2013 IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS)

November 3-7, 2013. Tokyo, Japan

Aerial Manipulation Using a Quadrotor with a Two DOF Robotic Arm

Suseong Kim1 , Seungwon Choi2 and H. Jin Kim3

Abstract— This paper presents aerial manipulation using a

quadrotor with a two-DOF robot arm. By considering a quadro-
tor and robot arm as a combined system, the kinematic and
dynamic models are developed, and an adaptive sliding mode
controller is designed. With the controller, an autonomous flight
experiment is conducted including picking up and delivering
an object, which requires accurate control of a quadrotor and
robot arm. Overall result shows that the proposed approach
demonstrates satisfactory performance as a potential platform
which can be utilized in various applications such as inspection,
manipulation, or transportation in remote places.

I. INTRODUCTION Fig. 1: Aerial manipulation using a quadrotor with 2DOF

robotic arm
Aerial robots that can hover and move around in three-
dimensional Euclidean space are popular research platform.
Due to their superior mobility, much interest is given to
regulated directly because manipulation is achieved using a
utilize them for mobile manipulation such as inspection of
cable which cannot always drive the motion of the payload as
hard-to-reach structures or transportation in remote areas.
desired. The last cases are applicable to specialized missions
Previous research on aerial manipulation can be divided
such as wall inspection or applying normal force to a surface.
into three categories. The first approach is to install a gripper
To cope with these limitations, one alternative approach is
at the bottom of a UAV to hold a payload. In [2] and [3],
to equip an aerial vehicle with a robotic manipulator that
a quadrotor with a gripper is used for transporting blocks
can actively interact with the environment. For example,
and constructing structures. In [4], a team of quadrotors are
in [12], a testbed including four-DOF robot arms and a
employed for collaboratively picking up a relatively large
crane emulating an aerial robot is proposed. By combining
object using a similar gripper. The stability of a helicopter
the mobility of the aerial vehicle with the versatility of a
with a gripper was investigated in [5]. The second approach
robotic manipulator, the utility of mobile manipulation can
is to suspend a payload with cables. In [6], an adaptive
be maximized.
controller is presented to avoid swing excitation of a payload.
When employing the robotic manipulator, the dynamics of
Another approach for a similar problem can be found in [7],
the robotic manipulator is highly coupled with the aerial ve-
where the movement of a suspended payload is suppressed
hicle, which should be carefully considered in the controller
by trajectory optimization. In [8], specific attitude and posi-
design for the aerial vehicle. Also, an aerial robot needs to
tion of a payload is achieved using cables connected to three
tolerate the reaction forces from the interactions with the
quadrotors. In [9], the problem of carrying a payload under
object or external environment. These reaction forces may
uncertainties such as strong wind is studied using single and
affect the stability of an aerial vehicle significantly.
multiple helicopters. The other types of research is concerned
Very few reports exist on handling the aerial vehicle and
about interaction with existing structures, for example, for
the robotic arms as an integrated system. Kinematic and
contact inspection. Research has been conducted on utilizing
dynamic models of the quadrotor combined with a multi-
force a sensor [10] or a brush [11] as a manipulator. However,
DOF robot arm are derived using the Euler-Lagrangian
the above approaches have limitations for manipulation.
formalism in [13]. Based on that, simulation results using
In the first category using a gripper, payloads are rigidly
the Cartesian impedance control for the combined system
connected to the body of an UAV. Accordingly, not only the
is presented in [14]. In [16], effects of a manipulator on
attitude of the payload is restricted to the attitude of the UAV,
the quadrotor are simulated based on the dynamic model
but also the accessible range of the end effector is confined
which considers a quadrotor and robotic arms separately,
because of the UAV body and blades. In the second type
treating the arm as the disturbance to the quadrotor control
using cables, the movement of the payload cannot be always
loop. In [15], a quadrotor with light-weight manipulators
1 Suseong Kim and Seungwon Choi are with the Department of Mechan- are tested, although the movement of manipulators is not
ical and Aerospace Engineering, Seoul National University, Seoul, South explicitly considered during the design of the PID controller.
Korea suseongkim, cso103 at snu.ac.kr This paper presents the design and experimental results
2 H. Jin Kim is with the Department of Mechanical and Aerospace
Engineering, Seoul National University, Seoul, South Korea hjinkim of aerial manipulation using a quadrotor that has a two-
at snu.ac.kr DOF robotic manipulator attached. To utilize the combined

978-1-4673-6358-7/13/$31.00 ©2013 IEEE 4990

system, a dynamics model which considers a quadrotor The angular velocity of the quadrotor in the inertial and
and a robotic manipulator as a unified system is presented. body fixed frame are defined as ω = [ωx ωy ωz ] and
Also, an adaptive sliding mode controller is proposed to ω b = [ωxb ωyb ωzb ] respectively. ω and ω b are also related by
control the unified system. To verify the proposed approach, Rt . Also, the time derivative of the Euler angles Φ̇ can be
a hardware test result is presented, which involves picking mapped to ω by the transformation matrix T . The relations
up and delivering an object during autonomous flight. among Φ̇, ω and ω b are
This paper is organized as follows: In section II, the
kinematic and dynamic models of a quadrotor with a two- ω = Rt ω b (3)
DOF robotic manipulator attached are presented. Also, the ω = T Φ̇ (4)
adaptive sliding mode controller for the entire system is ωb = RtT T Φ̇ = QΦ̇ (5)
proposed. In section III, experimental setup and a mission
scenario are described. Section IV presents experimental Let pbi be the position of the center of mass of link i = 1, 2
results and section V contains concluding remarks. in the body-fixed frame Ob . Then, pi , which is the position
of the center of mass of the link i in the inertial frame OI ,
II. K INEMATIC AND DYNAMIC M ODEL
is related to pbi by
This section presents the kinematic and dynamic models
of the quadrotor combined with a two-DOF manipulator. pi = p + Rt pbi (6)
General version of deriving kinematics relations and dynam-
In addition, the translational and angular velocity of each
ics equations for the similar combined system can be found
manipulator link are related by the time derivative of the
in [13].
joint variables η̇ and the Jacobian matrix Jt ∈ R2×2 and
A. Kinematics for the Combined System Jr ∈ R2×2 respectively.
ṗbi = Jt η̇ (7)
ωib = Jr η̇ (8)
With equations (7) and (8), the translational and angular
velocity of each link in the inertial frame are computed as

ṗi = ṗ + Ṙt pbi + Rt p˙bi (9)

ṗi = bb Rt pbi + Rt Jt η̇
ṗ + ω (10)
ωi = ω + Rt Jr η̇ (11)
For simplicity, equations (2),(4),(9) and (10) are rewritten in
the following matrix form.
Fig. 2: Configuration of the coordinates for the combined ṗ = [I3×3 03×3 03×2 ]q̇ = Mt,b q̇ (12)
system.
ω = [03×3 T 03×2 ]q̇ = Mr,b q̇ (13)
b ∧
Fig. 2 shows the coordinate frames defined to derive p˙i = [I3×3 − (Rt pi ) T Rt Jt,i ]q̇ = Mt,i q̇ (14)
the kinematic model of the combined system. The inertial ωi = [03×3 T Rt Jr,i ]q̇ = Mr,i q̇ (15)
coordinate frame is denoted as OI , and Ob and Oi represent
the body coordinate frames fixed to the quadrotor and links. where the subscript · × · represents the size of I and 0. Also,
∧
The subscript i = 1, 2 denotes the link number. All body- is the operator that converts a vector into a skew-symmetric
fixed coordinate frames are located at the center of mass of matrix. With the above relations, the time derivative of the
their respective rigid bodies. The position of the center of generalized coordinate variable vector q̇ is easily mapped
mass of quadrotor in the inertial frame p = [x y z]T , the into the translational and angular velocities of the quadrotor
Euler angles of the quadrotor Φ = [φ θ ψ]T and the joint and each link in the inertial coordinate frame.
angles of the two-DOF manipulator η = [η1 η2 ]T are chosen
B. Dynamics Derivation for Combined System
as the generalized coordinates. The joint angles η1 and η2
are defined about the positive yb axis. The vector containing To derive the dynamics of the combined system, the
all the generalized coordinate variables is defined as Lagrange-D’Alembert equation is used.

q = [pT ΦT η T ]T (1) d ∂L ∂L
− = τ + τext (16)
dt ∂ q̇ ∂q
The time derivative of p represents the translational ve-
locity of the center of mass of the quadrotor, and the L=K−U (17)
translational velocity in the body-fixed frame is denoted as
where K and U are the total kinetic and potential energy of
ṗb . They are related by the rotation matrix Rt ∈ SO(3) as
the combined system and τ represents the generalized force.
ṗ = Rt ṗb (2) τext indicates external disturbance applied to the system.

4991
 The total kinetic energy is the sum of the kinetic energy Generalized torque related to the joint variable η can
of individual components, which are the quadrotor and two be actuated directly through joint actuators, such as servo
links. The total kinetic energy K and its components are as motors. However, the generalized torque corresponding to p
follows: and Φ need to be transformed to be usable as the quadrotor
2
X input. The actual quadrotor inputs are given in the PWM
K = Kb + Ki (18) signals to motors, and those signals are mixed to generate
i=1 thrust or torques. But for the description in this paper, we
1 T 1 can treat the input to the quadrotor as the following four
Kb = ṗ mb ṗ + Φ̇T T T Rt Ib RtT T Φ̇ (19)
2 2 values: F , the total thrust value of the four propellers of
1 T 1 the quadrotor, and τx , τy and τz , the torque values applied
Ki = ṗ mi ṗi + ωiT (Rt Ri )Ii (Rt Ri )T ωi (20)
2 i 2 to the quadrotor body coordinate in x, y and z direction,
where m is the mass and I is the inertia matrix. The respectively.
subscripts b and i indicate the corresponding values are with Later, in section II.C, an adaptive sliding mode controller
respect to Ob and Oi respectively. will be described to design the τ . Then the computed control
Likewise, the total potential energy is the sum of potential signal τ can be converted to the input values of the quadrotor
energy of each component. The total potential energy U and and joint actuators. Let τη ∈ R2×1 denote the joint torque
its components are described by the equations below: values of the manipulator, and the numbers in parenthesis
indicate the element in the corresponding matrix or vector.
2
X Then,
U = mb geT3 p + mi geT3 (p + Rt pbi ) (21)
 
i=1 F  −1  τ (3) 
where the first and last terms are the potential energies of  τx  R (3, 3) 0 0
  t
 
the quadrotor and link i, respectively. e3 is the unit vector  τy  = 
  0 Q−1 0   ...  (26)
[0 0 1]T and g is the gravity.  τz  0 0 I2×2 τ (8)
By substituting equations (17),(18) and (21) in equation τη
(16), the dynamics equation which includes all components
as one system can be derived as the following The other two elements of τ , i.e. τ (1) and τ (2), are modu-
lated to the desired φd and θd of a quadrotor as
M (q)q̈ + C(q, q̇)q̇ + G(q) = τ + τext (22)     
θd 1 cos ψ sin ψ τ (1)
where M (q) ∈ R8×8 is the inertia matrix which is positive = (27)
φd F sin ψ − cos ψ τ (2)
definite and symmetric. C(q, q̇) is the Coriolis matrix and
C(q, q̇)q̇ represents the Coriolis and centrifugal force terms. This relation is derived based on the small roll and pitch
In addition, G(q) includes gravity effects at each joint [1]. angle assumption.
The total kinetic energy can be calculated using the inertia
matrix M (q) as C. Adaptive Sliding Mode Controller
1
K = q̇ T M (q)q̇ (23) This subsection describes the control design of the com-
2
bined system (22). In this research, aerial manipulation
By substituting equations (12), (13), (14) and (15) into
includes motions involving both picking and releasing an
equation (23), the inertia matrix M (q) is computed as the
object during hovering flight. Position holding, one of the
following equation.
most important functions for accurate manipulation, can be
T T
M (q) = Mt,b mb Mt,b + Mr,b Rt Ib RtT Mr,b (24) significantly disturbed by a sudden quadrotor attitude change
2 due to the additional torque caused by grabbing or releasing
X
+ T
Mt,i T
mi Mt,i + Mr,i (Rt Ri )Ii (Rt Ri )Tt Mr,i an object. In other words, robustness of the controller is a
i=1 key issue for using the aerial vehicle to manipulate an object
as desired. Also, the controller should be able to handle
where Ri is the coordinate transformation matrix from link
problems such as battery drains, miscalculated mechanical
i of the manipulator to the quadrotor body-fixed coordinate.
properties, measurement bias or noise. To cope with these
The Coriolis matrix C(q, q̇) can be derived by calculating
uncertainties, the controller needs to be adaptive. In this
each element with the following relation [1].
paper, to control the combined system, an adaptive sliding
X8   mode controller is used.
1 ∂mkj ∂mki ∂mij
ckj = + − To design a controller for equation (22), the error e and
i=1
2 ∂qi ∂qj ∂qk
sliding surface s are defined as the following:
where mαβ indicates the element αβ of the inertia matrix
M (q). G(q) is calculated with the following partial derivative e = q − qd (28)
∂U s = ė − Λe (29)
G(q) = (25) q̇r = q̇ − s = q̇d − Λe (30)
∂q
4992
where λ is a positive diagonal matrix. With these definitions, The shape of each link is shown in fig. 3. The length of the
the proposed control law for τ is arm when it is fully stretched is 0.32 meter and the total
ˆ − As − Ksgn(s) weigh is about 0.370 kg. The ground station and robot arm
τ = M̂ q̈r + Ĉ q̇r + Ĝ + ∆ (31)
communicate via bluetooth. To grab a small object, a parallel
Here, M̂ , Ĉ and Ĝ represent the estimation of each matrix. gripper connected with an AX-12 servomotor is used [17].
A and K are positive gain matrices. Also, ∆ ˆ describes the Also, each servomotor has an encoder which measures joint
estimated uncertainty. By substituting the proposed control angles. Vicon [19], an indoor GPS system, is used to estimate
law to the dynamics equation (22), we can derive the
following relations:
M ṡ + Cs + As = ˆ − Ksgn(s) + τext (32)
−∆ + ∆
M ṡ + (C + A)s = ˜ − Ksgn(s) + τext
∆ (33)
where
∆ = −M̃ q̈r − C̃ q̇r − G̃ (34)
and ˜ indicates the estimation error of the corresponding
matrices and vectors. With these equations, stability analysis Fig. 3: Specification of each link of the robotic manipulator
can be performed as the following.
Proof. Let a Lyapunov candidate function be the position and attitude of the quadrotor. With this device,
1 1 ˜T ˜ the estimated position, attitude and their derivatives are
V = sT M s + ∆ ∆>0 (35) computed up to 100Hz. Measured quadrotor and joint states
2 2
with directional time derivative are used in the adaptive sliding mode controller. The desired
configuration qd is obtained by solving inverse kinematics
1 T ˜˙
˜T∆
V̇ = s Ṁ s + sT M ṡ + ∆ (36) with the known position of the object for grasping and
2 releasing.
= sT (−As + ∆ ˜ − Ksgn(s) + τext )
Overall hardware configuration is summarized in fig. 4.
+∆ ˜T∆ ˜˙ (37)
T T
= −s As − s (Ksgn(s) − τext )
+∆ ˜˙
˜ T (s + ∆) (38)
T T
= −s As − s (Ksgn(s) − τext ) (39)
< 0
In the derivation, skew symmetricity of Ṁ − 2C is used.
Negative definitiveness of equation (39) can be satisfied with
arbitrarily large matrices A and K.
˜˙ = 0. From this
In equation (38), it is assumed that s + ∆
relation, the following adaptation law can be derived.
ˆ˙ − ∆
∆ ˙ = −s (40) Fig. 4: Configuration of the overall system.

with the assumption that ∆ changes very slow compared

with the adaptation rate B. Mission Scenario
ˆ˙ = −s
∆ (41) To emphasize the capability of aerial manipulation with a
two-DOF robot manipulator over using an gripper or cable,
As a result, by integrating −s, estimation of the uncertainty an experimental scenario is devised.
ˆ can be computed.
∆ After taking off, the quadrotor moves to an object. It is
assumed that the position of the object is known in advance
III. E XPERIMENTAL S ETUP
(measured by Vicon beforehand in this case), which will
A. Hardware setup soon be relaxed using vision. Once the quadrotor arrives at
The quadrotor used in this paper is a Smart Xcopter [20] the designated position, the robotic arm moves to the position
whose arm length is 25cm and weight is approximately 0.9 of the object and grabs it.
kg. The carrying capacity of the quadrotor is about 0.45 kg. While grabbing the object, the quadrotor moves to the
To transmit the control signal to the quadrotor, Spectrum next place to reach where the object is laid down. When the
DX7 transmitters is connected to the ground station through quadrotor gets close enough to the designated location, the
Endurance R/C PCTx. The robotic arm is customized with a robotic arm manipulates the object to release it at the desired
Bioloid frame set and Dynamixel MX-28 servomotors [18]. position. Here, the releasing point is located inside of a shelf

4993
where only a two or higher DOF robot arm can reach. After
5
completing the above missions, the quadrotor goes back to

φ (deg)
the place where it took off and lands.
0
IV. R ESULTS (1) (2)
−5
The proposed aerial manipulation using the adaptive slid- 0 10 20 30 40 50 60
ing mode controller is validated by an experiment. The time (sec)
results are summarized in figs. 5-9. 5

θ (deg)
0

1
−5
0 10 20 30 40 50 60
x (m)

0 time (sec)
(1) (2) 5
−1

ψ (deg)
0 10 20 30 40 50 60
time (sec) 0

0.2 −5
0 10 20 30 40 50 60
y (m)

0 time (sec)
−0.2
−0.4
0 10 20 30 40 50 60 Fig. 6: Attitude history of the quadrotor. Solid-green line is
time (sec)
the desired attitude and dashed-black line is the measured
0.8
attitude. (1): pick-up phase, (2): release phase
z (m)

0.6
0.4
0.2
0 10 20 30 40 50 60 0.25
time (sec) 0.2
xb (m)

0.15 (1) (2)

0.1
Fig. 5: Position history of the quadrotor. Solid-green line is 0.05
0 10 20 30 40 50 60
the desired position and dashed-black line is the measured time (sec)
position. (1): pick-up phase, (2): release phase 0.05
0.1
zb (m)

The first mission for the combined system is picking up an 0.15

object. The object used in the experiment is an wood block of 0.2
0.25
demension (H)7.5 × (D)5 × (W )1.5 cm. Given that the gap
0 10 20 30 40 50 60
of the end effector is only 3.2 cm, it is clear that grabbing time (sec)
the wood block requires accurate control of the quadrotor
0
and robot arm concurrently. It is shown in figs. 5-7 that
zb (m)

0.1
quadrotor position and attitude, and manipulator’s position
0.2
follow the reference trajectory precisely during the pick-up
0.3
phase. Even though no information such as mass and moment −0.4 −0.2 0 0.2 0.4 0.6 0.8
of inertia about the wood block is given to the adaptive xb (m)
sliding mode controller, overall manipulation is successful
because of robustness of the sliding mode controller. During
Fig. 7: End effector position history in the quadrotor body
the pick-up phase, root-mean-square error of the quadrotor
fixed frame. The third figure is plotted for full flight duration.
position is only 2.08 cm. In fig. 8, a sequence of images
Solid-green line is the desired position and dashed-black line
taken in the pick-up task is shown.
is the measured positionm (1): pick-up phase, (2): release
The second mission is releasing the wood block inside
phase
a shelf. The size of the gap of shelf is about (H)16 ×
(W )22cm. To lay the wood block down inside the shelf,
most part of the second link with gripper needs to be put
into the shelf. Again, given the size of the second link with From the entire trajectory histories in figs. 5-7, it
respect to the size of the shelf, it is a difficult task that is shown that the quadrotor and robotic arm are con-
demands exact control of the combined system. As shown in trolled as desired during the full flight including ma-
fig. 9, the task is smoothly completed. In the release phase, nipulations of an object. These results also means that
root-mean-square error of the quadrotor is 2.56 cm. the provided approach can be utilized in expanded ma-

4994
 (a) Quadrotor is approaching the (b) Robot arm is grabbing the (c) The wood block is picked up and
designated position wood block quadrotor moves to the next position
Fig. 8: Pictures taken during pick-up phase

(a) Quadrotor is approaching the (b) The robot arm is releasing the (c) Robot arm is retreating to go
designated position wood block at the desired spot back to the origin position
Fig. 9: Pictures taken during release phase

nipulation and transportation applications. A video clip [6] M. Bisgaard, A. la Cour-Harbo, and J. Bendtsen, ”Adaptive control
of the experiment is posted on the following URL: system for autonomous helicopter slung load operations,” Control
Engineering Practice, vol. 18, no. 7, pp. 800811, 2010.
http://icsl.snu.ac.kr/video/iros2013/aerialManipulation.mp4. [7] I. Palunko, R. Fierro, and P. Cruz, ”Trajectory generation for swing-
free maneuvers of a quadrotor with suspended payload: A dy-
V. C ONCLUSION namic programming approach,” in IEEE International Conference on
This paper presents a successful application of a quadrotor Robotics and Automation, pp. 2691-2697, May 2012.
[8] N. Michael, J. Fink and V. Kumar, ”Cooperative manipulation and
with a two-DOF robot arm attached for picking up and transportation with aerial robots,” Autonomous Robots, 30(1), pp. 73-
delivering an object. Kinematic and dynamic models for the 86, 2011.
combined system are analysed and an adaptive sliding mode [9] M. Bernard, K. Kondak, I. Maza, and A. Ollero, ”Autonomous
transportation and deployment with aerial robots for search and rescue
controller is designed to control them all together. This is missions,” Journal of Field Robotics, vol. 28, no. 6, pp. 914931, 2011.
demonstrated in an experiment with a scenario consisting [10] A. Torrem, D. Mengoli, R. Naldi, F. Forte, A. Macchelli, and L.
of picking up an object, moving, and releasing it inside of Marconi, ”A prototype of aerial manipulator,” IEEE/RSJ International
Conference on Intelligent Robots and Systems, Algarve, Portugal, pp.
a shelf. The experimental results demonstrates satisfactory 2653-2654, Oct 2012.
performance to perform the designed tasks. [11] A. Albers, S. Tautmann, T. Joward, T. A. Nguyen, M. Frietsch, and
C. Sauter, ”Semi-autonomous flying robot for physical interaction
ACKNOWLEDGMENT with environment,” IEEE conference on Robotics Automation and
Mechatronics, pp. 441-446, 2010.
This work was supported by the National Research Foun- [12] C. M. Korpela, T. W. Danko, and P. Y. Oh, ”MM-UAV:Mobile
dation of Korea(NRF) grant funded by the Korea govern- manipulating unmanned aerial vehicle,” Journal of Intelligent and
ment(MEST)(No. 20110001276, 2013013911). Robotic Systems, 1-9, 2012.
[13] V. Lippiello and F. Ruggiero, ”Cartesian impedance control of a UAV
R EFERENCES with a robotic arm,” in 10th International IFAC Symposium on Robot
Control, Dubrovnik, Croatia, Sep 2012.
[1] R. M. Murray, Z.Li, and S. S. Sastry, A Mathematical Introduction to [14] V. Lippiello and F. Ruggiero, ”Exploiting redundancy in Cartesian
Robotic Manipulation. Boca Raton, FL: CRC, 1994. impedance control of UAVs equipped with a robotic arm,” IEEE/RSJ
[2] Q. Lindsey, D. Mellinger, and V. Kumar, ”Construction of cubic International Conference on Intelligent Robots and Systems, Vilam-
structures with quadrotor teams,” in Proc. Robotics: Science and oura, Portugal, Oct 2012.
Systems (RSS), Los Angeles, CA, June 2011. [15] M. Orsag, C. Korpela, P. Oh, ”Modeling and control of MM-UAV:
[3] J. Willmann, F. Augugliaro, T. Cadalbert, R. DAndrea, F. Gramazio, Mobile manipulating unmanned aerial vehicle,” J Intell Robot System,
and M. Kohler, ”Aerial robotic construction towards a new field of pp. 227-240, 2013.
architectural research,” International Journal of Architectural Comput- [16] M. Orsag, ”Mobile manipulating unmanned aerial vehicle(MM-UAV):
ing, 10(3), pp. 439-460, Sep. 2012. Towards aerial manipulators,” System, 13, 14
[4] D. Mellinger, M. Shomin, N. Michael, and V. Kumar, ”Cooperative [17] Gripper, http://www.trossenrobotics.com/p/phantomx-parallel-ax12-
grasping and transport using multiple quadrotors,” Distributed Au- gripper.aspx
tonomous Robotic Systems, 2010, 545-558 [18] Bioloid Frame and Dynamixel, http://www.robotis.com/xe
[5] P. Pounds, D. Bersak, and A. Dollar Grasp from the air: hovering [19] VICON, http://www.vicon.com
capture and load stability, IEEE/RSJ International Conference on [20] Smart Xcopter, http://www.xcopter.com
Intelligent Robots and Systems 2011, San Francisco, CA Sep 2011.

4995