muir_banner

In a new book Christian Wolmar argues that issues are being glossed over and full automation may never be possible

 

 

We do not agree about everything, Christian Wolmar and I, but on driverless cars we do.

Myself, I have long been a driverless cars sceptic and have earned many patronising putdowns from people high on the hype cycle. Graham Dalton, then head of the Highways Agency, was amongst them. “You don’t understand, George. They will be here sooner than you think,” etc., etc.

Oh, yeah!?

Of course it depends what “be here” means. If it means driving in test conditions on well controlled roads in good visibility, then fine. They are probably doing it now. But if it means getting me back home from the pub, then no. Dream on, Graham.

Christian Wolmar’s book, Driverless Cars: On A Road to Nowhere, spells this out well. He writes: “My test of true driverlessness is a vehicle that would be capable of taking a passenger to their office before returning home to whisk the kids to school. We are nowhere near that. It is decades away and it may never be possible because of some of the issues raised in this book.”

Google et al out-do each other in hype. Chris Urmson, Google’s former project head, said his aim was that his son need never learn to drive. Pure hype. This son will be 18 in four years’ time. If he is to leave his dad’s house, he will have to learn to drive or catch a bus.

Paul Campion, chief executive, Transport Systems Catapult, speaking at a recent Westminster Energy Environment and Transport seminar, also seemed to be high on the hype cycle.

I asked him whether, if I was to keep my car for another five years, I would, putting aside legal and cost issues, then be able to replace it with an autonomous car able to drive me home to Fulham. “Yes”, he said, and then added in a quiet voice, “depending on the route”.

But this is exactly the sort of disingenuous half-truth which people in Paul Campion’s position should go out of their way to avoid. A car that is restricted to just some routes, and probably some times of day, and some weather conditions, is hardly a replacement for my old Rover. He should be helping policy makers, not adding to the hype. He should be spelling out, if he knows, what is and is not achievable in the next few years and why.

My very strong sense is that a general purpose autonomous car – one which I would buy to replace my old Rover – will not be available for decades.

My very strong sense is that a general purpose autonomous car – one which I would buy to replace my old Rover – will not be available for decades.

My neighbour at the seminar from a well known strategy consultancy was also a sceptic. “All the progress has been on good roads with good visibility. Google has been working on this for years. They may have got 99% of the way to a workable product, but the last 1% is proving very difficult indeed”.

That was certainly the view of other seminar speakers. Daniel Clarke, Smart Cambridge programme manager, was interested in autonomous pods, guided busways, even drones. But autonomous cars, “we are not so interested”.

In his book, Christin Wolmar does what Paul Campion should have done. He has looked coldly at what has actually been achieved, which is somewhat limited, and at the many, many issues in the way of producing a fully autonomous car.

This first issue is the difference between Level 3 and Level 4 automation. In Level 3, the car does lots of clever things – in the right traffic conditions it will drive itself – but it relies on a human “to respond appropriately to a request to intervene”.

Quite a few cars claim to be Level 3. Tesla cars can drive for “minutes at a time” without human intervention. The new Audi A8 claims Level 3 capability, but only in stop/go, slow driving conditions and on dual carriageways – ie. no oncoming traffic – at speeds up to 37mph. Outside that, bells ring and if the human does not take control, the car stops.

The crucial matter is the circumstances when the human must intervene “appropriately”. This is not clear. In theory, as I understand it, the human driver need not be so alert that he/she is ready to intervene in an instant to stop an accident – the car itself should do that – but must simply be ready to take over when the traffic conditions change. But this is a grey area.

In any event, Level 3 is not much use. Humans want either to drive or go to sleep. They do not want to have nothing to do, but still be ready to take over on very little notice. This is widely recognised. Level 3 is not a useful product.

A Level 4 car, none now exist, would take control of all driving functions but might still request a human to intervene. It would however handle the situation if the human did not. It must be able to “abort the trip”, which might just mean stopping in the middle of the road. Ford and Volvo and others claim they will offer such a car in 2021, in four years’ time.

It would, to be clear, only be operable in specific locations and probably in certain specific weather and visibility conditions. Despite all their sensors and much vaunted AI, these vehicles are not looking around and interpreting their environment in a general way, as a human would. They are not looking ahead and detecting a road junction and saying “this looks tricky, there is a van obscuring the view, better go slowly” or noticing a school and thinking “it is about 3.30pm in the afternoon; it looks quiet, but better be on guard for children running into the road”.

No. They are controlled by very detailed maps and photography of the road, constantly updated for road works or anything else. These define the boundary within which the car’s systems make decisions. So long as what it sees conforms to the maps and photography, the system drives the car. It avoids other cars, stops at halt signs (which the maps have told it about), proceeds at junctions (which the maps have told it about) if nothing is coming, and tries not to hit pedestrians or cyclists.

As an AI challenge, this is difficult enough, but it is not impossible. If you can adequately and narrowly define a boundary and if the tasks within that boundary are simple and clear, then an AI system can eventually learn how to carry them out. The games of Chess and Go have clear boundaries and simple tasks and it is no surprise that AI has learnt to defeat a human.

What politicians and those in charge of roads need to know is how narrowly and tightly must they define road space to make the driving task so simple that autonomous cars, at Level 4, can use them.

Radio connections between cars are increasingly thought to be part of the answer. Something to watch is the series of trials now underway in Coventry into seven different messages which can be sent between cars and infrastructure, such as “Intersection Priority Management”.

Beyond that narrow boundary, away from clear roads with reasonable visibility, on busy local side streets, human drivers are needed.

Only a Level 5 car, according to the definition, can do without any human intervention, in all circumstances, and meets Christian Wolmar’s definition of a driverless car.

Seeing so little information, as opposed to hype, coming from Google and the motor companies as to what has really been achieved, he began to ask questions. How would autonomous vehicles cope with pedestrians on the road? Or cyclists – a great difficulty as they behave sometimes like pedestrians and sometimes like vehicles? How would they distinguish between a traffic jam and a line of parked cars? Would they be programmed to never break the law, such as speed limits? Or, more controversially, would they be allowed to drive illegally? Would bad people or pranksters be able to stop cars at will or would they be programmed to run them over?

More broadly, would the urban realm have to be completely redesigned?

He asks also if the advantages claimed will be real. Will driverless cars really be safer? Will parking and congestion be avoided? Will people share cars instead of owning them? Well, possibly, he says. “Driverless cars are pitched as an exciting technological solution to social problems… but it is not the technology that can deliver the changes they forecast, it is transport policy, irrespective of any shift to autonomy”.

All of which helps our understanding of what future transport is likely to look like.

 

This article appears inside the latest issue of Passenger Transport.

DON’T MISS OUT – GET YOUR COPY! – click here to subscribe!