First, thoughts on the "Thoughts on the "Big Picture":

I like to think big, then work at the little steps that brings me
closer to the big picture.  Unfortunately, this tendency, coupled with
my relatively recent introduction to the entire subject may mean that
I may be far too optimistic or even absurd (optimism coupled with ignorance
can be very dangerous, yet also rewarding).  Nevertheless, it helps me to
focus at the task at hand.

Comments and critism are, as always, most welcome.  It is entirely
possible that I am repeating things that are already stated/thought out,
in which case, stating them again wouldn't hurt (except for wasting the
poor reader's time).  Another possibility is that I am stretching the
viability of this architecture to its limit, in ways that it was never
intended.  You have thus been warned.


Thoughts on the "Big Picture":

I envision a complete, autonomous, adaptive, mobile robotics architecture
that incoporates three main phases:

1.  Exploration
	Could be very similar to the ELDEN system.  Doesn't necessarily
	have to be effective (if it stays in a certain area, it only
	means that more readings, and thus better occupancy grids
	will be created for that place; if it doesn't stay in that
	certain area, more places are learned); tradeoff between
	quality of places learned versus quantity of places learned.

2.  Place Learning
	Upon some determination (say if it entered a "doorway", a
	marked location, or even at random intervals), the place-
	learning module kicks in.
	
	We can utilize the Moravec code to generate the o.g. of the
	current place.  The robot would then attempt to classify
	(on its own) the generated grid with the previous grids it has
	in its library -- if good match, "Bayesian" merge new grids with
	old grids -- if not good match, add a new place to place library.

	(It is, thus, no longer "lazy."  It actively tries to improve
	its place library with every new o.g....)

	(As the places gets more and more developed, is it worthwhile to
	attempt random classification between places in the place library
	itself?  Tradeoff between computation now versus efficiency later,
	when we are actually using the grids.)

	(If we utilize "doorway" mechanism, we can use the doorway to
	"align" ourselves before we generate our map.  This might result
	in new places generated by slanted furnitures that happen to
 	simulate a doorway.  That might be okay.. we will just have
	places that are rarely visited... or we could keep our trusty
	compass system and avoid this problem (with the compass, even
	if furniture signals a "doorway" the grid that is built would
	still be correctly "classified" in the place that
	contains the offending furniture))

	Once we have a reasonably adapt place recognizer, we might want
	to build up some sort of connectivity graph between places so
	we can later use prior knowledge of being in place A to tell use
	likelihood of being in place B.

3.  Dynamic Adaptation
	After the first two phases (maybe it ran overnight when
	everyone was home), the office starts to become chaotic as people
 	and furniture moves around during the day.  Our poor robot will
 	still be moving around (perhaps using the same exploration
	program as in phase one).  Again, when it reaches a place, 
	it tries to classify the place as a place it has seen before
	(the threshold may have to lower than from phase two, though
	since we are talking about a much more chaotic environment).
	Once we have classified and registered a place, we may have
	noticed that (1) things that were in our "historical" grids
	aren't in our "current" grids (2) things that are not in our
	"historical" grids are in our "current grids.  Slowly, but surely,
	we would "Bayesian" anneal out (again, more slowly than in phase
	two, since we are talking about a dynamic environment)
	such dynamic elements from our "historical" grids such that
	(1) only relevant, permenent features remain (such as walls)
	(2) we have areas of "dynamic uncertainty" (i.e. places where
	    people would like to place furniture.)
	Furthermore, we can refine this phase by selecting a subset of
	pertinent features or landmarks from among the "permenent" aspects
	of each "place" to speed registration and recognition.

If (a big if) everything works out correctly, we now have a robot that
can:
	(1) entirely and autonomously build up knowledge about a place
		with no previous knowledge of its location nor
		topology (and improves with increased exposure)
		(applications include exploration of underground caverns
		on other planets or robot inflitration of enemy military
		bases)
	(2) can adapt to dynamic environments consisting of mixtures
		of permenent (walls) and dynamic (furniture) elements
		and be able to utilize permenent features and ignore
		dynamic features
		(obvious bonus)
	(3) from any current location, correctly "classify" both the
		"place" it is in and its relative location within
		that "place"
	(4) have reasonable knowledge of the connectivity from place
		to place
		(3 and 4 allows for planner to make their plans)

Unfortunately, since the robot is autonomous, it might have to stumble
through errors on its own.  And it might take many training sets
to correct for such errors.  Furthermore, the robot might run into
difficulty with similar looking rooms (as might be the case in many
offices)... perhaps utilzing connectivity to resolve conflicts?  Only
experimentation can tell how far this will go.

Phase one seems to be domain of Mr. Yamauchi.  Phase two is relatively
complete with the work of Mr. Pfleger, Mr. Langley, and Mr. Sahami.  The
few touches to phase two only require a few coding that I could
do, especially with data/code that already exists.  Phase three looks
fairly complicated, but doable from the standpoint of someone that
has not done it yet.  The techniques exists... it just requires a lot
of experimentation to nail down the exact methologies.  I see my main
goal as filling the gaps that exists in phase two and three and leaving
phase one and the complete incoporation of the total architecture to
future brave souls.  Other directions exists, of course.