Self-driving cars, meet your nemesis: the London roundabout.
This strange piece of geometry, with tentacles shooting off at odd angles and cars nudging into impossible spaces, is one of the many headaches that will plague computer brains as the city’s autonomous vehicle (AV) trials accelerate.
In the US, Waymo and others boast fleets of self-driving cars that have racked up millions of miles of public road trials, across more than 25 cities. Billions of dollars of investment is flowing into AV units run by Uber and General Motors. Tesla is making bold promises about “robo taxis”, and Ford plans to start building AVs in 2021. If you believe the latest McKinsey report, China will see mass deployment of fully autonomous vehicles within a decade.
On this cautious island, start-ups are “dipping their toes” into the waters of big city trials, according to Jolyon Carroll, a vehicle safety and technology consultant at the Transport Research Laboratory (TRL). So far, these have been relatively minor and low-key, done under the watchful eye of a government trying to encourage innovation and apply safety brakes simultaneously. To date, all tests have included human drivers ready to take back control.
FiveAI and Oxbotica are two players with AV fleets operating in London and ambitious plans for the future. Both recognise that a megacity like London presents a unique set of problems for driverless cars.
Humans are good at understanding the abstract concept of safe driving and can apply their experience to new locations. It may take time and plenty of nerves, but a Moscow driver will eventually adapt to driving in Lagos. But algorithms trained in America can’t adapt to new environments in the same way – at least not yet – and have to be given very specific maps and information to do the three things required of them: perceive, predict and act.
The first challenge is to input enough information about London’s road rules – things like traffic signs and speed limits – and generate accurate maps. FiveAI’s Ben Peters calls this the “static scene”, which is the laid-back brother of a city’s “dynamic scene”.
The dynamic scene comprises different kinds of vehicles (a pick-up truck in Texas looks nothing like a London black cab), vulnerable road users such as pedestrians and cyclists, along with “hyperlocal peculiarities” including London’s ‘Boris Bikes’, which add an extra layer of complexity as they’re often used by inexperienced riders. Pedestrians and cyclists don’t behave in predictable ways, and AVs will encounter vast numbers of them in London.
Then there’s the capital’s layout, with its narrow streets in residential areas – where cars must perform an elaborate dance to squeeze past each other – and its chaotic roundabouts and gyratories. These will push a car’s perception and prediction abilities to their limits.
“You have to negotiate your way onto a roundabout,” explains Peters. “There is no gap. You have to force a gap by game theorising with other drivers. Those are very hard challenges for AVs.”
Drivers have to read human cues, make assumptions, take risks and resort to tactics which, strictly speaking, bend or break the rules. Unlike neat 90-degree intersections, roundabouts can have half a dozen arms flowing in and out of them at random angles, with cars speeding up, slowing down and cutting across lanes.
Anyone who’s had to negotiate the Hyde Park Corner loop, the curl around Elephant and Castle or Piccadilly Circus – all flooded with tourists, cyclists, cabs and double-decker buses – will know what AVs are up against.
Oxbotica’s Graeme Smith says London is one of the toughest cities for testing driverless cars. If he had to rate its difficulty level, he’d give it eight out of 10. (For all its chaos, London is governed by strict traffic laws and is easier to manage than cities in, say, India.) “It’s a great opportunity for us to learn how to work in a very congested place, originally designed for the horse and cart. There are lots of tight streets, tight turns and lots of specific traffic laws,” he says. “There’s no margin for error in terms of where you place your car on the road and how you deal with junctions.”
London’s weather makes things even messier. “It never rains in Phoenix, Arizona,” says Peters. “In London, we have to live with the rain. It means we have to take different decisions about how we build the systems.”
Smith says fog, rain or snow can confuse a vehicle’s sensors. To make matters worse, there are also connectivity issues. “It’s quite easy to run autonomy on a big freeway or a test track miles from anywhere, where you have a clear view of the sky and good connectivity,” he explains. But the cars don’t cope so well with big ‘urban canyons’ created by tall buildings, in underground car parks, or in the shadows of the Hammersmith flyover.
Both companies are working on solutions to these problems. FiveAI has eight cars on the streets between Croydon and Bromley, hoovering up data and sharpening their digital eyes (six pairs of high-megapixel stereo cameras per vehicle). Oxbotica has run tests in London’s Milton Keynes and Greenwich, and now has a fleet of six cars operating around Heathrow and Hounslow. It’s also working with Addison Lee, which allows it to place sensors on taxis and use their miles to generate maps. It’s trying to cut reliance on GPS and other forms of connectivity.
The general view from developers is that full autonomy – where everyone is doing every journey in an autonomous vehicle – is a long way off. Deloitte’s 2019 global consumer survey shows there are very serious concerns about safety and cyber security, while a London Assembly study predicts the city won’t be ready for mass rollout of AVs until at least 2030.
In the meantime, fully autonomous cars are expected to appear in safe “islands” around cities, says Smith. These islands will, over time, swell, merge and possibly swallow up inner cities. The approach is to test in controlled spaces (airports, campuses or quiet suburbs), demonstrate safety and move towards complexity. The Department of Transport says it is investing over £250 million to support trials in “safe testing environments”.
Safety is a major concern says Lucy Yu, policy director at FiveAi, who advocates strict rules and a regulator to make sure they’re adhered to. “We strongly believe that creating an independent body to enforce those standards is important to give citizens confidence,” she says. “We and any other competitors should not be allowed to mark our own homework.”
Automotive expert David Bailey, of Birmingham Business School, says there are also big conversations to be had about how AVs will fit into wider transport systems. Benefits include safer roads, cleaner air, more mobility and people spending less time trapped behind a steering wheel. But the last thing anyone wants is city streets clogged up with driverless cars. “We need to talk about what we want from AVs,” he says.
FiveAI's policy director Lucy Yu is speaking at WIRED Pulse: AI at the Barbican on June 15, 2019. Find out more at wired.uk/ai-event
This article was originally published by WIRED UK
