Sei sulla pagina 1di 22

J Intell Robot Syst (2017) 88:597–618

DOI 10.1007/s10846-017-0475-z

A Novel Landing System to Increase Payload Capacity


and Operational Availability of High Altitude Long
Endurance UAVs
Tin Muskardin · Georg Balmer · Linnea
Persson · Sven Wlach · Maximilian Laiacker ·
Anibal Ollero · Konstantin Kondak

Received: 14 September 2016 / Accepted: 5 January 2017 / Published online: 26 January 2017
© Springer Science+Business Media Dordrecht 2017

Abstract Unmanned stratospheric aircraft capable of significantly reduces the system’s operational avail-
staying aloft for long periods of time have become a ability. To address these drawbacks a novel landing
topic of interest in the past years. Several problems are system is proposed in this paper. The landing gear
still to be solved to allow for a profitable commercial can be removed from the aircraft and a ground-based
use of such aircraft. The inherent lightweight design mobile landing platform is introduced. The main tech-
leads to fragile structures with low payload capacities nical challenges consist in the precise relative state
and a high wind sensitivity. The weather dependence estimation and cooperative control of the involved
vehicles. A reliable simulation model of the overall
system was developed and a number of simulation
experiments performed before the actual landing was
T. Muskardin () · G. Balmer · S. Wlach · M. Laiacker · attempted with an experimental system setup. Mul-
K. Kondak tiple successful landing experiments demonstrate the
German Aerospace Center DLR, Institute of Robotics
and Mechatronics, Muenchener Str. 20, 802234 Wessling, validity of the proposed system.
Germany
e-mail: tin.muskardin@dlr.de Keywords Aerial robotics · Cooperative control ·
G. Balmer UAV · UAS · Multi-robot systems
e-mail: georg.balmer@dlr.de
S. Wlach
e-mail: sven.wlach@dlr.de 1 Introduction
M. Laiacker
e-mail: maximilian.laiacker@dlr.de 1.1 Motivation and Challenges
K. Kondak
e-mail: konstantin.kondak@dlr.de In the recent years, Unmanned Aerial Vehicles (UAVs)
have gained widespread interest, and they are now
L. Persson
KTH Stockholm, Department of Automatic Control, used in many areas. High altitude, long endurance
Osquldas väg 10, SE-100 44, 11428 Stockholm, Sweden (HALE) UAVs are promising in terms of poten-
e-mail: laperss@kth.se tial applications; however, the development has not
yet reached the stage where a commercial use is
A. Ollero
University of Seville, ETS Ingenieria, Camino
viable. Intended for a use in the stratosphere, at alti-
Descubrimientos s/n, 41092 Sevilla, Spain tudes in the order of 20km, HALE UAVs typically
e-mail: aollero@us.es have large wingspans and very lightweight structures.
598 J Intell Robot Syst (2017) 88:597–618

Equipped with solar cells and sufficient battery capac- with a human driver. The demonstrator setup is shown
ity, such airplanes can stay airborne day and night, in Fig. 1 and will be described in detail in Section 2.
which gives them a theoretically unlimited endurance. In a first step the dynamical models of the air-
This opens up possibilities for various applications, craft and ground vehicle have been derived using the
including earth observation, atmospheric science and methods described in [5, 7, 13], and [14] and a realis-
communication networks. Some of these tasks are tic simulation environment was created. The different
now performed by satellites, which are very expen- aspects of the landing manoeuvre were then sequen-
sive to build, launch and operate. Apart from saving tially analyzed before the first landing was attempted.
the cost of a rocket launch, HALE UAVs have more The concept of a ground based landing gear can be
advantages: they are independent of an orbit, and can beneficial to all types of aircraft and has already been
operate anywhere, either staying at a given location, studied in the scope of different research projects.
patrolling over a large area, or relocating as neces- Two mentionable projects can be found in [1] and
sary, e.g. in response to a natural disaster. Unlike [2], where the focus was on heavy commercial avia-
satellites, aerial vehicles would be able to return to tion and its inherent requirements. Consequently, the
the earth surface for maintenance or updates. One of resulting landing systems substantially differ from
the main limitation of HALE platforms is their typ- the solution proposed here, where the focus is on
ically low payload capacity. In order to fly at high ultra-lightweight HALE UAVs.
altitudes, a low wing loading is necessary, which trans-
lates to a large wing area and a low gross weight. 1.2 Main Contribution
However, the batteries required for overnight flight are
heavy, and further reduction of the structural mass is The main contribuntion of this work consists in the
very difficult, given the already extreme lightweight development of an integrated landing system for ultra-
design. New approaches are required to lower the light unmanned aircraft without landing gear. The
empty weight and in turn increase the payload. focus was thereby on practicality and reliability of the
The approach suggested in this paper is to remove solution. As described in [3], a first version of the pro-
the landing gear. With mission durations of several posed landing system was successfully tested using
months, the landing gear, which is only used for take- a hardware in the loop (HIL) test setup. Thereby a
off and landing, is merely dead weight for most of virtual aircraft was landed on the actual ground vehi-
the time. Without an aircraft-mounted landing gear, cle. In [4] the landing manoeuvre was successfully
new solutions need to be found for take-off and land- demonstrated for the first time using a smaller scale
ing. Take-off could be realized using some sort of a test setup. The landing was thereby executed multiple
temporary gear, which is dropped once the aircraft is times, proving the validity of the chosen approach. In
airborne. The landing, however, poses a bigger chal- the present paper, the results of the subsequent devel-
lenge. The core idea is to have a landing platform opment work and additional landing experiments are
mounted onto a ground vehicle, which is able to accel- presented. The inclusion of vision-based state esti-
erate to the UAV’s landing speed and to cooperatively mates into the control loops and the modification of
with the UAV control position and velocity, so that the the high-level mission control structure for a safer
UAV can touch down on the landing platform without
a horizontal velocity component relative to the plat-
form. This would not only eliminate the need for a
landing gear, but also facilitate landings in crosswind
situations, as the aircraft does not need to align with
the runway direction. HALE UAVs are typically very
sensitive to wind, due to their size and low weight,
and this proposed landing procedure could greatly
improve their operational availability. The overall goal
is to experimentally demonstrate the feasibility of such
a landing. For this purpose, a small UAV is used
together with a landing platform mounted on a car Fig. 1 Demonstrator setup for landing experiments
J Intell Robot Syst (2017) 88:597–618 599

Fig. 2 The different phases of the landing manoeuvre


600 J Intell Robot Syst (2017) 88:597–618

and smoother landing performance represent the most


important improvements. The locking mechanism pre-
sented in this paper was adapted to the demonstrator
setup and has yet to be modified for a use on large
scale aircraft as will be explained in Section 6. For bet-
ter readability, some basic information from [3] and
[4] is summarized here.
Fig. 3 Aerial vehicle for first landing experiments and mobile
ground vehicle with mounted landing platform
1.3 The Landing Manoeuvre
ground vehicle, referred to as UAV and UGV, respec-
The landing manoeuvre has been divided into six
tively. Both vehicles are shown in Fig. 3. The UAV
distinct phases, as illustrated in Fig. 2:
is a commercially available product1 with a wingspan
1. Initial approach of 3.3m and a maximum takeoff mass of 21.5kg.
2. High precision approach It has been equipped with a flight control system
3. Guided descent and sensors including RTK-GPS (Novatel FlexPak6),
4. Final descent and touchdown IMU (MicroStrain 3DM-GX3-25), pitot tube (Simtec
5. Deceleration to full stop PSS8), and an optical camera (Allied Vision Prosilica
6. Ground Lock GC1380). The ground vehicle is a modified passen-
ger car equipped with an on-board computer as well
Phases 1 and 2 guide the UAV to a pre-defined
as RTK-GPS, IMU, a Graphical User Interface (GUI),
start point about 300m behind the ground vehicle (will
and optical markers on the roof of the car.
also be referred to as UGV in the following), which is
The landing platform is mounted on the car’s roof
waiting at its start-/idle position at the runway thresh-
and consists of an aluminum frame and a net. It is 5m
old. The final approach consists of phases 3 and 4,
wide and 4m long. The net tension is adapted to the
where the UGV starts accelerating and then moves
weight of the UAV and the expected maximum touch-
cooperatively with the UAV, until the aircraft touches
down impact. The mesh spacing has been chosen in
down on the landing platform. In phases 5 and 6 the
accordance with the UAV wheel dimensions, so that
UGV is the main actor, decelerating the joined vehi-
upon touchdown the wheels push through the net and
cles and safely terminating the manoeuvre. In these
the aircraft fuselage comes to rest on the net. The high
phases, the UAV’s control surfaces are used to assist
level cooperative control loops are closed through a
in keeping the UAV safely locked down on the landing
wireless data connection on 5.8GH z, using a lag com-
platform.
pensation term for the generation of smooth signals
The transitions between the different landing
despite communication lags.
phases are managed via a state machine logic, which
also includes a failsafe mechanism. If the UAV devi-
ates significantly from the flight path (e.g. due to a
3 Relative State Estimation
strong gust), or if the operator gives an abort com-
mand, the UAV is commanded to go around and fly
Before any cooperative control task can be performed,
aerodrome circuits following pre-defined waypoints,
the relative vehicle states must be known. The qual-
while the UGV is commanded to a full stop. The
ity of the relative state estimates directly influences
implemented state machine will be explained in detail
the achievable control precision. A number of differ-
in Section 9. For further information on the landing
ent vision-based state estimation methods, including
manoeuvre please refer to [3].
incremental feature based motion tracking, RANSAC
based motion tracking, as well as techniques based on
geometrical models [10] and optical marker tracking
2 Experimental Setup
have been analyzed and tested. In the end a combi-
nation of GPS-based networked state estimation and
The proposed landing system used as technology
demonstrator consists of an aerial vehicle and a 1 Penguin BE UAV (http://www.uavfactory.com/)
J Intell Robot Syst (2017) 88:597–618 601

optical multi marker tracking showed the best perfor- analyzed and tested. In the end a very simple method
mance and was applied to increase the reliability and showed the best performance results. In the following,
precision of the relative state estimation system. the method will be explained for the UAV side, but
was implemented in the same way for the ground vehi-
3.1 Networked State Estimation and Lag cle. The locally measured UAV position is subtracted
Compensation from an obsolete UGV position received via the wire-
less network. The relative position is calculated and
Each vehicle’s sensor data needs to be transmitted via the signal lag corrected for, using the above described
a wireless network and fused with the other vehicle’s lag estimation together with the ground vehicle veloc-
local sensor data. In our case GPS position and veloc- ity that was measured together with its position (at the
ity are exchanged together with the corresponding same moment in time):
timestamp (GPS Time). This data needs to be synchro-
nized and communication delays compensated for. For x rel (t) = x U GV (t−tlag )+v U GV (t−tlag )·tlag−x U AV (t)
effective lag compensation, the amount of signal lag
(2)
needs to be determined, which is done via the GPS
timestamps. This calculation is performed on both
with:
vehicles, where the obsolete measurement received
⎡ ⎤ ⎡ ⎤
via the wireless network is subtracted from the local x u
measurement. In the case of the UAV the lag estimate x(t) = ⎣ y ⎦ v(t) = ⎣ v ⎦ (3)
is calculated as follows: z w
tlag = tlocal U AV − tnetwork U GV (1)
The reference frame used to describe the above
The calculation is performed in the same way on quantities is pointing in runway direction with its
the UGV. A typical lag profile encountered during x-axis, the y-axis is pointing to the right and the z-
landing tests is illustrated in Fig. 4. axis downwards. Due to the lack of synchronized
As can be seen in Fig. 4, the signal lag estimates acceleration measurements, lag compensation was not
are only available in discrete steps of 0.05s, which implemented for the calculation of relative veloci-
corresponds to the used GPS sensor’s measurement ties. Here it was assumed that the respective vehicle’s
frequency of 20H z. At a landing velocity of around velocity did not change from its previous value when
21 ms , this would result in a minimum value of 1.05m lag occurred.
for the correction of the relative position. The signal
lag typically ranges from 0.05s to 0.15s with occa- 3.2 Optical Marker Tracking
sional outliers at 0.2s and more. The main factors
influencing the signal lag are the wireless data trans- The relative state estimation method described in the
mission and the increased amount of signal routing previous chapter is based on data exchange via a
present in the system, as positioning data is exchanged wireless network. This induces significant signal lag
between the two vehicles via the ground station. which needs to be compensated for. The lag compen-
Different fusion methods (including Kalman filters sation method proposed in Section 3.1 is only valid for
and buffer based synchronization methods) have been smaller deviations from the linear forward motion, as
obsolete velocities are used for extrapolation. A virtu-
ally lag-free optical multi marker detection and local-
ization method was therefore developed and imple-
mented to further increase the robustness and preci-
sion of the target estimation. These directly measured
relative state estimates are then fused with the net-
worked GPS-based estimates as will be described in
Section 3.3. This leads to a more reliable system,
which is less sensitive to a degraded GPS quality or
Fig. 4 Typical lag profile during flight tests complete sensor outages.
602 J Intell Robot Syst (2017) 88:597–618

Artificial markers are a good tool to detect a known


object and localize it relative to a monocular cam-
era. The open-source software suite ARToolkit [11]
was chosen for object detection and localization. This
algorithm was chosen because it can run in real
time on the on-board image processing computer. The
method also has a very low false positive detection
rate because a checksum is included in the binary
pattern used for the markers. This is very important
for the reliability and safety of the automatic land-
ing system. A problem can occur when the marker
is in frontal direction of the camera. In this position
the orientation estimate gets unstable for the angles
around the camera x and y axis (z axis pointing in the
viewing direction) [12]. This limitation doesn’t affect
our application because the unstable angles are not
needed for the landing algorithm. Only the angle of
the marker around the camera z axis would be needed Fig. 5 Reference marker pose estimate and transformation to
desired reference point at GPS antenna mounting position
to align the UAV heading to the landing platform on
the UGV. Another drawback of this method is that
a partly occluded marker might not be detected and measurements using two RTK-GPS receivers and the
localized. To overcome this drawback multiple mark- vision-based multiple marker system.
ers are used. A reference marker is specified and the The vision measurements are more accurate the
measurements will be with respect to this marker. If closer the camera gets to the marker. The detection of
the position of the reference marker together with the marker is possible from a distance of up to 16m
one or more other marker is measured the relative when the target is in the field of view. Between 10s
position and orientation with respect to the reference and 15s the target is not inside the camera field of view
marker will be calculated and stored. If the markers so no measurements are available. Below 10m relative
are measured multiple times the measurements will altitude the target is reliably detected by the vision sys-
be averaged. If the new measurement differs from the tem. The relative measurements of the two systems are
stored more than a specified value (in our case 5cm) very close together during the critical landing phase
the old measurements will be discarded and only the below 5m. The vision system cannot detect the marker
new measurement is used. If multiple known mark- during the last phase below 1m. The visual relative
ers and not the reference marker are in the image position estimate is a good method to compensate for
a RANSAC algorithm is used to eliminate possible
outliers. After the RANSAC is performed the mea-
surements are averaged. Figure 5 shows the estimated
reference marker pose (large marker in the middle of
the roof) and the transformation to a pre-defined point
of interest (here the GPS antenna location) during an
exemplary landing attempt.
As can be seen, the vision-based estimation of the
GPS antenna location corresponds well to the actual
sensor location (white antenna in the middle of the
front crossbar). The advantages of this algorithm are a
more reliable optical target detection and also higher
accuracy of the measured position of the reference
marker because multiple measurements are fused. Fig. 6 Vision-based relative state estimates compared to RTK-
Figure 6 shows a comparison of the relative position GPS measurements
J Intell Robot Syst (2017) 88:597–618 603

GPS or communication problems during the most crit- the measurement noise covariance matrices for GPS
ical phase of the landing which significantly increases and marker tracking are:
the reliability of the landing system. Figure 7 shows a ⎡ ⎤ ⎡ ⎤
series of images, illustrating the continuous function- 0.4 0 0 40 0
ality of the optical multi marker tracking algorithm R GP S = ⎣ 0 0.4 0 ⎦ R vision = ⎣0 4 0 ⎦
until touchdown. 0 0 1.6 0 0 16
(6)
3.3 Sensor Data Fusion
The relative velocity estimate is used to estimate
To use both the marker detection and the relative GPS the relative position when no new measurement is
position the sensor data is fused using a linear Kalman available. The Kalman filter is running at 20H z.
filter. The fused realtive position is then used in the
control loops. The state of the filter consist of the three
relative position components and relative velocity x = 4 Cooperative Control
[x, y, z, u, v, w]. The state transition matrix is
⎡ ⎤ In the following subsections a right handed coordinate
1 0 0 Δt 0 0 system is used to describe the motion of the aircraft
⎢ 0 1 0 0 Δt 0 ⎥
⎢ ⎥ and the ground vehicle. The reference frame origin is
⎢ 0 0 1 0 0 Δt ⎥
A=⎢ ⎢0 0 0 1 0 0 ⎥
⎥ (4) located at the northern end of the runway (black home
⎢ ⎥ icon in Fig. 24), with the x-axis pointing in southward
⎣0 0 0 0 1 0 ⎦
runway direction, the y-axis to the right, and the z-
000 0 0 1 axis downwards. It will be referred to as local runway
The filter is updated with the measurement that is coordinate system.
available at the update timestep. If both sensors have
valid measurements only the relative GPS position is 4.1 Longitudinal cooperative control
used to update the filter. The system noise covariance
matrix is: Two main phases were considered during the land-
⎡ ⎤ ing. At the beginning of the landing manoeuvre, the
0.005 0 0 0 0 0
⎢ 0 0.005 0 0 0 0 ⎥ UGV is at rest at the runway threshold, while the
⎢ ⎥
⎢ 0 0 0.0005 0 0 0 ⎥ UAV is far behind, approaching the runway. While
Q=⎢ ⎢ 0
⎥ (5)
⎢ 0 0 0.15 0 0 ⎥⎥
the UAV is catching up, the UGV already needs to
⎣ 0 0 0 0 0.15 0 ⎦ accelerate to match the UAV’s speed. For this initial
0 0 0 0 0 0.01 acceleration phase, the UGV speed command is cal-
culated under the assumption that the UAV maintains
its speed, while the UGV has a constant acceleration.
With these assumptions, and the constraint that speed
and position shall be aligned at the same time, the
following UGV speed command is derived:

VU GV = VU AV − 2 · aU GV · (xU GV − xU AV ) (7)

where aU GV is the desired (constant) acceleration


of the UGV, and VU AV is the current UAV veloc-
ity, assumed to remain constant. This control law
implicitly defines the point at which the UGV starts
accelerating, which can be found by setting VU GV =
0, and solving for the relative position:

VU2 AV
xU GV − xU AV = (8)
Fig. 7 Continuous optical marker tracking until touchdown 2 · aU GV
604 J Intell Robot Syst (2017) 88:597–618

The start of the UGV acceleration phase depends However, as there are relatively tight constraints on
on the UAV speed and the UGV acceleration. Assum- the UAV’s airspeed, especially during approach and
ing a typical approach speed of 19 ms for the UAV landing, the control authority is rather limited for the
and roughly half the maximum UGV acceleration for UAV.
aU GV , a relative position of around 190m is found. A stability analysis with the linearized system mod-
Since VU AV in this case is the UAV’s ground speed, els of the UAV and UGV has been performed for the
effects of head- or tailwind are automatically consid- longitudinal direction. This analysis has shown that
ered and the UGV will adjust its acceleration phase the coupled system remains stable in a wide range of
accordingly, as defined in Eqs. 7 and 8. gains k1 and k2 .
While a constant, relatively high acceleration is The total longitudinal system consisting of UAV
ideal for the initial acceleration phase of the UGV, it is and UGV can be expressed as one single decoupled
not suitable for close proximity operations in the sec- system of the form:
ond phase, where small position and velocity errors

need to be corrected. ẋU AV AU AV 0 xU AV


=
For this second phase we propose an algorithm ẋU GV 0 AU GV xU GV

based on a feedback controller which coordinates the BU AV 0 uU AV


motion of both vehicles. The scheme of this controller + (11)
0 BU GV uU GV
is shown in Fig. 8. A position difference in x-direction
(along the runway) is fed through the controller and where the inputs are the desired velocity and path
translated into a velocity command. This additional angles of the vehicles, and the outputs are the relative
velocity command is then added to the desired land- distance and the velocities of the vehicles. Given that
ing airspeed of the aircraft and the current aircraft we are considering the longitudinal dynamics isolated
ground speed and sent to the UAV and UGV veloc- from the lateral dynamics, reasonable signals to feed-
ity controllers. Different approaches could be used back for the longitudinal control could be velocities,
for the implementation of the UAV and UGV cou- velocity difference and position difference. The system
pled controller blocks. In our experiments, we used is closed by choosing the input as one of these Equations:
a basic implementation of the two blocks shown in
Vdes U AV = Vland
Fig. 8, where a velocity command for the longitudinal
direction is calculated of the form: Vdes U GV = VU AV + k1 · (xU GV − xU AV ) (12)
VU GV = VU AV + k1 · (xU GV − xU AV ) (9)
Vdes U AV = Vland + k2 · (xU GV − xU AV )
VU AV = Vlanding + k2 · (xU GV − xU AV ) (10) Vdes U GV = VU AV (13)

UGV Coupled Controller 1


UGV
v _des_UGV vel _UGV s pos_UGV

UAV ground speed

19 pos_rel

desired landing airspeed

1
UAV
UAV Coupled Controller v _des_UAV vel _UAV s pos_UAV

Fig. 8 High-level cooperative control strategy


J Intell Robot Syst (2017) 88:597–618 605

Vdes U AV = Vland + k2 · (xU GV − xU AV )


Vdes U GV = VU AV + k1 · (xU GV − xU AV ) (14)

where Vland is a ground speed appropriate for land-


ing. In the first structure given in Equation 12 the
ground vehicle takes over longitudinal positioning
control while the UAV keeps a constant airspeed.
In the control structure from Equation 13, the UAV
takes over positioning control while the UGV adapts
its speed to the ground speed of the UAV. The final
control structure as given in Equation 14 combines
the two previous approaches and makes both vehicles
simultaneously react to differences in position.
Figures 9–11 show how the respective root locus
plots change with the different feedback structures for
the longitudinal controller. In Figs. 9 and 10, the rel-
ative positioning information is only fed back to one
vehicle. In Fig. 11, both vehicles use the positioning Fig. 10 Root locus plot for linearized overall system (UAV and
information simultaneously. UGV) with position feedback only to UAV
As can be seen in these plots, the overall dynam-
ics of the coupled system are very similar for all three For this controller structure the aircraft dynamics
controller structures. The time constants and stability are not dependent on the gain k1 and therefore not dis-
margin are also similar. played in the plot. The controlled aircraft is considered
In Fig. 9, the pole locations of the first coupled sys- a stable system and no further analysis was performed
tem defined in Equation 12 are shown in dependence at this point. Here the influence of the controller gain
of the controller gain k1 . on the overall coupled system dynamics was analyzed.

Fig. 9 Root locus plot for linearized overall system (UAV and Fig. 11 Root locus plot for linearized overall system (UAV and
UGV) with position feedback only to UGV UGV) with position feedback to both vehicles
606 J Intell Robot Syst (2017) 88:597–618

As can be seen, the blue pole remains on the real axis The modified airspeed command, which is sent to
for all values k1 . the UAV flight controller, can be derived as follows:
Figure 10 shows the root locus plot for the con-
troller structure of Equation 13 where the relative VA des = (Vk des − Vk ) + VA (15)
positions are fed back only to the UAV. The ground
If such command was fed to the current aircraft
vehicle is commanded to adjust its speed to the current
velocity controller (based on airspeed control error
UAV ground speed, automatically compensating for
eVA ), only the control error in flight path velocity eVk
changing wind conditions. As can be seen, an imag-
would be considered for the generation of actuator
inary part is developed with an increasing controller
commands, as VA would cancel out:
gain k2 .
Figure 11 shows the root locus plot of the fully cou- eVA = VA des − VA = (Vk des − Vk ) + VA − VA
pled system as defined through Equation 14, where the
= Vk des − Vk = eVk (16)
positioning error is fed back to both vehicles. Again,
an imaginary component is developed with increas- The desired ground speed command is then deter-
ing controller gains k1 and k2 (changing both gains mined as follows:
simultaneously).
In the end the first controller structure as shown Vk des = Vk land + k · (xU GV − xU AV ) (17)
in Fig. 9 was considered to be favorable to the other
approaches. As mentioned before all three approaches
lead to similar overall dynamics with comparable time Vk = VA land + Vk − VA = VA land + Vw (18)
land
constants and stability margin. For the first approach
the real axis pole does not develop an imaginary Here the wind estimate Vw represents the averaged
component with an increasing controller gain. This difference between the UAV ground speed and air-
implies lower oscillations and consequently less con- speed (denoted as Vk − VA ), read out at the moment
trol activity for both systems. It also represents a landing mode is activated (shortly after reaching WP2
simpler practical realization, with the UAV velocity in Fig. 24). As a result, Vk land represents the UAV
being constant and independent of the ground vehi- ground speed command that would approximately
cle state. Additionally all three approaches were tested produce the airspeed present at landing mode initi-
in flight experiments where the first control structure ation. A low-pass filter with a cut-off frequency of
was confirmed to be the best choice for the current 0.2H z was used to obtain an averaged value for the
experimental setup. wind estimate Vw . The resulting modified airspeed
command required for ground speed tracking is then
4.2 Ground Speed Tracking Mode found as:

VA des = VA land + k · (xU GV −xU AV )+Vw −Vk +VA


Additionally, a ground speed tracking mode has been 
implemented, where the UAV airspeed command is previous airspeed command (19)
modified to control the current UAV ground speed
instead of airspeed. The aircraft will then actively The results from the linear analysis presented in the
compensate for atmospheric disturbances (wind gusts, previous section indicate that a constant UAV ground
UGV flow effect) while maintaining the commanded speed (k = 0 in Equation 17) is the best choice for a
ground speed. With the ground vehicle controlling safe landing manoeuvre.
ground speed as well, both vehicles then directly con-
trol the same physical quantity, thus facilitating the 4.3 Lateral cooperative control
alignment process. It should be noted that all air-
craft envelope protection features remain active and For the lateral direction a similar approach as the one
safe flight is still guaranteed, as only the airspeed described for the longitudinal direction in Section 4.1
command is modified while the actual UAV flight was used. Here, a given lateral deviation is corrected
controller remains unchanged. for by commanding an appropriate corrective course
J Intell Robot Syst (2017) 88:597–618 607

action. The desired course for the UAV and the ground is performed on top of a moving landing platform, the
vehicle are now calculated as follows: touchdown vertical velocity is adjusted through the
retard and ground lock settings, as will be described in
χU AV = χrunway + k1 (yU GV − yU AV )
Section 6. The flare initiation altitude hf lare was set
1
+k2 (yU GV − yU AV ) (20) to 5m (relative altitude).
s
+k3 (vU GV − vU AV )

χU GV = χrunway + k4 (yU GV − yU AV ) 5 Aircraft Flight Control


+k5 (vU GV − vU AV ) (21)
Conventional flight control systems consist of sev-
Where vU AV and vU GV represent the y-component eral single-input, single-output (SISO) control loops,


of the UAV- and UGV velocity vector Vk = where the engine is typically used for speed control
T
( u v w ) in runway coordinates. As can be seen and the elevator for altitude control. Figure 12 illus-
from these equations, an integral part has been added trates such a classical control structure for altitude
for the aircraft, as well as a damping term for both control.
vehicles. On the ground vehicle side, an integrator has A separate SISO-controller is used to control air-
not been used, due to the inherent integrating behav- speed via the engine, typically consisting of a single
ior of the human actuator. With this control approach, PI-loop with a feedforward term:
the vehicles eliminate lateral position differences and
keep the course parallel to the runway. 1
δt = δtF F + (KVa + KiVa ) · (Vdes − Va ) (24)
s
4.4 Vertical Control and Flare Law
with throttle command δt , feed-forward term δtF F ,
The vertical flight path is currently implemented as a proportional gain KVa , integral gain KiVa , com-
constant 3◦ glideslope, combined with a classical flare manded airspeed Vdes , and current airspeed Va .
law as described in [8] and [9]. The sink rate during However, this approach does not account for
the flare is commanded as: the strong inherent coupling in altitude and speed
responses of elevator and thrust commands. Further-
ḣf lare = k · (h + hB ) (22)
more, large elevator deflections may drive the airspeed
The constants k and hB are calculated under the outside the permissible range, due to the limited con-
following constraints: trol authority of the speed control system.
ḣf lare (hf lare ) = ḣdescent To overcome these limitations of conventional
SISO controllers, an improved flight controller, con-
ḣf lare (0) = 0 (23)
sidering energy principles as proposed in [8, 15], and
This would lead to a relative vertical velocity of [16], was implemented. Such controller is commonly
zero at touchdown, which in principle is not desirable referred to as total energy control system (TECS) in
for a runway landing. In our case, where the landing literature. It consists of an integrated control system

Fig. 12 Simplified diagram of a conventional altitude con- h, flight path angle γ , angle of attack α and pitch rate q,
troller. It consists of four cascaded loops controlling altitude respectively, using P, PI and PD controllers
608 J Intell Robot Syst (2017) 88:597–618

Fig. 13 Simulation of a 50m climb at constant commanded airspeed (left) and a 8 ms speed increase at constant commanded altitude
(right) for the SISO and energy based (TECS) flight controllers

using both throttle and elevator in a coordinated man- 1 e V̇


ner to decouple altitude and airspeed responses. The θcmd = KEI (eγ − V̇ ) − KEP (γ − ) (26)
s g g
basic idea is to use throttle commands to control the
aircraft’s total energy, i.e. the sum of potential and
where eγ and eV̇ represent the flight path angle and
kinetic energy, and elevator commands to control the
acceleration errors respectively.
distribution between potential and kinetic energy. In
A qualitative analysis has been performed by means
practice, the dimensionless energy rates γ = Vḣ = of simulation and flight experiments for both con-
Ėpot Ėkin
mgV and V̇g = mgV are used, leading to the following trol approaches. The controllers were provided with
core algorithm: the same set of flight manoeuvres. Figure 13 shows
two exemplary simulation results for an altitude step
FT 1 e V̇ at constant commanded airspeed and a speed-up
= KT I (eγ + V̇ ) − KT P (γ + ) (25)
mg cmd s g g manoeuvre at constant commanded altitude.

Fig. 14 Simplified diagram of the lateral/directional flight controller. It consists of three cascaded loops controlling course, bank
angle and roll rate p, respectively, using P and PI controllers
J Intell Robot Syst (2017) 88:597–618 609

Fig. 15 Horizontal flight with the energy based flight con- Fig. 17 A 3 ms speed increase with the energy based flight con-
troller. The altitude error remains within ±1m, the airspeed troller. The altitude command is tracked just as well as during
error within ±1.5 ms (commanded values are displayed in blue, level flight, showing the good coordination of elevator and throt-
actual sensor data in red) tle (commanded values are displayed in blue, actual sensor data
in red)
As can be seen on the left side of Fig. 13 the SISO
controller is not able to maintain the airspeed dur- Steps in airspeed and altitude were followed with
ing the climb, despite a full throttle command. The only small overshoots, and with no coupling effect
energy based control logic however reduces the climb visible in the other variable. Figure 16 shows an exem-
angle as necessary to maintain the airspeed. On the plary climb and descent manoeuvre, while Figs. 17
right image in Fig. 13 a 8 ms speed increase at con- and 18 show different step commands in airspeed.
stant altitude is shown. The coordination of throttle
and elevator is clearly noticeable for the energy based
controller. In the SISO case, the altitude error reaches 6 Retard and Ground Lock
2.5m, while the energy based controller keeps it under
1m. Before a touchdown on top of the landing platform
For the directional control of the aircraft, a conven- can be attempted, the aircraft engine has to be shut
tional SISO controller using the ailerons as actuators down in order not to damage the aircraft or the landing
was implemented as shown in Fig. 14. platform. This so-called “retard” takes place during
The flight control system was tested in flight exper- a critical phase of the landing manoeuvre, where the
iments and performed as desired [5]. In straight and aircraft is very close to the landing platform. A sud-
level flight, the controller was able to keep the alti- den loss of thrust force (throttle cut-off) will produce
tude within ±1m and the airspeed within ±1.5 ms of a deceleration of the aircraft, consequently a decrease
the desired value, as shown in Fig. 15.

Fig. 18 A 3 ms speed decrease with the energy based flight con-


Fig. 16 Climb and descent with the energy based flight con- troller. Again, good tracking of altitude and airspeed. At t = 14s,
troller. The manoeuvres are nicely coordinated. The airspeed the airspeed increases, probably due to a wind gust, but it is
tracking works just as well as during level flight (commanded quickly brought back to the desired value (commanded values
values are displayed in blue, actual sensor data in red) are displayed in blue, actual sensor data in red)
610 J Intell Robot Syst (2017) 88:597–618

Fig. 19 Ground tests with aircraft on top of landing platform to determine the best strategy for the ground lock mechanism

of lift force and result in an increasing relative veloc- As mentioned before in Section 4.4, a non-zero ver-
ity and position in x- and z-direction (backward and tical velocity is required for the landing and needs
downward motion). to be configured correctly (ideally 0.2 ms at touch-
At the same time the aircraft is already close enough down). Therefore, the effects of both modes have
to the ground vehicle to experience the aerodynamic been analyzed during virtual landing approaches at
disturbances caused by the ground vehicle (increased higher altitudes and the produced additional down-
airspeed and angle of attack). This so-called “UGV ward velocities ḣ(t) have been determined through
flow effect” will partly compensate the downward direct measurement. The final (relative) activation
motion but increase the backward motion due to the altitude was set to 0.6m for “retard mode” and 0.2m
higher aerodynamic drag. for “ground lock mode”.
These complex effects have to be compensated by It should be noted, that the presented locking mech-
the flight controller, while now (after engine cut-off) anism has been adapted to the demonstrator setup and
only the elevator is available to control either altitude is not directly transferable to a larger aircraft. For
or velocity. The decision was taken in favor of velocity the final large scale landing system, the ground lock
control; the desired vertical velocity at touchdown is mechanism has to be reworked, since a landing on a
achieved by adjusting the altitude at which retard is stretched net will not be possible anymore. Among
activated. other things, this might include the introduction of pre-
In the last moments before touchdown the so-called defined attachment points on the aircraft aswell asanattitude
“ground-lock” configuration is activated. Thereby the alignment and locking system on the ground side.
ailerons are deflected to the maximum upward posi-
tion symmetrically to reduce lift, the elevators are
deflected downward to half of the maximum travel 7 Ground Vehicle Control
to generate a nose down moment, and the flaps are
fully retracted. This configuration showed the best Early simulations have shown that relatively slow
performance results during a number of ground tests, ground vehicle dynamics are sufficient for the execu-
where the actual aircraft was placed on top of landing tion of a successful cooperative landing manoeuvre.
platform at around 70 kmh . It was also found that a pitch The choice was therefore taken in favor of a manned
damper had to remain active in order to reduce possi- “semi-autonomous” ground vehicle. In this solution
ble oscillations once the aircraft is lying on the net of a human driver has to execute the control commands
the landing platform, as well as a locking mechanism, provided by the ground vehicle controller. The human
assuring that the front landing gear remains locked-in basically replaces the mechanical actuator or driv-
once it crossed the landing platform net, as illustrated ing robot. A Graphical User Interface (GUI) mounted
in Fig. 19. inside the car is used for displaying the corrective
J Intell Robot Syst (2017) 88:597–618 611

Fig. 20 Graphical User Interface (GUI) for human driver commands

action the “human actuator” has to perform. There- the last 15s before touchdown remains within a bound
fore a screen was mounted inside the ground vehicle of around 0.6 ms .
as shown in Fig. 20. The delay in the beginning of the acceleration phase
The GUI has been designed as a two-colored at around 12s into the landing manoeuvre is due to
crosshair with the horizontal line indicating throttle the delay introduced by the ground vehicle’s auto-
commands and the vertical line representing steer- matic transmission and the human reaction time. The
ing wheel deflection commands. Both, the ground notch in actual UGV velocity at around 15s is due to
vehicle velocity and course controllers were imple- the automatic transmission shifting gears. The ground
mented as proportional controllers, directly translating vehicle starts to decelerate and at around 38s due
velocity/course command into throttle/steering wheel to the fact, that only a relatively short taxiway with
deflection commands, allowing for the adjustment of a length of only 500m was used for the performed
the closed-loop ground vehicle dynamics despite the experiments. Only around one second after the virtual
human driver. The ground vehicle has been equipped touchdown, the ground vehicle reached the end of the
with the same real-time computer and sensor system available runway and had to decelerate, leading to an
(RTK GPS, IMU) as present on the UAV. Data is increase in desired UGV velocity.
exchanged via a wireless network. During first landing
experiments with a simulated UAV as described in [3],
the human driver was capable of following the given 8 Ground Vehicle Aerodynamics
throttle commands reasonably well. As can be seen in
Fig. 21, the actual ground vehicle velocity follows the The aerodynamic disturbance caused by the ground
desired velocity, where the maximum deviation during vehicle was identified as an important issue and

Desired vs. Actual UGV Velocity


25
desired UGV velocity [m/s]
UGV ground speed [m/s]

actual UGV velocity [m/s]


20

15
Touchdown at 37s
10

0
0 10 20 30 40 50 60
manoeuvre time [s]

Fig. 21 Ground vehicle desired vs. actual velocity during a virtual landing experiment (real ground vehicle, virtual aircraft)
612 J Intell Robot Syst (2017) 88:597–618

UGV (velocity vector magnitude and direction at a


specific point in space), as well as the optimal landing
trajectory.
Figure 22 illustrates the resulting flow situation as
well as the selected final landing trajectory. A vertical
descent above the middle area of the landing platform
induces the smallest variations in airflow magnitude
and direction.
Figure 23 shows the distribution of both, the induced
angle of attack (black lines) and the induced rela-
tive velocity (red lines) in the longitudinal and lateral
Fig. 22 Visualization of disturbed flow around the ground
vehicle and optimal landing tracectory direction of the ground vehicle at 1m above landing
platform. A detailed analysis of the results and their
impact on the landing system design can be found in [3].
needed to be analyzed in detail. Therefore a numerical
CFD simulation (Computational Fluid Dynamics) has
been performed using a realistic geometrical model of 9 High-Level Mission Control
the landing platform. The airflow characteristics have
been calculated in three dimensions in order to cap- A state machine has been developed to provide high-
ture the highest possible detail of the present airflow level mission control. The initial state corresponds to
conditions. The goals of the CFD simulation consisted waypoint mode, where the UAV flies a 8-shaped pat-
in the determination of the change of airflow veloc- tern above the airfield (Fig. 24). The waypoints were
ity and angle of attack induced by the presence of the set manually, taking into account pilot line-of-sight,

Flow angle and velocity distribution -3m<x<5.5m, y=0m, z=3m


4 1.08
alpha
3
rel. velocity 1.06
landing platform

rel. velocity [-]


4%
alpha [deg]

2 1.04
2.5%
1 1.02

0 1

-1 0.98
-3 -2 -1 0 1 2 3 4 5 6
x-position [m]

Flow angle and velocity distribution x=2m, -3m<y<3m, z=3m 4%


1 1.04
alpha
rel. velocity
rel. velocity [-]
alpha [deg]

0.5 1.02
landing platform

0 1
-3 -2 -1 0 1 2 3
y-position [m]

Fig. 23 Longitudinal (x) and lateral (y) distribution of induced angle of attack (black) and relative velocity (red) at 1m above landing
platform
J Intell Robot Syst (2017) 88:597–618 613

described earlier is active, both in longitudinal and


lateral direction, and the UAV is descending. Above
the flare height of hf lare = 5m, the flight path
angle is constant at γapproach = −3◦ . Once the flare
height has been reached, a flare law is active, reducing
the vertical speed proportionally to the height above
the landing platform. At 0.6m above the platform,
the motor is turned off, and at 0.2m ground lock is
activated. Once the motor is turned off, the UAV is
committed to the landing and will not perform any
go-around manoeuvres.
As can be seen in Fig. 25, during the flare the inner
fence sets relatively tight constraints of ±0.8m in both
longitudinal and lateral direction. This is necessary to
ensure a safe landing, since larger deviations from the
center of the landing platform may result in a part of
the UAV hitting the aluminum frame of the platform
and getting damaged. Above hf lare , the allowed devi-
ation increases as a linear function of the altitude, with
a very flat slope behind the car, because that is the
Fig. 24 Waypoint pattern used for the landing manoeuvre direction from which the UAV is approaching.
If the UAV leaves the boundary defined by this
inner fence before reaching the flare height, the
obstacles (buildings, trees) and wind conditions. Most descent is stopped (altitude hold) while the cooper-
of the pattern is flown at an altitude of 110m above ative control in lateral and longitudinal direction is
ground and an airspeed of 23 ms , except for the section still active. As soon as the UAV is inside the inner
between WP1 and WP2, where the UAV initiates a fence again, the descent is resumed. If the inner fence
rapid descent to 30m at an increased airspeed of 28 ms . is crossed after hf lare has been reached, the cooper-
The increased airspeed command allows the UAV to ative control remains active as well, but the aircraft
reach the necessary descent angle. The UAV then stays is commanded to climb to hf lare + 1m, as it is
at 30m altitude for a short distance, which gives the considered unsafe continue flying at such a low alti-
operator the opportunity to command the start of the tude. Flare is only resumed when the UAV is inside
landing manoeuvre. If no such command is given, the inner fence and at least 3m above the landing
the UAV will climb back to 110m and resume the platform.
waypoint pattern. While the UAV is in flare mode, an abort is also
Once the landing is initiated by the operator, the triggered if the communication lag exceeds 0.3s,
UAV’s actions are determined by its position relative resulting in the same actions as if the UAV has left
to the landing platform. Figure 25 gives an overview the inner fence. A too high lag would result in unreli-
of the different zones and their corresponding actions. able data for the position of the ground vehicle, and a
The main divisions consist of an “inner fence” and an landing would be too risky.
“outer fence”. As long as as the UAV is within the If, at any point in the manoeuvre, the UAV crosses
inner fence, the landing proceeds normally. Between the outer fence, i.e. reaches the red zones in Fig. 25,
the inner fence and the outer fence, the approach the manoeuvre is aborted altogether, and the UAV
is suspended, but the UAV still tries to reach the flies in runway direction while climbing to an altitude
inner fence in order to resume the landing. Outside of of 20m, before resuming the waypoint pattern. The
the outer fence, the manoeuvre is aborted, the UAV UAV will not resume the landing manoeuvre without
performs a go-around and resumes waypoint flight. a command from the operator.
As long as as the UAV is within the inner fence, the During the entire landing manoeuvre, after the start
landing proceeds normally, i.e. the cooperative control command was given by the operator, no further input
614 J Intell Robot Syst (2017) 88:597–618

Fig. 25 Geometrical conditions for state machine logic (side view)

is required from the operator. However, at any point The landing was successfully performed multiple
the operator is able to initiate a go-around by a manual times. The required landing precision of ±0.8m was
command, resulting in the same go-around manoeuvre consistently achieved.
that is performed when the UAV leaves the outer As described in Section 9, the landing manoeu-
fence. During the go-around, whether it was initiated vre was automatically controlled via a high-level state
automatically or manually, the operator can also com- machine logic. During a typical successful landing
mand a landing again, but this command only has an manoeuvre sequence, the different states are passed
effect if the UAV is inside the outer fence.

10 Landing Experiments

As described in [3] and [6], the proposed landing


system has been successfully tested in a number of
simulation experiments, including a so-called Hard-
ware In the Loop (HIL) test setup, where the real
ground vehicle was used together with a virtual air-
craft. The concept allowed for an investigation of the
expectable landing system performance with very lit-
tle risk involved, since only a virtual aircraft was
performing the landing manoeuvre. With the suc-
cessful execution of the HIL-experiments, the next
set of experiments was initiated, including the final
experimental setup as illustrated in Fig. 26. Fig. 26 Experimental setup for final landing experiments
J Intell Robot Syst (2017) 88:597–618 615

issued. The aircraft is thus commanded to climb above


6m in order to start the second approach. During the
second approach the relative x- and y-error remain
within the required bounds and the landing is achieved
with a touchdown at around 48s. The exponential flare
law is nicely visible in the altitude profile between sec-
ond 26 and 46. The effect of retard and ground lock
mode activation can be seen during the last two sec-
onds before the touchdown at around 46s, where the
relative altitude decreases rapidly. After second 48 the
aircraft is lying in the landing platform (slightly nega-
tive relative altitude due to downwards extended net),
the small oscillations occurred during the decelera-
tion process. The relative vehicle velocities during the
exemplary landing attempt are shown in Fig. 28:
Fig. 27 Relative vehicle positions during exemplary landing As can be seen, the manoeuvre starts with a rela-
attempt tive x-velocity (blue line) of around 21 ms , where the
ground vehicle starts accelerating at around 10s into
through successively in the following order: Way- the landing manoeuvre and reaches the UAV veloc-
pointMode, Descent, Flare, Retard, and GroundLock. ity at around 18s. The aircraft is descending with a
This sequence will generate a descent path within the constant vertical velocity (red line) of around −1 ms
previously defined “outer fence” (non-red areas of between second 6 and 25, which corresponds to the
Fig. 25). The desired landing airspeed was set to 21 ms commanded 3◦ descent slope. The retry command
throughout all performed landing experiments. In the can be clearly seen at around 25s, where the air-
following one exemplary landing attempt is described craft is commanded to climb above 6m for a second
in detail. attempt. The vertical velocity profile during the flare
The relative vehicle positions during the exemplary is also visible as an exponential curve between se-
landing attempt are shown in Fig. 27. cond 30 and 48 (see right image in Fig. 28). The
As can be seen, the state machine has entered the small oscillations after the touchdown at around 48s
flare (relative altitude ¡ 5m) at around 25s into the are due to the aircraft bouncing slightly on top of
landing manoeuvre. At that moment the required pre- the landing platform during the deceleration process.
cision of ±0.8m in x- and y-direction (red dotted lines) Figure 29 shows the successful touchdown onto the
was not given and an automatic retry command is landing platform.

Fig. 28 Relative vehicle velocities during exemplary landing attempt (right: vertical zoom)
616 J Intell Robot Syst (2017) 88:597–618

Fig. 29 Successful touchdown on top of the landing platform

The ground vehicle velocity tracking performance fence margins just after crossing hf lare at 5m altitude,
during the selected landing attempt is shown in it will try to recover the horizontal misalignment while
Fig. 30, demonstrating the validity of the human actu- maintaining altitude and minimizing parasitic effects
ator concept. on the horizontal controller performance.
After the first set of landing experiments were con-
cluded, a couple of changes have been implemented
to further improve the performance of the landing sys-
tem. The initially applied constant climb angle of 10◦
during a low-retry (below hf lare ) was affecting the
system performance, as it would induce a rather abrupt
climb manoeuvre, even if the aircraft was still high
above the landing platform. Although the flight con-
troller is able to decouple aircraft responses in velocity
and altitude, sudden changes are not always perfectly
handled. An adaptive climb angle was therefore intro-
duced, where the UAV would only climb abruptly
when close to the landing platform and smoother when
further away:
π
γclimb = (hf lare − h) · (27)
72
This results in a hold altitude behavior at h =
hf lare and a commanded climb angle of 12.5◦ at h = Fig. 30 Ground vehicle desired velocity vs. actual velocity for
0m. Therefore if the aircraft is pushed out of the inner the selected landing attempt
J Intell Robot Syst (2017) 88:597–618 617

vre have been performed, and it was shown that the


proposed cooperative control algorithms are effec-
tive and the landing manoeuvre in general is feasible.
The inclusion of the relevant effects, such as the
aerodynamic influence of the ground vehicle, greatly
increased the confidence in the simulation results. The
simulation results have been validated by first cooper-
ative landing experiments using a HIL-test setup with
the real ground vehicle and a simulated aerial vehicle
[3].
The final landing manoeuvre has been success-
Fig. 31 Comparison of relative state estimates based on 1. GPS fully executed multiple times. The landing was pos-
data exchanged via the network (red), 2. pure optical marker sible with an accuracy of less than 0.8m, which is
tracking (green), and 3. combined solution by sensor data fusion precise enough to perform safe and repeatable land-
(blue). At t = 0s the touchdown is performed
ing manoeuvres with the current setup. Synchronized
close proximity flight was possible with position-
ing accuracies of less than 0.8m despite the ground
Another important modification consisted in the vehicle induced aerodynamic disturbance, which was
introduction of vision-based state estimation data into found to correspond well to the predictions from CFD
the active control loops. The methods have been simulations (as presented in [3]). Thereby the aircraft
described in Section 3. Figure 31 shows the results of was fully immersed in the ground vehicle induced
a recently performed landing experiment where vision aerodynamic disturbance. The flight controller effi-
data was used for control. ciently compensated for the UGV flow effect. The
As can be seen, all three solutions lie close together relative state estimation system worked reliably, the
with very small deviations during the last meters inclusion of vision data resulted in an increased oper-
before touchdown. The optical marker tracking works ational safety. The ground lock mechanism performed
well down to a relative altitude of around 1m, where as expected and secured the aircraft in the landing plat-
the markers cannot be tracked anymore. It should form during the deceleration phase until a full stop was
be noted, that even though a very precise RTK-GPS reached. Nevertheless, the locking mechanism needs
system was used during these experiments, the rel- to be reworked for a use on large scale aircraft as
ative position estimates depend on the accuracy of explained in Section 6.
the lag estimate which could possibly induce smaller It can be concluded that automated landings on the
errors. The comparison of the GPS-based solution to mobile platform are feasible with the current exper-
the purely vision-based estimates confirms the proper imental setup and that the system is suitable for a
functionality of the lag correction method. The fused practical application. The next steps will include the
solution is optimally suited for a use in the coupled adaptation of the system to a large scale aircraft, where
controller, as it accounts for GPS or vision system an appropriate mechanical interface is required for the
outages and ensures a continuous mode of operation. final ground lock phase, possibly including active atti-
A landing with degraded GPS performance is there- tude tracking. This task will be initiated once the final
fore also possible, which significantly increases the aircraft specifications are available, as the ground lock
reliability and safety of the landing system. mechanism will significantly differ from on aircraft to
another.

11 Conclusions and Outlook


Acknowledgments The first author would like to thank Felix
The landing of a fixed-wing UAV on top of a mobile Huber, Jonatas Sant’Anna Santos, Stojan Stevanovic, and Dr.
Tobias Praßler for their strong and essential support during the
ground vehicle has been analyzed and algorithms for landing experiments. This work was partially supported by the
the cooperative control of both vehicles were pro- European Commission FP7 ICT Programme under the project
posed. Extensive simulations of the landing manoeu- EC-SAFEMOBIL 288082.
618 J Intell Robot Syst (2017) 88:597–618

References Georg Balmer received his Master’s degree in aerospace engi-


neering from KTH Royal Institute of Technology, Stockholm in
2015. He is currently a member of the Flying Robots research
1. Lütjens, K.H., et al.: Airport2030 – Lösungen für den
group at the Robotics and Mechatronics Institute of the German
effizienten Lufttransport der Zukunft Deutscher Luft- und
Aerospace Center DLR. His main research interests are flight
Raumfahrtkongress (2012)
control of fixed-wing unmanned aircraft and mission planning
2. Rohacs, D., et al.: Evaluation of Landing Characteristics
under consideration of weather influences.
Achieved by Simulations and Flight Tests on a Small-
scaled Model Related to Magnetically Levitated Advanced
Take-off and Landing Operations. International Council of
the Aeronautical Sciences (2014) Linnea Persson received her Master’s degree in control and
3. Muskardin, T., et al.: Landing of a Fixed-wing UAV on a robotics from KTH Royal Institute of Technology, Stockholm in
Mobile Ground Vehicle. IEEE International Conference on 2016. She is currently a doctoral student at the Automatic Control
Robotics and Automation (2016) department at KTH, working within the topic of autonomous
4. Muskardin, T., et al.: A Novel Landing System to Increase transportation systems. Her main research interests are in the
Payload Capacity and Operational Availability of High optimization and control of distributed, multi-agent systems.
Altitude Long Endurance UAV. International Conference
on Unmanned Aircraft Systems (2016)
5. Balmer, G.: Modelling and Control of a Fixed-wing UAV
Sven Wlach received his diploma in aerospace engineering
for Landings on Mobile Landing Platforms. Master’s thesis,
from the University of Stuttgart in 2015. He is currently a
KTH Royal Institute of Technology (2015)
member of the Flying Robots research group at the Robotics
6. Persson, L.: Cooperative Control for Landing a Fixed-Wing
and Mechatronics Institute of the German Aerospace Center
Unmanned Aerial Vehicle on a Ground Vehicle. Master’s
thesis, KTH Royal Institute of Technology (2016) DLR. His main research interests are high altitude and flexible
7. Kloeckner, A.: Geometry based flight dynamics modelling unmanned aircraft.
of unmanned airplanes. AIAA Modeling and Simulation
Technologies (MST) Conference Boston (2013)
Maximilian Laiacker received his diploma in computer sci-
8. Lambregts, A.A.: Vertical flight path and speed control
ence from the Technical University of Berlin in 2009. He is
autopilot design using total energy principles. AIAA paper
currently a member of the Flying Robots research group at the
83–2239 (1983)
9. Lambregts, A.A., Creedon, J.F.: Development and flight Robotics and Mechatronics Institute of the German Aerospace
evaluation of automatic flare laws with improved touch- Center DLR. His main research interests are vision based state
down dispersion. AIAA paper 80–1757 (1980) estimation, collision avoidance and sensor fusion for flying robots.
10. Laiacker, M., et al.: Vision Aided Automatic Landing Sys-
tem for Fixed Wing UAV. Intelligent Robots and Systems
(IROS) Tokyo (2013) Anibal Ollero is Full Professor at the University of Seville,
11. Wagner, D., Schmalstieg, D.: Artoolkitplus for pose track- Spain. He is also the head of the University of Seville Robotics,
ing on mobile devices. 12th Computer Vision Winter Vision and Control research group and the Scientific Director
Workshop (CVWW’07), pp. 139–146 (2007) of the Center for Advanced Aerospace Technologies (CATEC)
12. Tanaka, H., Sumiand, Y., Matsumoto, Y.: A high-accuracy in Seville, Spain. He received his PhD in Engineering and his
visual marker based on a microlens array. Intelligent Robots Electrical Engineering degree from the University of Seville in
and Systems (IROS), pp. 4192–4197 (2012) 1980 and 1976 respectively. He was also Full Professor at the
13. Drela, M., Youngren, H.: Athena Vortex Lattice. http:// Spanish Universities of Santiago and Malaga; and researcher at
raphael.mit.edu/avl (2004) LAAS-CNRS in Toulouse, France, and at Carnegie Mellon Uni-
14. Drela, M., Youngren, H.: XFOIL. http://raphael.mit.edu/ versity in Pittsburgh, PA, USA. He is the coordinator of several
xfoil (2000) national and international projects.
15. Lambregts, A.A.: Generalized automatic and augmented man-
ual flight control. BerlinTechnicalUniversity Colloquium (2006)
16. Kastner, N., Looye, G.: Generic TECS based autopilot Konstantin Kondak is the head of the Flying Robots research
for an electric high altitude solar powered aircraft. CEAS group at the German Aerospace Center (DLR) Institute of
EuroGNC (2013) Robotics and Mechatronics in Wessling, Germany. He received
his Computer Science degree from the Technical University
of Berlin (TUB) in 1999. He also received his PhD and post-
Tin Muskardin received his diploma in aerospace engineering doctoral degree in Robotics from the same university in 2001
from the University of Stuttgart in 2009. He is currently a mem- and 2006, respectively. He was also the head of the Real-Time
ber of the Flying Robots research group at the Robotics and Systems and Robotics research group at TUB (representative
Mechatronics Institute of the German Aerospace Center DLR. full professorship) working on different subjects in computer
His main research interests are flight control of fixed-wing science and robotics, including exoskeletons and UAVs. He
unmanned aircraft and cooperative control of multi-robot systems. participates in several national and international projects.

Potrebbero piacerti anche