Portland, Ore. -- Think the desert terrain of the previous Grand Challenge autonomous-vehicle races was unforgiving? This year, on Nov. 3, the bot-mobiles will take to the mean streets of a U.S. city.
Eighty-nine teams, nationwide, are gearing up to compete in this year's Urban Grand Challenge, sponsored by the Defense Advanced Research Projects Agency (Darpa). The goal is for driverless vehicles to navigate 60 miles of closed urban course in the company of other vehicles. The race will pit the nearly 90 driverless vehicles against one another and against human-driven vehicles that will present "moving obstacles" to the racing robots. The fastest qualifying vehicle will net the $2 million grand prize; $1 million will go to the second-place car and $500,000 to third place.
"We were the first to announce we were entering the race, but we are still recruiting volunteers--we need EEs," said Paul Grayson, chief engineer at American Industrial Magic LLC, an all volunteer organization. "We are outfitting an Army cargo truck with a combination of sonar, radar and computer vision as well as some trick software written by engineers like your readers," Grayson told EE Times.
Darpa's ultimate goal is to meet the Defense Authorization Act's 2001 mandate to develop the autonomous and remote-controlled technologies needed to take one-third of combat-zone vehicles driverless by 2015. Earlier Grand Challenges spawned vehicles that could navigate in off-road environments; now urban delivery is being added. Eventually, all of these capabilities are predicted to trickle down to consumer vehicles.
"The U.S. Army has been ordered by Congress to buy and install 400,000 driverless systems as soon as possible and before 2015," Grayson. "This race is to prove it can be done. On U.S. roads and highways, drivers will be able to benefit from these technologies, where they will be used as driver assistance systems and active safety systems that prevent traffic accidents."
The Urban Grand Challenge also adds simulated "supply missions" that will involve, for example, finding and parking in specified spots, as if for loading and unloading cargo. The waypoints will be specified by GPS coordinates in the flash memory of a USB "key" just five minutes before the race, although the course itself, sans waypoints, will be revealed the night before the race.
Tasks will include autonomously merging into moving traffic, reading traffic signs, navigating through traffic circles and honoring right-of-way rules at busy intersections--all while avoiding collisions with fixed obstacles, other robotic contestants and cars driven by human referees.
"This year's is a much, much harder competition. Before, it was one robot on its own out in the middle of the desert, but now there are other moving objects. The hardest part will avoiding collisions with them," said Mike Montemerlo, senior research engineer at Stanford University's Artificial Intelligence Laboratory.
Montemerlo was on the Stanford team that won Darpa's 2005 Grand Challenge with a Volkswagen Touareg nicknamed Stanley. This year, the Touareg will be traded in for "Junior," a Volkswagen Passat that's a better fit for the urban course. The brains, however, will continue to be supplied by Intel Corp.
Stanford announced details of its Intel-sponsored entry at the recent annual meeting of the American Association for the Advancement of Science in San Francisco (www.eetimes.com, search article ID: 197006955).
The team chose a Passat "because it's built for urban streets and has many more drive-by-wire features," said Montemerlo. "For instance, when you turn the steering wheel, the output of a torque sensor is amplified by a motor that turns the front wheels--the advantage being that you can take over that motor and steer the car with software."
During testing, Junior always has a human behind the wheel who can take over at a moment's notice to ensure safety. The servos that turn the wheel are weak enough, according to Montemerlo, that a human can easily overpower them. During the race, however, there will be no safety drivers allowed; the difference between collision and avoidance will be entirely up to the software reading the sensors and driving the actuators.
According to Montemerlo, there are two parts to this software problem: perception and decision making. "Perception means being aware of all the objects around you--other vehicles, telephone poles, curbs, stop signs, where the lanes are in the road," said Montemerlo. "The second part is choosing a course of action: Where do you steer the car?"
For perception, Junior has 360° laser range finders--or lidars--plus an automated visual system comprising six video cameras. The range and visual sensor data streams will be fused by software now being composed by Stanford and Intel engineers. Algorithms then take over the car's actuators to steer, accelerate and brake the vehicle.
Stanford tapped Intel as its corporate sponsor to leverage Intel's fastest dual- and quad-core microprocessors. Intel engineers are also helping with the computer programming.
"We have huge amounts of sensor data streaming in this year," said Scott Ettinger, a researcher with Intel's Corporate Research Group who has been assisting the Stanford team since the last Grand Challenge win. "It takes a lot of computing just to determine where you are in the world and what other objects are around you in the environment. Then, once you figure that out, you have to do higher-level reasoning regarding where you want to go, and finally you have to send commands to the actuators to turn the wheel, push on the gas and apply the brakes."
Junior's brain will be composed of dual- and quad-core Intel processors, with both high-speed and low-power versions performing different tasks as appropriate. Higher-level computing functions will be performed on the highest-performance processors, while multiple signal-processing data streams will be processed with low-power processors. All must fit into a tight power budget, since the only power available on Junior comes from the engine's alternator.
"We have to squeeze every bit of performance out of every watt," said Ettinger. "Because there are lots of different tasks going on at the same time, we can really take advantage of our multicore processors. I wouldn't say it was easy to do, but the problem is well-suited to multiprocessing."
Stanford programmers are also using Intel's performance primitive libraries and its computer vision libraries to simplify partitioning into multithreaded algorithms for execution on multiple cores.
"By using Intel's compilers, which perform automatic vectorization and multithreading of some code, programs can take advantage of multiple cores without having to always explicitly write multithreaded code," said Ettinger.
But no matter how much parallel processing is thrown at the problem, everyone seems to agree, the hardest part of the Urban Grand Challenge will be the rubber-meets-the-road decision-making software. That's where the dynamic ability of multiple robotic vehicles to share the roadway with each other and other human-driven vehicles will be tested. Any collisions will be judged as failure.
The rules of the road that human drivers take for granted also have to be programmed into the autonomous vehicles. Even on the open road, there are many signs informing drivers about conditions, as well as a host of implied rules for performing such feats as merging with traffic and changing lanes safely. To win the Urban Grand Challenge, a robotic vehicle will have to process any number of inputs and then make informed decisions on the fly.
"Understanding and classifying what you are seeing, rather than just reacting to it, is the real challenge. The concept of right of way is a very difficult concept to master even for some people," said Montemerlo. "We think the difference between winning and losing is going to boil down to software. Given all this noisy sensor data, the team that properly interprets it and makes the best decisions based on that interpretation will win the competition."
The innovations pioneered to win the Urban Grand Challenge will likely end up in consumer vehicles, after the military proves out the concepts with the vehicles built to meet its 2015 mandate.
"In the future, automobiles are going to have a lot more computing power on board, and we are just scratching the surface" with the technologies developed for the Darpa contest, said Ettinger. "Our work here will end up adding a lot to driver safety, driver assistance and keeping drivers better informed than they are today."
Of the 89 teams slated to compete this year, 60 made applications under Darpa's "fast track" program to qualify for up to $1 million in financial aid per team. Only 11 teams won awards, however, including all of the winners of the previous contest.