International Journal of Aerospace and Mechanical Engineering 2:2 2008

Robust Control of F-16 Lateral Dynamics

Hoa Vo* and Sridhar Seshagiri, Member, IEEE

Abstract— This paper considers the application of a conditional specifically, we wish to design the aileron and rudder controls
integrator based continuous sliding mode control design for robust to asymptotically track desired references for the side-slip and
regulation of MIMO minimum-phase nonlinear systems to the control roll angles. The application of SMC to flight control has been
of the lateral flight dynamics of an F-16 aircraft. The system is
non-affine in the input but can be rewritten as the perturbation of pursued by several others authors, see, for example, [4], [5],
a control affine system with matched (input-dependent) disturbances. [19]. Our work differs from earlier ones in that it is based on
A parameter dependent transformation brings the system to normal a recent technique in [14], [16] for introducing integral action
form, for which an output-feedback control can be designed to in SMC, that we refer to as “conditional integrators”. The
achieve robust regulation. Analytical results for stability are provided, controller that we present has a very simple structure; aside
and it is also shown through extensive simulations that the inherent
robustness of the SMC design provides a convenient way to design from an “input-decoupling” term, it is simply two saturated
controllers without adaptation for the unknown parameters, with a high-gain PID controllers, with the (anti-windup) conditional
transient performance that is comparable to discontinuous SMC, but integrator, and the derivative estimated by a high-gain ob-
without suffering from the drawback of control chattering. server. This controller structure is a special case of a general
Keywords— Integral Control, Conditional Integrators, Sliding design for robust output regulation for multiple-input multiple-
Mode Control, Lateral Dynamics, High-gain Observers. output (MIMO) nonlinear systems transformable to the normal
form, with analytical results for stability and performance
described in [14], [16]. The inclusion of integral control in
I. I NTRODUCTION
SMC design on the one hand allows us to use smaller gains,

T HE dynamic response characteristics of aircraft are

highly nonlinear. Traditionally, flight control systems
have been designed using mathematical models of the aircraft
while the conditional design means that integral action takes
place only inside the boundary layer, allowing us to recover
the transient performance of ideal (discontinuous) SMC, but
linearized at various flight conditions, with the controller without suffering from the trade-off between tracking accu-
parameters or gains “scheduled” or varied with the flight racy and robustness to unmodeled high-frequency dynamics.
operating conditions. Various robust multivariable techniques A preliminary version of this paper has been accepted for
including linear quadratic optimal control (LQR/LQG), H∞ publication in the Proceedings of the 34th Annual Conference
control, and structured singular value μ-synthesis have been of the IEEE Industrial Electronics Society (IECON’08) [22].
employed in controller design, an excellent and exhaustive The rest of this paper is organized as follows. In Section
compendium of which is available in [9]. In order to guaran- 2, we describe the nonlinear mathematical aircraft model,
tee stability and performance of the resulting gain-scheduled while in Section 3, we present our controller design, based
controllers, analytical frameworks of gain scheduling have on the results in [14], [16]. Simulation results showing the
been developed, including the powerful technique of linear- efficacy of the design, including demonstrating the robustness
parameter-varying (LPV) control [3], [8], [18], [21]. Nonlinear to parameter uncertainties, disturbances and time-delays, are
design techniques such as dynamic inversion have been used presented in Section 4. Finally, a summary of our work
in [1], [11], [20], while a technique that combines model and some suggestions for possible extensions are provided in
inversion control with an online adaptive neural network to Section 5.
“robustify” the design is described in [12], and a nonlinear
adaptive design based on backstepping and neural networks in
II. A IRCRAFT L ATERAL M ODEL
[7]. A radial basis functions neural network (RBFNN) based
adaptive design with time-scale separation between the system The model of the lateral dynamics that we use for control
and controller dynamics, with applications to control of both design is described below, and is extracted practically verbatim
longitudinal (angle-of-attack command systems) as well as from [23].
lateral (regulation of the sideslip and roll angles) dynamics γ0 cos γ0 ⎫
φ̇ = cos cos θ0 ps + cos θ0 rs ⎪
⎪
is described in [23]. A succinct “industry perspective” on Y ⎪
⎪
β̇ = Vβ β + YVr rs + gcos V
θ0
φ − rs ⎪
⎪
⎬
flight control design, including the techniques of robust control
p˙s = Lβ β + Lp ps + Lr rs + δl (ps , rs )+
(H∞ , μ-synthesis), LPV control, dynamic inversion, adaptive (1)
Lδa (β, δa ) + Lδr (β, δr ) ⎪
⎪
control, neural networks, and more, can be found in [2]. ⎪
r˙s = Nβ β + Np ps + Nr rs + δn (ps , rs )+ ⎪ ⎪
⎪
⎭
Our interest is in the design of robust sliding mode control
Nδa (β, δa ) + Nδr (β, δr )
(SMC) for the lateral flight dynamics of a F-16 aircraft. More
The system has 4 states φ, β, ps and rs , which are respectively
H. Vo and S. Seshagiri are with the ECE Dept., San Diego State the roll angle, the sideslip, and the stability axis roll and
University, San Diego, CA 92182, USA e-mail: hoaquang36@gmail.com,
seshagir@engineering.sdsu.edu. yaw rates, and two inputs δa and δr , which are respectively
*Financially supported in part by SDSU under UGP Funds. the aileron and rudder control. The other variables in (1)

80
 International Journal of Aerospace and Mechanical Engineering 2:2 2008


are the gravitational constant g, the trimmed pitch angle βref (t)
smooth commanded reference input yref (t) = ,
θ0 , the trimmed flight path angle γ0 , the true airspeed V , φref (t)
robustly in the presence of roll/yaw parametric uncertainties
the aerodynamic stability and control derivatives Yβ and Yr
and aileron/rudder unknown control effects.
(which can be taken as being approximately constant), the
incremental rolling and yawing moments δl and δn that Our approach to control design is based (see [14], [16]) on
are unknown functions of the roll and yaw rates, and the minimum-phase systems transformable to the normal form. To
rolling and yawing moments due to aileron and rudder de- simplify the presentation, we start with the SISO control-affine
flections Lδa , Lδr , Nδa , Nδr . Following [23], we assume that case of a nonlinear system with full relative degree ρ = n (i.e.,
the nonlinear terms δl (ps , rs ) + Lδa (β, δa ) + Lδr (β, δr ) and with no zero-dynamics), that can be transformed to the normal
δn (ps , rs ) + Nδa (β, δa ) + Nδr (β, δr ) in (1) can be expressed form
as shown below ξ˙ = Ac ξ + Bc [a(ξ) + b(ξ)u], y = Cc ξ
δl + Lδa + Lδr =Lδa0 (δa + f1 (·)) + Lδr0 (δr + f2 (·)) where ξ ∈ Rρ the output and its derivatives up to order ρ − 1,
δn + Nδa + Nδr =Nδa0 (δa + f1 (·)) + Nδr0 (δr + f2 (·)) and the triple (Ac , Bc , Cc ) a canonical form representation of
(2) a chain of ρ integrators, and the functions a(·) and b(·) are
where the “linear terms” Lδa0 , Lδr0 , Nδa0 , and Nδr0 are unknown, but b(·) is globally bounded away from zero, and a
known (i.e., we use nominal values for these parameters) lower bound on its magnitude is known. Under the assumption
and all uncertainty is lumped into the nonlinear functions that all exogenous signals (that include the reference yref (t))
f1 (β, ps , rs , δa ) and f2 (β, ps , rs , δr ). Consequently, system are asymptotically constant, we proceed with an integral
(1) can be written compactly in standard state-space form as control based design. In particular, for the class of systems
⎡ ⎤ ⎡ ⎤ considered here, it is shown in [14], [16] that a continuous
β̇ β 
⎢ φ̇ ⎥ ⎢φ⎥ sliding mode controller of the form
⎢ ⎥ = A ⎢ ⎥ + B δa + f1 (β, ps , rs , δa )  
⎣p˙s ⎦ ⎣ps ⎦ δr + f2 (β, ps , rs , δr ) k0 σ + k1 e1 + k2 e2 + · · · + eρ
u = −ksign(b(·)) sat (7)
r˙s rs μ
⎡ ⎤ (3)
 β  can be designed to achieve robust regulation, where e1 = y −
1 0 0 0 ⎢ ⎥ yref is the tracking error, and e2 , ..., eρ are its derivatives up
y= ⎢ φ ⎥ def= Cx =
C1
x
0 1 0 0 ⎣ ps ⎦ C2 to order ρ, the positive constants ki , i = 1, · · · , ρ − 1 in the
rs sliding surface function
ρ

where
⎡ Yβ gcos θ0 ⎤ ⎡ ⎤ s = k0 σ + ki ei + eρ (8)
V V 0 −1 0 0 i=1
⎢ 0 0 cos γ0 sin γ0 ⎥ 0 ⎥⎢ 0
A=⎢ ⎣L
cos θ0 cos θ0 ⎥ , B = ⎢⎥ are chosen such that the polynomial
β 0 L p Lr ⎦ L ⎣ Lδa
δr0 ⎦ 0

Nβ 0 Np Nr Nδr0 Nδa0 λρ−1 + kρ−1 λρ−2 + · · · + k1

(4)
In [23], it is assumed that based on available wind-tunnel is Hurwitz, and σi is the output of the “conditional integrator”
 
data, the unknown nonlinear functions f1 (·) and f2 (·) have s
σ̇ = −k0 σ + μ sat , σ(0) ∈ [−μ/k0 , μ/k0 ] (9)
the following form: μ
β2
− where k0 > 0, and μ > 0 is the “width” of the boundary
f1 (·) = ((1 − E1 )e 2δ1 2 + E1 )(tanh(δa + h1 )
(5) layer. From (8) and (9), it is clear that inside the boundary
+ tanh(δa − h1 ) + 0.001δa)
layer |s| ≤ μ,
+ D1 cos(A1 ps − w1 )sin(A2 rs − w2 ) + D2
β2 σ̇ = k1 e1 + k2 e2 + · · · + eρ
−
f2 (·) = ((1 − E2 )e 2δ2 2 + E2 )(tanh(δr + h2 )
+ tanh(δr − h2 ) + 0.001δr ) (6) which implies that ei = 0 at equilibrium, i.e., (9) is the
+ D3 cos(A3 ps − w3 )sin(A4 rs − w4 ) + D4 equation of an integrator that provides integral action “con-
ditionally”, inside the boundary layer. As shown in [14], [16],
A physical interpretation of the constants 0 < E1,2 < 1 as the such a design provides asymptotic error regulation, while
percentage of the control effectiveness available at high angles not degrading the transient performance, as is common in a
of sideslip is discussed in [23]. As opposed to [23] where the conventional design that uses the integrator σ̇ = e1 . In the
design is based on adaptive approximation using RBFNNs, we case of relative degree ρ = 1 and ρ = 2, the controller (7)
do not require to know the specific form of the functions in is simply a specially tuned saturated PI/PID controller with
our design, except to (numerically) verify an assumption that anti-windup (see [16, Section 6]).
we will make later on. The control (7) can be extended to the output-feedback
III. C ONTROL D ESIGN case 1 ei by its estimate êi , obtained using the high-gain

 Our control
objective is to design a control
 signal
δ = 1 This might be required even when the original state x is available, since
δa (t) β(t) ξ and hence ei , which is required in the control, depend on the state through
such that the output y(t) = tracks a
δr (t) φ(t) possibly unknown parameters.

81
 International Journal of Aerospace and Mechanical Engineering 2:2 2008

observer (HGO) can be found as [16, Theorems 1 & 2].

 With δfi (·) = 0, which really is the case of interest here,
ê˙ i = êi+1 + αi (e1 − ê1 )/i , 1 ≤ i ≤ ρ − 1
(10) an additional assumption is required, which we state below.
ê˙ ρ = αρ (e1 − ê1 )/ρ
To see the motivation for this assumption, we differentitate
where  > 0, and the positive constants αi are chosen such (11) and with the control as defined in (12), we rewrite the
that the roots of resulting equation in the form
 
λρ + α1 λρ−1 + · · · + αρ−1 λ + αρ = 0 sz
ṡz = Δz (x, σ, e, z̈ref , v) − γz sat (13)
μz
have negative real parts. To complete the design, we need   
to specify how k, μ and  (in the output-feedback case) are σβ eβ e
where σ = , and e = , with ez = 1z , z = β, φ.
chosen. The parameter k is chosen “sufficiently large” (to σφ eφ e2z
overbound uncanceled terms in ṡ) while μ and  are chosen From (13), it is clear that for the control term vz =
“sufficiently small”, the former to recover the performance −γz sat μszz to “overbound” the term Δz (x, σ, e, z̈ref , v),
of ideal (discontinuous) SMC (without an integrator) and which in turn depends on both vβ and vφ , some restriction
the latter to recover the performance under state-feedback needs to be imposed on the way Δz (·, v) depends on v. The
with the continuous SMC. Analytical results for stability and next assumption states this restriction precisely.
performance are given in [14], [16]. Assumption 1:
 
While we abstracted the design from [14], [16] above for  Δz (·) 
max   ≤ z + κz1 |vβ | + κz2 |vφ | , z = β, φ
γz 
SISO systems for the sake of clarity of presentation, the design
in those papers was done for MIMO systems. Consequently,
our next step is to extend the design presented above for where the maximization is taken over some compact set S of
the MIMO system (3), keeping in mind the basic features interest, the constants 2 z and κzi are known and I − K,
κ κβ2
presented above. To that end, we first rewrite where K = β1 , is an M-matrix.
κφ1 κφ2
f1 (·, δa ) = f1 (·, 0) + δf1 (·, δa ) The constants γz in the control (12) are chosen as follows.
f2 (·, δr ) = f2 (·, 0) + δf2 (·, δr ) The fact that I − K is an M-matrix implies that (i) (I − K) is
 nonsingular, (ii) the elements of (I − K)−1 are non-negative,
def f1 (·, 0)
and define f (·, 0) = . Then, it is easy to check and (iii) there exists a vector l with lz > 0 such that the
f2 (·, 0)
that with δfi (·) ≡ 0, the control-affine system ẋ = Ax +  of b = (I − K)l are all positive. For z = β, φ, let
elements
¯
B[u + f (·, 0)] has strong vector relative degree ρ = {2, 2}, ¯ = β , with ¯z ≥ z , and take γz = ψz + lz , where ψ =
since C1 B = C2 B = 0 and T = CAB is nonsingular, and  ¯φ
that with δfi (·) = 0, the change of variables ψβ
= (I − K)−1 ¯. With this choice of γz , the analysis in
ψφ
e1β = β − βref , e2β = ė1β , e1φ = φ − φref , e2φ = ė1φ [14] shows that the (state-feedback) control (11), (12) achieves
robust regional regulation, with S a subset of the region of
transforms the system to a normal form very similar to the attraction. The design is then extended to the output-feedback
one in the SISO case, that is identical to the general case case by replacing e2z = ė1z , z = β, φ, in (11) (and hence also
considered in [14]. In particular, it can be verified that the in the control (12)) with their estimates using HGOs. While
equations for ė2β and ė2φ take the form we have not verified Assumption 1 rigorously, our simulation
ė2z = bz (x)−z̈ref +ti1 (δa +δf1 (·))+ti2 (δr +δf2 (·)), z = β, φ results, which we present next, validate (the assumption and)
our control design.
where T = CAB = {tij }. As done in the SISO case, for
z = β, φ, we define

sz = kz0 σz + kz1 e1z +e2z  IV. S IMULATION R ESULTS
(11)
σ̇z = −kz0 σz + μz sat μszz
For our simulations, we use the same values of the aircraft
where the second set of equations represent conditional inte- parameters that are used in [23]. In particular, for the trim
grators that provide integral action only inside the boundary values corresponding to an airspeed of V = 502f t/s and
layers |sz | ≤ μz . When δfi (·) = 0, a standard sliding mode angle of attack α = 2.11◦ , we have
argument (see, for example, [16]) shows that the control ⎡ ⎤
   −0.3220 0.0640 0.0364 −0.9917
−1 vβ sz ⎢
−1 def
δ=T v = T , vz = −γz sat 0 0 1 0.0393 ⎥
vφ μz
(12) A=⎢ ⎣−30.6490
⎥,
0 −3.6784 0.6646 ⎦
achieves semi-global regulation, provided the gains γβ and 8.5395 0 −0.0254 −0.4764
γφ can be chosen “sufficiently large”, and the boundary layer
widths μβ and μφ are chosen “sufficiently small”. More 2 The function  can be chosen to be error dependent, and is assumed
z
precise analytical results for stability and performance of the constant only for convenience. It can always be chosen to be constant over
above control design in the control-affine case (i.e., δfi (·) = 0) compact sets, but at the cost of conservatism.

82
 International Journal of Aerospace and Mechanical Engineering 2:2 2008

⎡ ⎤
0 0
⎢ 0 0 ⎥
B=⎢ ⎥
⎣−0.7331 0.1315 ⎦
0.2 μβ=1

Error in β (deg)
0.1 μβ=0.1
−0.0319 −0.0620 0
−0.1
The values of the constants that appear in the expressions
for f1 (·) and f2 (·) are computed using averaged (and curve- −0.2

fitted) wind-tunnel data for α = 0◦ and α = 5◦ , with 5 10 15 20 25 30 35 40

the curve fit approximation done over β ∈ [−30◦ , 30◦ ],

ps ∈ [−180◦, 180◦ ], rs ∈ [−90◦ , 90◦ ], δa ∈ [−21.5◦ , 21.5◦],
0.2
and δr ∈ [−30◦ , 30◦ ]. From [23], the corresponding values

Error in φ (deg)
0
for the constants then are A1 = A2 = A3 = A4 = 0.1,
D1 = 0.075, D2 = 0.0016, D3 = 0.45, D4 = 0, ω1 = ω3 =
−0.2 μφ=1
μφ=0.1
1.5, ω2 = ω4 = 0, E1 = E2 = 0.3, h1 = 7, h2 = 4, −0.4

σ1 = 0.015, σ2 = 0.15. The commanded reference input is −0.6

10 15 20 25 30 35 40 45
Time (sec)
taken as
⎡   ⎤

βref (t) 0.2 − 0.5
t−8 + 1
t−30 − 0.5 Fig. 1. Tracking errors without actuator dynamics and no integral control.
=⎣  1+e  ⎦
1+e
φref (t) 0.2 − 1+e + 1+et−30 − 0.2
0.5 1
t−8

All initial conditions are taken as zero. Throughout, we 2, and it is clear that there is chattering in the control as μ is
assume the
 (nominal) value of the decoupling matrix T = made small in order to obtain smaller errors4 .
0.005 0.0663
CAB = , and for the controller param-
−0.7344 −0.1291 μβ = 0.1, no actuator dynamics μφ = 0.1, no actuator dynamics

eters, kβ0 = kφ0 = kβ1 = kφ1 = 5, and kβ = kφ = 10 3 . The 1 20

numerical values of the HGO parameters (for both the HGOs) 0.5 10

are chosen as  = 0.1, α1 = 15, and α2 = 50. For clarity of

δ (t) (deg)

δr(t) (deg)
0 0

presentation, we consider the following three cases:

a

−0.5 −10

−1 −20
10 20 30 40 10 20 30 40

A. The linear case: ẋ = Ax + Bu, fi (·) ≡ 0 μ = 0.1, with actuator dynamics

β
μ = 0.1, with actuator dynamics
φ
40 40

Our first simulation shows the performance of the control 20 20

δ (t) (deg)

designed in Section 3, but without integral action and when

δr(t) (deg)

0 0
there are no actuator dynamics. Since there is no need to use
a

−20 −20
SMC when the parameters are exactly known, in order to make
the simulation meaningful, we assume that the values of A −40
0 1 2 3 4 5
−40
0 1 2 3 4 5

and B are randomly perturbed from their nominal values by 1

μ = 1, with actuator dynamics
β
20
μφ = 1, with actuator dynamics

20%. The results are shown in Figure 1, for 2 different values

of both μβ and μφ , namely μ = 1 and μ = 0.1. From the
0.5 10
δ (t) (deg)

δr(t) (deg)

figure, it is clear that the controller achieves good performance, 0 0

a

even without integral control. The figure also shows that the −0.5 −10

errors are non-zero without integral control, but decrease as μ −1

10 20 30 40
−20
10 20 30 40
decreases. Time (sec) Time (sec)

The previous simulation shows that in the absence of

integral control, |e| = O(μ), and we must decrease μ in order
to achieve smaller steady-steady state errors, and this is clear Fig. 2. Effect of decreasing μ in the presence of (unmodeled) actuator
dynamics.
from the simulation results. However, smaller values of μ can
induce chattering when there are switching imperfections such
as delays or unmodeled actuator dynamics. On the other hand, As mentioned before, with integral control, we don’t need
the inclusion of integral action means that we don’t need to to decrease μ in order to obtain smaller errors. To that end, we
make μ very small to achieve small errors, only small enough repeat previous simulation with μ = 1, but with the conditional
to stabilize the equilibrium point. In order to emphasize this, integrator, and the results shown in Figure 3. It is clear that
we repeat the previous simulation, but assume first order the inclusion of integral control makes the error asymptotically
actuator lag dynamics with time constant τ = 20.2sec, i.e., converge to zero.
H(s) = s+20.2
20.2
. The simulation results are shown in Figure It is important to note that the results presented above
4 Robustness of the conditional integrator based SMC design to time
3 Inaddition to the saturated values of kβ and kφ of vβ and vφ respectively, delays has also been demonstrated by simulations for the case of control
we also saturate the controls δa and δr at their physically allowable limits of of longitudinal dynamics (specifically the pitch-rate) of an F-16 in a related
21.5 and 30◦ .
◦ paper [10].

83
 International Journal of Aerospace and Mechanical Engineering 2:2 2008

−3
x 10
0.4 4 0.05

with integrator 2
0.2 0
Error in β (deg)

Error in β (deg)

Error in φ (deg)
no integrator
0
0 −0.05
−2

−0.2 −0.1
−4

−0.4 −6 −0.15
0 10 20 30 40 10 20 30 40
5 10 15 20 25 30 35 40 45

2
1 15

0 10
0.5
Error in φ (deg)

with integrator 5

δa (deg)

δr (deg)
no integrator
−2 0
0
−5
−4 −10
−0.5
−15
−6
−1 0 10 20 30 40 0 10 20 30 40
5 10 15 20 25 30 35 40 45 Time (sec) Time (sec)
Time (sec)

Fig. 4. Asymptotic regulation in control-affine case with unknown nonlin-

Fig. 3. Recovery of asymptotic regulation with conditional integrator.
earities f (·, 0).

are only valid under output feedback. In particular, if one allowable value, and demonstrate the efficacy of the design
uses the original state x, then the disturbances are no longer through simulations. To do so, we repeat the simulation of the
matched when A is perturbed (any perturbation in B clearly previous subsection, but now with f (·, u), and not just f (·, 0).
constitutes a matched disturbance), so that if one uses ė1β = We also do so with several different initial conditions, with
C1 Ax − β̇ref to compute e2β and use it in the control (and the simulation results plotted in Figure 5. The figure clearly
similarly for ė1φ ), then the error will not be zero even with shows that the initial error is larger as we move from the first
integral control when A is different from its nominal value row to the last (the initial conditions for β and φ were chosen
(since our design explicitly uses the fact that the disturbances that way) but that the controller achieves good performance in
are matched). However, with the transformation to normal the presence of the control-dependent unknown nonlinearities
form, the disturbance is matched, and it is well-known (see, f (·, u).
for example, [6, Chapter 14]) that the HGO can be used
for this class of systems (where the states are the output
and its derivatives) to achieve asymptotic error regulation. A 0.05 2
Error in β (deg)

Error in φ (deg)

discussion of the effect of measurement noise on the HGO is 0

0
discussed in our related work in [17] on the control of F-16 −2
longitudinal dynamics, where we show that the performance is −0.05 −4
not degraded significantly when the HGO is used with noisy 0 10 20 30 40 0 10 20 30 40

measurements, and so we do not repeat similar simulations in 0.5 2

Error in β (deg)

Error in φ (deg)

this current work. 0

0
−2

B. The control affine case: ẋ = Ax + B(u + f (·, 0)) −0.5 −4

0 10 20 30 40 0 10 20 30 40
As shown in Section 3, the controller for this case has 5 10
Error in β (deg)

Error in φ (deg)

the exact same structure as that in the preceding case. Con- 0 5

sequently, we simply repeat some of the simulations in the
−5 0
preceding subsection, with some minor differences. There is
no need to perturb A and B in this case, since now the system −10
0 10 20 30 40
−5
0 10 20 30 40
already contains the unknown functions f (·, 0). Figure 4 shows Time (sec) Time (sec)

the tracking performance in the presence of the unknown

functions f (·, 0), and it is clear that the controller achieves Fig. 5. Asymptotic regulation in control non-affine case with unknown
good performance, and that the errors asymptotically converge nonlinearities f (·, u) for several different initial conditions.
to zero with the integral control.
Before we present our conclusions, we mention again that
our controller has a very simple structure, that of a saturated
C. The control non-affine case: ẋ = Ax + B(u + f (·, u)) PID controller with “anti-windup” like integrator, and an
When the function f (·) depends on u, we need to verify input-decoupling matrix. The only precise knowledge that
Assumption 1, and design the gains γz , z = β, φ accordingly. the controller requires is the relative degree of the system
As mentioned in the preceding section, we do not rigorously and the signs of the high-frequency gains. It is robust to
verify (numerically) that this assumption holds in our work, unmodeled actuator lag dynamics, and also to time-delays
but simply choose the gains to be their physically maximum and measurement noise (as demonstrated in related work). For

84
 International Journal of Aerospace and Mechanical Engineering 2:2 2008

the non-affine in the control case, our simulation results are [13] S. Seshagiri. Robust multivariable PI control: Applications to process
comparable to the ones presented in [23], where the controller control. In 17th IFAC World Congress, Seoul, S. Korea, July 2008.
[14] S. Seshagiri and H.K. Khalil. On introducing integral action in sliding
is much more complex, and uses RBFNNs in conjunction with mode control. In 2002 Conference on Decision and Control, 2002.
time-scale separation in an adaptive control design. We note [15] S. Seshagiri and H.K. Khalil. Position control of a PMSM using
that the design presented in this paper is flexible enough to conditional integrators. In 2005 American Control Conference, Portland,
Oregon, U.S.A, June 2005.
allow for error-dependent or time-varying gains γz and also [16] S. Seshagiri and H.K. Khalil. Robust output feedback regulation of
allow cancellation of any known/nominal terms in ṡ. This is minimum-phase nonlinear systems using conditional integrators. Auto-
elaborated upon in both [14], [16], but we did not mention it matica, 41(1):43–54, 2005.
[17] S. Seshagiri and E. Promtun. Sliding mode control of F-16 longitudinal
here for clarity of presentation. Finally, the design presented dynamics. In 2008 American Control Conference, Seattle, Washington,
here is valid for a more general class of nonlinear systems U.S.A, June 2008.
than the F-16 lateral dynamics considered in this work, and [18] J.S. Shamma and J.S. Cloutier. Gain-scheduled bank-to-turn autopilot
design using linear parameter varying transformations. Jnl. Guidance
has been successfully applied to the control of machines [15], Control and Dynamics, 9(5):1056–1063, 1996.
process control [13], and F-16 longitudinal dynamics [10], [19] Y. Shtessel, J. Buffington, and S. Banda. Tailless aircraft flight control
[17]. using multiple time scale reconfigurable sliding modes. IEEE Trans.
Ctrl. Sys. Tech., 10(2):288–62, 2002.
[20] S.A. Snell, D.F. Enns, and Jr. W.L. Garrard. Nonlinear inversion flight
V. C ONCLUSIONS control for a supermaneuverable aircraft. Jnl. Guidance, Control, and
This paper presents a new SMC design for control of the Dynamics, 15(4):976–984, 1992.
[21] M.S. Spillman. Robust longitudinal flight control design using linear
lateral dynamics of an F-16 aircraft, based on the conditional parameter-varying feedback. Jnl. Guidance, Control, and Dynamics,
integrator design of [14], [16]. The idea is based on rewriting 23(1):101–108, 2000.
the non-affine in the input system as a perturbation of a [22] H. Vo and S. Seshagiri. Robust control of F-16 lateral dynamics
(accepted). In 2008 IEEE IECON08, Orlando, Florida, U.S.A, Nov
control-affine system, with guaranteed analytical results for 2008.
stability and performance. The robustness of the method to [23] A. Young, C. Cao, N. Hovakimyan, and E. Lavretsky. An adaptive
modeling uncertainties and unmodeled actuator lag dynam- approach to nonaffine control design for aircraft applications. In AIAA
Guidance, Navigation, and Control Conference and Exhibit, Keystone,
ics is demonstrated through simulation, with the transient CO, USA, August 2006.
and steady-state performance comparable to that of a more
complex adaptive controller. In related work, we have shown
(via simulations) the robustness of the design to time-delays
and measurement noise for the output-feedback case. Conse-
quently, the results presented in this paper are a promising
start to demonstrate the efficacy of the conditional integrator
based SMC design to flight control.

R EFERENCES
[1] R.J. Adams, J.M. Buffington, and S.S Banda. Design of nonlinear
control laws for high-angle-of-attack flight. Jnl. Guidance, Control, and
Dynamics, 17(4):737–745, 1994.
[2] G.J. Balas. Flight control law design: An industry perspective. European
Jnl. of Ctrl., 9(2-3):207–226, 2003.
[3] J.M. Biannic and P. Apkarian. Parameter varying control of a high
performance aircraft. In Proc. AIAA, Guidance, Navigation and Control
Conference, pages 69–87, 1996.
[4] Y.J. Huang, T.C. Kuo, and H.K. Way. Robust vertical takeoff and
landing aircraft control via integral sliding mode. Control Theory and
Applications, IEE Proceedings-, 150:383–388, 2003.
[5] E.M. Jafarov and R. Tasaltin. Robust sliding-mode control for the
uncertain MIMO aircraft model F-18. IEEE Trans. Aerospace Electronic
Sys., 36(4):1127–1141, 2000.
[6] H.K. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River,
New Jersey, third edition, 2002.
[7] T. Lee and Y. Kim. Nonlinear adaptive flight control using backstepping
and neural networks controller. Jnl. Guidance, Control, and Dynamics,
24(4):675–682, 2001.
[8] B. Lu, F. Wu, and S. Kim. LPV antiwindup compensation for enhanced
flight control performance. Jnl. Guidance, Control and Dynamics,
28:495–505, 2005.
[9] J-F. Magni, S. Bennani, and J. Terlouw (Eds). Robust Flight Control: A
Design Challenge. Lecture Notes in Control and Information Sciences
- Vol 224. Springer, 1998.
[10] E. Promtun and S. Seshagiri. Sliding mode control of pitch-rate of an
F-16 aircraft. In 17th IFAC World Congress, Seoul, S. Korea, July 2008.
[11] W.C. Reigelsperger and S.S. Banda. Nonlinear simulation of a modified
F-16 with full-envelope control laws. Control Engineering Practice,
6:309–320, 1998.
[12] R. Rysdyk and A.J. Calise. Robust nonlinear adaptive flight control for
consistent handling qualities. IEEE Trans. Aut. Ctrl., 13(6):896–910,
2005.

85