Orchestrate all the Things - Wayve is working with Microsoft to tackle end-to-end deep learning-based autonomous driving. Featuring Wayve CEO / Co-founder Alex Kendall

Episode Date: May 18, 2022

Circa 2017, there was a lot of hype around autonomous driving. If one were to take that at face value, it would mean that by now autonomous driving would have been a reality already. Apparently that's not the case, and Alex Kendall claims to have known that all along. Still, that did not stop him from setting out then, and he's still working on it today. Kendall is the co-founder and CEO of Wayve, a company founded in 2017 to tackle the challenge of autonomous driving based on a deep learning approach. Today, Wayve announced a partnership with Microsoft to leverage the supercomputing infrastructure needed to support the development of AI-based models for autonomous vehicles on a global scale. We caught up with Kendall to discuss Wayve's philosophy and approach, its current status, as well as where the partnership with Microsoft fits in. Spoiler alert: this is not your typical "commercial application provider partners with cloud vendor" story. Article published on VentureBeat

Welcome to the Orchestrate All The Things podcast. I'm George Amadiotis and we'll be connecting the dots together. Around 2017, there was a lot of hype around autonomous driving. If one were to take that at face value, it would mean that by now, autonomous driving would have been a reality already. Apparently, that's not the case and Alex Kendall claims to have known that all along. Still, that did not stop him from setting out then, and he's still working on it today. Kendall is the co-founder and CEO of Wave, a company founded in 2017 in London, to tackle

the challenge of autonomous driving based on a deep learning approach. Today, Wave announced a partnership with Microsoft to leverage the supercomputing infrastructure needed to support the development of AI-based models for autonomous vehicles on a global scale. We caught up with Kendall to discuss Wave's philosophy and approach, its current status, as well as where the partnership with Microsoft fits in. Spoiler alert, this is not your typical commercial application provider partners with cloud vendor story. I hope you will enjoy the podcast.

If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook. So my name is Alex Kendall, I co-founded and I'm CEO at Wave. I'm really passionate about building intelligent machines that can really add a lot of value to our lives. And this, for me, really involves building embodied intelligence, co-designing the hardware and software to build systems that have the ability to reason in complex environments. And I think there's no better place to start than autonomous driving.

Autonomous driving is going to be the first widespread example of intelligent machines that really transforms the cities we live in. When you think about the way we move people and goods around, I think it's going to be truly exciting and usher in the age of autonomy. It's something that at Wave we've been working on for five years. We founded the company in 2017. And if you think back to 2017, that was a time when there was a lot of, you know, a lot of announcements and media cries that autonomous driving was here. It was just a few years away. And a lot of the commercial efforts were really pushing forward autonomous driving with what you might consider classical robotics technology, what we think of as AV 1.0 technology. At that time, we started to see a quiet revolution

happening in machine learning. We started to see the adoption of end-to-end deep learning methods start to really transform what was possible to be able to do with really challenging high dimensional dataset problems. We saw the solution of very complex challenges in natural language processing and protein structure prediction in the games of Go and other really difficult artificial intelligence tasks. And for us, autonomous driving was going to be no different. The challenge, though, is to take this technology out of the virtual world, out of simulation and put it on a real life physical robot. So that's what we started to do. We started in a house in Cambridge and started prototyping our robot in the garage and managed to, for the very first time, teach a car to drive around the block using reinforcement learning. This was, I think, hugely exciting for us seeing the system learn in front of our eyes, to be able to lane follow, to be able to navigate

around its environment. And that's really spurned us on. And since then, we've simply scaled this technology up and really started to build it out at fleet scale. Okay, great. Thanks for the introduction. And, well, you know, I have a number of questions lined up for you, both in terms of your technical underpinning and direction, but also in terms of your business trajectory and growth. So I think we may as well start with the latter just because looking at your timeline I saw that you recently received a pretty good amount of funding. So if you'd like to just expand a little bit on your co-founder team and well the business side of things and how that's how that has developed over the years and if you can also share a little bit of metrics like head count and you know that kind of thing

yeah it's been a story of i think uh fairly rapid but disciplined growth um we have uh we've really set out to build a full stack self-driving company but, but one where we really think about things from a machine learning perspective and build directly for what we think of AV 2.0, the next generation approach to autonomy, systems that we think will be the first to get to deployed in 100 cities. And so this has required us to build out all of the different components that you might need for an autonomous driving stack. We don't look to manufacture the vehicles. We don't look to operate them at scale, but we really build this embodied intelligence technology. So the compute, the sensors, the software, the simulation, the machine learning infrastructure, the data infrastructure, this is all core to what we do.

And we've built up, I think, one of the premier teams in the world to really go and pioneer this. Today, we're just over 150 people. We've raised $260 million. And we're headquartered in London with a small office over in the San Francisco Bay Area as well. So it's an exciting time for us. I feel like we, in many ways, have all of the ingredients now to go and build the system at scale. We've got the capital, the team, the data resources, and now also the compute with our new partnership with Microsoft as well. And so this is really all the ingredients that you need

to go and build large-scale embodied AI system. For us, now we're starting to look at some of the incredibly challenging machine learning problems around how do you actually go and build how do you go and build a a interpretable system one that we can really understand what it's doing and make sure that it's robust so that society can really trust it when it's deployed at scale you asked a little bit more about our business model. And so we really see this technology as being valuable and being able to be deployed with fleets, commercial fleets. And so we've partnered with three of the largest commercial fleets in the UK, Ocado, Asda and

DPD. And these fleets for us represent a fantastic way to see autonomous driving deployed at scale. These fleets are technology-enabled. They are world-class at what they do in operations. And for us, seeing our driver deployed as a service to be able to provide them with improved safety, increased asset utilization, improved economics of mobility, this is really where we see autonomous driving going at scale.

Okay, great. You already touched upon something that I intended to ask you as a follow-up question. So, and to be honest with you, I don't think I've had a conversation with anyone working in the autonomous vehicle space before, but well, I'm generally, let's say, familiar with, reasonably familiar with the domain. So looking up a little bit on the company's background in what you do, there was one thing that attracted my attention, which you already sort of alluded to the fact that it looks to me like you have a specific focus, a specific direction which is focused around, well, commercial fleets, as you put it. And that has, the way I saw it, it has a two-fold repercussion.

Let's say on the commercial side of things, as you already mentioned, it lets you focus your efforts on a specific area of applications, which kind of makes sense. But I think that it also has an effect on the technical side of things, because that means that you're limiting the domain of application and also the challenges in a way. I think most other companies that I've seen that are focused on autonomous vehicles, they kind of try to tackle the broad issue, let's say. So tackling the issue of controlling vehicles in many different situations, such as in highways, in urban environments and so on. While what you do allows you to sort of minimize, let's say, the surface of the problem you have to tackle. And so in a way, I guess that makes it easier. Is this why you chose? First of all, I wonder if my impression is correct and if indeed that is the

area that you target. And then if the interpretation is what went through your minds as well and when deciding to take this direction. Actually, commercial fleets have extremely large coverage of our world. For example, our partner DPD covers 96% of UK roads every month. So the coverage is enormous. In fact, the reason why we work with commercial fleets is really because it lets us move faster. It lets us scale our deployment and get access to larger amounts of experience and training data than we could otherwise. So for example, we have our autonomous

sensor stack and our data collection devices deployed on the manually driven fleets of our partners to be able to get the experience and understand the edge cases that we need to be able to learn to drive. And we see some really weird and wonderful things on the road. And to be able to have this experience is what's necessary to build a safe and robust representation of the world for driving. For example, we've seen some some really interesting uh what you might consider edge cases for example uh a truck carrying a bunch of um uh roadwork temporary traffic lights uh there are things like this you have to actually understand that those aren't traffic lights uh but they're actually just being transported or for example and thankfully no one

was hurt but we saw a motorcyclist fall off a motorbike, and that motorbike continued like a torpedo straight into one of our partner's vehicles. And so you can see all of these kind of edge cases that you would have to drive for many, many thousands of hours of driving to start to experience that we can get through our fleet partnerships. So having access to that is the first thing. The second thing is being able to really gain leverage in our ability to deploy at scale. Our partners, their expertise is in fleet operations. Our expertise is in building embodied intelligence. And to be able to bring these

together produces a result that's greater than the sum of the parts and so we see this as really being a way to focus on our expertise to build faster and ultimately achieve scale faster with our partners um and i guess you know finally we see uh shared mobility and fleet uh and commercial fleet ownership as a much richer way for the future to get the benefits of mobility at scale, moving away from private car ownership and having more safe and sustainable transportation. For example, operating autonomous vehicles

are really safety-credible task. And I think the expertise that commercial fleets have in maintenance and in operations of vehicles will mean that these vehicles are operated to a high degree of safety. Perhaps leaving that in the hands of consumers from the early nascent days of autonomous driving, we think might be a harder route to seeing autonomy deployed at scale initially and so that's why we pursue working with fleets. Okay. I think that that makes sense in a way. I see an analogy with the way, for example,

innovation in software is diffused as well. So usually you have initially the enterprise technology applications, and then this may trickle down to consumer-facing applications as well. So I guess it's the same principle in your case I want to follow up question I had there is do your collaboration with does your collaboration with commercial fleets also include data collection so I guess you basically equip at least a part of the fleets that you are collaborate with with the equipment that lets you collect data. That's right. So we have a data collection collaboration

with all the fleets that we work with. And, you know, I think if you think about what it takes to build and pioneer a fantastic, really boundary pushing artificial intelligence, if you think about all of these different breakthroughs we've seen over the last five years, some of them have been extraordinary and they all come from the same recipe

of having the data, the compute, and the right people to be able to push this forward. And so for us, positioning our team in a way that we have access to extraordinary amounts of driving data to be able to learn representations of the world that are safe and robust at scale. They have the right compute partnerships that we and the strategic partnership we now

have in place with Microsoft that gives us access to world class levels of computation. And the culture that we've built up um to really bring together a premier team in the world at the intersection of machine learning and robotics i think these are the kind of ingredients that we've made strategic bets on that we believe will produce scalable adaptable robust autonomous driving technology Well, I would like to return to the data collection and processing part and the architecture and all of the technical underpinnings a little bit later. But now, since the occasion for actually having this conversation is the fact that you're

about to announce collaboration with Microsoft, I would like to kind of shift gears and touch on that part as well. So let's talk about what exactly does this collaboration entail and how did it come about? Because I think there is a sort of ongoing relationship with Microsoft and I think it would be interesting to trace that back to its origin? Look, this is a fantastic opportunity for Wave and Microsoft. We've been partnered together for a couple of years now, since 2020, really working on scaling machine learning technology for autonomous driving with the commercial Azure cloud offering. Now, what we've seen is, as I was talking about earlier,

an absolute acceleration in our performance with more scale of training, more data, more compute, more parameters in our machine learning model. And this is really starting to push the boundaries of what is possible with the commercial cloud offering today or any commercial cloud offering today. If you think about the kind of training we have to do in autonomous driving, our cars from a raw data perspective collect in excess of a terabyte a minute. There's an extraordinary amount of data that comes through these systems. Now, if you think about a lot of the supercomputing technologies that are developed today, a lot of them are around large-scale text or natural language processing.

But moving from kilobytes of text data to petabytes or exabytes of video data is really what's required to make mobile robotics or autonomous driving work at scale with machine learning. So that's what Wave and Microsoft are setting out to build. And we believe that there's going to be a, you know, this kind of technology is going to be and create trillion dollar market opportunities through machine learning for autonomous vehicles and mobile robotics. I think video scale reinforcement learning is going to completely change many of these sorts of industries. I also think this represents a further signal that there is a growing shift in the way that we're thinking about autonomous driving, moving away from the classical robotics problem and shifting towards this AV 2.0, this machine learning thinking of solving the problem um through technology like deep learning and this is something that certainly microsoft is

recognizing with this partnership with us uh and so essentially uh you know this week we're really excited to be announcing that wave and microsoft are partnering together to build the super computing technologies to make this possible um to build the technology that lets us train, you know, peta or exabyte scale machine learning, machine learning technology for autonomous driving and for other other related fields. This infrastructure is going to really, you know, make a lot of what we are envisioning possible and we're really excited to see it implemented at the scale that Microsoft can unlock.

Microsoft also participated in our Series B, our $200 million Series B, we announced in January 2022. And so we're really excited to see a deeper strategic partnership here. And in particular, leveraging the expertise of both Wave and Microsoft to build such an innovative technology okay um actually that's that's a little bit more interesting than i thought it was if i

got it right i mean in the sense that i thought it was um a sort of typical uh let's say collaboration in which basically um you know some cloud provider in that case Microsoft sort of gives away access to their infrastructure to certain partners usually you know in particular in cases such as yours where said cloud vendor has some in some way invested in the technology partner that they work with them typically that comes in the in the form of of free credit for the cloud infrastructure and so on, but the way you framed it, it sounds to me like this may actually be something a little bit different than that. So you said something about building infrastructure to make that possible.

So does that mean that you're going to be collaborating with Microsoft on the technical level as well to develop new infrastructure that is suitable to your needs? Absolutely, the former you described is what the capacity we've been working with Microsoft and for the last couple of years. And this is really an announcement that,

we're hitting the boundaries of what is possible in the commercial cloud offering. And so Microsoft and Wave are really looking to join forces now to build out the necessary infrastructure, the super community infrastructure to make this possible. And this is going to involve some technology developments, some hardware and software innovation to produce compute and computing infrastructure

that lets us train at the scale that I've been describing. I think this is going to be an incredibly innovative and exciting journey for us to move on and perhaps the value of it is maybe unseen today, but in a couple of years time, I think this is going to be a core thing that underpins what makes AV 2.0 possible what makes it possible to train incredibly large-scale machine learning models that lead us our deployments of say you know ride hailing in London or grocery delivery in in Munich or public transport in Los Angeles or wherever you are in the world, bring together all the experience from our autonomous driving deployments, bring it into the cloud and to be able to train and learn so that we learn from the collective experience of all of our deployments. This is a

scale of machine learning and a scale of real-world embodied intelligence deployment that we haven't seen yet in the history of technology. And this kind of computational infrastructure is what's going to be making this possible. Microsoft sees this as a huge opportunity to unlock this market in this industry. And working together towards this is for both of us, I think, a very exciting next step. Yeah, yeah, indeed. I had, well, admittedly, I think, a very exciting next step. Yeah, indeed. I had admittedly underestimated what you were about to announce. I thought it was a more typical, let's say, scenario. But as you said as well, I think it sounds like it's also a good deal for Microsoft in the sense that they will get to test the limits of what's possible and work

closely with a use case that is sort of out of the ordinary in many ways. We see it as a really close collaboration and one that we're excited to continue working closely with Microsoft's team. So I think that's a good segue to talk a little bit in more detail about the tech side of things and again doing a little bit of background research one of the things that came out as a differentiator for the way that you do things is the fact that well you don't use a leader technology so you seem to rely exclusively on on video basically and visual cues. And I guess that also

that actually has many repercussions as well on the tech side of things. So that means that you don't need the expensive kind of equipment that that leader leads to and that also means that on the data processing side of things you have like a constant flow of well presumably high quality and high demand in terms of bandwidth data that you need to process and that also kind of leads to the need to have this super computing infrastructure that you talked about. Yeah, so to talk through our sensor strategy.

So, I mean, fundamentally going back to what I said at the start, we see building embodied intelligence as a joint hardware software problem. I think the wrong way to think about this is to think about, okay, what are the sensors that we need from a first principles perspective and let's design that in isolation because the intelligence you put behind a sensor hugely, you know, really defines what you're able to do with it. For example, the mantis shrimp is an animal that has the best eyes

in the animal kingdom, much better, you Much better visual dynamic range and resolution than, say, human eyes. But I think you'll agree with me that humans have a much richer and more advanced ability to perceive the world. And even though we have worse eyes and worse senses, our embodied intelligence, our joint hardware software, is much more capable of understanding our environment. So what I'm trying to make is that we need to build the most intelligent system. And at Wave, we really think about how do we build the most intelligent and scalable system that requires a lean set of hardware. It requires us to really focus on the intelligence first. And rather than when you have an edge case or a problem

putting in a different sensor to overcome it, we see that we primarily focus on making the system more intelligent first, if the signal is there in the sensor, and only when we hit fundamental limits will we think about expanding our sensing. So for us, this looks like a camera first approach. We think visual sensors have the richest information around the world, about the world around them.

And we also have a, so we also use radar in our system and radar gives us a complimentary redundant sensor that has different characteristics to the visual spectrum that say cameras or light hours have now a camera radar approach is great because both of these sensors are already manufactured at scale. They have the supply chain. Most modern vehicles have them in them today and for us really focusing on putting the right intelligence behind it is what will produce a safe and performance system that can be scaled okay i see and well since we're in many ways talking about uh things such as well visual

object detection and uh well collision detection and all of those things. I saw recently, and actually there's been a number of, there's been lots of heated debate, let's say, around those issues. And there's also been a number of cases in which critics, let's say, have identified flaws in existing systems. So the most recent of those that I saw was, well, someone who actually used to work with Tesla that took one Tesla for a test drive and kind of literally threw different objects at it

and recorded whether the other car was capable of identifying first and then driving around those objects. And while, you know, if you think about it, you would say that, well, in principle, it's the same kind of sequence that needs to happen. So you see something coming at you on the road, you have to swerve and then just drive around it.

But it turns out that depending on the type of object that this person tried to throw at the car, there was a different reaction. So some of the objects were identified properly and the car had the proper response. Some others that were sort of unexpected and put out of context in some way. So an example was like an office chair. So the guy just threw an office chair in the middle of the road and the car just didn't know what to do because my interpretation is that because, well, that kind of object was never in any training set. You don't normally expect to see an office chair in the middle of the road.

So the car had trouble identifying that. So how is your approach able to deal with that type of situation? So unexpected situations, edge cases is the term for those. When you think about building a representation of the world, something that you've learned to be able to drive with, it really comes down to how well you're able to generalize to be able to overcome edge cases is the first thing. And the second thing is about to understand what you don't know. So for example, if you've learned to drive and you might see some horses in your training data,

let's say you drive along and you see a zebra, you should be able to generalize from your horse experience of how a zebra should behave and, you know, and behave in a safe way around and navigate around that kind of situation. But if you then ran across a dragon, you know, that's completely different. And, you know, your system should be aware that it has never seen anything like this before and doesn't know how to handle it. You know, it doesn't know how to handle a flying fire breathing thing, but it might know how to generalize to a zebra. So maybe using that as a bit of a silly example, but the point I'm trying to make is that it's really important that you can generalize because you're never going to see the same thing you've seen in your training data before. You're never going to see that same exact

example. And so the ability for you to be able to see something new and understand it is incredibly important. How we've done that at Wave is we have trained our system to drive in London on passenger vehicles. We've then taken these vehicles to five other UK cities, cities like Liverpool, Manchester, Cambridge and in these cities we've driven there for the first time with no experience of their cities. It's gone to a new city and it's been able to take the experience of learning how to drive through traffic lights, roundabouts, roadworks in London and be able to demonstrate that behavior in a new city. That's an example of generalization. Another example is we've trained our system to drive on a passenger vehicle and we're now deploying our system on light commercial vans to be able to do last mile grocery delivery. Again

we're generalizing from one platform to another and doing something that can adapt. So adaptability is really important. The second thing I mentioned is around knowing what you don't know. This is actually a lot of the focus of my own PhD research, is building Bayesian deep learning networks, machine learning networks that are aware of what they don't know, aware of their uncertainty. So for example, if you do see a dragon or something you've never seen before, your neural network will understand that it's novel, that it's never seen this, and it is really uncertain about the scenario. In this instance, our autonomous vehicle will put itself in a minimal risk position. For example, it'll pull over and it will seek assistance and it will seek to be able to learn from that example with human supervision and so this kind of process really is what i think will let us scale but the most

important thing is you create the self-supervised fleet learning loop that with more data with more experience you continually re-optimize the system and this is the power of of machine learning if you think about a classical robotic stack where each edge case requires you to to tune your parameters and your thresholds and your rules, that becomes incredibly brittle at scale where everything becomes conflicting. With machine learning, machine learning excels at scale. The more data just lets us build more more powerful and rich representations that really let us generalize to the world. Yeah. You refer to that type of thinking and architecture as AV 2.0, as opposed to AV 1.0,

which was the kind of, well, more modular, let's say architecture that preceded what you're trying to do. I wonder if it would be possible to go a little bit into more detail into the components of the AV 2.0 architecture and have like a part-to-part, let's say, comparison with the previous one. And as a follow-up, I was also wondering if, in your opinion, Wave is the only one that's kind of following that approach. So the AV 1.0 system breaks the problem down into a number of sub-components like perception, traffic light detection, lane detection, mapping, control, prediction, motion

planning, all of these different components. Now each component may likely be solved with machine learning. For example, everyone builds a traffic light detection system today with convolutional neural networks. But the way that all these components are put together is largely hand-coded with set designed interfaces. And this makes the system very difficult to optimize globally, very difficult to put together and requires extraordinary amounts of people, time and capital. Now, to explain the AV2.0 system it's easy, there's only one component, it is an end-to-end neural network, a neural network that learns all the way from perception input to action output and this is very much, you know,

puts us in a paradigm where we can use machine learning to optimize this entire system and let data do the hard work. Use machine learning to uncover patterns in the data that are more complex than humans can hand code and hand engineer. This is what we believe will produce a system that can, well today we're showing it can drive in cities like London that are incredibly complex and do things that no other autonomous vehicle has shown today, like generalizing to a new city it never seen before like driving in an urban environment with just cameras um uh you know uh this this this is really what we think's the future this is technology that that we are entirely focused as a company is building and building i described um uh when we started at the very beginning of this conversation uh the first time where we put a reinforcement learning system on an autonomous vehicle and

showed that, you know, and that was a world first showing that we could learn to drive with reinforcement learning. And now we're scaling it up, really standing on the shoulders of giants and building a system with this technology at scale. And I think this is certainly something that, you know I'm really proud that of our team of pushing the boundaries here but fundamentally we have scale in mind and we're very early on that journey but we see this as being the first first technology that will reach 100 cities and we're really excited to continue to build towards that yeah my the impression I

get is that in a way you're trying to tackle some of the grand challenges of deep learning. So generalization obviously is the challenge, but I think earlier you also referred to explainability. And that sort of alludes to causation and the associated type of reasoning, which, as far as I know, was probably the two more challenging things to solve in deep learning. So if nothing else, you're not sort of ambition, but this is kind of the obstacles that you need to get out of the way to go where you want to go. So I'm wondering how close do you feel you are to actually making that happen and to kind of add some context to that

with the announcement that you're just about to make. How will working with Microsoft help you get where you want to go? So autonomous driving is one of the richest environments for really challenging and interesting research problems. And we do think of it like what you described, that we're tackling the grand challenges of autonomous driving. Things like generalizability, interpretability, how do we design rewards and optimization functions, all of these things. And you're right, they go beyond autonomous

driving and also just really push the boundaries of machine learning and embodied intelligence in general. I think it's an absolutely rich problem space for really, really stimulating and impactful challenges. We actually, my colleague Jeff wrote a, Jeff Hawke wrote a research, a positioning paper describing it was titled, it's titled Reimagining Autonomous Vehicle and talks about the seven grand challenges of autonomous driving. So if anyone wants to read that further, we really talk about in depth some of these challenges. So you mentioned a few of them.

Generalization, which we've just talked through. Interpretability is an interesting one. I don't think we strictly need to solve causality and causal reasoning in deep learning to bring this technology to market. You know, causal reasoning is not something that human brains can give. It's not something that AV 1.0 can, you know, simply give either. It's a really challenging problem. I think the key thing that we need is a system where as engineering teams, we can understand and fault triage and ultimately improve the system so we don't make the same mistake twice.

That's incredibly important. But I think trust really, if you think about some of the research that's been done on what produces trust in intelligent machines or trust in anything, it really comes down to two things. It's performance and predictability. So we need to build intelligent machines that are first and foremost performant, that are safe, that provide value, that provide impact in our lives. But then also predictable, that don't do erratic things, that they were accurate with what they can and can't do and they meet or exceed expectations. That's really what we need to think about. Interpretability is really important from a development perspective, from a validation perspective, from a, I guess, these kind of perspectives. But I, you know, the research shows that it's not strictly important for building trust. If you think about if you go in

an aircraft today, in an airline, you don't get an interpretable understanding of how the airline, of how the aircraft works, but you trust it because it is performant and predictable. And primarily these are the things that we need to make sure are solved at scale to see this technology trusted and adopted. Okay, well, hearing you elaborate on how you are approaching the problem, I had another question that just popped up. Well, obviously some of the things that you are dealing with are sort of on the edge of what's possible, let's say. So I wonder if besides the obvious focus that you have on commercial applications, I wonder if people in your company are also active in research and publications

and things like that. You did mention publication on Rxiv, which I sort of went through just very superficially and I want to return to that. I wonder if there are people in your company that are actively publishing in scientific venues as well, and what's the company position on that? I'm absolutely thrilled to be working with an absolutely world-class team. A team that's, I mean, there's a couple of things that I think I've noticed in my career that have been really important for us to build away. Firstly, machine learning at scale is really 90% an engineering challenge and 10% about

tinkering around with algorithms and so the the team we have that builds our benchmarks, our data infrastructure, our visualization systems, our compute, I think are really enabling a lot of what we're able to achieve today and are fantastic. We also have a brilliant operations team that lets us test our systems on the road and a simulation team that lets us accelerate our learning through simulation as well. But in the research domain I'm fortunate enough to work with a number of really world-class researchers in their own right. For example, our VP of AI Vijay, he's been someone who has really pushed the boundaries of a lot of perception technologies around recognition and semantic segmentation and you know or our chief scientist Jamie

Shotton who has done a lot in leveraging synthetic data to push the boundaries of machine learning and has seen his research realized in products like the Microsoft Connect and the HoloLens and now contributing to our AB 2.0. But I could name many more. We've got a brilliant set of researchers. The final thing I'll say is that building the future and building deep technology really takes more than just commercializing state of the art you need to build uh for what's possible in five years time for what's possible in the future to be in a position to really um really pioneer this when we started in 2017 you know we were under no

illusions that it's not possible it wasn't possible then to create a machine learning system that could drive at scale but everything we saw was that this was the trajectory that we were going in. If you think about the compute infrastructure, the simulation, the machine learning technology, the scale of deep learning, everything was going to conflate in five or 10 years' time from when we started, such that it was possible.

And that's what we're seeing now. And that by building for what was possible in five years time we're now in a position where we have the team the resources ready to go and and create this now doing that in isolation is also uh i think not the right way to go about it it's really important to us that we are open and transparent about what we do we are accurate and don't create unnecessary hype around the technology, accurate around expectations, and a lot of that is we get some very valuable feedback from the research community.

So our team is very active in publishing. We've had a number of papers published at venues like CVPR, the largest computer vision conference, or NeurIPS, one of the largest machine learning conferences, or ICRA, the International Conference of Robotics. And so we spend a lot of time getting feedback from the community and participating in these kind of events, giving talks. And ultimately, I think, you know, I'm really excited to see the collective industry push us forward in this and propel us into the future of autonomy. But I think Wave has a role to play in leading that conversation. Another thing that I think came up a couple of times during this conversation was synthetic

data. And I wonder, to be honest with you, that's also something that I became acquainted with quite recently. I mean, I was familiar with the possibility, let's say, of doing that, of using synthetic data, but I think I just recently came in contact with people that actually work on that, on the provider side of things. And so I wonder what's WAVE's approach to that? Do you have your own framework for simulating and generating that kind of data? Or do you work with some third party provider or some off the shelf platform? Yeah, I mean, the motivation for synthetic data is simple.

We want to run over thousands of children in simulation, so we never will in the real world. We want to be able to learn from experiences that are unethical or impossible or too rare to collect in the real world, to be able to build the most that are unethical or impossible or too rare to collect in the real world to be able to build the most robust and performant driver. It's a fantastic source of this information because you can get accurate ground truth data, you can get unlimited possibilities of combinations and with today's technology you can get near photo realistic results. So what have we done? Well we've built an internal simulation system. We spent a lot of time looking at the market for simulation and

synthetic data and nothing was designed for AV 2.0. Nothing was designed that was really set up for creating an end-to-end driving simulator. And so this is something we've built in-house. We have a large-scale procedural system that can create rich and diverse synthetic worlds that have the right content that something we've built in-house. We have a large-scale procedural system that can create rich and diverse synthetic worlds that have the right content that has the right domain randomization that gives us this kind of power. We've built our own renderer, and this is something that lets us simulate at very, very large distributed scale in Microsoft Azure's cloud, and even use technologies like neural rendering to be able to increase and improve the photorealism of this.

All of this as a package gives us the ability to learn from diverse simulated and synthetic worlds and ultimately improve the performance of our system and understand the performance of our system faster than we can from real world data alone. Okay, so I guess that plays into the theme which you mentioned earlier that when you started out or at the point that you actually needed that synthetic data there was nothing around that could stand up to what you needed so you had to build it

yourselves and I guess that also makes sense. It's a large upfront investment but again eventually the investment will pay out. And so I guess we're close to wrapping up so let's let's do that by asking you about your roadmap basically so again earlier you mentioned the fact that you had an initial successful as I understand deployment of your technology in different cities than London in which that you use for for training and so you know if somebody wanted to be critical about it, there were a couple of things that you may pick up. So you chose to deploy in other cities in the UK, and that obviously

makes it a little bit easier. So, for example, it would be probably much harder if you were to deploy in, let's say, Lagos, for example, where people drive on the other side of the street, the environment is much, much different and so on. So to frame it that way, so when, what's your timeline? When do you think you may be able to successfully do a test drive in a city like Lagos, for example? We've got an exciting roadmap ahead of us. So we're looking forward to shortly deploying our technology in our customers' operations to pilot last mile grocery.

So last mile grocery delivery. This is going to be exciting for us because we'll see this technology, you know, seeing commercial value. We will also see increased experience from these commercial operations. And, you know, and most importantly, get that feedback to continue to learn and grow performance. So we're gonna be starting last mile grocery delivery, driving commercial vans with a safety operator behind them,

starting in London. Beyond then, we're excited to grow this technology with more scale and experience. So as we think about adapting to new cities, to new use cases, we will see this continue to evolve over the coming years with the backing of the supercomputing technology that we're developing with Microsoft, with this data, with the machine learning research that we've been talking about. Ultimately, this roadmap leading us towards deployment in 100 cities and beyond. That's really what we're chasing and the vision that we're building towards at Wave.

I hope you enjoyed the podcast. If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook.