Autonomous Indoor Hovering with a Quadrotor
G. Angeletti, J. R. Pereira Valente, L. Iocchi, D. Nardi
Sapienza University of Rome,
Dept. of Computer and System Science,
Via Ariosto 25, 00185 Rome, Italy
http://www.dis.uniroma1.it
Abstract. Mini and micro UAVs are very promising platforms for se-
curity and surveillance applications, because of their increased mobility
in the environment. Moreover, they can effectively employed also in in-
door environments. On the other hand, the limited payload for carrying
sensors and the limited computational power on-board make the devel-
opment of autonomous UAVs very challenging. In this paper we present
hardware and software development of a quadrobot that can reliably
navigate in indoor environments. In particular, we focus on the problem
of indoor hovering by controlling the 6 DOF of the vehicle with different
on-board and off-board sensors.
Key words: quadrotor, UAV, indoor, hovering
1 Introduction
Recently, the quadrotor has become a standard platform in the experiments and
applications of mini unmanned aerial vehicles (mini-UAVs). The great manoeu-
vrability and small size of this platform make it suitable for indoor use, where the
development of other kinds of UAVs is still limited. In this scenario, a quadrotor
could be used in security and surveillance tasks, where the capacity of flying
above ground obstacles give it a great advantage over ground robots. Indoor un-
supervised flight of aerial vehicles is a hard challenge, because it is not possible
to use global positioninng systems (such as GPS) as in outdoor applications.
Moreover, quadrotors have usually limited payload that implies a selection of
sensors that can be placed onboard. Finally, the dynamics of a flying vehicle is
more complex than that of a ground robot, so that even the hovering becomes
a difficult task.
In this paper, we propose a vision based and a laser based approach for the
quadrotor indoor hovering. Our purpose is to propose two strategies that use
different sensors in order to experiment the benefits and disadvantages of both
the approaches.
2 Related Work
Among the several kinds of mini-UAVs, quadrotors are probably the most com-
mon. This platform has a “plus” shape with one rotor per corner and has been
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
2 Autonomous indoor hovering with a quadrotor
widely developed by many Universities and research centers. “X3D-BL” [1], “X4-
flyer” [2] and “OS4” [3] are examples of quadrotors which are entirely designed
and created by a University. In these works, a customized modeling and design
approach was employed, in order to create an effective platform.
Parallel to the hardware development, several efforts have been made in order
to develop autmotic control systems, by considering both on-board and off-board
sensors. Among others, several vision-based approaches for autonomous flight of
UAVs have been studied. In these works the vision system is used to estimate
the quadrotor position and orientation, in order to let it fly autonomously. Usu-
ally, a camera is placed on the quadrotor and the image processing is done on
a ground station [4, 5] or directly on-board by a dedicated device [6]. The main
issue in this approach concerns the (usually) bad quality images, resulting from
the camera instability and signal interferences (when the image processing is
done by a ground station). Altug and collaborators [7] used a two cameras ar-
chitecture, with one camera placed onboard and the other one on the ground; a
marker system allows the application to control the helicopter. However, a vision
system presents some intrinsic problems, such as the high sensitivity to environ-
ment lighting conditions. Therefore, in some works an approach purely based
on distance sensors was used. Roberts and collaborators [8] present a quadrotor
that is autonomously able to take off, fly at a constant altitude and land, using
only an ultrasonic sensor and four infrared sensors.
In this paper we analyze different combinations of sensors and software for
controlling a quadrotor in order to maintain a stationary pose (hovering). We
present both a vision-based approach, using an external camera placed on the
ground and a a laser-based control using an on-board laser range finder.
3 System Architecture
3.1 Hardware
The hardware architecture is made of three distinct computational elements:
the quadrotor, with its integrated control; the Gumstix boards, which provide
an ARM processor, a set of input-output ports and a wi-fi connection; a ground
station, used to execute the most complex computations (e.g.: run the vision
algorithm). Furthermore, three different kinds of sensors were used to let the
quadrotor sense the environment.
The Quadrotor The aerial platform we have used in our experiments is a
quadrotor (Fig. 1) developed by Ascending Technologies [1]. The main features
of this platform consist of a great robustness – due to the quality materials
used – and an on-board 1 KHz controller, which provides a good flight stability.
Moreover, the opportunity to fly the quadrotor and receive its status via a serial
port available on-board, makes this platform an excellent base to start developing
any autonomous behavior.
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
Autonomous indoor hovering with a quadrotor 3
As we said, the quadrotor offers a very-fast embedded controller, which –
using data derived from a triaxial accelerometer – makes the quadrotor able to
keep the roll and pitch angles commanded via the radiocontroller or the serial
interface. As a consequence, if the roll and pitch stick in the radiocontroller is left
centered (or, in the same way, if a roll and pitch zero angle is commanded through
the serial port), the quadrotor will try to keep an horizontal orientation1 . Since
the quadrotor has not a compass, the yaw command will be an angular velocity
rather than an angle; using the on-board gyroscope, the control system will try
to keep the current yaw (if the yaw command is neutral), being unable to keep
directly a yaw angle. However, as the datafusion is done with an update rate of
1 KHz, in the practice the quadrotor needs only occasional yaw corrections.
Fig. 1: The quadrotor X-3D-BL, developed by Ascending Tecnologies.
The Gumstix Boards In order to interact with the quadrotor electronics,
three small size boards were added, named respectively: Gumstix, Robostix and
Wifistix. All these boards are manufactured by Gumstix Inc., specialized in
mini-boards 8 x 2 cm.
The Gumstix board provides a 400 MHz ARM processor, without the float-
ing point unit, which can execute C and C++ code. Moreover, an optimized
Linux distribution is installed on this board, so that it is possible to use all the
features (e.g. multithreading) of an advanced operating system. This board can
be considered as a motherboard, to which one can link many expansion cards,
in order to have a wide range of extra functionalities.
In this work, we have used two expansion boards, the Robostix and the
Wifistix. The first one provides a microcontroller which can drive serial (UART)
ports, PWM ports, analog to digital converters and GPIO (General Purpose
1
Obviously, due to the sensors measurement errors, the quadrotor will not keep the
position.
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
4 Autonomous indoor hovering with a quadrotor
Input Output) ports. The connection with the Gumstix is achieved via an I2C
bus. The second expansion board allows, as the name “Wifistix” suggests, the
use of the 802.11b/g protocol.
Since the Gumstix boards need a 5V voltage whereas the quadrotor battery
provides a 11.1V voltage, a Tracopowerr DC-DC voltage converter was used.
The Ground Station Due to the Gumstix hardware limitations, it is possible
to execute only the simplest computations on-board. For the more demanding
computations, we have used a ground station, that is a high-performance PC
which keeps the communication with the quadrotor via the Wi-Fi interface.
Sensors. In order to achieve any autonomous behavior, it is necessary to have
a feedback of the environment. In our work we have used three kinds of sensors:
a sonar, a monocular camera and a laser. The context in which each sensor was
used will be specified in the ”Implementation” section.
Onboard sensors. The sonar we decided to use is a model manufactured
by MaxBotixr (Fig. 2a). This is a widely used sensor, characterized by a good
resolution (2.54cm) and a very good distance range (0.16m–6.45m), with a small
size package (19.9mm x 22.1mm x 16.4mm). This sensor has an onboard micro-
controller that calculates the distance and converts it to an analogue voltage,
PWM signal and USART. Readings can occur up to every 50ms (20Hz rate).
Considering the analog interface (the one we have used), the output consists
of a voltage proportional to the measured distance, equal to V512 CC
V per inch.
This analog output is then converted to digital by a 10-bit Analog to Digital
Converter (ADC) available on the Robostix.
The Hokuyor URG (Fig. 2b) laser range finder has a 240◦ measuring angle
and can range objects between 60mm and 4095mm. The accuracy is ±10mm for
objects between 60mm and 1000mm, 1% of measurement between 1000mm and
4095mm. The scanning time is about 100ms (10Hz rate). There are three inter-
faces available: USB, RS-232C and NPN open-collector (synchronous output).
External sensors. The monocular camera used in the vision system (Fig. 2c)
has in the fisheye lens its main feature. This lens has a very large field angle
(180◦ ), which lets the camera capture a large area; on the other hand, this lens
introduces a (easy-to-correct) distortion, which is quadratic with the distance
from the center of the image. The camera can be connected with a PC via a
FireWire interface.
3.2 Software
In this section, an overview of the software architecture (with the interaction
among components) is presented. As shown in Fig. 3, there is a single approach
to control the quadrotor height, whereas two distinct strategies for roll, pitch
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
Autonomous indoor hovering with a quadrotor 5
(a) (b) Hokuyor URG (c) The
MaxBotixr laser monocular
EZ1 sonar fish-eye
camera
Fig. 2: The three kinds of sensors used.
and yaw control are presented; these two strategies are discussed in depth in
Section 4.2 and 4.3. A fusion of these solutions into a single one is currently
under investigation.
As it is possible to see in Fig. 3, the throttle command generated by the
sonar-based control goes directly to the actuators, whereas the commands com-
ing from the laser-based control and the vision-based one pass through the
quadrotor control jurisdiction (see 3.1). In summary, we have a two-layer control
architecture: an inner control, with a fast cycle time, which provides a low level
stabilization and an outer control, much slower than the former, which aims to
keep the quadrotor flying autonomously above a desired area.
Fig. 3: Main software elements.
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
6 Autonomous indoor hovering with a quadrotor
4 Implementation
4.1 Altitude control through sonar
In order to control the quadrotor height, we have used an ultrasonic sensor linked
to the Robostix ADC (see 3.1). This sensor was pointed at the ceiling, which
present less obstacles (e.g.: chairs, tables etc.) than the ground. Data coming
from the sonar are processed by a simple PI (Proportional-Integrative) control
(see equation 1), which – proportionately to the gap between the current height
and the desired one – produces a quadrotor throttle command. The integrative
term – being a purely additive term – was added to compensate the battery
dischargement, which makes the throttle command less effective.
correction at the step n
0-error output z }|
n
{
(n)
z}|{ (n) (n)
X
uP I = M0 + (Kp · (hdes − hmis )) + (Ki · Ei ) . (1)
| {z } i=0
proportional term | {z }
integrative term
The control cycle time is 20Hz, the same of the sonar readings updating.
As shown in Fig 4, the PI control runs on the Gumstix while the Robostix
manages the communication with the sonar and the quadrotor.
Fig. 4: Height control architecture.
4.2 Position control through off-board vision
The position and yaw control within indoor environments presents more diffi-
culties than the outdoor one, because it is not possible to use the GPS signal to
track the quadrotor position.
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
Autonomous indoor hovering with a quadrotor 7
As mentioned in Section 2, there are different approaches that can be used
in a vision-based setting. Our approach is based on a single camera architecture.
This camera (see 3.1) – placed on the ground – is connected with a PC where a
vision algorithm is running. Moreover, two markers (Fig. 5) are placed just below
the base and the front rotor of the quadrotor. The vision algorithm corrects the
distortion introduced by the fisheye lens and extracts the position and the yaw
angle of the quadrotor, relatively to the image coordinates. These information are
sent via Wi-Fi to a two threads application running on the Gumstix: one thread
is a server listening for data from the PC; the other one implements three (roll,
pitch, yaw) PI controls (see equation 1) and communicates with the first one via
shared memory. In order to let the user decide the position that the quadrotor
has to keep, a graphic interface (Fig. 6) was implemented on the ground station.
This graphic interface is split into two sub-windows: the left window shows the
image as the camera has captured it; the right one shows the vision algorithm
elaboration. In this way, the user can – pushing the “L” (as lock! ) key on the
keyboard – decide the point of the image which will be considered as a reference-
point by the control application. When the “L” key is pressed, the system makes
a freeze of the current position and height (considering in the same architecture
the height control too) and send them – as a reference 3D point – to the controls
implemented on the Gumstix.
Fig. 5: Markers.
Fig. 6: The graphic interface implemented on the ground station.
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
8 Autonomous indoor hovering with a quadrotor
Since in this application we focus on the quadrotor position keeping, the yaw
angle does not matter when the quadrotor achieves the goal. Consequently, as
shown in Fig. 7, the camera image was divided in four reference angles: 0◦ , 90◦ ,
180◦ and 270◦ . The quadrotor performs a starting manoeuvre to reach the closer
angle, with a 45◦ rotation in the worst case. This manoeuvre produces a great
control simplification (Fig. 8), because – with the quadrotor directed to one of
this four angles – we can correct the position error separately for the coordinates
x and y, as the roll corrects directly the x-axis error and the pitch the y-axis error
or viceversa (it depends on the reference angle). After reaching the reference yaw
angle, this one is kept by the embedded quadrotor control (see 3.1); since the
quadrotor has not a compass, it can happen that the reference yaw angle is lost.
Therefore, when the yaw angle measured by the vision algorithm is out of a fixed
interval centered on the reference angle, the outer control is activated and the
quadrotor returns within the interval.
Fig. 8: An example of the reference
angle strategy benefits.
Fig. 7: An example of the quadrotor
starting manoeuvre.
The vision algorithm uses a blob extraction function to identify the markers
and the Hough transform to compute lines corresponding to the axis of the
quadrotor and to determine the current yaw angle. In the right window of the
graphic interface (Fig. 6), the centers of the two markers and the line indicating
the quadrotor direction are highlighted.
4.3 Position control through on-board laser
In this approach, the quadrotor position and heading (i.e. the yaw angle) are
retrieved by using data provided by an horizontal laser range finder and a vertical
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
Autonomous indoor hovering with a quadrotor 9
ultrasonic sonar (see Section 3.1 for details on the hardawre). In this setting,
we implemented a control system for autonomous hovering, combining the sonar
based height control described in Section 4.1 with a position based control based
on on-board laser.
Fig. 9 shows an example of the environment. When the hovering phase starts
(indicated by the human operator), the quadrotor memorizes the current laser
reading and use this as a reference scan. During the hovering task, a 2D scan
matching procedure (i.e., the current scan is compared to reference scan) is
performed and X, Y, θ displacement with respect to the target pose is thus com-
puted. This procedure ignores role and pitch angles of the vehicle, relying on the
low-level stabilization implemented on the quadrotor.
Fig. 9: Example of laser-based hovering.
The control of the vehicle is based on three controllers, one for each move-
ment: pitch, roll and yaw.
The pitch control is responsible for the aircraft displacement on the y-axis
and it is achieved through a PD control law. The roll control influences the
displacement on the x-axis and it is implemented using a cascade control archi-
tecture (Fig. 10), where a simple P control was nested inside a PD control loop.
Also the yaw control strategy is implemented through a standard PD control law
based on the orientation estimation provided by the scan matching procedure.
5 Discussion
As already mentioned, the functionality for a quadrotor of autonomously hover-
ing over a target is very important in many security and surveillance applications.
Moreover, it is a first step towards a complete autonomous control.
The analysis and the implementation described in this paper is the first step
towards a detailed experimental analysis that has the goal of evaluating the
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
10 Autonomous indoor hovering with a quadrotor
Fig. 10: Roll cascade control architecture.
different performances of the two proposed control methods, and in general the
effectiveness of the proposed approaches.
Besides such an experimental analysis, future work include the study and
the development of other more complex autonomous behaviors in indoor envi-
ronments, like wall following, point-to-point navigation, etc.
References
1. Gurdan, D., Stumpf, J., Achtelik, M., Doth, K.M., Hirzinger, G., Rus, D.:
Energy-Efficient Autonomous Four-Rotor Flying Robot Controlled at 1 KHz. In:
IEEE International Conference on Robotics and Automation. Rome (2007)
2. Pound, P., Mahony, R., Hynes, P., Roberts, J.: Design of a four-rotor aerial robot.
Australasian Conference on Robotics and Automation, November 2002
3. Bouabdallah, S.: Design and Control of Quadrotors with Application to Autonomous
Flying. PhD thesis, Ècole Polytechnique Fédérale De Lausanne, 2007
4. Kemp, C.: Visual Control of a Miniature Quad-Rotor Helicopter. PhD thesis, Uni-
versity of Cambridge, 2006
5. Tournier, G.P., Valenti, M., How, J.P.: Estimation and Control of a Quadrotor
Vehicle Using Monocular Vision and Moiré Patterns. Massachusetts Institute of
Technology, Cambridge. AIAA Guidance, Navigation and Control Conference and
Exhibit, 2006.
6. Lillywhite, K., Lee, D.J., Tippetts, B., Fowers, S., Dennis, A., Nelson, B., Archibald,
J.: An embedded vision system for an unmanned four-rotor helicopter. Robotic
Vision Laboratory, Department of Electrical and Computer Engineering, Brigham
Young University
7. Altug, E., Taylor, C.: Vision-based pose estimation and control of a model helicopter.
Istanbul Technical University and University of Pennsylvania
8. Roberts, J.F., Stirling, T.S., Zufferey, J.C., Floreano, D.: Quadrotor Using Min-
imal Sensing For Autonomous Indoor Flight. 3rd US-European Competition and
Workshop on Micro Air Vehicle Systems (MAV07) & European Micro Air Vehicle
Conference and Flight Competition (EMAV2007), September 2007
Workshop Proceedings of SIMPAR 2008
Intl. Conf. on SIMULATION, MODELING and PROGRAMMING for AUTONOMOUS ROBOTS
Venice(Italy) 2008 November,3-4
ISBN 978-88-95872-01-8
pp. 472-481
