c
Distributed Coordination and Control
Experiments on a Multi-UAV Testbed
by
Ellis T. King
Bachelor of Engineering
The State University of Bu?alo, 2002
Submitted to the Department of Aeronautics and Astronautics
in partial ful?llment of the requirements for the degree of
Master of Science in Aeronautics and Astronautics
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2004
? Massachusetts Institute of Technology 2004.
Author . .... .... .. .. .... .... .... .... .... .... .... .... . ... ... . .... .... .
Department of Aeronautics and Astronautics
August 20, 2004
Certi?ed by. ... .... . ... .... .... .... .... .... .... .... ..... .... .. .. .... .
Jonathan P. How
Associate Professor
Thesis Supervisor
Accepted by . .... .... .... .... .... .... .... ... . .... .... .... .... .... ....
Jaime Peraire
Professor of Aeronautics and Astronautics
Chair, Committee on Graduate Students
Chapter 2
Hardware In the Loop Modeling
and Simulation
The hardware-in-the-lo op (HWIL) simulation is only useful if it accurately portrays
the vehicle dynamics and if the behaviors observed during ?ight tests can be repli-
cated on the ground. This chapter focuses on identifying some of the dynamical
modes of ?ight for the 60 ARF Trainer aircraft, and verifying that the HWIL sim-
ulations re?ect the dynamics expected from the aircraft being employed. Reduced
order models for 4 of the 5 dynamical modes are determined for the trainer ARF
60 aircraft using identi?cation techniques on experimental ?ight data and analytical
predictions based on aircraft geometry and aerodynamic data. Section 2.1 describes
the simulation settings used to create the hardware-in-the-lo op (HWIL) simulations,
and Section 2.2 details the procedures used to create models of the aircraft dynam-
ics from data collected during ?ight tests and hardware-in-the-lo op simulations. In
Section 2.3, the Cloud Cap autopilot is tuned for the trainer ARF 60 aircraft and the
closed loop response for several of the modes is measured using the HWIL simulator.
2.1 Hardware in the loop simulations
2.1.1 Aircraft simulation Model Parameters
Aerodynamic, inertial and engine calibration information is provided to the Cloud
Cap HWIL simulation application to model the aircraft being ?own. For simply
31
(a) The Clark YH airfoil geometry.
?5 0 5 10 15 20 25 30
?0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
α [deg]
C
L
C
D
(b) The Clark YH airfoil lift and drag curves.
Figure 2-1: The Clark YH airfoil closely resembles the airfoil used on the
trainer ARF 60 aircraft and is used to model wing aerodynamics.
con?gured aircraft such as the tower trainer 60 ARF used in the testbed, many of the
performance characteristics can be obtained using the geometry of the aircraft, such
as the data found in Table 2.1. Detailed descriptions of the surface geometry, wing
lift curves, and engine performance curves enable simulations of the aircraft under
realistic ?ight conditions, providing the input parameters are con?gured accurately.
For example, the Clark YH airfoil closely resembles the trainer ARF 60 airfoil and
is used to describe the aerodynamic properties of the main wing on the aircraft [25].
Some of the data is shown in Figure 2-1. A more detailed description of the simulator
input ?les is given in Ref. [26].
32
Table 2.1: Trainer 60 ARF measurements, experimentally determined iner-
tias as shown in Subsection 2.1.1. Symbolic notation is borrowed
from Ref. [27]
Measurement Value Units Symbol
(SI)
Wing Span
Wing Area
Chord Length
Wing Incidence
Wing Dihedral
Wing Sweep
Tail Area
Tail Span
Tail O?set X
Tail Sweep
Fin Area
Fin Span
Fin O?set X
Fin O?set Z
Fin Sweep
Fin Volume Ratio
Fuselage CX Area
Fuselage Length
Gross Mass
Empty Mass
Roll Inertia*
Pitch Inertia*
Yaw Inertia*
1.707
0.5200
0.305
1
5
0.0
0.0879
0.606
1.14
9
0.0324
0.216
1.143
0.120
53
0.0189
0.130
1.270
5.267
4.798
0.31
0.46
0.63
m
2
m
m
deg
deg
deg
2
m
m
m
deg
2
m
m
m
m
deg
-
2
m
m
kg
kg
kg m
2
·
kg m
2
·
kg m
2
·
b
S
cˉ
Γ
Λ
Λ
l
b
S
t
t
t
t
S
f
b
f
l
f
Λ
h
f
f
ˉ
V
f
I
I
I
m
l
S
b
b
m
e
xx
yy
zz
Aircraft Inertia Experiment
The aircraft pitch, roll, and yaw inertias are important parameters for the accurate
HWIL simulation of the aircraft dynamics. Fortunately, due to the small scale of the
aircraft, experimental measurements can be easily made for each axis of the aircraft.
The experimental setup for the roll axis is shown in Figure 2-2. From the aircraft free
body diagram, the tension in each cable, T ,is
2T = mg (2.1)
33
Figure 2-2: Torsional pendulum experimental setup to determine roll axis
inertia, I
xx
, for the trainer aircraft. The period of oscillation
of a roll angle perturbation, φ, is measured to parameterize the
aircraft inertia. The angle α is the small angle deviation of the
supporting cables from the vertical position. This experiment
was also repeated for the pitch and yaw axes to determine I
yy
and I
zz
respectively.
For rotational perturbations applied to the airframe, the product of interior angles
and distances must be constant
Rφ = Lα (2.2)
where φ is the aircraft roll angle perturbation and α is the small angle deviation of
the supporting cables from the vertical position. The di?erential equation describing
the motion of the torsional pendulum is governed by a torsional inertia term and the
restoring moment due to tension forces
¨
I
xx
φ +2TR sin α = 0 (2.3)
34
?
Table 2.2: Experimentally determined aircraft inertias [kg-m
2
]
Aircraft Roll Axis Pitch Axis Yaw Axis
No. I
xx
I
yy
I
zz
1 0.28
2 0.30
3 0.33
Mean 0.30
Std Dev. 0.029
0.46
0.44
0.47
0.42
0.011
0.65
0.61
0.63
0.63
0.021
Using the small angle approximation for α since L ? R, and substituting known
values from Eqs. 2.1 and 2.2, Eq. 2.3 reduces to
¨
mgR
2
I
xx
φ + φ = 0 (2.4)
L
which is characterized by the undamped natural frequency, ω, and period of oscilla-
tion, T
p
ω =
mgR
2
I
xx
L
(2.5)
2π
T
p
=
ω
(2.6)
By ?nding averaged values for the period of oscillation, T
p
, in each of the pitch, roll,
and yaw axes, the inertia about each axis can be approximated. This experiment
neglects aerodynamic and other forms of damping as well as the cross-axis inertias
(e.g., I
xz
, I
yz
). The experimental results are summarized for each of the axes and three
di?erent aircraft in the same con?guration in Table 2.2, showing agreement between
di?erent vehicles used in the tests. The largest variation was found in the roll axis
due to the di?culty of mounting the aircraft through the center of gravity, which is
essential in this experimental setup.
2.1.2 Actuator Models
The servos used on the aircraft have saturation limits, limited bandwidth, and limited
slew rates which are captured in the actuation models of the Cloud Cap hardware-
35
Figure 2-3: Actuator models used the Cloud Cap Hardware-in-the-lo op sim-
ulations
in-the-loop simulator. As shown in Ref. [26], the actuator transfer function, G
act
(s),
is given by specifying the bandwidth limit, B
W
G
act
(s)=
ω
2
n
s
2
+2ζω
n
s + ω
2
n
(2.7)
ω
n
=2πB
W
(2.8)
ζ = ζ
c
=0.707 (2.9)
where the damping ratio, ζ, is selected at the critical value to set the actuator band-
width equal to the natural frequency (ω
b
= ω
n
) . The aileron, elevator, and rudder
channels all respond with approximately the same characteristics (B
W
= 10 Hz), but
the throttle is modeled with less dynamic range (B
W
= 2 Hz) as the engine RPM
requires added time to ramp up to produce thrust. The input/output saturation
and slew rate limits are determined as per manufacturer speci?cations (±60
?
,2 Hz
respectively), and applied as shown in Figure 2-3.
2.1.3 Sensor Noises
The Cloud Cap hardware-in-the-lo op simulator includes detailed sensor models based
on information from the manufacturer to corrupt the simulation measurements. For
the purposes of simulation, noises on the pressure, rate gyros and accelerometers
onboard the aircraft are modeled using band-limited white noise and speci?ed drift
rates [26]. Although the same noise and drift models could be applied to GPS position
and velocity measurements, this information is typically assumed to be perfect in the
HWIL tests. The values used to parameterize the Piccolo
TM
pressure sensors, the
Crista
TM
IMU angle-rate sensors, and the accelerometers are shown in Table 2.3.
36
Table 2.3: Crista IMU HWIL Sensor Noise Models
Sensor PDynamic PStatic Gyro Accel.
[unit] [Pa] [Pa] [deg/s] [m/s/s]
Resolution [unit]
Min [unit]
Max [unit]
Noise Gain
Butterworth Order
BW Cuto? Freq. [Hz]
Drift Rate [unit/s]
Max Drift value [unit]
Drift Update Rate [s]
3.906
-300
4000
20.0
2
11.0
0.05
15.0
5.0
2.00
0.0
110,000
20.0
2
11.0
1.0
100.0
5.0
1.6E-4 6.0E-3
-5.20 -100.0
5.20 100.0
0.10 0.0
2 2
20.0 20.0
1.5E-4 2.0E-3
0.01 0.20
1.0 1.0
2.1.4 Dryden Turbulence
Stochastic turbulence disturbances are required for accurate HWIL simulations, as
real world experiments are characterized by unpredictable winds acting on the vehicle.
The Dryden turbulence model is one of the accepted methods for including turbulence
in aircraft simulations [28]. By applying shaped noise with known spectral properties
as velocity and angle rate perturbations to the body axes of the vehicle, the e?ect
of turbulence is captured during discrete time simulations. The noise spectrum for
each of the perturbations is predominantly described by a turbulence scale length
parameter, L, the airspeed reference velocity, V
o
, and the turbulence intensity, σ.
The selection of these parameters allows for the turbulence to be modeled according
to the prevailing wind conditions.
The spectral frequency content for generalized aircraft turbulence have been well
studied [29, 28] and are given for each of the aircraft body axes:
S
ug
(ω)=
2σ
2
u
L
u
πV
o
·
1
1+ (
L
u
V
o
ω)
2
(2.10)
S
vg
(ω)=
σ
2
v
L
v
πV
o
·
1+ 3(
L
v
V
o
ω)
2
?
1+ (
L
v
V
o
ω)
2
?
2
(2.11)
S
wg
(ω)=
σ
2
w
L
w
πV
o
·
1+ 3(
L
w
V
o
ω)
2
?
1+ (
L
w
V
o
ω)
2
?
2
(2.12)
37
S
σ
2
0.8
?
πL
w
?
1/3
pg
(ω)=
w 4b
?
2
(2.13)
L
w
V
o
·
1+
?
4b
ω
πV
o
??
ω
2
V
o
S
qg
(ω)=
?
2
· S
wg
(ω) (2.14)
1+
?
4b
ω
πV
o
?
ω
?
2
V
o
S
rg
(ω)=
?
2
· S
vg
(ω) (2.15)
1+
?
3b
ω
πV
o
where ω is the spectral frequency of the turbulence and the aircraft wingspan, b,is
used as a parameter in the angle rate ?lters to scale the e?ect of rotation on the
main lifting surface. The subscripts u, v, w and p, q, r refer to the familiar body frame
aircraft wind velocities and angle rates, respectively, thereby allowing independent
classi?cation of the turbulence in each axis. These spectral shaping functions are
used to form shaping ?lters to give the body axis noise transfer functions [30]
?
L
u
1
H
ug
(s)= σ
u
2
1+
L
u
(2.16)
πV
o
·
s
1+
√
3
L
v
V
o
s
?
L
v
V
o
??
1+
L
v
2
(2.17)H
vg
(s)= σ
v
πV
o
·
s
V
o
s
?
L
w
1+
√
3
L
w
V
o
??
1+
L
w
2
(2.18)H
wg
(s)= σ
w
πV
o
·
s
V
o
?
1/6
?
0.8
?
π
4b
(2.19)
1/3
H
pg
(s)= σ
w
V
o
L
w
?
1+
?
4b
s
??
πV
o
s
H
wg
(s) (2.20)H
qg
(s)=
?
V
o
4b
?
·
1+ s
πV
o
s
H
vg
(s) (2.21)H
rg
(s)= ?
V
o
3b
? ·
1+ s
πV
o
The block diagram for the full 6 DOF Dryden turbulence model is shown in Figure 2-4.
Note the cross axis couplings of the angle rate ?lters q
g
and r
g
.
Example turbulence perturbations values are plotted in Figure 2-5 as a function of
the scale lengths and intensities for each of the body axes. The same 4×1 white noise
input was used for each trial set. Larger scale lengths, L, increase the time constant of
38
Figure 2-4: Block diagram for the 6 DOF Dryden Turbulence model. The
velocity perturbations u
g
,v
g
,w
g
are independent outputs of the
?ltered values of the turbulence scale lengths, L, intensity values,
σ and the white noise input sources. The principle axis coupling
of the aircraft is taken into account through the inputs to the
angle rate perturbation ?lters.
the turbulence seen for a given airspeed, while larger σ values increase the deviation
about zero. L and σ typically vary with altitude in the lower atmosphere as ground
e?ects become more prominent, but for HWIL simulations they are usually ?xed.
The frequency response of the Dryden ?lters are shown for the same three cases in
Figure 2-6. The ?lter cuto? frequency is determined by the ratio of the scale length to
airspeed, and this e?ectively produces lower bandwidth ?lters for larger scale lengths.
The scale length parameter is chosen according to one of several speci?cations, all
of which take into account the variation of L with altitude. The military reference
MIL-F-8785C provides one such model of the scale length at low altitudes, h, which
39
0.5
0
?0.5
Vu
Vv
0 1 2 3 4 5 6 7 8 9
1
0
?1
0 1 2 3 4 5 6 7 8 9
?1
?0.5
0
0.5
Vu
Set 1
Set 2
Set 3
Qr
Qq
Qp
0 1 2 3 4 5 6 7 8 9
Time [s]
(a) Velocity perturbations
0.4
0.2
0
?0.2
?0.4
0 1 2 3 4 5 6 7 8 9
0.2
0
?0.2
0 1 2 3 4 5 6 7 8 9
0.1
0
?0.1
?0.2
0 1 2 3 4 5 6 7 8 9
Time [s]
(b) Angle rate perturbations
Figure 2-5: The output of the Dryden velocity and angle rate ?lters for
di?erent selections of the intensity and scale lengths. Set 1:
L = 150,σ =0.5, Set 2: L = 1500, σ =0.5, Set 3: L = 150,
σ =1.5
40
is valid up to 1000 feet [29].
L
w
= h (2.22)
h
L
u
= L
v
= (2.23)
(0.177 + 0.000823h)
1.2
The turbulence intensity is a gain factor that scales the magnitude plots in Fig-
ure 2-6 to values appropriate for di?erent wind levels (i.e., light, moderate, severe).
The intensitylevel has been de?ned for low altitude ?ight according to MIL-F-8785C
as
σ
σ
w
=0.1W
20
(2.24)
w
σ
u
= σ
v
= (2.25)
(0.177 + 0.000823h)
0.4
where W
20
is the wind speed as measured at 20 ft in altitude. According to MIL-
F-8785C, W
20
< 15 knots is classi?ed as “light” turbulence, W
20
30 knots is≈
“moderate”, and W
20
> 45 knots is “heavy”. Other military speci?cations such as
MIL-HDBK-1797 exist for the low altitude cases [29], and di?erent types of models
are used for other regions of the atmosphere. For the purposes of the UAV application,
the low altitude models are su?cient.
The utility of the Dryden turbulence model is that it allows the expected turbu-
lence levels to be described for an aircraft ?ying at a given reference speed for more
realistic HWIL simulations. Turbulence is applied to the vehicle body axes consistent
with the known parameterized values for scale length and intensity, which e?ectively
de?nes the appropriate ?lters with cuto? frequencies and magnitudes needed for sim-
ulation. Note that in addition to turbulence, wind is also usually modeled with a
static component, W , that represents a prevailing magnitude and direction in an
inertial axis. Together these de?ne an arbitrary three-axis wind vector
W = W + δW
I
(2.26)
where δW
I
is the e?ective Dryden wind turbulence in each axis after being rotated
through the appropriate body to inertial transformation direction cosine matrix.
For small scale aircraft, the static wind component is usually a gross disturbance
relative to the aircraft airspeed, and it can have a large e?ect on high level planning
41
Bode Diagram
20
10
0
?10
?20
Phase (deg)
Magnitude (dB)
Phase (deg)
Magnitude (dB)
?30
?40
?50
?60
0
?45
?90
Set 1
Set 2
Set 3
?4 ?3 ?2 ?1 0 1
10 10 10 10 10 10
Frequency (rad/sec)
(a) Velocity ?lter for w
g
Bode Diagram
0
?20
?40
?60
?80
?100
?120
?140
?160
?180
?200
0
?45
?90
?135
?180
?3 ?2 ?1 0 1 2 3
10 10 10 10 10 10 10
Frequency (rad/sec)
(b) Angle rate ?lter for q
g
Figure 2-6: The Bode plots of the Dryden velocity and angle rate ?lters,
given white noise inputs to H
vg
(s) in (a) and H
qg
(s) H
wg
(s)·
in (b). Various selections of the intensity and scale lengths are
shown in di?erent sets. Set 1: L = 150,σ =0.5, Set 2: L = 1500,
σ =0.5, Set 3: L = 150, σ =1.5
42
algorithms. The e?ect of this type of disturbance on the planning system and aircraft
dynamics is discussed in more detail in Chapters 4 and 5.
2.2 Open Loop Aircraft Modeling
2.2.1 Longitudinal Dynamics
A common model for the longitudinal motion of the aircraft is [27, 31]
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
u˙ x x x x u x
θ δ
e
u w q
w˙?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
z z z z w z
θ δ
e
u w q
δ
e
(2.27)+
=
q˙
θ
˙
m m m m q
θ
m
θ δ
e
u w q
?
?
?
?? ?
?
?
0 0 1 0 0
x˙ = Ax + Bu (2.28)
where the state variables x =[uwqθ]
T
refer to the longitudinal velocities, u and
w, the pitch rate, q and the angle of inclination, θ. The elements of the A matrix
in Eq. 2.27 represent the concise form aerodynamic stability derivatives referring to
the airplane body axis. Tables of values relating the concise form derivatives to the
dimensionless or dimensional derivatives are available in numerous sources [27, 32].
The control input u = δ
e
is the elevator defection angle with the engine thrust ?xed
and is input to the dynamics through the aerodynamic control derivative matrix B.
The Longitudinal Dynamics in Eq. 2.27 are typically resolved into two distinct
phugoid and short period modes, which represent dynamics of the aircraft on di?erent
timescales. The short period is characterized by high frequency pitch rate oscillations
and can have high or low damping, depending on the dynamic stability of the aircraft.
In contrast, the phugoid mode is characterized by lightly damped, low frequency
oscillations in altitude and airspeed with pitch angle rates, q, remaining small.
Short Period Mode
A simple approximation for the short period mode of the aircraft can be obtained by
assuming the speed of the aircraft is constant over the timescale of the short period
dynamics (˙u = 0), that the aircraft is initially in steady level ?ight and that the
43
derivatives refer to a wind-axis system (θ = α = 0). The equations of motion then
reduce to
?
w˙
? ?
z
w
z
q
??
w
? ?
z
δ
e
?
?
δ
e
(2.29)
? ?
=
? ?? ?
+
?
q˙ m
w
m
q
q m
δ
e
Following further approximations shown in [27] which make assumptions about the
relative size of the m
q
,z
q
and z
w
derivatives, the transfer functions for the two short
term equations describing the response to elevator are:
w(s)
z
n
?
s+ V
o
m
δ
e
?
?
k
w
(s+1/T
α
)
(2.30)
z
δ
e
=
δ
e
(s) (s
2
?(m
q
+ z
w
)s+(m
q
z
w
?m
w
V
o
)) s
2
+2ζ
s
ω
s
s+ ω
s
2
q(s) m
n
(s?z
w
)
?
k
q
(s+1/T
θ
)
(2.31)=
δ
e
(s) (s
2
?(m
q
+ z
w
)s+(m
q
z
w
?m
w
V
o
)) s
2
+2ζ
s
ω
s
s+ ω
s
2
where k
q
,k
w
,T
θ
,T
α
,ζ
s
, and ω
s
represent approximate values for the short period mode
and V
o
is the vehicle reference speed.
One of the most accurate ways to obtain models for the aircraft data is to use
actual ?ight data. Identi?cation algorithms such as those in the Matlab System Iden-
ti?cation Toolbox [33] can be used to used to obtain open loop models of the system
dynamics from ?ight data collected during experiments. These models can then be
used to validate the HWIL simulation environment as well as to help determine the
gain settings for the autopilot control loops as shown in Section 2.3. Input-output
data was collected by disengaging all of the autopilot loops and performing a series
of maneuvers to measure the aircraft response to de?ections from the elevator and
aileron control surfaces. Example data from two experiments are shown in Figures
2-7(a)-(b) depicting the longitudinal and lateral modes, respectively.
To capture the longitudinal dynamics, the bank angle was held ?xed at zero
degrees, while a series of pitch oscillations were commanded using the elevator. Figure
2-7(a) shows a sample of data that was collected on one run of the pitch test. From
the plot it is clear that the longitudinal modes are being excited due to input from
the elevator, while the lateral motions in the roll and yaw axes are essentially ?xed.
Sample data from a roll excitation run is plotted in Figure 2-7(b). This plot also
clearly shows coupling in the yaw axis due to the dihedral angle of the wing.
44
(a) Pitch test data induced with elevator de?ections.
100
1055 1060 1065 1070 1075 1080 1085
?100
?50
0
50
Roll Axis
[deg/s]
Aile
Rudd
Elev
Yaw Axis
[deg/s]
40
20
0
?20
?40
1055 1060 1065 1070 1075 1080 1085
0
0
100
?100
Pitch Axis
[deg/s]
?100
1055 1060 1065 1070 1075 1080 1085
time [s]
306 308 310 312 314 316 318
?100
?50
0
50
100
Roll Axis
[deg/s]
?100
?50
0
50
Yaw Axis
[deg/s]
Aile
Rudd
Elev
306 308 310 312 314 316 318
?20
?10
0
10
20
Pitch Axis
[deg/s]
306 308 310 312 314 316 318
Time [s]
(b) Roll test data induced with aileron de?ections. Note the yaw axis coupling due
to the large dihedral angle of the wing
Figure 2-7: Sample ?ight test data used in the estimation algorithms. Dash-
dot lines represent rate gyro data output for each axis. Control
surface inputs are plotted for each corresponding axis in degrees
of de?ection
45
Parametric models for the input-output transfer functions can be formed using
experimentally collected data and used to determine the unknown coe?cients in
Eq. 2.31. Then from the characteristic equation, the longitudinal short period dy-
namics can be inferred from models of the transfer function from elevator angle to
pitch rate. Also note that qualitative predictions about the values of the parameters
in Eq. 2.31 are available, since they depend on the concise stability derivatives which
all have physical signi?cance. Once models are obtained the approximate parameters
can then be veri?ed against these qualitative predictions.
Figure 2-8 shows the output of a parametric subspace model based on experimental
data such as that shown in Figure 2-7(a). The model output (solid line) tracks the
actual measurements of pitch rate (dashed line) quite well and was validated on data
sets from di?erent test days and aircraft. Model residuals within the 99% con?dence
bounds for the auto- and cross correlation functions are plotted in Figure 2-9 and
indicate that the 3
rd
order model is su?cient to describe the input-output dynamics.
This model is represented by the continuous transfer function in Eq. 2.32 and has
zero-pole pairs as indicated in Figure 2-10(a). The short period is well represented by
the high frequency, oscillatory mode of the system, while there is one low frequency
pole located near the imaginary axis. The transfer function is:
q(s) 9.539s
2
?1440s +60.52
T
qδ
e
(s)= =
δ
e
(s) s
3
+21.77s
2
+ 325.8s +29.94
9.539(s ?150.9)(s ?0.0420)
(2.32)=
(s
2
+21.7s + 323.7)(s +0.0925)
ζ
The third order model T
qδ
e
(s) was selected because it provided the best ?ts to
a large number of data sets and residuals that remained below the 99% con?dence
intervals in Figure 2-9. Approximate values for the short period dynamics can be
obtained by resolving the oscillatory mode in Eq. 2.32 to determine the corresponding
values for ω
s
and ζ
s
. A step response for this mode is plotted in Figure 2-10(b),
indicating a reasonable short period response time with settling time 0.4 seconds, and
s
=0.6. Second order models produced from the same data set were found to have
di?culty reproducing the outputs of the experiment, and exceeded the con?dence
bounds as shown in Figure 2-9. Higher order models (4
th
and higher) tended to
46
place additional pole-zero pairs near the origin without achieving better tracking or
residual bound performance, therefore indicating the 3
rd
order model as the best
representation for the system.
The zero locations in Eq. 2.32 are not consistent with the expected values for
the stability derivatives in Eq. 2.31, which is an indication of the delay acting on
this input-output channel and consistent with the servo response time in an actual
physical system. In addition, it should be noted that the zero location for these types
of models is generally more uncertain than the pole location. As a result, the zeros
in Eq. 2.31, are di?cult to identify without more sophisticated validation techniques.
In order to validate the HWIL model the same set of tests was performed in
simulation and the results are compared in Figure 2-11. As shown in Figure 2-11(a),
both models reproduce the experimental outputs with excellent tracking. Since the
data set used to validate these two models is di?erent from either of the sets used
to generate them, this model can be relied upon with much higher con?dence. The
transient responses are shown in Figure 2-11(b) and the fast acting short period mode
is shown to agree well, however there is variation on the longer timescales. The HWIL
model for pitch response is given by
q(s) 2.232s
2
?1265s +2.449
T
qδ
e
(s)= =
δ
e
(s) s
3
+19.2s
2
+ 283.7s ?2.3
0.0022(s ?566.92)(s ?0.0019)
(2.33)=
(s
2
+19.209s + 283.8924)(s +0.0081)
which has a short period characteristic equation similar to Eq. 2.32, identi?ed using
experimental ?ight test data. The discrete P-Z plot in Figure 2-12 shows the pole and
zero locations for the longitudinal models. The short period mode is shown to agree
well, however the slower frequency dynamics are not as well modeled.
Phugoid Mode
The data from the preceding experiments captures the short period mode of the
system well, however the phugoid mode is not well represented. This is physically
consistent because the phugoid mode excites the airspeed and pitch response over
longer timescales, speci?cally when q˙ =˙w ≈ 0. Setting the corresponding terms to
47
?1
?0.5
0
0.5
1
[rad/s]
Sim Output
Exp Data
236 238 240 242 244 246 248 250 252 254 256
Time
Figure 2-8: Simulated and measured outputs from elevator to pitch rate. Val-
idation on a di?erent data set than that used to create the model.
Autocorrelation of residuals for output q
0.1
0.05
0
?0.05
?0.1
?20 ?15 ?10 ?5 0 5 10 15 20
Cross correlation for input δ and output q residuals
e
?0.2
?0.1
0
0.1
2
nd
Order
3
rd
Order
?15 ?10 ?5 0 5 10 15 20
Samples
Figure 2-9: Auto- and cross-correlations with 99% con?dence intervals
(dashed lines) for the elevator to pitch rate models obtained from
experimental data. 2
nd
order model has high cross correlations
for positive sample delays, indicating unmodeled higher order
dynamics, but 3
rd
order models stay within the 99% bounds.
48
10
15
System: untitled1
Pole : ?10.8 + 14.4i
Damping: 0.602
Overshoot (%): 9.34
Frequency (rad/sec): 18
System: untitled1
Zero : 0.042
5
Damping: ?1
Overshoot (%): 0
Frequency (rad/sec): 0.042
0
System: untitled1
Pole : ?0.0925
System: untitled1
Zero : 151
?5
Damping: 1
Overshoot (%): 0
Frequency (rad/sec): 0.0925
Damping: ?1
Overshoot (%): 0
Frequency (rad/sec): 151
?10
?15
?20 0 20 40 60 80 100 120 140 160
(a) The open loop pole and zero locations indicate the short period dynamics
identi?ed by the estimation algorithms.
Step Response
0 0.1 0.2 0.3 0.4 0.5 0.6
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Amplitude
Time (sec)
(b) Short period approximation step input
Figure 2-10: Estimated longitudinal model, T
qδ
e
(s), from elevator input to
pitch rate.
49
Measured and simulated model output ?? Pitch Axis
?1
?0.5
0
0.5
1
HWIL Sim Model
Flight Data
Exp Flight Model
Pitch Axis [rad/s]
?1.5
188 190 192 194 196 198
Time [s]
(a) Both models created from the actual ?ight data and the HWIL simulator
are able to track the experimental data well.
?5
?4.5
?4
?3.5
?3
?2.5
?2
?1.5
?1
?0.5
0
Pitch Rate Response [rad/s]
HWIL Model
Flight Model
0 0.5 1 1.5 2
Time [s]
(b) Transient response for the longitudinal models.
Figure 2-11: Transient response for the 3
rd
order models are agree in time
constant and damping, showing agreement in the short period
dynamics of the HWIL and ?ight test models.
50
0.6
0.5
0.4
0.3
0.2
0.1
0
?0.1
0.3π/T
0.2π/T
0.1π/T
0.9
0.3π/T 0.2π/T 0.1π/T
0.4π/T
0.3
0.1
0.4π/T
0.2
0.8
0.4
0.5
0.6
0.7
HWIL Poles
HWIL Zeros
Flight Poles
Flight Zeros
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
Figure 2-12: Discrete P-Z plot for the experimental and HWIL sim models.
Damping and time constant of short period mode agree to a
certain margin, but the slow dynamics are not captured well in
this experiment.
zero in Eq. 2.27, the dynamics can be simpli?ed to a second order system involving
only the body x-acceleration, u˙, and the pitch rate, θ
˙
[27]
?
u˙
? ? ?
m
u
V
o
?m
q
z
u
?
?g
?? ? ? ?
m
δ
e
V
o
?m
q
z
δ
e
? ?
u
?
θ
˙
?
=
?
x
u
?x
w
m
w
V
o
?m
q
z
w
?? ?
+
?
x
δ
e
?
m
w
V
o
?m
q
z
w
?
δ
e
?
m
u
z
w
?m
w
z
u
? ?
m
δ
e
z
w
?m
w
z
δ
e
?
0 θ
m
w
V
o
?m
q
z
w
m
w
V
o
?m
q
z
w
x˙
p
= A
p
x
p
+ Bu (2.34)
The phugoid mode can then be approximated by ?nding the poles of Eq. 2.34,
Δ(s)= sI ?A
p
= s
2
+2ζ
p
ω
p
+ ω
2
p
|
?
|
?
m
u
V
o
?m
q
z
u
?? ?
m
u
z
w
?m
u
z
u
?
2
= s ? x
u
?x
w
s + g
m
w
V
o
?m
q
z
w
(2.35)
m
u
z
w
?m
u
z
u
m
For conventional aircraft in subsonic ?ight several approximations can be made [27]
u
≈ 0; m
w
z
u
|; m
w
(2.36)|m
u
z
w
|?| | V
o
|?|m
q
z
w
|
51
These simpli?cations yield closed-form expressions for the damping and natural fre-
quency of the mode
x
u
(2.37)2ζ
p
ω
p
≈?
gz
u
?
?
(2.38)ω
p
≈
V
o
As outlined in [27], the concise stability derivatives can be transformed into their
dimensionless equivalents and related to the aerodynamic parameters of C
L
and C
D
.
The phugoid frequency and damping can then be expressed in terms of the aerody-
namics and airspeed, assuming lift is equal to the weight in the trimmed condition
g
√
2
(2.39)ω
p
≈
V
o
C
D
total
C
D
w
+ C
D
b
+ C
D
p
(2.40)ζ
p
≈ √
2C
L
w
≈ √
2C
L
w
The subscripts w, b and p denote the contributions due to the wing, body and par-
asitic forms of the drag for the aircraft, respectively. Eq. 2.39 is known to provide
reasonable approximations for the phugoid natural frequency, but due to the simpli-
fying assumptions and uncertainty in the lift and drag coe?cients, Eq. 2.40 is less
accurate in determining the damping ratio, ζ
p
.
V
In order to observe the phugoid dynamics, an experiment under trimmed condi-
tions with airspeed and/or pitch angle dynamics is required. Unfortunately due to
space constraints, an experiment requiring long term “hands o?” dynamics on the
actual aircraft is not feasible, however this test can easily be performed on the HWIL
simulator. The results of a 3 m/s airspeed impulse response with initial airspeed,
o
= 28 m/s and zero angle of attack are shown in Figure 2-13. The best ?t curve for
the oscillation was found to have damping and natural frequency
ω
p(?t)
=0.0732 rad/s and ζ
p(?t)
=0.20 (2.41)
which are close to the predictions from the phugoid approximations in Eqs. 2.39 and
2.40 with predicable uncertainty in the damping term
ω
p(theory)
=0.0786 rad/s and ζ
p(theory)
=0.14 (2.42)
52
?2.5
?2
?1.5
?1
?0.5
0
0.5
1
1.5
Δ
Airspeed [m/s]
HWIL Exp data
Best Fit
Theory
14.3 s 28.6 s 41.0 s
ω
P(theory)
= 0.0786ζ
P(theory)
= 0.14
ω
P(fit)
= 0.0732
ζ
P(fit)
= 0.20
?10 0 10 20 30 40 50 60 70
Time [s]
Figure 2-13: HWIL simulation of the phugoid mode using an airspeed per-
turbation con?rms analytical predictions from the initial speed,
V
o
= 28 m/s and Lift/Drag curves. C
L
w
and C
D
w
are deter-
mined from the wing lift curve at α = 0. The values calculated
for the body and parasitic components of drag are C
D
b
=0.021
and C
D
p
=0.01, respectively.
For these values, the wing lift and drag, C
L
w
and C
D
w
terms were determined from
the wing lift curves with α = 0 in Figure 2-1(b), while the body drag coe?cient can
be estimated from the power curve at the operating speed and the fuselage geometry.
2.2.2 Lateral Dynamics
Similar to the longitudinal model, the dynamics describing lateral perturbations about
an equilibrium trim condition can be written in concise state space form [27]
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
v˙
p˙
y
v
y
p
y
r
y
φ
l
v
l
p
l
r
l
φ
v
p
y
δ
a
y
δ
r
l
δ
a
l
δ
r
n
δ
a
n
δ
r
?
δ
a
?
?
?
?
?
?
?
?
?
?
?
?
?
? ??
?
?
?
?
(2.43)
+
=
?
?
?
r˙
φ
˙
?
?
?
?
?
?
n
v
n
p
n
r
n
φ
0 1 0 0
?
?
?
?
?
?
r
φ
?
?
?
?
?
?
δ
r
0 0
53
where v is the sideslip velocity, p and r represent the roll and yaw rates, and φ is the
roll angle. The aileron control de?ection angle is denoted by δ
a
, and since the e?ect
of the rudder cannot be neglected, it provides an additional control input δ
r
.
Roll Subsidence Mode
While the two longitudinal modes decouple rather easily, this is not the case in lateral
aircraft dynamics. The lateral motion is characterized by one oscillatory mode known
as the Dutch roll mode, and two ?rst-order lags known as the roll subsidence and
spiral modes. Because all of the lateral motion is coupled to some extent, it makes
identifying individual modes more di?cult. Simplifying approximations can still be
made for small roll angles, since there is little yaw or sideslip induced, particularly
for fast dynamics. In this case the lateral-directional model can be reduced to [27]
? ? ? ?? ? ? ?? ?
p˙ l
p
l
φ
p l
δ
a
l
δ
r
δ
e
(2.44)
?
=
? ?? ?
+
? ?? ??
φ
˙
10 φ 0 0 δ
a
If aircraft wind axes are assumed (l
φ
= 0), then Eq. 2.44 further simpli?es to a ?rst-
order equation representing the roll subsidence mode
p(s) k
p
T
pδ
a
(s)= = (2.45)
δ
a
(s) s +
T
1
r
Similar to the longitudinal modes, data can be sampled in the roll and yaw axes
and transfer function models can be formed to represent the input-output relations.
The ?rst-order dynamics in Eq. 2.45 represent the initial response of the system after
the ailerons are actuated. Figure 2-14 presents the experimental and simulated output
for the roll axis – the ?rst-order model (solid line) is shown to track the output (dashed
line) well. The output of the model in Figure 2-14 is represented by the ?rst order
system
106.48
T
pδ
a
(s)= (2.46)
s +14.36
where the time constant T
r
=
14
1
.36
=0.070 sec. identi?es the dominant physical
properties for the roll mode.
Figure 2-15(a) demonstrates the tracking of two models for the roll subsidence
mode created from experimental ?ight test data and the same test performed in HWIL
54
1
[rad/s]
0.5
0
?0.5
?1
?1.5
100 102 104 106 108 110 112 114 116
Sim Output
Exp Data
Time
Figure 2-14: Simulated and measured outputs from aileron to roll rate.
tests. The data sets show strong agreement indicating that the HWIL simulation
provides an accurate representation of the testbed aircraft in roll response. As shown
in Figure 2-15(b), the time constant value for both models was found to be roughly
0.1 seconds, which is reasonable for aircraft of this size. Subtle model di?erences
created a discrepancy in the overall gain factor of the roll-rate transfer function, but
this was corrected in the HWIL simulator by adjusting the aileron chord length from
its true physical value until the transfer functions responses were in better agreement.
The roll subsidence transfer function for the HWIL simulator is
108.10
T
pδ
a
(s)= (2.47)
s +14.27
which has time constant of T
r
H
W
=
14
1
.27
=0.07 sec. This model agrees well with
Eq. 2.45, indicating strong correlation with the true roll response of the aircraft.
Spiral Mode
The roll subsidence mode can be simply decoupled from the lateral dynamics, how-
ever the spiral and Dutch roll modes are more di?cult to identify. The spiral mode
55
Measured and simulated model output ?? Roll Axis
?2
?1.5
?1
?0.5
0
0.5
1
1.5
Roll Axis [rad/s]
Experimental Model
Flight Data
Simulated HWIL Model
?2.5
?3
14 16 18 20 22 24 26 28 30
Time [s]
(a) Models created from both the actual ?ight data and HWIL simulator data
can track the experimental data.
Step response roll subsidence mode
0
1
2
3
4
5
6
7
8
9
10
Roll Rate [deg/s]
Flight Model
HWIL Model (aile chord 18%)
HWIL Model (aile chord 100%)
0 0.2 0.4 0.6 0.8 1
Time [s]
(b) Transient response for roll subsidence models. Time constant is ≈0.1 sec.
Aileron chord length was varied from true value to obtain better agreement
with model obtained from ?ight test data.
Figure 2-15: Agreement between the models from aileron input to roll rate
output for the HWIL and ?ight data models.
56
is a non-oscillatory, slowly acting mode that captures complex motions in roll, yaw,
and sideslip. It is characterized by the interaction between directional stability (also
known as weathercock or ?n stability) and the lateral stability (dihedral) of the air-
craft. These two e?ects counterbalance one another resulting in the slowly acting
spiral mode, which can be stable, neutrally stable or unstable depending on the rel-
ative strengths of the lateral directional e?ects.
Over long periods, the motion variables v, q, and p can be assumed steady, and
Eq. 2.43 can be reduced using the approximations v˙ =˙q =˙ 0. This allows thep ≈
reduced order model for the spiral mode to be formed as shown in [27]
? ? ?
(l
v
n
p
?l
p
n
v
)
??
p
?
0
(2.48)
?
φ
˙
?
=
?
y
r
l
r
n
v
1
?l
v
n
r
)
y
0
φ
??
φ
?
Since φ
˙
= p, Eq. 2.48 can be reduced to a ?rst order di?erential equation describing
the unforced spiral mode dynamics. With reference to Appendix 1 of [27], the time
constant of the mode, T
s
can be expressed in terms of the dimensionless derivatives
T
s
=
V
o
(L
v
N
p
?L
p
N
v
)
(2.49)?
g(L
r
N
v
?L
v
N
r
)
Due to the timescales over which the spiral mode acts, it typically is not possible
to identify the mode using input-output estimation techniques. Instead, a series of
analytical approximations based on aircraft geometry and aerodynamic data can be
used to identify the parameters of the spiral mode, and this prediction is used to
validate the HWIL settings. Approximations for the derivatives in Eq. 2.49 were
found in terms of the aerodynamic and geometric properties of the ARF 60 aircraft
[27]
ˉ
N
v
= a
1F
V
f
(2.50)
1
?
N
p
= (2.51)?
12
C
L
?
?
?
α=0
1
?
l
f
N
r
=(N
r
)
wing
+(N
r
)
fin
= ?
6
C
D
?
?
?
α=0
?N
v
(2.52)
b
1
L
p
= (C
L
α
+ C
D
|
α=0
) (2.53)?
12
1
ˉ
h
f
l
L
v
=(L
v
)
wing
+(L
v
)
fin
= ?
4
C
L
α
Γ ?a
1F
V
f
(2.54)
f
57
Table 2.4: Dimensionless derivatives for the ARF 60 aircraft contributing to
the lateral spiral mode
Derivative Approx.
ARF 60
N
v
N
p
N
r
L
p
L
v
L
r
Yaw Sti?ness
Yaw Moment due to Roll Rate
Yaw Damping
Roll Damping
Dihedral E?ect
Roll Moment due to Yaw Rate
0.0792
-0.0167
-0.0547
-0.3550
-0.1103
0.0389
1 l
f
L
r
= C
L
α=0
?(L
v
)
fin
(2.55)
a
6
|
b
1F
denotes the wing lift curve slope for the tail, which is analytically approximated
as a thin wing in subsonic ?ow [32]
2πAR
f
a
1F
=(C
L
α
)
fin
= (2.56)
2+ AR
f
?
1 + tan
2
Λ
f
+4
where AR
f
is the aspect ratio of the ?n. Using the measurements of the aircraft
geometry from Table 2.1, the derivatives for the ARF 60 aircraft are computed and
summarized in Table 2.4. The time constant of the spiral mode can then be computed
from Eq. 2.49 as T
s
=28. 95. From this rough analysis it is con?rmed that the spiral
mode will be a very slow, but stable mode with the positive root given by
Δ
s
(s )=(s +0. 0345) (2.57)
This mode was di?cult to observe in ?ight tests or HWIL simulations, but it was
con?rmed that the trimmed open loop ARF 60 model does not exhibit divergent
behavior over extended periods of time (> 10 min), showing that the low frequency
modes are stable as predicted.
2.3 Autopilot Tuning
The Cloud Cap autopilot has a prede?ned set of controllers that need to be tuned to
suit the 60 ARF aircraft modeled in the previous sections. The HWIL simulator is a
58
useful tool that can be used to tune the autopilot gains and control settings before
?ight testing. Previous sections have veri?ed the accuracy of the HWIL simulations,
it is expected these predictions will map well to the actual aircraft being ?own. This
section deals with the tuning of the autopilot to suit the requirements of the MILP tra-
jectory planner, which pushes the limits of the waypoint tracking strategy employed.
By uploading closely spaced waypoints and making frequent changes to the plan the
autopilot is tracking, the closed loop system is required to be responsive, however as
with any control system, care needs to be taken to protect against instability.
In addition to path following, the closed loop tracker is required to reject any
disturbances acting on the system in order to ?y the tra jectory plans designed. Later
sections incorporate methods for estimating the disturbance levels and accounting for
them on higher planning levels, however these estimates will be imperfect, requiring
that the autopilot be successful in rejecting bounded levels of wind and measurement
noise error. Subsection 2.3.1 addresses the lateral autopilot controllers, the waypoint
tracker is described in Subsection 2.3.2, and Subsection 2.3.3 describes the process
in tuning the airspeed/altitude controllers. All of the tunings completed using the
HWIL simulator are also validated during actual ?ight testing of the ARF 60 aircraft.
2.3.1 Lateral Autopilot
E?ective strategies for autonomous control are often accomplished by nesting numer-
ous control loops to build up to a complicated desired level of functionality. However
the overall performance of the system is then limited by the bandwidth and perfor-
mance of the innermost loops. Because the autopilot will eventually be used for tight
tra jectory following, the tuning of the inner loops is important to gain adequate sys-
tem performance. Quick, tight response is needed on the turn rate loops on the order
of the roll dynamics response (from Subsection 2.2.2, T
r
=14.27), however settings
that are too aggressive quickly lead to roll oscillations and eventual instability.
59
Figure 2-16: Lateral Autopilot block diagram
Turn Rate
The turn rate of the aircraft is regulated by maintaining a desired bank angle, giving
an acceleration in the radial direction as shown in Figure 4-3. Assuming a coordinated
turn, the desired bank angle, φ
d
,is
φ
d
= (2.58)?
v
g
a
ψ
˙
d
where the desired turn rate, ψ
˙
d
, is the output of the waypoint tracker controller, v
a
is the aircraft airspeed, and φ is the bank angle (positive for right wing down). The
bank angle is regulated by generating a control signal to the ailerons, u
aile
, using
PI control on the bank angle error, e
φ
(t)= φ
?
(t)?φ
d
(t), where φ
?
is the bank angle
estimate obtained from ?ltered roll rate measurements with bias correction,
t
?
1
? ?
T
u
aile
(t)= K
φp
e
φ
(t)+ e
φ
(τ )dτ (2.59)
φ
0
and K
p
and T
φ
are the controller gain and reset rate, respectively. This PI control
strategy is appropriate for this application, as integral action is needed to ensure
the tracking of the desired turn rate command from the waypoint tracker, while the
separate roll damper loop handles upsets caused by turbulence. Note that another
60
(a) Tail view, with the roll angle, (b) Overhead view, with the air-
φ, and lift forces, L, shown. speed vector, v
a
, and yaw an-
gle, ψ, shown.
Figure 2-17: Aircraft in a coordinated turn undergoing lateral acceleration,
a. The ?gure assumes that the airspeed, v
a
, is constant.
strategy using strictly yaw rate measurements is available, if the roll angle estimates
cannot be used due to high noise levels.
Using the HWIL simulator the turn rate controller can be tuned to provide good
performance before ?ight testing. Figure 2-18 shows the response to a bank angle
command as given by Eq. 2.58 for varying K
p
gains and constant reset rate, 1/T
φ
=
0.1. The plot shows the increased rise time and settling time for increasing K
P
gains,
however at the cost of increased turn rate oscillations. The most appropriate gains
found for the HWIL simulation model are shown with the thicker lines, indicating
reasonable settling time of about 5 seconds, and only slight oscillations in the turn
rate.
Roll damper
K
The roll damper is used to counteract roll turbulence by direct feedback of the roll rate
measurement, φ
˙
to the ailerons before it integrate to the e?ect the turn rate controller.
A low pass ?lter (10 rad/s bandwidth) is applied to the roll rate measurement, and
GPS is used to correct for the drift bias before feeding it through the roll damper,
φ
. Figure 2-19 shows the e?ect of applying the added derivative action to the turn
rate loop, with increasing damper gain, K
φd
. The derivative action softens the roll
response of the autopilot and increases the bank angle oscillations. Since the 60 ARF
trainer aircraft already has a high degree of roll stability, very little, if any, damper
61
Response to Bank Angle Cmd
[deg/s]
[deg]
Turn Rate Controller Tuning
30
25
20
15
10
5
0
Roll angle φ
Response [deg]
Yaw rate ψ
Response [deg/s]
.
φ
d
[deg]
K
p
=?0.1, K
i
=?.01
K
p
=?0.1, K
i
=?.01
K
p
=?0.3, K
i
=?.03
K
p
=?0.3, K
i
=?.03
K
p
=?0.5, K
i
=?.05
K
p
=?0.5, K
i
=?.05
K
p
=?0.7, K
i
=?.07
K
p
=?0.7, K
i
=?.07
?5
0 2 4 6 8 10 12 14 16
Time [s]
Figure 2-18: The performance variation of the turn rate loop with varia-
tion of the proportional and integral gains, K
φp
, and K
φi
=
K
p
T
φ
respectively. The bank angle estimate, φ
?
, and a ?ltered yaw
?
˙
rate measurement, ψ, is shown in response to a step change in
desired bank angle. The thick line provides the best tradeo?
between roll response and undamped oscillations in turn rate.
gain is needed to prevent turbulence upsets. The thicker line in Figure 2-19 shows
the response for the slight damper settings that might be used in ?ight. Note that
from the time constant of the graph the bank angle dynamics of the aircraft may be
determined. Modeling the bank angle as a ?rst order lag of the desired command,
the dynamics can be written
φ(s) 1
= (2.60)
φ
d
(s) T
c
s +1
with T
c
1.25 seconds from Figure 2-19. This measurement will be later used in ≈
Section 4.2 to modify the MILP dynamics formulation to account for the lag in accel-
eration that would be seen as a result of the step change.
62
Roll Damper Tuning
30
Response to Bank Angle Cmd
[deg/s]
[deg]
25
20
15
10
5
0
Roll angle φ
Response [deg]
Yaw rate ψ
Response [deg/s]
.
K
d
= 0.0 [deg]
K
d
= 0.0 [deg/s]
φ
d
[deg]
K
d
= 0.02 [deg]
K
d
= 0.02 [deg/s]
K
d
= 0.05 [deg]
K
d
= 0.05 [deg/s]
K
d
= 0.1 [deg]
K
d
= 0.1 [deg/s]
?5
0 2 4 6 8 10 12 14 16
Time [s]
Figure 2-19: Variation of Turn Rate performance with selection of the appro-
priate roll damper gains, K
φd
with constant proportional and
integral gains, K
φp
=0.3, K
φi
=0.01 respectively
Rudder Mixing
Aileron-Rudder mixing rules are employed to assist in making coordinated turns for
the aircraft, reducing the sideslip motion while banking the aircraft. For this trainer
aircraft this is achieved by setting the aileron to rudder mixing ratio, K
ar
=0.15,
which helps to reduce the errors in GPS heading when making turns. The e?ective-
ness of the mixing ratio could not be determined without additional measurements
on the aircraft true heading, however by including an onboard magnetometer the
sideslip motion can be measured through comparison of the GPS and true heading
estimates. Added true heading measurements would facilitate the selection of the
optimum rudder mixing ratio for a variety of ?ight conditions and airspeeds, thereby
reducing the GPS errors due to sideslip motion in bank turns.
63
(a) Waypoint tracker geometry.
(b) Tracker Response Depicted
Figure 2-20: Lateral track control law for the Cloud Cap autopilot.
The yaw damper loop as shown in Figure 2-16 is typically needed on aircraft with
less lateral directional stability, however the trainer ARF 60 aircraft have a large
enough dihedral angle, Γ = 5
?
, and vertical ?n to reject perturbations in yaw (note
the relative stability of the derivatives in Table 2.4). Yaw damper augmentation of the
plant dynamics is therefore not needed and disabled for the trainer ARF 60 aircraft.
2.3.2 Waypoint Tracker
The Piccolo
TM
autopilot uses a robust nonlinear waypoint tracker, originally designed
for implementation on the Aerosonde UAV [34], which is capable of tracking the
vehicle through series of inertially ?xed waypoints using GPS position and velocity
64
measurements. The strategy employs PD control on the error signal e
ψ
= ψ
?
?ψ
d
,
where ψ
?
is the aircraft heading estimate and ψ
d
is the desired heading vector as shown
in Figure 2-20(a). The heading estimate, ψ
?
, is approximated using the GPS velocity
vector, which is a valid assumption for low wind conditions and small aircraft bank
angles (i.e., coordinated turns),
?
v
ay
?
v
ψ
?
= arctan (2.61)
ax
The output of the waypoint tracker is a turn rate command, which is subsequently
related to the bank angle of the aircraft as shown in Eq. 2.58 [35].
Convergence Parameter Scaling
As shown in Figure 4-5, the desired heading vector, ψ
d
, is determined in the intrack
reference frame by selecting a point L
d
meters ahead of the current along track po-
sition. From Figures 2-20(a) and 2-20(b), the distance L
d
, (also referred to as the
tracker convergence parameter ), determines how sensitive the waypoint tracker will
be to cross-track errors and is selected to be approximately equal to the vehicle turn
radius. For an aircraft in a coordinated turn at a speci?ed maximum bank angle,
φ
max
=30
?
, with airspeed v
a
= 25 m/s the vehicle turn radius is given by
2
v
L
d
≈ρ
min
=
g tan
a
φ
max
≈ 120m (2.62)
Simulations con?rm that selecting L
d
< 120 leads to instability in the waypoint
tracker, as noted by highly oscillatory ?ight patterns of the vehicle in Figure 2-21.
With the selection of the convergence parameter, the desired heading vector is given
by
?
y
IT
?
L
ψ
d
= ?arctan (2.63)
d
where y
IT
is the cross track distance of the vehicle and the negative sign is required
for correct operation.
The convergence parameter, L
d
, is in fact not constant for all operating conditions,
as there are certain situations which require adjustment to the closed loop dynamics.
The control law described in Figure 2-20(a) works well when the GPS vector provides
65
Latitude [deg]
41.967
L
p
L = 100
p
41.966
41.965
41.964
= 125
Wind
41.963
5 m/s
41.962
41.961
41.96
?70.892 ?70.89 ?70.888 ?70.886 ?70.884
Longitude [deg]
Figure 2-21: Groundtrack of a ?xed counter-clockwise waypoint plan for dif-
ferent selections of the intrack convergence parameter, L
d
. Dif-
ferent selections determine the autopilot sensitivity to cross-
track error, creating unstable oscillations for small values of L
d
.
Headwinds tend to excite oscillations for the marginally stable
system, as the turn rate command provides too much authority
to the aileron actuators.
an accurate estimate of the vehicle wind-relative heading. However, due to the e?ects
of wind acting on the vehicle, the aircraft will experience changes in the e?ectiveness
of the turn rate command with changes in the wind relative to the vehicle body axis.
By scheduling L
d
with varying airspeed to increase the tracker performance in the
case of winds, the sensitivity the tracking algorithm is mitigated over the variation
in the no-wind case.
The convergence parameter is also scaled according to the vehicle cross-track error,
in order to provide higher performance for small cross-track errors, y <L
d
. In order
IT
to achieve this one of several proprietary methods is employed in the Cloud Cap
66
autopilot including a nonlinear scaling of L
d
with cross-track error [36]. This scale
factor smoothly transitions the convergence parameter to small values for y <L
d
,
IT
providing a means to keep the waypoint tracker sensitive for all regions of operation.
While detailed simulations of the system in Figure 2-16 can be performed using
the aircraft models developed in earlier sections, the nonlinear scalings and ambi-
guities in the autopilot internal control loops make accurate predictive simulations
di?cult to achieve. The simulations lead to mismatches between the predicted and
actual gains needed for HWIL or ?ight testing, although general trends and limits
can be con?rmed. In practice these tests do not capture all the dynamics required to
accurately simulate the closed loop for controller tuning purposes. As a result, the
most e?ective way to select the control gains is to test on the HWIL simulator and
subsequently verify on actual ?ight tests.
Tracker Tuning
The performance of the tracker loop depends primarily on the inner turn rate loop
performance as well as the airspeed of the aircraft being ?own. Gains that are set at
low airspeeds will typically not have as much authority as the same gains at higher
airspeeds, while gains set at the high end of the airspeed range can lead to oscillations
if left unchecked. For this reason, the HWIL simulator is exercised at several points
of the ?ight envelope to ensure proper operation.
K
One suitable method for ?nding autopilot gains is to increase the proportional
term, K
p
, until the critical point for sustained oscillations is reached. Once this
critical point is found, the proportional term can be reduced by 20-30% in magnitude,
depending on the airspeeds the aircraft will be ?ying. The derivative gain, K
d
, is then
used to soften the autopilot response and further reduce oscillations as needed. The
results of this process are shown in Figure 2-22, for an aircraft being ?own at 26
m/s, close to the upper limit of the speed range. With the proportional gain set at
p
= ?0.25, the derivative gain of K
d
= ?0.5 is shown to signi?cantly reduce the
undamped response, with a settling time of 17 seconds and 10% overshoot.
67
Tracker Performance for derivative settings, K = 0.25, L
d
=120 m
p
120
Crosstrack Error [m]
100
80
60
40
20
0
?20
K
d
= 0.0
K
d
= ?0.1
K
d
= ?0.2
K
d
= ?0.3
K
d
= ?0.4
K
d
= ?0.5
0 5 10 15 20 25 30
Time [s]
Figure 2-22: The response of the closed loop system is shown for varying
derivative tracker gains, reducing the undamped oscillations of
the system. Closed loop settling time from a step disturbance of
100 m in cross-track error is shown to be reduced to 17 seconds
with 10% overshoot.
Dynamic MILP Trajectory Conversion
The output of the MILP planner is a series of ordered waypoint lists (?ight plans )
that are uploaded to the vehicle in real-time as each optimization completes. These
waypoint lists represent the physical locations in (x, y) space that the MILP planner
assumes the vehicle will be passing through at a given time, and they need to be
converted into a form that is consistent with the closed loop response of the waypoint
tracker. The waypoint lists are updated at a rate of roughly 1 plan every 4-5 seconds,
depending on the wind conditions. As shown in Ref. [8], the process is repeated when
the vehicle reaches the horizon point of the current plan.
68
Figure 2-23: Conversion of MILP waypoints to vehicle ?ight plans, utilizing
the waypoint tracker control law to keep bank angle constant
throughout a turn. With a small enough waypoint spacing,
Γ ≈ 100m, the desired heading vector, H, will point as shown
for the tra jectories with large changes in heading angle, Δψ.
Figure 2-23 shows the tra jectory points {A,B,C,D} with spacing Γ that were
designed by the MILP solver. To prevent the vehicle from experiencing undesirable
transients when new plans are uploaded mid-turn, the modi?ed plan {A
?
,C,D} is
uploaded to the vehicle. A
?
is selected as the straight line projection of the line CB
of distance Γ, as shown in Figure 2-23. The point B is not needed and is therefore
removed from the plan upload to the vehicle, however B is used as the reference
to the start of the next optimization (i.e., the Horizon Point ). The modi?ed plan,
{A
?
,C,D}, closely resembles the situation in Figure 4-5, and allows the vehicle to
maintain a constant bank angle as each point in the plan is reached. Note that
in the current implementation the vehicle receives new plans approximately every
4 seconds (Γ = 100m), meaning that large cross-track errors will not have time to
accumulate before new plans are uploaded to the vehicle. As a result, the convergence
parameter is scaled using the nonlinear scaling law for small cross-track, and for most
of the MILP tra jectory design the waypoint tracker will be operating in the scaled
nonlinear region. This provides bene?ts to the performance of the system, but makes
the closed loop response more di?cult to predict.
69
Figure 2-24: The longitudinal altitude, h, and airspeed, v
a
, control loops
shown for the Cloud Cap autopilot. The total and static pres-
sure, pitch rate and air density P
t
,P
0
,θ
˙
, and ρ, respectively,
are measured and used to determine the throttle and elevator
control inputs δ
throt
and δ
e
. There is also an optional PD loop
from altitude to elevator not shown here.
2.3.3 Airspeed/Altitude Control
Due to the inherent coupling between airspeed and altitude, the Cloud Cap autopilot
utilizes several coupled air, pitch and altitude control loops to maintain the desired
reference values. Altitude is not controlled directly, rather using an energy expres-
sion to help damp phugoid oscillations. Due to the coupling between altitude and
airspeed, the tuning of the various controller settings in Figure 2-24 is an involved
process requiring several iterations to ensure satisfactory performance in both vari-
ables. Later sections will utilize the reference speed commands to perform timing
control of the vehicle, however the altitude will remain essentially ?xed for the most
?ight experiments.
The true airspeed of the aircraft is estimated by measuring the total and static
70
pressures, P
t
and P
0
, of the incident air?ow
?
2P
t
P
0
v
a
= and ρ ≈
ρ 287T
where ρ is the air density estimated using onboard measurements of the air tem-
perature, T . The airspeed is regulated using PID control on the dynamic pressure
error signal, e(t)= P
t
(t) ?P
tD
, where P
tD
=1/2ρv
2
is the desired total pressure
a
corresponding to the target airspeed of the aircraft.
The tuning of the airspeed loop is shown in Figure 2-25, which shows the airspeed
response to a step input of 6 m/s for varying cases of control selections, I, II, III,
and IV. Case I shows the e?ect of a high proportional gain in the airspeed loop, as
the elevator is excited too heavily producing undamped oscillations. The airspeed
proportional gain is shown reduced by 25% in II, and setting the altitude derivative
gain 25% higher in Step III reduces some of the oscillation due to throttle. Finally
in step IV, the pitch damper term, K
θp
, is used to further decrease the oscillations,
and the closed loop response is shown to have a settling time of approximately 25
seconds (note that the initial step at t = 5 seconds). In Figure 2-25(b), the altitude
variation for the same series of tests is shown to lie within ±10 m of the nominal for
these control settings. These control settings provide adequate performance for the
tests required, however the entire space of options for control selection in this axis
was not fully explored. It is therefore possible that higher performance solutions exist
for the loops shown, and could be found with more detailed modeling of the closed
loop dynamics. The closed loop response of the airspeed loop will be required in later
chapters to develop an additional control loop on the relative timing of the mission
plan being ?own.
2.4 Conclusions
This chapter presented dynamics models for the trainer 60 aircraft which were used
to validate the Cloud Cap HWIL simulation environment. By analyzing the response
to speci?c modes of lateral and longitudinal aircraft dynamics, the reduced order
71
Airspeed Step Response
28
27
26
Altitude [m]
Airspeed [m/s]
25
24
23
22
21
CMD
I
II
20
III
IV
19
0 10 20 30 40 50
(a) Airspeed step response.
Altitude Responses to Step in Airspeed Cmd
165
160
155
150
145
140
CMD
I
II
III
IV
0 10 20 30 40 50
Time [s]
(b) The altitude response to a step change in desired airspeed, showing the
coupling between the airspeed and altitude loops.
Figure 2-25: Airspeed step response input and output of the airspeed con-
troller. Note the timescale indicates the initial step at t =5
seconds.
72
models developed here for the ARF 60 aircraft were shown to have similar responses
and characteristics to the models created from HWIL simulations. While a complete
model of the ?ight dynamics could not be formed, it was shown that the HWIL simu-
lator does represent certain key aspects of the dynamics in both the fast (short period,
roll subsidence modes) and slow (phugoid and spiral modes) regions, indicating its
suitability to represent the ARF 60 aircraft for autopilot controller tuning and high
level HWIL simulations.
The dutch roll mode was not modeled due to the di?cultyin collecting experimen-
tal data isolating this response. The inherent roll-yaw axis couplings are notorious for
being di?cult to experimentally verify, and reduced order analytical models require
making gross approximations in the dynamics and aerodynamic derivatives. Had dif-
?culties arisen in making the transitions from HWIL simulation to ?ight testing more
detailed models could have been considered.
Section 2.3 described the Piccolo
TM
autopilot tunings set up to be utilized in the
trainer ARF 60 aircraft. The closed loop system performance has been identi?ed
for the lateral track waypoint controller, as well as the airspeed controller shown in
Subsection 2.3.3, which will allow planning control loops to be closed in later sections
with greater accuracy. The performance of the inner loops has been veri?ed on
repeated occasions during actual ?ight tests.
73