Grand Challenge

SOLAR CAR

Imagine:

https://www.eere-pmc.energy.gov/nasc05/

You are cruising down the highway in a shiny new

convertible, enjoying a beautiful, sun-drenched California morning. Your car, however, bears only a small resemblance to the automobiles of today. Though you are the vehicle’s only occupant, you are not its driver: the car is driving itself, powered by sophisticated artificial intelligence and an array of navigational sensors. And as the miles fly by, your vehicle consumes no gasoline and emits no exhaust; instead, its onboard solar panels harvest clean, renewable energy from the sun’s abundant rays. Stanford visionaries have seen this future, and they’re racing toward it at full throttle. In these two articles, we explore the University’s victories in two recent contests: the DARPA Grand Challenge and the North American Solar Car Challenge. Winning each competition required significant engineering breakthroughs – innovative technologies that may ultimately revolutionize the future of transportation.

http://cs.stanford.edu/group/roadrunner/presskit/high_res_pics/Stanley_Image2.jpg

GRAND CHALLENGE

The Stanford team’s solar car, Solstice.

Stanford’s entry for the DARPA Grand Challenge, “Stanley,” gears up for the race. Volume IV 19

https://www.eere-pmc.energy.gov/nasc05

A SHINING ACHIEVEMENT

by Inna Vishik

Stanford Solar Car Team Wins First in Class

L

ike many others, the Stanford Solar Car team spent the summer of 2005 soaking up the sun. However, unlike the beach-crawlers and sun-bathers, they used solar power for purposes other than improving their tans: to win first place in their class in the North American Solar Car Challenge. In this ten-day, 2,500 mile race from Austin, Tex. to Calgary, Canada, the team and their solar-powered car, Solstice, encountered many challenges including a Texas thunderstorm, cold Canadian nights, and a minor fender bender. They were able to cross the finish line in 68 hours and 4 seconds, an achievement that resulted from not only the engineering teamwork and ingenuity necessary to build their vehicle, but also the strategy employed during the race itself. The entirety of Solstice, from its body to its electrical and mechanical parts, was designed by the Stanford Solar Car team. The team collectively came up with the vehicle’s design, and then divided into groups that focused on its different components. Designing a solar car is a https://www.eere-pmc.energy.gov/nasc05/ unique challenge: it must be ^The team works on developing their harvest maximum energy from solar panels. the sun while using this energy to drive in the most efficient way possible. Even the best solar panels have less than 30% efficiency for converting solar energy into usable energy, so the car must have a large surface area covered by solar panels. Furthermore, it must run as long as possible on the energy it harnesses, so it cannot have extraneous weight and drag forces slowing it down. Finally, the vehicle must be built to adhere to the size and

Workings of the Solar Car

Photo by John Shen

Stanford’s winning solar car is covered in commercially-available solar cells, which absorb sunlight and convert it into electricity. Some of this energy is used immediately to drive the car, and some is stored in a battery for later use. The motor is connected both to the solar cells and to a battery, and the amount of energy it takes from each is controlled by computer. During the day, the car primarily uses solar power at lower speeds and battery power at higher speeds. The vehicle relies on stored battery power in overcast conditions and on hilly terrain. A key part of the team’s strategy is anticipating these unfavorable conditions, and saving enough battery power to be able to drive through them.

20 Stanford Scientific

safety restrictions imposed by the North American Solar Car Challenge to compete in the race. While designing the vehicle and building the parts was a challenge in itself, integrating all these components—the electrical and mechanical systems, the drive train, the solar arrays, the vehicle body, as well as a computer interface—was even more time consuming. The final stage in preparing for the race was extensive testing of Solstice. Much of the testing—essentially running the car until it broke, and then correcting the problem—was done in the Central Photo by John Shen Valley, because the terrain is flat. The Winning Car: Specs & Strategy The team covSolstice is the fifth generation vehicle built ered thousands by the Stanford Solar Car team. It can go up of miles in test to 80 miles per hour, and can run 250 miles drives. Testing on a fully charged battery. It has a thumb not only allowed throttle and arm steering, and the driver them to troubleis almost completely laying down (as seen shoot, but also above). Because vehicle weight and availprovided them a able power are scarce, there are no amenimeans of learning ties, such as radio and air conditioning. about the vehicle. They got a good feel for how the car operates in different driving conditions, which was invaluable in formulating a winning strategy during the race. As fossil fuels become more limited, one might ask if we will soon be driving solar powered vehicles in our everyday lives. The truth is that they are likely to remain an engineering novelty: one simply cannot harness enough energy from the sun to power a practical vehicle. Nevertheless, competitions like the North American Solar Car Challenge are vitally important because they encourage engineers to push the limits of solar power technology. Solar energy is a vast resource, and we are only beginning to realize its potential. *Special thanks to Ad Shelton, a recent Stanford graduate and longtime Stanford Solar Car Project member, for an illuminating discussion about the race and the team. *More information about the Stanford Solar Car project can be found at http://www.stanford.edu/group/solarcar/ Inna Vishik is a senior majoring in Physics. She likes long walks at SLAC, problem sets at three in the morning, and high temperature superconductors.

http://upload.wikimedia.org/wikipedia/en/9/9b/Stanley_at_the_finish_line.jpg

by David Gobaud

Stanford Racing SPEEDS through the DARPA Grand Challenge

D

r. Sebastian Thrun and the members of the Stanford Racing team waited anxiously alongside their competitors for the first car to cross the finish line of the DARPA Grand Challenge. It was October 8, 2005, in Primm, Nevada, and this was no ordinary race: the cars entered in the competition were piloted not by human drivers, but by robots comprising autonomous computers and sophisticated sensory devices. Thrun nervously squinted into the distance under the midday sun of the Mojave Desert—the loudspeakers had announced his vehicle was in the lead. At approximately 1:50 PM it came into view, charging toward the finish line. Stanford Racing watched with growing exhilaration as their robotic car, nicknamed “Stanley,” traversed the final leg of the desert race course. Minutes later, Thrun and his teammates were celebrating the victory: Stanley had zoomed past the finish line and into the history books as the first robot to complete the 131-mile DARPA Grand Challenge.

The Challenge In 2004, the Defense Advanced Research Projects Agency (DARPA) issued the DARPA Grand Challenge to spur development in autonomous vehicle technology-- a new field dedicated to the creation of self-driven, robotic vehicles. DARPA challenged competitors to develop a vehicle that could navigate a 150-mile route through the Mojave Desert from Barstow, California to Primm, Nevada. The route was specified using GPS coordinates; the vehicles were expected to navigate it using onboard sensors that included cameras and a radar. The prize for winning the race was set at $1 million. On March 13, 2004, the first DARPA Grand Challenge was held. Of the fifteen vehicles that started their journey from Barstow, none of them completed the course. Carnegie Mellon’s Red Team vehicle, Sandstorm, made it the farthest: a mere 7.5 miles, less than 5% of the length of the course. Some viewed the race as a failure, but because the event spurred innovative developments in the field of autonomous vehicle technology, others hailed it a moderate success. To encourage further progress, DARPA resolved to hold the race again the following year.

2004 race, as the team’s software development leader. Stanford Racing saw the underlying problem in the competition as a software design challenge, not a hardware one. According to Montemerlo, the teams competing in the 2004 DARPA challenge spent too much time developing their hardware and focusing on extreme off-road conditions, and too little time refining their vehicles’ software navigation systems. In Montemerlo’s eyes, this was due to the lack of prior knowledge about the competition and the potentially demanding race parameters set by DARPA. For example, the participants were only informed that the course could be up to 300 miles long and were given sparse details about the types of terrain or obstacles they would encounter. Several teams developed custom vehicles capable of traversing huge distances and extreme terrain but lacking in software capabilities. As a result, the computers controlling the vehicles were not able to navigate the course. “The robots failed for reasons of perception or control. [When] they…were presented with good terrain and bad terrain, they either didn’t see the bad terrain or they saw it and made the wrong decision and then drove off the road,” Montemerlo recalled. The results of the first DARPA Grand Challenge were impressive given the race’s requirements, he said, but there was significant room for improvement. In order to focus Stanley’s “brain” consists of their full attention on 100,000 lines of code with their vehicle’s motion planning and navitwo main components: a gation algorithms, pose estimator and a motion rather than tinkering planner. It bases its deciwith hardware, Stansions on information gathford Racing formed ered from the three sensors: a partnership with Volkswagen. They radar, a set of lasers, and a received a modified camera. Touareg R5 with a custom drive-by-wire system developed by

Stanford Enters the Race In 2005, Dr. Sebastian Thrun, Director of Stanford University’s Computer Science Artificial Intelligence Lab, decided to enter the second DARPA Grand Challenge. He formed a team called Stanford Racing, a true interdisciplinary endeavor drawing talented Stanford students, staff, and professors from the fields of computer science, mechanical engineering, aeronautics-astronautics, and optimization. Thrun selected Dr. Michael Montemerlo, a research associate in the AI lab and a spectator at the

http://cs.stanford.edu/group/roadrunner/presskit.html

Volume IV 21

The DARPA Grand Challenge Volkswagen of America’s Electronic Research Lab. This vehicle, ultimately used as Stanford’s entry in the DARPA Grand Challenge, was nicknamed “Stanley.”

Stanley’s Hardware

performance in navigating faster or more accurately. As a result, Stanley’s computing power was on the lower end of the spectrum among the vehicles entered in the Grand Challenge. One feature that Stanford Racing considered adding to Stanley’s hardware design, but ultimately chose to omit, was a redundant array of computer processors, so that the same operations could be performed using different hardware. This way, the vehicle could continue driving even if it suffered a computer hardware failure in the middle of the race. Unfortunately, redundant designs are extremely difficult to test; the number of ways to distribute the software increases exponentially as more processors are added. Furthermore, Stavens noted that Stanley might have encountered other failures, such as a flat tire or a broken suspension, which would have been even more catastrophic. This made redundancy a less critical feature to implement.

Stanley perceives the world through five lasers, one camera, and a radar system. Each of these sensors is static (unable to rotate). David Stavens, a PhD student in the Stanford Computer Science Department who worked on the project, notes that having static sensors is one of the major differences between Stanley and some of Stanford Racing’s competitors. Carnegie Mellon, for example, implemented a gimbal—a device consisting of two rings mounted on axes at right angles to each so that an object will remain suspended in a horizontal plane between them regardless of whether the support structure moves. This gimbal was utilized to point and stabilize The robot-controlled cameras and laser sensors. Stanley’s Brain vehicle completed the Stanford Racing decided to use static sensors Stanley’s “brain,” or central processing system, Grand Challenge course early in the development process. Stavens recalls consists of approximately 100,000 lines of code that the team made the decision since there are “a with two main components: a pose estimator and in a record-breaking 6 lot of reliability problems with developing a gima motion planner. The pose estimator determines hours, 53 minutes, and bal because you have very heavy sensors that have Stanley’s position, roll, pitch, and yaw. These coto be pointed very precisely. … So being able to ordinates are calculated from data provided by an 58 seconds. design a system that can point these very heavy inertial measurement unit, a GPS compass, and a sensors reliably while the vehicle is bouncing and wheel odometer. The planner “rolls out paths into shaking is very hard to do.” Indeed, gimbals would the future and decides where to go,” according to later prove unreliable: one of Carnegie Mellon’s two vehicles in the Grand Montemerlo. It bases its decisions on information gathered from the three Challenge suffered from a broken gimbal during the race. Stavens also re- sensors: radar, a set of lasers, and a camera. marks that the function of five lasers pointed at different ranges is similar The lasers constitute Stanley’s primary obstacle avoidance system. With to that of a gimbal. a range of 30 meters and a high degree of accuracy, the lasers create a twoStanley’s computing power consists of seven 1.6GHz Pentium M blade dimensional “grid” charting the occupied and unoccupied areas of the terservers donated by Intel. These processors, Stavens explained, are extreme- rain. Stanley’s obstacle avoidance algorithms were “trained” using a techly reliable as they are built to withstand natural disasters such as fires and nique called machine learning, in which a computer discovers solutions by earthquakes. To simplify Stanley’s software implementation, Stanford Rac- analyzing past experience. To train Stanley, a research assistant drove the ing opted for a design where only six of the seven processors were active robot-vehicle down the middle of a road and had it mark the terrain in the during the contest and only four of the six were used. The team designated center of the road as “free” and the terrain a fixed distance to the left or right two of the CPUs to execute the main software, a third to process imag- of the center as “occupied.” In this manner, Stanley would be able to tell es, and a fourth to record a log of the race for future analysis. Additional where it was safe to drive and where it wasn’t. The machine learning techprocessing power, the team realized, would not have improved Stanley’s nique was a fantastic success: it required only fifteen minutes of training

The Gimbal GAMBLE Designers of an autonomous robotic vehicle are faced with a decision when mounting sensors on the machine’s exterior. They must either affix the sensors as static, immovable devices, or use a gimbal to compensate for the automobile’s motion. A gimbal consists of two or three concentric rings, oriented http://ic.arc.nasa.gov/people/mdh/gimbal.jpg at right angles to each other, that con^A gimbal allows objects tinuously monitor the yaw, pitch and to remain suspended even if their supporting struc- roll of their support. Objects (such as sensors) situated within the rings will ture is moving. remain suspended in a horizontal plane regardless of the vehicle’s own motion.

A gimbal allows the robot’s sensors to rotate freely, allowing a single sensor to monitor many different parts of the surrounding terrain. Unfortunately, the weight of Stanley’s cameras and lasers, as well as the constant jolting motion experienced during the race, made constructing a gimbal a difficult proposition. Computer Science Ph.D student David Stavens explains that building a system “that can point these very heavy sensors reliably while the vehicle is bouncing and shaking is very hard to do.” Stanford Racing ultimately opted for a design consisting of multiple static sensors, each set to detect a different portion of the terrain. The result was a much more reliable system that ^ Gimbals compensate for three adequately matched the per- types of vehicular motion: yaw, formance of gimbal-mounted pitch, and roll. http://upload.wikimedia.org/wikipedia/en/1/19/ sensors. Flight_dynamics.jpg

22 Stanford Scientific

The DARPA Grand Challenge time and reduced the number of false positives from 1 in 8 to 1 in 50,000. hours later, the Oshkosh Truck Corporation’s TerraMax. Stanley’s laser-generated maps allow it to drive at approximately 25 Despite having crossed the finish line first, Stanley was not immediately miles per hour while safely avoiding obstacles. To drive faster, however, declared the winner. Additional calculations were necessary to account for it must acquire additional data from its onboard camera. Terrain marked as the robots’ staggered start times. “We almost didn’t…care because hav“safe” by the lasers is mapped onto a corresponding image from the camera. ing driven 130 miles [autonomously]… was a great achievement,” Stavens This portion of the image is then used to locate other safe terrain outside said. The next day brought more good news: Stanley was crowned the ofthe range of the lasers. Stanficial winner of the 2005 DARley utilizes the knowledge PA Grand Challenge. it garnered from the team’s machine learning techniques Future Research to categorize the image and Though the DARPA Grand locate safe terrain. If part of Challenge lived up to its ~ David Stavens the map contains a “hole”— name—proving a profound Computer Science Ph.D Student an unresolved segment that challenge for the field of armay be either safe or unsafe—the vehicle slows down in preparation for tificial intelligence—it was in fact a simpler version of a more general possible evasive action. Radar plays a similar role: if an obstacle is de- problem: “How does a robot navigate from point A to point B?” Removtected, Stanley reduces its speed, allowing the lasers to guide it to safety. ing the course’s speed limit or eliminating the precise route specification With its full array of lasers, cameras, and radar units, Stanley is able complicates the task immensely, Montemerlo said. If a vehicle is provided to safely avoid obstacles at a top speed of 35 miles per hour. Traveling at with only start and end locations, it must not only avoid obstacles but also speeds in excess of 35 miles per hour, however, is an extremely difficult perform “path planning,” identifying roads and charting a course to its deschallenge for Stanley. Since a vehicle’s stopping distance is proportional tination. to the square of its speed, Stanley must look ahead four times as far if it is One key feature Stanley lacks is the capability to avoid moving obstato travel at twice its present velocity. Fortunately for Stanford Racing, they cles. Such ability would require “another fundamental jump forward in the determined that completing the entire course of the Grand Challenge at a technology,” said Stavens. Currently, Stanley’s lasers scan a portion of the maximum speed of 35 miles per hour would not leave the robot at a sig- road and assume it will not change after the data is gathered. Moving obnificant disadvantage. In fact, this restriction would result in a completion stacles leave “streaks” in Stanley’s perception of the world. To remove this time only twenty minutes more than the optimal completion time because “visual clutter,” the vehicle must rescan the road at rapid intervals. Stavens only a small fraction of the course had a speed limit over 35 miles per hour. suggests that one nascent technology that will be a “fantastic” solution for Given this analysis, the team decided to limit Stanley to a maximum speed this problem is raster scanning lasers that scan both horizontally and vertiof 35 miles per hour for the entire race. The speed restriction proved highly cally. effective. Stanley was the only robot to complete four perfect runs during the qualification event and to navigate the Grand Challenge course without Autonomously Driving to Work Tomorrow? a single collision. Montemerlo and Stavens believe that it is impossible to predict if—and Once Stanley’s maximum speed was set, the vehicle had to determine when—fully autonomous cars will become commonplace on public roads. how fast it would drive each moment of the race. Again, the Stanford Rac- Montemerlo believes there are two factors determining when autonomous ing team employed machine learning to “train” Stanley how fast to move cars will become standard: technology and culture. Do people want cars in real time. This approach was another major difference between Stanley to drive themselves? Should there be separate roads for autonomous cars? and its competitors. The Carnegie Mellon team, for example, manually ana- Although bits of autonomous technology such as adaptive cruise control lyzed satellite data two hours before the race—when the route information systems are appearing in cars today, it may take years to answer these was released—and marked speeds for each section of the course “by hand.” questions and develop the necessary technology for fully autonomous Stavens argued that this procedure is not always reliable because of the cars. In the meantime, look to races like the Grand Challenge to continue dynamic nature of the course: a boulder, for example, might fall into the driving the innovation that could one day bring autonomous vehicles to a middle of the road, blocking a route that the team had marked. In addition, street near you.S aligning satellite data is a difficult challenge that complicates this process; a half-meter error on a road bordered by a cliff is enough to be fatal. Recognizing the limits of manually-determined speeds, Stanford Racing chose the machine learning approach instead, allowing Stanley to make its calculations in real time.

Having driven 130 miles [autonomously]... was a great achievement.

Crossing the Finish Line With hundreds of hours of research, design, construction, and testing behind Stanley, the robot-controlled vehicle completed the 131.6 miles of the Grand Challenge course in a record-breaking 6 hours, 53 minutes, and 58 seconds. Out of the 23 robots that entered the competition, it was one of only five robots that made it to the finish line, and one of only four that finished within the 10 hour time limit. The other vehicles that completed the Grand Challenge were Carnegie Mellon’s Sandstorm and Highlander robots, the Gray Insurance Company’s GrayBot, and approximately seven

http://smadness.com/personal/mike/2005_stanley-gc_709.jpg

^ The Stanford team rejoices after their victory in the DARPA Grand Challenge. David Gobaud is a junior majoring in Computer Science. Gobaud is the President of the Stanford ACM Chapter and was a member of the 2005 Stanford ACM Programming Contest Team. He is interested in artificial intelligence research and programming.

Layout designed by Anita Verma

Volume IV 23