Slashdot Mirror

← Back to Stories (view on slashdot.org)

Stanford's Stanley wins DARPA Grand Challenge

Posted by ryuzaki0 on from the congrats-to-all-for-competing dept.

tonyquan writes "DARPA has just announced that Stanford's "Stanley" autonomous ground vehicle has won the Grand Challenge, a $2 million contest for driverless vehicles over a 132 mile course in California's Mohave Desert. Stanley's winning time over the course was 6 hours, 53 minutes and 58 seconds, for an average speed of 19.1 mph. Second was Carnegie Mellon's Sandstorm (7:04:50), third went to another CMU vehicle "H1ghlander" (7:14:00) and fourth to the Gray Team's KAT-5 (7:30:16) More info from DARPA."

6 of 239 comments (clear)

  1. How few remain by necro81 · · Score: 5, Informative

    Looking at the final stats on the Grand Challenge website, it would seem that only five teams, out of the 23 that made the finals, were able to finish the course. The team that got the farthest before calling it quits managed about 80 miles, which means that the cut between those who made it and those who didn't was still pretty big. Another interesting thing about the final results is that, if you look at the pretty red and blue graph lines, they describe what looks like a sort of decaying function...

    Or perhaps I'm just a dork.

    1. Re:How few remain by zurmikopa · · Score: 5, Informative

      They blew a tire and were somewhere around 60 feet off course when they were eliminated, if I remember correctly.

      I know that a good number of the teams were actually still moving when they were eliminated; they had generally just wandered far enough off course that it was determined that they would be unable to finish.

      There were a number of reasons why people did so much better this year than last year.

      The biggest reason I think is that people knew a little better what to expect this year, and focused development on more important items for the race. For instance, for the first race I had done work on using a terrain database for path planning, but it turned out that the waypoints are so close together that it ends up just being a waste of CPU cycles for the most part.

      Another important reason is there was a rather large jump in the quality of the software running on the bots, and a moderate jump in the quality of the hardware. The integration was much more refined.

      Finally, the course was easier overall this year and the difficult part was put near the end. There was nothing in the course really comparable to Daggett ridge from the first race. Also, pretty much the entire course was graded along with the edges of the road often had banks. We had cliff detection that pretty much went unused this year due to this.

      Overall, it was a pretty good race this year. Stanford did an awesome job and really deserved the win. Not that you guys have that much interest, but we (Axion) ended up in 7th place (right after Ensco) with about 66 miles. We ended up getting stuck in some sand. The current candidate for the cause is a broken sway arm bracket that caused us to pull to the right a bit. Further analysis will be required to determine if that's actually the case.

  2. Re:19.1? by necro81 · · Score: 5, Informative

    The course did have a fair number of twists and turns in it. There were some places, like dried lake beds, where the cars could open up a bit, but for the most part it was bumpy dirt tracks one which even you or I couldn't do more than, say, 40 mph. There were also, intentionally, a fair number of obstacles designed to throw the computer systems off. You and I wouldn't have much difficulty in recognizing a cattle gate on a road, but imagine trying to teach a computer vision system to distinguish that. In other cases, the robots had to drive through tunnels that would not only be dark (making vision systems less accurate) but also lack any GPS signal.

    So, yes, it did average out to a pretty slow "race." But, on the other hand, it is a marked improvement over last time, when no one even came close to finishing. I think that, in the interests of trying to ensure that they safely finished the course, let alone win, the various teams were playing it a little conservatively, and not trying to go for pedal-to-the-metal performance. Maybe next year, now that they have some confidence.

  3. More info by Zathrus · · Score: 5, Informative

    For far better info than the anemic (and completely flash based) gc.org site:

    http://www.darpa.mil/grandchallenge/discussion.htm l -- DARPA's GC message boards
    http://www.tgdaily.com/2005/10/08/darpagrandchalle nge2005/ -- Was updated throughout the actual event. Best coverage I've seen yet.
    http://www.popsci.com/popsci/darpachallenge/ -- Popular Science's rather disorganized site

    I'm still looking for "highlight" video myself... or pretty much any non-bland video (seeing them cross the finish line is nifty and all, but that was not a challenging part of the race). I particularly want video of Alice trying to take out some reporters!

  4. Re:so wait.. by EEJD · · Score: 5, Informative

    It's not so much an improvement in the AI as it is an improvement in the sensors. These vehicles look ahead about 30 feet and plot their course based on very simple logic. If there is a negative obstacle (a hole), it is more difficult for sensors to detect than if there is a rock sticking up in the path. Last race, the only thing that stopped red team was a hairpin turn. Their sensors looked straight ahead and only a little to the sides, but when faced with the hairpin turn, the vehicle almost fell off the side of the mountain! But the rules of the AI haven't changed much- just the sensors. If you're driving through jungle, for example, you have to have sensors that don't see leaves as obstacles. Otherwise the path will look totally impassable.

  5. cmu won all three by Anonymous Coward · · Score: 5, Informative

    sort of.

    the stanford leader (thrun) and their lead software developer
    (mike montelermo (sp?)) were originally from cmu.
    they only recently moved to stanford. although thrun claims it's coz of his wife, some people think it was coz of too much competition and bad blood at cmu which has lots of people working in mobile robots (wittaker, simmons, nourbaksh, choset, ...) while i think palo alto has much better weather than pittsburgh :)

    the particle filter based localizer and mapper was developed while at CMU. Frank Dellaert (now at georgia tech) first introduced that to mobile robotics after reading about the
    condensation algorithm in computer vision (i like to believe that i had a part in that last bit :) I would'nt be surprised if they also use large parts of the basic control and command software infrastructure (TCX) written by thrun and others while at cmu. if it is, no wonder they required
    7 PCs for redundancy, that is some of the worst spaghetti code i've ever had the displeasure of working with. it's easier to make it fault-tolerant by just throwing more hardware at it.

    i'm not trying to belittle stanford in any way, but i just thought people might be interested in knowing that the real story in this case is a lot more complicated. the relationship between the winning teams were a lot more incestuous :)

    thrun BTW is an amazing all-round guy with an infectious smile all the time.