Google's driverless cars 'can't use 99pc of roads'
Published 02/09/2014 | 09:56
Detractors of Google's driverless cars criticise the technology because it can't use 99pc of roads, park itself, avoid potholes or drive in the rain, but they miss the real problem: most humans are also utterly incapable of driving a car safely
“Would you buy a self-driving car that couldn’t drive itself in 99 per cent of the country? Or that knew nearly nothing about parking, couldn’t be taken out in snow or heavy rain, and would drive straight over a gaping pothole?”
That’s the rather loaded question posted by Lee Gnomes, writing in the MIT Technology Review, laying out current flaws in Google’s driverless car software. He isn’t alone in his pessimism; a recent survey in the UK, US and Australia showed that three quarters of people were sceptical about safety.
But to criticise autonomous cars for only being able to traverse one per cent of US roads is to miss the point: ten years ago they couldn’t drive any of them. And one day soon they will be able to manage two per cent, then 20 and, eventually, all of them.
To step in now and rip them to shreds, during R&D and before a commercial launch, seems needlessly pessimistic. It’s also disingenuous.
The crux of this 'one per cent problem' is mapping. Google has to build incredibly detailed models of roads, complete with any obstacles and signs, before its car can autonomously drive them. Currently it has only covered a tiny minority of the country's 3,980,817 miles of road, for good reason.
You need variety for testing, but also the ability to endlessly repeat the same stretches of road to measure, analyse and modify. The project involves dozens of engineers and labs full of equipment and it makes sense to keep all that in one area. Why traipse from one end of the country to the other? Is it better that the tricky junction which keeps stumping the car's evolving 'brain' is three miles from your HQ, or 3,000?
When the company needs total coverage, it’ll get it. Street View is testament to its ability to map large areas. But the sensible thing to do would be to get the autonomous cars – equipped with banks of cameras, sensors and computers – to do it themselves. The first few owners to drive every street would need to do so manually, beaming information invisibly and seamlessly back to Google HQ. After that, nobody would ever have to touch the wheel or pedals again.
It would take just weeks before every tiny, obscure mile of the US was meticulously mapped. Perhaps the company will offer Google Play rewards to the first people to map a section of road?
No, the main problem is not maps, it's software. Right now the code is easily confused. If it comes across a set of roadworks it has a wobble, slows to a crawl and occasionally gives up. But every time it does this Google’s engineers track down the problem, tweak the software and re-test. Lessons are learned and the car becomes more capable.
MIT’s article also points out, quite correctly, that the car cannot currently tell the difference between a rock and a piece of crumpled paper, so will steer to avoid both. But this is hardly a fault – I was always taught to do this, when safe, with even the most innocuous-looking obstacle. You never know what’s hiding within.
Right now it would ride straight over an open manhole, or carelessly sink a wheel into a deep pothole. But the off-road DARPA Grand Challenge races of the early 2000s shows that this is largely a solved problem.
Snow and rain may confuse it, but Chris Urmson, director of the Google car team, told the website that “safety concerns” have stopped the company testing in poor weather conditions thus far – an entirely sensible precaution, you would think, at this early stage. Of course this will be fixed too. Humans, meanwhile, tend to have more accidents on wet roads - it seems that our own software also needs some work.
Which is perhaps the most important point. To criticise automation software for being “worse than a human” is disingenuous because humans vary wildly in their ability to drive safely. Compare a veteran police driver with someone who needs glasses but refuses to have an eye test, a texting teenager, a wobbling drunk or a minicab driver at the end of a 13-hour shift.
The inconvenient truth is that if Google released an autonomous car which drove as well as the average human it would be immediately banned for the chaos it caused. But we're accustomed to the "accidents" caused by humans and, for a few more years at least, have no better option.
As a cyclist I’d rather share the roads with a fleet of robots than most humans. For me, they can’t come soon enough.
Independent News Service