarily in the purest approach

to that, meaning that you don't have any kind of way to manipulate the robot deterministically,

like, for example, you're driving on a split, what's called a split friction surface, which is

what dot drives on, because, because the payment ends at one point. And then there's as a bicyclist,

you're probably completely aware, there's like this little gutter area that is not usually

surfaced. And it has, so there are two different frictions there. So you're pure end to end person

would say, no, no, no, don't tell us anything about that, or just train it and it will learn,

right, it'll figure it out, right. And so you get this black box, which has figured stuff out,

maybe like, you know, I have a now a five month old baby, it's very interesting to watch what he's

figuring out on his own. But that means you're now dealing with a black box. And so how do you

know, especially in the current state of things, how do you know this is fully like safe at some

level, right, how do you validate and verify that it's doing what it is supposed to do?

You can try it out a lot. And I don't know, would we would do that with, would we have done that

with flight? I suspect not, right, we'll say, oh, we've tried, we've flown up, you know,

towns of times, it's just fine, don't worry about it. So I think there is clearly that question

around it, like, how do you go about even debugging something, okay, something it was off,

why did you do that? And the answer is well, train it some more, give it negative examples,

and so on, right. And I'm oversimplifying things. But conceptually, it is essentially

what I described. So then, so there's that end to it. Then the other question is, can you actually

use this notion of end to end, but use it in a very targeted, directed way, along with the system

that is also saying, hey, you might be end to end, and that's great. But I'm going to check

your trajectories, I'm going to check for collision, and you already have a baseline that

you're built. That's the approach we took. We are building exactly fully, and by the way,

it helped us a lot as a second more. We started building the stack from scratch in 22.

And a lot of these ideas are already there. And we kind of had a much better idea. So I sympathize

a lot with some of the older, you know, it's much harder to undo a stack and say, okay,

we're going to switch over immediately and go do more, learn things and not. But so, but what we

also did was we built a baseline stack that was a geometric search-based planner. And that now

becomes our safety net, if you want to call it that. And so it always, no matter what you can do

with a box, black box, we can check. And so now we can target that black box at very specific

things. For example, a construction zone might pop up. A search-based planner may or may not

be great at saying, oh, this is how to handle a construction zone, because there's so much variation

in what the construction zone could look like. And this is what, I mean, at some level, that's

what humans are great at, right? You put us in, you don't have to go relearn a city, a new city,

right? We just know how to drive. And so there is, that's why I meant the directionally, that's

where you do want to go. But if you can build it in a way that is, you know, you can in real-time

check what it's doing and ensure that if there is something off, for example, then you can, you know,

again, this makes far more sense for delivery. Sorry to sort of hype or go on and on about it,

but it's, we have such a nice feedback loop of everything just works just right. Because

the delivery, you can say, hey, okay, you know what, I can't solve this construction zone.

I can come to a safe stop, right? And say, okay. And then with the dot-sform factor, we're not

impeding traffic. So, so anyway, so back to answer your question. I, the short deal here, I guess,

is yes, it is the right direction, but the big, the big sort of $64,000 back used to be,

question is, yeah, how do you validate something? And then if you're a provider of such technology,

are you going to willing to take that liability on, you know, and to me, that is actually the

question I would ask lots of people out there is, yeah, how do you, how do you prove this is safe?

And how do you, how do you, and if something does happen, who has a, you know, especially if

you're like throwing the technology and giving it to other people. So, nice. Alex, take us home

with the last question. I know you must have something. Well, I have to ask this of everybody

who deploys any delivery robot sidewalk or bike lane, how many issues have you had with vandalism?

And what was the craziest one? Because when I see them now, I'm like, oh, that's cool.

But if I were 12 years old, and I saw that, like I'd be after it with a flamethrower and

like, you wouldn't be hugging it and like putting a cool sticker on it or something.

I was a bad kid. So what have been, what has been your experience with them?

So actually, we were also, you know, that was a concern and also a thing we were looking out for.

We're built by the lots of technology into the robot that it, the momentum, by definition,

right, it's an autonomous vehicle. So we know when something is too close. So it can give,

send an alert to our, you know, control center, so to speak. And, and so we will know what's

going on. We also have all the video. And so yeah, we had, you know, you have the usual sort of kids

that are curious. I remember we had a Christmas party several years ago and we invited people

very good to bring their family and the kids in particular, and they all wanted to run up and

hug, hug dot because, you know, it's been decided to be cute. And I'm like, oh man,

did we overdo this because, you know, that could be an interesting challenge on the street.

But generally speaking, it's been actually pretty, pretty good. We have had issues, of course,

we had a homeless person sort of who's not maybe, you know, in the best condition or health,

attack it, but felt very random. Like it wasn't clear what, what the intent was. Maybe,

maybe she thought there's food in there. But the awesome thing about that was

other people call the police. We realized it was happening. So we sent a dispatch,

but already people had called the police. And I want to think that because Dot is part of the

community, maybe they felt like, you know, maybe a man to formorphizing it too much here. But I

felt like, oh, yeah, this cute robot's getting assaulted. Let's, let's, let's help it out. So,

so, but we haven't had too many of those incidents so far. But, you know, you never know.

And I think it depends on where you deploy. You must not be in many college campuses then.

We are operating near Tempe or in Tempe. So, you know, they're not that far.

Oh, okay. Well, that will be the true test. ASU students. There you go.

I better go down in disguise. I'm confident that, you know, at some point, that's going to happen.

And people are, you know, you know, especially after a game or something and people up or

there'll be a, there'll be a fraternity initiation stunt or something.

Oh yeah. For sure. For sure.

Quite possible. Yes.

Well, on that note, thank you for joining us on another episode of The Atonic Cast.

Thank you so much. This was a blast. And yes, come over and help me with some interior design

tips anytime. Alex.

I have no tips for you, my friend.

Great. Thank you.

Tesla Model 3

About this episode

Original notes