The overall system architecture greatly affects the final system performance as demonstrated by two approaches to a single problem. As an example of the diversity of approaches to a single system challenge, we look at the robotics programs at Stanford University and at Carnegie Mellon University. Both schools were involved in the DARPA grand challenge and urban challenge programs.

The US military has always been interested in autonomous vehicles for two reasons. First to reduce the number of support staff versus combat troops. Second, to reduce the exposure of any troops to hostile activities. Some of these programs have been started as early as 1988 but were achieving little or no useful results.

In 2003, DARPA instituted the grand challenge, to have autonomous vehicles navigate from Los Angeles to Las Vegas, a distance of over 130 miles at speeds up to 60 mph across the desert. A $1 million prize was promised to the first vehicle to traverse the distance in under 10 hours. In 2004, 195 teams showed up in the desert outside of Barstow and raced using a staged rally car start to their destination outside of Las Vegas. This first event resulted in all vehicles disabled or incapacitated inside of the first 7 miles. Most of the vehicles got off the road and turned over.

After this debacle, DARPA renewed the challenge and gave all potential participants 18 months to develop better autonomous vehicles. In the second grand challenge, five vehicles finished the course, three completely unaided. The vehicles all incorporated numerous sensors, including GPS, laser rangefinders, RADAR, and video cameras. The vehicles were outfitted with controls for operations and maneuvering

Keep your eyes on the road

The first approach to system design is illustrated by comments from Sebastian Thrun, director of the artificial intelligence research lab at Stanford University, who described their efforts in a keynote from the Hot Chips Conference on August 25, 2008 at Stanford University.

The Stanford team elected to mimic as closely as possible a human driver. Therefore, they made video the primary input and used algorithms for edge detection to identify isosceles triangles or trapezoidal figures which represented a view of the road ahead. Laser sensors and radar information were used to provide information on vertical dimensions to indicate objects in the road and features outside the road to ensure the vehicle stayed relatively centered on the roadway. GPS was used for longer-range navigation and general mapping coordinate information.

When the car, based on a VW Toureg, was driving itself it "looked" at the road features ahead and plotted appropriate speed and path vectors to advance to the next GPS waypoint. When the car approached a turn in the road, it observed a distortion in the trapezoidal figures and slowed to make the turn. This approach seems to have worked, since Stanford won the second grand challenge.

On the other hand, the Carnegie Mellon team, according to Chris Urmson, director of technology for the urban challenge, based their primary correctional and speed control in human-assisted training. Later, when the car was on its own, it used this detailed mapping to establish its upper limit for speed while direction was fairly fixed by the previous training. Essentially this limited the vehicle to a maximum of what the human driver could do on a given run as any new obstacles or other changes caused the car to slow down and try to determine its next GPS waypoint.

One of the changes the military made to the course was to construct a tunnel on a portion of the road to block off GPS signals causing vehicles to drive "blind" for some distance. Together, the limitations of the human-based training and the reliance on the constantly updated position information forced the Carnegie Mellon at run at much less than optimal speeds, so they came in second.

After the first grand challenge was achieved, DARPA initiated a second challenge, the urban challenge. This would require an autonomous vehicle to go 60 miles in less than six hours. The course included intersections, merging, passing situations, parking in a parking lot, and other ordinary traffic conditions. The urban challenge did not include traffic lights, pedestrians, bicycles, and drunk drivers. Eleven teams made it through the preliminaries and attempted the final course. Six teams completed the course and three of these cars completed the testing without interventions.

Practice makes perfect

The biggest difficulty for the urban challenge is the dynamic nature of the data sets. The controller has to continuously monitor static and dynamic data and determine the relationship between the car, road, and other vehicles. The data from the sensor arrays must be parsed to separate those features that are due to vehicle motion versus the movements of other vehicles. One of the hardest tasks is to "see" the car next to you, since it behaves like a shadow. The sensors generate over 1 M elements per second.

The Carnegie Mellon team spent over 6 hours a day refining their programming and training the vehicle for the fixed course. Since the pathway was fixed, only the on-site traffic changed. The wining team used lasers to look for shape and height and video to locate moving objects. The programming had to make projections in time to plot trajectories of the moving objects and plot a clear path that stayed on the course. The extensive training included over 45 days of acceptance testing, 4 test sites that had various traffic features, and over 3200 kM in the overall testing.

All of this testing and training highlighted the weak areas in the algorithms and controls, and allowed the team to optimize the programming. But the work was worth it: in the final result, they won by 20 minutes over Stanford.

What's next?

These technologies will sneak into our daily lives. Most of the necessary hardware is either already in modern cars or can be incorporated without major redesign. The abilities of semi-autonomous cars will improve safety as the internal active and passive restraint systems are reaching their limits. Robotic cars will enable elderly and handicapped people to improve their mobility and reduce their isolation.