Anda di halaman 1dari 7

2014 IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS 2014)


September 14-18, 2014, Chicago, IL, USA

Simulating Quadrotor UAVs in Outdoor Scenarios


Andrew Symington, Renzo De Nardi, Simon Julier and Stephen Hailes
{a.symington,r.denardi,s.julier,s.hailes}@ucl.ac.uk

AbstractMotivated by the risks and costs associated with


outdoor experiments, this paper presents a new multi-platform
quadrotor simulator. The simulator implements a novel secondorder dynamic model for a quadrotor, produced through evolutionary programming, and explained by domain knowledge.
The model captures the effects of mechanics, aerodynamics,
wind and rotational stabilization control on the flight platform.
In addition, the simulator implements military-grade models
for wind and turbulence, as well as noise models for satellite
navigation, barometric altitude and orientation. The usefulness
of the simulator is shown qualitatively by a comparing how
coloured and white position noise affect the performance of
offline, range-only SLAM. The simulator is intended to be
used for planning experiments, or for stress-testing application
performance over a wide range of operating conditions.
Figure 1. A screen shot of a QRSim search and rescue simulation.

Index TermsUnmanned Aerial Systems, Simulation

I. I NTRODUCTION
Unmanned aerial vehicles (UAVs) have received significant
research attention as a result of the numerous military and
civilian applications that they enable. Quadrotors have proven
themselves as favourable platforms for research, since they
are inexpensive, easy to fly, agile and safe to operate [1].
Conducting outdoor experiments with quadrotors requires
ideal weather conditions and compliance with local aviation
regulations, uses significant resources, and puts platforms at
risk. It is therefore preferable to evaluate scenarios in simulation prior to experimentation, where possible. However,
the suitability of such an approach depends largely on the
fidelity of the simulator. The work in this paper is therefore
motivated by the need for a quadrotor simulator that (i)
enables multiple platforms to be simulated at the application
level, (ii) provides realistic models for dynamics, wind and
sensor noise, and (iii) balances accuracy with speed, thereby
enabling real-time, or near real-time, performance.
A great number of simulators exist as tools to for training
radio control enthusiasts, but not as platforms for research.
Research simulators based on first principles have been developed for USARSim [2] and Gazebo [3]. Both capture the behaviour of an ideal platforms mechanics and aerodynamics,
but not of the low-level controller used to perform rotational
stabilization control. The design of the stabilization control
algorithm is usually proprietary, and therefore cannot be
modelled by first principles. Although effort has been made
to investigate the effect of wind and turbulence on quadrotor
dynamics [4], such research remains to be integrated into the
widely-used simulators. More recent simulators are based on
the Robotic Operating System (ROS) and include SwarmSim
X [5] and Hector Quadrotor [6], which use a PhysX and
978-1-4799-6934-0/14/$31.00 2014 IEEE

Gazebo as simulation cores respectively. The contribution of


this work is the open-source1 QRS IM simulator depicted in
Fig. 1 that distinguishes itself by:
1) A general second-order dynamic model that captures
the behaviour of mechanics, aerodynamics, wind and
stabilization control. The model is derived from flight
data using the co-evolutionary programming method,
with model parameters fitted to a specific platform.
2) Models for environment and sensor noise, including a
military-grade model for turbulence that is validated
against a third-party implementation, and a satellite
navigation model that is validated against raw data.
3) Implemented as self-contained Matlab code that is
controllable over a TCP/IP interface, with the goal of
being a modular and extensible tool for research.
The remainder of this paper is structured as follows. Sec.
II discusses the dynamic model used in the simulator, and
shows how model parameters were obtained for a specific
flight platform. Sec. III then describes the environmental and
sensor noise models, as well as how they were validated. Sec.
IV then provides a qualitative illustration of the usefulness of
the simulator by comparing the effect of white and coloured
sensor noise on offline, range-only simultaneous localization
and mapping (SLAM) for a single platform. Sec. V concludes
the paper, and discusses directions for future work.
II. Q UADROTOR DYNAMICS AND C ONTROL
A flight control system (FCS) performs sensing to determine orientation or position, and then selects corrective motor
changes to stabilize the platform [1]. Orientation is usually

3382

1 Source

code is available at https://github.com/UCL-CompLACS/qrsim

Z,W, r,

4J

Figure 2. The state of a quadrotor comprises of a position [x, y, z], Euler


orientation [c/J, B, 7jJ] , velocity [u,v, w] and angular velocity [p, q,r].

corrected up to 1kHz, whereas positron can be controlled


at a much lower rate. Closing the position control loop
however requires measurements to external reference points.
Indoor experiments use motion capture systems to provide
millimeter-accuracy positioning [7], enabling the evaluation
of complex platform [8] and team [9] maneuvers. For outdoor
experiments this is not currently possible; GPS waypoints are
fused with barometric altitude [10] to provide corrections.
A number of first order dynamic models for quadrotors
exist in the literature [11], [12], [13], which only capture
the effects of mechanics and aerodynamics. Closing the gap
between simulation and experimentation requires a model
learned from flight data that captures the behaviour of proprietary stabilization control algorithms.

A. Co-evolutionary Programming Approach


The fundamental idea behind co-evolutionary programming is to iteratively refine a set of proposal functions based
on fitness. Each proposal is modelled by an algebraic tree,
where the inner nodes represent one of four mathematical
operators { +, -, x , ~ } and the leaf nodes represent either the
input variables or constant parameters. The reader is referred
to De Nardi and Holland [14] for further details.
An Acscending Technologies Hummingbird quadrotor was
flown indoors and fitted with reflective markers, so that
a motion capture system was able to track its state with
millimeter precision. Co-evolutionary programming was then
used to find the dynamic equations describing the relationship
between the input control (roll, pitch, yaw rate, throttle)
and state. These equations include a number of constant
parameters, which are unique to a platform. Non-linear least
squares optimization was then used to fit these constant
parameters to the Ascending Technologies Pelican shown in
Fig. 2 using data from the experiments shown in Fig. 3.

(a) Static experimental setup

(b) Dynamic experimental setup

Figure 3. The static experimental set up (left) was used to obtain data for
calibrating the thrust model, while the dynamic experimental setup (right)
was used to obtain data for calibrating the rotation and translation models.

B. Second-Order Dynamic Model


Let X, = [x,y,z,,B,'if;,u,v,w,p,q,r,Fth]T denote the
state of the platform at time t, where F th is the thrust force,
and the first four element triples correspond to position, Euler
orientation, velocity and angular velocity. Control is accepted
as the vector Ilt = [Upt; Uri; Uth; Uya; Vb], comprising of
desired pitch Upt and roll Uri angles, throttle Uth, yaw angular
velocity u ya and current battery voltage Vb.
The platform dynamics are captured by a Markovian
process X = f(X, Il), which propagates the state forward in
time using the control vector and state. The state update is
therefore separable into the thrust, rotational and translational
update equations, shown in Eqn 1-3. Note that the position
and angle update equations are simply velocity and angular
velocity, rotated into the navigation frame and integrated with
respect to time, and so are not explicitly listed.

Fth

fthr (X, Il)

(1)

[p,q,r]T

frat (X, Il)

(2)

ftra (X, Il)

(3)

' , ']T
[u,v,w

1) Thrust: The relationship between throttle and thrust


force F T is shown as Eqn 4, and agrees with results in
the literature stating proportionality to the square of blade
speed [15], [13]. However, as Eqn 5 shows, this force
cannot exceed a maximum amount F M supported by the
battery. C thO, C th1 , C th2 , C tbO and C tb1 are platformspecific constants. For the sake of notational convenience,
the difference between the current and old thrust force is
denoted F D = min (FT, F M) - F th.

3383

C thO + C th1Uth + C th2UZh


c.: + CVb1Vb

(4)
(5)

A 100s dynamic test was carried out to validate the


rotational model. The same control control input (roll, pitch,
yaw rate) was applied to a hardware platform in a windless
indoor environment, and a simulated platform. The resulting
orientation (roll, pitch, yaw) was derived from fusing inertial
and magnetic data on-board the platform, and compared to
the predicted quantities (see Fig. 5a).
3) Translational : The relationship ftra between control
and translational velocity is given by Eqn 7. The first
two terms represent a body to navigation coordinate frame
change, while m is the quadrotor mass, and Kuv and Kw
are constant parameters specific to a platform representing
T
aerodynamic drag. The wind strength [wu , wv , ww ] is added
post-hoc as a perturbation to linear velocity, and is simply
the constant wind field plus platform-specific turbulence (see
Sec. III-A), rotated into the body frame.
Figure 4. Proof of match between measured (black), simulated (red) and
residual (blue) thrust over a 40 second period, for the same throttle control
input.

For small absolute (less than a threshold umin ) or relative


throttle values, thrust changes approximate instantaneously.
In all other cases it changes at a constant rate 1 , as a result of
inertia and drag. Combining this behaviour leads to the nonlinear relationship fthr in Eqn 6, where T is the time step,
and 0 and 1 are platform-specific constants. The literature
shows also that thrust depends not only on throttle control,
but also on the freestream velocity [13] and blade bending
[16]. Including this factors will likely lead to improved
performance, but experimental validation is challenging as
freestream and thrust force are difficult to measure in flight.

0 FD T 1 uth < umin

FD T 1
|FD | < 1 T
(6)
F th =

1
D <0

1
FD 0

Fig. 4 shows the proof of match for the thrust model,


by comparing the output thrust from the simulator and real
platform, as a function of the same input throttle control.
2) Rotational : The relationship frot between the control
and rotational velocity is given by Eqn 7. In general the
propellers are controlled to limit maximum rotational pitch
pm and roll qm , so as not to breach the sensing or control
envelope. The constants Kpq0 , Kpq1 and Kpq2 dampen the
difference between the current and desired orientation in
order to keep the control smooth. The roll and pitch models
are equivalent because of rotational symmetry, while the
model for yaw requires only two dampening constants Kr0
and Kr1 because its is specified as a rate, and not an angle.

Kpq1 (Kpq0 url ) + pKpq2 |p| < pm
p =
0
|p| pm

Kpq1 (Kpq0 upt ) + qKpq2 |q| < qm
q =
0
|q| qm
r = Kr0 uya + rKr1
(7)

qw + rv + gu + Kuv (u wu )
ru + pw + gv + Kuv (v wv )
Fth
w = pv + qu + gw
+ Kw (w ww ) (8)
m
Fig. 5b shows the proof of match for the translational
dynamic model. The platform was flown for 100s, and its
position was compared to post-processed GPS measurements.
The large translational residual error may be explained by the
fact that noisy accelerations are integrated twice with respect
to time through a noisy orientation to update position.
u
v

=
=

III. E NVIRONMENTAL AND S ENSOR M ODELS


Error is simulated as either a zero-mean Gaussian or
Ornstein-Uhlenbeck process. The latter process is stationary, Gaussian and Markovian, making it an analyticallyconvenient model for short-term time-correlated error. Given
a standard deviation , mean reversion rate , and unit White
noise process Wt , the zero-mean Ornstein-Uhlenbeck process
is defined by the stochastic differential equation
dt = t dt + dWt .

(9)

If white noise perturbations are denoted by wt N (0, 1),


then sample paths may be generated recursively by
p
(10)
t = e|ts| s + 1 e2|ts| wt .
If the starting state 0 is unknown, then 0 N (0, 2 ).
The expected value and covariance are given by
E(t |0 = c)

Cov(s , t |0 = c)

cet
(11)

2  |ts|
e
e|t+s| .(12)
2

A. Wind
Deterministic wind models require significant computational resources, and so the simulator adopts a simpler model
that is comprised of a mean field plus stochastic turbulence.
Both components are defined in a right hand coordinate
system, where the Z axis is parallel to gravity and the Y

3384

(a) Difference between measured angle (black) and predicted angle (red)
over a 100 second period, for roll (top), pitch (middle) and yaw (bottom).

(b) Difference between measured position (black) and predicted position


(red) over a 100 second period, for X (top), Y (middle) and Z (bottom).

Figure 5. Proof of match for the rotational (left) and translational (right) models.

axis points in the direction of the mean wind field. Take note
that the wind vectors must therefore be rotated into the body
frame to be used in the dynamic model.
1) Mean Field: A mean wind field is described by a
magnitude u20 at 20 feet from the ground. As a result of
wind shear, the resultant wind speed depends logarithmically
on altitude. The navigation-frame wind vector wc at some
altitude pz can be modelled [17] by Eqn 13.
h
iT
z /0.15)
0
0
wc = u20 log(p
(13)
log(20/0.15)

2) Turbulence: At altitudes lower than 1000 feet, which


is well above the ceiling of a quadrotor, turbulence can
be modelled as a stochastic process defined by its velocity
spectra [17]. The turbulence field is assumed to be frozen
in time and space, with the implication that the response of a
flight platform results from its motion relative to the turbulent
field. Under this assumption, turbulence can be modelled as a
one-dimensional field that involves just the three orthogonal
velocity components taken at the platforms centre of gravity.
The Dryden spectrum is a stochastic process that models
turbulence based on the platforms altitude pz and net
velocity V , the mean wind magnitude u20 . These are used
to calculate the turbulence scale lengths Lu , Lv and Lw and
intensities u , v and w as per Eqn 14. Lower altitudes experience turbulence of shorter lengths and greater intensities.
Lw

Lv = Lu

pz

pz
(0.177 + 0.000823h)1.2
w = 0.1u20
u
v
1
=
=
(14)
w
w
(0.177 + 0.000823h)0.4
The Dryden spectrum can be implemented [18] as the filter
in Eqn 15, where V is the airspeed of the platform, T is

the discretization period, and u , v and w are white noise


2
perturbations with variances u2 , v2 and w
.
q



T
t+T
1 VLTu wut + 2V
u
L
u
wu


q


T
wvt+T =

1 VLTv wvt + 2V
(15)
v
Lv

q


t+T
ww
2V
T
VT
t
1 Lw w w
+

Lw w
The simulators turbulence model was experimentally
validated against the Dryden implementation in Matlabs
Aerospace Toolbox. Fig. 6 shows 200s of sample turbulence,
with a mean intensity and bearing of 2 feet per second and
30 degrees respectively. The graph on the left shows that the
intensity of both processes decay similarly with frequency,
while the graph on the right shows that the individual wind
direction components exhibit similar statistical properties.
B. Global Navigation Satellite System
Global navigation satellite system (GNSS) position error
depends on the total number, configuration and accuracy of
the satellite vehicle positions, the dielectric constants of the
ionosphere and troposphere, and noise at the receiver. GNSS
satellite vehicle orbits change periodically and are difficult
to simulate, so recordings of real Global Positioning System
(GPS) orbits are used in the simulator. This data is freelyavailable from the International GPS Service (IGS)2 .
Modelling GNSS error is simplified by working with
measured distances to satellites (pseudo-ranges) rather than
the position solution. Each satellite has time-correlated error,
as a result of the ionosphere, troposphere, ephemeris and
clock drift. These errors are bundled into a single zero-mean
Ornstein-Uhlenbeck process si , specific to satellite i.

3385

2 GPS

Orbit file location: http://igscb.jpl.nasa.gov/components/prods_cb.html

(a) Power Spectral Density of Turbulence

(b) Wind turbulence along each axis

Figure 6. Proof of match between the Dryden model in Matlab (black) and in QRSim (red).

A pseudo-range measurement i,r of the distance between


a given satellite i and receiver r may be written as Eqn
16. Here, di,r is the true distance between the receiver and
satellite, si is the satellite-dependent error term, Tr is a fixed
clock delay error, and r is a white noise term expressing the
accuracy to which the C/A code can be tracked.
i,r = di,r + si Tr + r

(16)

A navigation solution pr for some receiver r is calculated


by solving the minimization problem in Eqn 17 using ordinary least squares. The solver itself uses the pseudo-ranges
to a subset of satellites S, which are collectively known as
the satellites in view (SV). In practice the SV is difficult to
calculate, so it is specified manually in the simulator.
X
2
(kpr pi k2 i,r )
(17)
pr = arg min
pr

iS

Fig. 7 illustrates the performance of the GPS error model,


by comparing its statistical properties to real data. To mitigate
multipath effects, a quadrotor was placed stationary on the
ground outdoors for 350s. The actual and simulated error
random walks were recorded for the same date, location,
satellite positions and SV. Fig. 7a compares a synthetic
and real GPS random walk, and the corresponding Allan
Deviation plot for the data shown in Fig. 7b illustrates that
the two data sets share statistically similar noise properties.
C. Barometric Altitude
Barometric pressure sensors recover altitude from pressure
measurements, using the fact that atmospheric pressure is
inversely proportional to height above sea level. Such sensors
exhibit decimeter-level resolution, making them more accurate than GNSS for estimating altitude. However, the main
drawback of a barometric altimeter is that the air pressure

varies slowly with time as a result of meteorological effects


[10]. The simulator models this error by a single zero-mean
Ornstein-Uhlenbeck process b for each flight platform.
D. Orientation
Orientation noise cannot be modelled from first principles
when the data fusion algorithm is unknown. However, orientation is typically estimated by integrating angular velocity
measurements with respect to time. The estimate is periodically corrected against a the gravitational and geomagnetic
vectors. This implies that time-correlated orientation error
arises from uncorrected gyroscope bias errors. Further timecorrelation may be introduced during high acceleration maneuvers due to a simplified model adopted by the filter. To
replicate this behaviour roll, pitch and yaw are perturbed by
the zero-mean OrnsteinUhlenbeck processes , and .
IV. A PPLIED E XAMPLE : R ANGE - BASED SLAM
Each simulation component was either implemented from
a well-known model or validated experimentally from real
measurements. A complete evaluation of the simulator requires a mechanism for measuring the platform position and
wind field in-flight, both of which are difficult to achieve
in-flight. This section will therefore focus on an example
application illustrating the usefulness of the simulator.
The Simultaneous Localization and Mapping (SLAM)
problem describes the joint estimation of static landmark
positions and a mobile vehicles trajectory, using measurements. The full, or offline, variant of SLAM optimizes over
all measurements, rather than filtering them recursively. State
of the art approaches in offline SLAM use non-linear least
squares optimization, and exploit the block structure of the
problem to improve computational performance. However,
one of the fundamental assumptions made by this approach

3386

(a) Position error random walk.

(b) Allan deviation of position error.

Figure 7. Proof of match between real (black) and synthetic (red) GPS errors, for the same date, location and satellites in view.

is that the measurement noise is white. However, in real


world problems this is seldom the case, but the approach
is nonetheless widely used. The objective of this section is
to qualitatively demonstrate the usefulness of the quadrotor
simulator through an investigation of the effect of coloured
GPS noise on offline, range-only SLAM.
To achieve this, two simulations were carried out. In
both simulations one quadrotor collected range measurements
to ten landmarks, which were distributed randomly over a
40 40 meter search region. The quadrotor moved between
randomly chosen waypoints drawn uniformly from the region, at an altitude of 20m. Each measurement was perturbed
by white noise with a standard deviation of 5m. The sensing
range was a sphere with radius 25m which, as a result of the
quadrotors altitude, results in a 15m sensing range on the
ground plane. In both simulations the only source of error in
the state estimate originated GPS noise. In the first coloured
noise experiment the GPS pseudoranges were perturbed by
errors with a standard deviation and time constant of 0.003m
and 1.005s, and a white receiver noise error with a standard
deviation of 0.002m. The second white noise simulation
used a truthful position estimate perturbed by white noise.
Both simulations used the same landmarks and sequence of
waypoints, and the latitude and longitude were perturbed in
the white noise simulation with a standard deviation equal
to what was measured in the coloured noise simulation.
x0
xinit

xt

xt+1

dt,i

ai

dt+1,i

xT

The offline range-based SLAM problem can be expressed


as the graphical model in Fig. 8. Unknown state values
are shown as circles, while measurements are shown as
rectangles. Control t (comprising of roll, pitch, yaw rate,
throttle) relates two sequential states xt and xt+1 through
the second order dynamic model function described in Sec
II. A range measurement is a distance d in meters between
of
the position ai of some anchor i, and the position xpos
t
the quadrotor at some time t. A prior measurement is a
constraint placed on the initial state of the quadrotor.
For both of the simulations a graph was built using
the control, range estimates and prior, and a non-linear
least squares solution was obtained over 10 iterations using
CERES - SOLVER3 . The initial trajectory estimate was taken
as the estimate produced by the simulator, i.e. corrupted by
either white or coloured noise, depending on the simulation.
The same set of random coordinates were used for initial
landmark positions in both simulations.
The two resulting trajectories and anchor points are shown
in Fig. 9. It is evident that time-correlated position error
affects the accuracy to which one is able to localize and
map using a standard SLAM approach the trajectory and
anchor points lie further from the truth in the case of timecorrelated position error. The underlying reason for this is
that time-correlated position error manifests as control error.
The result being that measurements relating sequential states
violate the white noise assumption made by the least-squares
solver, which leads to a sub-optimal solution. The Allan
Deviation plot in Fig. 10 provides a graphical summary of
the position error; in the white noise experiment noise can be
reduced by successive averaging of measurements, whereas
the same does not apply to time-correlated GPS error.

Figure 8. The offline, range-based SLAM problem expressed as a graph.


Circles represent unknown states, while rectangles represent observations.

3387

3 Available:

https://code.google.com/p/ceres-solver

(a) Result

Figure 10. Position Allan Deviation for both SLAM experiments.

(b) Allan Deviation


Figure 9. Results of Gaussian (red) and coloured (blue) SLAM experiments.

V. C ONCLUSION AND F UTURE W ORK


This paper presents a quadrotor simulator featuring a
generic dynamic model and comprehensive environment and
sensor error models, implemented as a standalone MATLAB
toolbox. The usefulness of the simulator is demonstrated
through an experimental analysis of the effect of coloured
GPS noise on offline, range-only SLAM. The simulator
has several drawbacks, which are currently being addressed.
Firstly, the performance of the visualisation engine degrades
as more platforms are added. Secondly, the simulator is
currently restricted to quadrotor platforms, and the inclusion
other vehicles would be beneficial. These issues will be
resolved by moving to a ROS/Gazebo environment.
R EFERENCES
[1] R. Mahony, V. Kumar, and P. Corke, Multirotor aerial vehicles:
Modeling, estimation, and control of quadrotor, IEEE Robotics &
Automation Magazine, pp. 2032, Sept. 2012.
[2] A. Visser and N. Dijkshoorn, Closing the gap between simulation and
reality in the sensor and motion models of an autonomous AR. Drone,
in Proceedings of the International Micro Air Vehicle Conference and
Flight Competition (IMAV11), (t Harde, the Netherlands), 2011.

[3] N. Koenig and a. Howard, Design and use paradigms for gazebo,
an open-source multi-robot simulator, 2004 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS) (IEEE Cat.
No.04CH37566), vol. 3, pp. 21492154, 2004.
[4] S. L. Waslander and C. Wang, Wind Disturbance Estimation and Rejection for Quadrotor Position Control, in AIAA Infotech@Aerospace
and Unmanned Unlimited Conference, (Seattle, WA), Apr. 2009.
[5] J. Lchele, A. Franchi, H. H. Blthoff, and P. Robuffo Giordano,
Simulation, Modeling, and Programming for Autonomous Robots,
vol. 7628 of Lecture Notes in Computer Science. Berlin, Heidelberg:
Springer Berlin Heidelberg, Nov. 2012.
[6] J. Meyer, A. Sendobry, S. Kohlbrecher, and U. Klingauf, Comprehensive Simulation of Quadrotor UAVs Using ROS and Gazebo, Lecture
Notes in Computer Science, vol. 7628, pp. 400411, 2012.
[7] B. Y. N. Michael, D. Mellinger, Q. Lindsey, and V. Kumar, Experimental Evaluation of Multirobot Aerial Control Algorithms, IEEE
Robotics & Automation Magazine, pp. 5665, Sept. 2010.
[8] M. Muller, S. Lupashin, and R. DAndrea, Quadrocopter ball juggling, 2011 IEEE/RSJ International Conference on Intelligent Robots
and Systems, pp. 51135120, Sept. 2011.
[9] A. Kushleyev, D. Mellinger, C. Powers, and V. Kumar, Towards
a swarm of agile micro quadrotors, Autonomous Robots, vol. 35,
pp. 287300, July 2013.
[10] J. Seo, J. G. Lee, and C. G. Park, Bias suppression of GPS measurement in inertial navigation system vertical channel, in Proceedings of
the Position Location and Navigation Symposium, 2004. PLANS 2004,
pp. 143147, Apr. 2004.
[11] S. Bouabdallah, Design and control of quadrotors with application to
autonomous flying. PhD thesis, EPFL, 2007.
[12] C. Balas, Modelling and Linear Control of a Quadrotor, Masters
thesis, School of Engineering, Cranfield University, 2007.
[13] G. M. Hoffmann, H. Huang, S. L. Waslander, and C. J. Tomlin,
Quadrotor Helicopter Flight Dynamics and Control: Theory and
Experiment, in Proceedings of the AIAA Guidance, Navigation and
Control Conference (GNC), 2007.
[14] R. De Nardi and O. Holland, Coevolutionary modelling of a miniature
rotorcraft, in 10th International Conference on Intelligent Autonomous
Systems (IAS10), (Baden Baden, Germany), pp. 364 373, 2008.
[15] R. De Nardi, Automatic Design of Controllers for Miniature Vehicles
through Automatic Modelling. PhD thesis, University of Essex, 2010.
[16] P.-J. Bristeau, P. Martin, E. Salan, and N. Petit, The Role of Propeller
Aerodynamics in the Model of a Quadrotor UAV, in European Control
Conference 2009, pp. 683688, Aug. 2009.
[17] Military Spec., Flying Qualities of Piloted Airplanes. MIL-F-8785C.
[18] S. Gage, Creating a Unified Graphical Wind Turbulence Model
from Multiple Specifications Table 1 : Dryden Longitudinal Spectra,
Simulation, no. August, 2003.

3388