Copyright 1994 Elsevier Science Lid Printed in Gr eat Britain. All rights reserved 01105-1098/94 $7.011 + 0.00 Survey Paper A Survey of Models, Analysis Tools and Compensation Methods for the Control of Machines with Friction* B R I AN AR MS T R ONG- HI ~ L OUVR Y, t P I E R R E DUPONT~ t a n d C AR L OS C ANUDAS DE WI T This survey addresses contributions f rom the tribology, lubrication and phys- ics literatures, as well as the controls literature, which are important f or the understanding and compensation of friction in servo machines. Key Words--Friction; friction compensation; friction modeling; identification; adaptive control; control applications; feedback control; feedforward control; modeling. Ala~trKt--While considerable progress has been made in friction compensation, this is, apparently, the first survey on the topic. In particular, it is the first to bring to the attention of the controls community the important contributions from the tribology, lubrication and physics literatures. By uniting these results with those of the controls community, a set of models and tools for friction compensation is provided which will be of value to both research and application engineers. The successful design and analysis of friction compensators depends heavily upon the quality of the friction model used, and the suitability of the analysis technique employed. Consequently, this survey first describes models of machine friction, followed by a discussion of relevant analysis techniques and concludes with a survey of friction compensation methods reported in the literature. An overview of techniques used by practising engineers and a bibliography of 280 papers is included. 1. INTRODUCTION FRICTION IS PRESENT in all machi nes i ncor por at i ng par t s wi t h r el at i ve mot i on. Al t hough fri ct i on may be a desi r abl e pr oper t y, as it is for br akes, i t is gener al l y an i mpedi ment for ser vo cont r ol . The l i t er at ur e r el evant t o fri ct i on and cont r ol is ver y wi del y scat t er ed; i mpor t ant i deas ar e t o be f ound in t he j our nal s of cont r ol s, t r i bol ogy, l ubr i cat i on engi neer i ng, acoust i cs, and gener al engi neer i ng and physi cs. I t is t he ai m of this survey t o synt hesi ze t he cont r i but i ons of sever al * Received in revised form 17 December 1992; received in final form 25 July 1993. The original version of this paper was not presented at any IFAC meeting. This paper was recommended for publication in revised form by Editor K. J. Astr6m. Corresponding author Professor B. Armstrong- H61ouvry. Tel. + 1 414 229 6916; Fax + 1 414 229 6958; e-mail bsra@ee.uwm.edu. t Department of Electrical Engineering and Computer Science, University of Wisconsin, Milwaukee, P.O. 784, Milwaukee, Wl 53201, U.S.A. ~t Department of Aerospace & Mechanical Engineering, Boston University, 110 Cummington Street, Boston, MA 02215, U. S. A. Laboratoire d' Automatique de Grenoble, ENSIEG, B.P. 46 38402 St Martin d' H~res, France. 1083 hundr ed art i cl es f r om t he sever al di sci pl i nes, a n d t he i nput of engi neer s in i ndust r y who have wor ked wi t h fri ct i on and cont r ol , t o pr oduce a gr and pi ct ur e of model s and met hods i mpor t ant for fri ct i on and cont rol . Tr i bol ogy is t he sci ence of r ubbi ng cont act s. The field is act i ve, with 1000 i nvest i gat or s in Nor t h Amer i ca and a l i t er at ur e t hat grows by some 700 art i cl es per year ; and gr eat pr ogr ess has been made t owar ds under st andi ng t he physi cal pr ocesses of sliding machi ne cont act s: bear i ngs, t r ansmi ssi on el ement s, brushes, seal s, et c. For t he cont r ol s engi neer , it is f r i ct i onal dynamics whi ch is of gr eat est i nt er est . One chal l enge of t hi s revi ew has been t o bri ng t oget her f r om t he t r i bol ogy l i t er at ur e an under st andi ng of fri ct i onal dynami cs. Tr i bol ogy is concer ned wi t h fri ct i on; but in r ecent year s t he field has been most concer ned wi t h i ssues of wear and machi ne life on t he one hand, and of surface chemi st ry and physi cs on t he ot her . Dynami cs has not been a focus. St udi es in fri ct i onal dynami cs car r i ed out over t he past five decades ar e br ought t oget her in this survey. I nvest i gat i ons wi t hi n t he field of cont r ol s have not capi t al i zed adequat el y on t he fri ct i on model s avai l abl e from t he exper i ment al and t heor et i cal wor k of t ri bol ogy. Many i nvest i gat i ons have br ought t oget her powerful t ool s f r om st abi l i t y t heor y, nonl i near cont r ol , nonl i near syst em i dent i fi cat i on, adapt i ve cont r ol and ot her ar eas; but t hese i nvest i gat i ons have been based on t he fri ct i on model s of Le ona r do Da Vinci or el ement ar y physics. I t is no wonder t hat consi st ent resul t s have been el usi ve and t hat t he analysis t ool s capabl e of pr edi ct i ng st i ck sl i p and ot her fri ct i onal behavi or ar e not fully r el i abl e. Wi t hi n t r i bol ogy t her e is consi der abl e under st andi ng of t he fri ct i onal dynami cs of l ubr i cat ed met al - on- met al cont act s; and, whi l e per haps somewhat mor e compl ex t han Le ona r do' s s t a t i c + Co u l o mb fri ct i on model , 1084 B. ARMSTRONG-HI~LOUVRY et al. models are available to account for the dynami cs observed in a broad range of t ri bol ogy experiments, some conduct ed with remarkabl e resolution of sensing. In Section 2 of this paper, friction model i ng is addressed. Results from a range of experi ment s report ed in the t ri bol ogy, mechani sm, physics and controls literatures are present ed and assimilated. At the end of Section 2, an i nt egrat ed friction model is presented. In Section 3, analysis tools are present ed for studying servos with friction. Many of t he met hods present ed have been applied in the cont rol s literature, including analytic met hods, the describing function and phase plane analysis; but investigations have also been carried out in the areas of acoustics and mechanics, where frictional instability may be a maj or cont ri but or t o processes of interest. In Section 4, compensat i on met hods for machines with friction are presented. Here the cont rol s literature is the maj or cont ri but or. The br oad classes of compensat i on strategy are probl em avoi dance, non-model -based control and model -based cont rol . Probl em avoi dance deserves special consi derat i on because, as we will see in Section 2, mi nor modification of the lubrication may have a t remendous impact on the frictional instability, and friction modification is not always a priority of the lubrication engineer. Par amet er identification and adapt i ve cont rol strategies are also addressed in Section 4, as is input from engi neers in industry. In Section 5, we concl ude with a program for tackling t he challenging probl ems posed by friction in servo-controiled machines. 2. FRICTION IN MACHINES When met hods of feedback cont rol are applied t o moving bodies, friction is inevitably among t he forces of mot i on. The field of cont rol has long i ncorporat ed sophisticated investigations of ot her cont ri but i ons t o the forces of mot i on, such as mul t i body dynamics, electromagnetics, and aero- or fluid dynamics. But t he forces of mot i on cont ri but ed by friction are oft en studied with simplified model s, similar t o t hose empl oyed by Leonar do Da Vinci. The English language literature of t ri bol ogy grows at a rate of 700 articles per year and represent s a vast, moder n effort to underst and these phenomena. While oft en academically pursued, t ri bol ogy is hardly academically motivated: energy loss due t o friction and t he failure of equi pment due t o wear represent a consi derabl e percentage of every moder n economy. Feedback cont rol is oft en applied t o mechanical arrangement s involving met al -on-met al cont act with grease or oil lubrication. Issues of manufact ure and performance mot i vat e t he choice of metals for working members; and issues of service life mot i vat e the use of fluid lubricants. This study will concent rat e on what t ri bol ogy has t o offer t owards the model i ng of friction in fluid lubricated met al -on-met al junctions. Specialized tribological studies are available which address ot her combi nat i ons of engi neeri ng materials, such as plastics on metal and dry lubricated and electrical contacts. The classic model of fri ct i on--fri ct i on force is proport i onal to load. opposes the mot i on, and is i ndependent of contact ar ea- - was known t o Leonar do da Vinci, but remained hidden in his not ebooks for centuries. Rabinowicz (1965) argues t hat the scientific study of friction must have been subsequent to the elucidation of Newt on' s first law (Newt on, 1687) and the moder n concept i on of force. This is not quite true. Da Vinci' s ideas on the nature of force, of which he knew friction t o be an example, provi de a fascinating insight into probl ems of pre-Newt oni an natural philosophy (Da Vinci, 1519). Da Vinci' s friction model was rediscovered by Amont ons (1699) and devel oped by Coul omb (1785) among others. Amont ons ' claim t hat friction is i ndependent of cont act area (the second of Da Vinci' s laws) originally at t ract ed skepticism, but was soon verified. Morin (1833) i nt roduced the idea of static friction and Reynol ds (1866) the equat i on of viscous fluid flow, completing the friction model t hat is most commonl y used in engineering: the st at i c+ Co u l o mb + viscous friction model (Morin, 1833; Reynol ds, 1886) and shown in Fig. l (b). The science of tribology (Greek for the study of rubbing) was born in Engl and in the 1930s. Basic questions of wear mechanisms, t rue cont act area, relationships between friction, material propert i es and lubricating processes were addressed and answered. It is not possible here to give tribology its due. The interested reader is referred t o Bowden and Tabor (1956, 1973), Suh and Sin (1981), Czichos (1978), which provi de excellent and readabl e introductions t o the field. Dowson (1979) is an engaging work which illuminates the 3000 year history of man' s at t empt s t o underst and and modify friction. Hamr ock (1986) is a brief handbook survey of the relevant met hods of t ri bol ogy; and Hailing (1975) provides a survey that is rigorous but not overly detailed and sufficiently sweeping t o address such issues as friction induced instability and solid lubrication. Ludema (1988) is an interesting critique of tribology and cultural barriers t o interdisciplinary pursuits; and Rabinowicz (1978), a discussion of priorities for tribology. 2.1. The Tri bol ogy of Machine Friction The maj ori t y of servo-controlled machines, of the eart h-bound variety at least, are lubricated with oil or grease. Tribologically, greases and oils have mor e in common than not. Grease is essentially a soap matrix that carries oil, which is released under stress into load bearing junctions. These lubricants are widely used because t hey provide a fluid barrier between rubbing metal parts that exchanges dry friction for viscous friction and vastly reduces wear. The fluid barrier can be maintained by forcing lubricant under pressure into the load bearing interface, a t echni que called hydrostatic lubrication. This, however, entails great mechanical complexity and is not applicable to many bearing or transmission designs. The more common t echni que is t hat of hydrodynami c lubrica- tion, wherein the lubricant is drawn into the interface by the mot i on of the parts. Hydr odynami c lubrication is simple to implement, requiring only a bath of oil or A survey of friction and control s 1085 (a) Level of Friction Force Coulomb Friction 'Slope Due to Viscous Friction Velocity Friction Force Extra Friction at Zero Velocity, Static Friction (b) - - Velocit~ ( c ) Friction Force Negative Viscous Friction (The Stribeck Effect) Velocit~ FIG. 1. Friction models: (a): Coulomb+viscous friction model; (b): st at i c+Coul omb+ viscous friction model; (c): negative viscous + Coulomb + viscous friction model (Stribeck friction). gr ease or per haps a fluid spr ay, but suffers t he l i mi t at i on t hat t he fluid film is mai nt ai ned onl y above some mi ni mum vel oci t y. Bel ow t he mi ni mum vel oci t y sol i d- t o- sol i d cont act occurs. 2.1.1. The topography of contact To under s t and t he t r i bol ogy of engi neer i ng surfaces it is necessar y t o consi der t he surface t opogr aphy. Ear l y model s of fri ct i on f ai l ed because t he surface t opogr aphy was mi sunder st ood. The i nt er act i ons at cont act i ng surfaces will be exami ned by consi der i ng pr ogr essi vel y smal l er cont act s. I n Fi g. 2 a conf or mal cont act is shown schemat i cal l y; par t A rest s on par t B. Ki nemat i cal l y, such cont act s ar e i dent i fi ed as ar ea cont act s: t he appar ent ar ea of t he cont act is de t e r mi ne d by t he size of t he par t s. Par t s t hat do not enj oy a mat chi ng radi i of cur vat ur e meet at nonconf or mal cont act , as shown in Fig. 3. These cont act s ar e cal l ed poi nt or l i ne cont act s when consi der ed ki nemat i cai l y; but t hi s is an i deal i zat i on. I n fact t he par t s def or m t o cr eat e an appar ent ar ea of cont act , an ar ea t hat i ncr eases wi t h Conformal Contact , , / / / / i / / Stress Propoaional to Force / Area Part B ~ Ma c r o s c o p i c Contact Area "1 / / / / / / Proportional to Dimensions of Part Fro. 2. Conformal contact, such as machine guide ways or journal bearings. i ncreasi ng l oad. The one mi l l i met er cont act wi dt h suggest ed in Fi g. 3 is t ypi cal of smal l machi ne par t s, such as t he t ransmi ssi on gear s of an i ndust r i al r obot . Tr i bol ogy as a field is sophi st i cat ed in t he use of si mi l i t ude. One wi del y used t r ansf or mat i on maps a nonconf or mal cont act of t wo r adi i t o one of a flat surface and a single cur ved par t , as suggest ed in Fi g. 3 ( Dowson and Hi ggi nson, 1966; Ha mr oc k, 1986). Thi s t r ansf or mat i on gr eat l y simplifies t he st udy of noncon- f or mal cont act s. Nonconf or mal cont act s ari se f r e- quent l y in machi ner y and may be r ef er r ed t o as Her t zi an cont act s, af t er t he ori gi nal anal ysi s ( Her t z, 1881). The st resses f ound in conf or mal cont act s bet ween st eel par t s ar e r ar el y hi gher t han 7 MP a (7 MPa = 1000 psi ), wher eas i n nonconf or mal cont act t he peak st ress can be 100 t i mes gr eat er ( Hamr ock, 1986). A st ress of 700 MPa cor r es ponds t o 100,000 psi , which is gr eat er t han t he yi el d st r engt h of many t ypes of st eel . Thi s is possi bl e in Her t zi an cont act because t he stress is compr essi ve. I n a BBC r adi o pr ogr am, t r i bol ogy pi oneer F. P. Bowden obs er ved t hat " put t i ng t wo sol i ds t oget her is r at her l i ke t ur ni ng Swi t zer l and upsi de down and st andi ng i t on Au s t r i a - - t h e ar ea of i nt i mat e cont act will be smal l " ( Bowden, 1950). Cr yst al l i ne surfaces, even appar ent l y smoot h surfaces, ar e mi cr oscopi cal l y rough. The pr ot uber ant f eat ur es ar e cal l ed asper i t i es and, as shown schemat i cal l y in Fi g. 4, t he t r ue cont act occurs at poi nt s wher e asper i t i es come t oget her . I n this way, t he t r ue ar ea is much smal l er t han t he appar ent ar ea of t he cont act ( Bowden and Tabor , 1939). Over a br oad r ange of engi neer i ng mat er i al s, t he asper i t i es will have sl opes r angi ng f r om 0 t o 25 degr ees and concent r at ed in t he band f r om 5 t o 10 degr ees ( Dowson, 1979). 1086 B. ARMSTRONG-HI~LOUVRY et al. Nonconforrnal Contact Ideal: Point or Line , , Contact, Zero Area S j .... Stress / ~' ~ 10 -3 m, Typical, Steel ~ / / / / / / Part B ! Macroscopic Contact Area Proportional to Load and Material Strength FIG. 3. Nonconformal contact, such as a gear tooth mating or roller bearings. When asperities come into cont act , the local loading will be det ermi ned by t he strength of the materials. The asperities deform t o generat e the cont act area necessary t o t ake up the total load. As a first approxi mat i on, we may consi der the local stress at an asperity j unct i on t o be in proport i on t o t he yield strength of the material. The cont act area, on the ot her hand, is in direct pr opor t i on t o t he total load. As a rule of t humb, t he t rue cont act area, A, is given by A = W/ 3Y, where W is t he load and Y is t he yield strength of the material. Cont act stress at the asperity is t aken, by this rule of t humb, t o be t hree times the yield strength. As with t he nonconf or mal cont act , stress great er t han yield strength is possible because the asperities are under compression. Friction is proport i onal t o t he shear strength of t he asperity junctions. As t he load grows, the j unct i on area grows; but, t o first-order, the shear strength (measured per unit area) remai ns const ant . In this way, friction is proport i onal t o load. If truly clean metal surfaces are brought i nt o cont act , t he shear strength of the j unct i on (friction) can be as great as the shear strength of the bulk material, and t he friction coefficient can be much great er t han one (Bowden and Tabor, 1973; Hamr ock, 1986). Fort un- ately for the operat i on of machines, truly clean surfaces are all but impossible t o achieve. Even in t he Tr ue Cont act Bet ween Engi neeri ng Surfaces Surface Film (Boundary Layer) ~ (10-7m Typical) True \ ~ ~ b Cont a c t Site Junction Wi d t h ~ s p e r i t i e s 1o 5m, Typical, Steel FIG. 4. Part-to-part contact occurs at asperities, the small surface features. absence of lubricants, oxide films will form on the surface of steel and ot her engineering materials, producing a boundary layer. In t he presence of lubricants, additives t o the bulk oil react with t he surface t o form t he boundary layer. The boundar y layer additives are formul at ed t o cont rol the friction and wear of the surface. The boundar y layer is a solid, but because it has the lower shear strength, most shearing occurs in this film. If the boundar y layer has a low shear strength, friction will be low; if it has good adhesion t o t he surface and can be replenished from the oil, wear will be reduced. Boundar y layer thickness varies from a few at omi c thicknesses t o a fraction of a micron. As suggested in Fig. 4, a tenth of a micron is a typical thickness of the boundar y layer formed by the lubricity additives of industrial oil (Wills, 1980; Booser, 1984). Not e t hat this is perhaps two orders of magni t ude less t han t he typical dimension of an asperity in steel junctions. The boundary layer is exactly that, and does not markedl y influence the area or local stresses of cont act . 2.1.2. Friction as a function of velocity: four dynamic regimes There are four regimes of lubrication in a system with grease or oil: static friction, boundar y lubrication, partial fluid lubrication and full fluid lubrication. These four regimes each cont ri but e t o the dynami c t hat a cont rol l er confront s as the machine accelerates away from zero velocity. Figure 5 is known as the Stribeck curve and shows the three moving regimes (Stribeck, 1902; Biel, 1920; Czichos, 1978). The interesting characteristics of regime I, static friction, are not dependent on velocity. 2.1.2.1. The first regime: static friction and preslid- ing displacement. In Fig. 4, cont act is shown t o occur at asperity junctions. From the st andpoi nt of cont rol , these j unct i ons have t wo i mport ant behaviors: t hey deform elastically, giving rise t o presliding displace- ment ; and bot h the boundar y film and the asperities def or m plastically, giving rise t o rising static friction, discussed in Section 2.1.4 below. A survey of friction and controls 1087 Re g i me I. No Sl i di ng, ~ S - El ast i c De f o r ma t i o n , ~ __ ~.- Sl i di ng Ve l oc i t y b FIG. 5. The generalized Stribeck curve, showing friction as a function of velocity for low velocities. Part A Part B / / / / / / ~ ~ ~ , ~ - - - - Idealized Asperity Junctions / / / / / / FIG. 6 Idealized contact between engineering surfaces in static friction. Asperity contacts behave like springs. I t is o f t e n a s s u me d whe n s t udyi ng f r i ct i on t ha t t h e r e is no mo t i o n whi l e in st at i c f r i ct i on, whi ch is t o s ay no mo t i o n wi t hout sl i di ng; but in me c ha ni c s it is wel l known t ha t c ont a c t s a r e c ompl i a nt in bot h t he n o r ma l a nd t angent i al di r ect i ons , e. g. J o h n s o n (1987). Da h l (1968, 1976, 1977), s t udyi ng e x p e r i me n t a l o b s e r v a - t i ons o f f r i ct i on in smal l r ot a t i ons o f bal l be a r i ngs , c onc l ude d t ha t f o r smal l mot i ons , a j unc t i on in st at i c f r i ct i on be ha ve s l i ke a s pr i ng a nd c ons i de r e d t he i mpl i cat i ons f o r cont r ol Th e r e is a di s pl a c e me nt ( pr esl i di ng di s pl a c e me nt ) whi ch is an a p p r o x i ma t e l y l i near f unct i on o f t he a ppl i e d f or c e , up t o a cri t i cal f or ce, at whi ch b r e a k a wa y occur s. Th e el ast i ci t y o f as per i t i es is s ugge s t e d s chemat i cal l y i n Fi g. 6. Wh e n f or ces a r e a ppl i e d, t he as per i t i es will d e f o r m, as s ugges t ed by Fi g. 7, but r e c o v e r whe n t he f or c e is r e mo v e d , as doe s a spr i ng. I n t hi s r e gi me , t he t angent i al f or c e is g o v e r n e d by: F , ( x ) = - k , x , ( 1 ) whe r e F, is t he t angent i al f or c e , k, is t he t angent i al For c e /> Br e a k - Awa y Fr i ct i on P A / / / / / / Break-Away ~ ~ ~ B / / / / / / FIG. 8. At breakaway true sliding begins. st i ffness of t he c ont a c t a nd x is d i s p l a c e me n t a wa y f r om t he equi l i br i um pos i t i on. F, a nd x r e f e r t o t he f or ce and di s pl a c e me nt in t he c ont a c t b e f o r e sl i di ng begi ns, as i ndi cat ed in Fi gs 7 a nd 8. Wh e n t he a ppl i e d f or ce exceeds t he r e qui r e d b r e a k a wa y f or c e , t he j unct i ons b r e a k (in t he b o u n d a r y l ayer , i f pr e s e nt ) a nd t r ue sliding begi ns, as s ugges t ed in Fi g. 8. Po l y c a r p o u and Soom (1992) ha ve p o i n t e d out t ha t st at i c f r i ct i on is not t r ul y a f or ce of f r i ct i on, as it is ne i t he r di ssi pat i ve nor a c ons e que nc e o f sl i di ng; but is a f or c e o f cons t r ai nt , and e mp l o y t he t e r m t a nge nt i a l f or ce. Thi s issue is i mp o r t a n t f or bot h s i mul at i on a nd anal ysi s. Th e t angent i al st i ffness, k,, is a f unct i on o f a s pe r i t y g e o me t r y , ma t e r i a l el ast i ci t y and a ppl i e d n o r ma l f or c e ( J ohns on, 1987) No t e t ha t t he t angent i al st i ffness due t o pr esl i di ng di s pl a c e me nt is qui t e di f f er ent f r o m ( a nd ma y be s ubs t ant i al l y less t ha n) t he st i ffness of t he me c ha ni s m itself Th e as per i t i es , not t he me c h a n i s m c o mp o n e n t s , a r e def or mi ng Wh e n n o r ma l f or c e is changi ng, t he b e h a v i o r ma y be qui t e c ompl e x, becaus e n o r ma l f or ce, n o r ma l st i ffness a nd t angent i al st i ffness a r e nonl i near , i nt e r a c t i ng f unct i ons o f n o r ma l di s pl a c e me nt ( Ma r t i ns e t a l . , 1990). To first a ppr oxi ma t i on, it is act ual l y t he b r e a k a wa y di s pl ace- me nt t hat is c ons t a nt ; a nd t he st i ffness is t he n gi ven by: F~ k, = - - , (2) Xb whe r e F~ is t he b r e a k a wa y f or c e a nd xh is t he ma x i mu m d e f o r ma t i o n of t he as per i t i es b e f o r e br e a ka wa y. I f n o r ma l f or c e is va r yi ng a nd t he coeffi ci ent of st at i c f r i ct i on is a p p r o x i ma t e l y c ons t a nt , t hen k, b e c o me s p r o p o r t i o n a l t o n o r ma l f or ce. Th e b r e a k a wa y di s pl a c e me nt ma y be mi nut e in engi neer i ng ma t e r i a l s , b r e a k a wa y is o b s e r v e d t o oc c ur wi t h def l ect i ons on t he o r d e r o f 2 - 5 mi c r ons i n st eel j unct i ons ( Rabi nowi cz, 1951; Da hl , 1968; Bu r d e k i n e t For ce < Br e a k - Awa y Fr i ct i on q Displacement is ~oportional to Force F- / / / / / / / / / / / / For c e < Br e a k - Awa y Fr i ct i on P Displacement is Proportional to Force 5 x 10 6m, Typical Maximum P . A I'- / / / / / / P . B / / / / / / FIG. 7. Asperity deformation under applied force, presliding displacement. 1088 B. ARMSTRONG-Hi~LOUVRY et al. al. , 1978; Cheng and Kikuchi, 1985; Villanueva-Leal and Hi nduj a, 1984; Armst rong-HEIouvry, 1991). But elsewhere in a mechani sm a much great er displace- ment may be observed, di spl acement significant on the scale of feedback control. This will arise, for example, in robot s, where the arm itself acts as a lever t o multiply micron mot i ons at the gear teeth to millimeter mot i ons of t he out put ( Ar mst r ong- Hr l ouvr y, 1991). Presliding displacement has long been studied in the mechanics communi t y, and is somet i mes t er med micro-slip (Johnson, 1987). The transition from elastic cont act t o sliding is not simple. Sliding is observed t o originate first at the boundar y of a cont act and t o propagat e t oward the cent er (Johnson, 1962). Thus there is no abrupt transition t o sliding. Presliding displacement is of interest t o t he cont rol s communi t y in ext remel y high precision pointing applications (Dahi 1977; Wal rat h, 1984) in dynami cs (Canudas de Wit et al. , 1993) and in simulation (Haessig and Friedland, 1991); and may also be i mport ant in establishing t hat t here are no discontinuities in friction as a function of time. 2.1.2.2. The s econd regime: boundar y lubrication. In the second r egi me- - t hat of very low velocity sliding--fluid lubrication is not i mport ant , the velocity is not adequat e t o build a fluid film between the surfaces, e.g. Fuller (1984). As described, the boundar y layer serves t o provi de lubrication. It must be solid so t hat it will be mai nt ai ned under the cont act stress, but of low shear strength t o reduce friction (Bowden and Tabor , 1973). In Fig. 9 sliding in boundar y lubrication is shown. Because there is solid-to-solid cont act , t here is shearing in the boundar y lubricant. Because boundar y lubrication is a process of shear in a solid, it is oft en assumed t hat friction in boundar y lubrication is higher t han for fluid lubrication, regimes t hree and four. This, however, is not always the case; it is not necessary t hat the shear st rengt h of a solid be great er t han the viscous forces of a fluid. Consi der t hat glass is a fluid, with a viscosity great enough t hat centuries are requi red for it t o flow t o the bot t om of the wi ndow frame. Many solids will yield t o a lower shear force than the forces of viscous / f l o w in this fluid. Cer t ai n boundar y lubricants do reduce static friction to a level below Coul omb friction and entirely eliminate stick-slip. Some aspects of these and ot her boundar y lubricants are described in Section 2.1.3 below. 2.1.2.3. The t hi rd regime: part i al f l ui d lubrication. Shown in Fig. l 0 is the process by which lubricant is Sliding Or Rolling ~% Motion P i : i i : i : i : i : i : i if21 i i : i i : i i : i l i : i i i ' i i i i I i ~ i ~ i i ~ Lubricant, ' - Lubricant. Extruded by Pressure Entrained by Motion FIG. 10. Motion brings fluid lubricant into the contact zone. drawn into the cont act zone. Lubri cant is brought into the load bearing region t hrough mot i on, either by sliding or rolling. Some is expelled by pressure arising from the load, but viscosity prevent s all of t he lubricant from escaping and thus a film is formed. The ent rai nment process is domi nat ed by the interaction of lubricant viscosity, mot i on speed and cont act ge- omet ry. The great er viscosity or mot i on velocity, the thicker the fluid film will be. When the film is not thicker than the height of the asperities, some solid-to-solid cont act will result and there will be partial fluid lubrication. When t he film is sufficiently thick, separation is compl et e and the load is fully support ed by fluid. Partial fluid lubrication is shown schematically in Fig. 11. The dynami cs of partial fluid lubrication can perhaps be underst ood by anal ogy with a water skier. At zero velocity the skier is support ed buoyant l y in the water. Above some critical velocity the skier will be support ed dynamically by his mot i on. Bet ween floating and skiing there is a range of velocities wherein the skier is partially hydrodynami cal l y supported. These velocities are anal ogous t o the regime of partial fluid lubrication. The anal ogy is imperfect in t hat the buoyant support is not like solid-to-solid cont act ; and the dynami c support of the skier is due t o fluid inertia as opposed t o viscosity, t he domi nant force in lubrication. In one aspect, however, the anal ogy is valid: for bot h t he wat er skier and the machine, the regime of partial dynami c support is manifestly unstable. As the skier is elevated by his increased velocity, his drag is reduced, allowing him t o go even faster. As partial fluid lubrication increases, solid-to-solid cont act decreases, reducing friction and increasing the acceleration of the moving part. Partial fluid lubrication is the most difficult t o model of the four regimes. In t he case of nonconformal cont act , even full fluid lubrication (Elasto- Hydr odynami c Lubri cat i on, or EHL) must be Sliding - Boundary Layer Part B K~ Shearing Takes Place in the Softer Boundary Layer, Boundary Layer Strength Determines Friction FIG. 9. Boundary lubrication, regime lI of the Stribeck curve. Partial Support by Mot i on Fluid Lubricant - ~ .~ / - ~ , O l l d t o S o l i d C o n t a c t FIG. 11. Partial fluid lubrication, regime III of the Stribeck c u r v e . A s ur vey of fri ct i on and cont r ol s 1089 investigated numerically. For these contacts, steady state flows over smooth surfaces are well understood (Dowson and Higginson, 1966; Booser, 1984; Pan and Hamrock, 1989); but these are not the true conditions of partial fluid lubrication. Work is proceeding toward an understanding of the interaction of surface roughness and EHL in steady state motion (Zhu and Cheng, 1988; Sadeghi and Sui, 1989). From these papers it appears that the details of surface roughness, asperity size and orientation, have significant impact on the lubricant film characteristics, complicating a general analysis. Of principal interest to the controls engineer is the dynamics of partial fluid lubrication with changing velocity. Theoretical study of this problem is beginning (Sroda, 1988; Rayiko and Dmytrychenko, 1988). These numerical investigations show a time lag between a change in the velocity or load conditions and the change in friction to its new steady state level. This time or phase lag is called frictional memory and has been observed experimentally in a wide range of circumstances (Rabinowicz, 1958; Bell and Burdekin, 1969; Rice and Ruina, 1983; Wairath, 1984; Hess and Soom, 1990; Armst rong-Hrl ouvry, 1991; Polycarpou and Soom, 1992; Dupont and Dunlap, 1993). The observed delay may be on the order of milliseconds to seconds, and its impact on stick-slip motion may be substantial (Rice and Ruina, 1983; Dupont, 1994; Dupont and Dunlap, 1993; Armst rong-Hrl ouvry, 1991, 1992, 1993). Continuing the analogy of the water skier, frictional memory is a consequence of state in the frictional contact, just as the height of the skier is a state variable that does not come to its new equilibrium instantly. Indeed, new work in triboiogy suggests that frictional memory in fact arises from the normal separation in the frictional interface (see Section 2.3). 2.1.2.4. The fourth regime: full fluid lubrication. Hydrodynamic or elasto-hydrodynamic. Hydrodynamic and elasto-hydrodynamic lubrication (EHL) are two forms of full fluid lubrication. Hydrodynamic lubrication arises in conformal con- tacts, and EHL in nonconformai contacts. As Fig. 12 shows, solid-to-solid contact is eliminated. In this regime, wear is reduced by orders of magnitude and friction is well behaved. The object of lubrication engineering is often to maintain full fluid lubrication effectively and at low cost. Reynolds (1886) and Sommerfeld (1904) laid the ground work for the investigation of hydrodynamic lubrication, which has Full Support by Motion Fluid Lubricant ~ , , ' . ' . ' . ' . ' . . . . . - . - . - . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FIG. 12. Full fluid lubrication, regime IV of the Stribeck c u r v e . . ~ 1 . 0 - - ca 2 0 . 1 - - 0.01- t a 0 . 0 0 1 - U n l u b r i e a t e d B o u n d a r y E l a s t o - H y d r o d y n a m i c L u b r i c a t i o n C o n d i t i o n Fro. 13. The range of friction levels [Adapted from Bowden and Tabor 0973)]. been worked out in great detail (see, for example Hersey (1914, 1966), Hailing (1975)). EHL is common in servo-controlled machines. As mentioned, it is studied numerically: there is no analytic solution simultaneously satisfying the surface deformation and fluid flow equations. Generally speaking, EHL will give higher friction and wear than hydrodynamic lubrication, as suggested by Fig. 13. General predictive models of the steady state lubricant film thickness are available, e.g. Hailing (1975), Hamrock (1986). The film thickness, which determines friction as well as protection from wear, is a function of surface rigidity and geometry, lubricant viscosity and velocity. For control, the value of these results will lie in predicting the velocity of transition to full fluid lubrication. Work is beginning in the exploration of the transient dynamics of elasto- hydrodynamic lubrication (Xiaolan and Haiqing, 1987; Harnoy and Friedland, 1994). 2.1.3. Boundary lubricants, a domain of many choices Boundary lubrication is important to the controls engineer because of the role it plays in stick slip. The key to effective boundary lubrication is the discovery of a molecule that binds with reasonable strength to the metal surface, but is not corrosive; that has sufficient strength to withstand the forces of sliding and yet has a low shear strength to give low friction. Such molecules are added to the bulk lubricant, often comprising only a per cent or two of the total. Lubrication additives may be divided into three broad classes: lubricity agents; extreme pressure agents; and anti-wear agents. Long chain hydrocarbons with a polar group at one end are commonly used as lubricity agents. The polar group bonds to the metal and the long chain sticks away from the surface, creating, in effect, a mat of bristles (Merchant, 1946; Bowden and Tabor, 1973; Fuller, 1984); the longer the chain (bristle) the lower the friction. These additives are sometimes called oiliness agents, anti-friction agents or friction modifiers. Friction modification refers to reducing the static friction and friction in boundary lubrication. The polar hydrocarbons attach themselves to the metal surface by charge exchange in a process called ' physi-adsorption' . Their application is limited to situations of moderate temperature. At approximately 100C the polar hydrocarbons desorb and boundary 1090 B. ARMSTRONG-HI~LOUVRY et al. lubrication is lost (Bowden and Tabor, 1973; Fuller, 1984). For this reason the use of long chain hydrocarbons is restricted t o applications t hat generat e little frictional heating, which is generally a restriction to conformal contacts. Use of these pol ar hydrocarbons as friction modifiers is wide spread in the form of ' way oils' , oils specially formul at ed to eliminate stick slip in machine slideways (Merchant , 1946; Wolf, 1965; Mobil, 1978). Machine slideways are conformal , and thus less affected by frictional heating. A premi um is placed on eliminating st i ck-sl i p in precision machine tools and great at t ent i on has been give to the probl em (Merchant , 1946; Wolf, 1965; Bell and Burdekin, 1966, 1969; Kat o et al., 1972, 1974). The level of static friction can, in fact, be reduced below the level of Coul omb friction so t hat there is no destabilizing negative viscous friction and stick-slip is eliminated (Merchant , 1946; Wolf, 1965; Mobil, 1978; Wills, 1980). There are st andard procedures for measuring the lubricity of way oils, one is the Cincinnati Milacron stick-slip test (Cincinnati Milacron, 1986). This test procedure measures the friction at breakaway and at a velocity of 0.5 inches per minute. The Cincinnati Milacron test procedure is quite similar t o t hat described in Wol f (1965). The test manual indicates that when ( E/ Fc <0 . 8 5 ) , stick-slip will be eliminated. ELF,. values as low as 0.55 are observed (Millman, 1990). The possibility of wider use of these lubricants in servo machi nery is an intriguing one . Ext reme pressure (EP) agents chemically react with the metal t o form a film t hat will prot ect the surface from wear. The principal issue in their formul at i on has been the reduct i on of wear and seizure (Papay, 1974, 1988; Papay and Di nsmore, 1976), but most EP additives also provi de a degree of friction modification and some will pass st andard stick-slip tests (Facchi ano and Vinci, 1984; Lubrizol, 1988; Cincinnati Milacron, 1986). EP agents are available in a vast variety, and are universally present in gear and ot her machine lubricants and thus in many servo-cont rol l ed mach- ines. EP agents bond with the surface by chemically reacting with the metal, or ' chemi -adsorpt i on' . For this reason they tend t o be metal specific. EP additives function at higher t emperat ures t han do lubricity additives and so are serviceable under more severe loading, such as in nonconformal contacts. The chemi-adsorption also offers a generally st ronger bond to the surface and thus great er prot ect i on against wear. The principal limitations of EP additives are a weaker friction modification than is achieved by the lubricity agents, and chemical reaction with the surface, which is by its nat ure corrosive. With EP agents one in effect acquires greatly reduced mechanical wear at the price of slow corrosi on (Wills, 1980; Fuller, 1984; Papay, 1988). Ant i -wear agents ext end the service life of machine parts t hrough a remarkabl e chemistry t hat can repair some forms of wear i nduced surface damage (Estler, 1980; Booser, 1984). The issue of principal concern for the controls engi neer is t hat the anti-wear agent can interfere with the friction modification of lubricity or EP additives. Lubri cant additives also perform a host of ot her functions, including viscosity modifica- tion, foam cont rol , corrosion prot ect i on, and oxida- tion stabilization (Papay, 1988). These functions are key to machine and lubricant life, but do not bear directly on mechani sm dynamics or control. Lubricant additives must stay in suspension or solution in the bulk lubricant. In this way t hey are available to replenish sites on the surface where t he lubricant film is damaged by rubbing (Bowden and Tabor, 1973; Hamr ock, 1986). Repl eni shment of the boundary layer from the bulk lubricant may be required after each pass (Vi nogradov et al., 1967; Camer on, 1984). Gitis (1986a,b) has studied the relationship bet ween the rates of attrition and repl eni shment of boundary lubricants and the impact on stick-slip. Boundar y lubricants are st andard additives in machine grease or oil; there is a great range of formulations, and t hey typically constitute less t han 2% of the total. Systems with high loading and low relative velocity, such as gear teeth, may operat e entirely in boundary lubrication (Mobil, 1971; Wellauer and Hol l oway, 1976; Wilson, 1979). Much of the attention in boundar y lubricant formul at i on has been focused on reduct i on of wear. In the design of lubricants, ot her t han way lubricants, friction modification has played a secondary role. Dry lubricants, such as Teflon , operat e by a variety of mechanisms. Thei r principal liability is the loss of the protection against wear provi ded by full fluid lubrication, A good survey of dry lubrication may be found in either Hailing (1975) or Fuller (1984). Gassenfeit and Soom (1988) exami ne and contrast breakaway friction in dry and lubricated contacts. Pope et al, (1989) address many issues of concern in space applications. From a cont rol perspective, dry lubricants may offer the substantial advantage of eliminating destabilizing partial fluid lubrication, al t hough negative viscous friction and stick slip may still be present (Martins et al., 1990). 2. 1. 4. Rel axat i on oscillations Stick slip was apparent in early studies of low speed mot i on. The first at t empt s at explanation were carried out within the static plus Coul omb friction model of Fig. l (a) (Thomas, 1930). Using a sensitive displace- ment measuri ng apparat us, phot omi crographs of the rubbi ng surfaces and hydraulically produced steady mot i on, Bowden and Leben (1939) demonst rat ed that sticking occurs and coined the term stick-slip (Rabi nowi cz, 1956b). They observed welding in t he phot omi cr ogr aphs and, using the t hermocoupl e effect bet ween dissimilar metals, t hey found wide t empera- t ure fluctuations that are correlated with the stick cycle. Br owden and Leben posited local melting of one rubbi ng metal as a mechanism for decreased friction during sliding. They found t hat a similar stick-slip occurs in many lubricated systems, even if t here is no welding; and t hat no stick-slip occurs when long chain fatty acids are used as a lubricant. At the time boundar y lubricants were not well under- A survey of friction and control s 1091 stood. The fatty acids used by Bowden and Leben (1939) are now commonly used as lubricity agents. In 1940 experiments had not yet been conducted which could observe the details of friction during a stick-slip cycle, but it became evident from macro- scopic observations, in particular the range of speeds and structural conditions over which stick-slip will occur, that the static plus Coulomb friction model w a s inadequate to explain the observed phenomena. Dudley and Swift (1949) employed phase plane analysis to study the possible oscillations in slider mechanisms, that is mass-spring-damper systems equivalent to PD control. A negative viscous friction, as shown in Fig. 1(c), was posited and efforts were directed at elucidating its character by fitting predicted oscillations to observed stick-slip (Dudley and Swift, 1949). Experiments grew progressively more sensitive (Sampson e t a l . , 1943; Dokos, 1946; Rabinowicz, 1951, 1956, 1958; Rabinowicz and Tabor, 1951; Rabinowicz e t a l . , 1955) and evidence mounted both for negative viscous friction, Fig. 1(c), and indicating that changes in friction do not coincide exactly with changes of mechanism state. That is to say that dynamics were found to exist within the surface processes that determine friction. Using experiments designed to directly determine the properties o f breakaway (the transition from static to Coulomb friction), Rabinowicz (1951) found that breakaway is not instantaneous, and proposed a model involving translational distance to account for decreasing friction as motion progressed. Rabinowicz (1958) reports an experiment capable of measuring the acceleration of a slider during stick-slip, and observes that the acceleration and deceleration curves are n o t symmetric. Rabinowicz (1958) is a landmark paper because the two temporal phenomena in the stick-slip process are integrated into a friction model that will at least qualitatively predict the range of speeds and structural conditions over which stick-slip will occur. The temporal phenomena are: (1) a connection between the time a junction spends in the stuck condition, i.e. dwell time, and the level of static friction (rising static friction); and (2) a time delay or phase lag between a change in velocity and the corresponding change in friction (frictional memory). 2.1.4.1. R i s i n g s t at i c f r i c t i o n a n d e x t i n g u i s h i n g s t i c k s l i p b y i n c r e a s i n g v e l o c i t y . To understand the role played by rising static friction and frictional memory, it is necessary to consider the stages of a stick-slip cycle; this discussion and Figs 14-16 follow Rabinow- icz (1958). In Fig. 14 a pin-on-flat friction machine is sketched. Here the pin is held in place by a spring and the flat moves at a constant velocity. The mechanism is analogous to a servo machine moving with a desired velocity, k~, a proportional control gain, kp, and damping, k,,. The discussion assumes moderate values of damping; extremely large values of damping will influence the qualitative behavior, but moderate values will not (Bell and Burdekin, 1969). Under some conditions, a system such as that of Fig. 14 will exhibit stick-slip. The spring force (control Motion of S t i c k - S l i p Cycle Se ns e F , -~ -- ~ Spring.Force _~or ma l F o r c e k p Dampe, [ ~_ ? v ~ 1 Pin j Flat I : , , , , , , . . . . . . I ~ X d Flo. 14. Pin on fiat friction machine, schematic; fiat slides under-pin. S t a l k F r k t l e a D e p e a d mt UIMm Rat e Time . g gh F s , o o b - A Stick d Slip f \ \ _ / S~T~. v~. T~=, V~ d TI T2 Time A Velocity I n c r e a s e d FIo. 15. Spring force profile during sti ck-sl i p motion at t wo velocities; spring force decreases whe n velocity increases. =i i_ _ d b Fstatic F coulomb Dwell Time FIO. 16. Static friction ( breakaway force) as a f unct i on o f dwel l t i me, schemati c; wi t h sti ck-sl i p cycl e s hown. Dwe l l t i me is the t i me in static fri cti on, s hown as T 2 in Fig. 15. 1092 B. ARMSTRONG-Hi~LOUVRY e t a l . action) observed during mot i on is sket ched in Fig. 15. Duri ng the stuck, interval, interval a - b , the force rises at a rate /~wr~,~ = kr- f a. At point b the force reaches Fs~, the level of static friction when the system has been at rest for considerable time, and slip begins. During interval b - c slip occurs; t he exact mot i on is governed by the mass spring dynami cs plus the details of the friction forces. A rapid transit is qualitatively indicated here. At point c, the pin is arrested on the fiat and the spring force again begins to rise at rat e i~spri,g = k p Xd , ent eri ng a stable limit cycle of points c - d - e . Point d is somewhat lower than point b because the system has onl y been at rest for dwell time c- d. At point g the velocity -fa is increased. The i mport ant empirical fact is that as the velocity is increased, the size of the limit cycle, i - j - k , diminishes (Dokos, 1946; Rabinowicz, 1958; Kat o e t a l . , 1972, 1974). I f the condi t i on at point j were identical t o t he condition at point d, a decrease in the slip distance would not be observed, and an analysis based on the static plus Coul omb friction model will not predict that the limit cycle will decrease. In Fig. 16 the limit cycles c - d - e and i - j - k are shown on a plot of static friction as a function of dwell time. The dwell time is the time during which the surfaces are in fixed cont act , the time intervals a - b , c - d, e - f , g - h and i - j in Fig. 15. The static friction increases with dwell time and this account s for t he larger limit cycle at lower velocity. Figure 17 is a plot of rising static friction measured directly by Kat o e t al . (1972) who provi de a t horough analysis of the processes relating static friction and dwell time. Lubri cant s A, B, C and D are, respectively, viscous mineral oil, commerci al slideway lubricant, castor oil and paraffin oil. Not e that in Fig. 16 the time scale is linear, as opposed t o logarithmic in Fig. 17. The empirical model of ( Kat o e t a l . , 1972), relating static friction and dwell time is: E( t ) = F~= - ( V~ - Fc)e r ' ' , (3) where F~. is the ultimate static friction; Fc is the Coul omb friction at the moment of arrival in t he stuck condition; y and m are empirical paramet ers. Kat o e t al . (1972) exami ne conformal contacts and find y to range from 0.04 t o 0.64, and rn from 0.36 t o 0.67. , , . , ~ L_ , I , t t , t ~ A 0. 5 S I G I ~ B C ' - - - O l I I 0 I 0 0 I 0 0 0 I 0 0 0 0 7"2 seconds FIG. 17. Measurements of the static friction coefficient (/~o in Kato's notation) as a function of Tz, the dwell time or time spent in static friction [from Kato et aL (1972), courtesy of the publisher]. 0.3 t T I t L.~ o Steel On steel, unlubncated J ~ _ L = 1750grn/.~. = 0.50 0 . 2 o ~ 1 ..~ / k = 0.23 cm/kg : - ~ c m/ M < 0.1 . ~ Il k = 0 , 0 0 ,~ I l / k d. 006 c mt ~ ,5 ~ l 10 -3 10 -~ 10-1 Velocity (cm/ see) FIG. 18. Stick-slip amplitude as a function of velocity, for several values of spring stiffness [from Rabinowicz (1965), courtesy of the publisher]. Armst rong-Hdl ouvry (1991) examines a non- conformal cont act and finds 7 = 1.66 and m = 0.65. A small y indicates a long rise time and thus resists stick slip. Armst rong-Hdi ouvry (1991, 1992) presents a model of rising static friction which is useful for analysis and solves some probl ems associated with using Fc as t he starting point of the static friction rise. The model , which has one fewer par amet er than t hat of equat i on (3), is: t, F,.h,,(t2) = F~ ...... + (E~.~ - E ..... ' ) t - ~ y ; (4) where F,.h, is the level of Stribeck friction at the beginning (breakaway) of the nt h interval of slip; and Fs . . . . . is the Stribeck friction at the end (arrival) of the previous interval of slip. Not e t hat y, still an empiric fact or, will be different in physical dimension from t hat of equat i on (3). Figure 18 presents the amplitude of t he spring force cycle during stick-slip, shown as a function of machi ne velocity, ka, for several values of spring stiffness, kp (Rabi nowi cz, 1965). Rabi nowi cz' s experi- ment is shown schematically in Fig. 14. The amplitude of t he spring force cycle is a decreasing function of velocity until stick-slip is abrupt l y extinguished. The amplitude is also a decreasing function of stiffness. These dat a represent values of several stiffnesses in unl ubri cat ed contacts. Brockl ey e t al . (1967), Brockl ey and Davis (1968), Ko and Brockl ey (1970), present dat a observed in an experi ment with several levels of dampi ng and Kat o e t al . (1972) present dat a collected with various lubricants. The analysis and dat a of Kat o e t al . (1972) are the most germane t o servo mechanisms as t hey i ncorporat e engineering materials and lubricants. All of these dat a present the same pat t ern: slip amplitude as a decreasing function of velocity up t o an abrupt elimination of stick-slip. The process is one of increased velocity leading to reduced dwell time, which lowers the static friction at breakaway, this furt her reducing the dwell time. At some critical velocity t he dwell time is insufficient t o A survey of friction and controls 1093 build up destabilizing static friction and stick-slip is extinguished. Derjaguin et al. (1957), Singh (1990) and Armstrong-H61ouvry (1991) present theoretical treatments that predict the critical velocity for termination of stick-slip as a function of system paramet ers and rising static friction. For the controls engineer these analyses provide an approach to the question of how slow a machine may be driven before the onset of stick-slip, and on what parameters this limit depends. Richardson and Nolle (1976) point out that in the experiments of Rabinowicz, Kat o and others, force was applied at a steady rate, as shown by the slope of the line from c- d in Fig. 15, creating a connection between force rate and dwell time: the higher the force rate the shorter the dwell time. Johannes et al. (1973) and Richardson and Nolle (1976) report experiments designed to allow independent variation of force rate and dwell time. They find that the reduction of static friction is not so much a consequence of short dwell time as of rapid force application rate, posing a challenge for explanations based on creep. Martins et al. (1990) propose an explanation based on normal penetration of the friction surfaces. For linear feedback control the distinction is perhaps not great; but for impulsive control designs the implications may be both considerable and favorable. 2.1.4.2. Fri ct i onal me mo r y and ext i ngui shi ng s t i c k - sl i p by i ncreasi ng stiffness. In Fig. 18 one observes that the trial with the stiffest spring did not exhibit stick-slip at any velocity. It is widely observed that stick-slip can be eliminated by stiffening a mechanism (Bell and Burdekin, 1966, 1969; Rabinowicz, 1965; Armstrong, 1989; Armstrong-H61ouvry, 1991). A stiffness above which there will be no stick-slip is not predicted by a model like that of Fig. l(a); but increased stiffness is the key to eliminating stick-slip in many mechanical situations (Hailing, 1975). The Stribeck curve, Fig. 19(a), shows a dependence of friction upon velocity. If there is a change in velocity, one might presume the corresponding change in friction to occur simultaneously, as suggested in Fig. 19(b). In fact there is a delay in the change in friction, as suggested by Fig. 19(c), (Sampson et al . , 1943; Rabinowicz, 1958, 1965; Bell and Burdekin, 1966, 1969; Rice and Ruina, 1983; Hess and Soom, 1990; Polycarpou and Soom, 1992). Returning to the image of partial hydrodynamic lubrication as a water skier with partial dynamic support, if we imagine the water skier half out of the water, his drag will be a decreasing function of velocity. If the tow boat suddenly increases speed, the skiers drag will decrease, but, as in Fig. 19(c), some time will pass before the new steady state drag is observed. Figure 19 is schematic. Experimental data corresponding to the observation of Fig. 19(c) is presented in Fig. 20. Rabinowicz (1951) showed that friction level lags a change in system state with an experiment that related delivered impulse to translation distance in a sliding contact. He ascribed the frictional memory to a necessary translation distance for a change in friction, on the scale of surface asperities (Rabinowicz, 1951, 1958, 1965). In fluid lubricated contacts, there is evidence that a simple time lag better describes the effect (Hess and Soom, 1990). At extremely low velocities, evidence supports a state variable model (Rice and Ruina, 1983; Dupont and Dunlap, 1993); see Section 2.1.5. Bell and Burdekin' s (1966, 1969) data are particularly applicable to common machine Friction a > Friction b > Q o tL a b ~ ~ j A A Velocity b Velocity Velocity a (a) Stribeck Friction versus Velocity Curve Friction a - Friction b - Velocity b - Velocity a - nctnon Friction a - Friction b - V o l o c i t y u - ! Velocity Velocity a T i m e ~ (b) Friction and Velocity vs Time, No Frictional Memory Frictional Memory A t - * L I ' - - Fri cti on ~ Velocity Time b (c) Friction and Velocity vs Time, With Frictional Memory FIG. 19. Time relation between a change in velocity and the corresponding change in friction. 1094 B. ARMSTRONG-Hi~LOUVRY e t a l . 14.0 F (N) ~ 3.0 ; t i , i i ; ! i 0.1 : v (~/sec) i 0.0 , , , , o [ , 0.5 t (see) 0.6 FiG. 20. Typical friction-speed time shift; contact l oad= 250 N, lubricant viscosity = 0.322 Pa - s, frequency = l Hz. F(N): friction, N; V(m s-t): velocity [from Hess and Soom (1990), courtesy of the publisher]. configurations. Figure 21 is from (Hess and Soom, 1990) and shows friction dat a for one oscillation of an oscillatory mot i on t hat brings t he system i nt o partial fluid lubrication. This experi ment was conduct ed by superimposing a velocity oscillation on st eady sliding. Af t er first stabilizing t he average mot i on, t he magni t ude of the velocity oscillation may be chosen t o probe t he very low velocity regime wi t hout arriving at zero velocity or static friction. " # " in Figs 21 and 22, as well as Fig. 17, is t he friction coefficient, friction force divided by the normal load. Not e t he vertical separation bet ween t he friction curves. The upper friction curve is given during t he acceleration away from zero velocity and t he lower during decel erat i on. The solid line of Fig. 21 was generat ed model i ng frictional memor y as a pure lag, such t hat F I ( t ) = t v c , ( i ( t - At)), (5) where Fr(t ) is t he i nst ant aneous friction force, Fvc~(') is friction as a function of st eady state velocity, see Fig. 5, and At is t he lag paramet er, t he time by which a change in friction lags a change in velocity. Hess and Soom (1990) carefully measure At and find it t o range from 3 t o 9ms in a range of load and lubricant combinations; the lag increasing with increasing lubricant viscosity and with increasing cont act load. The lag appears t o be i ndependent of oscillatory frequency (Hess and Soom, 1990). When the period of the oscillation is short relative t o At, t he hysteresis, t hat is the separation bet ween the friction levels 0.2 o,~ ,:o o o.o ~ ~ o . o ' V C m / s e ) - - - ' 0 . 2 FIG. 21. Friction as a function of velocity; O: experimental; - - : theoretical, from equations (7) and (5) [from Hess and Soom (1990), courtesy of the publisher]. 0. 2 # 0. 0 (a) i \ 0. 0 0.2 (b) 0. 0 x x . . . . . . . . - - 0.0 V (m/sec) 0.5 FIG. 22. Friction as a function of velocity; for three different frequencies of oscillation: .: 0.1 Hz; : 1 Hz; : 5 Hz. (a): experimental; (b): theoretical, from equations (5) and (7) [from Hess and Soom (1990), courtesy of the publisher]. during acceleration and deceleration, is greatest. This is illustrated in Fig. 22, also from Hess and Soom (1990). The dat a present ed were acqui red driving their pin-on-disk cont act at t hree different fre- quencies. Figure 22(b) shows t he friction curves predicted by their model with frictional memor y model ed as a pure lag and should be compar ed with the experimental dat a illustrated in Fig. 22(a). Indicative of t he progress of triboiogy, the friction model of Hess and Soom (1990) which account s for contact geomet r y and loading, material properties, velocity, lubricant viscosity and Stribeck friction, is t o a large degree based on cont act and lubricant parameters, only three paramet ers are fit a p o s t e r i o r i to the data. Evi dence for frictional memor y is available from a range of experimental sources: Sampson e t a l . (1943), Rabinowicz (1958, 1965), Bell and Burdeki n (1966, 1969), Wal rat h (1984), Rice and Rui na (1983), Hess and Soom (1990). Tri bol ogy is not yet able t o offer a theoretically mot i vat ed model of the frictional memory, t hough Xiaolan and Hai qi ng (1987) numeri- cally investigate transient el ast o-hydrodynami c lubri- cation using an analysis t hat starts with Reynol d' s equat i on and Hert zi an cont act analysis; with this t hey find a time lag of 3 ms between velocity and friction changes in simulated sliding contact. The physical process giving rise t o frictional memor y appears t o relate t o the time required to modi fy the lubricant film thickness, a process measured by several investigators (Tolstoi, 1967; Bell and Burdeki n, 1969; Bo and A s ur ve y o f f r i ct i on a nd c ont r ol s 1095 Pavelescu, 1982). A peri od of time requi red t o obt ai n a new film thickness may be one of several contributing processes, as frictional memor y is also observed in dry cont act s (Rabi nowi cz, 1951). 2.1.5. St at e v a r i a b l e f r i c t i o n mo d e l s An alternative t o t he pure time lag model is provi ded by t he state variable model s devel oped by the rock mechani cs communi t y (Rui na, 1980; Rice and Rui na, 1983; Gu e t al . , 1984; Okubo, 1986; Dieterich, 1991; Li nker and Di et eri ch, 1992). Interest in rock friction stems f r om t he hypot hesi s t hat eart hquakes are fault-line st i ck-sl i p events. While these model s have been devel oped f r om friction experi ment s on rocks, their propert i es have recent l y been observed for a range of materials (Dieterich, 1991; Dupont and Dunl ap, 1993). These include lubricated steel, Teflon on steel, glass, plastic and wood. To date, these experi ment s have been limited t o velocities within the boundar y lubrication regime. The state variable model s i ncorporat e a dependence on displacement history. They typically possess t he following t hree propert i es (assuming const ant normal stress): (1) a steady-state dependence on velocity; (2) an i nst ant aneous dependence on velocity; and (3) an evol ut i onary dependence on characteristic sliding distances. The steady-state effect, (1), represents t he general- ized Stribeck curve. The i nst ant aneous effect, (2), means t hat an i nst ant aneous change in velocity results in an i nst ant aneous change in t he friction force in t he same direction. The third pr oper t y indicates t hat following a sudden change in velocity, t he steady-state curve is approached t hrough an exponent i al decay over characteristic sliding distances. This t ype of model can reproduce t he friction behavi or depi ct ed in Fig. 21 ( Dupont , 1994). For const ant normal stress, t he general model including t he n state variables, 0i, is given by: Ft ( t ) = f ( V , Or, 02 . . . . . 0 , ) Oi =g ~ ( V, 0 , , Oz . . . . . 0 , ) , i = 1, 2 . . . . . n. (6) This form implies t hat a sudden change in velocity cannot produce a sudden change in t he state, 0, but does affect its time derivative. Hence, t he instan- t aneous velocity effect takes place at const ant state. The evolution of t he state variables in response t o changes in velocity, t oget her with t he i nst ant aneous velocity effect, dictate t he dynami c behavi or. Physical i nt erpret at i ons of t he state variables are possible. Consi der a st andard dry friction model in which friction stress depends on t he yield stress of asperity junctions. For a single state variable and constant normal stress, the state variable can be related t o the mean lifetime of an asperity j unct i on. Recent l y, these models have been enhanced t o include dependence on normal stress. I n this case, t he state variables can be rel at ed t o t he t i me-dependent growth of the l oad-beari ng j unct i ons (Li nker and Dieterich, 1992). The functional f or m of t he state variable model s was deduced from the response t o step changes o. o o o. o ~ 9 o o go LL C) 0 STEP CHANGE tN NOMINAL S L I P RATE 1 0 - 2 - 1 0 o / j . m / s 1 0 0 I 1 0 - 1 I 1 0 - 2 I 1 0 - 1 1 0 0 I 1 0 - 2 I 1 0 0 R u i n a , ~ m / s p, r n / s p.m/s p m / s I p.m/s p.m/s p m / s q u a r t z i t e 0 . 0 i i i i i i i 1 0 . 0 2 0 . 0 3 0 . 0 4 0 . 0 5 0 . 0 6 0 . 0 7 0 . 0 DISPLACEMENT (u.m) FiG. 23. Friction stress as a function of displacement in trials with unlubricated quartzite. Step changes in velocity produce the instantaneous-effect spikes and subsequent evolution to the new steady-state level [from Ruina et al. (1986), courtesy of the authors]. imposed on the velocity at t he friction interface. These experiments are a significant i mpr ovement over standard tribology experi ment s because t hey involve control of the friction interface mot i on instead of t he act uat or mot i on. Figure 23 depicts friction stress versus di spl acement dat a obt ai ned by Rui na e t al. (1986). The t hree model ed effects are clearly visible in t he data. The fact that very small, steady velocities were achieved t hrough closed-loop cont rol in these experi ment s is additional evidence t hat stable, low-velocity cont rol is possible. State variables models (or additional internal states) have also been pr oposed whose behavi or resembl es t hat of a connect i on with a stiff (nonlinear) spring (Dahl, 1977). The Dahl model predicts a frictional lag between velocity reversals and leads t o hysteresis loops. The mat hemat i cal propert i es of t he Dahl model are studied in Bliman (1993). However , this model can onl y predict Coul omb friction steady-state velocity characteristics; t he Stribeck effect is not included. An interesting i nt erpret at i on of this model by using linear space invariant models (instead of nonl i near differential equat i ons) is present ed in Bliman and Sorine (1991). With this new model it becomes clear how frictional forces, predicted by the Dahl model , depend on t he curve length associated with t he t raj ect ory of relative mot i ons (integral of t he velocity absolute value). To i nt roduce t he Stribeck effect, it is possible t o ext end t he Dahl model (which is first-order) t o a model with a high degree of differentiability (Bliman and Sorine 1991, 1993). The second-order Dahl model can show t he Stribeck phenomenon by produci ng an over shoot in t he response of t he friction forces. Anot her possibility is t o modify t he original Dahi model so as t o include t he Stribeck effect wi t hout increasing the system state dimension (Canudas de Wit e t al. 1993). I n this modified Dahi model , t he internal states have a physical i nt erpret at i on. They describe t he Bristles average deformat i on. The state variable model s of Rice and Rui na (1983), t he translation distance of Rabi nowi cz, and the pure lag of Hess and Soom (1990) are all represent at i ons of 1096 B. ARMSTRONG-HI~LOUVRY e t al . frictional memor y. The effect of frictional memor y is a delay in the onset of the destabilizing drop in friction. Fr om a cont rol st andpoi nt , the frictional memor y reduces the destabilizing influence of Stribeck friction. If t he time const ant s of a system are short in relation t o t he frictional memor y, which is to say that the mechani sm (cont rol ) is sufficiently stiff, the stick-slip limit cycle will not be stable (Rabi nowi cz, 1965). (For t he range of frictional memor y time constants, see Tabl e 1). Thi s is t he p r o c e s s wh e r e b y i n c r e a s i n g s t i f f n e s s e l i mi n a t e s s t i c k - s l i p . 2.1.6. Fr i c t i o n as a f u n c t i o n o f s t e a d y s t at e v e l o c i t y : v a r i a n t s o f t he S t r i b e c k c u r v e Friction is a function of velocity because the physical process of shear in the j unct i on changes with velocity. Figure 24 presents several fri ct i on-vel oci t y curves. Details of t he ( f - v) curve depend upon the degree of boundar y lubrication and the details of partial fluid lubrication. Curves such as (a) arise when lubricants t hat provi de little or no boundar y lubrication are empl oyed. The dat a of Bell and Burdekin (1966, 1969) and Hess and Soom (1990) indicate such a curve. When boundar y lubrication is more effective, t he friction is relatively const ant up t o the velocity at which partial fluid lubrication begins t o play a role. Vi nogr adov e t al. (1967) and Khitrik and Shmakov (1987) present dat a support i ng a flat ( f - v) curve t hrough t he region of boundar y lubrication, as suggested by curve (b) of Fig. 24. Fuller (1984) cites dat a cont rast i ng a specific lubricating oil with and wi t hout a lubricity additive. The plain oil gives a curve of t ype (a); with the lubricity additive a curve of type (b) is observed [see Fuller (1984), Figs 11-14; the reference offers considerable discussion of boundar y lubrication]. One must be careful in discussing friction as a function of steady state velocity. Dat a collected during velocity transients will exhibit t he effects of frictional memor y, equat i on (5), and a curve of t ype (b) may be observed even if the underl yi ng steady state ( f - v) curve is of t ype (a). Bell and Burdeki n (1969) present a t hor ough analysis of this phenome- non. A curve of t ype (c) is given by way lubricants (Merchant, 1946; Wolf, 1965). The boundar y lubrication provi ded by the additives t o these oils : Limited Boundary Lubrication : Substantial Boundary Lubrication Way Lubri cant Velocity ,, FIG. 24. Friction as a function of steady state velocity for various lubricants; the (f-v) curve [after Fuller (1984)]. reduces static friction t o a level below Coul omb friction. For analysis or simulation it is i mport ant to have a mathematical model of the steady-state fri ct i on- velocity dependence. Hess and Soom (1990) empl oy a model of the form F ( Y c ) = F, - + ] ( g F c) + ( ~ I L ) " + F,,~ (7 ) and show a systematic dependence of .f,. and F. on lubricant and loading paramet ers. Bo and Pavelescu (1982) review several models proposed in the literature and adopt and t hen linearize an exponential model of the form: F(Yc) = Fc + ( ~ - Fc ) e (s/x,)~ + F,,Yc, (8) where E is the level of static friction, Fc is t he minimum level of Coul omb friction, and ,f,. and 6 are empirical paramet ers. The viscous friction paramet er, F,,, is added here; a viscous term was not i ncorporat ed by Bo and Pavelescu (1982). In t he literature surveyed by Bo and Pavelescu (1982), t hey find 6 t o range from 1/2 to 1. Armst rong-H61ouvry (1990, 1991) empl oys 6 = 2; and the dat a cited by Fuller (1984), observed in a system with an effective boundar y lubricant, would suggest 6 very large. The exponential model (8), with 6 = 2, is a Gaussi an model. The Gaussian model is nearly equivalent t o the Lorent zi an model of Hess and Soom (1990), equat i on (7). The exponential model (8), is not a strong constraint. By appropri at e choice of paramet ers, curves of types (a), (b) and (c) can be realized. What is needed are dat a such as t hat of Hess and Soom (1990) over a broad range of engineering materials, conditions and lubricants. For specific lubricant formulations, lubrication engineering firms can pro- vide measures of lubricity and ot her qualities based on standard industrial tests. The st andard tests of lubricant qualities are not t he equivalent of t he dat a of Hess and Soom (1990), but are none-the-less useful. Industrial testing for iubricity is still evolving (Ludema, 1988). 2.2. An Int egrat ed Friction Model This discussion of friction has focused on sliding between hard metal parts lubricated by oil or grease. For reasons of machine life and performance, these engineering materials make up many of the machines encount ered by controls engineers. When these materials are used, the state of underst andi ng supports a friction model t hat is compri sed of four velocity regimes, two time dependent properties and several mechani sm dependent properties. (1) The four velocity regimes. (I) Static Friction: displacement (not velocity) is proport i onal t o force [see Fig. 7 and equat i on (1)]. (II) Boundar y Lubrication: friction is dependent on surface properties and lubricant chemistry. ( I I I ) Partial Fluid Lubri cat i on: if static friction is great er than Coul omb friction, friction decreases with increasing velocity. A s u r v e y o f f r i c t i o n a n d c o n t r o l s 1 0 9 7 (IV) Full Fluid Lubri cat i on: friction is a funct i on of velocity, a viscous plus Coul omb friction model may model t he friction quite accur- ately. [Regi mes I I - I V in Fig. 5, see also equat i on (8).] (2) The t wo t i me- dependent propert i es. (I) Rising Static Friction with Increasing Dwell Ti me [see Fig. 16 and equat i on (3)]. (II) Frictional Memor y: in partial fluid lubrica- tion, friction is dependent upon velocity and load; a change in friction will lag changes in velocity or l oad [see Fig. 20 and equat i on ( 5) 1. 2.2.1. T h e s e v e n p a r a m e t e r f r i c t i o n m o d e l . Theor et - ically mot i vat ed model s for t he component s of friction are not yet available, and a variety of empirically mot i vat ed forms have been present ed. One choice of model is t he seven par amet er model , where t he friction is given by: N o t s l i d i n g ( p r e - s l i d i n g d i s p l a c e m e n t ) . Fr(x ) = - k , x ( 9 ) S l i d i n g ( C o u l o m b + v i s c o u s + S t r i b e c k c u r v e friction wi t h f r i c t i o n a l m e m o r y ) . 6 ( ~ , t ) = / - [ F c + F ~, l i l + F ~( y, t2) 1) ( 2 ( t - rL)] 2 sgn (i ). 1 + - - - - - - - ~- - - / ( l o ) R i s i n g s t at i c f r i c t i o n ( f r i c t i o n l e v e l a t b r e a k a w a y ) . t2 F~(y, t2) = F, ~ + (Fs. - F , . , ) - - , (11) ' t 2 + where: FI(. ) is t he i nst aneous friction force; Fc ( * ) is t he Coul omb friction force; F~ ( * ) is t he viscous friction force; F, is t he magni t ude of t he Stribeck friction (frictional force at breakaway is Fc + F,); F,., is the magni t ude of t he Stribeck friction at t he end of t he previ ous sliding peri od; F, (*) is the magni t ude of t he Stribeck friction after a l ong time at rest (with a slow application of force); k, (*) is t he tangential stiffness of t he static cont act ; i , (*) is the characteristic velocity of t he Stribeck friction; rL (*) is the time const ant of frictional memor y; }, (*) is t he t empor al par amet er of t he rising static friction; h is t he dwell time, time at zero velocity; (*) marks friction model paramet ers, ot her vari- ables are state variables. The magni t udes of t he seven friction paramet ers will naturally depend upon the mechani sm and lubrica- tion, but typical values may be offered. Ranges suggested elsewhere in this section, originating TABLE 1. APPROXIMATE RANGES FOR THE PARAMETERS OF THE SEVEN PARAMETER FRICTION MODEL Parameter range Parameter depends principally upon 0.001 - 0.1 * F, Lubricant viscosity, con- F~ F,, Fs.~ k, 0-very large 0 - 0.1*F. 1 ~ * ( ~ + F c ) ; a ~ = 1 - 50[ um] 01[ meter] 0 . 0 0 0 0 1 - tse--S0-dodJ t t 1 - 5 0 [ m s ] ; , 0 - 2 0 6 [ s ] tact geometry and loading Lubricant viscosity, con- tact geometry and loading Boundary lubrication, F,- Material properties and surface finish Boundary lubrication, lubricant viscosity, Material properties and surface finish, Contact geometry and loading Lubricant viscosity, con- tact geometry and loading Boundary lubrication principally with Bowden and Tabor (1973), Kat o e t al. (1974), Fuller (1984), Armst rong-H61ouvry (1991), Hess and Soom (1991a, b), Pol ycarpou and Soom (1992), are summari zed in Tabl e 1. The friction force magnitudes, F o Fo and Fs are expressed as a funct i on of normal force, i.e. as coefficients of friction. Ax is t he deflection before breakaway resulting from cont act compliance. Each of t he seven paramet ers of the model represent s a different friction phenomenon. The seven rows of Tabl e 2 indicate t he effect of these TABLE 2. FRICTION MODEL CAPABILITIES Friction model Predicted/observed behavior Viscous Coulomb Static + Coulomb + Viscous Stribeck Rising static friction Frictional memory Presliding displacement Stability at all velocities and at velocity reversals. No stick-slip for PD control; No hunting for PID control Predicts stick slip for certain initial conditions under PD control; predicts hunt- ing under PID control. Needed to correctly predict initial conditions leading to stick-slip. Needed to correctly predict interaction of velocity and stick-slip amplitude. Needed to correctly predict interaction of stiffness and stick-slip amplitude. Needed to correctly predict small displacements while sticking (including velocity reversals). AUTO30-7-B 1098 B. ARMSTRONG-HI~LOUVRY et al. phenomena on sliding behavior. Al t ernat i vel y, and mor e appropri at el y, the table can be used to select a friction model based on experimental observations. Pol ycarpou and Soom (1992) have recently report ed dynamic measurement s of friction in lubricated metal contacts made with a remarkabl y sensitive apparatus. Except for viscous and rising static friction, each of the component s of the seven par amet er model is evident in the dat a of Pol ycarpou and Soom (1992); and the aut hors observe t hat rising static friction may have been present on a time scale ot her than t hat observed. Furt hermore, al t hough a detailed para- met er identification is not present ed, the aut hors are able to account for all of the qualitative phenomena with reference to presliding displacement, Coul omb and Stribeck friction, and frictional memory. In practical machi nes t here t end t o be many rubbi ng surfaces---drive elements, seals, rot at i ng electrical contacts, bearings et c- - whi ch cont ri but e t o the total friction. In some mechanisms, a single interface may be the domi nant cont ri but or, as transmission el ement s often are. In ot her cases where t here are several elements contributing at a comparabl e level, it may be impossible to identify their individual cont ri but i ons without machine disassembly. In these cases, a model , such as the one above, can be used t o represent t he aggregate friction. 2. 2. 2. Speci al mechani cal consi derat i ons Much of this survey has dealt with sliding lubricated metal cont act s; but ot her cont act s may be i mport ant . This section provi des a brief overview of rolling friction as well as ot her friction phenomena which may arise in compl ex machines. 2.2.2.1. Rol l i ng f ri ct i on. Rolling el ement s typically generat e much less friction t han sliding elements at comparabl e loads and speeds. For this reason, the friction cont ri but i on of roller bearings is usually insignificant in compari son with t hat of the sliding cont act s in a machine and, thus, oft en plays a mi nor role in machi ne design. Some i mport ant exceptions include disk drives; ball screws ( Ro and Hubbel , 1993) and ball-bearing slideways (Futami et al. , 1990) used in precision engi neeri ng; and the gimbal bearings of pointing and tracking devices (Gilbart and Winston, 1974; Wal rat h, 1984; Himmell, 1985; Maquei ra and Masten, 1993). To gain an appreci at i on of the level of friction involved, consider t hat for ball and roller bearings operat i ng at typical loads and speeds, the friction coefficients range bet ween /~ = 0.001 and 0.005 (Eschmann, 1985). For roller bearings, the friction coefficient is related t o friction t orque by: r/ (12) t~ = Fd / 2 " Here, ~/ is t he friction t orque, F is the resultant bearing load, including bot h radial and axial component s, and d is the bearing bore di amet er. Starting from rest, a slightly higher stiction level of ' rolling' friction may exist, but in ball bearings this effect is usually quite small (Pal mgren, 1945). Several friction models have been proposed over the years. Roller bearing texts typically provide semi-empirical equat i ons of the basic form: r / = r,, + r , . (13) where ro is the no-load component of friction t orque and r, usually depends strongly on bearing load, but only lightly on velocity (Eschmann, 1985). While the model described above is meant t o apply to a broad range of operat i ng conditions, the Dahl model was devel oped t o explain the hysteretic behavi or of precision ball bearings undergoi ng very small ampli- tude oscillations (Dahl, 1968, 1977). The Dahl model has been widely used to study the simulation and control of machines. Mechanisms of rolling friction. There are two effects associated with the elasticity of the contact zone which cont ri but e t o rolling friction (Harris, 1984). These effects, however, make up a small portion of the total rolling friction. It is a surprising fact t hat most of the friction in roller bearings is due to sliding mot i on. This sliding is one of the maj or reasons t hat roller bearings must be lubricated with oil, grease or sometimes, and with less effect, a dry lubricant. To underst and how sliding can occur, first consider that pure rolling would require point contacts or line contacts parallel t o the bearing axis of rotation. Owi ng t o elastic deformat i on, ball bearings on flat or curved raceways have curved contact regions. In addition, rollers and raceways are usually crowned in or der to prevent edge loading (Harris, 1984). Thus the cont act region is curved. Consider Fig. 25. With the ball rolling at a particular velocity, there will be only two curvilinear segments within the elliptical cont act zone at the proper radius t o undergo pure rolling. The velocity profile for the maj or axis of the cont act ellipse is shown. The points D and D' lie on the rolling O r i g i n a l r a c e w a y f o r m B a l l a x i s - - - - C o n t a c t a r e a - I , O r i g i n a l ~ i i r o l l i n g e l e m e n t - i t i f o r m ~ I 2 0 , D i r e c t i o n o f r o t a t i o n I - S l i d i n g s p e e d ] ~ C o n t a c t elllp~ L I I , , , I I / FIG. 25. Sliding in the contact ellipse of a ball rolling on a curved raceway [from Eschmann (1985), courtesy of the publisher]. A s ur ve y o f f r i ct i on a nd c ont r ol s 1099 segments. Bet ween these points, slip will occur opposite t he direction of rolling. Out si de t he points, slip occurs in the direction of rolling. Accordi ng t o t he t ype of roller beari ng, sliding friction will also arise f r om cont act bet ween t he rolling elements and t he cage, bet ween t he rolling el ement s themselves and bet ween t he roller faces and t he raceway lips. Ther e is also viscous drag on t he rollers caused by the lubricant and friction due t o t he bearing seals. Seal friction can be considerable and can far exceed t he total of all ot her sources of beari ng friction (Harris, 1984). 2.2.2.2. Ot her machi ne elements. The precedi ng discussions apply t o simple sliding or rolling friction; in compl ex machi nes t here may be additional consi derat i ons. One such consi derat i on is different friction magni t udes in different directions of mot i on. Different Coul omb and viscous friction levels in t he left and right rot at i on directions have been observed experi ment al l y on many occasions, e.g. Mukerj ee and Ballard (1985), Canudas de Wit et al. (1987), Ar ms t r ong- Ht l ouvr y (1991). Theoretically, this may be due t o anisotropies in material or geomet ry (Zmi t rowi cz, 1981; Ibrahi m, 1992a). And t he phenomenon is a sufficient consi derat i on t hat a st andard stick-slip test calls for separat e measure- ment s in t he left and right directions (Cincinnati Mi l acron, 1986). Some mechani sms will exhibit posi t i on-dependent friction (Mukerj ee and Ballard, 1985; Candas de Wit et al. , 1987; Ar mst r ong, 1988; Ar mst r ong- Ht i ouvr y, 1991). This is particularly t rue of transmissions with spatial inhomogeneities, i.e. cont act geomet r y or loading which varies as a function of position. Gear drives are a common exampl e and give rise t o posi t i on-dependent friction. Wi t h accurat e friction measurement s, Ar ms t r ong- Ht l ouvr y (1991) was able t o count t he transmission gear teeth, and i ncorporat - ing this fact or i nt o t he friction model substantially increased t he accuracy of predi ct ed friction. I n part t o eliminate posi t i on-dependent friction, Salisbury et al. (1988) and Townsend (1988) study designs with homogeneous transmissions. 2.2.2.3. Ext ernal f ri ct i on. Sources of internal friction, such as bearings, are oft en designed so as t o minimize friction. Mechanisms which must make cont act with their envi ronment , however, can have quite different design goals. In robot i c dext rous mani pul at i on, for exampl e, high friction coefficients are desirable. Ver y soft fingers, made of rubber or el ast omeri c material, can provi de friction coefficients great er t han one. As a result, obj ect s can be grasped gently while inhibiting bot h tangential sliding and rot at i on about the cont act normal (Cut kosky and Wri ght , 1986). Due t o t he compl exi t y of the dext rous mani pul at i on probl em, many simplifying assumptions are made in t he system modeling. For instance, most studies involving sliding assume quasistatic conditions (Kao and Cut kosky, 1992; Peshkin and Sanderson, 1988; Tri nkl e, 1989). This is done under t he assumption t hat fine assembly operat i ons are typically performed slowly (Trinkle, 1989). Recent l y, at t ent i on has been given t o issues of cont rol arising f r om t he details of friction in grasp ( Howar d and Kumar 1993; Schimmels and Peshkin 1993). The modeling of friction in cont act s involving rubber or elastomers has received at t ent i on, but its description is beyond t he scope of this paper. The following references on this topic are provi ded by Cut kosky and Wri ght (1986), Cut kosy et al. (1987), Howe et al. (1988), Moor e (1972, 1975) and Schallamach (1971). The issue of stick slip as it affects mot i on planning and cont rol in dext rous mani pul at i on has apparent l y not been studied. Ot her exampl es of external friction, such as deburri ng or drilling operations, pose quite different model i ng and cont rol challenges as these tasks involve deliberate oper at i on within t he severe wear regime for one surface (Smith, 1989). 2.2.2.4. Run- i n and f ri ct i on noise. In developing our friction model , we have, for t he most part, dwelt on factors such as velocity and load which can be consi dered as exogenous variables. There are also internal fact ors at work which depend on time, sliding cycles or total sliding distance. These effects are due t o such things as loss of lubricant, deformat i on of surface material, change in t emperat ure due t o generat ed heat or accumul at i on of wear debris. These factors all cont ri but e t o pr oduce changes in the mean friction force even while t he exogenous variables of velocity and load are held const ant . These effects are perhaps most evident at t he beginning and end of the life of a tribo-system. Duri ng t he run-in period, the friction level of a new machi ne may increase or decrease until a l ong-t erm steady-state condition of mild wear is reached. The end of a tribo-system' s useful life is mar ked by a transition t o severe wear. In addition t o variations in t he mean friction level, the ' noi se' level can also vary over time dependi ng on such properties as surface roughness and accumul at i on of wear debris. Oft en, variation in t he friction force is highest during t he run-in peri od and after t he transition t o severe wear (Blau, 1987). These factors are i mport ant in terms of friction identification and cont rol for t he following reasons: A new machine may exhibit a higher or lower (and noisier) level of friction t han t he ' st eady-st at e' level achieved after runqn. Af t er a peri od of machi ne inactivity (at t he start of the day, for example), it may be worthwhile t o perform machine calisthenics. This will allow for circulation of the lubricant, t emper at ur e stabi- lization and thus stabilization of friction level. The average friction level obt ai ned from very noisy dat a may not be correct. While t he maxi mum friction magni t udes may well be due t o t he microscopic geomet ri c and structural propert i es of the interface, the mi ni ma may depend mor e on t he machine stiffness and sensor response (Blau, 1987). In distributed par amet er systems, such as a violin string or railway wheel, friction can induce chaot i c motions. Popp and Stelter (1990) have investigated frictionally i nduced chaos in l umped and distributed paramet er systems, and find t hat PI D cont rol of a 1100 B. ARMSTRONG-HI~LOUVRY e t al . single mass with s t a t i c +Coul omb friction is not expected t o exhibit chaotic mot i on; but that a two-mass, spring system under the same conditions will, as will a distributed mass system (such as a railway wheel or break dr um) under a broad range of conditions, t hey present bot h theoretical and experi ment al results, including a proposed met hod for distinguishing chaos from noise in empirical data. 2.2.3. N o r m a l f o r c e a n d t he coef f i ci ent o f f r i c t i on For much of this discussion, friction has been addressed as a force, rat her than as a coefficient of friction, and normal force has not been addressed in depth. The frictional force, normal force and coefficient of friction are, of course, rel at ed t hrough: Ff ( t ) = I~iF,(t), (14) where Ff(t) is t he i nst ant aneous force of friction, F, ( t ) is the i nst ant aneous normal force and ~u r is the coefficient of friction. The coefficient, /~I, is not constant, but may depend upon velocity, velocity history, normal force and normal force history (Pavelescu and Tudor , 1987; Martins et al . , 1990). In cont rol applications, situations exist in which it is possible t o know the normal force, such as in a machine way carryi ng a known load; t here are situations in which it may or may not be possible t o know the normal force, such as in a bearing where t he external load is known but internal force may not be; and there are situations in which it is not at all straight forward t o know the normal force, such as in a prel oaded gear train or mot or brushes. In some cases the normal force may be const ant and in ot hers it may vary. In systems which exhibit stable friction, such as joint 1 of the PUMA r obot (Armst rong, 1988; Armstrong-H61ouvry, 1991), normal force, along with t emperat ure and ot her factors, must be well behaved. In mechanisms where t he normal force is varying, the prediction of friction becomes mor e complicated. This is particularly t rue where normal force is det ermi ned by cont rol effort, as will be the case in transmissions t hat are not prel oaded. The characteris- tic velocity of the Stribeck curve, state associated with frictional memor y and the stiffness of presliding displacement are all influenced by i nst ant aneous normal force, and by the history of applied normal force (Martins et al . , 1990; Soom, 1992). Nor mal force history has been shown t o influence friction in geophysical systems (Li nker and Dieterich, 1992). It is beyond the current state of t he art t o compl et el y model t he influence of changi ng normal force, t hough attention within t ri bol ogy is t urni ng t o what appears to be the central issue: the normal displacement, e.g. Toistoi (1967), Oden and Martins (1985), Martins et al. (1990), Hess and Soom (1991a, b). For the moment , t he most viable approach t o probl ems of dynami c normal force empl oys the integrated model , with Coul omb, viscous and Stribeck friction com- ponent s represent ed as coefficients of friction and the stiffness of presliding proport i onal t o normal force. 2.3. Future Trends in Tri bol ogy and Implications for Cont rol The overwhel mi ng maj ori t y of t reat ment s of friction have viewed the part -t o-part interaction as a one degree of freedom mot i on: tangential, sliding mot i ons are considered. Normal force has always been considered, but normal mot i ons have been neglected. A school of t hought is devel opi ng that normal mot i ons play a central role in det ermi ni ng friction; including the realization of frictional memor y and the Stribeck curve (Tolstoi, 1967; Tudor and Bo, 1982; Oden and Martins, 1985; Martins et al . , 1990; Goyal et al . , 1991). Tolstoi and ot hers have made careful observat i ons of friction and sub-micron normal displacements and find a strong correlation between i nst ant aneous friction and instantaneous normal displacement, as shown in Fig. 26 (Tolstoi, 1967; Budanov e t al . , 1980). Described heuristically, as the cont act begins t o slide, impacts bet ween the contacting asperities increase the separation between surfaces. Because the friction is a strong and nonl i near function of asperity penet rat i on (normal separation), friction is modified by the changing normal separation (Martins et al . , 1990). The fri ct i on-vel oci t y curve and frictional memor y are thus in part manifestations of t he normal dynamics. Different mechanisms, such as prel oaded gears or a slider on a machine way, may have very different normal stiffness and dampi ng, giving different frictional dynamics, even t hough material, N (kgm)! J , 2 ! i . o t o8 06 ~ - - - - 0 . 4 F (kgrn) 0 . 2 0. 0 O0 04 0 . 8 1 2 FIG. 26. Normal load (N, curve 1) and static friction force (F, curve 2) versus the normal separation,/~ (arbitrarily,/; is taken to be zero for the maximum normal load used in the experiments). Dry steel surfaces [from Oden and Martins (1985), courtesy of the publisher, adapted from data reported in Tolstoi (1967)]. A s ur vey of fri ct i on and cont r ol s 1101 lubrication, geometry and normal loading may be the s a me . As a demonstration of the potential of this framework, Martins e t a l . (1990) have been able to account qualitatively for a broad range of previously irreconcilable experimental observations, using com- puter simulations based on a simple model of friction physics and a more detailed model of the normal direction contact dynamics. To date, this work has concentrated on dry friction contacts; the impact of fluid lubricants must be considered for the results to be directly applicable to common control situations. The payoff for controls is the possibility of predictive, physically motivated models for machine friction. At issue are the physics underlying the Stribeck curve and frictional memory, both of which play leading roles in determining stick slip. 2.4. A Final Word on Models As evidenced by the recent works of Oden, Martins, Soom and others, tribology has found a renewed interest in frictional dynamics, and new paradigms that may overcome conundrums left by the investigations of the 1950s. The direct motivation often stems from vibrationally induced noise, fatigue and wear--act i ve feedback is never addressed in the tribology l i t erat ure--but the possibility of spin-off technology for the controls community seems great, especially in as much as both camps are concerned with interfaces of engineering materials and mass- spring-damper systems. Even if predictive models of friction are never genuinely achieved, benefit for mechanism design and controls will come in the forms of more certain model structure, bet t er identification strategies, bounds on paramet er ranges, a broader range of frictional interfaces which are understood, and a richer pallet of design strategies for friction modification. All of which will contribute to better price/performance in machines. 3. ANALYSIS TOOLS THAT HAVE BEEN APPLIED TO SYSTEMS WITH FRICTION Analysis of the motions of machines with friction have been made employing four types of tools: describing functions, algebraic analysis, phase plane analysis and simulation. Simulation is not normally considered an analysis tool; but when sufficient trials, perhaps thousands of trials, are conducted, the structure of the system behavior may be illuminated or empiric relations identified. We include simulation as an analysis tool here because its use is common in applications. In almost all cases where these tools have been applied, the goal has been to predict the conditions for stick slip. The character of the result depends heavily upon the friction model, task and control structure considered. For a slip cycle during which velocity does not reverse---the common case with tracking tasks (Derjaguin e t a l . , 1957)---a Coulomb friction model permits an exact integration of the acceleration through a slip cycle and thus exact algebraic results. In all other cases approximations are involved. The character of the approximations and their impact on the validity of the conclusions drawn are important issues in all analyses. A relatively small number of investigators have verified their analysis with either experiment or extensive simulation. The works applying nonolinear analysis techniques to systems with friction are all relatively specific in their focus. As a general introduction to analysis techniques for these systems several books have been written in the last decade, such as Slotine and Li (1991), Vidyasagar (1991) and Khalil (1992). Among older texts Atherton (1975) is often cited. Mees (1984) provides an interesting discussion of recent results regarding the describing function. 3.1. Describing Functions The application of describing function analysis to study the motions of machines with friction has a long history (Tou and Schuitheiss, 1953; Satyendra, 1956; Siiverberg, 1957; Shen, 1962; Woodward, 1963; Brandenburg, 1986; Brandenburg and Schiller, 1987, 1988a, b, 1989, 1991: Sch~ifer and Brandenburg, 1990, 1993; Townsend and Salisbury, 1987; Wallenborg and ,~str/Sm, 1988; Canudas de Wit 1988; Canudas de Wit and Seront, 1990; Canudas de Wit 1987, 1991; Ehrich, 1991). The technique is an approximate one consisting of representing the i nput -out put map of a single- input/single-output nonlinear element by the mag- nitude and phase relationship between a sinusoidal input and the fundamental harmonic of the cor- responding output. This relationship constitutes a sort of transfer function; it can be represented by a complex number, and will in general be frequency and magnitude dependent. Underlying describing function analysis is the requirement that the element be single-input/single-output, and the assumption that investigation of the first harmonic provides a reasonable approximation to the behavior to the true system (Brogan, 1991). The advantage of the describing function is that it permits the use of frequency domain tools for the analysis of control. There is an important special case in the study of describing functions: the memoryless, odd function. A memoryless element is one without state; and for such an element, the describing function will depend solely upon magnitude of the input. When the function is odd (friction as a function of velocity, with symmetric friction in the left and right directions, is a memoryless, odd function), the describing function will be strictly real (Brogan, 1991). Recent authors Brandenburg (1986), Brandenburg and Schiller (1987), Townsend and Salisbury (1987), Wailenborg and Astr6m (1988), Canudas de Wit (1988), Canudas de Wit and Seront (1990), Canudas de Wit e t al . (1987, 1991) and Ehrich (1991) have taken advantage of this special case, which we will call the ' memoryless element' construction. Earlier authors Tou and Schuitheiss (1953), Satyendra (1956), Silverberg (1957), Shen (1962) and Woodward (1963) formed describing functions of the combined plant with friction. This construction is not memoryless because 1102 B. ARMSTRONG-HI2LOUVRY et al. k o ' ~ ~,, ~ .,~ (Friction vs. Velocity) [ FIG. 27. A single mass system with friction, friction modeled as a function of velocity. it pl aces pl ant st at e within t he model ed nonl i near el ement ; this we t er m t he ' i nt egr at ed f r i ct i on/ pl ant ' const ruct i on. The descri bi ng funct i on anal ysi s of a posi t i on cont r ol l ed one degr ee of f r eedom ri gi d mass syst em with sl i di ng fri ct i on may be const r uct ed as shown in Fig. 27. Si mpl e or compl i cat ed syst ems may be consi der ed with appr opr i at e choi ces of GAs), G~(s) and G2(s), [see, for exampl e, Townsend and Sal i sbur y (1987), Br andenbur g (1986), Br andenbur g and Sch~ifer (1987)]. The bl ock di agr am must be mani pul at ed t o ar r ange all of t he l i near el ement s as one bl ock, l eavi ng t he nonl i near el ement as a second bl ock giving t he bl ock di agr am of Fig. 28. The t r ansf er funct i on of t he syst em of Fig. 27 may be wri t t en: G~ X(s) = Gr 1 + G, Gt G2 G2 (15) St i ck- s l i p is a l i mi t cycle, t hat is a mot i on giving a cl osed pat h in t he phase pl ane. A l i mi t cycl e may be ei t her st abl e, which means t hat near by pat hs conver ge ont o it; or unst abl e, which means t hat near by pat hs do not converge t o t he l i mi t cycle, but does not i mpl y t hat t he syst em is exponent i al l y unst abl e. Usi ng t he descri bi ng funct i on, a l i mi t cycl e is det ect ed when t her e exists an ampl i t ude, A, and f r equency, s = Ro, such t hat t he denomi nat or of (15) goes t o zer o; giving: 1 G~ (16) N(A) 1 + G, GtG2 or, r ef er r i ng t o Fig. 28: 1 G~ - - - = G, ( s ) ; GL(s)- (17) N(A) 1 + G,G~G2 Thi s condi t i on is easi l y t est ed by dr awi ng a Nyqui st pl ot , with -1/N(A) as one br anch, and Gt.(s) as a second branch; as seen in Fig. 29. An i nt er sect i on predi ct s a limit cycle. The l i mi t cycl e must addi t i onal l y +-- f GL(S ) j - F r X . . . . N(A) ~ , i FIG. 28. A single mass system with friction, block diagram manipulated to lump linear elements. - I/N(A) ~" 0,50 Limit Cycle 5 Indicated -~ f Increasing A / \ 0 1 5 0 ( lncreasing~..~ 5 -0.50 - -- ~ P I D; Ki=10 ~ ~ Real Axis 0 t / ~ riD; K i = I00 FIG. 29. Nyquist plot for limit cycle detection in a system with integral control [following Townsend and Salisbury (1987)]. be t est ed for st abi l i t y, [see e. g. Br ogan (1991)]. Thr ee cont ours of Gt.(s) ar e pl ot t ed in Fi g. 29, cor r espond- ing to a damped mass with a PD or PI D cont r ol l er . For t he syst em under consi der at i on, equat i on (16) gives: 1 s PI D: N(A) Ms2+(b+k,,)s+k,, (18) 1 s z PI D: N(A) Ms 3+(b+k,,)s 2+kvs+kf Defi ni t i ons and val ues for t he par amet er s of equat i on (18) and Fig. 29 ar e given in Tabl e 3. Of t he t hr ee Gl.(s) cont our s of t he Nyqui st di agr am, Fig. 29, two pr edi ct st abl e mot i on. Those ar e t he PD cont our and t he PI D cont our wi t h K; = 10. Fol l owi ng Townsend and Sal i sbury (1987) a cont our is shown with a l arge i nt egral cont r ol gai n; t hi s cont our i ndi cat es stick slip. Not e t hat t he cont our s of Fig. 29 ar e pl ot s of Gt.(s) as gi ven by equat i on (17), r at her t han t he t r ansf er funct i on cust omar i l y pr esent ed in Nyqui st pl ot s. The memor yl ess el ement descri bi ng funct i on is st r ai ght f or war d t o appl y. Several aut hor s have made ext ensi ons t o this anal ysi s. Wal l enbor g and ,~,str6m (1988) pr esent an i nt er est i ng pr oof t hat a syst em with Coul omb fri ct i on (no st at i c fri ct i on) and st at e f eedback can be unst abl e onl y for t he speci al case of an unst abl e cont r ol l er . Ami n (1993) has st udi ed Coul omb and Co u l o mb +s t a t i c fri ct i on wi t h t he memor yl ess el ement and i nt egr at ed pl ant / f r i ct i on descri bi ng funct i on, coupl ed t o a single mass and PI D TABLE 3. DEFINITIONS AND VALUES FOR THE PARAMETERS OF EQUATION (18) AND FIGURE 29 Quantity Symbol Value in Figure 29 Mass M 1.0 Damping b 0.1 Position gain k v 10 Velocity gain k,, 5 Integral gain k, 10 or 100 A s ur ve y o f f r i ct i on a n d c ont r ol s 1103 control. He derives t he predictions of t he describing function analyses for all possible par amet er combi na- tions, and compares these with analytic results. Canudas de Wit et al. (1991) empl oy describing functions t o study t he behavi or of adapt i ve systems in the presence of negative viscous friction and find t hat over compensat i on can lead t o instability. Branden- burg and Schiller (1987, 1988a, 1991) have applied describing functions t o t he case of unidirectional sliding by local iinearization of the Stribeck curve and application of t he Popov stability criterion. The ease of use seems t o come at consi derabl e cost however. Br andenbur g (1986), Br andenbur g and Schiller (1987) and Townsend and Salisbury (1987) bot h report extensive simulations done t o verify t he predictions of describing funct i on analysis, implicitly recognizing t hat t he describing function analysis adds a layer of approxi mat i on above and beyond t he approxi mat i ons within t he friction model itself. Townsend and Salisbury (1987) report : "Dynamic simulations show that the SIDF (single input describing function) predictions for Coulomb friction are qualitatively useful though quantitatively inaccurate. The (single input describing function) predictions for stiction become even qualitatively incorrect." Brandenburg and Schiller have carried out an extensive series of studies of a t wo mass, flexible system with multi-loop feedback; and report t hat t he describing function analyis (ext ended t o t he harmoni c balance technique) qualitatively agrees with t he results of simulation for t he case of Coul omb friction and unidirectional sliding, but fails substantially in t he cases of Coul omb + static friction or integral cont rol (Brandenburg, 1986; Br andenbur g and Schiller, 1987, 1988a, b, 1989, 1991; Schiller and Brandenburg, 1990, 1993). The aut hors at t ri but e this t o t he infinite-valued branch at ~ = 0 and, as Townsend and Salisbury also point out , the oscillation is not well approxi mat ed by its fundament al harmoni c. Ami n (1993) shows t hat , for the case of a single mass and PI D cont rol , t he memoryless el ement describing funct i on will predict stick slip if and only if t he system wi t hout nonl i near friction is unstable. Figure 29 is an exampl e of this, the linear port i on of t he system is unstable when ki = 100. A furt her challenge arises in represent i ng static friction by a memoryl ess el ement describing function. The describing funct i on is t he compl ex gain bet ween a sinusoidal input and the fundament al harmoni c of t he out put of a nonl i near el ement . The sinusoidal input spends zero time at zero, thus any phenomena occuring precisely when t he input is equal t o zero will be model ed as having zero extent. When friction is model ed as a function of velocity, static friction will make no cont ri but i on t o t he fundament al harmoni c of the out put , and thus cannot be represent ed by t he describing function. These challenges t o application of t he memoryl ess el ement describing function are consequences of a deeper issue, illuminated by Ehr i ch' s effort t o construct two describing functions: one a funct i on of applied force and t he ot her a funct i on of velocity Command Velocity (Input to the Nonlinear Element) Applied Force (Output o f the Nonlinear Hew, eat ) Flo. 30. A single mass system with friction, friction mo de l e d as a function of velocity and applied force. (Ehrich, 1991). The issue is this: Friction is neither a function of velocity nor a function of applied force, but of both. When the body is sliding, friction---in a simplified model - - i s a single val ued function of velocity; but when t he body is at standstill, friction is not friction at all, properl y speaking, but is a constraining force: a function of applied force (Pol ycarpou and Soom, 1992). Virtually all aut hors who have under t aken simulation have reflected this fact with some f or m of switching function t hat handles t he case of ~ = 0. The early investigators (Tou and Schultheiss, 1953; Sat yendra, 1956; Silverberg, 1957; Shen, 1962; Woodwar d, 1963) did not empl oy t he memoryl ess element const ruct i on, but devel oped describing functions for t he composi t e of t he el ement s from t h e force input t o t he velocity out put , as shown in Fig. 30. With this approach a describing function can be worked out by piecewise i nt egrat i on of t he response. This procedure, illustrated in Fig. 31, gives rise t o a describing function t hat is a curve on t he compl ex plane, or a family of curves for t he case of Coul omb + static friction, as illustrated in Fig. 32. Her e t he family of complex valued describing funct i on curves is paramet eri zed by t he ratio of the static + Coul omb friction and pl ot t ed as - 1 / N ( A , to), as described in equat i on (16). This analysis predicts t hat the s e r i e s compensat or will exhibit stick-slip while the parallel c,A TEL#/ \ x ",t I / \ - ~ / D \/ / i / ,, \ b c d f i j FIo. 31. Piecewise evaluation of the out put o f a s ys t e m pl us friction element wi t h s i nus oi dal applied force and periods o f standstill [from Tou and Schultheiss (1953), c our t e s y o f t he publisher]. 1104 B. ARMSTRONG-HI~LOUVRY e t a l . T 1 ~. 1. 5 ~ " 1 . 2 5 1 1 - 5 ~1 - 3 - 2 1 FIG. 32. Nyquist diagrams showing complex valued describ- ing function of an integrated plant/friction model. The family of curves corresponds to different ratios of static: Coulomb friction [from Tou and Schultheiss (1953), courtesy of the publisher]. c o mp e n s a t o r will not ; a pr e di c t i on ver i f i ed in t he l a b o r a t o r y ( To u , 1953). Ami n (1993) has r ecent l y r evi si t ed t he i nt e gr a t e d pl a nt / f r i c t i on des cr i bi ng f unct i on of T o u a nd Schui t - hei ss and c o mp a r e d its pr e di c t i ons wi t h e xa c t anal yt i c resul t s. He fi nds t ha t in spi t e of t he r e p r e s e n t a t i o n of st andst i l l and st at i c f r i ct i on, To u ' s c ons t r uc t i on is not , in ge ne r a l , abl e t o c or r e c t l y pr edi ct t he p r e s e n c e o f a limit cycl e in t he i deal i zed si ngl e ma s s , Co u l o mb + st at i c f r i ct i on, P I D c ont r ol l e d s ys t em. Th e b r e a k d o wn a p p e a r s t o r e l a t e t o t he p o o r r e p r e s e n t a t i o n o f t he f or ces and mot i ons dur i ng st i ck slip by a si nusoi dal a ppr oxi ma t i on. Us i ng t he i nt e gr a t e d f r i c t i on/ pl a nt c ons t r uc t i on, To u and Schul t hei ss (1953) s t udy a s ys t e m whi ch i ncl udes a t ype I c ont r ol l e r c ons t r uc t e d of a na l og c o mp o n e n t s , a nd Co u l o mb or Co u l o mb + st at i c f r i ct i on. Li mi t cycl i ng is pr e di c t e d; t he a ut hor s concl ude t ha t it can be ext i ngui s hed by t he addi t i on o f sufficient vel oci t y f e e d b a c k , whi ch t he y d e mo n s t r a t e in t he l a bor a t or y. Sa t y e n d r a (1956) s t udi ed me c h a n - i sms wi t h ba c kl a s h as wel l as Co u l o mb + st at i c + vi scous f r i ct i on. As di d To u , Sa t ye ndr a s t udi ed st abi l i t y at z e r o vel oci t y. Si l ver ber g (1957) st r i ves t o s e p a r a t e t he descr i bi ng f unct i on o f a Co u l o mb + st at i c + vi scous f r i ct i on e l e me n t i nt o f r e que nc y d e p e n d e n t a nd a mpl i t ude d e p e n d e n t pa r t s , whi ch gr e a t l y ai ds i nt e r pr e t a t i on. Thi s ef f or t is ne c e s s i t a t e d by t he f act t hat he is wor ki ng wi t h an i nt e gr a t e d pl a nt / f r i c t i on e l e me nt , such as in Fig, 30. Shen (1962) a nd Br a n d e n b u r g a nd Sch/ i fer (1987, 1988a, 1991) ha ve s t udi ed s ys t e ms wi t h s t a t i c + Co u l o mb f r i ct i on t r a c ki ng a r a mp pos i t i on i nput . T h e shi ft in e mpha s i s f r o m t he s t a t i ona r y t o t he sl owl y movi ng p r o b l e m is i mp o r t a n t . Th e des cr i bi ng f u n c t i o n - - wh e t h e r me mo r y l e s s or i nt e gr a t e d wi t h t he p l a n t - - c h a n g e s whe n a s t e a dy vel oci t y is i nt r oduc e d. Thi s occur s b e c a u s e t he si nusoi dal i nput ( vel oci t y or f or c e ) mus t be s u p e r i mp o s e d on a DC l evel , shi f t i ng its r el at i on t o t he ( f r i c t i on- ve l oc i t y) cur ve. For t he me mo r y l e s s e l e me n t c ons t r uc t i on o f Fi g. 27, t he shi ft o f l evel will ups e t t he o d d f unct i on p r o p e r t y o f t he nonl i ne a r e l e me n t , a nd t he des cr i bi ng f unc t i on, N ( A ) , will be c ompl e x r a t h e r t ha n st r i ct l y r eal . Br a n d e n b u r g get s a r o u n d t hi s p r o b l e m by f or mi ng t he des cr i bi ng f unct i on of a local l i near i zat i on o f t he St r i beck cur ve. He finds t he r esul t s t o gi ve pr edi ct i ons in qual i t at i ve a g r e e me n t wi t h o b s e r v e d be ha vi or . Shen (1962) pr e s e nt s a des cr i bi ng f unct i on f or nonl i near , l ow- vel oci t y f r i ct i on. He veri fi es t he rel i abi l i t y of his ( a p p r o x i ma t e ) descr i bi ng f unct i on anal ysi s by c ompa r i s on wi t h t he r esul t s of al gebr ai c a na l ys i s - - i nt e gr a t i on o f e qua t i ons of mo t i o n t hr ough a slip c y c l e - - f o r t he case Co u l o mb f r i ct i on. He finds r oughl y 10% di f f er ence in t he r ange of t he p a r a me t e r s f or whi ch st i ck slip is pr e di c t e d by t he i nt e gr a t e d pl ant / f r i ct i on des cr i bi ng f unct i on and al gebr ai c anal yses. None o f t he a ut hor s o f t hi s pe r i od car r i ed out ext ens i ve s i mul at i on. 3. 2. Al ge br a i c Anal ys i s An a l t e r na t i ve a p p r o a c h t o pr edi ct i ng st i ck slip lies in i nt e gr a t i ng t he e qua t i ons of mo t i o n t hr ough a slip cycl e a nd t he n d e t e r mi n i n g wh e t h e r t he s ys t em ar r i ves agai n in t he st uck condi t i on. Ar r i va l in t he st uck condi t i on ma y be d e t e r mi n e d by a t est on s ys t em st at e ( De r j a g u i n e t a l . , 1957; Shen, 1962; Shen and Wa ng, 1964; Co c k e r h a m and Col e, 1976; Ami n , 1993), or by a t es t on s ys t e m e ne r gy ( Ar ms t r o n g - Hr l o u v r y , 1992, 1993). Th e possi bi l i t y of st abl e mot i on can al so be d e t e r mi n e d by i nvest i gat i ng t he st abi l i t y of equi - l i br i um poi nt s of t he s ys t em dynami cs ( Du p o n t , 1994). Wh e n a Co u l o mb , or Co u l o mb + vi scous f r i ct i on mo d e l is a ppl i e d dur i ng sliding t he e qua t i ons of mo t i o n ma y be i nt e gr a t e d exact l y ( De r j a gui n e t a l . , 1957; Shen, 1962; Shen and Wa ng, 1964; Co c k e r h a m a nd Col e , 1976). St at i c f r i ct i on ma y t hen be mode l e d as a modi f i cat i on of t he initial condi t i on of t he t r a j e c t or y. Cons i der i ng s t a t i c + C o u l o m b + vi scous f r i ct i on, De r j a gui n e t al . pr ovi de p e r h a p s t he cl ear es t e xpos i t i on o f t hi s t echni que. Th e a ut hor s focus on a s e c o n d - o r d e r s ys t em ( a na l ogous t o a si ngl e mas s wi t h PD cont r ol ) and s t udy t he i nf l uence of rising st at i c f r i ct i on. Appl yi ng di mens i onal anal ysi s, i nt egr at i ng t he e qua t i ons o f mo t i o n dur i ng sliding and de t e c t i ng s t i c k - s l i p by e xa mi ni ng wh e t h e r t he t r a j e c t or y ar r i ves agai n at z e r o vel oci t y; t hey de t e r mi ne t hat t he r e is a cri t i cal l evel o f st at i c f r i ct i on bel ow which s t i c k- s l i p will be e xt i ngui s he d, and t hat t hi s critical l evel can be e xpr e s s e d as an i mpl i ci t f unct i on of des i r ed vel oci t y a nd s ys t e m p a r a me t e r s : 0 ~ ( ~ S ( 1 - ~ t a n . - - ~ . 0 _- O2) ) = In ( k/ q~ - 2q~0 + 1), (19) whe r e , f ol l owi ng t he not a t i on o f De r j a g mn , k , , / m t o = ~ / k p / m ; 0 = 2 t o AF, AF,. m v t o v V ~ / , and kp is s ys t em st i ffness; m is mass; k,, is vel oci t y f e e dba c k or da mpi ng; AF is t he excess o f b r e a k a wa y fri ct i on o v e r Co u l o mb f r i ct i on and AFt is t he cri t i cal val ue r e qui r e d f or st i ck slip; 0 is di mens i onl es s da mpi ng; and qL is di mens i onl es s excess o f b r e a k a wa y A s ur ve y o f f r i ct i on a n d c ont r ol s 1105 24 22 20 l g 16 14 10 $ 6 4 -- ~'t~--a 2 1 i r a 0 0 0.2 0.4 0.6 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0 L5 L 0 I I I I 3.5 2.0 2.5 1.0 1.5 0 l o g O Flo. 33. Variation of critical parameter, $~, with damping parameter 0; : according to exact equation (19); . . . . : according to the first-order approximation $, ~ ~' 4"~; [from Derjaguin et al. (1957), courtesy of the publisher]. friction over Coul omb friction. The relationship bet ween c and 0 given implicitly by equat i on (19) is pl ot t ed in Fig. 33. Rising static friction is model ed as a modification of AFt. Derj agui n e t al. (1957) find, as t he experi ment s of Rabi nowi cz (1951, 1958) and ot hers had shown, t hat t he abrupt extinction of stick slip with increasing average velocity is strongly influenced by t he t i me const ant of t he rising static friction. This work is also not abl e for its use of dimensional analysis. Shen (1962) carries out an exact analysis as a met hod of verifying t he results of a modified describing funct i on analysis. He finds t hat f or low velocity ramp inputs t o a PD cont rol l er, i.e. slow tracking, stick slip will be observed under a wide range of conditions and may be extinguished by sufficient derivative feedback. Shen and Wang (1964) ext end this result t o systems with an integral cont rol t enn. They observe t hat t he dampi ng requi red for stabilization is a decreasing funct i on of desired velocity and pr opose a variable st ruct ure system compri si ng a very high-gain velocity feedback with saturation. The saturation serves t o reduce t he i mpact of high-gain velocity feedback on t he system above those velocities where stick slip is observed. Int egral feedback with a deadband is also consi dered. Because it is useful t o achieve high t racki ng accuracy at velocities above t he stick-slip range, a deadband is proposed t hat is a funct i on of system velocity. Const ruct i ons such as these are also f ound in industrial applications. Cocker ham and Col e (1976) consi der a model t hat is based on t he dat a of Bell and Burdeki n (1969); which shows a friction curve as shown in Fig. 34 [linearized by Cocker ham and Col e (1976)]. This friction curve is an approxi mat i on t o t he Stribeck curve with frictional memor y. As a linear approxi ma- tion, it has t he i mport ant pr oper t y of maki ng analytic results possible. Ami n (1993) has consi dered Coul omb + static + viscous friction, PI D cont rol and the pointing task (see Section 4). By consi deri ng mot i ons bet ween t he instant after breakaway and t he instant before stick, Friction t ore UC T! ! ! I I ' Slip s . velocity FIG. 34. Friction curve showing linearized Stribeck friction and frictional memory ]from Cockerham and Cole (1976), courtesy of the publisher]. and establishing t he solution t o t he equat i ons of mot i on for all possible combi nat i ons of paramet ers, Ami n finds t hat a PI D, position-control system will hunt when t here is static friction. And t hat it will do this for all combi nat i ons of paramet ers for which t he linear port i on of the system is stable. He also finds t hat t he system will not hunt for any stable combi nat i on of paramet ers and Coul omb friction alone. The result establishes an i mpor t ant baseline for t he study of describing funct i on and ot her approxi- mat e analysis met hods. Armst rong-H61ouvry (1991, 1992, 1993) has applied first-order pert urbat i on t heory t o a two degree of freedom system with PD cont rol and friction described by a Coul omb + Stribeck + frictional memor y + rising static friction model . He studies a tracking probl em comparabl e t o t hat of Derj agui n e t al. (1957) or Shen (1962). In this analysis t he t raj ect ory of t he system unper t ur bed by friction is used to det ermi ne t he energy transactions due t o t he full friction model. The influence of frictional memor y and rising static friction are mapped out and t he required dampi ng for stable mot i on is f ound as a function of system stiffness (position feedback gain) and desired velocity. The predictions of t he analysis are verified by simulation and experi ment . The frictional memor y is f ound t o domi nat e t he extinction of stick slip in very stiff systems, and rising static friction t o domi nat e as t he average velocity increases. A test on the total change in system energy during a slip is used t o detect t he presence of stick slip; and gives the prediction t hat steady mot i on will obt ai n when: "2 2 _ 2 : r p ~2 + , , 1 ( F * ~ F s . h n ~ d A ~ < 7 2 \ "'h,, 1 + $, ~/ (21) where, following t he not at i on of Armst rong- H61ouvry (1993), 19 is dimensionless dampi ng; Sd is dimensionless desired velocity; F*hn is dimensionless excess breakaway friction and is influenced by rising static friction; and A~t, reflects t he ener gy cont ri but i on due t o Stribeck friction with frictional memor y, and is given by an integral t hat is a funct i on of t wo paramet ers and must be eval uat ed numerically. P, 6d and F~b~.are respectively anal ogous t o 0, v and AF in equat i on (19). The resultant stick-slip extinction boundar y f or t he case of a specific r obot 1106 B. ARMSTRONG-HF.LOUVRY e t a l . A O- r
]D 0 (/3 2 0 0 0 . 0 1500.0 i , - i Data; No Stick Slip Data; Stick Slip - Theoretical Extinction Boundary 1000.0 500.0 ~ . 0.0 a , ~, 0. 0 5. 0 10.0 t 5. 0 Dimensionless Velocity FIG. 35. Experimental stick-slip data; the contour is the stick-slip extinction boundary calculated with equation (21). Dimensionless velocity is given by ~d = kd/SC~, cf. equation (10) ]from Armstrong-H6louvry (1993)]. arm, along with experimental dat a, are present ed in Fig. 35. Sout hward e t al . (1991) propose a variable st ruct ure controller comparabl e t o sliding mode cont rol (Ut ki n, 1977). A hard switch in cont rol action about ~ = 0 is used t o overcome the frictional nonlinearity. A Lyapunov pr oof of stability is achi eved with the use of Dini-derivatives t o handle the discontinuity in friction at k = 0. The analysis requires onl y t hat an upper bound on friction be known. I mpl ement at i on problems, such as chat t er or the i mpact of cont rol l er and act uat or bandwi dt h limitations, are not ad- dressed. But sliding mode cont r ol - - sui t abl y modified t o be i mpl ement abl e- - has been remarkabl y successful elsewhere (Siotine, 1984). Heck and Ferri (1991) apply singular pert urbat i on t heory to a fourt h-order model of a t urbi ne system with a Coul omb friction model , and find t hat a first-order correction t erm is sufficient to produce a system simulation t hat agrees well with a full model simulation. Their effort is focused on model or der reduction. The investigations not ed above have sought t o establish the presence of stick slip by testing for the existence of a stable limit cycle. For constant-velocity sliding mot i on, an alternative approach is t o consi der the stability of the steady-sliding fixed point. I f t he fixed point is stable and its domai n of attraction is large enough, the possibility of smoot h sliding is established. Rice and Rui na (1983) and Dupont (1994) have pursued this approach to illuminate t he interaction of stiffness and frictional memory. Thei r contribution is the demonst rat i on of the role of frictional memor y in the stability of steady sliding. 3.3. Phase Plane Analysis The phase plane may be used t o graphically represent the trajectories of systems, linear and nonlinear. An exampl e is shown in Fig. 36. In cases where theoretical restrictions on t he charact er of trajectories are not possible, the phase plane serves as a graphical present at i on of the results of simulation; and can be said t o demonst rat e general results about the system when trajectories spanni ng a sufficient range of initial conditions are explored. In ot her FIG. 36. Phase trajectories of a PD regulator and mass with Coulomb friction [from Southward et al. (1991), courtesy of the publisher]. cases, where theoretical restrictions on t he behavi or of trajectories are possible, the phase plane serves t o organize and clarify the interactions and consequences of these restrictions. In Fig. 36, for example, all trajectories lying sufficiently near t he origin can be shown t o ent er t he set of multiple equilibria, E e o . The dimensionality of t he phase plane is t he or der of the system state. A mass with PD cont rol thus has a two-dimensional phase plane, one that can be easily realized and interpreted. The same system with integral cont rol , however, has three dimensions, and might already be difficult t o visualize. A two mass system with flexibility and integral cont rol , such as t hat consi dered by Br andenbur g e t a l . , e.g. Branden- burg (1986) has five states and is probabl y not usefully studied with the phase plane. In two dimensions there are a number of strong t heorems belonging t o phase plane analysis; for example, the Poi ncar r - Bendi xson t heorem (Slotine and Li, 1991): If a trajectory of a second-order system remains in a finite region r , then one of the following is true: (a) the trajectory goes to an equilibrium point, (b) the trajectory tends to an asymptotically stable limit cycle, (c) the trajectory is itself a limit cycle. This result and several ot hers, however, do not ext end to higher dimensions. Many aut hors have used phase planes t o illustrate various points about system trajectories (Shen, 1962; Shen and Wang, 1964; Kubo e t a l . , 1986; Townsend e t a l . , 1987; Radcliffe and Sout hward, 1990; Sout hward e t a l . , 1991; Ar mst r ong- Hr l ouvr y, 1992). Radcliffe and Sout hward (1990) go beyond illustration and use the phase plane t o make several i mport ant points. They are concerned with t he ability of various friction models to predict stick slip, and show t hat all trajectories of an otherwise stable linear system with Coul omb friction and PD cont rol will converge t o a set of multiple equilibria, and will thus fail t o show stick slip. Ext endi ng the phase plane t o t hree dimensions, the aut hors proceed t o show t hat a system with Coul omb friction, and typical cont rol l er gains and friction values, cannot exhibit hunt i ng (back and forth oscillation about t he set point) with PI D cont rol . A s ur ve y o f f r i ct i on a nd c ont r ol s 1107 Y Yu H XL" . " ~ FIG. 37. Parallel trajectories inside the sticking band S have zero acceleration until they reach the border ]from Radcliffe and Southward (1990), courtesy of the publisher]. for p equals 0.01, 0.7, 2.0 and 4.0 1 500.0 , , , 200 1 o o o . o c 5 0 0 . 0 0 . 0 ' ' ' 0. 0 2. 0 4. 0 6. 0 8. 0 0. 0 Di mens; onl ess Vel oci t y FI6. 38. Plot of boundary above which there is no stick-slip, determined by 8000 simulation trials, p is the dimensionless damping coefficient, equation ( 21) ]from Armstrong- H~louvry (1991)]. This is done by investigating the properties of a two- dimensional slice t hrough the three-dimensional phase space. The slice yields t he x - y plane, as shown in Fig. 37; y is the integral er r or state. The phase space is partitioned i nt o six regions, trajectories in each region are shown t o ent er anot her regi on; and t he map from one region ont o itself, similar t o a Poincar6 map, is established. Specifically, t he posi t i on-error coordi nat e at breakaway is mapped t o the posi t i on-error coordi nat e at t he next breakaway event. Conver- gence, limit cycling or divergence are det ermi ned by establishing cont ract i on, stability or expansion in this one-di mensi onal map. Int egrat i ons in the met hod are per f or med numerically, and so onl y for specific combi nat i ons of paramet ers. In t he cases studied, it is shown t hat excess breakaway friction over Coul omb f r i ct i on- - bot h static and Stribeck friction model s are consi der ed- - i s requi red t o excite stick slip. The authors propose t hat friction models more sophisti- cated t han t he Coul omb friction model need t o be studied t o predict even t he qualitative behavi or of frictional limit cycles. 3.4. Analysis by Simulation For cont rol l er analysis and synthesis, the tools described in t he previ ous sections are applied t o system models. When the models permi t exact analysis, the validity of the results depends onl y on t hat of the model. When mor e compl i cat ed systems are analyzed with approxi mat e techniques, such as describing functions, t he validity of t he results rests not onl y on the validity of model , but also on t he assumptions of the analysis technique. Simulation provides a means of verifying bot h models and analysis techniques. Compari son of experi ment with simulation can be used t o validate models. Compari son of analytic results with simula- tion can be used t o validate approxi mat e analysis techniques. While simulation can certainly be abused, when used in conj unct i on with bot h experi ment and analysis, it is a powerful tool f or closing t he l oop between t hem. A great deal can oft en be learned by simulating a handful of carefully chosen trajectories. Alternatively, by performing hundreds or t housands of simulation trials, one can numerically map stability boundari es in paramet er space or domai ns of at t ract i on in state space. Figure 38 provi des an exampl e of t he former. Here, t he "stick-slip extinction boundar y" is pl ot t ed as a function of system paramet ers. A valid analysis technique should be used t o verify these numerically comput ed cont ours. By extensive simulation, one gains a measure of par amet er sensitivity and experiment/analysis verification which is lacking when only a few trajectories are simulated. The increasing availability of massively parallel comput ers makes this approach practical and efficient. I n t he following paragraphs, t he special probl ems associated with simulating friction are present ed along with a discussion of solution techniques. For systems with multiple degrees of f r eedom, such as robot s, the rigid-body dynami c equat i on including friction can be written in the form: = D(x)Ji + h( x, 2) + f ( x , Yc, it). (22) The vectors of joint displacements and act uat or t orques are x and ~, respectively. Their di mensi on equals t he number of degrees of freedom of t he mechanism. The confi gurat i on-dependent inertia matrix is denot ed by D. It is bot h symmet ri c and positive definite. The vect or h consists of centrifugal, Coriolis and gravity terms. The vect or f includes all friction t erms and is a function of joint positions, velocities and accelerations. The forward dynami cs probl em is t o solve for t he joint positions, velocities and accelerations given t he input torques or forces and t he initial conditions. This is the probl em of simulation. At each time step, t he known joint t orques, positions and velocities are used t o comput e t he joint accelerations. I n the absence of friction, this typically involves solving a set of linear algebraic equat i ons for t he accelerations. Usi ng t he values of acceleration and velocity, numerical integration yields the velocity and position at t he next time step. By considering t he st andard Coul omb friction equat i on, we can gain insight i nt o t he comput at i onal issues involved in simulating a br oad class of friction models. I ndependent of t he area of cont act , t he 1108 B. ARMSTRONG-HI~LOUVRY et al. Coul omb friction force always opposes relative mot i on and is proport i onal t o the normal force of cont act . This force can be expressed as Fc =/~ IFNI sgn (-fr), (23) where /l is the coefficient of friction, FN is the normal force and 2r is the relative sliding velocity. The signum function is defined: +1, . f >O sgn (.~) = O, 2 = O. (24) - 1 , 2 < 0 Due t o its dependence on t he sign of velocity, the friction force is discontinuous at zero velocity. This indicates t hat the gover nmg differential equat i ons are discontinuous in the highest or der derivative terms. In addition, as indicated by f(x, k, $), the normal forces in machine component s can depend not onl y on link positions and velocities, but also on accelerations. Because this dependence is oft en nonlinear, equat i on (22) can onl y be solved explicitly for $ in special cases. Bot h difficulties are discussed in detail in t he following paragraphs. 3.4.1. Di s cont i nuous f r i ct i on mode l s When integrating discontinuous ordi nary differential equat i ons, t he appropri at e value of the derivative must be used on each side of a discontinuity. Unf or t unat el y, discontinuities generally occur inside i nt egrat i on subintervals. A st andard t echni que is t o empl oy switching functions which flag the presence of a discontinuity in t he last subinterval. For the initial value probl em, 2 = f ( x , t ) x ( 0 ) = x . a switching function, q)(x, t), is defined such t hat q~(x, t ) = 0 when f ( x , t ) is discontinuous and (p(x, , t , ) . dp(x,+l, t , +~) <0 implies a discontinuity in the subinterval x , ~- x <- x, + ~ (Fatunla, 1988; Dupont , 1993). For a velocity zero crossing, sgn (2,) is such a function. In addition t o detecting a discontinuity, the integrator must also provi de a mechani sm for locating the point of discontinuity within t he subinterval. Next, integration up t o the point of discontinuity is r epeat ed and t hen the integration rout i ne is rest art ed from t he discontinuity using the appropri at e derivative value. Initially, small steps should be t aken t o accurately capture any transients which follow the discontinuity. If the friction model includes stietion, the i nt egrat or must also include tests t o det ect when sticking occurs. Variable-step-size, vari abl e-order met hods are ap- propri at e for integrating di scont i nuous equations. For synchronization purposes in real-time simulation and cont rol , however, Morgowi cz (1988) suggests the use of fixed-step-size met hods. By choosi ng the cont rol l er period as a multiple of the fixed-step-size, the simulated machine state is available at cont rol l er sampling times. To locate discontinuities occurri ng during the previous subinterval, Morgowi cz (1988) uses linear interpolation. Approxi mat e values of the state derivatives on bot h sides of the discontinuity are comput ed. They are used t o reintegrate the subinterval in one step. Unless very fine mot i ons are under consideration, simplifications of this type can give quite adequat e simulation results. The time saved in simulating a given t raj ect ory will depend on the number of velocity zero crossings involved. 3.4.2. Al t er nat e f ri ct i on model s A number of researchers have proposed alternate friction models with the goal of produci ng accurate results while minimizing algorithm complexity and simulation time. A general approach is t o replace the discontinuity of the s t at i c+ Coul omb model by a curve of finite slope (Threlfall, 1978; Bernard, 1980; Rooney and Deravi, 1982; Haessig and Friedland, 1991). This t ype of model eliminates the need t o search for the switching point within an integration subinterval. I f the slope is large, however, small step sizes are needed and t he numerical integration remains slow. More i mport ant l y, these models do not provi de a t rue stiction mode. The system creeps t hrough zero velocity instead of sticking. This effect may be i mport ant when the period of the stick-slip limit cycle is long. Several techniques have been proposed to include stiction while still avoiding the search for the switching point. In one met hod, best described in Karnopp (1985), friction is given by: - s g n (k)Fc I,~1 > Dv F r ( ~ , F) ( 2 5 ) / t _sgn ( F) max (F, F, ) I.tl - Ov" A small nei ghborhood of zero velocity is defined by Dr , as shown in Fig. 39 (Bernard, 1980; Kar nopp, 1985; Johnson and Lorenz, 1991; Younki n, 1991). Outside this nei ghborhood, friction is a function of velocity. Inside the nei ghborhood, velocity is con- sidered t o be zero and friction is force dependent . As long as the resultant force is less t han the maxi mum stiction force, the small velocity remains const ant within the +Dr nei ghbor hood or is set to zero. Kar nopp' s (1985) i mpl ement at i on of this met hod for a block of mass m sliding on a flat frictional surface with an applied force, F, is depi ct ed in t he block diagram of Fig. 40. The deadband of the gain block bet ween the two integrators forces the velocity t o exactly zero during sticking. These models allow discontinuity of static t o Coul omb friction force at the nei ghborhood boundary. ~ z D V ~ FIG. 39. Karnopp's friction-velocity model, By allowing stiction within the interval +Dr, the integrator does not have to search for velocity zero crossings [from Karnopp (1985), courtesy of the publisher]. A s ur ve y o f f r i ct i on a n d c ont r ol s 1109 l:m F ( a) (b) P = ! S I F s ~ k ~ 7 (c) FIo. 40. Block diagram of Karnopp's friction simulator for a force F, applied to a block of mass m moving on a flat frictional surface [from Karnopp (1985), courtesy of the publisher]. A second appr oach is based on experi ment al observat i on of presliding displacement, t hat is, at velocity reversals, friction may be mor e appropri at el y model ed as a cont i nuous function of displacement. This can be represent ed graphically with a hysteresis l oop as shown in Fig. 41. For small displacements, this can be i nt erpret ed as t he straining and event ual rupt ure of many small bonded cont act s bet ween t he t wo sliding or rolling surfaces (Dahl , 1968, 1977; Threifall, 1978; Haessig and Friedland, 1991). I f very Iza P.R Fr i ct i onol | o r c e / changes J si gn Displacement J Fio. 41. Friction-displacement hysteresis loop [from Threl- fall (1978), courtesy of the publisher]. small displacements are t o be accurately simulated, these posi t i on-dependent model s could be mor e accurate t han a vel oci t y-dependent model. Threlfali (1978) and Haessig and Fri edl and (1991) propose switching bet ween position- and vel oci t y-dependent models. For large displacements, Coul omb friction t oget her with spring stiffness can somet i mes be model ed in a similar fashion (Bernard, 1980). 3.4.3. L o a d - d e p e n d e n t f r i c t i o n The significance of varying applied loads in machine friction has not been t horoughl y studied. Some experi ment s have suggested that, in certain mechan- isms, friction does not exhibit a measurabl e dependence on the transmission and reaction loads which det ermi ne t he frictional normal forces (Armst rong-H61ouvry, 1991; Canudas de Wit e t a l . , 1991). Ot her experi ment s indicate t hat it can be an i mport ant fact or in transmission el ement s (Newman e t a l . 1992; Dohr i ng e t a l . 1993; Dupont , 1993). The situations are distinguished by the degree of prel oadi ng of t he elements. I n general , t he inclusion of a l oad-dependent Coul omb or static friction t erm, f ( x , / c , J i ) , in the dynami c equat i on (22), renders it implicit in t he joint accelerations, $. The forward solution requires an iterative root-finding procedure (such as a modified Newt on met hod) at each step of the integration t o comput e the accelerations. The cause of t he implicitness is the dependence of friction on t he magni t ude of t he normal force. The normal force itself is a function of the resultant force and moment at a joint. Expressed in a local coordi nat e frame, t he component s of the resultant force and moment can be formul at ed in terms of t he joint positions, velocities and accelerations. These component s will be affine transformations of t he accelerations ( Dupont , 1993). If the direction of t he normal force is const ant in the local coordi nat e frame, t he normal force can be expressed as a function in which t he net force and moment component s appear linearly. This is t rue for friction in translational joints and certain transmis- sions. Since t he sign of the normal force can change, its absolute value must be used t o obt ai n its magnitude. When the direction of the normal force is not const ant in a local joint coordi nat e frame, the magni t ude of the normal force will involve t he square root of sums of squares of net force and moment component s. As an exampl e, consider a radially l oaded revol ut e joint. More generally, there are multiple bearings at a joint and one must consider joint geomet r y and reaction t orques as well as reaction forces (Gogoussis and Donat h, 1988). Thus, l oad-dependent friction involves t he absolute value or square root of sums of squares of accel erat i on-dependent terms. Substituting ei t her t ype of expression into t he original dynami c equat i ons renders t hem implicit in t he accelerations. Conse- quently, it is necessary t o solve iteratively for joint accelerations at each time step of a simulation. To minimize t he number of iterations, Threlfail (1978) proposes a pr edi ct or - cor r ect or met hod which 1110 B. ARMSTRONG-HI2LOUVRY e t a l . uses the reaction forces from the previous seven time steps to predict the new values. Morgowi cz (1988) uses only the reaction forces from the precedi ng time step. Bot h aut hors indicate t hat convergence is usually obt ai ned in two iterations. The cost of implicitness is significant, however. Using two iterations effectively requires two forward solutions in addition to solving the inverse probl em once for those force and t orque component s needed to comput e the normal forces. To avoid an iterative solution, Gogoussis and Donat h (1990) propose a hybrid analog/digital comput er i mpl ement at i on solv- ing the implicit port i ons in t he anal og comput er. Anal og simulations are discussed by Cocker ham and Cole (1976) and by Dahl (1968) al t hough nei t her paper considers varying normal force. For those cases in which the magni t ude of frictional normal force involves onl y an absolute value, there is a finite number of possible solutions based on the combi nat i ons of signs of normal forces. By adopt i ng an efficiency formul at i on and tracking the signs of the normal forces using switching functions, the number of iterations can be minimized ( Dupont , 1993). This is an i mport ant case because it applies t o transmission el ement s which, when present, often domi nat e machi ne friction. 3.4.4. E x i s t e n c e a n d u n i q u e n e s s In t he absence of friction, t he rigid-body forward dynami c equat i ons can be shown t o possess a unique solution for t he accelerations at each time step. For most cases of interest involving low-friction mechan- isms, rigid-body model s yield a unique solution as well. For sufficiently high coefficients of friction, however, systems with l oad-dependent friction can be shown to possess ei t her multiple consistent solutions or none at all. The probl em arises because the normal -force reactions cannot be written as functions of t he system state, but instead must be expressed in terms of t he unknown accelerations. Since the resulting set of equat i ons is nonl i near in the unknowns, it is not surprising t hat solution existence and uniqueness is problematic. The existence and uniqueness probl em associated with Coul omb friction bet ween rigid bodies has been studied in Painlev6 (1895), Lr s t edt (1981), Rooney and Deravi (1982), Raj an e t al . (1987), Mason and Wang (1988), Wang e t al . (1992) and Dupont (1992a, b). Dupont (1992a) shows t hat a single degree of freedom is sufficient t o exhibit these probl ems and that, accordi ng t o the value of input force or t orque and velocity, there can be either zero, one, two or t hree feasible solutions. Rooney and Deravi (1982) demonst rat e similar behavi or with a quasistatic analysis of a slider-connecting rod mechanism. Mason and Wang (1988) address t he case of no consistent solution and model it as an impact with zero approach velocity. The hypot heses used to govern frictional impact can somet i mes violate t he principles of dynami cs such as energy conservat i on. Wang e t al . (1992) provi de a good discussion of this topic and propose discretizing the normal -force cont act zones into compl i ant patches with lumped linear stiffness t o ensure a unique solution. As did Wang e t al . (1992), Dupont (1992b) points out that the existence and uniqueness probl ems associated with l oad-dependent friction can be resolved by relaxing the rigid-body assumption. He proposes to make the otherwise ambi guous normal reactions functions of system state by introducing a lumped compliance with a component normal to the friction cont act surface. Using this approach, he addresses the case of multiple dynami c solutions and shows that for a system of finite stiffness, the "ext r a" solutions are dynamically unstable. In many situations, existence and uniqueness is not a concern. In these cases, it is a straightforward task t o develop comput er code for simulating the appropri at e friction model. Al t ernat i vel y, several commercial software packages allow for the numerical integration of nonl i near differential equat i ons includ- ing those with discontinuities. Friction simulation can be a valuable analysis tool when used judiciously and in combi nat i on with experiments and ot her analysis techniques. 3.5. Summary of Analysis Tools A general analysis t ool - - one which will illuminate both the positioning and tracking tasks; PD, PI D and ot her control structures; and i ncorporat es an adequat e friction model - - has not yet been presented. But as shown in Table 4, many partial steps have been taken. Radcliffe and Sout hward (1990) have shown t hat systems with a broad range of friction models and PD control will not exhibit hunting. This result has been arrived at by a number of routes, including those of Kubo e t al . (1986) and Wal l enborg and As t r r m (1988). Furt hermore, as Kubo e t al . (1986) have established, a system with Coul omb friction alone will not stick slip while tracking with PD control. This point was made long ago in the mechanics communi t y, and was the motivation for early proposals in the tribology literature that, based on the observat i on of self-exciting stick slip during sliding, the Stribeck friction model might apply t o a wide range of situations (Thomas, 1930). The algebraic analyses of Derj agui n e t al . (1956, 1957), and others, consider Coul omb + static friction and the tracking problem in a general way, but only a PD control; and the possibility of stick-slip is found. Ar ms t r ong- Hr l ouvr y (1993), likewise, presents a general analysis, incor- porat i ng a rich friction model , but considers only PD cont rol . The result of Wal l enborg and ,~str0m (1988) is quite rigorous and extends to full state feedback, but rests on a describing function description t hat is limited to Coul omb friction and the positioning task. The results of Tou (1953) and Shen (1962), among ot hers are specific t o particular systems and are not t horoughl y verified, but over come t he difficulties of reflecting standstill or steady motion with the describing function. Too much attention has been focused on t he simplest probl em: positioning of a machine with Coul omb friction and PD control. The repeat ed demonst rat i on t hat this case will nei t her hunt nor stick slip, coupl ed with the evident presence of limit-cycles in the l aborat ory and field, should provi de adequat e A s u r v e y o f f r i c t i o n a n d c o n t r o l s 1111 TABLE 4. PAPERS PROVIDING ANALYTIC PREDICTION OF STICK SLIP IN MACHINES; THE INDICATIONS NO OR YES REFER TO THE POSSIBILITY OF STICK SLIP DETERMINED BY THE CITED ANALYSES Task/controller Positioning Tracking Friction model PD PID or Lag* PD PID or Lag* NO NO Coulomb Kubo, NO Radcliffe, Radcliffe, 9 Amin Kubo Wallenborg NO YES YES YES Coulomb + Static Tout Tout , Amin Radcliffet Radcliffet Derjaguin:~ Shent YES NO YES Dupont YES Coulomb + Stribeck Radcliffet Radcliffet Armstrong, * Parameters giving stable linear portion assumed. t Specific system parameters considered. $ Incorporates rising static friction. Incorporates frictional memory. Tou: (Tou, 1953; Tou e t a l . , 1953) Derjaguin: Derjaguin e t al . (1956, 1957) Shen: (Shen and Wang, 1964) Radcliffe: (Radcliffe e t a l . , 1990) Kubo: (Kubo e t a l . , 1986) Wailenborg: (Wallenborg e t a l . , 1988) Dupont: (Dupont, 1994) Armstrong: (Armstrong-HAIouvry, 1991, 1993) Amin: (Amin, 1993) mot i vat i on t o move on t o t he mor e chal l engi ng cases of PI D cont r ol , t he t r acki ng t ask and a St r i beck fri ct i on model . Fur t her mor e, t o be useful in pr act i ce, anal ysi s t ool s must compr ehend not onl y t he nonl i near i t i es of fri ct i on, but t he nonl i near i t i es of cont r ol as well. Even for nomi nal l y l i near PI D cont r ol , nonl i near i t i es such as de a dba nd and sat ur a- t i on oper at i ons ar e common. The possi bi l i t i es have not been exhaust ed. The pr oj ect i ve phase pl ane t echni ques of Radcl i ffe and Sout hwar d (1990), for exampl e, coupl ed wi t h t he anal yt i c t echni ques of Der j agui n e t a l . (1957) or Ar mst r ong- HAl ouvr y (1993), mi ght yi el d mor e gener al anal ysi s t ool s. Per haps t hese r out es, or ot her s, will br i ng t ool s for t he most gener al case: t he t r acki ng t ask wi t h a gener al cont r ol st r uct ur e and adequat e fri ct i on model . It seems l i kel y t hat an anal ysi s capabl e of handl i ng t hi s case woul d compr ehend t he ot her possi bi l i t i es as speci al cases, and gr eat l y ext end our abi l i t y t o anal yt i cal l y pr edi ct st i ck sl i p in machi nes. 4. COMPENSATION TECHNIQUES FOR MACHINES WITH FRICTION An ext ensi ve body of l i t er at ur e exists r el at i ng t o fri ct i on compensat i on; t hi s sur vey, for exampl e, ci t es over one hundr ed paper s in t hi s ar ea. Thi s body of l i t er at ur e can be vi ewed as cor r es pondi ng t o t he mul t i di mensi onal char act er of t he pr obl em. As suggest ed by Fi g. 42, t he coor di nat e axes mi ght be t he mechani sm, t ask, fri ct i on model , anal ysi s t echni que, and compensat i on t echni que under consi der at i on. Many paper s in this ar ea ar e ver y specific, such t hat t hey can be vi ewed as a single pi xel in t hi s mul t i - di mensi onal space. Al l owi ng for di scussi on t hat t her e mi ght be: f our classes of mechani sms- - - r ol l er bear i ngs, sl i di ng bear i ngs with dr y or fluid l ubr i cat ed cont act s, and hydr ost at i c or magnet i c bear i ngs; f our di fferent t asks- - - pr eci si on posi t i oni ng, smoot h vel oci t y r ever sal , l ow vel oci t y t r acki ng and hi gh vel oci t y t racki ng; five di fferent fri ct i on mo d e l s - - Co u l o mb fri ct i on, Coul omb + st at i c fri ct i on, Coul omb + St r i beck fric- t i on, sliding fri ct i on wi t h fri ct i onal memor y, sl i di ng fri ct i on with rising st at i c fri ct i on; f our di fferent anal ysi s t echni ques----descri bi ng func- t i on, exact i nt egr at i on, phase pl ane and ext ensi ve si mul at i on; and seven di fferent compens at i on t echni ques- - - l ow fric- t i on machi ne desi gn or l ubr i cant choi ce, stiff posi t i on cont r ol , i nt egr al cont r ol wi t h de a dba nd, di r ect force f eedback, i mpul si ve cont r ol , Coul omb fri ct i on f eedf or war d, pos i t i on- dependent fri ct i on f eedf or war d; one finds t hat 2240 specific paper s ar e possi bl e. Despi t e t he ext ensi ve l i t er at ur e in t hi s ar ea, we ar e not awar e of any compr ehensi ve at t empt t o pi ece t oget her t he many t hr eads r unni ng t hr ough t hi s space. In compi l i ng t hi s sect i on, we had sever al goal s. The Compensation Technique FIG. 42. Axes of friction compensation space. 1112 B. ARMSTRONG-HI~LOUVRY e t a l . first was t o provi de an overvi ew of the maj or friction compensat i on techniques. More i mport ant l y, we have tried to assemble the many pieces of this probl em into their whole and to point out where work remains t o be done. In the next section, we propose a classification of friction cont rol tasks. For each task, the probl emat i c behavi or is described and the contributing friction component s identified. Following that, we describe how the cont rol probl em can somet i mes be avoi ded t hrough the selection of alternative mechanisms, materials or lubricants. Next, bot h model -based and non-model -based compensat i on met hods are present ed along with t echni ques for friction identification and adapt i ve control. The section concludes with a survey of the compensat i on techniques in current use by industry. 4.1. Compensat i on Tasks A classification of compensat i on tasks appears in Tabl e 5. Of the four tasks, one is the regul at or and t he remai ni ng t hree are versions of the tracking probl em. They are listed along with the associated cont rol l er er r or and the domi nant friction effect. A specific machi ne application could involve several of these tasks. Task !, t he Regul ator, is encount ered with positioning and pointing systems. Appl i cat i on ex- amples include telescopes, ant ennas, machine tools, disk drives and robots. In this case, a system spends most of its time ei t her near or within the stiction regime. When the fri ct i on-vel oci t y curve of a system is negatively sloped at the origin, the equilibria of a PD position regul at or consist of an interval on the position axis in phase space. Onl y one point in this set corresponds t o the desired goal position as was shown in Fig. 36. By addi ng integral cont rol , t he equilibria set consists onl y of points with the desired position (and velocity); however, this set can be unstable with nearby trajectories diverging away to a limit cycle (Radcliffe and Sout hward, 1990). This integral- induced stick-slip oscillation about t he goal position is referred t o as "hunt i ng". This task is discussed in Blackwell e t al . (1988), St ockum e t al . (1988), Yang and Tomi zuka (1988), Ost ert ag e t al . (1989), Radcliffe and Sout hward (1990), Ausl ander and Dass (1990), Brandenburg and Sch/ifer (1991) and Sout hward e t al . (1991). As t o its frictional cause, Task !I, Tracki ng with Vel oci t y Reversals, is closely allied with Task I. Due t o a higher static level of friction, mot i on t hrough zero velocity is not smoot h. A system may pause at zero velocity until sufficient force is applied t o exceed the maximum stiction level. This task is encount ered with machine tools, tracking mechanisms and robot s under position or force control. An i mport ant example of the effect of friction on this task occurs for machine tool slideways, where it is known as stand still or quadrant glitch. In multiple degree of freedom mot i on, the joint undergoi ng velocity reversal pauses while the ot hers cont i nue uni mpeded. The resulting mot i on manifests itself as an aberrat i on in the workpiece cont our. Suzuki and Tomi zuka (1991) address this probl em in machining circular paths. In many papers, tracking with velocity reversal is studied for a sinusoidal reference signal, including Gilbart and Winston (1974), Canudas de Wit e t a L (1987), Canudas de Wit and Seront (1990), Wal rat h (1984), Schiller and Brandenburg (1990) and Br andenbur g and Schiller (1991). Maquei ra and Masten (1993) have investigated tracking with spectrically broad band inputs. Task 111, Tracking at Low Vel oci ti es, differs from Task II in t hat the desired mot i on is of constant direction and perhaps const ant velocity. This task arises for machi ne tools, tracking mechanisms and robot s under position or force control. It is the task most oft en associated with stick slip. The common pictorial represent at i ons include the pin on flat apparat us of Fig. 14 and its kinematic inversion depi ct ed in Fig. 43. The potential for stick-slip limit cycling exists when the operat i ng point, Vo, lies on a negatively sloped port i on of the steady-state fri ct i on-vel oci t y curve such as Regi me I I I of Fig. 5. This task is oft en studied by addressing one of two criteria for smoot h mot i on: (1) Does a stick-slip limit cycle exist and for what system paramet er values will it be stable? V 0 tAAAA~ I T H| | k, 1/ H / / / / / / / / / / / friction interface FIG. 43. Tracking at low velocities. The free end of the spring and damper are moved at constant velocity, V~). TABLE 5. COMPENSATION TASKS Compensation task Control error Dominant frictional contributor I. Regulator (pointing or Steady-state error, Stiction. position control) hunting (limit cycle around fixed point) Stand still, lost motion II. Tracking with velocity reversal III. Tracking at low velocities IV. Tracking at high velocities Stick-slip Large tracking errors Stiction. Negatively-sloped Stribeck curve; stiction. Viscous behavior of lubricant. A sur vey of fri ct i on and cont r ol s 1113 (2) For what system parameter values is the equilibrium point i = Vo stable? The former has been addressed most often in the literature. Examples include Harmer (1952), Singh (1960), Shen and Wang (1964), Kato and Mat- subayashi (1970), Kato e t al . (1972, 1974), Armstrong-H61ouvry (1991, 1993). Investigations of equilibrium-point stability have been carried out by Rice and Ruina (1983) and Dupont (1994). Analyses involving the determination of stable stiffness and damping values can be applied in controller design by relating these quantities to position and velocity gains. Controller design for tracking at low velocities is also studied in Gilbart and Winston (1974), Walrath (1984), Kubo e t al . (1986). Task IV, Tracking at High Velocities, arises for machine tools, position-controlled robots and tracking mechanisms. High-speed operation not only increases productivity, it may actually be necessary to meet process constraints. For example, in high-speed machining, a critical cutting velocity must be exceeded in order to avoid the excessive tool temperatures which lead to premature failure (Suzuki and Tomizuka, 1991). This task is significantly different from the previous three tasks because high-velocity friction is dominated by viscous effects. The friction-velocity curve is positively sloped and stability is usually not a problem. Instead, tracking error is observed to increase as a function of velocity. For example, the radial error in machining circular contours is approximately proportional to the square of the angular-velocity feed rate (Suzuki and Tomizuka, 1991). Often, machines performing high-velocity tracking must also cope with velocity reversals. Due to the nonlinearity of friction, a linear fixed-gain controller that is tuned for low velocities may perform poorly at high velocities and vice versa. This suggests the need for nonlinear compensation, as described in Section 4.4, and for variable structure controllers, whose industrial use is addressed in Section 4.5. The effectiveness of a particular compensation technique depends strongly on the task. This is due in a large part to the task defining the dominant frictional effect. Figure 44 connects task and compensation technique pairs which have been investigated in the literature. Each of these compen- sation methods is described in the following sections. Figure 44 also lists the common applications associated with each task. The first item listed under compensation tech- niques, Friction Problem Avoidance, is not truly a compensation technique. In this method, one attempts to replace the given system with one which is easier to control. Model-based methods are distinguished from non-model-based methods in that they employ a TASK COMPENSATION TECHNIQUE TASK Friction Problem Avoidance Tracking with Mechanical Design* Vel0cl.ty.. Reversal f - - ~ - I Material Choice * "~. ~-----/ A; B 1, n; D ~ t Lubricant Choice* J / [ Non-Model-Based Compensation U I~--~--~ Stiff Position Control (1) ~ Integral Control / Deadb~ * ~ Regulator --'JJV}I ) / (Position) ~ Joint Torque Control A, C; D - ~ l / ~ f Impulsive Control / / / / / / / / o = . + j - III// o m p e n . t i o n I l l / ~ " Coulomb Friction Feedforward/~ I I / \ F b=k" I / \ ~. . ~ General Friction Feedforward/ Adaptive Feed forward/Feedback* Tracking; Tracking; High Velocity A', B ii; C Applications: A. Machine Tools B. Robotics i) Force Control ii) Position, Trajectory Control C. Disk Drives D. Gimbals; Telescopes; Military Pointing * Found in Industrial Applications (I) Feasible in gimbals, difficult in robotics, etc. (2) Principally hydraulic systems. FIG. 44. Tasks and t hei r associated compensat i on techniques as r epor t ed in the literature. Typical task applications appear in t he legend. AUTO 30-7-C 1114 B. ARMSTRONG-HI~LOUVRY e t al . friction model in a feedback or feedforward loop. The model paramet ers may be identified once t hrough a set of identification experiments. Otherwise, adaptive control techniques may be applied. 4.2. Probl em Avoi dance: Design for Cont rol While the latest cont rol t echni ques are finding acceptance in industry, probl em avoi dance, or design for cont rol , is usually the first st rat egy empl oyed t o defeat friction probl ems. Many studies have shown that the amplitude of stick slip can be r educed- - even to zero under some circumstances---by decreasing t he mass, increasing the dampi ng or increasing the stiffness of a mechanical system (Rabi nowi cz, 1959; Singh, 1960; Kat o e t al . , 1974). Dampi ng is usually controlled t hrough selection of the lubricant and t he sliding surfaces. The latter may be of different composition t han the underl yi ng bulk material ei t her applied as a coating or liner or embodi ed as a bearing. Inertia and stiffness are det ermi ned t o a great ext ent by the geomet ry and composi t i on of t he mechani sm' s bulk material. The selection of actuators, bearings and sensors can affect system dampi ng, stiffness and inertia. For example, rolling el ement bearings may possess dampi ng and stiffness characteristics which are considerably different t han sliding bearings. These topics are discussed briefly below. 4.2.1. L u b r i c a n t s e l e c t i o n In a cont rol cont ext , the goal of lubricant selection is usually t o reduce the negative slope of the fri ct i on-vel oci t y curve near zero velocity. Slope reduct i on is equi val ent t o increased damping. I f the slope remains negative, the system is still unstable, but is, however, easier t o stabilize by active control. Dr y lubricants, such as pol yt et rafl uoroet hyl ene ( PTFE or Teflon ) and mol ybdenum disulfide, can be used t o produce a stabilizing positive slope at very low velocities. For example, the fri ct i on-vel oci t y curve for PTFE has a positive slope up t o 10c ms 1. One limitation of these soft lubricating films, however, is their high wear rate. In addition, while these solids may stabilize and decrease friction at low velocities, t hey may generat e more friction t han would ot herwi se be encount ered at higher velocities. In certain applications, such as machi ne tools, this ext ra dampi ng is desirable for stability. Solids, such as Rulon , are often used in conj unct i on with liquid lubricants as slideway liner materials. 4.2.2. Be a r i n g s Bearing friction can be a probl em in high-precision positioning, pointing and tracking systems. Several schemes have been proposed t o over come mot i on errors from ball bearings. One involves active cont rol of the bearing out er race (Bi fano and Dow, 1985). In anot her met hod, t he out er bearing race is not rigidly mount ed t o t he machine frame. Inst ead, it is connect ed to the frame t hrough torsional springs as shown in Fig. 45 (Clingman, 1991). At mot i on initiation and direction reversal, the friction t orque and spring t orque act in series reducing t he effective slope of the fri ct i on-di spl acement curve. A controller of ball ~ / ~ ] ~ I ~ N~ Outer race of ball Outer bearing ring Q Flexing elements FIG. 45. Outer race of ball bearing is attached to static outer ring through torsional flexing elements [from Clingman (1991), courtesy of publisher]. can be added t o the bearing which drives the torsional springs t o null displacement (Clingman, 1991). In or der t o avoid t he nonlinearity of low-velocity friction, oil or air hydrostatic bearings can be used. While produci ng ext remel y low friction, air bearings exhibit low stiffness and dampi ng in the normal direction which can make t hem sensitive t o profile errors of guideways and t o external disturbance forces. To actively cont rol stiffness and dampi ng, lubricant supply pressure can be controlled or piezoelectric act uat ors can be placed in series with t he air pads (Hori kawa e t al . , 1991). Alternatively, active magnetic bearings represent a promising non-cont act t echnol ogy which is currently used only in high- velocity applications. In addition t o produci ng minimal friction, their ability t o provi de active dampi ng t o high speed rot ors eliminates critical speed vibrations ( O' Connor , 1992). 4. 2. 3. S t i f f n e s s a n d a c t u a t i o n The most common source of excessive machine compl i ance is transmission elements. The obvious solution, which has been considered by a number of researchers, is t o eliminate the transmission (Asada and Youcef-Toumi , 1984), or at least t o locate it at the joint (Townsend and Salisbury, 1987). Even in the absence of transmission elements, machine com- pliance due t o shafts and couplings can be significant (Brandenburg and Sch~ifer, 1991). Very high machine stiffnesses are sought when designing experimental apparat us for studying friction (Dieterich, 1979; Rui na, 1980; Pol ycarpu and Soom, 1992). Elimination of the transmission can necessitate the use of oversized or special hi gh-force/ t orque motors. For fine displacement applications, piezoelectric actuators with or wi t hout sliding cont act s can be used. A s u r v e y o f f r i c t i o n a n d c o n t r o l s 1115 To o v e r c o me t he i r s t r o k e l i mi t at i on, t he y ma y be us ed in c onj unc t i on wi t h c onve nt i ona l l ong- s t r oke act uat or s . 4 . 2 . 4 . I n e r t i a Whi l e i ner t i a r e duc t i on ha s a st abi l i zi ng ef f ect on s t i ck- s l i ppi ng s ys t e ms , it is not al ways pos s i bl e or pr act i cal t o r e duc e i ner t i a a nd, in s ome cases, r e duc e d i ner t i a can b e d e t r i me n t a l t o s ys t e m p e r f o r ma n c e . An i mp o r t a n t e x a mp l e is t he st abi l i zat i on of gi mba l l e d poi nt i ng s ys t ems . Cons i de r , f or e x a mp l e , a c a me r a mo u n t e d t h r o u g h a gi mbal t o a bas e whe r e t he bas e is s ubj e c t t o di s t ur ba nc e s . At hi gh di s t ur ba nc e f r e- que nc i e s , t he ba ndwi dt h of t he c ont r ol l e r ma y pr e c l ude act i ve c o mp e n s a t i o n . I n t he s e cases, t he me c h a n i s m mus t be des i gned f or pas s i ve i ner t i al s t abi l i zat i on. I n t he si ngl e d e g r e e of f r e e d o m case, t he gove r ni ng e q u a t i o n is r r = 1 0 , (26) whe r e ~r is t he f r i ct i on t o r q u e g e n e r a t e d by t he ba s e di s t ur bance, I is t he mo me n t o f i ner t i a o f t he c a me r a a b o u t t he gi mbal axi s a nd 0 is t he r es ul t i ng c a me r a accel er at i on. T o mi ni mi ze 0, i t is t he r a t i o o f f r i ct i on t or que t o i ner t i a, r e ~I , whi ch is i mp o r t a n t ( El l i son a nd Ri chi , 1983; St o c k u m e t a l . , 1988). I n t hi s c a s e , r educi ng i ner t i a wi t hout r educi ng f r i ct i on as wel l can wor s en hi gh- f r e que nc y p e r f o r ma n c e . I n addi t i on t o this di r ect c onne c t i on wi t h f r i ct i on, i ner t i al c om- pone nt s a r e s o me t i me s a d d e d t o poi nt i ng s ys t e ms t o passi vel y cancel e i t he r i ner t i al di s t ur ba nc e t e r ms ( El l i son a nd Ri chi , 1983) or i ner t i al r e a c t i ons dur i ng act i ve mo t i o n c ont r ol ( Ge r ma n n a nd Br acci o, 1990). Whi l e desi gn f or c ont r ol does not g u a r a n t e e t he pas s i ve e l i mi na t i on o f s t i c k- s l i p, it usual l y pr oduc e s a s ys t e m whi ch is eas i er t o c ont r ol a nd whi ch pos s es s es b e t t e r p e r f o r ma n c e char act er i st i cs. Fur t he r i mp r o v e - me n t can be a c hi e ve d by f r i ct i on c o mp e n s a t i o n as de s c r i be d in t he f ol l owi ng sect i ons. 4. 3. No n - mo d e l - b a s e d Co mp e n s a t i o n f or Fr i ct i on 4. 3. 1. S t i f f P D c o n t r o l Whi l e t he r e gul a t or p r o b l e m is s t abl e u n d e r PD c ont r ol , t he t r a c ki ng p r o b l e m doe s exhi bi t st i ck slip at l ow vel oci t i es. Fo r ma n y year s , it has be e n known t ha t by i ncr eas i ng t he d a mp i n g or t he st i ffness of a s ys t e m, s t i c k- s l i p can be e l i mi na t e d. I n a c ont r ol c ont e xt , t hi s is a c c ompl i s he d by i ncr eas i ng t he PD gai ns. PD cont r ol , a l ong wi t h i nt egr al a c t i on, is wi del y us ed in i ndust r y. Pe r h a p s be c a us e of its wi d e s p r e a d use, f ew p a p e r s a p p e a r r e por t i ng t he e x p e r i me n t a l p e r f o r ma n c e of PD and P I D cont r ol l er s in ma c hi ne s wi t h f r i ct i on. I ns t e a d, mos t r e l a t e d p a p e r s use e x p e r i me n t t o obt a i n a s ys t em mo d e l , i ncl udi ng f r i ct i on, whi ch is t he n a na l yz e d t o p r o d u c e st abi l i zi ng PD gai ns f or a gi ven i nput t r a j e c t or y. Si nce t he s e r esul t s d e p e n d on t he f r i ct i on mo d e l us ed in t he anal ysi s, f r e q u e n t r e f e r e nc e t o f r i ct i on mode l i ng in this sect i on is una voi da bl e . Th e success of stiff PD c ont r ol can onl y be ful l y unde r s t ood by cons i der i ng f r i ct i onal me mo r y . Whi l e t he t r i boi ogy c o mmu n i t y has be e n a wa r e of t hi s f r i ct i onal ef f ect f or a l mos t 50 ye a r s , it has onl y r ecent l y be e n i n c o r p o r a t e d i nt o f r i ct i on mo d e l s us ed f or cont r ol . Cons e que nt l y, mo s t o f t he l i t e r a t ur e on stiff PD c ont r ol , whi l e c or r e c t f or t he f r i ct i on mo d e l s c ons i de r e d, is o f l i mi t ed pr act i cal val ue. Whi l e t he s e si mpl i f i ed mo d e l s can mi mi c s t i c k- s l i p l i mi t cycl i ng, t he y pr e di c t s t e a dy sl i di ng onl y t hr ough der i vat i ve act i on, e ve n t h o u g h e x p e r i me n t s ha ve shown t ha t p r o p o r t i o n a l c ont r ol is al s o ef f ect i ve. Anal ys es ba s e d on f r i ct i on mo d e l s whi ch a r e si ngl e- val ued f unct i ons of vel oci t y f or n o n z e r o vel oci t y i ncl ude Ha r me r (1952), De r j a g u i n e t al . (1957), Co c k e r h a m a nd Sy mmo n s (1976), Br oc kl e y e t al . (1967), Ba n n e r j e e (1968), Ar ms t r o n g - Hr l o u v r y (1990) a nd Ga o a nd Ku h i ma n n - Wi l s d o r f (1990). Th e s e anal ys es ma y ha ve di r ect i mpl i cat i ons f or s ys t ems in whi ch f r i ct i onal me mo r y is of such a ma gni t ude t ha t its negl ect is j ust i fi ed. Begi nni ng wi t h Sa mp s o n e t al . (1943), r e s e a r c he r s obs e r vi ng st i ck-sl i p l i mi t cycl es n o t e d t ha t f r i ct i on was hi gher dur i ng t he a c c e l e r a t i on pha s e o f t he sl i p cycl e t han dur i ng t he de c e l e r a t i on phas e. ( Recal l Fi g. 21). Thi s l ed t o t wo- va l ue d mo d e l s whi ch we r e e mp l o y e d in t he st abi l i t y anal ys es of Bel l a nd Bu r d e k i n (1969), Co c k e r h a m and Sy mmo n s (1976), Co c k e r h a m a nd Col e (1976) and Bo a nd Pa ve l e s c u (1982). Thi s wo r k r e pr e s e nt s a si gni fi cant s t e p t o wa r d t he r e c ogni t i on o f f r i ct i onal me mo r y . Al t e r na t i ve l y, Ka t o and Ma t - s uba ya s hi (1970) o b s e r v e t he mul t i - val ued be ha vi or , but p r o p o s e t he us e of a me a n f r i ct i on coef f i ci ent dur i ng t he slip phas e. Re c e nt l y, it has b e c o me pos s i bl e t o expl ai n t he st abi l i t y of hi gh- st i f f ness s ys t e ms whi ch exhi bi t ne ga t i ve s t e a dy- s t a t e d a mp i n g t hr ough t he i ncl usi on of f r i ct i onal me mo r y ef f ect s. For e x a mp l e , Ar ms t r o n g - Hr l o u v r y (1992) cons i der s a f r i ct i on mode l c o mp o s e d of St r i beck, vi scous a nd r i si ng- st at i c- f r i ct i on c om- pone nt s . Th e St r i beck c o mp o n e n t is s ubj ect t o t he p u r e t i me l ag p r o p o s e d by He s s and Soom (1990). Ar ms t r o n g - Hr l o u v r y e mp l o y s a p e r t u r b a t i o n me t h o d in c o mb i n a t i o n wi t h nume r i c a l i nt e gr a t i on t o de t e r - mi ne t he st abi l i t y of a st i ck-sl i p l i mi t cycl e. Hi s anal ysi s first cons i der s t he f r i ct i onal me mo r y a nd ri si ng st at i c f r i ct i on mo d e l s s e pa r a t e l y t o obt a i n st abi l i t y cr i t er i a. Th e s e a r e t he n c o mb i n e d t o d e t e r mi n e t he ef f ect of each. Cons i der i ng onl y t i me l ag wi t h a f r i ct i on mo d e l exhi bi t i ng negat i ve d a mp i n g at l ow vel oci t i es, a si ngl e de gr e e o f f r e e d o m s ys t em consi st i ng o f a sl i di ng ma s s , M, will not e xpe r i e nc e st i ck slip f or mo d e r a t e f r i ct i on i f t he s ys t em st i ffness ( i ncl udi ng p r o p o r t i o n a l gai n) me e t s or e xc e e ds a cri t i cal va l ue , k , r wh e r e kc, = M ~-" (27) r ~" Cl ear l y, as t he t i me l ag, rL, a p p r o a c h e s z e r o , t he critical st i ffness a p p r o a c h e s i nfi ni t y. T h e addi t i on o f t he r i si ng- st at i c- f r i ct i on mo d e l is of gr e a t e s t i mp o r t f o r ver y l i ght l y d a mp e d s ys t e ms , i ncr eas i ng t he cr i t i cal st i ffness at low vel oci t i es a nd de c r e a s i ng i t at hi gher vel oci t i es. Th e c o mb i n e d mo d e l pr ovi de s a good ma t c h wi t h e x p e r i me n t a l d a t a f r o m t he ba s e j oi nt o f a P UMA r obot . 1116 B. ARMSTRONG-HI~LOUVRY et al. In contrast to studying the stability of stick-slip limit cycling, Rice and Rui na (1983) empl oy the concept of a nonl i near friction model which depends on slip history and consider the stability of the equilibrium point associated with steady sliding. This class of friction models includes the state variable models described in Section 2.1.5. They consi der small pert urbat i ons by linearizing t he system about st eady sliding. The steady-state friction-velocity slope is assumed t o be negative while t he i nst ant aneous slope is positive ei t her due t o the friction itself or external damping. Using a root-locus argument , t hey obt ai n a critical stiffness above which the equilibrium point is stable under small pert urbat i ons. For a friction model involving a single exponential decay over the characteristic length, L, the critical stiffness, kcr, can be expressed as kc, = Vd T, s ( V) J dV 1 + (28) L L ' where V is velocity, r is frictional shear stress, r~.~ is steady-state frictional shear stress and m is slider ma s s . Dupont and Bapna (1992) have devel oped expressions for critical stiffness in systems where frictional memor y is also associated with changes in normal stress. In summary, stick slip can be eliminated t hrough ei t her high derivative (velocity) feedback or high proport i onal (position) feedback. They are best used t oget her as t hey are compl ement ar y. While derivative feedback is additive with i nherent system dampi ng, this is not t he case with proport i onal (position) feedback. System stiffness acts in series with controller stiffness. Thus, high gain proport i onal cont rol is most successful in systems which can be designed for high rigidity. This topic is discussed furt her in Section 4.5. 4.3.2. Integral control While stiff PD cont rol can be used t o achieve stable tracking, integral cont rol of position or velocity is al most always i nt roduced t o minimize steady-state errors. Usi ng integral action, systems are found to limit cycle when tracking at low or zero velocities. Int egral action, and the limit cycling it induces, are rarely discussed in the cont rol s literature except as mot i vat i on for mor e compl i cat ed cont rol met hods. This is in direct cont rast with its widespread use and with the vari et y of techniques devel oped t o circum- vent its short comi ngs. To over come limit cycling, one st andard t echni que is t o empl oy a deadband as the input t o the i nt egrat or block. This, of course, imposes its own steady-state er r or - - l ess, it is hoped, t han t hat before the integral action was added. Shen and Wang (1964) empl oy a st at i c-Coul omb model t o study posi t i on-ramp inputs. They find t hat the size of t he stabilizing deadband decreases al most linearly with ramp rate. To avoid sluggish response and large steady-state errors at high ramp rates, t hey propose setting the deadband limit as a funct i on of input r amp rate. In addition t o inducing limit cycling, integral cont rol can be ineffective and even deleterious at velocity reversals (Tung et al., 1993). Integral windup from prior motion can actually inhibit breakaway. To prevent this, the integral t erm is typically reset at velocity reversals. While this eliminates the windup probl em, the ensuing integral action produces minimal effect when needed t he most t o overcome stiction. In a system with multiple degrees of freedom, ot her joints may be moving at high velocity during the reversal. Consequent l y, the reduced effect of the integral action in concert with t he higher level of static friction at reversal can lead t o significant t racki ng errors. Suzuki and Tomi zuka (1991) consider this issue in the context of high speed machining of circular contours. As an alternative to integral cont rol , t hey propose a model -based cont rol l er which applies a pulse to over come stiction at breakaway. Hansson et al. (1993) apply a fuzzy rule system t o controlling windup in PI D controllers. Br andenbur g and Sch~ifer (1988a, 1989, 1991) consider integral control in the context of a variety of tasks and compensat i on techniques. Industrial use of integral cont rol is discussed furt her in Section 4.5. 4.3.3. Dither Di t her is a high frequency signal i nt roduced into a system t o modi fy its behavior. Di t her can stabilize unstable systems (Bogol i ubov and Mitropolsky, 1961), and is used t o i mprove performance by modifying nonlinearities in adapt i ve cont rol ( Ander son et al., 1986), communi cat i on systems (Chou, 1990), optics (Hirel, 1990) and image processing (Chau, 1990). For machi nes with friction, the controls communi t y has focused on the capability of dither t o smoot h the discontinuity of friction at low velocity. Earl y discussions of di t her empl oy the describing function, (MacColl, 1945; At her t on, 1975). Lat er t reat ment s bring t o bear averaging t heory (Mossaheb, 1983), functional analysis and ot her met hods of nonlinear system analysis (Zames and Shneydor, 1976, 1977; Cebuhar , 1988; Bent sman, 1990; Lee and Meerkov, 1991). How smoot hi ng arises with dither can be seen in an example [following Cebuhar (1988)]: t he relationship y( t ) = sgn ( u( t ) ) (29) is discontinuous. However , when a dither of amplitude 0~ and frequency oJ is added t o the input, the averaged out put becomes: ~(t ) = sgn ( u( r ) + ~ sin (oJ~)) d~'. (30) Through averaging, )7(0 can be a cont i nuous function of u(t). 4.3.3.1. Tangential and normal dither. The analyses present ed in the cont rol literature focus on a dither signal added t o the command input, which, for t he configuration shown in Fig. 46, will give rise t o vibrations t hat are tangential t o t he sliding contact. In the t ri bol ogy literature, on the ot her hand, the impact of vibrations normal t o the cont act have been considered (Fri edman and Levesque, 1959; Godf r ey, 1967; Oden and Martins, 1985; Hess and Soom, 1991a, b). The distinction bet ween normal and A survey of friction and controls 1117 Tangential Vibration Modifies Influence of Friction (e.g. Control Input Dither) ~ Normal Vibration t Modifies Friction u(t) ~ ~ (e.g. External Vibrator) Friction S / / / / Interface Fro. 46. Direction and effect of dither. tangential di t her in a friction cont act is a consi derabl e one: t he effect of tangential di t her is t o modi fy t he influence of friction (by averagi ng t he nonl i neari t y); t he effect of vibrations normal t o t he cont act is t o modi fy t he friction [by reduci ng t he friction coefficient, ( Godf r ey, 1967; Martins et al., 1990)]. This survey is apparent l y t he first t i me t hat these t wo possibilities have been consi dered t oget her. Wor ki ng with a simple Coul omb + viscous friction model one woul d not expect friction t o be reduced by normal vibrations, so long as cont act is not br oken; but when cont act compl i ance (t he origin of presliding displacement) and asperity cont act s are consi dered, more sliding is seen t o occur during peri ods of reduced loading and less during peri ods of i ncreased loading. This arises because t he mechani cal bandwi dt h of individual asperities may be orders of magni t ude higher t han t he bandwi dt h of t he macroscopi c mechanical elements. Godf r ey (1967) report s a reduct i on of t he coefficient of friction f r om 0.15 t o 0.06 in a lubricated steel cont act with t he addition of 1000 Hz normal vibrations. For t he cont rol s engi neer, t he i mport ance of this distinction lies in how di t her is t o be applied. The original applications of di t her involved external mechanical vi brat ors (Bennet , 1979). And on gun mount s and ot her large pointing systems, such vibrators, somet i mes called "di ppl er s, " are still used. While any f or m of di t her will result in bot h tangential and normal forces t hrough t he coupl i ng arising out of asperity cont act s (Martins et al., 1990), mor e f r eedom exists for ori ent i ng t he di t her when an external vi brat or is used. I t is likely t hat mor e t han one servo functions because of a badly bal anced fan in t he vicinity is provi di ng dither. 4.3.3.2. Dept h o f discontinuity and dither in hydraulic servos. When di t her is applied, filtering exists bet ween t he source of t he vibration and t he poi nt where it is t o have its effect, as shown by the t ransfer funct i on Gl ( s) in Fig. 47. I n all cases, but particularly when di t her is applied at t he cont rol input, this filtering is i mport ant . Cebuhar (1988) and eval uat es t he filtering in terms of the ' dept h' of t he discontinuity, and defines a formal measure of this dept h relating t o t he number of integrators in G~(s). He finds t hat when t he dept h is great , t he designer is F1G. 47. A transfer function between the input and the nonlinearity where dither is to have its influence. mor e restricted in t he application of a dither. Bot h Cebuhar (1988) and Lee and Meer kov (1991) investigate the opt i mal di t her paramet ers with a const ruct i on like t hat of Fig. 47. The ratio Gl ( ito ) (31) G2(io9) p l a y s a n i m p o r t a n t r o l e in t h e e f f e c t i v e n e s s o f t h e d i t h e r a n d C e b u h a r p r o p o s e s t h a t co,i, t h e d i t h e r f r e q u e n c y , s h o u l d b e c h o s e n t o m a x i m i z e a slightly m o d i f i e d f o r m o f e q u a t i o n ( 3 1 ) . H o r o w i t z e t al. ( 1 9 9 1 ) consider limitations on range of applicable rod and propose an adaptive const ruct i on which allows l ower frequency dither. Based on input f r om engi neers in industry (see Section 4.5), it seems t hat di t her is onl y occasionally applied t o mot or servos, but is oft en applied and with great effect t o spool valves in hydraulic servos. The larger Gl (i t o)/ G2(i t o) achievable in the hydraulic servo may account for t he great er success of dither in these systems. 4.3.4. I mpul si ve control A number of investigators have devised controllers which achieve precise mot i ons in t he presence of friction by applying a series of small impacts (Yang and Tomi zuka, 1988; Suzuki and Tomi zuka, 1991; Ar mst r ong, 1988; Ar mst r ong- Hr l ouvr y, 1991; De- weert h et al., 1991; Hoj j at and Higuchi, 1991). These are ' impulsive cont rol l ers' . Impulsive cont rol is distinguished from dither in t hat it is the impulses themselves which are t o carry out t he desired mot i on. The impulses used are not zero mean and must be calibrated t o produce the desired result. Impulsive cont rol is also distinct from st andard pulse width modul at i on ( PWM) controllers, where voltage pulses are applied t o a mot or. In PWM controllers, the mot or i nduct ance averages the relatively high frequency (perhaps 20 kHz) voltage pulses t o pr oduce a nearly const ant mot or current , and t herefore nearl y constant t orque. I n Yang and Tomi zuka (1988), Suzuki and Tomi zuka (1991) Ar mst r ong (1988), Armst rong- Hr l ouvr y (1991), Hoj j at and Higuchi (1991), the impulses are applied when the system is at rest, i.e. in t he stuck condition. The effect of the impulse is a small di spl acement ( Ar mst r ong- Hr i ouvr y, 1991; Hoj - jat and Higuchi, 1991) or a cont rol l ed breakaway, leading t o transition t o anot her cont rol l er which regulates macroscopi c movement s (Yang and Tomi - zuka, 1988; Suzuki and Tomi zuka, 1991). By maki ng the impulses of great magni t ude but short durat i on; the static friction is over come and sensitivity t o t he details of friction is reduced. A typical behavi or is shown in Fig. 48. Hoj j at and Higuchi (1991) present an apparat us designed especially t o demonst r at e impulsive control. They reliably achieved a remark- able 10 nm per impulse mot i on and speculate t hat repeat abl e 1 nm per impulse mot i ons may be possible. Thei r mechani sm was not unlike a machi ne slideway; the slider measured about 3 cm on a side and weighed 155g. Hoj j at and Higuchi (1991) cont rol t he amplitude of their impulses, typically applying a force about 10 times the static friction for about 1 ms, and 1118 B. ARMSTRONG-HI~LOUVRY et al. z o 4 0 0 I mp u l s i v e force Di s p l a c e me n t 0 10 Ti me (ms) 400 p2 v 2 ~M2g \ j x M g o 2o FIG. 48. Behavior of motion under impulsive control [from Hojjat and Higuchi (1991), courtesy of the publisher]. show t hat di spl acement is given by the square of the ampl i t ude times an empirical constant. Yang and Tomi zuka (1988) present a variable structure controller. Away from zero velocity, a standard linear controller operat es. At or near zero velocity, the adaptive pulse width control takes over. The impulse is t uned by varying the pulse width. A current pulse is applied, as opposed to the voltage pulses used in st andard PWM control. The impulses are of a predet ermi ned force which must be great er than the static friction; Yang and Tomi zuka (1988) chose a force about four times the level of Coul omb friction. The durat i on is t hen selected to achieve the desired displacement. The aut hors present a rigorous demonst rat i on of t he stability of bot h their controller and the adapt i ve el ement . Suzuki and Tomi zuka (1991) focus on the challenges of high speed machi ni ng and demonst rat e a cont rol l er which includes an impulsive el ement which operat es near zero velocity. Deweert h et al. (1991) describe a neural net work based controller, with functionality comparabl e to a PD controller, but which generat es a pulse train. The impulses are a natural consequence of the neural structure. The aut hors demonst r at e t hat t he cont rol l er is effective in controlling low-speed, friction-limited motions. Ar mst r ong (1988) and Armst rong-H61ouvry (1991) demonst r at e impulsive cont rol of a PUMA robot . Thei r objective was very high precision force cont rol , needed t o manipulate an obj ect with a crush strength of only 1/60th the level of static friction in the mechanism. Using a calibrated table of impulse magni t ude and durat i on, Armst rong-H61ouvry (1991) applies impulses t hat are onl y 10- 20% great er in magni t ude t han the static friction and achieves 10 mi cro-radi an per impulse mot i ons of the industrial mani pul at or. The impulsive controllers of Hoj j at and Higuchi (1991), Yang and Tomi zuka (1988), Ar mst r ong (1988) and Armst rong-H61ouvry (1991) have in common the use of a long sampling interval. Underl yi ng these impulsive controllers is the requi rement t hat the system be in the stuck condition when each impulse is applied. The met hod is in essence a small bang followed by an open-l oop slide. Ret urni ng t o the stuck condition imposes a number of limitations, but improves the predictability of the response t o t he impulse. The variable structure controller dem- onstrated by Yang and Tomi zuka (1988), which employs the impulsive control onl y at zero velocity, perhaps best exploits the capability of impulsive control. The interaction of force application rate and breakaway friction has not been considered by the authors presenting impulsive controllers. But if force application rate is at the heart of the observed rising static friction (Ri chardson and Nolle, 1976; Johannes et al. , 1973; see Section 2.1.4.2), the implications may be substantial for impulsive control. 4.3.4.1. Dither, i mpul si ve cont rol and cont act compl i ance. None of the aut hors describing either di t her or impulsive cont rol address contact com- pliance, discussed in Section 2.1.2. In the above cited works, it is presupposed t hat high frequency forces applied to the sliding elements act fully on the friction contact. But the elastic response of the junction serves as a low pass filter, increasing the depth of the frictional discontinuity by adding one or perhaps two poles t o G, ( s ) of equat i on (31). As the works of Cebuhar (1988) and Lee and Meerkov (1991) make clear, these met hods will be influenced by G, ( s) at the applied frequencies. The i mport ance of contact compliance is illuminated by the fact that the micro-radian movement s of Ar mst r ong (1988) and Armst rong-H61ouvry (1991) and the nanomet er movement s of Hoj j at and Higuchi (1991) are both smaller t ha n- - a nd in the later case three orders of magni t ude smaller t han- - t he typical presliding dis- placement. This suggests that actual motion is occurri ng onl y in some percent age of asperity contacts, with movement of the body a consequence of the shifted equilibrium of the stresses of all of the contacts. Cont act compliance depends upon contact loading, surface finish and the material properties of the parts, indicating t hat the success of either impulsive cont rol or dither will depend upon these aspects of machine design. 4.3.5. Joi nt t orque cont rol Joi nt t orque cont rol is a sensor-based technique which encloses the act uat or-t ransmi ssi on subsystem in feeback l oop to make it behave more nearly as an ideal t orque source (Wu and Paul, 1980). Distur- bances due t o undesirable act uat or characteristics (friction, ripple, etc) or transmission behaviors (friction, flexibility, inhomogeneities, etc) can be significantly reduced by sensing and high gain feedback. The basic structure is shown in Fig. 49; an inner t orque l oop functions t o make the applied t orque, T, follow t he command t orque, T+.. Joint t orque control has been i mpl ement ed as a means of compensat i ng for act uat or and transmission friction (Wu and Paul, 1980; Luh et al., 1983; Pfeffer et al., 1986, 1989; Vischer and Khatib, 1990a, b; Karlen et al., 1990; Hashi mot o et al., 1992), as a means of compensat i ng or mor e precisely controlling transmission flexibilities (Karlen et al. , 1990; Furusho et al., 1990; Hashi mot o et al., 1992), and as a means of sensing and compensat i ng for the nonl i near rigid body dynamics and gravitational loads experi enced in A s ur vey of fri ct i on and cont r ol s 1119 Xd T Actuator Cmrent T Torque Applied to u Command a the Mechanism X e T c Te / / C~ ( ~ ~ t ~ ' / Mot or, ~___4:~,.Q.J Mechanism ~ _ _ L ( I - - r ' ~ ' I - - ~ ~ " t ' ~ ' ] ]Transmission I t~t,-I I I [ - - - - I \ T o r q u e | I Torq Saner ! (High Bandwidth Inner Loop) j / (Lower Bandwidth Out*r Loop) FIG. 49. Block diagram of a joint torque control (JTC) system. robotics (Kosuge e t a l . , 1988; Hashimoto, 1989). The above implementations all incorporate one axis of sensing per joint. Six axes of force/ t orque sensing per joint has been proposed in robotics (Mukerjee and Ballard, 1985) for inertial paramet er identification. Implementation of joint torque control requires torque or force sensing as near as practical to the output element of the system so that all or nearly all of the actuator and transmission friction will be enclosed in the joint torque feedback loop. A number of authors have employed strain gages, typically mounted on supports for transmission elements (Wu and Paul, 1980; Luh e t a l . , 1983; Pfeffer e t a l . , 1989; Hashimoto, 1989; Hashi mot o e t a l . , 1992). Hashi mot o (1989) and Hashi mot o e t al . (1992) focus specifically on harmonic drives. Vischer and Khatib (1990a) present a sensing package that may offer advantages of sensitivity, robustness and stiffness. It is based on differential inductive sensing. In joint torque control, the sensor and actuator are often non-collocated, separated by the compliance of the transmission and perhaps that of the sensor itself. This gives rise to the standard challenges of non-collocated sensing. All of the authors cited above who demonstrate joint torque control discuss the issues of system identification and Laplace or Z domain compensator design. As shown in Fig. 49, the controller design is a nested one, with an inner loop which maintains applied torque and compensates for friction, and an outer loop which governs the execution of the mechanism task. The multi-loop structure would in general require a full multi-loop analysis; the authors cited, however, have focused their attention on the inner torque loop, and have required that the frequency domain separation between the inner and outer loops be sufficient to protect against dynamic interactions. Eismann, who has substantial experience with the commercial Robotics Research arm, suggests that a 4: 1 ratio is required between the cross over frequency of the inner joint torque controlled loop and that of the outer task control loop (Eismann, 1992). 4.3.5.1. P e r f o r m a n c e . Luh e t al . (1983) report that the Coulomb friction in their Stanford robot arm is 1072 oz-in, and that the apparent friction with joint torque feedback is 33.5 oz-in, a reduction of 32:1. Pfeffer e t al . (1989) report that the apparent friction in joint 3 of a PUMA robot arm was reduced to 3% of its uncompensated level; they demonstrate free swinging motions of the robot link (To = 0 in Fig. 49, but T, quite active to compensate for actuator and transmission friction during the motions). Hashi mot o e t al . (1992) evaluate joint torque control coupled with feedforward Coulomb-friction compensation, in both fixed and adaptive forms. Step torque input and sinusoid position input responses of their system are presented in Figs 50 and 51. The four presented responses are: open-loop, joint torque control with no friction feedforward, joint torque control with fixed friction feedforward, and joint torque control with adaptive friction feedforward. Numeric values for the data are not presented, but it is evident that the influence of friction is very greatly reduced. Hashimoto e t al . (1992) also present vibration suppression data which show a factor of 10 reduction of vibration at the mechanical resonant frequency during a medium velocity motion. The resonant vibration couples the inertia with the compliance of the transmission, and is reduced because the transmission compliance appears within the joint torque control loop. Robotics Research Corporation implemented joint torque control to compensate for the compliance of their harmonic drives. Using only sensing at the actuator, the drive compliance would result in lightly-damped, low frequency modes. Figure 52 is adapted from Karlen e t al . (1990), and shows the magnitude of response to a sinusoidal disturbance at the toolplate (distal end) of the robot arm. Without joint torque control, there are two lightly damped modes between 10 and 20Hz; with joint torque control the response is nearly flat through this frequency range. In all of the trials reported with joint torque control, the quality of the system behavior as measured at the out put --i n terms of apparent friction, oscillatory behavior or mechanism nonlinearities---is very greatly improved. The data reported are not directly comparable to the results of other compensa- tion techniques, such as the 5:1 i mprovement in telescope pointing achieved by adaptive control (Walrath, 1984), but show that when friction is sensed, it can be compensated by feedback control. 4.3.6. D u a l m o d e c o n t r o l High precision applications, such as semiconductor manufacturing and diamond turning of optical 1120 B. ARMSTRONG-HF_.LOUVRY e t al . i---. t , a o i 10 5 0 Open 100p command response [ l a : a : t , a 7a t.--. g 10 Type A 0.00 0 . 2 5 0 . 5 0 0 . 0 0 0 . 2 5 0 . 5 0 TIME [see] TIME [sec] 10 5 Type B V y l q 7 . . . . "rl . . . . r-,'- -" . . . . . ~,r" ",u',',~ ". 10 z ':, 5 [ . . . T y p e C O r i 0 0.00 0.25 0.50 0.00 0.25 0.50 TIME [sec] TIME [sec] FIG. 50. Step response in torque control; open-loop, (a) joint torque control with no friction fcedforward, (b) joint torque control with fixed friction feedforward, (c) joint torque control with adaptive friction feedforward [from Hashimoto et al. (1992), courtesy of the author]. elements, require nanomet er positioning accuracy over millimeters of motion range. The standard technology for nanomet er positioning involves two stage mechanisms. The coarse positioning stage might be a ball screw; and the fine motion stage might comprise a piezoelectric actuator [see Futami et al. (1990) and references for a brief discussion of two stage mechanisms]. The liabilities of the two stage mechanisms are weight, size and complexity: two actuators and two controllers are required per degree of freedom. By capitalizing on presliding displacement, referred to as microdynamics in the nanotechnology literature, it is possible to achieve two modes of control in a - 1 ~ 0 0.0 wi t hout com.,and torque I o o ~ A I i 1.2 TIME {sec] 2A - 1 0 with Type A - 1 2 . . ~ 0 , . c z with Type C J I I i I J 0.0 1.2 2.4 0.0 1.2 2. 4 TIR [secJ TIME {sec} FIG. 51. Sinusoid tracking in torque control; (a) open-loop, (b) joint torque control with no friction feedforward, (c) joint torque control with adaptive friction feedforward [from Hashimoto et aL (1992), courtesy of the author]. A sur vey of fri ct i on and cont r ol s 1121 2:h FFT RNRLYZI~ I, a I B! O~0.0000~ R'-ZOdB, ,, B: - l Ocl B. FREC': 100Xz 7 . ~ O s I I t I ' I ' i I _ 1 ! t i i : 1 1 I t ' F t t I t r . R E s p t t I t t C O N T R O L S E N A e L ~ : e a ~ E S o ~ inch t ~ l w ' ' 1 l t M A C H I N E D Y N A M I C S ' '1 "" I I MPROVED BY TORQUE L OOP j ~ ! " ]i_ I I , . . L L o . o o o . k - ; ' . 9 o ~ t t I t 1 I " i t . . . t . , f r ' z ~ s P t t t t _ t _ _ _ ~ a , I L l t { l 1 I { l l L . . 1 _ , : h i t t tl;t o . o o o ~ "~-~-P" 1 I t t I - t 1 . 1 f - t - 0 10 20 3 0 4 0 50 6 0 7 0 80 90 100 L I N E A R F R E Q U E N C Y - Hz FIG. 52. Dynami c response of a robot wi t h (above) and wi thout (bel ow) j oi nt torque control [from Karlen et al. (1990), courtesy of the publisher]. single mechanism: gross motion in the standard way, and fine motion in the presliding displacement. Referring to Fig. 7 above, presliding displacement is motion that occurs by the deformation of asperities in the sliding interface. In this motion, position is a function of applied force; the junction appears to be a stiff spring rather than a sliding (or rolling) bearing. The result is two markedly different mechanism dynamics: ' macrodynamics, ' the ordinary dynamics of the mechanism, and ' microdynamics, ' which governs motions that depend upon elastic deformation in the frictional contact. Because the dynamms are drasti- cally different, two different controller structures are required, thus dual mode control. Futami e t al . (1990) demonstrate accurate tracking of 1 nm step inputs with a slider on an air table and a linear AC mot or and ball screw. The stroke of the mechanism is 250mm, giving a dynamic range of 2. 5x108: 1. The authors identify three friction regimes, as seen in Fig. 53. Regime I exhibits the behavior of a linear spring, regime II the behavior of a nonlinear spring with damping, and regime I I I the normal sliding friction. The authors present control- lers for regimes I and II. The coarse motion (regime III) controller is capable of positioning within 100 nm, which is the range of operation of the fine controller (regime I). Aft er a coarse motion, 50 ms are allowed for vibrations to settle out, and the fine motion controller takes over. The authors have observed that the force-di spl acement characteristic of regime I is consistent throughout the travel of the mechanism. Ro and Hubbel (1993) demonstrate dual mode control for a similar ball screw mechanism. They argue that the frictional characteristics are not constant throughout the workspace, and present a model reference adaptive control scheme. Demon- strating step input tracking, they find that the microdynamic controller is satisfactory for motions smaller than 400 nm, and the macrodynamic controller satisfactory for motions larger than 1000 nm. Inter- mediate size steps they find difficult to accurately control. Armst rong (1988) and Armstrong-H61ouvry (1991) has observed presliding displacement in an industrial manipulator and proposed dual mode control for that mechanism, suggesting that the method may be suitable for standard as well as high precision equipment. It appears that two and perhaps three orders of magnitude of i mprovement in positioning accuracy is possible with sufficiently accurate sensing and dual mode control. FORCE IN) 0.1 I - I 0 0 - 0 . 1 J . . . . . . . . . . . . . . . . . t m D 0.1 -0.I -2 [I m DISPLACEMENT (Urn) IO0 FIo. 53. Outline of the force-to-displacement relationship of a ball screw driven slider [from Futami et aL (1990) courtesy of the author]. 1122 B. ARMSTRONG-HI2LOUVRY e t a l . 4.4. Model -based Compensat i on for Friction 4.4.1. F i x e d c o m p e n s a t i o n When a model of friction is available, it is possible to compensat e for friction by applying a f or ce/ t or que command equal and opposi t e to the instantaneous friction force. This presumes t hat force or t orque act uat i on of adequat e bandwi dt h is available and is stiffly coupl ed to the friction element. In many cases the domi nant friction sources are the mot or and transmission, and so adequat e stiffness is assumed. Her e compensat i on is addressed; questions of tuning the friction model paramet ers either off-line or adaptively, will be taken up below. Model-based compensat i on has been report ed in a number of studies (Gilbart and Wi nst on, 1974; Walrath, 1984; Henrichfreise, 1985, 1992; Acker mann and Miiller, 1986; Kubo e t a l . , 1986; Craig, 1987; Canudas de Wit e t a l . , 1987, 1989, 1991; Tomi zuka e t a l . , 1988; Rat t an e t a l . , 1989; Br andenbur g e t a l . , 1988b, 1989, 1991; Ar mst r ong, 1988; Armst rong- H61ouvry, 1991; Johnson and Lorenz, 1991; Suzuki and Tomi zuka, 1991; Dupont , 1993; Maquei ra and Masten, 1993); and is found in industrial cont rol applications (see Section 4.5). All of these studies report experimental results, and many of t hem report significant i mprovement in performance when feed- forward compensat i on is applied. The basic construc- tion of model -based friction compensat i on is shown in Fig. 54. The model -based schemes can be classified according to what estimate of velocity is used to evaluate the friction model and what port i ons of the friction model are applied. In all of the studies above, as well as the report ed industrial cont rol applications, compensat i on for Coul omb friction is included; many of the studies include additional friction terms as well (Wal rat h, 1984; Ar mst r ong, 1988; Armst rong- H61ouvry, 1991; Br andenbur g and Sch~fer 1991; Canudas de Wit e t a l . , 1989, 1991; Johnson and Lorenz, 1991). Typical of the approaches where onl y Coul omb friction is compensat ed, Gilbart and Winston (1974) apply their scheme to optical tracking devices and report a fact or of six reduct i on in RMS tracking error. The linear compensat or is not affected by the feedforward compensat i on. Canudas de Wit e t a l . (1987) show a significant reduct i on in stand still during zero crossings of velocity with Coul omb friction compensat i on; as does Walrath (1984). Because the Coul omb friction model is discon- tinuous, the choice of estimate of velocity with which to evaluate the model is significant. Gilbart and Winston (1974), Wal rat h (1984), Kubo e t a l . (1986) and Canudas de Wit e t a l . (1987) empl oy sensed velocity. Though these aut hors do not report undue difficulty with stability, there are no doubt sensor noise and stability issues which must be considered when an infinite gain oper at or is introduced in the feedback loop. Canudas de Wit e t a l . (1989) point out that model-based Coul omb friction may reduce the need for high servo gains, and thus reduce the impact of sensor noise. None-the-less, to reduce the impact of sensor uncertainties on the friction compensat i on a state est i mat or may be used. Br andenbur g and Schiller (1988b, 1989) and Sch~ifer and Br andenbur g (1990), who term the feedforward Coul omb friction compensat i on a "di st urbance observer", have shown remarkable performance with a system empl oyi ng a state est i mat or for velocity. Here the principal concern is with the presence of limit cycles rat her than tracking accuracy; the aut hors study a two mass system with backlash and report the presence of high and low frequency limit cycles. By use of the Coul omb friction compensat i on, the high frequency limit cycle is eliminated compl et el y and the range of velocities over which a low frequency limit cycle is observed is halved (Brandenburg and Schiller, 1988b). Johnson and Lorenz (1991) and Armst rong- H61ouvry (1991) have consi dered the use of commanded velocity t o generat e the friction compen- sation. Johnson and Lor enz (1991) present ex- perimental results for bot h feedforward and feedback systems; the two perform comparabl y, except near zero velocity. They report a fact or of four reduct i on in position errors during step input trials for bot h systems. Armst rong-H61ouvry (1991) reports two different types of results: open-l oop motions i nt ended t o demonst r at e the potential for accurate friction modeling, and high precision force control with an Xd . ( System Under Control ? : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 % i+ + + ' ~ + ~ i - - ~ '+'++'+~ F - - - 1 ~';+'+ x ) ~ GrCs ) i~+} - +_ Go(S ) G,(s) (Actual Friction) Friction Compensation - - I Friction Predictor ] _ Velocity, J ~ Sensed, Estimated or Desired FIG. 54. Basic construction of model-based friction compensation. A s ur ve y o f f r i ct i on a n d c ont r ol s 1123 2. 0 m 1.5 + 1 . 0 0 . 5 0 l l I I I 0 . 1 0 . 5 0 . 9 1 . 3 1 . 7 T i m e ( s e c ) F l o . 5 5 . P o s i t i o n a s a f u n c t i o n o f t i m e f r o m e x p e r i m c n t a l trials: ( a ) r c f c r c n c c m o d c l ; ( b l ) w i t h o u t n o n l i n c a r c o m p e n s a t i o n ; ( b 2 ) w i t h C o u l o m b f r i c t i o n c o m p e n s a t i o n ; ( b 3 ) C o u l o m b p l u s K a r n o p p - l i k c s t a t i c f r i c t i o n c o m p e n s a t i o n w i t h a n a d d i t i o n a l d y n a m i c t e r m ] f r o m B r a n d c n b u r g a n d S c h ~ i f c r ( 1 9 9 1 ) , c o u r t e s y o f t h e a u t h o r s ] . impulsive controller. In these cases it was necessary t o use command velocity. Tracki ng errors of 10% or less are demonst rat ed in t he open-l oop mot i ons, cor- responding t o a very high fraction of t he friction compensat ed by the model term. While good per f or mance has been shown with straight Coul omb friction compensat i on, i mprove- ments have been shown with t he use of richer friction models. Br andenbur g and Sch/ifer (1991) and Johnson and Lor enz (1991) present systems with static friction modeling and compensat i on. Bot h empl oy a Kar nopp- like model (see Section 3.4, Fig. 39), with a higher level of friction compensat i on provi ded near zero velocity. In Fig. 55, experimental results f r om Brandenburg and Sch/ifer (1991) are present ed. Thei r system is a complicated one, involving t wo masses, compliance, backlash and a multi-loop cont rol l er t hat simulates an industrial servo drive controller. Model reference adaptive compensat i on is used (a block di agram of this system is provi ded with t he discussion of adapt i ve cont rol , see Fig. 59). The figure shows t hat standstill at zero-crossings of velocity are substantially reduced duri ng a sawt oot h mot i on. Curve (a) is the track of t he reference model , curve (bl ) shows the case where no friction compensat i on is used. The curve (b2) indicates Coul omb friction modeling and feedforward compensat i on; and t he curve (b3) Coul omb + static friction model i ng and feedforward compensat i on, using a discrete time, single zero differentiating filter which compensat es for frictional memory In Fig. 56, taken from Johnson and Lor enz (1991), their results are present ed for bot h feedforward and feedback compensat i on and position and velocity errors. One of the maj or difficulties in performi ng friction compensat i on is the difficulty in model i ng friction at very low velocities. Several practical probl ems can appear as a consequence of doing friction compensa- tion on the basis of a di scont i nuous model . Conceptually, state variable model s are bet t er adapt ed to describe and hence t o compensat e for friction for very small velocities. They bet t er reflect the fact t hat friction (or any force t ransmi t t ed by a v u~ 8 :> -1 0 ,:, ;~ FF FB No Comp. ~ i . I 0.1 0.2 0.3 T~e( s ec) 0.5 0 O :=_ -0.5 0 ,~ " \ \ , , F F F B , . j : Comp? \ "/ " No 0.1 0.2 Time (see) 0.3 FIG. 56. State errors for three controllers: feedback friction compensation, feed forward friction compensation and uncompensated [from Johnson and Lorenz (1991), courtesy of the publisher]. compliant member) is a cont i nuous function of time. Examples of such models are t he Dahl model (Dahl , 1977), t he second-order Dahl model (Bliman and Sorine, 1991) and t he modified first-order Dahl model (Canudas de Wit e t a l . , 1993). When these model s are used as a basis for friction compensat i on, one maj or difficulty arises because t he internal states of these models are not measurable. However , internal state observers can be designed and stability of t he observer-based cont rol schemes can be shown (Bliman and Sorine 1991; Canudas de Wit e t a l . , 1993). I n connect i on with cont rol design, these model s enj oy the i nput - out put propert y of passivity or, mor e precisely, dissipativity. These propert i es can be explicitly exploited during the cont rol design leading t o an explicit det ermi nat i on of t he class of compensat ors t hat render the feedback l oop stable. Walrath (1984) report s a particularly sensitive and detailed study of friction phenomena duri ng velocity reversal of a tracking telescope. The technical challenge is considerable because t he telescope must be accurately poi nt ed from a movi ng pl at form. Response t o pl at form mot i ons involves frequent velocity reversals. Unlike t he machi ne t ool or robotics systems studied above, most of t he friction in Wal rat h' s mechani sm arises in the bearings. Wal rat h (1984) extends his model with an acceleration dependent decay of t he friction from one Coul omb friction level t o t he ot her during zero crossing; thus, there is no i nst ant aneous friction transition in his friction model or applied compensat i on. Wal rat h develops his model f r om t he Dahl model and does not at t empt t o explain t he underl yi ng physics; t he model , however, is consistent with the presence of frictional memory. Wal rat h report s a fact or of five i mpr ovement in RMS pointing error. 1124 B. ARMSTRONG-HELOUVRV e t al . Tung e t al. (1993) report an interesting variant on the t heme, they apply a corrective signal t o the input of the compensat or, G, ( s ) in Fig. 54. In this way compensat i on may be applied t o st andard industrial controls. 4.4.2. Fr i ct i on i dent i f i cat i on a n d a d a p t i v e c o n t r o l "Ther ef or e always when you wish t o know the quant i t y of the force t hat is requi red in or der t o drag the same weight over beds of different slope, you have t o make the experi ment and ascertain what amount of force is required t o move the weight along a level road, t hat is t o ascertain the nature of its fri ct i on. " Leonar do da Vinci (1452-1519), T h e N o t e b o o k s , F I I 106 r. Al t hough progress has been and cont i nues t o be made t oward the ability t o anticipate friction forces based on features of the mechanical design, e.g. (Kragelskii, 1988), this challenge remains a for- midable one and for t he foreseeabl e fut ure it will cont i nue t o be necessary, as Leonar do has suggested, t o ascertain by experi ment the friction paramet ers of a particular machine. The friction paramet ers may be det ermi ned either off-line, following a dat a gat heri ng experi ment , or continuously, on-line as part of operat i on of the machine. I f the paramet ers are t hen used in a model -based friction compensat i on, on-line identifica- tion becomes adapt i ve cont rol . The off-line and on-line identification schemes are most distinguished by the t ype of experimental mot i on: the designer of the off-line identification is oft en free t o specify t he mot i ons t hrough which dat a will be gat hered, whereas the on-line identification must normal l y use dat a from mot i ons dictated by t he operat i on of t he machine. The distinction is i mport ant in t hat there are issues of excitation: with common friction models, the terms of the model , or basis functions, are not automatically i ndependent or ort hogonal but may be t o a large degree ort hogonal i zed by pr oper choice of ex- perimental trajectories. Stated anot her way: when not using special trajectories, t he friction model terms may be strongly coupl ed and difficult t o accurately identify. But by pr oper choice of machi ne mot i on it is possible t o more nearly ort hogonal i ze the model basis functions, i.e. t o increase t he excitation. Several of the aut hors presenting schemes for off-line identifica- tion have capitalized on this possibility. The adapt i ve friction compensat i on has been pr oposed and demonst rat ed in many forms. The principal advant age of an adaptive scheme relative t o an off-line scheme lies in its ability to track changes in friction. 4.4.2.1. Of f - l i ne i dent i f i cat i on. The t ri boi ogy litera- t ure is rich in experi ment al work which might be descri bed as off-line identification of friction models and model paramet ers, e.g. Rabinowicz (1965) and Bowden and Tabor (1973). A large part of this work has been carried out with simplified friction contacts (e.g. sphere or cylinder on flat), with controlled normal forces, and direct and precise sensing of cont act displacements and forces, e.g. Pol ycarpou and Soom (1992). The controls engineer, on the ot her hand, must cont end with a machine designed for ot her purposes and sensing bet t er suited to the task of the machine than to friction identification. For these reasons, careful at t ent i on must be given t o the design of the off-line friction identification experiment. Friction identification, bot h off- and on-line, is made more challenging by the need for acceleration sensing to directly observe friction in mechanism mot i on. Accel erat i on sensing is not uncommon in the t ri bol ogy literature, and has been empl oyed by Ar mst r ong- Hr l ouvr y (1991) to identify friction in complex motions. Johnson and Lorenz (1991) t ake the alternative tack of using the desired acceleration as an estimate of the true acceleration. Their technique consists in a stepwise procedure for individually identifying the paramet ers of an inertia/friction model. The t orque command signal from the feedback compensat i on is used. Following the not at i on of Johnson and Lorenz (1991), their compensat i on is given by: [Total Command] T* = Tj~ + Te + (T,,,L~Ih or T,,,,~e) (32) [Standard Feedback] T~ = k r ( O* - O) + k,,(to* - to) (33) [Linear Feedforward] V,, r = Jib* +/~to* (34) [Friction Compensat i on/ Feedback] { ~ sgn (to) + bto; Itol > Ato (35) T , , ~ , = MI N ( L, , , ~ sgn (to)); Jtol < Ato [Friction Compensat i on/ Feedf or war d] / ~ sgn (to*) + bto*; [to*l > at o (36) T"'d~H= t MI N ( L, , , ~ sgn (to*)); Ito*l < Ato where T* is the total command t orque; Trh is the feedback t orque from the er r or correct i on process; k p , k,,, 0 and to are position and velocity gains and position and velocity, respectively; T~ is a linear feedforward term t hat compensat es for inertia and known dampi ng t hrough the paramet ers J and/ ~; tb* is t he desired acceleration; T.t,~rb or Tntdsif are t orque, nonl i near decoupl i ng state feedback, and t orque, nonl i near decoupling state feedforward, respectively. Onl y one of the terms T, Id.,~ or T,,a,,~ would be used at a time. The aut hors i mpl ement t he Kar nopp friction model, where Ato marks the small range of velocity over which static friction is taken t o apply; T,,,, is t he applied mot or t orque, estimated from commanded mot or current T*; ~ is the estimate of Coul omb friction; and ~ is the estimate of static friction. T,,a~, or T, la,rt differ as t o whet her measured velocity, to, or desired velocity, to*, is used. In the procedure of Johnson and Lorenz, the inertial paramet er, J, is identified first, t hen the Coul omb and viscous friction paramet ers, T, . , b, and finally the static friction paramet er, ~. As each paramet er is identified, that t erm is added t o a A survey of friction and controls 1125 feedforward or a feedback compensation. Ideally, as the entire model is identified the error correcting t erm of feedback control, T~, would go to zero. If the model is adequate, Tm can indeed become quite small (Armst rong-Hrl ouvry, 1991). Once the inertial paramet er is identified, its presence in the feedfor- ward path brings the desired acceleration into the estimation of the subsequent parameters. The parameters are identified by minimizing the error correcting t erm of the feedback control, T~. The procedure is distinctly off-line in that the parameters are identified sequentially. Care is taken to select experimental trajectories which will maximize the sensitivity of the identification to each paramet er as it is sought (Johnson and Lorenz, 1991). The stepwise identification can achieve higher excitation [coupling of the sought parameter(s) to the observed signal(s)] than can a simultaneous identification. While the lower excitation of a simultaneous identification might be partially compensated by allowing a longer run than the individual paramet er identification experi- ments of Johnson and Lorenz; the biasing influences of sensor noise---as opposed to process noise,--and systematic disturbances are not reduced by longer experimental runs (Armstrong, 1989a). Johnson and Lorenz apply their technique to friction in the joint of a motorized robotic gripper. Their apparatus allows t hem high bandwidth torque (current) control and sensing of actuator position The coupling of the actuator to the friction contacts, the mot or and a preloaded transmission element, is presumed to be quite rigid. This combination of features is adequate for accurate friction identification (Armst rong-Hrl ouvry, 1991); and is common to many of the report ed friction identification experiments, e.g. (Kubo e t al . , 1986; Townsend and Salisbury, 1987) Johnson and Lorenz (1991) employ the Karnopp friction model, with viscous friction added, and different friction paramet ers in the two movement directions. They also observe rising static friction. Their identification technique, coupled with model- based compensation, is quite successful at reducing the impact of friction on the closed-loop performance of their machine, as shown by Fig. 56. Armst rong-Hrl ouvry (1991) presents an experi- ment for identifying Coulomb + viscous friction that does not require acceleration, either measured or estimated. His experiment consists of long, open-loop #i des at constant torque and measuring average velocity. The technique was applied to a PUMA manipulator; initial transients were reduced with a manually tuned acceleration torque. The experiment is repeated at several torque levels, corresponding to several velocities; and a Coulomb + viscous friction model is fit to the (friction-velocity) data points obtained, as shown in Fig. 57. The technique depends upon viscous friction in the machine to create a stable gliding velocity. Using a breakaway experiment, Armst rong-Hri ouvry had identified position depend- ency in the friction and constructed a table lookup compensation. This compensation was applied in feedforward during the open-loop glides to better achieve a constant velocity. Armst rong-Hrl ouvry (1991) also demonstrated a technique involving closed-loop constant velocity glides and measuring average torque. This technique resulted in a higher noise content than did the open-loop glides, but could be applied to a machine without viscous friction. The closed-loop technique is like that of Johnson and Lorenz with the (average) acceleration taken to be zero during the constant velocity glides. Using a friction model comprising Coul omb+ viscous + position-dependent friction, Armstrong (1988) and Armst rong-Hrl ouvry (1991) demonstrated three axis open-loop motions of the robot using joint-wise spline trajectories. Model-based, open-loop control was not proposed as a viable architecture, rather the experiments were made to demonstrate the possibility of accurate friction identification. Spatial motions with an accumulated error of less than 10% were demonstrated. 4.4.2.2. Ful l m o d e l i dent i f i cat i on. In a Coulomb + viscous + static friction model, all of the param- eters enter the model in a linear fashion and may thus be identified by standard techniques. This is, for example, true of ~, /~, and ~ in equations (32)-(36). When Stribeck or rising static friction are incorpor- ated, the model parameters, specifically ~s and y in equation (10), bear a nonlinear relationship to the friction torque. To identify these parameters, nonlinear techniques are appropriate. Cheok e t al. (1988) employ the simplex method to determine the parameters of a Karnopp friction model, including the width paramet er Dv of equation (25). Like J/., in equation (10), Dv of the Karnopp model appears in a nonlinear relationship to the friction torque. Cheok e t al. (1988) point out that multiple minima are possible and pose a problem for gradient techniques. Armst rong-Hrl ouvry (1991) identified the parameters of a Coulomb + viscous + Stribeck + frictional memory + rising static friction + position-dependent friction model from friction data collected with the PUMA 560 robot. The frictional memory, rising static friction and position-dependent friction were identified separ- ately, in a step-wise fashion comparable to that of Johnson and Lorenz (1991). The remaining friction I S Z I O c o s U 0 (... I ~ - - S O ) C - t O 1 0 g . ! | B r e o L - A w o y L~ B r o o k - A w o y . ~ - ~ P r o J o c t o c l - ~ . . f . - ! ! f ~ " ~ P r o J o c t o d L o a o n d :
901 Con( I donco I I . 5 S I ~ t d . d o v . ) ' "' , i . . . . u . . . . i . . . . u . . . . ! - 2 " 1 0 1 2 V e L o c t ty ( R a d l s e c ) FIG. 57. Friction torque as a function of velocity for joint 1 of a PUMA 560 robot. The crosses indicated (friction- velocity) data points; the breakaway friction level has been recorded in a separate experiment [from Armstrong- Hriouvry (1991)]. 1126 B. ARMSTRONG-HI~LOUVRY e t a l . versus velocity model cont ai ned bot h nonl i near and linear paramet ers. To identify the paramet ers which ent er in a nonl i near fashion, an exhaustive search t echni que was used. At each point in the space of (two or four) nonl i near paramet ers, the four paramet ers entering linearly in the model were fit by regression analysis (the normal equat i on) and the residual er r or det ermi ned. The paramet er set giving the smallest residual was selected. A search technique was certainly necessary; multiple local minima were observed and would have arrest ed a gradient technique. The presence of noise in the data is t hought t o have i nt roduced local minima where intuitively t here ought to have been none. In addition to nonlinearity in certain paramet ers, some of the paramet ers of a full friction model may be very difficult t o measure. Armst rong-H61ouvry (1991) report s a measur ement of Stribeck and frictional memor y paramet ers. Using sensitive acceleration and force instruments and statistical techniques, he was able t o observe the Stribeck friction, which is domi nant at velocities near 5 milliradians per second, a velocity range in which stable mot i on is not possible for t he PUMA. Even with acceleration and force sensors, he was not able t o identify the frictional memor y par amet er directly. Rat her, he investigated the correl at i on bet ween t he best fit friction model and the actual friction signal (obt ai ned with measurement of acceleration). A 10% i mprovement in correlation was observed with the introduction of a 51ms frictional lag. Frictional memor y has been measured directly in sensitive t ri bol ogy experiments (Hess and Soom, 1990; Pol ycarpou and Soom, 1992), but with a quality of i nst rument at i on and cont rol of conditions t hat are probabl y unachievable in practical machines. It has been shown t hat all of the paramet ers of the seven par amet er model , equat i on (10), influence the presence of stick slip (Derj agui n e t a l . , 1957; Armst rong-H61ouvry, 1991); it may be possible t o indirectly identify the full friction model paramet ers by mappi ng the presence or absence of stick slip across a range of system stiffness, dampi ng and velocity conditions (Armst rong, 1988). If possible, such an identification t echni que would function with not hi ng more t han force actuation and position sensing. 4.4.2.3. A d a p t i v e c o n t r o l . The challenges t o adap- tive cont rol of a machi ne with friction are not unlike the general challenges of adapt i ve cont rol : probl ems of stability, the need for persistent (indeed, sufficient) excitation, difficulties t hat arise when the true model is not in the model set, the want of met hods for setting rat e and ot her paramet ers in the adaptive algorithm, etc. With friction, the motivation for adapt i ve cont rol is also not unlike the general mot i vat i on: friction will, in many cases, be a variable quant i t y which the cont rol l er must track. A very substantial number of papers concerni ng adaptive cont rol in robotics and elsewhere have t ouched on friction; here we focus on several papers where friction has been a maj or concern in the design of an adapt i ve cont rol algorithm. All of t he papers surveyed empl oy a model -based friction compensat i on and adaptively updat e the model parameters. The adaptive algorithms, however, span a great range; including the recursive-least-square (RLS) and least- mean-square (LMS) algorithms of Walrath (1984) Craig, (1986, 1987), Canudas de Wit (1988) and Canudas de Wit e t al . (1987, 1991); the model reference adapt i ve control ( MRAC) algorithms of Gilbart and Wi nst on (1974), Brandenburg and Sch~ifer (1988b, 1989), Sch~ifer and Brandenburg (1990), Yang and Tomi zuka (1988) and Maquei ra and Masten (1993); or the Lyapunov function based algorithm of Friedland and Park (1992); which has been extended t o multi-mass systems by Friedland and Ment zel opoul ou (1993). Adapt i ve control has been richly described in the literature, e.g. (Widrow and Sterns, 1985; Ast r 6m and Wi t t enmark, 1989). Onl y brief descriptions of algorithms will be given here. An interesting experimental compari son of several adap- tive algorithms and PI D control in a low-velocity position-tracking task with velocity reversal is present ed in Leonar d and Krishnaprasad (1992). The Recursive Least Squares (RLS) and Least Mean Squares (LMS) algorithms Intuitively the simplest algorithms, the recursive least squares (RLS) and least mean squares (LMS) algorithms function by determining the correlation between an error signal and the state dependent basis functions that make up the model. The model is constructed: y ( t ) = q~, ( t ) O, + q~z(t )02 + q0~(t)03 + - . . , (37) where y is an out put signal, the q~i(t) are model basis functions of the system state; as in the case of Coul omb friction, the basis functions need not be linear; and 0, are the (time invariant or slowly varying) model parameters. Not e t hat t he 0~ make a linear contribution to y, whet her or not the q0~(t) are linear. Following an example from Canudas de Wit e t al . (1991) (with gravity and Stribeck friction terms omi t t ed for simplicity) one might have for a single mass system: where ~( t ) is the regressor vector, the vect or of qgi. m = r - F,,k - Fc sgn (2) [System Dynamics] r ( t ) = O r ~ ( t ) [Model in 0/q~ Form] q~(t) = [(t), k(t), sgn ( ~ ( t ) ) ] " 0 r = [ m, F , , E ] (40) and where r(t ) is the applied t orque; x ( t ) is the position variable; and m, F, and E are mass, viscous and Coul omb friction paramet ers, respectively. An error signal is const ruct ed: e ( t ) = r ( t ) - Or ~ ( t ) , (41) where 0 is the vect or of estimated parameters. An updat e equat i on is const ruct ed: O r = ) ~ ( t ) P ( t ) ~ ( t ) e ( t ) = A ( t ) P ( t ) ~ ( t ) ( r ( t ) - 0r ~( t ) ) , (42) A s ur ve y o f f r i ct i on a nd c ont r ol s 1127 where ).(t) is a rat e gain; and P ( t ) is t he inverse of the input correlation matrix for t he RLS al gori t hm or t he identity matrix for t he LMS algorithm. G o o d convergence propert i es can be shown for t he RLS a n d LMS algorithms when conditions are ideal, i.e. IP(t)12 is bounded (persistent excitation) and not t o o big (sufficient excitation); and the t rue system is in the model set; t hat is t o say, t here exists a choice of paramet ers, 0", such t hat r ( t ) = o * r ~ ( t ) . Canudas de Wit e t al . (1987) study a basic case and show convergence and per f or mance i mpr ovement for a DC mot or/ t ransmi ssi on mechanism. Canudas de Wit e t al . (1991) ext end this result t o t he i mport ant case of Stribeck friction, and show t he possibility of destabilizing over-compensat i on when adapt at i on occurs at low velocities. The probl em arises when a Coul omb friction t erm adapts t o compensat e for high friction at very low velocities and t hen over- compensat es at higher velocities. Because models for Stribeck friction cont ai n paramet ers which do not appear linearly in t he out put , such as 2s in equat i on (10), Canudas de Wi t e t al . (1991) propose the use of a squar e- r oot velocity t erm: rs(2 ) = ~ r ~ ; ~ = [sgn (2), 1211'2sgn (~), 2] r (43) b T = [re, ~, F~,], where t r ( t ) is t he friction term. They show t hat t he square-root -vel oci t y t erm can he used t o closely mat ch (fri ct i on-vel oci t y) curves pr oposed by Tustin (1947) and ot hers; as seen in Fig. 58. It appears, however, t hat this will only be t he case if the break in the Stribeck friction curve lies in a velocity range comparabl e t o t hat of the break (the greatest second derivative) in the square-root -vel oci t y basis funct i on; and t hat the basis functions might have t o be replaced with = sgn( 2) , 2 , ~2sgn(2), 2 , (44) where 2~ is a scaling par amet er in velocity t hat can be used t o locate t he velocity range of the break in t he curve. Even t hough 2s could not be adj ust ed by a linear scheme, if it could be set a p r i o r i t o an 1.5 N m 1.0 - ~ %fmodel ( 2 . 2 ) ~f mo de l ( 2 . 3 ) 0.5 I I I I I 0 0.I 0.2 0.3 0.4 0.5 ( r a d / s ) FIG. 58. Comparison of the exponential friction model [of Tustin (1947)] and the estimated linear model, equation (43) [from Canudas de Wit et al. (1991)]. appropri at e range, t he linear identification scheme might perform satisfactorily. Craig (1986, 1987) presents an LMS al gori t hm comparabl e t o t he RLS algorithm of Canudas de Wit e t al . (1987). Crai g' s (1986, 1987) compensat i on for friction is part of a larger identification of the inertial paramet ers of the first t wo degrees of f r eedom of an Adept robot . Experi ment al results are present ed, including identified friction paramet ers; but from the results it is difficult t o det ermi ne whet her the adapt ed paramet ers either reflect true friction values or significantly i mprove performance. In practical i mpl ement at i ons, the use of accelera- tion in ~( t ) poses a considerable challenge. Versi ons of t he algorithm with filtered signals have been presented which do not empl oy acceleration signals, but require a high pass filtering step on t he velocity, (Hsu e t a l . , 1987). As did Johnson and Lor enz (1991) with the off-line identification, SIotine and Li (1987, 1989) have empl oyed command acceleration rat her than actual acceleration in the regressor vect or of a sliding mode adaptive algorithm. The aut hors present a pr oof of t he stability of their algorithm using Lyapunov' s direct met hod; and t hey present very successful adaptive control of a robot mani pul at or, of which friction compensat i on is a part. Model Reference Adapt i ve Cont rol ( MRAC) Gilbart and Wi nst on (1974) were t he first t o implement adaptive friction compensat i on, and do so with a model reference adapt i ve controller. , ~str0m and Wi t t enmark (1989) present a good discussion of the MRAC. Br andenbur g and Schiifer (1988b, 1989), Schiifer and Br andenbur g (1990), Hel d and Mar on (1988) and Maron (1989a, b) also present friction compensat i on in an MRAC framework. Giibart and Wi nst on (1974) i mpl ement an MRAC with a first-order reference model and a Lyapunov function chosen t o eliminate acceleration terms. The result is Coul omb friction compensat i on given by: u = (PD Cont rol ) + K3 sgn (2) K3 = B3 ~ sgn (2) dt + C3 ~ sgn (2) (45) =, t , , - 2 , where k is velocity; 2m is model velocity; K3 is a Coul omb friction compensat i on paramet er; and B3 and C3 are positive const ant s which are det ermi ned by the paramet ers of t he Lyapunov function. As opposed t o equat i on (40), (45) does not involve acceleration. Gilhart and Winston present a remarkabl e anal og comput er i mpl ement at i on t hat achieves a 6X reduc- tion in RMS pointing er r or of a t racki ng telescope. Brandenburg and Sch~ifer (1988b, 1989) and Sch~ifer and Brandenburg (1989, 1990, 1993) empl oy an MRAC structure t o adapt the paramet ers of a Coul omb friction compensat i ng di st urbance observer for a two mass flexible system with backlash. A block di agram is shown in Fig. 59. Wi t hout friction compensat i on, t he system exhibits two stick-slip limit cycles. Though t he details are not specified, it is 1128 B. ARMSTRONG-HI~LOUVRY e t a l . Dl w [- . . . . . . . L r ~nc 1 2 nlM 4 _ 5 e y "7 J . . . . . E l 1 i l = m l -,1, reference model adaptation speed controller current controller disturbance observer T o , r . . . . . ~ _ L /,~ mRL C12 m12 f ' / 7 ~ F KB t VB ~ 7 - ~ 1 I I d , 2 /7~R5 n 2 FiG. 59. Speed-controlled elastic two-mass system with reference model [from Brandenburg and Sch~ifer (1988b), courtesy of the authors]. indicated t hat t he MRAC design is based upon a Lyapunov funct i on; t he result is a friction compensa- tion made by appl yi ng a lag (PI) filter t o d, t he difference bet ween model and t rue velocity; just as in equat i on (45). Combi ned with i nt egrat or deadband, their al gori t hm eliminates stick slip and reduces standstill intervals at speed reversals by a fact or of five. Lyapunov and Lyapunov-l i ke met hods Fri edl and and Park (1992) present an adapt i ve friction compensat i on scheme which is based upon a Lyapunov-l i ke ar gument involving t he position error. T h e result is an updat e law t hat does not depend upon acceleration measur ement or estimation. The friction compensat i on is given by [modifying the not at i on of Fri edl and and Park (1992)]: m e = u - f ( yc, E * ) [System Dynami cs, single mass] f ( ~ , ) = Pc s ~ n ( 2 ) [Friction Model] u ( t ) = (St andard Cont rol ) + Pc sgn (2) [Control Law] Fc = z - k 121" [Friction Est i mat or] i = k# I ~l "- ' 1 (u - f ( - L Fc)) sgn (2) m [Friction Est i mat or Updat e Law], wher e m is syst em mass; u is t he cont rol input; z is a defined funct i on; Pc is t he est i mat ed Coul omb friction; and # and k are tunable gains. Defining the model misadjustment: e = f * - F~ (51) one finds t hat = - k # 12l"-' e (52) making e = 0 the stable fixed point of the process. Friedland and Park (1992) do not present experi men- tal results, but report simulations of a single degree of freedom system with Coul omb friction and the adaptive algorithm run with several values of # and k. The algorithm significantly i mproves dynami c response. The RLS/ LMS, MRAC and Lyapunov-based adapt i ve laws have in common a relationship between the integral of acceleration er r or and t he est i mat ed friction constants. In the RLS and LMS algorithms, acceleration appears explicitly and, scaled by mass, is subtracted from t he applied t orque t o give t orque or acceleration error, which det ermi nes t he rat e of change of the estimated paramet er. The est i mat ed paramet er is thus in pr opor t i on with the integral of acceleration error. In terms of physical dimensions, this situation holds whet her t he actual acceleration, desired acceleration or filtered velocity is used. In the MRAC algorithms, a PI filter is applied to the velocity error signal. The ' P' t erm gives a contribution proport i onal t o the integral of accelera- tion error. In the algorithm of Friedland and Park (1992), the t erm 2 is proport i onal t o t he time derivative of k [J?[", where ( 1 / m ) ( u - f ( 2 , Fc)) has t he physical dimension of . Since 2 is in proport i on with the derivative of a function of velocity, z - k 12[" is in proport i on with the integral of acceleration error. The RLS algorithm is a recursive version of t he normal equat i on from statistics, which gives t he squared-error optimal estimate of the paramet ers, but A s ur ve y o f f r i ct i on a n d c ont r ol s 1129 requires t hat t he basis functions, i.e. accel erat i on, be known. The ot her algorithms may be seen as implementing filtered versions of t he normal equa- tions; versions which are perhaps not squared-error optimal, but which are i mpl ement abl e wi t hout measur ement of acceleration. A t hor ough analysis of t he statistical implications of t he several estimators has not been presented. Learni ng Cont rol Learni ng cont r ol - - somet i mes called repetitive cont r ol - - i s a process of devel opi ng feedforward corrections for a specific t raj ect ory. It has been pr oposed for robotics and ot her servos (Craig, 1986; Kuc e t a l . , 1991) and is available on some industrial controllers (see Section 4.5). General l y, t he learning cont rol compensat i on t akes t he form of a table of corrections t o be added t o t he cont rol signal during the execut i on of a mot i on. The table must be learned during executions of t he precise mot i ons t o be correct ed, somet hi ng which will not be a limitation when t he application involves repetitive mot i ons. A correction table devel oped in this way will include frictional forces. 4.5. I nput f r om Engi neers in Indust ry who have Cont rol l ed Machines with Friction An informal survey has been conduct ed of engineers in industry who have cont rol l ed machi nes with friction. Engi neers were cont act ed t hr ough t he professional societies, references in t he literature and referrals. While not all cont act ed were equally concerned about privacy, it was somet i mes an issue; and so the results of this survey will be r epor t ed without reference t o t he specific cont ri but or. The cont ri but ors include engi neers with 23 com- panies in Eur ope, Japan and t he Uni t ed States, and several government laboratories. The machi ne t ool industry was t he most represent ed among t he respondent s, followed by precision pointing applica- tions (oft en for gover nment systems), robot i cs and all others. A great breadt h of t echni ques were f ound t o be in service: system hardware modification; high servo gains (stiff position and velocity cont rol ); modifications t o integral cont rol ; linear adaptive cont rol ; model -based friction compensat i on: - - f i xed; - - adapt i ve; dither; table l ookup compensat i on; learning cont rol ; joint t orque cont rol ; and variable structure cont rol . System hardware modification is by far t he most common and successful appr oach t o over comi ng frictional disturbances. As opposed t o t he cont rol s literature, cont rol s engi neers oft en ment i oned issues of machine design and particularly lubrication as t he first and perhaps onl y necessary step t o achieving t he needed performance. Machi ne modification did not always consist of reduci ng t he overall friction. I n t he machine t ool industry, for exampl e, "fri ct i on mat e- rials" such as Rul on , are used t o provi de a high Coul omb friction with a reduced excess of static over Coul omb friction. The friction materials consume substantial energy, but serve t o reduce mass-spring oscillations as well as frictional limit cycles, and t hereby increase the robustness of t he controller. A number of cases were encount er ed in which a new lubricant choice solved a stick-slip probl em, reflecting the fact t hat friction modification is not always considered in t he initial lubricant selection. Di t her is most successful in cases where t he vibration can be applied wi t hout great at t enuat i on t o t he friction contact. Hydraul i c systems are suitable, and several st andard hydraulic servo controllers include provisions for di t her in t he spool valve command signal. The earliest applications of di t her were not in the cont rol signals, but were made by attaching a vibrating el ement - - t ypi cal l y an eccentric wheel - - t o the machi ne (Bennet , 1979). No applica- tions of this t ype were report ed. I n each field, cases can be f ound where machine design and lubrication have been pushed t o their nat ural limits and i mpr ovement in performance requires enhancement s t o t he servo control. I n stiff systems, high position and velocity servo gains are used t o over come frictional disturbances. The t echni que is applied with t he greatest success in poi nt i ng devices, which do not carry a payl oad and can be designed specifically t o meet mechanical rigidity requi rement s. When high gains have been applied successfully in robotics, it has been observed t hat many feat ures of the system must function harmoni ousl y, including sensor characteristics and properl y t uned l ead- l ag compensat i on. In some applications, variable structure systems have been engi neered which apply higher servo gains near zero velocity, in or der t o meet t he demands of nonl i near friction at low velocities and t he desired performance characteristics at higher velocities. Joint t orque cont rol has been discussed in Section 4.3.5. Engi neers worki ng with joint t or que cont rol in commercial robotics report t hat t he residual friction is imperceptible, and t hat t he cl osed-l oop controllers function well in position, force and transitional tasks. Integral cont rol , or lag compensat i on, is certainly very common in practice; much mor e so t han is suggested by its st andi ng in the cont rol s literature. The integral cont rol t erm aggravates stick slip and introduces hunting; and a great many mechani sms for modifying integral cont rol are available in off-the-shelf servo controllers. A deadband in the posi t i on-error input t o t he i nt egrat or eliminates hunting, but introduces a threshold in t he precision with which a servo can be positioned. The use of lag compensat i on with high but finite DC gain will accomplish somet hi ng of t he same end wi t hout i nt roduci ng a nonlinearity. Int egral cont rol terms are also modified by resetting the i nt egrat or when mot i on is det ect ed (appropri at e when short mot i ons wi t hout over shoot are sought ), and saturating t he integral t erm as a AUTO30-7-D 1130 B. ARMSTRONG-HI~LOUVRY et al. function of posi t i on or vel oci t y er r or t er ms, a deadband- l i ke const r uct i on which allows full i nt egral cont rol act i on dur i ng mot i on. One very subst ant i al chal l enge i nt r oduced by i nt egral cont r ol ari ses because dur i ng mot i on in one di rect i on t he i nt egr at or wi nds up t o compens at e for Coul omb fri ct i on in t hat di r ect i on, and when t her e is a reversal , t he i nt egral cont r ol suddenl y compounds r at her t han compensat es for Coul omb fri ct i on. Thi s behavi or enl ar ges ( but does not cr eat e) quadr at ur e glitches in mul t i -axi s machi ni ng s t a t i ons - - pr oba bl y t he fri ct i onal di st ur bance of gr eat est economi c i mpact . One appl i cat i on was f ound in which t he i nt egral cont rol t er m is mul t i pl i ed by t he sign of vel oci t y. Thi s t echni que will l i kel y be successful so l ong as t he i nt egral cont r ol is not compensat i ng for anyt hi ng ot her t han Coul omb fri ct i on. A mor e common t echni que is to reset t he i nt egr at or at vel oci t y reversal . Adapt i ve cont r ol is pr esent in a growi ng number of i ndust ri al appl i cat i ons. In its l i near f or m, t he adapt i ve al gor i t hms ar e cer t ai nl y r espondi ng to fri ct i on. The possi bl e danger s of over - compens at i on are not i nsubst ant i al , because t he t rue model is not in t he model set for t hese syst ems. Model - bas ed fri ct i on compensat i on is al so empl oyed and offsets some of t he l i mi t at i ons of bot h i nt egral and l i near adapt i ve cont rol . The Coul omb fri ct i on t er m expl i ci t l y mul t i pl i es t he sign of vel oci t y in t hese syst ems, al l owi ng t he i nt egral cont r ol t er m, if pr esent , t o accommodat e ot her di st ur bances. Appl i cat i ons wer e found in which t he Coul omb fri ct i on compens at i on was based on desi r ed vel oci t y or upon a model vel oci t y, in a model r ef er ence adapt i ve cont r ol l er . No cases were f ound in which Coul omb fri ct i on compensat i on was based on sensed or es t i mat ed vel oci t y, suggest i ng t hat pr obl ems ari se when t he abr upt t ransi t i on is based on possi bl y noi sy measur ement s. In pract i ce, it seems, t he Coul omb fri ct i on compensat i on is always coupl ed t o an adapt i ve, or t o an aut o- t uni ng cont r ol l er , which adj ust s t he Coul omb friction par amet er . I ndi cat i ons of t he range over whi ch fri ct i on woul d vary spanned from t ens of per cent for machi ne t ool s t o hundr eds of per cent in aer ospace appl i cat i ons t hat funct i on over a very wi de t emper at ur e r ange. The quest i on of r ange of fri ct i on var i at i ons to expect , however , r emai ns a r emar kabl y el usi ve one. Al t hough pr obl ems with over - compens at i on wer e r epor t ed, t he range of appl i cat i ons t o whi ch model - based Coul omb fri ct i on compens at i on has been appl i ed suggests t hat it is an effective t ool . Model - based f eedf or war d schemes using mor e compl - et e friction model s wer e not r epor t ed, but i nt er est was f ound in such a possi bi l i t y. Tabl e l ookup compensat i on has been used in machi ne t ool s for some t i me, t o compens at e for vari abl e backl ash. The t abl es ar e filled in l engt hy aut o- t uni ng pr ocedur es; and cer t ai nl y, in conj unct i on with t he servo stiffness and t uni ng pr ocedur e, i ncl ude a f act or t hat is dependent on t he fri ct i on. Compens a- t i on t abl es t uned speci fi cal l y t o fri ct i on wer e not r epor t ed. Lear ni ng cont r ol schemes are comi ng on-l i ne in commer ci al cont rol l ers. These syst ems i dent i fy and compensat e for a di st ur bance signal which cert ai nl y i ncl udes friction. Because so many appl i cat i ons in i ndust ry are r epeat ed, t he l earni ng cont r ol l er s will cert ai nl y become wi de spr ead if t he per f or mance enhancement is near what t he manuf act ur er ' s cl ai m. Al t hough consi der abl e research at t ent i on has been given t o t he anal yt i c t r eat ment of stick slip, engi neer s in i ndust ry wor ki ng with preci si on machi ni ng, opt i cal t racki ng, di sk head posi t i oni ng or r obot i c force cont rol r ar el y r epor t stick slip as t he pr obl emat i c fri ct i onal behavi or . St i ck sl i p, wher e it ari ses, is most oft en di spached with appr opr i at e choi ce of l ubri cant , or modi fi cat i ons to t he mechani sm or i nt egral cont rol . Fri ct i onal forces, none- t he- l ess, i nt r oduce di st ur- bances such as lost mot i on at vel oci t y reversal or over shoot due to i nt egr at or wi ndup. These di st ur- bances are most i mpor t ant when accur at e t r acki ng is r equi r ed. Thus, at t ent i on needs t o be given to accurat e fri ct i on model i ng in pract i cal machi nes, and t o aggressi ve, and in many cases model - bas ed, di st ur bance r ej ect i on. Many of t he engi neer s sur veyed r epor t ed rel i ance upon, and di ssat i sfact i on wi t h, by- guess- and- by- gosh met hods used for fri ct i on compensat i on desi gn. By far t he most syst emat i c t echni ques lie in t he real m of pr obl em avoi dance, e. g. machi ne desi gn for f avor abl e friction charact eri st i cs. When it comes t o cont r ol , t her e seemed t o be no t ool s useful even for choosi ng bet ween t he var i et y of deadbands and sat ur at i on funct i ons avai l abl e for i nt egral c ont r ol - - not t o ment i on t uni ng t hese syst ems. No one cont r i but i ng to this survey r epor t ed an anal ysi s t echni que consi st ent l y giving resul t s useful for appl i cat i on. Of t he anal ysi s t ool s, i nt erest was gr eat est in ext ensi ve si mul at i on. But sel ect i on of cont r ol l er desi gn and par amet er t uni ng r emai n l engt hy, hand pr ocedur es. In cont r ast t o t he avai l abi l i t y t ool s for anal ysi s, however, was t he f r equency with which successful appl i cat i on of t he vari ous compensat i ons were r epor t ed. In r obot i cs, in preci si on poi nt i ng and posi t i oni ng devi ces and in sophi st i cat ed machi ne t ool s, cont r ol funct i ons i ncl uded specifically for fri ct i on compensat i on ar e not uncommon. 5. CONCLUSION Li ke t he el ephant encount er ed by six bl i nd men, fri ct i on in machi nes is a mul t i f acet ed phenomenon, i ncor por at i ng Coul omb and viscous fri ct i on, nonl i near fri ct i on at low vel oci t i es, t empor al phenomena and t he el ast i ci t y of t he i nt erface. In any given ci rcumst ance, some f eat ur es may domi nat e over ot hers, and some f eat ur es may not be det ect abl e with t he avai l abl e sensing. But all of t hese phenomena ar e pr esent all of t he t i me in fluid l ubr i cat ed met al cont act s and, in many cases, pr esent in dr y cont act s as well. The use of a mor e compl et e fri ct i on model will ext end t he range of appl i cabi l i t y of anal yt i c resul t s and resol ve di scr epanci es t hat ari se when di fferent i nvest i gat i ons ar e based on di fferent phenomena, each of which domi nat e under some ci rcumst ances. A s ur ve y o f f r i ct i on a n d c ont r ol s 1131 Anal yses based on a compl et e friction model will be able t o consistently account for t he observed behavi or of systems. Successful servo-cont rol t echni ques have been demonst rat ed which compensat e for friction in machines. Stiff cont rol , model -based feedforward cont rol , adaptive strategies and variable st ruct ure control have all received at t ent i on in bot h t heor y and application ori ent ed efforts. I ndeed, t he compensat i on techniques seem furt her advanced t han t he analysis techniques for cont rol l er design. Friction compen- sators are t oo oft en t uned by hand. Tool s, as yet , do not exist which can correct l y and over a br oad range of conditions predict t he presence of stick slip, let alone per f or mance in subtler measures, such as tracking error or optimality. 5.1. A Paradi gm for Fut ure Wor k The oppor t uni t y exists for very substantial progress in t he cont rol of machi nes with fri ct i on--progress which may ultimately identify and reach the limits of per f or mance achievable with a given machine. But t o realize this progress, a new paradi gm must be applied. 5.1.1. Friction model s must have a theoretical and experimental f oundat i on The mechani cs of sliding cont act s is certainly less well established t han t he mechani cs of inertial bodies. None-t he-l ess, friction models and frictional behavi or must be selected with justification. In some cases in t he past, el ement s of a friction model have been selected because t hey are able t o account for behavi or under i mmedi at e investigation. While this might be adequat e if t he slate were blank the triboiogy, mechani cs and lubrication engineering literatures provi de a br oad theoretical and experimental foundat i on upon which details of a friction model may be based. I n many cases, the experi ment s present ed have been conduct ed with a level of sensing beyond what is normal l y available t o the cont rol s engineer. This survey of the t ri bol ogy literature is certainly not exhaustive; and t he i nt egrat ed friction model present ed will certainly be superceded as new ideas and results emerge. Thus t he model present ed is neither the onl y model support ed by t he literature, nor t he final wor d on t he question. But alternatives t hat are proposed or empl oyed as a basis for analysis must be gr ounded in t he science of friction. 5.1.2. In experimental work, lubrication must be addressed The details of lubrication absolutely domi nat e frictional instability and play a leading role in all ot her frictional behavi or. Yet papers in the controls literature rarely ment i on the lubricant empl oyed. An experi ment r epor t ed wi t hout reference t o t he lubricant is not a reproduci bl e experiment. Even if the reference is not hi ng more t han a manufact urer' s part number, the possibility will remai n open for ot her investigators t o reproduce the results or investigate t he role pl ayed by t he lubricant. The issue of lubrication is i mport ant on mor e t han one level. Lubri cat i on is not normal l y a concern of t he controls engi neer; but t hen stick slip is not normal l y a concern of t he lubrication engi neer, who is mor e oft en concerned with limiting machi ne wear and issues of the lubricant envi ronment , service life and delivery. Someone must give at t ent i on t o t he stick-slip properties of the lubricant. One lubrication engi neer, interviewed in t he course of this proj ect , went so far as t o say t hat stick slip could be eliminated in any application by appropri at e choice of lubricant. This perhaps overstates t he possibilities, but the poi nt is made t hat the impact of lubricant cannot be neglected. 5.1.3. Anal yses must be verified An analysis t ool cannot reach its potential utility until it is established t hat t he t ool correct l y predicts behavi or across t he i nt ended range of application. Theor y, simulation and experi ment are mutually support i ve in this regard; and analysis tools should, at a mi ni mum, be verified against many simulations carried out across a range of paramet ers. While t he analysis t ool / si mul at i on synergy cannot verify t he correctness of a friction model , it can verify t hat t he approxi mat i ons i ncorporat ed within t he analysis itself are valid. This is an issue with describing function analysis, f or exampl e, where very general results may be possible, but t he utility of t he results quite limited. Verification by experi ment is mor e involved t han simulation, particularly across a br oad range of operat i ng conditions. Ther e are opport uni t i es, how- ever, in the literature t o find experi ment s report ed in sufficient detail t hat new analysis tools may be applied and verified using the results of report ed experiments. This is perhaps an appropri at e st andard for fut ure articles report i ng experimental investigations: t hat t he experiments be adequat el y described so t hat ot her investigators can use t he results t o apply and verify new and developing analyses and models. The work of Brandenburg et al. demonst rat es this possibility. 5.1.4. Data showi ng the repeatability o f friction are needed One of the assumptions underl yi ng model -based cont rol , and perhaps any cont rol analysis, regards t he repeatability of t he process in question. Quest i ons of t he feasibility of model -based compensat i on tech- niques, t he i mport ance of adaptive cont rol , and t he robustness requi red of controllers all hinge on t he repeatability of friction. And yet the literature is nearly silent on this question. Ever y report of experimental work with friction and cont rol should include ment i on of the repeatability of t he observed frictional behavior. It is clear t hat a number of fact ors influence friction, principal among t hem const ancy of lubrication, t emperat ure of operat i on and t he state of wear. And t hat these factors will give rise t o variations over time or operat i ng conditions, and perhaps t o variations which appear random. As report s of repeatability under specific experi ment al conditions become available, it will be possible t o synthesize an i nformed view of t he anticipated repeatability of friction in a mechanism. Such a result will mark very 1132 B. ARMSTRONG-HI~LOUVRY et al. cons i der abl e pr ogr ess i n t he ar ea of f r i ct i on and cont r ol . Wi t h a t r i bophysi cal l y j ust i f i ed model , r i gor ousl y veri fi ed t ool s for anal ysi s, c ons i de r a t i on of all of t he syst em at t r i but es t hat affect f r i ct i on, a nd avai l abl e dat a r egar di ng such i ssues as r epeat abi l i t y; t her e is gr eat possi bi l i t y of pr ogr ess i n t he cont r ol of machi nes wi t h fri ct i on. By c ombi ni ng t he s t r engt hs of exi st i ng anal ysi s t ool s, such as t he i nt egr at ed pl a nt / f r i c t i on descr i bi ng f unct i on, wi t h a mor e c ompr e he ns i ve f r i ct i on model , mor e gener al t ool s may be devel oped. Ot he r avenues , such as gener al i zi ng t he pr oj ect i ve phase pl ane t echni ques of Radcl i ffe and Sout hwar d (1990), may al so l ead to t hi s goal . Bet t er anal ysi s t ool s will per mi t bet t er ut i l i zat i on of de mons t r a t e d com- pe ns a t i on t echni ques , such as adapt i ve, mode l - ba s e d Co u l o mb f r i ct i on c ompe ns a t i on; and c ompe ns a t i on t echni ques on t he hor i zon, such as i mpul s i ve cont r ol , will ope n new possi bi l i t i es. Be t t e r model s a nd anal ysi s t ool s, coupl ed wi t h c ompe ns a t i on st r at egi es, will l ead t o pr eci si on mot i on achi eved by l ower cost ma c h i n e s - - t h e or i gi nal a nd c ont i nui ng obj ect i ve of ser vo cont r ol . Acknowledgments--Partial support for this project has been provided by the University of Wisconsin, Milwaukee, by Beamon, Inc. and by the National Science Foundation under grant MSS-9302190. We would like to thank Dr Michael K. Masten for his careful review of the original manuscript. Our appreciation also goes out to all those who passed along their papers or suggested those of their colleagues. Many engineers in industry have contributed generously of their time to talk with us about their experiences with friction and control. The contributors include engineers with Adept Technologies, Inc., Allen Bradley, Inc., American Phauter, Inc., Asea Brown Bovari, AB, Astronautics Corporation of America, Inc., dSpace, GmbH, FMC, Inc., Galil, Inc., GEC-Marconi Corp., GE Fanuc, Inc., Giddings & Lewis, Inc., Hauser Elektronik, GmbH, Hewlett Packard, Inc., IBM Corp., Integrated Motion, Inc., Johnson Controls, Inc., Kistler-Morse Corp., Martin Marietta, Inc., Moog, Inc., NASA, Ora Corp., Robotics Research, Inc., Systems Engineering Corp., Texas Instruments, Inc., and Trellis Motion Control, Inc. We appreciate their input and believe it an important contribution to this survey. REFERENCES Ackermann, J. and P. C. Miiller (1986). Dynamical behavior of nonlinear multibody system due to Coulomb friction and backlash. IFAC/IFIP/IMACS Int. Syrup. on the Theory of Robots, Wein, pp. 289-95. Ami n, B. (1993). The reliability of describing function predictions for systems with friction. M.S. Thesis, Dept. of Electrical Engineering and Computer Science, University of Wisconsin, Milwaukee. Amont ons, G. (1699). On the resistance originating in machines. Proc. of the French Royal Academy of Sciences, pp. 206-22. Anderson, B. D. O. , R. R. Bitmead, C. R. Johnson, Jr., P. V. Kokotovic, R. L. Kosut, I. M. Y. Marells, L. Praly and B. D. Riedle (1986). Stability of Adaptive Systems, Passivity and Averaging Analysis. Cambridge. MIT Press. Armstrong, B. (1988). Dynamics for robot control: friction modeling and ensuring excitation during parameter identification. Ph.D. Thesis, Dept. of Electrical Engineer- ing, Stanford University, Stanford Computer Science Memo STAN-CS-88-1205. Armstrong, B. (1989). Control of machines with non-linear low-velocity friction: a dimensional analysis. Proc. of the First International Symposium on Experimental Robotics, Montreal, Quebec, pp. 180-95. Armstrong, B. (1989a). On finding exciting trajectories for identification experiments involving systems with nonlinear dynamics. Int. J. of Robotics Research, 8(6), 28-48. Armstrong-H61ouvry, B. (1990). Stick-slip arising from Stribeck friction. Proc. 1990 Inter. Conf. on Robotics and Automation, Cincinnati, IEEE, pp. 1377-82. Armstrong-HElouvry, B. (1991). Control of Machines with Friction. Kluwer Academic Publishers, Norwell, MA. Armstrong-H61ouvry, B. (1992). A perturbation analysis of stick-slip. In R. A. Ibrahim and A. Soom (Eds), Friction-Induced Vibration, Chatter, Squeal, and Chaos, Proc. ASME Winter Annual Meeting, Anaheim, DE-Vol. 49, ASME, NY, pp. 41-48. Armstrong-HElouvry, B. (1993). Stick-slip and control in low-speed motion. IEEE Trans. on Automatic Control, 38(10), 1483-1496. Armstrong-H61ouvry, B. and P. Dupont (1993). Friction modelling for controls, and Compensation Techniques for Servos with Friction, Proc. 1993 American Control Conference, AACC, San Francisco; CA, pp. 1905-1915. Asada, H. and K. Youcef-Toumi (1984). Analysis and design of a direct-drive arm with a five-bar-link parallel drive mechanism. J. of Dynamic Systems, Measurement and Control, 106(2), 225-3[). ,&str6m, K. J. and B. Wittenmark (1989). Adaptive Control. Menlo Park; Addi son-Wesl ey, Reading, MA. Atfierton, D. P. (1975). Nonlinear Control Engineering. Van Nostrand Reinhold Co., London. Auslander, D. M. and K. Dass (1990). Stopping: the endpoint of mechanical motion control. Electrosoft, 1(3), 166-182. Bannerjee, A. K. (1968). Influence of kinetic friction on the critical velocity of stick-slip motion. Wear, 12(2), 107-116. Bell, R. and M. Burdekin (1966). Dynamic behavior of plain slideways. Proc, of the lnstn, of Mechanical Engineers, Vol. 181, pt. 1, no. 8, pp. 169-83. Bell, R. and M. Burdekin (1969). A study of the stick-slip motion of machine tool feed drives. Proc. of the lnstn, of Mechanical Engineers, Vol. 184, pt. 1, no. 29, pp. 543-60. Bennet t , S. (1979). A History of Control Engineering, Peter Peregrinos, Sterenge. Bentsman, J. (1990). Oscillations-induced transitions and their application in control of dynamical systems. J. of. Dynamic Systems, Measurement and Control, 112(3), 313-19. Bernard, J. E. (1980). The simulation of Coulomb friction in mechanical systems. Simulation, 34(1), 11-16. Biel, C. (1920). Die Reibung in Glietlagern bie Zusatz von Voltool zu Mineralol und bie Veranderung der Unlaufzahl und der Temperatur. Zeitschrift des Vereines Seutscher lngenieure, 64 (1920), 449-83. Bifano, T. and T. Dow (1985). Real time control of spindle runout. Optical Engineering, 24(5), 888. Blackwell, C. C., S. W. Sirlin and R. A. Laskin (1988). Precision pointing of scientific instruments on space station: The LFGGREC perspective. Proc. IEEE National Aerospace and Electronics Conf. NAECON 1988, IEEE, Dayton, pp. 566-73. Blau, P. J. (1987). A model for run-in and other transitions in sliding friction. J. of Tribology, 109(3), 537-44. Bliman, P. A. (1993). Mathematical study of the Dahl' s friction model. Euro. J. Mech., A/Solids, 11(6), 835-848. Bliman, P. A. and M. Sorine (1991). Friction modelling by hysteresis operators, applications to Dalai, Stiction and Stribeck effects. Proc. Conference on Models of Hysteresis, Pitman Research Notes in Mathematics, Trent o, Italy. Bliman, P. A. and M. Sorine (1993). A system-theoretic approach of systems with hysteresis. Application to friction modelling and compensation. European Control Conference, ECC'93, Groningen, The Netherlands, pp. 1844-t845. Bo, L. C. and D. Pavelescu (1982). The friction-speed relation and its influence on the critical velocity of the stick-slip motion. Wear, 82(3), 277-289. Bogoliubov, N. N. and Y. A. Mitropolsky (1961). Asymptotic Methods in the Theory of Non-linear Oscillatiott~. Gordon and Beach. NY. A survey of friction and controls 1133 Booser, E. R. (Ed.) (1984). CRC Handbook of Lubrication. CRC Press, Boca Raton, FL. Bowden, F. P. (1950). BBC Broadcast. Bowden, F. P. and L. Leben (1939). The nature of sliding and the analysis of friction. Proc. of the Royal Society, Series A, Voi. 169, pp. 371-91. Bowden, F. P. and D. Tabor (1939). The area of contact between stationary and between moving surfaces. Proc. of the Royal Society, Series A, Vol. 169, pp. 391-413. Bowden, F. P. and D. Tabor (1956). Friction and Lubrication, John Wiley and Sons, NY. Bowden, F. P. and D. Tabor (1973). Friction---an Introduction to Tribology. Anchor Press/Doubleday; Reprinted 1982, Malabar: Krieger Publishing Co, NY. Brandenburg, G. (1986). Stability of a speed-controlled elastic two-mass system with backlash and Coulomb friction and optimization by a load observer. Proc. IMACS-IFACS Syrup. on Modelling and Simulation for Control of Lumped and Dist. Parameter Systems, IFAC, Lilly, pp. 371-381: (In P. Borne and Tzafestas (Eds), Applied Modelling and Simulation Technological Systems. Elsevier, Amsterdam. Brandenburg, G. , and U. Schiller (1987). Influence and partial compensation of simultaneously acting backlash and Coulomb friction in a position- and speed-controlled elastic two-mass system. Proc. 2nd European Conf. on Power Electronics and Applications, EPE, Grenobl e, pp. 1041-1047. Brandenburg, G. and U. Sch~ifer (1988a). Stability analysis and optimization of a position-controlled elastic two-mass- system with backlash and Coulomb friction. Proc. 12th IMACS World Congress, IMACS, Paris, pp. 220-223. Brandenburg, G. , and U. Sch~ifer (1988b). Influence and partial compensation of simultaneously acting backlash and Coulomb friction in a position- and speed-controlled elastic two-mass system. Proc. 2nd Int. Conf. on Electrical Drives, ICED, Poicina Brasov. Brandenburg, G. , and U. Schiller (1989). Influence and adaptive compensation of simultaneously acting backlash and Coulomb friction in elastic two-mass systems of robots and machines tools. Inter. Conf. on Control and Applications, paper WA-4-5, I EEE, Jerusalem. Brandenburg, G. and U. Sch~ifer (1991). Influence and compensation of Coulomb friction in industrial pointing and tracking systems. Proc. of the Indus. App. Soc. Annual Meeting, I EEE, Dearborn, MI, pp. 1407-1413. Brockley, C. A. , R. Cameron and A. F. Potter (1967). Friction-induced vibration. J. of Lubrication Technology, 89(2), 101-108. Brockley, C. A. and H. R, Davis (1968). The time- dependence of static friction. J. of Lubrication Technology, 90(1), 35-41. Brogan, W. L. (1991). Modern Control Theory. Prentice- Hall, Englewood Cliffs, NJ. Budanov, V. B., V. A. Kudinov and D. M. Tolstoi (1980). Interaction of friction and vibration. Soviet J. of Friction and Wear, 1(1), 79-89. Burdekin, M., N. Back and A. Cowley (1978). Experimental study of normal and shear characteristics of machined surfaces in contact. J. of Mechanical Engineering Science, 20(3), 129-32. Cameron, A. (1984). On a unified theory of boundary lubrication. Proc. of 11th Leeds-Lyon Syrup. on Tribology, Leeds, Butterworths, London. Canudas de Wit, C. (1988). Adaptive Control for Partially Known Systems--Theory and Applications, Elsevier, Amsterdam. Canudas de Wit, C., K. J. ,~str~im and K. Braun (1987). Adaptive friction compensation in DC motor drives. IEEE J. of Robotics and Automation, RA-3(6). Canudas de Wit, C., P. Noel, A. Aubi n and B. Brogliato (1991). Adaptive friction compensation in robot man- ipulators: low-velocities. The Inter. J. of Robotics Research, 10(3), 189-99. Canudas de Wit, C., P. Noel, A. Aubi n, B. Brogliato and P. Drevet (1989). Adaptive friction compensation in robot manipulators: Low velocities. In Proc. Inter. Conf. on Robotics and Automation, I EEE, Scottsdal.e, pp. 1352-57. Canudas de Wit, C., H. Olsson, K. J. AstriSm and P. Lischinsky (1993). Dynamic friction models and control design. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1920-1926. Canudas de Wit, C., and V. Seront (1990). Robust adaptive friction compensation. Proc. Inter. Conf. on Robotics and Automation, I EEE, Cincinnati, OH, pp. 1383-1389. Cebuhar, W. A. (1988). Smoothing and approximate linearization of discontinuous control systems. Ph.D. Thesis, Dept. of Applied Mathematics, Harvard University. Chau, W. K. (1990). A critical analysis of dithering algorithms for image processing. IEEE Region 10 Conf. on Computers and Communication Systems, I EEE, Hong Kong, pp. 309-313. Cheng, J. H. and N. Kikuchi (1985). An incremental constitutive relation of unilateral contact friction for large deformation analysis. J. of Applied Mechanics, 52(3), 639-648. Cheok, K. C., H. Hu and N. K. Loh (1988). Modeling and identification of a class of servomechanism systems with stick-slip friction. J. of Dynamic Systems, Measurement and Control, 110(3), 324-328. Chou, W. (1990). Dithering and its effects on sigma delta and multi-stage sigma delta modulation. IEEE Inter. Syrup. on Orcuits of Systems, Part 1, I EEE, New Orleans, pp. 268-271. Cincinnati Milacron (1986). Revised Stick-Slip Test Procedure. Clingman, W. D. (1991). Flexure bearing reduces startup friction. NASA Tech Briefs, 15(11), 72-74. Cockerham, G. and M. Cole (1976). Stick-slip stability by analogue simulation. Wear, 36(2), 189-198. Cockerham, G. and G. R. Symmons (1976). Stability criterion for stick-slip motion using a discontinuous dynamic friction model. Wear, 40(1), 113-120. Coulomb, C. A. (1785). Th6orie des machines simples, en ayant 6gard au frottement de leurs parties, et a la roideur dews cordages. M~m. Math Phys., x, 161-342.Craig, . I. J. (1987). Adaptive Control of Mechanical Manipulators, Addison-Wesley, Reading, MA. Craig, J. J. (1987). Adaptive Control of Mechanical Manipulators, Addison-Wesley, Reading, MA. Craig, J. J. (1986). Adaptive control of mechanical manipulators. Ph,D. Thesis, Electrical Engineering Dept . , Stanford University, MA. Cutkosky, M. R. , 1. M. Jourdain and P. K. Wright (1987). Skin materials for robotic fingers. Proc. 1987 Inter. Conf. on Robotics and Automation, I EEE, pp. 1649-1654. Cutkosky, M. R. and P. K. Wright (1986). Friction, stability and the design of robotic fingers. The Int. J. of Robotics Research, 5(4), 20-37. Czichos, H. (1978). Tribology, Elsevier, Amsterdam. Da Vinci, L. (1519). The Notebooks, Dover, NY. Dahl, P. R. (1968). A solid friction model. TOR-158(3107- 18), The Aerospace Corporation, El Segundo, CA. Dahl, P. R. (1976). Solid friction damping of mechanical vibrations. AIAA J., 14(12), 1675-1682. Dahl, P. R. (1977). Measurement of solid friction parameters of ball bearings. Proc. of 6th Annual Symp. on Incremental Motion, Control Systems and Devices, University of Illinois, ILO. Derjaguin, B. V., V. E. Push and D. M. Tolstoi (1956). A theory of stick-slip sliding of solids. J. of Technical Physics (Moscow), 6. Derjaguin, B. V., V. E. Push and D. M. Tolstoi (1957). A theory of stick-slip sliding of solids. Proc. of the Conference on Lubrication and Wear, Instn. Mech. Engs., London, pp. 257-268. DeWeerth, S. P., L. Nielsen, C. A. Mead and K. J. ~,str6m (1991). A simple neuron servo. IEEE Trans. on Neural Networks, 2(2), 248-251. Dieterich, J. H. (1979). Modeling of rock friction: 1. experimental results and constitutive equations. J. of Geophysical Research, 84(B5), 2161-2168. Dieterich, J. H. (1991). Micro-mechanics of slip instabilities 1134 B. ARMSTRONG-HI~LOUVRY et al. with rate- and state-dependent friction. Eos, Trans. Am. Geophys. Union. Dohring, M. E., E. Lee and W. S. Newman (1993). A load-dependent transmission friction model: theory and experiments. Proc. 1993 Inter. Conf. on Robotics and Automation, I EEE, pp. 430-436. Atlanta, GA. Dokos, S. J. (1946). Sliding friction under extreme pressures-l. J. of Applied Mechanics. 13 (Series A), 148-56. Dowson, D. (1979). History ol" Tribology, Longman Ltd, London. Dowson, D. and G. R. Higginson (1966). Elasto- hydrodynamic lubrication--the fundamentals of roller and gear lubrication. Pergamon Press, Oxford. Dudley, B. R. and H. W. Swift (1949). Frictional relaxation oscillations. Philosophical Magazine, 40 (Series 7), 849-861. Dupont, P. E. (1992a). The effect of Coulomb friction on the existence and uniqueness of the forward dynamics problems. Proc. 1992 Inter. Conf. on Robotics and Automation, Nice, France, pp. 1442-1447. Dupont , P. E. (1992b). The use of compliance to resolve the existence and uniqueness of the forward dynamics solution with Coulomb friction. Proc. of the CSME Forum 1992, CSME, Montreal, Canada, pp. 537-542. Dupont, P. E. (1993). The effect of friction on the forward dynamics problem. Int. J. Robotics Research, 12(2), 164-179. Dupont, P. E. (1994). Avoiding stick-slip through PD control. I EEE Trans. on Automatic Control, 39(5), 1059-1097. Dupont , P. E. and D. Bapna (1992). Stability of sliding frictional surfaces with varying normal force. Submitted to Journal of Vibration and Acoustics, 116(2), 237-242. Dupont , P. and E. Dunlap (1993). Friction modeling and control in boundary lubrication. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1915-1919. Ehrich, N. E. (1991). An investigation of control strategies for friction compensation. M.S. Thesis, Dept. of Electrical Engineering, University of Maryland, MA. Eismann, P. (1992). Private Communication. Ellison, B. and J. Richi (1983). Inertial stabilization of periscope sights band driven three axle gimbal. In W. H. Taylor (Ed. ), Optical Systems Engineering II1, SPIE, Los Angles, pp. 107-120. Eschmann, P. (1985). Ball and Roller Bearings Theory. Design and Application, 2nd ed., revised by L. Hasbargen and J. Brandlein, John Wiley and Sons, NY. Estler, R. B. (1980). What' s new in preventive maintenance for gears and bearings. Conf. of the American Society of Lubrication Engineers, ASLE, Baltimore, MA. Fatunla, S. O. (1988). Numerical Methods f or Initial Value Problems in Ordinary Differential Equations, Academic Press, Boston, MA. Facchiano, D. L. and J. N. Vinci (1984). EP industrial gear oi l s- - a look at additive functions and a comparison of sulfur phosphorus and leaded gear oils. Lubrication Engineering, 40(10), 598-604. Friedland, B. and S. Mentzelopoulou (19931. Friction estimation in multimass systems. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1927-1931. Friedland, B. and Y.-J. Park (19921. On adaptive friction compensation. Proc. 30th CDC, 1991, I EEE, Brighton, pp. 2889-2903; I EEE Trans. on Automatic Control, 37(10), 1609-1612. Friedman, H D. and P. Levesque (1959). Reduction of static friction by sonic vibrations. J. Appl. Physics, 30(10), 1572-1575. Fuller, D. D. (19841. Theory and Practice of Lubrication f or Engineers. John Wiley and Sons, NY. Furusho, J. , H. Hagao, A. Sano and M. Naruse (1990). Vibration control of robot arm using joint torque feedback. Trans. Japan Soc. of Mech. Eng., C-56(528), 119-126. Futami, S., A. Furutani and S. Yoshida (1990). Nanometer positioning and its micro-dynamics. Nanotechnology, 1(1), 31-37. Gao, C. and D. Kuhl mann-Wi l sdorf (1990). On stick-slip and the vclocity dependence of friction at low speeds. J. of Tribology, 112 (April), 354-3611. Gassenfeit, E. H. and A. Soom (19881. Friction coefficients measured at lubricated planar contacts during start-up. J. of Tribology, 110(3), 533-538. Germann, L. and J. Braccio (1990). Fine-steering mirror technology supports 10 nanoradian systems. Optical Engineering, 29( 11 ), 1351 - 1359. Gilbart. J. W. and G. C. Winston (1974). Adaptive compensation for tin optical tracking telescope. Automatica, 10(21, 125-31. Gitis, N. V. (1986a). Assessing the anti-stick properties of slideway materials. Soviet Engineering Research, 57(3), 21-22. Gitis, N. V. (1986b). Study of anti-stick-slip properties of machine tool guideway materials. Soviet J. of Friction and Wear, 7(5), 72-76. Godfrey, D. (19671. Vibration reduces metal to metal contact and causes an apparent reduction in friction. AS L E Transactions, 10(2), 183-192. Gogoussis, A. and M. Donath (1988). Modeling robots: a real time method for solving the forward dynamics problem incorporating friction. Proc. of the US--Japan Symposium on Flexible Automation, ASME, Minneapolis, MN. Gogoussis, A. and M. Donath (1990). A method for the real time solution of the forward dynamics problem for robots incorporating friction. Trans. ASME, J. Dyn. Sys., Meas. and Control, 112(4), 630-639. Goyal, S., A. Ruina and J. Papadopouls (1991). Planar sliding with dry friction, part 1. Limit surface and moment function. Wear, 143(2), 3117-330. Gu, J. , J. Rice, A. Ruina and S. Tse (1984). Slip motion and stability of a single degree of freedom elastic system with rate and state dependent friction. J. Mech. Phys. Solids, 32(3), 167-196. Haessig, D. A. and B. Friedland (1991). On the modeling and simulation of friction. J. of Dynamic Systems, Measurement and Control, 113(31, 354-362. Halting, J. (1975). Principles of Tribology, Macmillan, London. Hamrock, B. J. (1986). Lubrication of machine elements. In M. Kutz (Ed. ), Mechanical Engineers" Handbook, John Wiley and Sons, NY. Harmer, J. D. (1952). The jerking motion caused by static friction in position control systems. Automatic and Manual Control, Academic Press, NY. Hansson, A. , P. Gruber and J. Todtli (1993). Fuzzy anti-reset windup for PID controllers. Proc. World Congress on Automatic Control, IFAC, Sidney, Australia. Harnoy, A. and B. Friedland (1994). Dynamic friction model of lubricated surfaces for precise motion control. Tribology Transactions, in press. Harris, T. A. (1984). Rolling Bearing Analysis, 2nd ed., John Wiley and Sons, NY. Hashimoto, M. (1989). Robot motion control based on joint torque sensing. Proc. Inter. Conf. on Robotics and Automation, I EEE, Scottsdale, pp. 256-261. Hashimoto, M., K. Koreyeda, T. Shimono, H. Tanaka, Y. Kiyosawa and H. Hirabayashi (19921. Experimental study on torque control using harmonic drive built-in torque sensors. Proc. Inter. Conf. on Robotics and Automation, I EEE. Nice, pp. 2026-2031. Heck, B. S. and A. A. Ferri (19911. Model reduction of a Coulomb friction damped system using singular perturba- tion theory. Proc. 1991 American Control Conference, AACC, Boston, MA, pp. 539-541. Held, V. and C. Maron (1988). Estimation of friction characteristics, inertial and coupling coefficients in robotic joints based on current and speed measurements. Robot Control 1988 ( SYROCO '88), Selected Papers f rom the 2rid I FAC Symposium, Karlsruhe, Germany, pp. 207-212. Henrichfreise, H. (1985). Fast elastic robots. Control of an elastic robot axis accounting for nonlinear drive properties. A survey of friction and controls 1135 Proc. l l t h IMACS World Congress, Oslo, I MACS, Oslo, Norway, pp. 23-26. Henrichfreise, H. (1992). Observed-based Coul omb friction torque compensation for a position control system. Proc, PCIM'92, PCIM, Nurnberg, Germany. Hersey, M. D. (1914). The laws of lubrication of horizontal journal bearings. J. Wash. Acad. Sci., 4, 542-552. Hersey, M. D. (1966). Theory and Research in Lubrication, John Wiley and Sons, NY. Hertz, H. (1881). On the contact of elastic solids. J. Reine und Anges. Math., 92, 156-171. Hess, D. P. and A. Soom (1990). Friction at a lubricated line contact operating at oscillating sliding velocities. J. of Tribology, 112(1), 147-152. Hess, D. P. and A. Soom (1991a). Normal vibrations and friction under harmonic loads: part I - - Her t zi an contacts. J. of Tribology, 113(1), 80-86. Hess, D. P. and A. Soom (1991b). Normal vibrations and friction under harmonic loads: part l I - - r ough planar contacts. J. of Tribology , U3(1), 87-92. Himmell, L. C. (1985). Examination of adaptive control- based approaches to friction compensation in spacecraft gimbal control systems. Proc. 24th Conf. on Decision and Control, Ft. Lauderdale, FL, pp. 642-646. Hirel, P. (1990). Adaptive optics, dither optimization method in an adaptive optic model. Proc. SPIE Inter, Soc. Opt. Eng., SPIE, Hague, pp. 22-32. Hoj j at , Y. and T. Higuchi (1991). Application of electromagnetic impulsive force to precise positioning. Int. J. Japan Soc. Precision Engineering, 25(1), 39-44. Horikawa, O. , K. Yasuhara, H. Osada and A. Shimokohbe (1991). Dynamic stiffness control of active air bearing. Int. J. Japan Soc. Precision Engineering, 25(1), 45-50. Horowitz, I., S. OIdak and A. Shapiro (1991). Extensions of dithered feedback systems. Inter. J. of Control, 54(1), 83- ! 09. Howard, S. W. and V. Kumar (1993). A minimum principle for the dynamic analysis of systems with frictional contacts. Proc. 1993 Inter. Conf. on Robotics and Automation, I EEE, Atlanta, pp. 437-442. Howe, R. D. , I. Kao and M. R. Cutkosky (1988). The sliding of robot fingers under combined torsion and shear loading. Proc. 1988 Inter. Conf. on Robotics and Automation, I EEE, Philadelphia, pp. 103-105. Hsu, P., M. Bodson, S. Sastry and B. Paden (1987). Adaptive identification and control for manipulators without using joint accelerations. Proc. Inter. Conf. on Robotics and Automation, I EEE, Raleigh, pp. 1201-1215. lbrahim, R. A. (1992a). Friction-induced vibration, chatter, squeal, and chaos: part l --mechani cs of friction. In R. A. Ibrahim and A. Soom (Eds), Friction-Induced Vibration, Chatter, Squeal, and Chaos, Proc. ASME Winter Annual Meeting, Anaheim, DE-Vol. 49, ASME, NY, pp. 107-122. lbrahim, R. A. (1992b). Friction-induced vibration, chatter, squeal, and chaos: part II---dynamics and modeling. In R. A. lbrahim and A. Soom (Eds), Friction-Induced Vibration, Chatter, Squeal, and Chaos, Proc. ASME Winter Annual Meeting, Anaheim, DE-Vol. 49, ASME, NY, pp. 123-138. Johannes, V. I. , M. A. Green and C. A. Brockley (1973). The role of rate of application of the tangential force in determining the static friction coefficient. Wear, 24(5), 384-385. Johnson, C. T. and R. D. Lorenz (1991). Experimental identification of friction and its compensation in precise, position controlled mechanisms. Proc. of the Indus. App. Soc. Annual Meeting, Dearborn, Michigan, pp. 1400- 1406. Johnson, K. L. (1962). Tangential tractions and microslip. Rolling Contact Phenomena, Elsevier, Amst erdam, pp. 6-28. Johnson, K. L. (1987). Contact Mechanics. Cambridge University Press, Cambridge. Kao, I. and Cutkosky, M. R. (1992). Dextrous Manipulation with compliance and sliding. Int. J. Robotics Research, 11(1), 20-40. Karlen, J. P., J. M. Thompson, H. I. Void, J. D. Farreli and P. H. Eismann (1990). A dual-arm dexterous manipulator system with anthropomorphic kinematics. Proc. 1990 Inter. Conf. on Robotics and Automation, I EEE, Cincinnati, pp. 368-373. Karnopp, D. (1985). Computer simulation of stick-slip friction in mechanical dynamic systems. ASME J. of Dynamic Systems, Measurement and Control, 107(1), 100-103. Kato, S. and T. Matsubayashi (1970). On the dynamic behavior of machine tool slideway. Bulletin of the Japanese Soc. of Mechanical Engineers, 13(55), 170-198. Kato, S., N. Sato and T. Matsubayashi (1972). Some considerations of characteristics of static friction of machine tool slideway. J. of Lubrication Technology, 94(3), 234-247. Kato, S., K. Yamaguchi and T. Matsubayashi (1974). Stick slip motion of machine tool slideway. ASME J. of Engineering f or Industry, 96(2), 557-566. Khalil, H. (1992). Nonlinear Systems, Macmillan, NY. Khitrik, V. E. and V. A. Shmakov (1987). Static and dynamic characteristics of friction pairs. Soviet J. of Friction and Wear, 8(5), 112-115. Ko, P. L. and C. A. Brockley (1970). The measurement of friction and friction-induced vibration. J. of Lubrication Technology, 92(4), 543-549. Kosuge, K. , H. Takeuchi and K. Furuta (1988). Motion control of a robot arm using joint torque sensors. Proc. 27th CDC, I EEE, Austin, TX. Kragelskii, I. V. (1988). Handbook of Friction Units of Machines. ASME Press, NY. Kubo, T. , G. Anwar and M. Tomizuka (1986). Application of nonlinear friction compensation to robot arm control. Proc. 1986 Inter. Conf. of Robotics and Automation, I EEE, San Francisco, CA, pp. 722-727. Kuc, T. -Y. , K. Nam and J. S. Lee (1991). An iterative learning control of robot manipulators. Trans. on Robotics and Automation, 7(6), 835-842. Lee, S. and S. M. Meerkov (1991). Generalized dither. I EEE Trans. on Information Theory, 37(1), 50-56. Leonard, N. E. and P. S. Krishnaprasad (1992). Adaptive friction compensation for bi-directional low-velocity tracking. 31st Conference on Decision and Control, I EEE, Tucson, pp. 267-273. Linker, M. F. and J. H. Dieterich (1992). Effects of variable normal stress on rock friction: observations and constitu- tive equations. J. of Geophysical Research, 97(B4), 4923-4940. LOtstedt, P. (1981). Coulomb friction in two-dimensional rigid body systems. Zeitschrift fiir Angewandte Mathematik un Mechanik, 61, 605-615. Lubrizol, Inc. (1988). Product Data Sheet, Lubrizol 5346. Ludema, K. C. (1988). Engineering progress and cultural problems in tribology. Lubr&ation Engineering, 44(6), 500-509. Luh, J. Y. S., W. D. Fisher and R. P. C. Paul (1983). Joint torque control by a direct feedback for industrial robots. I EEE Trans. on Automatic Control, AC-28(2), 153-161. MacColi, L. A. (1945). Fundamental Theory of Servomechanisms, Van Nostrand, Princeton, NJ. Maron, J. C. (1989a). Identification and adaptive control of mechanical systems with friction. Adaptive Systems in Control and Signal Processing, Selected Papers from the 3rd I FAC Symposium, Glasgow, U. K. pp. 325-330. Maron, J. C. (1989b). Nonlinear identification and observer based compensation of friction in mechanical systems. In A. Isidori (Ed. ), Nonlinear Control Systems Design, Pergamon Press, Oxford. Martins, J. A. C. , J. T. Oden and F. M. F. Sim6es (1990). A study of static and kinetic friction. Int. J. Engineering Science, 7,,8(1), 29-92. Mason, M. T. and Y. Wang (1988). On the inconsistency of rigid-body frictional planar mechanics. Proc. 1988 I EEE Int. Conf. on Robotics and Automation, I EEE, Philadel- phia, PA, pp. 524-528. Maqueira, B. and M. K. Masten (1993). Adaptive friction compensation for line-of-slight pointing and stabilization. 1136 B. ARMSTRONG-Hr~LOUVRY et al. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1942-1946. Mees, I. A. (1984). Describing functions: ten years on. IMA J. of Applied Mathematics, 32(1-3), 221-233. Merchant, M. E. (19461. Characteristics of typical polar and non-polar lubricant additives under stick-slip conditions. Lubrication Engineering, 2(2), 56-61. Millman, G. (1990). Cincinnati Milacron, Private Communication. Mobil Oil Corporation (1971). Gears and their lubrication. Technical Bulletin 1-92-003. Mobil Oil Corporation (19781. Way l ubri cat i on--machi ne tools. Technical Bulletin 8-93-001. Moore, D. F. (1972). The Friction and Lubrication of Elastomers. Elsevier Scientific, NY, Moore, D. F. (1975). The Friction of Pneumatic Tyres. Elsevier Scientific, NY. Morgowicz, B. (1988). Techniques for real-time simulation of robot manipulators. Ph. D. thesis, Aerospace Engineer- ing, Univ. of Michigan. Morin, A. J. (1833). New friction experiments carried out at Metz in 1831-1833. Proc. of the French Royal Academy of Sciences, 4, 1-128. Mossaheb, S. (1983). Application of a method of averaging to the study of dithers in non-linear systems. Inter. J. of Control, 38(3), 557-576. Mukerj ee, A. and D. H. Ballard (1985). Self-calibration in robot manipulators. Proc. 1985 Inter. Conf. on Robotics and Automation, I EEE, St. Louis, MS, pp. 1(150-1057. Newman, W. S., Y. Zhang, W. J. Anderson and W. Shipitalo (1992). Experimental evaluation of a new traction-drive robotic transmission. Proc. IEEE Int. Conf. Robotics and Automation, I EEE, Nice, France, pp. 652-657. Newton, I. (1687). Philosophiae Naturales Principia Mathe- matica. S. Pepys, Reg. Soc. Praeses, 5 Julii, 1686. O' Connor , L. (19921. Active magnetic bearings give systems a lift. Mechanical Engineering, 114(7), 52-57. Oden, J. T. and J. A. C. Martins (1985). Models and computational methods for dynamic friction phenomena. Comput. Meth. Appl. Mech. Eng., 52(1-3), 527-634. Okubo, P. (1986). Experimental and numerical model studies of frictional instability seismic sources. Ph.D. Dissertation, Depart ment of Earth, Atmospheric and Planetary Sciences, MIT. Ostertag, E. , N. Bakri and N. Becker (1989). Functional disturbance observer for simultaneous control and dry friction compensation. Proc. IFAC Advanced Information Processing in Automatic Control, I FAC, Nancy, France, pp. 421-426. Painlevh, P. (18951. Sur les Lois du Frottement de Glissement. Comptes Rendus de l'Acad(mie des Sciences, 121, 112-115. Palmgren, A. (19451. Ball and Roller Bearing Engineering, S. H. Burbank, Philadelphia, PA. Pan, P. and Hamrock, B. J. (1989). Simple formulas for performance parameters used in elastohydrodynamically lubricated line contacts. J. of Tribology, 111(2), 246-251. Papay, A. G. (1974). EP gear oils today and tomorrow. Lubrication Engineering, 30(9), 445-454. Papay, A. G. (1988). Industrial gear oi l s--st at e of the art. Lubrication Engineering, 44(3), 218-229. Papay, A. G. and D. W. Dinsmore (1976). Advances in gear additive technology. Lubrication Engineering, 32(5), 229-234. Pavelescu, D. and A. Tudor (1987). The sliding friction coefficient--its evolution and usefulness. Wear, 120(3), 321-336. Peshkin, M. A. and A. C. Sanderson (19881. Minimization of energy in quasistatic manipulation. Proc. 1988 Inter. Conf. on Robotics and Automation, I EEE, Philadelphia, PA, pp. 421-425. Pfeffer, L., O. Khatib and J. Hake (1986). Joint torque sensory feedback in the control of a PUMA manipulator. Proc. 1986 American Control Conference, AACC, Seattle, WA. Pfeffer, L., O. Khatib and J. Hake (1989). Joint torque sensory feedback in the control of a PUMA manipulator. IEEE Trans. on Robotics and Automation, 5(4), 418-425. Polycarpou, A. and A. Soom (1992). Transitions between sticking and slipping. In R. A. lbrahim and A. Soom (Eds), Friction-Induced Vibration, Chatter, Squeal, and Chaos, Proc. ASME Winter Annual Meeting. Anaheim, DE-Vol. 49, ASME, NY, pp. 139-148. Pope, L. E., L. L. Fehrenbacher and W. O. Winer (1989). New Materials Approaches to Tribology: Theory and Applications, Materials Research Society. Pittsburgh, PA. Popp, K. and P. Stelter (19901. Stick-slip vibrations and chaos. Phil. Trans., R. Soe. London series A, 332(1), 89-105. Rabinowicz, E. (19511. The nature of the static and kinetic coefficients of friction. J. of Applied Physics, 22(11), 1373-1379. Rabinowicz, E. (1956a). Autocorretation analysis of the sliding process. J. of Applied Physics, 27(2), 131 - 135. Rabinowicz, E. (1956b). Stick and slip. Scientific American, 194(5), 109-118. Rabinowicz, E. (1958). The intrinsic variables affecting the stick-slip process. Proc. Physical Society of London, 71(4), 668-675. Rabinowicz, E, (19591. A study of the stick-slip process. In R. Davies (Ed. ), Friction and Wear. Elsevier, NY. Rabinowicz, E. (1965). Friction and Wear of Materials, John Wiley and Sons, NY. Rabinowicz. E. (1978). Friction, especially low friction. Int. Conference on Fundamentals of Tribology, (1978: MIT) edited by N. P. Sub and N. Saka, MIT Press, Cambridge, pp. 351-365. Rabinowicz, E. , B. G. Rightmire, C. E. Tedholm and R. E. Williams (1955). The statistical nature of friction. Trans. of the ASME, 22, 981-984. Rabinowicz, E. and D. Tabor (19511. Metallic transfer between sliding metals: an autoradiographie study. Proc. of the Royal Soc. of London, Series A, 208 (Series A), 455-475. Radcliffe, C. J. and S. C. Southward (1990). A property of stick-slip friction models which promotes limit cycle generation. Proc. 1990 American Control Conference, ACC, San Diego, CA, pp. 1198-1203. Rajan, V. T. , R. Burridge and J. T. Schwartz (1987). Dynamics of a rigid body in frictional contact with rigid walls: motion in two dimensions. Proc. 1987 IEEE Int. Conf on Robotics and Automation, I EEE, New York, pp. 671-677. Rattan, K. S., B. Chiu, V. Feliu and H. B. Brown (1989). Rule-based fuzzy control of a single-link flexible manipulator in the presence of joint friction and load changes. Proc. 1989 American Control Conference, AACC, Pittsburgh, PA, pp. 2749-2750. Rayiko, M. V. and N, F. Dmytrychenko (1988). Some aspects of boundary lubrication in the local contact of friction surfaces. Wear, 126(1), 69-78. Reynolds, O. (1886). On the theory of lubrication and its application to Mr. Beauchamp Tower' s experiments, including an experimental determination of the viscosity of olive oil. Phil. Trans. Royal Soc., 177, 157-234. Rice, J. R. and A. L. Ruina (1983). Stability of steady frictional slipping. J. of Applied Mechanics, 50, 343-349. Richardson, R. S. H. and H. Nolle (19761. Surface friction under time-dependent loads. Wear, 37, 87-101. Ro, P. I. and P. I. Hubbel (19931. Model reference adaptive control of dual-mode micro/macro dynamics of ball screws for nanomet er motion. ASME J. of Dynamic Systems, Measurement and Control, !13 (March), 103-108. Rooney, G. T. and P. Deravi (1982). Coulomb friction in mechanism sliding joints. Mechanism and Machine Theory, 17(3), 207-211. Ruina, A. (1980). Friction laws and instabilities: a quasistatic analysis of some dry frictional behavior. Ph.D. Disserta- tion, Division of Engineering, Brown University. Ruina, A. , Y. Katzman, G. Conrad and F. Horowitz (1986). Some theory and experiments related to frictional behavior of rocks at low normal stress. Unpublished manuscript. A survey o f friction and cont rol s 1137 Sadeghi, F. and P. C. Sui (1989). Compressible elas- tohydrodynamic lubrication of rough surfaces. J. of Tribology, 111(1), 56-62. Salisburty, J. K. , W. T. Townsend, B. S. Eberman and D. M. DiPietro (1988). Preliminary design of a whole-arm manipulation system (WAMS). Proc. of the 1988 Int. Conf. on Robotics and Automation, I EEE, Philadelphia, pp. 254-260. Sampson, J. B., F. Morgan, D. W. Reed and M. Muskat (1943). Friction behavior during the slip portion of the stick-slip process. J. Applied Physics, 14(12), 689-700. Satyendra, K. N. (1956). Describing functions representing the effect of inertia, backlash and Coulomb friction on the stability of an automatic control system. AIEE Transactions, 75(II), 243-249. Sch/ifer, U. and G. Brandenburg (1989). Compensation of Coulomb friction in industrial elastic two-mass systems through model reference adaptive control. Proc. 3rd European Conf. on Power Electronics and Applications, EPE, Aachen, Germany, pp. 1409-1415. Schiifer, U. and G. Brandenburg (1990). Model reference position control of . an elastic two-mass system with backlash and Coulomb friction using different types of observers. Power Electronics and Motion Control, PEMC, Budapest, vol. 3, pp. 797-801, Schiifer, U. and G. Brandenburg (1993). Model reference position control of an elastic two-mass system with compensation of Coulomb friction. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1937-1941. Schallamach, A. (1971). How Does Rubber Slide? Wear, 17(4), 301-312. Schimmels, J. M. and M. A. Peshkin (1993). The space of admittance control laws that guarantees force-assembly with friction. Proc. 1993 Inter. Conf. on Robotics and Automation, I EEE, Atlanta, pp. 443-448. Shen, C. N. (1962). Synthesis of high order nonlinear control systems with ramp input. IRE Trans. on Automatic Control, AC-7(2), 22-37. Shen, C. N. and H. Wang (1964). Nonlinear compensation of a second- and third-order system with dry friction. IEEE Trans. on Applications and Industry, 83(71), 128-136. Silverberg, M. Y. (1957). A note on the describing function of an element with Coulomb, static and viscous friction. AIEE Trans., 75(II), 423-425. Singh, B. R. (1960). Study of critical velocity of stick-slip sliding. J. of Engineering for Industry, 393-398. Singh, S. K. (1990). Modified PID control with stiction: periodic orbits, bifurcation and chaos. Report CAR-90-04, Dept. of Mechanical Engineering, Dartmouth University. Siotine, J.-J. E. (1984). Sliding controller of robot manipulators. Int. J. Control, 40(2), 421-434. Slotine, J.-J. E. and W. Li (1987). On the adaptive control of robot manipulators. Int. J. Robotics Res., 6(3), 49-59. Slotine, J.-J. E. and W. Li (1989). Compomsite adaptive control of robot manipulators. Automatica, 25(4), 509-520. Slotine, J.-J. E. and W. Li (1991). Applied Nonlinear Control. Prentice-Hall, Englewood Cliffs, NJ. Smith, G. T. (1989). Advanced Machining: The Handbook of Cutting Technology. IFS, Springer Verlag, NY. Sommerfeld, A. (1904). Zur Hydrodynamischen Theorie der Schmiermittehreibung. Zeitschrift Fur Mathematic und Physik, 50(1904), 97-155. Soom, A. (1992). Private correspondence. Southward, S. C., C. J. Radcliffe and C. R. MacCluer (1991). Robust nonlinear stick-slip friction compensation. ASME J. of Dynamic Systems, Measurement and Control, 113(4), 639-645. Sroda, P. (1988). Analysis of the shape of the contact geometry during meshing of involute gears. Wear, 121(2), 183-196. Stockum, L., J. Profeta and L. Bailou (1988). Precision stabilization system design to reduce the effects of friction. Proc. SPIE Vol. 887 Acquistion, Tracking and Pointing 11, SPIE, 159-167. Stribeck, R. (1902). Die Wesentlichen Eigenschaften der Gieit- und Rol l enl ager--t he key qualities of sliding and roller bearings. Zeitschrift des Vereines Seutscher Ingenieure, 46(38), 1342-1348, 46(39), 1432-1437. Suh, N. P. and H. C. Sin (1981). The genesis of friction. Wear, 69(1), 91-114. Suzuki, A. and M. Tomizuka (1991). Design and implementation of digital servo controller for high speed machine tools. Proc. 1991 American Control Conference, AACC, Boston, WA, pp. 1246-1251. Thomas, S. (1930). Vibrations damped by solid friction. Philosophical Magazine, 7(9), 329-345. Threlfall, D. C. (1978). The inclusion of Coulomb friction in mechanisms programs with particular reference to DRAM. Mechanism and Machine Theory, 13(4), 475-483. Tolstoi, D. M. (1967). Significance of the normal degree of freedom and natural normal vibrations in contact friction. Wear, 10(3), 199-213. Tomizuka, M., R. Horowitz, G. Anwar and Y. L. Jia (1988). Implementation of adaptive techniques for motion control of robotic manipulators. ASME J. of Dynamic Systems, Measurement and Control, 110(1), 62-69. Tou, J. (1953). Coulomb and static friction in servo- mechanisms. Ph.D. Thesis, Electrical Engineering Dept., Yale University. Tou, J. and P. M. Schultheiss (1953). Static and sliding friction in feedback systems. J. of Applied Physics, 24(9), 1210-1217. Townsend, W. T. (1988). The effect of transmission design on the performance of force-controlled manipulators. Ph.D. Thesis, Mechanical Engineering Dept. , Mas- sachusetts Institute of Technology, MA. Townsend, W. T. and J. K. Salisbury (1987). The effect of Coulomb friction and sticktion on force control. Proc. 1987 Inter. Conf. on Robotics and Automation, I EEE, Raleigh, pp. 883-889. Trinkle, J. C. ( 1989) . A quasi-static analysis of dextrous manipulation with sliding and rolling contacts. Proc. Inter. Conf. on Robotics and Automation, I EEE, Scottsdale, pp. 788-793. Tung, E. D. , Y. Urushisaki and M. Tomizuka (1993). Low velocity friction compensation for machine tool feed drives. Proc. 1993 American Control Conference, AACC, San Francisco, CA, pp. 1932-1936. Tudor, A. and L. C. Bo (1982). The squeeze film under boundary lubrication conditions and its effect on the vertical displacement of sliding bodies. Wear, 80(1), 115-119. Tustin, A. (1947). The effects of backlash and of speed-dependent friction on the stability of closed-cycle control systems, lEE Journal, 94(2A), 143-151. Ut ki n, V. I. (1977). Variable structure systems with sliding mode: a survey. IEEE Trans. on Automatic Control, 22(2), 212-222. Vidyasagar, M. (1991). Nonlinear System Analysis. Prentice- Hall, Englewood Cliffs, NJ. Villanueva-Leal, A. and S. Hinduja (1984). Modeling the characteristics of interface surfaces by the finite element method. Proc. of the lnsm. of Mechanical Engineers, 198C(4), 9-23. Vinogradov, G. V., I. V. Korepova and Y. Y. Podolsky (1967). Steel-to-steel friction over a very wide range of sliding speeds. Wear, 10(5), 338-352. Vischer, D. and O. Khatib (1990a). Design and development of torque-controlled joints. In V. Hayward and O. Khatib (Eds), Experimental Robotics !, pp. 271-286. Springer- Verlag, Heidelberg. Vischer, D. and O. Khatib (1990b). Performance evaluation for force/torque feedback control methodologies. Proc. Romansy ' 90, Cracow, Poland. Wallenborg, A. and K. J. Astr6m (1988). Limit cycle oscillations in high performance robot drives. Proc. lEE International Conference CONTROL 88, l EE, pp. 444-449. Walrath, C. D. (1984). Adaptive bearing friction compensa- tion based on recent knowledge of dynamic friction. Automatica, 20(6), 717-727. 1138 B. ARMSTRON6-HELOUVRV et al. Wang, Y., V. Kumar and J. Abel (1992). Dynamics of rigid bodies undergoing multiple friction contacts. Proc. I EEE Int. Conf. Robotics and Automation, I EEE, Nice, France, pp. 2764-2769. Wellauer, E. J. and G. A. Holloway (1976). Application of EHD oil film theory to industrial gear drives. J. of Engineering f or Industry, 98B( 1 ), 626-634. Widrow, B. and S. D. Stearns (1985). Adaptive Signal Processing. Prentice-Hall, Englewood Cliffs, NJ. Wills, G. J. (1980). Lubrication Fundamentals. Marcel Dekker, NY. Wilson, A. R. (1979). The relative thickness of grease and oil films in rolling bearings. Proc. of the Insm. of Mechanical Engineers, 193(17), 185-192. Wolf, G. J. (1965). Stick-slip and machine tools. Lubrication Engineering, 21(7), 273-275. Woodward, J. L. (1963). Describing functions for nonlinear friction in relay servos. Trans. on Automatic Control, AC-8, 260-262. Wu, C. H. and R. P. Paul (1980). Manipulator compliance based on joint torque control. 19th Conference on Decision and Control, I EEE, Al buquerque, NM, pp. 89-94. Xiaolan, A. and Y. Haiqing (1987). A full numerical solutions for general transient elastohydrodynamic line contacts and its application. Wear, 121(2), 143-159. Yang, S. and M. Tomizuka (1988). Adaptive pulse width control for precise positioning under the influence of stiction and Coulomb friction. AS ME J. od Dynamic Systems, Measurement and Control, 110(3), 221-227. Younkin, G. W. (1991). Modeling machine tool feed servo drives using simulation techniques to predict performance. I EEE Trans. on Industry Applications, 27(2), 268-274. Zames, G. and N. A. Shneydor (1976). Dither in nonlinear systems. I EEE Trans. on Automatic Control, AC-21(5), 660-667. Zames, G. and N. A. Shneydor (1977). Structural stabilization and quenching by dither in nonlinear systems. I EEE Trans. on Automatic Control, AC-22(3), 352-361. Zhu, D. and H. S. Cheng (1988). Effect of surface roughness on point contact EHL. J. of Tribology, 110(1), 32-37. Zmitrowicz, A. (1981). A theoretical model of anisotropic dry friction. Wear, 73(1), 9-39.