perm filename NSF.FOO[DOC,TOB] blob
sn#051538 filedate 1973-06-26 generic text, type T, neo UTF8
00100 ABSTRACT
00200
00300
00400 The Stanford Artificial Intelligence Laboratory proposes to undertake
00500 a program in advanced automation over a period of one year at a cost
00600 of $200,000. This proposal is to continue and extend programs
00700 sponsored under ARPA Contract SD-183.
00800
00900 The laboratory proposes a one year investigation in the area of
01000 computer integrated assembly systems. This project unites two
01100 dominant forces in industrial automation: computerization of
01200 manufacturing and programmable assembly automation. The laboratory
01300 brings to the project experience in robotics research and an existing
01400 robot system which has performed a simple assembly task and which
01500 integrates simple touch and force sensing, and rudimentary visual
01600 skills.
01700
01800 This one year proposal is intended to be part of a three or four year
01900 "proof of concept" effort. The first year will be devoted to
02000 extending our capabilities, demonstrating them to industry and
02100 establishing collaboration with industrial firms. This information
02200 will be used to shape the larger effort.
02300
02400 Extensive industrial and governmental consultation has shaped the
02500 proposal. Several industrial needs have emerged from studies now
02600 underway. One is a widespread desire of major industries for
02700 programmable automation. Another is the desire to cut costs by
02800 better overall organization and understanding of the manufacturing
02900 process, using computers as a tool for organization and information.
03000
03100 The proposed project would: a) extend the capability of our existing
03200 robotic system by investigating the use of two arms in typical tasks,
03300 by improving touch and force sensing techniques and developing
03400 necessary software in carrying out assembly tasks. b) demonstrate the
03500 utility of the assembly system by performing assembly tasks of
03600 moderate complication and demonstrate that the system could be
03700 replicated at acceptable cost to industry by completing a stand-alone
03800 system on a minicomputer. c) attack the problem of interfacing
03900 programmable assembly devices to a computer-aided manufacturing
04000 system and the problem of programming general purpose devices for
04100 assembly tasks. We will implement an example of an interactive robot
04200 programming system which will be an exercise for design of an
04300 APT-like programming system for assembly devices.
04400
04500
00100 I. THE NEED FOR ADVANCED AUTOMATION
00200
00300
00400 The contributions that advanced automation can make to enhance
00500 productivity in manufacturing are now widely appreciated. Current
00600 interest in advanced computer-oriented automation stems from two
00700 sources. The first is the decreasing cost of computation combined
00800 with an improved understanding of how to apply computation to various
00900 aspects of manufacturing. The second cause is the realization that
01000 many current manufacturing operations, primarily assembly and
01100 inspection, will require complex automation methods which are now
01200 research techniques. The same considerations apply to the automation
01300 of tasks such as laboratory and medical procedures, but the emphasis
01400 has been (as ours will be) on manufacturing. A recent ARPA-sponsored
01500 study [Anderson 1972] in which we participated suggests the following
01600 major benefits from advanced automation:
01700 improved national productivity
01800 reduced prices
01900 increased product quality and uniformity
02000 customization of products
02100 increased job satisfaction.
02150 We would like to add just one comment. As we have seen with other
02250 technological developments such as numerical-controlled tools, there
02350 will be things made which were previously impractical to make. There
02450 is widespread agreement in the technical, industrial and government
02550 communities on the potential benefits of advanced automation. There
02650 is much less agreement on the level of sophistication needed in
02750 automation hardware and software. While there are large application
02850 areas in which specially constructed machines or simple general
02950 purpose devices will prove optimal, we believe that the most
03050 practical solution for intermediate quantities (500 to 100,000) will
03150 be to develop general purpose assembly and inspection devices
03250 controlled by sophisticated programs in digital computers.
03400
03500 These conclusions are supported by our consultation with industrial
03600 management who have manufacturing responsibilities. Programmibility
03700 is necessary even in the automobile industry which has large volume,
03800 but essentially custom assembles each car. This has been repeatedly
03900 emphasized in workshops.
04000
04100 Another conclusion is that the process of improving the organization
04200 of manufacturing can make heavy use of computer systems.
04300 Computer-aided design is already in heavy use in semiconductor
04400 industries, and making a beginning in other industries. Inventory
04500 and parts tracking are important facilities in the aircraft industry.
04600 Direct numerical control of several N/C machines by a central
04700 computer is also extensively used. Some effort has been directed in
04800 Europe towards the "group technology" approach of Opits, which
04900 attempts to emphasize commonality by using parts descriptions to
05000 group items which are alike. Evolution of computer-aided
05100 manufacturing depends on integrating these facilities and bringing
05200 more facilities under effective control.
05300
05400 These two systems, programmable automation and computer-aided
05500 manufacturing interact in a beneficial way. An important part of the
05600 development process for programmable automation is the development of
05700 programming systems which actually program assembly devices. Part of
05800 that can be done by "learning by doing". However, many operations
05900 are only small modifications of other operations, so that it is
06000 desirable to be able to modify the assembly program instead of
06100 reprogramming entirely. It would appear that most assembly
06200 operations could be strung together from modifications of modules
06300 from a library of standard assembly subroutines like inserting and
06400 screwing. Inspection combined with assembly (which is made necessary
06500 for quality control and because of the implicit inspection in
06600 assembly operations) and error and contingency conditions lead to an
06700 assembly control structure like a conventional computer program.
06800
06900 The central computer will have parts designs and assembly operations.
07000 Why should these not be used to supply information to assembly
07100 subroutines? And should the designer have a system which allows him
07200 to design for automatic assembly by interacting with the system which
07300 programs the actual assembly device? That would provide a useful
07400 experimental system for cost and design automation, in addition to
07500 shifting some of the responsibility for programming of assembly
07600 devices from production employees to professional staff with
07700 technical sophistication. We emphasize the separation into an
07800 assembly system which could operate stand-alone, but is better
07900 interfaced with a central computer system, and the robot programming
08000 system which depends on more facilities but is used infrequently when
08100 jobs are changed.
08200
08300 There is one further concept in common among presentations of
08400 automation. Automation must be incremental. This is supported by
08500 numerous considerations: invested capital in existing plants, design
08600 costs for conversion, and the necessity for a sequence of
08700 understandable and predictably profitable steps along the path.
08800
08900 We contend that inserting programmable assembly devices incrementally
09000 into existing factories or new factories with only slight
09100 modifications, demands more from the programmable device. We see
09200 this process at work as industrial robots of the Unimate class slowly
09300 develop more facilities. The simple robot described in the ISI study
09400 [Anderson] does not exist yet; it must still be proven feasible by
09500 demonstration. We outline below some arguments about cost and
09600 productivity factors which follow from programmable assembly. These
09700 arguments follow from compatibility with current manufacturing
09800 design, from economies of scale, and from commonality of technology,
09900 and from desires for quality of product and predictability of the
10000 process. Several studies have indicated some priorities among
10100 directions of automation. The ISI study [Anderson] indicates that
10200 less than 20% of discrete manufacturing is amenable to special
10300 purpose automation. A Chrysler Corporation study indicates that car
10400 assembly labor has first priority in automation; assembly plant
10500 direct labor is 40% of direct corporate labor. The ISI study quotes
10600 a figure of 75% as the fraction of batch produced discrete
10700 manufactures. One conclusion is that "the emphasis in development
10800 should be on simple vision devices and on low-bandwidth tactile and
10900 force sensing feedback". That study indicates that the assembly
11000 process typically accounts for 50 percent of production cost.
11100
11200 We assume that present manipulators have a small but stable market,
11300 and that most will have minicomputers soon (Sundstrand does now).
11400 Will the incremental cost of touch, force and visual sensors be
11500 offset by the opening of new capabilities, and will the cost be
11600 offset by reduced cost of system design for machines which allow
11700 loose alignment tolerances? We estimate that touch and force sensing
11800 will add about 5% to the price of current manipulators of the Unimate
11900 class. We estimate that by 1977, the incremental cost of visual
12000 sensors, a laser ranging device such as we use now,and additional
12100 memory requirements, will add about 15%-20% to current manipulator
12200 prices. If labor costs continue to rise, part of the increase will be
12300 offset, but cost comparisons will not change appreciably over that
12400 time scale unless significant new mechanical designs are made. A
12500 conclusion is that sensor-equipped robots will be available at slight
12600 additional cost. We will discuss the benefits of incremental
12700 improvements to manipulators.
12800
12900 We do not expect mechanical production costs to drop as dramatically
13000 as electronics costs fall with integration. The most promising cost
13100 reduction development would be high-production standard modules such
13200 as robots. As an example, consider the Stanford arm; produced in
13300 unit quantity, parts, manufacturing and assembly costs are
13400 approximately $10K. The arm is much simpler than an automobile; there
13500 is no reason why it should cost more. A smaller, more recent design
13600 costs $3500 each in quantities of two.
13700
13800 At what cost are robots economically viable? This varies enormously;
13900 if they are serving a large and expensive machine, the cost can be
14000 relatively high. Unimation now projects labor costs at $22,000 to
14100 $24,000 per man year averaged over the next 10 years, and assumes a
14200 payoff of 2 years for a projected acceptable cost level of $46,000. A
14300 robot with two arms, minicomputer control, force and touch sensors
14400 could be constructed for under $10K in quantities of 1000 now. To
14500 summarize: two arms (new small Scheinman design) $2000 with hand; arm
14600 drivers and arm interface electronics $1000; force and touch $800;
14700 small computer 8k memory $4500; total $8300. Assuming that the
14800 computer price falls to $1200 by 1978, this will at least compensate
14900 for inflation. Of course, these are small arms, suited for parts
15000 assembly, not automotive body assembly. The larger arms, assuming
15100 slight redesign for cost, would cost about $2500 in units of 1000,
15200 for a total system cost (not price) of $12300. This is the range of
15300 price of present transfer machines which have limited market because
15400 they have limited motion, no control, and are thus locked out of the
15500 economy of scale. There is a dynamic which leads to economy of scale;
15600 the greater the versatility, the wider is the market for
15700 applications. This will be true particularly in small production and
15800 job shop operations which have been difficult to automate. The
15900 larger the production, the more versatility for the same initial cost
16000 as special devices. Also, with high volume, standardization of
16100 maintenance, and availability of parts and service are improved. The
16200 quality of the device is also improved.
16300
16400 The economies of scale in production of assembly modules are matched
16500 by economies of intellectual scale in development of design for
16600 automation. The design process is modularized and easily
16700 communicable; there is a common discipline which can be widely
16800 absorbed; developments of many sources can be integrated. Programming
16900 and training done once for one machine can be transferred to all
17000 machines in a plant, and even transferred nationwide. Functional
17100 modularity has been emphasized in our studies. Modularity is clearly
17200 possible and there is a powerful economic incentive to assemble
17300 widely varying systems from one, two, or three arm modules, force and
17400 touch modules, vision and ranging modules, with planning and
17500 monitoring modules. This would also entail software modularity with
17600 computer-aided mechanical design and job specification for robots on
17700 a local or network time-sharing service.
17800
17900 The alternative is special design for each operation. We expect that
18000 at some point, design will be cheaper with standard robot modules
18100 than with special purpose designs, just as it is now more economical
18200 to use mini-computer controllers than special purpose controllers.
18300 Modularity has the potential to reduce design time and improve
18400 predictability of design time and costs, permitting better planning
18500 and response to competition. To some extent, the dehumanized
18600 industrial worker now fills the role of standard module, so that we
18700 do not expect discontinuous change. The continuity with decades of
18800 experience with design for human versatility in assembly can be
18900 relied upon; robots with some versatility, sensing, contingency
19000 actions, and dexterity will provide a transition to newer schemes
19100 which will evolve more slowly. Another advantage of general purpose
19200 techniques is the ease with which the size of the robot can be scaled
19300 up or down. There appears to be no difficulty in making very large or
19400 very small assembly stations which use essentially the same programs.
19500
19600 Versatile assembly devices have an efficiency advantage in medium
19700 scale and job shop production; their utilization efficiency can be
19800 higher because versatility allows a variety of operations to be
19900 scheduled at one station. The availability of computers for
20000 production scheduling will allow the flexibility to be used to
20100 advantage. Robots would also have longer life over which to amortize
20200 costs, since they could function with greatly changed production
20300 operations, and evolve with new hardware and software modules. They
20400 would have higher reliability, since they would have monitoring,
20500 contingency action, and error recovery. They would be self-aligning,
20600 self-calibrating, and self-testing.
20700
20800 With simple devices, inspection and assembly operations must be
20900 broken down into small operations with transfer operations between.
21000 This introduces extra transfer devices with orientation-preserving or
21100 re-orienting steps, and overlapping operations which can be combined
21200 in an advanced device. The extra cost of jigs and fixtures, and of
21300 stopping assembly lines adds to the cost of special purpose
21400 automation. Robots can cut down on real estate, since several
21500 operations can be done at one station.
21600
21700 Because of the adaptibility of work stations, production runs can be
21800 smaller, with less capital tied up in inventory and unused machines.
21900 There can be better response to the needs of high technology industry
22000 which often requires small quantities. There has been difficulty in
22100 meeting the tight tolerances necessary for operation of
22200 non-intelligent devices. It is hard to estimate the value of systems
22300 which work within current industrial tolerance, rather than requiring
22400 radical improvements in engineering practice, but that continuity
22500 does seem important.
22600
22700 Serious issues arise in the development of robots with simple sensory
22800 systems. Some evaluation issues should be settled by demonstration,
22900 since much sharper analyses can be made this way.
23000
23100 There is already a significant degree of industrial involvement in
23200 computer-aided assembly. GM Research Lab has demonstrated a wheel
23300 mounting operation using visual alignment. IBM has a small robot
23400 project. Hitachi has demonstrated a robot with touch and a system
23500 with visual identification of objects on a conveyor belt, with visual
23600 control of manipulation. Can industry be expected to adequately
23700 support this development? Development of primitive robots is an
23800 evolutionary process, not a single product development. The area is
23900 too big; the payoffs are huge and the investment level is large.
24000 Also, traditionally, industry has been dependent on suppliers. But
24100 with the exception of IBM, the suppliers are smaller and less able to
24200 risk the development. Management techniques require inordinate
24300 return on capital. An alternative support could come again from the
24400 military. The ARPA study and the history of numerical controlled
24500 machine tools are natural consequences of the vast military
24600 procurement in technologically advanced products. In the case of N/C,
24700 funds could be diverted from an outmoded technology to development of
24800 devices which were essential to production plans for the future. It
24900 is high technology products which are most competitive
25000 internationally in the civilian economy. The record of government
25100 investment in numerical control, solid state electronics, computers,
25200 and electro-optical devices has been very productive. However, in
25300 large scale development of civilian products and international
25400 markets, foreign governments have taken the lead. The development of
25500 general purpose assembly devices will benefit manufacturers in a
25600 broad spectrum of industry, particularly medium scale manufacturing.
25700 It is an appropriate area for governmental action to help meet a
25800 national need.
25900
26000 Given sensor-equipped robots at little extra cost over current
26100 senseless robots, they will have visual ability to sense color,
26200 strong contrasts in brightness, recognize known objects from two
26300 dimensional outlines (top view of flat object) or three dimensionally
26400 using depth, to a certain extent they will be able to recognize three
26500 dimensional objects using monocular recognition. They will be able
26600 to visually align mating parts in assembly operations, increase
26700 reliability by self- calibration and flaw and disaster detection, and
26800 permit large tolerances in parts transfer, even permitting parts to
26900 be taken from a bin. Several of these abilities are not yet
27000 available in the laboratory, but with known conditions and objects,
27100 and some engineering (e.g. color coding) are achievable within the
27200 time span of the proposed project.
27300
27400 The major determinant of the economic impact of assembly robots will
27500 be the extent of their capabilities. Assessing and extending these
27600 capabilities is the principal goal of the proposed project.
27700
00100 II. RESEARCH PLAN
00200
00300 A. THE PILOT ASSEMBLY STATION
00400
00500 A basic goal of the proposed project is to extend our existing robot
00600 system, to specialize on assembly tasks, and meanwhile, to transfer
00700 the system to a standard minicomputer, in order to demonstrate its
00800 applicability to real problems. The project will be considered a
00900 success if it has major impact on an industrial assembly system or or
01000 if an enterprise chooses to manufacture and market the general
01100 purpose station.
01200
01300 Our current robot system has one arm of our design, two TV cameras
01400 and a laser ranging system. A variety of vision modules have been
01500 standardized, and a hand language (HAL) has been used for several
01600 tasks [Paul]. A demonstration of the robot's capabilities has been
01700 made by assembling a water pump, using vision, touch and a tool.
01800
01900 We propose to extend that system by: a) doing more assembly
02000 experiments to guide research in developing the hand language and
02100 manipulation techniques. b) design and implement visual operations
02200 especially for assembly, which would not be programmed in our general
02300 vision effort, again using assembly experiments to guide research. c)
02400 adding a second arm which is nearly completed, and improved touch and
02500 force sensing; extending the hand language to two arm operations.
02600 demonstrate practical speed and dexterity with two arms and tool
02700 using. d) after some progress with b, implementing a language which
02800 combines vision with the hand language. show the commonality of
02900 operations among similar assembly tasks.
03000
03100 Simultaneously, we would transfer our robot system to a minicomputer
03200 which we are acquiring. One benefit of this transfer is for our own
03300 research. Another is that we can show the relevance of our research
03400 to systems which are cost effective for industrial applications. Yet
03500 another is that we can transfer our research results more directly to
03600 industrial observers and industrial laboratories. In many cases,
03700 this can be direct transfer, since we have the beginning of a
03800 robotics user's group sharing the same minicomputer (GM research lab,
03900 3M research lab, JPL, and Artificial Intelligence lab, MIT). Also,
04000 our hand language, arm control and arm trajectory planning software
04100 are suitable for a variety of arms, including the Unimate.
04200
04300 The result of this effort is a small system we call a pilot assembly
04400 station, based on a minicomputer. Two related projects complement
04500 this proposed project. These projects at SRI and MIT, Draper Lab
04600 have industrial participation to work with commercial manipulators of
04700 a type already in use in factories (Unimate). They propose to
04800 develop systems which perform simple two dimensional inspection tasks
04900 and systems for manipulation at about our current level of dexterity.
05000 With their industrial participation, and a commercial device, their
05100 impact may be short term. Our project is longer term in that while
05200 we have those accomplishments in manipulation and assembly, we work
05300 with equipment which is not a commercial product, so that there may
05400 be a delay in converting our techniques to a commercial manipulation
05500 system, or marketing our manipulator. Finally, we are aiming
05600 toward a higher level of evolution, something like the simple robot
05700 described in the ISI report [Anderson].
05800
05900 In one configuration, the assembly station is assumed to be connected
06000 to a large time-shared computer facility. Not only does this appear
06100 to be a good design for our problem, but it is also necessary for an
06200 integrated programmed automation facility. The only emphasized
06300 statement in the ARPA study report [Anderson 1972] is: "The maximum
06400 impact of computers in manufacturing will come from complete,
06500 real-time computer cognizance and control of all processes and
06600 resources, allowing the precise scheduling and allocation of these
06700 resources."
06800
06900 Each assembly station will consist of a small digital computer, one
07000 or more manipulators and cameras. Manipulators and cameras will be
07100 combined as modules in systems to suit a given task, and will be
07200 digitally controlled by the station computer. The station computer
07300 may be connected to a central computer or may stand alone. The latter
07400 configuration is less efficient in the long run, but allows for
07500 incremental automation.
07600
07700 The program in the station computer will be flexible as to the number
07800 and position of manipulators and cameras. It will control input from
07900 the cameras, provide for some pre-processing of the picture and the
08000 application of simple tests to the visual data. In the case of
08100 manipulators it will provide for simultaneous servo control. It will
08200 also process other inputs, such as from touch sensors on the
08300 manipulators, photocell interrupt devices, limit switches, etc. Both
08400 analog and binary outputs will be available for the control of other
08500 devices.
08600
08700 Each station computer will have access to a task program specifying
08800 the actions and tests it should perform to carry out its task. This
08900 task program will be compiled by the central computer. Compilation
09000 will consist of the analysis of predicted scenes in order to specify
09100 simple tests that the station computer can carry out, the calculation
09200 of collision free trajectories for the manipulators, the sequence of
09300 operations and simple contingency plans. etc. We are planning for
09400 tasks whose normal execution does not require detailed scene
09500 analysis. If scene analysis becomes necessary (e.g. a dropped part)
09600 we envision the station sending an image to the central computer for
09700 processing.
09800
09900 The central computer will perform overall control of the individual
10000 assembly stations and the relative rates of production. If at any
10100 station a planned test fails, for which a contingency plan does not
10200 exist, or any unexpected sensor signals are received, the central
10300 computer will also be interrupted and appropriate actions taken. If
10400 the assembly station is not connected to a central computer, these
10500 operations will require human intervention.
10600
10700
10800 Most of the hardware required for the assembly station has been
10900 developed for the laboratory. The major additional problem is to
11000 refine the existing designs to meet the demands of an industrial
11100 environment.
11200
11300 We are now acquiring a small computer, a PDP-11, and will interface
11400 the existing arms to it together with one or more cameras. A software
11500 package for the small computer will then be developed in order to
11600 control the manipulators, read the TV cameras and to perform low
11700 level visual analysis. Programs to plan tasks, calibrate the system
11800 and to perform higher level visual analysis, error recovery and
11900 strategy planning will be written in a high level language for the
12000 time-sharing processor.
12100
12200 The most difficult part of the assembly station, and the heart of the
12300 project is a demonstration of the feasibility of the station. We plan
12400 to do this by developing programs which actually assemble and inspect
12500 a variety of fairly complex manufactured items.
12600
12700 Our first task was assembly of an automobile water pump. A film
12800 shows the process in detail. The assembly consists of four sets of
12900 operations: Visually locating the pump body to move it to standard
13000 position; inserting guide pins and the gasket, then visually
13100 verifying that the gasket is in place; picking up the pump head (at
13200 an approximate standard position) and placing it over the guide pins;
13300 using a tool to pick up screws from a feeder and insert them, by
13400 searching near the hole if necessary. Tools, vision, touch and force
13500 feedback were used. The assembly was at approximately human speed,
13600 but not as fast as a practiced worker. The reliability is relatively
13700 high. Assembly was programmed in the hand language, with a small
13800 program, in terms of macros like "insert a screw". Visual operations
13900 were special case programs. A major part of the development of the
14000 pump assembly was implementing tools and finding manufacturing
14100 engineering techniques. Speed and reliability in repetitive tasks
14200 could be increased by using errors found in searching to insert
14300 screws to update target positions for screw insertion, etc., so that
14400 these operations were almost always immediately successful.
14500
14600 As we move to more difficult tasks, the sophistication required goes
14700 up quite rapidly. The current manipulation routines seem to be
14800 expandable to cover anticipated problems, especially since we intend
14900 to employ special purpose manipulators (tools) for some tasks. A
15000 major part of our effort will be vision for assembly; an assembly
15100 task will be performed using vision, touch, and two hands. The
15200 vision effort will be integrated with other abilities, and while no
15300 great generality will be stressed, new modules will be incorporated
15400 into the system, and a set of techniques will evolve. These will
15500 depend heavily on techniques similar to those of our recent studies
15600 in visual feedback and hand/eye coordination [Gill 1972]. The
15700 internal model of the world will be used to predict visual appearance
15800 of selected areas of the scene, and a set of verification procedures
15900 will confirm or contradict the predictions. Many of the operations
16000 will be connected with supplying feedback to manipulation, for
16100 alignment and insertion. These visual procedures will have models for
16200 the parts and approximate locations.
16300
16400 This and other visual tasks connected with assembly will initially be
16500 performed as special case problems. In addition, to make these
16600 techniques available, we will pursue development of strategy
16700 programs. This step is essential if intelligent assembly devices are
16800 to be conveniently programmed for complex tasks. Within the division
16900 of labor between large and small machines, this facility resides in
17000 the central time-sharing facility which has large memory, file
17100 storage, and problem-solving languages. The system must have a model
17200 of the parts for assembly,and a sequence of instructions for
17300 assembly. It will be necessary to have a representation of shape and
17400 models for usual assembly operations such as insertion and alignment,
17500 and the ability to solve simple problems in connection with these
17600 operations, all this to enable convenient communication with the
17700 user. When these facilities exist, the addition of new features and
17800 operations can be incorporated smoothly, so that users can call on a
17900 growing library of planning modules. The result of a planning effort
18000 is a cycle of machine planning and man-machine interaction which
18100 results in a plan network for the mini computer which controls
18200 individual machines.
18300
00100 PROGRAMMING ASSEMBLY DEVICES AND INTERFACING WITH THE CENTRAL FACTORY
00200 SYSTEM
00300
00400
00500 The second part of the proposed project is to design and implement an
00600 initial version of an interactive system for programming assembly
00700 devices and interfacing with a central factory computer system.
00800
00900 The most obvious form of programming robots is that in use with
01000 current industrial robots, "learning by doing". The robot is led to
01100 end points of a trajectory or along a path by an operator. We are
01200 able to do that already; positions for the pump assembly were
01300 specified in just that way. The process could be improved by better
01400 teleoperator control, for example the resolved rate control system at
01500 MIT.
01600
01700 Those techniques are quite limited, however. They are limited to
01800 tasks with positional control only. Where force control is
01900 necessary, the control program must know in what way to control
02000 forces, i.e. what forces to exert or null along what direction
02100 vector. For example, in inserting a rod into a hole, transverse
02200 forces should be nulled. Such tasks require specifying a trajectory
02300 with additional symbolic force control information. Since arm
02400 control will be primarily based on minicomputers in new designs, a
02500 natural mode to accomplish this partially symbolic programming is
02600 with an interactive program, which can be small.
02700
02800 For any task involving a sequence of such operations, the interactive
02900 system allows the operator to choose from a menu of operations and
03000 proceed step by step to assemble a sequence of these operations which
03100 have been verified. Checking the success of these operations could
03200 use simulation if necessary to avoid tying up a production machine.
03300 Useful modules like inserting a screw would be saved as macros so
03400 that subsequent programming sessions would be much simplified,
03500 dealing with combinations of such high level modules. The macros can
03600 have arguments, so that one basic operation can be used in a variety
03700 of situations, and further, our software allows modifying
03800 trajectories so that a planned trajectory allows considerable
03900 variation in execution, without replanning. We have built and used
04000 such an interactive system for the hand.
04100
04200 Current work is involved with extending the hand progrmamming
04300 language to include more than one arm. Problems of synchronization
04400 and coordinated motion will be provided for.
04500
04600 Later this work will become the output language for a higher level
04700 processor, in which tasks such as inserting testing and mating will
04800 be expressible in a simple and direct manner.
04900
05000 As visual feedback is included, the human operator must be able to
05100 interact with the visual system through a graphics console. These
05200 are becoming inexpensive, but it is a resources shared among many
05300 assembly stations. Unlike the hand language, no interactive visual
05400 system exists. We must emphasize that development of the language is
05500 not primarily a language effort, nor is it a statistical learning
05600 system. Rather, we plan that it resemble the hand language, i.e.
05700 embedding the useful set of visual operators in a convenient
05800 interactive system. During the early stages of the effort, the
05900 language will be at a low level, with macros, much as the hand
06000 language is now. That early emphasis will be on building a trial
06100 system based on a lot of experiments done as special cases. This can
06200 be regarded as an extended learning by doing. Meanwhile, our present
06300 efforts on representation of objects will evolve to a state at which
06400 the modelling of objects is integrated with our graphics capabilities
06500 to begin to automate some of the programming of feedback operations,
06600 that is to allow programming at a higher level. It is at this level
06700 that a non-technical worker might accomplsh the programming of visual
06800 feedback tasks.
06900
07000 We will meanwhile work on development of strategies for assembly.
07100 The goal is to produce a language at about the level of an
07200 instruction manual: put the cover (part a) on the opening of part b.
07300 At such a level, the system begins to be applicable to design for
07400 automation. The designer can use the system to simulate the assembly
07500 process and modify designs which are marginal for automatic assembly.
07600 He will have an assembly program as part of the design. Thus, a
07700 large part of programming for assembly may in fact be done by
07800 designers.
07900
08000 Several factors may decrease the burden of programming robots for
08100 production line personnel:
08200 libraries of operations-fastening, inserting, etc.
08300
08400 interactive systems which do bookkeeping and
08500 prompt the human operator- an extended sort
08600 of learning by doing
08700
08800 programming by designers.
08900
09000 The process of building the APT-like language for assembly should be
09100 based on design, not on just random accretion, and should be based on
09200 computer science inputs, to avoid forseeable and unnecessary
09300 limitations on syntax and implementation. It should be based on a
09400 wider perspective than just one industry or the market strategy of a
09500 dominant computer company. With APT, the language was developed with
09600 air frames in mind, and has been clumsy in other applications.
09700 Further, the development should be based on extensive experience,
09800 which can be obtained sooner in a research environment where a
09900 variety of tasks and skills can be investigated.
10000
00100 B. RESEARCH IN VISION
00200
00300 Research in computer vision has progressed so that the larger part of
00400 our effort is directed toward vision in real scenes. We outline past
00500 research, including work on visual feedback, development of useful
00600 modules, analysis of some outdoor scenes, interpretation of scenes
00700 with curved objects, and a study of sensor systems.
00800
00900 Our past work on visual feedback [Gill 1972] in stacking blocks and
01000 inserting objects in holes is directly relevant to vision for
01100 assembly. The issues involved are: coordinating hand and eye,
01200 utilizing incremental accuracy of vision and manipulation, accurate
01300 prediction to limit search, and limiting accumulation of errors.
01400 Visual guidance in picking up objects and stacking or inserting
01500 objects overcomes errors in the visual and arm coordinate systems,
01600 and errors in correspondence between the two coordinate systems. This
01700 was done under realistic conditions, with the hand still holding the
01800 object and in view; the system predicts the appearance of edges and
01900 corners within a small window around the edges to be aligned, and
02000 uses a corner/edge finder to determine alignment error. The arm makes
02100 accurate incremental error correction. If predictions can be made
02200 accurately, the search window can be small, which increases the speed
02300 of response and decreases possibility of confusion with other areas
02400 of the scene. A major limitation to accuracy of prediction (and to
02500 long-term reliability in the industrial environment) is drift of
02600 calibration. That work incorporates calibration updating by
02700 self-calibration to coordinate hand and eye coordinate systems and
02800 provide for reliability, as well as to limit search by accurately
02900 predicting expected features. Visual feedback with stereo vision is
03000 much more powerful than without; we have demonstrated stereo visual
03100 feedback in a program [Binford unpublished].
03200
03300 We have a set of visual modules to be used in a goal-directed way.
03400 These modules answer questions which arise in practical visual
03500 feedback tasks, and in analysis of outdoor scenes. Are there color
03600 edges in a selected window? Are there edges in brightness? What is
03700 the color of the region in the window? Several preliminary color
03800 region analysis programs have been developed [Binford, unpublished,
03900 Bajcsy 1972]. An edge operator [Hueckel 1969] and a region-based
04000 corner/edge finder [Gill 1972] both operate on small windows of the
04100 scene, as directed by a higher level. An edge verifier is used to
04200 confirm or reject edges predicted from context [Tenenbaum 1970].
04300 Color identification has long been demonstrated. Recently, a program
04400 has been implemented [Thomas unpublished] based on Land's Retinex
04500 model of color constancy in vision, which defines a color triangle
04600 from ordering of brightnesses. The center of the triangle is
04700 designated as white. The program shows a color constancy similar to
04800 humans; table top scenes are assigned colors reasonably well under a
04900 variety of lighting with different spectra: sunlight, flourescent,
05000 and incandescent.
05100
05200 Understanding outdoor scenes is a more complex task than many
05300 industrial visual operations, but it has utility for planetary
05400 exploration; it is also a useful exercise for developing color,
05500 texture, and stereo operations. Bajcsy analyzed outdoor scenes with a
05600 combination of color and texture region analysis [Bajcsy 1972]; depth
05700 cues were taken from the gradient of texture. An attempt was made to
05800 obtain consistent interpretations by making three dimensional
05900 interpretations, and confirming boundaries by independent properties.
06000 Natural objects and natural scenes have been used for our work with
06100 motion parallax [Nevatia in draft] and stereo vision
06200 [Quam,Hannah].These are complementary techniques of depth perception
06300 which are applicable for planetary exploration and industrial tasks.
06400 Motion parallax can be used for planning of navigation and
06500 characterizing large scale terrain features, since it provides large
06600 baselines.
06700
06800 Another aspect of vision with direct application to industrial
06900 environments is explored in our work with curved objects. Current
07000 programs use depth information to make useful descriptions of objects
07100 like a doll, a toy horse, a snake, or a glove [Agin]. Programs can
07200 segment these objects into parts like legs, arms or fingers, and link
07300 these parts together to form descriptions of entire objects
07400 [Nevatia]. These operations are precursors to a recognition process.
07500 The work depends on a new representation of shape [Binford 1971] in
07600 terms of volume components and a part/whole segmentation. Primitive
07700 parts are described by an axis which is a space curve, and normal
07800 cross sections along the axis. The representation is based on a
07900 formalism, "generalized translational invariance". There are
08000 significant theoretical reasons why this representation leads to a
08100 natural choice of parts by simple procedures, in distinction to other
08200 suggested representations. We developed a laser ranging device
08300 [Binford 1970,Agin] which was operational two years ago. It is
08400 particularly interesting that the device adds a cost of only about
08500 $400 to the cost of a visual system.
08600
08700 A study of visual and depth ranging sensors has been carried out
08800 which provides a background for choosing sensors and interface, and
08900 making cost estimates [Binford 1972, Earnest 1967]. We have
09000 maintained a calibrated hand and vision system for several years
09100 [Sobel 1970], and have developed self-calibration techniques with one
09200 and two eyes [Yakimovsky unpublished, Hannah 1972, Gill 1972].
09300
09400 A system for understanding scenes of blocks from single views has
09500 been implemented [Falk 1970]; it segments objects, matches against
09600 internal models, predicts appearance from its assumptions, and
09700 searches for verification and contradiction based on the predicted
09800 image. Support and spatial relations are also obtained. There has
09900 been past work on programs to obtain line drawings as input to
10000 recognition systems. An earlier system was based on limited context
10100 of expecting continuation of edges at vertices and predicting missing
10200 edges in certain expected configurations [Grape 1970]. A more recent
10300 approach has been to use models within the process of generating line
10400 drawings, so that the most efficient next guess can be made and
10500 context used intrinsically in the formation of line drawings [Grape
10600 1973].
10700
10800 VISION FOR ASSEMBLY
10900
11000 This is a key part of the proposed project. We propose to evaluate
11100 and implement visual facilities required by a robot with simple
11200 vision, in conjunction with force and touch. These facilities must be
11300 modularized by working out several examples, and made fast enough and
11400 small enough to be practical for a minicomputer-based robot. We have
11500 considerable experience with such standardization, and with visual
11600 feedback techniques in the domain of blocks. We are now at work on
11700 extending visual feedback techniques to typical assembly tasks. We
11800 are also evaluating methods of automating visual feedback. We have
11900 well-developed systems for modeling objects from the console and to
12000 predict their appearance using techniques from computer graphics
12100 [Baumgart 1972a,b]. We can predict curved edges from our models. We
12200 will use simple region analysis programs and edge operators to locate
12300 curved edges predicted within a small window. Automated visual
12400 feedback would combine with evolving strategy programs in an
12500 interactive system. We will maintain a calibrated system and improve
12600 calibration updating so that reliability is maintained at a high
12700 level.
12800
12900 We have developed a special case program for locating the pump body
13000 in the pump assembly and moving it to the standard location. We are
13100 at work on several other such tasks involved in the pump assembly
13200 which contain components of general utility for visual feedback.
13300 These tasks are typical of the simple visual operations which we will
13400 implement or adapt.
13500
13600 VISION RESEARCH
13700
13800 While continuing research activities are not to be supported by the
13900 proposed project, in all likelihood, the research will suggest useful
14000 programs for the assembly system, particularly work in recognition of
14100 complex objects using depth data. The computer now has symbolic
14200 internal models of objects. We are now in a position to recognize
14300 objects such as a glove, a hammer, etc. We will immediately deal
14400 with matching descriptions for recognition. The important extensions
14500 we wish to deal with are occlusion and access to a large visual
14600 memory. These will involve matching of partial descriptions and
14700 encoding of partial descriptions as indices to a small class of
14800 similar shapes, which can then be compared in more detail. We are
14900 working toward making the hardware stable enough that it can maintain
15000 calibration, and software streamlined so that the depth measurement
15100 system can be integrated with other hand/eye facilities. We intend
15200 to apply large parts of our programs for description and recognition
15300 of curved objects using only monocular intensity information. This is
15400 a much harder problem, but within our representation, shares much of
15500 the same code, and relies on the same conceptual basis of local
15600 three-dimensional models.
15700
15800
15900 There is a growing effort on navigation of vehicles in outdoor
16000 scenes. While this is not directly relevant to the industrial
16100 environment, it does reflect the growing involvement in complex
16200 visual tasks, and will lead to further development of an integrated
16300 visual system. One approach which has already given initial results
16400 [Yakimovsky] is to combine region growing with interpretation using a
16500 world model. Thus, regions are combined not simply on the basis of
16600 uniformity but because combining them improves the interpretation. At
16700 the same time, we will be combining more facilities into these visual
16800 systems, and more usable knowledge in the world model.
16900
17000 C. MANIPULATOR DEVELOPMENT
17100
17200 Manipulator research began at the Stanford Artificial Intelligence
17300 Laboratory in 1965 with the acquisition of a Rancho Orthotic Arm.
17400 This arm was interfaced in such a way that the computer could read
17500 the joint positions by means of potentiometers and drive the arm by
17600 means of a voltage, pulse rate modulated signal for each joint. Two
17700 solution programs were written: one to give the rectangular
17800 coordinate position of the hand knowing the six joint angles, and one
17900 to give the joint angles knowing the desired x,y,z position and
18000 orientation of the hand. These programs were then combined with a
18100 servo program to move the hand.
18200
18300 The servo program represents two key concepts of all subsequent
18400 manipulator research at this laboratory: 1) To provide a digital
18500 computer with all sensory inputs, initially joint positions, but
18600 later velocity, force and touch measurements. 2) The utilization of a
18700 programable digital computer to take the sensory inputs and to
18800 provide basic control output to the manipulator, in our case joint
18900 torque. Initially the sensory input of joint angles was to provide a
19000 position servo with a sampling rate of 60 cycles per second. The arm
19100 control has been changed many times since, a simple task in the case
19200 of a program, but difficult in the case of a hard wire servo system.
19300 This system was used by Wichman, Pingle and Singer in 1968 to provide
19400 a program which could pick-up, sort and stack blocks, and in a first
19500 attempt at visual feedback.
19600
19700 At this time Piper considered the kinematics of manipulators and
19800 developed a classification scheme for all possible arms and their
19900 analytic solvability. He concluded his work by implementing a
20000 trajectory based collision avoiding program. Although slow, this work
20100 demonstrated the need for trajectory control of the arm.
20200
20300 Kahn followed Piper with a study of the dynamics of manipulators and
20400 developed a "bang-bang" servo for a new hydraulic arm. This arm,
20500 still at the laboratory, was developed as a high performance
20600 manipulator.
20700
20800 By 1970 a general purpose program capable of manipulating plane faced
20900 bodies was implemented. This program was used in the initial version
21000 of the "Instant Insanity" program where all possible reorientations
21100 of the four colored cubes were required in the puzzle solution.
21200
21300 Our present arm, the result of a continuing design project, replaced
21400 the Rancho Arm in 1970. This arm, like its predecessor is controlled
21500 by a digital servo program running under the time sharing system. The
21600 new arm was used in the "Instant Insanity" demonstration in 1971.
21700
21800 As this arm is driven by electric motors through high efficiency gear
21900 reductions it is possible to predict the torque exerted at each joint
22000 in terms of drive level. Based on kinematic analysis it is possible
22100 to calculate the joint torques necessary to produce a given force at
22200 the hand. Further, for the hand to exhibit degrees of freedom of
22300 motion, various joints are selected to be "free," that is, to have no
22400 position feedback applied to them. By exerting a force at the hand
22500 and providing that certain degrees of freedom exist it is possible to
22600 perform a much wider class of tasks than simple positional control,
22700 for instance screwing in screws, mating parts, inserting tools and
22800 turning cranks.
22900
23000 Positional accuracy at the hand is typically 0.1 inches with
23100 repeatability of 0.02 inches. Force control at the hand is within 10
23200 ounces. The data rate to run the arm from an assembled arm program is
23300 of the order of 100 bits/second.
23400
23500 Touch sensors on the fingers allow the hand to sense objects without
23600 disturbing them. When activated the arm will stop when any of the
23700 sensors touch. Branching arm instructions can test the state of the
23800 sensors and modify the execution of an arm program. When screwing a
23900 nut onto a bolt, for instance, the hand is first centered over the
24000 bolt by touching it.
24100
24200 A second arm of the same type, with only minor modifications, is
24300 currently being interfaced to the computer. Both arms will share a
24400 common work space and will be able to be used individually, or
24500 together for two-handed tasks. This should increase the scope of our
24600 manipulation work by some orders of magnitude. It may be possible for
24700 many assembly workers to work with their eyes closed but it would not
24800 be possible, in most cases, to work with only one arm.
24900
25000 Control of the arms is achieved by a series of high level
25100 instructions; these instructions may be executed directly, or may be
25200 assembled together into an arm program. Arm programs may then be
25300 executed, as desired, by the control program. The control
25400 instructions may be accessed directly by other programs, using the
25500 message procedure mechanism, or in the form of a symbolic source
25600 file.
25700
25800 Manipulator research moves ahead simultaneously on three fronts:
25900 control, primitives and devices. Tasks are translated into an
26000 existing hand language and where the task cannot be accomplished in
26100 terms of existing primitives, new primitives are formulated in a
26200 consistent manner. When a new primitive is formulated, real time
26300 routines are developed to perform the action. This also leads, in
26400 many cases, to the development of new hardware. Also as new
26500 developments occur in hardware design, existing real time routines
26600 are either re-written or modified.
26700
26800 Research is now concentrating on the following areas: To provide
26900 strain gauge instrumentation of the wrist, in order to allow for more
27000 sensitive interaction of the hand with a task. The development of
27100 high resolution, high sensitivity touch sensors, to increase tactile
27200 recognition and control abilities. These developments will lead in
27300 turn to new control concepts.
27400
27500 Development of tools for the hand. These tools can be instrumented
27600 such that the computer can feel and or see at the end of the tool.
27700 This should be a very important area of research as it goes beyond
27800 present human capabilities.
27900
28000 Two handed arm control, the planning and execution of tasks requiring
28100 the coordinated motion of both arms, will require an extension of the
28200 existing servo program and the analysis of synchronization problems.
28300 The development of collision avoiders, so that arms will not collide
28400 with other objects and with each other.
28500
28600 Modification of the existing arm design to provide for better
28700 incremental motions, more degrees of freedom and a more rugged design
28800 suitable for industrial use. Development of arm and eye computer
28900 interfaces. These contain synchronized analog/digital converters for
29000 input and current controlled drivers for output.
29100
29200 The development of new hands is expected as more complex tasks are
29300 attempted. At present, with only one hand, we have been performing
29400 the positioning type of hand actions. When the second arm is
29500 operational we will be able to perform holding and pulling actions as
29600 well.
29700
29800
29900 D. SYSTEM REQUIREMENTS FOR ADVANCED AUTOMATION
30000
30100 A key element in the development of any large programming effort is
30200 the software support. There has been continual effort along these
30300 lines at the Stanford Artificial Intelligence Laboratory and many
30400 useful ideas have been developed. We propose to extend the research
30500 along two separate lines: support for research and system software
30600 for operating automated systems. We will briefly describe our past
30700 efforts in this area and what we propose to undertake.
30800
30900 The Stanford Artificial Intelligence Laboratory has pioneered in
31000 several areas of system programming, including timesharing systems,
31100 interactive graphics, and programming languages.
31200
31300 A great deal of additional effort has gone into direct support of
31400 hand-eye research. Very early, we established a set of standard
31500 formats and library routines [Pingle 72a]. A programming language,
31600 SAIL[Swinehart], was implemented in 1968. The language incorporates
31700 special features for large real-world-oriented problem solving
31800 systems and is currently in use at a number of Artificial
31900 Intelligence and other laboratories.
32000
32100 As we began to assemble larger and more sophisticated collections of
32200 programs, we felt a need for flexible mechanisms for cooperating
32300 sequential processes. The resulting monitor and language features
32400 [Feldman & Sproull 1971] have worked well for us and have been
32500 influential on other projects. Recently, we have added an
32600 additional set of features to SAIL to facilitate non-deterministic
32700 and event-directed programming[Feldman 1972]. The early reactions
32800 to this work, both within the laboratory and without have been very
32900 encouraging.
33000
33100 We expect relatively little work to be needed on further systems
33200 support for basic research. There will undoubtedly be efforts to
33300 take advantage of new hardware and of new programming ideas that may
33400 develop in the course of this program. Almost all of this work
33500 should be covered by other funding.
33600
33700 The development of software for the assembly station will require
33800 more of our effort. The central-local computer configuration
33900 requires code to be resident in the local computer and in the shared
34000 machine. The low level programs in the local machine will not be
34100 modified often, thus we can assume that a compiler resident in the
34200 central machine will suffice. The planning programs and general
34300 vision programs will be resident in the shared system. There are two
34400 possible configurations for the shared computer system: A PDP-10
34500 equivalent machine or a software package in ASA Standard Fortran IV
34600 which will run on any sufficiently large machine. One of the tasks
34700 under this proposal would be to evaluate these two alternatives.
34800
34900 A major share of the programming research effort will be concentrated
35000 on developing a programming system for automated assembly. This will
35100 combine ideas from our past work on languages, representation,
35200 computer graphics and arm control. We view this task as a branch of
35300 the larger effort on automatic programming being undertaken in this
35400 laboratory and elsewhere.
35500
00100 TIMETABLE
00200
00300
00400
00500
00600 The full program is designed for a four year period, including the
00700 initial one year program proposed here. We will set up the timetable
00800 only for the one year grant period. During that period we will
00900 accomplish the following:
01000
01100 interface the second arm and develop two arm software;
01200
01300 transfer arm software and hardware to the PDP-11;
01400
01500 interface improved touch and force feedback for the arms;
01600
01700 extend our library of vision primitives and perform several
01800 visual feedback experiments to test their utility;
01900
02000 begin the vision language;
02100
02200 perform an assembly task (possibly assembly of a
02300 two stroke engine) to aid in development of our system;
02400
02500
02600
02700 For the following timetable, we assume the project to begin in September,
02800 1973.
02900
03000
03100 Month
03200 1 begin transferring arm hardware to PDP-11; begin development of new
03300 visual feedback modules;
03400 2 Second arm operational; begin on two arm software
03500 3
03600 4 use of visual feedback for inserting screws into holes
03700 5
03800 6 force sensing wrist operational; two arm software initially usable
03900 7
04000 8 system software for minicomputer usable;
04100 begin combining vision with hand language
04200 9
04300 10 completion of transfer of arm control software to PDP-11
04400 11 completion of demonstration of assembly task
04500 12 improved touch sensing
00100 III. UTILIZATION PLAN-DISSEMINATION OF RESULTS AND RELATION WITH
00200 INDUSTRY
00300
00400
00500 There are several other efforts, at varying levels of sophistication,
00600 in the area of computer controlled advanced automation. The Stanford
00700 Artificial Intelligence Laboratory is the oldest and largest of these
00800 and has provided the cutting edge of much of he development. We
00900 intend to continue this role and to lead in the development of a
01000 practical general purpose assembly system and the associated
01100 software.
01200
01300 All of the discoveries made under this grant will, of course, be made
01400 public. We view our role in information transfer in a much broader
01500 way. Having no desire to sell products and services and no clients
01600 with interests to protect, a university laboratory should be
01700 primarily a place for acquiring and spreading knowledge.
01800
01900 The Stanford Artificial Intelligence Laboratory has always considered
02000 information transfer to be one of its primary goals. In addition, to
02100 our published papers and reports, we have made a number of film
02200 reports which have been extensively circulated. There have been
02300 thousands of visitors, ranging from elementary school classes to
02400 major scientific, governmental and industrial missions from many
02500 countries. We have also encouraged extended visits by workers from
02600 other laboratories and from industry. Members of the laboratory staff
02700 have played a major role in the organization of journals, conferences
02800 and seminars in artificial intelligence. Many of our algorithms,
02900 systems, and hardware designs have been adopted by other groups. Our
03000 students are an important contribution to information transfer.
03100
03200 We have close contacts with SRI (several former students are employed
03300 there) and with Draper Lab, MIT. These laboratories, with industrial
03400 participants in their research programs, provide the right setting
03500 for transfer to industry. We have close contact with the computer
03600 science group at GM Research Labs.
03700
03800 A development of consierable interest is that five groups doing
03900 similar research have chosen the PDP-11 as a research minicomputer.
04000 We will have immediate relevance to members of this informal robot
04100 user's group. Our programs will be usable by members and any others
04200 who acquire the PDP-11. It would be advantageous for future
04300 purchases of minicomputers for automation research, that they be
04400 compatible with this group. Our efforts are exportable in a further
04500 sense; the arm servo and planning programs are applicable to a wide
04600 range of arms, to the Unimate in particular.
04700
04800 The success of the program proposed here depends heavily on
04900 interaction with industry. We will need additional information on
05000 the opportunities for automation and will try to involve manfacturers
05100 in the development of the pilot assembly station. We have good
05200 working relationships with a number of firms and we expect that a
05300 concrete project will help forge the bonds necessary for the transfer
05400 of advanced automation technology from the laboratory. During the
05500 initial proposed grant period, we intend to negotiate a cooperative
05600 relation with one or more firms.
05700
05800 One promising approach to technology transfer is a staff associates
05900 program. This involves residence of about a year's duration at the
06000 laboratory by industrial scientists. Such a program was used
06100 successfully at the MIT, Electronic's Research Laboratory in the area
06200 of programming systems for numerically controlled tools and design
06300 automation. We have initiated a similar program in the laboratory
06400 (cf Appendix A) and are pursuing additional forms of cooperation.
00100 Appendix A
00200
00300 The Stanford A.I. Lab has been engaged in research on
00400 advanced automation technology. Our work has progressed to the stage
00500 where we are seriously engaged in developing and disseminating
00600 practical applications. To do this effectively, we will need the
00700 help of industrial organizations with computer sophistication and an
00800 understanding of practical problems. One form of cooperation which
00900 we are engaged in is a staff associates program. A staff associate
01000 is a member of another organization in residence at the A.I. Lab. We
01100 have had a number of staff associates over the years, mostly from
01200 foreign countries. The enclosed letter from Japan is typical of the
01300 inquiries we receive. The purpose of the new program is to invite a
01400 few domestic inductrial organizations to submit candidates for the
01500 staff associates positions.
01600
01700 We expect to be able to have three to five staff associates
01800 in the laboratory at a time. The associate would remain for about a
01900 year and would be paid by his sponsoring organization. He would be a
02000 regular member of the laboratory staff, working on automation
02100 problems.
02200
02300 There are, of course, academic programs at Stanford which
02400 could also involve work at the Artificial Intelligence Lab. The
02500 staff associate program differs in that the selection criteria are
02600 different, the timing more flexible, and the expected time devoted to
02700 course work much less. We believe that it is the most effective
02800 mechanism for us to learn from and help industrial practitioners of
02900 computer-aided advanced automation.
03000
03100
03200 Jerome A. Feldman