Stanley Tang on DoorDash DOT: 7 years to build, L4 autonomous delivery live in Phoenix today

the thing. Once it's out, they can they can top ticket. Yep. Or they can ride it up. But um he heard people loud and clear. Yep. Uh, well, we have our next guest. We're going back to Door Dash World with Stanley from Door Dash. Welcome to the show. Welcome to the show, Stanley. Yeah, thanks for having me. Absolutely.

Massive week for you guys. We got uh a lot of the updates uh from Tony earlier, but we're pumped to talk to you as well. Yeah. No, excited excited to be here. Thanks for having me. What out of um yesterday's announcements are you personally most excited about?

Well, the announcement I'm most excited about is is our autonomy program, which we are finally unveiling for the first time. So, uh, yesterday we announced Door Dash Dot, which is the the first ever commercial and autonomous delivery robot to travel on bike lanes, on roads, and sidewalks.

And and the key to what's different about DOT is that it's really purpose-built for local delivery. It's a tenth of the size of a car. It's about 350 pounds, but it can still go up to 20 20 miles per hour.

So, it's it's much faster than a traditional delivery robot that only drives on sidewalks, but a fraction the footprint of a autonomous car. And it's something I'm really excited about announcing because we spent many, many, many years developing, seven or eight years. Yeah. Yeah. Exactly.

So we so we started Door Dash Labs which is the team I I get to work with have the privilege to get to work with and um many years getting to this point.

I think we've built one of the the the most sophisticated autonomy stacks out there and and more importantly we scaled it on this super novel form factor and and and that's the reason why we're announcing DOT is we think we have something that's ready ready for scale. We're went live um we're live today in in Phoenix.

Um full autonomy, full L4 doing real. Yeah. Break down the we we've covered the levels um autonomy levels on the show before, but but uh could you kind of summarize it or kind of give your uh framework for how you think about autonomy?

Uh yeah, I mean I think I think well I think for us like building autonomy well I think I think there's a lot of nuances when it comes to autonomy. Building autonomy for like local delivery our use case is actually a pretty different challenge than let's say autonomous vehicles.

I mean you're obviously seeing a lot of progress in the past few years in the robo taxi space uh with with the likes of Whimos. Yep. um and and things like that. But autonomy for delivery is a slightly different problem.

I mean, it's like it's it's it's it's it's not just um and and the reason is because there there's this third party involved. It's called the merchant. It's not just you and a robo tax, the hardware. Yeah. Yeah. Yeah. Right. And there there's a merchant. Uh it's it's about it's about delivering packages, not people.

Uh and really you have to figure out how do you solve that solving that first and last 50 feet problem becomes super super critical because people expect packages to be delivered right to their doorstep.

Merchants you have to make it really easy for a merchant to integrate into robots and you guys had to build your own map your own map system just because of even how challenging last mile is when you have a human that's doing the delivery. Right. Exactly right. Exactly.

I mean that's it's it's so so it's I mean like like you said I mean we've done over 10 billion deliveries in our lifetime. We we have so much data on what works, what doesn't insights into like you said where that pickup point is, drop off point is like the exact door you need to go into.

Is it a long distance delivery short distance like pizza delivers different like ice cream groceries versus restaurant food? So it's like it's it's really all that data that we've we've gathered and you kind of have to distill all that down into in your guys's re research.

How important did you find or did you think it would be to make the robot cute? Uh it's very cute robot and uh that feels important for the way in which people out in the real world engage. I've seen people messing with all different types of robots.

I mean, the the the number one example being here in Los Angeles, there was an there was an era where people celebrated throwing scooters off of buildings. Um, you remember this? Yeah. Bird graveyard. Yeah. Bird graveyard account was birds were they just looked like normal scooters.

They weren't really cute, but they also didn't look dystopian to me. But people they weren't dystopian, but they weren't they weren't um at all like you know if if somebody was going to mess with it's like you're messing with a baby carrot. Yeah, it does look like that. a burrito baby car. It doesn't feel Yeah.

It would feel very wrong to to mess with it. And I think that's important for durability of the fleet, right? And and making the environment better, right? If you see one of these, you're going to smile and think, "Oh, that's cute. " For sure. Totally.

I mean, that that was that and it was it was very very intentional the way we we came up with the design because the last thing you want is, "Oh, this is like an alien spaceship that landed in your neighborhood. " Yeah. It's like this is your friendly neighborhood delivery delivery robot.

I mean, you guys are you guys are integrating with other other uh types of fleets as well over time? I know uh drones are a category that you guys are partnering on.

How should um how should companies be thinking about integrating into the Door Dash network if they're building building actual hardware that they want to um participate in this physical delivery economy? Yeah, I mean totally.

I mean I think I think the other announcement we made yesterday is is also our autonomous delivery platform which is really the the secret sauce that makes all this possible including dot like dot is really part of the bigger broader vision of creating this multimodal fleet that hopefully we that that would consist hopefully will consist of all types of modalities right like like and the reason is because the again the scale at which door dash is operating today you know we've done 10 billion deliveries the variety the the the the use cases that we have on Door Dash now from groceries to food to restaurant to, you know, you order toothbrush and iPads on on Door Dash now.

Like we're we're going to need all the modalities we can get. So So our our vision here is is really creating a multimodal future of, you know, some orders might be great for dashers.

Uh if it's like a complex grocery order that requires pick and pack or you're navigating through apartment complexes, you still need a dasher. If it's a five minute, you want five minute coffee delivered in in a rural area, drones are great for that.

Y if it's a dense urban core, college campuses, we still want sidewalk robots.

And of course, there's DOT, which we think is really ideal for the the kind of the dense suburbs that that that 3 to five mile delivery in these in the dense suburbs, places like Phoenix, places like markets that you guys dominated in from the very beginning. Exactly.

the mark is that where a lot of the delivery that Door Dash happened and we're known to be the the the delivery company that was built off of the fact that we went after the suburbs, the the kind of the underserved delivery markets.

In the history of the chatbot, I think of the transformer paper as a key turning point, then the common crawl scrape the entire web into a database as an important milestone, and then the scaling of compute and the larger and larger clusters to get from GPT3 to GPT4. And then maybe you put RLHF in there as well.

uh what's going on in in delivery robots like is I is the transformer paper important? Are there is there another kind of algorithmic turning point? Was there is it a data acquisition like we need to get to a certain scale of data to make this work? Is it a certain amount of training compute?

Are we training these models on massive clusters? Like what what happened in the tech tree to get us to this point? Totally. I mean I it's probably a little the honest answer it's probably a little bit of all of the above. Yeah.

I mean I I think the big I'll say the big inflection point in in our space or kind of my version of what I call like the four-minute mile version breaking the four-minute mile barrier for autonomy is definitely seeing Whimo making it possible in autonomous ride share.

Um and and and and I think that was kind of the the proof point that no like that the tech is possible and and all of a sudden the problem shifted from it's no longer a science fiction R&D uh kind of initiative now it's it's fully shifted into a how do you commercialize operationalize scale um kind of hardware on Whimo Whimo and Tesla have have taken like hard lines on liar versus camera only have you had to take a hard line there are thinking that to be more flexible there like what's your take?

Yeah. No, well our our stack is is is uh we use cameras lighters and radar but okay it's definitely much it's definitely a camera primary approach and I think it's going to be increasingly more camera heavy as as especially as as um sensors get better and better. Yeah. ML gets better.

I think I think there's also the shift from like you know 10 years ago perhaps the autonomy approach was a lot more kind of heruristics or rule based and over the past few years like the entire industry I mean obviously Tesla is kind of the kind of the the lead on that is kind of shifted to kind of ML AI kind of more endto-end kind of approach.

Um, so, so that I mean that's definitely I think that's definitely where which is why you're saying like autonomy that the software piece of autonomy is is now finally coming to fruition and now you're seeing the problem now it's shifted towards okay now you you have the software figured out how do you actually scale hardware how do you actually scale operations how do you scale scale commercialization and I think that's kind of where we are in in that moment right now and I think that's also where Door Dash gets to play to its trends I mean like That's the piece where the core that that's that's the piece where the core businesses get at.

That's our bread and butter. How do you how do you see operations? Um how do you scale fleet management and integration restaurants etc? Last kind of oddball question. Um do you think that humanoid robots have any place in the uh uh delivery chain?

Like you think I'm sure you guys have been pitched from various companies trying to say, "Hey, you guys work with a lot of restaurants. Maybe we can integrate there. " Or is there a world where an apartment complex has uh some type of humanoid that can function and help with that uh last 100 feet of delivery?

But or is it just not not even that great of a form factor? I definitely see it. No, I I can see that. I mean, again, it goes back to what's the use case you're going after, what's the problem you're going after.

I think a lot of I think one of the biggest mistakes a lot of these people in who go into hard tech or or AI or autonomy end up doing is is they end up doing it end up becomes a research problem like R&D like they're building technology for the sake of building technology instead of okay let's start with what is the use case what's the problem right just like how we came up with dot we start with the problem first work our way backwards um and then come came up with the idea kind of design.

Yeah. Right. So again it's like with humanoid again it starts with like what what's the use case? Is it is there something to do with our dashs and micr fulfillment centers? Is it solving that last 100 feet? Is it something around rational? I uh I like to start with the use case first.

Um and then we can see if if humanoid is is is is the right solution. And what kind of humanoid? It's pretty funny to think of a humanoid trying to do like let's say like a four mile delivery sprinting with with a with sprinting with a coffee and a bunch of food is just flying everywhere. Yeah.

So so that's probably not the right use case or maybe the better use case at our maybe it's at our dash marks right like that could be you know like these again have you have to we have to just start let's start let's start with what the use case is and work your way backwards. Yeah that makes sense.

Uh super super exciting. I don't think Door Dash gets nearly I mean obviously you guys get plenty of attention but but not nearly enough on on how many different areas you guys are innovating. So would love to have you both back on in the near future. Thanks so much for stopping by. Yeah, we'll talk to you soon. Cheers.

All right. All right. And if you want to put your sleep on autonomy on autonomous mode, head over