Machines of Loving Grace

by John Markoff

  Google and Mobileye have taken very different approaches to solving the problem of making a car aware of its surroundings with better-than-human precision at highway speeds. Google’s system is based on creating a remarkably detailed map of the world around the car using radars, video, and a Velodyne lidar, all at centimeter accuracy, augmenting the data it collects using its Street View vehicles. The Google car connects to the map database via a wireless connection to the Google cloud. The network is an electronic crutch for the car’s navigation system, confirming what the local sensors are seeing around the car.

  The global map database could make things easier for Google. One of the company’s engineers confided that when the project got under way the Google team was surprised to find how dynamic the world is. Not only do freeway lanes frequently come and go for maintenance reasons, but “whole bridges will move,” he said. Even without the database, the Google car is able to do things that might seem to be the province of humans alone, such as seamlessly merging into highway traffic and handling stop-and-go traffic in a dense urban downtown.

  Google has conducted its project with a mix of Thrun’s German precision and the firm’s penchant for secrecy. The Israelis are more informal. On that hot spring afternoon in suburban Jerusalem there was little caution on the part of the Mobileye engineers. “Why don’t you drive?” Eyal suggested to me, as he slid into the passenger seat behind a large display and keyboard. The engineers proceeded to give a rapid-fire minute-long lesson on driving a robot car: You simply turn on cruise control and then add the lane-keeping feature—steering—by pulling the cruise control stick on the steering wheel toward you. A heads-up display projected on the windshield showed the driver the car’s speed and an icon indicated that the autonomous driving feature was on.

  Unlike the Google car, which originally had a distinctive Star Trek start-up sound, there is only a small visual cue when the autopilot is engaged and the Mobileye Audi takes off down the highway by itself, at times reaching speeds of more than sixty miles per hour. On the road that snakes down a desolate canyon to the Dead Sea, it is difficult to relax. In an automated car, it is very challenging for a novice driver when, before long, a car ahead begins to slow for a stoplight. It takes all of one’s willpower to keep a foot off the brake and trust the car as, sure enough, it slows down and rolls smoothly to a stop behind the vehicle ahead.

  The Google car conveys the detached sense of a remote and slightly spooky machine intelligence at work somewhere in the background or perhaps in some distant cloud of computers. By contrast, during its test phase in 2013, the Mobileye car left a passenger acutely aware of the presence of machine assistance. The car needs to weave a bit within the lane when it starts to pull away from a stop—not a behavior that inspires confidence. If you understand the underlying technology, however, it ceases to be alarming. The Audi’s vision system uses a single “monocular” camera. The third dimension, depth, is computed based on a clever algorithm that Shashua and his researchers designed, referred to as “structure from motion,” and by weaving slightly the car is able to build a 3-D map of the world ahead of it.

  Knowing that did little to comfort a first-time passenger, however. During the test ride, as it passed a parked car, the Audi pulled in the direction of the vehicle. Not wanting to see what the car was “really thinking,” I grabbed the wheel and nudged the Audi back into the center of the lane. The Israeli engineers showed no signs of alarm, only amusement. After a half hour’s drive along an old road that felt like it was still a part of antiquity, the trip was over. The autonomous ride felt jarringly like science fiction, yet it was just the first hint of what will gradually become a broad societal phase change. “Traffic jam assist” has already appeared on the market. Technology that was remarkable to someone visiting Israel in 2013 routinely handles stop-and-go freeway traffic around the world today.

  The next phase of automatic driving will start arriving well before 2020—vehicles will handle routine highway driving, not just in traffic jams, but also in the commute from on-ramp to off-ramp. General Motors now calls this capability “Super Cruise,” and it will mark a major step in the changing role of humans as drivers—from manual control to supervision.

  The Google vision is clearly to build a vehicle in which the human becomes a passenger and no longer participates in driving. Yet Shashua believes that even for Google the completely driverless vehicle is still far in the future. Such a car will stumble across what he describes as the four-way stop puzzle—a completely human predicament. At intersections without stoplights there is an elaborate social dance that goes on between drivers, and it will be difficult for independent, noncommunicating computer systems to solve anytime in the foreseeable future.

  Another complication is that human drivers bend the rules and ignore the protocols frequently, and pedestrians add massive complications. These challenges may be the real barrier for a future of completely AI-based automobiles in urban settings: we haven’t yet been able to wrap our heads around the legal trouble posed by a possible AI-caused accident. There is a middle ground, Shashua believes, somewhere short of the Google vision, but realistic enough that it will begin to take over highway driving in just a couple of years. His approach wraps increasingly sophisticated sensor arrays and AI software around the human driver, who stays in the loop, a human with superawareness who can see farther and more clearly, and perhaps switch back and forth from other tasks besides driving. The car can alert the driver when his or her participation is beneficial or necessary, depending on the driver’s preferences. Or perhaps the car’s preferences.

  Standing beside the Audi in the Jerusalem suburbs, it was clear that this was the new Promised Land. Like it or not, we are no longer in a biblical world, and the future is not about geographical territory, but rather about a rapidly approaching technological wonder world. Machines that will begin as golems are becoming brilliant and capable of performing many human tasks, from physical labor to rocket science.

  Google had a problem. More than three years into the company’s driverless car program, the small research team at the Mountain View–based Internet search company had safely driven more than a half-million miles autonomously. They had made startling progress in areas that had been generally believed impossible within the traditional automotive industry. Google cars could drive during the day and at night, could change lanes, and could even navigate the twisty Lombard Street in San Francisco. Google managed these advances by using the Internet to create a virtual infrastructure. Rather than building “smart” highways that would entail vast costs, they used the precise maps of the world created by the Google Street View car fleet.

  Some of the achievements demonstrated an eerie humanlike quality. For example, the car’s vision system had the ability to recognize construction zones, slow down accordingly, and make its way through the cones safely. It also could adjust for vehicles partially blocking a lane, moving over as necessary. The system had not only been able to recognize bicyclists, but it could identify their hand signals and slow down to allow for them to change lanes in front. That suggested that Google was making progress on an even harder problem: What would a driverless car do when it was confronted by a cop making hand signal motions at an accident or a construction zone?

  MIT roboticist John Leonard had taken particular joy in driving around Cambridge and shooting videos of the most confounding situations for autonomous vehicles. In one of his videos his car rolls up to a stop sign at a T-intersection and is waiting to make a left turn. The car is delayed by a long line of traffic passing from right to left, without a stop sign. The situation is complicated by light traffic coming from the opposite direction. The challenge is to persuade the drivers in the slow lane to give way, while not colliding with one of the cars zipping by at higher speed in the other direction.7

  The video that was perhaps the toughest challenge for the Google vision system was taken at a busy crosswalk somewhere downtown. There is a crowd of people at a pedestri
an crossing with a stoplight. The car is approaching when suddenly, completely ignoring the green light for the cars, a police officer’s hand abruptly shoots out on the left of the frame to stop traffic for the pedestrians. Ultimately it may not be an impossible problem to solve with computer vision. If today’s systems can already recognize cyclists and hand signals, uniforms cannot be far behind. But it will not be solved easily or quickly.

  Intent on transforming education with massive open online courses, or MOOCs, and not wishing to compete for leadership of X Lab with Google cofounder Sergey Brin, Thrun largely departed the research program in 2012. As is often the case in Silicon Valley, Thrun had not been able to see his project through. After creating and overseeing the secret X Laboratory at Google for several years, he decided it was time for him to move on when Brin joined the effort. Brin proposed that they be codirectors, but Thrun realized that with Google’s cofounder in the mix he would no longer be in control and so it was time for a new challenge.

  In the fall of 2011 Thrun and Peter Norvig had taught one of several free Stanford online course offerings, an Introduction to Artificial Intelligence. It made a big splash. More than 160,000 students signed up for the course, which was almost ten times the size of Stanford’s nonvirtual student body. Although only a fraction of those enrolled in the course would ultimately complete it, it became a global “Internet moment”: Thrun and Norvig’s class raised the specter of a new low-cost form of education that would not only level the playing field by putting the world’s best teachers within reach of anyone in the world, but also threaten the business models of high-priced elite universities. Why pay Stanford tuition if you could take the course anyway as a City College student?

  Thrun was still nominally participating one day a week at Google, but the project leadership role was taken by Chris Urmson, a soft-spoken roboticist who had been Red Whittaker’s chief lieutenant in the DARPA vehicle challenges. He had been one of the first people that Thrun hired after he came to Google to start the then secret car program. In the summer of 2014 he said he wanted to create a reliable driverless car before his son reached driving age, which was about six years in the future.

  After Thrun departed, Urmson took the program a long way toward its original goal of autonomous driving on the open road. Google had divided the world into highway driving and driving in urban conditions. At a press conference called to summarize their achievements, Google acknowledged that their greatest challenge was figuring out how to program the car to drive in urban areas. Urmson, however, argued in a posting he made on the company’s website that the chaos of the city streets with cars, bicyclists, and pedestrians moving in apparently random fashion, was actually reasonably predictable. The Google training experiment had encountered thousands of these situations and the company had developed software models that would expect both the expected (a car stopped at a red light) and the unexpected (a car running a red light). He and his team implied that the highway driving challenge was largely solved, with one caveat—the challenge of keeping the human driver engaged. That problem presented itself when the Google team farmed out some of their fleet of robotic vehicles to Google employees to test during their daily commute. “We saw some things that made us nervous,” Urmson told a reporter. The original Google driving program had involved two professional drivers who worked from an aircraft-style checklist. The person in the driver’s seat was vigilant and ready to take command if anything anomalous should happen. The real world was different. Some of the Google employees, on their way home after a long day at the office, had the disturbing habit of becoming distracted, up to and including falling asleep!

  This was called the “handoff” problem. The challenge was to find a way to quickly bring a distracted human driver who might be reading email, watching a movie, or even sleeping back to the level of “situational awareness” necessary in an emergency. Naturally, people nodded off a lot more often in driverless cars that they had come to trust. It was something that the automotive industry would address in 2014 in the traffic jam assist systems that would drive cars in stop-and-go highway traffic. The drivers had to keep at least one hand on the wheel, except for ten-second intervals. If the driver didn’t demonstrate “being there,” the system gave an audible warning and took itself out of its self-driving mode. But automobile emergencies take place during a fraction of a second. Google decided that while in some distant future that might be a solvable problem, it wasn’t possible to solve now with existing technology.

  A number of other automakers are already attempting to deal with the problem of driver distraction. Lexus and Mercedes have commercialized technology that watches the driver’s eyes and head position to determine if they are drowsy or distracted. Audi in 2014 began developing a system that would use two cameras to detect when a driver was inattentive and then bring the car to an abrupt halt, if needed.

  For now, however, Google seems to have changed its strategy and is trying to solve another, perhaps simpler problem. In May of 2014, just weeks after they had given reporters an optimistic briefing on the progress of their driverless car, they shifted gears and set out to explore a new limited but more radical solution to autonomous transportation in urban environments. Unable to solve the distracted human problem, Google’s engineers decided to take humans out of the loop entirely. The company de-emphasized its fleet of Prius and Lexus autonomous vehicles and set out to create a new fleet of a hundred experimental electric vehicles that entirely dispense with the standard controls in a modern automobile. Although it had successfully kept it a secret, Google had actually begun its original driverless vehicle program by experimenting with autonomous golf carts on the Google campus very early in the self-driving program. Now it was planning to return to its roots and once again autonomously ferry people around the Google campus, this time with its new specially designed vehicle fleet. Riding in the new Google car of the future will be more like riding in an elevator. The two-seat vehicle looks a bit like an ultracompact Fiat 500 or Mercedes-Benz Smart car, but with the steering wheel, gas pedal, brake, and gearshift all removed. The idea is that in crowded downtowns or on campuses passengers could enter the desired destination on their smartphones to summon a car on demand. Once they are inside, the car provides the passengers only a Trip Start button and a red E-Stop panic button. One of the conceptual shifts the engineers made was limiting the speed of the vehicle to just twenty-five miles per hour, allowing the Google cars to be regulated like golf carts rather than conventional automobiles. That meant that they could forgo air bags and other design restrictions that add cost, weight, and complexity. These limitations, however, mean that the new cars are suited only for low-speed urban driving.

  Although 25 miles per hour is below highway standards, the average traffic speeds in San Francisco and New York are 18 and 17 miles per hour, respectively, and so it is possible that slow but efficient fleets of automated cars might one day replace today’s taxis. A study by the Earth Institute found that Manhattan’s 13,000 taxis make 470,000 trips a day. Their average speed is 10 to 11 miles per hour, carrying an average of 1.4 passengers an average distance of 2 miles, with a 5-minute average wait time to get a taxi. In comparison, the report said, it would be possible for a futuristic robot fleet of 9,000 automated vehicles hailed by smartphone to match that capacity with a wait time of less than 1 minute. Assuming a 15 percent profit, the current cost of today’s taxi service would be about $4 per trip mile, while in contrast, it was estimated, a future driverless vehicle fleet would cost about 50 cents per mile. The report showed similar savings in two other case studies—in Ann Arbor, Michigan, and Babcock Ranch, a planned community in Florida.8

  Google executives and engineers have made the argument that has long been advocated by urban planners: A vast amount of space is wasted on an automotive fleet that is rarely used. Cars used to commute, for example, sit parked for much of the day, taking up urban space that could be better used for housing, offices, or parks. In urban areas, automated
taxis would operate continuously, only returning to a fast-charging facility for robots to swap out their battery packs. In this light, it is easy to reimagine cities not built around private cars, with more green space and broad avenues—which would be required to safely accommodate pedestrians and cyclists.

  Thrun evoked both the safety issues and the potential to redesign cities when he spoke about the danger and irrationality of our current transportation system. In addition to squandering a great deal of resources, our transportation infrastructure is responsible for more than thirty thousand road deaths annually in the United States, and about ten times more than that in both India and China, which amounts to more than one million annual road deaths worldwide. It is a compelling argument, but it has been greeted with pushback both in terms of liability issues and more daunting ethical questions. An argument against autonomous vehicles is that the legal system is unequipped to sort out the culpability underpinning an accident that results from a design or implementation error. This issue speaks to the already incredibly complicated relationship between automotive design flaws and legal consequences. Toyota’s struggle with claims of unintended acceleration, for example, cost the company more than $1.2 billion in damages. General Motors also grappled with a design flaw involving sudden stops because of a faulty ignition switch that resulted in the recall of more vehicles than they manufactured in 2014, and may ultimately cost several billion dollars. Yet there is potentially a simple remedy for this challenge. Congress could create a liability exemption for self-driving vehicles, as it has done for childhood vaccines. Insurance companies could impose a no-fault regime when only autonomous vehicles are involved in accidents.