Link

oping Studies in Science and Technology

Thesis No. 875

Flight Control Design

Using Backstepping
Ola Harkeg
ard

REGL

AU

ERTEKNIK

OL
TOM
ATIC CONTR

LINKPING

Division of Automatic Control

Department of Electrical Engineering
Linkopings universitet, SE581 83 Linkoping, Sweden
WWW: http://www.control.isy.liu.se
E-mail: [email protected]
Linkoping 2001

Flight Control Design Using Backstepping

c 2001 Ola H

arkeg
ard
Department of Electrical Engineering,
Linkopings universitet,
SE581 83 Link
oping,
Sweden.

ISBN 91-7219-995-4
ISSN 0280-7971
LiU-TEK-LIC-2001:12
Printed by UniTryck, Link
oping, Sweden 2001

To Eva

Abstract
Aircraft flight control design is traditionally based on linear control theory, due
to the existing wealth of tools for linear design and analysis. However, in order
to achieve tactical advantages, modern fighter aircraft strive towards performing
maneuvers outside the region where the dynamics of flight are linear, and the need
for nonlinear tools arises.
In this thesis we investigate backstepping as a new framework for nonlinear flight
control design. Backstepping is a recently developed design tool for constructing
globally stabilizing control laws for a certain class of nonlinear dynamic systems.
Flight control laws for two different control objectives are designed. First, general
purpose maneuvering is considered, where the angle of attack, the sideslip angle,
and the roll rate are the controlled variables. Second, automatic control of the
flight path angle control is considered.
The key idea of the backstepping designs is to benefit from the naturally stabilizing aerodynamic forces acting on the aircraft. The resulting state feedback control
laws thereby rely on less knowledge of these forces compared to control laws based
on feedback linearization, which today is the prevailing nonlinear design technique
within aircraft flight control.
The backstepping control laws are shown to be inverse optimal with respect to
meaningful cost functionals. This gives the controllers certain gain margins which
implies that stability is preserved for a certain amount of control surface saturation.
Also, the problem of handling a model error appearing at the input of a nonlinear dynamic system is treated, by considering the model error as an unknown,
additive disturbance. Two schemes, based on adaptive backstepping and nonlinear
observer design, are proposed for estimating and adapting to such a disturbance.
These are used to deal with model errors in the description of the aerodynamic
moments acting on the aircraft.
The designed control laws are evaluated using realistic aircraft simulation models and the results are highly encouraging.

Acknowledgments
First of all, I want to thank Professor Lennart Ljung for drafting me to the Automatic Control group in Link
oping, and hereby giving me the opportunity to perform
research within a most professional, ambitious, and inspiring group of people. I
also want to thank my supervisors Professor Torkel Glad and Karin St
ahl Gunnarsson for their guidance and expertise within nonlinear control theory and aircraft
control, respectively.
Besides these key persons in particular and the Automatic Control group in general, a few people deserve an explicit Thank you!: Fredrik Tjarnstr
om, Mikael
Norrl
of, and Jacob Roll proofread the thesis and provided valuable comments, significantly increasing the quality of the result. Anders Helmersson shared his practical flight experience and commented on the computer simulation results. Ingegerd
Skoglund and Mikael R
onnqvist at the Department of Mathematics suggested the
numerical schemes which were used for control allocation in the implementation of
the controllers.
And now to something completely different: Ett stort tack till sl
akt och v
anner,
och inte minst till k
arestan, som uth
ardat den senaste tiden d
a jag levt i ett socialt
vakuum, och som st
ottat mig i v
att och torrt!
This work was sponsored by the graduate school ECSEL.

Ola H
arkeg
ard
Link
oping, March 2001

iii

Contents

1 Introduction
1.1 Introductory Example: Sideslip Regulation . . . . . . . . . . . . . .
1.2 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Aircraft Primer
2.1 The Impact of Automatic Control . . . . . . . . .
2.2 Control Objectives . . . . . . . . . . . . . . . . . .
2.3 Control Means . . . . . . . . . . . . . . . . . . . .
2.4 Aircraft Dynamics . . . . . . . . . . . . . . . . . .
2.4.1 Governing physics . . . . . . . . . . . . . .
2.4.2 Modeling for control . . . . . . . . . . . . .
2.5 Current Approaches to Flight Control Design . . .
2.5.1 Gain-scheduling . . . . . . . . . . . . . . . .
2.5.2 Dynamic inversion (feedback linearization)
2.5.3 Other nonlinear approaches . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

1
2
4
4
7
7
9
12
12
13
16
20
20
21
24

2.A Wind-axes Force Equations . . . . . . . . . . . . . . . . . . . . . . . 25

v

vi

Contents

3 Backstepping
3.1 Lyapunov Theory . . . . . . . . . . . . . . . . . . . .
3.2 Lyapunov Based Control Design . . . . . . . . . . .
3.3 Backstepping . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Main result . . . . . . . . . . . . . . . . . . .
3.3.2 Which systems can be handled? . . . . . . . .
3.3.3 Which design choices are there? . . . . . . . .
3.4 Related Lyapunov Designs . . . . . . . . . . . . . . .
3.4.1 Forwarding . . . . . . . . . . . . . . . . . . .
3.4.2 Adaptive, robust, and observer backstepping
3.5 Applications of Backstepping . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

29
30
32
33
34
36
37
42
43
43
43

4 Inverse Optimal Control

4.1 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Inverse Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Robustness of Optimal Control . . . . . . . . . . . . . . . . . . . . .

45
45
46
48

5 Backstepping Designs for Flight Control

5.1 General Maneuvering . . . . . . . . . . . . . . .
5.1.1 Objectives, dynamics, and assumptions
5.1.2 Backstepping control design . . . . . . .
5.1.3 Flight control laws . . . . . . . . . . . .
5.1.4 Practical issues . . . . . . . . . . . . . .
5.2 Flight Path Angle Control . . . . . . . . . . . .
5.2.1 Objectives, dynamics, and assumptions
5.2.2 Backstepping control design . . . . . . .
5.2.3 Flight control law . . . . . . . . . . . .
5.2.4 Practical issues . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

51
51
52
54
62
65
65
66
67
72
72

6 Adapting to Input Nonlinearities and Uncertainties

6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Problem Formulation . . . . . . . . . . . . . . . . . . . .
6.3 Adaptive Backstepping . . . . . . . . . . . . . . . . . . .
6.4 Observer Based Adaption . . . . . . . . . . . . . . . . .
6.4.1 The general case . . . . . . . . . . . . . . . . . .
6.4.2 The optimal control case . . . . . . . . . . . . . .
6.5 A Water Tank Example . . . . . . . . . . . . . . . . . .
6.6 Application to Flight Control . . . . . . . . . . . . . . .
6.6.1 General maneuvering . . . . . . . . . . . . . . . .
6.6.2 Flight path angle control . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

75
75
76
78
79
79
81
82
86
86
86

7 Implementation and Simulation

7.1 Aircraft Simulation Models . . . . . . . . . . . . . . . . . . . . . . .
7.1.1 GAM/ADMIRE . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.2 HIRM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

87
87
87
88

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

Contents

7.2
7.3

7.1.3 FDC . . . . . . . . . .
Controller Implementation . .
7.2.1 Control allocation . .
Simulation . . . . . . . . . . .
7.3.1 Conditions . . . . . .
7.3.2 Controller parameters
7.3.3 Simulation results . .

vii

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

88
89
90
92
92
92
93

7.A Aircraft Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

8 Conclusions

101

Bibliography

103

Notation

Symbols
R
x1 , . . . , xn
x = (x1 xn )T
u
k(x)
V (x)
xref
1
xdes
1

e
e

the set of real numbers

state variables
state vector
control input
state feedback control law
Lyapunov function, control Lyapunov function
reference value of x1
desired value of x1
multiplicative control input perturbation
unknown additive input bias
estimate of e

Operators
p
kxk = x21 + . . . + x2n
V = dV
dt

Euclidian norm
time derivative of V

ix

Notation

V 0 (x1 ) = dVdx(x11 )
Vx (x) = ( Vx(x)

V (x)
xn )

derivative of V w.r.t. its only argument x1

gradient of V w.r.t. x

Acronyms
clf
GAS
NDI
TVC
GAM
HIRM

control Lyapunov function

global asymptotic stability, globally asymptotically
stable
nonlinear dynamic inversion
thrust vectored control
Generic Aerodata Model
High Incidence Research Model

Aircraft nomenclature
State variables
Symbol

= (, , )T

= (p, q, r)T
s = (ps , qs , rs )T
p
q
r
p = (pN , pE , h)T
pN
pE
h
V = (u, v, w)T
u
v
w
VT
M
nz

Unit
rad
rad
rad
rad
rad
rad

rad/s
rad/s
rad/s
m
m
m
m/s
m/s
m/s
m/s
g

Definition
angle of attack
sideslip angle
flight path angle
aircraft orientation (Euler angles)
roll angle
pitch angle
yaw angle
body-axes angular velocity
stability-axes angular velocity
roll rate
pitch rate
yaw rate
aircraft position
position north
position east
altitude
body-axes velocity vector
longitudinal velocity
lateral velocity
normal velocity
total velocity
Mach number
normal acceleration, load factor

Notation

xi

Control surface deflections

Symbol

es
ed
cs
cd
r

Unit
rad
rad
rad
rad
rad

Definition
collective representation of all control surfaces
symmetrical elevon deflection
differential elevon deflection
symmetrical canard deflection
differential canard deflection
rudder deflection

Unit
kg

Definition
aircraft mass

kg m2

aircraft inertial matrix

m2
m
m
m

wing planform area

wing span
mean aerodynamic chord
zb -position of engine thrust point

Unit
kg/m3
N/m2

Definition
air density
dynamic pressure

Unit
m/s2
N
N
N
N
Nm
Nm
Nm

Definition
acceleration due to gravity
engine thrust force
drag force
lift force
side force
rolling moment
pitching moment
yawing moment

Aircraft data
Symbol
m
Ix
I= 0
Ixz
S
b
c
ZT P

0
Iy
0

Ixz
0
Iz

Atmosphere
Symbol

Forces and moments

Symbol
g
FT
D = qSCD
L = qSCL
Y = qSCY
= qSbCl
L
M = qS
cCm
N = qSbCn

Coordinate systems
Symbol
(xb , yb , zb )
(xs , ys , zs )
(xw , yw , zw )

Definition
body-axes coordinate system
stability-axes coordinate system
wind-axes coordinate system

xii

Notation

1
Introduction

During the past 15 years, several new design methods for control of nonlinear dynamic systems have been invented. One of these methods is known as backstepping.
Backstepping allows a designer to methodically construct stabilizing control laws
for a certain class of nonlinear systems.
Parallel to this development within nonlinear control theory, one finds a desire
within aircraft technology to push the performance limits of fighter aircraft towards
supermaneuverability. By utilizing high angles of attack, tactical advantages
can be achieved, as demonstrated by Herbst [32] and Well et al. [77], who consider
aircraft reversal maneuvers for performance evaluation. The aim is for the aircraft
to return to a point of departure at the same speed and altitude but with an
opposite heading at minimum time. It is shown that using high angles of attack
during the turn, the aircraft is able to maneuver in less air space and complete the
maneuver in shorter time. These types of maneuvers are performed outside the
region where the dynamics of flight are linear. Thus, linear control design tools,
traditionally used for flight control design, are no longer sufficient.
In this thesis, we investigate how backstepping can be used for flight control
design to achieve stability over the entire flight envelope. Control laws for a number
of flight control objectives are derived and their properties are investigated, to see
what the possible benefits of using backstepping are. Let us begin by illustrating
the key ideas of the design methodology with a concrete example.
1

Introduction

Y ()

x 10

1.5
1

Y (N)

0.5

rudder, r
VT

0
0.5
1

FT

1.5
2
20

(a) Top view of aircraft

10

0
(deg)

10

20

(b) Side force vs. sideslip

Figure 1.1 The sideslip, , is in general desired to be kept zero. The aerodynamic side force, Y (), naturally acts to fulfill this objective.

1.1

Introductory Example: Sideslip Regulation

An important aircraft variable to be controlled is the sideslip angle, , which is

depicted in Figure 1.1(a). Nonzero sideslip means that the aircraft is skidding
through the air. This is unwanted for several reasons: it is unsuitable for the
engines, especially at high speeds, it increases the air resistance of the aircraft, and
it is uncomfortable for the pilot. Thus, = 0 is in general the control objective.
To regulate the sideslip, the rudder, situated at the back of the aircraft, is used.
The (somewhat simplified) sideslip dynamics are given by
1
(Y () FT sin )
mVT
r = cN (r )

= r +

(1.1a)
(1.1b)

Here, r = yaw rate of the aircraft, m = aircraft mass, VT = aircraft speed, Y =

aerodynamic side force due to the air resistance, FT = thrust force produced by
the engines, c = constant related to the aircraft moment of inertia, N = yawing
moment, and r = rudder deflection.
The side force, which is the component of the air resistance that affects the
sideslip, is a nonlinear function of . Figure 1.1(b) shows the typical relationship
between the two entities. In particular we see that for a negative sideslip, the side
force is positive and vice versa for a positive sideslip. Thus, considering (1.1a),
the side force is useful for bringing the sideslip to zero since, e.g., for < 0 its
contribution to is positive. The same is true for the component due to the
engine thrust force. For < 0, FT sin is positive, bringing towards zero.

1.1 Introductory Example: Sideslip Regulation

Pilot
inputs Backstepping

k(x)

control laws
(Ch. 5)

udes

Control
allocation

(Ch. 7)

(Ch. 2)
u

e
Bias
estimation
(Ch. 6)

Aircraft state, x
Figure 1.2 Controller configuration.

The important implication of these discoveries is that since the nonlinear terms
act stabilizing, they need not be cancelled by the controller, and hence complete
knowledge of them is not necessary. This idea of recognizing useful nonlinearities
and benefitting from them rather than cancelling them is the main idea of this
thesis.
In Chapter 5 we use backstepping to design state feedback laws for various flight
control objectives. A common feature of the derived control laws is that they are
linear in the variables used for feedback, considering the angular acceleration of
the aircraft as the input. In our example this corresponds to designing a control
law
u = k(, r) = k1 k2 r
where
u = r = cN (r )

(1.2)

To realize such a control law in terms of the true control input, r , requires
perfect knowledge of the yawing moment, N . Since typically this is not the case,
we can remodel (1.2) as
(r ) + e
u = cN
is our model of the yawing moment and e is the bias to the actual yawing
where N
moment. An intuitively appealing idea is to compute an estimate, e, of the bias,
e, on-line and realize the modified control law
r ) = k(, r) e
cN(

(1.3)

Introduction

which, if the estimate is perfect, cancels the effect of the bias and achieves u =
k(, r) as desired. In Chapter 6, two such estimation schemes are proposed and
shown to give closed loop stability.
The remaining problem is that of control allocation, i.e., how to determine r
such that (1.3) is satisfied. This will be discussed in Chapter 7 where two numerical
solvers are proposed.
The overall control configuration, along with chapter references, is shown in
Figure 1.2.

1.2

Outline of the Thesis

Chapter 2 Contains basic facts about modern fighter aircraft such as their dynamics, the control inputs available and what the control objectives are.
Chapter 3 Introduces the backstepping methodology for control design for a class
of nonlinear system. Discusses design choices and contains examples showing
how some nonlinearities can actually be useful and how to benefit from them.
Chapter 4 A short chapter on inverse optimal control, i.e., how one can decide
whether a given control law is optimal w.r.t. a meaningful cost functional.
Chapter 5 Contains the main contributions of the thesis. Backstepping is used
to design state feedback control laws for various flight control objectives.
Chapter 6 Proposes two different methods for adapting to model errors appearing
at the input and investigates closed loop stability in each case.
Chapter 7 Proposes numerical schemes for solving the control allocation problem.
Also presents computer simulations of the designed aircraft control laws in
action.
Chapter 8 Concludes the thesis by evaluating the ability to handle important
issues like stability, tuning, robustness, input saturation, and disturbance
attenuation within the proposed backstepping framework.

1.3

Contributions

The main contributions of this thesis are the following:

The ideas in Sections 5.1.2 and 5.2.2 on how to benefit from the naturally stabilizing aerodynamic forces using backstepping, rather than cancelling them
as in feedback linearization. The resulting control laws rely on less knowledge
of the aerodynamic forces than the feedback linearizing designs which have
been previously proposed (reviewed in Section 2.5.2).
The backstepping designs for the two general nonlinear systems in Sections
5.1.2 and 5.2.2.

1.3 Contributions

The discovery in Section 5.1.3 that the angle of attack control law used by
Snell et al. [71], based on feedback linearization and time-scale separation,
can also be constructed using backstepping and is in fact optimal w.r.t. a
meaningful cost functional.
The adaptive schemes in Chapter 6 for handling model errors appearing at
the control input.
The computer simulations in Section 7.3 showing the proposed control laws
to work satisfactory using realistic aircraft simulation models.
Parts of this thesis have been published previously. The backstepping designs
in Chapter 5 originate from
Ola H
arkeg
ard and S. Torkel Glad. A backstepping design for flight
path angle control. In Proceedings of the 39th Conference on Decision
and Control, pages 35703575, Sydney, Australia, December 2000.
and
Ola H
arkeg
ard and S. Torkel Glad. Flight control design using backstepping. Technical Report LiTH-ISY-R-2323, Department of Electrical
Engineering, Link
opings universitet, SE-581 83 Link
oping, Sweden, December 2000. To be presented at the 5th IFAC Symposium Nonlinear
Control Systems (NOLCOS01), St. Petersburg, Russia.
The results in Chapter 6 on how to adapt to a model error at the input can be
found in
Ola H
arkeg
ard and S. Torkel Glad. Control of systems with input nonlinearities and uncertainties: an adaptive approach. Technical Report
LiTH-ISY-R-2302, Department of Electrical Engineering, Linkopings
universitet, SE-581 83 Link
oping, Sweden, October 2000. Submitted to
the European Control Conference, ECC 2001, Porto, Portugal.

Introduction

2
Aircraft Primer

In this chapter, we investigate modern fighter aircraft from a control perspective.

The aim is to introduce the reader to the aircraft dynamics, the control objectives,
and how to assist the pilot to achieve these objectives using automatic control.
Much of what is said applies to aircraft in general, not only to fighters.
Section 2.1 discusses the possibilities of using electric control systems to aid the
pilot in controlling the aircraft. In Sections 2.2 and 2.3, the control objectives are
presented along with the control means at our disposal. In Section 2.4 we turn to
the dynamics of flight, and in Section 2.5 some existing approaches to flight control
design are presented.

2.1

The Impact of Automatic Control

The interplay between automatic control and manned flight goes back a long time,
see Stevens and Lewis [74] for a historic overview. At many occasions their paths
have crossed, and progress in one field has provided stimuli to the other.
During the early years of flight technology, the pilot was in direct control of the
aircraft control surfaces. These where mechanically connected to the pilots manual
inputs. In modern aircraft, the pilot inputs are instead fed to a control system.
Based on the pilot inputs and available sensor information, the control system
computes the control surface deflections to be produced. This information is sent
7

Aircraft Primer

Manual inputs
Mission
Pilot

Actuator settings

Control
System

Aircraft behavior

Aircraft

Sensor data

Cockpit displays, visual information, etc.

Figure 2.1 Fly-by-wire from a control perspective.

through electrical wires to the actuators located at the control surfaces, which in
turn realize the desired deflections. Figure 2.1 shows the situation at hand. This
is known as fly-by-wire technology. What are the benefits of this approach?
Stability Due to the spectacular 1993 mishap, when a fighter aircraft crashed
over central Stockholm during a flight show, it is a widely known fact, even to people
outside the automatic control community, that many modern aircraft are designed
to be unstable in certain modes. A small disturbance would cause the uncontrolled
aircraft to quickly diverge from its original position. Such a design is motivated by
the fact that it enables faster maneuvering and enhanced performance. However,
it also emphasizes the need for reliable control systems, stabilizing the aircraft for
the pilot.
Varying dynamics The aircraft dynamics vary with altitude and speed. Thus
without a control system assisting him, the pilot himself would have to adjust his
joystick inputs to get the same aircraft response at different altitudes and speeds.
By hiding the true aircraft dynamics inside a control loop, as in Figure 2.1, the
varying dynamics can be made transparent to the pilot by designing the control
system to make the closed loop dynamics independent of altitude and speed.
Aircraft response Using a control system, the aircraft response to the manual
inputs can be selected to fulfill the requirements of the pilots. By adjusting the
control law, the aircraft behavior can be tuned much more easily than having to
adjust the aircraft design itself to achieve, e.g., a certain rise time or overshoot.
Interpretation of pilot inputs By passing the pilot inputs to a control system,
the meaning of the inputs can be altered. In one mode, moving the joystick sideways
may control the roll rate, in another mode, it may control the roll angle. This

2.2 Control Objectives

nz

VT
Figure 2.2 Pitch control objectives.

paves the way for various autopilot capabilities, e.g., altitude hold, relieving the
pilot workload.

2.2

Control Objectives

Given the possibilities using a flight control system, what does the pilot want to
control?
In a classical dogfight, whose importance is still recognized, maneuverability is
the prime objective. Here, the normal acceleration, nz , or the pitch rate, q, make
up suitable controlled variables in the longitudinal direction, see Figure 2.2. nz ,
also known as the load factor, is the acceleration experienced by the pilot directed

10

Aircraft Primer

yb
xb

zb

ps

xs

VT
xw
Figure 2.3 Lateral control objectives and coordinate systems definitions.
In the figure, and are both positive.

along his spine. It is expressed as a multiple of the gravitational acceleration,

g. nz is closely coupled to the angle of attack, , which appears naturally in
the equations describing the aircraft dynamics, see Section 2.4. Therefore, angle
of attack command control designs are also common, in particular for nonlinear
approaches.
In the lateral direction, roll rate and sideslip command control systems are
prevalent. The sideslip angle, , is depicted in Figure 2.3. Typically, = 0 is
desired so that the aircraft is flying straight into the wind with a zero velocity
component along the body y-axis, yb . However, there are occasions where a certain
sideslip is necessary, e.g., when landing the aircraft in the presence of side wind.
For the roll rate command system, a choice must be made regarding which axis
to roll about, see Figure 2.3. Let us first consider the perhaps most obvious choice,

2.3 Control Means

11

Elevons, e

Rudder, r

Leading-edge flaps
Canard wings, c

TVC

Figure 2.4 A modern fighter aircraft configuration.

namely the body x-axis, xb . Considering a 90 degrees roll, we realize that the initial
angle of attack will turn into pure sideslip at the end of the roll and vice versa.
At high angles of attack this is not tolerable, since the largest acceptable amount
of sideslip during a roll is in the order of 35 degrees [14]. To remove this effect,
we could instead roll about the wind x-axis, xw . Then and remain unchanged
during a roll. This is known as a velocity-vector roll. With the usual assumption
that a roll is performed at zero sideslip, this is equivalent to a stability-axis roll,
performed about the stability x-axis, xs . In this case, the angular velocity ps is the
variable to control.
There also exist situations where other control objectives are of interest. Autopilot functions like altitude, heading, and speed hold are vital to assist the pilot
during long distance flight. For firing on-board weapons, the orientation of the
aircraft is crucial. To benefit from the drag reduction that can be accomplished
during close formation flight, the position of the wingman relative to the leader
must be controlled precisely, preferably automatically to relieve the workload of
the wingman pilot [26]. Also, to automate landing the aircraft it may be of interest
to control its descent through the flight path angle, , see Figure 2.7.

12

2.3

Aircraft Primer

Control Means

To accomplish the control tasks of the previous section, the aircraft must be
equipped with actuators providing ways to control the different motions. Figure
2.4 shows a modern fighter aircraft configuration.
Pitch control, i.e., control of the longitudinal motion, is provided by deflecting
the elevons and the canard wings symmetrically (right and left control surfaces
deflect in the same direction). Conversely, roll control is provided by deflecting the
elevons, and possibly also the canard wings, differentially (right and left control
surfaces deflect in the opposite directions). Therefore, it is natural to introduce
the control inputs
elef t + eright
2
elef t eright
=
2
clef t + cright
=
2
clef t cright
=
2

es =

symmetrical elevon deflection

ed

differential elevon deflection

cs
cd

symmetrical canard deflection

differential canard deflection

Yaw control, i.e., control of the rotation about the body z-axis, is provided by the
rudder. The leading-edge flaps can be used, e.g., to minimize the drag.
Recently, the interest in high angle of attack flight has led to the invention of
thrust vectored control (TVC). Deflectable vanes are then mounted at the engine
exhaust so that the engine thrust can be directed to produce a force in some desired
direction. The NASA High Angle-of-Attack Research Vehicle (HARV) [30] uses this
technology.
When convenient, we will let represent all the above control surface deflections.
Finally, the aircraft speed, or rather the engine thrust force, is governed by the
engine throttle setting.

2.4

Aircraft Dynamics

We now turn to the aircraft dynamics, and present the governing equations that tie
the variables to be controlled to the control inputs available to us. The presentation,
based on the books by Stevens and Lewis [74] and Boiffier [7], is focused on arriving
at a model suitable for control design, consisting of a set of first order differential
equations. For a deeper insight into the mechanics and aerodynamics behind the
model, the reader is referred to the aforementioned books or, e.g., Etkin and Reid
[16], McLean [54], or Nelson [57].

2.4 Aircraft Dynamics

2.4.1

13

Governing physics

We will use the assumptions that Earth is flat and fixed, and that the aircraft body
is rigid (as opposed to flexible). This yields a 6 degrees of freedom model (rotation
and translation in 3 dimensions). The dynamics can be described by a state space
model with 12 states consisting of
p = pN p E
nate system;
V= u v
system;

T

T

, the aircraft position expressed in an Earth-fixed coordi-

, the velocity vector expressed in the body-axis coordinate

T
= , the Euler angles describing the orientation of the aircraft
relative to the Earth-fixed coordinate system;

T
= p q r , the angular velocity of the aircraft expressed in the bodyaxes coordinate system.
The task of controlling the aircraft position p is typically left entirely to the pilot,
formation flight being a possible exception. The only coupling from p to the other
state variables is through the altitude dependence of the aerodynamic pressure
(2.4). Since the altitude varies slower than the rest of the variables, it can be
regarded as a constant during the control design. Therefore the position dynamics
will be left out here.
The equations governing the remaining three state vectors can be compactly
written as
+ V)
F = m(V
M = I + I
= E()

force equation
moment equation

(2.1)
(2.2)

attitude equation

(2.3)

where

1 sin tan
cos
E() = 0
0 sin / cos

cos tan
sin
cos / cos

m is the aircraft mass and I is the aircraft inertial matrix. The force and moment
equations follow from applying Newtons second law and the attitude equation
spurs from the relation between the Earth-fixed and the body-fixed coordinate
systems.
F and M represent the sum of the forces and moments, respectively, acting on
the aircraft at the center of gravity. These forces and moments spring from three
major sources, namely

14

Aircraft Primer

gravity,
engine thrust, and
aerodynamic efforts.
Introducing
F = FG + FE + FA
M = ME + MA
we will now investigate each of these components and express them in the bodyfixed coordinate system.
Gravity
Gravity only gives a force contribution since it acts at the aircraft center of gravity. The gravitational force, mg, directed along the normal of the Earth plane, is
considered constant over the altitude envelope. This yields

sin
FG = mg sin cos
cos cos
Engine thrust
The thrust force produced by the engine is denoted by FT . Assuming the engine
to be positioned so that the thrust acts parallel to the aircraft body x-axis (not
using TVC) yields

FT
FE = 0
0
Also assuming the engine to be mounted so that the thrust point lies in the bodyaxes xz-plane, offset from the center of gravity by ZT P in the body-axes z-direction
results in

0
ME = FT ZT P
0
Aerodynamic efforts
The aerodynamic forces and moments, or the aerodynamic efforts for short, are
due to the interaction between the aircraft body and the incoming airflow. The
size and direction of the aerodynamic efforts are determined by the amount of air
diverted by the aircraft in different directions (see [3] for an enlightening discussion
on various explanations to aerodynamic lift). The amount of air diverted by the
aircraft is mainly decided by

2.4 Aircraft Dynamics

15

the speed and density of the airflow (VT , );

the geometry of the aircraft (, S, c, b);
the orientation of the aircraft relative to the airflow (, ).
The aerodynamic efforts also depend on other variables, like the angular rates (p,
but these effects
q, r) and the time derivatives of the aerodynamic angles (,
),
are not as prominent.
This motivates the standard way of modeling aerodynamic forces and moments:
M, . . . )
,
Force = qSCF (, , , p, q, r, ,
M, . . . )
Moment = qSlCM (, , , p, q, r, ,
,
The aerodynamic pressure,
q =

1
(h)VT2
2

(2.4)

captures the density dependence and most of the speed dependence, S is the aircraft
wing area, and l refers to the length of the lever arm connected to the moment.
CF and CM are known as aerodynamic coefficients. These are difficult to model
analytically but can be estimated empirically through wind tunnel experiments
and actual flight tests. Typically, each coefficient is written as the sum of several
components, each capturing the dependence of one or more of the variables above.
These components can be represented in several ways. A common approach is to
store them in look-up tables and use interpolation to compute intermediate values.
In other approaches one tries to fit the data to some parameterized function.
In the body-axes coordinate system, we have the expressions

X
X = qSCx

FA = Y where
Y = qSCy
Z = qSCz
Z

= qSbCl
L
rolling moment
L

pitching moment
cCm
MA = M where M = qS
yawing moment
N = qSbCn
N
These are illustrated in Figure 2.5. The aerodynamic forces are also commonly
expressed in the wind-axes coordinate system (related to the body-fixed coordinate
system as indicated in Figure 2.3) where we have that

D
drag force
D = qSCD

side force
Y = qSCY
(2.5)
FA,w = Y where
lift force
L = qSCL
L
where the lift and side force coefficient, CL and CY , mainly depend on and
respectively.

16

Aircraft Primer

Y
M

L
Z

Figure 2.5 Aerodynamics forces and moments in the body-axes coordinate

system.

Essentially, only the aerodynamic moments are affected when a control surface
is deflected. This is a key feature without which some nonlinear control design
methods, including backstepping and dynamic inversion, would not be applicable.
Figure 2.6 shows the lift force and pitching moment coefficients, CL and Cm , as
functions of angle of attack and symmetrical elevon deflection. The aerodata comes
from the HIRM model [56].
In Section 2.2, the normal acceleration nz was introduced. We now have the
setup to define nz more precisely and find its relationship to . We have that
nz =

qSCz (, , , . . . )
Z
=
mg
mg

Given an nz command, the above equation can be used to solve for a corresponding
command.

2.4.2

Modeling for control

We will now collect the equations from the previous section and write the result in a
form suitable for control design, namely as a system of first order scalar differential

2.4 Aircraft Dynamics

17

0.4

1.5
0.2
1
Cm

CL

0.5

0
0.2
0.5
1
40
30
20
10

0.4
40
30
20
10

0
es (deg) 10 10 0

10

20

30

40

50

(deg)

0
es (deg) 10 10 0

10

20

30

50

40

(deg)

0.5
0.4

1.5

0.3
1
0.2
0.1
Cm

0.5
0

0.1

0.5

0.2
1
0.3
1.5
2
10

0.4
0

10

20
30
(deg)

40

50

0.5
10

10

20
30
(deg)

40

50

Figure 2.6 The aerodynamic coefficients CL and Cm as functions of

and es . In the lower figures, the shaded area represents the
dependence on es . Apparently, elevon deflections primarily
produce aerodynamic moments rather than forces.

18

Aircraft Primer

equations. Expanding Equations (2.1)(2.3) yields

$

'
Force equations (body-axes)
1
u = rv qw g sin + (X + FT )
m
1
v = pw ru + g sin cos + Y
m
1
w = qu pv + g cos cos + Z
m

Moment equations (body-axes)

+ c4 N
p = (c1 r + c2 p)q + c3 L

(2.6a)

q = c5 pr c6 (p r ) + c7 (M + FT ZT P )
+ c9 N
r = (c8 p c2 r)q + c4 L

(2.6b)
(2.6c)

Attitude equations (body-axes)

= p + tan (q sin + r cos )
= q cos r sin
q sin + r cos
=
cos

&

Here we have introduced

2
c1 = (Iy Iz )Iz Ixz

c4 = Ixz
c7 =

1
Iy

c2 = (Ix Iy + Iz )Ixz
Iz Ix
c5 =
Iy

c3 = Iz
Ixz
c6 =
Iy

2
c8 = Ix (Ix Iy ) + Ixz

c9 = Ix

where

Ix
I= 0
Ixz

0
Iy
0

Ixz
0 ,
Iz

2
= Ix Iz Ixz

In Section 2.2, , , and VT were mentioned as suitable controlled variables.

We can rewrite the force equations in terms of these variables by performing the

2.4 Aircraft Dynamics

19

following change of variables:

p
u2 + v 2 + w 2
w
= arctan
u
v
= arcsin
VT

VT =

This gives us
'

Force equations (wind-axes)

1
V T = (D + FT cos cos + mg1 )
m
1
(L FT sin + mg2 )
= q (p cos + r sin ) tan +
mVT cos
1
(Y FT cos sin + mg3 )
= p sin r cos +
mVT
&

(2.7a)
(2.7b)
(2.7c)
%

where the contributions due to gravity are given by

g1 = g( cos cos sin + sin cos sin + sin cos cos cos )
g2 = g(cos cos cos + sin sin )

(2.8)

g3 = g(cos cos sin + sin cos sin sin sin cos cos )
See Appendix 2.A for a complete derivation. A pleasant effect of this reformulation
is that the nonlinear aerodynamic forces L and Y mainly depend on and ,
respectively. This fact will be used for control design using backstepping.
A very common approach to flight control design is to control longitudinal
motions (motions in the body xz-plane) and lateral motions (all other motions)
separately. With no lateral motions, the longitudinal equations of motion become
'
$

&

Longitudinal equations
1
V T = (D + FT cos mg sin )
m
1
(L FT sin + mg cos )
= q +
mVT
= q
1
q = (M + FT ZT P )
Iy

(2.9a)
(2.9b)
(2.9c)
(2.9d)
%

Here, = is the flight path angle determining the direction of the velocity
vector, as depicted in Figure 2.7.

20

Aircraft Primer

Figure 2.7 Illustration of the longitudinal aircraft entities.

2.5

Current Approaches to Flight Control Design

In this section we survey some of the proposed design schemes with the emphasis
on nonlinear control designs. Flight control design surveys can also be found in
Magni et al. [53] and Huang and Knowles [35].

2.5.1

Gain-scheduling

The prevailing flight control design methodology of today is based on gain-scheduling.

The flight envelope that one wants to conquer is partitioned into smaller regions.
For each region, a steady state operating point is chosen around which the dynamics are linearized. Linear control tools can then be used to design one control law
for each operating point. Between the operating points, the controllers are blended
together using gain-scheduling to make the transitions between different regions
smooth and transparent to the pilot.
Dynamic pressure or Mach number and altitude together are typical scheduling
variables. Note that if the closed loop behavior is designed to be the same throughout the envelope, gain-scheduling can be seen as a way of cancelling the nonlinear
behavior due to variations in the scheduling variables. However, the nonlinear effects due to high angles of attack or high angular rates are not dealt with. Whether
or not closed loop stability holds when the aircraft state is far from the operating
point at which the linearization was performed, depends on the robustness of the
control design.
Let us summarize the pros and cons of gain-scheduling.

2.5 Current Approaches to Flight Control Design

21

+ Using several linear models to describe the aircraft dynamics allows the control designer to utilize all the classical tools for control design and robustness
and disturbance analysis.
+ The methodology has proven to work well in practice. The Swedish fighter
JAS 39 Gripen [63] is a flying proof of its success.
The outlined divide-and-conquer approach is rather tedious since for each
region, a controller must be designed. The number of regions may be over
50.
Only the nonlinear system behavior in speed and altitude is considered. Stability is therefore guaranteed only for low angles of attack and low angular
rates.

2.5.2

Dynamic inversion (feedback linearization)

The idea behind gain-scheduling was to provide the pilot with a similar aircraft
response irrespectively of the aircraft speed and altitude. This philosophy is even
more pronounced in nonlinear dynamic inversion (NDI), which is the term used
in the aircraft community for what is known as feedback linearization in control
theory. In this thesis, we only deal with feedback linearization through examples
and intuitive explanations. For an introduction to feedback linearization theory,
the reader is referred to, e.g., Slotine and Li [70] or Isidori [36].
Using dynamic inversion, as the name implies, the natural aircraft dynamics are
inverted and replaced by the desired linear ones through the wonders of feedback.
This includes the nonlinear behavior in speed and altitude as well as the nonlinear
effects at high angles of attack and high angular rates.
To make things more concrete, consider the simplified angle of attack dynamics
(cf. (2.9b), (2.9d))
= q
q =

1
L()
mVT

1
M (, , q)
Iy

(2.10a)
(2.10b)

The speed is assumed to vary much slower than and q so that V T 0 is a good
approximation. Now, introduce
z=q

1
L()
mVT

and compute z:

z =

1
1
M (, , q)
L0 ()z
Iy
mVT

(2.11)

22

Aircraft Primer

ref

u
k0

NDI
(2.12)

, q
Aircraft

Choord.
change

, z

(k1 k2 )
Figure 2.8 Illustration of a dynamic inversion control law. The inner feedback loop cancels the nonlinear dynamics making the dashed
box a linear system, which is controlled by the outer feedback
loop.

Introducing
u=

1
1
M (, , q)
L0 ()z
Iy
mVT

(2.12)

we can now rewrite (2.10) as a chain of two integrators:

= z
z = u
Through this transformation, we have cast a nonlinear control problem onto a
linear one. Again, linear control tools can be used to find a control law
 


ref
(2.13)
u = k0 k1 k2
z
giving a satisfactory response from the commanded value ref to . Using (2.11)
and (2.12) we can solve for the true control input , given that (2.12) is invertible
w.r.t. . Invertability is lost, e.g., when the control surfaces saturate.
Equation (2.12) can be interpreted as an inner control loop cancelling the nonlinear behavior while (2.13) is an outer control loop providing the system with the
desired closed loop characteristics, see Figure 2.8.
There exists a large number of publications devoted to flight control design
using feedback linearization. The 1988 paper of Lane and Stengel [48] is an early
contribution on the subject containing designs for various control objectives.
Enns et al. [15] outline dynamic inversion as the successor to gain-scheduling
as the prevailing design paradigm. This bold statement is supported by discussions regarding the ability to deal with, e.g., handling quality specifications, disturbance attenuation, robustness, and control surface allocation in a dynamic inversion
framework.

2.5 Current Approaches to Flight Control Design

23

Robustness analysis has often been pointed out as the Achilles heel of dynamic
inversion. Dynamic inversion relies on the complete knowledge of the nonlinear
plant dynamics. This includes knowledge of the aerodynamic efforts, which in
practice comes with an uncertainty in the order of 10%. What happens if the true
nonlinearities are not completely cancelled by the controller? [78] contains some
results regarding this issue.
One way of enhancing the robustness is to reduce the control law dependence on
the aerodynamic coefficients. Note that computing from (2.12) requires knowledge of the aerodynamic coefficients Cm and CL as well as dCL /d (recall from
(2.5) that L = qSCL ). Snell et al. [71] propose a dynamic inversion design which
does not involve dCL /d, thus making the design more robust. The idea is to use
time-scale separation and design separate controllers for the - and q-subsystems
of (2.10). Inspired by singular perturbation theory [42] and cascaded control design, the system is considered to consist of slow dynamics (2.10a) and fast dynamics
(2.10b). First, the slow dynamics are controlled. Assume the desired slow dynamics
to be
= k1 ( ref ),

k1 > 0

Then the angle of attack command ref can be mapped onto a pitch rate command
q ref = k1 ( ref ) +

1
L()
mVT

(2.14)

We now turn to the fast dynamics and determine a control law rendering the fast
dynamics
q = k2 (q q ref ),

k2 > 0

This can be achieved by solving

1
1
M (, , q) = k2 (q q ref ) = k2 (q + k1 ( ref )
L())
Iy
mVT

(2.15)

for , which obviously only requires knowledge of Cm and CL . The controller

structure is shown in Figure 2.9.
The remaining question to be answered is whether this time-scale separation
approach indeed is valid is the closed loop system guaranteed to be stable? Snell
et al. [71] use linear pole placement arguments to confirm stability. Interestingly
enough, in Section 5.1.3 we will show that the above design is a special case of a
general backstepping design, and stability can be shown using Lyapunov theory. We
also show that using backstepping, the knowledge of the aerodynamic coefficients
required to guarantee stability can be further reduced.
Further efforts to robustify a dynamic inversion design against model errors can
be found in Adams et al. [1] and Reiner et al. [60, 61]. Here, the idea is to enhance
robustness by using a synthesis controller in the outer, linear loop in Figure 2.8.
Let us summarize the pros and cons of dynamic inversion.

24

Aircraft Primer

ref

Outer ctl.
(2.14)

q ref

Inner ctl.
(2.15)

q = . . .
(2.10b)

= . . .
(2.10a)

Figure 2.9 Cascaded dynamic inversion control design based on time-scale

separation.

+ One single controller is used throughout the whole flight regime.

+ Stability is guaranteed even for high angles of attack, provided that the model
is accurate at those angles.
+ Closed loop performance can be tuned using linear tools.
Dynamic inversion relies on precise knowledge of the aerodynamic coefficients
to completely cancel the nonlinear dynamics.

2.5.3

Other nonlinear approaches

In addition to dynamic inversion, many other nonlinear approaches have been applied to flight control design. Garrard et al. [23] formulate the angle of attack control problem as a linear quadratic optimization problem. As their aircraft model
is nonlinear, in order to capture the behavior at high angles of attack, the arising
Hamilton-Jacobi-Bellman equation is difficult to solve exactly. The authors settle
for a truncated solution to the HJB equation.
Mudge and Patton [55] consider the problem of pitch pointing, where the objective is to command the pitch angle while the flight path angle remains
unchanged. Eigenstructure assignment is used to achieve the desired decoupling
and sliding mode behavior is added for enhanced robustness.
Other approaches deal with the problem of tracking a reference signal, whose
future values are also known. Levine [50] shows that an aircraft is differentially flat
if the outputs are properly chosen. This is used to design an autopilot for making
the aircraft follow a given trajectory.
Hauser and Jadbabaie [31] design receding horizon control laws for unmanned
combat aerial vehicles performing aggressive maneuvers. Over a receding horizon,
the aircraft trajectory following properties are optimized on-line. The control laws
are implemented and evaluated using the ducted fan at Caltech.

Appendix
2.A

Wind-axes Force Equations

This appendix contains the details of the conversion of the aircraft force equation
from the body-axes to the wind-axes coordinate system. The result is a standard
one, but the derivation, which establishes the relationship between the forces used
in the two different representations, is rarely found in textbooks on flight control.
The body-axes force equations are
1
(X + FT )
m
1
v = pw ru + g sin cos + Y
m
1
w = qu pv + g cos cos + Z
m
u = rv qw g sin +

The relation between the variables used in the two coordinate systems is given by
w
u
v
= arcsin
VT
p
VT = u2 + v 2 + w2
= arctan

u = VT cos cos

v = VT sin
w = VT sin cos

Differentiating we get
uu + v v + ww
1
1
V T =
=
u(rv qw g sin + (X + FT ))
VT
VT
m
1

1
+ v(pw ru + g sin cos + Y ) + w(qu pv + g cos cos + Z)
m
m
= g( cos cos sin + sin sin cos + sin cos cos cos )
|
{z
}
g1

1
+ (FT cos cos + X cos cos + Y sin + Z sin cos )
|
{z
}
m
D

25

26

Aircraft Primer

uw wu
VT cos (w cos u sin )
=
2
2
u +w
VT2 cos2
1
1
(qVT cos cos pVT sin + g cos cos + Z) cos
=
VT cos
m


1
(rVT sin qVT sin cos g sin + (X + FT )) sin
m
= q tan (p cos + r sin )
1
(m g(cos cos cos sin sin ) FT sin + Z cos X sin )
+
{z
}
|
|
{z
}
mVT cos

g2

v(u
2 + w2 ) v(uu + ww)
vV
T v V T

=
=
2
3
VT cos
VT cos
vV
T2 cos2 VT2 sin cos (u cos + w sin )
VT3 cos
1
1
((pw ru + g sin cos + Y ) cos
=
VT
m
1
(rv qw g sin + (X + FT )) cos sin
m
1
(qu pv + g cos cos + Z) sin sin )
m
= p(sin cos2 + sin sin2 ) r(cos cos2 + cos sin2 )
|
|
{z
}
{z
}
=

cos

sin

1
(m g(cos sin cos + cos sin sin sin sin cos cos )
+
|
{z
}
mVT
g3

FT cos sin X cos sin + Y cos Z sin sin )

|
{z
}
Y

Summing up, we have the following transformed system of equations.

1
V T = (D + FT cos cos + mg1 )
m
= q tan (p cos + r sin ) +
= p sin r cos +

1
(L FT sin + mg2 )
mVT cos

1
(Y FT cos sin + mg3 )
mVT

The orientation dependent gravitational components are

g1 = g( cos cos sin + sin cos sin + sin cos cos cos )
g2 = g(cos cos cos + sin sin )
g3 = g(cos cos sin + sin cos sin sin sin cos cos )

2.A Wind-axes Force Equations

27

The relationship between the aerodynamic forces expressed in the two coordinate
systems is given by
D = X cos cos Y sin Z sin cos
L = X sin Z cos
Y = X cos sin + Y cos Z sin sin
These equations are relevant since often the available aerodata relates to the bodyaxes system while it is the wind-axes forces that appear in the control laws.

28

Aircraft Primer

3
Backstepping

Lyapunov theory has for a long time been an important tool in linear as well as
nonlinear control. However, its use within nonlinear control has been hampered by
the difficulties to find a Lyapunov function for a given system. If one can be found,
the system is known to be stable, but the task of finding such a function has often
been left to the imagination and experience of the designer.
The invention of constructive tools for nonlinear control design based on Lyapunov theory, like backstepping and forwarding, has therefore been received with
open arms by the control community. Here, a control law stabilizing the system is
derived along with a Lyapunov function to prove the stability.
In this chapter, backstepping is presented with the focus on designing state
feedback laws. Sections 3.1 and 3.2 contain mathematical preliminaries. Section
3.3 is the core of the chapter where the main backstepping result is presented along
with a discussion on the class of systems to which it applies and which choices
are left to the designer. In Section 3.4, some related design methods based on
Lyapunov theory are outlined, and in Section 3.5 we survey applications to which
backstepping has been applied.
29

30

Backstepping

3.1

Lyapunov Theory

A basic requirement on a controlled system is that it should end up at the desired

equilibrium without taking a too big detour getting there. Let us formalize this
requirement in terms of the properties of the desired equilibrium, following Slotine
and Li [70].
Definition 3.1 (Lyapunov stability)
Consider the time-invariant system
x = f (x)

(3.1)

starting at the initial state x(0). Assume xe to be an equilibrium point of the

system, i.e., f (xe ) = 0. We say that the equilibrium point is
stable, if for each  > 0 there exists () > 0 such that
kx(0) xe k < kx(t) xe k <  for all t 0
unstable, if it is not stable;
asymptotically stable, if it is stable and in addition there exists r > 0 such
that
kx(0) xe k < r x(t) xe as t
globally asymptotically stable (GAS), if it is asymptotically stable for all initial
states.

Global asymptotic stability (GAS) is what we will strive towards in our control
design. We will refer to a control law that yields GAS as globally stabilizing,
dropping the word asymptotically for convenience.
How can we show a certain type of stability? The definitions above involve
x(t), the explicit solution to (3.1), which in general cannot be found analytically.
Fortunately there are other ways of proving stability.
The Russian mathematician and engineer A. M. Lyapunov came up with the
idea of condensing the state vector x(t) into a scalar function V (x), measuring how
far from the equilibrium the system is. V (x) can be thought of as representing the
energy contained in the system. If V (x) can be shown to continuously decrease,
then the system itself must be moving towards the equilibrium.
This approach to showing stability is called Lyapunovs direct method (or second method) and can be found in any introductory textbook on nonlinear control
such as those by Slotine and Li [70], Khalil [40], Isidori [36], and Vidyasagar [76].
Lyapunovs original work can be found in [52]. Let us first introduce some useful
definitions.

3.1 Lyapunov Theory

31

Definition 3.2
A scalar function V (x) is said to be
positive definite if V (0) = 0 and
V (x) > 0,

x 6= 0

positive semidefinite if V (0) = 0 and

V (x) 0,

x 6= 0

negative (semi-)definite if V (x) is positive (semi-)definite

radially unbounded if
V (x) as kxk

We now state our main theorem for proving stability.
Theorem 3.1 (LaSalle-Yoshizawa)
Let x = 0 be an equilibrium point for (3.1). Let V (x) be a scalar, continously
differentiable function of the state x such that
V (x) is positive definite
V (x) is radially unbounded
V (x) = Vx (x)f (x) W (x) where W (x) is positive semidefinite
Then, all solutions of (3.1) satisfy
lim W (x(t)) = 0

In addition, if W (x) is positive definite, then the equilibrium x = 0 is GAS.

Proof See Krstic et al. [46].

For showing stability when V (x) is only negative semidefinite, the following
corollary due to LaSalle is useful.
Corollary 3.1
Let x = 0 be the only equilibrium point for (3.1). Let V (x) be a scalar, continously
differentiable function of the state x such that
V (x) is positive definite
V (x) is radially unbounded

32

Backstepping

V (x) is negative semidefinite

Let E = {x : V (x) = 0} and suppose that no other solution than x(t) 0 can stay
forever in E. Then x = 0 is GAS.


Proof See Krstic et al. [46].

Note that these results are non-constructive, in the sense that they give no clue
about how to find a V satisfying the conditions necessary to conclude GAS. We
will refer to a function V (x) satisfying the itemized conditions in Theorem 3.1 as
a Lyapunov function for the system.

3.2

Lyapunov Based Control Design

Let us now add a control input and consider the system

x = f (x, u)

(3.2)

Given the stability results from the previous section, it would be nice if we could
find a control law
u = k(x)
so that the desired state of the closed loop system
x = f (x, k(x))
becomes a globally asymptotically stable equilibrium point. For simplicity, we will
assume the origin to be the goal state. This can always be achieved through a
suitable change of coordinates.
A straightforward approach to finding k(x) is to construct a positive definite,
radially unbounded function V (x) and then choose k(x) such that
V = Vx (x)f (x, k(x)) = W (x)

(3.3)

where W (x) is positive definite. Then closed loop stability follows from Theorem
3.1. For this approach to succeed, V and W must have been carefully selected, or
(3.3) will not be solvable. This motivates the following definition:
Definition 3.3 (Control Lyapunov function (clf ))
A smooth, positive definite, radially unbounded function V (x) is called a control
Lyapunov function (clf) for (3.2) if for all x 6= 0,
V = Vx (x)f (x, u) < 0 for some u


3.3 Backstepping

33

Given a clf for the system, we can thus find a globally stabilizing control law. In
fact, the existence of a globally stabilizing control law is equivalent to the existence
of a clf. This means that for each globally stabilizing control law, a corresponding
clf can be found, and vice versa. This is known as Artsteins theorem [4].
If a clf is known, a particular choice of k(x) is given by Sontags formula [72]
reproduced in (3.5). For a system which is affine in the control input
x = f (x) + g(x)u
we can select
u = k(x) =

a+

a2 + b 2
b

(3.4)

(3.5)

where
a = Vx (x)f (x)
b = Vx (x)g(x)
This yields
 a+


V = Vx (x) f (x) + g(x)u = a + b

p
a2 + b 2 
= a2 + b 2
b

(3.6)

and thus renders the origin GAS.

A closely related approach is the one by Freeman and Primbs [22] where u is
chosen to minimize the control effort necessary to satisfy
V W (x)
for some W . The minimization is carried out pointwise in time (and not over some
horizon). Using an inequality constraint rather than asking for equality (as in
(3.6)) makes it possible to benefit from the systems inherent stability properties.
If f (x) alone drives the system (3.4) towards the equilibrium such that
V |u=0 = Vx (x)f (x) < W (x)
it would be a waste of control effort to achieve V = W (x).

3.3

Backstepping

The control designs of the previous section rely on the knowledge of a control
Lyapunov function for the system. But how do we find such a function?
Backstepping answers this question in a recursive manner for a certain class
of nonlinear systems which show a lower triangular structure. We will first state
the main result and then deal with user related issues like which systems can be
handled using backstepping, which design choices there are and how they affect the
resulting control law.

34

Backstepping

3.3.1

Main result

This result is standard today and can be found in, e.g, Sepulchre et al. [66] or
Krstic et al. [46].
Proposition 3.1 (Backstepping)
Consider the system
x = f (x, )
= u

(3.7a)
(3.7b)

where x Rn , R are state variables and u R is the control input.

Assume that for the subsystem (3.7a), a virtual control law
= des (x)

(3.8)

is known such that 0 is a GAS equilibrium of (3.7a). Let W (x) be a clf for (3.7a)
such that
|=des = Wx (x)f (x, des (x)) < 0, x 6= 0
W
Then, a clf for the augmented system (3.7) is given by
V (x, ) = W (x) +

2
1
des (x)
2

(3.9)

Moreover, a globally stabilizing control law, satisfying

2
V = Wx (x)f (x, des (x)) des (x)
is given by
u=

des (x)
f (x, ) f (x, des (x))
f (x, ) Wx (x)
+ des (x)
x
des (x)

(3.10)

Before presenting the proof, it is worth pointing out that (3.10) is neither the
only nor necessarily the best globally stabilizing control law for (3.7). The value
of the proposition is that it shows the existence of at least one globally stabilizing
control law for this type of augmented systems.
Proof We will conduct the proof in a constructive manner to clarify which design choices
that can be made during the control law construction.
The key idea is to use the fact that we know how to stabilize the subsystem (3.7a) if we
were able to control directly, namely by using (3.8). Therefore, introduce the residual
= des (x)
By forcing to zero, will tend to the desired value des and the entire system will be
stabilized.

3.3 Backstepping

35

the system dynamics (3.7) become

In terms of ,

x = f (x, + des (x)) , f (x, des ) + (x, )

= u

(3.11a)

des

(x)
f (x, + des (x))
x

(3.11b)

where
=
(x, )

f (x, + des (x)) f (x, des (x))

In (3.11a) we have separated the desired dynamics from the dynamics due to =
6 0.
To find a clf for the augmented system it is natural to take the clf for the subsystem, W ,
Let us select
and add a term penalizing the residual .
1 2

2
and find a globally stabilizing control law by making V negative definite.
h
i
h
i
des
+ u (x) f (x, + des (x))
V = Wx (x) f (x, des (x)) + (x, )
x
h
i
des
des
+ u (x) f (x, + des (x))
= Wx (x)f (x, (x)) + Wx (x)(x, )
x
= W (x) +
V (x, )

(3.12)

The first term is negative definite according to the assumptions. The second part, and
thereby V , can be made negative definite by choosing
des
+ (x) f (x, + des (x))
u = Wx (x)(x, )
x

This yields
V = Wx (x)f (x, des (x)) 2
which proves the sought global asymptotic stability.

The key idea in backstepping is to let certain states act as virtual controls
of others. The same idea can be found in cascaded control design and singular
perturbation theory [42].
The origin of backstepping is not quite clear due to its simultaneous and often
implicit appearance in several papers in the late 1980s. However, it is fair to say
that backstepping has been brought into the spotlight to a great extent thanks to
the work of Professor Petar V. Kokotovic and his coworkers.
The 1991 Bode lecture at the IEEE CDC, held by Kokotovic [43], was devoted
to the evolving subject and in 1992, Kanellakopoulos et al. [39] presented a mathematical toolkit for designing control laws for various nonlinear systems using
backstepping. During the following years, the textbooks by Krstic et al. [46], Freeman and Kokotovic [21], and Sepulchre et al. [65] were published. The progress of
backstepping and other nonlinear control tools during the 1990s were surveyed by
Kokotovic [41] at the 1999 IFAC World Congress in Beijing.
Let us now deal with some issues related to practical control design using backstepping.

36

3.3.2

Backstepping

Which systems can be handled?

Input nonlinearities
An immediate extension of Proposition 3.1 is to allow for an input mapping to be
present in (3.7b):
= g(x, , u) , u

u
can now be selected according to (3.10) whereafter u can be found given that
g(x, , u) = u

can be solved for u. If this is possible, we say that g is invertible w.r.t. u.

Feedback form systems
Also, we see that the system (3.7) can be further augmented. Assume that u is not
the actual control input to the system, but merely another state variable obeying
the dynamics
u = v

(3.13)

Then (3.10) becomes a virtual control law, which along with the clf (3.9) can be
used to find a globally stabilizing control law for (3.7) augmented by (3.13).
Now, either v is yet another state variable, in which case the backstepping
procedure is repeated once again, or v is indeed the control input, in which case
we are done.
Thus, by recursively applying backstepping, globally stabilizing control laws
can be constructed for systems of the following lower triangular form, known as
pure-feedback form systems:
x = f (x, 1 )
1 = g1 (x, 1 , 2 )
..
.
i = gi (x, 1 , . . . , i , i+1 )

(3.14)

..
.
m = gm (x, 1 , . . . , m , u)
For the design to succeed, a globally stabilizing virtual control law, 1 = 1des (x),
along with a clf, must be known for the x subsystem. In addition, gi , i = 1, . . . , m
1 must be invertible w.r.t. i+1 and gm must be invertible w.r.t. u.

3.3 Backstepping

37

Systems for which the new variables enter in an affine way, are known as
strict-feedback form systems:
x = a(x) + b(x)1
1 = a1 (x, 1 ) + b1 (x, 1 )2
..
.
i = ai (x, 1 , . . . , i ) + bi (x, 1 , . . . , i )i+1
..
.
m = am (x, 1 , . . . , m ) + bm (x, 1 , . . . , m )u
Strict-feedback form systems are nice to deal with and often used for deriving
results related to backstepping. Firstly, the invertability condition imposed above
is satisfied given that bi 6= 0. Secondly, if (3.7a) is affine in , the control law (3.10)
reduces to
u=

des (x)
(a(x) + b(x)) Wx (x)b(x) + des (x)
x

Dynamic backstepping
Even for certain systems which do not fit into a lower triangular feedback form,
there exist backstepping designs. Fontaine and Kokotovic [18] consider a two dimensional system where both states are affected by the control input:
x 1 = (x1 ) + x2 + (u)
x 2 = u
Their approach is to first design a globally stabilizing virtual control law for the
x1 -subsystem, considering = x2 + (u) as the input. Then backstepping is used
to convert this virtual control law into a realizable one in terms of u. Their design
results in a dynamic control law, and hence the term dynamic backstepping is used.

3.3.3

Which design choices are there?

The derivation of the backstepping control law (3.10) leaves a lot of room for
variations. Let us now exploit some of these.
Dealing with nonlinearities
A trademark of backstepping is that is allows us to benefit from useful nonlinearities, naturally stabilizing the system. This can be done by choosing the virtual
control laws properly. The following example demonstrates this fundamental difference to feedback linearization.

38

Backstepping

x = x
x 0

6
2

x
Figure 3.1 The dynamics of the uncontrolled system x = x3 + x. The
linear term acts destabilizing around the origin.

Example 3.1 (A useful nonlinearity I)

Consider the system
x = x3 + x + u
where x = 0 is the desired equilibrium. The uncontrolled dynamics, x = x3 +
x, are plotted in Figure 3.1. For the system to be stable, the sign of x should be
opposite that of x for all x. This holds for large values of x where the cubic term
x3 dominates the dynamics, but near the origin, the linear term x dominates
and destabilizes the system.
Thus, to make the origin GAS, only the linear dynamics need to be counteracted
by the control input. This can be achieved by selecting
u = x

(3.15)

A clf is given by
W =

1 2
x
2

(3.16)

3.3 Backstepping

39

which yields
= x(x3 + x + u) = x4
W
proving the origin to be GAS according to Theorem 3.1.
Applying feedback linearization would have rendered the control law
u = x3 kx,

k>1

(3.17)

Obviously, this control law does not recognize the beneficial cubic nonlinearity
but counteracts it, thus wasting control effort. Also, the feedback linearizing
design is dangerous from a robustness perspective what if the true system
dynamics are x = 0.9x3 + x + u and (3.17) is applied...
Weighting the clf
When constructing the combined clf (3.9), we can choose any weighted sum of the
two terms,
1
V = cW + ( des )2 , c > 0
2
In our designs, we will use the weight c to cancel certain terms in Equation (3.12).
A technical hint is to put the weight on W since it yields nicer expressions.
Non-quadratic clf
Although quadratic clf:s are frequently used in backstepping, they do not always
make up the best choice as the following example demonstrates.

Example 3.2 (A useful nonlinearity II)

Consider the system in Example 3.1 augmented by an integrator:
x 1 = x31 + x1 + x2
x 2 = u
To benefit from the useful nonlinearity x31 , let us reuse (3.15) as our virtual
control law, i.e.,
xdes
2 (x1 ) = x1
As it turns out, the choice of clf for the x1 -subsystem will strongly affect the
resulting control law. To see this we first reuse (3.16) and pick
W =

1 2
x
2 1

40

Backstepping

Introducing
x
2 = x2 xdes
2 (x1 ) = x2 + x1
we can rewrite the system as
2
x 1 = x31 + x
x2 = u x31 + x
2
Following the proof of Proposition 3.1, we add a quadratic term to W , to
penalize the deviation from the suggested virtual control law:
V (x1 , x2 ) = W (x1 ) +

2
1
1
1 2
x2 xdes

= x21 + x
2 (x1 )
2
2
2 2

Differentiating w.r.t. time yields

2 ) + x
2 (u x31 + x
2 )
V = x1 (x31 + x
2 (x1 + u x31 + x
2 )
= x41 + x
To render V negative definite, u must clearly dominate the x
2 term using a
control input of, e.g., 3
x2 . In addition, since the mixed terms between x1 and
x
2 are indefinite, there seems to be no other choice than to cancel them using
the control law
u = x31 x1 3
x2 = x31 4x1 3
x2
We note that this control law does not recognize the fact that x1 -subsystem
is naturally stabilized for high values of x1 but instead counteracts this effect,
thereby wasting control effort.
So how should we pick W (x1 ) to avoid this cancellation? One idea is not to
specify the clf beforehand, but instead let the choice of W be decided by the
backstepping design. Thus, we let W be any function satisfying Definition 3.3.
As before, use
1 2

V (x1 , x2 ) = W (x1 ) + x
2 2
and compute its time derivative.
2 ) + x
2 (u x31 + x
2 )
V = W 0 (x1 )(x31 + x
2 (W 0 (x1 ) + u x31 + x
2 )
= W 0 (x1 )x31 + x
We now use our extended design freedom and select a W so that the indefinite
mixed terms cancel each other. This is satisfied by
W 0 (x1 ) = x31 ,

W (x1 ) =

1 4
x
4 1

3.3 Backstepping

41

which indeed is a valid choice. We now have

V = x61 + x
2 (u + x
2 )
Clearly, the control u no longer has to cancel the cubic x1 term but can be
chosen linear in x1 and x2 .
u = 3
x2 = 3x1 3x2
renders V = x61 2
x22 negative definite and thus makes the origin GAS.
This refinement of backstepping is due to Krstic et al. [45]. The technique of
designing a non-quadratic clf will be used for flight control design in Chapter 5,
where some of the aerodynamic forces also have the property of being nonlinear
but still stabilizing.
Implicit residuals
In (3.9) the deviation from the desired virtual control law is penalized through the
difference des (x). Another way of forcing towards des is to instead penalize
the implicit residual
() ( des (x)) , () des (x)
where () is an invertible and strictly monotone mapping. This leads to the clf
V (x, ) = W (x) +

2
1
() des (x)
2

An equivalent way of reaching this clf is to replace by () in the system

description (3.7):
x = g(x, )
= 0 ()u
Here, g(x, ()) = f (x, ) has been introduced. Applying (3.10) now gives us
u=

i
g(x, ) g(x, des (x))
1 h des (x)
g(x, ) Wx (x)
+ des (x)
des
x
(x)
(3.18)

0 ()

Example 3.3
Consider the system
x 1 = x31 + x52 + x2
x 2 = u

42

Backstepping

A clf for the x1 -subsystem is given by W (x1 ) = 12 x21 which gives

= x1 (x31 + x52 + x2 )
W
In terms of = x52 + x2 , a globally stabilizing virtual control law is easy to find.
= x41 negative definite. Inserting
Pick, e.g., des (x1 ) = 2x31 which yields W
this into (3.18) gives the control law
u=

i
1 h
6x21 (x31 + x52 + x2 ) x1 2x31 x52 x2
+1

5x42

In [2], this technique was used for speed control of a switched reluctance motor
where it was convenient to formulate the virtual control law in terms of the square
current i2 .
Optimal backstepping
In linear control, one often seeks control laws that are optimal in some sense, due
to their ability to suppress external disturbances and to function despite model
errors, as in the case of H and linear quadratic control [79].
It is therefore natural that efforts have been made to extend these designs to
nonlinear control. The difficulty lies in the Hamilton-Jacobi-Bellman equation that
needs to be solved in order to find the control law.
A way to surpass this problem is to only require the desired optimality to hold
locally around the origin, where the system can be approximated by its linearization. In the global perspective, one settles for optimality according to some cost
functional that the designer cannot rule over precisely. This is known as inverse
optimality, which is the topic of Chapter 4.
Contributions along this line can be found for strict-feedback form systems.
Ezal et al. [17] use backstepping to construct controllers which are locally H optimal. L
ofberg [51] designs backstepping controllers which are locally optimal
according to a linear quadratic performance index.
One advantage of using an optimality based approach is that the designer then
specifies an optimality criterion rather than virtual control laws and the clf:s themselves. This enhances the resemblance with linear control.

3.4

Related Lyapunov Designs

Besides state feedback backstepping, several other constructive nonlinear control

designs have been developed during the last decade. We will now outline some of
these.

3.5 Applications of Backstepping

3.4.1

43

Forwarding

The backstepping philosophy applies to systems of the form (3.7). Another class
of nonlinear systems for which one can also construct globally stabilizing control
laws are those that can be written
x = f (x, u)
= g(x, u)

(3.19a)
(3.19b)

A clf and a globally stabilizing control law for the x-subsystem (3.19a) are assumed
to be known. The question is how to augment this control law to also stabilize the
integrator state in (3.19b). This problem, which can be seen as a dual to the one
in backstepping, can be solved using so called forwarding [67].
By combining feedback (3.7) and feedforward (3.19) systems, interlaced systems
can be constructed. Using backstepping in combination with forwarding, such
systems can also be systematically stabilized [66].

3.4.2

Adaptive, robust, and observer backstepping

So far we have only considered the case where all the state variables are available
for feedback and where the model is completely known. For the non-ideal cases
where this is not true, there are other flavors of backstepping to resort to.
For systems with parametric uncertainties, there exists adaptive backstepping
[46]. Here, a parameter estimate update law is designed such that closed loop
stability is guaranteed when the parameter estimate is used by the controller. In
Section 6.3 we will see how this technique can be used to estimate and cancel
unknown additive disturbances on the control input.
Robust backstepping [21] designs exist for systems with imperfect model information. Here, the idea is to select a control law such that a Lyapunov function
decreases for all systems comprised by the given model uncertainty.
In cases where not all the state variables can be measured, the need for observers
arises. The separation principle valid for linear systems does not hold for nonlinear
systems in general. Therefore, care must be taken when designing the feedback law
based on the state estimates. This is the topic of observer backstepping [39, 46].

3.5

Applications of Backstepping

Although backstepping theory has a rather short history, numerous practical applications can be found in the literature. This fact indicates that the need for
a nonlinear design methodology handling a number of practical problems, as discussed in the previous section, has existed for a long time. We now survey some
publications regarding applied backstepping. This survey is by no means complete, but is intended to show the broad spectrum of engineering disciplines in
which backstepping has been used.

44

Backstepping

Backstepping designs can be found for a wide variety of electrical motors [2, 10,
11, 33, 34]. Turbocharged diesel engines are considered in [20, 37] while jet engines
are the subject of [45].
In [25, 75], backstepping is used for automatic ship positioning. In [75], the
controller is made locally H -optimal based on results in [17].
Robotics is another field where backstepping designs can be found. Tracking
control in considered in [38] and [9] where the latter is a survey of approaches valid
for various assumptions regarding the knowledge of the model.
There also exist a few papers, combining flight control and backstepping. [68]
treats formation flight control of unmanned aerial vehicles. [69] and [73] use backstepping to design flight control laws which are adaptive to changes in the aerodynamic forces and moments due to, e.g., actuator failures. Also, the Lyapunov
functions used contain a term penalizing the integral of the tracking error, enhancing the robustness.

4
Inverse Optimal Control

This chapter is preparatory for the upcoming control designs. The tools we develop
in this chapter will be used to show that the control laws derived in Chapter 5 each
minimize a certain cost functional. This route of first deriving a control law and
then determining which cost it minimizes, and thus in which sense it is optimal, is
known as inverse optimal control.
The material in this chapter will be presented in a rather intuitive manner.
A mathematically strict treatment of the subject can be found in, e.g., Sepulchre
et al. [65]. In Section 4.1 the general infinite horizon optimal control problem
is introduced. In Section 4.2, systems which are affine in the control input are
considered, and some standard inverse results are derived for cost functionals which
are quadratic in the input. Finally, the well known gain margin result of optimal
control is shown in Section 4.3.

4.1

Optimal Control

A general idea within control design is to select a control law which is optimal in
some sense. Given a dynamic system
x = f (x, u)
45

46

Inverse Optimal Control

where x Rn is the state vector and u Rm is the control input, we seek the
control law u(x) that minimizes the cost functional
Z
J=
L(x, u)dt
0

By choosing L properly, the system is guaranteed to reach steady state as t .

J, representing the cost to get there, the cost-to-go, will depend on the initial state
of the system, x(0). We therefore write this cost J(x).
The optimal control law is denoted by u (x). When this optimal control law is
applied, J(x) will decrease along the trajectory, since the cost-to-go must continuously decrease by the principle of optimality [6]. This means that V (x) = J(x) is
a Lyapunov function for the controlled system. At steady state it must hold that
V = 0. Hence, the following holds:
Z
Z


V (x(0)) =
L(x, u )dt = V (x()) V (x(0)) =
V (x)dt
| {z }
0
0
=0

Clearly, when the optimal control law is used, L and V coincide. This motivates
the Hamilton-Jacobi-Bellman (HJB) equation


(4.1)
0 = min L(x, u) + Vx (x)f (x, u)
u

for finding the optimal control law u along with a Lyapunov function V (x) for the
controlled system.

4.2

Inverse Optimal Control

It is well known that in general, it is not feasible to solve the HJB equation (4.1).
We therefore restrict our discussion to dynamic systems of the form
x = f (x) + g(x)u

(4.2)

For these systems, the HJB equation is greatly simplified if L is chosen quadratic
in u according to
L(x, u) = q(x) + uT R(x)u
where q(x) > 0, x 6= 0 and R(x) is symmetric matrix, positive definite for all x.
Inserting this into (4.1) yields


(4.3)
0 = min q(x) + uT R(x)u + Vx (x)(f (x) + g(x)u)
u

The equation is solved in two steps. First we find the minimizing u, and then we
solve for equality to zero. The minimization can be done by completion of squares:
q + uT Ru + Vx f + Vx gu =

T 
 1
1
1
q + Vx f + u + R1 (Vx g)T R u + R1 (Vx g)T Vx gR1 (Vx g)T
2
2
4

4.2 Inverse Optimal Control

47

The control input u only appears in the square, positive definite term. The
minimum therefore occurs when this term is set to zero, which is achieved by
1
u = k(x) = R1 (Vx g)T
2

(4.4)

What remains is to insert this control law into (4.3). This gives us
1
0 = q + Vx f Vx gR1 (Vx g)T
4

(4.5)

Equations (4.4) and (4.5) provide the connection between the cost functional,
given by q(x) and R(x), and the optimal control strategy, in terms of k(x) and
V (x). As for the general problem in the previous section, it is in general not
feasible to solve for k(x) and V (x) given q(x) and R(x) of the designers choice.
However, we see that the reverse task is simpler. Given a control law k(x) and a clf
V (x) (corresponding to a Lyapunov function for the controlled system), q(x) and
R(x), determining the cost functional that is minimized, can be found by solving
1 1
R (x)(Vx (x)g(x))T
2
1
q(x) = Vx (x)f (x) + Vx (x)g(x)k(x)
2

k(x) =

(4.6)
(4.7)

For a single input system we can explicitly solve for R(x):

R(x) =

Vx (x)g(x)
2k(x)

(4.8)

Example 4.1 (Exploiting useful nonlinearities is optimal)

Let us return to the system considered in Example 3.2 and show that the derived
backstepping control law is indeed optimal w.r.t. a meaningful cost functional.
The original dynamics,
x 1 = x31 + x1 + x2
x 2 = u
can be written in the form (4.2) with
 3

x1 + x1 + x2
f (x) =
,
0

 
0
g(x) =
1

In Example 3.2,
u = k(x) = 3(x1 + x2 )
was shown to be globally stabilizing using the clf
V (x) =

1 4 1
x + (x1 + x2 )2
4 1 2

48

Inverse Optimal Control

which satisfies
Vx (x) = x31 + x1 + x2

x1 + x2

Inserting this into (4.7) and (4.8) yields

R(x) =

x1 + x2
1
=
2 3(x1 + x2 )
6

1
q(x) = (x31 + x1 + x2 )(x31 + x1 + x2 ) + (x1 + x2 ) 3(x1 + x2 )
2
1
6
2
= x1 + (x1 + x2 )
2
Thus, the suggested control law minimizes the cost functional
Z
1
1
J=
(x61 + (x1 + x2 )2 + u2 )dt
2
6
0

4.3

Robustness of Optimal Control

Besides optimal control being an intuitively appealing approach, the resulting control laws inherently possess certain robustness properties [24]. One important property regards the gain margin.
Assume that the prescribed optimal control input (4.4) cannot be produced
exactly, but that the actual control input is
u = (x)u

(4.9)

where (x) > 0 is a scalar, see Figure 4.1. Actuator saturation, for example, can
be modeled as gain reduction, (x) < 1. Are optimal controllers robust to such
changes in the gain? The control law (4.9) is globally stabilizing provided that
V = Vx f + Vx gu = Vx f + Vx gu
is negative definite. From the assumptions and (4.7) we know that
1
q = Vx f + Vx gu
2
is negative definite. Combining these two equations yields
1 1
1
V = q + ( )Vx gu = q(x) ((x) ) Vx (x)g(x)R1 (x)(Vx (x)g(x))T
2
2 |2
{z
}
positive (semi-)definite

Apparently, V is negative definite (at least) for all (x) 12 . Thus, all state
feedback control laws which solve an optimal control problem of the type considered
in Section 4.2, have a gain margin of [ 12 , ].
Note that the actual tolerable gain reduction may be more than 50%. In Example 3.2, any control law u = k
x2 where k > 1 makes V negative definite and
hence is globally stabilizing. The selected control law u = 3
x2 thus has a gain
margin of ] 13 , ].

4.3 Robustness of Optimal Control

u = k(x)

(x)

49

x = f (x) + g(x)u

Figure 4.1 The optimal control law u = k(x) remains globally stabilizing for any gain perturbation (x) 12 .

50

Inverse Optimal Control

5
Backstepping Designs for
Flight Control

In the previous chapters we have introduced the aircraft dynamics, the backstepping design procedure and inverse optimality tools for evaluating state feedback
laws. We now have the toolbox we need to live up to the title of the thesis, and do
flight control design using backstepping. This chapter is the core of the thesis.
We will design flight control laws for two different objectives for general maneuvering (control of , , and ps ) and for flight path angle control (control of ).
The two presentations in Sections 5.1 and 5.2 follow the same outline. First, the
relevant dynamics from Chapter 2 are reproduced and the assumptions needed for
making the control design feasible are stated. The flight control problem of interest is then viewed as a more general nonlinear control problem and backstepping is
used to derive globally stabilizing control laws, whose properties are investigated.
We finally return to the flight control context and investigate which practical consequences applying the derived control laws leads to.

5.1

General Maneuvering

We begin by designing flight controllers for general purpose maneuvering, useful,

e.g., in dogfight situations where maneuverability is very important in order to
get a first shot opportunity. Sharp turns imply high angles of attack and a prime
feature of our design is that stability is guaranteed for all angles of attack, including
51

52

Backstepping Designs for Flight Control

post-stall, given that enough pitching moment can be produced.

5.1.1

Objectives, dynamics, and assumptions

Following the discussion in Section 2.2 on relevant variables to be controlled for

this type of general maneuvering we set up the following objectives:
Fore-and-aft control stick deflections should demand angle of attack.
Side-to-side control stick deflections should demand stability axis roll rate.
The sideslip should be kept zero at all times.
The controlled variables are illustrated in Figure 2.3. With these objectives, the
task of the control system is to make
= ref
=0
ps = pref
s
a globally asymptotically stable equilibrium. Speed control is assumed to be handled separately and will not be dealt with.
The relevant dynamics consist of the angular rate dynamics in (2.6) and the
and dynamics in (2.7b) and (2.7c). Trading the scalar moment equations (2.6)
for the compact vector description (2.2) yields
= q (p cos + r sin ) tan +
= p sin r cos +

1
(L FT sin + mg2 )
mVT cos

1
(Y FT cos sin + mg3 )
mVT

M = I + I

(5.1a)
(5.1b)
(5.1c)

where
g2 = g(cos cos cos + sin sin )
g3 = g(cos cos sin + sin cos sin sin sin cos cos )
For backstepping to be applicable, the system dynamics must comply with the
pure-feedback form (3.14). We therefore need the following assumption:
A1. Control surface deflections only produce aerodynamic moments, and not
forces. Also neglecting the dependence on the angular rates, the aerodynamic force coefficients can be written
lift force coefficient: CL ()
side force coefficient: CY ()
whose characteristics are shown in Figure 5.1.

5.1 General Maneuvering

53

Side force coefficient

0.3

1.5

0.2

0.1
CY

CL

Lift force coefficient

2

0.5

0.1

0.5

0.2

1
10

10

20
30
(deg)

40

50

30

20

10

0
10
(deg)

20

30

Figure 5.1 Typical lift force coefficient vs. angle of attack and side force
coefficient vs. sideslip characteristics.

To simplify the design we also make the following assumptions:

A2. The speed, altitude, and orientation of the aircraft vary slowly compared to
the controlled variables. Therefore, their time derivatives can be neglected.
A3. Longitudinal and lateral commands are assumed not to be applied simultaneously. Hence, when is controlled, and ps are considered constant.
Conversely, when and ps are controlled, is considered constant.
A4. The control surface actuator dynamics are assumed to be fast enough to be
disregarded.
Assumption A3 certainly does not hold in practice but allows us to decouple the
control design. As we will show using simulations in Chapter 7, the resulting control
laws work well in practice.
Since the roll rate to be controlled, ps , is expressed in the stability-axes coordinate system, we need to establish the relationship to the body-axes angular
velocity, . The stability-axes angular velocity,
s = ps

qs

rs

T

is related to the body-axes angular velocity

= p

T

54

Backstepping Designs for Flight Control

through the transformation

s = Rsb ,

Rsb

cos
= 0
sin

0 sin
1
0
0 cos

(5.2)

1
T
Note that the transformation matrix Rsb satisfies Rsb
= Rsb
.
Inspecting the dynamics (5.1), we see that it is more convenient to work with
s rather than . Introducing

T
u = u1 u2 u3 = s

we can rewrite the aircraft dynamics (5.1) as

p s = u1

(5.3a)

1
(L() FT sin + mg2 )
= qs ps tan +
mVT cos
qs = u2
1
(Y () FT cos sin + mg3 )
= rs +
mVT
rs = u3

(5.3b)
(5.3c)
(5.3d)
(5.3e)

These are the dynamics we will use for the control design.
The relationship between u, which will be considered as the control input during
the control design, and the actual control input, , can be found by combining (5.1c)
and (5.2). Under assumption A3 above, , and thereby also Rsb , are considered
constant while realizing u1 and u3 which relate to lateral control. This yields
u = s = Rsb = Rsb I 1 (M I)

(5.4)

Solving for the net moment, M, which depends on the control surface deflections,
, we get
T
M() = IRsb
u + I

(5.5)

We will postpone the discussion on how to practically solve for given u until
Chapter 7.

5.1.2

Backstepping control design

The nonlinear control problem

To begin with, note the structural similarities between the angle of attack dynamics
(5.3b)(5.3c) and the sideslip dynamics (5.3d)(5.3e). Both these second order
systems can be written
w 1 = f (w1 , y) + w2
w 2 = u

(5.6)

5.1 General Maneuvering

55

General system
(5.6)
R
w1
w2
u
y
f (w1 , y)

dynamics
(5.3b)(5.3c)
ref

qs
u2
ps , , VT , h, ,
f (, y )

dynamics
(5.3d)(5.3e)
0

rs
u3
, VT , h, ,
f (, y )

Table 5.1 The relationships between the general nonlinear system (5.6)
and the angle of attack and sideslip dynamics in (5.3).

where y represents the influence of variables that we regard as constant, and w1 = R

is the desired equilibrium. We assume that the reference, R, is constant, which is
the best we can do without making any assumptions regarding the maneuvers that
the pilot wants to perform. Table 5.1 summarizes the relationships between (5.6)
and the original aircraft dynamics (5.3). Here,
f (, y ) = ps tan +
f (, y ) =

1
(L() FT sin + mg2 )
mVT cos

1
(Y () FT cos sin + mg3 )
mVT

(5.7)
(5.8)

have been introduced.

For notational convenience it is favorable to make the origin the desired equilibrium. We therefore define
x1 = w1 R
x2 = w2 + f (R, y)
(x1 ) = f (x1 + R, y) f (R, y)

(5.9)

which yields (0) = 0. In these coordinates, the dynamics (5.6) become

x 1 = (x1 ) + x2
x 2 = u

(5.10a)
(5.10b)

The relationships between (5.10) and the original aircraft dynamics (5.3) can be
found in Table 5.2. If the aircraft is assumed to be built symmetrically such that
the side force, Y , is zero for zero sideslip, we get
f (0, y ) =
using (2.8).

1
g cos sin
VT

(5.11)

56

Backstepping Designs for Flight Control

General system
(5.10)
x1
x2
u
(x1 )

dynamics
(5.3b)(5.3c)
ref
qs + f (ref , y )
u2
f (, y ) f (ref , y )

dynamics
(5.3d)(5.3e)

rs + f (0, y )
u3
f (, y ) f (0, y )

Table 5.2 The relationships between the translated general nonlinear system (5.10) and the angle of attack and sideslip dynamics in
(5.3).

In the remainder of this section, we will use backstepping to construct globally

stabilizing state feedback control laws for the system (5.10) assuming a general
nonlinearity . In Section 5.1.3 we will return to the flight control context and
express the control laws in the original coordinates.
The backstepping design
In the case of and control, will contain the negative lift force, L, and
the side force, Y . As seen from Figure 5.1, the gradients of these contributions
are negative (note the sign on the lift force) in large parts of the operating range.
The only exception is the post-stall behavior of the lift force. The key property of
backstepping is that it allows us to benefit from these inherently stabilizing forces
and not cancel them. If we do not have to cancel them, we need less information
about them, which is appealing from a robustness point of view.
Step 1: In the spirit of backstepping, we start by finding a globally stabilizing
virtual control law for the for the x1 -subsystem (5.10a). Not to constrain ourselves
at this stage, we pick a general control law
xdes
2 = (x1 )

(5.12)

The idea is to let the necessary demands on be revealed by the design below.
Temporarily pick the clf
W (x1 ) =

1 2
x
2 1

negative definite. Thus, we select a

To make x1 = 0 GAS we must make W
satisfying


|x =xdes = (x1 ) (x1 ) x1 < 0, x1 6= 0
(5.13)
W
2
2
This condition is natural, since for x1 = 0 to be GAS, the desired net x1 dynamics,
x 1 = (x1 ) (x1 ) must lie in the second and fourth quadrants.

5.1 General Maneuvering

57

Step 2: Continue by introducing the residual

x
2 = x2 xdes
2 = x2 + (x1 )
2 .
and rewrite the dynamics (5.10) in terms of x1 and x
2
x 1 = (x1 ) (x1 ) + x


0
x
2
2 = u + (x1 ) (x1 ) (x1 ) + x

(5.14)

Remembering the benefits of using a non-quadratic clf (cf. Example 3.2), we select
1 2

2 ) = F (x1 ) + x
V (x1 , x
2 2
as the clf for the total system (5.14), where F is any valid clf for the x1 -subsystem.
Specifically this means that


0
F (x1 )|x2 =xdes
=
F
(x
)
(x
)

(x
)
= U (x1 )
(5.15)
1
1
1
2
where U (x1 ) is positive definite. Differentiating V w.r.t. time we get





2
2 u + 0 (x1 ) (x1 ) (x1 ) + x
V = F 0 (x1 ) (x1 ) (x1 ) + x2 + x




x2
= U (x1 ) + x2 F 0 (x1 ) + u + 0 (x1 ) (x1 ) (x1 ) + 0 (x1 )
We can reduce the complexity of the second term by selecting F such that the x1
terms inside the brackets cancel each other. This is achieved by


F 0 (x1 ) = 0 (x1 ) (x1 ) (x1 ) , F (0) = 0
Inserting this into (5.15) yields

2
U (x1 ) = 0 (x1 ) (x1 ) (x1 )
For U (x1 ) to be positive definite, must satisfy
0 (x1 ) > 0,

x1 6= 0

(5.16)

Note that (x1 ) 6= (x1 ) holds according to (5.13). An intuitive interpretation of

this condition is that the virtual control law (5.12) must provide negative feedback
everywhere, even locally. We now have that


2 u + 0 (x1 )
x2
V = U (x1 ) + x
To make V negative definite, u must dominate the positive definite, destabilizing
term 0 (x1 )
x22 . If 0 (x) has an upper bound, u can be selected linear in x
2 .
Choosing


(5.17)
u = k
x2 = k x2 + (x1 )

58

Backstepping Designs for Flight Control

()

xdes
2

x2

x 2 = u

x1
x 1 = (x1 ) + x2

Figure 5.2 The nonlinear system 5.10 can be globally stabilized through
a cascaded control structure.

where
k > max 0 (x1 )
x1

(5.18)

renders
V = U (x1 ) (k 0 (x1 ))
x22
negative definite, thus making the origin of (5.10) GAS. Note the cascaded structure
of the control law as illustrated in Figure 5.2. This way of viewing the control law
motivates the condition (5.18). The condition apparently states that the inner
control loop must have a higher feedback gain than the outer loop.
Before we conclude, let us investigate which system nonlinearities can be
handled using the control law (5.17). If 0 (x1 ) is upper bounded by k as in (5.18),
then the growth rate of must also be bounded in the sense that
(x1 )
max 0 (x1 ) < k
x1
x1
and hence, must be confined to the sectors shown in Figure 5.3(b). Dividing
(5.13) by x21 and inserting this inequality yields
(x1 )
(x1 )
<
<k
x1
x1
Thus, for the control law (5.17) to be applicable, the growth of the system nonlinearity must also be linearly upper bounded as depicted in Figure 5.3(a). Conversely, given an upper bound on (x1 )/x1 we can always find a and a k such
that (5.17) is globally stabilizing.
Let us summarize our findings as a proposition.
Proposition 5.1
Consider the system
x 1 = (x1 ) + x2

(5.19a)

x 2 = u

(5.19b)

5.1 General Maneuvering

59

0.5

0.5

0.5

1
1

0.5

0.5

x1

0.5

1
1

(a) System nonlinearity

0.5

x1

0.5

(b) Virtual control law

Figure 5.3 and must be sector bounded for a globally stabilizing

control law of the form (5.17) to exist.

where (0) = 0. Assume there exists a constant such that

(x1 )
for all x1 6= 0
x1

(5.20)

Then, a globally stabilizing control law

u = k x2 + (x1 )

(5.21)

can be found, where

(x1 ) (x1 ) x1 < 0,

x1 6= 0

(5.22)

and
0 < 0 (x1 ) < k

(5.23)

Moreover, a clf is given by

Z x1


1
0 (y) (y) (y) dy + (x2 + (x1 ))2
V =
2
0
which satisfies

2
x22
V = 0 (x1 ) (x1 ) (x1 ) (k 0 (x1 ))


60

Backstepping Designs for Flight Control

Inverse optimality
Proposition 5.1 gives us a family of control laws which all globally stabilize the
system of interest. Before discussing which choices of might be of interest, let
us further examine the properties of the derived control law. Using the tools of
Chapter 4 we will show that (5.21) is actually optimal with respect to a meaningful
cost functional, provided that k is chosen properly.
The control input enters the system (5.10) affinely and thus the tools of Chapter
4 can be used. However, we will use the transformed system (5.14) to compute the
cost functional that is minimized since the expressions become simpler. Comparing
(5.14) with (4.2), we have that


 
(x1 ) (x1 ) + x
2
0
f (x) =
, g(x) =
0 (x1 )((x1 ) (x1 ) + x
2 )
1
We also have that
Vx = 0 (x1 )((x1 ) (x1 )) x2

Inserting this into (4.7) and (4.8) yields

1
x
2
=
2k
x2
2k
q(x) = 0 (x1 )((x1 ) (x1 ))((x1 ) (x1 ) + x
2 )
1
2 k
2 ) x
x2
+x
2 0 (x1 )((x1 ) (x1 ) + x
2
k
= 0 (x1 )((x1 ) (x1 ))2 + ( 0 (x1 ))
x22
2

R(x) =

Apparently, to make q(x) positive definite, which is required for the cost functional
to be meaningful, k should be chosen such that
k > 2 max 0 (x1 )
x1

Note that this lower limit for inverse optimality is twice the limit in (5.23) regarding
global stability. This is natural considering the 50% gain reduction robustness of
all optimal controllers, cf. Chapter 4.
Proposition 5.2
The control law (5.21) is optimal w.r.t. a meaningful cost functional for
k > 2 max 0 (x1 )
x1

The control law then minimizes

Z 
k
1 2
u dt
0 (x1 )((x1 ) (x1 ))2 + ( 0 (x1 ))(x2 + (x1 ))2 +
2
2k
0


5.1 General Maneuvering

61

Two special cases

So how should be chosen? Let us investigate two special choices. One has control
law simplicity in focus while the other aims at linearizing the dynamics.
Corollary 5.1 (Linear control)
Consider the system (5.19). A globally stabilizing control law, linear in the states,
is given by
u = k2 (x2 + k1 x1 )
where
k2 > k1 > max{0, }
with as defined in (5.20). In addition, for k2 > 2k1 the control law minimizes a
meaningful cost functional given by
Z 
k2
1 2
k1 ((x1 ) k1 x1 )2 + ( k1 )(x2 + k1 x1 )2 +
u dt
2
2k2
0
Proof Selecting (x1 ) = k1 x1 and k = k2 yields u = k2 (x2 + k1 x1 ) and leads to the
conditions
(5.22): ((x1 ) k1 x1 )x1 ( k1 )x21 < 0,

x1 6= 0

k1 >

(5.23): 0 < k1 < k2

The cost functional follows directly from Proposition 5.2.

Note how little information about the system nonlinearity this control law is
dependent of. Only an upper bound on its growth rate, , is needed. In particular,
if is known to lie in the second and fourth quadrants only, thus intuitively being
useful for stabilizing x1 , we do not need any further information, since then < 0
and the parameter restriction k1 > 0 becomes active.
Corollary 5.2 (Linearizing control)
Consider the system (5.19). A globally stabilizing control, partially linearizing the
dynamics, is given by
u = k2 (x2 + k1 x1 + (x1 ))
where
k1 > max{0, min 0 (x1 )}
x1

and
k2 > k1 + max 0 (x1 )
x1

(5.24)

62

Backstepping Designs for Flight Control

provided that such upper and lower bounds on 0 exist. In addition, for k2 >
2(k1 + maxx1 0 (x1 )) the control law minimizes a meaningful cost functional given
by
Z 
k2
1 2
(k1 + 0 (x1 ))k12 x21 + ( k1 0 (x1 ))(x2 + k1 x1 + (x1 ))2 +
u dt
2
2k2
0
Proof Selecting (x1 ) = k1 x1 + (x1 ) and k = k2 yields u = k2 (x2 + k1 x1 + (x1 ))
and the conditions
(5.22): ((x1 ) k1 x1 (x1 ))x1 = k1 x21 < 0,

x1 6= 0

k1 > 0

(5.23): The first part of the inequality becomes

0 < k1 + 0 (x1 ),

x1 6= 0

k1 > min 0 (x1 )

x1 6=0

The second part of the inequality becomes

k1 + 0 (x1 ) < k2 ,

x1 6= 0

k2 > k1 + max 0 (x1 )

x1 6=0

The cost functional follows directly from Proposition 5.2.

This control law corresponds to choosing the virtual control law

xdes
2 = (x1 ) k1 x1
which, as in feedback linearization, cancels the nonlinear dynamics of x1 , (x1 ),
and replaces them with linear dynamics, k1 x1 . Condition (5.24) states that it is
alright to cancel the natural dynamics, (x1 ), as long as the new dynamics, k1 x,
provide at least the same amount of negative feedback. This is sound also from
an optimality point of view. Intuitively, it must be suboptimal to waste feedback
effort in order to slow down the natural system dynamics.

5.1.3

Flight control laws

Let us now return to the flight control context and apply the control laws derived
in the previous section to the aircraft dynamics in (5.3). The boxed control laws
below are the ones that will be implemented and evaluated in Chapter 7.
Angle of attack control
Let us first apply the linear control law in Corollary 5.1. Using Tables 5.1 and 5.2
for translation from x to the proper aircraft variables yields
u2 = k,2 (qs + k,1 ( ref ) + f (ref , y ))

(5.25)

5.1 General Maneuvering

63

u2

ref

Prefilter

, qs
dynamics

(k1 k2 k2 )
Figure 5.4 The backstepping control law (5.25) moves the dependence on
CL outside the control loop, thereby enhancing the robustness.

with f from (5.7). The control law is illustrated in Figure 5.4. Although implementing this control law requires knowledge of the lift force, and thereby the
lift force coefficient, CL , we note that the lift force dependent computation is performed in the prefilter outside the feedback loop. Therefore, imperfect knowledge
of CL does not jeopardize closed loop stability but only shifts the equilibrium.
The parameters should satisfy
k,2 > 2 k,1 ,

k,1 > max{0, }

for the control law to be globally stabilizing and also minimize a meaningful cost
functional. Here,
= max
x1

(x1 )
f (, y ) f (ref , y )
= max
x1
ref
,ref ,y

The maximum occurs when ref is chosen to be the point where f has the highest
positive slope and is selected infinitely close to ref . Then, the fraction above
turns into the derivative w.r.t. , i.e.,
= max
,y

f (, y )

In practice, f in (5.7) is dominated by the lift force term. The contribution to

from the lift force becomes
max
,y

1
dCL
(h)VT S
L
() = max
max
()

,V
,h
mVT cos
2m cos
d
T

recalling from (2.5) that

L = qSCL =

1
(h)VT2 SCL
2

Since dCL /d is negative in the post-stall region, see Figure 5.1, will be positive.
This means that the higher speed VT , and the larger sideslip one wants the control

64

Backstepping Designs for Flight Control

law to handle, the higher becomes, and the higher the control law parameters
k,1 and k,2 must be chosen.
To solve this, we impose that y y , where y R6 is selected to represent
the flight envelope of interest. For instance, a practically valid assumption may be
that the sideslip is always less than 10 degrees. The final expression for then
becomes
f (, y )
= max

Let us now instead apply the linearizing control law in Corollary 5.2. Using
Table 5.2 we get
u2 = k,2 (qs + k,1 ( ref ) + f (, y ))

(5.26)

The only difference to (5.25) is that now f takes as its first argument rather
than ref . This causes the feedback loop to depend on CL , and robustness against
model errors in CL becomes more difficult to analyze.
Somewhat surprisingly, the control law in (5.26) is identical1 to the one in (2.15),
which was derived using dynamic inversion and time-scale separation arguments.
Using our Lyapunov based backstepping approach, we have thus shown this control
law to be not only globally stabilizing, but also inverse optimal w.r.t. a meaningful
cost functional, according to Corollary 5.2.
Sideslip regulation
Applying the linear control law in Corollary 5.1, using Table 5.2 along with (5.8)
and (5.11) yields
u3 = k,2 (rs + k,1 +

1
g cos sin )
VT

(5.27)

The parameters restrictions

k,2 > 2 k,1 ,

k,1 > max{0, }

ensure the control law to be globally stabilizing and to be optimal w.r.t. a meaningful cost functional. Here,
= max
x1

(x1 )
f (, y ) f (0, y )
= max
,y
x1

The first two terms of f in (5.8) both give negative contributions to . Since these
typically are superior to the gravity contribution, < 0 holds and the parameter
restrictions above simply reduce to
k,2 > 2 k,1 > 0
Thus, precise knowledge of the side force is not necessary to implement the globally
stabilizing control law (5.27).
1 The

slight difference is due to that (2.15) was derived using somewhat simplified dynamics.

5.2 Flight Path Angle Control

65

Stability axis roll control

Finally, controlling the stability axis roll is straightforward. Given the dynamics
from (5.3a) and the roll rate command pref
s , simply assign
u1 = kps (pref
s ps )

(5.28)

where kps > 0. This corresponds to ordinary proportional control.

5.1.4

Practical issues

Let us now turn to some practically relevant issues regarding the application of the
derived flight control laws.
Tuning
How should the control law parameters kps , k,1 , k,2 , k,1 , and k,2 be selected?
For roll control, kps can be chosen to satisfy a given requirement on the roll time
constant, which becomes 1/kps .
For control, the closed loop system will not be linear since the nonlinear lift
force L is not cancelled by the control law (5.25). However, since the control law is
linear in and qs , it is tempting to still use linear techniques. A natural procedure is
to linearize the angle of attack dynamics (5.3b)(5.3c) around a suitable operating
point and then select k,1 and k,2 to achieve some desired linear closed loop
behavior locally around the operating point.
For regulation, the situation is the same. Here, k,1 and k,2 , determining the
control law (5.27), can be selected by choosing some desired closed loop behavior
using a linearization of the sideslip dynamics (5.3d)(5.3e).
Saturation
The control laws (5.25), (5.27), and (5.28) all can handle a certain amount of
gain reduction and still remain stabilizing. Thus, even in cases where actuator
saturation makes the moment equation (5.5) infeasible, the control laws remain
stabilizing within certain bounds. The maximum moment that can be produced
depends on the aircraft state, see Figure 2.6. This makes it difficult to determine the
exact part of the state space (in terms of , , etc.) where stability is guaranteed,
and we will not further pursue this issue.

5.2

Flight Path Angle Control

We now turn to flight path angle control. This is not a standard autopilot function, but may still be of interest, e.g., for controlling the ascent or descent of an
unmanned aerial vehicle.

66

5.2.1

Backstepping Designs for Flight Control

Objectives, dynamics, and assumptions

The objective is for the flight path angle, depicted in Figure 2.7, to follow a given
command, ref . In other words, we want to make
= ref
a globally asymptotically stable equilibrium.
We consider only the longitudinal motion of the aircraft, assuming that the
roll and sideslip angles are zero. Again, speed control is assumed to be handled
separately. The relevant dynamics are then given by (2.9b)(2.9d). Using the
definition of the flight path angle, = , yields
1
(L + FT sin mg cos )
mVT
= q
1
q = (M + FT ZT P )
Iy

(5.29)

Preparatory for the backstepping design, we make the following assumptions

and simplifications.
A1. The lift force coefficient, CL , is assumed to be a function of alone.
A2. The time derivatives of the aircraft speed and altitude are neglected.
A3. The contribution to due to gravity is replaced by its value at the desired
equilibrium, ref . This means cos cos ref is used. The idea is to make
a function of = alone.
A4. The control surface actuator dynamics are assumed to be fast enough to be
disregarded.
Also introducing
u = q =

1
(M + FT ZT P )
Iy

(5.30)

the aircraft dynamics (5.29) become

1
(L() + FT sin mg cos ref)
mVT
= q
q = u

(5.31a)
(5.31b)
(5.31c)

The control law will be derived considering u as the input. The relationship to
the true control input, , affecting the pitching moment, M , is given by
M () = Iy u FT ZT P
In Chapter 7 we discuss how to solve for given u.

(5.32)

5.2 Flight Path Angle Control

5.2.2

67

Backstepping control design

The nonlinear control problem

We begin by posing the control problem above as a more general nonlinear control
problem. We first define new variables such that the origin becomes the desired
equilibrium.
x1 = ref
x2 = ( ref + 0 )
x3 = q
(x2 x1 ) = ( 0 ) =

(5.33)

1
(L() + FT sin mg cos ref)
mVT

0 is the angle of attack at steady state, solving = (0) = 0. This gives us the
dynamics
x 1 = (x2 x1 )
x 2 = x3

(5.34a)
(5.34b)

x 3 = u

(5.34c)

considering the reference ref as a constant.

The nonlinearity is decided by the lift force, see Figure 5.1, and the contribution due to the thrust force. For . 90 degrees, a valid conclusion is therefore
that belongs to the first and third quadrants only. Mathematically, this can be
written
x(x) > 0 x(x) < 0,

x 6= 0

(5.35)

In the forthcoming design, we will show that this scarce information suffices to
construct a globally stabilizing control law.
The backstepping design
The key idea of the forthcoming design is the following. For x2 = 0, we get
x 1 = (x1 ), which acts stabilizing since (x1 ) is (x1 ) mirrored about the
y-axis, and thus lies in the second and fourth quadrants. Using backstepping, we
will show how to utilize this inherent stability property.
Below, ki are constants which parameterize the control law, while ci are dummy
constants whose values will be assigned during the derivation to simplify various
expressions.
Step 1: As usual, start by considering only the x1 -subsystem (5.34a). To find a
globally stabilizing virtual control law, we use the clf
V1 =

1 2
x
2 1

(5.36)

68

Backstepping Designs for Flight Control

Considering x2 as our virtual control input

V 1 = x1 (x2 x1 )
= x1 ((1 + k1 )x1 + x2 + k1 x1 )
= x1 ((1 + k1 )x1 ) < 0,

x1 6= 0

can be achieved by selecting

x2 = xdes
2 = k1 x1 ,

k1 > 1

(5.37)

The fact that k1 = 0 is a valid choice means that x1 feedback is not necessary
for the sake of stabilization. However, it provides an extra degree of freedom for
tuning the closed loop performance.
Step 2: Since we cannot control x2 directly, we continue by introducing the
deviation from the virtual control law.
x
2 = x2 xdes
2 = x2 + k1 x1
Including the x2 dynamics (5.34b) we get
x 1 = ()
x
2 = x3 + k1 ()
where
2
= (1 + k1 )x1 + x

(5.38)

has been introduced. We will also need

= (1 + k1 )() + x3 + k1 () = () + x3
A regular backstepping design would proceed by expanding the clf (5.36) with a
term penalizing x
2 . We do this, but also add a term F () as an extra degree of
freedom, where F is required to be positive definite. Hence,
V2 =

c1 2 1 2
x + x
+ F (),
2 1 2 2

c1 > 0

We compute its time derivative to find a new virtual control law, xdes
3 .
2 (x3 + k1 ()) + F 0 ()(() + x3 )
V 2 = c1 x1 () + x
= (c1 x1 + k1 x
2 F 0 ())() + (
x2 + F 0 ())xdes
x2 + F 0 ())(x3 xdes
3 + (
3 )
Although it may not be transparent, we can again find a stabilizing function independent of . Choosing
xdes
2 ,
3 = k2 x
0

F () = c2 (),

k2 > 0
F (0) = 0,

(5.39)
c2 > 0

(5.40)

5.2 Flight Path Angle Control

69

where (5.40) is an implicit but perfectly valid choice of F , yields

V 2 |x3 =xdes
= (c1 x1 + (k1 k2 c2 )
x2 )() c2 ()2 k2 x
22
3
|
{z
}
(k1 k2 c2 )

To make the first term negative definite using (5.35), we select c1 to make the
factor in front of () proportional to , see (5.38). This is achieved by
c1 = (1 + k1 )(k1 k2 c2 ),

k2 c2 > k1

(5.41)

With this choice,

V 2 |x3 =xdes
= (k1 k2 c2 )() c2 ()2 k2 x
22
3
becomes negative definite. The benefit of using the extra term F () shows up in
Equation (5.41). F () 0 (or equally, c2 = 0) leads to c1 = (1 + k1 )k1 > 0 and
the severe restriction 1 < k1 < 0 implying positive feedback from x1 .
Step 3: The final backstepping iteration begins with introducing the third residual
x
3 = x3 xdes
2
3 = x3 + k2 x
We also update the system description
x 1 = ()
x
2 = x
3 k2 x
2 + k1 ()
x
3 = u + k2 (
x3 k2 x
2 + k1 ())

(5.42)

Furthermore,
= () + x
3 k2 x
2
V3 is constructed by adding a term penalizing x3 to V2 .
1 2
,
V3 = c3 V2 + x
2 3

c3 > 0

We get


22 + x
3 (
x2 + c2 ())
V 3 = c3 (k1 k2 c2 )() c2 ()2 k2 x
|
{z
}
negative definite



x3 k2 x
2 + k1 ())
+x
3 u + k2 (


c2 c3 2 () k2 c3 x
22 + x3 u + k2 x
3 + (c3 k22 )
x2 + (k1 k2 + c2 c3 )()
2 x
3 cross-term and try yet
once again using (5.35). Select c3 = k22 to cancel the x
another linear control law.
u = k3 x
3 ,

k3 > k2

(5.43)

70

Backstepping Designs for Flight Control

is a natural candidate and with this we investigate the resulting clf time derivative.
V 3 k22 c2 2 () k23 x
22 (k3 k2 )
x23 + (k1 k2 + k22 c2 )
x3 ()
In order to investigate the impact of the last cross-term, we complete the squares.
k1 k2 + k22 c2
())2
22 (k3 k2 )(
x3
V 3 k23 x
2(k3 k2 )
(k1 k2 + k22 c2 )2 2
) ()
(k22 c2
4(k3 k2 )
V 3 is negative definite provided that the 2 () coefficient is negative, which is true
for
(k1 + k2 c2 )2
)
(5.44)
k3 > k2 (1 +
4k2 c2
We now pick c2 to minimize this lower limit under the constraints c2 > 0 and
k2 c2 > k1 .
For k1 0, we can make k1 + k2 c2 arbitrarily small whereby Equation (5.44)
reduces to
k3 > k2
i.e., the same restriction as in (5.43). For k1 > 0 the optimal strategy can be shown
to be selecting c2 arbitrarily close to the bound k1 /k2 . This yields
k3 > k2 (1 + k1 )

(5.45)

Let us summarize this lengthy control law derivation, which resulted in a globally stabilizing control law (5.43) for the system (5.34), under the parameter restrictions in (5.37), (5.39), (5.43), and (5.45).
Proposition 5.3
Consider the system
x 1 = (x2 x1 )
x 2 = x3
x 3 = u

(5.46)

where x(x) is positive definite. A globally stabilizing control law is given by

u = k3 (x3 + k2 (x2 + k1 x1 ))

(5.47)

k1 > 1
k2 > 0

k1 0
k2
k3 >
k2 (1 + k1 ) k1 > 0

(5.48)

where


The cascaded structure of the control law is illustrated in Figure 5.5.

5.2 Flight Path Angle Control

k1

xdes
2

k2

xdes
3

71
u

k3

x2

x3
x 3 = u

x 2 = x3

x 1 = (x2 x1 )

x1

Figure 5.5 The nonlinear system (5.46) can be globally stabilized through
a cascaded control structure.

Robustness
For this design we refrain from investigating inverse optimality due to the complicated expressions involved. Regardless of this, it is clear that the control law
(5.47) gives a certain gain margin. E.g., in the case of k1 > 0, we can afford a gain
reduction of
>

k2 (1 + k1 )
k3

(5.49)

at the input (see Figure 4.1) without violating k3 > k2 (1 + k1 ) from (5.48).
Backstepping vs. feedback linearization
It is rewarding to compare the preceding backstepping design with a control design
based on feedback linearization. Such a design makes the open loop system a chain
of integrators by defining new coordinates according to
2
x 1 = (x2 x1 ) = x
0
x
2 ) = x3
2 = (x2 x1 )(x3 x
00
2 )2 + 0 (x2 x1 )(u x
3 ) = u

x
3 = (x2 x1 )(x3 x
We can now select
u
= k1

k2

k3

x1
x
2
x
3

to achieve any desired linear relationship between u

and x1 .
However, solving for the actual control input u to be produced, we get
u=x
3 +

u
00 (x2 x1 )(x3 x
2 )2
0 (x2 x1 )

(5.50)

72

Backstepping Designs for Flight Control

Two things are worth noting about this expression. Firstly, it depends not only
on (through x
2 ), but also on its first and second derivatives, which therefore
must be known. In the aircraft control case, this corresponds to very accurate
knowledge of the lift force, see (5.33). Since L() in practice comes with certain
model error, especially at high angles of attack, the estimates of L0 () and L00 ()
may be poor. This means that the nonlinear system behavior cannot be cancelled
completely. Unfortunately, it is difficult to analyze the robustness of (5.50), i.e.,
how incomplete cancellation of the nonlinearities affects the controlled system.
Secondly, 0 is in the denominator of (5.50) implying that the control law has
a singularity where 0 = 0. In the aircraft case, this occurs around the stall angle,
where the lift force no longer increases with , see Figure 5.1. Thus, global stability
cannot be achieved using feedback linearization.
Our backstepping design did not suffer from any of the problems above, since
all we required from was for x(x) to be positive definite, see (5.35).

5.2.3

Flight control law

We now return to the flight control context. Expressing (5.47) in the original
coordinates using (5.33) gives us
u = k3 (q + k2 ( + k1 ( ref ) ref 0 ))

(5.51)

This control law is globally stabilizing provided that k1 , k2 , and k3 satisfy (5.48).
Recall that 0 is the angle of attack at steady state, solving =0

in (5.31a).

5.2.4

Practical issues

We now investigate some practical issues regarding the application of this control
law.
Tuning
For selecting the controller parameters, k1 , k2 , and k3 , we can use the same strategy
as for tuning the general maneuvering control laws, cf. Section 5.1.4. This means
linearizing the dynamics (5.31) around a suitable operating point and selecting the
controller parameters to achieve some desired linear closed loop behavior locally
around this operating point.
Saturation
As shown in Section 5.2.2, the control law (5.51) remains stabilizing in the presence of a gain certain reduction, given by (5.49) in the case k1 > 0. Thus, even
for a certain amount of control surface saturation, such that the desired pitching
moment in (5.32) cannot be produced, the closed loop system is stable. As noted

5.2 Flight Path Angle Control

73

in Section 5.1.4, the maximum moment, M , depends not only on the control surface deflections, , but also on the angle of attack, see Figure 2.6. Again, this
makes is difficult to determine the part of the state space within which stability is
guaranteed.

74

Backstepping Designs for Flight Control

6
Adapting to Input
Nonlinearities and Uncertainties

The flight control laws developed in Chapter 5 consider the angular accelerations
as the control input, u. To find the corresponding control surface deflections, , the
mapping from to u must be completely known. Typically this is not the case in
practice. For example, the aerodynamic moment coefficients suffer from inevitable
model errors, and the moment contributions from, e.g., the engine thrust, may
not be measurable. Hence, it is necessary to add integral action in some form to
the control laws to reach the desired equilibrium despite such model imperfections.
This is the topic of the chapter.
In Section 6.1, we further illustrate the problem, and in Section 6.2, we present
its mathematical formulation. Two solutions, based on adaptive backstepping and
nonlinear observer techniques, respectively, are proposed in Sections 6.3 and 6.4.
The two solutions are evaluated in Section 6.5 using a water tank example, and in
Section 6.6 the adaptive schemes to be used for flight control are explicitly stated.

6.1

Background

Many of todays constructive nonlinear control design methods assume the control
input to enter the system dynamics affinely, i.e., for the model to be of the form
x = f (x) + g(x)u
75

76

Adapting to Input Nonlinearities and Uncertainties

In many practical cases this is not true. A common solution, see, e.g., [37, 47], is
to find some other entity, a virtual control input v, that does enter the dynamics
linearly, and that depends on the true control input u through a static mapping.
Using, e.g., backstepping or feedback linearization, a globally stabilizing control
law v = k(x) can then be derived. These virtual control inputs are often physical
entities like forces, torques, or flows, while the true input might be the deflection of
a control surface in a flight control case or the throttle setting in an engine control
case.
The remaining problem, how to find which actual control input u to apply, is
often very briefly discussed, typically assuming that the mapping from u to v is
completely known and invertible. In this chapter we investigate the case where the
mapping is only partially known. It might be that the true mapping is too complex
to identify, or that other sources than u contribute to v. Friction might for example
reduce the net torque in a robot control case. Here, we will pragmatically model
the discrepancy between the model and the true mapping as a constant bias. We
propose two different ways of adapting to the bias, and for each case, the issue of
closed loop stability is investigated.

6.2

Problem Formulation

Consider a single input system of the form

x = f (x) + Bv
v = g(x, u)

(6.1)

x Rn is the measurable state vector and u R is the true control input.

B= 0

T
1

(6.2)

is such that only the last state, xn , is directly affected by the control input through
x n = fn (x) + g(x, u)
Assume that the mapping, g(x, u), from the true control input, u, to the virtual
control input, v, is not completely known but only a model, g(x, u), such that
g(x, u) = g(x, u) + e
The model error, e = g(x, u) g(x, u), is modeled as a constant. This pragmatic
assumption may be more or less realistic but allows us to correct for biases and
reach the correct equilibrium at steady state. With this we can rewrite (6.1) as
x = f (x) + B(w + e)

(6.3a)

w = g(x, u)

(6.3b)

w is the part of the virtual control input, v, that we are truly in control of.

6.2 Problem Formulation

77

e
+

k(x)

+
v

x = f (x) + Bv

Estimator

Figure 6.1 Illustration of the certainty equivalence controller (6.7). The

shadowed box represents the dynamic system to be controlled.
The dashed box makes up the system that the control designer
actually faces, since e is not known.

From a preceding control design, a control law

v = k(x)

(6.4)

is assumed to be known such that the origin is a GAS equilibrium of

x = f (x) + Bk(x)
We also assume that a Lyapunov function V (x) for the closed loop system is known,
such that
V (x) = Vx (x)(f (x) + Bk(x)) = W (x)

(6.5)

where W (x) is positive definite.

Given e, (6.4) could be realized by solving
w = v e g(x, u) = k(x) e

(6.6)

for u. How do we deal with the fact that e is not available? A straightforward solution is to rely on one of the corner stones of adaptive control and use the certainty
equivalence [46] of (6.6). This means that we replace the unknown parameter e by
an estimate e and form
w = g(x, u) = k(x) e

(6.7)

Figure 6.1 illustrates the approach. The strategy is intuitively appealing but leads
to two important questions:

78

Adapting to Input Nonlinearities and Uncertainties

How do we estimate e?
Can we retain global stability using e for feedback?
Two approaches to the problem will be pursued. In Section 6.3, we will use
standard adaptive backstepping techniques to find an estimator that will guarantee
closed loop stability without having to adjust the control law (6.7). In Section 6.4,
the starting point is that a converging estimator is given. The question then is
how to adjust the control law to retain stability. This approach is due to the
author, but was inspired by the observer backstepping techniques introduced by
Kanellakopoulos et al. [39].

6.3

Adaptive Backstepping

Adaptive backstepping [46] deals with the unknown parameter e by extending the
Lyapunov function V (x) with a term penalizing the estimation error e = e e:
Va (x, e) = V (x) +

1 2
e ,
2

>0

By cleverly selecting the update rule

e = (x, e)
closed loop stability can be guaranteed. To see this we investigate V a when (6.7)
is used as feedback. Using (6.2) and (6.5) we get

1
V a = Vx (x) f (x) + B(k(x) e + e) e (x, e)

V (x)
1
= W (x) + e(
(x, e))
xn

(6.8)

The first term is negative definite according to the assumptions, while the second,
mixed term is indefinite. Since e is not available, the best we can do is to cancel
the second term by selecting
V (x)
xn

(6.9)

x = f (x) + B(k(x) + e)

(6.10)

(x, e) = (x) =
The resulting closed loop system becomes

V (x)
e =
xn
which satisfies
V a (x, e) = W (x)

6.4 Observer Based Adaption

79

Despite V a only being negative semidefinite, the origin, x = 0, e = 0, is GAS

according to Corollary 3.1. Since only x = 0 solves V a = 0, e = 0 must also hold
for x to be zero.
If V is quadratic in xn , V = . . . + 12 x2n , the control law (6.7) becomes
Z
w = k(x)

xn (s)ds
0

In this case, estimating e and using the estimate for feedback corresponds to adding
integral action from xn .

6.4

Observer Based Adaption

In adaptive backstepping, the estimator was a consequence of assigning a negative

Lyapunov time derivative. In this section, we first design an estimator and then
investigate how to possibly adjust the control law (6.7).

6.4.1

The general case

The idea is to regard e as an unknown but constant state variable. Augmenting

the original dynamics (6.3a) with this extra state yields
x = f (x) + B(w + e)
e = 0
Although this system is nonlinear, we can design an observer for e with linear error
dynamics, since the nonlinearity, f , is a function of the measurable states x only
[44].
We will use xn alone to measure of the goodness of the estimate. In the presence
of measurement noise, it would be advantageous to utilize the entire state vector,
which is also fully possible. The nonlinear observer becomes
  
  
d x
n
k
f (x) + w + e
+ 1 (xn x
n )
(6.11)
= n
0
k2
dt e
With this, the dynamics of the estimation error


n
xn x
=
e e
become linear:

k1
 =
k2


1
 = A 
0

(6.12)

80

Adapting to Input Nonlinearities and Uncertainties

For any asymptotically stabilizing observer gains k1 and k2 , making A Hurwitz1 ,

we can find a positive definite matrix P such that
d T
 P  = T Q q
e2
dt
by solving the Riccati equation
AT P + P A = Q,

Q = qI,

q>0

according to basic linear systems theory, see Rugh [62]. A is Hurwitz if and only
if
k1 > 0,

k2 > 0

To investigate the closed loop stability, we combine the original Lyapunov function V (x) with T P  and form
Vo (x, ) = V (x) + T P 
We also augment the control law (6.7) with an extra term, l, to be decided, to
compensate for using e for feedback.
w = k(x) + l(x, e) e

(6.13)

yields


V o = Vx (x) f (x) + B(k(x) + l(x, e) e + e) T Q
W (x) +

V (x)
(l(x, e) + e) q
e2
xn

By choosing
l(x, e) = l(x) =

V (x)
,
xn

>0

(6.14)

we can complete the squares.

V (x)
1 2
1 2
e) (q
)
e
V o W (x) (

xn
2
4
1
To achieve GAS, we must satisfy q 4
> 0, which can always be done once in
(6.14) has been selected, since q is at our disposal.
Let us summarize our discussion.

Proposition 6.1 (Observer based adaption)

Consider the system
x i = fi (x),

i = 1, . . . , n 1

x n = fn (x) + w + e
1A

matrix is said to be Hurwitz if all its eigenvalues are in the open left half plane.

(6.15)

6.4 Observer Based Adaption

81

where x R is the measurable state vector, w R is the control input, and e R

is an unknown constant. Let w = k(x) e be a globally stabilizing control law
and let V (x) be a Lyapunov function for the controlled system such that V (x) is
negative definite.
Then, the control law
w = k(x)

V (x)
e,
xn

>0

where e is produced by the observer

  
  
d x
n
k
fn (x) + w + e
+ 1 (xn x
n ),
=
0
k2
dt e

k1 > 0, k2 > 0


is also globally stabilizing.

An interesting feature is displayed by computing an explicit expression of the

estimate produced. Using (6.12) we have that


e(t) = e e = e 0 1 eA t (0)
This means that e evolves independently of the control input u(t) and the state
trajectory x(t). This is not true for the adaptive backstepping estimate.

6.4.2

The optimal control case

Let us consider the case where the original, unattainable control law (6.4) solves
an optimal control problem of the form
Z
V (x) = min
q(x) + r(x)v 2 dt
(6.16)
v

where r(x) r0 for all x. According to (4.6) it then holds that

k(x) =

1 V (x)
2r(x) xn

We recall from Section 4.3 that a fundamental property of control laws minimizing
a criterion like (6.16) is that they have a gain margin of [ 12 , ]. This inherent
robustness means that we do not need to modify the certainty equivalence control
law (6.7) to retain stability, since l(x) in Equation (6.14) is proportional to k(x).
To show this we make the split
V (x)
V (x)
1 V (x)
+

k(x) =
2r(x) xn
xn
xn
|
{z
} | {z }

k(x)

l(x)

82

Adapting to Input Nonlinearities and Uncertainties

x2

x2

x1

x1
Figure 6.2 Two tanks connected in series.
Now,

v = k(x)
= (

V (x)
1
)
2r(x)
xn

is guaranteed to be globally stabilizing, i.e., to make W (x) in (6.5) negative definite

given that the gain reduction due to is less than 50%. Thus,
1
1
1
1
1

2r(x)
2 2r(x)
4r0
4r(x)
must hold which does not contradict the only previous requirement from (6.14)
that > 0.
An intuitive interpretation of this result is that some of the optimal control effort
can be sacrificed in order to compensate for using the estimate e for feedback.

6.5

A Water Tank Example

Let us apply the two strategies to a practical example to investigate their pros
and cons. Consider the two tanks in Figure 6.2. The control goal is to achieve a

6.5 A Water Tank Example

83

certain water level r in the bottom tank. Using Bernoullis equation and setting
all constants to unity, the system dynamics become

x 1 = x1 + x2

x 2 = x2 + v
where x1 = water level of the lower tank, x2 = water level of the upper tank, and
v = incoming water flow. v is produced by changing the aperture of the valve of
the input pipe.
We assume the dynamics of the valve to be very fast compared to the dynamics
of the tanks, so that the relationship between the commanded aperture radius, u,
and the water flow, v, can be regarded as static. Assuming some external water
supply to keep a constant pressure, v will be proportional to the aperture opening
area, which in turn depends on u2 . Again setting all constants to unity we would
have v = u2 . In order to be able to account for a possible model error in this static
relationship and for other sources contributing to the net inflow, e.g., leakage, we
assign the model
v = u2 + e
in accordance with (6.3a).
The first step is to find a globally stabilizing control law v = k(x). We do this
using an ad hoc Lyapunov approach. At the desired steady state, x1 = x2 = r.
Therefore consider the control Lyapunov function
V (x) =

1
a
(x1 r)2 + (x2 r)2 ,
2
2

a>0

Compute its time derivative:

V (x) = a(x1 r)( x1 + x2 ) + (x2 r)( x2 + k(x))

By collecting the beneficial terms and cancelling the indefinite ones, one finds that
k(x) =

a
(r x1 ) + b(r x2 ),
r+
x2 + r

a > 0, b 0

(6.17)

yields V (x) = W (x) where

W (x) = a(x1 r)( x1 r) + (x2 r)( x2 r) + b(x2 r)2

is positive definite.
Let us now evaluate the expressions involved with the two approaches for adapting to the leakage. The adaptive backstepping update rule (6.9) for estimating e
becomes
V (x)
= (x2 r),
e =
x2

>0

(6.18)

84

Adapting to Input Nonlinearities and Uncertainties

Adaptive
backstepping
= 0.3

k(x)
a=1
b = 0.5

Observer
based adaption
k1 = 1
k2 = 0.5
=0

Table 6.1 Controller parameter values used in the simulations.

With this, the implicit control law (6.7) becomes

Z

(r x2 (s))ds

w = u = k(x) +
0

Using the observer based approach, the estimator can be designed according to
(6.11). For the actual implementation, we can rewrite this as
  
   

d xn
k1 1
xn
1
=
+
( x2 + w)
e
k2 0
0
dt e
which can be implemented using, e.g., Simulink. The implicit control law (6.13)
becomes
w = u2 = k(x) + (r x2 ) e,

>0

If b > 0 was selected in the control law (6.17), we do not have to add the term
(r x2 ) for the sake of stability, since it can be seen as a part of k(x) already. As
in the optimal control case treated in Section 6.4.2, closed loop is then guaranteed
using the original certainty equivalence control law (6.7) without any modification.
In the simulations, the parameter values for the two adaptive controllers were
selected according to Table 6.1. The initial water level, which is also fed to the
observer, is 1 in both tanks. The control goal is to for x1 to reach the reference level
r = 4 and maintain this despite the leakage e = 3 starting at t = 25 s. Figure
6.3 shows the actual control input and the water level of the lower tank when
no adaption is used. Figures 6.4 and 6.5 show the results of applying adaptive
backstepping and observer based adaption, respectively.
There is a striking difference between the initial behaviors of the two leakage
estimates. As pointed out in Section 6.4.1, the observer e estimate evolves independently of u and x. Since (0) = 0, the estimation error remains zero until the
leakage starts. The adaptive backstepping estimate on the other hand depends on
the integral of r x2 over time, causing an oscillatory behavior due to the initial
error of the upper tank water level.
Also, in the presence of actuator saturation, adaptive backstepping will suffer
from the windup problems that generally occur when using integral action in the
feedback loop. This is avoided with the observer based approach if the observer is
fed with the true, saturated value of the control input.

6.5 A Water Tank Example

85

u
2.6

x1

4.5

2.4

2.2
3.5

2
1.8

1.6

2.5

1.4

1.2
1.5

1
0.8
0

10

20
30
time (s)

40

50

1
0

10

20
30
time (s)

40

50

Figure 6.3 No adaption, pure state feedback.

u
2.6

x1

4.5

2.4

e
1.5
1

2.2

0.5

3.5

2
1.8

1.6

2.5

1.4

0
0.5
1
1.5

1.2
1.5

1
0.8
0

2.5

10

20
30
time (s)

40

50

1
0

10

20
30
time (s)

40

50

3.5
0

10

20
30
time (s)

40

50

40

50

Figure 6.4 Adaptive backstepping.

u
2.6

4.5

2.4

x1

e
1.5
1

2.2

0.5

3.5

2
1.8

1.6

2.5

1.4

0
0.5
1
1.5

1.2
1.5

1
0.8
0

2.5

10

20
30
time (s)

40

50

1
0

10

20
30
time (s)

40

50

3.5
0

Figure 6.5 Observer based adaption.

10

20
30
time (s)

86

6.6

Adapting to Input Nonlinearities and Uncertainties

Application to Flight Control

Considering the appealing properties of the observer based adaption, as demonstrated in the previous section, this will be our choice of adaptive scheme for flight
control. All of the state feedback laws to be used along with the observers, (5.25),
(5.27), (5.28), and (5.51), can afford a certain amount of gain reduction and still be
globally stabilizing, as shown in Chapter 5. As shown in Section 6.4.2, this means
that the certainty equivalence control law (6.7) is globally stabilizing.
Let us now state the resulting observers when Proposition 6.1 is applied to the
flight dynamics used for control design in Chapter 5.

6.6.1

General maneuvering

The relevant dynamics are given by (5.3). To handle unmodeled nonlinearities and
uncertainties in the mapping (5.4) from to u, we redefine (5.3a), (5.3c), and (5.3e)
as
ps = u1 + e1
qs = u2 + e2
rs = u3 + e3

Comparing these dynamics to (6.15) yields the observers

  

d ps
u1 + e1
=
+ L1 (ps ps )
0
dt e1
  

d qs
u2 + e2
=
+ L2 (qs qs )
0
dt e2
  

d rs
u3 + e3
=
+ L3 (rs rs )
0
dt e3

(6.19)

L1 , L2 , and L3 are 2 1-vectors whose entries must be positive for the estimates
to converge.

6.6.2

Flight path angle control

The dynamics are given by (5.31). As above, (5.31c) is redefined as

q = u + e
to handle unmodeled nonlinearities and uncertainties in the mapping (5.30) from
to u. Using Proposition 6.1, an observer for e is given by
  

d q
u + e
=
+ L(q q)
(6.20)
0
dt e
where L is a 2 1-vector with positive entries.

7
Implementation and Simulation

The control designs of the preceding chapters were based on a number of pragmatic,
simplifying assumptions regarding the aircraft dynamics and the types of model
errors and uncertainties to be handled. Using computer simulations we will now
evaluate the control laws experimentally.
The aircraft models used for simulation are presented in Section 7.1. Some
implementation details are covered in Section 7.2, while Section 7.3 is devoted to
the actual computer simulations.

7.1

Aircraft Simulation Models

Due to the complexity of the dynamics and the military nature of the field, there
exist only a few available aircraft simulation models. Three of these are listed in
Table 7.1.

7.1.1

GAM/ADMIRE

The Generic Aerodata Model (GAM) contains aerodynamic data for a small fighter
aircraft, not unlike JAS 39 Gripen [63]. The model was produced and made available by Saab AB, Linkoping, Sweden. A model description is given by Backstr
om
[5], and the package can be downloaded from [64]. A disadvantage of the GAM is
87

88

Implementation and Simulation

Simulation
model
GAM/ADMIRE
HIRM
FDC

Publicly
available
Yes
Not in general
Yes

Fighter aircraft
dynamics
Yes
Yes
No

Table 7.1 Three existing environments for aircraft simulation.

that aerodynamic data is only recorded for 10 30 and 20 20 .
In this flight envelope, the aerodynamic efforts are not significantly nonlinear, see
Figure 2.6. Basic facts regarding the GAM can be found in Appendix 7.A.
To perform flight simulations based on the GAM, there exists a Simulink
wrapper called ADMIRE. We will use the GAM/ADMIRE environment to evaluate
the general maneuvering control laws developed in Section 5.1. The GAM was
recently used by Cronander [12] to evaluate a pitch rate controller based on dynamic
inversion.

7.1.2

HIRM

The High Incidence Research Model (HIRM) was developed by DERA1 of the
United Kingdom. The model is based on aerodynamic data from wind tunnel
tests and drop tests of a small-scale model. The HIRM was then derived by scaling up these data to create an aircraft of F-18 proportions, see Appendix 7.A.
Aerodynamic data exist for a wide range of angles of attack and sideslip angles,
50 120 and 50 50 .
The HIRM was used as a benchmark fighter aircraft model in the robust flight
control design challenge [53] initiated by GARTEUR2 . A technical description of
the HIRM, which is implemented as Simulink model, can be found in [56].
We will use the HIRM to evaluate the flight path angle control law developed
in Section 5.2. DERA is gratefully acknowledged for granting permission to use
the model for this purpose.

7.1.3

FDC

The Flight Dynamics and Control toolbox (FDC) by Rauw [59] is a Simulink
toolbox for general flight dynamics and control analysis. The toolbox comes with
aerodata from a small, non-military aircraft, but the modular structure of the
toolbox allows the user to plug in external aerodata of his or her choice. Although
promising, FDC has not been used for simulation since for the GAM and HIRM
aerodata, well functioning interfaces already exist.
1 Defence
2 Group

Evaluation and Research Agency

for Aeronautical Research and Technology in EURope

7.2 Controller Implementation

ref , pref
s

89

Backstepping k(x)
control laws

udes

Control

allocation
u

e
Bias
estimation

Figure 7.1 Controller configuration.

7.2

Controller Implementation

The controller configuration is shown in Figure 7.1. The backstepping block contains the state feedback control laws derived in Chapter 5, while the bias estimator
block contains the nonlinear observers from Chapter 6, which are used to estimate
and adapt to unmodeled moments acting on the aircraft. The control allocation
block will be discussed in Section 7.2.1. Its function is to translate the desired
angular acceleration,
udes = k(x) e
into actual control surface deflections, . The signal u, which is fed to the estimator,
is the actual angular acceleration, which differs from udes when udes is not feasible.
By feeding the actual value u to the estimator we avoid wind-up problems.
The observers used are compactly stated in Section 6.6. Let us for convenience
also gather the control laws from Chapter 5. For general maneuvering the control
laws are made up by (5.25), (5.27), and (5.28). Using k(x) to denote these control
laws, we get

T
k(x) = k1 (x) k2 (x) k3 (x)
where
k1 (x) = kps (pref
s ps )
k2 (x) = k,2 (qs + k,1 ( ref ) + f (ref , y ))
1
g cos sin )
k3 (x) = k,2 (rs + k,1 +
VT
with f from (5.7). The flight path angle control law (5.51) reads
k(x) = k3 (q + k2 ( + k1 ( ref ) ref 0 ))

90

Implementation and Simulation

Let us make two remarks regarding the implementation of these controllers.

In the control designs in Chapter 5, the force effects due to the control surface deflections and the angular rates were neglected. Here, these effects are
included when f and 0 are computed in order to improve the accuracy.
0 , the steady state angle of attack, is obtained by solving = 0 in (5.31a) at
each time step, using current values of the variables involved. Although this
ad hoc approach to determine 0 is not guaranteed to converge, successful
simulations provide an alibi.

7.2.1

Control allocation

The control designs in Chapters 5 and 6 considered the angular accelerations of the
aircraft as the control inputs, see (5.4) and (5.30). In (5.5) and (5.32) we solved
for the actual moments to be produced. In most nonlinear aircraft designs it is
assumed that the control surface deflections, , affect these moments linearly. In
practice this is not true, since the aerodynamic moments produced by the control
surfaces also suffer from stall effects similar to the ones that are well known for the
lift force. This can be seen in Figure 2.6 where the pitching coefficient, Cm , tends
to saturate for low and high values of the symmetrical elevon deflection, es .
Here, we will take into account the nonlinear mapping from to the aerodynamic moments and propose numerical algorithms for finding the proper given
the moments to be produced. The implementations used are due to Press et al.
[58].
General maneuvering
The equation to be solved for is given by (5.5):
T des
M() = IRsb
u + I

Using the definition of M from Section 2.4, M() can be translated into the desired
aerodynamic moment coefficients,
C des = Cldes

des
Cm

Cndes

T

Introducing

T
C() = Cl () Cm () Cn ()

(7.1)

the control allocation problem now becomes to find such that

C() = C des

(7.2)

For simplicity we will only make use of the elevons and the rudder and ignore the
canard wings, see Figure 2.4. Thus, = (es , ed , r ). To handle cases where these

7.2 Controller Implementation

91

control surfaces saturate, and (7.2) cannot be satisfied, we reformulate the control
allocation problem as an optimization problem:
= arg min J()

where J() = kC() C des k2

where refers to the set of feasible control surface settings.
Nonlinear optimization is a complex matter and in general no guarantee can be
given that the global optimum is found. However, since in our case the optimization
is performed at each sample time, we can use the solution at one time step as the
initial guess at the next time step. This strategy of each time starting very close
to the optimal solution yields very good performance.
In the implementation, the Broyden-Fletcher-Goldfarb-Shanno (BFGS) variable
metric method [13], a quasi-Newton method, was chosen. This takes J() and
the gradient J() as arguments, where the latter can be numerically computed.
It recursively estimates the inverse Hessian of J and performs a line search in
this direction to update . Reusing the inverse Hessian at the next time step
significantly improves the performance.
Flight path angle control
The equation to be solved for is given by (5.32):
M () = Iy udes FT ZT P
where M corresponds to the aerodynamic pitching moment to be produced. Recalling that M = qS
cCm we can explicitly solve for the desired pitching coefficient:
des
Cm
=

1
(Iy udes FT ZT P )
qS
c

Again we ignore the canard wings for simplicity, and only make use of the symmetrical elevon deflection, es . The control allocation problem then becomes solving
des
des
Cm (es ) Cm
=0
Cm (es ) = Cm

for es . Unlike above, we will use this formulation for finding es and not use an
optimization framework. In this scalar case, actuator saturation can be handled
separately.
The pitching coefficient, Cm , is usually a complicated function of the arguments
involved. However, given measurements of , q, etc., Cm becomes a close to monotone function of es , see Figure 2.6. The only exception is when the aforementioned
stall effects occur and the moment produced saturates. However, given the aerodata, these regions can be manually removed by virtually saturating es when Cm
saturates.
Many numerical solvers are well adapted to finding the zero of a monotone scalar
function. For the implementation, the Van Wijngaarden-Dekker-Brent method

92

Implementation and Simulation

[8, 19] was chosen. This method is a happy marriage between bisection, which ensures convergence, and inverse quadratic interpolation, which provides superlinear
convergence in the best-case scenario.

7.3

Simulation

We now turn to the actual simulations performed. Before showing the actual plots
in Section 7.3.3, Section 7.3.1 gives an account of the conditions surrounding the
simulations, and the controller parameters used are listed in Section 7.3.2.

7.3.1

Conditions

The simulations are performed under the following general conditions:

The simulation models, GAM/ADMIRE and HIRM, are both equipped with
sensor dynamics and actuator dynamics mimicing those encountered in a real
aircraft.
No noise is added to the models, i.e., we assume perfect sensing (besides the
dynamics of the sensors) and no external disturbances like wind gusts.
The dynamics of the aircraft model are the fully nonlinear dynamics introduced in Section 2.4 and not the simplified ones used for control design in
Chapter 5 (see Sections 5.1.1 and 5.2.1). Specifically this means that the
control surface deflections affect not only the aerodynamic moments but also
the aerodynamic forces.
The aerodynamic moment coefficients used for control allocation as described
in Section 7.2.1 agree with the ones in the aircraft models as far as possible.
Since, e.g., and are not measured, the effects of these are neglected in
the control allocation schemes.
Speed control is handled by separate controllers. For the HIRM, the speed
controller by Larsson [49] is used, and for the GAM/ADMIRE, a simple PIcontroller has been implemented.
The sensor information used by the flight control system is updated at a
frequence of 50 Hz for the GAM/ADMIRE and 80 Hz for the HIRM. The
control surface demands are sent to the aircraft at the same frequency.

7.3.2

Controller parameters

The controller parameters used for the simulations are given in Tables 7.2 and 7.3.
The state feedback parameters kps , k,1 , k,2 , k,1 , and k,2 have been chosen
according to the guidelines in Sections 5.1.4 and 5.2.4, to achieve suitable linear
dynamics around the initial states of the flight cases presented below. The observer
gains L1 , L2 , and L3 place the poles of the error dynamics (6.12) in 8 i, while
L gives the observer poles 2 i.

7.3 Simulation

kps
L1

93

ps control
2.0

T
16.0 65.0

k,1
k,2
L2

control
2.0
5.0

T
16.0 65.0

k,1
k,2
L3

control
2.0
5.0

T
16.0 65.0

Table 7.2 General maneuvering control parameters.

k1
k2
k3
L

control
0.4
1.2
2.1

T
4.0 5.0

Table 7.3 Flight path angle control parameters.

7.3.3

Simulation results

General maneuvering
To evaluate the general maneuvering control laws from Section 5.1 we use the
GAM/ADMIRE environment. The simulations are performed at an initial speed
of 0.5 Mach and at an altitude of 1000 m. The assessment maneuvers are the
following:
M1. Roll rate demand, pref
s = 150 deg/s. See Figure 7.2.
M2. Angle of attack demand, ref = 15 deg. See Figure 7.3.
M3. M1 and M2 performed simultaneously. See Figure 7.4. During this maneuver,
both and ps vary rapidly which means that assumption A3 in Section 5.1.1
is violated.
Let us make some comments regarding the simulation results.
During M1 and M2, the controlled variables, ps , , and all follow their
reference trajectories (given by the dashed curves) well.
The increase in the aircraft speed, VT , during M1 is due to our strategy of
rolling about the stability x-axis, rather than the body x-axis, as discussed in
Section 2.2. Since is small, the stability x-axes coincides with the velocity
vector. Initially, = 2.8 deg is necessary to have the lift force make up for
gravity. But after rolling 180 deg without changing , the same amount of
lift force is directed towards Earth, which causes the aircraft to dive and the
speed to increase.
In M2, we see that ps and are not completely unchanged despite the maneuver being constrained to the body xz-plane. The small perturbations are

94

Implementation and Simulation

caused by the feedback from the bias estimators. This is also the reason for
the small perturbations of the ed and r input signals.
Shifting our attention to M3 we see that the resulting aircraft trajectory is
not quite the superposition of the M1 and M2 trajectories, in terms of the
controlled variables. The roll rate response is still satisfactory, but and
oscillate more. The reason is (at least) twofold:
1. According to assumption A3 in Section 5.1.1, the controller assumes
to be constant, and vice versa. Since this is not the case here, the
roll axis will not coincide exactly with the stability axis as desired, see
Figure 2.3. As outlined in Section 2.2, this means and are no longer
decoupled, but during the roll, part of the angle of attack turns into
sideslip and vice versa.
2. During the initial phase of the step, the parameter estimates, e1 , e2 , and
e3 oscillate. The 0.2 peak in e2 causes the controller to believe some
external source is contributing to the pitching moment. Consequently,
the pitching moment produced by es is reduced, which leads to a reduction in q and that the increase in is temporarily stopped. Efforts
have been made to tune the observers to avoid these oscillations but the
ultimate solution is yet to be found.
Flight path angle control
The flight path angle control law from Section 5.2 is evaluated using the HIRM.
Here, the initial aircraft state is level flight at Mach 0.3 at an altitude of 1524 m
(5000 ft). A model error in the pitch coefficient, Cm , of 0.03 is introduced on
purpose to examine the robustness of the controller. This is the same error that was
used for evaluating the controllers in the GARTEUR robust flight control design
challenge [53]. The following assessment maneuver is used:
M4. Two consecutive flight path angle demands, ref = 25 deg followed by ref =
15 deg. See Figure 7.5.
Let comment on the simulation results.
Overall, the flight path angle, , follows its reference trajectory well.
The small initial dip of is caused by the model error introduced, which
makes the true pitching moment less than the controller expects. Soon however, the model error is estimated by the observer, as seen in the e plot, and
the controller compensates for the error and brings back to zero after 4
seconds.
The maximum angle of attack is 38 deg which is greater than the HIRM stall
angle, see Figure 5.1. In accordance with the GAS property of the control
law shown in Section 5.2.2, this does not cause any stability problems.

7.3 Simulation

95

During the steep ascent just after 5 s, the elevons saturate at 40 deg, which
means that the desired angular acceleration, udes , cannot be produced. The
actual u can be computed by numerically differentiating q from sensor data.
Doing so, the maximum gain reduction during the saturation period is given
by
min
t

u
0.55
udes

In Section 5.2.2, the backstepping control law was shown to have a certain
amount of gain margin. Using the parameter values of Table 7.3, the bound
(5.49) becomes
k2 (1 + k1 )
= 0.8
k3
Thus, despite violating (5.49), the system converges to the desired state3 .
This indicates that the parameter restrictions (5.48) are sufficient but not
always necessary for closed loop stability.

3 One should also note that the robustness results in Chapter 5 only regard the state feedback
laws, not including the observer feedback.

96

Implementation and Simulation

200

3.5

0.5
3
(deg)

100
50

(deg)

ps (deg/s)

150

2.5
0.5

0
50

5
time (s)

10

182

5
time (s)

10

180

6
4

168

5
time (s)

10

5
time (s)

0.6

10

0.2

5
time (s)

10

5
time (s)

10

5
time (s)

10

0.4

0.8

172
170

10

1.2
nz (g)

q (deg/s)

VT (m/s)

174

5
time (s)

1.4

178
176

r (deg)

es (deg)

ed (deg)

0.6
0.8

1
0

1.2
0
2

1.4
0

5
time (s)

1.6

10

1.5

5
time (s)

10

0.05

0.05

0.5

3 (rad/s2)

2 (rad/s2)

0.5

1 (rad/s )

0.1

0.15

0.1

0.15

1
1.5

5
time (s)

10

0.2

5
time (s)

10

0.2

Figure 7.2 Assessment maneuver M1: roll rate demand, pref

s = 150 /s.

7.3 Simulation

97

0.1

16

0.01

14

0.005

(deg)

0.1
0.2

(deg)

ps (deg/s)

12
10
8
6
0.3
0.4

0.005

4
0

5
time (s)

10

170

40

160

30

150

20

5
time (s)

10

0.01

5
time (s)

10

5
time (s)

10

5
time (s)

10

5
time (s)

10

140

nz (g)

VT (m/s)

q (deg/s)

10

130

120

10

110

5
time (s)

20

10

0.4

5
time (s)

10

20

0.4

15

0.2

0.3

0.2

r (deg)

es (deg)

ed (deg)

10
5

0.2
0.1

0
0.4
0.6

5
time (s)

10

10

5
time (s)

10

0.1

x 10

2 (rad/s2)

1 (rad/s2)

0
5

0.3

0.2

0.2

0.1

0.1

0.1

10
15
20

0.3

3 (rad/s2)

5
time (s)

10

0.1

0.2

0.2

0.3

0.3

0.4

5
time (s)

10

0.4

Figure 7.3 Assessment maneuver M2: angle of attack demand, ref = 15 .

98

Implementation and Simulation

(deg)

100
50
0

50

5
time (s)

14

12

10
8

40

170

30
q (deg/s)

180

VT (m/s)

160
150

5
time (s)

10

5
time (s)

15

5
time (s)

10

20

10

15

r (deg)

es (deg)

ed (deg)

5
time (s)

10

10

5
time (s)

25

10

0.4

0.5

0.2

0.2

0.2
0.4

0.4

1.5

0.6

0.6

0.8

10

5
time (s)

10

0.2

5
time (s)

3 (rad/s )

0.6

0.4

2 (rad/s )

0.6

10

20

5
time (s)

15

1.5

0.5

10

1 (rad/s )

10

0
0

5
time (s)

10

20

10

10

10

10

10

5
time (s)

20

140
130

10

nz (g)

ps (deg/s)

150

16

(deg)

200

5
time (s)

10

0.8

Figure 7.4 Assessment maneuver M3: M1 and M2 combined.

7.3 Simulation

99

30

50

25
40
(deg)

(deg)

20
15
10
5

30

20

0
5

10
time (s)

15

10

20

40

10
time (s)

15

20

10
time (s)

15

20

10
time (s)

15

20

10
time (s)

15

20

40
30
20

q (deg/s)

(deg)

30

20

10
0

10
10
0

10
time (s)

15

20

20

100

2.5

95

2
n (g)

90

1.5

VT (m/s)

105

85

80

0.5

75

10
time (s)

15

20

10

0.1
0.05

0
es (deg)

(rad/s )

0
2

10
20

0.05
0.1
0.15

30
40

0.2
0

10
time (s)

15

20

0.25

Figure 7.5 Assessment maneuver M4: flight path angle demands, ref =
25 followed by ref = 15 . The dashed line in the plot
represents 0 .

Appendix
7.A

Aircraft Data

Aircraft model properties for the GAM [5] and for the HIRM [56].
Entity
mass
moment of inertia

wing planform area

wing span
mean aerodynamic chord

m
Ix
Iy
Iz
Ixz
S
b
c

100

kg
kg m2
kg m2
kg m2
kg m2
m
m
m

GAM
9100.0
21000.0
81000.0
101000.0
2500.0
45.0
10.0
5.20

HIRM
15296.0
163280.0

37.16
11.4
3.511

8
Conclusions

In this thesis, we have investigated backstepping as a possible framework for aircraft

flight control design. The main result is that by recognizing the naturally stabilizing
parts of the aerodynamic lift and side forces, globally stabilizing control laws can
be designed which rely on less information of these forces than corresponding NDI
designs, based on feedback linearization.
To conclude, let us evaluate the proposed control laws with respect to five
major issues, relevant to all control designs, namely stability, tuning, robustness,
input saturation, and disturbance attenuation.
Stability Backstepping deals explicitly with stability, through the construction
of a Lyapunov function for the closed loop system along with the construction of the
control law itself. The derived state feedback flight control laws yield asymptotic
stability within the entire flight envelope of practical interest, including high angles
of attack.
Tuning Since our backstepping designs focus on utilizing useful nonlinearities
rather than cancelling them, the resulting closed loop systems are not linear. This
means that the aircraft response to the pilot inputs will not be independent of, e.g.,
the angle of attack. However, we have shown that the nominal performance, valid
around an operating point of interest, can be tuned according to ones requirements.
101

102

Conclusions

Robustness Robustness profits greatly from our backstepping approach. Since

cancellation of the nonlinear aerodynamic lift and side forces, L() and Y (), is
avoided, complete knowledge of these forces is not required. We have also proposed
adaptation techniques for handling model errors in the aerodynamic moments.
Input saturation Our backstepping control laws all possess a certain gain margin. Thus, stability is retained for a certain amount of control surface saturation
leading to a reduction of the moments produced.
Disturbance attenuation A main weakness of nonlinear control design in general, including backstepping, is the lack of tools to quantitatively analyze the effects
of disturbances, e.g., wind gusts and sensor noise in our aircraft control case. Fortunately, as reported by Enns et al. [15], a flight controller giving a properly selected
bandwidth generally does a good job of suppressing such disturbances.

Bibliography

[1] Richard J. Adams, James M. Buffington, and Siva S. Banda. Design of nonlinear control laws for high-angle-of-attack flight. Journal of Guidance, Control,
and Dynamics, 17(4):737746, 1994.
[2] Muthana T. Alrifai, Joe H. Chow, and David A. Torrey. A backstepping
nonlinear control approach to switched reluctance motors. In Proceedings of the
37th IEEE Conference on Decision and Control, pages 46524657, December
1998.
[3] David Anderson and Scott Eberhardt. How airplanes fly: A physical description of lift. Sport Aviation, February 1999.
http://www.allstar.fiu.edu/aero/airflylvL3.htm.
[4] Zvi Artstein. Stabilization with relaxed controls. Nonlinear Analysis, Theory,
Methods & Applications, 7(11):11631173, 1983.
[5] Hans Backstr
om. Report on the usage of the Generic Aerodata Model. Technical report, Saab Aircraft AB, May 1997.
[6] Dimitri P. Bertsekas. Dynamic Programming and Optimal Control. Athena
Scientific, 1995.
[7] Jean-Luc Boiffier. The Dynamics of Flight: The Equations. John Wiley &
Sons, 1998.
103

104

Bibliography

[8] R. P. Brent. Algorithms for Minimization Without Derivatives. Prentice-Hall,

1973.
[9] M. M. Bridges, D. M. Dawson, and C. T. Abdallah. Control of rigid-link,
flexible-joint robots: A survey of backstepping approaches. Journal of Robotic
Systems, 12(3):199216, 1995.
[10] James J. Carroll, Andrew J. Geoghan, Darren M. Dawson, and Praveen Vedagarbha. A backstepping based computed torque controller for switched reluctance motors driving inertial loads. In Proceedings of the 4th IEEE Conference
on Control Applications, pages 779786, September 1995.
[11] James J. Carroll, Manfred Schneider, and Darren M. Dawson. Integrator
backstepping techniques for the tracking control of permanent magnet brush
DC motors. In Conference Record of the 1993 IEEE Industry Applications
Society Annual Meeting, pages 663671, October 1993.
[12] Carin Cronander. Control applications in aeronautics. Licentiate thesis 99-23,
Department of Aeronautics, Royal Institute of Technology, 1999.
[13] J. E. Dennis, Jr. and Robert B. Schnabel. Numerical Methods for Unconstrained Optimization and Nonlinear Equations. SIAM, 1996.
[14] Wayne C. Durham, Frederick H. Lutze, and William Mason. Kinematics and
aerodynamics of the velocity vector roll. Journal of Guidance, Control, and
Dynamics, 17(6):12281233, NovemberDecember 1994.
[15] Dale Enns, Dan Bugajski, Russ Hendrick, and Gunter Stein. Dynamic inversion: an evolving methodology for flight control design. International Journal
of Control, 59(1):7191, January 1994.
[16] Bernard Etkin and Lloyd Duff Reid. Dynamics of Flight: Stability and Control.
John Wiley & Sons, third edition, 1996.
[17] Kenan Ezal, Zigang Pan, and Petar V. Kokotovic. Locally optimal and robust
backstepping design. IEEE Transactions on Automatic Control, 45(2):260
270, February 2000.
[18] Dan Fontaine and Petar Kokotovic. Approaches to global stabilization of a
nonlinear system not affine in control. In Proceedings of the 1998 American
Control Conference, pages 16451647, June 1998.
[19] G. E. Forsythe, M. A. Malcolm, and C. B. Moler. Computer Methods for
Mathematical Computations. Prentice-Hall, 1976.
[20] Jonas Fredriksson. Nonlinear control of turbocharged diesel engines. Licentiate thesis 329L, Control Engineering Laboratory, Department of Signals and
Systems, Chalmers University of Technology, 1999.

Bibliography

105

[21] Randy A. Freeman and Petar V. Kokotovic. Robust Nonlinear Control Design:
State-Space and Lyapunov Techniques. Birkh
auser, 1996.
[22] Randy A. Freeman and James A. Primbs. Control Lyapunov functions: New
ideas from an old source. In Proceedings of the 35th Conference on Decision
and Control, pages 39263931, December 1996.
[23] William L. Garrard, Dale F. Enns, and S. Anthony Snell. Nonlinear feedback
control of highly manoeuvrable aircraft. International Journal of Control, 56
(4):799812, 1992.
[24] S. Torkel Glad. Robustness of nonlinear state feedback - a survey. Automatica,
23(4):425435, 1987.
[25]
Aslaug Grvlen and Thor I. Fossen. Nonlinear control of dynamic positioned
ships using only position feedback: An observer backstepping approach. In
Proceedings of the 35th Conference on Decision and Control, pages 33883393,
December 1996.
[26] James K. Hall and Meir Pachter. Formation maneuvers in three dimensions. In
Proceedings of the 39th Conference on Decision and Control, December 2000.
[27] Ola H
arkeg
ard and S. Torkel Glad. A backstepping design for flight path angle
control. In Proceedings of the 39th Conference on Decision and Control, pages
35703575, Sydney, Australia, December 2000.
[28] Ola H
arkeg
ard and S. Torkel Glad. Control of systems with input nonlinearities and uncertainties: an adaptive approach. Technical Report LiTH-ISYR-2302, Department of Electrical Engineering, Link
opings universitet, SE-581
83 Link
oping, Sweden, October 2000.
[29] Ola H
arkeg
ard and S. Torkel Glad. Flight control design using backstepping.
Technical Report LiTH-ISY-R-2323, Department of Electrical Engineering,
Link
opings universitet, SE-581 83 Link
oping, Sweden, December 2000.
[30] HARV. The NASA High Angle-of-Attack Research Vehicle homepage.
http://www.dfrc.nasa.gov/Projects/HARV.
[31] John Hauser and Ali Jadbabaie. Aggressive maneuvering of a thrust vectored
flying wing: A receding horizon approach. In Proceedings of the 39th Conference on Decision and Control, December 2000.
[32] W. B. Herbst. Future flight technologies. Journal of Aircraft, 17(8):561566,
August 1980.
[33] J. Hu, D. M. Dawson, and K. Anderson. Position control of a brushless DC
motor without velocity measurements. IEE Proceedings of Electric Power
Applications, 142(2):113122, June 1995.

106

Bibliography

[34] Jun Hu, Darren M. Dawson, and Yi Qian. Position tracking control for robot
manipulators driven by induction motors without flux measurements. IEEE
Transactions on Robotics and Automation, 12(3):419438, June 1996.
[35] Chien H. Huang and Gareth J. Knowles. Application of nonlinear control
strategies to aircraft at high angles of attack. In Proceedings of the 29th
Conference on Decision and Control, pages 188193, December 1990.
[36] Alberto Isidori. Nonlinear Control Systems. Springer, third edition, 1995.
[37] Mrdjan Jankovic, Miroslava Jankovic, and Ilya Kolmanovsky. Constructive
Lyapunov control design for turbocharged diesel engines. IEEE Transactions
on Control Systems Technology, 8(2):288299, March 2000.
[38] Zhong-Ping Jiang and Henk Nijmeijer. Tracking control of mobile robots: A
case study in backstepping. Automatica, 33(7):13931399, 1997.
[39] I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. A toolkit for nonlinear
feedback design. Systems & Control Letters, 18(2):8392, February 1992.
[40] Hassan K. Khalil. Nonlinear Systems. Prentice-Hall, second edition, 1996.
[41] Petar Kokotovic. Constructive nonlinear control: Progress in the 90s. In
IFAC 1999 Proceedings, pages 4977, 1999.
[42] Petar Kokotovic, Hassan K. Khalil, and John OReilly. Singular Perturbation
Methods in Control: Analysis and Design. Academic Press, 1986.
[43] Petar V. Kokotovic. The joy of feedback: Nonlinear and adaptive. IEEE
Control Systems Magazine, 12(3):717, June 1992.
[44] Arthur J. Krener and Alberto Isidori. Linearization by output injection and
nonlinear observers. Systems & Control Letters, 3:4752, June 1983.
[45] Miroslav Krstic, Dan Fontaine, Petar V. Kokotovic, and James D. Paduano.
Useful nonlinearities and global stabilization of bifurcations in a model of jet
engine surge and stall. IEEE Transactions on Automatic Control, 43(12):
17391745, December 1998.
[46] Miroslav Krstic, Ioannis Kanellakopoulos, and Petar Kokotovic. Nonlinear
and Adaptive Control Design. John Wiley & Sons, 1995.
[47] Miroslav Krstic and Petar V. Kokotovic. Lean backstepping design for a jet engine compressor model. In Proceedings of the 4th IEEE Conference on Control
Applications, pages 10471052, 1995.
[48] Stephen H. Lane and Robert F. Stengel. Flight control design using non-linear
inverse dynamics. Automatica, 24(4):471483, 1988.
[49] Tomas Larsson. Linear quadratic design and -analysis of a fighter stability
augmentation system. Masters thesis, Link
opings universitet, 1997.

Bibliography

107

[50] J. Levine. Are there new industrial perspectives in the control of mechanical systems? In Paul M. Frank, editor, Advances of Control: Highlights of
ECC99, chapter 7, pages 197226. Springer, 1999.
[51] Johan L
ofberg. Backstepping with local LQ performance and global approximation of quadratic performance. In Proceedings of the American Control
Conference, pages 38983902, June 2000.
[52] A. M. Lyapunov. The General Problem of the Stability of Motion. Taylor &
Francis, 1992. English translation of the original publication in Russian from
1892.
[53] Jean-Francois Magni, Samir Bennani, and Jan Terlouw, editors. Robust Flight
Control: A Design Challenge. Springer, 1997.
[54] Donald McLean. Automatic Flight Control Systems. Prentice Hall, 1990.
[55] S. K. Mudge and R. J. Patton. Variable structure control laws for aircraft
manoeuvres. In International Conference on Control, 1988, pages 564568,
1988.
[56] E. A. M. Muir et al. Robust flight control design challenge problem formulation and manual: The high incidence research model (HIRM). Technical
Report TP-088-4, Group for Aeronautical Research and Technology in EURope GARTEUR-FM(AG08), 1997.
[57] Robert C. Nelson. Flight Stability and Automatic Control. McGraw-Hill,
second edition, 1998.
[58] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. Numerical Recipes in C. Cambridge University Press, second edition,
1992.
[59] Marc O. Rauw. FDC 1.2 - A SIMULINK Environment for Flight Dynamics
and Control Analysis. Zeist, The Netherlands, February 1998.
http://www.dutchroll.com.
[60] Jacob Reiner, Gary J. Balas, and William L. Garrard. Robust dynamic inversion for control of highly maneuverable aircraft. Journal of Guidance, Control,
and Dynamics, 18(1):1824, JanuaryFebruary 1995.
[61] Jacob Reiner, Gary J. Balas, and William L. Garrard. Flight control design
using robust dynamic inversion and time-scale separation. Automatica, 32(11):
14931504, 1996.
[62] Wilson J. Rugh. Linear System Theory. Prentice Hall, second edition, 1996.
[63] SaabBAE Gripen AB. JAS 39 Gripen homepage. http://www.gripen.se.

108

Bibliography

[64] Saab Aircraft AB. Generic Aerodata Model (GAM).

http://www.flyg.kth.se/edu-dir/downloads-dir.
[65] R. Sepulchre, M. Jankovic, and P. V. Kokotovic. Constructive Nonlinear Control. Springer, 1997.
[66] R. Sepulchre, M. Jankovic, and P. V. Kokotovic. Interlaced systems and recursive designs for global stabilization. In Proceedings of the 1997 European
Control Conference, 1997.
[67] Rodolphe Sepulchre, Mrdjan Jankovic, and Petar V. Kokotovic. Integrator
forwarding: A new recursive nonlinear robust design. Automatica, 33(5):979
984, May 1997.
[68] Sahjendra N. Singh, Phil Chandler, Corey Schumacher, Siva Banda, and Meir
Pachter. Nonlinear adaptive close formation control of unmanned aerial vehicles. Dynamics and Control, 10(2):179194, 2000.
[69] Sahjendra N. Singh and Marc Steinberg. Adaptive control of feedback linearizable nonlinear systems with application to flight control. Journal of Guidance,
Control, and Dynamics, 19(4):871877, JulyAugust 1996.
[70] Jean-Jaques E. Slotine and Weiping Li. Applied Nonlinear Contol. Prentice
Hall, 1991.
[71] S. Anthony Snell, Dale F. Enns, and William L. Garrard Jr. Nonlinear inversion flight control for a supermaneuverable aircraft. Journal of Guidance,
Control, and Dynamics, 15(4):976984, JulyAugust 1992.
[72] Eduardo D. Sontag. A universal construction of Artsteins theorem on nonlinear stabilization. Systems & Control Letters, 13:117123, 1989.
[73] Marc L. Steinberg and Anthony B. Page. Nonlinear adaptive flight control
with genetic algorithm design optimization. International Journal of Robust
and Nonlinear Control, 9(14):10971115, 1999.
[74] Brian L. Stevens and Frank L. Lewis. Aircraft Control and Simulation. John
Wiley & Sons, 1992.
[75] Jann Peter Strand, Kenan Ezal, Thor I. Fossen, and Petar V. Kokotovic.
Nonlinear control of ships: A locally optimal design. In Proceedings of the
IFAC NOLCOS98, pages 732737, July 1998.
[76] M. Vidyasagar. Nonlinear Systems Analysis. Prentice Hall, second edition,
1993.
[77] K. H. Well, B. Faber, and E. Berger. Optimization of tactical aircraft maneuvers utilizing high angles of attack. Journal of Guidance, Control, and
Dynamics, 5(2):131137, MarchApril 1982.

Bibliography

109

[78] Bing-Yu Zhang and Blaise Morton. Robustness analysis of dynamic inversion control laws applied to nonlinear aircraft pitch-axis models. Nonlinear
Analysis, Theory, Methods & Applications, 32(4):501532, 1998.
[79] Kemin Zhou, John C. Doyle, and Keith Glover. Robust and Optimal Control.
Prentice-Hall, 1996.

110

Bibliography