Robotic Cars: The DARPA Urban Challenge (Part 1)

This blog comes to you from Scott Ettinger of our Applications Research Lab. Scott is on-site at the DARPA Urban Challenge, working with the Stanford team to try and win a unique automobile race where there are no drivers. Photos are courtesy of Jan Becker.

George A.F.B. – Victorville California. This once abandoned air force base has been transformed into a busy tent city filled with some of the top technical talent in the world. Super high-tech vehicles – most covered with arrays of cutting-edge laser sensors can be seen around every corner. With the mass of equipment on board, many look like they are designed to explore distant planets. Instead, these vehicles are designed to explore much closer to home on our own city streets. This entire gathering is in fact dedicated to developing cars that drive themselves. It is the pinnacle of years of work by the best robotics experts from around the world.

Darpa1.jpg

[Photo 1: Me with “Junior” – Stanford’s Darpa Urban Challenge vehicle]

It is called the DARPA Urban Challenge. 35 teams, selected through a series of preliminary testing are here to compete for a two million dollar prize sponsored by the U.S. government’s Defense Advanced Research Projects Agency. These extreme technology vehicles will “race” each other through the streets of the abandoned Air Force base to claim the prize – with no drivers inside. These are not “remote controlled” vehicles. The vehicles themselves make all the decisions – no humans involved. Armed with a digital map of the area supplied by DARPA, each vehicle will be required to complete a given “mission” – a set of places within the city they must visit in order. Basically, at each segment of the mission, the vehicles are given point A and point B, but not any information on how to get from one to the other. In between they must follow all traffic laws and detect and deal with any obstacles or vehicles, just as you or I would when driving downtown.

DARPA has not specified exactly how they will score the contest, but they have said that teams will be penalized for violating any of the rules in the California driver’s handbook. That’s right, the same document you got tested on if you have a California license.

Although this is a contest with high stakes, nobody is keeping any secrets- in fact quite the opposite. As with most scientific endeavors, a spirit of sharing knowledge is everywhere here at the event. Nearly everyone is happy to show you and discuss the intimate details of the inner workings of their robots. In between official practice runs and inspections here on the first two days of the event, participants spend most of their time visiting with other teams and comparing notes. Tonight, here at the Stanford Racing Team tent we invited all participants and press over for a presentation describing the software technology developed for “Junior” – Stanford University’s robot vehicle.

Intel is a sponsor of the Stanford team, and I have been lucky enough to be working with them to help develop some of this very futuristic technology. Sebastian Thrun (project leader) and Mike Montemerlo (lead software developer) described the software technology behind Junior’s brain. Mike first showed a number of videos to illustrate some of the techniques and processes behind Junior’s decision making including a number of traffic examples. He went on to show some really impressive visualizations of the perception technology developed by team member and code guru Dirk Haehnel. These videos effectively show what Junior “sees” through its laser sensors. The Terminator 2 scenes of robot perception have nothing on these 3d reconstructions Junior’s world.

Darpa3.jpg

[Photo 2: Press talk with Sebastian Thrun of Stanford University]

Darpa4.jpg

[Photo 3: Mike Montemerlo describes Junior’s decision process]

I myself gave a brief demonstration of the simulation technology we developed in order to better understand Junior’s behavior. The demonstration showed off technology to let Junior drive in the virtual world – complete with 3d models of himself and numerous other cars. In this demonstration I drove a virtual Lamborghini around with Junior and 4 other cars in a virtual reconstruction of the actual race site built using data given to us by DARPA including maps and aerial imagery (similar to what you get on Google or Yahoo maps). I acted as a bad driver (that guy on a cell phone driving while eating) to put Junior into difficult traffic situations to watch his behavior. This technology is used extensively to uncover any problems with Junior’s behavior first in the virtual world, so that we do not have to encounter them in the real one. The simulation runs the exact same software as Junior does, but feeds it with simulated perception (processed sensor data) and simulates the vehicle dynamics given Junior’s controller output.

Darpa5.jpg

[Photo 4: Georgia Tech’s Porsche Cayenne “Sting” with Junior in the background]

After the technical talks, we headed over to the Red Bull tent for a great catered reception (Red Bull is a Stanford financial sponsor) to catch up with the rest of the community. Qualifying tests begin tomorrow and continue through Monday to determine which teams will move on to the final event on Nov. 3rd.

Darpa6.jpg

[Photo 5: Insight Racing’s Lotus Esprit]

There are many impressive looking vehicles here (our tent is across from the Georgia Tech team with their robotic Porsche Cayenne). Most have custom paint jobs with sponsorship logos. MIT’s Land Rover has a very impressive array of sensors. Other interesting vehicles include Insight Racing’s Lotus Esprit, University of Florida’s hybrid Toyota Highlander painted with some nice green alligator textures, and MIT’s mobile computing powerhouse with 40 intel cores (10 DP dual-core xeons) if I remember correctly. I’m looking forward to seeing them in action tomorrow.

[Update 10/28/07]

Today saw the first official tests of the DARPA Urban Challenge vehicles. When DARPA puts together an event, they go big. In order to set up an environment to put the vehicles to the test, DARPA assembled over 200 (it may be closer to 300) Ford Taurus sedans reinforced with roll cages and safety harnesses. They also hired professional drivers (some with flames on their helmets) to drive them along with the robot vehicles in their simulated driving environments.

The most visible test today was over at the merge testing area. This area consists of a rectangular road loop with a one-way road segment dividing in two. Running along the outside of the loop in both directions is a swarming mass of professional human driven Taurus vehicles moving in a precisely orchestrated pattern. The robot being tested must continually make left turns into and out of the loop traffic. This is no easy task – as the results showed today. While some robots performed fairly well, many of the robots made some very dangerous maneuvers, but they were handled very well by the human drivers. These guys have nerves of steel. I watched a few robots attempt to pull into the lane with oncoming traffic only a few feet away or even worse right in front of the intersection. The professional drivers calmly took evasive maneuvers. I didn’t see any robots perform this task flawlessly. The humans were instructed to honk when the robots tried an unsafe merge, and there was a good amount of honking going on. Only a third of the teams ran through this test today. The big excitement of the day came when a robot had a fairly spectacular failure and accelerated directly into one of the large concrete barriers lining the course.

Over at the intersection testing area, Junior (Stanford’s robot entry) performed nearly perfectly. He was run through a number of 4 way intersections with varying numbers of cars arriving in turn. Junior properly waited his turn and followed all proper traffic procedures. Junior will face the merge testing gauntlet tomorrow around noon. We ran simulations of the merge testing task today and after some minor parameter tuning, Junior performs exactly how we would like him to in the simulation. Tomorrow we will find out how he does in the real test.

The third and final test is the mysterious “Area B” test. This test has no spectator viewing. All we know is that the vehicles drive off into the largest portion of the city test area and return about 30 minutes later. Junior will perform this task on Monday. It was impressive to see the vehicles in action today. More soon…

[Ed. Note 10/29/07: The story continues in Scott's next post]

5 Responses to Robotic Cars: The DARPA Urban Challenge (Part 1)