EVERY day for two months, Sebastian Thrun drove Stanley, his customised Volkswagen SUV, through the Sonoran desert, Arizona, burning down dusty roads and avoiding boulders and potholes. With every twist and turn, Stanley鈥檚 on-board computers watched and learned. Then one day Thrun did something most people would find rather scary. He turned over control of the car to the computers. 鈥淚 learned to trust it, and pretty soon the car was driving me,鈥� says Thrun.
Thrun, director of the artificial intelligence lab at Stanford University, California, was preparing Stanley for the ultimate race for robot cars. On 8 October Stanley took part in the Grand Challenge, sponsored by the Pentagon鈥檚 Defense Advanced Research Projects Agency (DARPA), in which 23 self-guiding vehicles were pitted against one another to navigate a 212-kilometre course over dirt roads, around narrow mountain tracks and through tunnels to the finish line just outside Primm, Nevada.
Stanley won, in just 6 hours and 53 minutes, netting Thrun and his team a cool $2 million in prize money for their efforts. In the end, five of the 23 entrants finished the race (see Graphic). That might sound like a poor success rate, but consider this: in the only previous Grand Challenge, in March last year, none of the entrants managed to complete the course. The most successful car only made it 11 kilometres into the race before disaster struck (快猫短视频, 20 March 2004, p 24). This year, four of the five finishers made it in less than the prescribed 10 hours.
Advertisement
The question on everybody鈥檚 lips is simple: what changed in the intervening year that had such a dramatic effect on the results? Was there a miracle breakthrough in the quality of the AI guidance systems in the cars? Or did a combination of luck and more prosaic changes 鈥� a more forgiving course and more time to refine sticky engineering problems 鈥� help more competitors to the finishing line? Most importantly, what can we learn from the event to shape the future of autonomous vehicles?
In the weeks since the race, the teams have been poring over the data from their cars in an attempt to understand what went wrong and what went right. 鈥淭he biggest distinction between 2005 and 2004 was the quality of design, integration and testing,鈥� says William 鈥淩ed鈥� Whittaker, leader of the team from Carnegie Mellon University in Pittsburgh, Pennsylvania, that fielded two of the five cars that finished. Thrun claims it was his attentive AI system that gave his car the edge: 鈥淲e won with good luck and the most precise software,鈥� he says.
Some observers of the race have been quick to link the successes with general advances in AI research. 鈥淚t dramatically demonstrates a set of research advances in machine vision, probabilistic modelling, estimation and path planning that have evolved in laboratories and field experiments over the past decade,鈥� says Ken Goldberg, a roboticist at the University of California, Berkeley, who was not involved in the race. But others remain to be convinced that breakthroughs in AI made a real difference this time. 鈥淭he robotic road race community owes us a sound, clear and useful explanation of what, if anything, they have learned,鈥� says Marvin Minsky, a pioneer in AI at MIT.
One thing that did change this year is that the organisers gave the teams three times as much preparation and testing time for the second race. They were also helped along by a new course, 16 kilometres shorter and with fewer hills. Last year the most difficult portion came near the beginning, where the leader of the race got stuck. This time the wickedly narrow and windy cliff-side road with a 30-metre drop, called Beer Bottle Pass, came towards the end of the course, meaning the cars got plenty of easy ground to cover before reaching the really challenging parts.
But don鈥檛 underestimate the abilities of this year鈥檚 finishers. The winner, Stanley, is a tricked-up Volkswagen Touareg R5 SUV with GPS sensors, radar, accelerometers, five laser range finders, video cameras and six computers to crunch all the data. The key to victory, says Thrun, was Stanley鈥檚 software 鈥� a computer program that was trained to drive on the sort of terrain the racecourse crossed.
To train for the race, Stanley was driven across more than 2000 kilometres of desert tracks. The car鈥檚 sensors observed the terrain it was passing through and what its human drivers were doing to stay on course. Then it used this data as a guide for what to do when it started driving itself. For example, Stanley learned from observation that when faced with uneven terrain or a steep slope it should slow down. When confronted with unprecedented situations, Stanley鈥檚 computers used a complex algorithm devised by Thrun鈥檚 team to combine training data with inputs from sensors, sampled 10 times a second, to make an educated guess about how to proceed.
But was it, as Thrun says, this sophisticated AI that clinched the race for Stanley and the other finishers? It鈥檚 hard to say, but other competitors were put out of the race by failures in more commonplace matters. The best of the non-finishers, a vehicle from technology company Ensco, got a flat tyre after 130 kilometres. Others suffered sensor malfunctions: a laser range finder on the rooftop of the Digital Auto Drive car came loose and lost power. The lesson here? Keep it simple and sturdy. 鈥淢ore vehicle hardening and a less sophisticated sensor would have won us the race,鈥� says Ann Gargiulo, spokesperson for Digital Auto Drive, whose robot made it 42 kilometres.
Although DARPA has declared the Grand Challenge conquered and is not planning to sponsor a third race, the problems of robot navigation for land vehicles remain far from solved. For example, how does a robot know how to avoid new kinds of obstacles reliably? 鈥淚t鈥檚 hard to distinguish true obstacles from stuff the robot can easily drive through,鈥� says Larry Matthies, head of the computer vision group at NASA鈥檚 Jet Propulsion Laboratory in Pasadena, California. Being able to tell a pit from a shadow, for instance, could mean the difference between a happily functioning planetary rover and a useless heap of metal millions of kilometres away.
Nevertheless the researchers and sponsors are united behind the idea that the race results will promote the use of AI and autonomous vehicles beyond next-generation military vehicles and planetary rovers. In the next few years, they say, expect to see AI-based systems for improving car safety: things like avoiding moving-traffic collisions and detectingout-of-lane swerving. These would go far beyond today鈥檚 most advanced commercial systems, such as proximity sensors for parking. Eventually, your car might even observe your bad driving habits and learn to take corrective action. But Thrun concedes that practical systems for completely self-driving cars and automated public transit are still decades away.
So what conclusions can we draw from the winners? The most obvious one is that the challenge was well within today鈥檚 technological capabilities. And it鈥檚 also clear that the difference between this year鈥檚 five finishers and last year鈥檚 wannabes is not down to any single breakthrough. Rather, the improvements in AI and engineering that allowed cars to succeed this year were brought about by a longer design and testing period. Once again, slow and steady wins.
