Humans spend much of their lives piloting cars from place to place. For millions of people who spend time stuck in traffic, the promise of one day being able to take their eyes off the road, read a book, or catch some extra sleep took a big step closer on November 3rd, 2007. The DARPA Urban Challenge saw teams of researchers, students, and their corporate partners, primarily from the US and Germany, field their latest generation of autonomous robotic vehicles. The object of the challenge is to produce and race a car that navigates through complex roadways, operates in traffic, parks, merges, signals, passes, and moves at traffic speed; completely through its own software intelligence and sensor systems.
Carnegie Mellon's BOSS charges up before Saturday's race.
Surprisingly, off-the-shelf sensors are the eyes of these advanced vehicles. Single line scanning LIDAR sensors from Sick adorned the corners and rear of the vehicles, and well-funded teams sported the rotating Velodyne 64-beam LIDAR scanner that covers a 25 degree wedge completely around the vehicle. Data from all of these sensors are collected and combined into a complete picture of the environment around the vehicle. Millions of distance and reflectance data points are generated every second and fed into the artificial intelligence (AI) systems. At this 3rd DARPA challenge, most teams had discarded direct imaging machine vision due to the difficulty of identifying obstacles and lane markings in harsh lighting conditions. While the LIDAR suffers some difficulties in very dusty environments and needs to be recalibrated for different humidity and environmental effects, it is the system of choice for rapid robotic vision.
It is hard to over-estimate how complex and dynamic the driving environment can be. Every situation that a vehicle encounters on the road must be handled by decision-making software, and as other vehicles and obstacles change position, these decisions need to be re-evaluated and updated continuously. For example, three program modules control the Tartan Racing Team's "Boss" - one handles the 2 million data points per second from the LIDAR sensors, one controls the moment-to-moment behavior of the robot, and the last module does route planning and navigation. Since the sole job of the sensor module is to create a projected 2 dimensional map of hazards from whatever data is available, the robot can actually gain or lose a sensor, or have their configuration changed, without affecting any of the other aspects of the AI. The overall speed of the race is determined by the speed at which the AI can maintain accurate decision-making for the vehicle. Throughout the contest, software engineers continuously retuned the parameters which governed the obstacle avoidance and vehicle speed to prevent a contest-ending accident. In the end, it really was what Stanford Team leader Sebastian Thrun called, "a software race."
Dirk Fabian inspects the Insight Racing team's Lotus packed with gear.
University teams dominated the race, especially those teams who harnessed the expertise of numerous experts and sponsors to supplement their internal knowledge. Volkswagen was highly visible in the contest, since their Passat is specifically designed to take a variety of electronic inputs, including steering, making it easier to convert AI steering commands into actual wheel positions. Vehicles of all shapes and sizes, from the enormous TerraMax truck to Insight Racing's Lotus bodies, were customized to accept computer control, mount sensors, and house the computers. Some customizations had the mark of hard-won engineering knowledge. Insight Racing found they had to pipe all the air conditioning from the tiny Lotus passenger compartment back to the trunk where the 9 Mac Minis need to be cooled.
Engineering feats were not the exclusive property of the well-funded. The University of Central Florida's entry, a slightly beat up 1996 Subaru donated by the wife of one of the team leads, caught attention, and not just because it is called "Knight Rider." This 'bot competed well in the 2005 DARPA off-road challenge and with some significant work on the artificial intelligence; the team got it ready for the urban environment. While other large teams enjoyed significant sponsorship and sported $75,000 Velodyne LIDAR sensing systems, UCF fielded a robot in the finals
Remo Pillat poses with his rotating LIDAR scanner on the Univ. of Central Florida's "Knight Rider."
with a team of just six and a total budget of just $150,000. Remo Pillat, Sensor Specialist for UCF, says that the team has had to "eat, sleep, and breathe robotics for the last year and a half." Pillat did a significant bit of improvisation to turn a pair of much cheaper sensors into the near equivalent of the Velodyne. A single beam LIDAR scanner is mounted on each roof corner and rotated back and forth through about 270 degrees. The scanned beam creates a dense point cloud and gives the robot a good wedge of information about its environment. Figuring out the optimum angle for the two scanners and how to turn the data into a usable point cloud is a significant part of Pillat's thesis on 3D terrain modeling and machine vision.
Pillat found the practical challenges to be the most motivating, despite the incredible investment of time required. "For the last three years it was my real passion to work on that car. Each weekend - Friday, Saturday, Sunday - we were out the whole day testing the car and writing software. And then during the week, I am a full-time PhD student on the side. My wife never liked it; she didn't get much time out of it." The opportunity to produce something that actually works in the real world was significant for Pillat. "I strive for practical application. That's what a university environment should do - have projects that students get excited about and actually achieve something in the real world. I think that is the embodiment of engineering." This sentiment was echoed often by other Urban Challenge engineers. In fact, many schools reported that applications to their engineering departments had risen since the last challenge, with many applicants specifically citing participation in the DARPA Challenge contest as their motivation for applying.
Everyone expects that practical experience from the Urban Challenge will be very useful for job hunting, especially with so many potential employers sponsoring the competing teams. This is where the Urban Challenge students will have an edge, Pillat believes, "a lot of students stop at some point. They publish some papers and their professor says, 'okay, you have some results, let's move on.' They never bring it to a level where they could put a real vision algorithm on a car." Challenge veterans have probably not only built one, but have debugged and optimized it, too. Now that the contest is over, Pillat has his sights set on the space robotics program at JPL (and writing a thesis, of course). Sponsors like Volkswagen and General Motors have their sights set on building the next generation of vehicles, probably after hiring students from the university racing teams. Immediate spin-off technology from the challenge will be the technologies and computer algorithms needed to make adaptive cruise control and obstacle avoidance systems more robust. Some of these advances are already available on high-end production cars, and LIDAR sensors will become more prevalent as time goes on.
The Urban Challenge did more than push autonomous vehicle design; it demonstrated a good working model of how a specific science challenge can be used as a strong motivator for research and innovation. During the first DARPA Challenge, the winning vehicle completed just 7 miles of the 150 mile course, but successive events have seen vehicle capabilities grow by leaps and bounds. This is good because DARPA has a congressional order to turn one quarter of all Defense vehicles into autonomous vehicles by 2015.
Bill Whittaker, proud own er of BOSS and recent winner of $2 million.
Besides the Urban Challenge, the Space Elevator Challenge, the Lunar and Vehicle X-prizes and the Solar Decathlon are well-known contest projects that students can get involved in. The Space Elevator specifically is optics and nanotechnology related, but all of the others touch optics technology. Taking on one of these challenges is a big commitment, but one where the rewards are proportionally large. The spirit of competition, and the publicity for the winning teams, often can provide access to resources unavailable otherwise. Even on a smaller scale, this kind of project based learning has many advantages in giving students valuable practical experience. Of course, without a major government sponsored prize to back it up, there wouldn't be over-sized checks to give out.