Re: DARPA Grand Challenge

humiguel@xxxxxxx wrote:
> HMS Beagle <bgates@xxxxxxxxxxxxx> wrote:
>
> >It's that time of year again....
> >http://www.darpa.mil/grandchallenge/discussion.html
> >
> >For an eye-opening experience, go through the red tape and read the
> >messages in this forum. Note the messages are posted backwards in
> >terms of date. So the "tops" of the threads are posted at the ends of
> >the message-pages.
> >
> >The ability to make a wheeled autonomous vehicle pass 175 miles of
> >off-road terrain in less than 10 hours is DARPA's Grand Challange. It
> >is obvious that this is a gigantic exercise in sensory-motor
> >coordination on a grand scale.
>
> I see that DARPA's Grand Challenge is going the GOFAI route. Will
> people never learn from the mistakes of the past? People are so used
> to make programs that run on limited environments where they can
> control every single variable that they forget that our natural
> environment is too complex and unpredictable to be treated in the
> same way.
>
> I hope that nobody at DARPA thinks that GOFAI will ever produce a
> vehicle that can be used in a real world situation, because if they
> do, sooner or later, they will get a big disappointment. A vehicle
> produced using that paradigm will work correctly until the day it
> encounters a situation that was not predicted by the programmers and
> at that moment it will fail catastrophically.

Last year, the barbed wire fence seemed to do that to a bunch of them.
It's exactly the type of thing you might not think about to code for
because it's small enough that you might think you could ignore it (like
you would ignore flys buzzing around) but deadly enough for the car that it
can't get through it. Many of them ran into problems with the fence:

http://www.off-road.com/race/events/2004darpa/images/hernandez-DARPA19.jpg

> In a real world
> environment this is not a matter of if, it's a matter of when
> something occurs that has not been predicted. This can be something
> as improbable as a meteorite hitting the side of the road or as
> silly as a mosquito dancing in front of the camera.

Mud splasing up on a camera or water spray off the wheels is another thing
that can require a lot of special code to deal with.

> Many people make the mistake of thinking that if we insist enough in
> the GOFAI route, we will some day reach intelligence in small
> incremental steps. This is not possible and people should need no
> further proof than the fact that for the past 50 years or so we have
> been trying to produce an intelligent conversation robot and
> currently we are not nearer that goal than we were at the beginning.
> GOFAI and intelligent systems use two completely different and
> incompatible models that run in parallel routes; they never meet.

Yeah, well, that just shows how slow people are to learn. :) But all the
GOFAI work is not wasted because it has given us a large body of evidnece
about what has not worked. In the long run, all the failures is what is
going to lead us to strong AI.

> Antonio Esteves

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.