Name: Wayve
Uploaded: 2024-12-28T16:54:27.276Z
Duration: 2 min 33 s
Description: A UK-based company pioneering AI-driven autonomous vehicle technology through end-to-end deep learning solutions.

wayve

A UK-based company pioneering AI-driven autonomous vehicle technology through end-to-end deep learning solutions.

Wayve

You out this morning hope you enjoy it. Wow I really did. I was really really impressed actually really really impressed. That is a hard stretch up the Caledonian road. Yeah so zero interventions I saw that. Yeah yeah it's really difficult. 

Yeah so here with uh with Alex Kendall he's the founder and CEO of Wayve. Hi Alex is name. Uh I mean this is quite an interesting experience being in um what a jaguar pce with a human driver and a powerful computer sitting in the back. All right here we go. So the uh the AI is now controlling everything about the car. It's controlling the speed the steering the brakes the indicator. Not yet the horn not yet. Oh will it control the horn eventually? But uh but but not yet. It's a it's a polite and considerate driver. 

Um but yeah you see how we slowed down for that pulling away vehicle and now we're moving forward. I mean none of this behavior has been hand programmed. It's all uh all emerged from the training data. Yeah and the training data is something that we'll definitely have to explore because you know not every driver is a great driver. So if you were to just learn from the average London driver I wonder exactly how nice a driver The Wayve system would be. 

Yeah exactly we we see it all here in London. But I mean look at this we've got a cyclist ahead we've got some um road works on the side of the road. Yeah this is actually a complicated moment right cu the road narrowed. Anyway we've got roadworks and we've got a cyclist and it's uphill so he's wobbling quite a lot. 

Yeah so it's quite interesting to see how the car is dealing with this. Very very polite I have to say and he's on his phone on their phone and now as they pull away we drive off. I mean that uh I cycle to work every day and I'm glad that our car is uh able to be safer on cyclists. So I things I'm looking for and trying to experience here are just how um how sensitive and empathetic the car is both to its passengers and to the other road users. Because you know obviously we're starting to see things like adaptive cruise control and other technologies appear in cars that you can you know buy today. 

And sometimes you feel that they're not the most careful drivers. I mean if I drove the way some cars drive when they they're doing their.

Automated braking and automated acceleration. I would think of myself as a race driver in a way very aggressive this but the Wayve seems to be you know it's kind of decent in terms of and thoughtful in the the way in which it's accelerating. Yeah look there's two Paradigm shifts here uh that are worth worth talking about. Uh the first one is just the complexity of the environment we driving driving in. I mean if you look at this narrow space here it's very unclear where the lanes are. In fact there's not really a center line here it's an unstructured environment. Uh there's lots of busy road agents around whether it's the other cars the cyclist we just went past. Uh the complexity of this environment is is such that you know you get an AI system that's capable of deploying here it it should be able to go worldwide. I think that's right. I think just to describe what we we what we're in on it's what I would call a fairly typical run London Road which means it's narrow it's bumpy or just gone over a big bump there are speed bumps and potholes cars parked on both sides big vehicles approaching us. 

Uh in all the time I've driven in the United States I don't think I've come across would have come across a road that is as complex as and unpleasant actually as this one is for a driver. And yet this is probably a pretty average London Road it's much worse than this kicking around. So did was that a deliberate Choice then to to start here because sometimes when people are developing Technologies they'll say let's start in a really easy environment with in the case of autonomous vehicles with big wide roads low traffic levels and nice weather. You've chosen narrow roads aggressive drivers bad traffic frequent Road Works diversions uh poor Road quality and pretty terrible weather is your environment. 

Well I love London. Me too. Me too I've been here uh what 5 years now but when we started uh 6 years ago in Cambridge um at that time that's exactly what you saw in the autonomous driving industry. I mean people were building uh systems that were told how to drive in very clean and structured environments and if I thought okay what's the future of Robotics going to be what's going to enable intelligent machines to exist worldwide it's not systems that

Are hand programmed how to behave but ones that have the onboard understanding to be able to drive through the rich complex and diverse and worlds that we live in. Uh so the the Bold um decisions that we made at Wayve were we want to build a car that has the onboard intelligence how to drive a car that isn't following a set map but can understand the environment in real time and deal with the dynamic complexities that we we have on the streets. 

And those decisions of what I mean with London's been a good forcing function where if you don't do that I mean you've got no chance of operating here. Um but it's allowed us to to Pioneer a Next Generation approach and look at what we've got in front of us that was amazing. 

Yeah so so there that was interesting we had two pedestrians older women who decided to cross in the middle of the road really without looking. Um interesting the car slowed down but didn't give way to them what was what was that about because I think some drivers might have stopped and let them pass. 

Yeah I mean it would project where they're going and it it looks to balance both being training to of course firstly be safe and make decisions it's confident in but then secondly uh to you know to make sure you're moving um you're moving through effectively whether it's I mean here we've got a okay so this is a great example right. 

Yeah so here of course you need to yield to a pedestrian on the crossing but I mean just up in front of us we've got a a truck that's parked and sticking out into the road and so you know we shouldn't be waiting behind it. Uh hopefully the AI recognizes that and is able to actually drive around it once the bus comes past and and lets us through and then look it safely comes around. 

I I this is a really really fascinating Tech demonstration. Waving not sure if he's going to cross but like look how busy this is right you've got so much going on. For for listeners who are in in the UK and particularly in London we should say we're driving up the Caledonian road which is running between Kings Cross and Islington. 

Um it's a sort of light Industrial Road it's not particularly wide it's really really busy with buses industrial traffic big lorries double parking Road Works tried to come onto our.

Side of the road and try and cut us off and we had to be defensive and pull over absolutely. On top of that because it's sitting between these areas where a lot of people live, there's a lot of pedestrian traffic as well. I mean it is a really tough road, uh, for anyone. I mean I know because I live around here I don't really really enjoy driving down the Caledonian Road for just that reason. 

Um, so I'm quite curious though about um, we're seeing it in this environment and to be fair that the speed limit is 20 mil hour. Right, the car doesn't even get a chance to get that fast uh in London. Um, you know what's the experience when you go into will this same car happily transition up through to the highways and then the freeways or motorways as we we call them here? And then with the long at 70 mph and does it just do that in the seamless way a human driver might? 

Yeah, the way that the system grows and expands is all through training data. So, uh, you know the same system should be intelligent enough to drive highways, freeways, other countries, other cities, different road rules whether it's I don't know anything from south of Italy or India or even the United States. Um, but the beauty of the system is that so much of the knowledge is shared between those different tasks. 

I mean the way that pedestrians act, the way that traffic lights work, these are fairly common throughout the world, right? Um, and so what that means is that you don't need to recollect the entire amount of training data to adapt to a new place. In fact, to give you some statistics we saw that uh when our AI was able to learn to adapt to go from driving a car like the one we're in right now to a van, uh, we drive uh like commercial vans, 3.5 ton vans to do grocery delivery. You know we only needed about 3% of the training data in order to bring the van up to the same level of performance. 

We've also seen that the system learns to drive here in London but we've taken it to over 15 UK cities now and it's able to drive there without ever having any experience in those cities. It's because it doesn't need a high definition map to tell you where to drive but it's able to take that knowledge and learn the general behaviors of how do you go through a.

Traffic light how do you go through a roundabout and apply those to new new places it hasn't seen before. I so let's go look at the example of teaching the system to drive a van. A light commercial vehicle you said there was 3% training data required. So essentially is that does that mean you took the the model that we're using in this this lovely uh electric uh vehicle and then you just had to train it for 3% additional van specific circumstances? 

Yeah so we collected a small amount of data on the van and we trained the model on both sets of data on the Union and dat majority of dat car experience. But it was able to you know could recognize the van as a larger vehicle it needs to take wider turning circles and things like that. But it also was able to pull knowledge that you know we only see emerge after thousands of hours of of data and I was able to take that over from the the car data set and learn to to get the best of both. 

I I mean I my son is uh learning to drive uh at the moment so I have some experience as a as as teaching. Uh how many hours has it taken your biological uh intelligence ex exactly? I mean you know and I think that in general it's measured in the the tens rather than hundreds of hours right and that's that's typically where you you end up going um in terms of of actual formal tuition and then and then practice. 

So when we think about what the comparison is with your existing systems do you require two orders of magnitude more that you require thousands rather than the tens or is it is it more than that? Yeah the the interesting it's a really interesting comparison because yes it took it well it took me I think 10 maybe 10 to 20 hours to drive. I actually failed my first driving test uh as an aside but I eventually got there. 

But it's not just that experience it's the um you know the 15 or 16 years of of experience you have growing up sitting in a car and observing the road or not. Yeah or even dad playing video games or even just talking about risk and and learning. You know we have the same eyes as a 5-year-old versus an adult right um but it's the ability to judge risk and understand the world that um makes us confident to give someone a driver's license. 

And so in a similar way we're training in AI not just on that Driving Experience but also the broad knowledge base that you need to make safe decisions. Um our latest models are in fact trained on internet Text data. We feed in the literal PDF document the UK government publishes on the highway road code on how to drive. We feed that information into our system and it learns from that. It learns from broad pre-training on, uh you know, wide internet scale Text data and other sources. From all of that information brings together, um you know, we want to make sure it can make safe decisions with that.

So, so so presumably you're not going to train it on, uh Fast and Furious, uh films and Grand Theft Auto. Well, the interesting thing is you actually want to see if I draw analogy to large language models. You know when they're first trained on, oh what do we got here? We got this is great. Okay, we got what, 20 police officers? Yeah, a protest around here with or some investigation. We got van coming around. Wow, 20-30 police officers? Yeah, that's interesting. Yeah, um anyway crazy stuff you see on the roads. 

Uh what was I? We were talking about like Grand Theft or C Furious, a large language model. You, um you train these systems on all the internet text, right? And that includes bullying, profanity, abuse, all of that kind of stuff. And then after you've got that pre-trained model, you then look to align it with trusted and safe human values. But that pre-training is important to build a general representation of what can happen before you align it with what you want. 

In a similar way, we want to see as much diverse driving experiences we can so that we can understand in those scenarios how Society plays out. But then make sure that we fine-tune and prompt and adapt our system to be as safe and compliant as possible. Okay, so that makes sense. During pre-training, we want to see Grand Theft Auto. We want to see that kind of data. But during, um fine-tuning and production, it's very safety critical and careful with what we train our system to do. 

We'll get into some of those questions. I mean I'm curious about, you know, those steps and is there an RHF reinforcement learning with human feedback step afterwards as well in terms of getting the behaviors that you really want?

From the vehicle but but maybe I can come back to um to to your that that observation you made which is that while it takes a human 50 years 50 hours to learn how to drive and and pass their test that's sitting on top of 17 years of experience in the real world and experience of things like uh you know physics and uh risk assessment and observing parents drive and being driven in places and uh watching Cars on TV or on YouTube and playing video games right. So there's a a lot of um knowledge that's been acquired over that over that period of time. 

So I think that that that allows us to maybe put however many thousands or tens of thousands of hours of training you need um to train the vehicle um into context. One of the things that's surprising me though is that you're talking about thousands of hours not millions of miles and tens of millions of miles which is sometimes what we hear from the autonomous vehicle industry whether it's you know a company that's got cars out there and is is like Tesla that is gathering data or whether it's one of the um you know Grassroots uh from the bottom up startups that's been around for a long time. 

They're always talking in terms of millions if not billions of of miles which means you know hundreds of thousands and millions of hours. So is is there something different about your your approach that makes it more efficient in the way that it it's learning? Well there's two ways of looking that first one is for us to see this level of performance with I mean the fleet of 10 vehicles and uh the team of 200 people that we have uh is you know is extraordinarily efficient. 

But then imagine where the system is going to go when we do get access to millions of hours and our road map is going to see us grow into that whether it's our Fleet Partnerships we have with some of the UK's largest fleets or uh you know through larger scale of deployment that we have planned. Um the scaling curves we see of performance against data are very similar to what the large language model Community has experienced and so we will grow that substantially. 

Um the reasons why we've been able to be so efficient I mean machine learning is is is very effective at pulling out the right signals with data and even thousands of hours.

Is enough to build on top of a pre-trained representation and and have this experience. But also the generative Ai and simulation work uh that we've done has enabled us to to get uh you know broader representations to be able to train on. 

But look uh we're just coming back to the office here uh just finishing the drive and we'll make your left turn and just uh just duck back into the office. But hope you enjoyed enjoyed that as it this has been a remarkable remarkable demonstration. 

And uh it is you've taken the C through I think some of the toughest daytime driving that you would get in London which is not an easy car uh City to drive in. Uh the car drives pretty well and I think on this journey we probably saw a couple of things the police the ongoing oncoming truck that you wouldn't expect to see in a given day journey. 

I mean it's really fascinating Alex well done you're taking out this morning. Wow I really did I was really really impressed actually really really impressed that is a hard that stretch up the Caledonian road. 

Yeah so zero intervention I saw that yeah yeah it's really difficult.

Note, that there may be slight differences in text and video. Watch video for the original content.

Wayve Video