The take-away

• Swiss start-up BestMile has developed software to control a fleet of autonomous vehicles.
• An automatic anti-skid technology has reduced the toll in loss-of-control accidents by 25% per year in the UK and US.

Browse any online technology news website, catch a newscast or even read a newspaper and you could be forgiven for thinking that self-driving cars are about to appear in a showroom near you. With Google’s experimental cars having driven hundreds of thousands of kilometres on bright, broad California boulevards, and car makers like Tesla gradually automating their expensive executive vehicles, we’re told it won’t be long before drivers will be able to sleep in their cars, read books, or watch movies – while smart software does all the work.

Meanwhile, experts warn that the driverless propaganda is getting way, way ahead of reality: there are yawning gaps in the capabilities of self-driving vehicles that mean carmakers are a long way from being close to offering a trouble-free, safe ride. “The media and other commentators feel we are further along than we really are,” says John Leonard, an engineer who studies the mapping, localization and navigation of autonomous vehicles at the Massachusetts Institute of Technology (MIT). “I do feel the technology is being over hyped.”

He is far from alone in believing that. While the likes of Google may argue that few accidents have occured to date in their tests, all the geek hype in the world is nowhere near convincing the organisations that need to back novel automotive technologies with their own cold, hard cash. Those organisations? Insurance companies.

“There are thousands of random little situations and unforeseen conditions that self-driving cars will have to be programmed to cope with before they can be launched,” warns Matthew Avery, research head at Thatcham Research, the European motor insurance industry’s technology analysis centre in Newbury, UK. “As a result, we do not think driverless cars can be expected to arrive until 2025.” Those situations range from getting software to mimic the arcane intricacies of human road etiquette, to coping with low-sun obscuring traffic lights, to snow or truck spray concealing lane markings – or simply having the computer break the law to save human life. On top of this, there are also major ethical questions about the priorities of a self-driving car’s software: should it be to protect occupant’s lives at all costs? Or those of pedestrians? The answers have vast implications and are nowhere near being solved.

Sion, Switzerland

Since early 2016, the city tests autonomous busses.

Piecemeal automation

All this means that the first driverless systems we will see – and pretty soon – will be simple, slow-moving multi-passenger electric “pods” plying well-mapped public transport routes. In Sion, Switzerland, and in Milton Keynes, UK, such services are due to kick off in 2016. The Swiss system will use a fleet of eight-seat vehicles intelligently choreographed by swarm-control software developed by BestMile, a spinoff of the École Polytechnique Fédérale de Lausanne (EPFL).

“One vehicle is smart, but a fleet of vehicles is stupid,” explains BestMile founder and CEO Raphaël Gindrat. “Our platform lets people operate a fleet of vehicles in a smart and efficient way, optimising the schedule, traffic, demand and charging strategy – all in real-time.”

This eminently sensible, slow start to the driverless era does not mean fully self-driving cars are not needed: once perfected they could help seriously reduce the jaw-dropping toll of 1.2 million road deaths worldwide per year – that’s nearly 3,300 killed per day.

The potential for new technology to make a difference is already evident as piecemeal car automation has already begun to reduce casualty figures. Over the last decade, says Avery, Electronic Stability Control (ESC), an automatic anti-skid technology, has reduced the toll of people killed and injured in loss-of-control accidents by 25% per year in the U.S. and the UK. And Autonomous Emergency Braking (AEB), which senses a stopped vehicle or a pedestrian in front of a car, has reduced crash rates overall by 10% and crashes involving casualties by around 30%. The VW Golf Mk7 is the star car here, says Avery: recent models are involved in 42% fewer crashes since AEB arrived.

The introduction of such technologies is paving the way for the driverless car by providing the sensors and automatic actuators that are its building blocks – variously adding wheel-motion sensors, laser radars, microwave radars, ultrasound sensors, cameras and automatic braking and steering servos.

These have led to a gradual increase in automatic operation, moving from the assisted automation of AEB and ESC, to more active, partially automated systems like those that keep a car in its lane, rather than just annoyingly warning of lane departure. After that come automatic parking and traffic-jam assist systems from the like of Ford, Volvo and Mercedes, which brake and steer in slow-moving traffic. And a notch above all this in terms of automation power are the auto-driving systems that steer for you on straight autobahn or motorway runs, like Daimler’s nascent Highway Pilot for trucks, and Tesla’s recently launched Autopilot. But the driver must still monitor the controls with both these systems.

The next level is full automation. “That will be the Google car in 2025 – with no steering wheel, no controls at all”, says Avery. The arrival is years away because of the sheer number of problems that need solving. MIT’s Leonard has been driving the streets of Boston video-recording road situations that are likely to confuse a self-driving car. For example, imagine simply turning left (or right in the UK) into a constant stream of traffic that will not allow your car in. As a human at the wheel you can gently nose your car out, make eye contact with another driver and “hope somebody shows mercy,” says Leonard. “That’s a human interaction problem and I don’t see how the computer-vision folks are going to solve that in a self-driving car.”

Double trouble

In another situation, Leonard approaches a traffic light with the sun low behind it so that lights are obscured by glare. But what is not immediately obvious is that there is a traffic cop below the lights – and he is also obscured by glare. The lights have failed and the officer is trying to direct the traffic by hand – a double problem for a driverless computer to interpret. Or a traffic cop steps out into the road and halts traffic just as the lights turn green: what does the robot car do?

Meanwhile lane markings are crucial to some types of driverless navigation – yet Leonard shows easily how one day a Boston bridge was well marked, yet the next day the lane markings had gone as the bridge deck had been repaved. And on still another day the lines were back but obscured by snow.

The answer is to map the world very precisely, as Google does, to allow the car’s position to be localised. But what if unexpected objects – traffic cones say – are hidden by snow? Lidar (laser radar) may someday solve that problem, but in the meantime the number of problem permutations is a major stumbling block.

But it will be worth the candle, says Avery. “As soon as people get used to driving highly automated vehicles they will really want to move to driverless operation. It will be a gradual move, but uptake will be high because it will be so liberating.”

Ethical dilemmas

Will a car programmed to obey the laws of the road drive up on the pavement to allow an ambulance to pass? It’s against the law, though it’s what any sane human would do. What about the choice between running over a child and sacrificing a car’s own occupants to avoid an oncoming lorry? A driver often has no time to think in an emergency, but there will always be a thought process behind a computer’s decision. Such ethical issues have been probed in depth by Jean-François Bonnefon at the Toulouse School of Economics in France. “These moral algorithms will need to accomplish three potentially incompatible objectives: being consistent, not causing public outrage, and not discouraging buyers,” he says. At the root of the problem is a very simple fact: a completely driverless car is a robot – one that people sit inside. And if there is one thing roboticists can agree on, it’s that it is devilishly hard to get a robot to do anything for very long in a safe, consistent way. So conquering the issues will take time.