Cofounder, Zoox
Inside The Ride: Scaling Zoox | Episode One
Building a robotaxi from the ground up means making decisions that no one has ever had to make: from sensor architecture to ...
About Tim Klay
Tim Klay, cofounder of Zoox, appeared in two recent videos discussing developments at Zoox and his work with a new AI system called Hyperdrive. In a July 2026 episode of the show "Shotgun with Marwan," Klay demonstrated a system his company Hyperlabs built that uses an Nvidia AGX computer running at about 33 watts to drive a Tesla through San Francisco, calling it "hyperdrive." He stated that the system was created by four software engineers, three algorithmic architects, and one info engineer over about two years. Klay said his motivation for moving from computer graphics to robotics was the belief that around 2012 a transition from the automobile era to the robotics era was beginning, and that the power-to-weight ratio of Hyperlabs' system is "probably unmatched on the planet right now." In a June 2026 episode of "Inside The Ride: Scaling Zoox," Klay described Zoox's latest software release, stating that it quadrupled the size of the geofence in San Francisco, added major pickup and drop-off locations in Las Vegas including the airport, and began driving robotaxis in Austin and Miami for the first time. He said the release improved the vehicle's ability to handle dense junctions, lane changes, and unpredicted turns. Klay noted that when Zoox first drove on public roads the vehicle could get stuck every few hundred trips, but that with recent progress it now gets stuck in "many, many tens of thousands of miles." He attributed this improvement to custom hardware that is "very redundant" and "fail operational." In a selected quote, Klay stated that Zoox's approach "took longer, took more money, but we actually end up with a better, safer, more comfortable, more enjoyable experience that actually will have better unit economics."
Transcript (82 segments)
H
Host0:00
Hello and welcome to our first episode of Inside the Ride. Where will we go today?
T
Tim Klay0:04
I hear we're going to do a Fisherman's Wharf.
H
Host0:06
Let's do it. Okay. Should we go ahead and start the ride?
T
Tim Klay0:08
Let's do it.
H
Host0:09
You have the privilege. There we go. And off we go. Recently, we achieved our latest release. Tell us a little bit about what that release is about and it actually enabled us to go from where we are to all the way up there.
T
Tim Klay0:21
In our new release, we've more than quadrupled the size of the geofence. It contains more than half of all ride-hailing trips that start and end in San Francisco, which is super cool. In Las Vegas, we significantly expanded where we can drive. We added tons of major pickup and drop-off locations, major hotels like Aria, Bellagio, we added the convention center, we added the Sphere, and we're actually adding the airport as well in the next few weeks. And then for the first time, we now have our robotaxis driving in Austin and in Miami.
H
Host0:49
Let's talk a little bit about the technology, right? Like what step did we do to enable that big of an expansion?
T
Tim Klay0:59
So there were a few things about these new geofences that we had to solve in order to comfortably and safely go that far. In San Francisco, it's things like steep hills. We got much better at assertive driving. We have traffic-aware routing now. The ride is just a lot smoother. And that's everything from the low-level braking control and sort of vehicle dynamics all the way up to the higher-level AI stack. So better perception, better prediction. And so you're able to be safer and smoother at the same time, which is kind of hard to pull off.
H
Host1:29
Tell us a little bit about the operation in the different cities, right? Like why is it copy and paste to go from city to city? What is similar from city to city? What is pretty different?
T
Tim Klay1:40
Yeah. So the cool thing is we run the exact same software in all of our different cities. So we don't have to build a separate Las Vegas AI and a San Francisco AI. Our AI is smart and generalized. It knows how to drive everywhere. Having said that, there are some differences between the cities, especially operationally. Las Vegas is unique because it's really all about these major destinations with their pickup and drop-off. You have the ride-hail areas, you have the main entrances. How do you pick up and drop off passengers in these super complicated, very dense pickup and drop-off areas? Which you don't have as much of in most other cities. Las Vegas also has higher speed roads. So, the first time we ever did 45 mph in the robotaxi, that was in Las Vegas.
H
Host2:21
Give me an overview of our approach to expansion and scaling.
T
Tim Klay2:26
Yeah. So the cool thing is there's only so many different types of topologies that exist and as you kind of cross off more road features and intersection types, then you actually do automatically get an exponentially large area that you can cover. And so at the beginning we really focused on, okay, we built a ground-up robotaxi, nobody's ever done that before. How do we build and validate a safety case even in a very simple environment? Once we did that then we can say, okay, we have our test vehicles driving much more complicated large areas. How do we get that to a level of safety that we can put it on the robotaxi that doesn't have any manual controls? So, you can't take over. There's no steering wheel. There's no pedals.
H
Host3:03
How does it feel to now sit here and experience that?
T
Tim Klay3:07
It's pretty special. I mean, it is true. When we started Zoox in 2014, if you had told me it would take more than a decade to pull this off, I would have actually assumed that it was just not going to happen.
H
Host3:21
Yes, because you know that's a long time for a startup to be patient and the employees have to be patient and you have to be able to fund the company for that long. So I probably would not have thought it was very likely that it would take that long and that we would make it. It does make it all the more rewarding I think to have gone on that entire journey. I've been working on self-driving cars I think similar to you for just like two decades now.
T
Tim Klay3:47
Which is kind of amazing, right? And so, you know, you make progress every year, but you sort of wonder like, is it really all going to come together so like regular people can just actually take rides with no driver all throughout a city as complicated as San Francisco? And the fact that we're doing that in a purpose-built robotaxi, right? This is just mind-blowing. Just really proud of, you know, we did it. It's safe. And it's really comfortable and just like this experience, right? We're sitting, we can see each other, we're having a conversation. Having a conversation with the person sitting next to you in a car is like not a great experience. So, here we can just hang out. We can see each other. We can talk. You can put three people, you can put four people in the vehicle. And nobody's ever done that before. The great thing is now we know that it's possible and we know that we can do it because we have gotten to a level of safety and geofence that it's starting to be useful.
Now, this is a fun scenario. So, there's a little bit of a construction situation going. There's a flagger. He's saying we can go slow and we're going, but there's actually somebody blocking us. They're getting out of the way. Yeah. This is smooth. There wasn't even the slightest jerk or brake tap. Like, that was like butter. And that's a tricky situation. We always said, right, like once we get to our first paying customer, it'll only be like the invitation for a whole new set of problems that we then need to take care of. And I think we're all excited to knock on that door and take care of those challenges.
H
Host5:21
When we look at the Zoox journey, one of the criticisms that some people had about Zoox, oh, you know, Zoox is like hitting three home runs. We have to solve autonomous driving.
T
Tim Klay5:30
Yep. We have to build the ground-up robotaxi and be able to, you know, manufacture it. It has to be reliable. And then we have to build a kind of a ride-hailing, you know, network with the app, the dispatch, routing, customers, all that. And so, to be fair, that was a pretty compelling argument, right? It's like, okay, that does sound like three really hard things to do, but it's not impossible.
H
Host5:48
But it's not impossible.
T
Tim Klay5:49
And our view was also that they actually reinforced each other, right? The vehicle that we've built actually makes it easier to solve autonomy because we put the sensors in exactly the right locations. We built more redundancy, more safety, four-wheel steering. There's things we can do that a retrofitted car can't. So, it took longer, took more money, but we actually end up with a better, safer, more comfortable, more enjoyable experience that actually will have better unit economics. Now that we've started to solve these pieces, and you can see the result, it really is just super exciting because we're really close to being a big part of, I think, a lot of our customers' daily lives. We're just going around a Waymo. This is fun. So doubly parked Waymo dropping somebody off as it should and then we smoothly go around as we should. So friendly robotaxis coexisting in San Francisco.
H
Host6:38
Who would have thought, right? As we are about to arrive.
T
Tim Klay6:43
This is beautiful water.
H
Host6:45
My mind is blown, I cannot comprehend that this is actually happening. It's my first new release ride all the way up to here. And now here we are, man. I would like to have hot chocolate. How are you?
T
Tim Klay7:03
Sounds amazing.
H
Host7:03
Look at this beautiful Golden Gate Bridge.
T
Tim Klay7:06
And the Zoox.
H
Host7:08
All right. So, should we head back?
T
Tim Klay7:10
Let's do it.
H
Host7:12
It fits in so well. What is one of the common misconceptions about Zoox? Let's start there.
T
Tim Klay7:20
When we started in Las Vegas, which was the first time we were open to the public, we started with just a handful of major destinations and you could use our app to hail between them. Some people thought, oh, you know, Zoox is like a shuttle. Of course, that's not the case. In San Francisco, you can pretty much go anywhere within our geofence. In fact, in our new geofence, we have over 5,000 pickup and drop-off locations that we're able to go to. These are actually like zones, too. They're not just like a specific point. So, yeah, we do true point-to-point in San Francisco, in Austin, in Miami.
H
Host7:55
And then tell us a little bit about the purpose-built vehicle as a solution, right? Like what advantages do you see and why does it make Zoox stand apart?
T
Tim Klay8:06
Yeah. Well, we're still the only one. I keep waiting for other companies to kind of get with the program and make their own purpose-built robotaxis. We know that's going to happen eventually, but so far everybody else basically started with a car. Some companies are talking about maybe we remove the steering wheel, maybe we remove the pedals. We don't see anybody truly building a new fundamental platform where every aspect of it was designed just to be a robotaxi.
We see a lot of advantages. Some of them are very obvious to the customer like we get to sit and face each other, right? That's not happening in a regular car. The form factor, our vehicle is very short. This is great for dense urban mobility. So, it's about a meter shorter than a Toyota Corolla, which is not a very big car to begin with. So, super compact, but look how much room we have on the inside, right? And then there's some really great advantages for AI and for safety and for redundancy. So, we have our sensor pods in the corners. Each sensor pod can see 270 degrees, so you get overlapping field of view. If a sensor fails for any reason, that's okay. We can still finish the mission because we still have enough coverage.
And then there's just a lot of redundancy. Like we have two steering racks, we have two motors, we have two batteries, we have two braking systems. And so anything can fail. Yeah, they don't fail very often. These are automotive grade components. They're very reliable. And then from an economic perspective, we realized, you know, it's a robotaxi. We want these driving all day, all night long. So we put a really big battery. Zoox's robotaxis has a 133 kWh battery. So, we can really drive a long time on a single charge. Otherwise, you're having to drive back and forth to your charging and you're having to charge the battery too many times, then your battery wears out and you have to replace a whole battery after a couple years, which is very bad environmentally, very bad economically. So, by thinking from first principles and starting from the ground up, we just built a much better solution to moving people around cities.
My favorite thing is in a regular car, even steer by wire, you know, humans driving it, you don't need crazy precision, right? Well, we're a robot, so we have AI controlling it. We can do way better than a human. So, we have this special firmware on the steering rack that lets us control the exact position much more accurately than in a regular car. And then we have two of them. And so by controlling each axle independently with this super precise controller, we are able to get a level of trajectory tracking that is about an order of magnitude better than anything you can ever do on a retrofitted car. On a regular car using the fanciest best algorithms just due to the inherent limitations of the hardware, typically the best you can do is around 5 to 10 cm, which is not bad, right? That's still better than a human.
H
Host10:53
So it's not like that's not decent. We get 5 mm to 1 cm in accuracy. Wow.
T
Tim Klay11:00
So that means that on average when we're driving, we are usually much less than a centimeter away from where we wanted to be.
H
Host11:08
Yeah. That is insane.
T
Tim Klay11:10
And again, do you need that all the time? Usually not. But if you are in a very difficult situation where you really need to precisely control the vehicle either because you're going through a very narrow gap or there's some emergency situation and you really need to precisely control your trajectory. It's amazing to be that precise and you just cannot do that on a regular car.
H
Host11:29
Oh, we have an emergency vehicle.
T
Tim Klay11:31
Oh yeah, I hear it.
H
Host11:32
Yeah, there's an ambulance with its lights going the other way.
T
Tim Klay11:34
Ambulance. We're at a red light anyway, so we don't have to do very much right now. But the cool thing is we're very good at detecting emergency vehicles. We use the cameras and we use also eight microphones around the vehicle so we can triangulate where it's coming from and we know if we have to pull over. Sometimes we can even do that faster than a human can, which is pretty cool.
H
Host11:53
Tell us a little bit about how did we achieve the more assertive driving.
T
Tim Klay11:59
Yeah. Well, I'll let our viewers in on a little bit of inside Zoox. The previous release, we had a target geofence and we were driving it with the test vehicles and driving it in simulation and it was all looking good. We were passing all of our safety metrics. So, we're like, okay. But as we were getting closer to the release, there was one part of that geofence that when we would go drive it, we would find that sometimes it was so busy, we would sometimes struggle to make a lane change that we had to make. We would get kind of stuck because we couldn't shove our way in to these narrow gaps and we were too polite essentially. And as a human driver you would just be a little bit more assertive. And so we realized we weren't quite ready for that part of the geofence. And so we chopped it off and we said, 'Okay, we don't drive there yet.'
But in this release, we had to drive that and way way way more. So we're like, 'We can't mess around with this assertive driving anymore.' But it's a hard problem, right? Because if you're being more assertive, then you still have to make sure you're being really, really safe. And it might feel like these are a little bit contradictory. You also have all of our validation pipelines. And it's easier if you just have one way of driving that you drive all the time, right?
And we agreed that we could be a little more assertive. But then we said, you know, you don't want to be really assertive all of the time because that could be a little bit obnoxious, right? Like under normal circumstances, you want to prefer to be more polite, let people in and so on. But if you're always that way, sometimes you could get stuck. So, we realized, we kind of bit the bullet and we're like, you know what? We are going to make our vehicle a little more assertive in general, but we also need a special mode where it can essentially decide to be more assertive than normal and that will allow it to make progress, still be safe, but not be kind of annoying in the normal situation where you don't have to be quite so assertive. And so this was like a little bit of a mindset shift for not only the motion planning team but also our safety teams because we have to validate no matter which mode we're in we have to make sure it's safe. But there could be a bunch of cues in the scene using AI that tell us how assertive do we want to be. So that is in this version now and it's really nice.
So that's great as the baseline. But then if you're in a situation, I'll give you an example. Let's say we're trying to make an unprotected right turn. And you have people coming from the left, right? So normally you don't want to cut anybody off, right? You want to look for a nice big gap. And so we do that. But eventually if there isn't a nice big gap and it's just a stream of traffic, you don't want to sit there for the rest of your life, that doesn't work either. So, we kind of have this concept that as we're getting a little bit more impatient, in a constructive way, we have to be willing to take a smaller gap. And that means maybe a little bit higher acceleration, higher lateral acceleration while still being safe, but you don't leave as much of a buffer as you would normally do under normal circumstances. And we had to then validate that behavior, including all of our safety pipelines. And what's really nice is in this release, we've significantly improved our overall safety even though we're driving in much harder, denser geofences. So the team's done a really nice job of finding that balance.
H
Host15:14
Yeah. Giving the robot some confidence, right?
T
Tim Klay15:17
Yes. Yeah. And I think even though the previous release had some initial versions of that, like when you're in a junction and the traffic light turns yellow, we did start on it a little bit before in some more limited situations. But for example, when you go through an intersection, it's interesting because you can read the California driver handbook or whatever, and there's all these rules, but sometimes you actually can't follow every single rule at the same time, right? So, as an example, you do not want to enter an intersection if you don't think you can clear it because you're blocking the junction. But if it's super super crowded, sometimes you have to, right? And otherwise you're just sitting there for 3 hours. It doesn't work. And so there's kind of an expectation that eventually you're going to go in there. And so we did in the previous release work on that sort of dense junction handling. We got much better at that. And then in this version, taking it to the next level and also things like lane changes and unpredicted turns, that kind of stuff.
H
Host16:21
Yeah, that is also tricky. Let's talk a little bit about the machine learning side, foundation model side, right? What are your favorite advancements there in our latest release on that end?
T
Tim Klay16:31
Yeah, it's an amazing time in AI in general, right? When we started Zoox, there was deep learning. It was still pretty new. Obviously, things have advanced way beyond that. You can now give algorithms different types of multimodal sensor data and they kind of just figure out how to use it, which is really cool. One of the things that we're doing at Zoox is we're not going all the way to end-to-end, sensors to driving, for several reasons. It's not good enough yet to be safer than humans. It's also not very easy to understand always why did it do what it did. It's also if you want to change the behavior can be a little bit tricky. But at the same time, these are very powerful techniques. So at Zoox, we're saying how do we get the best of both worlds.
So, one of the things we're doing in this release, first of all, there's just a lot more machine learning in the prediction and planning. That's really cool. It's giving us more natural smooth trajectories. You want to make sure that the machine learning doesn't go do something that's too creative or not safe, right? So, we did a nice job with that. But, there's also something we're running now. We call it Scene IQ. This is kind of like a vision language action model that's running a little bit in the background. And what's cool about it is it's not directly driving the vehicle, but it's essentially giving some hints to the AI stack that is driving the vehicle. It's kind of running in the background and it's giving some hints to the planner about the scene.
So, as an example, this VLA can basically suggest maybe we want to slow down a little bit. And what's nice about that is if you have a false positive every few thousand miles, it's not so bad, right? The vehicle will slow down a little bit and maybe it didn't have to. So you get to your destination 5 seconds slower, 1% of the time, not the end of the world. But if there is something really weird going on, you might be very happy that you proactively slowed down a little bit. And it's having pretty good precision and recall rates already, meaning it doesn't trigger very often when there wasn't a good reason to. But when there is a good reason to, it's turning out to be pretty helpful. And we're working on more sophisticated versions that can even help us get some hints about sort of laterally where we might want to drive. And so what's cool about this approach is that over time these models will get larger and better and we'll start trusting them more and more.
H
Host18:47
But you don't have to completely hand over the reins and just sort of trust, you know, like vibe driving, right? You know, you say vibe coding, like this vibe driving thing. It is amazing most of the time, but the problem is most of the time is not good enough for a safety critical system. So I think we've come up with some really creative and clever ways to combine the best of both worlds.
T
Tim Klay19:06
Also at Zoox on our research teams, and at Zoox we don't really have segregated research teams over here who only do research. We rather try to have our really talented engineers, many of whom have a research background, work on projects that could pay off in one or two or three years but we don't need them for this next milestone necessarily. So, we're doing some incredible new work on world models, really foundational driving models. We have double digit millions of miles of driving data now to train these models on. And also some super cool work tying that with simulation.
So, imagine, you take a log file that we've driven before and you want to say, well, let's make it nighttime. Let's make it pouring rain. Let's remove that car. Let's add a bicycle and actually be able to simulate the raw sensor data that you would get in that new scenario or even imagine whole new scenarios we've never seen before. So we have some really great cross-functional projects between simulation and AI teams that are working on this. It's also great for validation of corner cases as well as you can even train models for scenarios that are very rare or objects that are very rare. You don't have enough real world examples but you still want to be able to handle them.
So if you're interested, anybody watching this, if you're interested in foundation models, world models, generative AI, next-gen 3D simulation, sim to real, any of these topics, we have some amazing people working on this and some of the best data pipelines and infrastructure to train these large models. And because Amazon is our parent company, we have a wonderful relationship with AWS. So we have access to the biggest latest compute to train these models on. So it's a really cool place to do cutting-edge research and then apply that research to a product that everyday people can use and enjoy. So if you're into that stuff, please come talk to us. We're really investing heavily in it.
H
Host21:09
Why do you think Zoox is in a unique and special position to build a foundation model that fundamentally understands driving but also can generatively create scenes?
T
Tim Klay21:21
Well, it's interesting. You know, one of the big buzzwords recently is physical AI. And I was thinking about it. I was like, hm, it feels like in 2014, we started a company to build a robot to run AI to move people around. In some ways, we might be the largest physical AI company. At least in terms of a company that was really started just to do that.
And so we have incredible data, right? We have data and it's not just camera data. Camera is great, but we have the camera, radar, LiDAR data from all over so many cities, so many scenarios, so many corner cases. And the number of miles we're driving is really growing exponentially. We have 100 robotaxis driving around now, but that's going to be thousands by next year, literally, because we built our factory. We will by the end of the year be building three robotaxis an hour. So every 20 minutes a new robotaxi will come out of our factory. So we will have thousands of these things in multiple cities commercially next year. And so that's just a profound amount of data. Now of course it's not the biggest fleet in the world. I mean there are car companies who have even bigger fleets, but they don't have the richness of the data, right? They don't have all the radar. They don't have the LiDAR. Sometimes you see them driving around with these giant LiDAR rigs to get certain validation data, but all of our vehicles have that. So we have an incredibly rich data source. And then as I mentioned, we have a great partnership with AWS.
H
Host22:53
Oh, there's another Zoox.
T
Tim Klay22:54
Oh, hey.
H
Host22:55
Hi Zoox.
T
Tim Klay22:57
You know, one of the secrets about training large models is there's a lot of sausage making behind the scenes to be able to reliably train these models and keep these jobs running and have metrics and KPIs and dashboards. And so we have an awesome software infrastructure team. A couple years ago, we were like, oh, big model would be a couple billion parameters. Right now, we're training 32 up to 100 billion parameter models. Now, these aren't going to run on the vehicle. We're not going to be running a 100 billion parameter model real-time on the vehicle anytime soon, but for offline, you can use them to generate scenarios. You can use them to triage scenarios. So, we used to have to have a lot of people hand-looking at real and simulated scenarios. That's not very scalable. It's also really slow.
I mean, you can talk about this because you run the autonomy team. So, how is it like for your team, you know, over the years, one of the frustrations in the early days is like, hey, I have a new idea for my new motion planner or I want to change this parameter, this algorithm, and it's like, oh, it takes so long to find out if it's actually better. And so, maybe you can tell us a little bit about how is that changing the developer experience?
H
Host24:11
Yeah, absolutely. And I think on multiple levels AI is helping us. So on the one hand, developing the foundation model that fundamentally understands driving, not just based purely on camera data or on video, but also how our robot sees the scenario and with that the interpretation capabilities, you can make heavy use in your data loop of how you get better and better, speeding that up, helping out in finding scenarios. Hey, I have a new idea and now I want to deploy that on relevant scenarios and get answers, interpret the results already. But also in terms of AI tooling, it is amazing what is happening out there and how that accelerates developers as well. At Zoox, you can point your AI agent to a problem ticket and say, hey, go figure out what the root cause is. And that thing will go off and it'll come back and more often than not it gives you something really insightful.
T
Tim Klay25:17
And more often than not it gives you something really insightful, right? Or it asks for like, hey, can you help me, give me a direction? Should I go this way or that way? So you say go this way. That is amazing to see. And I think we invest a lot into making that AI agent be capable to tie all the different cues together because it's not only enough that you would have access to the problem catalog, you need to have access to the source code, you need to have access to how can I reimulate that thing, how can I read the reimulation data, what's going on on Slack, like are people talking about some stuff. Pulling that all together is very powerful and that really accelerates our development.
A few years ago sometimes some of our software engineers would be a little bit frustrated, they're like, hey, I love Zoox but it takes a long time to develop and to iterate. And that is now more one of the things people are complimenting us on rather than complaining about. We can always keep making it better and better but that's been really rewarding to see.
H
Host26:28
You mentioned like the word end-to-end driving, software from sensor to actuator. Those models are not good enough yet. What do you think about if you say 'yet' like we'll get there at some point, right? So that path?
T
Tim Klay26:42
Yeah, I mean, look, I think if you examine the progress of these generally intelligent models and you look at the frontier models, you look at Gemini, you look at Claude, you look at GPT, every few months they're getting meaningfully smarter at a variety of tasks, which is super cool. So, those same techniques do apply to driving. It's starting to show some pretty incredibly insightful behavior in complicated scenes. On the other hand, I think we've all experienced asking AI something and we're like, is that really good? And then you tell the AI, you're like, are you sure? And then it's like, oh, no, you're right. You're absolutely right. I was totally full of it. So that's not, you don't want that driving your car, right? You don't want it to be like 99% of the time it's awesome and then 1% of the time it does something just egregious.
These models, again, they're amazing. They are not yet superhuman performance, even if you give them multimodal sensor data which does help. But they're not yet good enough to do all of the driving all of the time. Will they ever be? I do think so. But I also think that for a very long time the best systems will be hybrid systems. So you're going to have these kind of end-to-end very large models, although remember to run them in real-time they can't be as large as, you know, GPT 5.5 Pro people are saying this is probably like a five trillion parameter model. People are saying Claude Mythos is possibly like a 10 trillion parameter model. You're not running that on a robotaxi in real-time anytime soon. And so what you can distill and run on a vehicle is not going to be quite as generally intelligent.
And when you're running on a robotaxi, you want to be making decisions within a couple hundred milliseconds. Sometimes you ask, like I have the ChatGPT Pro account, it's really amazing. Sometimes you ask it a hard question, it thinks for half an hour, right? That's not very helpful in a self-driving car. So I think that you will absolutely see more and more of these techniques applied to the on-vehicle software, but there's also something nice about the explicit representation of objects in 3D space and the idea that you have some software running that's explicitly doing collision checking as opposed to, oh, I've trained on however many examples. And so I think you're going to have hybrid architectures for a very long time. I just think that you're going to start biasing it more and more towards these larger, more generally intelligent models, but I think we're very far away from removing all of the guardrails. Like very very very far away from that, which is fine. Like the guardrails are good, right? And you can kind of get the best of both worlds, which is what we're starting to do in this release. But in future releases, we will make more and more use of these end-to-end type models with again still very very strong safety guarantees.
H
Host29:47
Yeah, that makes sense. And before we arrive, let's talk about what's next, right? Like what's ahead of us?
T
Tim Klay29:54
Oh man, it's so exciting. You know, it's funny at Zoox, we work so hard on our next release, right? And then it's out and we're like, 'This is awesome.' And then like a couple weeks later, we're ready for the next one. Okay, the next one from a deployment perspective is just amazing because for the first time we will be able to drive all of San Francisco and that's been a dream, right? That's been a dream forever. And it will also enable our production robotaxis. So as we discussed we have just over a hundred of these driving around mostly in San Francisco and Las Vegas, a little bit in Austin and Miami and even a couple in Foster City where our headquarters are. But we have way too much demand already and we know that's just going to increase. So we will be making these thousands of robotaxis. Our next release will support that new hardware version.
It also has some great new AI advancements handling things like very narrow roads. Single-family residential, getting really good at that type of driving, a little bit more on maybe like stacked roads. I mean we do stacked roads now but there's certain very complex topologies that we haven't done yet. Getting it better at handling parking lots is coming even in the release after that. Then we get like basically can handle just about any unstructured parking lot. Continued improvements to comfort and safety is sort of driving all of these things. More environmental conditions. So this release, we didn't talk about it yet, but this got much better at handling medium and heavy rain. But you know there's always degrees of rain. So like handling even more extreme rain or handling heavy fog. So that's also things that we're working on for future releases. And then new cities. So we just deployed this version in Austin and Miami, but a relatively small geofence, especially the Miami one is very small. Have big geofence expansions in all of our cities and then we'll also be finally putting the robotaxis in Atlanta and Los Angeles.
Which is going to be amazing. Not for customers in the first version but then in the future version starting to open up to customers. So it's going to be an amazing rest of the year where we really go commercial, we can start charging in multiple markets and really making Zoox a part of people's daily lives.
H
Host32:06
Yeah, so what's the biggest challenge you see for our next release that we have to overcome?
T
Tim Klay32:12
Well, as we start going from 100 robots to thousands, you continually expand the safety and the comfort and the reliability. So we are already many many times safer than a human but we're never satisfied. We really just, we don't see a particularly, oh we have to get this many times and then we're done. We have some really ambitious safety goals for next release. We want to continue to make the ride even more comfortable. Although I have to say we're now at the point where it's pretty damn comfortable. But we know we can keep making it even better.
The reliability is an important one. So, we've made so much progress on that over the last couple years. When we first started driving on public roads, we would drive in our little Foster City kind of just from one office to another. You know, the vehicle could get stuck every few hundred trips and it wasn't the end of the world. That's not working anymore. So, we've made a couple of orders of magnitude improvement in that. So, now we're talking about this thing is only getting stuck in the many, many tens of thousands of miles. But we want to take that to hundreds of thousands as we scale the fleet to thousands of robotaxis. So, I don't know if that sounds easy or hard, but it turns out it's really hard, because there's hundreds of things that could cause a robotaxi to get stuck, and you have to make sure that almost none of them does. And because we have this custom hardware that is very redundant and in theory very fail operational, we have the opportunity to be the best in the world at not getting stuck.
H
Host33:41
And look at that. Right in time. Let's open those doors. Thank you for all the watchers and the listeners. Excited to see you for the next one.