This article was taken from the January 2012 issue of Wired magazine. Be the first to read Wired's articles in print before they're posted online, and get your hands on loads of additional content by subscribing online.
At 8.20am on a Sunday morning, on a ten-lane carriageway that runs towards the Brandenburg Gate from west Berlin, hundreds of drivers are doing what comes naturally: mostly travelling above the speed limit of 55kph, drifting out of their lanes and fighting for space. This morning, however, one car will be doing its own thing. A heavily modified VW Passat, developed by a 20-strong team, will twice drive itself up and down the Kaiserdamm -- a 29km journey -- without human intervention.
There are four of us in the car, including Tinosch Ganjineh, a PhD student and research assistant, who sits in the driver's seat.
Ganjineh is the safety driver: he will intervene at the slightest sign of trouble from the car's systems. Beside him sits researcher Miao Wang. On his lap is a quad-core notebook fed by an ethernet cable from the car's boot.
As we pull out into the traffic, Wang's voice interrupts the classical music on the car's sound system. "Engage," he says.
Ganjineh presses a button and the car's autonomous systems take over. Ganjineh is mostly silent. When he speaks, it's in the low, flat, tones of someone intensely focused. "I have to watch the traffic closely," he says. "When something goes wrong I can press the button to take control, or simply touch the steering wheel or one of the pedals. When you hear a beep, I have control."
Our ride is more fitful than with a human driver. Twice during the 13-minute journey down the Kaiserdamm, Ganjineh's hands hover millimetres from the steering wheel. And it's tense: first a bus, then a lorry, start moving out of adjoining lanes, veering towards us. On both occassions, the car pulls sharply to the left, but then reverts to its original trajectory as the human drivers correct their course. Wang warns Ganjineh each time the car seems likely to make a decision. The verbal feedback loop, explains Raúl Rojas, professor of artificial intelligence at the Free University of Berlin and the leader of the project, is designed to reduce the stress experienced by the safety driver. "Watching the steering wheel moving by itself is a little uncomfortable,"
Rojas says. "You wonder if it's going to be OK."
On his notebook screen, Wang is monitoring the car's vital signs and a 3D representation of the outside world. A green line carves out the car's projected route. Obstacles such as the back of a bus are outlined in red. On the road, a woman on a bicycle comes into view. She is drifting across the lanes, into the traffic, carrying a child in a rear baby seat. Back on Wang's screen, the swarming lines have resolved into a solid red outline. The car has noticed another, more vulnerable, road user approaching. The car passes safely and the image disappears from the screen.
The Passat is the result of a €2.3 million (£2 million) investment by the German Federal Ministry of Education and Research. One of a handful of autonomous-driving projects in Europe and the US, Rojas's project has the potential to transform our relationship with that crowning achievement of 20th-century capitalism: the motorcar.
Self-driving cars are no longer science-fiction fantasy. The technology is mostly proven. And, as Jaron Lanier, the San Francisco-based computer scientist who has written extensively on the interface between the virtual and psysical worlds, puts it: "If you get in a driverless car, you are less likely to die and more likely to get where you want to go faster. You'll also do less damage to the environment." As a result, pushed by a need to unclog our cities, reduce pollution and save lives, government agencies around the world have started to think seriously about a future in which robotic precision replaces flawed human driving skills on our roads.
After the test drive, Rojas and Ganjineh offer a tour of the car's workings. It starts with an array of yellow and blue plugs in the boot. The car's software sends its instructions into this drive-by-wire gateway, ordering it to adjust its speed, or change lanes. Before it can decide what to do, the car needs to analyse data harvested from outside. Some of that data comes from six sensors on the car's sides. On top of the vehicle is a revolving scanner equipped with 64 laser beams that assimilate 1.3 million data points every second. "It's highly accurate," says Ganjineh. "It gives you a really good picture of the car's surroundings, including pedestrians, animals and small vehicles."
Behind the windscreen, a mono camera detects road markings, allowing the car to centre itself in its lane. Two colour cameras identify traffic lights, sending the appropriate stop or go message to the car's software platform. On the roof sits a €90,000 GPS receiver. The final piece of kit is an electronic route map on Wang's notebook.
Expensive as it may be, this hardware isn't revolutionary.
Instead, it's the data-interpreting software -- written by Rojas's team -- that gives this car the ability to "see" a cyclist. The team programs the car to respond to this data. And, says Ganjineh, one thing matters more than most: "Professor Rojas always says my grandma should be able to drink a cup of coffee on the back seat."
When Google's self-driving cars started attracting attention in September 2010, you could have been forgiven for thinking that the visionaries from Mountain View had invented autonomous driving. Not so. Research into self-driving cars has a lengthy pedigree in the US and Europe. Yet the story of how Google became a big player in the small world of autononomous driving certainly testifies to the freewheeling commercial imagination of
Larry Page, Google's cofounder and chief executive.
In 2007, Page visited the Stanford University offices of German academic Sebastian Thrun. The youngest-ever director of the Stanford Artificial Intelligence Laboratory (Sail), Thrun had led Stanford's autonomous driving team to victory at the second DARPA Grand Challenge two years earlier. Page had been present as the team's heavily modified VW Touareg beat 22 other vehicles to complete the 212km race across the Mojave Desert in six hours and 53 minutes. By 2007, Google was rolling out Street View in the US. At their meeting that year, according to one source, Page and Thrun spoke for an hour about the future of cars and mapping. Google agreed to fund Thrun's research.
Thrun, in turn, would work on Street View. The way was clear for Google to create its own driverless car.
In September 2010, Google went public with its efforts. The following June, Nevada became the first US state to allow autonomous vehicles on public roads. Recently, however, Page told investors to discount media speculation about "the latest crazy thing that Google did". "We are not betting the farm on a lot of those things," he added. Google's self-driving cars are tiny part of the company's R&D budget -- the team has only recently exceeded 15 staff.
Rojas believes that Google is intrigued by the possibility of vehicles forming an intelligent swarm. A spokesperson says that "how to get the right information to the cars at the right time" is a major preoccupation. Google sees its data centres as the key.
Processing what a spokesperson calls "the enormous amounts of information gathered by these cars when mapping terrain" is a major exercise.
Mapping isn't unlike search. Destinations are analogous to the web pages that Google indexes and identifies in response to user queries. In theory at least, there's no reason why self-driving cars equipped with Google's maps can't generate income in a similar way to search, a business that generates tens of billions of dollars each year. Invited by Wired to comment on its plans, Google simply responded that: "We are a technology company, and as such our executives are always interested in cutting-edge ideas."
In his office the next morning, Rojas acknowledges that there is big money to be made. He says that the German Ministry of Education and Research want him to produce "patents or startups or licensed technology". Rojas wants to secure further government funding for a startup. The team is building software to help Berlin's refuse trucks manoeuvre more safely. A company that makes $3m (£1.8m) supertrucks for open-cast mines is interested. Aircraft could taxi more safely if they were equipped with $10,000-worth of sensors, says Rojas. "It's a sensible investment," he adds. "It would be easy to automate an airport."
Ultimately, however, the AutoNOMOS Project aims to make mass-produced self-driving cars a reality by slashing their cost. To do this, says Rojas, the industry will need plug-and-play AI systems. There remain many technical problems.
Traffic lights and signs pose challenges. "Bright sunlight can saturate the camera image and video sensors need to improve," says Rojas. Understanding pedestrians' intentions isn't easy. "When you are driving and you see someone, you can interpret by their posture. If a pedestrian is smoking, for example, he is probably not going to cross the street immediately." Computers find this kind of interpretation tricky.
Research into autonomous driving, Rojas says, is dominated by a dozen or so universities in Europe and the US, where trials are under way. In this world, simultaneous localisation and mapping (Slam) has emerged as the standard approach. In this respect, Rojas's project is very similar to the Google scheme. "We use the same kind of sensors," says Rojas, "the same kind of algorithmic approach, as well as maps and GPS."
When Rojas's car pilots itself down the Kaiserdamm, GPS provides its co-ordinates. The car then locates itself on the map, which also tells it where hazards such as traffic lights are located. "We also need to know the number of lanes, and which of them are open to traffic," says Rojas. "So we take a map of the city and we enrich it, we add our own annotations." In Google's case, this approach is augmented by what Professor Jürgen Schmidhuber of the Switzerland-based Dalle Molle Institute for Artificial Intelligence calls "the massive brute-force approach of Street View".
But if Slam represents the mainstream, it's not universally loved. A former Nasa employee called Ernst Dickmanns pioneered a rival method in the 80s that relies heavily upon interpreting the output from video sensors in real time. In 1995, Dickmanns rigged up a Mercedes that drove itself at up to 175kph along 1,758km of motorway between Germany and Denmark.
Fifteen years before Google, Dickmanns' Mercedes achieved it without GPS or maps.
Today, Dickmanns' heir, Joe Wuensche, persists with the same approach. Talking by phone from his office in Munich, Wuensche points to the obvious weaknesses of GPS. The low-strength satellite signals on which it relies are jammable, courtesy of £20 hardware devices readily available on eBay. "And to use GPS you need maps that are very accurate," he adds. "If you have road construction going on, no one is going to tell you that until they pop up on your way to work." Google's approach, Wuensche says, is "typical of computer scientists" who like to "deploy massive computing power".
Not only is the Munich approach lightweight, but it also mimics nature. "How many creatures on Earth rely on ultrasonic distance measurement?" asks Wuensche. "Bats and dolphins: that's about it.
In nature, vision has won. There must be a reason for that.
Building self-driving cars that rely upon vision is harder and it will take longer," he says. "But it will be the cheaper and better solution in the long term."
According to the World Health Organisation, 1.2m people die each year in car accidents globally. In 2009, 2,337 deaths in the UK were a result of road accidents. The deaths that resulted from accidental drowning (205), fire (279), assault (318) and alcohol poisoning (179) is tiny by comparison. Road accidents are the leading cause of death among people aged between 16 and 24.
Google's Thrun believes that the vast majority of road-based fatalities are caused by human error and could be prevented by machines.
Humans aren't just dangerous when they get behind the wheel; they're also inefficient. Today, according to research from Texas A&M University, the average American commuter spends 34 hours a week stuck in traffic jams. Congestion, it suggests, costs the US economy $100bn a year, or roughly $750 per commuter. So clearly, self-driving cars could do a lot to change this. Rojas foresees commuters hopping into an autonomous car that has driven itself to their front door. The car picks up other passengers on the way to the nearest suburban rail station, and from there commuters travel into the city. Rojas has forecast the potential impact using a digital model of Berlin's transport system. "You could move the entire city with one-tenth of the cars that you have now," he says.
Further, sales of autonomous vehicles to the developing world, whose conurbations are increasingly clogged with traffic, could provide a billion-dollar revenue stream, Rojas says. "I think it will work out just fine for the car industry," he says, smiling.
In the meantime, other projects are winning funding. In France, for example, a small VC-backed company called Induct is touting the Cybergo, a self-driving electric vehicle designed for what the company's founder Pierre Lefevre calls "the last kilometres that lie between public transport and your destination in pedestrianised city centres". This suggests that Rojas's vision may yet play out in much the same way as London's " Boris Bikes", or the recently launched €200m electric-car-sharing scheme in Paris.
The Frankfurt Motor Show is a temple to what Volkswagen used to call Fahrvergnügen ("driving enjoyment"). The clubby, male atmosphere at the city's hotels at night disguises a radical shift that is transforming the car industry. By one estimate, electronics and software now account for 80 per cent of innovation in the industry. Today, the average luxury car contains up to 100 microprocessors; software and electronics account for 35 to 40 per cent of its cost. This proportion will rise as more hybrid and electric vehicles become available. The Volt, a hybrid launched by Chevrolet last year, contains ten million lines of code. By comparison, a Lockheed Martin F-35 fighter relies upon 5.7 million lines of code to fly at speeds of more than 1,900kph. Wuensche suggests that most of the vehicles launched at Frankfurt this year qualify as robots.
When the history of autonomous driving is written, its authors will probably choose the introduction of adaptive cruise control (ACC) a decade ago as their starting point. Cars equipped with ACC can detect vehicles ahead of them. When this happens, an ACC-equipped car slows down and maintains a safe distance, only returning to its original speed when the road ahead is clear.
Each year, automation pushes a little further. If an inattentive driver gets behind the wheel of Volvo's current-model S60, the car can detect an oncoming pedestrian and apply braking power of up to 1G -- enough to bring a vehicle travelling at 35kph to a halt in one second. Peter Miller, from UK-based automotive engineering consultancy Ricardo, says this is a turning point. "We're now going beyond systems that react at the last minute to reduce the damage in a collision," he says.
Technology such as this has begun filtering into the mass market. Ford's latest Focus keeps inside lane markings, recognises road signs and automatically brakes at less than 32kph if the car in front stops unexpectedly. Such innovations are being boosted by a Moores Law-like fall in the cost of radar sensors. "The downward price curve on radar sensors is very aggressive," says Mike Thoeny, director of electronic controls engineering at US component supplier Delphi. "Our prices will fall by another 50 per cent by 2013."
Echoing Joe Wuensche, Miller says: "You could run most of these safety systems without radar. The price of vision sensors has fallen more rapidly in price than radar. We have the huge demand for these sensors in smartphones to thank for that."
At Frankfurt, there's evidence of ambivalence. Some executives talk about technology "entering" the vehicles, as if an alien invader. Is there a split in the car industry?
Not quite, says Wuensche. "There are people who look ahead and openly speak out, and there are others who are still afraid of speaking out." Jürgen Leohold, director of research at VW Group, is one of the few senior industry figures willing openly to discuss autonomous driving. "We will see further innovation," he says. "On an autobahn, you often follow a crew of cars and it's really boring," he says. "The next step is to automate the car in this kind of situation."
Beyond that, there is VW's Temporary Auto Pilot, a semi-automated driving platform unveiled by Leohold last summer, which he expects to take "some years" to reach the market. "In the long run we will have an autopilot in the car," he says.
Back in Berlin, Rojas thinkshe knows how to overcome the weakness of Slam. "What Google and other companies would like to have," he says, "is a kind of social network on our streets.
So when lanes on the highway are suddenly closed, for example, your car already knows what is coming up." Rojas is anticipating the arrival of the internet of things -- whereby physical objects are tracked on the network -- on our streets. To make this happen, autonomous cars will need to communicate with one another and with traffic lights, road signs and transceivers on the roadside.
Alan Winfield of the University of the West of England foresees vehicles behaving like a "biological swarm, a distributed autonomous system". Cars will drive themselves in the same way that starlings fly in a murmuration. To carve the optimal route, Winfield says, each connected car "will only need to interact with the ones next to it, or nearby". "Within this decade, we will see this technology rolled out more or less worldwide," says Leohold. Within two years, he says the penetration rate may be five to ten per cent.
On a test track outside Gothenburg, Stefan Solyom, a technical specialist at Volvo, is working to realise these benefits. For the past few years, his team has been collaborating with researchers from the UK, Spain and Germany on ways of allowing lorries to travel in self-organising convoys.
Using radar, cameras and the Wi-Fi standard 802.11p for real-time communication between vehicles, Solyom's lorries mimic the speed and trajectory of the convoy's lead vehicle. "The other drivers can have breakfast or watch a movie," Solyom says, only half-joking. Drivers travelling at 90kph typically like to remain at least one second -- or up to six vehicle lengths --behind the vehicle in front. In Solyom's convoys, lorries travel much closer: as little as half a vehicle length from one another. As a result, each lorry travels in its predecessor's slipstream. "You can cut fuel consumption by 50 per cent and reduce congestion," says Solyom.
Fixating on such benefits is easy. But there will be side effects. In San Francisco, Jaron Lanier predicts millions of job losses. "Driving a lorry or a taxi is one of the basic ways of earning a living around the world," he says. "This is non-trivial."
Lanier can see new business models emerging. To understand how this might happen, he recommends thinking about Rojas's scenario: a road on which three lanes unexpectedly narrow to two because of roadworks. "There's always a point at which there's no choice but for human drivers to get in line," says Lanier. "But every time two cars meet, there's a tiny time loss. Cumulatively, the results are insane. We all pay for this." The solution involves intelligently networked cars organising themselves -- because humans, it seems, can't. "The cars will save time by merging as far back down the freeway as necessary to maintain speed all the way through," Lanier adds.
Some road users may decide to "pay more money to end up further up the queue", says Lanier. Among the results would be the birth of what he describes as "a little stock market, or a bidding market" for levels of access and rights amid the swarm. "You will have brokers and analysts and this whole little world of people organising this. There will be people on top of the algorithms, making sure it all works because they'll have liability if it doesn't. This will create jobs."
For more than a century, our relationship with the internal combustion engine has proceeded along the lines of a Faustian pact.
We've gained much, but the costs can be measured in millions of deaths, endemic gridlock and a degraded environment. Autonomous driving could help to reset this pact. Most of the challenges aren't technical; they're political, economic, legal and social.
The question is whether we're ready for the machines to take control.
This article was originally published by WIRED UK
