Pmeeedings of the 42nd IEEE

Conference on Decision and Control

Maui, Hawaii USA, December 2003 WeM08-1
A New Nonlinear Observer for Tire/Road Distributed Contact Friction
CARLOS CANUDAS-DE-WIT, MORTEN LlND PETERSEN A N D ANTON SHlRlAEV
Labratoire d’Automatique de Grenoble, UMR CNRS 5528
ENSIEG-INPG, B.P. 46,38 402, ST. Martin d’H&es, FRANCE
The Maenk Mc-Kinney Moller Institute for Production Technology
University of Southem Denmark, Odense, DENMARK

Abslracl-In this paper we present a new nonlinear observer differ signifcantly from its steady-state value. Experiments,
structure for estimating the tire-road friction force in a vehicle performed with commercial vehicles, have shown that the
by use of only the known angular wheel velocity. The observer tirelroad forces do not necessarily vary along the steady-
developed here is shown feasible for a more complex friction
model accounting for the distributed nature of contact, and state curves are most likely to exhibit hysteresis loops, clearly
a modifed observed is proposed relieving some uncontrolled indicating the dynamic nature of friction.
necessary preconditions for observer convergence, of previous In Canudas-de-Wit and Horowitz [ 3 ] , and in Canudas-
works in Canudas-de-Wit and Horowitz, 131, and in Canudas- de-Wit et a/ [ 2 ] , it has been proposed to use dynamical
de-IVit el al [Z]. punctual tirelroad friction model together with a nonlinear
a ) Keywordr: Tirehad Friction estimation, nonlinear
observers. observer to estimate a parameter in the model that renects
changes in the tirelroad characteristics. Since in general the
linear velocity cannot be computed from existing sensors, the
I. INTRODUCTION proposed observer also estimates the linear wheel (or car)
Literature for tirelroad friction modeling and estimation is velocity using only rotational wheel velocity information.
numerous. lirelroad friction models are usually described by This paper extends and improves the previously authors
steady-state maps of the wheel slip s, and the the normalized cited works in the following two main directions:
friction force, i.e. First, we reformulated the problem using a distributed
p=-=
F Friction force model instead of using the punctual one. Namely, we
F, Normal force explicitly consider the distributed nature of the contact,
Examples of such a maps, are given hy B a k e r et al. [I], and more explicitly we use the approximated mean
and Pacejka and Sharp [9].The model proposed by the latter, lumped model recently derived in Canudas-de-Wit et
known as the “magic formula”, is derived heuristically from a/ [5]that suitably approximates the PDE by an ODE.
experimental data to produce a good f t . Secondly, we re-design the observer structure to cope
Estimation methods for /I are often based on this type of with a theoretic fundamental limitation of the previous
steady-state models, and often intent to estimate the “slope” results in 131. At one hand this new structure captures
of this curves in the vicinity of zero slip. Some examples the modifcation due to the new model, and at the other
can be found in Kiencke [71, Gustafsson 161. Ray [ l l ] hand it provides suffcient conditions for asymptotic
estimates fi based on a different approach. Instead of using convergence in term of the observer gains magnitude.
the slip information to derive a characteristic curve, Ray [ I l l These conditions are easily verifed by simply increasing
estimates the forces on the tires with an extended Kalman these gains.
flter.
One diffculty in using steady-state maps, is that they are 11. PROBLEM FORMULATION
not invariant with respect to the vehicle velocity, the road In this paper we consider the simplifed motion dynamics
characteristics, the normal .load, among other factors. An of a quarter-vehicle model. The system is then of the form
attempt to consider the velocity dependence for ABS control
is presented in Liu and Sun [SI.The authors assume the mu = F , JG = -rF iU
tirelroad characteristics to be known. Nevertheless, such ide-
where m is 114 of the vehicle mass and J, r are the inertia
alized steady-state conditions are not reached except during
and radius of the wheel, respectively. U is the linear velocity
the rather particular case of cruising at constant speed. The
of the vehicle, w is the angular velocity of the wheel, U is
development of the friction force at the tirelroad interface
the accelerating (or braking) torque, and F is the tirdroad
is very much a dynamic phenomenon. In other words, the
friction force.
friction force does not reach its steady-state value instan-
bJ 7ire/Road friction mode/.: Dynamic friction models
taneously, hut rather exhibits transient behavior which may
can he adapted to suitably describe the road-tire contact
Corresponding author Emait: [email protected]. inpg. fr friction. A potential advantage of such models is their ability
Visiting Ph.D. student from the Maenk Institute. Emait: to describe closely some of the physical phenomena found in
[email protected]
Assistmt Professor at [he Maenk Institute. Email: ar- roadtire friction (i.e. hysteresis loops, pre-sliding displace-
[email protected] ment, etc.), and to depend on a parameter directly related

0-7803-7924-1/03/$17.00 02003 IEEE 2246

with the observed phenomena, like for instance the change on
the road characteristics (i.e. dry, wet, etc.). Dynamic models
can be formulated as lumped or distributed ones.
Distributed models are described by parlial differential
equation and derived by assuming a contact patch between
the tire and the ground, and hence a distribution of the
friction force along the patch. In Canudas-de-Wit er ai [SI,
an ordinary differential equation (the lumped model) for
the friction force is developed based on the patch boundary
conditions and the normal force distribution along the contact
patch. This lumped model is shown to closely approximate
the distributed friction model. Contrary to common static
frictiodslip maps, this model is able to accurately capture Fig. 1. Steadystate profles computed for the uniform force
the transient behaviour of the friction force observed during distribution (braking caw). These curves show the normalized
transitions between braking and acceleration. It also captures friction p = P(s)/F,, as a function of the slip velocity s. Date
similar steady-state characteristics, as the ones reponed in the used for this plots are: 00 = 181.54 = [l/m], 01 = l.O[s/m],
literature. U* = 0.0018[s/m], 0, = 1.0 = [KgmZ/s],pc = 0.8, p 3 = 1.55.
U* = 6.57[m/s],L = 0.2[m],I( = lO[l/m].
Under certain hypotheses on the force distribution, the
lumped model proposed in [SI has the following simplifed
form:
shapes match reasonably well the experimental data reponed
i ( t ) = ur - B---Z(t)
,JOlU t - r;rlwlz(t) (1) in the literature.
dV?)
d ) The mimation problem.: We consider the one-wheel
+
F ( t ) = (uoz(t) mi(t) + 02%) F, (2) model with the lumped tirdroad friction model, i.e.
with
mC +
F,(OOO; ~ i i ) F"UZV? + (3)
JJ = -rF,(uoz +
uii) - uUw UT + (4)
where, uo is the normalized rubber longitudinal lumped
stiffness, 01 the normalized rubber longitudinal lumped t = ur - ow2+lz(q
S(%)
- (5)
damping, U* the normalized viscous relative damping, pc
the normalized Coulomb friction, pa the normalized static where we have neglected the term 02 in equation (4). a, is
friction,(p, 5 pa, E [0, l]),U. the Striheck relative velocity, the viscous rotational friction, and B is the parameter related
Fn = jt fn(C) dC the total normal force (of a particular to changes in the road conditions. Assume that the lumped
force distribution), up = (rw - U) the relative velocity, and 2 friction parameters with B = 1, has been identifed off-line.
the intemal weighted mean friction state. The parameter B is The problem is to design an on-line observer for adhesion
introduced to capture the changes in the road characteristics. coeffcient 0 and for linear vehicle velocity U, that allows
It can be interpreted as being the coeffcient of road adhesion. the controller to monitor the eventual changes in the road
Finally, the constant K = L f 2 > 0 captures the effect in the conditions, by using only the measure of the rotational wheel
lumped model of the force distribution. It is worth to notice velocity w.
that K is in general not a constant, but that it can be suitably This problem is frst studied in more general manner, and
approximated by a constant under certain hypotheses (see afterwards we will apply the result to a panicular case.
~51).

-
c ) Steadystate properties: The steady-state character-
istic of the lumped model (1)-(2) is formally given as a map 111. GENERAL
OBSERVER DESIGN
from ( v , w ) FSS, i.e.
Consider the following nonlinear system:

Formally speaking, this is a two-dimensional map, hut for

{. X=Ax+B[B4(y,x)+$(y,x)] + R u + E y
B=O , y = c x
(6)

reasons of comparison, it is sometimes illustrative to project

with 0 ; 4 , 4 . y E R, x ; B , R , E , C E R3,a n d A E R3"3We
this map into a one-dimensional one. This is done by fxing
assume that system states and inputs are bounded, and that
v or U , and introducing the sliding velocity s = U,.f v if
the following holds:
U > rw for the braking case, or s = v T / ( r w ) ,if TW > U for
the traction case. This yields a family of curves s c FSS Assumption 1 (A ,I?) is an observable pair.
for different values of U'S (or w's). This is shown at the left
of Figure 1. Variations of FSS with respect to the coeffcient Assumption 2 The 4 and 6 functions satisfy the Lipschitz
of road adhesion B are shown at the right of Figure 1. These property with respect to the variable x with the positive

2247
bounded functions p,(y) 5 L, and p+(y) 5 L*, i.e
5 P d Y ) 11x1 - XOII
I 4 ( Y > X l ) - d(Y,XO)l
5 L,llXl - xoll
I*(Y,Xl) - *(Y,XO)/ 5 P d Y ) 11x1 - XOII (9)
(7)
(9)
in equation (15), w e s :

- FC?]
+sS d + a y +B
x] B9 r--
$-pc
ac\
(16)
, g=ei
5 L* 11x1 - xoll (10)
As an important consequence of this we observe that a Now, let us consider the following Lyapunov function
vector function, Do(y,xo.xl), can he defned as
= $PC +lS2
Y
Evaluation the time derivative along the error equations, and
the use of the properties of in assumption 3 we get:
such that llDoll < p, and

+(Y,xI)- ~ ( Y , x o )=J$(~,xi,xo)[xi -x01 (11)

Note that lim.,,,,D&xl,xo) = 0. This can be done
likewise for the 4 function, defning a corresponding vector
function aL. with respect to i .
Assumption 3 The map G(s) : U
k=[A -KC?]i+Bv
- of the system
(12)
where the last expression is obtained by choosing the adap-
tation law for v(y,i), given by

4Y,~)=Yc[d(Y,i)+ael

c=Fi (13) Using (1 I), the last two terms in V , are bounded as follows:

with [A - KC?] Hunvitz, is strictly passive, i.e..3Q > 0,

P = PT > 0, and K, such that
P[A-KC?]+[A-KC?rP=-Q"
PB=C?
Assumption 4 The 4 function has non-zero limit for any
solution trajectory, (y(t),x(t)). i.e., and, by a similar derivation 26 [q - 861 5 $ lli112Finally,
lim 4(y(t):x(t)) # 0
t-m using -ZQ%
5 -Xm6"(Q) lli1/2.
we have
e ) Observer Srructure and Ermr Dynamics. : In view
of the previously given model and assumptions the following
ObSeNeI structure is proposed

1.
x =Ax + B8 [$+ a [?/ - $11 Letting a = 0, and e E (0: Xmin(Q)) a constant describing
the speed of convergence for i,makes V semi-defnite
+ s [d + o [Y- el] + RU + EY (14) negative with a selected as the following function of y
+ K IY - 91
e=" ~

Here a, 8, and v are design functions, which may eventually In practice we simply use a constant and chosen as
depend only on 2, y. and U. For simplicity in the notation
we will drop the arguments in the above functions, i.e. 4 = a >>
9"L; + L:
O(y,i), etc. ZXmin(Q)
Defne the error variables With the choice in equation (17) we have ensured that V 5
.
i=x-i , e=e-e , ~=~-c=ci -E Ilillz,
and hence we have ensured the convergence of the
observer state to the system state
Subtracting (14) from (6) gives the following error dynam-
ics equations x ( t ) + x(t) for t + 03

It remains now to prove the convergence of the 8 parameter

to 9. To this end we ohserve from the error dynanucs,
equations (16), that all terms are ensured to vanish as t --t w,
except the term 64. As a consequence, this term must also
be zero in the limit, ie. limt,, 94 = 0. Using assumption 4
Using the following relation

e+ - e {$+ ac} = e { 4 - .e} + s {$+ S(t) -

this clearly implies the convergence of 6
9 ( t ) for t - M

2248
 We have thus proved the following theorem: Comparing to the structure in theorem 1, we identify
y = w and U = ur and pose equations ( 2 2 ) in the form
Theorem 1 Consider the system in equations ( 6 ) and rhe of equations (6) by identifying A , B, R , E, and C

[ < -: i]
ji)//owing observer
-L.Q 0
i = A i + B [8$+4 + a ( y ) { 1 + 8 ) { y - c}] A = -_ g],BT=[ r,],RT=[
+ Ru + Ey + K[y - Y ] 7m
{ .i = 76 [ ~ ( y . i +
) ac] , 6 = Ff
r 2 F m [I + A] - G 0
(Y)t P 2 (Y)
E=[ h
0, -o w
with a(y) = 8.""q xI P:
_,_ c,Q,-e, and6 E ( O : L " ( Q ) )
Under the assumptions 1-3 it is ensured that .[]+A1
It remains now to show that the assumptions 1-4 hold for
lim f = x, the system with the Lucre model.
t-m
.f) Linear 00serva6iliiy: The observability matrix for
Assurnpiion 4 on persistence of excitation funher ensure that
the pair (A ,c') can be computed as
lim 0 = 0.
d

1
t-m 0 1 -rF,ol
Ed9 0
IV. APPLICATION TO AVERAGED
LUMPED L U C R E O = [ CF AA2 ] = + [ - m 0
MODEL
The determinant of the observability matrix is
The physical modeling of the wheel leads to the dynamics
given in equations (3) and (4), and with the friction model rF;uo
D e t ( 0 ) = -[F,a,o, - moo]
given by equation (5). J3m2
To apply theorem 1 to the friction model in equation ( 5 ) 0 has full rank, ie. r a n k ( 0 ) = 3, if and only if F,,020, #
we identify the 4 and $ functions as moo, and hence we have linear observability; except for the
very special case of Fnu2u1 = moo.
Q(% 3) __
=u o o U r ' 2, $(W,Z)=I(T/WIZ (18) g) Lipschiti Pmpenies of Q and $ : It is immediately
dUT) clear that the $ function is Lipschitz with the following
To further put the dynamics equations in the form given donunating function and
in theorem 1, we apply the following coordinate transform P ~ ( Y )= nr Ivl , h = KT Idmaz
between (1). x) and (U, U). leaving z unchanged
A tedious but straight forward derivation shows that the b
1) = rmu + Jw , x = Jw + rF,alz (19) function is also Lipschitz, with the constant given by

The inverse transform is

Details are given in the appendix.

hJ Passivio; The map of the system in equation (12)
and equation (13) is given by
Taking derivative of equations (19), we arrive at the
following dynamic equations for the state variables
C ( S )= c
' [IS - A + K d ] - l B
We will fnd the vector KT = [ k , .k2, lis] such that the U 0
rj = rmii + JW , = JW + rF,oli (21) map is strictly passive. This means that K is to he found
such that for all r~ > 0 we have iR(G(p) > 0.
Substituting equations (3). (4), and ( 5 ) into equations (21)
Assuming kz = 0, computation of the map results in
and using equations (20) suitably, one can anive at the
following, more explicit state dynamics equations, where we
have added the dynamics of z for completeness
with kc = y,0 =
- ~ ~ O T F ,+
,Uk ~i h
e, 0- el = k3rF,ol, a2 =

A suffcient con&ion for C ( S )to he SPR is that all

coeffcients be positive, and that a1 > p. These conditions
are met if we chose
ki>O , ks<O
Thus assumption 3 is fulflled.

2249
 iJ Persistence of Excitation: To fulfll assumption 4 it
is necessary for the relative velocity to he different from
zero, 71, f 0. This is obtained by continued braking or
acceleration. Attimes where v, = 0, we must expect poor
convergence of 8.

A. Shnulation Result5
Simulations have been performed with the one-wheel
system and the lumped Lucre model. The friction parameters
used in the simulations are the ones given in Figure 1.
The remaining physical parameters for the wheel and the
car used in the simulation are r = 0.25m, m,, = 5Kg,
J = 0.2344 Kg m2, m = 350 Kg, F,, = 3.500 Kg m2/s2. Here
we have used formulas .I = 0.75 mwrz and F,, = 9.82s m,
and m is chosen approximately at the size of quaner of the
mass of a loaded light car. The observer gains are set to
k ~ = - k ~ = l O O O ,y=200000, and a=20.
The emphasis of the simulation is on identifcation and (a) Road adhesion parameter.
tracking of 0 at low slip velocities; and hence, low relative
velocities, where 0 should have poor convergence. The input

-
torque is kept low enough that skidding hardly occurs. In
fact the slip velocity varies smoothly between 0 and 0.05,
which is quite low.
The simulation has frst an acceleration phase, then a
breaking phase, and fnally another acceleration phase. Plots
of some results are shown in Fig. 2. In Fig. 2(b) is shown
the time-profle of the slip velocity. The estimation of linear
velocity U is shown in Fig. 2(c).
Since the ultimate goal of this work is to he able to on-
line estimate varying road conditions, the simulation was (b) Slip velocity. (c) Linear velocity.
done under variations of the parameter 8, representing the
road variation conditions (see Fig. I(h)). Fig. 214 shows Fig. 2. -Simulation results. Evolution of road adhesion parameter
with a solid curve the value of 0, -which evolves within four R and 8, slip velocity s, and linear velocity U and 6. Estimated
different conditions in OUT simulation. The frst quaner of quantities shown with dashed lines and true quantities with solid
lines.
the simulation corresponds to, dry asphalt conditions. The
second quarter corresponds IO a sudden change from dry to
wet. During the third quarter, there is a smooth variation from ity information. To achieve this goal we have used a lumped
wet to snowy. The last quarter is kept constant at the snowy friction model that suitable approximates the distribute fric-
conditions. tion model. This model also allows us to characterize road
The dashed line in Fig.-2(a) shows the evolution of the condition variations via a single parameter.
estimated road condition, B ( t ) . As we can observe, a good We have introduced a model-base ObSeNer that ensures
parameter tracking is obtained as long as the relative contact asymptotic tracking of road condition, under a mild condition
velocity is different from zero. At t = 7.5s g d t = 12.5s implying a non-vanishing evolution of the slip rate. This
are marked two vertical dotted lines. These are the times at condition is quite natural in this context (it implies that
which v, crosses zero, and it is interesting to OhSeNe the the vehicle should operate away from the ideal pure rolling
lack of convergence in 8 in their neighborhood. At t = 7.5s condition).
it is not very pronounced, since 8 does not chang? in that
region. However, in the region around t = 12.5s. 0 oattens VI. REFERENCES
while 0 changes signifcantly. At t = 15.0s. the relative
velocity has grown signifcantly away from zero again, and [ I ] Bakker, E., Nyborg, L.. and Pacejka, H. (1987). Tyre
Modelling for Use in Vehicle Dynamic Studies. Society of
8 quickly converges to 8. This behaviour is not observed to Automotive Engineers Paper # 870421.
the same extent when simulating with higher input torques, 121 Canudas de Wit, R. Horowitz, and P. Tsiotras, (1999). Model-
where higher relative velocities ensures rapid convergence of based observers for tirelroad contact friction prediction, In
0. " N e w Directions in Nonlinear Obsenrr Design". Nijmeijq
H . nnd 21 Fossen (Eds), Verlag, Lectures Notes in
V. CONCLUSIONS control and Information Science, May 1999.
131 Canudas-de-Wit, C., and R. Horowitz. (1999). Observen for
We have presented a new method to estimate on-line the Tirelroad Contact Friction using only wheel angular velocity
changes in road condition using only wheel rotational veloc- information.. In 3Srh IEE-CDC 1999.

2250
 141 Canudas-de-Wit, C. , and P. Tsiotrds, (1999). Dynamic tire We proceed with
friction models for vehicle traction., In 38rh IEE-CDC 1999,
alru in Vehicle Dynamical Sysirm, 2003.
151 Canudas-de-Wit, Tsiotras P. , Velenis E. , Basset M. and
Cissinger G. (2002). Dynamic Friction Models for Roamire
Longitudinal Interaction, To appear in Vehicle Spiem Dy-
namics
161 Gustahon, F. (1997). Slipbased Tire-mad Friction Estima-
tion. Auromarica, 33(6):1087-1099.
171 Kiencke, U. (1993). Realtime Estimation of Adhesion Char-
acteristic Between Tvres
,~~~ and Road. In Pmceedinm of the I~I

IFAC World Congress, volume 1.

-
181 Liu, Y. and Sun, 1. (1995). Target Slip Tracking Using Gain-
, -
Scheduline for Antilock Brakine Svstems. In The American
~~

Corirml Curference, pages I 178-82, Seattle, Washington.

191 Pacejka. H.B. and Sharp R.S. (1991). Shear Force Devel-
opments by Psneumatic tires in Steady-state conditions: A
review of Modeling Aspects.. Vehicle Sysrems Dywmics. Vol.
20, pp.121-176. The existence of [-g'],,, is evident since, taking g as a
[IO] Pasterkamp, W. R. and Pacejka, H.B. (1997). The Tire as a function of 1 ~ ~ we
1 , have
Sensor to Estimate Friction. Vehicle Sysrems Dynamics, Vol.
29,(1997) pp.409-422. 9'(lurl) = --
@* - l l c
e
-y
1111 Ray, L. R. (1997). Nonlinear Tire Force Estimation and U.
Road Friction Ldentifcation: Simulation and Experiments. Hence
Auromarica, 33(10):I8 19-1 833.
lls - Pc
[-9'l,,,, 5 -
APPENDIX U8

We wish to prove the existence of the L+ constant sat- Summarizing, we have expressed
isfying equation (7) for the LuGre model with $ given by I~(Y,xl)-d(Y.xo)/ < k I Z l -~01+k,1171-vol
equation (18). At frst we consider $ as a function of y. e and
U , and intmduce a positive bounded function a as follows
5 [ h+ k,,]llxll

g o IUr I
$(y, i:U ) = a(y, u ) z = -
9(%)
Notice from equations (20) that U , = r y - U is a function Following this we have shown that the Lipschitz constant
of q and U. since can be set to
L+ = I;, + k,
We introduce the following abbreviations
Qi = $(y:zi,ui) I ai = a(y,u,,i)
si = d u r , i ) , ur,i = r y - ui
1
U; = -qi -Jy
rm
The expression for the desired difference can now be
written
I~(Y,xI)-$(Y,~o)/=~$I -401
= btzl - w o l
SUI121 - 201 + IZOIla1 - 001
5%" 121 - 201 + Mmar la1 - aol
The existence of a maximum of a relies on properties of
the g function. Clearly g is positive and bounded as
0 < llc < g < Ps < 'x
Hence a is bounded as
<a 60 lu~lmaz
k
The existences of lzImaz and come from the
assumption that the system states are hounded.

2251