Preparing Self-Driving Cars for the Wild World of Developing Cities
Self-driving cars are no longer confined to controlled test tracks or even to placid suburban streets—they’re tackling real traffic in US cities such as New York, San Francisco, and Pittsburgh. They’re honing their skills amidst humans in Europe, South Korea, Singapore, and Japan. They’re preparing for the day they can purify our chaotic streets with their robotic perfection.
Learning how to drive in places like unruly Boston, a land of creative left turns and seemingly optional yields, comes with its challenges. But the aggressive driving and the complexity of the city’s twisting streets pale in comparison to the developing world. Even Patriots fans look like goody two-shoes compared to drivers who have little to zero respect for lanes, traffic signals, warning signs, and speed limits.
On wide roads without lanes and huge, anarchic intersections all over the world, human interaction dictates traffic flows, with each driver adjusting to others’ maneuvers on the spot, regardless of what the rule book says.
These informal systems work for the most part, but they come at a high cost. Of the 50 countries with the deadliest roads, 44 are in Africa or the Middle East, according to 2013 figures from the World Health Organization (the most recent data available). Together these nations accounted for nearly 250,000 deaths in 2013—a fifth of the world’s total.
Yet the factors that make these places the most likely to benefit from autonomous cars also make them the least likely to get the technology anytime soon.
“Many of the things that we’re doing in self-driving at the moment probably wouldn’t work if we were trying to do it in a third-world country,” says Ram Vasudevan, codirector of the University of Michigan’s Ford Center for Autonomous Vehicles.
Autonomous driving requires understanding the intent and trajectory of everyone and everything on the road: vehicles, cyclists, pedestrians, construction workers, playing children, pets, an errant dart from a Nerf gun. In driving environments governed by a set of rules that people actually follow, the law limits the sorts of behaviors an autonomous vehicle should expect in the world around it.
The fewer formal rules in place, the more the ability to predict intent matters. Around wild humans, cars can’t rely on shared guidelines to dictate behavior. Basic driver assists that keep cars inside painted lanes, for example, are only useful if everyone else on the road respects lane markings. Otherwise they’re useless, or even dangerous.
Compared to suburban and even urban America, driving environments in many Middle Eastern and African countries have all the structure of a jellyfish. In Lebanon, where I live, it’s common to see cars driving the wrong way, running red lights, and zigzagging across wide roads without the slightest regard to lane markings, among other shenanigans.
“There are no rules here. Everything is possible,” said Daniel Asmar, a computer-vision expert and engineering professor at the American University of Beirut. “Humans can deal quite well with that, even if they get frustrated and honk at each other.” For computers, the chaos would be an enormous challenge.
Even in relatively orderly environments, a confusing situation such as a freeway merge can make a self-driving car hesitate long enough to hold up traffic or even cause an accident, Vasudevan says. This might be because the car’s software, erring on the safe side, isn’t willing to merge in front of a speeding car, or because the car needed more time to understand the scene around it and the intent of other drivers. Put the same car on a road where stop signs, traffic signals, and yielding rules don’t exist or are routinely ignored, and its reaction times will need to be a great deal sharper to survive.
What’s more, self-driving cars need the help of mapping data that doesn’t yet exist in most parts of the world. Autonomous driving requires highly detailed street maps that contain everything from the height of street curbs, to the location of temporary construction detours, to the exact position of street signs and traffic lights in 3-D space. Those maps have already been developed for cities with self-driving fleets, and they’re constantly being updated using data that autonomous cars capture as they drive around.
In places like Lebanon, where two-dimensional Google and Apple Maps contain basic mistakes, missing data is an enormous disadvantage. Even if detailed maps existed, they would require intensive upkeep. “In a structured environment, you wouldn’t have to do it that often, because things are pretty much staying the same,” Asmar says. “In an unstructured environment, where things are changing all the time, you can imagine how many times you have to keep building this platform over and over again. It’s a really daunting task.”
A few wealthy countries in the Middle East are already moving toward autonomous driving. Israeli companies are behind important developments in autonomous driving software, and the country opened its first test track for driverless cars last month. In Dubai, a 10-seater driverless shuttle began trundling through a riverside business district last year. City officials are aiming for a quarter of local trips to be made without a driver by 2030, and Dubai’s police force is planning to roll out tiny self-driving patrol cars by the end of the year.
But it appears India and China are the only countries that contain both driving chaos and local companies developing autonomous vehicles. Unsurprisingly, their efforts face extra hurdles. India’s Tata has created a testing track outside Bangalore to simulate local roads, complete with fearless pedestrians and stray cattle, Bloomberg reported. The company still has a long way to go: Its computer-vision systems currently fail to identify 15 percent of vehicles on Indian roads, a senior vice president at Tata told Bloomberg, because of the sheer variety in their shapes and sizes. (When former Uber CEO Travis Kalanick visited India last year, he joked that the country would be “the last one on earth” to get self-driving cars. “Have you seen the way people drive here?”)
China’s Baidu, meanwhile, is openly working on autonomous driving, teaming up with more than 50 international companies to develop its software. In a recent video demo, Baidu CEO Robin Li sat in a self-driving car as it wound its way through Beijing traffic—making a few unsafe maneuvers along the way. Since self-driving cars aren’t currently road-legal in China, Chinese police said they’d investigate whether Li broke any laws. (India is moving toward a similar ban, citing concerns about job losses.) Despite the regulatory hurdles, Baidu’s president, Ya-Qin Zhang, told Bloomberg that he’s confident that the company’s autonomous cars will be on the road “as early as next year.”
Didi Chuxing, the reigning ride-hailing company in China, is taking a much more measured approach. Although it opened an office in California earlier this year to develop autonomous driving technology, the company’s president, Jean Liu, said in a recent interview with Charlie Rose that a sudden, “disruptive” switch to autonomous driving would be dangerous. “I think people should be more, you know, focusing on how safe it is [rather] than how soon it can come out,” Liu said.
In China, autonomous vehicles won’t just have to learn to deal with cars, electric scooters, and pedestrians that don’t follow the rules, a Didi spokesperson said—they would need to be able to understand regional differences in signage and traffic signaling, which aren’t standardized in China like they are in the US or Europe. There, Didi’s size offers it an advantage. The company says its human drivers give 25 million rides every day, generating more than 70 terabytes of data daily that it can mine to develop its autonomous driving capabilities.
Following the Leader
For now, many companies are testing their autonomous vehicles by throwing unexpected scenarios at them on controlled tracks. At Castle, Waymo’s secret compound for training its cars, human assistants cut off self-driving minivans at high speed, back out of blind driveways into their path, and throw basketballs at them, all to test and improve the cars’ reactions.
But artificial intelligence that’s trained on one set of assumptions can fail when it meets a different set. Studies have found that facial-recognition algorithms trained on Caucasian test subjects perform poorly on African American faces, and algorithms trained on East Asian subjects perform poorly on Caucasian faces. The same might go for self-driving cars. Software trained on worst-case scenarios that involve flying basketballs and dicey merges might freak out at the sight of two dudes hanging out the back of a station wagon on a fast-moving highway.
Despite vast regional variations in how people drive, manufacturers might not have to create a Ghana version and an Iran version and a Southwest India version of their driving software. “It’s really the same math and the same software that’s going to exist in every cultural context,” says Matthew Johnson-Roberson, a University of Michigan engineering professor and the Ford Center’s other codirector.
What matters most is that cars are trained to react to all of them. A spokesperson for Uber, which is testing self-driving cars in the US and Canada, said that its cars have driven more than a million autonomous miles in multiple cities, under different conditions and during different times of day, in order to improve its software’s adaptability.
Even if self-driving software understands unruly drivers and can predict how they’re likely to break the law, autonomous vehicles will probably be constrained by it. Uber’s cars will always follow local traffic laws, a company spokesperson says. Stephan Hoenle, senior vice president of automated driving at Bosch, agrees. “You can drive more aggressively or defensively without breaking the rules,” Hoenle says. An autonomous vehicle’s driving style might vary from one market to another based on demand and expectations, but violating the law isn’t an option—it’s too great a liability for a manufacturer.
The problem is that in some places, driving according to the letter of the law could be more dangerous than aping law-breaking human drivers. Failing to adjust when impatient commuters turn a two-lane road into a four-lane highway by driving on the shoulder during rush hour can quickly lead to an ugly pileup.
Back of the Line
To someone steeped in the day-to-day work of teaching computers to drive better than humans, the details of where self-driving cars will end up might not seem very pressing. “It doesn’t even work here, right?” said the University of Michigan’s Johnson-Roberson. “From an engineering perspective, I don’t know anyone who’s working on this, because some of the fundamentals are still not there.”
Putting off these questions risks shunting the very regions that most need self-driving technology to the very end of the line. Hoenle claims no part of the world will be excluded from self-driving cars’ eventual rollout but acknowledges it won’t happen all at once. Compared to the US and Europe, he says, “normally some of these other continents have a slower technology ramp-up curve.”
The developing world will eventually catch up, predicts Carlo Ratti, the director of MIT’s Senseable City Lab. “Every technology needs to start somewhere—and often it starts at the cutting edge,” he wrote in an email. “At the beginning, new technologies can increase existing societal gaps between the haves and have-nots. However, the subsequent dissemination of technology can cause interesting ‘leapfrogging’ effects and help reduce gaps.”
Mobile phones, for example, were at first only available to rich Westerners. Now they’re abundant in Africa, where startups are coming up with new ideas for mobile banking and healthcare provision. “There is no reason to think that self-driving cars will follow a different path,” Ratti said.
The gap between introduction and the “leapfrog” stage might be considerably longer for self-driving cars, which have to adapt to their surroundings, need gobs of data specific to each street they drive, and have the potential to kill if poorly designed.
Developers that put off questions about regional differences and leave matters to the “ramp-up curve” will be locked out of an immense market. And as their lifesaving autonomous technology rolls onto friendly roads in places such as North America, Europe, and Singapore, it may leave behind the developing countries that most desperately need that technology.