Starship's Ahti Heinla: 8 million deliveries, zero stolen robots, and why sidewalk bots don't need lidar

repetition, some flashcards or something. Um but uh next up we have the founder of Starship coming in to the building. Welcome to the show. How you doing? Boom. What's going on? Hello. Good to meet you all. I am doing fine. super exciting thing and I like it. Yes, thank you.

Uh would you mind uh kicking us off with just a little bit of background on the company and explain uh not just what Starship is but kind of uh the the the footprint, the rollout, the strategy, all of that. Yeah. Yeah. Yeah. All right. Yeah. Sure.

So what we are doing we are uh we are a company del developing and building and operating delivery robots. Yes. So robots that transport stuff like you know and stuff that means you know burgers, milk, you know could be you know packages and so forth, right?

So in science fiction movies you know you don't see UPS guys knocking on your door. Mhm. You see things coming to you, things flying to you. That's the future that we are building. That's the present that we are building.

And uh uh we have many competitors as well who have who we have inspired to do similar things you know with drone deliveries and so forth.

Beardwing is on the ground but robots that drive on the ground small robots that drive on the on the ground and uh uh we are we are actually not like a like but many people think that you know this sort of futuristic thing is like a pilot or test you know somewhere in some limited area or something like that.

We are actually in full commercial operation with thousands of robots in hundreds of locations.

It might not be you know we are not not there not yet in every place that you know all of this audience that's listening to this right now is but we are in some places and in in some of these hundreds of places that we we are in uh you know our robotic delivery is completely commonplace.

It's something that people are completely accustomed to. It's something that uh you know people don't really use you know like human couers to to to deliver things they use robots. Yeah. Um can uh can you talk about all the different things that you've seen around the way humans treat robots?

I try to I try I I try to say thank you to robots you know when when they help me out with different things.

But some humans, you know, we saw this with a lot of vandalism of We saw this, you know, bird, the big thing with bird vandalism in LA is I think a lot of birds issues locally here just came from a culture of vandalism. You did the hard thing in making sort of cute robots, which I I think uh probably helps.

But uh how is how do humans kind of present a challenge to trying to like automate delivery, which should be in everyone's best interests? We get that question a lot but we don't actually get a lot of vandalism itself. Uh you know the the the the truth actually is that people really love our robots.

Yes I see that you know with like scooters and you know some of the other modes of transport people don't really treat them nicely but they actually do treat our robots really nicely. Like for example you know you know we often you know get the question that oh you know are your robots stolen?

And uh actually none of your ro our robots have ever been stolen from the street and we have done 8 million deliveries like we are doing millions of deliveries and it's just not happening you know kids feed our robots bananas so the the the treatment is completely different than with you know like scooters for example I understand you know that's you know that's actually a lot we get that that question actually a lot like people you know sitting in you know sitting at the table very often and say that you know I treat things nicely but there are some other humans out there that do not treat things nicely just like you said right you know but it's not actually happening it's not actually happening you know like how often do you vandalize like a you know UPS vehicle you don't really right you know sure you can do it yeah you can puncture puncture the tires of a UPS vehicle you don't really do it right you know and you know our robot actually our robot has 10 cameras it's constantly connected to the internet It's really It has a really loud siren when it's tampered with and so forth.

It's not actually really happening.

Well, can you can you uh talk a little bit about the roll out maybe because um I I think that the locations in which the robots are doing deliveries is like really going to impact right if you're rolling through a suburb and it's a you know familyfriendly neighborhood something like that the robot is going to run into different challenges than you know rolling down you know a street in in Manhattan for example.

So how how have you guys approached your kind of like go to market uh and kind of picking what regions uh are are are you know sort of top of the list in terms of uh getting robots on the ground. Yeah, great question. Uh we are operating our robots in at 60 cities and 60 college campuses. Mhm.

So we have learned a lot about different environments and uh yeah like different neighborhoods are completely different like there's a lot of difference in the world and we operating in six different countries like five countries in Europe and you know US right and you know US of course you know is varied as well right uh so we have learned a lot on how the different you know sidewalks look like how crossings look like how traffic light lights look like you know what the traffic patterns are All of these things we've learned a lot of these things a lot of data for our machine learning algorithms and so forth right but in terms of actually how people treat robots there's not much difference actually people are actually really friendly towards towards our robots I understand it's really hard to believe but that is actually true it's completely true and uh we are operating a completely commercial quality service we are we are cooperating with you know most of the major delivery apps in the world and with with a number of you know tier one retail retailers as well and uh it just works.

Can you talk about the progression of the technology specifically uh the path to end to end there's a single AI model but there's probably tea operation in the early days then there's uh some mixed uh there's some C++ for pathf finding but there's some AI for you know image processing and world modeling uh how have you thought about developing the technology and do you see uh tea operation and endtoend AI systems playing nicely together in kind of a centaur mode for a long period or is this are these specific like gates that you have to go through?

Yeah, we are operating a quite the hybrid model. Sure. We have been in we working on this for 10 years. Yeah. Right.

And uh you know we built this at the at the time where you know AI was not actually as developed and we have obviously reap the benefits of the all of the AI development that has happened has happened since. Uh but uh but we also recognize that safety for example is super important. Yeah. Super important.

You know like for example our robots you know cross roads. They they are generally sidewalk robots. They drive on the side of the road as well when there is no sidewalk but generally they drive on drive on sidewalks but they cross the roads.

They cross roads in similarly as a pedestrian does you know with using crosswalk right? A robots cross roads 100,000 times a day. 100,000 times a day, right? We have thousands of robot robot robots operating, right? You know, we have been speaking right now for eight minutes.

During these eight minutes, our robots have probably crossed the road, you know, over thousand times. Wow. Probably probably it's like that about a thousand times during these eight minutes that we have we we have we have talked, right? You know, safety is super important.

Suppose we have something going wrong in like 1% of the crossing. That's too much. Mhm. That's about 1% or 0. 99% too much, right? So we need safety. We need to prove that it's safe. We are actually not operating an end to end neural network but we are operating a combination of yes C++ and neural network.

When when when you see uh hard techch founders claiming online that they're operating an endtoend neural network which some people have done. Does that does that surprise you? Is it almost unbelievable or do you think it's not surprising to me at all that you can do it?

Uh the downside with that it's very hard to prove that it is safe. Uh it's it's very hard hard to prove that prove that it is safe and you know we are actually operating in also some pretty challenging regulatory regimes. I mean not just you know loop is actually beneficial there right?

Like if you have the ability to remotely take over, that's going to make your safety case so much easier because you're going to say, "Hey, yeah, it is kind of crazy. There is a robot piloting this a little bit, but at any moment we can hop in and beam in and there's a human. " Exactly. Exactly.

For for us, it's a combination of remote assistance, which happens, you know, more and more rarely all the time, right? But it still happens.

It's there there there is a human you know somewhere that can take over in difficult situations or complex situations unusual situations then there is C++ and there is end to end neural network as well. Sure. Sure. Combination of all of these Yeah.

has the transformer architecture or any of the other uh like kind of foundational innovations in AI been important to Starship and what you're building.

Uh obviously we see hype around the studio Gibli moment and diffusion in image processing but does that actually make it easier for you to understand the stuff of like where am I in the world? Where am I going? I need to plan a path.

Uh how has how should we be tying all the amazing progress in LLMs and AI to your business? Yeah, it is definitely helping and improving our stuff. Mh. uh I think we were in commercial quality operation uh already before that. Mhm. But it is helping us tremendously for sure.

Uh primarily it is clearly something that is is dramatically reducing the need for this you know human uh somewhere in the loop. Sure. Uh absolutely. What about um have you thought about uh giving you know embedding some sort of LLM or voice voice model into the robot?

So if a if a pedestrian bumps into the robot, they can have a little conversation and it can kind of explain like, hey, I'm just going across the street. I'm delivering a burrito. Like it can answer some basic questions that seems like I'm just the burrito guy. I'm just the burrito guy.

Uh that seems like maybe silly but also like maybe great from a user perspective but also like wildly extra. I don't know. Have you thought about that? Yeah. Yeah. Uh yeah. So we do not have like a personified LLM in the robot right now. But I think it is conceivable that that that will be the case. Sure.

Our robot does speak. Yeah. But but the speaking is not actually driven by an by an LLM. But it's more like business logic decision. Exactly. It's a little bit more business logic that Yes. you know, every time the robot, you know, does a delivery delivery, it says thank you. Yeah. Right. Doesn't take an LLM to do that.

Yeah. Um, well, can you tell me more about tradeoffs in robot production? Uh, Elon has been like anti-lidar from a cost perspective. I'm sure there's a bunch of different trade-offs in terms of size, weight, speed, battery life, all these different things.

You have probably like a base hub where these things go to charge and then they have a certain amount of range. uh what are you optimizing for? What are some of the pitfalls to avoid? Yeah, great great question. So, uh our robots actually do not have a lighter. Mhm. But that does not mean that we are anti- lighter. Mhm.

Uh the thing is though that the reason we have not used lightars is that lighters is effectively lighters are perfect sensor for uh for an autonomous vehicle. Mhm.

Autonomous vehicle needs to needs to see quite far away and it needs to see you know it doesn't really well it it does need to be need to have short range sensors as well but you know it does need to you know see you know like you know 200 yards 300 yards because it's moving fast right our robots are moving much slower than a car right they don't actually need to see that far so light are actually not perfect sensor for us and the downside with lighters is that lighters typically have a narrow vertical field of view.

M they have like this narrow you know you know uh rays effectively that see very far but it's a very narrow the angle the vertical field of view is very narrow like a couple of degrees or so right we actually need to have perfect vision from immediate vicinity of the robot like very wide nar wide you know vertical field of view we need to see you know down in front of the robot and also you know up and you know we need to have have that sort of sort sort of vision.

So lighters are not perfect center for us. That's the reason we are not using lighters but the moment that that a lighter with suitable spec appears on the market, we will absolutely use it. So we are not like religiously anti anti- lighter at all. Jord, you have a question. Uh changing gears a little bit.

Uh Skype is shutting down on May 5th. Uh and end end of an era. you were the the founding uh engineer there. Uh is that emotional uh for you the the shutdown or at this point you've been you know uh in doing something you know else for for uh Yeah. Yeah. Yeah. I've been doing something else for for a long time.

I'm not actually an active user of Skype anymore. Sure. Or actually quite quite some time. I still have the app, you know, in my phone somewhere, but you know, not really using it all that much any anymore. I mean Skype was an amazing ride.

I mean it was one of these startups you know which is actually rare you know one of these startups that just took off like a wildfire immediately from day one right from outside perspective all startups seem like that that they come out of nowhere and they just boom you know it's like that but you know but in reality as a startup founder most startups are not like that most startup is hard work hard work before things actually start you know getting off the ground right kind was not like that you know it was like for me it was easy work I was like I just did what I love to do and boom, you know, loads of users came and you know, it just kind of kind of happened, right?

So, it was an amazing journey, but also, you know, frankly, I mean, it was a journey that happened 20 years ago. Yeah. 20 years ago. Literally 20 years ago. What what what are the key lessons that you've t that you took from the Skype story and applied to building this business?

Is it just the engineering culture, the pace of play, or is it wildly different because it's a completely different growth curve?

It's both of these things, but it's also I would say one fundamental thing is that uh with with a lot of products and a lot of services, you kind of uh you know, it's a very simple service really like you know one one delivery delivery app a major delivery app founder you know told me that you know look after you know if your robots really work and you can give me cost savings because it needs to cost less than than than human delivery mhm for me then I will use your service any day and you know that's how it is for them that's how it is is for us you know we have no demand problem right if we prove to them that it really works and that we give them cost savings they will use us in like for like billions of deliveries that's how it works and that's the traction we are we are seeing on the market and Skype was like that as well you know if you actually actually put out a product that exactly fits what people want and it just works.

It just works. It doesn't have some sort of major downside, right? You know, then it just takes off like wildfire. Granted, it's harder to do with the self-driving robot, right?

clearly much harder to do, you know, like we built we built Skype, you know, like we we were like a team of like how many engineers we were like, you know, 15 maybe 15 engineers, nine months of work and we have a beat had a beat out there that everyone loved and just took off like wildfire, right? Yeah. Sure.

You can't do like a self-driving robot with like a team of 15 engineers in nine months, right? You know, it take it does take longer longer, right? Yeah. But it's also harder for competitors, right?

You know, so we are, you know, I said, you know, we have done, you know, 8 million deliveries I'm not sure you know our closest competitor there are lots of competitors tens of competitors our our closest competitor probably has done 200,000 maybe or 300,000 not sure something like that right like an order of magnitude difference so we started this trend let's say you know 10 years ago and we're still number one because whatever is hard for us it's also hard for our competitors.

I have one last question then we'll let you get out of here. Uh why is Estonia so successful in producing technology entrepreneurs? Uh great question. I think uh uh I don't have a full answer. No. Um I don't know.

But the but uh but but one thing the thing out there is that um you know Estonia was occupied by Soviet Union for for for like 50 years or so, right? and and I was you know my age is such that I just turned 19 when we regained independence from the occupation right. Uh so I I I spent my childhood in Soviet Union.

Uh but my adult life has been like in in a in a free country, right? and u uh you know turning 19 and finding yourself in a free country that actually doesn't have a lot of establishment. Mhm. Built in that means that you kind of grow up with an assumption that there are no obstacles. So no obstacles for you.

There are no big companies out there that you need to kind of compete with. You're just free to do whatever you want, right?

So that's the that's the culture I think you know overall you know successful startups can be built by any but it helps that if you don't know what's impossible you knowing you don't know that it's impossible you you're you're going to have a better success than you just think completely agree that's great that's amazing that yeah that that's amazing thank you so much for joining the show congratulations on your first 8 million deliveries and uh looking forward to the 800 million we'll have to have you back on Then next year, yeah, next year.

Let's hear it. Amazing. Thank you so much for joining. We'll talk to you soon. Cheers. Bye. Um, before we bring in Yeah, Daniel. Uh, well, I will bring him in and I will let him give his introduction. Daniel, if you're in the studio, welcome to the show. Good to see you. It's been uh a few days since we hung out.

Still looking great. How you doing? What's the