Automatica, Vol. 16, pp. 595-613 Pergamon Press Ltd. 1980. Printed in Great Britain International Federation of Automatic Control 0005- 098/80/ 0 -0595 502.00/0 Assembly Research* J. L. N E V I N S t and D. E. W H I T N E Y t Assembly is a process performed by people and is poorly understood. To automate assembly requires an information-control approach which accurately describes the process, the workpieces and the assembly equipment so that reproducible systems can be designed. Key Words--Computer controlled assembly; robot assembly; manufacturing automation. Abstract--Assembly research comprises the definitive description of how parts interact during assembly (called Part Mating Science), the collection of part mating processes into systems that assemble products, and the integration of this process into the factory as a whole. Products can be assembled by four techniques: manual labor, special purpose machines, programmable systems, or hybrids of the above. Each system includes material handling, parts feeding, assembly, and inspection. Issues of design include technology, part mating and product system design tools, and economics. lie OnOOn~_~n~n ~ i,o ~o/ ............ .... n MO~i I W ~ k,A~~ PAad~iea,oqr~ n~n ~I1N . . . . . . . . . . INTRODUCTION C O N T I N U I N G inflation, competition from other countries, and record deficits in international trade have created a widespread awareness in various countries of the need to increase productivity in manufacturing, which usually means decreasing the man-hours, materials, energy or capital required to produce industrial goods of all kinds. An additional stimulus for increasing productivity arises from the desire to improve the quality of life, including the life of workers now engaged in stultifying, repetitive and sometimes hazardous tasks. For most of the past century growth in manufacturing productivity has been maintained by the substitution of power-driven machines and new technological methods for labor. Today there are pressures to use power, materials and capital more efficiently. The conventional remedies need to be reexamined and new solutions need to be sought. To illustrate, Fig. 1 gives the distribution of direct labor in four major U.S. durable-goods industries and demonstrates the large role that manual-assembly labor still plays, even in the ifN O00nnnnn~n ..... ....... [1 Fl(;. 1. Distribution of direct labour. highly mechanized motor-vehicle industry. On first impulse automation would be best applied by attacking the functions with the highest labor percentages. However, a better perspective is obtained if these percentages are scaled by the value of shipped goods, as Figs 2 and 3 illustrate. These figures are averaged over the indicated specific industries. Thus, for the farm machinery and equipment industries 30.2 per cent of direct labor is involved in assembly, but when this is scaled by the value of shipments it represents only 7 per cent of costs. Similarly, 33.4 per cent of direct labor is assembly in the automobile and parts industry, but again this is only 4.7 per cent of the value of the shipped goods. If we were to completely automate the assembly functions in these two industries we would only change our total costs by 7 per cent for farm machinery and 4.7 per cent for the auto industry. This, of course, assumes we know how to do so. ]~igure 4 illustrates this perhaps more *Received October 17 1979; revised May 9 1980. The original version of this paper was presented at the 2nd IFAC/IFIP Symposium on Information Control Problems in Manufacturing Technology which was held in Stuttgart, Federal Republic of Germany during October, 1979. The published Proceedings of this IFAC Meeting may be ordered from: Pergamon Press Limited, Headington Hill Hall, Oxford OX3 0BW, England. This paper was recommended for publication in revised form by associate editor A. Longmuir. tCharles Stark Draper Laboratory, Inc., 555 Technology Square, Cambridge, MA 02139, U.S.A. 595 596 J.L. NEVINS and D. E. WHITNEY SIC 3522 FARM MACHINERY & EQUIPMENT BREAKDOWN OF VALUE OF SHIPMENTS % TOTAL LABOR % VALUE OF SHIPMENTS Function Co,If 100.0 23.4 Nonproductlon 42.7 10.0 Pro¢lucti=~ 560 ~3 1 ~%i TOTAL l LABOR %SHIPMENTS VALUE OF 4 1426 59A 84 47 16.6 2.4 Asmmbly ~0.2 71 33.5 47 1.3 9 0 1.3 5.0 4.6 0.3 M=terlal Colt| 54.0 Oull~ty Cc~. 15 Othl~ CO¢~, Profit, etc. . . . . . 21 1 . . . SIC 3541 - - MACHINE TOOLS. METAL CUTTING TYPES SIC 3601 RADIO AND TV RECEIVING SETS BREAKDOWN OF VALUE OF SHIPMENTS % TOTAL TOTAl LABOR % VALUE OF SHIPMENTS % TOTAL LABOR % VALUE OF SHIPMENTS F~tkon 100,0 37.5 100.0 15.g Cost~ Nonproduction 32.3 12.1 52.7 8.4 Production 665 24.9 45.g 72 Plwts Fal~i~tion 50.1 10.8 84 1.3 ~bly 11.0 41 239 3.8 2.0 5.4 35.4 IBll~r ial Coltl i Quality Co~t~ . 13.3 17 0E 1.2 MilcJdlar4ou$ I s AND P A R T ~ 202 Inspectio. . ~ic: 37,7- ' V IWO lETO HRI C L E Parts F#;)ri~ t i ~ Mi*cllllmmlou s . F- ~ts, Profit,~t¢. . F'l(is 2 a n d . 3. 2.1 0.3 615 15 1.5 25.6 21.0 . ] . Labor costs as a percentage of value of shipments. ILLUSTRATIVEEXPERIMENTS F=rm M ¢ h i ~ MI~i~T~I=' f and ECmipcnent C~I Elects ble~tCu~ing Types IS*C a5221 tSIC 3541> Radi°al~V ReceivmgSets Motm Vehicle~ ~d parts ISIC3651} (SIC 37171 +. *5O la~t Stiff E~p~V~ R,lat,om Finan~ Ml.k,t,ng Co~p~t,, Op~aUO,, En~i~,,ng +20 0 i +5 0 100 Mamri=l Movem~t m0 A~mblv mo Inl~tlon 900 5.0 100 500 20 50 In~lory +0 i ~han~= m P,oduct Cost Ch~np mA ~ Re(IUir~ n t i NOTE All v a l ~ are e x p ~ 100 100 ~0 38 as a ~ -25 07 O0 2~ c h a ~ In t ~ ratio of c~t5 Io val~ of thipm~ts FI(i. 4. hnpact calculation. dramatically. Here, not only are production functions changed, but also non-production functions "white collar workers"-- like engineering. Again dramatic changes in the specific function cause only minor a few per cent -modification in the final sum. Clearly, no one single device or technique can be used to accomplish such huge changes. Neither will a haphazard application of singular solutions accomplish the goal. Rather, what is needed is a systematic multi-discipline approach if one is going to achieve the kind of changes needed to cause a significant effect on industrial productivity. Although there are many ways of increasing manufacturing productivity financial, fiscal, and social we shall focus here on the possibilily o f raising productivity through the application of science and advanced technology to an old field: assembly. Technology has brought about radical changes in many areas power generation, transportation, chemical manufacturing, communications, and data processing but it has had only a minor effect on the way the broad spectrum of consumer goods, from electric toasters to automobiles, are actually assembled. The rest of this paper will be devoted to describing assembly research in this context. That is, we will generate a n overall s c h e m e o r system required for automatic assembly a n d then identify the research directions currently m process and how they appear to fit into this overall system. In particular, with this kind of approach, systematic tools that illuminate these relationships or help define the performance requirements of the supporting subsystems appear at the moment to be the items most needed. Later, of course, with good agreement on the outlines of the system defined, the emphasis will shift to the refinement of the supporting systems needed. By then, the priority of this supporting work will also be fairly clear. In the simplest view, assembly can be categorized into two principal activities, namely: (a) part mating or what occurs when two or more parts interact during the assembly process, and (b) the aggregation of this process activity into the assembly of an entire product. If one is to consider the "assembly system" then additional things must be considered. Specifically, assembly cannot be considered without the integration of both the feeding of the parts and the inspection steps necessary during the process. Finally, the integration of the above activities into the context of the total factory produces the desired assembly system. At the present time research has been carried out on part mating and product assembly. F o r "'assembly systems" only unrelated items a r e being pursued. Integration of these systems into the factory arc restricted to the present tools of industrial engineers or simple heuristic techniques. However, the application of operational research techniques or knowledge based systems to these last two problems offer more interesting solutions which will be illustrated later. PART MATING In order to define assembly systematic way must be used to functions, group them in some categorize the statistics of their processes, some identify tasks or generic way and occurrence a s a function of various classes of products. An early Assembly Research study by Kondoleon (Kondoleon, 1976) of ten products with various part counts and sizes (Fig. 5) yielded the range of tasks shown by Fig. 6. The statistics for the assembly functions for those products are illustrated by Fig. 7. Additionally, Fig. 7 illustrates the number of degrees-offreedom required in order to move the parts from some location to the vicinity where assembly takes place. These are pertinent data necessary to specify the number of articulations required if linkage type machines (industrial robots) are used in product assembly. ELECTRIC TIMER COVER SUBASSEMBLY* 7 45- ELECTRICTtM~R CASE AND FINAL ASSEMBLY* 18 45" REFRIGERATORCOMPRESSOR FAMILY 26 10" 15 6 to 8 6" 10' 29 7' BICYCLE COASTER BRAKE TRANSFORMER ELECTRIC BUSHING FAMILY END CAP SUBASSEMBLIES FOR SMALL INDUCTION MOTORS INDUCTION ivtOTOR MAIN BODY SUBASSEMBLY AND FINAL ASSEMBLY ELECTRIC JIGSAW TOASTER OVEN 21 15- 58 41 12" 15- AUTOMOBILE ALTERNATORFAMILY*" 17 8' *CURRENTLY ASSEMBLED AUTOMATICALLY **CURRENTLY ASSEMBLED BY A MIXTURE OF MANUAL AND AUTOMATIC WORKSTATIONS FIG. 5. Ten products analysed. 597 Additional statistics are also needed for weights, part tolerances, and force levels occurring at assembly, etc. Using these techniques, General Motors found that 90 per cent of the pieces used in the assembly of an automobile weigh less than 2kg. John Deere and Company, who manufacture large agricultural machines like combines and tractors as well as large earth moving equipment, in a similar study found that 80 per cent of their pieces weigh less than 4 kg. Round peg-hole insertion tasks with chamfered corners are by far the most frequent in the assembly of metal products with machined or cast parts (Fig. 7). This task has been extensively analyzed, and a complete statement of the requirements on relative errors is now available. The important design variables have been identified and for many parts it has been determined that conditions for successful assembly can be met more easily than the clearances between the parts would make it at first appear. These analyses have been amply verified experimentally and it has been shown that the conditions can be obtained from the blue prints for the parts (Nevins and co-workers, 1976). The analysis and experiments also form a valuable pattern which can be followed in the analysis of other tasks. In particular, the extension of this work to compliant parts has proved to be particularly useful in the design of parts used in the electronics industry. ¢~.MF ON~ miNI ACT tWO POq~r ¢aNWAe~ D~LAYEe Te r ANGULARERRORIGRE~N) FiG. 6. Assemblytasks. DURINGtb*K)POINTC O N T A C t THE T~ OF THE pE~ MUST ~ [ U A I ~ W I 1 H I N TNE IHE F ~ N E L IS REO FUNNEt SMALLE. IF T N t R A ~ I O OF CLt~ ~NCETOPE¢ mAMEVERqS ~ * t t e ~ IgG. 8. Basic geometry. ............ .,,.~,,,, . . . . . . . . D I m E { ~ I O ~ OF" A T T A C H ~ I L ~ [ OF" IPAR T~*~ u .R~l~.*~l m the FIG. 7. Task ' ':,~-~,~ . ~ " ~ , ~', ~,,~ .,,~ " ~ n ~r statistics. ~ ~ I~ , Cases where the peg touches one or both sides of hole have been considered (Fig. 8). Two types of "jamming" during assembly have been identified (Fig. 9), and the conditions for preventing them specified in terms of geometry, friction coefficient (Fig. 10), and-arrangement of the applied forces (Fig. 11). The geometry of 598 J.L. NEVINS and D. E. WHITNEY . d ~ PEG Fz DIFFERENCE BETWEEN WEDGING AND JAMM/~G was chlrtlJed din'in8 the d e v e l o p mlml ~ cemtldJl~t-lp, Jppe~ moelbma/m~. ~ foe' m p l e , • INIreml dllwe~ be¢omm wedled (IeJ~), N b Ute8111~ IIo~..keclL A j a r ~ spt~lkmffea eG force wUI d d o r m the dJrawe~ mr die b u r m u e r b o 4 ~ 1~ee¢7 8bows t k t t w e d s ~ arbes w b e a *ke drltwe¢ Js imes-ted at such ~ ~ ~ e c4--t the r u t l o eli L / D b lees t l m t t k e c o d ~ J e a ( o~r IrrklJoa (;,) wltea I w o - p o i n t coutact I r n t occurs. 1['be e a ~ r e m e d y b to p u U 1be d r a w e r o u t m d ~ 8 p b L If, however, t k e retio L / D b larse~ t k m p at d i e time o( ~ (wo-pebst eoutact (nEht), wedJrln8 c l n m o t result, mlt b o u l b furtbe~ movement ~ m be impeded by j J m m l n 8. l r b e remedy Is to break the t w o - p o i n t c o n l l c t by p u ~ I t A, Ihereb¥ c l ~ Ibe d h ' e ~ o( be4tk the a p p l i e d force t u d the applied enomemL - r~ Z JAMMING REG(ON Fl(i. 9. Wedging and jamming. 1 QX rj JAMMING R Wt4EN THER| iS k~ FRiCTiONCON. TACT r~f~c| (IILUEI IS FELT|QQALL¥ AND A rUnWAYS RE~TION (RED) LATERAL FORC~CAN ~IE t.~EO AJ A CUE TO THE OEI~RED CORRECTIVE MOTION(REO) 4 ~REGION OF NO JAMMING ~A WHEN THERE IS MO FRiCTiON COil. TACTFO~qcESI|LUE)CAU~EAI4OM~N T /~DUT THE TIP IRE(}I. WP*41CHOlVI[I A CUE/UStO T.E O~R~O CO~lCTWE MOT~C~InED) F'IG. l l. Force moment relationships to avoid ammmg. ....L . CO'TACTFaeCESe~TW~E~PARTSCANU SaNStO~VO~ 0 TO auloE C O 4 ~ C T W E N O T ~ FR~TION ca~¢ CO~TRmUTE FORCIES ~ M4J~ TNE FC~CE ff4FOgtMAT~N CONTACT FORCES ARE FIEFIPENOCCUt.AR TO TNE CONTACTING ILmFACES FRK:T~D4~ FOp,CE8 RItE PARALLEL TO C~ITACTING St,reFACES MAQI4f~ OF I')'NE FRK:TION P ~ C E i i ~q~qOXIMATELY p ~ l FO~ITIOklAL TO TNIE NJU3N4TUOIE Of THIE C~WTACT F,I~IC| THE CIQWSTANTOF P ~ O F O I N T I ~ I T Y I I CALLED TI,I~ECOIFFICI|NT OF FRICTION ITJVALUERAIIIH]iE$FNONO;' FORIWE| L PART|TO I o ~ LARGER F ~ AL~ 1 ~ p~TI • FI(;. 10. Eorce friction relationships. angular errors in screw thread mating has also been analyzed and shown to have looser error tolerances than the typical peg-hole problem. To test these theories a number of carefully constructed experiments have been carried out. One of the first experiments (Fig. 12)involved the use of a pair of Unimate 5000 manipulators to test the capability of these kinds of manipulators to perform the assembly of simple, rigid, machined pieces without adaptability. What was shown was that assembly was possible, but the probability of success depended on unspecified compliances unique to each machine that were at least time and temperature dependent. A second group of experiments was carried out to test friction and jamming theories (Fig. 13). These early experimenrs were conducted on a second generation pedestal force sensor, but the sensitivity of the unit to moment information was inadequate. This same sensor was used FIG. 12. Programmability experiment. Two Unimatc 5000 Industrial Manipulators were used to assemble pieces of a small gasoline engine where nominal clearances of parts was 0.001 0.002 in. (0.025 0.05mm). Positioning error of the manipulators was 0.020--0.028 in. (0.525 0.72mm) and about 0.5' in angle (neglecting backlash). The experiment showed that this kind of system would work for pieces with slightly lubricated mating surfaces, chamfers, bevels and rotationally compliant grippers, but the probability of successful assembly could not be stated. 599 Assembly Research FIG. 13. First friction and j a m m i n g test. Experimental system shown was used to verify coefficient of friction for various materials and to test j a m m i n g theories. Although system used second generation force sensor system (pedestal configuration), its m o m e n t sensitivity was almost two orders of magnitude less sensitive than force readings. successfully to perform the first experiments for very large errors (Error<radius of hole) where estimation techniques might be useful. To compensate for the lack of sensitivity in the sensor, the test pieces were made ten times oversized (Fig. 14) (Simunovic, 1979). The first definitive part mating experiments required the development of a more sensitive force sensor array and a wrist with only one carefully specified compliance. This device was attached to a milling machine base and specific relative errors were imposed on insertion tasks using special test pegs and holes. All data were gathered and processed on-line by a computer. The results verified the conditions for one of the two predicted types of lamming and showed that unambiguous force and moment data could be obtained (Fig. 15). A newer more sensitive instrument for studying small electrical components with a minimum threshold of 4g in the x and y direction and 10g in the axial direction has recently been developed to support these activities (Fig. 16). {WLL~E~t *27 ~ l t ~ E ~ 6 G~ oF ~ Aer a To iim~mN ~lm~ ~ } Flcl. 15. First definitive part mating cxpcrimcm. FIG. 14. Tilt strategy estimator. Use of estimator for part mating errors approximately equal to radius of hole. Pedestal force sensor measured tip-over moment from which estimator calculated the azimuth of tilt, and from this, the direction to the center of the hole from the center of the peg. The peg was then driven sideways until it struck the opposite side of the hole. The tilt was then removed and the peg inserted into the hole. The sequence of moves constitutes an information strategy without search. The inaccuracies in the U-5000 manipulator and the lack of moment sensitivity in the force sensor required that the pieces be ten times over size. FIG. 16. 4 G r a m 6 Axis force sensor array. 600 J . L . NEVINS and D. E. WHITNt ', The combination of geometric and friction analyses, plus the stimulus of the above data gave rise to the invention of a totally new device for accomplishing assembly--the Remote Center Compliance (RCC) (Watson, 1976) (Figs 17 24). This device accomplishes mechanically what active force feedback does with sensors and servos. It allows chamfered peg-hole insertions with 0.0003 clearance ratio* with 1 mm (0.04in.) or larger initial position errors with a few degrees of angular misalignment. The error tolerance of / MOMENT ,',,,, '*<IMF~,I C F N T E N :l~ C O M ~ L I A N C I ~LATERALrOOC E nOT CfNTER ~ORC~ AV OF COMPLIANCF I RCC APPLICATIONS Categories and examples: Clearance fits Bearings in housings Shafts in bearings Splines Castings or forgings onto locating pins Noncircular pegs into broached holes FIG. 19. RCC clearance fit apolication~ -,q FORCE ~,~ CENTER ~ CO~PL I a n c ~ t J(;. 17. Requircment for remote center, A II -7 L1 I. IG. 20. Model 4A R C C The Mod 4A has been designcd as a rugged relatively foolproof unit with overall protection in all directions - done to provide a unit for evaluation by industr,, m the industrial environment. Several models of the 4A ha~e been placed in industry. Its nominal specifications are: Mass (weight), 1.36kg (31b); focal length from base, 20cm (Sin): lateral stiffness, lOONcm ~ (551bin 1); torsional stiffness. 0.1NM mrad ~ (0.9in lbs/mrad). _A- / CONNEC/ LINK NGC~T i / J I I INTERFERENCE FITS -? t-it;. 18. Concept RCC. The vertical compliance center effectively at infinity. produce only lateral motion. The other compliance center at the end of the peg angular rotation in response to a moment Locating pins Calibrated orifices links (B) give a So lateral forces links (C) give a which allows only applied at the tip. Bearings in bores Shafts in bearings Pins in bores Noncircular geometries lq~. 21. RCC interference fit application. *Clearance Ratio dia. of Hole - dia. of Peg Diameter of Hole Assembly Research 601 MANUFACTUR ING Reduction of tooling forces Reduction of tooling breakage Tool-guide bushing mating Tool mating to bushing I" IG. 22. RCC applic~ltion~ ill inzlnul'zlctiuit/g. l-t(i. 23. Designed to provide a lighter weight (1 lb or 0.45 kg massl unit capable of easier repair and lower cost. The focal length and elastic parameters are nominally the same as the model 4A. Fifteen units are currently being evaluated by . industry. this passive compliance device is much larger than was originally thought possible for passive devices. Further, the device works for both negative as well as positive clearances. To illustrate the usefulness of the RCC device, a comparison can be made of a number of techniques for absorbing the error occurring between industrial robots and the pieces they might be required to perform tasks on (Fig. 25). Industrial devices tend to be quite imprecise compared to a rigid machine tool. Therefore, some kind of device is necessary to absorb the error that arises due to the lack of repeatability of these machines when they are required to interact with piece parts. Figure 25 shows two classes of systems, namely passive and active devices, that could be used for a variety of tasks that include material handling, assembly, and manufacturing. Figure 26 compares the estimated probability-of-success of accomplishing tasks for three passive techniques and one active technique vs the initial position error. The active system shown--the Hitachi Hi-T-Hand (Fig. 27) (Goto, Inoyama and Takeysau, 1974}--is also capable of solving the chamferless insertion problem. Figure 28 further compares the active FIG. 24. Fixture for aiding rapid, reproducible manual instrument bearing assembly. Instrument ball bearings are a light interference fit at operating temperatures and are assembled into a heated housing to provide assembly clearance. The assembly must be done quickly and at light force levels. It is c o m m o n in unaided assembly that a ball bearing will j a m and have to be driven home. A manual assembly stand assisted by the RCC to avoid jamming, and a six degree-of-freedom force sensor to document the assembly force history, is expected to yield significant improvements in the process. A brake is included in the design; activated by the force sensor, it can retard the drive and limit the seating forces exerted on the bearing. Any one of the six forces and moments measured may be selected for analog display (on the meter to the left of the assembly stand). Each or all of the measured quantities may be stored on a computer file or recorded for permanent record. ERROR ASSONBER PASSIVE SYSTEM No Complilmce Undocurmmed C o m p ~ Enwn.erld Compl~nce RCC x - Y T.bM TASKS 1. MATERIAL HANDLING L old/Unlold Machines Stick/P~:k Tnzr,s~r Pzrtz Hold Puts 2. ASSEMBLY ACTIVE SYSTEM F~c~ Fa~Jb~ck Hi-T-Hand ViRon Clear eeoc Fits Inlerf~en~ Fits Other Tasks 3. MANUFACTURING Drilling R~Im9 I-'l(:. 25. Error absorbers for industrial robots. • I .:/' "L~ o.ooi " \ ! i I o01 1 dl ERROR (ram) ERROR ABSORBERS PROBABILITYOF SUCCESSFULASSEMBLY ~Y TECHNIQUEvmu$ INITIAL POSITRON ERROR FIG. 26. 1.0 ' - - ' - LOCATION OEPENOENT ON CHAMFER SIZE 602 J . L . NEVINS and D. E. WHIINEY POST IO I NN I Gi PROGRAMMABLE > ~ MECHANISM I I ~'Fz ~,~,, 1> I MEOHAN[SM I 1 I "+:-> Fx 0.,c Two dimensional representation of inserting operation. FIG. 27. Hi-T-Hand. •////// . . ~ / / / / L - Z ~ I RCC 0,', d, ,% 4 tRROR (~1 ERROR A B ~ R B E ~ $ ~LUTION TIME ~,ut INITIAL ~SITIONAL ERROR 0 Fk~. 28. system (Hi-T-Hand) and the passive system (RCC) by solution time and the initial position error. For the class of errors it can handle, the totally passive system is quite simple, cheap, and very fast. The research challenge thus is to create a new device capable of solving the chamferless ins&tion problem, but is simpler and not quite as slow as the Hi-T-Hand. Compared to the above systems, the use of vision (Agin, 1977) is limited by cost, and the x y passive table (Rosen and co-workers, is limited to certain applications. The above devices are only the start. Since the general requirements for solutions have been determined (Simunovic, 1979; Nevins and co-workers, 1978a, b; Drake, 1977) it will only be a matter of time before we are flooded with a whole group of new devices. SYSTEMS To identify and quantify the kinds of assembly tasks required to assemble products, the study shown earlier was made of ten products and subassemblies to answer two questions: what assembly tasks occur, and what gross directions relative to the assembly do the parts approach from? Eight of these products were of cast or machined metal, one of molded plastic, and one of a variety of plastic, sheet metal stampings, and wires. The latter product was the exception to the findings from the others: 70 per cent of the parts arrive from one direction, and 35 per cent of all tasks are single peg-hole insertions. Adding a second direction antiparallel to the first picks tip another 20 per cent of the parts. Screw insertions represent 25 per cent of all tasks, and ten other tasks make up the rest. This study justified the choice of tasks to analyze, although more extensive surveys could alter the results. It can be concluded that the first nine products form a group with quantifiable characteristics. In particular, because most parts arrive from one direction, it should be possible to assemble them with devices which have much fewer than six degreesof-freedom. This has important economic and system configuration ramifications. A second survey, Fig. 29, examined industrial design practice with respect to sizes and clearances between parts. Here it was found that particular types of parts, made by certain manufacturing techniques, reliably fall into predictable clearance ratio ranges from 0.001 to 0.01. This confirms the choice of assembly task difficulty to analyze. It categorizes parts by their geometric properties and allows the analytical tools developed to be applied to general manufactured parts. One of the principal problems facing programmable assembly researchers is the fact that there are many technological options to consider for these very complex systems. For our early work, we chose a fairly straightforward, but very powerful, economic modeling technique (Lynch, 1976) to help systematize these issues. Economic models have been made of programmable assembly machines and systems of such machines. Simple assumptions have been made concerning machine component costs and cycle times. Comparison models of special purpose assembly machines and manual assembly have been developed from similar assumptions so relative costs can be obtained (Figs 30-33). The economic model for programmable assembly systems identified a price time product relationship as an important tool for the selection or design of assembly robots to be used in programmable assembly systems. The price time product is defined Assembly Research 603 INDUSTRIAL PRACTICE-- DIAMETER VS CLEARANCE RATIO I0 HALL BEARING FIT (OUTER DIA. TO HOUSING) ~ - r - ~ . -~ -~---~z--~l ID I ~ ~ \ AUTO VALVES 1974 AUTO CAMSHAFTS 1974 --GENERAL MACHINE BEARING PRACTICE . . . . ELECTRIC MOTOR PRACTICE~ --'--PRECISION LAPPED BEARING • BRONZEBUSHING BEARINGS \ \ \ ~ WASHERS • \, \ \ \ I [ BALL BEARING Fir (SHAFT TO RING) 0.1 I I I I I~ 1 I I1~1-- I 1 I I 1111 I I O.Ol 0.001 0.0001 I " CLEARANCE RATIO I I I II I 0 1 D FIG. 29. Industrial practice. where Manual MCPU = MCPU MATP - assemb1~ cost: MATP * LABCST * NPART manual assembly cost per u n i t manual assembly tlme per p a r t , sec ASSUMPTIONS LABCST = cost rat e of l a b o r , S/see NPART - number of part s in the product a .to mat lnn I t r a n ; f | Ft~d TCPU - where r machtne~ assembly c o s t : TSCST PAYPER * VOL TCPU - t r a n s f e r machine assembly cost per u n l t TSCST - t r a n s f e r machine t o t a l c o s t PAVPER * payback p e r i o d tn years Next, TSCST - NRART * TMCPP where TMCPP " t r a n s f e r machine cost per p a r t Therefore NPART * TMCPP TCPU = PAYPER * VOL T h i s model i s based on the payback p e r i o d method r a t h e r c o u n t e d cash f l o w f o r simplicity. worked o u t u s i n g the more a c c u r a t e Any p a r t i c u l a r than d i s - 10 PROGRAMMABLE STATION PRICE $30,000 TOOLING PER PART $7500 TRANSFER MACHINE COST PER PART $30.000 PART STATION TIME 3 s PAYBACK PERIOD 2 years to 4 years LABOR COST $ 7 . 5 0 / h r to $10.OO/hr MANUAL ASSEMBLY STATION TIME 7 s NUMBER OF SECONDS PER YEAR 1.152 x TO7 case can be method. FIG. 30. Economic models for manual assembly and fixed automation. Programmable s~stem cost: PSCST - NSTA * STAP + NPART * TOLPP where PSCST - programmable system cost NSTA = number of assembly stations STAP - s i n g l e s t a t i o n price • TOLPR = t o o l i n g p r i c e per p a r t The number of s ta tio n s required is NHTA = VO6 * NPART * PARTTIME NHPY where NUMBER OF PARTS PARTTIME- assembly tlme per p a r t , sec RSPY - 1.152 X I07 sec/yr f or an uptime f r a c t i o n of O.E, a 250 day year and a 2 s h i f t , i6 hour day. The assembly system cost Is then PSCST = STAP * VOL * NPART * PARTTIME + NPART * TOLPP NSPY and, using the same payback period model as in e q ( I I - 2 7 ) , the assembly cost per u n i t is TOLRP ] . _ _ _ _ PCPU = NPART [ STAP * PARTTIME PAYPER NSPY + VOL *Includes basic feeding mechanism (bowl feeders, hoppers, magazines. e t c ) , feed tracks and chutes, and placement or escapement devices or conveying mechanisms t h a t l i n k the parts together. FIG. 31. Economic model for programmable assembly. FIG. 32. Cost assumptions for comparing manual, fixed automation and programmable assembly systems. as the cost of the assembler (including installation in the system) multiplied by the average single part assembly time for that assembler. Thus, the price-time product measures the effects of the cost of the assembler and its speed. Assemblers capable of comparable motions which have equal price-time products will, when combined into programmable assembly systems, result in the same cost of assembly for that product. For example, if assembler B costs $100,000 and assembler A costs $50,000, but assembler B can assemble one part, on the average, in 2s while assembler A takes 4s, then assemblers A or B, when used in the correct quantity in programmable assembly systems will result in the same cost of assembly for that product since both have price-time products of 200,000 $-s. 604 J . L . NEVINS and D. E. WHITNEY PRODUCTION VOLUME REGIONS WHERE PROGRAMMABLE ASSEMBLY MIGHT BE ECONOMIC COST TO ASSEMBLE ONE UNIT OF PRODUCT WITH TEN PARTS, S FACTORS • lO PRICE-TIME PRODUCTS • L A B O R COST • R E Q U I R E D R A T E OF R E T U R N SlOlh MANUAL LABOR = = i a ~ @ S7 SO/h iiI mi • I| 0.01 01i S • MANUAL ASSY i 1 0 D AUTO.'~ =~III~/XE %,, MANUAL me # l l l l l ~ { n ~ PROGRAMMARLEAssYm~ m~ ANNUAL PRODUCTION VOLUME. MILLIONS FIXED AUTO ~ . . , D I ~ [Q ~l(i. 33. l-;conomic comparison of manual, fi×ed automaliom and progranlmable assembly systems. Also, the economic model can be used to obtain an upper bound on the price--time product for which an assembler cannot be economically competitive with either manual assembly or fixed automation assembly. For an assembler to be significantly economic, with a relatively wide range of production volumes, its price-time product value should be well below this bound. For assumptions based on typical 1976 values of the cost of system components, labor rate, and cost of capital used, the upper bound was estimated at 293,000 $-s. The structure of this model has been examined, and conclusions have been drawn concerning the sensitivity of unit assembly costs to various factors. The sensitivity is equally important for identifying research issues or directions that need to be examined. Without an analytical model, these vital sensitivity analyses cannot be performed. To study adequately the area of programmable systems, it was found necessary to construct a programmable system test bed. This test bed consists of a single electric arm originally constructed by the Bendix A & M Division (Dayton, Ohio) as a prototype arm for industrial assembly. Supporting the arm are sensors, tooling, parts feeding, minicomputer systems and a specifically developed software system. To test the knowledge developed to date an automobile alternator with 17 parts was assembled on the system (Fig. 34). I~[(;. 34. I)rogrammable system test bed. At present it can assemble the alternator m 2min and 42s* using six tools and eight tool changes. In principle it can assemble products about one foot cube in size requiring insertion motions vertically down, or (with appropriate tools) horizontally, or vertically down with some spin about the vertical axis. Electric motors, some types of gear boxes, electric entrance boxes and terminal board arrays are examples. This test has yielded much knowledge of a technical and *Applying routine industrial engineering this number can be reduced to about 70s tNevins and co-workers, 1978a). Assembly Research e c o n o m i c n a t u r e c o n c e r n i n g p r o g r a m m a b l e assembly. S o m e of the results are s h o w n in Figs 35-37 a n d are s u m m a r i z e d in T a b l e 1. P~RTA~EMeLYCYCLE ~'~ GRO~ MOTIONTIME (FEEOEATOWORKAREA) ~ G~OSS E MrmON TDME. i ,N~ERFACE MORIONT*M~feR APPRaACH & WITHDRAWAL A B - ~ DOWNWABDMOTqoN UPWARD~OTION B - PART"~L~ASE C U~WARDMO~tO~ - O~WAROMOTION U~AflDMOTION , FINEMOTIONTIMEFORPARTPbCKUP B ¢ ~ A ~ pickup U~ARO MOTION FIG, 35. Robot arm trajectory definitions. TABLE I. 605 RESULTS FROM EXPERIMENTAL PROGRAMMABLE TEST BED (ll Adaptable, programmable assembly of a real industrial product was accomplished: adaptability was demonstrated by the system's ability to absorb position error of 1.25ram (0.05in.) or more due to part variations and uncertainty in the location of part feeders; programmability was demonstrated by ability to program different assembly sequences using substitute tools in place of broken tools or new "tricks" to improve system performance; approximately 75 alternators were assembled to test the system. (21 Performance of the system was thoroughly documented (Figs. 35-37). t3) Concept of engineered compliance was shown to be sufficient to perform a variety of difficult single-direction tasks that could either not have been performed at all, or at least with the necessary reliability. (4) The software and teaching technique appear adequate for this class of assembly, For this problem planning was done off-line and explicit locations of tools and feeders was taught on-line, i.e. on the shop floor. ASSEMBLY CYCLE TIME BREAKDOWN ESTIMATED TIME (158.6) [~ MEASURED TIME (16~ II 1 60 7-50 =, 40 30 i I \ 0 GROSS MOTION TIMEFOR ASSY INTERFACE AND FINE MOTIONTIME FOR ASSY GROSS MOTION TIMEFOR TOOL INTERFACE AND FINE MOTIONTIME FOR TOOL CHANGE CHANGE CYCLE TIME FOR ASSY CYCLE TIME FOR TOOL CHANGE FIG. 36. Assembly cycle time analysis from programmable system experiment. A l t h o u g h the previously described e x p e r i m e n t was the first p r o g r a m m a b l e assembly e x p e r i m e n t to generate any real d a t a on the p e r f o r m a n c e of these classes of systems, it w o u l d only be considered m a r g i n a l l y e c o n o m i c by industry. Clearly m o r e definitive tools are necessary to define system p e r f o r m a n c e in a m o r e specific way. T h e early e c o n o m i c m o d e l i n g suffers from simplifying a s s u m p t i o n s and the assembly stations are treated as if they were all the same kind. F u r t h e r , other technologies like parts feeding, m a t e r i a l handling, a n d inspection must be i n c o r p o r a t e d if the tool is to be of any general use. T o this end, a new tool, based on mixed linear and i n t e g e r - p r o g r a m m i n g , has been d e v e l o p e d (Whitney and Graves, 1979) which on initial tests holds p r o m i s e in this area. To illustrate, using the assembly of the a l t e r n a t o r as a reference case, a n u m b e r of e x a m p l e systems were analyzed for the range of a n n u a l volume of 60,000 units to 480,000 units in eight steps (Figs 38 42). At each step the p r o g r a m selected the system most suitable from a cost a n d p r o d u c t i o n constraint. ANALYTICAL DESIGN TOOL • Mixed linear/integer programming formulation 8 I0 TIMEILf FIG. 37. Trajectory data for the up down axis for six assembly tasks. • INPUT: Robot capabilities, task requirements • OUTPUT: Selected robots, task assignment fractions, task assignments when one or another machine is broken, utilizations, costs FIG. 38. Analytical design tool for manufacturing assembly systems. 606 J . L . NEVINS and D. E. WHITNEY 'ij I$) TIME/OPERATION I PAX. 3 ~A, $~0.000+ ~.000 2 PAX slz,ooo RESOURCEi 4 2 PUMA 5 6 PROGRAM TASK I NUT LOCKWASNER PULLEY FAN SPACER FRONTHOUSING BEARING RETAINEF~ 3 SCREWS USES • System preliminary • Technology • System • Transport topology • Parts flow management • Phased introduction needs design assessment I breakdown resistance 3 S~ACER 1~ )2 13 14 ROTOR TIGHTENNUT REARhOUSING 3 SCREWS I 8 100 100 8 :(,. 39. Input example using alternator as an cxamplc. ~Ax a 4 s X X ........ ....!!!!I!! e 7 8 X X X x = UNIT IS FULLY x = U N I T I~ A L ~ T LOADE0 FULLY 1-tG. 42. List of p r o g r a m uses. Many simplifying assumptions and approximate data were used in these runs. The results should therefore be taken as an indication of the program's capabilities, but not used to generalize on the suitability of any particular robot. Although robot systems have been shown for illustration, the program will work equally well for complex systems involving manufacturing, material handling, parts feeding, and inspection assuming that the necessary data base on these systems has been established. The necessary data bases (Fig. 43) include the equipment data base (available technology) as well as the operations data base for the manufacturing and assembly constraints. LOAOEU PUMA 2 PUMA 5pSCR DR AUTOPL I . . . . . . . . .I AUTOPL 2 ........ ~NUAL i X ...... b . . . plan . ON~SHIF~UNIT *f ~. ~AVBACK F-I(~. 40. Typical results. L I M I T I N G V A L U E S OF C O S T / U N I T vs V O L U M E FOR S A T U R A T E D C O N F I G U R A T I O N S ONE Y E A R P A Y B A C K $2.00 $1.00 I- 8 w < $0.10 10000 I J i I I I I I I l 100,000 UNITS/YEAR FIG. 41. E c o n o m i c comparisons. I I I I I J I t 1,000.000 Assembly Research 607 A DESIGN SYSTEM (FUTURE) (AVAILABLE TECHNOLOGY) (MANUFACTURING & ENGINEERING) PROGRAM TO CALCULATE OPERATION TIMES ANO EQUIPMENT COSTS .EGR,STIC INPUTS l I ,N'EGER L.NEAR1 PROGRAMS ~ J f REL,AB,L,TY, _l -L 8REAKOOWN s uoY MACHINE SELECTIONS AND TASK ASSIGNMENTS FIG. 43. Completedesign showing required data bases (economicand technical) and method of inputinginnovativedesignideas. ASSEMBLY SYSTEMS The previous discussions dealt with tools and techniques of importance to the assembly of a product in a very limited way. That is, only assembly functions were considered and the supporting systems like part feeding, inspection and material handling were ignored. However, the latest design tool based on mixed linear and integer-programming does offer the potential for handling these systems as well as the ability for examining specific technology items like vision system. So if the problem can be reduced to one of annual volume and costs, then this technique will work. Let us consider more complex problems. For example, consider a product where assembly steps and manufacturing steps are intertwined and may actually share the same machines with different setupts. Here it may be necessary to cluster the operations into separate subsystems or to consider possible redesign of the product. The choice of clusters depends not only on the technological capabilities of machines, but also the need to support different manufacturing strategies. Depending on the market for the product, one may have to build to order, or build subassemblies to stock and final assemblies to order. Or one may, for large enough production volume and order quantity, build entirely to stock. Economic studies show that increased volume will be best served by less programmable but more efficient technologies. At present there are no design tools for systematically attacking such problems. Pieces of the necessary tools have been developed principally by the management science groups for solving the above class of problems. However, these tools must be developed and integrated. The artificial intelligence people have also been examining related issues and have created deductive planning and path finding techniques for relational data bases (Kellog and co-workers, 1978). These systems are quite big and expensive to operate. Further, they, and for that matter all the tools mentioned so far, need extensive development of data bases, both technological and economical, if they are to be at all useful. We are just now beginning to apply these tools and beginning to appreciate the complex set of questions that arise when one seeks to apply them in industry. But, only through this kind of process is there any hope of creating the understanding necessary for designing the newer systems of the future. As the result of the PSP study indicated (Figs 2-4), only a wide spectrum, multidisciplined approach can achieve the kind of productivity gains industries and nations are seeking. ASSEMBLY TECHNOLOGY Assembly technology can be classified into two broad groupings, namely: (1) system issues, and (2) component issues. System issues Product design. Although the importance of this work has not really been highlighted, it is quite obvious that maximum system efficiency will be realized only when system and product design are truly integrated. It does not matter what the system is. It can be a manual or special machine, or a programmable system. The only thing that matters is that the limitations of the 608 J . L . NEVINS and D. E. WHITNEY particular assembly techniques are compensated for in the product design. It has often been found in special machine design that once a product has been redesigned for that assembly technique, it is often significantly cheaper to assemble manually. The part mating studies at the present are our best technique for indicating system design requirements. But this work is not complete, so a complete programmable system product design specification cannot be written. Analytical tools and simulation. Three types of analytical tools are needed in this field: (1) part mating analysis techniques, especially to treat problems of non-rigid parts; (2) assembly system design and operational tools, to specify system needs in the face of uncertain demands as described earlier; and (3) assembly system financial analysis and justification methods which adequately take account of system flexibility. Parts fi, eding. The parts to be assembled must be presented to the assembly system somehow. There are two crucial issues, as noted below. (1) Should the parts be fed oriented in the proper way for the robot, as they are for other automated processes, or should they be fed (more cheaply) unoriented, with the robot having the responsibility for orienting them'? The former makes for a simple robot system and pushes the responsibility off onto other parts of the factory. The latter requires a complex system, but one that can more easily be introduced into present factories one work site at a time. 12) How to keep bad parts from entering the assembly system and how to carry out in process and final inspection'? People are very good at both of these when their performance is up to par. Machines usually detect bad parts by jamming, taking time and requiring a person to free them. The required inspections can be quite subtle and often have not been quantified ("click means it's good, clunk means it's bad"). So far this problem has been addressed only by proposing presorting of the parts by people or by supposing better and better computer TV vision systems. Table 2 lists some alternative feeding methods and comments on them. Inspection systems. Currently there are no systematic ways of looking at inspection only singular solutions and devices. Some description of this current work is described below in this section on sensors under component issues. Sol?ware and hardware control systems. This section discusses design of computer controls for robot arms, while the next section discusses philosophies for teaching new assembly tasks to a robot. T A B L E 2. A L I E R A N I F Method Conventional feeder tracks, vibratory bowls, etc. I>ARF I I [{DIN(i MI I l l ( ) l ) S Comments Examples The least programmable. Paris arrive oriented but often a person docs the orienting. ('an use simple robot lfewer than 6 axes could be sufficientl [)raper PAX U-6000 SIGMA ('onvcntional assembly machines Parts dumped randomly onto a vision station Paris arrive disoriented. Requires TV/Computer. Requires 6 axis arm and time to reorient Bendix PAX U. of Rhode Island SRI Palletized kits of identical parts Parts arrive oriented. A person orients them (could be oriented}. Simple robot Weslinghouse design Palletized kits of parts needed for one product unit Parts arrive oriented. A person orients them (could be automated). Simple robot ? The same kit plus all the necessary tools Better for model mix assembly. Simple robot ? A computer control system for a robot[s) must consist of one or more computers plus programs to run the arms, read sensors, input new assembly instructions, execute these instructions, and interact with the operator. Data must be inputed, consisting of points in space, times and speeds, and values for forces. The arms may be controlled in a variety of endpoint coordinate systems. Separate computers may be used for sensor processing, operator interaction and so on. Three approaches have been taken so far, as described below. (1) Complex systems consisting of compilerbased languages, IBM 360 or similar host computer, IBM System 3 or System 1 real time controller, and perhaps additional sensor processors. Examples include work at tile Stanford University Artificial Intelligence Laboratory (Finkel and co-workers, 1974) and IBM (Lieberman and Wesley, 1975). (2) True minicomputer based systems (PDP 11-34 or Data General NOVA) stand alone, usually with interpreter based languages which are simpler than the above compilers (examples include Bendix PAX, DL DIAL, SIGMA) (D'Auria and Salmon, 1975). (3) The simplest: hardwired logic set up to execute a sequence of taught spatial points one at a time. A computer may be used to input these points. Such systems have quite limited adaptability. The most they can do is stop if conditions are observed to change. The Unimate class of machines are an example. Assembly Research Naturally, the simplest is the cheapest and most reliable. The major issue again is how much complexity is necessary. It is nice to contemplate an adaptable system, but one must be able to show exactly what events the system is expected to be able to recover from. The hard fact is that many simple events (dropped parts is a severe one) which a person could handle easily would require a dazzling robot system. A second major issue is how to obtain the desired arm motions. In our work we have identified three motion regimes: gross motions (high speed, low accuracy transfers), fine motions (when parts are touching during assembly), and interface motions (between gross and fine when the arm is lining up the parts) (Fig. 35). At present the robot performs all of these except if a Hi-T-Hand or Remote Center Compliance (RCC) are in use. Exceptions occur where binary force sensing (detecting contact) (D'Auria and Salmon, 1975) or visual servoing (Agin, 1977) (using wrist TV to locate a hole) are used, but even here the main control technique is for the computer to drive the arm to a position or through a sequence of positions. Forces, velocities, even frequencies, could be used as control references, but so far have not. The required computer complexity for such methods has not been evaluated. Table 3 compares a number of current systems. 609 Teaching methods. Teaching a robot a new assembly task involves determining the assembly actions required, their nominal sequence, and their detailed breakdown into robot motions, tests, loops, tool actions, contingency actions, and so on. This breakdown must then be communicated to the robot controller. At present all robot programming is geometric. That is, points and orientations in space are the fundamental data used in programming. Force data is used mostly as a check or to conform to load requirements of the assembly itself ("torque nut to 10N-m'). Intellectually this means that all programming is physical, giving detailed "how to" guidance to the robot. There is, as yet, no functional programming ("this is what I want, you figure out how"), although various approaches have been suggested in research laboratories (Whitney, 1969). The major current research issue in robot teaching involves what is termed explicit versus implicit programming. Explicit programming consists of expressing the desired robot actions in terms of explicit points in space and a sequence of logical actions. This technique is common among the simpler computer architectures. The implicit technique requires much more computer power and assumes that the robot and work comprise a surveyed environment which has been stored in the computer as a "world model". TABLE 3. COMPUTER CONTROL SYSTEM OF VARIOUS ROBOTS Bendix PAX Draper PAX U-6000 U-500 Shuttle SRI Hardware architecture One mini plus disk One mini plus disk External servo hard-wired control 6 micros plus diskette One computer PDP-10 host plus 2 minis Control complexity Several coordinate systems, Stored points, Control of 2 arms. Program seq. with conditionals Two coordinate systems for teaching. Stored points. Force feedback teaching aid Extra computer for teaching, Stored points, Stored points Two arms Seve.ral coordinate systems, Stored points. Several coordinate systems, Stored points, Manual control Program sequence with conditionals visual servo No Multiple arm control Trajectory control Vibration suppression? Use of sensors Yes No Yes Nominal velocity between taught points Nominal velocity between points. Position error No No No Force, vision Force monitoring None No No Velocity ref. Nominal velocity No No No None yet None Contact, force vision NOTE: Draper PAX is an assembly station built around the prototype model of the Bendix PAX arm, but having a computer system and tooling developed by Draper Laboratory. 610 J . L . NEVINS and D. E. WHITNEY TABLE 4. Bendix PAX COMPARISON O I~ R O B O I ( O N I R O L I.ANGUA(iES Draper PAX SIGMA U-500 Stanford A.I. Lab SIGLA VAL AL M ix Implicit Language Name ? FSYS Explicit/Implicit?* Implicit Explicit Compiler/Interpreter ? Interp. lnterp, lnterp, lnterp. Compiler Features Conditionals. Multi-arm interlocks. Link to vision station. Macros. Link to force monitoring. Conditionals. Loops. Interlocks. Link to force monitoring, Conditionals. Interlocks. Immediate execution of single commands. World model coord, frames. Block structure. Conditionals. Two arm coordination. *Judgment here grossly oversimplified. Locations and orientations can be referred to by name, and displacements caused by the robot's actions and the buildup of parts are kept track of automatically. If this is combined with trajectory planning, then something closer to functional programming results, because the computer will calculate all the missing data. Figure 37 illustrates the complicated trajectory data that would have to be described and stored. Explicit programming can be compared to NC programming in APT, whereas implicit is more like projected CAD/CAM techniques in which the computer's internal model of the part is converted automatically into the data needed to support the NC programmer's description of the sequence of cuts. There is general agreement that implicit programming makes the programmer's job easier, but only if he is a programmer by training. Explicit programming requires less of the programmer and the computer, but it will probably limit the degree of adaptability which can be programmed in and incorporated into the program. Table 4 compares several current systems. errors, very close clearances, and no chamfers or lead-ins. (1) Active (motorized) wrist with sensing this device holds the workpiece, the robot carries it to the work site and the device maneuvers the pieces together. The best known of these is the Hi-T-Hand (Goto, Inoyama and Takesau, 1974). It can assemble parts that have no chamfers, and do so with very light touch. Assembly usually takes 3 to 5s, which is relatively long, and the device is rather fragile and in need of continual adjustment (Fig. 27). (2) Active wrist with estimation theory this is, as yet, a theoretical technique with little laboratory verification. Modern control theory and mathematical models of the assembly operation are combined to yield a closed loop technique for assembling parts (Simunovic, 1979: Nevins and Whitney, 1973). In all of these approaches, the major issues involve how much complexity is enough to handle common assembly tasks and how can these tasks be modeled adequately so that effective devices can be designed. If one assumes the existence of chamfers and enough repeatability Component issues Articulating devices such as robot arms. There are many concerns in this area. They range from kinematic design, number of degrees-of-freedom (in the arm or in the workplace), size, actuators, actuator location (shoulder or at respective joints) and accuracy. A paper by Whitney (Whitney, 1979) goes into great detail in this with Table 5 summarizing the issues and Table 6 listing the various assembly robots presently available. Adaptive devices. We describe here a number of other devices used in assembly. With the exception of the conventional fixed mechanical workhead, all these are recent developments. They are mostly wrist or hand devices and their goal is to help a robot perform assembly under difficult conditions, such as large part-to-part TABLE 5. RESEAR( H ISSUESFOR ROBO] ARMS il) Kinematic design and number of axes. Included in thi~ discussion is the question of, should you have an elbo,a or slide axes like machine tools'? I21 Size two issues here. Namely, how big a reach should an arm have. and why are they so inefficient? Typically it takes a 500 to 1000 kg system consuming 40kW to carry 2kg loads a distance of one meter in three to five seconds. 13) Actuators should they be hydraulic or electric'? (4} Actuator location at tbe shoulder'? should they be located at tile joints or (5) Accuracy- this issue is irrelevant if the arm is used only in a playback or incremental mode. In this case only repeatability and resolution are important. An arm, when coupled with an RCC, requires only enough accuracy to get to the chamfer (1 to 2mm) which is much cruder than present systems, and therefore possibly cheaper. Assembly Research 611 TABLE 6. COMPARISON OF ASSEMBLY ROBOTS Bendix PAX Unimate 6000 Unimate 500 SIGMA Shuttle manipulator Kinematics Cylindrical plus wrist Elbow Elbow XYZ Elbow No. of axes Six Six Five Three-five Six Size 0.5m × 0.5m × 200 ° ~ l . 3 m reach ~lm ~0.5m 3 ~18m Actuators Electric/gears, ball screws Hydraulic/gears, ball screws Electric/gears Electric/ ball screws Electric, gears Repeatability ~0.2m ~0.2 m m ~0.1 m m ~0.1 m m 5cm Weight of moving parts 200 kg ? 900 kg ? 60 kg 50 kg ? 500 kg Payload 2 kg 10 kg? 3 kg ? 32 tons NOTES: (1) (2) (3) (4) (5) (6) PAX was a research effort by Bendix, now abandoned. Unimate 6000 is a cooperative effort between Unimation, Inc., and Ford Motor Co. Unimate 500 is being built to General Motors Co. specifications. S I G M A is marketed by Olivetti, s.P.a. The Shuttle Manipulator is being built by SPAR, Ltd., Toronto, Canada. All data in this table are estimates by the author. in robots, parts, grippers, jigs and feeders so that the chamfers will meet during assembly, then the simple Remote Center Compliance will assemble the parts. When parts have 0.01 mm clearances and no chamfers, the Hi-T-Hand suffices, but often so does simple vibration (Dunne, 1977) so the issue is still open. What does seem clear is that visual imaging is less useful in the 0.01 to 1.0mm range than contact-based techniques (Fig. 44). Further description of this problem can be found under sensing in the next section. 0.01 am 0.1 I am I sm lO am 10 cm I I | THE ROBOT OPERATES IR THIS REGION ,41 IF THE I m I L ARE CHAMFERS, THE REMOTE CENTER O~qpLIANCE OPERATES 1N THIS REGION THE H I - T I - H A N D OPERATES IN THIS REGION EVEN NITH NO CHAMFERS COMPUTER-T¥ CAN AID L ":ASSEMBLY IN THIS 'REGION PROXIMITY SENSORS "OPERATE IN THIS ~EGIOR NEGATIVE CLEARANCE reach EAR&RCE BETWEEN PARTS OR POSITIONING ERROR OF ROBOT, ~K PANTS, Off JIBS ~ I I RANGE OF MOTION REQUIRED I I FIG. 44. Hierarchy of sensors and error absorbers vs linear distance or error. Sensing systems. It is generally agreed that sensing, especially monitoring, is essential to much of automation, and assembly is no exception. A conclusion of Draper research is that quite a lot of assembly can be done without direct sensing of any kind, except monitoring. In other areas of sensing, especially TV vision, there has not yet emerged an adequate balance between cost and effectiveness. An idea of the relative applica.tion areas of passive wrists, active wrists, robot arm inherent accuracy and vision is emerging, however, and this is expressed in Fig. 44. A list of some current sensing techniques, with comments, are listed in Table 7. For the time being, it would appear that force sensing is confined to a monitoring role while vision can be used for gross inspection and as an aid in part feeding, if one does not ask too much (a variety of parts, all jumbled together, is too much for people sometimes). Tools and grippers. An assembly robot will need at least one tool or gripper to do its job. How many more than one is the question. There are several options in use and others have been proposed (see Table 8). The issue is whether each part or operation should have its own tool, the problem being cost and time for tool changing. While both of these can be attacked, the nagging philosophical issue is whether one has a reprogrammable system and is there a large fixed investment in specialized tools? On the other hand, it is doubtful that a truly universal tool can be achieved. Even the human hand needs aid from other tools. 612 J . L . NEVINS and D. E. WHITNEY TABLE 7. SENSING OPTIONS Option Comments Examples Active force sensing to guide assembly Remote center compliance, if applicable, will be much faster. Can handle large errors Hi-Ti-Hand Draper demo, 1972, 1975 Stanford A.I. Lab IBM Force sensing to monitor the progress of the assembly Record history Detect drift Prevent damage Verify operations Abort failed operation Draper bench unit SIGMA Proximity sensor Collision avoidance Jet propulsion Lab TV "vision" Part Orientation Gross inspection Product integrity Feeder/tool operation monitor. Visual servo for endpoint guidance of robot Bendix PAX U. of Rhode Island SRI Stanford A.I. Lab N o v r : The Draper bench unit is a manually operated inserter equipped with a load cell and a remote center compliance. Electronic logic will halt insertion if the forces rise above a set level. This unit was built to aid close clearance assembly of delicate gyroscope components. TAt~t,F. 8. TOOL AND GRIPPER r)PTIONS Options currently in use Comments Examples Universal socket on robot plus tool changing and external tool storage Takes time to change tools. Really universal within range of tools themselves Draper PAX U-6000 Multiple arms or arm at each work station, each a n n with one "'permanent" tool Limited range of tasks suited to intermittent changeover U-6000 Bendix PAX Westinghouse design SIGMA Small tool turret on robot wrist Limited range of tasks. Could be heavy U-6000 degree of specialization. Hopefully, this specialization will occur within a carefully designed general environment (see below l. Software Tools General apparatus Compiler/Editor Debugger The Robot Interface Real time operator system Robol wrist with tool socket Specific items Programs for specific assembly operations, and specific products Specific tools "'Libraries'" of useful programs and tool concepts can be expected to emerge. For example, four types of tools (six tools in all) were needed by Draper Laboratory to assemble the alternator: (1) expanding gripper for small object I.D.; (2) contracting gripper for large object O.D.; (3) contracting gripper for small O.D.: and (4) screwdriver. Part feeder and inspection devices. There are many devices for part feeding and inspection. Some of these were described in previous sections and tables. Singular solutions and devices will be the order of the day until systematic ways of looking at lhe problems and defining the requirements (such as those being worked on by Birk (Kelley, Birk and Wilson, 1977) and Boothroyd (Abraham et al., 1977) are fully developed. CONCLUSIONS We have tried to list here (but make no claim for completeness), the reserve issues in this area. The intent was to show that this is truly a virgin research area. Our small successes, while significant, have mainly helped us to get visibility into the larger issues, and to help focus our research. Other possibilities Parts fed with tool Extends universality of wrist socket, but takes time Truly universal tool Possible'? Product redesign to reduce number and variety of tools needed Not done as often as it could be A similar problem arises in software system design. Should one expect future robots to come preprogrammed for any assembly operation, or is it more reasonable to expect software facilities which can be used to compose programs for specific operations? In both tools and software it appears more reasonable to accept a certain REFERENCES Abraham, R. G. et al. t1977). F'rogrammabie Assembly Research Technology. Transfer to Industry. Final Report, Westinghouse Corporation, Pittsburgh, PA. Agin G. (1977). Servoing with visual feedback. Proc. 7th ISIR, Tokyo, 551 560. D'Auria, A. and M. Salmon (1975). SIGMA An integrated general purpose system for automation manipulation. Proc. 5th International Symposium on Industrial Robots. Chicago. Drake, S. H. (1977). The use of compliance in a robot assembly system. Presented at the IFAC Symposmm on Information and Control Problems in Manufacturing Technology, Tokyo, Oct. 1977, and Using compliance in lieu of sensory feedback for automatic assembly, Sc.D. Thesis, MIT Mechanical Engineering Department, CSDL Report No. T-657, Sept. Dunne, M. J. (1977). An assembly experiment using programmable robot arms. Proc. 7th International Symposium on Industrial Robots, Tokyo, 387 397. Assembly Research Einkel, R. and co-workers (1974). AL--A Programming System for Automation, Report No. AIM-243, STAN-CS74-456, Stanford University, Artificial Intelligence Laboratory. Gibbons, J. B. and J. L. Nevins (1976). Product system productivity, a collaborative study by the General Electric Co. and The Charles Stark Draper Laboratory, Inc. Goto, T., T. Inoyama and K. Takeysau (1974). Precise insert operation by tactile controlled robot "Hi-T-Hand" expert 2. Proc. 4th International Symposium on Industrial Robots, 209-218. Johnston, A. R. (1973). Optimal proximity sensor for manipulators, Jet Propulsion Laboratory, Report No. JPL TM33-612, Pasadena, CA. Kelley, R., J. Birk and L. Wilson (1977). Algorithms to visually acquire workpieces. Proc. 7th ISIR, Tokyo, 497 506. Kellog, C. and co-workers (1978). In Herve Galleries and Jack Minker (Eds.), Deductive Planning and Pathfinding for Relational Data Bases, Logic and Data Bases, 179-200. Plenum Press, New York. Kondoleon, A. S. (1976). Application of technology-economic model of assembly techniques to programmable assembly machine configuration. S.M. Thesis, MIT Mechanical Engineering Department. Lieberman, L. I. and M. A. Wesley (1975). AUTOPASS, An automatic programming system for computer-controlled mechanical assembly. IBM Research Dept. No. RC-5606, T. J. Watson Research Center, Yorktown Heights, New York. Lynch, P. M. (1976). Economic technological modeling and design criteria for programmable assembly machines. CSDL Report No. T-625, Ph.D. Thesis, MIT Mechanical Engineering Department, June 1976. See also, Lynch, P. M. (1977). An Economic Guideline for the Design of Programmable Assembly Machines, ASME Paper No. 77WA/Aut-2. Nevins, J. L. and D. E. Whitney (1973). The force vector 613 assembler concept. Proc. 1st International Conference on Robots and Manipulator Systems, Udine, Italy, September 1963, and published as CSDL Report No. E-2754, March 1973. See also, Whitney, D. E. (1977). Force feedback control of manipulator fine motions. ASME J. Dynamic Systems, Measurement Control U99G (2), 91 97. Nevins, J. L. and co-workers (1976). Exploratory research in industrial modular assembly. 4th Report, CSDL Report No. R-996. Nevins, J. L. and co-workers (1978a). Exploratory research in industrial modular assembly. 5th Report, CSDL Report No. R-1111. See also Nevins, J. L. and D. E. Whitney (1978), Computer controlled assembly. Scient. Amer. 238 (2), 62 73. Nevins, J. L. and co-workers (1978b). Exploratory research in industrial modular assembly. 6th Report, CSDL Report No. R-1218. Rosen, C. and co-workers, Machine intelligence applied to industrial automation. Research Report Series, SRI International, Menlo Park, CA. Simunovic, S. (1979). An information approach to parts mating, Sc.D. Thesis, MIT Mechanical Engineering Department, and published as CSDL Report No. T-690. Watson, P. C. (1976). A multidimensional system analysis of the assembly process as performed by a manipulator. Presented at the 1st North American Robot Conference in Chicago. Whitney, D. E. (1969). State space models of remote manipulation tasks. Trans. IEEE Automatic Control AC-14, 617 623. Whitney, D. E. (1979). Discrete parts assembly automation .... An overview. ASME J. Dynamic Systems, Measurements Control 101 (1), 8- 15. Whitney, D. E. and S. Graves (1979). 7th NSF Grantees' Conference on Production Research and Technology, Cornell University.