To inspire progress in robots, the U.S. government's Department of Defense Advanced Research Project Agency (DARPA) sponsors a grand challenge for autonomous vehicles to complete a 100-plus mile, off-road course in the Mojave Desert (www.grandchallenge.org). In 2004, the first year of the event, Sandstorm from Carnegie Mellon's Red Team traveled the furthest, but that was just 7.36 miles, so no one won the $1M prize.
The publicity surrounding the 2004 competition, plus the doubling of the award to $2M, led to 195 entrants from 36 states and four countries for this year's event. DARPA whittled the list down to 43 in July using a series of two-mile obstacle-course trials.
A final time trial was held at the California Speedway Sept. 27Oct. 5 to reduce the competition to 23 finalists announced Oct. 6. CMU's Red Team Too edged the Stanford Racing Team for first place in this trial, followed by Sandstorm from CMU's original Red Team. These university entries leverage the latest advances in statistical and machine learning and in vision to automate their vehicles. In particular, the leader of the Stanford Racing Team, Sebastian Thrun, is also one of the developers of FastSLAM, a factored solution to the simulation localization and mapping problem, which was critical to Stanley's success .
If we could reduce the costs of automobile accidents by $30B per year in the U.S., a single year's savings would exceed the government's investment in academic computer science research for the last 50 years.
CMU's Red Team Too entry, named H1ghlander, is a 1999 H1 Hummer with a turbocharged diesel engine, traction control, and embedded drive-by-wire technology. Sandstorm, from the CMU Red Team, started life as a 1986 HMMWV. Stanford's entry, Stanley, is a stock Volkswagen Touareg R5 modified with full-body skid plates and a drive-by-wire system controlled by seven Pentium M computers. Other bots that performed well were Team Terramax, a combined effort from Oshkosh Truck and Collins Radio, and the Gray Insurance Ford Hybrid SUV.
After the trials, the finalists had one day to repair their robots. The race began Oct. 8, with each entrant given a CD defining a 132-mile, obstacle-laden course; they were positioned according to their lap times in the final time trial. The results of the finals are listed in the table here.
Of the 23 entrants, 18 were disabled, but all but one exceeded the maximum distance from 2004. Five finished the course; although only four did it under the 10-hour limit needed to win the award. Stanford's Racing Team took home the $2M prize. Given last year's results, I was certainly surprised that many teams did so well. Many AI experts consider this milestone a more significant achievement than when IBM's Deep Blue computer beat the reigning world chess champion in 1997.
The success of the challenge could help accelerate the deployment of driver-assistance devices, potentially reducing the loss of life and property from traffic accidents. In just the U.S. alone, traffic accidents cost 40,000 lives and $55B. Worldwide, nearly 1.2 million die each year, 20 to 50 million more are injured or disabled, and more than $500B is spent annually on traffic accidents . Indeed, ACM A.M. Turing Award winner Butler Lampson suggests a worthy grand challenge for our field is to create the technology to eliminate all automobile accidents. Even if the technology prevented half the accidents, it would save $25B to $30B and prevent tens of thousands of deaths and injuries in the U.S.
Upon further reflection, it's clear the DARPA challenge tested the importance of academics participating in a research mission in an open competition against all comers. The agency recently shifted the emphasis toward more classified research and research with short go/no-go deadlines that discourage participation from universities. The result is a 50% decline of DARPA funding of contracts with academics as principal investigators .
University teams were 30% to 40% (depending if you count joint industry-academic teams) of the 43 semifinalists and 40% to 50% of the 23 finalists. Not only did Stanford University's team win first place, university teams swept the top three places, and only one other entrant completed the course within the 10-hour window. These results jibe with the NRC report illustrating the important role of academic research in helping create 17 multibillion-dollar industries . Whether you look to history or to this competition, what conclusion can you reach other than academic research is vital to technological advances in IT and to the creation of new IT industries?
If we could reduce the costs of automobile accidents by $30B per year in the U.S., a single year's savings would exceed the government's investment in academic computer science research for the last 50 years. Given the tremendous progress over the last year as a result of this competition, doesn't it seem plausible that if we invested, say, 5% of the potential savings in computer science research for five to seven years, that we could at least halve traffic accidents down the road? Wouldn't almost all drivers urge such an investment? If so, what are we waiting for?
1. Montemerlo, M., Thrun, S., Koller, D., and Wegbreit, B. FastSLAM: A factored solution to the simulation localization and mapping problem. In Proceedings of the AAAI National Conference on Artificial Intelligence, 2002; robots.stanford.edu/papers/montemerlo.fastslam-tr.html.
©2005 ACM 0001-0782/05/1200 $5.00
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2005 ACM, Inc.