See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/325247172

Control of Nonlinear Systems using Terminal Sliding Modes

Article · May 2018

CITATIONS READS

3 421

1 author:

Venkat Shastri
University of San Diego
18 PUBLICATIONS   770 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Venkat Shastri on 17 July 2018.

The user has requested enhancement of the downloaded file.

 1992 ACC/WM1 3
Control of Nonlinear Systems Using Terminal Sliding Modes
S. T. Venkcataraman S. Gulati
JPL/California Institute of Technology
4800 Oak Grove Drive, Pasadena, CA 91109
Abstract
ity properties. A primary concern with sliding mode robust
The primary focus of this paper is development of a new control approaches is the large switching gains required, and
approach to control synthesis for robust robot operations in the consequent chattering around the sliding surface. Control
unstructured environments. To enhance control performance switching, even as a phenomenon, cannot be recommended
with full model information, we introduce the notion of terrni- for several robotic tasks. Chatter elimination with interpo
nao convergence, and develop control laws based upon a new lation manifold designs [2] could potentially be complex in
class of sliding modes, denoted terminal sliders. We demon- unstructured environments.
strate that terminal sliders provide robustness to parametric
uncertainty without having to resort to high frequency con- The paper focuses is on two key conitrol requirements,
trol switching, as in the case of conventional sliders [2]. In control (convergence) performance and control ro-
addition, stability analysis that is conducted to demonstrate bustness. We argue that when perfect model information
terminal slider approach results in improved control perfor- is available, the closed loop convergence must be control-
mance and allows for simple robust design of control param- lable, depending, of course, upon environmental characteris-
eters. Further, improved (guaranteed) precision of terminal tics. For exanple, during astronaut assistance, robotic tasks
sliders is argued for through an analysis of steady state be- must be completed in any specified time interval. The latter
havior. property has been referred to as finite time control systems
in the control literature [5]. The second issue pertains to the
INTRODUCTION development of robust control laws that do not require high
frequency control switches. With this motivationi, a theoret-
Many robotic systems would, in the future, be required ical framework that allows terminal control convergence is
to operate in environments that are highly unstructured with developed, wherein the convergence time is finite and con-
varying dynamical properties, and active. Although a signif- trollable. A terminal sliding mode robust control law is pro-
icant volume of results may be found in literature for such posed to deal with model uncertainties. It is shown that the
systems, many issues pertinent to robust/adaptive control al- proposed method leads to greater guaranteed precision in all
gorithms for large magnitude and higlh frequency parametric control cases discussed in here.
uncertainties remain unresolved.
The primary intent of this paper is robust nonlinear con-
TERMINAL SLIDERS
trol synthesis for robotic operations in unstructured environ- In this section, we develop the concept of terminally slid-
ments. Our approach is based upon a new class of sliding ing surfaces from first principles, and apply them to con-
modes, denoted Terminal Sliders. Terminal sliding mode trol synthesis for nonlinear systems. To enhance convergence
controllers enforce closed loop system convergence to equilb- properties of dynamical systems, the concept of terminal at-
rium in finite time. Improved performance of terminal sliding tractors was introduced in [9]. Since then it has demonstrated
mode controllers also results from the elimination of high fre- considerable success in neural learning. It has the basic form
quency control switching function, employed previously for of a cubic parabola: I
robustness to parametric uncertainties [2]. Improved perfor- X = -X3 (1)
mance results from the dependence of terminal slider stability with an equilibrium point at r = 0. Integrating between
upon the rate of change of uncertainties over the sliding sur-
face, rather than the magnitude of the uncertainty itself. The twii and teuiubrium,
terminal slider approach also yields improved (guaranteed) 3 2
precision. In the latter part of this paper, we apply terminal (twiim- tjni4 = ~x4 (2)
sliders to robot manipulator control and benchmark perfor- This implies that Equation 2 settles into equilibrium in finite
mance with the traditional computed torque control method
[4]. -Insights are also provided for the design of control pa- time! This property has also been applied in [5] for finiite
rameters.
time control of distributed parameter systems. The better
convergence results from increased local stability. A detailed
discussion on terminal attractor may be found in [23].
CONTROL PROBLEM
State-of-the-art control approaches are surveyed in [7]. TERMINAL CONTROL
PD control cannot guarantee system stability during the We now utilize the tools developed in the preceding sec-
above mentioned robotic tasks. PID control, while accomo- tion to develop a new class of slidinig mode controllers. Their
dating model uncertainties, is limited in applicability due to performance is compared with the conventional sliding mode
its sluggish transient bahvior. Conventional compliance and
stiffness control [8] methods are limited by their local stabil- summarized in the previous section. The system i = f(z)+ u
(S) will be used as running example, where z denotes state
'This work was supported in part by t}e NASA Office of Aeronautics and u, the control. error e = (x - Xd), where Xd is the desired
and Exploration Technology, and a JPL Director's Discretionary Fund. trajectory. Consider a control law of the form

891
 where a denotes the configuration variables (joint angles, op-
ut = xd-aSe~e-i (3) erational space coordinates, etc.) and I(x),C(±,x) and
where, a,4, and 3d are defined as before. Substituting G (z) denote the inertia matrix, Coriolis matrix and the grav-
Equation 3 for u into system (S), we derive the closed loop itational vector, respectively. Conventional computed torque
system method utiliz the following control law
r = -H (z) (i + K,,i+ Ke) + C (r,x) i + G (z) (11)
i+a-e Ad
Od
e = O (4)
Substituting this control back into equation 10 yields a dosed
For the above system design a sliding surface s, such that loop system of the form e+ K.e+K,e = 0. Assuming K., K.
and H to be positive definite, the closed loop error system
As' can be made exponentialy stable at i = e = 0. This implies
,i= i Pda ? (5) successful tracking of arbitrary trajectories using computed
and the subscript i denotes initial conditions. Equations 5 torque control schemes. Note that control laws that utilize
and 4 together establish the terminal stability of system (S) H (zg) ,C (:d,zd) and G(rd) have also been suggested in the
under the control law in Equation 3. We denote such a surface past [4]. A terminal slider counterpart would apply a control
s as a Terminal Slder, and control laws similar to Equation 3 law of the form
as Terminal Sider ControL It is obvious that for feedback lin-
earizing control with ful model information, terminal slider r = -0(q) (+aL4et +7( + aet))
control wil outperform its conventional counterpart. Subti-
tuting for e in Equation 3 in terms of e, we have the following +CQ(q)4 +GC(q) (12)
V U= Xd + OfR e2 _l f
_
rd+a~-.--et (6) This would result in A+7ys'i 0, which is terminally stable.
=

When model information is not accurate, the issue of con-

Now for the control signal ut to be bounded, for a bounded e, trol robustness must be considered. Only additive uncertain-
the exponent of e must be positive. This implies tie are considered in the following discussions. Multiplcative
4 >O 1 uncertainties (e.g., in robot inertias) are deferred to future
_l 2 (7) work. Such effects are typically reflected in the system's con-
trol gains. For additional information refer to [2], [4]. We
For the initial condition si to be always zero, the slider pa- analyze the example system (S) and extrapolate its results
rameters will require continuous redign for each trajectory. to control of robot manipulators with uncertainties in the
In conventional sliders, this deficiency is overcome with a high coriolis and gravitational terms. Consider the dynamic slider
gain switch that forces the system from any initial condition control law in equation 8, derived estimate on f (denoted by
to the sliding surface [2). We propose a new approach to the f)
problem that models the behavior between initial condition
and the sliding surface as a dynamical system. Consider the u = Xd-a -e' e-ys'd-f (13)
following control law
When substituted in system (S), the control law in Equation
u = id-aonetie -yst - f (8) 13 yields a non-homogeneous dynamic tenninal slider equa-
tion of the form
Substituting Equation 8 into system (S) yields
A=
i = -7Ys",+A (14)
p = 8+7'Jd = 0 (9)
where A = (f -f). To analyze the stability properties of
For any initial condition si, Equation 9 specifies convergence the dosed loop system in Equation 14, consider a Lyapunov
to the steady state of s in finite time, after which the system function V of the form
will reach e = 0 on the surface s as before. We refer to the
surface p in Equation 9 as a dynamic terminal slider. Note
that a conventional slider counterpart of the form i + 75 = 0 V = A -9 Jd (15)
would also provide the attraction from arbitrary initial condi-
tion. However, the system would approach the sliding surface where %.7 e are as defined previously, liV(A,s)ll = 0 and
only exponentially. In contrast, by modeling the dosed loop A = 7$s, and IIVII > 0 everywhere else. For bounded
system in the form of a dynamic terminal slider, we achieve a, IIVII is bounded since A the uncertainty, ie. bounded.
finite time reaching of the slider surface from any arbitrary Differentiating Equation 15, we get
initial condition and without the application of any high gain
switch.
= [(A7S)2s l]f(a§ 1z-s'I) (16)
ROBOT CONTROL
In this section, we apply terminal sliding modes to robot An important observation that may be made is that if A is a
control of robots with rigid links. In general, the dynamics time invariant constant, the closed loop system is terminally
of the robot may be expressed as stable since = 0, driving V = -cV&" in the preceding
empression c > 0, v, and d are as defined previously. The
H(r) i+C(i,)i+G(z) = r (10) *
stability analysis when $ 0, may be performed as follows:

892
 It is well known that robots are passive systems, and that the The above condition ensures terminally stable robust control
terms I, C and G in Equation 10 are such that the joint ve- of robots with prior knowledge on the bounds uncertainty in
locities and transcendental functions involving joint positions link masses, moments, inertia, ,y and 8.
are linearly separable from link lengths, masses and inertias
[3]. It is reasonable to assume that the model uncertainties The implications of the proposed terminal slider robust
arising from incorrect estimation of link lengths, linik masses control law may be analyzed as follows. Since the stability
and link inertias would have a substantially larger effect than properties of the system in Equation 14 depend on 3", the
the uncertainties in the computation of transcendenital func- system is robust to configuration and/or velocity indepen-
tions. Consequently, A may be expressed using dent effects such as Coulomb friction. Secondly, the stability
of the system depends upon the term AAf43y (denoted hence-
A = ( Q-i()
xY(, ) (17) forth by 9) in Equation 19, rather than the magnitude of
A itself, as in conventional sliders. This implies an accom-
where, Y is a matrix of joint velocities and transcendental modation of the controller to the first order dynamics of the
functionis of joint positionls Al is denoted a vector of link uncertainty itself. Further, as the system converges its ability
masses, moments and inertias, while Al is the estimated value to accommodate 9 increases. This is due to the fact that the
of M. Let A31. denote A! - MT . The term ' in Equation stability properties of Equation 19 depends upon the magni-
16 may be expressed as tude of Vsld-1 For values of llsll < 1, sd l increases as
OA OY 13d e l- d n' )
(18)
llsll decreases. V
Knowledge of the bounds on and _ can
=y- AM --"e
Ix 0dd+C)
iA3f1 easily be obtained using kinematic information on degrees
of freedom of the robot. Bounds on Am may be evaluated
ay 8al and Y are bounded and the derivatives are known based upon the degree of uncertainty in the estimation of re-
to be continuous. Since AMf in the above empression, repre- spective parameters.
sents the effect of mass, moment and inertia it is bounded. DISCUSSION
Substituting Equation 18 inito 16, we get
In this paper, we introduced a radically different approach
v 22-1T aay 76-lz,,.AM- I to nonlinear control synthesis, based upon a class of sliding
V~= Rtnj
I J
iAmir-
\U5~a
- -5'd1
6d
(19) modes denoted dynaniic terminal sliders. Our analysis
demonstrated that the proposed slider formalisms yield
where - (A7S) . - n e d + y. Note that superior performance with both complete as well as partial
model information.
Equation 18 implied that as the error reduces, MY- decreased
since the term a is multiplied by a decaying error term.
References
When AM-9Y_ s 6d < 0 (in Equation 19). The closed
[1] Arimoto, S., Miyazaki, F., "Stability and robustness of
loop system becomes terminally stable, since v cVd PID feedback control of robot manipulators of sensor
where c > 0, rM and 'd are as defined previously. When capability", Proc. International Symposium on Robotics
(w_ s 1 0, the system is neutrally stable and Research, MIT Press, Cambridge, MA, 1983.
remains within bounded error for bounded 9A. [2] Asada, H. and Slotine, J.E., Robot analysis and control,
John Wiley and Sons, New York, NY, 1986.
We now discuss the design of parameters a4,,,nl3d,6,7n
and Yd for terminally stable, closed loop system response. [3] Atkeson, C.G., An, C.H., Ifollerbach, J.M., "Estima-
Consider the expression tion of inertial parameters of robot loads and links",
Proc. International Symposium on Robotics Research,
DY Pd DY
09zaDY lA,~
i-k
=
D9i
-- Od + (20) MIT Press, Cambridge, MA, 1986.
[4] Bayard, D.S., Wen J.T., "New class of control for robotic
The first step is to choose the terminal exponent parameters manipulators, Part 2. Adaptive Case", International
ih and Aid. Subsequently, a choice of the terminal slider gain, Journal of Control, Vol. 47, No. 5, 1988, pp 1361 - 1385.
a, may be made using:
Pd i-^
[5] Jayasuriya, S., Dia, A.R., "Performance enhancement
a > -e Od (21) of distributed parameter systems by a class of nonlinear
controls, Proceedings Conference on Decision and Con-
Note that the norm of the error has not been used in Equation trol, LA, CA, Dec 1987, pp. 2125-2126.
21 since the exponent term removes the sign anyway. Also, a
strictly greater than has been used instead of > since ; must [6] Venkataraman, S.T., Gulati, S., "Termrinal Slider Con-
be bounded for all e. The second step is to choose appropriate trol of Nonlinear Systems", (to appear) ASAME Journal
values need to be chosen for fi and bd. For terminal stabifity, of Dynamic Systems, Afeasurement and Control, 1992
[7] Whitney, D., 'Historical perspective and state of the art
7 > [II Ii I(a X + 1)1]iij 4s (22) in robot force control', Proceedings, IEEE Conference
on Robotics and Automation, San Francisco, CA, April
For robust control, the choice of y must be made using the 1985, pp. 269-274.
bounds on values of various parameters. Say,
[8] Zak, M., "Terminal attractors for content addressable
r [l^3fi
>
7 >1IAAtIItaxII O
(DY
+ Dy )
7)II 1 d1-J'd memory in neural networks", Physics Letters, Vol. 133,
(23) 1988, pp 218 - 222.

893

View publication stats