World Academy of Science, Engineering and Technology

Vol:7 2013-05-23

Modeling and Control of a Quadrotor UAV with

Aerodynamic Concepts
Wei Dong, Guo-Ying Gu, Xiangyang Zhu, Han Ding

Abstract—This paper presents preliminary results on modeling simulation. In this paper, we aim to obtain this kind of
and control of a quadrotor UAV. With aerodynamic concepts, a model for a special case, which is under the assumptions
mathematical model is firstly proposed to describe the dynamics that the quadrotor is well controlled at a fixed height and the
of the quadrotor UAV. Parameters of this model are identified by
experiments with Matlab Identify Toolbox. A group of PID con- angle of attack is small. We focus on studying the dynamics
trollers are then designed based on the developed model. To verify of rate loop and speed loop. By the linear estimation and
the developed model and controllers, simulations and experiments for identification, a parameterized transfer function for a selected
altitude control, position control and trajectory tracking are carried quadrotor platform is obtained. A group of PID controllers
out. The results show that the quadrotor UAV well follows the are then designed by the root locus analysis and Ziegler-
referenced commands, which clearly demonstrates the effectiveness
Nichols tunning rules based on the developed model. Finally,
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

of the proposed approach.

simulations and experiments are carried out to verify the model
Keywords—Quadrotor UAV, Modeling, Control, Aerodynamics, and the control strategy.
System Identification.
The remainder of this paper is organized as follows. Section
II gives a brief introduction to the experimental platform. The
I. I NTRODUCTION model is derived in Section III, and the controller is designed
in Section IV. Then the results are shown in Section V, and
W ITH the development of micro-electronic technologies,
the quadrotor UAV draws a lot of attention in recent
years due to its advantages of compact size, agility and
Section VI concludes this work.

autonomous flight. A mature quadrotor system can be used II. Q UADROTOR P LATFORM
for reconnaissance, rescue, photography and works that are A so called Hummingbird quadrotor is chosen from As-
dangerous or space limited for human beings [1], [2]. cending Technologies, GmbH, which has a mature hardware
Many researchers involve themselves into this area in recent system and sufficient software packages. With SDK, control
years. As in [3]–[5], Samir Bouabdallah et al. carried out a algorithms can be easily simulated in Matlab Simulink and
series of researches on the control system of the quadrotor converted to embedded codes with no gap. The two level
UAV. In their works, the rigid body dynamics of a quadrotor processors structure and safe landing mode can protect the
UAV and several effects of aerodynamics were studied. They quadrotor from damages in most cases. For these reasons,
implemented several controllers, including the PID controller, many researchers choose it as their experimental platform [7],
backstepping controller, sliding-mode controller, and so on. [9], [10].
By comparison, a so called integral backstepping method was This platform is shown in Fig. 1. The tip-to-tip wingspan
proposed [5]. The backstepping controller and sliding-mode is 50cm and the height is 8cm. It weighs 500g with a
controller have great influences on subsequent researchers. E. capability of additional 200g payload. The attitude is sensed
I. Pounds et al. presented a delicate method to identify the by three gyroscopes along with a 3D magnetometer. And three
rigid body dynamics of the quadrotor UAV and an attitude accelerometers are equipped in the system to measure the
control algorithm was adopted by root locus analysis [6]. In acceleration in body fixed coordinates. The height is sensed by
[7], a PD controller was presented by V. Kumar et al., and a pressure sensor. The speed and position information come
performs well in their robotic team [8]. from a GPS unit. All of these characteristics provide enough
From the brief review, although the rigid body dynamics and hardware basis for outdoor experiments.
several aerodynamic effects have been analyzed by Bouabdal-
lah et al. [3], [5], a complete dynamic model, especially that III. M ODELING
can be adopted for position control, is still missing. In order to derive a complete dynamic model, the rigid body
For this reason, to develop a dynamic model that can fully dynamics, motor dynamics and the effects of aerodynamics are
describe the quadrotor UAV is necessary. Such model can studied in this section.
greatly help researchers to design controllers in an analytical
way and predict the behaviors of the quadrotor UAV via
A. Rigid body dynamics
Wei Dong, Guo-Ying Gu, Xiangyang Zhu and Han Ding are with Rigid body dynamics of the quadrotor UAV governs the
the State Key Laboratory of Mechanical System and Vibration, School response of attitude control. The expressions are derived in
of Mechanical Engineering, Shanghai Jiao Tong University, Shang-
hai 200240, China; email:[email protected], [email protected], two coordinate systems: an inertial coordinates and a body
[email protected], [email protected] fixed coordinates.

377
 World Academy of Science, Engineering and Technology
Vol:7 2013-05-23

F4

Z
2 F3 M4
1
YB
ZB F1
M3 XB
F2 M1
X L
Y mg
r
3 M2 ZI
4 YI
XI
Fig. 1. Hummingbird quadrotor
Fig. 3. Force diagram for quadrotor

z z Z z
Z
Z
Y θ
ϕ
Y dynamics is governed by
ϕ Y
ψ ⎧ ⎡ ⎤ ⎡ ⎤
0 0
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

y y y ⎪
⎪
x
X
x
θ x ψ
⎪
⎪
⎪ mr̈ = R ⎣ 0 ⎦ − ⎣ 0 ⎦ − q̇ × mṙ
X X
⎪
U1 mg
⎨
⎡ ⎤ (3)
⎪ U 3
Fig. 2. Attitude definition ⎪
I q̈ = ⎣U2 ⎦ − q̇ × I q̇
⎪
⎪
⎪
⎪
U4
⎩

The body fixed coordinates is defined as follows. As indi- where

cated in Fig. 1, the lever marked with orange strip is chosen
m mass of the quadrotor g local gravity constant
as the Y axis, and the perpendicular lever is the X axis. Then r position in inertial frame q attitude in body fixed frame
the Z axis is defined by the right hand rule. I rotary inertia
For inertial coordinates, the point where the quadrotor starts
its flight is set as the origin, and an east-north-up orthogonal Since the rotary inertia is small and the quadrotor UAV is
coordinate system is established by the right hand rule. symmetric, an estimation for (3) can be expressed in the form
With attitude angles defined as in Fig. 2, the transformation of (4), which is also adopted by other researchers [3], [5], [7]
matrix from inertial coordinates to body fixed coordinates is
ẍ = U1 (sin ψ sin φ + cos ψ sin θ cos φ)/m
⎧
⎪
ÿ = U1 (sin ψ sin θ cos φ − cos ψ sin φ)/m
⎪
⎪
⎡ ⎤ ⎪
cψcθ cψsθsφ − sψcφ cψsθcφ + sψsφ
⎪
⎨ z̈ = U cos φ cos θ/m − g
⎪
1
R(φ, θ, ψ) = ⎣sψcθ sψsθsφ + cψcφ sψsθcφ − sφcψ ⎦ φ̈ = U2 /Ixx
(4)
cθsφ cθcφ
⎪
−sθ ⎪
θ̈ = U3 /Iyy
⎪
⎪
(1) ⎪
⎪
ψ̈ = U4 /Izz
⎩
where s stands for sin, c stands for cos, and φ, θ, ψ represent
attitude angles of roll, pitch, yaw respectively. where Ixx , Iyy , Izz are rotary inertia around X, Y, Z axis
In the body fixed coordinates, the direct inputs are RPM respectively.
(revolutions per minute) commands for the motors. The resul-
tant outputs are Z directional thrusts in body fixed coordinates.
However, the concerned outputs are attitude and position. To B. Motor Dynamics
eliminate this gap, four control variables are defined as The dynamics of DC motor is generally described as [11],
[12]
di

Li dt + Ri + ke ωm = u
⎧
U1 = F1 + F2 + F3 + F4 (5)
dωm
⎪
J dt = τ − τd
⎪
U2 = (F3 − F1 )L
⎨
(2)
U 3 = (F2 − F4 )L
⎪
⎪ where
U4 = M1 + M3 − M2 − M4
⎩
Li : coefficient of inductance i: armature current
R: armature resistance ke : back emf constant
where F1 , F2 , F3 and F4 are thrusts, M1 , M2 , M3 and M4 ωm : speed of motor u: armature voltage
are momentums, and L is the lever length as illustrated in Fig. J: inertia of motor τ: torque of motor
3. The subscripts correspond to the ordinal numbers in Fig. 1. τd : load

According to the Newton-Euler formalism, the rigid body

378
 World Academy of Science, Engineering and Technology
Vol:7 2013-05-23

TABLE I
C. Effects of Aerodynamics I DENTIFICATION R ESULTS FOR C( S )
Two main effects are taken into consideration. One concerns
Item Ka τm Experiment Time(s)
how thrust is generated while the other deals with the drag
22.75 0.001
force. 1 144
18.00 0.001
The thrust T produced by each motor is calculated as 2
13.05 0.001
122
17.61 0.001
2 17.53 0.001
T = ρCT Aωm R2 (6) 3
18.29 0.001
72
15.48 0.001
4 62
where 15.48 0.001
25.58 0.001
5 65
CT : thrust coefficient ρ: air density 16.49 0.001
A: rotor disk area R: blade radius 13.71 0.001
6 72
32.30 0.001

According to (4) and (6), φ̈ and θ̈ are related to motors’

speeds in the form of A. Attitude and Position Control
1) System identification: The block diagram of this system
φ̈ = ρCT AR2 (ω32 − ω12 ) ≈ 2ρCT AR2 ω1 δωφ

(7) is shown in Fig. 4, which consists of a rate loop(R), an attitude

θ̈ = ρCT AR2 (ω22 − ω42 ) ≈ 2ρCT AR2 ω4 δωθ loop(A), a speed loop(S) and a position loop(P).
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

where ω1 , ω2 , ω3 , ω4 are rotary speeds of four rotors, ω3 = The identification of those subsystems is processed as
ω1 + δωφ , and ω2 = ω4 + δωθ . follows.
Rate Loop According to (7), with the consideration that the
Considering the flight at a fixed height, the aerial drag force
changes of ω1 and ω4 are small considering the large values
is expressed as [13]
of themselves in practice, φ̈ and θ̈ are governed by δωφ and
1 δωθ with linear relationship respectively. And from equation
D= CD ρv̄ 2 S (8)
2 (5), the motor is a first order system. Thus, the overall transfer
function of C is a first order system which is expressed as
where
Ka
D: drag force CD : drag force coefficient C(s) = (11)
v̄: speed of the quadrotor S: effective drag area
τm s + 1
where Ka stands for the overall gain of this subsystem and
τm represents the delay. To identify the parameters in (11),
In order to ensure that the quadrotor UAV flies in a steady six experiments are carried out. With a preliminarily designed
speed, (6) should be equal to (8). Thus, one can obtain controller, the quadrotor flies with a time varied attitude.
1 References and responses for the rate loop are recorded in
ρCT Aωe2 R2 sin α = ρCD v̄ 2 (St sin α + Ss cos α) (9) 50Hz. The data are then processed by Matlab Identify Toolbox.
2
The identified results are shown in Table I. Thus, C(s) is
where estimated as (12) by taking mean values from the results
St : planar area from top view α: angle of attack 19
Ss : planar area from side view ωe : equivalent motor speed C(s) = (12)
0.001s + 1
Attitude Loop The open loop behavior is in the form
Furthermore, when a quadrotor flies at a fixed height, the of integral as shown in Fig. 4, which is also proven by
equivalent thrust (U1 ) itself is a function of the angle of attack experiments.
according to (4) and (6). Substituting those two equations into Speed Loop According to (10), even for a quadrotor flying
(9), the flying speed is a function of the angle of attack as at a fixed height, the relation between attitude and flying speed
 is nonlinear. In this work, we only consider a special case
v̄ = cos αρCD 2mg sin α
which is under the assumptions as follows:
  (St sin α+S s cos α)
(10)
2mg
≈ ρCD Ss − S 2 α St α 2 • The height of the quadrotor is well controlled
s
• The quadrotor flies at a fix height

From (10), when α is small, the side area Ss dominates the • The attitude angles are small

drag effect. However, the influence of St increases when α Thus, for a certain platform, (10) can be expanded by the
becomes larger. Taylor series at any nonzero point. Then a first order system
is chosen to give an estimation. Noticing the fact that time
delay always exists in the conversion from attitude to speed,
IV. C ONTROLLER DESIGN the transfer function of STF in Fig. 4 is proposed as
Based on the proposed dynamic models (4), (5) and (10), Ks
the controllers will be designed in this section. STF(s) = e−Ls (13)
τs s + 1

379
 World Academy of Science, Engineering and Technology
Vol:7 2013-05-23

Reference Position Control Speed Control Attitude Control Rate Control Body Dynamics

Angular acceleration
r +- P +- S +- A +- R C

1/s STF 1/s 1/s

To Position Attitude to Speed To Attitude To Angular speed

Fig. 4. structure of the control system

TABLE II
I DENTIFICATION R ESULTS FOR STF( S ) 10

Position(m)
Item Ks τs L Experiment Time(s)
19.61 2.69 0.47 0 Xreference
1 144
11.33 1.37 0.59 −5
Xresponse
6.81 0.91 0.79 Xsimulation
2 122
6.66 0.81 0.81 −10
0 5 10 15 20 25 30 35 40 45
6.13 0.63 0.84
3 72
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

15.50 2.12 0.57

10
15.50 1.92 0.58 Y
4 62 reference
7.47 0.77 0.81 Position(m)
5 Yresponse
13.63 1.84 0.65
5 65 Y
13.07 2.17 0.83 0 simulation

11.45 1.26 0.72

6 72 −5
12.72 2.15 0.70
−10
0 5 10 15 20 25 30 35 40 45
Time(s)

To identify the parameters, six experiments are also carried

out and results are shown in Table II. Fig. 5. Comparison between simulation and experiments for position control

Taking mean values from the results and substituting them

into (13), STF(s) is estimated as For rate and attitude loop, by the root locus analysis, an
11.7 −0.7s undamped and fast response system can be obtained when
STF(s) = e (14)
1.6s + 1 Krp = 10 and Kap = 12 are chosen.
For speed control, the Ziegler-Nichols(Z-N) [11] rules is
Position Loop The open loop behavior is in the form
adopted to tune the parameters. A proportional controller is
of integral as shown in Fig. 4, which is also proven by
implemented at the beginning. The critical gain Kcr = 0.4 and
experiments.
critical period Tcr = 2.5 are obtained. Then PID parameters
2) Control Approach: With its three-term functionality cov-
are selected as Kps = 0.24, Tis = 1.25 and Tds = 0.31.
ering treatment to both transient and steady-state responses,
For position loop, the Z-N technique is also adopted, and
the PID control provides the simplest and yet most effcient
Kpp = 0.7, Tip = 2.3, Tdp = 0.28 are set for the system.
solution in many applications. It has been successfully applied
3) Verification: In order to verify the analysis, a simulation
for control of the quadrotor [14]. In this work, PID controllers
and a position control experiment implemented with totally
are adopted for those four loops as follows
same controllers are carried out. Fig. 5 shows the results.
R = Krp Both in experiment and simulation, the controllers follow the
⎧
⎪
⎨ A=K references well, which shows the accuracy of the model and
⎪
ap
S = Kps (1 + Tis 1
+ Tds s) (15) the effectiveness of the controllers.
⎩ P = K (1 + 1s + T s)
⎪
⎪
pp Tip s dp
B. Altitude Control
where proportional controllers are implemented for R and A
Similar to the reported controllers in [4] [15], a sliding-
while S and P are adopted by PID controllers.
mode control algorithm is implemented here.
Then close loop transfer functions for rate loop(RC ), at-
A linear sliding mode surface is selected as
titude loop(AC ), speed loop(SC ) and position loop(PC ) are
given as s1 = c1 e1 + e2 = c1 (zr − z) + (z˙r − ż) (17)
⎧
⎪
s2 +1000s
RC (s) = s2 +1000s+19000K where zr stands for the reference of altitude, z is the actual
rp
⎪
⎨ AC (s) = Kap RC
⎪
⎪ response, and c1 is a positive number.
s+Kap RC A so called exponential reaching law [16] is adopted
11.7Kps (Tds s2 +s+1/Tis s)AC e−0.7s (16)
⎪ SC (s) = 1.6s2 +s+11.7K 2
ps (Tds s +s+1/Tis s)AC e
⎪
⎪ −0.7s
s˙1 = c1 (z˙r −ż)+(z¨r −z̈) = −ε1 sgn(s1 )−k1 s1 , ε1 > 0, k1 > 0
Kpp (Tdp s2 +s+1/Tip sSC
⎪
PC (s) = s2 +K
⎩
2
pp (Tdp s +s+1/Tip s)SC
(18)

380
 World Academy of Science, Engineering and Technology
Vol:7 2013-05-23

where ε1 is a small positive number. 15

Substituting equation (4) into equation (18), one can obtain 10
Xcmd
Xact

Position(m)
the control input for altitude control 5
m 0
U1 = [c1 (z˙r − ż)+ z¨r +ε1 sgn(s1 )+k1 s1 +g] (19)
cos φ cos θ −5

−10
0 50 100 150
V. E XPERIMENTS
All the experiments are carried out in an outdoor environ- 20
Ycmd
ment. Rate, attitude, speed, position and altitude are controlled 10 Y

Position(m)
act
by onboard processors running at a frequency of 1000Hz. 0
Trajectory information is sent from a laptop to the position
−10
controller at a frequency of 50Hz. The reference and response
for each controller are collected to the laptop via a wireless −20
0 50 100 150
router. Time(s)

Fig. 7. Results of position control

A. Altitude Control
The results of altitude control are shown in Fig. 6. The
subscript cmd stands for command, act stands for response, 10
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

and they keep these meanings through the following text. The 5
steady state error is within 0.5 meter, and the delay is within Speed(m/s)
0
7 seconds. Compared with reported controllers [4] [15], it is X’cmd
−5
good enough in the outdoor environment. ’
Xact
−10
0 50 100 150
5
10
4
5
Speed(m/s)

3
Height(m)

0
2
’
Ycmd
−5
1 Zcmd Y’act
0 Z −10
act 0 50 100 150
Time(s)
−1
0 20 40 60 80 100 120 140
Time(s)
Fig. 8. Results of speed control
Fig. 6. Results of altitude control

θcmd
0.5
B. Position Control θact
Angle(rad)

The position control is carried out at a fixed flying height of 0

3 meters. The quadrotor UAV firstly hovers at the position of
(0,0,3) for 60 seconds, then it follows several ramp references. −0.5
20 40 60 80 100 120 140
The results are shown in Fig. 7. It is satisfactory when
compared with the reported results [3], [4], [15]. The steady
1 φ
errors are within 0.5 meter, overshoots are within 15%, and cmd
φact
Angle(rad)

the delays are within 3 seconds. The corresponding results of 0.5

speed control are shown in Fig. 8, and the results of attitude 0

control are shown in Fig. 9. −0.5
In Fig. 8, approximately 1 second delay exists in the speed 20 40 60 80 100 120 140
control. This clearly demonstrates the delay term in (14). Time(s)

The maximum error of 1.4m/s appears at 106s with reference

of 5.7m/s in X direction. However, there are no significant Fig. 9. Results attitude control
influences on the position control at that point.
In Fig. 9, when commands are small(≤0.2rad), which
corresponds to the case of hovering with speed less than The maximum error of 0.1rad(20%) appears in attitude control
1m/s, the maximum error of around 0.1rad arises. This can when the quadrotor accelerates the speed up and the maximum
be explained as follows. On one hand, the controller lacks overshoot of 0.25rad(40%) appears in attitude control when the
of integral terms, so it is not sensitive to small errors. On quadrotor decelerates the speed to zero. However, there are no
the other hand, disturbances such as wind affect the results. overshoots in speed control at these points due to regulation

381
 World Academy of Science, Engineering and Technology
Vol:7 2013-05-23

of the speed controller as in Fig. 8.

15
cmd
act
C. Trajectory tracking 10
A rectangular trajectory tracking and a circular trajectory
tracking are performed. The flying height is 3 meters. Each 5
experiment takes 140 seconds. The results are shown in Fig.
10 and Fig. 11. The trajectory errors are within 0.5 meter.
0

Y
15
cmd −5
act
10

−10
5

−15
0 −15 −10 −5 0 5 10 15
Y

−5
Fig. 11. Circular trajectory tracking
International Science Index Vol:7, No:5, 2013 waset.org/Publication/7686

−10

[2] D. Mellinger, M. Shomin, and V. Kumar, “Control of quadrotors for

−15
−15 −10 −5 0 5 10 15 robust perching and landing,” in International Powered Lift Conference,
X October 5-7, 2010, 2010.
[3] S. Bouabdallah and R. Siegwart, “Full control of a quadrotor,” in Intel-
ligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ International
Fig. 10. Rectangular trajectory tracking Conference on. Ieee, 2007, pp. 153–158.
[4] S. Bouabdallah, R. Siegwart, S. Bouabdallah, and R. Siegwart, “Back-
stepping and sliding-mode techniques applied to an indoor micro quadro-
tor,” in Robotics and Automation, 2005. ICRA 2005. Proceedings of the
VI. C ONCLUSION 2005 IEEE International Conference on. Ieee, 2005, pp. 2247–2252.
[5] S. Bouabdallah, “Design and control of quadrotors with application to
In this paper, a complete dynamic model of the quadrotor autonomous flying,” Lausanne Polytechnic University, 2007.
UAV is proposed for a special case which is under the [6] P. Pounds, R. Mahony, and P. Corke, “Modelling and control of a quad-
rotor robot,” in Proceedings Australasian Conference on Robotics and
assumptions that the fight is well controlled at a fixed height Automation 2006. Australian Robotics and Automation Association
and the angle of attack is small. By linear estimation and Inc., 2006.
parameter identification, the transfer function is developed [7] N. Michael, D. Mellinger, Q. Lindsey, and V. Kumar, “The grasp
multiple micro-uav testbed,” Robotics & Automation Magazine, IEEE,
for an off-the-shelf quadrotor UAV. This makes it possible vol. 17, no. 3, pp. 56–65, 2010.
to design controllers in an analytical way. A group of PID [8] Q. Lindsey, D. Mellinger, and V. Kumar, “Construction of cubic struc-
controllers are selected and the parameters are tuned according tures with quadrotor teams,” Proc. Robotics: Science & Systems VII,
2011.
to the developed model. Simulations and outdoor experiments [9] R. Mahony, V. Kumar, and P. Corke, “Multirotor aerial vehicles:
show that the controllers follow the references very well with a Modeling, estimation, and control of quadrotor,” Robotics & Automation
maximum error of 0.5 meter in position control and maximum Magazine, IEEE, vol. 19, pp. 20 – 32, 2012.
[10] T. Tomic, K. Schmid, P. Lutz, A. Domel, M. Kassecker, E. Mair, I. Grixa,
overshoot of 15%. Since the angle of attack is not always F. Ruess, M. Suppa, and D. Burschka, “Toward a fully autonomous uav:
small, height is not precisely fixed and random disturbances Research platform for indoor and outdoor urban search and rescue,”
arise in the outdoor environment, experiments show that the Robotics & Automation Magazine, IEEE, vol. 19, pp. 57–68, 2012.
[11] K. Ogata and Y. Yang, Modern control engineering. Prentice Hall
model can be adopted in more general cases than that under Upper Saddle River, NJ, USA, 1990, vol. 4.
the assumptions. [12] S. Bouabdallah and R. Siegwart, “Design and control of a miniature
Although only the PID algorithm is selected for analysis, quadrotor,” Advances in Unmanned Aerial Vehicles, pp. 171–210, 2007.
[13] J. Leishman, Principles of helicopter aerodynamics. Cambridge
the approach is validated for any other controller. However, the University Press, 2006.
dynamic model is derived from a special case, so more general [14] A. Salih, M. Moghavvemi, H. Mohamed, and K. Gaeid, “Modelling
cases will be studied in future. Another limitation of this work and pid controller design for a quadrotor unmanned air vehicle,” in Au-
tomation Quality and Testing Robotics (AQTR), 2010 IEEE International
is that the current test is only applied to one quadrotor UAV, Conference on, vol. 1. IEEE, 2010, pp. 1–5.
and more researches will be taken for different platform in [15] R. Xu and U. Ozguner, “Sliding mode control of a quadrotor helicopter,”
order to generalize the results. Nonetheless, the success of in Decision and Control, 2006 45th IEEE Conference on. IEEE, 2006,
pp. 4957–4962.
current work will motivate us to move on. [16] C. Fallaha, M. Saad, H. Kanaan, and K. Al-Haddad, “Sliding-mode robot
control with exponential reaching law,” Industrial Electronics, IEEE
Transactions on, vol. 58, no. 2, pp. 600–610, 2011.
R EFERENCES
[1] V. Kumar and N. Michael, “Opportunities and challenges with au-
tonomous micro aerial vehicles,” in Int. Symp. on Robotics Research,
2011.

382