Re: darpa show tonite


Yves Quemener wrote:
feedbackdroids wrote:
mimo_545@xxxxxxxxxxx wrote:
Stanford relied mainly on short-range laser scanning to build real-time
internal representations of the terrain directly ahead and immediately
to the side, and produced a probablistic mapping of good and bad zones
for travel. This was combined with a long-range camera image of the
road several 100s of meters ahead for course prediction. Then overlaid
the internal maps. Worked great, although at one point in the course
the
vehicle slowed greatly and went erratic back+forth across the road.
It wasn't explained, but the sensors were probably picking up false
positives on obstacles, and trying to steer around.

I didn't see the show, but I did intend to a conference where a Stanford
team member explained their technologies (it was 2 months before the
callenge, IIRC) He talked about the laser-scanner, but also explained
that they merged its information with a color camera in order to produce
the good/bad zones mapping. From what I understood, they only used the
color information of the pixels to identify the dangerous zones such as
water flakes and to find the orientation of the road.




Hi Yves. You might be interested that there are team technical papers
to be found online, and also an overview of Sebastien Thrun's technique
for probabliistic robotics.

http://www.darpa.mil/grandchallenge/TechPapers.html
http://robots.stanford.edu/papers/Thrun01h.html




He also insisted on the fact that they needed to use simple and reliable
technologies instead of fancy, smarter systems (to the dismay of the AI
researchers present). Memorably, one of the scientists asked a few
questioned and conclued "You know, when I came to this conference, I was
thinking that the Darpa challenge would not be won before two or three
years, but now, I think that you have a real chance to make it."

.