jawfish

Robot Cars in Traffic: DARPA Urban Challenge
john posted in rides, technologies on October 31st, 2007

The DARPA Urban Challenge is the third in a series of DARPA sponsored contests for robot cars. Except for a remote kill switch the cars are totally autonomous once turned loose. The first two contests were on remote dirt roads with no traffic. This time the test is in simulated urban traffic, with stop signs, turns, and moving vehicles in both directions.

Stanford’s entry waits to turn left ( blue Passat with Red Bull logo)

The qualifying rounds are just finished at abandoned George AFB in Victorville California and the finals are on November 3.

Video of the Stanford entry. The horn you hear is a safety device sounded by the robot car ( blue Passat with Red Bull logo) the other cars have human drivers.

All videos by Matt

More video, text and pictures after the jump…

To give an image of the difficulty of the test, imagine you are a nervous parent with a teenage driver taking the car out alone for the first time. Now instead of a human teenager, imagine a software program is driving your car. In real time it is deciding when it’s safe to turn in front of a moving car, and how to drive with oncoming traffic. The only way to stop the car if things go wrong is to hit the Big Red Button in the marshall’s booth. Now add an array of delicate $75,000 laser rangefinders on the bumpers and roof, and enclose the test in concrete freeway barricades, where a single error could smash your delicate equipment.

Georgia Tech showing sensors on roof and bumpers

The other cars on course are government issue Tauruses with roll cages and drivers wearing helmets. They run an intricate radio-controlled pattern so that the robot is confronted with cars behind, cars in front, and cars going the other way.

Course A with human-driven bogies

Can They Do It?

Well some entries can drive in traffic and stay in the lane, and obey traffic laws. Some had a hard time and got stuck. We were fortunate to see the Stanford car, Junior, in action as it completed Course A.

Stanford’s Junior has some trouble:

As you can see in the video, Junior gets confused when a stationary car is parked in the lane ahead, even though he is supposed to turn left.

Course B was a longer loop with traffic following, but not ahead of the bot. We saw the Georgia Tech car fail here when it made a right turn and nearly hit the barricade.

Georgia tech on Course B

Anatomy of a Robot Car

First you need a normal car that is street legal. Humans will have to drive it to the starting gate. Small SUVs and station wagons were the common choice.

Next you need to computer-control steering, throttle, brakes, gears.

Next you add several computers, power supplies, cooling fans, and an on-board data network.

University of Texas entry

Then you need some very high-tech sensors, super-accurate GPS, laser rangefinders, radar, cameras, wheel rev counters.

Roof-mounted sensors

Finally you need a lot of custom software to gather and process external information from the sensors, keep track of location on the map, and decide what to do.

The software has to be wrapped up in a package that puts safety first, and prevents runaway cars from ramming the barricades and driving wildly through the parking lots of the old base.

Oh my yes, you need a lot of money, even with sponsorship. The teams, mostly college kids, are large, and the sensors are really expensive, even allowing for discounts and freebies from the manufacturers.

The Princeton team

What’s Easy, What’s Difficult

From the first Challenge in 2004, all the teams were able to control vehicles with the computer, add computers and sensors, and run the bot safely. In the previous Grand Challenge test three vehicles were able to drive the long desert course at a reasonable speed, staying on the road and out of trouble. Many teams failed to navigate the course at all in the previous test, so you can see the basic problem is in the choice of sensors and the software.

Machine Vision

To understand the difficulty of getting a bot to see like a human or even simple animal, imagine that you are driving a car with no windows, no sense of direction, no sense of road feel, and no hearing. Now we place a series of abstract pictures in front of you, and give you commands from the GPS navigator which says turn left, turn right and so on. You have to decide how fast to drive and when to turn based on the abstract pictures. The GPS isn’t accurate enough to keep you from hitting other cars or the wall, but it does keep track of your route. Trouble is, the pictures look like gobbledygook until you laboriously translate them into your visual language. So the software is racing all the time to figure out if that shadow is a wall, or a hole, or just a shadow.

We talked with a member of the Stanford team about their sensor strategy. They are using the data from the LIDARs, laser rangefinders, to look for lines in the pavement. Because the LIDAR is returning an echo, it isn’t confused by shadows and colors. Their bot runs on two Intel quad processors, while other competitors using cameras for visual data are using much more computing power.

Stanford runs light

Cal Tech goes with heavy-duty blade servers

What Does it All Mean?

DARPA is trying to develop a military force that is 30% robots, but clearly recognizes that most military contractors are too bureaucratic and slow-witted for this sort of cutting-edge research, so they are piggy-backing on college students and academics.

The future of mil-transport? This one military-industrial group was third in the Grand Challenge 2005.

DARPA was able to jump-start the Internet in the past, and there are already military autonomous planes in service. Leaving aside the DOD’s Star Wars vision of fighting droids swarming over the battlefield, the civilian implications are vast and unpredictable. I have written about this before, and the success we saw in Victorville shows that success is coming sooner than I would have predicted two years ago.

For instance, robot parking garages, robot taxis and buses, robot subways, robot delivery service, robot security, software-based robot car trains on the highway, robot farm and construction and mining equipment come instantly to mind. True, this sounds like those breathless Popular Science articles about flying cars, and typical science fiction, but there are good economic reasons it may come sooner rather than later.

First, the cost of these vehicles is in the software and sensors. In both cases, the per-vehicle cost will plummet when mass production starts. Second, the green revolution in transport ( see Jawfish here and here too ) requires new energy-efficient re-working on many mature systems that use human operators. As we found with the PC computer and the cell-phone we can’t know what all the uses are of this robot technology until it gets out into the commercial world.Third the software and sensor technology is easily transferable to other types of ground-based vehicles.

For more pictures see our gallery.

The DARPA website has extensive information and up-to-date news.