COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002			Goals of Hand/Eye
C00011 00003			Low-Level Representations
C00013 ENDMK
C⊗;
		Goals of Hand/Eye


Our goal  is to  understand and  build intelligent machines,  smarter
tools, which  aid man in interacting with  the real world, perceiving
the   external   world   and   manipulating/controlling   it,    both
interactively and automatically.  Our  field is AI to the extent that
such a  goal requires intelligence.  It overlaps other areas of AI to
the extent that the mechanisms are shared between  hand/eye and other
areas of AI.  It is in the forefront of AI to the extent that:
  the  mechanisms  and  representations  necessary  for  progress  in
  perception  and manipulation/control are  particularly important to
  the progress of AI;

  and that  necessity  is particularly  clear and  intuitively  clear
  approaches to the solutions are available in hand/eye. 
Interactive systems are interesting to AI to the extent that they
require the mechanisms of intelligence.


SCIENTIFIC GOALS
Any program corresponds to a sequence of maps between representations
(not  usually  made explicit).   Our  goal  is to  make  explicit and
interesting representations which make intuitive sense to  humans and
which make possible efficient programs.  We further want to enumerate
the non-equivalent representations, absorb redundant representations,
and adopt standard  forms for the non-equivalent  representations and
develop  the mathematical  relations  between these  structures.   We
further want  to find  the computational  complexity of  programs  in
terms of the  math structure of  these representations.   When we say
that  we want  a  computer to  aid  a human,  one sense  is  that
the human should  be
able to speak  in a high level  language to it.  That  means that the
computer    must    have   the    representations    and   algorithms
(representations  and  maps)  to  go  from  the  desired  high  level
representation  input   to  low  level   perceptual  or  manipulation
representation.
Thus the computer can understand and keep track of details in doing
so.
   We see  such a  high  level language  as a  set  of
representations, a semantics.  We see this as  a paradigm for AI, the
detailed semantics (representation) of particular domains.  Moreover,
most of the other areas of AI have essentially limited the domains of
their representations to set theory, the most trivial of semantics. 

GUIDING QUESTIONS
STRUCTURE OF VISUAL ALGORITHMS
  representations
  maps between representations
  relations between representations
    which are non-equivalent
    lattice of partial order (inclusion on set of representations as support)
  computational complexity
  parallel realization?
WHAT IS THE WORLD THAT WE CAN SEE IT?
  The world did not design itself to be seen.  We designed ourselves
  to see it.  Speech is a set of conventions to be understood.
  What are the physical characteristics which we sense?
DESIGN AND BUILD LOW-LEVEL STRUCTURES, HIGH-LEVEL STRUCTURES
DESIGN AND BUILD PROGRAMMING SYSTEMS
  Vision depends on the utilization of a variety of perceptual facilities:
  stereo, color, edge, texture, geometry, shape representations.  The only
  way to use these cooperatively is to make an integrated system.  AL and
  the vision language are to perception what FOL is to the MTC effort.
  In so doing, we learn a lot about the relationship of visual structures.
HOW TO PROGRAM CLASSES OF PERCEPTUAL PROBLEMS?
  It has been typical to program by any means whatever, a particular task.
  Now we address the problem of generating programs to deal with a class
  of similar problems.
THE POWER TO SOLVE PERCEPTUAL PROBLEMS LIES IN THE SPECIFIC CONTEXT
  Accordingly, it is important in vision to investigate specific systems
  in detail, and encode the semantics for that domain.  We think that is
  a paradigm for AI as a whole.
TEST AND EXPERIMENT TO DEVELOP ADEQUATE THEORIES
  We cannot foresee all difficulties and we do find things we might not
  have reasonably have been able to foresee.
  May take longer to model the world than to USE THE WORLD AS A MODEL.
  Things we cannot foresee ahead of time are properties of the world itself
  which we would like to measure:  how uniform are surfaces,
  The world is so complicated that it takes a lot of effort to model it;
  It is so devious that we can't model it well.
I probably would not try to do assembly without hardware.
The specified problem really focusses on the primary problems.

PDQ makes the comment that nowhere do we make the case for actual manipulators
and actual sensors. I see that as coming from the need to experiment and test.
What sorts of things have been learned that were not obvious at the beginning.
With simulation, etc.  Things are overlooked.  What was learned in Paul's
thesis which could not have been learned in other ways?  
  dynamics prediction-variable gain --taking into account model of the arm
    in terms of gravity, friction, weights--feedforward and feedback
    All of this could be foreseen given some science approach, as opposed to
    the sitting down and programming approach.
  some of the assembly expts point out the need for things which would not
   be encountered without real world hardware
      wobble to get screws out in pump assembly
      what one can do with those errors
  what are necessary for the roving cart?
  what are we measuring about vision?
    reflectivity function, edge effects
    details of particular application domains
		Low-Level Representations


SENSOR:	joint	joint	wrist	motor	finger	clock
 	pots	tachs	gages	current	gages	
	  |	  |	  |	  |	  |	  |
	  ↓	  ↓	  ↓	  ↓	  ↓	  ↓
PERCEPT:joint	joint	wrist	joint	finger	time
INNER   angle	vel.	force	force	force	
	  |	  |	  | 	  |	  |
	  |	  |	  ------- | -------
	  ↓       ↓             ↓ ↓ ↓
STATE:	pos     veloc   	force
EXTERN  &orient





	goal 		sensor
	  |		  |
	  ----        -----
	     |        |
	     ↓	      ↓
	      dynamics
		 |
		 ↓
	      predicted
		 |
		 ↓
	       motor
	       drive






------- ai  ------ ai,comp sci ---------- programming ---------
|WORLD|----→|MATH|------------→|COMPUTER|------------→|PROGRAM|
-------     |REPR|             |  REPR  |             ---------
	    ------	       ----------