Professional Documents
Culture Documents
JEAN-JACQUES E. SLOTINE
To cite this article: JEAN-JACQUES E. SLOTINE (1984) Sliding controller design for non-linear
systems, International Journal of Control, 40:2, 421-434, DOI: 10.1080/00207178408933284
New result.s are presented OIl the sliding control methodology introduced by Siotine
and Sastry (l983) to achieve accurate tracking for a class of non-linear time-varying
rnultivariable systems in the presence of dist.urbences and parameter variations. An
explicit trade-off is obtained between tracking precision and robustness to modelling
uncertainty: tracking accuracy is set according to the extent of parametric un-
certainty and the frequency range of unmodelled dynamics. The trade-off is further
refined to account for time-dependence of model uncertainty.
1. Introduction
New results are presented on the sliding control methodology introduced
by Slotine and Sastry (198:3) to achieve accurate tracking for a class of non-
linear time-varying systems in the presence of disturbances and parameter
variations.
The classical 'sliding mode control' methodology has been developed
essentially in the literature from the Soviet Union (see, for example, Utkin
(1978) and references contained therein), where it has been used to stabilize a
class of non-linear systems. However, as discussed by Slotine and Sastry
(198:3), this methodology has several important drawbacks, particularly high
control authority and control chattering. Chattering is generally undesirable
in practice, since it involves extremely high control activity, and further may
excite high-frequency dynamics neglected in the course of modelling. Siotine
and Sastry (198:3) showed how to transform the classical sliding mode control
schemes and achieve accurate tracking control for a large class of non-linear
time-varying systems, while maintaining reasonable control activity. In
particular, the methodology involved a trade-off between tracking precision
and robustness to unmodelled high-frequency dynamics. Here we quantify
the precise nature of this trade-off, and complete the methodology of Slotine
and Sastry (198:3) by showing how to achieve optimal tracking performance
given bandwidth requirements and (perhaps time-varying) modelling un-
certainties.
Section 2 summarizes the methodology of Slotine and Sastry (198:3) and
illustrates, with an example, the simplicity of the preliminary design. Section 3
shows how chattering is indeed eliminated, and quantifies the corresponding
trade-off between robustness and tracking precision. The analysis of §:3 is
further refined in § 4, leading to significant improvements in tracking perfor-
mance at the price of modest additional complexity. Section 5 presents
concluding remarks.
where X = [x, d; ... , x(n-llJ'!' is the state and u is the control input, In (I), the
function 1 (in general non-linear) is not exactly known, but the extent of the
imprecision 111 on 1 is bounded by a known continuous function of X and t.
Similarly, control gain b(X; t) is not exactly known, but is of constant sign
and is bounded by known continuous functions of X and t. Disturbance
d(t) is unknown but bounded by a known continuous function of X and t. The
control problem is to get the state X to track a specific state X d = [x d, Xd, ... ,
~;d (,,-I )JI' in the presence of model imprecision 'on I(X; t) and b(X; t), and of
disturbances d(t). For this to be achievable using a finite control u, we must
assume
(2)
where
S(t): s(X; t) = 0
where
s(X; t) (:t + .\
= r x, .\> 0 (3)
Given the initial condition (2), the problem of tracking X = X,\ is equivalent
to that of remaining on the surface S(t) for all t > O. Now a sufficient condition
for such positive invariance of S(t) is to choose the control law u of (I) such that
outside'S(t) (sliding condition)
Ii
- S2(X; t).,; -'1ls(X; t)1 (4)
dt
where '1 is a positive constant (Slotine and Sastry 1983). Given the bounds of
the uncertainties on I(X ; t) and b(X; t), and of the disturbances d(t), obtaining
such a control law is quite straightforward, as we shall illustrate in the example
below (numerous examples can also be found in Slotine (1983)). Moreover,
satisfying eqn. (4) guarantees that if (2) is not exactly verified, i.e. if X II~o is
actually off X d I1-0' the surface S(t) will none the less be reached in a finite time
(inferior to 2s(X(0); 0)/'1), while (3) then guarantees that X.-O as t-s o».
We now illustrate the choice of control law u to satisfy (4), with a simple
multivariable example.
Sliding controller design for non-linear systems 423
Example 2.1
Consider the coupled multi-input system
(Jl = a(t)023 + 20,2 + '11, + «. (t) (5)
{
°
(J2=f2(Ov 0" 02' 2 ; t)+u 2+d2(t) (6)
where
"'~ la(t)-d(t)j, y>v+ih 1-(Jd,+d,(t)1
and thus the control law satisfies the sliding condition (4) for surface S,(t)
d -
dt (S,)2:;;; - 2[y - (v + dll Is,1
Letting
424 J.-J. E. Slotine
where X; = [x;, x;, , x/nj-')]T; we again assume that the imprecisions of the
I;, b, and d; are bounded by known continuous functions of the X; and of time.
Dynamics of the form (11) describe a large number of non-linear systems
encountered in practice, including a vast class of mechanical systems. More-
over, Hunt et al. (1983) have shown that a wider class of controllable non-linear
systems could be put in the form (II) by using appropriate ' global' trans-
formations; being able to deal explicitly with imprecision of the system model
allows for numerical conditioning problems, that may affect such transforma-
tions, to be easily accounted for.
As remarked in the example, control laws which satisfy the sliding condition
(4) are discontinuous across the surface S(t), thus leading to control chattering.
Chattering is, in general, highly undesirable in practice, since it involves
extremely high control activity, and may excite high-frequency dynamics
neglected in the course of modelling. Slotine and Sastry (1983) suggested
a solution to this problem by smoothing out the control discontinuity in a thin
boundary layer neighbouring the switching surface (Fig. I)
(12)
where. is the boundary layer width. This is achieved by choosing control law u
outside B(t) as before (i.e. satisfying the sliding condition (4), which guarantees
boundary layer attractiveness, and hence positive invariance), and then
interpolating u inside B(t)-for instance, replacing in the expression of u the
term sgn 8 by [8/(An - ' . ) ] , inside B(t), as illustrated in Fig. 2. As shown by
Slotine and Sastry (1983), this leads to tracking within a guaranteed precision e,
while allowing avoidance of control chattering. We now quantify the trade-off
thus achieved between tracking precision and robustness to unmodelled high-
frequency dynamics.
Remark
We have assumed full state feedback in the preceding development. Effects
of measurement uncertainty and state estimation are discussed by Slotine
Sliding controller design [or non-linear systems 425
5=0
satisfies sliding condition (4)t. Let us now smooth out the control discon-
tinuity inside a boundary layer B(t) of width E, defined by (12). Control law U
then becomes
U= -J(X; t) - "f (n P I)
p_1
,\px(n-p)-k(X ; t) sat [~J
"n-IE
(19)
t (n) = n!
m m! (n-m)!
for m c n,
Sliding controller desiqn. for non-linear 8Y8tern8 427
since !:>.f(X; t) and k(X; t) are continuous in X. We see from (21) that the
variable 8 (which is a measure of the algebraic distance to the surface S(t)) is the
output of a stable first-order filter whose dynamics only depend on the desired
state Xd(t) and perhaps explicitly on time, and whose inputs are (to first order)
, perturbations': disturbance d(t), uncertainty !:if(X d ; t) on the dynamics,
and nth order derivative xd(n) of the desired trajectory.
Equation (21) shows that chattering is indeed eliminated, as long as un-
modelJed high-frequency dynamics are not excited. The dynamic structure of
the closed-loop system is summarized in Fig. 3; perturbations are lowpass
filtered according to (21) to give 8, which in turn provides tracking error x by
further filtering, according to definition (3); control u is a function of 8, X and
X d as specified by eqn. (19). Now, since A is the break frequency of filter (3),
it has to be chosen to be ' small' with respect to frequencies of un modelled
dynamics (such as unmodelled structural modes, neglected time delays, etc).
x
-xdn)(t) J----'T--j 5 OYNAMICS )----r-T---1 DEFINITION
dl t} +O(E)
(l st ORDER) OF 5
Assume now that F(X d ; t), D(X d ; t) and v(t) can be a priori upper bounded,
so that k(X d ; t) can be upper bounded, say by k m a x • By abuse, kmax/(An-l.)
may be thought of as the' break frequency' of (21); as A, it must also be
chosen to be 'small' with respect to unmodelled high-frequency dynamics.
Thus, if A is set to be the largest acceptable break frequency of (3), we must
have
Erumple 3.1
Consider again the system (8) of § 2, with (discontinuous) control law u,
defined as in eqn. (10). The corresponding gain k,(X; t) is
t If f3ndn ,,;b(X : f)/b(X; /)"; f3mu" set 13: = (f3lnux!f3min)1/2 and use bncw(X ; /): =
ux , f)!(f3mi"f3n",xl as the new estimate of us , I).
fJ can be timc-depcndcnt.
l /2
Note that fJmin. fJmux and
Sliding controller design {or non-linear systems 429
Remark
The effect of data sampling can also be interpreted as part of unmodelled
, high-frequency dynamics'. It is shown by Slotine (1983) that the corres-
ponding upper bound that sampling rate "sampling imposes on I. is
Let us now summarize our sliding control methodology for the class of
coupled non-linear systems considered in (11). For each j (j = 1• . .. ,111.)
(i) Choose bandwidth I. j according to the frequency range of unmodelled
dynamics.
(ii) Define a set of 111. time-varying sliding surfaces Sj(t)
Sj(t) : = {X j: Sj(X j; t) = O}
where Sj depends only on tracking error xj
Consider again the dynamics (21) of the variable s inside the boundary layer
B(t)
instead of (4)-the additional term ,\n-1ilsl in (26) reflects the fact that the
boundary layer attractiveness condition is more stringent during boundary
layer contraction (i < 0), and less stringent during boundary layer expansion
(ti > 0).
To satisfy eqn. (26), we simply replace control gain k(X; t) in (17) by
k(X; t) - ,\n-1i (note that ti is well known since .(t) is defined deterministically).
Equation (21) is then replaced by
. , _ k(X d ; t)
.+I\E- ,\n-1 ' Vt~O (28)
k(X<I(O); 0)
.(0) = An (29)
Sliding controller design for non-linear systems 431
so that the term ,\n-l. does not introduce discontinuity in the control law at
t = 0. With such a choice in initial conditions, (25) is then indeed achieved,
thus leading to improved tracking precision
and by replacing. by .(t) in eqn. (19), with .(t) defined by (28) and (29).
We now extend the previous results to the case that control gain b(X; t) is
estimated by D(X; t), with a gain margin f3
1 D(X; t)
.:«; ,;:. f3
f3" b(X; t)"
In this case, simply subtracting ,\n-l. from the control gain k(X; t) would not
necessarily satisfy attraction condition (26), since it would only guarantee
However, we remark that for all f3' such that I '" f3' '" f32
I d I d ,\n-I·lsl
2 «rs2+>:n lsl- ,\n-I'lsl "<::.-
=> -- nl s l- - =-'--'-
Ij
.>-0
~ 2a s2+
- : >: f3'
I d I d
.",0 =>"2 dt s2+ TJl sl-,\n-I·l s l "'"2 at s2+ TJ IS I- ,\n-l.jslf3'
where
.",o=>k(X; t)=k(X; t)--f3-
\
so that to satisfy (26) we can simply use gain ia , t) in place of k(X ; t),
"n-I£'
that is
J.-J. E . Slotine
'\"< ,\< I
t) ~ 7f => i+ f32 = f3,\"-1 k(X d ; t) (32)
f3
«0) = ,\n k(Xd(O), 0) (33)
As expected, when f3= I the dynamic balance conditions (3\) and (32)
reduce to (28), while (30) and (34) then give identical expressions for the control
gain. Also, eqns. (3\ )-(33) again lead to better tracking precision than that
provided by the static balance condition (23).
This particular type of sliding control, which uses time-varying boundary
layer widths to account for time-dependence of parametric uncertainty, will be
referred to as suction control. It is often the case in practice that k(X d ; t)
shows large variations along a desired trajectory. In these instances, the use
of time-varying boundary layer widths, as specified by the dynamic balance
conditions, greatly improve tracking performance while only introducing
modest additional complexity. In robot manipulator control, for instance,
k(X; t) involves centripetal terms 0,2 and Coriolis terms (Oi OJ); the dynamic
balanco conditions then allow to trade-off speed efficiently against tracking
precision, while preserving system robustness to un modelled dynamics. Also,
recalling that k(X; t) reflects uncertainty on system dynamics, k(X'I; t)
may be decreased as the result of a parameter estimation process, for instance,
thc dynamic balance conditions allow us to account for such on-line improve-
mcnt on modelling precision. Our suction control methodology is thus likely
to provide robust' adaptive' schemes, since it guarantees stability and fixes
control system bandwidth, while achieving the best tracking precision given
current modelling uncertainties.
Remark
In the case that the dynamics of k(X d : t) are 'slow' with respect to
break frequcncy ,\1f3 2 of eqn. (32), the dynamic balance conditions (31)-(33)
lead to
f3k(X,I; t)
«t) ::::; ,\n
Further, for some practical instances desired bandwidth itself may vary
with time. In robot manipulator control, for example, structural resonant
frequencies decrease as the load mass at the tip of the arm gets larger (see, for
example, Paul (1981)). The control law, initially tuned not to excite the lowest
expectable mode (i.e. to handle maximum load), can thus exploit on-line load
estimation by increasing control bandwidth, in addition to decreasing k(X d ; t).
Similarly, it is desirable to monitor mechanical compliance when performing
automatic assembly of close-tolerance parts. Changes in desired bandwidth
are also encountered in the control of high-performance aircraft, for instance.
In these cases attraction condition (26) should be replaced by the more general
form
(35)
where e and ;\ are now possibly time-dependent. Let us call ;\0 the constant
value of ;\ that we previously used, i.e. the desired cut-off frequency based on a
uniform lower bound on frequencies of unmodelled dynamics. Since ;\(t);;,
;\0> 0 for all t;;, 0, the definition (3) of s, now written
(
d
de + ;\(t) )n-I x =: s
</>(0) = f3k(Xd(O); 0)
;\(0)
Again, in the case that the dynamics of k(X,,; t) are' slow' with respect to
Ao/{12, we can simply use
tax , t): = k(X; t)
A(t)q,(t) = {1k(X d ; t)
i.e.
An(t)e(t) = {1k(X d ; t)
s. Concluding remarks
A sequence of transformations led us from classical sliding mode control to
our suction control scheme:
(i) By replacing fixed sliding surfaces in the state space by time-varying
sliding surfaces we avoided high-gain' reaching phases' (Slotine and
Sastry 1983) and achieved perfect tracking of desired trajectories for a
large class of multivariable non-linear time-varying systems.
(ii) By substituting smooth transitions across a boundary layer to control
switching at the sliding surface we eliminated chattering and obtained
a trade-off between tracking precision and robustness to unmodelled
dynamics.
(iii) By allowing boundary layer width to be time-varying we refined the
above trade-off to account for time-dependence of parameter uncer-
tainties, thus improving tracking precision while still maintaining
robustness to unmodelled high-frequency dynamics.
(iv) Finally, we monitored the orientation of the boundary layer in the
state space (defined by A) to account for possible time-dependence of
the desired bandwidth (whether due to actual changes in the plant or to
on-line modelling improvements), thus further improving tracking
performance.
The foregoing methodology has been demonstrated in the high-speed
control of two-link manipulators handling variable loads (Slotine 1983, Slotine
and Sastry 1983). It is currently being implemented on a large hydraulic
manipulator (Slotine and Yoerger 1983), and on a fast direct-drive robot.
ACKNOWLEDGMENTS
The author is deeply grateful to Professor W. S. Vander Velde for valuable
advice and suggestions. This research also benefited from stimulating
discussions with Dr. P. K. Chapman, Professor B. K. P. Horn, Professor 1\1.
Pelegrin, Professor S. S. Sastry, Professor W. S. Widnail and Dr. D. Yoerger.
REFERENCES
HUNT, L. R., Su, R., and G., 1983, Differential Geometry and Control Theory
MEYER,
Conj., Birkhauscr. Boston, pp. 268-298.
PAUL, 1{,. P., 1981, Robot Manipulators : Mathematics, Programming, and Control
(Cam bridgc: IIHT Press),
SLOTINE, .J. J., 1983, Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge,
Massnohusetts.
SLOTINE, .J. J., and SASTRY, S. S., 1983, Int. J. Control, 38,465.
SLOTD1E, J. J., and YOERGER, D., 1983, Suction Control 0/ a Large Hydraulic lIfanipu-
laior, IIHT-MMSL Report (to be published).
UTKIN, V. L, 1978, Sliding Mode Control and its Appl-ications to Variable Structure
Systems (Mir: Moscow}.