Denis Naberezhnykh and Stan Boland gave evidence.

Welcome. Would you like to introduce yourselves, please?

Stan Boland:

I am Stan Boland. I am the CEO of a start-up company called FiveAI. We are building a driverless car system, which we hope to trial in London by the end of 2019.

Denis Naberezhnykh:

My name is Denis Naberezhnykh. I am head of ultra low emission vehicles and energy at the Transport Research Laboratory. We work with industry and Government to help to introduce new technologies such as electric and automated vehicles.

Q I have one really quick question on safety. What are the safety implications for blind pedestrians, runners and people of that nature?

Stan Boland:

Safety is the start and finish of whether we can bring these cars on to the streets. A huge amount of attention will be focused on making these vehicles safe, in our case, for use in urban environments, where we will have all sorts of obstacles and agents with all sorts of different behaviours. That really centres on having systems that are able to perceive what is in the scene accurately in 360° and three dimensions and classify what those objects are.

This also talks to predicting what will happen next. We actually have to predict human behaviour, and we have to learn what those behaviours might be ahead of time. Our vehicles will certainly have to be state of the art for perception, but they will also have to be very good at predicting human behaviours. In the case where we identify an object and can tell, just like a human can, that this person, cyclist or whatever it turns out to be has a certain type of behaviour, we will have learnt those ahead of time, and if we are not sure, we will have to propagate that uncertainty through our software and slow down.

The behaviour of these vehicles will be slightly different to that of human drivers, but it will be possible to attain the levels of human safety, and in the long term surpass them, by applying technology. Our systems can pay attention in 360° all the time, and that makes it a bit different to human drivers.

Q What will be the limit of the speed that your vehicles can travel at in an urban area?

Stan Boland:

We are kind of hoping that we can operate at normal driving speeds. To be able to do that, it is important that we can predict behaviours. We cannot have a system that is collision-avoidance only, because that would result in frozen robots all over the city and would make congestion worse. What we humans do is anticipate human action. We actually run more than one world in our heads, and are constantly looking to see whether that world is turning into reality or some other world is going to happen. That allows us to merge on to full lanes of traffic, for instance. We cannot just have a system that is collision-avoidance only, because we would make traffic worse. The idea is that we are operating in normal streets with normal road signs at normal road speeds and obtaining and exceeding human levels of safety.

Q In 2019, your vehicles would be the only automated vehicles on the road.

Stan Boland:

At that point it is a trial, so there is a safety driver in the car. The safety driver is able to take control of the vehicle immediately.

Q Immediately?

Stan Boland:

Yes. The safety driver has to be there, literally able to take control of the car instantly.

Q All the research that I have seen in preparing for this Committee shows that there is a time lag in the transition from the vehicle being under automated control to being under driver control. Your vehicles will not have that.

Stan Boland:

You are describing what is called level 3 autonomy, which is a system where the car is under automated control and then there is a warning to give a human driver time—there is a debate about what that warning time should be—and then the human is meant to take over. We think that system is intrinsically unsafe. It is much better if either the human is in control or the system is in control—that is a fully automated, level 4 or level 5 system. We are building a system where the cognitive capability of the car is in control, but for the purpose of testing, until it is actually legal to offer that service, there will always be a driver in the car who can take over instantly.

Q But there is an issue about when drivers take over from automated vehicles: they are over-cautious and slow down. There are concerns about whether that is dangerous at that moment in time and also it increases congestion because you suddenly have loads of slow-moving vehicles. You seem to be suggesting a flick of a switch and that it goes from driver to automated.

Stan Boland:

While we are testing it. We are talking about a period when we are testing the capability of the vehicle in our existing cities. It is level 4—a highly automated or fully autonomous system—but for the period between now and a regulatory capability of doing this and, moreover, underwriting the risk of it, we have to have a driver in the car to take over.

Q As the legislation is drafted, as long as somebody insures you, there is nothing stopping you putting that vehicle on the road for an experimental period in 2019.

Stan Boland:

As long as there is a safety driver who can take over the car. That is not the same as somebody watching a Harry Potter movie while the car is self-driving. We are talking about a qualified driver who is paying full attention to the road scene all the time and can take over.

To test—

Stan Boland:

To test the vehicle.

Q To follow on from the hon. Member for Eltham, you are talking about level 4 rather than 3. Do you envisage that the machine’s design will be able to engage in autonomous conduct on every kind of road from the very first moment that they are launched?

Stan Boland:

No, that would be a definition of level 5 in our parlance: something that could literally drive anywhere on the planet and be able to work out what every object was, what the semantics of every scene was, and the human behaviour in that part of the world, so we are definitely not saying that.

Q So a level 4 vehicle has some of the aspects of a level 3 vehicle, in the sense that there will be certain road conditions and certain kinds of roads where it needs to hand back to the human driver.

Stan Boland:

No, I think it would stop in that case. In the behavioural model, we were able to bring up a system that works in a defined geography and in defined driving conditions, but if one day the place is completely wiped out with snow, we probably would not drive on that day. Our business model is to deliver a service. It is a service model.

Q So I am driving along a motorway where your system is happy. I diverge on to a trunk road where your system is happy. I move from the trunk road on to a B road where your system is happy. It is a somewhat rainy or snowy day. I get on to a one-track road in my constituency and your machine suddenly decides it is not going to get me to my home.

Stan Boland:

No, that is not the model at all. First, you would not buy one of these cars. This is a shared form of mobility that is offered in cities. You would not buy it because the sensors and the compute you have to put in that car make that prohibitively expensive. It adds some £30,000 to the car.

Q So these are city-only cars.

Stan Boland:

City-only vehicles.

Q And they are available to drivers not as a vehicle, so to speak, but as a service on certain days of the year under certain conditions.

Stan Boland:

Probably all days of the year, but there may be times when humans probably should not be driving, frankly. In those conditions our vehicles would not stop just like that, because that would be unsafe, but they would be able to be carefully brought to a slowdown to stop safely.

Q It is basically like not being able to get a cab.

Stan Boland:

If you like, yes. We think there will in any case be remote supervision so that it would be possible for a control centre to be able to monitor any cars that are stopped and then perhaps carefully move them to some other place. We are expecting a remote control room with perhaps one per 30 cars or something that would be able to take over and carefully manage the car. We are also expecting the cars to have a limp-home system, so if there is a catastrophic failure there would be a limited amount of capability where the vehicles could—at quite a low speed and with warnings—find their way back to a service centre.

Q I have three questions. First, you made an interesting case in response to questions from my right hon. Friend the Member for West Dorset and the hon. Member for Eltham about what you see as the likely development in autonomous vehicles. I will put two scenarios to you, and I would ask you to describe the most likely one. There is an instinctive, intuitive view that autonomy will grow gradually, partly because that is more likely to lead to public acceptance. So rather in the way that assisted parking or sat nav or cruise control have become increasingly routine, other aspects of autonomy will be added to that. Autonomous vehicles will sort of creep on us.

There is another view that we may go straight to a kind of autonomous vehicle. Indeed I have looked at some of the R and D on that. As you may know, there is an entirely autonomous vehicle at Greenwich supported by Greenwich council, with some Government funding too. That is a vehicle that travels on a straight run of road that is entirely autonomous. You get into it, and it does what it says on the tin. Which of the two scenarios is the most likely, in your view? Or are they most likely to develop in parallel?

Stan Boland:

They are developing in parallel today, so I think that is the state of affairs. The first of those can be characterised as the view of the German car industry, which is that these things will happen, but in 2035 or 2040. In the meantime we can just keep adding these features, keep selling people more features, and keep selling cars that people buy. However, I think the world was really shaken up by the challenges we saw in the 2000s and the emergency of Google cars and so on, as well as the idea that it was within touching distance for science to deliver fully autonomous capability in a relatively meaningful timeframe.

That really is the difference between level 2 and level 3 autonomy and what is really a huge jump to level 4 and level 5. Our entire business is predicated on level 4 and level 5 being the dominant model. We think that that is the dominant model for getting to a situation of safety in an urban environment. Significant amounts of algorithms, computer models, training data and sensors are involved in achieving this, which will considerably increase the cost of the car. We estimate that getting the car to human levels of safety will add a further £30,000 to £40,000 to its cost. That is not a car that people buy. That is definitely a service, and if it is a service then it is fully autonomous.

Q My second question is about software. We debated this to some degree today. Presumably, the challenge is to develop sufficiently sophisticated software to anticipate scenarios encompassing all kinds of different eventualities. An interesting question was put earlier by the hon. Member for Eltham about the element of human judgment in driving when faced with a dilemma, where you would hit traffic rather than a child, or you might swerve and possibly cause a more serious accident. The way that scenario planning is written into software would require hundreds of thousands of scenarios being tested. I know there is R and D being undertaken in this area. Would you briefly describe that to us?

Stan Boland:

It is impossible to test all of that in the real world, and it would not be safe to do so. It has to be done as a simulation, which is the key to getting to the point where we have safe systems that can operate in our cities. We have to be able to simulate all the sensors on the car and all the different failure modes and so on. We have to simulate all the cases where our predicted models break down, or where somebody in the distance who is wearing a green pullover against a green wall with a reflective window near it cannot, for whatever reason, be seen in our systems. We have to be able to simulate those kinds of things—perception failures. We also have to simulate the extent to which we may not be able to predict human behaviours. We may never have seen a particular behavioural type before, and it may be dissimilar to anything we have seen before.

We have to do all that in simulations, so the money is invested in creating a simulated world that may be like the whole of London, photo-accurate for example, and it may be that we create generative models that allow us to create every angle of a road—instead of 43 degrees, it is 44, 45 or 46. Instead of five objects, there are six; instead of a certain kind of road markings, they are slightly different. We can basically generate all that in simulations, so we can drive potentially billions of miles in simulation ahead of that software actually going in a vehicle and being sent out on the road. That is the way we can really assure the safety of those vehicles—a heavy investment in simulation. It turns out that the UK is good at that. The UK is good at artificial intelligence, gaming and simulation, so we are in a good position to do that.

Q So you are not simulating what is, as I put it, a tens of thousands scenario; it is millions. It is the whole of London, in all weather conditions, in all circumstances, for all vehicles, in all eventualities.

Stan Boland:

Exactly. We will find real cases in the real world which we will codify. We are working with TRL to do that, to deliver a curated set of regression test cases.

Q Thank you for that; that was fascinating. My final question is about skills. We debated skills earlier; Mr Turner raised the issue and I then amplified it. From what you are describing, and I discovered this in my early research, the car mechanic of the future will be a software engineer as much as he is a mechanic, and that is going to require a step change in skills. Are we ready for that, and if not, what are we going to do about it?

We will have to have Ministers with proper skills in future too. Sorry, Mr Boland, please answer the question—that was just a facetious remark. This must be the last answer, because we might have multiple Divisions.

Stan Boland:

We definitely need more software engineers as a nation anyway, so we are probably not ready for any of this in terms of the total number of skills that we need to go alongside companies the size of Silicon Valley companies, but I think there is a kind of rarity about what—[Interruption.]

Order. Thank you. I apologise, but we have been interrupted.

Ordered, That further consideration be now adjourned. —(Andrew Stephenson.)

Adjourned till Thursday 2 November at half-past Eleven o’clock.