Modelling and Control of a Quad-Rotor Robot

Paul Pounds, Robert Mahony, Peter Corke

Australian National University, Canberra, Australia
CSIRO ICT Centre, Brisbane, Australia
[email protected], [email protected], [email protected]
Abstract
To date, most quad-rotor aerial robots have
been based on ying toys. Although such
systems can be used as prototypes, they are
not suciently robust to serve as experimental
robotics platforms. We have developed the X-4
Flyer, a quad-rotor robot using custom-built
chassis and avionics with o-the-shelf motors
and batteries, to be a highly reliable experi-
mental platform. The vehicle uses tuned plant
dynamics with an onboard embedded attitude
controller to stabilise ight. A linear SISO con-
troller was designed to regulate yer attitude.
1 Introduction
A major limitation of helicopters is the need for exten-
sive, and costly, maintenance for reliable ight. Un-
manned Air Vehicles (UAVs) and Micro Air Vehicle
(MAV) rotorcraft are no exception. Simplifying the me-
chanical structure of a ying machine produces clear
benets for the logistics of operating these devices.
Quad-rotors are robust and simple helicopters as they
do not have the complicated swashplates and linkages
found in conventional rotorcraft. The majority of four-
rotor aerobots are constructed from remote-control toy
components. As a result, these craft lack the necessary
reliability and performance to be practical experimental
platforms.
1.1 Existing Quad-Rotor Platforms
Several quad-rotor craft have been developed recently,
for use as a toy or for research. Many research quad-
rotors began life as a commercially available toy, such
as the HMX-4 and RCtoys Draganyer. Unmodied,
these craft typically consist of light airframes with plastic
rotors. They are powered by NiCd or Li-Poly cells and
use rate feedback from MEMS gyros. These quad-rotors
generally have no attitude stability.
Research quad-rotors add automatic stability and use
a variety of hardware and control schemes. CSIROs
Figure 1: X-4 Flyer Mark II.
quad-rotor yer, for example, is a Draganyer deriva-
tive that uses visual servoing and an Inertial Measure-
ment Unit (IMU) to stabilise the craft over a blob tar-
get. Other quad-rotors include Eidgenossische Technis-
che Hochschule Zurichs OS4 [Bouabdallah et al, 2004],
a belt-driven yer with low-aspect ratio blades; CEAs
X4-yer
1
, a small quad-rotor with four blades per mo-
tor [Guenard et al, 2005]; and Cornells Autonomous
Flying Vehicle, a large craft using hobby aeroplane pro-
pellers.
The Australian National Universitys (ANU) X-4
Flyer quad-rotor MAV (cf. Fig. 1) aims to address the
problems faced by small-scale UAVs. The X-4 is much
heavier than similar robots: it weighs 4 kg total and is
designed to carry a 1 kg payload. It has a strong carbon-
bre and aluminium chassis and a high thrust-to-weight
ratio. The motors and cells used are o-the-shelf compo-
nents. The motors directly drive the rotors, eliminating
the need for a gearbox the robot has only eight moving
1. Although similarly named, the ANU X-4 Flyer and CEA
X4-yer are quite dierent craft
parts. As a result, the yer is rugged and reliable with
little scope for catastrophic failure in ight. It promises
a practical payload capacity with a substantial ight du-
ration.
1.2 Goals of Current Development
High-performance rotors and speed controllers have been
developed for the X-4 Flyer. These have adequately
solved the problems of thrust generation and dynamic
motor speed performance [Pounds et al, 2005], [Pounds
et al, 2007]. In addition, a model of the ight dynamics,
including rotor apping eects, was derived. A 3D simu-
lator of the craft generated state trajectories of the robot
for a variety of congurations, subjected to disturbances.
Current work on the yer aims to stabilise the aircraft
in roll, pitch and yaw. Continuous ight requires the
pitch and roll angles to remain around zero, except when
actively translating. The natural instability of ying sys-
tems requires active compensation. The special design
for the chassis results in purely divergent instability in
pitch and roll that a controller can readily correct.
In this paper we present the X-4 Flyer as a fully-
functional aerial robot. The dynamics of quad-rotor
helicopters with blade apping are studied. We estimate
the system parameters from data to produce a numerical
plant model. Based on a 6DOF aerodynamic model we
derive decoupled dynamics in longitudinal (pitch/roll)
and azimuthal modes. The control approach is to opti-
mise the mechanical design for control of these dynamics
and implement linear SISO control in the decoupled dy-
namics. We describe the controller used to stabilise the
craft in simulation and then go on to demonstrate the
function of the roll and pitch compensation in tethered
ight.
2 X-4 Hardware and Construction
The X-4 Flyer is set apart from other quad-rotor vehi-
cles by its larger construction. It consists of a chassis,
motors and power cells, and attitude control and com-
munications avionics. Each subsystem is described in
detail below:
2.1 Chassis
The X-4 has an aluminium centre frame with carbon
bre-foam sandwich arms. Regularly spaced mounting
points allows the CoG to be shifted easily. Motors and
batteries are mounted as far from the central axis as
possible. The arms angle down slightly to provide more
clearance between the bottom of the arms and ap-
ping rotor tips. The rotor mounts are teetering hubs,
a freely pivoting joint between the drive shafts and ro-
tor blades, machined from aluminium. The blades are
screw-clamped between the rotor mount top and bottom
plates.
2.2 Drive System
The X-4s rotors are designed to lift the yer with an ad-
ditional 30 per cent control margin (greater than 5.2 kg).
The blades are three-ply carbon bre and were designed
and fabricated at the ANU. The geometry is designed so
that the rotor tips ex to the optimal operating angle
under load. The ANUX2 airfoil used is a custom section
made specially for the rotors.
The rotors are driven by Jeti Phasor 30-3 three-phase
brushless motors for radio-controlled aircraft. They oer
high torque performance that allows for direct drive of
the rotors, eliminating the need for gearing. The motors
can pass more than 300 W and are rated up to 35 A.
Custom motor control boards commutate the motors.
These were developed by the CSIRO Queensland Centre
for Advanced Technology ICT group. The boards are
based around the Freescale HC12D60A microprocessor
and Toshiba TB9060 brushless motor speed control chip.
Power is provided by 24 Li-Poly 2000 mAh high-
discharge cells. Each cell has a nominal voltage of 3.7 V,
ranging from 4.2 V fully charged and dropping to 3 V at
depletion. Each cell can deliver up to 20 A. The batter-
ies are connected to a power bus of six parallel sets of
four cells in series; that is, 14.8 V nominal voltage and
120 A of current draw per motor. This gives the yer an
expected ight time of 11 minutes at hover speed.
2.3 Control
The craft is stabilised by an onboard embedded HC12
controller. The controller reads attitude from a CSIRO
Eimu IMU that provides angular rate and acceleration
measurements and angular position estimates at 50 Hz.
The controller outputs rotor speed references to the mo-
tor control cards over the CANbus, also at 50 Hz.
2.4 Command and Telemetry
Human directions to the robot and information about
the X-4s state are transmitted over a long-range Blue-
tooth serial module connected to a laptop base station
running Linux. The Bluetooth unit has a range of up
to 100 m. Telemetry from the yer is logged by the
base station and displayed on-screen. The user can is-
sue commands via the laptop using the keyboard and a
JR-X3810 radio handset.
The radio handset can also trigger a safety kill switch
on the X-4, independently of the Bluetooth communi-
cations channel, using an onboard radio receiver. In an
emergency the kill switch can stop the rotors instantly
by disabling the motor control boards, even if data com-
munications is lost.
3 Quad-Rotor Dynamics
The dynamic model described in [Pounds et al, 2004]
added articulated apping rotors to the basic quad-rotor
Figure 2: Flapping Quad-Rotor Free-body Diagram.
rigid body dynamics model. The current conguration
of the X-4 Flyer does not incorporate the hub-springs
originally included in the model. As a result, the apping
equations can be substantially simplied:
The right-hand inertial frame is denoted by
I= {E
x
, E
y
, E
z
}, where x is aligned with the front of the
craft and z is in the direction of gravity, and = (x, y, z)
is the origin of the body xed frame A ={E
a
1
, E
a
2
, E
a
3
}.
The frame A is related to I by the rotation matrix
R : A I. V and are the linear and angular ve-
locities of the frame in A (cf. Fig. 2).
The equations are:

= RV (1)
m

V = m V +mgR
T
e
3
+

N,S,E,W
T
i
(2)

R = R sk () (3)
I

= I +

N,S,E,W
[Q
i
+M
i
] (4)
T
i
= C
T
Ar
2

2
i
_
_
s
a
1
s
i
c
a
1
s
i
s
b
1
s
i
c
b
1
s
i
c
a
1
s
i
_
_
(5)
Q
i
= C
Q
Ar
3

i
|
i
|e
3
(6)
M
i
= T
i
D
i
(7)
where m and I are the mass and rotational inertia of the
yer, g is acceleration due to gravity, rho is the density of
air, r is the rotor radius, and A is the rotor disc area. In
equation 6, is multiplied by its magnitude to preserve
the sign of rotation for counter-rotating rotors.
Here sk(x) is the skew-symmetric matrix such that
sk(a)b = a b for vectors in
3
. The s
x
and c
x
no-
tations represent sinx and cos x respectively. The ro-
tation matrix R is constructed with the yaw-pitch-roll,
= (, , ) Euler angles. Rotors are indexed by their
corresponding compass directions: North, South, East
and West (NSEW), where N indicates the front rotor.
Correspondingly, D
i
is the rotor displacement from the
yer centre of mass:
D
N
=
_
0 d h
_
(8)
D
S
=
_
0 d h
_
(9)
D
E
=
_
d 0 h
_
(10)
D
W
=
_
d 0 h
_
(11)
where d is the arm length of the yer and h is the height
of the rotors above the CoG.
Vectors T
i
and Q
i
are the rotor thrust and torque,
and M
i
is the moment due to the thrust vector of the
ith rotor for a teetering rotor, the moment produced
by the rotor apping is due solely to the thrust vector
acting around a displacement from the vehicles centre
of gravity. The rst harmonic of the longitudinal and
lateral apping angles of the ith rotor are denoted by a
1s
i
and b
1s
i
. The non-dimensionalised thrust and torque
coecients, C
T
and C
Q
, are treated as constants here.
The speed of the ith rotor is given by
i
. The non-
dimensionalised thrust coecient and apping equations
are discussed in more detail in Sections 3.1 and 3.2.
3.1 Pitch and Roll Rotor Damping
A quad-rotor necessarily has a horizontal displacement
between its masts and CoG. When the craft rolls and
pitches, the rotors experience a vertical velocity, leading
to a change in the inow angle. From Prouty [Prouty,
2002, pp 101], C
T
can be related to the vertical velocity,
V
c
, by:
C
T
/ =
a()
4
_

tip

v
i
+V
c
r
_
(12)
where a is the polar lift slope, theta
tip
is the geometric
blade angle at the tip of the rotor, v
i
is the induced
velocity through the rotor, and is the solidity of the
disc - the ratio of the surface area of the blades and the
rotor disc area.
The polar lift slope is itself a function of the rotor
blade angle of attack, . It is highly nonlinear for some
airfoils and so the relation can be better expressed as a
variation around a set point, C
T0
:
C
Ti
= C
T0
+ C
Ti
(13)
where C
T
is the change induced by the changing inow
conditions. From Equation 12, this is written as:
C
Ti
=
a
0
4

i
r
(V + D
i
)e
3
(14)
where a
0
is the lift slope at the set point.
Figure 3: Blade Flapping Angle Rotation.
3.2 Blade Flapping
When the rotors translate horizontally there is a dier-
ence in blade lift between the advancing and retreating
blades, which causes the rotor tip path plane to tilt. The
resulting angle of the rotor plane is obtained by simulta-
neously solving the constant and sinusoidal components
of the blade centrifugal-aerodynamic-static weight mo-
ment system. Flapping is important, as previous sim-
ulations of the X-4 have shown that the tilting rotor
can introduce signicant stability eects for the vehicle
[Pounds et al, 2004].
The dynamics of rotor apping are very fast, occur-
ring within one revolution of the rotor [Leishman, 2006],
compared to the rigid body dynamics of the helicopter.
Consequently, the blade apping equations can be writ-
ten as instantaneous functions of the crafts planar ve-
locity.
A quad-rotors ight is not limited to longitudinal mo-
tion when the vehicle moves arbitrarily, the apping
motions of the rotors need not be in line with the nomi-
nal front of the aircraft. When the craft yaws the linear
velocity of the rotor hubs about e
3
is added to the motion
of the vehicle.
The apping of the ith rotor due to planar motion is
found by calculating the magnitude and direction of ro-
tors translation and dening a local frame of reference,
B
i
, aligned in that direction. We calculate the longitu-
dinal and lateral apping angles in the rotor frame (u
1
s
i
and v
1
s
i
) and then re-express them in the body-xed
frame (a
1
s
i
and b
1
s
i
using a rotation matrix (cf. Fig. 3).
This allows us to avoid computational complexity by us-
ing standard apping equations in the local frame.
The per-rotor apping is found by rst computing the
advance ratio and azimuthal direction of the rotor. We
derive this as:
V
ri
= V + D
i
(15)

ri
=
V
r(1,2)i

i
R
(16)

ri
= arctan
_
V
r(2)i
V
r(1)i
_
(17)
where V
r(n)i
is the nth element of the ith rotors velocity
vector,
ri
is the ith rotors advance ratio and
ri
is the
azimuthal direction of motion.
The conguration of the X-4 Flyer dispenses with the
sprung virtual hinge osets used previously. Thus, the
equations describing this motion can be greatly simpli-
ed: the longitudinal and lateral apping angle solutions
of the ith rotor in the local frame, B
i
, are:
u
1
s
i
=
1
1

2
ri
2

ri
(4
t
2
i
) (18)
v
1
s
i
=
1
1 +

2
ri
2
4
3
_
C
T

2
3

ri

a
+
ri
_
(19)
respectively, where
i
is the non-dimensionalised inow
of the ith rotor, approximated by

i
=
_
C
T
/2 (20)
and is the Lock Number [Leishman, 2006]:
=
a
0
cr
4
I
b
(21)
where I
b
is the rotational inertia of the blade about the
apping hinge.
These are transformed back into the body-xed frame
by the frame mapping between A and B
i
, J
i
to derive the
body-frame apping angles due to motion of the yer:
A
J
B
i
=
_
cos
ri
sin
ri
sin
ri
cos
ri
_
(22)
_
a
1
s
i
b
1
s
i
_
=
A
J
B
i
_
u
1
s
i
v
1
s
i
_
(23)
The components of the apping angles produced by
the crafts pitch and roll rates [Prouty, 2002] are added
to those of the body-xed frame:
a
1
s
i
= . . . +

16

_
q

_
+
_
p

_
1

2
i
2
(24)
b
1
s
i
= . . . +

16

_
p

_
+
_
p

_
1

2
i
2
(25)
Value Error Unit
a
0
5.5 0.5
c
tip
0.012 0.001 m
m 4.34 5 10
3
kg
A 0.0855 0.1 10
3
m
2
C
T
0.0047 0.2 10
3
C
Q
0.228 10
3
0.015 10
3
I
b
40.887 10
6
3.655 10
6
kgm
2
R 0.165 0.5 10
3
m
1.184 Not available kgm
3
1.417 0.133
0.049 2 10
3

tip
4.4 0.5 deg
0.054 1 10
3

hover
850 5 rads
1
Table 1: Aerodynamic Parameters and Associated Error.
4 Model Parameterisation and Stability
Designing a controller based on this model requires pa-
rameters of the physical system to be specied. Most
of these values are dictated by the ight performance of
the system; some, most importantly h, can be chosen
freely. The error associated with each parameter denes
the envelope of the plant models dynamic response. We
analyse the system behaviour within this envelope to de-
termine the best value of h, the height of the rotor plane
above the CoG.
4.1 Measured Values and Uncertainty
We have a set of parameter estimates, taken directly
from measurements or derived from experiments, along
with the associated error. In the case of parameters com-
puted from other known values, the associated error was
also computed:
Aerodynamic parameters
Rotor, blade and aerodynamic parameters are ob-
tained through measurement, computation, simula-
tion or from references. These are listed in Table 1.
Masses and Displacements
Component masses and distances measured with re-
spect to the rotor plane, (masses 0.005 kg, dis-
tances 0.005 m) are given in Table 2. Note that
this table is not a complete listing of all masses, but
includes all major masses screws and fasteners are
omitted (cf. Fig. 4).
Rotational Inertia Computed from the previous val-
ues by treating the masses as point masses, the di-
agonal entries of the inertial matrix are given in
Table 3. The CoG is 0.0071 0.005 m above the
rotor plane.
Figure 4: X-4 Component Osets.
Part mass/kg d/m e/m h/m
A Avionics 0.242 0 0 -0.02
B Rotor 0.046 0.315 0 0
C Motor 0.288 0.315 0 -0.06
D ESC 0.074 0.15 0.035 -0.055
E Powerbus 0.099 0 0 -0.13
F Batt
long
0.165 0.0125 0.06 0.035
G Batt
lat
0.165 0.0 0.04 0.035
H Arm 0.039 0.157 0.035 0.04
I Hoop 0.200 0 0 -0.17
Table 2: Component Masses and Osets.
4.2 Unforced Stability Analysis
The dominant dynamics of a helicopter, or a quad-rotor,
are associated with the longitudinal dynamics of the ve-
hicle. Around hover, the motion of a helicopter is largely
decoupled in each axis. The symmetry of quad-rotors
means that the important attitude dynamics can be de-
scribed by a single equation. We analyse the natural
stability of these dynamics to provide insight into the
best airframe geometry for controllability of the system.
In earlier work [Pounds et al, 2004], we applied
Proutys stability derivation to analyse the near-hover
dynamics of quad-rotors. This treatment furthers that
analysis through the addition of terms specic to quad-
rotors and the elimination of apping due to hub springs
that are not used in the current X-4 Flyer.
From the basic dynamic equations for a helicopter con-
strained to translate in x and rotate in pitch only without
Value Error Unit
I
XX
0.0820 0.0025 kgm
2
I
Y Y
0.0845 0.0029 kgm
2
I
ZZ
0.1377 0.0059 kgm
2
Table 3: Diagonal Inertial Elements.
control inputs, the stability derivative matrix is
2
:

ms +
X
x
X

s mg

x
I
Y Y
s
2
+

= 0 (26)
This uses the standard stability derivatives given in
Prouty [Prouty, 2002, pp 564] X is the longitudinal po-
sition, theta is the pitch angle and s is the Laplace trans-
form of the dierential operator. We modify the stan-
dard treatment of helicopters by multiplying M/ x,
and X/ x by 4 for the four rotors, as well as adding a
term in /

due to the vertical motion of the rotors
in pitch and roll:

= . . . A(R)
2
2d
C
T

(27)
where
C
T

=
a
8

1
R
(28)
The characteristic equation of the system matrix de-
terminant becomes:
s
3

_
1
m
X
x
+
1
I
Y Y

_
s
2
+
g
I
Y Y

x
= 0 (29)
Solving for the roots of this polynomial gives the ex-
ponential components of the dynamic behaviour of the
system. Since
_
1
m
X
x
+
1
I
Y Y

_
> 0 (30)
and
g
I
Y Y

x
> 0 (31)
for any system, it is clear that the unforced, open-loop
dynamics can never be stable for the X-4.
Application of Rouths Discriminant, as outlined in
Prouty, uses the characteristic polynomial to determine
the nature of the instability. The Rouths Discriminant,
R.D., is given by
R.D. = AD BC (32)
where A, B, C and D are the coecients of 29. If it is
positive, the craft will exhibit pure divergence. If nega-
tive, the craft will exhibit unstable oscillation. If zero,
the pitch dynamic will be neutral. In this case:
R.D. = C
T
A(R)
2
h (33)
Of the composing terms, only h can change signs. For
a conventional helicopter, where h < 0, the craft has
an unstable pole pair. If the rotors are inverted (above
the CoG), the craft will diverge without oscillation. If
the rotors and CoG are coplanar, the craft is marginally
stable. This behaviour was demonstrated in a full 3D
simulation previously [Pounds et al, 2004].
Value Error
p
1
2.507 +2.671i 0.714 +1.244i
p
2
2.507 2.671i 0.714 +1.244i
p
3
2.578 1.129
z -0.015 0.003
Table 4: Poles and Zeros of the Open Loop Pitch Dy-
namics.
4.3 Parameterised Model Envelope
Using the physical values for the yer, the coupled pitch
and
X
translational dynamical equations can be com-
puted. The error range of the parameters maps the
roots of the plant into a space on the complex plane.
Linearised dierential equations for the yer are:
m x = mga
1
s
mg (34)
I
Y Y

= 4dC
T
AR
2

0
+Ta
1
s
h
a
2
AR
0
d
2

(35)
These can be solved for a single transfer function
H = / between pitch angle, , and the input change
in rotor speeds, :
H =
4dC
T
Rc
2
(s+gc
1
)
(s+gc
1
)(I
Y Y
s
2
hmgc
3
s+
a
0
2
c
2
d
2
s)+hmg(gc
3
sg))
(36)
where
c
1
=
4
t
2
R
(37)
c
2
= AR
0
(38)
c
3
=
16

0
(39)
We approximate the apping angle as a linear function
of x and

:
a
1
s
= c
1
x +c
3

(40)
Using the previously given parameters and errors, the
poles and zeros of the system are given in Table 4. The
rotor height above the CoG is the largest contributor to
error, producing more than 80 per cent of the error of
each pole calculation. Thus, accurate knowledge of the
rotor height is important to determining the dynamic
model.
The unforced stability analysis demonstrated that h
is also important in determining the behaviour of the
dynamic system. The root locus for h shows that the
structure of the open-loop poles changes signicantly as
h changes sign (cf. Fig. 5). Analogous to the unforced
case, the system exhibits an unstable oscillation when
2. We use Proutys aircraft coordinate notation in stability and
control analysis for clarity
Figure 5: Root Locus of Pitch Dynamics for Changing
Rotor Height Above CoG.
the CoG is below the rotor, pure divergence when it is
above the rotor, and neutral stability when coincident
with the rotor.
Prouty suggests that helicopters can benet from an
inverted rotor conguration, as pure divergence is easier
for a human pilot to correct for than unstable oscillation
[Prouty, 2002, pp 603].
4.4 Design for Optimal Sensitivity
The use of automatic compensators no longer requires
that a system be intuitive for a human pilot, and so
oscillatory systems are acceptable. Instead, we use the
fundamental limits of control to congure the plant for
controller performance.
For good performance we want strong disturbance re-
jection and fast response to input commands. However,
the Waterbed Eect of the Bode integral for the sensi-
tivity function imposes a limit on arbitrary design tar-
gets for the controller across all frequencies: it states that
any arbitrary reduction in the sensitivity of the system
implies a corresponding increase in sensitivity over other
frequencies [Seron et al, 1997].
For this reason, it it desirable to reduce the Bode in-
tegral of the underlying system, prior to the application
of any control. The Bode integral can be related directly
to the poles of the open-loop plant. From Seron et al :
_

0
log |S(e
j
)|d =
n
p

i=1
p
i
(41)
where S is the sensitivity function of the system, p
i
are
the poles of the open loop plant, and is frequency.
Calculating the Bode integral for a range of h from
0.05 to 0.05 m below the rotor demonstrates a sharp
notch at h = 0 (cf. Fig. 6). When the rotor plane is
coincident with the center of gravity, the bode integral is
zero. In this conguration, the pitch dynamic is neutral.
Figure 6: Bode Integral With Respect to Rotor Plane
Placement.
The magnitude of the integral changes sharply as the
rotor plane moves away from the CoG. Given the strong
correlation between h error and plant model error, and
the link between control sensitivity and h position, it is
clear that close attention to the correct tuning and ver-
ication of rotor height is essential for the performance
of the helicopter.
For the X-4 Flyer, the ideal rotor position is at h = 0.
However, as the root locus with changing h demon-
strates, the structure of the plant undergoes signicant
change with error around this point. For this reason, we
set the CoG slightly away from the rotor plane so that
small errors will not have an impact on stability.
3
5 Control and Simulation
A variety of control techniques have been implemented
successfully on quad-rotor UAVS these include PID
and LQ [Bouabdallah et al, 2004] and PD
2
[Tayebi and
McGilvray, 2004] control. Bouabdalla found that PID
performed favourably compared to LQ due to the simpler
methods tolerance for model uncertainty. This quality
is desirable for our full apping model which is especially
sensitive to changes in h.
In addition to the attitude dynamics, the X-4 Flyer
also has important motor dynamics. The motor dy-
namics act in series with the rigid body dynamics
fast motor response is important for authoritive atti-
tude control of quad-rotors. To this end, rotor speed
controllers have been developed to improve the natural
performance of the rotor-motor system [Pounds et al,
2007]. The linearised closed-loop motor system transfer
function, H
M-CL
, is:
H
M-CL
=
68.85(s + 0.42)
(s + 78.46)(s + 0.44)
(42)
3. In practice, the X-4s rotors are set slightly above the CoG at
h = 0.07 m so it may be optionally piloted by a human.
Figure 7: Disturbance Propagation Block Diagram.
where u is the input reference speed.
5.1 Discretised Model
The controller runs at 50 Hz, the maximum frequency
at which attitude data is updated, and so the dynam-
ics of the plant are discretised at t
s
= 0.02 seconds for
the control design. The IMU returns both angle and
rate information, which allows for an improper PID con-
troller to be realised. The complete discretised model,
G
c
= /u, is:
G
c
=
1.434310
5
(z0.9916)(z+1)(z0.9997)
(z0.2082)(z0.9914)(z1.038)(z
2
1.943z+0.9448)
(43)
where u is the dierential variation in rotor speed about
the operating condition, 850 rads
1
. The additional
zero at z = 1 comes from the matched pole-zero dis-
cretisation method.
5.2 Controller Design
The proposed controller consists of a pure integrator for
zero angle tracking error and a complex zero pair to sta-
bilise the plant. The transfer function of the controller,
C, is:
C =
2210(z
2
+ 1.9z + 0.9045)
z 1
(44)
As the motor dynamics are so fast, the dominant pole
has little interaction with the attitude mechanics. If it
were slower, the excess poles would diverge closer to the
unit circle, leading to oscillation and possibly instability.
The slow motor pole-zero cancellation is associated with
the dynamics of the lithium ion polymer cells used to
power the yer. Sucient gain causes the pole to close
with the zero, reducing the inuence of the eect.
5.3 Disturbance Rejection
The disturbances experienced by the attitude dynam-
ics are expected to take the form of aerodynamic eects
propagated through variations in the the rotor speed.
We use the sensitivity model developed for the motor
speed controller to predict the displacement in position
due to a motor speed output disturbance (cf. Fig. 7). We
desire to keep the X-4 Flyer position variation small, in
the order of 0.5 m.
The rotor speed noise is modelled by an output dis-
turbance to the rotor speed, d, characterised as white
Figure 8: Pitch Angle Sensitivity Function Bode Plot.
noise, w, passed through a coloured lter, F [Pounds et
al, 2005] :
F =
0.0143(s + 7)
(s + 0.1)
(45)
The pitch angle sensitivity due to w is given by:

W
=
H
1 +C
M
H
M
C
M
H
M
C
M
H
M
+CH(1 +C
M
H
M
)
F (46)
where H
M
is the motor plant and C
M
is the motor
compensator. The peak sensitivity in pitch angle is at
0.4 rads
1
(cf. Fig. 8).
Pitch angle is integrated to x position. Using equa-
tions 35 and 40, the transfer function, X, is:
X =
gc
3
g
s(s +gc
1
)
(47)
A unit disturbance at the peak angle sensitivity fre-
quency yields a positional variation of 0.01 m, well
within the target. However, due to the integral posi-
tion dynamics, the peak sensitivity in x occurs at low
frequencies to DC (
d
< 0.01 rads
1
) at 6.3 dB; a
unit sinusoid in this range will produce a corresponding
position deviation of 0.78 m with negligible angle devi-
ation. Note that this deviation is very slow a period
of 600 s and would be easily compensated for, given
position measurement.
5.4 Simulation
A complete simulation of the attitude control system was
coded in Matlab Simulink. This included nonlinearities
in the system arising from multiple sample times in the
microcontrollers, saturation of the motors, quantisation
of measurements and slew limitation in the motor con-
troller.
In simulation, the closed loop system has a unit im-
pulse response settling-time of 2 seconds and 0.2 rad
maximum anglular displacement. A unit sinusoid distur-
bance applied at w = 0.01 rads
1
produced small angu-
lar displacements that were subsumed by the small non-
linear eects of the model and did not propagate into
Figure 9: X-4 Flyer Stabilised in Pitch and Roll.
the output. It is likely that error measurements due to
slow disturbance eects will be lost in the quantisation
of the sensor readings.
6 Implementation and Performance
Prior to the designed controller being tested under ight
conditions, we tested the controller on a tether appa-
ratus. In this conguration, no apping occurs due to
horizontal motion since the yer is xed in space, free
only to rotate in pitch and roll. In practice, it was found
that the tethered X-4 exhibits two additional stable os-
cillatory poles at z = 0.9664 0.0331, from mechanical
cross-coupling with the test rig. The rotors can be oper-
ated at reduced speeds to conserve battery power during
initial testing at these speeds, the system gain changes
in proportion to the rotor speed. The resultant simpli-
ed full-speed system transfer function becomes:
G
rig
=
1.87 10
8
(z + 1)
3
(z 1.0)(z 0.953)(z
2
1.933z + 0.935)
(48)
This requires modication to the controller to place
the zeros on the real axis:
C
noap
=
2210(z
2
+ 1.9z + 0.9)
z 1
(49)
When implemented, it was found that the controller
worked reliably for low rotor speeds ( < 450 rads
1
).
The X-4 can stabilise itself in pitch and roll, and remain
within 2 degrees of level (cf. Fig. 9). To test the dy-
namic performance, twenty-two step experiments were
performed over 800 seconds, from which steps were av-
eraged for analysis (cf. Fig. 10). The step motion was al-
ternating 10 degrees pitch forwards and backwards from
level, to eliminate directional bias. Roll was held at
Figure 10: Low Speed Average Step Reference (black),
Data (blue) and Prediction (green).
0 degrees by the controller and yaw was locked in place
on the test rig.
From the data, the system has a 1.25 second rise-time,
30 per cent overshoot and a slow 40 second settling time,
compared with the 2.15 second rise-time, 30 per cent
overshoot and 15 second settling time the model predicts
at this rotor speed. The step clearly shows the inuence
of the two test-rig poles producing a 0.4 Hz oscillation
with 1 degree angular variation. This oscillation would
lead to horizontal displacement of 0.027 m, were the
X-4 in ight.
It was found that as the rotor speed increases, the sys-
tem displays chaotic semi-stable behaviour that would
make untethered ight impossible. We believe that the
instability is due to high-frequency noise from the rotors
destroying the validity of IMU accelerometer data. We
are condent that additional isolation of the sensors will
allow full-speed operation.
7 Conclusion
We have developed a larger quad-rotor platform than is
typically used in current robotics research. The anal-
ysis of yer attitude dynamics allowed us to tune the
mechanical design for best control sensitivity and dis-
turbance rejection. We designed a controller to stabilise
the dominant decoupled pitch and roll modes, and use
a model of disturbance inputs to estimate the perfor-
mance of the plant. It was found that the compensator
sucessfully regulates attitude at low rotor speeds.
8 Acknowledgements
The authors would like to thank CSIRO ICT Robotics
for its ongoing support of this project.
References
[Bouabdallah et al, 2004] S. Bouabdallah, A. Noth and
R. Siegwart. PID vs LQ Control Techniques Applied
to an Indoor Micro Quadrotor. In Proceedings of the
IEEE International Conference on Intelligent Robots
and Systems, Sendai, Japan, 2004.
[Guenard et al, 2005] N. Guenard, T. Hamel and V.
Moreau. Dynamic Modeling and Intuitive Control
Strategy for an X4-Flyer. In proceedings of 5th In-
ternational Conference on Control and Automation,
Budapest Hungary, June, 2005.
[Hamel et al, 2002] T. Hamel, R. Mahony, R. Lozano
and J. Ostrowski. Dynamic Modelling and Cong-
uration Stabilization for an X4-Flyer. In proceed-
ings of 15th Triennial World Congress of the Inter-
national Federation of Automatic Control, Barcelona,
July, 2002.
[Leishman, 2006] J. G. Leishman. Principles of Heli-
copter Aerodynamics, 2nd Ed. Cambridge University
Press, Cambridge, United Kingdom, 2006.
[Pounds et al, 2004] P. Pounds, R. Mahony, J. Gresham,
P. Corke and J. Roberts. Towards Dynamically-
Favourable Quad-Rotor Aerial Robots. In Proc. of
Australasian Conference on Robotics and Automation,
Canberra, Australia, 2004.
[Pounds et al, 2005] P. Pounds, R. Mahony and P.
Corke. Small-Scale Aeroelastic Rotor Simulation, De-
sign and Fabrication. In Proc. of Australasian Confer-
ence on Robotics and Automation, Sydney, Australia,
2005.
[Pounds et al, 2007] P. Pounds, R. Mahony and P.
Corke. System Identication and Control of an Aer-
obot Drive System. In Proc. of Information, Decision
and Control, Adelaide, Australia, 2007, submitted.
[Prouty, 2002] R. W. Prouty. Helicopter Performance,
Stability, and Control. Krieger Publishing Company,
2002, reprint with additions, original edition 1986.
[Seron et al, 1997] M. M. Seron, J. H. Braslavsky and
G. C. Goodwin. Fundemental Limitations in Filtering
and Control. Springer-Verlag, London, United King-
dom, 1997.
[Tayebi and McGilvray, 2004]
A. Tayebi and S. McGilvray. Attitude Stabilization
of a Four-Rotor Aerial Robot. In proceedings of 43rd
IEEE Conference on Decision and Control, Atlantis,
Paradise Island, Bahamas, pp14-17, December, 2004.