A student-led research group at Princeton University, PAVE [Princeton Autonomous Vehicle Engineering] has for the past three years pursued the goal of a car that can drive by itself. The team, which consists primarily of undergraduates with assistance from a diverse group of graduate students, has now competed in two DARPA [Defense Advanced Research Projects Agency] Grand Challenges.
Alain Kornhauser, Professor of Operations Research & Financial Engineering, Co-Director, of the Center for NJ TIDE (Transportation Information & Decision Engineering), Director of the Transportation Program and one of the faculty advisors of PAVE introduced members of the group [Derrick Yu ‘10, Christopher Baldassano ‘09, Jonathan Mayer ‘09, Ian Ferguson ‘09, Issa Ashwash ‘09, and Lindsay Gorman ‘10] at the December 5 Lunch ‘n Learn seminar “Computers Driving Down Nassau Street.”
Derrick Yu ‘10 provided a short history of the team’s efforts. As the team’s video shows, the effort to create an autonomous vehicle is not easy. PAVE entered the 2005 DARPA Grand Challenge, with a silver pickup truck named “Prospect Eleven” that relied upon a single stereo camera for navigation. Kornhauser emphasized that effort focused upon a simplicity of approach. Rather than throwing money and technology at every problem, the team sought straight-forward ways to control steering, brakes, shifting, and the throttle. Having successfully passed a site-visit evaluation and numerous runs at the semifinals event, Prospect Eleven was seeded 10th out of 23 entrants, the only undergraduate team to qualify for the finals. In the final October race, a memory leak caused the vehicle to wander off after approximately 9 miles of autonomous travel. After the conclusion of the Grand Challenge, the team fixed the bug by altering just one line of code. The group returned to the desert three weeks later and watched Prospect Eleven avoid collisions and navigate most of the 2004 and 2005 Grand Challenge Courses under autonomous control.
Following the challenge, the PAVE team improved the reliability of the systems in Prospect Eleven, added functionality, and exploring additional areas of research. The team decided to compete in the 2007 DARPA Urban Challenge, with the goal of completing an even more challenging course within the six hour time limit. The simulated urban environment included stop signs, four way intersections, merging, lane markers, and a variety of obstacles. In the spring of 2007, Ford Motor Company donated a 2005 Ford Escape Hybrid to the Princeton effort. PAVE modified the truck for drive-by-wire operation and outfitted it with computers and an array of stereo and monocular cameras, as well as RADAR for detecting obstacles. A GPS system permits the car to determine its location.
PAVE organized its effort within six interacting teams: Perception, Cognition, Actuation, Substrate, Environment, and Business. The environment team is responsible for identifying obstacles, lane markings, signs and conditions outside the vehicle. The Cognition team takes that information and is then responsible for decisions about navigation. Those decisions are passed to the Actuation team, which is responsible for moving the brake or adjusting the flow of gas. And when the car moves, the changing environment requires a revised assessment.
Christopher Baldassano ‘09, a junior in Electrical Engineering, reviewed the perception algorithms used on the vehicle. He emphasized that staying on the road within lane lines was the most important criterion. For that, the car relies upon a monocular camera, essentially a glorified web cam on the top of the car. To detect lanes, the car receives an image from the camera. Two well tuned filters help the car to perceive curbs as well the white and yellow lane lines. PAVE’s approach was to rely upon what the cameras could see for basic driving down the road; most other teams relied instead on expensive military-grade, precision GPS systems. To avoid collisions, the team relied upon a stereo vision system that generates a depth map. Algorithms produced by the team work to distinguish obstacles that need to be avoided. Once all the perception data is gathered, it is passed to navigation.
Jonathan Mayer ‘09, a junior in the Woodrow Wilson School, reviewed the navigation systems. He explained that there are two broad classes, Global and Local Navigation. The over-arching goal is bring the gap between what the sensors are telling the car and what actions are then taken. Global Navigation is in charge of the over-all goal: here are the GPS coordinates of the destination. Computer Science algorithms are used for route-planning, to find the shortest path to the destination. The path is dynamically re-planned around short-term obstacles. Navigation also had the responsibility of determining whether the vehicle was in a parking lot or an otherwise unmarked road. Local navigation generates the immediate path to be followed and a desired speed for the vehicle. There were also specialized controllers for K-turns and parking lot behavior (getting to the parking spot as quickly as possible).
Ian Ferguson ‘09, a junior in Mechanical Engineering, discussed the activation efforts, everything that makes the car move, including the devices that regulate the throttle, the steering, the brakes, the gear shifter, and even the horn. The Ford Escape proved to be an excellent vehicle for activation because it is a fully drive-by-wire vehicle. The team therefore developed a set of electronics under the hood that spoofed the car’s computer into believing that its sensors were activated. To activate the throttle, for example, the electronics simply signaled that the pedal had been lowered by some percentage. Activating the brakes required spoofing two sensors. Derrick Yu ‘10, a sophomore in Electrical Engineering, explained that the substrate team that developed the power and computing systems and architecture into which all of the perception equipment connects. A rack of five networked computers in the back of the car are dedicated to vision processing, navigation, and vehicle control. All systems run under Microsoft Robotics Studio, a framework that supports the development of real-time robotic control and services. There are powerful features for processing simultaneous commands and for taking advantage of the processing power on all of the computers.
Issa Ashwash ‘09, a junior in Electrical Engineering, noted that the Environment team was responsible for testing the vehicle and collecting data. Testing proved to be very important because systems that worked well “on the bench” did not always work as planned within the vehicle. The introduction of changes to one system required tests on that system as well as on systems that use that module. The team assembled design criteria for the tests and collected information on critical points of failure. Issa also reviewed the efforts of the business team, a critical part of the effort given the costs and the need for contributions and sponsorships from industry. Together, these teams designed, implemented and tested a complex autonomous system, passing a 5-minute video demonstration, Technical Paper submission and Site Visit evaluation in the process. The team qualified as one of 35 semifinalists.
The Princeton team did not make it to the finals of the 2007 DARPA Urban Challenge. 11 teams were accepted, and 6 finished the course, with top prizes going to teams from Carnegie Mellon University, Stanford University, and Virginia Tech. The students from PAVE drove the car to the talk, giving all comers an opportunity to look first hand at their amazing achievements.
Lunch ‘n Learn attendees examine PAVE in snow shower.
Posted by Lorene Lavora