00100	We  expect  a  short  term  payoff  in  industrial  applications   of
00200	now-standard  robotics  techniques.    A manipulator manufacturer has
00300	demonstrated a pump assembly. An industrial  laboratory  accomplishes
00400	visual recognition among a small number of more or less arbitrary 2-D
00500	shapes in one second  on  a  small  configuration  minicomputer;  the
00600	system  learns new shapes easily. Simple inspection systems are being
00700	marketed. Demonstrated techniques  of  touch,  joint  force  sensing,
00800	vision,  and  minicomputer  control  (perhaps  of  several  arms) can
00900	augment  current  industrial  manipulators.       These   incremental
01000	improvements  allow coarser assembly line tolerances, at lower system
01100	cost, with little increase in manipulator cost.
01200	
01300	Initially, when robots sense  misalignments  or  problems,  they  can
01400	branch among stored trajectories, or use simple search routines. When
01500	applications turn from repetitive transfer  operations  to  assembly,
01600	trajectory  planning  and  modification,  as demonstrated in our lab,
01700	will become important.           These middle range  and  long  range
01800	applications involve adaptive grasping, visual coordination, two hand
01900	cooperation, tool using,  and  complex  shape  recognition.       Arm
02000	prostheses  can  benefit  from  touch  and  joint force sensing, with
02100	reliable, compact electronics.  Visual prostheses  are  now  severely
02200	limited  in  bandwidth  at the human interface.     Our work in depth
02300	sensing suggests the value of visual depth prostheses.  Our  work  in
02400	motion  parallax  applies  to mobile automata, while manipulation has
02500	application to deep space missions  with  long  transmission  delays.
02600	Visual  studies  extend  the  knowledge  of intelligent systems, both
02700	natural and artificial.       We have developed  models  for  several
02800	psychological  phenomena,  and  useful  examples of representation of
02900	space  and  shape,  of  interest   in   representation   theory   and
03000	problem-solving.
03100	
03200	
03300				SUMMARY
03400	
03500	In  the  first  stage   of   development   of   computer   controlled
03600	manipulators,  we  have built fast and accurate hardware at low cost,
03700	and made computer control software to plan and execute  smooth,  fast
03800	motions.      The software has been systematised in an extensive user
03900	package, which is exportable.        This first stage might be called
04000	trajectory  control.           In contrast, a human plans motions for
04100	teleoperators and industrial manipulators. We are beginning a  second
04200	stage  of  arm  function which could be called the interactive stage,
04300	incorporating touch, visual and joint force sensing  and  interactive
04400	software for contingency planning,collision avoidance and reaction to
04500	sensed external conditions.
04600	
04700	We have  progressed  in  recognition  of  complex  three  dimensional
04800	shapes,  based  on  depth  data. We now make crude disconnected stick
04900	figure descriptions of some toys, including a doll and a horse. These
05000	descriptions are the basis for recognition and visual learning.   Our
05100	aim is to base recognition on these computer generated models, rather
05200	than externally given models; as a consequence, the system will learn
05300	new objects by seeing them.       An underlying shape  representation
05400	scheme and a formal mathematical representation of primitive elements
05500	support description of a useful category of object shapes.   We  have
05600	programmed  primitive motion parallax, separated objects in depth and
05700	achieved accuracy  much  better  than  initially  expected.    Motion
05800	parallax allows large stereo baselines and provides depth information
05900	beyond the range of useful stereo. The use of  texture  gradients  to
06000	provide  depth  cues  has  been  demonstrated; a program finds coarse
06100	boundaries corresponding to textured regions.
06200	
06300	We propose three tasks as goals for our research:
06400		simple assembly operations and tool-using
06500		description and recognition of complex objects
06600		interaction  with  a  problem-solver.  These tasks direct our
06700	concentration on the following subgoals:   for manipulation, improved
06800	touch  and joint force sensing, hands with three or more fingers, two
06900	hands, and software for collision avoidance  and  motions  guided  by
07000	force  constraints  of  the object (sliding a sleeve over a rod); for
07100	vision, complex shape description  from  the  laser  experiment,  and
07200	verification  techniques  using higher-level knowledge; for planning,
07300	plan reliable strategies and direct vision and  manipulation.     Two
07400	projects deserve special emphasis.    Touch sensing appears to be the
07500	most cost-effective development for advancing  manipulation  research
07600	and  applications.   While  we  will  continue  development  of touch
07700	sensing, broader support for adequate touch hardware  development  is
07800	important.    Also,  we stress the potential value of special purpose
07900	digital processors for  vision,  aimed  at  a  factor  of  100  speed
08000	increase over the PDP-10.
08100	
08200				MANIPULATION
08300	
08400	The  new arm, designed and built by Scheinman [1], was operational in
08500	December 1970.  The servo program, developed by R. Paul, was in use a
08600	few  weeks  later. In all respects, the performance of the new arm is
08700	dramatically better than that of  the  prosthetic  arm  it  replaces.
08800	After  extensive use,we are happy with it and the software.         A
08900	film demonstrates arm operation in real time.
09000	
09100	The servo program is unique  in  servoing  along  a  trajectory  [2],
09200	rather  than  point  by  point,  and  uses  a newtonian dynamic model
09300	instead of proportional servo. An extensive  user  package  has  been
09400	incorporated  into  the  hand/eye  system,  including an interpreter,
09500	WAVE.    A complete trajectory is pieced together  from  a  departure
09600	vector,  an approach vector, and a mid-segment which matches position
09700	and velocity with departure and approach.  Where  actions  cannot  be
09800	carried out in a single motion, an intermediate position is supplied.
09900	Trajectories may be modified by  patching  in  new  pieces  to  avoid
10000	collision or to remain within limit stops. These facilities have been
10100	standardized and  are  in  use  by  several  people.       Trajectory
10200	planning  software  is  largely  standard  ALGOL;  the arm model is a
10300	separate  module  and  could  be  replaced  by  another  arm   model.
10400	Translating  the  servo  program  to  a  standard small machine could
10500	provide a convenient interface to most laboratory systems.
10600	
10700	We plan minor improvements to trajectory control:  incremental  shaft
10800	encoders,  a velocity-control servo loop to augment computer control,
10900	and a sampled data system analysis.     We  plan  to  concentrate  on
11000	interactive manipulation; existing software provides touch interrupts
11100	and execution of a  functional  argument,  for  example,  incremental
11200	motion for visual alignment or grasping by touch.  The trajectory may
11300	be resumed, modifying it to smoothly compensate for the perturbation.
11400	Trajectory  modification  and  incremental  motion  make possible the
11500	versatile use of planned or stored trajectories.  A  PLACE  operation
11600	senses motor torques to place an object lightly on another.   We  now
11700	execute  motions  which  exert controlled forces, typical of assembly
11800	operations which we will attempt. We are building a wrist with strain
11900	gage  measurement  of force and torque; separation of components will
12000	be done by software.    We have built and used  two  forms  of  touch
12100	sensors,  a  4x4  array for each finger, and a simple switch. Initial
12200	experiments will use two finger hands  to  grasp  irregular  objects.
12300	That  task  properly  requires at least three fingers.  A five finger
12400	hand is under construction.  We plan to build improved touch  sensors
12500	and   complete   construction  of  a  second  arm,  to  be  used  for
12600	experimentation with hands, and two hand tasks.
12700	
     

00100				VISION
00200	
00300	
00400	
00500	As  discussed  in  the  last  technical  report,  we  have  developed
00600	apparatus for depth measurement and have ventured beyond the world of
00700	blocks, considering tools, toys, and rocks.          Agin and Binford
00800	conceived  and implemented a depth-measuring device using a laser and
00900	TV camera, operating in January 1971.  The accuracy is about  1/2  mm
01000	at 1 meter.  We are implementing description and recognition programs
01100	which generate their own representations,and do  not  use  externally
01200	given  models.         As  a  basis  for  the  description of complex
01300	objects,  Binford  [6]  has  developed  a   mathematical   formalism,
01400	"generalized local translational invariance", as a representation for
01500	a broad class of primitive elements, including  locally  cylindrical,
01600	locally  conical,  and  locally  helical,  to  be  used  in  compound
01700	representations.  We now separate elongated parts as "arms", each  of
01800	which  is described by a stick figure.  Our immediate goal is to join
01900	these "arms" together in compound descriptions, to view objects, then
02000	recognize them when seen again in different situations, with possible
02100	obscuration.
02200	
02300	Nevatia has used motion stereo to map depth  contours  of  rocks  and
02400	other   objects,   while   tracking  them  on  a  rotating  turntable
02500	(equivalent to moving the camera around the center of the turntable).
02600	Frames  are  taken over small intervals of motion; each frame closely
02700	resembles the previous frame.  Correlation of small areas is adequate
02800	to  match  successive images, without essential ambiguity.   Measured
02900	depth errors are about 2mm at 1 meter.  These errors are a factor  of
03000	eight  less  than  expected;  a  similar  phenomenon  is  observed in
03100	psychological  measurement  of  stereo  and   vernier   acuity.     A
03200	straightforward explanation is adequate for these phenomena.
03300	
03400	Bajcsy  [4]  has  used  the  Fourier  transform  as a spatial feature
03500	descriptor, particularly for  directionality,in  delineating  texture
03600	boundaries.   A  variety  of  outdoor and artificial scenes have been
03700	analyzed,  locating  large  areas  of  uniform  texture,  within  the
03800	resolution  of  its smallest window, 8 x 8.    Since possible texture
03900	features vary over a wide range of sizes, a range of window sizes  is
04000	used.      The  use  of  texture  gradients,  to  determine  observer
04100	orientation, as emphasized by Gibson [5], has been demonstrated.    A
04200	sheaf  formalism  has  been  evolved  as  a formal background for the
04300	analysis.
04400	
04500	The Instant Insanity puzzle solution [9] tested the hand/eye  system,
04600	arm software, color and other vision modules in an integrated task. A
04700	home movie shows a typical performance.  In  early  1971,  Winograd's
04800	BLOCKS  program  [3] was imported, support relations were generalized
04900	by Orban, to control hand and eye for building  stacks  and  arcades.
05000	Now  Moorer has a preliminary strategy program, with a miniature task
05100	language  and  a  planning  mechanism  which  evaluates  reliable  or
05200	cost-effective  solutions.  Our next system-wide task is execution of
05300	a connected set of  actions  requiring  strategy  and  monitoring  of
05400	execution,  in dialogue with the strategy program.   Monitoring gives
05500	particularly useful  examples  of  goal-directed  vision,  for  which
05600	knowledge  can  be used effectively. Gill has made several approaches
05700	to these problems using visual monitoring  of  arm  motion.   An  arm
05800	tracking  routine  makes  a  coordination  of the eye relative to the
05900	hand.  Routines for alignment of blocks in stacking are in progress.
06000	
06100	
06200	The  hand/eye  system  [7] has been the mechanism for integrating and
06300	standardizing diverse modules.  Both  the  Instant  Insanity  control
06400	program and the BLOCKS program are separate modules which communicate
06500	with sensors and the arm through a general  driver;  interfacing  was
06600	simple.      New  arm  routines,  the  COMPLEX body recognizer [8], a
06700	depth package, and a line drawing generator, have all been  added.  A
06800	number of changes to SAIL have been motivated by work in strategies.
06900	
07000	
07100	1.          V.Scheinman, Design of a Computer Controlled Manipulator,
07200	Stanford Artificial Intelligence Project Memo AI-92.
07300	
07400	2.     R.   Paul, Trajectory Control of a Computer Arm; International
07500	Joint Conference on Artificial Intelligence p.385 (1971).
07600	
07700	3.      T.     Winograd, Procedures as a Representation for Data in a
07800	Computer  Program  for Understanding Natural Language, MIT Artificial
07900	Intelligence Laboratory MAC TR-84.
08000	
08100	4.  R. Bajcsy, Computer Identification of Visual Textured Scenes, PhD
08200	Thesis, Computer Science Dept, Stanford Univ (1972).
08300	
08400	5.       J.J.      Gibson,  Perception  of the Visual World, Houghton
08500	Mifflin, 1950.
08600	
08700	
08800	6.      T.    O.    Binford, Visual Perception by Computer, IEEE Conf
08900	on Decision and Control, (1971).
09000	
09100	7.  J.A.  Feldman and R.F. Sproull, System Support for  the  Stanford
09200	Hand/Eye System, International Joint Conf on Artificial Intelligence,
09300	p.183 (1971)
09400	
09500	8.        G.      Falk, Scene Analysis Based on Imperfect Edge  Data,
09600	International Joint Conf on Artificial Intelligence, p.8 (1971).
09700	
09800	
09900	
10000	
10200	
10300	9.  The  Use  of  Vision  and  Manipulation  to  Solve  the  "Instant
10400	Insanity"  Puzzle;  J.  Feldman, K. Pingle, T. Binford, G. Falk, Alan
10500	Kay, R. Paul, R. Sproull, and J. Tenenbaum.
10600