Robotic Vehicle Competition - Introduction

This is the first article to be posted on the AI pages at RamaLila.net after a long hiatus. It is to be hoped that these pages will become a more regular feature again. The first few new articles will center around a recent robotic vehicle challenge sponsored by DARPA, the Defense Advanced Research Projects Agency. In this challenge, known as the DARPA Grand Challenge, a number of vehicles modified to be driven entirely by robots with no human intervention raced across a 132-mile course they were given through the Mojave desert. The teams of (human) AI experts who designed the robotics were motivated by the $2 million grand prize (doubled from the $1 million grand prize in the unsuccessful 2004 race). In 2004, none of the competing vehicles were able to complete more than 7.5 miles of the assigned course, although even this was probably a success as compared with the prior state of the art in robotics, and so the prize money was not awarded.

In 2005, on October 8th, a number of vehicles were able to successfully navigate the course. Because the start times were staggered, the vehicles did not compete directly against each other but against the clock. The winner of the race was a Volkswagen called Stanley designed by a team led by Stanford computer scientist Sebastian Thrun which was able to complete the course in less than 7 and a half hours.

Stanley, a modified Volkswagen, programmed by AI and robotics experts at Stanford University, finished the course in 7 and a half hours - Click here to see larger photo

This represents an important step forward for artificial intelligence. Robotic competitions have been a part of artificial intelligence conferences, for example the National Conference on Artificial Intelligence, for at least 15 years now. However, the earlier robots were very limited in their capabilities. Generally they would wander around a very small maze with carefully colored obstacles designed to make it as easy as possible for robots to make their way around. The course and the obstacles were designed with very simple shapes, simple colors, and generally over flat ground in a large convention center hall to make the robots’ algorithms as simple as possible. And, even then, the results were sometimes disappointing, at least to an outside observer. The robots would spend a great deal of time ‘thinking’ about what seemed quite simple decisions and then would often still end up making the wrong choice or getting stuck.

To be able to navigate successfully over real terrain in a tough, rugged environment (I’ve seen Rama students get stuck driving similar 4x4 vehicles off road on desert trips) should open up a whole new set of vistas for artificial intelligence. Although DARPA’s original domain of application is, of course, military, the participants in the competition envision applications such as automatically driven cars, household robots that actually work, yard work successfully completed without having to break out a sweat, and airplanes that can fly completely on their own. Now that we have a “proof of concept ” of a robot (multiple robots, actually) performing reasonably well outside the laboratory and in the “real world” it should only be a matter of time before the solution scales up to real robots performing real tasks for real people.

For the next set of articles on AI, then, it seems like it would be useful to look at what some of the challenges were that were faced by the different robot teams and how someone, if they were so inspired, might go about actually building a robot to perform similar tasks. This will be based on general knowledge of AI - here at RamaLila we don’t have direct knowledge of how Prof Thrun and his team, or the other teams, modified their vehicles to compete in the challenge. However, it seems likely that they would have followed a general approach that would have combined elements of AI planning, computer vision, machine learning, search, and reaction.

AI planning would be the highest level of the program that would drive Stanley or similar vehicles. At this level, the robot is interest in planning out a general course through which to take the vehicle. The robot is also interested in devising checkpoints to know whether it is keeping on plan or is getting off course and needs to replan or do a course correction. Closely related to AI planning is the problem of search. All AI programs search through a large space of possible solutions to find the most efficient one that fits the problem. In the case of Stanley, it would have needed to search through all possible routes through the Mojave desert to find the best way to reach the destination.

Computer vision is what Stanley would use at a lower level in actually executing that plan that it has devised through search. Stanley has to actually navigate through the Mojave desert and therefore has to recognize obstacles (rocks, vegetation, cliffs, impassable terrain, other vehicles) that it must avoid (and, if serious enough, possibly replan around). This requires the use of computer vision. It may also require the use of machine learning, especially during its many training runs, because its initial attempts to identify obstacles may not be sufficiently good to enable it to avoid them and so it has to refine its behavior through trial and error.

Reaction is the lowest level of the behavior of Stanley or similar robots. It is similar to the human stimulus-response reaction where, if you put your hand on a hot stove, you immediately pull it back. Stanley would need to be able to immediately react to obvious problems like getting into terrain that is too rocky, the wheels spinning indicating the possibility of getting stuck, feedback to indicate the speed of the vehicle is getting too high, and so on. The specific reactions might also change a little based upon the plan.

During the coming weeks, we will look at each of these areas of Stanley’s behavior in more detail to inspire people to learn how to actually build a robotic vehicle like Stanley. In the next article the focus will be on  search .


Home: Ramalila.NET



All copyrights are maintained by respective contributors and may not be reused without permission. Graphics and scripts may not be directly linked to. Site assets copyright © 2000 RamaLila.com and respective authors.
By using this site, you agree to relinquish all liabilities and claims financial or otherwise against RamaLila and its contributors. Visit this site at your own risk.