Direct Adaptive Tracking Control of Quadrotor Aerial
Vehicles
Yannick Morel
Department of Mechanical, Materials and Aerospace
Engineering
University of Central Florida
Orlando, Florida 32816
ymorel@gmail.com
ABSTRACT
This paper presents a novel adaptive control algorithm
solving the trajectory tracking problem for quadrotor aerial
vehicles. The control law is derived using a backstepping
procedure. A technique derived from dynamic surface con-
trol is used to simplify the expression of the obtained control
algorithm, with no significant loss in terms of performance.
Proof of stability is obtained using Lyapunov stability the-
ory. Results from numerical simulations illustrate the per-
formance of the obtained controller.
1 . INTRODUCTION
Numerous research group have in recent years focused
on control of Unmanned Aerial Vehicles (UAVs). This fam-
ily of vehicles can be divided in two distinct classes. Air-
planes can reach considerable speed, allowing them to cover
long distances in a relatively short amount of time. These
vehicles naturally lend themselves well to mission pertaining
to surveillance of extensive areas. They can be distinguished
from Vertical Take Off and Landing (VTOL) vehicles, as the
latter are not able to reach comparable speeds, but are physi-
cally able to hover in the vicinity of an arbitrary entity. They
are therefore particularly well suited for missions involving
the monitoring of stationary or slow moving targets.
Focussing on the VTOL category, another distinction
can be made between classical helicopters, which essen-
tially use a single main rotor for lift in conjunction with an
auxiliary tail rotor used to adjust the vehicle’s attitude, and
quadrotors, which use a combination of four rotors for both
lift and attitude. Although quadrotors feature more rotors
than classical helicopters, they are of a simpler mechanical
design, as the latter requires a tilting mechanism to orient the
main rotor to adjust pitch and roll, which is more simply ac-
complished, for quadrotors, using differential lift on pairs of
rotors. This simplified mechanical design however puts the
onus on the control system to provide levels of agility and
maneuverability comparable to that of classical helicopters.
Controlling the motion of a quadrotor UAV is a chal-
- 1-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006
Alexander Leonessa
Department of Mechanical, Materials and Aerospace
Engineering
University of Central Florida
Orlando, Florida 32816
aleo@mail.ucf.com
lenging task. The interaction of the air flows generated by the
four rotors contribute to complex aerodynamic forces affect-
ing the vehicle’s motion. The system’s dynamics are not only
nonlinear, but also difficult to satisfactorily characterize, due
to the complexity of the system’s aerodynamic properties.
Accordingly, due to the nonlinear nature of the system,
linear control techniques are expected to perform relatively
poorly ([1]), as they can not account nor compensate for non-
linear phenomenon affecting the vehicle’s dynamics. In [2],
a Proportional Integral Derivative (PID) controller and a Lin-
ear Quadratic (LQ) controller were implemented and proved
capable of regulating the system. However, these designs
rely on a linearized version of the vehicle’s model. The con-
trollers perform adequately if the vehicle is in a configura-
tion close to that the model was linearized about, but the per-
formance dramatically deteriorates if the vehicle state sig-
nificantly differs from this desired equilibrium. As a result,
these controllers have difficulties in handling perturbations,
and seem unlikely to allow for trajectory tracking.
As stated above, a nonlinear controller should be better
suited to handle and account for nonlinearities in the vehi-
cle’s model. Indeed, a backstepping technique ([3]) yielded
a control algorithm that, when implemented on the same sys-
tem as that considered in [2], outperformed the PID and LQ
controllers ([4]). Similarly, in [5, 6], feedback linearization
([7]) and a different backstepping approach were used and al-
lowed to satisfactorily control the vehicle’s altitude and yaw.
As mentioned above, due to the presence of intricate
aerodynamic forces, obtaining a mathematical model satis-
factorily describing a quadrotor dynamical behavior is a dif-
ficult task. To handle uncertainties in the model, a robust
feedback linearization approach was used in [8]. The ob-
tained control law proved to be robust to parameter uncer-
tainties in simulation.
Classically, model uncertainties are handled using adap-
tive techniques. In [9], a Single Hidden Layer Neural Net-
work (SHL-NN) is used to compensate for unknown dy-
namics and allows implementation of the obtained controller
with almost no knowledge of the system’s dynamics. The ap-
Miami, Florida, May 25-26, 2006
proach was extended in [10], allowing the control algorithm where
to satisfy control input saturation constraints, and success-
fully implemented on a physical system. The algorithm is  
cθ cψ sφ sθ cψ − cφ sψ cφ sθ cψ + sφ sψ
limited to control of the vehicle’s pitch. J1 (η2 ) ,  cθ sψ sφ sθ sψ + cφ cψ cφ sθ sψ − sφ cψ , (7)
In this paper, we propose a novel adaptive control algo- −sθ sφ c θ cφ cθ
rithm, solving the attitude and altitude tracking control prob-  
lem for quadrotor aerial vehicles. A direct adaptive control 1 sφtθ cφtθ
law is derived, using an integrator backstepping approach J2 (η2 ) ,  0 cφ −sφ , (8)
([3]). Section 2 details the considered mathematical model. 0 sφ /cθ cφ /cθ
A step by step description of the controller derivation is pre-
sented in Section 3. Results of numerical simulations illus- and sς , cς , and tς denote the sine, cosine and tangent of angle
trate the performance of the controller. ς. Equations (5) and (6) constitute the kinematic equations
of the considered system.

2 . MATHEMATICAL MODEL
2.1 Kinematics 2.2 Dynamics
Considering motion in six degrees of freedom, the vehi- The equations characterizing the dynamic behavior of
cle’s position is described by the system are of the form ([11])

£ ¤T 1 1£ ¤T
η1 , xN yE z , (1) ν̇1 (t) = C1 (ν2 (t))ν1 (t) + fg (φ(t), θ(t)) + 0 0 fz (t) ,
m m
ν1 (0) = ν10 , t > 0,(9)
where (xN , yE ) denotes the position of the vehicle in a two
ν̇2 (t) = I −1C2 (ν2 (t))ν2 (t) + I −1 τ(t), ν2 (0) = ν20 , (10)
dimensional frame, the xN axis is arbitrarily chosen to point
North, the yE axis East, while the z axis is pointing down-
ward. The vector η1 is expressed in an Earth Fixed Frame where m is the mass of the vehicle,
(EFF) of arbitrary origin. The vehicle’s attitude is described
by  
ix 0 0
£ ¤T I ,  0 iy 0  , (11)
η2 , φ θ ψ , (2) 0 0 iz

where φ, θ and ψ denote the vehicle’s roll, pitch, and yaw, is the rotational inertia matrix,
respectively. Position and attitude are regrouped in η ,
£ T T ¤T    
η1 η2 . The body-fixed velocities are introduced with 0 r −q 0 −iz r iy q
the following notation, C1 (η2 ), −r 0 p , C2 (η2 ), iz r 0 −ix p ,(12)
q −p 0 −iy q ix p 0
£ ¤T
ν1 , u v w , (3)
£ ¤T £ ¤T
ν2 , p q r , (4) are Coriolis matrices, fg (φ, θ) , g −sθ cθ sφ cθ cφ is the
vector of restoring forces, g is the gravity acceleration, Fz is
where u, v and w are the forward, lateral and downward ve- the thrust provided by the propulsion system, and τ regroups
locities, expressed in the Body Fixed Frame (BFF), while the torques and moments generated by the propellers.
p, q and r are the angular velocities in roll, pitch and yaw, We augment the dynamical system (9)-(10) with the fol-
respectively. Linear and angular velocities are regrouped in lowing equation, which models the delay between a change
£ ¤T in a Pulse Width Modulation (PWM) signal provided to the
the velocity vector ν , νT1 νT2 . The overall state of the
£ ¤T motors and the time at which it affects the remaining dynam-
vehicle is described by the state vector x , ηT νT .
ics of the vehicle,
The time derivatives of (1) and (2) are related to (3) and
(4) as follows ([11]), ¡ ¢
τ̃˙ (t) = T −1 −τ̃(t) + g2m B̃Λ−1 p2 (t) , τ̃(0) = τ̃0, t > 0, (13)
η̇1 = J1 (η2 )ν1, (5)
η̇2 = J2 (η2 )ν2, (6) where T ∈ R4×4 is a diagonal matrix of time constants, T >

- 2-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006 Miami, Florida, May 25-26, 2006
 £ ¤T
0, τ̃ , fz τT , gm ∈ R, Now consider the following filter,

    ³ · ¸T
k 0 0 0 1 0 2 1 03×1
0 lk 0 0 1 1 −2 0 −1  T1 χ̇1(t)+χ1(t)=J3−1(η2 (t)) ksz
B̃ ,  , Λ,   , (14) u(t)sθ (t)−v(t)sφ (t)cθ (t)
0 0 lk 0 
4 1 0 −2 1  ´ 1
0 0 0 d 1 2 0 −1 +G1 ep (t) + J3T (η2 (t))ep (t), t > 0, (19)
2

where k, l and d ∈ R, and, finally, p2 ∈ R4 regroups the where G1 ∈ R4×4 , G1 > 0, with initial condition χ1 (0) = χ10
squares of the PWM signals provided to the four motors. and associated filtering error

à · ¸T !
3 . DIRECT ADAPTIVE CONTROL ALGO- 03×1
α1 (xr1 , x, χ1 ) , J3−1(η2 ) ksz +G1 ep (xr1 , ηs )
RITHM usθ −vsφ cθ
In this Section, we will introduce the considered track- 1
ing errors. Then, following an approach similar to that pre- + J3T (η2 )ep (xr1 , ηs ) − χ1 , (20)
2
sented in [12], we will derive a nonlinear control law which
solves the considered tracking problem. Note that the ve- and define the velocity error
hicle does not directly track a desired trajectory, but that of
a reference system corresponding to a virtual vehicle. This
reference system, composed of third order oscillators ([13]), ev (xr2 , x, χ1 ) , J3−1(η2 )xr2 + χ1 − νs . (21)
will track the desired trajectory. The state of this reference
system is noted xr ∈ R12 . Using (20) and (21), we can rewrite (17) as

µ ¶
3.1 Tracking Error 1 T
£ ¤T V̇1 (ep , ev , η2 , α1 ) = −eTp G1 ep + ep J3(η2 ) α1 − J3 (η2 )ep
We define ηs , ηT2 ksz z , where ksz ∈ R is a scal- 2
ing constant. The position of the reference system is given
£ T ¤T +eTv J3T(η2 )ep , (22)
by xr1 , xra ksz xrz , where xra ∈ R3 corresponds to three
reference angles, and xrz ∈ R is a reference altitude. The the first term of which is negative quadratic, the second is due
considered position error is of the form to the filtering error, and third is due to the velocity error.

ep (xr1 , ηs ) , xr1 − ηs . (15)

3.3 Second Lyapunov Function Candidate
Consider the following function of ep and ev ,
3.2 First Lyapunov Function Candidate
Consider the following function of ep ,
1 1
V2 (ep , ev ) , eTp ep + eTv ev , (23)
1 2 2
V1 (ep ) , eTp ep , (16)
2
which is C 1 for all ep , ev ∈ R4 , V2 (ep , ev ) > 0 for all ep , ev ∈
note that V1 (ep ) ∈ C 1 for all ep ∈ R4 , V1 (ep ) > 0 for all ep ∈ R4 \ {0}, and V2 (0, 0) = 0. The time derivative of V2 (ep , ev )
R4 \ {0}, and V1 (0) = 0. The time derivative of V1 (ep ) is is given by
given by ³ ep ´
³ V̇2 (epv , xr , x, α1 , τ̃) = −eTp G1 ep + ep J3(η2 ) α1 − J3T(η2 )
£ ¤T´ ¡ ¢ 2
V̇1(xr , x) = eTp xr2 −J3 (η2 )νs+ksz 01×3 usθ −vsφ cθ ,(17) +eTv J3T(η2 )ep + ėv (xr , x, α1 , τ̃) , (24)

£ ¤T £ ¤T
where xr2 , ẋr1 , νs , νT2 krz w , 0i× j is the i × j dimen- where epv , eTp eTv , and
sional zero matrix, and
· ¸ ėv(xr , x, τ̃, α1)=M −1Θ∗1 fev(xr , x, α1)+J3−1(η2)xr3 −M −1B̂τ̃,
J2 (η2 ) 03×1
J3 (η2 ) , . (18)
01×3 cφ cθ (25)

- 3-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006 Miami, Florida, May 25-26, 2006
with The first term in (33) is due to the estimation error Θ̃1 , the
second one to the tracking error eτ , the third term is negative
fev (xr , x, α1 ) , quadratic, the fourth and fifth terms are related to filtering
" £ ¤T # errors.
qr pr pq pv − qu
£ ¤T ¡ ¢ , (26)
ksz g 01×3 −cφ cθ + dtd J3−1 (η2 ) xr2 + T1−1 α1
· ¸ 3.4 Third Lyapunov Function Candidate
I 03×1 £ £ ¤ ¤ Consider the following function of ep , ev and Θ̃1 ,
M, m , Θ∗1 , diag iz − iy ix − iz iy − ix 1 M .
01×3 ksz
(27) 1 1 ¡ ¢
V3 (epv , Θ̃1 ) , eTpv epv + tr M −1 Θ̃1 Γ−1
1 Θ̃1 ,
T
(35)
2 2
We now define a second filter as follows,
· ¸ where Γ1 ∈ R8×8 , Γ1 > 0. Note that V3 (epv , Θ̃1 ) is C 1 for all
T2 χ̇2(t)+χ2(t) = fev (t) + −1
04×1
, epv ∈ R8 and Θ̃1 ∈ R4×8 , V3 (epv , Θ̃1 ) > 0 for all (epv , Θ̃1 ) ∈
J3 (η2 (t))ep (t) + G2 ev (t) R8 ∪ R4×8 \ {0}, and V3 (0, 0) = 0. The time derivative of
χ2 (0) = χ20 , t > 0, (28) V3 (epv , Θ̃1 ) is given by

where G2 ∈ R4×4 , G2 > 0, with associated filtering error V̇3 (epvτ , xr , x, α, Θ̃1 ,t)=−eTpv G12 epv +eTτ B̂T M −1ev
· ¸ 1
04×1 +eTpv P(η2 )α − eTp J3 (η2 )J3T (η2 )ep ,
α2 (epv , xr , x, χ1 ) , fev (xr , x, α1 )+ −χ2, ¡ ¡ 2 ¢¢
J3−1 (η2 )ep + G2 ev −tr M −1 Θ̃1 ϕ1(xr3 , η2 , χ2)ev − Γ−1 1 Θ̇1 (t)
T
.(36)
(29)
Choosing
and define the new tracking error

Θ̇1 (t) = ev (t)ϕT1 (t)Γ1 − σ1 kev (t)k2 Θ1 (t),

eτ (xr , x, χ2 , τ̃, Θ1 ) , B̂T Θ1 ϕ1 (xr3 , η2 , χ2 ) − τ̃, (30)
Θ1 (0) = Θ10 , t > 0,(37)

where Θ1 is an estimate of Θ∗1 and

where σ1 ∈ R, σ1 > 0, we obtain
µ · ¸¶
04×1
ϕ1 (xr3 , η2 , χ2 ) , χ2 + . (31) V̇3 (epvτ , xr , x, α, Θ̃1 )=−eTpv G12 epv +eTτ B̂T M −1ev +eTpv P(η2 )α
J3−1 (η2 )xr3
1 ¡ ¢
− eTp J3 (η2 )J3T (η2 )ep − σ1 kev k2 tr M −1 Θ̃1 Γ−1 1 Θ1 .(38)
T
Using (29) and (30), we can rewrite (25) as 2

ėv (epvτ , xr , x, χ2 , α2 , Θ̃1 ) = −J3T (η2 )ep − G2 ev 3.5 Fourth Lyapunov Function Candidate
¡ ¢ Consider the following function of epvτ and Θ̃1 ,
+M −1 Θ∗1 α2 − Θ̃1 ϕ1 (xr3 , η2 , χ2 ) + B̂eτ ,(32)

£ ¤T 1 1 ¡ ¢
where Θ̃1 , Θ1 −Θ∗1 , and epvτ , eTpv eTτ . Combining (32) V4 (epvτ , Θ̃1 ) , eTpvτ epvτ + tr M −1 Θ̃1 Γ−1
1 Θ̃1 , (39)
T
2 2
and (24), we obtain

which is C 1 for all epvτ ∈ R12 and Θ̃1 ∈ R4×8 , V4 (epvτ , Θ̃1 ) >
V̇2 (epvτ , xr , x, α, Θ̃1 )=−ev M −1Θ̃1ϕ1(xr3 , η2 , χ2)+eTτ B̂T M −1ev 0 for all (epvτ , Θ̃1 ) ∈ R12 ∪ R4×8 \ {0}, and V4 (0, 0) = 0. The
1
−eTpv G12 epv + eTpv P(η2 )α − eTp J3 (η2 )J3T (η2 )ep ,(33) time derivative of V4 (epvτ , Θ̃1 ) is given by
2

where V̇4 (epvτ , xr , x, α, Θ̃1 ,t) = −eTpv G12 epv + eTpv P(η2 )α
1 ¡ ¢
· ¸ · ¸ − eTp J3 (η2 )J3T (η2 )ep − σ1 kev k2 tr M −1 Θ̃1 Γ−1
1 Θ1
T
G1 04×4 J3 (η2 ) 04×8 2 ¡ ¢
G12 , , P(η2 ) , . (34)
04×4 G2 04×4 M −1 Θ∗1 +eTτ B̂T M −1 ev + ėτ (t) , (40)

- 4-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006 Miami, Florida, May 25-26, 2006
where time derivative of V5 (epvτ , Θ̃) is given by

¡
ėτ (xr , x, Θ1 , χ2 , α2 , r̂, τ̃, p2 )= B̂T Θ̇1 (xr , x, Θ1 )ϕ1 (xr3 , η2 , χ2 ) V̇5 (epvτ , xr , x, α, Θ̃,t) = −eTpvτ G123 epvτ + eTpv P(η2 )α
µ ¶
d −1 1 ¡ ¢
−1
+Θ1 T2 α2 + Θ12 J (η2 ) xr3 − eTp J3 (η2 )J3T (η2 )ep − σ1 kev k2 tr M −1 Θ̃1 Γ−1 1 Θ1
T
dt 3 2¡ ¡ ¢¢
¢ −tr B̄Θ̃2 ϕ2(t)eTτ − Γ−1 2 Θ̇2 (t)
T
+Θ12 J3−1 (η2 )ẋr3 (xr , r̂) + T −1 τ̃ − B̄Λ−1 p2 , (41) . (48)

where Θ12 is the (1, 4) × (5, 8) block of Θ1 . Now consider Choosing

the control command
Θ̇2 (t) = eτ (t)ϕT2 (t)Γ2 − σ2 keτ (t)k2 Θ2 (t),
p2 (t) , ΛΘ2 (t)ϕ2 (t), t > 0, (42) Θ2 (0) = Θ20 , t > 0,(49)

where £Θ2 ∈ R4×12 ¤ is an estimate of Θ∗2 , where σ2 ∈ R, σ2 > 0, we obtain

1 −1
g2m
B̃ T T −1 B̂T M −1 I4 , I j is the jth dimensional
identity matrix, and V̇5 (epvτ , xr , x, α, Θ̃) = −eTpvτ G123 epvτ + eTpv P(η2 )α
1 ¡ ¢
£ ¤T − eTp J3 (η2 )J3T (η2 )ep − σ1 kev k2 tr M −1 Θ̃1 Γ−1 1 Θ1
T

ϕ2 (t) , τ̃T (t) eTv (t) ϕ̃T2 (t) , t > 0, (43) 2 ¡ ¢

³ −σ2 keτ k2 tr B̄Θ̃2 Γ−1 2 Θ2 .
T
(50)
ϕ̃2 (xr , x, Θ1 , χ2 , α2 , r̂, eτ ) , B̂T Θ̇1 (xr , x, Θ1 )ϕ1 (xr3 , η2 , χ2 )
µ ¶
d −1 Note that, by completion of the square,
+Θ1 T2−1 α2 + Θ12 J3 (η2 ) xr3
dt
´ 1
−1
+Θ12 J3 (η2 )ẋr3 (xr , r̂) + G3 eτ , (44) eTp J3 (η2 )α1 = − (J3T (η2 )ep − α1 )T (J3T (η2 )ep − α1 )
2
1 T 1
e J3 (η2 )J3T (η2 )ep + kα1 k2 . (51)
Combining (41) and (42), we obtain 2 p 2

In addition, for kTi k sufficiently small, kαi (t)k < ᾱi , t > 0,
ėτ (t) = −B̂T M −1 ev (t) − G3 eτ (t) − B̄Θ̃2 (t)ϕ2 (t), (45)
i = 1, 2, and we obtain the following upper bound on V̇5 ,

where Θ̃2 = Θ2 − Θ∗2 , and B̄ = g2m T −1 B̃. Hence, 1

V̇5 (epvτ , xr , x, α, Θ̃) 6 −eTpvτ G123 epvτ + ᾱ21 + kev kM −1 Θ∗1 ᾱ2
2
V̇4 (epvτ , xr , x, α, Θ̃,t) = −eTpvτ G123 epvτ + eTpv P(η2 )α σ1 ¡ ¡ ¢ ¡ −1 ¢¢
+ kev k2 tr M −1 Θ∗1 Γ−1 ∗T
1 Θ1 − tr M Θ̃1 Γ1 Θ̃1
−1 T

¡ ¢ 2
1
− eTp J3 (η2 )J3T (η2 )ep − σ1 kev k2 tr M −1 Θ̃1 Γ−1 σ2 ¡ ¡ ¢ ¡ ¢¢
1 Θ1
T
2 + keτ k2 tr B̄Θ∗2 Γ−1 ∗T −1 T
2 Θ2 − tr B̄Θ̃2 Γ2 Θ̃2 , (52)
2
−eTτ B̄Θ̃2 ϕ2 (t), (46)
Which guarantees convergence of (epvτ , Θ̃) ([14]) to the
£ ¤ £ ¤
where Θ̃ , Θ̃1 Θ̃2 , and G123 , diag G1 G2 G3 . compact set

n 1
3.6 Fifth Lyapunov Function Candidate M , (epvτ , Θ̃) : eTpvτ G123 epvτ 6 ᾱ21 + kev kM −1 Θ∗1 ᾱ2
2
Consider the following function of epvτ and Θ̃, ¡ ¢ ¡ −1 ∗ −1 ∗T ¢
tr M −1 Θ̃1 Γ−1
1 Θ̃1 6 tr M Θ1 Γ1 Θ1
T

¡ ¢ ¡ ∗ −1 ∗T ¢ o
tr B̄Θ̃2 Γ−1 Θ̃T
6 tr B̄Θ2 Γ2 Θ2 .(53)
1 1 2 ¡ ¢ 2 2
V5 (epvτ , Θ̃) , eTpvτ epvτ + ∑ tr M −1 Θ̃i Γ−1
i Θ̃i , (47)
T
2 2 i=1
4 . NUMERICAL SIMULATION
which is C 1 for all epvτ ∈ R12 and Θ̃ ∈ R4×20 , V5 (epvτ , Θ̃) > 0 The control law (42) presented in Section 3 was tested
for all (epvτ , Θ̃) ∈ R12 ∪ R4×20 \ {0}, and V5 (0, 0) = 0. The through numerical simulation. The reference system used is

- 5-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006 Miami, Florida, May 25-26, 2006
of the form bustness to unmodelled or mismodelled dynamics, and com-
pensation for external perturbations such as winds.
      
ẋr1 (t) 04×4 I4 04×4 xr1 (t) 04×4
 ẋr2 (t)  =  04×4 04×4 I4   xr2 (t)  +  04×4  r̂(t),
REFERENCES
ẋr3 (t) −ω20 −ω20 −Ar4 xr3 (t) ω20
[1] Waslander, S. L., Hoffmann, G. M., Jang, J. S., and
xr (0) = xr0 , t ≥ 0, (54) Tomlin, C. J., 2005. “Multi-agent quadrotor testbed
control design: Integral and sliding more vs. reinforce-
ment learning”. In Proc. 2005 IEEE/RSJ Int. Conf. In-
where ω0 = 10.2 I4 , and Ar4 = 38.8 I4 . In addition, the telligent Robots and Systems, pp. 3712–3717.
following values were used in the vehicle’s model, m = [2] Bouadballah, S., Noth, A., and Siegwart, R., 2004.
12, ix = 20, iy = 10, iz = 1, gm = 10, l = 0.25, k = 2, “PID vs LQ control techniques applied to an indoor
d = 0.1, g = 9.81, T = 0.01 I4 . The gain were arbitrar- micro quadrotor”. In Proc. 2004 IEEE/RSJ Int. Conf.
ily chosen to be G1 = G2 = G3 = 10 I4 , and the learning Intelligent Robots and Systems, pp. 2451–2456.
coefficients Γ1 = I12 , Γ2 = I12 . The initial conditions
£ ¤for
[3] Krstić, M., Kanellakopoullos, I., and Kokotović, P.,
T
the presented simulation are as follows, η10 = 0 0 10 , 1995. Nonlinear and Adaptive Control Design. John
£ ¤ £ ¤ Wiley and Sons, Inc, New York, NY.
η20 = π8 1 1 1 , ν = 06×1 , xr0 = 16 π
T T
1 1 1 01×9 , and the [4] Bouadballah, S., and Siegwart, R., 2005. “Backstep-
desired trajectory that xr1 is tracking is of the form xd (t) = ping and sliding-mode techniques applied to an indoor
£ ¤T π
a sin(bt) −a sin(bt) a2 sin(bt) −10 ,t > 0, with a = 12 , micro quadrotor”. In Proc. 2005 IEEE Int. Conf. on
2π Robotics and Automation, pp. 2247–2252.
b = 10 . This reference trajectory has the vehicle elevating
in altitude while its orientation is oscillating. [5] Altug, E., Ostrowski, J. P., and Mahony, R., 2002.
“Control of a quadrotor helicopter using visual feed-
40
actual
back”. In Proc. 2002 IEEE Int. Conf. on Robotics and
desired Automation, pp. 72–77.
20 reference [6] Altug, E., Ostrowski, J. P., and Taylor, C. J., 2003.
Angles [deg]

“Quadrotor control using dual camera visual feed-

back”. In Proc. 2003 IEEE Int. Conf. on Robotics and
0
Automation, pp. 4294–4299.
[7] Isidori, A., 1995. Nonlinear Control Systems. Springer-
-20
0 2 4 6 8 10
Verlag, New York, NY.
[8] Mokhtari, A., Benallegue, A., and Daachi, B., 2005.
-20 “Robust feedback lineariztion and GH ∞ controller for
a quadrotor unmanned aerial vehicle”. In Proc. 2005
-10 IEEE/RSJ Int. Conf. Intelligent Robots and Systems,
Altitude [m]

0
10
20
0 2 4 6
Time [s]
Figure 1. Orientation and altitude.
As seen on Figure 1 (top), the actual roll (in blue), pitch
(green) and yaw (red) converge rapidly and smoothly to their
corresponding desired trajectory. Similarly, the altitude con-
verges to its desired value (bottom).
5 . CONCLUSION
We presented a novel direct adaptive controller solving
the tracking control problem for quadrotor UAVs. A step
by step description of the algorithm derivation was provided,
and the obtained control law was tested through numerical
simulation. Future improvement on the presented result in-
clude the addition of a SHL-NN allowing for increased ro-
pp. 1198–1203.
[9] Hovakimyan, N., Nardy, F., Calise, A., and Nakwan,
zd K., 2002. “Adaptive output feedback control of uncer-
x rz tain nonlinear systems using single-hidden-layer neural
z
networks”. IEEE Trans. on Neural Networks, 13,
8 10
pp. 1420–1431.
[10] Kutay, A. T., Calise, A., and Hovakimyan, N., 2005.
“Experimental results on adaptive output feedback con-
trol using a laboratory model helicopter”. IEEE Trans.
on Control Systems Technology, 13, pp. 196–202.
[11] Nelson, R. C., 1997. Flight Stability and Automatic
Control. McGraw-Hill.
[12] Leonessa, A., VanZwieten, T. S., and Morel, Y., 2004.
“Neural network model reference adaptive control of
marine vehicles”. Current Trends in Nonlinear Systems
and Control.
[13] Morel, Y., 2002. “Design of an adaptive nonlinear con-
troller for an autonomous underwater vehicle equipped
with a vectored thruster”. Master Thesis, Florida At-
lantic University.
[14] LaSalle, J., and Lefschets, S., 1961. Stability by Lia-
punov’s Direct Method with Applications. Academic
Press.

- 6-
2006 Florida Conference on Recent Advances in Robotics, FCRAR 2006 Miami, Florida, May 25-26, 2006