IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO.

6, NOVEMBER 2000 753

Robust Backstepping Control of Nonlinear Systems

Using Neural Networks
Chiman Kwan, Senior Member, IEEE and F. L. Lewis, Fellow, IEEE

Abstract—A controller is proposed for the robust backstepping notice that the regression matrix almost covers one full page
control of a class of general nonlinear systems using neural net- in the Transactions on Control System Technology. In addition,
works (NNs). A new tuning scheme is proposed which can guar- the so-called linearity-in-the-parameter assumption may not be
antee the boundedness of tracking error and weight updates. Com-
pared with adaptive backstepping control schemes, we do not re- true in many practical situations. For example, friction in a robot
quire the unknown parameters to be linear parametrizable. No re- is a complicated nonlinear process that is hard to model as a
gression matrices are needed, so no preliminary dynamical analysis linear-in-the-parameter process.
is needed. One salient feature of our NN approach is that there is Parallel to fast development in adaptive and robust control
no need for the off-line learning phase. Three nonlinear systems, in- techniques, neural networks (NNs) have been applied to system
cluding a one-link robot, an induction motor, and a rigid-link flex-
ible-joint robot, were used to demonstrate the effectiveness of the identification [7], [18] or identification-based control [5], [34],
proposed scheme. [35]. Uncertainty on how to initialize the NN weights leads to
the necessity for “preliminary off-line tuning” [5], [10]. Re-
Index Terms—Adaptive, backstepping control, neural networks,
nonlinear systems, robust. cently, many NN controllers have been proposed for various
control applications that can provide closed-loop stability [6],
[17], [26]–[28], [31], [36], [38]–[40], [44].
I. INTRODUCTION In this paper, a unified and general approach to backstepping
control of nonlinear systems using NN is presented. We will
I N RECENT adaptive and robust control literature, there has
been a tremendous amount of activity on a special control
scheme known as “backstepping” [20], [23], [24]. When used
use neural nets in each stage of the backstepping procedure
to estimate certain nonlinear functions. This means that lin-
under some mild assumptions, many existing robust and adap- earity-in-the-parameter assumption is not needed, and no
tive control techniques can be extended to wide classes of appli- regression matrices need be found. Thus, a major problem
cations [23]. Recent papers in [4], [12], and [29] have applied with backstepping is cured. Recent papers in [26]–[28] and
such techniques to various kinds of robotic control schemes with [31] have initially applied this new idea to robots and motors.
the inclusion of motor dynamics. A major problem with back- The objective of this paper is to further generalize our work
stepping approaches is that certain functions must be “linear in to more general nonlinear systems with the goal of retaining
the unknown parameters,” and some very tedious analysis is the advantage of systematic design in backstepping control,
needed to determine “regression matrices.” For instance, even while eliminating its tedious and lengthy procedure of finding
for two robots within the same class (same number of links, the regression matrices. Compared with other NN approaches,
revolute joints) but with a different number of unknown param- the NN weights here are tuned on-line, with no learning phase
eters, minor changes in link lengths and masses, etc., one has to required. Most importantly, we can guarantee the boundedness
restart the whole tedious process of determining the regression of tracking error and weight updates.
matrix again. For a robot with six links, the job becomes even The paper is organized, as follows. In Section II, we will
more difficult. Although symbolic computation may offer some give a description of a class of nonlinear systems, system
help, one still has to manually manipulate and combine a lot of stability, and an example of standard backstepping design.
terms in the dynamical equations. In the case of backstepping Then, in Section III, we will introduce our NN backstepping
adaptive control, the problem of determining and computing the controller. Closed-loop stability of NN will also be stated and
regression matrices becomes even more acute. The complexity proven in Section III. Several practical applications, including a
of the regression matrices and the number of unknown parame- one-link robot tracking, speed control of induction motors, and
ters increase with each step of the backstepping process. If one rigid-link flexible-joint robot trajectory control, will be given
looks at a recent paper [13], which talks about the application of in Section IV. Finally, conclusions will be given in Section V.
backstepping technique to a simple DC motor control, one will
II. PRELIMINARIES
A. System Description
Manuscript received April 1, 1999; revised October 13, 2000. This work was
supported by NSF Grant IRI-9216545. This paper was recommended by Asso- Robust control of nonlinear systems with uncertainties is of
ciate Editor S. Lakshmivarahan. prime importance in many industrial applications. The model of
C. Kwan is with Intelligent Automation Inc., Rockville, MD 20850 USA many practical nonlinear systems can be expressed in a special
(e-mail: [email protected]).
F. L. Lewis is with the Automation and Robotics Research Institute, The Uni- state-space form
versity of Texas at Arlington, TX 76118 USA.
Publisher Item Identifier S 1083-4427(00)11082-3.

1083–4427/00$10.00 © 2000 IEEE

754 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

design can be easily used to tackle this problem. The procedure

consists of two steps.
Step 1—Treat as a Fictitious Control Input to (2.3a): We
can simply choose to be of the form
(2.1)
(2.4)
where denote the states
of the system, is the vector of control inputs, Substituting (2.4) into (2.3a) yields
are nonlinear functions that
contain both parametric and nonparametric uncertainties, and (2.5)
’s are known and invertible. Many systems can be expressed Define the following Lyapunov function candidate
in the above form. For example, rigid robots and motors [4],
[12]–[15], [26]–[31], power converters [37], and jet engines (2.6)
[24].
The requirement that ’s be invertible and known may be Differentiating (2.4) and using the dynamics of (2.3) gives
stringent. However, as will be illustrated in Section IV, one can
eliminate this requirement by exploiting physical properties of
a given nonlinear system. Hence, from Lyapunov stability theory, is globally asymp-
Equation (2.1) is known as strict-feedback form [24]. The totically stable. If the nonlinear function in (2.4) contains
reason for this name is that the nonlinearities depend unknown parameters, we can use a neural network to approxi-
only on , that is, on state variables that are “fed mate this function.
back.” Step 2—Realize the Fictitious Control Signal : Since
Other more general nonlinear systems such as “pure-feed- is only a fictitious control signal, we need to find
back” and “block-strict-feedback” forms [24] will be dealt with a way to realize this. Let us denote this desired signal by ,
in future papers. i.e., . Define
The control objective is to make to follow certain desired
trajectory . In the case of robotic control, denotes the joint (2.7)
angles of the robot. as the error between and . Differentiating gives
B. Stability of Systems [31] (2.8)
Consider the following nonlinear system
Choosing the following controller
(2.2)

with state . We say the solution is uniformly ulti- (2.9)

mately bounded (UUB) if there exists a compact set
such that for all , there exists an and a will make the error to go to zero exponentially since, after
number such that for all . substituting (2.9) into (2.8), the resulting equation becomes

C. An Example of Backstepping Design (2.10)

To illustrate the backstepping design procedure, let us con- Hence, will go to zero. Therefore, (2.5) will be valid
sider a very simple nonlinear system in “strict-feedback” form. and will go to zero. Here again, if (2.9) contains significant
The system is a second-order nonlinear system described by nonlinearities due to the functions and , we can also use
a second neural network to approximate them. Details of ex-
(2.3a) tending the above ideas to more general nonlinear systems will
(2.3b) be described in Section III.
Although the above backstepping procedure becomes more
where complicated when there exist parametric uncertainties in the
, scalar state variables; systems, the basic idea remains the same. The complications are
nonlinear functions with for all states; due to the following problems with the existing robust and adap-
control input. tive procedures. First, “regression matrices” in each step of the
The control objective is to make both to go to zero despite backstepping design must be determined. For example, it is well
the presence of nonlinearities. One important observation of this known that, in the first step of designing adaptive controllers
system is that the linearized system is uncontrollable since the for robots, one has to determine the regression matrix which is
linearized equation for (2.3a) is of the form a very tedious and time consuming task. This procedure gets
even more involved as the number of backstepping increases.
Second, one basic assumption in the current robust and adap-
which is clearly uncontrollable. Hence, linear control tech- tive backstepping methods is that the unknown system param-
niques cannot be used for (2.3). On the contrary, backstepping eters must satisfy the so-called linearity-in-the-parameter as-
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 755

sumption. This may not be true in many practical situations. For

example, friction in robots is a complicated nonlinear process
which is hard to model as a linear-in-the-parameter process. An-
other example will be that certain nonlinear functions may not
be linear parametrizable, i.e., if where is an un-
known system uncertainty.
The goal of this paper is to present a unified approach of back-
stepping control to a class of nonlinear systems using neural net-
works. While retaining the merit of systematic design in back-
stepping control, we strive to alleviate the disadvantage of the
tedious and lengthy process of determining and computing the
regression matrices. The proposed controller is also re-usable
in a sense that it is applicable to a class of, for instance, 6-link
Fig. 1. Two-layer NN.
revolute joint robots.

III. ROBUST BACKSTEPPING CONTROLLER DESIGN USING NNs controllers for to . Each is designed with the aim
to reduce error between in the previous design
A. NN Basics stage. Second, we design an actual controller for to force the
Let denote the real numbers, the real -vectors, error between and as small as possible. In each step of
the real matrices. Let be a compact simply connected the design process, NNs are used to approximate the nonlinear
set of . With map , define the functional functions in the error dynamics. The overall control structure
space such that is continuous. We denote by any suitable in shown in Fig. 2. Third, we perform an overall closed-loop
vector norm. When it is required to be specific we denote the stability and performance analysis of the on-line weight-tuning
p-norm by . Define as the collection of NN weights. algorithm.
Then the net output is Step 1—Design Fictitious Controllers for , and
: First of all, we design the fictitious controller for . Re-
(3.1) calling that
A general nonlinear function , can
(3.3)
be approximated by an NN as

(3.2) Choosing the following fictitious controller

with a NN functional reconstruction error vector. The (3.4)

structure of a two-layer NN is shown in Fig. 1.
For suitable NN approximation properties, must satisfy with a design parameter, the estimate of and
some conditions. substituting (3.4) into subsystem (3.3) yields the error dynamics
Definition 1 [38]: Let be a compact simply connected set
of , and be integrable and bounded. Then (3.5)
is said to provide a basis for if
with . The form of will be discussed in the next
1) A constant function on can be expressed as (3.2) for
section. The usual adaptive backstepping approach is to assume
finite .
that the unknown parameters in [24] are linear parametriz-
2) The functional range of NN (3.2) is dense in for
able (LP) so that standard adaptive control can be used in this
countable .
stage. As we will see in a moment, we will use a two-layer NN
It is emphasized that a basis is not difficult to find. The
to approximate . The advantage is that no linearity-in-the-un-
radial basis functions, for instance, provide a universal basis for
known-system-parameters assumption is needed and no regres-
all smooth nonlinear functions [40]. Barron [1] has shown that
sion matrix need be found.
the approximation error in (3.2) can never be made smaller than
The next step of backstepping design is to make the error be-
order where is order of the input space. Despite
tween and as small as possible. Differentiating defined
the lower bound achievable for , the tracking error will be
in (3.5) gives
proven to be bounded by a term as shown in (3.23), and the
bound can be made small by increasing the gain of certain co- (3.6)
efficients in the controller.
A fictitious controller for of the form
B. Controller Structure
Referring to (2.1), our control objective is to make to (3.7)
follow a desired trajectory . The idea of backstepping is
like this. First, we treat up to as fictitious control sig- can be chosen. Note that there is a coupling term in (3.5).
nals. In this stage, we use NN approach to design the fictitious The purpose of the term in (3.7) is to compensate the
756 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

Fig. 2. Backstepping NN control of nonlinear systems in the “strict-feedback” form.

effect of the coupling due to . Substituting the fictitious with a design parameter and the estimate of .
controller (3.7) into (3.6) gives Similar to Step 1, the usual backstepping design procedure is
to assume to be LP in unknown system parameters. How-
ever, in our controller design here, we will use a two-layer NN
to approximate which means no LP or regression matrix
with , a design parameter and the requirement is needed. Also note that a term is
estimate of . added in (3.9) which is necessary to compensate the coupling
In a similar fashion, we can design a fictitious controller for effects introduced in (3.9).
to make the error as small as Step 3—Closed-Loop Stability and Performance Analysis of
possible, i.e., NN Weight Tuning Algorithm: We will perform a detailed treat-
ment of stability and performance analysis of a weight-tuning
algorithm in Section III-C. Using Lyapunov stability theory we
(3.8) will carry out the stability analysis. We can show that all signals
including tracking error, NN weights are all UUB. The overall
The dynamics of is then governed by control structure is shown in Fig. 2. It is important to note the
simplicity of NN control when compared to adaptive backstep-
ping control. In adaptive backstepping control, it is assumed that
(3.9) (3.5), (3.6), (3.9), (3.10) are linear in terms of
known regression matrices. These regression matrices are very
with , a design parameter and tedious to find and must be computed for each specific system.
the estimate of . In fact, for some systems the LP assumption may not hold. For
Conventional backstepping design assumes that example, friction in robot is a complicated nonlinear process
are LP in unknown system parameters. that is hard to model as a linear-in-the parameter process. An-
Regression matrices are needed for all ’s, . other simple example is that the nonlinear function may be in
This is a very lengthy and tedious process which needs to be the form of which is clearly not LP. On the other hand, if
repeated for each new system. As we will see in Section III-C, bases in (3.13) are appropriately chosen, then
our design procedure uses NNs to approximate the complicated NN equations (3.13) are valid. No regression matrices need be
nonlinear functions ’s, . As a result, no re- computed. It has been shown that the sigmoid can form a basis
gression matrices are needed and the controller is re-usable for set [1], [11], [19]. In [40], it was shown that the radial basis
different systems within the same class of nonlinear systems. functions can form a basis. In [8], it was shown that a basis set is
Step 2—Design of Actual Control : After the fictitious con- particularly easy to choose for CMAC (Cerebellar Model Arith-
troller is designed, we need to find a way to realize them. metic Computer) neural network.
Differentiating defined in (3.9) yields
C. Bounding Assumptions, Error Dynamics, and Weight
(3.10) Tuning Algorithms
Assume that the nonlinear functions ’s,
Choosing the controller of the form in (3.5), (3.6), (3.9), and (3.12) can be represented by
2-layer neural nets for some constant “ideal” weights
(3.11)
, i.e.,
gives the following dynamics for error

(3.12) (3.13)
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 757

where ’s provide suitable basis functions for the NNs. The Theorem 1: Suppose Assumptions 1 and 2 are satisfied. Take
net reconstruction errors are bounded by known constants the control input (3.11) with NN weight tuning be provided by
.
Define the NN functional estimate of in (3.13) by (3.18)

(3.14) with constant matrices ,

and scalar positive constant . Then the errors
with the current NN weight estimates provided by the tuning , and NN weight estimates are UUB. The errors
algorithms. Then the error dynamics (3.5), (3.6), (3.9), (3.12) , can be kept as small as possible by
become increasing gains in (3.16).
Proof: Let the NN approximation property [i.e., (3.13)]
holds with given accuracy ’s for all in the compact set
with a positive constant. Let .
Now consider the following Lyapunov function candidate

(3.19)
(3.15) Differentiating (3.19) and using (3.16) gives

Define (3.20)

Applying the following inequality (also known as Schwartz in-

equality in [16]) to (3.20)
diag diag
diag
(3.21)

we have

(3.22)
The error dynamics (3.15) can be expressed in terms of the
above quantities as which is negative as long as the term in square bracket is pos-
itive. Here is the minimum eigenvalue of . Completing
(3.16) the square for the term inside the square bracket in (3.22) yields

Note that the term denotes the couplings between the error
dynamics in (3.16). The matrix is skew-symmetric. The
closed-loop stability analysis and the weight tuning algorithms
will be discussed in the next section. which is positive as long as
Two standard assumptions, which are quite common in the (3.23)
neural networks literature [17], [26]–[28], [31], [44] are stated
next. or
Assumption 1: The ideal weights are bounded by known pos-
itive values so that (3.24)

Thus, is negative outside a compact set. The form of the

right-hand side of (3.24) shows that the control gain , which
or equivalently are contained in , can be selected large enough so that

(3.17)
According to a standard Lyapunov theorem extension (Narendra
where diag and is known. The
and Annaswamy 1987), this demonstrates the UUB of both
symbol denotes the Frobenius norm, i.e., given a matrix
and . Q.E.D.
, the Frobenius norm is given by
Remarks:
a) A comparison with -modification [33] shows that the
NN reconstruction error increases the bounds on
and in a very interesting way. Note, how-
Assumption 2: The desired trajectory and its derivatives ever, that small tracking error bounds may be achieved
up to the th order are bounded. by selecting large control gain . On the other hand,
758 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

the NN weight error is fundamentally bounded by , these applications have been included in [27] and [28],
the known bound on the ideal weights . The parameter we present some partial results here to illustrate that
offers a design tradeoff between the relative eventual stability is still achievable even ’s are unknown.
magnitudes of and ; a smaller yields a To deal with more general systems with unknown ’s,
smaller and a larger , and vice versa. we propose to use an approach described in [31, p. 287].
b) Similar to -modification in [33], no persistency of ex- This research is still underway.
citation (PE) is needed to establish the bounds on NN
weight errors with the weight tuning algorithm (3.18). IV. APPLICATIONS
c) The contrast between the NN approximation property
In this section, we present three applications. The first one is
(3.2) and the adaptive control linear-in-the-parameter
a one-link robot system with the inclusion of motor dynamics.
(LIP) assumption should be understood.
As pointed out by Tarn et al. [43], the effects of motor dy-
1) Both are linear in the tunable parameters, but the namics will affect the performance of overall robot tracking. In
former is linear in the tunable NN weights, while the this case, the ’s are assumed to be known. The backstepping
latter is linear in the unknown system parameters. NN theory described in Section III can be directly applied. The
2) The former holds for all functions in , second application is the robust control of an induction motor.
while the latter holds only for a specific function . Here, the ’s are actually unknown. However, by exploiting
3) In the NN property, the same basis set suffices the physical properties of the motor dynamics, we circumvented
for all , while in the LIP assump- the problem of unknown ’s. In the third application, we ap-
tion the regression matrix depends on and plied the theory of backstepping NN control to rigid-link flex-
must be recomputed for different . That is, for ible-joint system. Properties of robot dynamics were used to al-
instance, one must recompute for each different leviate the unknown problems.
type of robot arm. Therefore, the two-layer NN con- These applications, especially the last two, clearly demon-
troller is significantly more powerful than adaptive strate that the proposed backstepping control using NN has great
controllers; it provides a universal controller for robot potential in many diverse applications.
arms within the same class. An example of “class” is
the class of 2-link revolute robot arms with flexible A. One-Link Robot Tracking
joints. Consider a one-link manipulator with the inclusion of motor
dynamics. The robot model is given by
d) It should be emphasized that our NN controller design
procedure was motivated by a technique known as back-
stepping control [24]. In backstepping design procedures, (4.1)
preliminary dynamical analysis to determine regression
matrices is crucial. The procedure becomes very tedious Equation (4.1) can be expressed in the form (2.1) by noting that
if we are dealing with a robot with multiple degrees
of freedom. In [29], a backstepping design for RLFJ
was given using sliding mode and adaptive control. The
derivation of regression matrices, even for the one-link
RLFJ simulation model, was very time-consuming and
tedious. An immediate advantage of NN design is the no The parameter values with appropriate units are given by
regression analysis is needed and the controller structure , , , , , . The
is reusable for different robots with different masses and desired trajectory is . The design procedure in
lengths within the same class. Section III was modified slightly. First, we defined a filtered
e) Note that the problem of neural net weight initialization tracking error with and .
does not arise, since if ’s are taken as zeroes the Second, we design a fictitious NN controller for , namely ,
linear proportional control term stabilizes the which drives to zero. Third, we design a second NN controller
system on an interim basis. Similar to other NN methods, for u to drive the error between and to zero. The con-
our control scheme does not guarantee that ’s will troller parameters are , , , and
converge to the true ’s. All we can say is that we can . The number of neurons used in each of the two 2-layer
guarantee the boundedness of ’s. NNs is 10. We used sigmoids for . The initial conditions for
f) In Section II, it was assumed that ’s are known and are 0.1, 6.28, and 0, respectively. The robust tuning
invertible. Although this may sound restrictive, it should algorithms (3.18) are used for the simulations. Simulation re-
be emphasized that, in many practical applications, sults are shown in Fig. 3. The performance is very good. The
the above mentioned restriction can be alleviated by tracking errors are reduced significantly when NNs are used.
exploiting the physical properties of the system. Two
practical applications will be described in Section IV: B. NN Control of Induction Motors
one for induction motor control and the other one for The nomenclature of induction motors can be found in [32]
rigid-link flexible-joint robot control. Although the and details of controller derivation, proof, and simulations can
details of controller derivation and simulation results of be found in [27].
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 759

Fig. 3. Performance of robust NN backstepping controller for 1-link robot tracking.

1) Model of Induction Motor: This model is known as field- with , , ,

oriented model, which was introduced by Blaschke [2]. It in- and . Since is unknown, , are also un-
volves a transformation from the stator fixed frame to a known. However, and are known. Blaschke [2] also devel-
frame , which rotates along the flux vector . The oped a feedback controller to control (4.4). Other nonlinear ap-
transformations between currents and flux magnitudes in dif- proaches include [3], [21], [22], and [25]. Marino et al. [32] used
ferent frames are given by adaptive input–output decoupling technique to tackle the control
problem. Our work here uses the same field-oriented model and
follows the same assumptions as those in [32]. In the next sec-
(4.2) tion, we will make use of a special structure of the above model
to perform our NN controller design.
(4.3) 2) Controller Structure, Error Dynamics, and Weight Update
Rules: If we define and , then (4.4) is in
where strict-feedback form (2.1). We first treat , as the ideal ficti-
tious control signals for a subsystem consisting of (4.4a)–(4.4b).
We design an NN controller for , . Finally, we use a second
2-layer NN to realize these fictitious signals. It should be noted
that special physical properties of the motor dynamics were ex-
The field-oriented model of induction motor in frame is ploited so that the problem of unknown ’s can be eliminated.
given by [2] Our control objective is to regulate the rotor speed and the mag-
netic flux magnitude. Denote and as the desired reference
(4.4a) levels of and , respectively.
Step 1—Selection of Desired and to Control Subsystem
(4.4b) (4.4a) and (4.4b): First, we rewrite (4.4a)–(4.4b) as

(4.4c) (4.5a)

(4.5b)

(4.4d) where
(4.4e)
760 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

Fig. 4. NN backstepping controller for induction motor.

Since in (4.5b) is unknown, this will cause some difficulties where

using the results of backstepping NN described in Section II. To
alleviate this difficulty, we apply the following trick. Dividing
both sides of (4.5b) by yields

(4.6)

Now the coefficient of is unity and known. Then we can

rewrite (4.5) as

(4.7)

where
It should be noted that is exactly known. To make as small
as possible, the following control is chosen:

(4.12)

Note that is the estimate of the unknown function . Similar

It should be noted that is exactly known and invertible. By to Step 1, we will use another two-layer NN to approximate
treating as a fictitious input, we design a controller for the ideal . Also note that a term is added in (4.12) which is
as necessary to cancel the effect of in (4.9) so that we will be
able to prove the closed-loop stability.
(4.8) Step 3—Closed-Loop Stability Analysis and On-Line Weight-
Tuning Algorithm: We now perform a detailed treatment of sta-
with a design parameter, , and the bility and performance analysis of a weight-tuning algorithm.
estimate of . Substituting (4.8) into (4.7) gives The overall control scheme is shown in Fig. 4.
Using (3.13) for and yields the following error dy-
(4.9) namics for :

where . The form of is given by (3.13), which is the (4.13)

output of a two-layer NN. (4.14)
Step 2—Realization of the Desired Reference Signals in
(3.8): In order to achieve the desirable result in Step 1, i.e., the Theorem 2: Let the desired trajectories and be
ideal fictitious control signal in (4.8), we need to find the error bounded. Take the control input (4.12) with NN weight tuning
dynamics of which is defined as be provided by

(4.10) (4.15a)
(4.15b)
Differentiating (4.10) and using the dynamics in (4.4) yields
with constant matrices , , and scalar
(4.11) positive constant . Then the errors are UUB. NN
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 761

Fig. 5. Performance of NN backstepping controller for induction motor.

weight estimates are bounded. The errors can be kept set diag , diag , ,
as small as desired by increasing gains diag in , . The applied voltage has the same mag-
(4.13) and (4.14). nitude as that of [32] and is well within inverter limits.
Proof: Consider the following Lyapunov function candi- It should be noted that the plots of , are not due to
date: switching in sliding mode control as it appears to be. There is
no switching term in the NN controller. Similar waveforms have
also been observed in [32]. This phenomenon is just a charac-
teristic of induction motor dynamics.
with defined in (4.13) and is a identity matrix. Then
the proof follows the same procedure described in Section III. C. NN Backstepping Control of N-DOF Rigid-Link
Details of the proof can also be found in [27]. Q.E.D. Flexible-Joint Robots
Finally, is related to the actual control , 1) RLFJ Robot Model and Its Properties: The model for an
through the following relation: -link RLFJ robot is given by [42]

(4.16) (4.17a)
(4.17b)
Simulation Results: Using the data in [32], we simulate the
robust backstepping NN controller. The model we used was with denoting the link position, velocity, and ac-
the original motor model before the state transformation from celeration vectors, respectively, the inertia ma-
stator frame to rotator frame was applied. In other words, the trix, the centripetal-Coriolis matrix,
field-oriented model was only used for controller design. The the gravity vector, representing the friction
results are shown in Fig. 5. We used four and ten neurons in the terms, the additive bounded disturbance,
two NNs which approximate , respectively. The inputs the motor shaft angle, velocity, acceleration, respectively,
to NN1 consist of , , , and . The inputs to NN2 con- the difference between motor and joint
sists , , , , , . The reference tra- angles, the positive definite constant diagonal ma-
jectories are the same as those in [32]. Reference is zero trix which characterizes the joint flexibility, a posi-
from 0 to 0.3 s., 220 r/s from 0.3 to 5 s., and 350 r/s from 5 s. tive definite constant diagonal matrix denoting the motor inertia,
onwards. Reference is 1.3 Wb from 0 to 5 s. and 0.8 Wb representing the natural damping term, the control
after 5 s. The discontinuities are smoothed by linear interpola- vector used to represent the motor torque, and repre-
tions. A load disturbance of 40 Nm is added at s. We senting an additive bounded torque disturbance.
762 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

If we define , , , , then (4.17)

is indeed in the form of (2.1). However, due to the presence of
uncertainties in the robot dynamics, we cannot directly apply the
backstepping NN control theory to RLFJ expressed in “strict-
feedback” form of (2.1). We need to exploit the following robot
properties in order to effectively control the robot system.
The rigid dynamics (4.17a) has the following properties [9],
[41]:
Property 1—Boundedness of the Inertia Matrix: The inertia
matrix is symmetric and positive definite, and satisfies the
following inequalities:

(4.18)

where and are known positive constants, and de- Fig. 6. Two-link robot with joint flexibility.
notes the standard Euclidean norm.
Property 2—Skew Symmetry: The inertia and cen-
Our controller design can be considered as consisting of three
tripetal-Coriolis matrices have the following property:
steps. The first step is to treat , the difference between motor
(4.19) shaft angle and joint angle , as a fictitious control signal to
the error dynamics (4.24). We call this fictitious signal . Then
where is the time derivative of the inertia matrix. (4.24) can be rewritten as
The joint elasticity matrix is bounded by
(4.26)
for arbitrary vector (4.20)
where is an error signal which we will try to make
where and are positive scalar bounding constants. The
as small as possible in the second step. The control objective of
motor inertia matrix is also bounded by
the first step is to design an NN controller for to make (and
for arbitrary vector (4.21) hence tracking error ) as small as possible. The structure of the
controller will be described below. The objective of the second
where and are positive scalar bounding constants. step is to design a second NN controller for another fictitious
Property 1 is very important in generating a positive definite signal such that the error signal is as small as possible.
function to prove stability of the closed-loop system. Property To achieve this, we need to derive the dynamic equation for .
2 will help in simplifying the controller. Many robust methods Differentiating and using (4.17b) yields
have incorporated Properties 1 and 2 in their controller designs
[9], [41]. It should be emphasized here that, unlike standard ro- (4.27)
bust and adaptive control schemes, we do not require linearity
in the unknown robot parameter assumption. where and is a very complicated nonlinear
2) Control Objective and Central Ideas of Our Controller function of , and . Now we need to derive a controller
Design: The control objective is to develop a link position- for u to make as small as possible. The error dynamics for
tracking controller for the RLFJ robot dynamics given by (4.17) is obtained by differentiating and multiplying the final
based on inexact knowledge of manipulator dynamics. To ac- expression by
complish this purpose, we first define the link position tracking
error as (4.28)

(4.22) where is another very complicated nonlinear function

, and . It should be noted that link acceleration
where denotes the desired link position trajectory. It is not needed in our controller. The reason is that, when-
is assumed that and its derivatives up to the fourth order are ever shows up in (4.27) and (4.28), it will be replaced by
bounded. In addition, we also define a filtered tracking error as . Finally, in the
third step, we will perform an overall stability analysis using
(4.23) Lyapunov stability theory.
where is a diagonal, positive definite control gain Now (4.26)–(4.28) are in modified “strict-feedback” form
matrix. Using (4.23) and (4.17a), we can derive the equation (2.1). The problem of unknown ’s is eliminated.
3) Control Design Procedure:
(4.24) Step 1—Design of NN Controller for : To design an NN
controller for the fictitious signal , we select the following
where the complicated nonlinear function is defined as structure:

(4.25) (4.29)
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 763

Fig. 7. NN backstepping controller for RLFJ.

where , is a positive definite matrix, is a ro- Differentiating yields

bustifying term to be defined shortly, and is defined in (4.20).
Note that will depend on certain update algorithms to be de- (4.37)
scribed in the next section. Substituting (4.29) into (4.26) gives
where
. Control is chosen to be

(4.30) (4.38)

The form of is chosen to be (Dawson et al. 1992 and Dawson where and is a control gain. Substituting
et al. 1995) (4.38) into (4.37) gives

(4.31) (4.39)

Step 3—Overall Stability Analysis: The stability analysis

where
will be proved by using Lyapunov stability theory. We can
show that all signals including tracking error, NN weights are
all UUB. The overall control structure is shown in Fig. 7.
(4.32)
Theorem 3: Let the desired trajectory and its derivative
up to the fourth order be bounded. Let ,
(sufficiently small number) (4.33) the control input be given by (4.29), (4.31), (4.34) and weight
tuning provided by
in (4.32) stands for the upper bound of .
Step 2—Design of and the Actual Control : For the
design of fictitious signal , we choose the following structure:

(4.34) (4.40)
where
with any constant symmetric matrices ,
(4.35) , , and scalar positive constant . Then
the errors , , , and NN weight estimates are. The
and . The weights will be generated from some errors , , can be kept as small as desired by in-
update algorithms to be described below. Inserting (4.34) into creasing gains , , in (4.29), (4.31), (4.34).
(4.27) gives Proof: Consider the Lyapunov function candidate

(4.36) (3.23)
764 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

Fig. 8. Performance of NN backstepping controller to RLFJ.

where is defined as diag and , are eters are diag , diag ,

defined in (4.17), is the identity matrix. Then the proof follows diag . The inputs to the NNs are given by
a similar procedure as was described in Theorem 1. Details can
also be found in [28]. Hence it is omitted. Q.E.D.
Simulation Results: Consider a simple 2-link manipulator
shown in Fig. 6. The model for this robot system can be de-
scribed in the form of (4.17) with [30] where , , is chosen as diag .
Here we use sigmoid functions in the NNs. The gains of the
NNs are chosen as , . The
gains are diag , diag ,
diag . Fig. 8 shows the results using the NN con-
troller. Both tracking errors go to small values. The desired and
actual trajectory almost overlap with each other. In addition, we
do not even need the explicit expressions for those three highly
complicated nonlinear functions , and in (4.26)–(4.28),
respectively. This is a significant advantage since our controller
can be applied to any type of RLFJ robots of different masses
and lengths within the same class.

V. CONCLUSIONS
We have presented a general NN controller for the robust
backstepping control of a class of nonlinear systems. The
The parameter values are m, m, method does not require the system dynamics to be exactly
kg, kg, m/s . The flexible-joint param- known. Compared with adaptive backstepping control, linearity
KWAN AND LEWIS: ROBUST BACKSTEPPING CONTROL OF NONLINEAR SYSTEMS 765

in unknown parameters is not needed. A major problem with [20] I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse, “Systematic de-
backstepping is corrected in that no tedious computation of sign of adaptive controllers for feedback linearizable systems,” IEEE
Trans. Automat. Contr., vol. 36, pp. 1241–1253, 1991.
“regression matrices” is needed. Compared with other NN [21] G. S. Kim, I. J. Ha, and M. S. Ko, “Control of induction motors via
approaches, we do not require an off-line “training phase.” All feedback linearization with input–output decoupling,” Int. J. Control,
errors and weight are guaranteed to be bounded. The tracking vol. 51, no. 4, pp. 863–883, 1990.
[22] G. S. Kim, I. J. Ha, and M. S. Ko, “Control of induction motors for
error can be reduced to arbitrarily small values by choosing both high dynamic performance and power efficiency,” IEEE Trans. Ind.
certain gains large enough. Applicat., vol. 39, no. 4, pp. 323–333, 1992.
Several practical systems, including an induction motor and [23] P. V. Kokotovic, “Bode lecture: The joy of feedback,” IEEE Contr. Syst.
Mag., no. 3, pp. 7–17, June 1992.
a RLFJ robot, were used to demonstrate the effectiveness of the [24] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adap-
proposed controller. tive Control Design. New York: Wiley Interscience, 1995.
[25] Z. Krzeminski, “Nonlinear control of the induction motor,” in 10th IFAC
World Congress, 1987, pp. 349–354.
[26] C. M. Kwan, F. L. Lewis, and D. Dawson, “Robust neural-network con-
REFERENCES trol of rigid-link electrically driven robots,” IEEE Trans. Neural Net-
works, vol. 9, no. 4, pp. PAGE NOS?–, 1998.
[1] A. R. Barron, “Universal approximation bounds for superposition of a [27] C. M. Kwan and F. L. Lewis, “Robust backstepping control of induc-
sigmoidal function,” IEEE Trans. Inform. Theory, vol. 39, no. 3, pp. tion motors using neural networks,” IEEE Trans. Neural Networks (AU-
PAGE NOS?–, 1993. THOR: UPDATE?), September 2000.
[2] F. Blaschke, “The principle of field orientation applied to the new [28] C. M. Kwan, F. L. Lewis, and Y. H. Kim, “Robust neural network control
transvector closed-loop control system for rotating field machines,” of rigid link flexible-joint robots,” Asian J. Control, vol. 1, no. 3, pp.
Siemens Rev., vol. 39, pp. 217–220, 1972. 188–197, 1999.
[3] M. Bodson, J. Chiasson, and R. Novotnak, “High performance induction [29] C. M. Kwan and K. S. Yeung, “Robust adaptive control of revolute flex-
motor control via input–output linearization,” IEEE Control Syst. Mag., ible-joint manipulators using sliding technique,” Syst. Control Lett., vol.
vol. 14, no. 4, pp. 25–33, 1994. 20, pp. 279–288, 1989.
[4] T. C. Burg, D. M. Dawson, J. Hu, and P. Vedagarbha, “Velocity tracking [30] F. L. Lewis, C. Abdallah, and D. Dawson, Control of Robot Manipula-
control for a separately excited DC motor without velocity measure- tors. New York: MacMillan, 1993.
ments,” in Proc. Amer. Contr. Conf., 1994, pp. 1051–1056. [31] F. L. Lewis, S. Jaganathan, and A. Yesildirek, Neural Network Control
[5] F. C. Chen and H. K. Khalil, “Adaptive control of nonlinear systems of Robot Manipulators and Nonlinear Systems. New York: Taylor and
using neural networks,” Int. J. Control, vol. 55, no. 6, pp. 1299–1317, Francis, 1998.
1992. [32] R. Marino, S. Peresada, and P. Valigi, “Adaptive input-output lineariza-
[6] F.-C. Chen and C.-C. Liu, “Adaptively controlling nonlinear contin- tion control of induction motor,” IEEE Trans. Automat. Contr., vol. 38,
uous-time systems using multilayer neural networks,” IEEE Trans. Au- pp. 208–221, 1993.
tomat. Contr., vol. 39, pp. 1306–1310, 1994. [33] K. S. Narendra and A. M. Annaswamy, “A new adaptive law for robust
[7] S. R. Chu and R. Shoureshi, “Neural-based adaptive nonlinear system adaptation without persisitent excitation,” IEEE Trans. Automat. Contr.,
identification,” in Intelligent Control Systems, ASME Winter Annu. vol. AC-32, pp. 134–145, 1987.
Meet., vol. DSC-45, 1992. [34] K. S. Narendra, “Adaptive control using neural networks,” in Neural
[8] S. Commuri and F. L. Lewis, “CMAC neural networks for control of Networks for Control. Cambridge, MA: MIT Press, 1991, pp.
nonlinear dynamical systems: Structure, stability and passivity,” in Proc. 115–142.
IEEE Int. Symp. Intelligent Control, 1995, pp. 123–129. [35] T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, and Y. Ushikawa, “Trajec-
[9] J. J. Craig, Adaptive Control of Mechanical Manipulators. New York: tory control of robotic manipulators,” IEEE Trans. Ind. Electron., vol.
Wiley, 1986. 38, pp. 195–202, 1991.
[10] X. Cui and K. G. Shin, “Direct control and coordination using neural [36] M. M. Polycarpou and P. A. Ioannou, “Identification and control using
networks,” IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. PAGE neural network models: design and stability analysis,” Dept. Elect. Eng.
NOS?–, 1993. Sys., Univ. of Southern California, Tech. Rep. 91-09-01, September
[11] G. Cybenko, “Approximation by superpositions of a sigmoidal func- 1991.
tion,” Mathematics of Control, Signals, and Systems, vol. 2, no. 4, pp. [37] A. Sabanovic, N. Sabanovic, and K. Ohnishi, “Sliding modes in power
303–314, 1989. converters and motion control systems,” Int. J. Control, vol. 57, pp.
[12] D. Dawson, Z. Qu, and J. Hu, “Robust tracking control of an induction 1237–1259, 1993.
motor,” in Proc. Amer. Contr. Conf., 1993, pp. 648–652. [38] N. Sadegh, “Nonlinear identification and control via neural networks,”
[13] D. Dawson, J. J. Carroll, and M. Schneider, “Integrator backstepping in Control Systems with Inexact Dynamic Models, ASME Winter Annu.
control of a brush DC motor turning a robotic load,” IEEE Trans. Contr. Meet., vol. DSC-vol. 33, 1991.
Syst. Technol., vol. 2, no. 3, pp. PAGE NOS?–, 1994. [39] G. A. Rovithakis and M. A. Christodoulou, “Adaptive control of
[14] D. Dawson, Z. Qu, and M. M. Bridges, “Hybrid adaptive control for unknown plants using dynamical neural networks,” IEEE Trans. Syst.,
tracking of rigid-link flexible-joint robots,” Proc. Inst. Elect. Eng., vol. Man, Cybern., vol. 24, pp. 400–412, 1994.
140, pt. D, no. 3, pp. 155–159, 1992. [40] R. M. Sanner and J. J. Slotine, “Stable adaptive control and recursive
[15] D. Dawson, M. M. Bridges, and Z. Qu, Nonlinear Control of Robotic identification using radial Gaussian networks,” in Proc. IEEE Conf. De-
Systems for Environmental and Waste Restoration. Englewood Cliffs, cision and Control, 1991.
NJ: Prentice-Hall, 1995. [41] J. J. Slotine and W. Li, “Adaptive manipulator control: A case study,”
[16] F. R. Gantmacher, The Theory of Matrices. New York: Chelsea, 1959. IEEE Trans. Automat. Contr., vol. 33, no. 11, pp. 995–1003, 1988.
[17] S. S. Ge and C. C. Hang, “Direct adaptive neural network control of [42] M. W. Spong, “Adaptive control of flexible joint manipulators,” Syst.
robots,” Int. J. Syst. Sci., vol. 27, no. 6, pp. 533–542, 1996. Control Lett., vol. 13, pp. 15–21, 1989.
[18] B. Horn, D. Hush, and C. Abdallah, “The state space recurrent neural [43] T. J. Tarn, A. K. Bejczy, X. Yun, and Z. Li, “Effects of motor dynamics
network for robot identification,” in Advanced Control Issues for Robot on nonlinear feedback robot arm control,” IEEE Trans. Robot. Automat.,
Manipulators, ASME Winter Annu. Meet., vol. DSC-39, 1992. vol. 7, no. ISSUE NO?, pp. PAGE NOS?–, 1991.
[19] K. Hornik, M. Stinchombe, and H. White, “Multilayer feedforward net- [44] T. Zhang, S. Ge, and C. C. Hang, “Direct adaptive control of nonaffine
works are universal approximators,” Neural Netw., vol. 2, pp. 359–366, nonlinear systems using multilayer neural networks,” in Amer. Contr.
1989. Conf., 1998, pp. PAGE NOS?–.
766 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART A: SYSTEMS AND HUMANS, VOL. 30, NO. 6, NOVEMBER 2000

Chiman Kwan (S’85–M’93–SM’98) was born on F. L. Lewis (S’78–M’81–SM’86–F’94) was born

February 19, 1966, in Jilian, China. He received the in Würzburg, Germany, and subsequently studied in
B.S. degree in electronics (with honors) from the Chile and at the Gordonstoun School in Scotland.
Chinese University of Hong Kong in 1988 and the He received the B.S. degree in physics/electrical
M.S. and Ph.D. degrees in electrical engineering engineering from Rice University, Houston, TX, and
from the University of Texas at Arlington in 1989 the M.S. degree in electrical engineering from Rice
and 1993, respectively. University, both in 1971. After spending six years
From April 1991 to February 1994, he was with in the U.S. Navy, serving as Navigator aboard the
the Beam Instrumentation Department of the Super- frigate USS Trippe (FF-1075) and Executive Office
conducting Super Collider Laboratory (SSC), Dallas, and Acting Commanding Officer aboard the USS
TX, where he was heavily involved in the modeling, Salinan (ATF-161), he received the M.S. degree in
simulation, and design of modern digital controllers and signal processing algo- aeronautical engineering from the University of West Florida, Pensacola, in
rithms for the beam control and synchronization system. He later joined the Au- 1977 and the Ph.D. degree from the Georgia Institute of Technology (Georgia
tomation and Robotics Research Institute, Fort Worth, TX, where he applied in- Tech), Atlanta, in 1981.
telligent control methods such as neural networks and fuzzy logic to the control From 1981 to 1990, he was a Professor at Georgia Tech, where he is currently
of power systems, robots, and motors. Since July 1995, he has been the Director an Adjunct Professor. He is also a Professor of electrical engineering at The Uni-
of Robotics Research at Intelligent Automation, Inc., Rockville, MD, where he versity of Texas, Arlington (UTA), where he was awarded the Moncrief-O’Don-
has been Principal Investigator/Program Manager for more than 20 different nell Endowed Chair in 1990 at the Automation and Robotics Research Institute.
projects such as modeling and control of advanced machine tools, digital control He has studied the geometric, analytic, and structural properties of dynamical
of high-precision electron microscope, enhancement of microscope images, and systems and feedback control automation. His current interests include robotics,
adaptive antenna arrays for beam forming, automatic target recognition of FLIR intelligent control, neural and fuzzy systems, nonlinear systems, and manufac-
and SAR images, fast flow control in communication networks, vibration man- turing process control. He is the author/coauthor of two U.S. patents, 124 journal
agement of gun pointing system, health monitoring of flight critical systems, papers, 20 chapters and encyclopedia articles, 210 refereed conference papers,
high-speed piezoelectric actuator control, fault tolerant missile control, active and seven books. He was selected to the Editorial Boards of the International
speech enhancement, fault detection isolation of various electromechanical sys- Journal of Control, Neural Computing and Applications, and the International
tems, and underwater vehicle control. His primary research areas include fault Journal of Intelligent Control Systems. He is currently an Editor for the flagship
detection and isolation, robust and adaptive control methods, signal and image journal Automatica.
processing, communications, neural networks, and fuzzy logic applications. Dr. Lewis is a Registered Professional Engineer in the State of Texas. He is the
Dr. Kwan is listed in the New Millennium edition of Who’s Who in Science recipient of an NSF Research Initiation Grant and has been continuously funded
and Engineering and is a member of Tau Beta Pi. He received an invention award by NSF since 1982. Since 1991, he has received $1.8 million in funding from
for his work at SSC. NSF and upwards of $1 million in SBIR/industry/state funding. He has received
a Fulbright Research Award, the American Society of Engineering Education F.
E. Terman Award, three Sigma Xi Research Awards, the UTA Halliburton En-
gineering Research Award, the UTA University-Wide Distinguished Research
Award, the ARRI Patent Award, various Best Paper Awards, the IEEE Control
Systems Society Best Chapter Award (as Founding Chairman), and the National
Sigma Xi Award for Outstanding Chapter (as President). He was selected as En-
gineer of the Year in 1994 by the Fort Worth IEEE Section. He was appointed
to the NAE Committee on Space Station in 1995 and to the IEEE Control Sys-
tems Society Board of Governors in 1996. In 1998, he was selected as an IEEE
Control Systems Society Distinguished Lecturer. He is a Founding Member of
the Board of Governors of the Mediterranean Control Association.