[From Bruce Abbott (941021.1050 EST)]
[Paul George 941021 10:00]
The example of MIT's "Collection Machine" is used. Apparently this machine
wanders aimlessly around a university lab at night, picking up empty soda
cans and dropping them into a recycling bin. This machine has no central
authority: its arms, wheels, and sensors find their own way to work together
without a central processor and no human intervention. Abstracting from this
experience, MIT developed this method for devising a system of distributed
control (how many of us are currently working on this subject?):
A non-technical book I recently read on the problems experienced by the field
of "artificial intelligence" (I can't seem to recall the name or author at the
moment) describes both this machine and the "Dante" machine that recently had
such a "glowing success" exploring a volcano and was funded by NASA as a
prototype Mars explorer. The Collection Machine uses what its builder calls
"subsumption architecture," which appear to be a collection of independent
low-level control systems imbued with a limited but useful ability to learn
"what works" from experience. The version I read about had legs, like an
insect, and not only learned to coordinate its legs, it could climb over
obstacles in its path such as a brick. The Dante machine, in contrast, uses
"heuristics" and other AI-type processing to determine what to do. It takes
forever to make a decision and is capable of moving only a few feet per hour.
High-level control is usually imposed from above via operators who communicate
to it via a radio link. Despite all this "sophistication," it and its
predecessor have both managed to fall on their backs and become trapped while
exploring. Ridiculous!
This was a topic I planned to bring up later as I did not want to get too many
sticks in the fire at once, but now that you brought it up, what I was going
to say is that it seems to me that some members of this group could no doubt
design a Mars robot that would do everything NASA could ever hope for, using
an HPCT-based design. The builder of the Collection Machine is on the
right track.
Bruce