Each robot in the Urban Challenge last weekend seemed to have a personality all its own. (See “Champion Robot Car Declared.”) Some cars, like Stanford’s Junior and Carnegie Mellon’s Boss, accelerated through stop signs and rounded curves confidently. Other vehicles, like the University of Pennsylvania’s Little Ben and MIT’s Talos (see “The Land Rover That Drives Itself”), appeared more tentative and cautious.
Talos, in particular, raised eyebrows with a series of jerky movements on a dirt road, slow driving and pausing on paved roads, one collision, and a near head-on crash with another car. While Talos was eventually named the fourth-place finisher, its behavior was surprising, especially since it had many more sensors and much more computing power than any other car on the course. Was Talos overengineered?
After the awards ceremony last Sunday, I caught up with John Leonard, the Talos team lead and a professor of mechanical and ocean engineering at MIT. I learned that, essentially, MIT had designed its car with a handicap that none of the winning cars had: it didn’t rely extensively on Global Positioning System (GPS) sensors. When the MIT engineers started working on the project, Leonard explained, they believed that there would be minimal access to GPS data. After all, he said, GPS can be unreliable in dense urban environments and jammed in military settings. “We first and foremost used perception [technologies such as laser range finders, radar, and cameras],” Leonard said. “GPS was a backup.”
In a press conference after the awards ceremony, the leaders of the winning teams said that they prepared for the final competition by programming the GPS coordinates of points along the course, once they knew it, into their vehicles, effectively giving their cars breadcrumbs to follow as they cruised the roads during the race. This was allowed by DARPA (the U.S. Defense Advanced Research Projects Agency), which sponsored the event, but Leonard said that he didn’t realize that this would be an option. “We were caught by surprise that DARPA let people go through the course and let the teams have an a priori look,” he said. “Our vehicle might have looked cautious, but what was going on inside was that it was looking at radar and other sensors” to find its way.
Leonard says that his team was also surprised that Talos was expected to drive on a dirt road. It was something the researchers hadn’t prepared for, and Leonard suspects that their software was mistaking the dirt-mound curbs for stationary objects that needed to be avoided. With respect to the near head-on crash, Leonard says that his team and the CarOlo team agreed that CarOlo was at fault. As for the collision with Cornell’s car, Skynet, DARPA director Tony Tether said that the race officials had not determined which team was at fault. Skynet had stopped in the road. Talos approached Skynet from behind, and concluding that it was a stationary object, decided to pass it. While Talos was curving back into the lane after passing, Skynet moved forward and hit it. Neither bot was severely damaged, and Skynet also went on to finish the race.
The Urban Challenge was MIT’s first try at a DARPA autonomous-vehicle race. In the first Grand Challenge, in 2004, no car completed a 150-mile course along a desert road; at the second Grand Challenge, in 2005, Stanford’s car won, and Carnegie Mellon’s came in second.
Why did MIT finally throw its hat in the ring? “To us, the Urban Challenge seemed an impossible problem,” said Leonard. “To try to navigate without GPS and detect static and moving objects seemed extremely challenging. We’re drawn to solving hard problems.” Also, Leonard adds, the previous races took place in a desert, and, in Boston, his team “didn’t have a lot of desert to practice in.” The Urban Challenge was a better fit, he said, because Boston provided the ideal testing ground.
It’s unclear whether or not there will be another DARPA-sponsored robotic-car race, but the technology for autonomous vehicles will surely continue to move forward, now that these teams have shown that it is possible for robotic cars to obey the rules of the road. To try to help seed innovation, Leonard said, his team is publicly releasing all of Talos’s software and the data logs collected during the race.
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
Data analytics reveal real business value
Sophisticated analytics tools mine insights from data, optimizing operational processes across the enterprise.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.