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

net/publication/224659845

Terminal sliding modes: A new approach to nonlinear control synthesis

Conference Paper · July 1991

DOI: 10.1109/ICAR.1991.240613 · Source: IEEE Xplore

CITATIONS READS
112 169

2 authors:

Venkat Shastri Sandeep Gulati

University of San Diego 46 PUBLICATIONS 847 CITATIONS
21 PUBLICATIONS 832 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Venkat Shastri on 19 April 2022.

The user has requested enhancement of the downloaded file.

 Richter, R., Lefebvre, S., and DeCarlo, R. A., 1982, "Control of a Class of
Nonlinear Systems by Decentralized Control," IEEE Transactions on Auto-
matical Control, Vol. AC-27, pp. 492-494.
Ryan, E. P., 1983, " A Variable Structure Approach to Feedback Regulation
of Uncertain Dynamical Systems," International Journal of Control, Vol. 38,
pp. 1121-1134.
Utkin, V. I., 1978, Sliding Mode and Their Application in Variable Structure
Systems, Moscow, MIR Publishers.
Vidyasagar, M., 1978, Nonlinear Systems Analysis, Prentice-Hall, NJ, pp.
58-58.
Xu, X., Wu, Y., and Huang, W., 1990, "Variable Structure Control Approach
of Decentralized Model—Reference Adaptive Systems," IEE Proceedings Pt.
D, Vol. 137, pp. 302-306.

Fig. 3 Phase-plane plots of the perturbed system

Control of Nonlinear Systems
<a> 2 Swi t o l l i n g variables
Using Terminal Sliding Modes
^ a • • x-

S. T. Venkataraman1 and S. Gulati

B fl.2 B.A 0.6 Q.8 JL
<b> 2

1 ^
"
±-
B —
• _
_ ^
f

^ *
Many robotic systems would, in the future, be reuqired to
* 1
operate in environments that are highly unstructured and ac-
tive, i.e., possessing means of self-actuation. Although a sig-
-2
nificant volume of results exist in model-based, robust and
Tine (5ec>
a e . ' 2 e . ' 4 0.6 a.B I adaptive control literature, general issues pertinent to the per-
Fig. 4 (a) The trajectory of switching variables. (1: s„ 2: s2); (o) the formance of such control systems remain unresolved, e.g.,
trajectories of V(t), (1: V, 2: V, controls start at / = 0) feasibility of implementing high gain switches for control ro-
bustness. It is also pointed out that in certain applications,
control switching can be very detrimental to the overall system.
The primary focus of this paper is development of a new
posed DVSC can not only stabilize the large-scale system approach to control synthesis for robust robot operations in
globally and robustly but also assign the hitting time of the unstructured environments. To enhance control performance
system and the convergence rate of a Lyapunov function formed with full model information, we introduce the notion of ter-
by the local switching vectors. Furthermore, the DVSC is in minal convergence, and develop control laws based upon a
a decentralized form which makes design simple and com- new class of sliding modes, denoted terminal sliders. We dem-
putation efficient. onstrate that terminal sliders provide robustness to parametric
uncertainty without having to resort to high frequency control
Acknowledgments switching, as in the case of conventional sliders [2]. In addition,
stability analysis that is conducted to demonstrate terminal
The authors would like to thank referees for their careful slider approach results in improved control performance and
review and valuable comments. The research sponsored by the allows for simple robust design of control parameters. Further,
National Science Council of Taiwan R.O.C. under contract improved (guaranteed) precision of terminal sliders is argued
NSC 80-0404-E008-05 is also sincerely acknowledged. for through an analysis of steady state behavior.
References
Barmish, B. R., and Leitman, G., 1982, "On Ultimate Boundedness Control
of Uncertain Systems in the Presence of Matching Assumptions," IEEE Trans-
actions on Automatic Control, Vol. AC-27, pp. 153-158. 1 Introduction
DeCarlo, R. A., Zak, S. H., and Matthews, G. P., 1988, "Variable Structure
Control of Nonlinear Multivariable Systems: A Tutorial," Proceedings of IEEE, Many robotic systems would, in the future, be required to
Vol. 76, pp. 212-232. operate in environments that are highly unstructured with vary-
Dorling, C M . , and Zinober, A. S. I., 1986, "Two Approaches to Hyperplane ing dynamical properties, and active, i.e., possessing means
Design in Multivariable Structure Control Systems," International Journal of of self-actuation. The development of control technologies for
Control, Vol. 44, pp. 65-82.
Gutman, S., and Palmor, Z., 1982, "Properties of Min-Max Controllers in unpredictable environments is a critical first step in our efforts
Uncertain Dynamical Systems," SIAM Journal of Control and Optimization, to architect robust autonomous robotic systems. Specifically,
Vol. 20, pp. 850-861. algorithms devised for such a purpose must exhibit (i) ro-
Khurana, H., Ahson, S. I., and Lamba, S. S., 1986, "On Stabilization of bustness to parametric uncertainties in dynamic models, and
Large-Scale Control Systems Theory," IEEE Transactions on Automatical Con-
trol, Vol. AC-31, pp. 176-178. (ii) the ability to adapt rapidly enough (to parametric varia-
Lefebvre, S., Richter, S., and DeCarlo, R. A., 1983, "Decentralized Variable tions) in order to ensure operational performance. Although
Structure Control Design for a Two-Pendulum System," IEEE Transactions on a significant volume of results may be found in literature for
Automatical Control, Vol. AC-28, pp. 1112-1114. the aforementioned twin problems (robust control and adap-
Madani-Esfahani, S. M., Hached, M., and Zak, S. H., 1990, "Ultimate
Boundedness and Asymptotic Stability of Uncertain Dynamical Systems via
Variable Structure Control Approach," Proceedings, 1990, Automatic Control 'Machine Intelligence Applications Group and Neural Computation and Non-
Conference, San Diego, CA, pp. 2229-2234. linear Science Group, respectively, Jet Propulsion Laboratory, California In-
Matthews, G. P., and DeCarlo, R. A., 1988, "Decentralized Tracking for a stitute of Technology, Pasadena, CA 91109.
Class of Interconnected Nonlinear Systems Using Variable Structure Control," Contributed by the Dynamic Systems and Control Division of THE AMERICAN
Automatica, Vol. 24, pp. 187-193. SOCIETY OF MECHANICAL ENGINEERS. Manuscript received by the Dynamic Sys-
Mohammad, J., 1983, Large-Scale Systems: Modeling and Control, North- tems and Control Division January 11, 1991; revised manuscript received July
Holland. 27, 1992. Associate Technical Editor: S. Jayasuriya.

Downloaded From:554 / Vol. 115, SEPTEMBER 1993

https://dynamicsystems.asmedigitalcollection.asme.org Transactions
on 11/28/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use of the ASME
Copyright © 1993 by ASME
 tive control respectively), many issues pertinent to robust/ have demonstrated the ability to stabilize closed-loop behavior
adaptive control algorithms for large magnitude and high fre- in the presence of bounded parametric uncertainties. In this
quency parametric uncertainties remain unresolved. paper, we limit our focus to sliding mode robust control [2],
The primary intent of this paper is robust nonlinear control [17], In addition, recent progress in adaptive control [16] has
synthesis for robotic operations in unstructured environments. enabled on-line accommodation of unknown robot mass and
Our approach is based upon a new class of sliding modes, inertial parameters and payloads.
denoted Terminal Sliders. Terminal sliding mode controlleres There are a number of applications in which state-of-the-
enforce closed loop system convergence to equilibrium in finite art control technologies would fall short of prescribed system
time. Improved performance of terminal sliding mode con- performance. Consider two scenarios and substantial relevance
trollers also results from the elimination of high frequency NASA application scenarios: robot-assisted extravehicular ac-
control switching function, ertiployed previously for robustness tivity and autonomous fresh sample acquisition during pre-
to parametric uncertainties [2]. Improved performance results cursor science [18]. In the former, voice-activated robots would
from the dependence of terminal slider stability upon the rate assist astronauts by fetching and returning tools, helping trans-
of change of uncertainties over the sliding surface, rather than port large and heavy payloads, etc. In the latter, robotic ele-
the magnitude of the uncertainty itself. The terminal slider ments interact with soil and rock media to extract core samples
approach also yields improved (guaranteed) precision. In the for scientific analysis. Note, however, that in both applica-
latter part of this paper, we apply terminal sliders to robot tions, the environment dynamics is partially or completely
manipulator control and benchmark performance with the tra- unknown. In addition, since astronauts are active (capable of
ditional computed torque control method ([2], Arimoto in [6], self-actuation) environments (or rocks could contain unknown
and [19]). Insights are also provided for the design of control crevices, shear planes and cracks) the rate at which the robot
parameters. control law accommodates the uncertainty must be explicitly
The paper is organized as follows. In Section 2, we motivate controllable, it needs to be adjusted according to the rate of
the control problem and argue for the development of a new change in the environment. Table 1 summarizes the environ-
class of control laws. Section 3 briefly describes control syn- mental effects in the two problem areas.
thesis with sliding modes. Following this, we derive from first Further, in both the above applications the actual dynamic
principles, the basic formalism for terminal sliding surfaces characteristics of the environments are highly nonlinear. For
and describe their convergence and stability properties in Sec- example, research in biomechanical systems [21], [22], has
tion 5. To address the latter objective, we introduce the concept indicated that the linear spring, dashpot model used in the Hill
of Terminal Stability, contradistinct it with asymptotic and equation [8] would be inadequate. Models comprising expo-
exponential stability using Lyapunov analysis [13], and derive nential and higher-order polynomial terms have been suggested
terminally convergent sliding modes. Section 6 applies terminal for the visco-elastic properties of the muscles and their con-
sliders to the control of a prototypical nonlinear system based tractile dynamics have been modeled using first-order ordinary
upon the computed torque scheme for feedback linearization differential equation [22]. Further, the strong interdependences
(Arimoto in [6]). Section 7 is devoted primarily to robust robot between a human's neuronal impedance (input impedance of
control. We apply terminal sliders within the framework of the human muscular structure), his mechanical impedance [21],
computed torque trajectory control (Arimoto in [6]) to dem- coupled with the variations caused by fatigue [14] cause the
onstrate superior convergence, robustness, and guaranteed environmental models to be nonautonomous. Similarly, en-
precision [2] properties. Section 8 concludes the paper with vironmental characteristics of rocks and regolith are extremely
some future research directions. complex. For example, the models suggested in [15] for dia-
mond coring suggest a nonlinear dependence of the normal
thrust force on the rock hardness, characteristics of the dia-
2 Control Scenario mond matrix, drill diameters, and the drilling rate. In addition,
We begin by summarizing state-of-the-art control ap- the relationship is nonautonomous due to the effect of tem-
proaches. While PD control schemes (with gravity compen- perature on drill characteristics, diamond wear, and chip re-
sation) have been shown to be globally stable for setpoint moval.
control [1], they cannot guarantee stability in the presence of
parametric uncertainties. PID control with gravity compen- 2.1 Control Problem. State-of-the-art PD control cannot
sation is globally stable for setpoint control, even in the pres- guarantee system stability during the above-mentioned robotic
ence of uncertainty [1], but the resulting closed-loop system tasks. The sluggishness in system response with PID control
is rather sluggish. This could potentially generate large tran- could result in tool damage during sampling tasks, and cause
sient errors, causing tool and robot damage during contact excessive human fatigue during astronaut assistance. The ef-
operations. Compliance and stiffness control schemes [20] fectiveness of conventional compliance and stiffness control
guarantee local stability, and function adequately in quasi- methods during autonomous sampling and astronaut assist-
static contact situations with well characterized stationary en- ance operations would be extremely limited, given the complex
vironments. Impedance control has been applied to a wide nonautonomous nature of environment dynamics. A primary
spectrum of contact applications [10]. The implementations, concern with sliding mode robust control approaches is the
are, however, limited to static environments. A detailed sta- large switching gains required, and the consequent chattering
bility analysis of impedance control is available in [9]. Many around the sliding surface. During autonomous sampling, this
robust control schemes have been proposed and implemented typically causes excessive tool wear, sample degradation, and
for pure motion and compliant motion control of robots, and actuator saturation. Control switching, even as a phenomenon,

Table 1 Environmental effects

Unmodeled Effect on
Application phenomena dynamics
Cracks in rocks Unstructured dynamics
Robotic High frequency variations
coring Holes/vesicles in rocks
Pressure variation in regolith High amplitude variations
Astronaut Unpredictable human movements Unstructured muscular dynamics
assistant Varying muscle stiffness Unpredictable muscular actuation

Downloaded From:Journal of Dynamic Systems, Measurement,

https://dynamicsystems.asmedigitalcollection.asme.org and Control
on 11/28/2018 SEPTEMBER 1993,
Terms of Use: http://www.asme.org/about-asme/terms-of-use Vol. 115 / 555
 cannot be recommended for man-machine systems. Although
chattering can be potentially eliminated by designing of inter-
polation manifolds [2], it is not recommended for the above
applications since their size cannot be determined accurately
apriori.
Our focus is on two key control requirements—control (con-
vergence) performance and control robustness. We argue that
when perfect model information is available, the closed-loop
convergence must be controllable, depending, of course, upon
environmental characteristics. For example, during astronaut
assistance, robotic tasks must be completed in any specificied
time interval. The latter property has been referred to as finite
time control systems in the control literature [11]. The second
issue pertains to the development of robust control laws that
do not require high frequency control switches. With this mo-
tivation, a theoretical framework that allows terminal control
convergence is developed, wherein the convergence time is
finite and controllable. A terminal sliding mode robust control INCREASING a , s.XtiX*'
law is proposed to deal with model uncertainties. It is shown
that the proposed method leads to greater guaranteed precision
in all control cases discussed here.
Fig. 1 Attractor behavior

3 Sliding Mode Control tems, the concept of terminal attractors was introduced in [23].
The concept of applying sliding modes to control emerged Since then it has demonstrated considerable success in neural
from earlier work on variable structure systems, notably [7]. learning [4], It has the basic form of a cubic parabola:
In principle, it revolves around the choice of a control law x=~xxn (3)
that forces the closed-loop system behavior to be identical to
a sliding surface. Typically, the closed-loop system dynamics with an equilibrium point at x = 0. Integrating between /initiai
represents controlled system error behavior. As a result, one and t.equilibrium*
can model closed-loop behavior through an appropriate choice
of sliding surfaces. If a sliding surface 5 is chosen such that s ttequilibrium" .,.••initial/
.. , \ _ £
~ *
y?/:
A
;
(4)
= e + Xe = 0, where e is the trajectory error, then exponential
error convergence occurs. Consider, for example, the system This implies that Eq. (4) settles into equilibrium in finite time!
[2] This property has also been applied in [11] for finite time
x=f(x)+u (1) control of distributed parameter systems. The better conver-
gence results from increased local stability. A detailed discus-
where x is the system state and u is the control. A control law sion on terminal attractor may be found in [23].
of the form u = u0 = -f + xd - Xe will result in s = 0. The most general form for a first order terminal attractor
If, s(t0), the initial condition is zero then system converges would be x + X(x) = 0, where X is bounded for bounded
exponentially. When s(t0) ^ 0, a control switch of the form x, Sgn(.Y) = Sgrw and dX/dx — oo as x — 0. Such systems
K Sgn(s), K > 0 may be added to u0 to force the system to are Lipschitzian in any e neighborhood of the equilibrium
converge toward the sliding surface [2]. At s = 0, the switch point, but are non-Lipschitzian at the equilibrium point itself.
is deactivated, and the previous arguments apply. It can be An intuitive argument about the dynamic behavior of such
shown that the system will reach the sliding surface in finite systems is provided in [24]. The actual extensions of GronwalPs
time [2]. conditions for a formal proof of the terminal stability are to
In the presence of parametric uncertainties where only/, an be dealt elsewhere. To analyze such systems using Lyapunov
estimate on/is available, control law takes the following form: methods, we postulate the following. Given a dynamical system
u = u- KSgn{s) of the form x + X(x) = 0, and a Lyapunov function candidate
V(x), where Fis bounded for bounded x, II V(x ^0)11 > 0,
= -f+xd-\e-KSgn(s) (2) II K(* = 0)11 = 0, if
which yields s = / - / - K Sgn(s). By choosing K > 11/ — V+V(V)=0 (5)
/ll, Lyapunov stability and convergence toward the sliding such that V(-) has the terminal attractor property elucidated
surface [2] can be ensured. The closed-loop system does not above, the dynamical system is terminally stable. For example,
actually stay on the sliding surface (upon reaching these), since the system x = -x[/5 would be terminally stable since V =
at 5 = 0, i ^ 0 it chatters in the neighborhood of the sliding x6/5 implies V = -6/5Vl/3. In this work, we exploit terminal
surface [2]. Conventional sliding mode control, therefore, attractors of the form x = ax0"/l3d, where, a > 0 and /3„, (3d
guarantees exponential stability with full model information, = (2/ + 1), where ;' € I, and 0d > /3„. Figure 1 describes
and asymptotic stability in the presence of uncertainties. The terminal attractors with various convergence rates. Curves 1
design of interpolation regions is typically performed off-line and 2 portray attractor behavior Galedon different initial con-
using bounds on uncertainty and the expected system response ditions. Note that (•)' is the conventional counterpart of curve
in the neighborhood of the sliding surface. 1. Curves 3 and 4 portray the effect of attractor gain a.

4 Terminal Sliders
In this section, we develop the concept of terminally sliding 5 Terminal Control
surfaces from first principles, and apply them to control syn- We now utilize the tools developed in the preceding section
thesis for nonlinear systems. The performance thus obtained to develop a new class of sliding mode controllers. Their per-
will be compared to those with the conventional sliding mode formance is compared with the conventional sliding mode sum-
control. To enhance convergence properties of dynamical sys- marized in the previous section. The system in Eq. (1) will be

556 / Vol. 115, SEPTEMBER 1993 Transactions of the ASME

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org on 11/28/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 used as running example. As before, e = (x - xd), where xd
is the desired trajectory. Consider a control law of the form
ft> J3„/pd- 1 DYNAMIC ATTRACTOR
U= Xd~a ~f (6)

where, a, ft,, and j3d are defined as before. Substituting Eq.

(6) for u into Eq. (1), we derive the closed-loop system

e + afV"^-'* -0 (7)
Pd
For the above system design a sliding surface s, such that
si=ei + aef»/<i<i = 0 (8)

and the subscript / denotes initial conditions. Equations (8)

and (7) together establish the terminal stability of system in
Eq. (1) under the control law in Eq. (6). We denote such a
surface 5 as a Terminal Slider, and control laws similar to Eq.
(6) as Terminal Slider Control. It is obvious that for feedback
linearizing control with full model information, terminal slider
control will outperform its conventional counterpart. Substi-
tuting for e in Eq. (6) in terms of e, we have the following

u = xd + a / (9)
ft/
Now for the control signal u to be bounded, for a bounded
e, the exponent of e must be positive. This implies Fig. 2 Dynamic terminal attractor behavior

fti>i (10) In summary, we have proposed a new approach applicable

ft/ 2 to a particular class of feedback linearizing controllers, Eqs.
For the initial condition s-, to be always zero, the slider (6) and (11), and have demonstrated superior closed-loop sys-
parameters will require continuous redesign for each trajec- tem performance for arbitrary initial conditions. In addition,
tory. In conventional sliders, this deficiency is overcome with we have eliminated the need for high gain control switches,
a high gain switch that forces the system from any initial thereby alleviating their detrimental effects during implemen-
condition to the sliding surface [2]. We propose a new approach tation.
to the problem that models the behavior between initial con-
dition and the sliding surface as a dynamical system. Consider
the following control law 6 Robot Control Synthesis
In this section, we apply terminal sliding modes to robot
Jn'Vd,e-ysd"/6d-f (11)
control of robots with rigid links. In general, the dynamics of
u = xd—a
ft/ the robot may be expressed as
Substituting Eq. (11) into Eq. (1) yields H(x)x+C(x, X)X + G(X) = T (13)
p=i+ysi»/dd =0 (12) where x denotes the configuration variables (joint angles, op-
erational space coordinates, etc.) a n d / / ( x ) , C(x,x), and G(x)
For any initial condition sh Eq. (12) specifies convergence to denote the inertia matrix, Coriolis matrix, and the gravitational
the steady state of s in finite time, after which the system will vector, respectively. It is well known that simple PD and PID
reach e = 0 on the surface s as before. We refer to the surface control schemes [1] provide global asymptotic stability for
p in Eq. (12) as a dynamic terminal slider. Note that a con- setpoint control. Wen and Bayard [19] have discussed the ex-
ventional slider counterpart of the form i + ys = 0 would tent of these control schemes. Computed torque schemes may,
also provide the attraction from arbitrary initial condition. on the other hand, be applied for controlling robot trajectories
However, the system would approach the sliding surface only [19], Using these results as a baseline, we examine robust com-
exponentially. In contrast, by modeling the closed loop system puted torque control, especially assuming to uncertainties in
in the form of a dynamic terminal slider, we achieve finite time the gravitational and Coriolis terms. Consideration of uncer-
reaching of the slider surface from any arbitrary initial con- tainties in the inertial parameters will be discussed elsewhere.
dition and without the application of any high gain switch.
To demonstrate Lyapunov stability for dynamic sliders, con- 6.1 Robust Computed Torque Control. We now consider
sider the Lyapunov function V = s2""/"d, where v,„ vd = (2/ the computed torque method for trajectory control. Given the
+ 1), i € I, and vd > vn. Differentiating the Lyapunov function robot dynamics in Eq. (13), the conventional computed torque
yields V = -2 v„/vd ys2"^"6 - 'sVw. Choosing v„ and vd such method utilizes the following control law
that 2 vn/vd (1 - i\Ji\d) = (1 - 5„/5rf), where rj„, ijrf = (2/ +
1), / € I, and i)d > ?/„, terminal stability of dynamic sliders can T=-H(x)(x + K„e+Kpe)+C(x, x)x + G(x) (14)
be easily demonstrated (e.g., vjvd = 1/3, t]„/i)d = 3/5 and Substituting this control back into Eq. (13) yields a closed-
t>„/5d = 11/15). Figure 2 displays the temporal behavior of loop system of the form e + Kve + Kpe = 0. Assuming Kp,
dynamic terminal sliders. Curves labeled 1, 2, 3, 4, 5, and 6 Ku, and H to be positive definite, the closed-loop error system
are dynamic sliders converging to their respective terminal can be made exponentially stable at e = e = 0. This implies
attractors (shown by thick lines). Curves 1, 2, and 3 show the successful tracking of arbitrary trajectories using computed
effect of modulating the attractor gain, y, while curves 4, 5, torque control schemes. Note that control laws that utilize
and 6 illustrate the effects of manipulating the exponent y„/ H{xd), C(xd, xd), and G(xd) have also been suggested in the
yd. Curves 4, 5, and 6 have different initial conditions (but past [19]. A terminal slider counterpart would apply a control
same starting time). law of the form

Journal
Downloaded From: of Dynamic Systems, Measurement,
https://dynamicsystems.asmedigitalcollection.asme.org and Control
on 11/28/2018 SEPTEMBER 1993,
Terms of Use: http://www.asme.org/about-asme/terms-of-use Vol. 115 / 557
 y= [(A_.ys'V6rf)2T!n/irf-1]7
T=-H(q)(q + a^ e^/fs"-' e +y (k + a A 7 ^
dY
$o ei-On/Od + ^ H - T ^ ^ - i
+ C(q,q)q + G(q) (15) ydx aj3„ dxj ' 8d
= [(A - ySd"/Sd) lr, ,r,
n d ~ '] T
This would result in i + yss"/dd = 0, which is terminally
stable.
When model information is not accurate, the issue of control x(AM^-^5V^-')(A-7^/6d) (22)
robustness must be considered. Only additive uncertainties are
considered in the following discussions. Multiplicative uncer- where,
tainties (e.g., in robot inertias).are deferred to future work.
Such effects are typically reflected in the system's control gains. /V0d. dY
For additional information refer to [2], [19], [16]. We analyze ds dx a/3„ dx'
the example system in Eq. (1) and extrapolate its results to
Note that Eq. (21) implied that as the error reduces, d Y/ds
control of robot manipulators with uncertainties in the coriolis
decreased since the term dY/dx is multiplied by a decaying
and gravitational terms. Consider the dynamic slider control
error term. When AM dY/ds - ybn/&d s V ^ 1 < 0 (in Eq.
law in Eq. (11), derived estimate on/(denoted by J)
(22)). The closed loop system becomes terminally stable, since
^gUn^d-'-
-.,.«.,/«,/
V= - cVn"'nd where c > 0, ?j„ and r]d are as defined previously.
u = xd- 'e-y<f„'<-<i-f
a (16) When
fid
When substituted in Eq. (1), the control law in Eq. (16) yields 9A -1^1 ,AA/-'
— = 0,
a nonhomogeneous dynamic terminal slider equation of the ds'
form
the system is neutrally stable and remains within bounded error
i= -ysSn/sd+^ (17)
for bounded dA/ds.
We now discuss the design of parameters a, f3„, f3d, S, y„,
where A = (/" — / ) . To analyze the stability properties of the and yd for terminally stable, closed loop system response. Con-
closed-loop system in Eq. (17), consider a Lyapunov function sider the expression
V of the form
dY dY pd
V={A-ys°> (18) ei-Pn/f>d + — (23)
ds dx a(5„ dx
where n„, -qd are as defined previously, \V(A, s)ll = Oand A The first step is to choose the terminal exponent parameters
= ys5" d, and II V\\ > 0 everywhere else For bounded s, II Kll ft, and /3</. Subsequently, a choice of the terminal slider gain,
is bounded since A the uncertainty, i.e. bounded. DiffBren- a, may be made using:
dating Eq. (18), we get
d_A a>— e /VPd (24)
t s [ ( A - y "•' n0 rd\f ln) ld-
s /s 2r /r
>1 A/5rf_ (19) Pa
dt~ ' Sd Note that the norm of the error has not been used in Eq. (24)
An important observation that may be made is that if A is a since the exponent term removes the sign anyway. Also, a
time invariant constant, the closed-loop system is terminally strictly "greater than" has been used instead of > since \/a
stable since dA/dt = 0, driving V = -cV"'"11 in the preceding must be bounded for all e. The second step is to choose ap-
empression c > 0, ?j„ and ijd are as defined previously. propriate values needed to be chosen for 8„ and 8d. For terminal
The stability analysis when dA/dt ^ 0, may be performed stability,
as follows: It is well known that robots are passive systems, dY
and that the terms / , C, and G in Eq. (13) are such that the ^1 s ^ d '<0 (25)
joint velocities and transcendental functions involving joint ds~ 8d
positions are linearly separable from link lengths, masses and A choice of 7 that satisfied the condition in Eq. (25) is
inertias [3] and [5]. It is reasonable to assume that the model
dY dY &d
cl"V«d
uncertainties arising from incorrect estimation to link lengths, 7> (26)
link masses, and link inertias would have a substantially larger dx dx 8,,'
effect than the uncertainties in the computation of transcen-
dental functions. Consequently, A may be expressed using For robust control, the choice of y must be made using the
T bounds on values of various parameters. Say,
A = (M-M) Y(x, x) (20)
dY dY &d
y> II A* -V«d (27)
where Yis a matrix of joint velocities and transcendental func- dx dx
tions of joint positions M is denoted a vector of link masses,
moments and inertias, while M is the estimated value of M. The above condition ensures terminally stable robust control
Let AM denote M - M'. The term dA/dt in Eq. (19) may be of robots with prior knowledge on the bounds uncertainty in
expressed as link masses, moments, inertia, dY/dx and dY/dx.
dYdx ___
dYdx_ The implications of the proposed terminal slider robust con-
d A
A trol law may be analyzed as follows. Since the stability prop-
a7 = A " dx ds dx ds erties of the system in Eq. (17) depend on dA/dt, the system
is robust to configuration and/or velocity independent effects
(dY (3rf , R /a dY\ such as Coulomb friction. Second, the stability of the system
(21) depends upon the term AM d Y/ds (denoted henceforth by 9A/
\dx ap„ dxj
ds) in Eq. (22), rather than the magnitude of A itself, as in
dY/dx, dY/dx, and Y are bounded and the derivatives are conventional sliders. This implies an accommodation of the
known to be continuous. Since AM in the above empression, controller to the first order dynamics of the uncertainty itself.
represents the effect of mass, moment and inertia it is bounded. Further, as the system converges its ability to accommodate
Substituting Eq. (21) into (19), we get dA/ds increases. This is due to the fact that the stability prop-

558 / Vol. 115, SEPTEMBER 1993 Transactions

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org on 11/28/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use of the AS ME
 erties of Eq. (22) depends upon the magnitude of yb„/dd
s6"/bd-\ For values of \\s\\ < 1, A / 6 " ~ ' increases as lis II de-
creases. Knowledge of the bounds on dY/dx and dY/dx can
easily be obtained using kinematic information on degrees of
freedom of the robot. Bounds on AM may be evaluated based
upon the degree of uncertainty in the estimation of respective
parameters. Note that the equilibrium point for the system in
Eq. (17), is at s = 0 and s = (A/y)s<i/s><. Therefore, the guar-
anteed precision with terminal sliding mode control becomes

(A\S«/5"\

(29)

Terminal slider control outperforms conventional sliders un-

der all of the above-mentioned conditions. Also, guaranteed Fig. 3 Dynamic terminal slider robustness
precision is better with terminal sliders due to the following
reasons. In conventional sliding mode control [2], the inter-
polation region $ is designed such that along the respective sliders to the origin of Fig. 3. Note that
$ + a$ = A (30) uncertainty term A and the term y Sd/8„ /«/6tf ~' have also been
marked along the vertical axis. Constant A's are represented
for some a > 0. This implies that at steady state,
by the lines marked 1 and 2. When A is constant, the closed-
_^£2 loop system is stable from any initial condition. For example,
(31) from an initial condition SAQ (marked by S and A coordinates),
the system will converge to an equilibrium SAE- dA/ds < 0 is
The best performance results from this system when the un- represented by the regions in the figure marked at /. If the
certainty is minimal at the closed-loop system's equilibrium system remains in region / at all times, it will converge from
point. The best guaranteed precision is then given by any initial condition for all a, 7 > 0 and appropriate choice
of terminal exponents. For example, if dA/ds is as shown by
-*min
"pc- (32) curve C, then from initial condition SBr the system will con-
a
verge to some steady state SBE- When dA/ds lies in the regions
where Amin represents the minimal value of the uncertainty. If marked //, the parameters a and 7 must be designed for control
conventional sliders were applied in a dynamic sense (as with robustness. Let the bound on dA/ds be denoted by the line D.
dynamic terminal sliders), such that Note that 60 = Tan[dA/ds\. The terminal slider parameters
p = s + ys = 0 where (33) must be chosen such that the initial value of s must be less
(in magnitude) than the value SBo, where SB is the projection
s = e + ae = 0 (34) of So on the s axis and S0 is the intersection between the 7 5d/
then ePc would be equal to A m i n /a7. This changed only by a 5n ls6"/6<?~1 versus s hyberbolic curve, and the line marked D.
factor of rj. Conversely, when terminal sliders are utilized, For such a design, the closed-loop behavior would resemble
from Eq. (28), the guaranteed precision becomes the curve marked G from S0 to SE. Note that as the value of
s grows smaller, for a constant value of 7, the system will be
robust to increasing values of dA/ds, i.e., larger values of 80-
In summary, we have demonstrated greater performance
(35) with terminal sliders when partial or complete model infor-
mation is available. For computed torque control schemes,
Equations (35) may be reunited at, terminal sliders have demonstrated significant payoffs, espe-
cially in the context of control convergence and robustness.
e^a'^-W A \ (36)

Typically Amin/"ya « 1, and the exponents, /3d//3„ and 5d/8nt 7 Conclusions and Future Directions
axe greater than unity. Thus In this paper, we have studied the control requirements for
two robotic operational scenarios, i.e., namely robot-assisted
extravehicular activity and autonomous sample acquisition.
Our analysis has revealed that current robust control ap-
Further, choosing proaches are rather limited, particularly in their (convergence)
properties and robustness to changing uncertainties in the dy-
a>l (38) namic models. We concentrated on situations where the desired
performance constraints preclude recourse to high frequency
(39) control switching, thereby obviating the applicability of con-
ft, <5„ ventional sliding mode based robust control algorithms. To
it can be ensured that eP/ < e , thereby suggesting that a ensure the robustness and guaranteed measures required by
greater guaranteed precision can always be maintained using the aforementioned applications, we introduced a radically
terminal sliders, as compared to conventional sliders. different approach to nonlinear control synthesis, based upon
Figure 3 illustrated the properties of robust terminal slider a class of sliding modes denoted and dynamic terminal sliders.
control. Curves A and B show the phase space behavior of The proposed framework was then exploited to synthesize tra-
dynamic sliders (s versus 5 plots). With full model information, jectory control of robot manipulators. Our analysis demon-
from any initial state {SAl or SB,) the system will converge strated that the proposed slider formalisms yield superior

Journal of Dynamic Systems, Measurement,

Downloaded From: https://dynamicsystems.asmedigitalcollection.asme.org and Control
on 11/28/2018 SEPTEMBER 1993,
Terms of Use: http://www.asme.org/about-asme/terms-of-use Vol. 115 / 559
 performance with both complete as well as partial model in- 21 Winters, J, M., Stark, L., and Seif-Naraghi, A-H., "An Analysis of the
Sources of Musculoskeletal System Impedance," Journal of Biomechanics, Vol.
formation. For example, we achieved terminal tracking with 21, No. 12, 1988, pp. 1011-1025.
the new approach as compared to exponential tracking when 22 Winters, J. M., Stark, L., "Estimated Mechanical Properties of Synergistic
state-of-the-art conventional computed torque control schemes Muscles Involved in Movements of a Variety of Human Joints," Journal of
were used. Biomechanics, Vol. 21, No. 12, 1988, pp. 1027-1041.
23 Zak, M., "Terminal Attractors for Content Addressable Memory in Neural
In the presence of uncertainties, conventional sliding mode Networks," Physics Letters, Vol. 133, 1988, pp. 218-222.
formalisms exhibit finite time reaching of the sliding surface 24 Zak, M., "Cumulative Effect at the Soil Surface Due to Shear Wave
only through control switching. Furthermore, dynamic ter- Propagation," ASME Journal of Applied Mechanics, Vol. 50, 1983, pp. 227-
minal attractor effect ensures time constrained convergence 229.
for any arbitrary condition. More significantly, terminal sliders
provide increased overall guaranteed precision (i.e., system
performance at the equilibrium point). The latter property is
known particularly enabling for in-space automated servicing Using Passive End-Point Motion
and assembly. Our future directions of research include for-
mulation of higher order coupled terminal sliders and their
Constraints to Calibrate Robot
application to fast adaptive force and impedance control. Manipulators
Acknowledgments M. R. Driels1
The work described in this paper was performed by the JPL,
Caltech under the sponsorship of National Aeronatics and
Space Administration, Office of Aeronautics and Exploration
Technology. This work was sponsored in part by NASA Code The concept of using a method to constrain the endpoint of
RM through the Sample Acquisition Analysis and Preservation a serial linkage {such as a robot manipulator) in order to
Project, and through a JPL Directors' Discretionary Fund identify the kinematic parameters is appealing, since it no
award. Thanks to D. Bayard and M. Zak for fruitful discus- longer becomes necessary to use ancillary equipment to meas-
sions. ure the pose, partial or otherwise. This means that the joint
angles obtained from the manipulator and knowledge of the
type of constraint, are all that is needed to perform the cali-
References bration. This concept is applied to the calibration of a six-
degree-of-freedom manipulator by connecting its end point to
1 Arimoto, S., and Miyazaki, F., "Stability and Robustness of PID Feed-
back Control of Robot Manipulators of Sensor Capability,'' Proc. International
a table by means of a ball bar. The kinematic equations of the
Symposium on Robotics Research, MIT Press, Cambridge, MA, 1983. manipulator and constraint are developed, and then used to
2 Asada, H., and Slotine, J. E., Robot Analysis and Control, Wiley, New simulate the calibration process in order to determine whether
York, NY, 1986. the noise level present in the measurements disrupts the con-
3 Atkeson, C. G., An, C. H., and Hollerbach, J. M., "Estimation of Inertial
Parameters of Robot Loads and Links," Proc. International Symposium on
vergence of the process, the effect of measurement noise on
Robotics Research, MIT Press, Cambridge, MA, 1986. the accuracy of identification and the number of experimental
4 Barhen, J., and Gulati, S., "Self-organizing Neuromorphic Architecture observations needed. The acquisition of experimental data is
for Manipulator Inverse Kinematics," NATO ASI Series, Ed. C. S. G. Lee, then described, and the results of the identification are dis-
Vol. 44, 1990. cussed. It is concluded that for certain types of kinematic
5 Bayard, D., Private Communication, March 1990.
6 Brady, M., Robotics Science, MIT Press, Cambridge, MA, 1989. structures, this method of calibration is particularly attractive
7 Fillipov, A. F., "Differential Equations with Discontinuous Right-Hand since it is rapid, simple to perform, and requires very little
Sides," Annals Mathematical Society Transactions, Vol. 42, 1964. precision equipment.
8 Hill, A. V., "The Heat of Shortening and the Dynamic Constants of
Muscle," Proceedings Royal Society, Vol. 126B, pp. 136-195.
9 Hogan, N., "On the Stability of Manipulators Performing Contact Tasks,"
IEEE Journal of Robotics and Automation, Vol. 4, No. 6, Dec. 1988.
10 Hogan, N., "Impedance Control: An Approach to Manipulation: Part
I—Theory, Part II—Implementation, Part III—Applications," ASME JOURNAL 1 Introduction
OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL, 1985, pp. 1-24.
11 Jayasuriya, S-, and Diaz, A. R., "Performance Enhancement of Distrib-
In this study, the intent is to improve the accuracy of a robot
uted Parameter Systems by a Class of Nonlinear Controls," Proceedings Con- manipulator arm through the use of a calibration process. The
ference on Decision and Control, LA, CA, Dec. 1987, pp. 2125-2126. term accuracy is defined to be the average precision with which
12 Kazerooni, H., "Human/Robot Interaction Via the Transfer of Power a manipulator will move its end effector to a demanded location
and Information Signals—Part I, I I , " Proceedings 1989 International Confer-
ence on Robotics and Automation, Scottsdale, AR, 1989, pp. 1632-1649.
in space. Repeatability, on the other hand, is defined to be the
13 La Salle, J., and Lefschetz, S., Stability by Lyapunov's Direct Method, average precision with which a manipulator will move its end
Academic Press, New York, 1961. effector to a previously taught location in space. Manipulators
14 Pousson, M., Van Hoecke, J., and Goubel, F., "Changes in Elastic Char- have repeatabilities of the order of 0.1 mm but have accuracies
acteristics of Human Muscle Induced by Eccentric Exercise," Journal of Bio- several orders of magnitude greater. In applications where
mechanics, Vol. 23, No. 4, 1990, pp. 343-348.
15 Rowley, D. S., and Appl, F. C , "Analysis of Surface Set Diamond Bit spatial data bases are used to program the motion of a robot
Performance," Society of Petroleum Engineers Journal, Sept. 1969, pp. 301- (off-line programming), the performance of the manipulator
310. will depend upon its accuracy and not its repeatability. The
16 Slotine, J-J., Li, W., "Adaptive Manipulator Control: A Case Study," resulting poor performance is a major impediment to using
IEEE Transactions on Automatic Control, Vol. 33, No. 11, Nov. 1988, pp..
995-1003.
robots in the way they were intended to be used, and has
17 Spong, M. W., and Vidyasagar, M., Robot Dynamics and Control, Wiley, contributed to their lack of acceptance in many manufacturing
New York, NY, 1989. applications.
18 Venkataraman, S. T., and Marzwell, N., "Telerobots for Robust Space
Operations," (to appear) Proceedings SPACE 92, Denver, CO, 1992.
19 Wen, J. T., and Bayard, D. S., "New Class of Control for Robotic 'Department of Mechanical Engineering, Naval Postgraduate School, Mon-
Manipulators, Part 1. Non-Adaptive Case," International Journal of Control, terey, CA 93943.
Vol. 47, No. 5, 1988, pp. 1361-1385. Contributed by the Dynamic Systems and Control Division of the THE AMER-
20 Whitney, D., "Historical Perspective and State of the Art in Robot Force ICAN SOCIETY OF MECHANICAL ENGINEERS. Manuscript received by the Dynamic
Control," Proceedings, IEEE Conference on Robotics and Automation, San Systems and Control Division January 15, 1991. Associate Technical Editor:
Francisco, CA, Apr. 1985, pp. 269-274. K. Youcef-Toumi.

Downloaded From: 560 / Vol. 115, SEPTEMBER 1993

https://dynamicsystems.asmedigitalcollection.asme.org Transactions
on 11/28/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use of the ASME

View publication stats