Alt-BEAM Archive
Message #13847
To: he11bender@hotmail.com, alt-beam@egroups.com
From: "John A. deVries II" zozzles@lanl.gov
Date: Fri, 28 Apr 2000 16:34:28 -0600
Subject: R2D2 & "The Manager" panic mode
Since I took the notes, let me see if I can explain. I might be slightly
heretical about my answers, though .
The last time I saw it (probably two years ago) the R2D2 vacuum cleaner
looked like a miniature version of the robot in the Star Wars movies. Mark
designed its software so that it did some kind of mapping -- in other
words, after it had bumped into enough things, it had a clue where stuff
like the furniture was and would avoid hitting it in the future. The
essential difficulty (from Mark's point of view) was that his cat was more
clever than his computer-driven robot. Over a period of time, the cat
would repeatedly get in the robot's way -- and continue to do so in a rough
circle. Eventually the robot's "map" of clear area was very small indeed
and the robot was pretty useless as a vacuum cleaner unless you wanted one
tiny spot very, very clean.
Now comes the heretical part: Mark blames the use of a computer as being
the limiting factor in his design. Personally, I think that the basic
-design- of his software was flawed. I can only guess, but I suppose that
once an obstacle had been marked, it stayed marked -- essentially a binary
decision. If a value larger than merely on/off were assigned, then the
obstacles could "age" -- in other words, be forgotten over a period of
time. When it was hit, you'd start at the highest value, but as time went
on the value would gradually decrease until it was zero -- not an obstacle
any more.
Far more important than this, however, is the notion that any mapping need
be done at all. I don't know of any BEAM robot that has anything even akin
to a physical map -- indeed, none of them have memory other than what can
be provided by Nu (integrating) neurons or -- and this hasn't actually been
researched as far as I know -- loops of Nv neurons. Yet, Mark considers
BEAM robots to be superior -- but they are, in fact, superior merely
because they "aren't trying to be so smart".
With regard to the "panic mode of the manager" bit -- this has to do with
the Lobster robot, described in "Living Machines (go to
http://www.solarbotics.com
and search around -- you can get a PDF copy of
it there). Basically, the design of this robot was a more-or-less
conventional Nv loop that ran the leg motors directly, surrounded by a Nu
loop or (loosely speaking) a net. The idea was that the Nu net would
process input from the outside world and then provide that processed
information to the Nv net.
One of the basic design questions was how one ought to determine the ratio
of "responsiveness" between the outer "neural net" and the inner "nervous
net". Apparently, although one solution -- which involved relatively
little input actually being applied to the robot -- worked ok, that
solution did -not- work when more/a lot of input was applied. In my
opinion, the Nu net ought to have reduced the amount of information, but
apparently it didn't and as a result far too many signals were being
applied to the Nv net (which, of course, run the motors). With the Nv net
in a "panic", the robot "fell on its ass". It was utterly uncoordinated
until the inputs went away and the low-information solution was able to
"regain control".
Mark seems to have solved at least a portion of this problem with the
Unicore-style robot. Rather than having an entire loop of "managerial"
stuff around a loop of "worker" stuff, the Unicore is far more like a
conventional (non-BEAM) robot. In other words, the "head" is a bicore that
produces a signal that more-or-less reacts to the light around the robot
and that signal is coupled to a fairly independent "motor" bicore. Part of
what is important here, in comparison to the Lobster, is that the coupling
isn't extremely tight anymore -- in fact, you could cut off the head and
the robot would still be able to wander around blindly, even to the point
of backing up because there is a second sensory system, the one connected
to the "bumper switches".
Well, I've run out of steam here. I sure would like it if anyone who knows
anything about what I -think- I'm writing about would make comments,
corrections, etc.
Zoz
Home