You are on page 1of 15

An Airborne Experimental Test Platform: From Theory to Flight

Andrei Dorobantu, Will Johnson, F. Adhika Lie, Brian Taylor, Austin Murch,
Yew Chai Paw, Demoz Gebre-Egziabher, and Gary Balas
AbstractThis paper provides an overview of the experimen-
tal ight test platform developed by the University of Minnesota
Unmanned Aerial Vehicle Research Group. Key components
of the current infrastructure are highlighted, including the
ight test system, high-delity nonlinear simulations, software-
and hardware-in-the-loop simulations, and the real-time ight
software. Recent ight control research and educational appli-
cations are described to showcase the advanced capabilities of
the platform. A view towards future expansion of the platform
is given in the context of upcoming research projects.
I. INTRODUCTION
Reliable and accessible experimental test platforms are
key enablers for the transition of theoretical research into
practice. In the aerospace domain, these platforms include
a ight test system as well as a simulation environment
for nonlinear aircraft dynamics, software-in-the-loop (SIL)
testing, and hardware-in-the-loop (HIL) testing. The related
software suite, which is often a custom implementation,
must be supported by version control and documentation
management systems. It is important to continue developing
these infrastructures so that new theory can be efciently
guided through a process of verication, validation, and,
ultimately, application. Over the past decade, xed-wing
unmanned aerial vehicles (UAVs) have become critical in
the aerospace community as experimental test platforms for
transitioning new control theory to real systems.
The University of Minnesota (UMN) UAV Research
Group has developed a low-cost experimental test platform
with advanced research capabilities in the of areas of real-
time control, guidance, navigation, and fault detection (see
Figs. 1 and 2). This paper provides an overview of the
Fig. 1. UMN UAV Research Group ight test.
Department of Aerospace Engineering & Mechanics, University of
Minnesota. Corresponding authors: dorob002@aem.umn.edu and
balas@aem.umn.edu.
Fig. 2. UMN UAV Research Group laboratory.
current infrastructure, recent applications, and future re-
search. The platform is centered around a ight test system
and a simulation environment. The ight test system (de-
scribed in Sec. II) includes three different vehicles equipped
with arrays of avionics and sensors. The simulation environ-
ment (described in Sec. III) includes high-delity nonlinear
aircraft models, a SIL simulation, and a HIL simulation. The
test eet operates using a custom real-time software suite
that is common across the vehicles. A baseline ight control
system, including attitude and navigation state estimation,
has been developed to enable autonomous ight. The real-
time software suite, simulation environment, and baseline
ight control system are version controlled and documented.
The UMN UAV Research Group is structured to operate
under a set of guiding principles. These principles include
supporting open-source development, providing freely avail-
able ight data, and enabling resource and information
sharing. Consequently, all development is publicly available
online through the UMN UAV Research Group website [1]:
www.uav.aem.umn.edu. The website includes wiki-style
documentation, the full subversion repository (SVN) for
access to the simulation environment and the real-time soft-
ware suite, and a database of ight test data. A dedicated
software package of tools has been created and placed on
the SVN that includes data les and scripts required to
generate the results presented in this paper [2]. We encourage
the reader to download this package and take advantage
of its resources. Open access to the test platform allows
for efcient cooperation. We look forward to expanding
interaction with the larger aerospace community.
II. FLIGHT TEST SYSTEM
The ight test system developed by the UMN UAV Re-
search Group relies on three integrated components: vehicles,
avionics and sensors, and ight operations. First, a eet of
test vehicles supports research experiments with respect to a
variety of airframe size and payload requirements. Second,
an avionics and sensor array supports fundamental ight and
communication needs, as well as specic research experi-
ments. Third, a set of ight operations ensures the safety
and reliability of the entire ight test system. This integrated
system is complemented with a simulation environment,
which is described in Sec. III.
A. Vehicles
The current eet includes three versions of conventional
xed-wing aircraft that belong to the Ultra Stick family. The
Ultra Stick family is a commercially available group of radio-
controlled (R/C) aircraft. Each airframe is modied to t and
carry the necessary avionics and sensors. The Ultra Stick 120
is the largest and heaviest airframe, with a 1.92 m wing span
and 7.4 kg mass. It is capable of carrying the most payload
and is equipped with the largest array of sensors. The Ultra
Stick 25e is a 66% scale version of the Ultra Stick 120,
with a 1.27 m wing span and 1.9 kg mass. It serves as the
primary ight test vehicle due to its convenient size and is
equipped with a core avionics and sensor array. The Ultra
Stick Mini is a 52% scale version of the Ultra Stick 120,
with a 0.98 m wing span. This aircraft is used as a wind
tunnel model and is not equipped with any ight avionics or
sensors. Fig. 3 shows an example of each airframe version
currently operating in the eet.
All three vehicles have conventional xed-wing airframes
with aileron, rudder, elevator, and ap control surfaces. Each
control surface is actuated using an electric servo, with a
(a) Ultra Stick Mini (b) Ultra Stick 25e
(c) Ultra Stick 120
Fig. 3. UMN UAV Research Group ight test vehicles.
maximum deection of 25 degrees in each direction. The
propulsion systems consist of electric motors (with varying
power depending on the airframe size) that drive xed-
pitch propellers. The aircraft systems are battery powered,
designed to allow for approximately 30 minutes of power
on a single charge. Some of the key physical properties of
the three Ultra Stick aircraft operated by the UMN UAV
Research Group are given in Table I.
TABLE I
KEY PHYSICAL PROPERTIES OF ULTRA STICK VEHICLES
Parameter Mini 25e 120
Mass 0.62 kg 1.90 kg 7.41 kg
Wing Span 0.98 m 1.27 m 1.92 m
Wing Chord 0.21 m 0.30 m 0.43 m
Wing Area 0.21 m
2
0.32 m
2
0.77 m
2
Length 0.87 m 1.05 m 1.32 m
Endurance 10 15 min 15 20 min 15 20 min
Cruise Speed 12 m/s 17 m/s 25 m/s
Cost $120 $170 N/A
Thrust for the vehicles is generated by electric outrun-
ner brushless DC motors, which require electronic speed
controllers. Ultra Stick 120 vehicles are powered by Actro
40-4 motors along with Castle Creations ICE2 HV80 speed
controllers. These motors require two 5S 5000 mAh lithium
polymer (LiPo) batteries. The avionics and servos are pow-
ered by a single 4S 1650 mAh LiPo battery. Ultra Stick
25e vehicles are powered by Eite Power 25 motors along
with Castle Creations ICE LITE 50 speed controllers. These
motors require a single 3S 3000 mAh LiPo battery, which
also powers the servos. The avionics are powered by a single
3S 800 mAh LiPo battery.
The Ultra Stick 120 was initially used as a low-cost
ight test platform at NASA Langley Research Center [3].
Aerodynamic modeling efforts have included extensive static
wind tunnel tests, which were later complemented with
dynamic wind tunnel tests [4], [5]. The aerodynamic model
is a nonlinear look-up table that includes effects due to the
basic airframe, control surfaces, thrust, and angular rates.
This high-delity model was made publicly available by
NASA. Unfortunately, the Ultra Stuck 120 model airframe
is currently out of production. To ensure the continuity
of the Ultra Stuck 120 as a ight test vehicle, the UMN
UAV Research Group has acquired a stock of three spare
airframes.
The Ultra Stick 25e was co-developed as a ight test
platform by the UMN UAV Research Group [6], [7] along
with researchers at the Budapest University of Technology
and Economics in Hungary. Over time, the needs of the two
groups have evolved, and, hence, the vehicles are currently
equipped with different avionics and sensors. However, the
similarity in airframes allows for cooperation in critical
research areas, such as control and navigation. The aero-
dynamic model was derived using frequency domain system
identication techniques based on ight test data obtained
by the UMN UAV Research Group [8], [9]. More details
on the system identication research for the Ultra Stick 25e
aerodynamics are provided in Sec. VII-C.
The Ultra Stick Mini is used primarily as a wind tunnel
model. It serves as an educational tool for undergraduate
courses and laboratories for the Department of Aerospace
Engineering & Mechanics (AEM). For example, undergrad-
uates use the airframe mounted on a sting in a wind tunnel
to derive basic aerodynamic coefcients.
B. Avionics and Sensors
The architecture of the core avionics and sensor array
is shown in Fig. 4. This hardware combination is installed
onboard each Ultra Stick 120 and 25e airframe and represents
the minimum requirement for research experiment ight
tests. Some individual airframes have additional sensors
to support specic experimental functions. These specic
sensor outts will be highlighted following a description of
the core avionics and sensor array.
Fig. 4. Core avionics and sensor array.
At the center of the avionics and sensor array is the
ight computer, a phyCore MPC5200B-tiny 32-bit PowerPC
microcontroller [10]. It has a clock frequency of 400 MHz,
760 MIPS of processing power, and performs oating point
computation. The ight computer utilizes a real-time oper-
ating system called eCos [11], and the ight software is
written in C. The ight software is modularized with standard
interfaces, allowing different modules (e.g. different control
or fault detection algorithms) to be easily interchanged. More
details on the software architecture are presented in Sec. IV.
The current software utilizes about 2% of CPU capacity and
runs at a framerate of 50 Hz.
The MPC5200B has a wide range of input-output (I/O)
capabilities. It supports communication with external devices
via TTL and RS232 serial, SPI, I
2
C, and Ethernet. Commu-
nication with servo actuators is handled with PWM. Flight
data is recorded at 50 Hz and stored in the 64 MB SRAM
available onboard. The ight data is downloaded after each
ight via Ethernet connection to a ground station laptop. The
Ethernet connection is also used to load ight software onto
the ight computer. The ight computer is mounted on an
interface board, which is a custom design and handles power
and the communication interface with external devices.
A failsafe board [12] is used to switch control of the
aircraft between manual mode (human R/C pilot stick-to-
surface control) and ight computer automatic mode. In both
modes, pilot commands are recorded and provided to the
ight computer. This enables the option for piloted closed-
loop control or signal augmentation experiments. Telemetry
is sent to a ground station laptop through a wireless radio
at 10 Hz [13]. The transmitted data is visualized on a
custom developed synthetic heads-up display (HUD). The
HUD provides real-time information about attitude, altitude,
airspeed, and GPS performance.
In addition to the ight computer and mode switch, each
ight test vehicle is equipped with a core set of onboard
sensors. Measurements of static and dynamic air pressure
from a Pitot probe are used to estimate airspeed and altitude.
Pressure transducers [14] communicate with the ight com-
puter over I
2
C. Angular rates and translational accelerations
are measured with an inertial measurement unit (IMU) [15],
which communicates through SPI. This sensor is comprised
of a gyroscope triad and an accelerometer triad. The IMU is
aligned with the body axis of the vehicle and located near
the center of gravity. A GPS receiver provides position and
velocity information at 1 Hz and communicates over a TTL
serial line [16], [17]. Table II summarizes the core avionics
and sensor array. This array is integrated into a single module
that is common to all ight test vehicles, shown in Fig. 5.
Digital production and fabrication documents for the module
are available on the SVN.
TABLE II
SUMMARY OF CORE AVIONICS AND SENSORS
Component Module Cost
Flight Computer Phytec MPC5200B Microcontroller $250
Failsafe Switch AcroName Robotics RxMux $300
Interface Board AEM Custom Design $250
Receiver Spektrum AR7010 $90
Telemetry Radio Free Wave MM2-T 900 MHz Modem $590
IMU Analog Devices iSensor ADIS16405 $860
GPS Receiver Hemisphere GPS Crescent Board $300
GPS Antenna GPS Outtters Titan 3 Antenna $70
Pitot Probe Eagletree $10
Pressure Trans. AMSYS AMS 5812 (x2) $60
Total Cost $2,780
Several vehicles are equipped with additional sensors
to enhance their research capabilities. An Ultra Stick 120
(code name UMN FASER), and an Ultra Stick 25e (code
name Thor), are equipped with 5-hole Pitot probes [18].
These probes were originally used as wind tunnel prototypes
by Goodrich Corporation in Burnsville, MN (now UTC
Aerospace), after which they were donated to the UMN
UAV Research Group. Each 5-hole Pitot probe takes 4
additional air pressure measurements along with the standard
2 (the name refers to 5 pressure taps on front of sensor).
(a) Digital representation
(b) First prototype
Fig. 5. Integrated avionics and sensor array module.
The additional measurements are used to estimate angle-of-
attack and angle-of-sideslip. UMN FASER is also equipped
with wingtip sensor booms that measure angle-of-attack
and angle-of-sideslip directly [19]. Another Ultra Stick 120
(code name GPS FASER) is equipped with 2 additional
GPS antenna/receiver systems. This aircraft is specialized
for navigation and GPS research. Table III summarizes the
additional sensor equipment currently available onboard each
vehicle.
TABLE III
SUMMARY OF ADDITIONAL SENSOR EQUIPMENT
Vehicle Type Code Name Sensor
Ultra Stick 120 UMN FASER 5-hole Pitot Probe
Ultra Stick 120 UMN FASER Wingtip Sensor Booms
Ultra Stick 120 GPS FASER GPS Antenna/Receiver (2x)
Ultra Stick 25e Thor 5-hole Pitot Probe
C. Flight Operations
Typical ight experiments are divided into three segments:
take-off, research experiments, and landing. Each ight be-
gins with a manual take-off by the pilot. For safety, winds
below 10 mph with no gusts are required. Once airborne, the
pilot ies the aircraft into a race track pattern with constant
altitude (below 200 m) and obtains a steady trim. The race
track pattern is generally used to maximize available straight
and level ight time. Dimensions of the pattern are dened
by line of sight requirements. In an emergency, the pilot
must always be able to visually guide the aircraft back to
safe operation. As a result of these safety constraints, the
Ultra Stick 120 and 25e can only achieve about a 20 second
maximum of straight and level ight.
Fig. 6 shows a satellite view [20] of a hobby R/C aireld
where the UMN UAV Research Group performs its research
experiments. The runway is centered at the origin of the
map, near the Aireld label, and oriented in a north/south
direction. A ight test was conducted to determine the
maximum line of sight range, and this data is shown in
Fig. 6. Direct GPS position measurements are shown along
with the rened position estimate from the navigation lter.
More details on the navigation lter algorithm are provided
in Sec. V. The range test indicates that the pilot can safely
y inside a semi-circle with approximate radius of 500 m.
Safety rules prohibit ying behind (or over) the ight-line,
hence the operating ight range is limited to a semi-circle. To
ensure a continued safe operation, a collection of documents
has been drafted to standardize procedures and maintenance
plans.
Fig. 6. Aireld and visual range for ight experiments.
III. SIMULATION ENVIRONMENT
A simulation environment provides an important comple-
ment to the ight test system. Simulation-based development
and validation prior to ight testing reduces the total de-
sign cycle time for experimental research algorithms. The
UMN UAV Research Group maintains three simulations,
illustrated by the block diagram in Fig. 7. A common
Matlab/Simulink [21] implementation of the aircraft dynam-
ics is shared between the three simulations. This shared
implementation includes ight dynamics, actuator models,
sensor models, and an environmental model. All experimen-
tal research algorithms must pass through a validation test
in each simulation before consideration for ight testing.
E
c
'
T
E
'
E
'
E
E
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Simulink
Flight Computer
Nonlinear
Aircraft
Model
Sensor
Models
Actuator
Models
Simulink
Controller
Flight Code
S-Function
Flight Code
Fig. 7. Three levels of simulation environments.
The lowest-level and most basic simulation allows for
control algorithms to be implemented in Simulink. This
frequently serves as a rst step in the design process of new
control algorithms. The mid-level simulation is a software-
in-the-loop simulation. The SIL simulation allows a research
algorithm, written as ight code in C and interfaced via S-
function, to be validated in Simulink. Finally, the highest-
level simulation is a hardware-in-the-loop simulation. The
HIL simulation allows a research algorithm, written as ight
code in C and implemented on a ight computer, to be
interfaced with Simulink and validated. The latest versions of
all three simulations can be downloaded as a package from
the SVN, which is directly accessible from the UMN UAV
Research Group website [1].
A. Nonlinear Aircraft Simulation
A 6 degree-of-freedom (DOF) nonlinear simulation model
of the aircraft dynamics is implemented in Simulink. This
model represents a set of conventional rigid-body equations
of motion for generic xed-wing aircraft. Forces and mo-
ments due to aerodynamics, propulsion, and the environment
are integrated numerically to solve the nonlinear differential
equations. The environmental model includes a detailed
model of Earths atmosphere, gravity, magnetic eld, wind,
and turbulence. Models of the aircraft subsystems, such as
actuators and sensors are also included.
Each test vehicle is associated with three simulation com-
ponents: physical properties, a propulsion model, and an
aerodynamic model. This allows the nonlinear simulation
model to be easily recongured for a particular test vehicle.
Physical properties for each airframe are determined in the
lab, where moments of inertia are found using bilar pendu-
lum swing tests. Wind tunnel tests are used to characterize
the motor and propeller thrust, torque, and power for each
aircraft.
The aerodynamic models vary depending on the airframe.
The Ultra Stick 120 aerodynamic model is derived from
extensive wind tunnel data obtained at NASA Langley
Research Center [3]. This is a high-delity model that
covers large ranges of angle-of-attack and angle-of-sideslip
aerodynamics, and is implemented as a look-up table. The
Ultra Stick 25e aerodynamic model is derived using ight
test data and frequency domain system identication tech-
niques [1], [8], [9]. This model is linear and assumes constant
aerodynamic coefcients. The Ultra Stick Mini aerodynamic
model is based strictly on wind tunnel data obtained by the
UMN UAV Research Group.
Linear models of the aircraft dynamics (about an operating
point) are frequently desired for the design of control algo-
rithms. The 6 DOF nonlinear simulation model is set up for
trimming and linearization. Automatic functions to perform
these tasks are provided in the simulation package. After the
performance of a typical control algorithm has been veried
using the linearized dynamics, it must be veried using the
nonlinear simulation. The gray-shaded controller shown in
Fig. 7 illustrates this verication process.
B. Software-in-the-loop
The SIL simulation uses the 6 DOF nonlinear simulation
model in feedback with a control algorithm implemented
as ight code in C. This implementation of the control
algorithm is interfaced with Simulink through an S-function
block. Fig. 7 represents the SIL simulation with the blue-
shaded controller. The ight control algorithm alone is linked
to the S-function; the remainder of the ight software is not
included. The primary purpose of the SIL simulation is to
verify the accuracy of a control algorithm transition from
Simulink (mathematical discrete-time model) to ight code
written in C.
C. Hardware-in-the-loop
The HIL simulation is an extension of SIL simulation
that includes the ight software and ight computer. In
Fig. 7, this simulation environment is represented by the
red-shaded controller. The entire ight software suite is
compiled and runs on the ight computer in sync with
the nonlinear simulation model. The MathWorks Real-Time
Windows Target toolbox [22] is used to ensure the simulation
runs in real-time on a Windows PC. This is crucial to obtain
meaningful results when the ight computer is included in
the simulation loop.
The nonlinear simulation model, in Simulink, interfaces
with the ight computer using a serial connection. The
ight software is modied in two ways in order to interface
correctly with the HIL simulation. First, the data acquisition
code (which normally solicits the onboard sensors) reads sen-
sor data from the nonlinear simulation. Second, the actuator
commands (which are normally delivered to the actuators
via PWM signals) are sent back to the nonlinear simulation.
Through the HIL verication process, any implementation
issues or bugs associated with a control algorithm are identi-
ed and resolved. The HIL simulation is also useful in testing
attitude and navigation state estimation algorithms, such as
the one described in Sec. V.
The HIL simulation provides an interface for an R/C
pilot through a USB R/C-style remote. The aircraft state can
be visualized via FlightGear [23], which is an open-source
ight simulator. This interface can be used to evaluate the
performance and handling qualities of a control algorithm
prior to ight testing.
IV. FLIGHT SOFTWARE
The software implemented on the MPC5200B ight com-
puter is programmed as a single-thread, real-time process
executing at 50 Hz. The entire real-time software suite and
simulation environment (described in the previous section)
are managed by a version control and documentation man-
agement system.
A. Real-time Software
The ight software is divided into code modules that
are called in sequence by the main function. Each module
is dedicated to a certain type of computation, e.g. atti-
tude/navigation estimation or control algorithm. All can-
didate modules use an interface layer, which allows the
software engineer to easily select which modules to compile
in order to build the full program. In general, each type of
module must be present in the compiled code, even if no
computation native to a given module type is required. Fig. 8
shows a schedule and order diagram of the code modules
implemented in the ight software.
Fig. 8. Real-time software schedule and order diagram of code modules.
A real-time clock is managed by an open-source, real-
time operating system (RTOS) called eCos [11]. The RTOS
provides alarms to the ight software that trigger code mod-
ules to execute. Three alarms are scheduled and validated in
order to allow enough time for the software to execute on
a 0.02 second frame. The data acquisition module (DAQ)
is triggered by the rst alarm immediately at the start of
a new frame. This module reads data from the onboard
sensors and is allowed the longest time to execute. For the
HIL simulation, a different DAQ module is compiled in
order to interface with Simulink. An INS/GPS algorithm is
executed in the navigation lter module (NAV). Commands
to the control system are generated in the guidance law
module (GL). Potential sensor faults, such as biases, can
be added in the sensor fault module (SF). The main control
algorithm is computed next (CL). Signals required for system
identication (SI) can be augmented after the control law.
Potential surface faults can be added in the surface fault
module (SF). Once this sequence of modules has executed,
the software waits for the next alarm.
The actuator module (ACT), which sends PWM signals
to the actuators, is triggered by the second eCos alarm.
For HIL simulation, a different ACT module is compiled
in order to interface with Simulink. It is important that the
time in between the DAQ and ACT modules is consistent;
it represents the time delay of the ight computer. After
the ACT module has executed, the software waits for the
next alarm. A third eCos alarm allows the data logging (DL)
and telemetry (TM) modules to execute. The sequence of
modules then repeats as the software waits for the DAQ alarm
from eCos, which indicates the start of a new frame.
The ight software is built and compiled using a
makefile. This makefile species which version of
each code module is included in the program. For example,
to compile software for HIL simulation instead of ight,
the HIL versions of DAQ and ACT are selected in the
makefile. This approach allows for modularity and soft-
ware exibility.
B. Version Control and Documentation
The ight software and simulation environments are man-
aged by a version control program. The UMN UAV Research
Group utilizes the open-source subversion [24] server to
manage a software development repository. This repository
is available publicly on the research group website [25].
The ight software is automatically documented using Doxy-
gen [26]. This utility allows documentation to be generated
directly from the source code. Each le in the ight software
suite has a special Doxygen header that allows the automatic
documentation to be generated.
V. NAVIGATION STATE ESTIMATION
The attitude state of an aircraft must be estimated from
measurements provided by the onboard sensors. Position
and velocity states, on the other hand, can be measured
directly with a GPS receiver. Typical receivers, however,
do not provide data at a sufciently high rate for use in
feedback control of aircraft. For example, the GPS receiver
used by the UMN UAV Research Group provides data to the
ight computer at 1 Hz. A sensor fusion algorithm is thus
required to provide accurate and high bandwidth estimates
of the aircraft attitude, position, and velocity. These states
are known as the navigation states.
The navigation state estimates are computed with an
algorithm that integrates an inertial navigation system (INS)
with GPS. More in-depth description of INS/GPS integration
can be found in the literature [27][30]. An INS provides
measurements at higher bandwidth than a typical GPS re-
ceiver. For example, the IMU used by the UMN UAV
Research Group provides measurements at 800 Hz. Due to
ight software limitations, however, the IMU is only sampled
at 50 Hz by the ight computer. Numerically integrating
IMU measurements to obtain the navigation state estimates
leads to unbounded errors that grow over time. A combined
INS/GPS algorithm provides a solution that has the high
bandwidth of the INS and the drift-free long-term stability
of the GPS measurement.
Attitude can be equivalently described by Euler angles or
the quaternion. The INS/GPS estimation algorithm utilizes
the quaternion for computation, and converts the solution
to Euler angles for control. Attitude determination using an
IMU calls for integrating the angular velocity measurement

B
to propagate the attitude forward in time. Rate gyro-
scopes measure inertial rotation and should be compensated
to account for the Earths rotation rate and the transport rate
due to the Earths curvature [28]. For consumer/automotive
IMUs used in low-cost UAV applications, however, these
terms are small ( 10
5
rad/s) compared to the noise level
in the sensors, and are thus neglected.
The INS uses measurement of the acceleration to generate
position and velocity estimates. A triad of accelerometers
in the IMU provides measurement of the force-over-mass
(specic force) acting on the aircraft. When the IMU is
rigidly attached to the aircraft, this specic force is measured
in the body frame (f
B
) and thus needs to be rotated into the
navigation frame (f
N
) before integration and propagation in
time. This rotation uses the aircraft attitude to formulate the
required transformation matrix C
N
B
. After compensating for
gravity and the Coriolis effect, the force can be integrated
once to yield velocity, and twice to yield position. The
structure of INS/GPS integration algorithm is shown by the
diagram in Fig. 9.
E
c
f
B

N
c
C
N
B
E E E
c
f
N
a
N
T T

c c
c
'
'
_
d
'
_

()dt

r
r
r
v
N
INS

()dt

r
r
r
p
N
INS
ba, bg
p
N
v
N
'
cc
v
N
GPS
p
N
GPS
GPS
Extended
Kalman Filter
Bias
Correction
IMU
Gyro Triad
Accel Triad
Attitude
Determination
Coriolis/Gravity
Correction
Fig. 9. Block diagram of INS/GPS integration.
An Extended Kalman Filter (EKF) [31] is used to correct
the attitude, velocity, and position estimates for errors. In
order to improve the navigation solution between GPS mea-
surements, and to allow coasting during short GPS outages,
the EKF makes frequent corrections to compensate for the
inertial sensor errors. Although more sophisticated sensor
error models exist, a simplied model presented in [32] is
used. This model is robust to parameters that are unob-
servable when the UAV is not accelerating [27]. Using this
model, the estimated sensor bias not only represents the true
bias corrupting the measurement, but also accounts for all
unmodeled errors that corrupt the sensor measurement.
The INS/GPS algorithm is initialized as soon as a valid
GPS measurement becomes available (when implemented
on the ight computer). Accordingly, position and velocity
estimates are initialized at the rst available position and
velocity measurements. Since the initialization is set to occur
on the ground before a ight test begins, the attitude is
initialized to an approximate attitude of the aircraft on the
ground. Fig. 10 shows a portion of navigation ight data
obtained during a ight test with Thor.
N
Airfield
East displacement [m]
N
o
r
t
h

d
i
s
p
l
a
c
e
m
e
n
t

[
m
]


50 0 50 100 150 200
200
150
100
50
0
50
100
150
Navigation Solution
GPS Measurement
Fig. 10. Navigation solution from INS/GPS integration algorithm.
The ight test result in Fig. 10 shows the navigation
solution (computed in real-time onboard the aircraft) along
with GPS measurements as markers. The INS/GPS inte-
gration algorithm connects the GPS measurements with a
smooth, high bandwidth position estimate. Innovations in the
position estimate from the EKF are small (typically below
2 m for position, and 0.5 m/s for velocity), which indicates
convergence in the algorithm. Although this ight test was
conducted with an Ultra Stick 25e test vehicle, numerous
data sets exist in the SVN of Ultra Stick 120 ight tests that
validate the same attitude and navigation algorithm.
VI. BASELINE FLIGHT CONTROL DESIGN
A simple and reliable baseline control algorithm is re-
quired as a benchmark for the platform. A classical controller
has been designed and validated for each vehicle through
simulation and ight tests. This controller serves as the
standard for any experiment that requires closed-loop control.
The design has a two-tiered structure: an inner-loop attitude
controller and an outer-loop ight management system. The
inner-loop controller tracks desired pitch and roll angles
of the aircraft while damping out oscillations present in
the open-loop dynamics. The outer-loop controller maintains
desired altitude, airspeed, and heading direction. The outer-
loop controller is built around the inner-loop controller.
A. Inner-Loop Controller
The inner-loop attitude controller tracks desired pitch and
roll angles while damping oscillations in the open-loop
dynamics. The desired closed-loop rise times are 1 second,
and the overshoot is no greater than 5%. Robustness of 6
dB gain margin and 45 degrees phase margin are satised.
Although the inner-loop control architecture is common
for the entire eet of test vehicles, the gains are tuned
individually in order to account for variations in the aircraft
dynamics. One set of gains is valid for the Ultra Stick 25e,
and another for the Ultra Stick 120. For the control design,
the nonlinear simulation model is linearized about a trim
condition for each vehicle. Trim conditions are dened as
level ight at 17 m/s and 23 m/s for the 25e and 120 models,
respectively. The linearized dynamics are decoupled into
longitudinal and lateral/directional subsystems, and inner-
loop controllers are designed separately.
The longitudinal controller is shown in Fig. 11, where
AC
lon
is the linearized aircraft model. The inputs to the
model are throttle setting (
thr
) and elevator deection
(
ele
). The outputs used in feedback are pitch angle () and
pitch rate (q). A proportional gain pitch rate damper (K
q
)
is applied to increase damping. The pitch angle tracking
controller (K

) uses proportional-integral gain. Integrator


anti-windup logic (A/W) is implemented to handle actuator
saturation. The inputs to the closed-loop system are pitch
angle reference (
ref
) and
thr
(throttle not controlled).
AC
lon
E

thr
K
q
q
'
T
E
K

T
E E

ref
E

ele
'
A/W
c
Fig. 11. Longitudinal dynamics control architecture.
The lateral/directional controller is shown in Fig. 12,
where AC
lat/dir
is the linearized aircraft model. The inputs
to the model are aileron (
ail
) and rudder (
rud
) deections.
The outputs used in feedback are roll angle (), roll rate
(p), and yaw rate (r). A proportional gain roll rate damper
(K
p
) is applied to reject disturbances in turbulent conditions.
A proportional gain yaw damper (K
r
) is implemented to
increase damping in the Dutch roll mode. A washout lter is
also required to avoid an adverse yaw effect during turns. The
roll angle tracking controller (K

) uses proportional-integral
gain and A/W logic handles actuator saturation. The resulting
input to the closed-loop system is roll angle reference (
ref
).
AC
lat/dir
K
r

rud
E
'
r
K
p
'
T
p

ail
E E
K

E E

ref
T

A/W
c
'
Fig. 12. Lateral/directional dynamics control architecture.
Performance of the combined inner-loop controller was
validated using simulation and ight tests. The test scenario
was a step reference pattern of different amplitudes and
lengths. Pitch axis commands were applied independently
from roll axis commands. This type of pattern was used
to excite the closed-loop dynamics over a broad frequency
range. Simulation and ight tests validated that the controller,
which was designed using a linear model, performed as
expected in the simulation and in ight. Fig. 13 shows the
combined ight test and SIL simulation results for Thor.
0 5 10 15
5
0
5
10
15
Time [s]
P
i
t
c
h

A
n
g
l
e


[
d
e
g
]


Reference
Flight Data
Simulation
(a) Pitch angle tracking pattern
0 5 10 15
20
0
20
40
Time [s]
R
o
l
l

A
n
g
l
e


[
d
e
g
]


Reference
Flight Data
Simulation
(b) Roll angle tracking pattern
Fig. 13. Inner-loop control ight test vs. SIL simulation.
The ight test results conrm that the inner-loop attitude
tracking controller objectives are satised. The rise times
for the pitch and roll axes are around 1 second. There is 5%
overshoot in the pitch response, and no overshoot in the roll
response. Strong agreement between the ight data and the
simulation result afrms the accuracy of the aircraft model
and the reliability of the controller.
B. Outer-Loop Controller
The outer-loop controller is a ight management system
that tracks altitude, airspeed, and ground track angle. Its
characteristics include no overshoot on the ground track
angle, and maintaining altitude and airspeed within 5 m and
2 m/s, respectively. The architecture is shown in Fig. 14.
Inner-loop
Control
System
c
E
K
V
E

thr V
E
V
ref
Velocity
Command
T
E
K
h
E

ref
h
E
h
ref
Altitude
Command
T

E
K
g
E

ref

g
E

g
ref
Guidance
Logic
Fig. 14. Aircraft outer-loop control architecture.
The inputs to the inner-loop control system are throttle
setting (
thr
), pitch angle reference (
ref
), and roll angle
reference (
ref
). The outputs utilized for outer-loop control
are indicated airspeed (V ), altitude (h), and ground track
angle (
g
). The ground track angle is dened as
g
=
arctan(v
e
/v
n
), where v
e
and v
n
are the east and north
velocities estimated by the navigation lter.
The three guidance blocks on the left provide the ight
management system with commands V
ref
, h
ref
, and
g
ref
.
The altitude controller (K
h
) produces a pitch angle reference
command, and the airspeed controller (K
V
) produces a
throttle command. Although non-ideal for engine-out sce-
narios [6], this architecture was selected for simplicity. Both
K
h
and K
V
use proportional-integral control and implement
integrator A/W logic (not shown in Fig. 14) in order to
safely limit the commands provided to the inner-loop control
system. The throttle command is constrained between 0 and
1, and the pitch angle reference is constrained to 20

. The
ground track angle controller (K
g
) uses proportional gain,
and, hence, the roll angle reference
ref
can be constrained
directly at 45

. This limiting is required to prevent the


aircraft from rolling over due to large ground track angle
step commands.
A ight test was conducted with Thor to verify the perfor-
mance of the ight management system. The test consisted of
a series of 90 degree ground track angle step commands. In
the absence of wind, this should result in a square pattern. In
the presence of wind, the airplane ies a rectangular pattern
that drifts in the direction of the wind. Fig. 15 shows a
comparison between ight data and SIL simulation of the
ground track angle, airspeed, and altitude signals throughout
the maneuver.
0 10 20 30 40 50 60 70
90
95
100
A
l
t
i
t
u
d
e

h

[
m
]
Time [s]
0 10 20 30 40 50 60 70
180
90
0
90
180
G
r
o
u
n
d

T
r
a
c
k


g

[
d
e
g
]


Reference
Flight Data
Simulation
0 10 20 30 40 50 60 70
15
17
19
A
i
r
s
p
e
e
d

V

[
m
/
s
]
Fig. 15. Validation of outer-loop ight management system.
The simulation and ight test results conrm that the
outer-loop ight management system objectives are satised.
The ground track angle is followed by the aircraft, while
altitude is held within 5 m, and airspeed within 1.5 m/s.
Fig. 16 shows a visualization of the aircraft trajectory across
a satellite image of the aireld.
The rectangular trajectory begins at the origin in Fig. 16.
Using the ight data, an estimate of wind direction and
speed is obtained by examining the drift. The estimated wind
components are applied in the nonlinear simulation, which
shows strong agreement with the ight test.
N
Airfield
East displacement [m]
N
o
r
t
h

d
i
s
p
l
a
c
e
m
e
n
t

[
m
]


100 50 0 50 100 150 200 250 300
100
50
0
50
100
150
200
Flight Data
Simulation
Fig. 16. Satellite view of ight management system validation results.
VII. RESEARCH APPLICATIONS
The UMN UAV Research Group supports and is involved
with a wide range of research projects. Brief descriptions
of these projects are provided in this section, as well as
references to the literature for more details.
A. Fault Detection and Identication
The UMN UAV Research Group has a history of develop-
ing fault detection and identication (FDI) algorithms and
analysis tools [33][37]. Recent research uses ight data
to evaluate the performance of two FDI algorithms [38].
A set of fault scenarios was tested on Thor with active
baseline inner-loop control, including step and ramp aileron
bias faults. The ight data was post-processed using both FDI
algorithms with the goal of detecting the actuator faults.
The rst FDI approach implements a model-based strategy
using a robust linear lter. The lter design requires a model
of the aircraft dynamics in order to estimate the aileron
faults. Standard H

model-matching optimization is used to


synthesize the lter, which minimizes the innity norm (i.e.
worst-case gain) from the aileron fault to a fault estimate
error. The performance of the robust lter is evaluated by its
ability to estimate that an aileron fault has occurred.
The second approach implements a data-driven strategy
that relies strictly on raw ight test data. This approach
does not depend on a model of the aircraft dynamics.
However, an extensive collection of un-faulted training data
is required. Statistical distributions of the un-faulted ight
data are generated and then compared to the statistics of a
faulted case. An anomaly score is computed that represents
the probability that the faulted statistics are distinct from
the un-faulted statistics. The anomaly score is compared to
a threshold value, and the performance of the data-driven
algorithm is evaluated by its ability to predict that an aileron
fault has occurred.
Fig. 17 shows the results of both FDI algorithms for an
un-faulted ight test scenario. In this case, the aircraft ies a
standard lateral maneuver with no aileron bias fault. Hence
the normalized fault remains zero for all time. The fault
estimate computed by the robust lter tracks the normalized
fault. The anomaly score never crosses the threshold value
of 1, which would have indicated the likely probability that
a fault has occurred.
0 5 10 15 20
1
0
1
2
Time [s]


Anomaly
Threshold
Fault Est.
Fault
Fig. 17. FDI algorithm test in un-faulted scenario.
Fig. 18 shows the results from a 5 degree step aileron
bias fault ight test scenario. The fault was injected after 8
seconds and the aircraft attempts to complete the lateral ma-
neuver (using the inner-loop controller). The fault estimate
computed by the robust lter tracks the normalized fault to
a value of -1. At the same time, the anomaly score exceeds
the marked threshold. The data-driven anomaly score drops
below the threshold at around 15 seconds as the integral gain
in the controller counters the bias. The robust lter estimate,
however, continues to track the fault.
0 5 10 15 20
1
0
1
2
Time [s]


Anomaly
Threshold
Fault Est.
Fault
Fig. 18. FDI algorithm test in step-fault scenario.
B. Synthetic Air Data Reconstruction
The UMN UAV Research Group has developed a system
to estimate synthetic air data, which includes airspeed, angle-
of-attack, and angle-of-sideslip [39]. These quantities are
traditionally measured directly using a Pitot system con-
sisting of pressure sensors, Pitot tube, wingtip vanes, and
an internal plumbing system. In the case where installation
of such systems is prohibitive either due to weight, cost,
or mission requirement (e.g. water landing), an alternative
system to provide these quantities is desirable. Fig. 19 shows
the architecture of the synthetic air data system.
Fig. 19. Synthetic air data system architecture.
.
The synthetic air data system is a cascade of two lters.
IMU and GPS measurements are combined with a dynamic
model of the aircraft to estimate airspeed, angle-of-attack,
and angle-of-sideslip. The UMN FASER vehicle, equipped
with wingtip sensor booms that measure angle-of-attack and
angle-of-sideslip directly, is used to evaluate the performance
of this algorithm. Fig. 20 shows the synthetic estimates of
the air data quantities.
415 420 425 430 435 440 445 450
0
2
4
6
8
Time [s]
A
n
g
l
e

o
f

a
t
t
a
c
k

(
d
e
g
)


Wingtip Vane
Synthetic
Fig. 20. Flight test result of angle-of-attack estimate.
.
C. System Identication
The aerodynamic model for the Ultra Stick 25e was gen-
erated based on ight test data obtained with Thor [8], [9].
Frequency domain system identication techniques [40] were
applied to the ight test data to obtain a linear model of the
aircraft dynamics. This type of system identication relies on
exciting the aircraft dynamics over a broad frequency range,
and then tting models to the ight test data in the frequency
domain. One of the main advantages of this approach is that
the model t can be emphasized over a frequency range of
particular interest.
Frequency sweep input signals were used to excite the
aircraft dynamics. These inputs are sinusoids with varying
frequency over a particular range. The frequency sweep in-
puts were augmented with a manual pilot command to ensure
that the aircraft oscillated about the trim condition. In order
to identify the aircraft dynamics, input-output frequency
sweep ight data is required from all the control surfaces to
all the measurements provided by the IMU. Fig. 21 shows a
single frequency sweep experiment for the elevator control
surface. Note that the pilot augmentation alters the computer-
generated frequency sweep input. The corresponding pitch
rate response is shown below in Fig. 21.
The input-output ight data was processed and converted
into a set of input-output frequency responses. Nonlinear
0 2 4 6 8 10 12 14
15
10
5
0
E
l
e
v
a
t
o
r

I
n
p
u
t

[
d
e
g
]
0 2 4 6 8 10 12 14
40
20
0
20
40
60
P
i
t
c
h

R
a
t
e

[
d
e
g
/
s
]
Time [s]
Fig. 21. Elevator and pitch rate signals from frequency sweep experiment.
optimization was used to t a parametrized linear state-space
model to these frequency responses. The resulting identied
linear model was used to extract constant aerodynamic
coefcients. Finally, the accuracy of the identied model
was validated in the time domain. Under manual open-
loop control, the pilot executed a doublet sequence for the
elevator, aileron, and rudder control surfaces. A recording
of the pilot inputs from the doublet sequence was used to
simulate the identied model for comparison. Fig. 22 shows
the model validation results for the angular rate response.
0 1 2 3 4 5 6 7
200
100
0
100
200
R
o
l
l

R
a
t
e

p

[
d
e
g
/
s
]


Flight Data
Linear Model
Aileron Cmd. (5xdeg)
0 1 2 3 4 5 6 7
100
50
0
50
100
P
i
t
c
h

R
a
t
e

q

[
d
e
g
/
s
]


Elevator Cmd. (5xdeg)
0 1 2 3 4 5 6 7
100
50
0
50
100
Time [s]
Y
a
w

R
a
t
e

r

[
d
e
g
/
s
]


Rudder Cmd. (5xdeg)
Fig. 22. Validation of system identication result using ight data.
The results in Fig. 22 indicate that the primary aircraft
dynamics are captured accurately by the identied model.
The roll and pitch axes, in particular, are modeled to a high
level of accuracy. A slight miss-match is noted in the Dutch
roll mode response of the yaw rate. The identied model did
not capture the full resonant peak of this mode in the rudder
to yaw rate frequency response. Nevertheless, the result
of the system identication provides a model of sufcient
accuracy for control design and research applications.
D. Waypoint Guidance
A GPS waypoint guidance algorithm was implemented
and ight tested by the UMN UAV Research Group. The
guidance logic was originally developed by researchers at
the Budapest University of Technology and Economics in
Hungary [41] and later ported to the UMN UAV platform.
After completing modications required for integration, a set
of demonstration ight tests were carried out using Thor to
validate the performance of the algorithm.
The waypoint guidance logic ensures the reachability of
a GPS waypoint by operating in two modes that account
for the closed-loop aircraft dynamics. Mode 1 is used when
the target GPS waypoint is within the line-of-sight of the
aircraft. Line-of-sight is a tuning parameter in the guidance
logic, which in this case was dened as 20
o
relative to the
nose of the vehicle. While operating in Mode 1, the baseline
ight control system is commanded to follow a ground track
vector pointed directly at the target GPS waypoint.
Mode 2 is used when the target GPS waypoint is outside
the line-of-sight of the aircraft. The worst-case scenario
occurs when the next waypoint is near and also behind the
aircraft. Mode 2 logic checks for this scenario rst. If the
waypoint is not reachable with respect to the closed-loop
turning radius of the aircraft (about 40 m for the Ultra Stick
25e), the aircraft continues in forward ight until reachability
is achieved. When the waypoint becomes reachable, Mode 2
logic commands the aircraft to turn. As soon as the waypoint
enters the line-of-sight of the aircraft, Mode 2 is terminated
and the logic switches to Mode 1. A waypoint is considered
to be captured when the aircraft ies within a 20 m safety
zone of the target.
Fig. 23 shows the ight test result from a 4 GPS waypoint
pattern, arranged in a counter-clockwise direction at the cor-
ners of a 100x100 m square. The 20 m safety zones around
each waypoint are also shown. The guidance algorithm is
initiated at the origin on the map (lower left waypoint), with
the aircraft pointing due east. Hence, the initial trajectory
towards the lower right waypoint differs from the steady-
state trajectory. The ight test result shown represents 4 loops
around the pattern.
N
Airfield
East displacement [m]
N
o
r
t
h

d
i
s
p
l
a
c
e
m
e
n
t

[
m
]


150 100 50 0 50 100 150
50
0
50
100
150
Navigation Solution
GPS Waypoint
Fig. 23. GPS waypoint guidance ight test result.
The ight trajectory in Fig. 23 is distorted by the steady
winds that prevailed during the test. Without the presence of
wind, the ight trajectory would have been symmetric about
the center of the pattern. The ight test results suggest that a
100 m distance between GPS waypoints is near the limit of
how aggressively the baseline ight controller can perform.
E. Undergraduate Education
The UMN UAV platform is frequently used by the De-
partment of Aerospace Engineering & Mechanics as part of
the undergraduate curriculum. The Ultra Stick Mini serves
as a wind tunnel model for the main laboratory exercise
associated with the Flight Dynamics & Control course.
Students use the vehicle mounted on a sting in the wind
tunnel to measure aerodynamic forces and moments. These
measurements are used to construct a linear model of the
longitudinal ight dynamics.
The Design, Build, and Flight Test a UAV course relies
heavily on the full UMN UAV platform. Students are orga-
nized into groups tasked with the development and design
of a simple controller for the Ultra Stick 25e vehicle. Each
group must verify its control design by going though the
process of SIL and HIL simulations. Finally, the controllers
are ight tested and the students use the data obtained to
validate their designs.
VIII. FUTURE DEVELOPMENT
The UMN UAV Research Group is continuing work to
expand its capabilities and research directions. Two areas of
particular interest are highlighted here. One of the research
directions investigates the formulation of new condence
level metrics for model validation and a set of associated
analysis tools. Another research direction looks to expand
the capabilities of GPS related navigation algorithms.
A. Condence Level Metrics
Nonlinear aircraft simulation models must be tested based
on ground, bench, and ight data prior to use in validating
safety-critical ight algorithms. It is a challenge to develop
simulation models with adequate delity for all stages of
ight, e.g. nominal and off-nominal conditions. These mod-
els have uncertainty in all components and coefcients to
a varying degree, which results in a family of models.
Validation of the simulation and uncertainty models is critical
to reducing the need for extensive ight tests of the system.
Condence level metrics must be used to validate aircraft
simulation models. These metrics provide quantitative mea-
sures of model accuracy given a limited number of experi-
mental observations. Currently, such metrics are not widely
used to evaluate simulation models. One candidate metric
is the Theil Inequality Coefcient (TIC), which measures
deviation between two time domain signals [42]. Evaluating
the TIC is simple, and its interpretation can be tailored
towards various forms of analysis. For example, suppose that
x
i
is a simulation time history, and that y
i
is a ight test time
history. Considering n time samples of the data, i = 1, . . . , n,
the TIC is dened as:
TIC =

1
n

n
i=1
(x
i
y
i
)
2

1
n

n
i=1
(x
i
)
2
+

1
n

n
i=1
(y
i
)
2
. (1)
To demonstrate the concept of analysis using this metric,
1000 Monte Carlo simulations of the UMN FASER nonlin-
ear model were executed. 10% parametric uncertainty was
included in the lateral aerodynamic coefcients. The control
surface inputs were taken from a ight test aileron bank
maneuver. Fig. 24 presents the results from the ight test
and Monte Carlo simulations, including the 95% bound and
the worst-case uncertainty from the simulation.
Time [s]
B
a
n
k

A
n
g
l
e

[
d
e
g
]


1003 1003.5 1004 1004.5 1005
20
0
20
Flight Data
Sim Worst Case
Sim Mean
95% Confidence
Time [s]
R
o
l
l

R
a
t
e

[
d
e
g
/
s
]
1003 1003.5 1004 1004.5 1005
100
50
0
Fig. 24. Segment of a bank maneuver ight test using UMN FASER.
The results in Fig. 24 show that the nonlinear simulation
model is able to predict the majority of the aircraft response.
It is unclear, however, what rigorous conclusions can be
drawn regarding the slight discrepancy between the signals.
One approach is to calculate the TIC for the nominal
simulation model response versus the ight test response.
This value of the TIC can be compared to the TIC found
between the nominal response and each of the 1000 Monte
Carlo simulations. The distribution of TICs obtained for
each Monte Carlo simulation run, worst-case, and nominal
responses are shown as a histogram in Fig. 25.
The TIC provides a quantitative metric of correlation
between the nonlinear simulation model and the ight data.
Markers highlight the 95th percentile TIC, worst-case re-
sponse TIC, and the ight response. The TIC analysis
mathematically indicates that the bank angle response is
captured by the simulation (with 95% condence), while the
roll rate response is not. One of the future goals of the UMN
UAV Research Group is to continue developing these types
of analysis tools and validate their performance using the
ight test platform.
B. Advanced Navigation Research
The GPS FASER test vehicle is equipped with dedicated
systems for advanced navigation and GPS research. One
of the central challenges with developing new navigation
algorithms is the problem of ground truth. This problem
refers to characterizing the magnitude and statistical nature
of errors in the navigation solution. Carrier phase differential
0 0.05 0.1 0.15 0.2
0
50
100
150
200
TIC
B
a
n
k

A
n
g
l
e

S
i
m
s


Flight TIC
95% Confidence
Sim Worst Case
0 0.05 0.1 0.15 0.2
0
50
100
150
200
TIC
R
o
l
l

R
a
t
e

S
i
m
s
Fig. 25. Model validation using ight test data and TIC condence metrics.
GPS is an excellent truth reference for position and velocity,
providing centimeter- and millimeter-level accuracy, respec-
tively. Until recently, GPS receivers capable of generating
accurate carrier phase solutions (referred to as RTK-systems)
were large and expensive, hence not ideal for small UAV
applications. Furthermore, another key parameter needed for
a truth reference is an accurate attitude solution. While high-
quality inertial navigation systems are available for this,
they are also very expensive and do not always meet the
requirements of small UAVs.
GPS FASER is instrumented with a low-cost, carrier phase
positioning and attitude determination system. The system
consists of three GPS receiver/antenna pairs mounted on the
airplane in the triangular conguration, shown in Fig. 26. In
conjunction with a fourth GPS system on the ground, seen in
Fig. 26. Multi-antenna conguration on GPS FASER.
Fig. 27. Carrier phase differential positioning with GPS FASER.
Fig. 27, the system generates a carrier phase position solution
in post-process with accuracies on the order of millimeters.
Unlike commercially available RTK systems, which cost
around $10-20K, this system costs around $1K.
The algorithm for processing the carrier phase measure-
ments and generating the position and velocity solution was
developed at the University of Minnesota and is available
as open-source code. This system also generates an atti-
tude solution with errors less than 0.2 degrees. Thus, in
conjunction with the INS/GPS system and air data system
described previously, GPS FASER has a complete, low-cost,
truth system that is indispensable in navigation research.
One of the key navigation research challenges today is
GPS-denied or -stressed navigation [43]. GPS FASER will
be used to evaluate various systems/algorithms. Examples of
some advanced navigation concepts being explored include:
1) Model-aided Inertial Navigation for GPS-denied Envi-
ronments: This research explores navigation strategies that
use a dynamic model of the vehicle to aid inertial navigation.
The vehicle model can be used in lieu of GPS to help arrest
some of inertial navigation system drift. While it has been
shown to be a successful approach to navigation in GPS-
denied environments [44], [45], its sensitivity to errors in
the aircraft dynamic model have yet to be quantied. An
accurate dynamic model for the Ultra Stick 120 vehicle is
available from wind tunnel data, and, therefore, GPS FASER
is an ideal platform for evaluating this type of approach to
GPS-denied navigation.
2) Synthetic Air Data System: The synthetic air data
system described previously relies on GPS position and
velocity information to estimate airspeed, angle-of-attack,
and angle-of-sideslip. The availability of an independent
attitude determination system allows the implementation of
different lter architectures [39] to synthetically estimate
these air data quantities. Such examination would lead to
development of a robust synthetic estimator that is capable
of providing accurate estimates of the air data quantities.
3) Vision-aided Navigation System: Low-cost UAVs rely
heavily on the availability of GPS for navigation, and there-
fore cannot operate in areas where GPS signals are poor,
such as due to jamming [43]. Vision-based navigation (e.g.
[46]) is an ideal candidate in such situations. GPS FASER is
currently being outtted with a high-denition camera that
will be used to validate vision-based navigation algorithms.
This system integrates vision and IMU signals to compute
an accurate navigation solution in the absence of GPS.
IX. CONCLUSION
It is important for the aerospace community to continue
developing exible and efcient ight test platforms to
support the testing and validation of new theory. The Uni-
versity of Minnesota operates such a platform with advanced
research capabilities in control, guidance, navigation, and
fault detection. Collaborating with researchers, using open-
source software, and ight testing with these platforms will
expedite the development and application of new theory that
could one day revolutionize aerospace technology.
ACKNOWLEDGMENTS
This research was partially supported under the NASA
Langley contract NNH077ZEA001N entitled Analytical
Validation Tools for Safety Critical Systems, NASA Langley
contract NNX08AC65A entitled Fault Diagnosis, Prognosis
and Reliable Flight Envelope Assessment, NASA STTR
subcontract NNX12CB02C from Tao Systems Inc. entitled
Robust Aeroservoelastic Control Utilizing Physics-Based
Aerodynamic Sensing, and NSF Grant No. NSF-CNS-
0931931 entitled CPS: Embedded Fault Detection for Low-
Cost, Safety-Critical Systems. Any opinions, ndings, con-
clusions, or recommendations are those of the authors and
do not necessarily reect the views of NASA or the NSF.
REFERENCES
[1] B. Taylor, A. Dorobantu, and G. Balas, University of Minnesota UAV
Research Group. www.uav.aem.umn.edu.
[2] A. Dorobantu, W. Johnson, and G. Balas, Package for 2013 ACC
tutorial session lead paper on UAV experimental test platforms.
http://svn.umnaem.webfactional.com/trunk/ACC2013/.
[3] D. B. Owens, D. E. Cox, and E. A. Morelli, Development of a Low-
Cost Sub-Scale Aircraft for Flight Research: The FASER Project, in
25th AIAA Aerodynamic Measurement Technology and Ground Testing
Conference, no. 2006-3306, (San Francisco, CA), 2006.
[4] E. A. Morelli and R. DeLoach, Wind Tunnel Database Develop-
ment using Modern Experiment Design and Multivariate Orthogonal
Functions, in 41st AIAA Aerospace Sciences Meeting and Exhibit,
no. 2002-2795, (Reno, NV), 2002.
[5] G. Hoe, B. Owens, and C. Denham, Forced Oscillation Wind Tunnel
Testing for FASER Flight Research Aircraft, in AIAA Atmospheric
Flight Mechanics Conference, no. 2012-4645, (Minneapolis, MN),
2012.
[6] Y. C. Paw, Synthesis and Validation of Flight Control for UAV. PhD
thesis, University of Minnesota, 2009.
[7] A. M. Murch, Y. C. Paw, R. Pandita, Z. Li, and G. J. Balas, A
Low Cost Small UAV Flight Research Facility, in Advances in
Aerospace Guidance, Navigation and Control, pp. 2940, Springer
Berlin Heidelberg, 2011.
[8] A. Dorobantu, A. Murch, B. Mettler, and G. Balas, Frequency
Domain System Identication for a Small, Low-Cost, Fixed-Wing
UAV, in AIAA Atmospheric Flight Mechanics Conference and Ex-
hibit, no. 2011-6719, (Portland, Oregon), 2011.
[9] A. Dorobantu, A. Murch, B. Mettler, and G. Balas, System Identica-
tion for Small, Low-Cost, Fixed-Wing Unmanned Aircraft, Accepted
to the AIAA Journal of Aircraft, 2013.
[10] Phytec, phyCORE MPC5200B-tiny microcontroller datasheet.
http://www.phytec.com/products/sbc/PowerPC/phyCORE-
MPC5200B-tiny.html.
[11] eCos, Real-time operating system. http://ecos.sourceware.org/.
[12] AcroName Robotics, RxMux failsafe manual/auto switch datasheet.
http://www.acroname.com/robotics/parts/R306-RXMUX.html.
[13] FreeWave, MM2-T 900MHz modem datasheet.
http://www.freewave.com/products/allproducts/mm2.aspx.
[14] AMSYS, AMS 5812 pressure transducer datasheet.
http://www.amsys.info/sheets/amsys.en.ams5812 e.pdf.
[15] Analog Devices, iSensor ADIS16405 IMU tri-axial
sensor datasheet. http://www.analog.com/static/imported-
les/data sheets/ADIS16400 16405.pdf.
[16] GPS Outtters, Titan 3 GPS antenna datasheet.
http://www.gpsouttters.com/Titan3.html.
[17] Hemisphere GPS, Crescent GPS receiver board datasheet.
https://www.navtechgps.com/Downloads/Crescent OEM.pdf.
[18] Goodrich Corporation, Model 0858 multi-function probes datasheet.
[19] D. D. Kershner, Miniature Flow-direction and Airspeed Sensor for
Airplanes and Radio-Controlled Models in Spin Studies, NASA Tech-
nical Paper, no. 1467, 1979.
[20] Google, Inc., Satellite view of acrc aireld.
[21] The MathWorks, Inc, MATLAB/Simulink, r2010a.
[22] The MathWorks, Inc., Real time windows target.
[23] FlightGear, Open-source ight simulator. http://www.ightgear.org/.
[24] Apache Subversion, SVN repository. http://subversion.apache.org/.
[25] University of Minnesota, UAV research group subversion repository.
http://svn.umnaem.webfactional.com.
[26] Doxygen, Automatic documentation generated from source code.
http://www.doxygen.org/.
[27] S. Gleason and D. Gebre-Egziabher, GNSS Applications and Methods.
Artech House, 2009.
[28] P. Groves, Principles of GNSS, Inertial, and Integrated Navigation
Systems. Artech House, 2008.
[29] J. Farrell and M. Barth, The Global Positioning System and Inertial
Navigation. McGraw-Hill Professional, 1999.
[30] D. Titterton and J. Weston, Strapdown Inertial Navigation Technology.
Institution of Engineering and Technology, 2004.
[31] D. Simon, Optimal State Estimation: Kalman, H, and Nonlinear
Approaches. Wiley, 2006.
[32] D. Gebre-Egziabher, Design and Performance Analysis of Low-Cost
Aided Dead Reckoning Navigator. PhD thesis, Department of Aero-
nautics and Astronautics, Stanford University, December 2001.
[33] R. Pandita, Dynamic Flight Envelope Assessment with Flight Safety
Applications. PhD thesis, University of Minnesota, 2010.
[34] R. Pandita, J. Bokor, and G. Balas, Closed-loop performance metrics
for fault detection and isolation lter and controller interaction,
International Journal of Robust and Nonlinear Control, 2011.
[35] P. Freeman, P. Seiler, and G. Balas, Robust fault detection for
commercial transport air data probes, in IFAC 18th World Congress,
(Milan, Italy), pp. 1372313728, August 2011.
[36] P. Freeman, Robust fault detection for commercial transport air data
probes, Masters thesis, University of Minnesota, November 2011.
[37] P. Freeman, P. Seiler, and G. J. Balas, Air data system fault modeling
and detection, Submitted to Control Engineering Practice, 2012.
[38] P. Freeman, R. Pandita, N. Srivastava, and G. J. Balas, Model-
based and Data-driven Fault Detection Performance for a Small UAV,
Acccepted to IEEE Trans. Mech., 2013.
[39] F. A. Lie and D. Gebre-Egziabher, A synthetic airdata system, in
AIAA Guidance, Navigation, and Control Conference, no. 2012-4704,
(Minneapolis, MN), August 2012.
[40] M. Tischler and R. Remple, Aircraft and Rotorcraft System Identi-
cation. AIAA, 2006.
[41] P. Bauer and J. Bokor, Tuning and improvements in a waypoint and
trajectory tracking algorithm, in AIAA Guidance, Navigation, and
Control Conference, no. 2012-4604, (Minneapolis, MN), 2012.
[42] H. Theil, Economic Forecasts and Policy. North-Holland Publishing
Company, 1970.
[43] D. Gebre-Egziabher and B. Taylor, Impact and Mitigation of GPS-
Unavailability on Small UAV Navigation, Guidance and Control,
Report # 2012-2, University of Minnesota, Department of Aerospace
Engineering & Mechanics, November 2012.
[44] M. Koifman and S. J. Merhav, Autonomously aided strapdown atti-
tude reference system, Journal of Guidance, Control and Dynamics,
vol. 14, no. 6, pp. 11641172, 1990.
[45] M. Koifman and I. Y. Bar-Itzhack, Inertial navigation system aided by
aircraft dynamics, IEEE Transactions on Control Systems Technology,
vol. 7, pp. 487493, July 1999.
[46] C.-C. Chu, F. A. Lie, L. Lemay, and D. Gebre-Egziabher, Per-
formance comparison of tight and loose INS-Camera integration,
in Proceedings of the 24th International Technical Meeting of The
Satellite Division of the Institute of Navigation (ION GNSS 2011),
(Portland, OR), 2011.