Pedal to the metal

LAS VEGAS - With a forward jolt, the ChevyTahoe begins navigating a course markedwith bright orange cones.The starts and stops are jerky, as though astudent driver were using two feet on theaccelerator and brake simultaneously. Theturns are deliberate and sharp, with little concernfor the comfort of backseat passengers.At a four-way stop, an encounter with ahefty Hummer H2 is treated cautiously; theTahoe waits patiently to make sure the othercar has indeed stopped.For a novice driver, the exercise is passable.For a driverless car depending 100 percent ontechnology, the trip is impressive.The vehicle ? Boss ? was built by TartanRacing, part of Carnegie Mellon University'sRobotics Institute. The vehicle won the 2007Urban Challenge, held by the DefenseAdvanced Research Projects Agency, and its$2 million top prize. November's race was thethird one DARPA sponsored.Stanford Racing's Junior, from StanfordUniversity, took the second-place, $1 millionprize, and Victor Tango's Odin, from VirginiaTech, received $500,000 for finishing third.The DARPA competitions are designed tofoster development of autonomous roboticground vehicle technology for the battlefield.As these technologies improve, their applicationwill provide business opportunitiesbeyond the battlefield. Boss was recentlydemonstrated in Las Vegas as part of theConsumer Electronics Show.Vehicles competing in the Urban Challengewere required to operate entirely autonomously,without human intervention, as theyobeyed California traffic laws and performedmaneuvers such as merging into moving traffic,navigating traffic circles and avoidingobstacles.The vehicles had to think like human driversand continually make split-second decisionsto merge into traffic, safely passthrough intersections and avoid collidingwith other vehicles.The first Grand Challenge was held in 2004on a 142-mile desert course between vehicles attempted the course, but none finishedand the $1 million cash prize wentunclaimed.The most recent race near Victorville,Calif., was held in a simulated urban settingand was the most difficult to date, said ChrisUrmson, director of technology at TartanRacing."The idea was to have the vehicles interactwith one another, and with human-driventraffic, much the same way drivers do whenthey commute to work in the morning," hesaid. "The vehicles had to be able to seewhere they were in the world and to see othervehicles."DARPA officials gave teams maps of theroads they could use ahead of time. The winningvehicle had to complete the course inthe least amount of time in a safe manner,while following the rules of the road.Since work began in 2003 for the firstchallenge, the technology used for perceivingand tracking vehicles has greatly evolved,Urmson said. Improved computing power isalso a critical factor in making the technologywork.All the teams relied on computers, laserrangefinders, the Global Positioning Systemand inertial measurement to navigate thecourse.Team Victor Tango is composed of VirginiaTech undergraduates, graduate students, facultyand Torc Technologies, a Virginia Techspinoff company that works with autonomoussystems. The Odin vehicle is a converted2005 Ford hybrid Escape.The Stanford team, which won the 2005Grand Challenge, used a converted 2006Volkswagen Passat for the Urban Challenge.The ability to avoid collisions and understandhuman driving concepts, such as rightof way, is of great interest to Pentagon officialswho want autonomous vehicles for dangerousmissions, such as delivering supplies."This has a component of prediction," saidMike Montemerlo, a senior research engineerin the Stanford Artificial Intelligence Lab."There are other intelligent robot drivers outin the world. Predicting what they are goingto do in the future is a hard problem that isimportant to driving. Is it my turn at theintersection? Do I have time to get across theintersection before somebody hits me?"All that prediction and evaluation getsmore complex in dense urban congestion."If you're in a lane and you're going to stayin that lane, that's easy enough to do,"Urmson said. "But when people are driving invery dense traffic, there's a lot of social activitythat happens. You make eye contact withsomebody that's going to change lanes, youmight gesture in some way that you're eitherletting them merge in, or that you want to gochange lanes. That kind of social interactionis hard to do."Understanding traffic lights is difficultbecause the configuration of one intersectionto the next is usually different.Predicting what another vehicle might dois difficult enough, predicting what pedestriansmight do is even harder, Urmson said."With vehicles, you have that constraintthat they can only go the way the wheels arepointed, they can't leap sideways," he said."If instead you had Emmitt Smith out in theroad, he can go anywhere. So modeling theway pedestrians move is difficult."Although it will likely be at least a decadebefore fully autonomous vehicles are available,General Motors' Larry Burns said thecompany is committed to investing in theresearch and development needed for thetechnology."OnStar, for example, is 12 years old, andwe're on our eighth-generation hardware,"said Burns, GM's vice president of researchand development. "The only way to positionyourself for leadership in technology in theauto industry is to get out there and createthose opportunities and get them to a generationone, level-of-proof concept and then seewhat really plays out."Business opportunities to develop new sensorsand computer systems to interpret themare expected to increase. Improved laser andradar are needed for the technology tobecome mainstream."The car is going to become much more ofan information appliance," Urmson said."You're actually going to have sensors on thevehicle that allow you to push informationback into the network, and that's a big deal.So we can start to look at how you use thatinformation and what applications thatinspires in the future."