Robot Central had the opportunity to talk with Austin Robot Technology yesterday and watch the smartest Isuzu in history do some autonomous driving. Lets talk about horsepower. No, I’m not referring to Marvin (the name of the aforementioned Isuzu), I’m referring to the world-class talent that makes up this 2007 DARPA Urban Challenge team.
Team President Arturo Martin de Nicolas explained, “When we assembled the team, we simply asked ourselves who the ideal person for each role would be. Who would make the best AI guy? Associate Professor Peter Stone would. He’s the recipient of the prestigious International Joint Conferences on Artificial Intelligence Computer and Thought Award–which is only given out once every two years. Who would make the best OS guy? Unix Guru Jack O’Quin would. He’s one of the founding engineers who wrote IBM’s AIX.” Arturo went down the list proudly citing credentials and decades of experience of teammates. The result is a formidable collective skill of of engineering, entrepreneurship, and vision. Above all, this team is delivering.
Arturo isn’t new to DARPA’s challenges. He was with Austin Robot Technology when it was one of 40 selected from a field of over 200 to participate in the 2005 DARPA Grand Challenge National Qualifying Event. He says that Stanford, Carnegie Mellon, and Team Gray are the teams to beat this year.
Art’s one of those super-smart guys that defines and implements chip-level instructions on microprocessors. He defined the instruction set that gave the PowerPC better x86 emulation and he subsequently spent another 10 years programming micro-code for VIA’s family of x86 processors. He now works at Intel.
After speaking with Art for a while I walked over and introduced myself to Peter Stone. I engaged him about Marvin’s architecture.
Professor Stone explained that Marvin’s AI is based on a 3T architecture and is comprised of the Pilot Layer, Navigator Layer, and Commander Layer. The Pilot Layer provides the basic abstraction from the car’s actuators while the higher level AI behaviors are encapsulated in the Navigator and Commander Layers.
A good architecture is more important in the 2007 Urban Challenge than it was in 2005’s Grand Challenge. The ’05 race was largely an obstacle avoidance exercise. While primitive behaviors such as “drive straight,” “go to the next waypoint,” and “don’t hit anything” were sufficient for the 05 ‘bots, this year, Marvin’s going to have to know the right-of-way rules at a four-way intersection, when it’s safe to pass a moving vehicle, how to park, and how to find an alternate route if the path he’s taking is blocked. Higher levels of abstraction are required to effectively implement these very specific behaviors.
Team Leader Dave Tuttle has something to say about all of that. In his vision of the future, autonomous transportation is enabled through smarter vehicles rather than through smarter highways. “It makes substantially more sense to leverage the existing infrastructure by making smart cars” rather than by reinventing the highway system. He went on to explain that “today, even the lowest end car has 20 to 30 computers. The last number I saw was that the … cost of hardware and software is more than the steel that makes up the car. It’s about 15% now but it’s going to double by 2015.” Dave’s the chairman of Austin’s Digital Horsepower Initiative and was the panel moderator of a CleanTX Forum session with automotive industry leaders speaking about Grid-Connected Vehicles.
Leading the way in creating the next generation of application developers, Patrick Beeson is a Graduate Research Assistant who lets his fingers do the walking. This is the man that turns the architecture and vision into actual running code. Patrick is also the technical lead over a team of undergraduates Ryan Madigan, Richard Edwards, Tarun Nimmagadda, Mickey Ristroph, Bartley Gillan, Sanju Ashok, and David Li all of whom are responsible for the implementation of everything from control code to perceptual code. Patrick made it a point to explain to me that David Li put in place a “nice algorithm” that takes waypoints and generates some navigational geometry.
Marvin’s mechanical and software plumbing is largely unchanged from his 2005 run, but the software has been re-architected and higher level functions in the abstraction stack have been completely rewritten and new ones added.
I made a short video that shows Marvin driving autonomously. (That’s Patrick behind the wheel by the way.) When I was watching the video playback I was shocked to see just how much I took for granted that Marvin could avoid obstacles. In the video you’ll notice Marvin drive by my pickup truck with only inches to spare and I of course stood right by just trusting he’d do the right thing. And he did.
“If there’s anything I want people to know about Austin Robot Technology, it’s that we feel good about where we are, we really enjoy working with great people, and we respect the challenge,” Tuttle said. “Our satisfaction comes from taking on the challenge itself. And to the degree that it helps save the lives of our troops, the challenge is a good thing.”
Other links:
Great article (and video)! Any plans for doing more of these team perspectives?