Car: Where's my dude?*
A 130-mile course in the Mojave Desert. Millions in prize money. Hundreds of mechanics. No drivers.
The Great Robot Race, a Nova documentary airing March 28 on PBS, covers the contest sponsored by the Defense Advanced Research Projects Agency (aka DARPA—they're the guys who actually invented the Internet). We won't spoil the ending, but one contestant, Stanford University robotics Prof. Sebastian Thrun, spoke with U.S. News about the event and the future of robotic cars.
When did you accept the Grand Challenge?
DARPA conceived of the Grand Challenge in 2003. One year later, the first race was held, and none of the teams was able to complete the course. I was moving out to California at the time and was not involved with the race, which I regret, especially after seeing what happened. Within the robotic community, there was mixed opinion after the first race on whether it could even be done. But in 2004, I talked to many journalists and said a robot would be able to finish the race next year. In hindsight, it looks like a good prediction.
The race seems very simple.
On the surface, you could say that it is an easy task: Put a computer in a car, put a few sensors on it, and let it go. A good number of people who competed in 2004 had that philosophy. The key players in the game, like the university teams, recognized the challenges and allocated their resources accordingly.
What are the challenges?
It's incredibly simple: If you stay on the road, you win. If you drive off the road, you lose. Now, the challenge is to find out where the road is. A child can look out a car window and see where the road is because our brains are very good at object recognition. A computer has great difficulty in determining what is the road and what is not. We have not even remotely matched the brain's ability to understand the physical world. So, it was distinguishing the road from obstacles that was important.
You approached the road recognition problem quite differently from your competition.
We used two things: multiple sensors and artificial intelligence. We put lasers, cameras, and radar all together on the car. And we invested very heavily in software that can tell with a great deal of accuracy what it knows and what it doesn't know and how it can acquire more information. We taught the computer to recognize the road and seek out more of the road. So our emphasis on software was what made us different from the competition.
Without speaking in techno-jargon, can you outline your scheme?
We used lasers to determine what the road looked like 20 meters out. This worked very well, but it does not allow you to go very fast because of the inertia of the vehicle. The question became: How can we look further ahead so we can go faster? We put a camera looking forward and tried to find where the road was on the image. This is a problem we have not solved. So, we combined the short-range system that could identify the road at the bottom of the image and then projected further toward the horizon on the image to see where the road was likely to be. The computer then looked at the image and at the road 10 times per second to find out where there was more road to follow.
The contest wasn't just an aimless road trip.
No, it was a guided race. We were given several thousand GPS points as breadcrumbs that the robots had to follow around the course. One never had to teach the vehicle to make decisions about which path to take—only how to follow the path and how to avoid obstacles at high speeds. In the 2004 challenge, the best-performing team didn't have a robot that perceived its environment; it just went full speed after the breadcrumbs. The problem was that the GPS is not that good. There is about a 2-meter resolution for GPS, and if you are 2 meters off the road, you fall off a cliff. In 2004, the Carnegie Mellon team went off the road and burned up a tire, which then caught on fire.
Could the race have been won with technology that existed in 2004?
My claim is, yes. It could have been done. We just needed more time to examine and understand the problem. It is a tedious argument and we have the benefit of hindsight, but The lesson is that robotics has reached a level where it can really do magic.
The movie makes the challenge look rather easy.
Driving is easy. People age 80 and age 16 do it all the time. You can be drunk and do it. How many brain cells does it take to drive a car?
What kinds of brain cells were inside your vehicle?
Inside, we had the equivalent of a high-end desktop computer. If we were to commercially package what we designed, it could be run off one laptop quite easily. It's not the number-crunching power or the speed—it's the algorithms.
What's the next challenge for the robotic community?
To change the world.
Starting small, eh?
You might be able to tell that I am a bit of an optimist. But here's an example. In 2007, we want to have a car that drives itself from downtown San Francisco to downtown Los Angeles, 100 percent autonomously. There will be a person inside, of course, but that person will not touch anything for six hours as the car drives. The limitation of the grand challenge is that it is essentially done in a vacuum with nothing interfering with the progress. In our next attempt, there will be moving traffic, which poses a considerable challenge.
*Headline recycled from a 2003 U.S. News story on the Cadillac XLR.