DARPA Announces 2005 Grand Challenge Semifinalists

Mockingbird writes "DARPA announced 40 semifinalists for the 2005 Grand Challenge autonomous robot race today. Notable remaining teams include the Carnegie Mellon University Red Team, Stanford Racing and a high school team, the Palos Verde Road Warriors. 78 teams missed the cut. The race, which will take place on Oct. 8, 2005 features a $2 million prize for the first team whose robot crosses 175 miles of the Mojave in under ten hours. The robots must be fully autonomous, with no team intervention allowed once the vehicle is launched. The first race was held in 2003, when the most successful team managed to log only 7.4 miles."

Comments

iRobot

[Anonymous Coward]

The Roomba actually only made it 4 miles, but it cleaned up the competition...

(groan)

Re:A few questions...

[neuroinf]

You need to look at the footage from the last attempts that showed how easily they become stranded on top of fence-posts. You wouldn't think it was possible to destroy a heavy vehicle like that, but a human will back off when hearing the gearbox squeal - not a computer. The vision stuff is an absolute nightmare - any sensor is better than vision. It sure is a serious challenge. I expect maybe 30 miles this time?

Re:code for winning robot

[Spy Hunter]

petal_to_the_metal_grandma();

Unfortunately the race was lost when, due to a tragic typo, the robot drove off-course to deliver flowers to a statue of an old lady.

Why 3D Computer Vision is HARD

[TERdON]

Ah, coder? Do you have good knowledge of math too? Then YOU could possibly make the breakthrough in 3D Computer Vision.

I'm studying a course in 3D Computer Vision right now, at TUHH. It's part of the Erasmus exchange program I'm having here - the eigth and last semester (excluding the thesiswork) of my master of engineering in automation and mechatronics at Chalmers in Gothenburg. I can easily say this course is the most difficult one of all I've been taking for all of my study time, hopefully the three weeks I have between that exam and the last of my others, will be enough to learn what doesn't stay in my head during the lectures...

In fact, I have the course book right beside me. To begin, the description of it would be more or less along the lines "an orgy in linear algebra, mathematical statistics, with some flavouring of image processing, geometry, optimization and algorithms". Basically, it's 30-40% mathematical formulas, 650 pages, some containing things not even all MSc even learn like tensor notations etc. Not something I'm even sure is a good thing to recommend to very many slashdotters, even. You'll get its name though - "Multiple View Geometry in Computer Vision", by Hartley & Zisserman. ISBN 0-521-54051-8.

What I see as problems in the book, is that almost everything is working on corner detection. This is great, if you want to make 3D-models of houses or other man-made objects (at least half of the examples in the book are architectural, I would say). It's not so great if you want to image bushes, rocks and other things with not so obvious corners on them. Also, the process involves quite heavy processing - both image processing, finding all those corners, statistical processing (to sort out outliers, which there will be), and optimization to find the best fitting backprojection of the image planes). I don't have a sure grip on the needed processing power but I doubt, when considering realtime demands in a car, that it'll hardly be easy to get it working.

Also, it's still to a big deal itself an area under research. The situation with using 5+ images (from different cameras och just consecutive images from the same, moved camera), isn't very well known. Using more images, of course would mean a bigger chance to get a decent 3D model of the scene...

And still, you would at least need two cameras to do anything useful. You can't reconstruct 3D space without having at least two images of the object to reconstruct. And probably you will need more - you would probably want to reconstruct all the way around (ie more cameras on the sides and backwards), and add extra sensors like radar etc for extra checks.

And then you really haven't solved the problem of driving the car. You have only built a decent mapping of the 3D surroundings of it. You have to add AI/some kind of steering logic, which only in itself is a demanding task. Just look at all FPS games out there - if it would be easy to construct good AI, with a known 3D-world, tailormade for the figures, would we really be seeing that many games with crap-AI? I'm happy I ain't taking an AI course too, for sure!