The Effect of the Product Cost Factor on Error by rwi74592


									        The Effect of the Product Cost Factor on Error Handling in
                           Industrial Robotics
                                           N.W. Hardy* and M.H. Lee
                                          Centre for Intelligent   Systems
                                         Department of Computer Science
                                                University of Wales
                       Abstract                                   The capability of dynamicreconfiguration allowing the
                                                                  changeover from assembling one product variant to an-
   The approachesto error handling in two contrasting in-         other within one cycle time.
   dustrial robot case studies are presented and discussed.
                                                               ¯ Reliability
                                                                  Lowequipmentfault rates.
                        Introduction                              Theability to deal with faults.
Weare interested in the role of advancedsoftware tools and     , Meeting user requirements
techniques in achieving robust operation of robotic systems       A safe and easy to operate user interface which accommo-
in industrial and commercial  applications. In this context,      dates existing industrial skill levels.
errors which halt production are costly and systems must          Ease of maintenance.
be carefully designed to cope with any shortcomingsduring         Animportant aspect to note is that the machinewouldbe
productionin waysthat are cost effective for the particular    highly integrated, covering all aspects of the assemblypro-
application.                                                   cess and that a reasonably high level of task programming
   Wepresent two case studies from industry to illustrate our  wouldbe required to allow non-specialist programmers       to
approach to error handling. The case studies showhowpro-       handle the diversity and numberof devices in the machine
duction constraints can influence the priorities for solution  and deliver robust performance. Coupledwith this, i) eco-
of given robot manipulation problems.                          nomicviability in the mass assemblymarket is crucial and
                                                               ii) errors mustbe dealt with (not necessarily corrected).
          The InFACT Assembly Machine                          implication of these two points is that throughput rates of
The project                                                    correctly assembled products wouldtend to be more impor-
The InFACT   project ran from 1989 until 1991 with a short     tant than the proportion of failed assemblies. The trade-off
feasibility study during 1988. It had a budget of £6.8M        betweenrate of production and rate of failure wouldclearly
and was funded under the multinational EUREKA/FAMOS be influenced by manyfactors but with reasonable failure
programas Project 321. Commercialand academic collab-          rates and the typical cost of parts used in the intendedprod-
orators from the UK,France, Italy and Austria were in the      uct range, 100%success is unlikely to be an appropriate
consortium [Loughlin, 1992]. Commercialexploitation of         target.
many  aspects of the project is nowunder way.                  The Hardware Configuration
   The aim of the project was to develop a flexible assem-
bly machineincorporating sensor systems, parts and com-        The machine has three zones. The central or main zone
ponents feeders, manipulators and assembly devices. The        houses a four degrees of freedomcartesian manipulator de-
machinewas to be suitable for the assemblyof a wide range      signed for rapid, light and precise operations. The power
of electo-mechanicalproducts of up to 300ram   cubedin size.   operations zone houses a three degrees of freedom manip-
The design criteria laid down the outset characterise the
                               at                              ulator designed for operations requiring high forces or the
machineand are as follows [InFACT      Group, 1992].           use of specialist tools. The supplyzone houses a specially
                                                               designed modular system allowing a number of stacks of
¯ Economic    viability                                        pallets. This provides the entry point for parts pallets and
   The ability to performthe basic tasks associated with the   the exit point for finished products and rejected parts and
   assemblyof products at a cost which is competitive with     assemblies. To link the three zones a linear shuttle device
   existing labour rates.                                      runs beneath them. The shuttle is able to carry two pallets,
¯ Flexibility                                                  raise theminto positions in the floors of the twomanipulator
   A high degree of reusable hardwareand software.             zones and later retrieve them. The main zone has five such
   Easy reconfiguration for different manufacturing   require- positions and the power operations zone two. The shuttle
   ments.                                                      can acquire and replace pallets at the base of stacks.

   Upto 20 linear vibratory feeders with exchangeable    tracks     standard circumstances, intended to overcomethe problems
can be sited around the two manipulator zones. These pro-           which wouldotherwise arise. Broadly three levels of such
vide an attractive alternative to pallet feeding for suitable       activity were recognised.
small parts.                                                       1. Halting the machinewith goodquality information for the
   The end of arm tooling design for the main manipulator              operator.
incorporated exchangeable tools and programmable       compli-
                                                                   2. Abandonment the current assembly but continuation of
ance with six degrees of freedom. [Selke, 1992] The demon-
                                                                       production without operator intervention.
stration machineused a device with a subset of the original
facilities. End of arm tooling on the poweroperations ma-          3. Salvaging of the assemblyin question.
nipulator again benefits from a standard physical and control          Level 1 was clearly undesirable within the InFACT  objec-
interface to the machine.                                           tives, thoughhigh quality diagnostics in an assembly context
   Many  sensors are provided,both for local control of partic-     were seen as valuable if this eventuality arose. Designpro-
ular actuators and to make sensor data are available to the         gressedon the basis of levels 2 and 3.
assembly task programmer.The only standard devices pro-                Six classes of technique to improveresilience were con-
vided for measurement    operations are simple binary touch         sidered (it wasrecognisedthat the classes were not all inde-
probes [Bell, 1992].                                                pendent).
   The design of the machine developed hand-in-hand with
a particular policy for internal scheduling. Assembly     takes    1. Correction for deviation from nominal positions The
place on specialised pallets knownas fixtures. Assembly            variability in positioning and dimensionsof fed parts must
will be proceedingin parallel on three fixtures. Onewill be        be handled. Three approacheswere considered. In the light
located in the poweroperations zone and one in the main            of experience with earlier machinesit was considered that
manipulatorzone. The third will be in the process of being         active compliance would not be a necessary or effective
repositionedin order that it be ready for the mainmanipulator      technique, A clear specification of the required passive
to move to it and not be idle while transportation takes           compliance capabilities was established. The type and level
                                                                   of measurement correction techniques was a matter
place. The most cost-effective operation of the machinewill
be whena numberof assemblies are producedon each fixture           for investigation during the project.
and the principle of addingone part to all the assemblieson        2, Explicitly programmed    user specified conditions and
one fixture before continuing to another part was adopted.         responses This technique, whichreflected the standard in-
                                                                   dustrial approachwasnot consideredappropriate. It requires
Task programming                                                   conditional constructs within the task programming  environ-
The machinedesign criteria concerning the ease and robust-         mentand typically requires significant analysis, design and
ness of task programming production staff implied that                                                 It
                                                                   codingeffort by the task programmer. fails to capitalise on
an alternative to a language based task programming      envi-     available assemblyexpertise for the benefit of future assem-
ronment was required. Using on experience gained in ear-           blies and clashed with the fundamental InFACT    philosophy
lier projects [Barnes et al., 1983, Hardyand Barnes, 1992,         of well engineered high level programming  primitives.
Orgili et al., 1989, Williams,1985, Williamsand Lill, 1987]
it was based on so called generic tasks. These can be con-         3. Gooddesign The mechanical and control components
sidered as highly parameterised subroutines embodying      well    of the machinewere designed from the outset with a clear
designed and tested software, assembly expertise and best          set of functional requirementslinked closely to the applica-
practice for exploiting the design and configuration of the        tions. It was anticipated that design for assemblywouldbe
assemblymachine. Further and highly effective assistance           employedwherever possible.
can be given through the provision of an inventory. This              The design of activity within the cell must be good. This
                                                                   was seen at two levels. First, the individual generic tasks
is a database containing details on all elements of the ma-
chine the tools the componentsand the product. Data are            should be well designed. Second, sequences of these tasks
added to the inventory in a systematic wayand maybe used           need to be well designed to workat least in cooperation and
repeatedly to provide parametersfor generic tasks.                 preferably synergistically. This was expected to be achieved
   In addition to a coherent database, an assembly   specifica-    through guiding and constraining the user.
tion consists of a sequenceof objects. Theseare instances of       4. Pre-programmed    generic responses A natural exten-
parts known the database and have associated sequences             sion to a systembased on a set of generic tasks is that tasks
of generic tasks with appropriate parameter values. The job        should include generic responses associated with particular
of the task programmer to create a sequence of objects             conditions. This approachhad been taken in earlier workand
which represents the assemblysequence and then to select                                                         Two
                                                                   it was expected to play a part in InFACT. broad types
and provide parameters for a sequence of task generics to          of response may be considered. Hidden responses which
specify the actions to be carried out with the part.               are broadlyimplicit in the functional definition of the action
                                                                   (such as retries and searches) and more apparent responses
Error handling - initial considerations                            wherethe effect of the action is altered (such as discard-
Fromthe earliest stages of the project the term resilience         ing suspected faulty componentsand obtaining newones or
was used. This covered behaviour of the machine, in non-           re-grasping).

out of tolerance, i.e. Pr > dr +t~ orpr < dr -tr should be         any error treatment is boundto be higher than the cost of the
rejected. Other product mayclosely matchthe pattern, i.e.          productthere is no justification for attemptingto process any
Pr = dr and should be markedfor acquisition, while a third         suspect product. Consequentlythe sole mechanism     available
class will fall within the tolerance band, i.e. p~ < dr +tx and    for dealingwitherrors is rejection. In other words,error rates
p,~ > dr - tr. These marginal cases will provide material          mustbe kept low, even at the expenseof the reject rate, i.e
for learning the scope of the grasping function mthey will         ’reject on any hint of error’. It is usually foundthat there
be processedfor acquisition but the results will provide in-       is a trade off between errors and rejects, as high quality can
formation that can be used to update the tolerance function.       often be gained at the expenseof severe rejection criteria.
Wenotice a fourth possibility: any miss-feed of different          In this application Epaekmust be minimisedby using Rproa
product, such as might occur in mixedbatches, maycause             and P~ackto prevent errors reaching the output.
product of type P~ to be recognisedduring a run of Px. This           Returning to the above four classes of errors, suitable
will be treated as a reject case and logged.                       recoveryactions are: for (a) ignore product (it will pass
                                                                   to the end of the belt) -- this increases Rprod;for (b) open
Acquisition (Grasping)                                             gripper, ignore and try next product -- this also increases
The present end-effectors consist of grippers with fixed-                 for
                                                                   Rproa; (c) and (d) send signal to reject the current packing
throw, two-state pneumatic fingers. This limits the out-           and start a newone mthis increases -Rpa,~.
comeof grasping operations to three possibilities. Either             From this analysis the following factors emerge as key
the grasp is successful (product position being under full         design parametersfor any proposederror handling processes:
control), grasping fails during retraction from the conveyor       Reject Propensity As the product has almost negligible
surface, or the product position is lost during transport or         value in its prepackagedstate it can be rejected at the
delivery. The last failure is much   moreserious than failing        slightest hint of an error. Withhigh volumethere will be
to pick from the conveyorbelt. Loss of positional control            a non-trivial proportionof faulty product that is presented
maymeaneither complete product loss (dropping the prod-              for handling. It is essential that this is rejected without
uct fromthe gripper) or errors in location and/or orientation        engaging handling; recovery is a high risk operation and
(partial gripping, twisted product, etc.)                            wouldbe extremely costly. Fortunately detection and re-
                                                                     jection of faulty product in the feed is easily accomplished
Error Handling                                                       by the vision sensor and conveyorcombination.
Wecan nowexaminethe scope for error detection and re-              Effective Grasping Oncea decision to grasp the product
covery. There are four cases where faulty product maycause                                         it
                                                                     with the robot has been made is vital that the grasp is
errors: (a) the product is discoveredto be faulty before ac-         optimised and is as effective as possible. Someform of
quisition is attempted,(b) the product is not picked correctly       additional sensing is necessaryif grasping errors are to be
from the conveyor,(c) the product is lost during transport,          detected.
and (d) the product is lost during packing. Wenotice that
only case (a) can be detected without additional sensors.          Product Variation Despite the controls inherent in mass
                                                                     production someuncertainty in product shape and size can
   The application factors which determine the form of ap-           still be manifest. Whilelarge variations will be treated as
propriate error treatmentare:
                                                                     rejects, smaller changes should be accommodated.     Grasp-
Lowproduct cost The product has negligible cost. Any                 ing strategies must be able to respondto local variations
  damaged, misshapen, oversized or undersized product                and longer term trends in product dimensions.
  should be rejected. Wedefine the incomingproduct rate
  as Fvroaand the reject rate as Rvrod.                            Learning and Adaptation
High production rates Comparedwith manufacturing ap-               There are two forms of learning that are appropriate to the
  plications, food processingbatch sizes are very large. Ben-      task: continuousadaptation to small variation in product di-
  efits are only gained for low cost product whenhugequan-         mensionsso as to optimise the success of grasping, and the
  tities are processed. Thus, it is essential for maintaining                                          of
                                                                   utilisation of previous knowledge gripper/product combi-
  viability and competitiveness that very high production          nations so as to find effective grasp sites for newproducts.
  rates are achieved. This meanserrors must be handled at          Weare attempting to incorporate both forms of learning.
  similarly rapid rates.                                                                                     as
                                                                     Weare using a learning methodknown the ’Collateral
                                                                   Architecture’ to fine tune the grasping strategy for a given
Lowerror rate in output packages Even when packed the                                        of
                                                                   product; this is a method adjusting the tolerance function,
  value of the product is still small. However, faulty             t~, to capture the best fit from the cumulative experience
  packingcould incur serious penalties if the productreaches       gained during production runs. Learning methodsare also
  the customer. It is important that no packed product             being used to select grasping parametersfor different prod-
  emerges which has missing items, bad orientation, dam-           uct types. Different grasping strategies are selected by this
  aged items, etc. Hence the error rate of the outgoing                                               to
                                                                   process and learning is employed capture and utilise ex-
  product, Epaekmust be minimised.
                                                                   perience of previous grasp performance. By this means a
  Fromthe above it is seen that there is no scope for any          decision surface, dx is selected for a given product type, P~,
significant processingduring error handling. As the cost of        using the database of gripper/product experience.

ics estimate corrections for this. All generics used after         Equipment Configuration
a measure automatically take the correction into account.          An Adept robot is stationed over a conveyor belt which
[Nicholls and Hardy, 1992]                                         supplies individual frozen product pieces. A vision system
The expectation that well designed application specific            is placed upstream of the robot and uses standard vision
jigs could be providedat acceptable cost. The provision            algorithms to recognise features of the product shape. The
of simple passive jigs is considered an acceptable cost in         robot then uses the coordinates of the product and the belt
the context of the development of a new product. If an             position to grasp the product and transport it to a packing
active jig or a jig incorporating sensing is required, this        location. Normallyseveral product are packedinto one box.
could significantly increase the cost. To reduce this a simple
general purposeinterface is providedat each fixture position.      Implicit Constraints
Appropriate and programmable      compliance in the grip-          The configuration of the work-cell is essentially fixed and
per mechanism. The compliance devices designed and                 all sensing and error recovery is to be carried out in the
used in InFACT described above.                                    context of existing equipment. The viewing angle of the
General purpose sensing to detect incorrect assemblies             camerarestricts vision to two dimensionaloverhead images,
provided as part of the task programming       environment.        consequently,in all that follows, wetreat productas laminae,
A numberof permanent monitors were provided to detect              i.e. fiat plates of uniform thickness with variable shape
general problemssuch as major collisions. Beyond    that each      outline.
generic task explicitly uses appropriate sensors to check as
far as possible that its intended purpose has been achieved.       Operational Requirements
The types of checks performed are those which would be
                                                                   The normal modeof operation of the handling cell requires
expectedfor standard current industrial applications.
                                                                   high speed production and rapid changes of product batch,
The abandonment an assembly where an error is de-
                                                                   with fast gripper changes and minimum      setup time. The
tected and its removal using a parameterised routine.
                                                                   controls and user interface are to be designed for absolutely
Any mis-match between a sensed value and the expected
                                                                   minimal operator involvement.
value causes the generic task to be abandoned the asso-
ciated assemblyto be markedas such. A record of the stage          The Robot Handling Task
at which the assembly was abandonedis also made.
                                                                   Thereare a range of different product types, PI...Pn, each of
   The final stage in an assembly involves the transfer of         whichcan be handled by one of the available gripping tools,
completedassemblies from the fixture (which will begin the         GI...G~, wherek _< n. Eachbatch run uses only one gripper
cycle again) to a pallet which will be taken to a product-         and deals with a single product type. Gripper changes are
out stack. At this stage any abandonedassemblies must be           madeat setup time.
disposed of. To achieve this, each object in the assembly            Analysisof the task identifies 7 stages:
sequence mayhave a disposal generic associated with it, in
addition to the generics specifying its normalmanipulation.        FEED  Present product to cell. This is achieved by the
The disposal generic is a simple pick and place task speci-          conveyor.
fying how pick up that object and transfer it to a disposal                The
                                                                   LOCATE vision system gains information on product
pallet. Thispallet is likely to be little morethan a box, so ac-     position, orientation, size and shape.
curate placement not required. Eachobject in the sequence
                                                                   ACQUIRE positional control over product. This is
must have a disposal sequence defined. This is essentially
                                                                     achieved by successful grasping.
a list of references to those objects whosedisposal generics
must be executedto removeall the parts fromthe fixture.            ORIENTAlign product with given reference frame.
                                                                     Achieved by Robot action.
   This approach appears to place a considerable burden on
the task programmer.This burden is not unacceptable heavy                   Move
                                                                   TRANSPORT product to spatial location.             Achieved
for two reasons. First, the parameterdata required are simple       by Robot action.
and probably already in use. Second, very rarely will it           DELIVERMate or place product at target           location.
be necessary to movean abandonedassembly part by part.              Achievedby Robot action.
Many  products have a base or enclosure which means that
most parts can be removed grasping it.                             INSPECT  Measure product/task parameters. None speci-
                                                                     fied other than rejection of faulty product.
                                                                     Of these operations only LOCATE ACQUIREand           pro-
                                                                   vide opportunities for sensing and correcting errors. Let us
               Robotic     Food Handling                           examinethe functions of these two stages.

This problemoriginates in the requirements of a food man-          Location (Vision)
ufacturer to automatethe packing of frozen food products.          Each product type P~ will have an associated decision sur-
It is being investigated under a project supported by the UK       face, d~ in the space of measuredparameters. Let there be
Science and Engineering Research Council.                          an associated tolerance function, t~. Then product that are

5. Flexible pre-programmed     responses The generic re-          unacceptablyhigh in manycases. The target time (achieved
sponse approach maybe enhanced by the use of contextual           for a range of products) for the inclusion of one part in
data to select and parameterisegenerics leading to less rigid     single assembly was 3s. Anylengthening of this begins to
and more relevant responses. This method follows natu-            challenge the economic   viability of the machine. A doubling
rally from the concept of tasks generics and the inventory        of the time (as might be caused by a completeand successful
database. Suitably designed responses could be selected and                                   of
                                                                  retry) for a small number parts wouldnot have a signifi-
executed to give a significant range of behaviour and this        cant overall effect on throughputwhenspread across all the
approach wouldbe investigated and exploited in InFACT.                                                       is
                                                                  parts in all the assemblies. This however not the real time
6. On-line recovery planning Techniques to plan cor-              cost of corrective actions whichis related to the schedul-
rective actions were considered. The cost of such planning        ing of the whole machine. High levels of activity can be
wouldbe important. From early stage it was proposedthat
                         an                                       achieved within the cell by relatively simple fixed sequenc-
                                                                  ing of the activity of the manipulatorsand shuttle. Inves-
such techniques should not be used in InFACT three rea-           tigations revealed that deviation from that schedule would
sons. First, they are relatively expensiveto implement  both
at the systems level and for the task programmer must
                                                   who            typically result in a large drop in throughput.Ifa pans pallet
                                                                  becomesempty during work on a fixture, the arm will be
assemble a large knowledgebase to support them. Second,
                                                                  inactive while a newpallet is introduced. A policy of car-
while certain aspects of the problemcould be tackled in this
                                                                  rying on parts pallets an integer multiple of the number     of
way, others could not and a significant research component
                                                                  parts required by a fixture was used and any strategy which
was implied in the approach.Mostcrucially, the value to In-
FACT terms of enhancedperformancewould be relatively
     in                                                           exhausted a pallet prematurely would disrupt the schedule
                                                                  by delaying the arm. Secondary   disruption is likely to result
                                                                  from the use of the shuttle for replenishment and from any
                                                                  attempt to use the armfor workon another fixture while more
Error handling, the development of understanding                  pans are fetched.
The process of design of the final InFACTmachine was                 Thereare twoalternatives to the abovesituation. First all
complexone and involved the resolution of a large set of          pans pallets could carry a small excess calculated to handle
often contradictory pressures. Tracingin detail the steps by      the expectedworst error rate for that part. This approachwas
whichinitial ideas and understanding were transformedinto         not exploredin detail but was expectedto lead to a high rate
the final decisions is a major task and beyondthe scope of        of pans recycling and a loss of space within the machine.
this paper. The following trends can howeverbe recognised         The second would be a dynamic scheduling regime. This
in retrospect.                                                    could not be investigated in the scope of the project and its
   The ease and economicviability of providing physical           potential for the machineas configured is not known.It
restraint within the machinecomparedto the problems and           seemsunlikely that such a system could achieve utilisation
cost of handlingpositional variation by means sensing the
                                                of                and throughputrates as high as those achieved by the fixed
software techniques were increasingly recognised.                 scheduling.
   The cost of rejection of a possibly faulty part or aban-
donmentof a failed assembly was taking increasingly into          Error handling - the final design
consideration. Many   componentsare cheap and in addition         The final InFACT  design included the following features as
such rejection has possible advantagesfor quality assurance.      a result of the aboveconsiderations.
Manualcorrection and completion of assemblies can be a
cost effective mechanism.                                         High but reasonable levels of constraint on the quality
   Therange of possible errors is related to the flexibility of   and positioning of fed parts. Twopans feeding methods
the machine. machinecapable of a wider range of activity
              A                                                   are provided. The linear vibratory feeders have replaceable
will clearly be opento a widerrange of errors.                    tracks which can be automatically designed and machined
   In establishing a range of activities of which InFACT          starting with basic designs for the parts.[Dick and Lo, 1989]
wouldbe capable there was a strong tendency to reduce and         Theyare able to reliably and repeatably present small parts
simplify. Many  proposedfacilities were recognised to be of       at a fixed location. The basic pallet can be enhancedby
use only in a small proportion of real assemblytasks. Their       simple and cheap means such as vacuumformed inserts to
omission wouldnot significantly reduce the utility of the         provide accurate pans specific presentation. Thesepart spe-
machinebut wouldsimplify implementation and reduce the            cific adaptations are consideredunacceptablecosts.
scopefor error.                                                   A careful characterisation of classes of remainingun-
   The probability of success of any corrective action will       certainties in feed positions andthe provision of sensory
obviously not be 100%.Whenthe possibilities of incor-             mechanismsto remove each class. The feeding mecha-
rect diagnosis of the problemand of having a fundamentally        nisms were recognised to leave positional uncertainties of a
insoluble problemat a given level of the assemblyare con-         limited and simple set of types. These were matched one
sidered success rates could be quite low. When       combined     to one by a set of generic measurement   tasks each designed
with what should be a low basic error rate the achievable         to removea particular uncertainty. All the tasks use the
increase in throughput maybe small.                               touch probes. Feed position errors result in deviations from
   The cost of corrective action was increasingly seen to be      nominal grasp configurations and the measurementgener-

   It is clear that detection of successfulgraspingis vital for    of the project and source of vital expertise and inspiration.
any treatment of those errors that affect Epaek. Withvision        Wealso acknowledgethe contributions of the team on the
as the only sensor it is not possible to detect any handling       food handling project:- Jem Rowland,MarcosRodrigues and
operation that fails. Hencethe tolerance bands must be held        YoufuLi.
very wide in order to reduce potential risk. This can lead to
an unacceptably  high reject rate.                                                         References
   By using a simple ’product present’ sensor in the grip-         [Barnesetal., 1983] D. P. Barnes, M. H. Lee, and N. W.
per it is possible to confirm correct gripping, detect faulty         Hardy. A control and monitoring system for multiple-
gripping and discover lost product during transport and deliv-        sensor industrial robots. In Proc. 3rd. Int. Conf. Robot
ery. Other more sensitive and sophisticated sensors provide           Vision and Sensory Controls, pages 471--479, Cambridge,
more information (such as twisted or skewed grasping of               MA.USA., 1983.
the product) but they also incur more expense, complexity
and computerprocessing power. These include: finger con-           [Bell, 1992] Steve Bell. Touchprobes: The vision argument
tact sensors, finger displacementdetectors and finger force           - sophistication vs simplicity. In Clive Loughlin, editor,
                                                                      InFACT."  Project, Concepts, Machine., pages 46--47. MCP
   It is also useful to consider the levels of actuation that         University Press Limited., Bradford, UK,1992.
are available for handling. Grasping and action strategies         [Dick and Lo, 1989] M.D. Dick and E.K. Lo. Computer
are constrained by the available motor functions. The ex-             aided design and manufacture of a feeding system for
isting gripper is a design basedon fixed-throw,fixed-centre,          robotic assembly. In Proc. Int. Conf. Manufacturing  Sci-
binary-actuation fingers. This gives three control variables:         ence and Technology, Sweden, 1989.
x, y, and 8, for defining the position and orientation of the      [Hardy and Barnes, 1992] N.W. Hardyand B.P. Barnes. Af-
target grasp site on the conveyor. The only strategies for            firm: Knowledgebased robot supervision. Technical
grasping consist of variations in approachby changingthese            Report UCW-CS-TR-RRG-200-92,       Department of Com-
values, (we note that re-grasping strategies mayhave some             puter Science, University College of Wales, Aberystwyth,
merit). Weare currently considering more dexterous grip-              UK,April 1992.
pers whichwill allow morescopefor flexible strategies. For
example, control over the initial aperture of the fingers would    [InFACTGroup, 1992] InFACTGroup. Infact. Integrated
allow more product ranges to be accommodated        within one        flexible assemblycell technology. Final report. Bristol
gripper, thus reducing the numberof gripper changes. Some             Polytechnic., ColdharbourLane, Bristol, UK., February
adaptation for product variation mightalso be facilitated, as         1992.
minoraperture adjustments can cause corresponding changes          [Loughlin, 1992] Clive Loughlin, editor. InFACT:   Project,
in gripping force. This wouldoffer four control variables:            Concepts, Machine., Bradford, UK, 1992. MCP      Univer-
x, y,0 and r, wherer is the initial opening radius for all            sity Press Limited. 59 pages.
fingers. Asthe throw,t, is fixed, all fingers close on a radius    [Nicholls and Hardy, 1992] H.R. Nicholls and N.W. Hardy.
of r - t. Anotherdesign involves linear-actuation fingers.            Distributed touch sensing. In H.R. Nicholls, editor, Ad-
In this variant the closing of the fingers is performed anby          vancedTactile Sensingfor Robotics, chapter 6, pages 107-
electric motoror other drive such that the fingers can be             122. WorldScientific, Singapore, 1992.
driven to any position betweenfully open and fully closed.
This gives five control variables: z, y, 0, r and t and much       [Orgill et al., 1989] C.H. Orgill, N.W.Hardy, M.H.Lee, and
moreopportunity for dexterous grasping strategies, owingto            K.A.I. Sharpe. Anapplication of a multiple agent system
the enhancedperformance and control of finger movement.               for flexible assemble tasks. In Knowledge  based envirn-
   Weare currently exploring the cost/benefit trade off be-           ments for industrial applications including cooperating
tweenreducederror rates and additional sensory-actuatorfa-            expert systems in control. IEE London,1989.
cilities. It has been shown  that the task requirements provide    [Selke, 1992] Klaus Selke. Endeffector. In Clive Loughlin,
constraints that may severely limit justifications for compli-        editor, InFACT:  Project, Concepts, Machine., pages 35-
cated error processing and the values of the various product          39. MCP  University Press Limited., Bradford, UK,1992.
cost factors are of great importance.In the present study we
                                                                   [Williams andLill, 1987] A.M. Williams and B.H. Lill.
haveargued that considerable potential benefit is offered by
                                                                      Commercially  available flexible assemblycell. In Proc.
the addition of somesimple sensing and learning procedures.           BRA-IO. Birmingham UK, 1987.
Wehope to report on experiments based on implementations
of these ideas in the near future.                                 [Williams, 1985] A.M.Williams. A flexible assemblyceil.
                                                                      In Proc. 6 th ICAA, BirminghamUK, 1985.
Although a position of overall responsibility for the han-
dling of errors in the InFACTmachinethe authors wish to
make clear that the developmentof the InFACT   design was
undertaken by a large team of which they were a part. In
particular we thank TonyWilliams, the technical manager


To top