Delft University of Technology

Position controller for a flapping wing drone using uwb

Gonzalez Archundia, G.; de Croon, G.C.H.E.; Olejnik, D.A.; Karasek, M.

Publication date
2021
Document Version
Final published version
Published in
Proceedings of the 12th International Micro Air Vehicle Conference

Citation (APA)
Gonzalez Archundia, G., de Croon, G. C. H. E., Olejnik, D. A., & Karasek, M. (2021). Position controller for
a flapping wing drone using uwb. In J. Martinez-Carranza (Ed.), Proceedings of the 12th International Micro
Air Vehicle Conference (pp. 85-92). [IMAV2021-10]
Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.

Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.

This work is downloaded from Delft University of Technology.

For technical reasons the number of authors shown on this cover page is limited to a maximum of 10.
 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

Position controller for a flapping-wing drone using UWB

Guillermo González*, Guido C.H.E de Croon, Diana Olejnik and Matěj Karásek
Delft University of Technology, Mekelweg 5, Delft

A BSTRACT drones (e.g. VICON/OptiTrack), but it is an expensive option

and its accuracy can be affected by lighting conditions [4].
This paper proposes an integral approach for Another option for IPS is ultra wide band (UWB) which
accurate ultra wide band indoor position control was already defined in 2006, but just recently started to gain
of flapping wing micro air vehicles. Three popularity. UWB is able to transmit in nano-second scale
aspects are considered to reach a reliable and periods. Thus it allows excellent timing for signal arriving,
accurate position controller. The first aspect which translates to centimeter level accuracy. Its low spec-
is a velocity/attitude flapping-wing model for tral density reduces the interference with other RF devices.
drag compensation. The model is compared However, it also has some drawbacks like high computational
with real flight data and shown to be applicable and memory capacity of the controller and its vulnerability
for more than one type of flapping wing drone. to multi-path effect when signals bounce with the physical
The second improvement regards a battery-level boundaries of the environment [5].
dependent thrust control. Lastly a characteri- Inertial navigation systems (INS) are usually added in in-
http://www.imavs.org/

sation of ground effects in flapping-wing flight door environments to complement IPS. Usually INS systems
is obtained from hovering experiments. The employ micro-electro mechanical systems (MEMS) or iner-
proposed controller improves position control tial measurement units (IMU). Furthermore, data fusion is
by a factor ∼ 1.5, reaching a mean absolute also applied for better accuracy, by means of a Kalman fil-
error of 10cm for position in x and y, and 4.9cm ter. Hence many combinations of integrated systems have
for position in z. been proposed. For example a GPS/UWB/MEMS navigation
system with Kalman filter [6] and INS/UWB system based
on a fuzzy adaptive Kalman filter [7]. Although most of
1 I NTRODUCTION these implementations provide a basis for designing an au-
The fact that drones are becoming increasingly popular tonomous position controller, it is important to keep in mind
is intimately related with the development of more sophis- they are specifically designed either for fixed-wing or quadro-
ticated automation resources. In the case of drones, weight, tor UAVs. In the case of flapping wing drones some extra
processing speed, and energy consumption are critical aspects constraints shall be considered like the physical limitations in
to accomplish fully autonomous flights. Therefore, elements terms of payload and energy consumption, or the noise influ-
such as motors, micro-processors, memory units and batteries ence in IMU measurements due to high frequency mechanical
need to be continuously improved to meet these requirements. vibration [8]. To overcome these circumstances, the existing
These conditions have fostered the development of flapping solutions must be adapted, leaving a potential research devel-
wing micro-air vehicles (FWMAV). This type of UAV is in- opment for position controllers of FWMAVs.
spired on the flight of birds and insects, and has become at- One of the challenges of working with FWMAVs comes
tractive in the field of small-scale micro-air vehicles, since it from the aerodynamics of the system. A reliable model of
provides both the ability of hovering and flying into any di- the dynamics can significantly increase the performance and
rection. After the foundational work from Ellington [1] and accuracy. Nevertheless, most aerodynamic models for flap-
Dickinson [2], many experiments have been carried out in or- ping wings require extensive system identification techniques
der to get an optimal physical design of FWMAVs, as well as for numerous parameters, whose values only remain valid
velocity and position controllers to keep a stable flight. for a specific drone. A widespread alternative is the use of
Regarding position feedback, the most common option is quasi-steady models where force coefficients are obtained ei-
an indoor positioning system (IPS). Some of the technolo- ther from experimental data or from theoretical principles [9].
gies used for for IPS are Wi-fi, Radio Frequency Identifica- The control strategy proposed in this paper follows a sim-
tion (RFID), and infrared (IR) motion tracking systems. Wi-fi pler solution where the model is obtained by directly aver-
and RFID can be found for tracking mobile devices, but not aging aerodynamic parameters as functions in terms of the
specifically for drones [3]. IR is used for indoor tracking of body velocities. Albeit the method may be considered just
a rough approximation compared to quasi-steady models, it
* Email address(es): g.gonzalezarchundia@student.tudelft.nl,
g.c.h.e.decroon@tudelft.nl,
tends to be more practical since just a reduced amount of pa-
d.a.olejnik@tudelft.nl, rameters is required. When validating averaging parameters
matej@flapper-drones.com with real flight data, several authors address the issues of un-

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 85

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

steady flight due to mechanical vibrations [10, 11]. However

just some models follow an approach based on drag compen-
sation. Within the few cases where drag compensation mod-
els are implemented, most of them are linear models, which
are only valid for a limited range of velocities [8, 12].
Similar to most MAVs, FWMAVs are also prone to short
flights due to the limited size of the batteries they are able
to carry. Thus it is common that voltage will significantly Figure 1: Cyberzoo structure where the UWB anchors are
change during flight, affecting as well the required throttle placed (left) and detail of UWB anchor mounted on the struc-
level for hovering [13]. This condition poses a challenge in ture (right)
height control, for both reaching and keeping a specific posi-
tion along the z axis. (Figure 2).
Another issue affecting height control is the ground effect. About the flapping-wing drone, referred in this paper as
This phenomenon particularly occurs when the drone is flying Flapper, is a design from the company Flapper-drones [19]. It
close to the ground [14]. An extra thrust is generated because is a 102g tailless FWMAV with a wingspan of 49cm, able to
the wind currents underneath the drone bounce against the keep flapping frequencies up to 12 Hz when hovering. The
ground, causing the drone to behave like sitting on a cushion on-board processing hardware consists of a Crazyflie Bolt
of air [15]. Thus, an extra lift is generated, causing the output autopilot board, including an IMU with 3-axis accelerom-
thrust to be higher than the input thrust. eter/gyroscope (BMI088). The data link between the au-
http://www.imavs.org/

This paper addresses the three aforementioned challenges topilot and the ground station is done with a Crazyradio PA
of FWMAVs. Sections 2, 3 and 4 provide the background (also from Bitcraze), which is a USB radio dongle based on
for the implementation. Section 5 describes how averaging nRF24LU1+ from Nordic Semiconductor. The system can be
parameters are applied for drag compensation. Section 6 powered with a 300-mAh two-cell 7.4V LiPo battery, reach-
discusses the issues of changing voltage and ground effects. ing a flight time between 4-6 minutes, depending on the in-
Lastly, Section 7 shows the results on how the transient re- ternal resistance of the battery.
sponse is improved in any of the three axes: forward (x-axis),
sideways (y-axis) and vertical (z-axis).

2 E XPERIMENTAL SETUP
Since the work on this project relies mainly on feed-
back through UWB, the experimental setting should be a grid
where UWB receivers are strategically located in the vertices
and the transceiver is mounted on the FWMAV. This way, the Figure 2: Flapper drone equipped with IR markers and UWB
drone flies inside the volume of a cube bounded by the posi- sensor
tion of the anchors. A similar set-up is used in [16], where the
eight anchors on the corners optimise the possibility for the
drone to have line-of-sight at least with one anchor. Such a 3 C ONTROL LOOP
grid was set-up at the Cyberzoo (shown in Figure 1), the flight For autonomous flight, the flapping-wing drone uses a
arena of the TU Delft Faculty of Aerospace, which is known cascaded PID-controller, based on the structure presented by
as the main facility of the university for performing tests on [20]. The controller consists of three loops. The output pro-
drones, and has a size of 10m × 10m × 7m. A reliable choice vides the control signals for the motors involved in the flight
for UWB sensor is the Decawave DWM1000, as it has proven dynamics of the drone: two brush-less motors in charge of
to give successful results for UAV tracking [17]. The UWB the flapping frequency for thrust and roll, one servo-motor
sensors are used as the anchors of the IPS and are set to work that modifies the angle of the dihedral for pitch, and another
using a time difference of arrival (TDOA) algorithm. servo-motor that changes the deflection of the wings for yaw.
In order to have a reference for the UWB measurements, a Regarding the structure of the cascaded loops, the inner-
motion tracking system is used for measurements of position most loop is in charge of attitude rate and runs at 500Hz. The
and rotations. It consists of 12 OptiTrack Prime 17W mo- intermediate loop also runs at 500Hz and is in charge of atti-
tion tracking cameras (set to resolution 1664 px × 1088 px, tude control. The outer loop is the position/velocity control,
50 fps) and has proven to deliver accurate results for MAV which runs at 100Hz and can take either position or velocity
test flights [18]. When working with the motion tracking sys- commands.
tem, the drone was equipped with four retro-reflective mark- Another remark is the trimming values for servos in
ers placed on the landing gear of the drone, and the UWB charge of pitch, roll and yaw. This procedure must be done
sensor was placed at the top to optimise direct line of sight in order to get the proper behaviour from the controller. For

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 86

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

trimming, one should fly the drone manually until it hovers at

a steady position. The trim values are the pitch, roll and yaw
commands given to keep the drone hovering. The values are
specific for each flapping-wing drone, as they change depend-
ing on manufacturing factors. Thus it is important to properly
set the appropriate trimming values before going further with
any flight test.
4 S TATE ESTIMATION
The state estimation is done by means of an Extended
Kalman Filter (EKF) which fuses the measurements of the
IMU and the UWB positioning system. The model of the
EKF is the same as the one proposed by [21], which consid-
ers a nine-dimensional state vector defined as:
 T
S = x E vB d (1)
Where xE is the position vector in global frame, vB is the ve- Figure 3: Effects in position estimation before using variable
locity vector in body frame and d is an attitude error vector, sensor noise (left) and after implementing it (right).
where the error is defined as the difference between the last
http://www.imavs.org/

measured attitude and the current attitude. The purpose for Once the variable sensor noise term was implemented
using attitude errors instead of the conventional Euler angles into the EKF, several flight paths were tested (linear, square,
is to simplify the state prediction equations because it only rhombus and hexagonal paths). The measurements from
considers the increments in roll, pitch and yaw. A more de- UWB using a TDOA algorithm proved to be reliable enough
tailed explanation on the implementation of the filter is given for position estimation. The mean absolute error obtained in
in [21]. all cases remained between 8-10 cm, when compared to the
The main advantage of fusing UWB with IMU through measurements from the Optitrack system.
the Kalman filter is the attitude correction to account for the
drift in pitch and roll caused by sensor noise and bias. For in- 5 D RAG COMPENSATION
stance, when the Flapper is left standing up, the Kalman filter
The first technique for yielding a better control strategy
resets correctly the roll and pitch to zero when UWB mea-
for FWMAVs was to obtain a model for drag compensation.
surement are coming in. On the other side, a complementary
Since the drag is typically neglected in aerodynamic models,
filter, which relies only on IMU data, will converge to a cer-
there tends to be an offset between the commanded veloc-
tain drift and thus propagates it through time.
ity and the velocity output by the controller. Usually a feed
One last adjustment done to the EKF aims to diminish the
forward term is used to compensate for these drag effects.
detrimental effects on position estimation caused by multi-
Following the structure of the controller in Section 3, the ve-
path UWB signals. Specially when going close to the ground,
locity loops provides the input for the attitude loop. Thus,
the position estimates tend to drift for more than 20 cm.
feed forward is modelled as a function of velocity:
A way to account for the wrong measurements when flying
close to the ground is to use a variable sensor noise value
θF F = f (vxE ) (3)
Ruwb for measurements from UWB, rather than a constant
value. Consequently, the sensor noise is defined as:
ϕF F = f (vyE ) (4)
2
Ruwb = λ(ẑ)σuwb (2)
Where Eq. (3) is the feed forward term for pitch and Eq. (4)
Where λ(ẑ) is a factor dependent on the estimated height ẑ, is for roll. In order to derive such a model, a system identi-
in meters. For the implementation, λ = 0.5, for z ≥ 1; fication experiment is proposed based on the supplementary
λ = 1.5 − ẑ, for 0.5 < ẑ < 1; and λ = 1, for z ≤ 0.5. materials from [22], where step inputs in roll and pitch are
The effect of variable sensor noise in position estimation was given to the flapping-wing drone and then the transient re-
tested by placing the drone static on the ground, at the origin sponse is recorded. The experiments are done using manual
(0,0,0) of the UWB IPS. The plots from Figure 3 show the flight via a Frsky RC-controller where the pitch and roll step
estimated position (est) in contrast to the real position (cmd), inputs are pre-programmed. The tracking data is recorded
Both values were sampled at 50Hz, during 8 seconds. Using with the motion tracking system mentioned in Section 2.
the variable sensor noise Ruwb significantly increases the ac- Figure 4 is given as an example of the transient response
curacy in estimation approximately by 0.2m in the three axes. in velocity obtained for different step-inputs in pitch (a sim-
ilar response is obtained for roll angles). Due to limitations

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 87

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

4
Velocity
As expected, results from Table 1 show on the angles of
Pitch = 12°
Pitch = 22°
Pitch = 33°
accurate ground truth (12º,22º and 33º) that the overall accu-
3
Pitch = 45°
Pitch = 52° racy of the extrapolation model, for both pitch and roll, in-
V xE [m/s]

2 creases as more data points are taken for the regression. Indi-
vidually, for each angle, accuracy remains approximately the
same with 50% of the data points or more. For illustration
1

0 purposes, Figure 5 shows the predicted steady-state velocity

for each angle when taking 70% of the data points for the
0 1 2 3 4 5 6
Time [s]

regression model. Nevertheless, the real steady-state values

Figure 4: Velocities attained at different pitch step inputs used were the ones obtained taking all of the data points since
those ones result in the lowest error.
on the size of the flight arena, only for the small angles (be- Once the steady-state velocity for each angle is available,
low 45°) it is possible to reach the steady-state velocities. For the pitch-velocity and roll-velocity models can be obtained.
higher angles, a common technique to reach the steady-state Given the mapping of the steady-state velocity/attitude pairs,
velocities is to perform wind-tunnel experiments [23]. In this different regression models were tested to approximate the
case, an alternative method is used where a non-linear regres- relationship the models (linear, quadratic and exponential).
sion model is obtained for the smaller angles and then it is Since the maximum roll and pitch angle is 90° for a FWMAV,
used to extrapolate the data at higher angles. From figure 4, it it is expected that the velocity will converge to a maximum
can be appreciated that the transient response approximates to value as the angle approaches 90°. Thus, the model is ap-
http://www.imavs.org/

the behaviour of a first-order system. Hence, the equation for proximated as the exponential function in Eq. (6). Linear or
the transient response is modelled as an exponential function quadratic models could be used as well, but they would only
of the form: be valid within specific ranges since they do not converge to
v(t) = a − be−ct+d (5) a constant value.

The parameters a, b, c and d from Eq. (5) have to be initialised v(θ) = a + be−cθ (6)
with certain values, depending on the shape of the curve ob-
tained from the measured data points. The final steady-state Using the obtained steady-state velocity, a nonlinear re-
values obtained from the extrapolation model strongly de- gression model is obtained using Eq. (6). For both pitch and
pend on how many data points are considered for the regres- roll the parameters are initialised as: a = −90, b = 90 and
sion model. Hence, Table 1 shows the absolute error between c = 0.3. Moreover, to test the applicability of the exponential
the last measured velocity for each angle, and the respective model, the same described procedure was applied to the data
value calculated from the extrapolation model. For the angles sets from [22] regarding the Delfly Nimble (a 33cm-wingspan
of 30° or less, this velocity approximates to the steady-state FWMAV of the same type). Then the nonlinear regression
velocity. For angles of 45° or above, the last measured veloc- model is applied initialising the parameters with exactly the
ity is used as ground truth equilibrium velocity, although it is same values. Figure 6 shows how the exponential curves of
still part of the transient response. the Flapper and the Delfly properly approximate to the given
data points for both roll and pitch.
Final velocity error at different pitch angles [m/s]
Data taken for regression 12° 22° 33° 45°* 52°*
10% 0.124 0.0752 0.1842 0.2641 0.276
6 H EIGHT CONTROL
30% 0.0318 0.0459 0.0347 0.1899 0.1259 Having a reliable height control strategy is essential to
50% 0.0166 0.0636 0.015 0.2048 0.0038
70% 0.0183 0.0632 0.0282 0.1379 0.0402
guarantee successful autonomous flight. Mainly it is used
90% 0.0142 0.0561 0.0289 0.0857 0.0185 to keep the drone flying at a certain altitude and for hov-
Final velocity error at different roll angles [m/s] ering. Nevertheless, it also plays a major role in landing,
Data taken for regression 12° 22° 33° 47°* 52°* which is known to be as the most challenging phase of flight
10% 0.1007 0.1479 0.0391 0.2972 0.3028
30% 0.10 0.1216 0.0452 0.2303 0.2366
for any aircraft. Height control is intimately related with
50% 0.0262 0.1427 0.0325 0.0499 0.0849 thrust. Therefore, two approaches for improving thrust com-
70% 0.0234 0.0802 0.0378 0.0486 0.095 mand were considered, first one is a voltage-dependent thrust
90% 0.0268 0.0247 0.0261 0.0266 0.0757 model, and second one is an analysis of the ground effects.
Table 1: Extrapolation errors taking different amounts of data 6.1 Voltage-thrust model
for regression model. *Indicates that for those angles the final Thrust is part of the position and velocity control of the
velocity is not considered as ground truth. Hence these errors drone, it also involves a PID controller which takes a com-
are overestimated. manded throttle and delivers an output signal for the motors.
The signal from the controller is summed with a feed forward

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 88

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

5.5
4
Data
5 Data for regression
Data
3.5 Regression model
Data for regression
Regression model 4.5
3
4

2.5 3.5

Velocity [m/s]
Velocity [m/s]

2 3

2.5
1.5

2
1
1.5
0.5
1

0 0.5
0 2 4 6 8 10 0 1 2 3 4 5 6 7 8 9
Time [s] Time [s]

Figure 5: Extrapolation models for pitch (left) and roll (right) at different velocities

Pitch-velocity Roll-velocity
10 10
Setpoints Flapper Setpoints Flapper
0 =0° Setpoints Delfly =0° Setpoints Delfly
=0° 0 =0°
Exponential Flapper Exponential Flapper
Exponential Delfly Exponential Delfly
-10 -10 =10°
=12°
=12°

-20 =17° -20

=22° =20°=20°

-30 =33° -30

=33° =30°
Pitch [deg]

Pitch [deg]
-40 -40
http://www.imavs.org/

=42°
=45°
=45°
-50 =45° -50
=52°
=52°
=57°
-60 =57° -60

-70 -70

-80 -80

-90 -90
0 5 10 15 0 5 10 15
Velocity [m/s] Velocity [m/s]

Figure 6: Pitch/velocity (left) and roll/velocity(right) models obtained from extrapolated data

term known as base thrust, which is the thrust required to trollers: a P and PI height controller, using both the constant
keep the drone hovering at a certain altitude. In most cases and variable base thrust. For each controller, the flight was
the base thrust is a constant value, but there is evidence that analysed in five different directions: X-motion, Y-motion, Z-
the required base thrust tends to increase as the battery volt- motion, XZ-motion and YZ-motion. For each motion, a flight
age drops off [13]. In order to observe this behaviour, a series test was done consisting on five repetitions of step-input com-
of experiments were done where the drone started flying with mands in the given motion, in order to get an average be-
a fully-charged battery and then let it hover until the battery haviour. Hence, Table 2 summarises the standard deviation
got discharged. all values were sampled at 50 Hz and mapped and the mean absolute error between the transient response
as depicted in Figure 7. and the commanded height for each motion and controller.
Throttle is used instead of thrust because this is the signal Notice from the table that for most of the motions, using a PI
that is directly input into the motors. In this case, throttle is a will result in lower error than a P controller, regardless of the
dimensionless value where 30000 sets the motors to the low-
est speed and 60000 indicates full speed. The data points are
then filtered and used to create a regression model. The im-
provement when increasing the order of the regression model
is not significant. First, second and third-order models yield
R2 ≈ 0.756. Albeit there is noisy data involved, it is not
considered for the regression model since it is already known
that the noise comes mainly from the feedback of the con-
troller and the IMU measurements. Nevertheless, there is the
option of using alternative regression models to account for
the stochastic behaviour observed in Figure 7, which may re-
sult in a better R2 value.
Once the model is implemented, the base thrust turns into
a variable base thrust whose value will tend to increase the
longer the drone keeps flying. To evaluate the model, a series Figure 7: Throttle and voltage mapping with tendency lines
of experiments was conducted using four different thrust con- for different regression models

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 89

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

variable or constant base thrust. For most conditions, vari- between thrust ratio and height can be modelled as:
able PI reaches the lowest mean absolute error and standard  2
deviation, showing the utility of varying thrust model. Tinput 1
=1−λ (7)
Toutput h−a
Figure 8 depicts the averaged transient response in height
for each controller, for the cases of X motion and XZ mo- After mapping the thrust ratio with their respective height,
tion. The reduction of steady-state error caused by the inte- a non-linear regression model using equation Eq. (7) is ap-
gral gain is clear in the X motion, where the drone should plied. Such model approximates the coefficients λ = 0.00093
keep flying at the same height when moving from one posi- and a = 0.4213. In Figure 9, the regression model stays
tion to another. The XZ motion shows a case in which vari- within the bounds of each data point. At the tested heights the
able thrust model performs less well. In motions where a ground effect still has relatively little influence on the thrust
change in height occur, performance is more similar between ratio. The thrust ratio only decreases to 92% at the lowest
P and PI controllers. From one side the integral gain tends to height of 0.33m. Thus, the ground effect is little up to 0.33m.
increase the settling time compared to the P controller, but at Lower heights are not considered relevant, because, due to
the same time the oscillations of the PI response reduces the the disposition the UWB anchors, the drone is unlikely to fly
error to minimum whenever it crosses the commanded value. lower except during take-off and landing. Nevertheless, the fit
shows a sharp drop-off below 0.33m. Whether this is correct
will have to be confirmed with future experiments.
http://www.imavs.org/

0.8
Thrust ratio [-]

0.6

0.4

Figure 8: Extrapolation models for pitch (left) and roll (right) 0.2
Data points

at different velocities Model

0
0.2 0.4 0.6 0.8 1 1.2
Distance to touch ground [m]

6.2 Ground effect Figure 9: Thrust ratio for each height

.
For the second approach in height control, the target was
to investigate whether the ground effect has a large effect 7 E XPERIMENTAL VALIDATION
when flying low. According to [14], the range of height in Once all the modifications were done, two versions of the
which ground effects usually occur is 3d > h > d, where controller were tested for validation. The first version is re-
d is the diameter of propellers in quadrotors. Similarly, for ferred as ”raw” version, which is the working implementation
a FWMAV, the range would be proportional to the wingspan with the default EKF and PID controller mentioned in Section
(0.5m). Hence, the experiments to model the ground effect 3, without the proposed enhancements. The second version is
consisted on keeping the drone hovering at different heights referred as the ”modified” version, and is the implementation
between 0.33m and 1.2m for over a minute. The lowest height with the adaptations mentioned in Sections 4, 5, and 6.
tested was 0.33m above the ground, since this was the lowest The validation consists of analysing the transient re-
height at which the drone can be kept hovering autonomously sponses in position and velocity when position step-inputs
due to the location of the UWB anchors. In this section, the were given in x, y, and z. For a wider perspective on the
height is considered to be the distance between the landing stability of the controller, the amplitude of the step input was
gear and the ground. For each different height, a mean thrust changed from 1m to 2m. Thus, any difference in the aggres-
and mean voltage is obtained. Using the thrust-voltage re- siveness of the response can be detected as well. For each
lationship found previously, the corrected thrust can be ob- motion and step-input, a series of five repetitions was done
tained. The ratio between the corrected thrust and the mean and averaged to obtain the general behaviour of the transient
thrust is specific for each different height, as depicted in fig- response. Table 3 presents how the mean absolute error, for
ure 9. Notice that for each data point an upper and lower both position and velocity, decreases when using the modi-
bound is also provided based on the standard deviation of the fied controller instead of the raw controller. In general, errors
thrust and battery voltage. According to [24], the relationship are approximately 1.5 times lower after the modifications.

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 90

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

X motion Y motion Z motion XZ motion YZ motion Overall

Controller Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev. Mean Std. dev.
Constant P 0.109 0.08 0.098 0.075 0.077 0.117 0.035 0.166 0.043 0.143 0.072 0.116
Constant PI 0.061 0.159 0.048 0.093 0.06 0.075 0.047 0.09 0.047 0.1175 0.053 0.107
Variable P 0.114 0.069 0.086 0.079 0.065 0.069 0.057 0.116 0.046 0.1038 0.074 0.09
Variable PI 0.049 0.107 0.025 0.083 0.06 0.056 0.05 0.092 0.037 0.061 0.044 0.08

Table 2: Errors and standard deviations for each controller. Best result for each condition is bold-cased.

Mean absolute error for position step input 1

X motion Y motion Z motion
Version Raw Modified Raw Modified Raw Modified
Position [m] 0.1706 0.105 0.1898 0.1048 0.0496 0.043
Velocity [m/s] 0.3945 0.1297 0.3001 0.1199 0.1001 0.0892
Mean absolute error for position step input 2
X motion Y motion Z motion
Version Raw Modified Raw Modified Raw Modified
Position [m] 0.3316 0.1978 0.2989 0.2055 0.1245 0.1097
Velocity [m/s] 0.6330 0.2859 0.4583 0.1802 0.2476 0.1613

Table 3: Mean absolute errors of validation experiments. Best (a) Absolute error in Velocity X
http://www.imavs.org/

performance for each case is bold-cased.

Lastly, Figure 10 illustrates how the absolute error ob-

tained in the transient response of the modified version is
smaller and more consistent than the one of the raw version.
For the sake of simplicity, only the transient responses of ve-
locity in x and y to the step input of 2m are given, as the
increase in performance in those two is the largest, according (b) Absolute error in Velocity Y
to Table 3. In Figure 10 the error is defined as the absolute
difference between the output value and the corresponding Figure 10: Comparison of velocity error during transient re-
commanded value. The plotted error corresponds to the tran- sponse to a step input of amplitude 2
sient response when the step input of amplitude 2 is given at
t = 1s. Notice that almost throughout the whole response the a more consistent performance through time. A more notori-
error of the modified version is lower than the one of the raw ous contrast can be obtained if the FWMAV flies for longer
version. Moreover, most of the peaks of the modified version periods.
are roughly 75% smaller than the ones from the raw version. Thirdly, the ground effect experiments prove that as long
as the drone’s wings fly above 0.5m from the ground, no sig-
8 C ONCLUSIONS nificant additional thrust will appear. Nevertheless, to know
The achievement from this work was defining a strategy how much extra thrust is produced below 0.5m, experiments
for enhancing position control for FWMAVs. From the esti- should be done through manual flight.
mation perspective, the strategy uses an extended Kalman fil- Lastly, in order to improve performance in autonomous
ter to fuse UWB and IMU data, and it also incorporates a vari- flight using UWB, further development can be done in the po-
able sensor noise term. Altogether, state estimation achieves sition estimation. For instance drag is not considered in the
accuracy between 8-10 cm error. From the control side, three EKF for the estimation of body velocities. An option would
specific aspects are considered. be deriving a drag model from IMU data. Other approach
Firstly, a velocity/attitude nonlinear model, which showed would be fusing data from other sources (e.g. barometer and
to be valid for two different kinds of FWMAVs. Thus, prov- magnetometer data) to compensate for the noisy accelerome-
ing that it can be adapted, as long as there is experimental ter measurements due to inherent mechanical vibrations from
data from which the steady-state velocity can be extrapolated flapping wing flight.
for a certain angle. Moreover, the model was validated using
real flight data and proved its efficacy for drag compensation ACKNOWLEDGEMENTS
as part of the feed-forward term in the velocity control loop. The realisation of this paper was possible thanks to the
Secondly, the lower standard deviation of voltage- support of the start-up Flapper Drones B.V., and the MAVLab
dependent thrust, compared to constant thrust, demonstrates from the TU Delft Faculty of Aerospace.

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 91

 IMAV2021-10 12th I NTERNATIONAL M ICRO A IR V EHICLE C ONFERENCE
http://www.imavs.org/papers/2021/10.pdf

R EFERENCES A tailless flapping wing mav performing monocular visual ser-

[1] Charles Porter Ellington. The aerodynamics of hovering insect voing tasks. Unmanned Systems, 8(04):287–294, 2020.
flight. IV. Aerodynamic mechanisms. Philosophical Transac- [14] IC Cheeseman and WE Bennett. The effect of the ground on a
tions of the Royal Society of London. B, Biological Sciences, helicopter rotor. R & M, 3021, 1957.
305(1122):79–113, 1984.
[15] Sanjukta Aich, Chahat Ahuja, Tushar Gupta, and P Arul-
[2] Michael H Dickinson, Fritz-Olaf Lehmann, and Sanjay P Sane. mozhivarman. Analysis of ground effect on multi-rotors. In
Wing rotation and the aerodynamic basis of insect flight. Sci- 2014 International Conference on Electronics, Communica-
ence, 284(5422):1954–1960, 1999. tion and Computational Engineering (ICECCE), pages 236–
[3] Juan J Pomárico-Franquiz, Moises Granados-Cruz, and 241. IEEE, 2014.
Yuriy S Shmaliy. Self-localization over RFID tag grid excess [16] Yuan Xu, Yuriy S Shmaliy, Xiyuan Chen, Yueyang Li, and
channels using extended filtering techniques. IEEE Journal of Wanfeng Ma. Robust inertial navigation system/ultra wide
Selected Topics in Signal Processing, 9(2):229–238, 2014. band integrated indoor quadrotor localization employing adap-
[4] Abdulrahman Alarifi, AbdulMalik Al-Salman, Mansour Al- tive interacting multiple model-unbiased finite impulse re-
saleh, Ahmad Alnafessah, Suheer Al-Hadhrami, Mai A Al- sponse/kalman filter estimator. Aerospace Science and Tech-
Ammar, and Hend S Al-Khalifa. Ultra wideband indoor posi- nology, 98:105683, 2020.
tioning technologies: Analysis and recent advances. Sensors, [17] Steven van der Helm, Mario Coppola, Kimberly N McGuire,
16(5):707, 2016. and Guido CHE de Croon. On-board range-based relative
[5] Bardia Alavi and Kaveh Pahlavan. Modeling of the TOA-based localization for micro air vehicles in indoor leader–follower
distance measurement error using UWB indoor radio measure- flight. Autonomous Robots, 44(3):415–441, 2020.
http://www.imavs.org/

ments. IEEE Communications Letters, 10(4):275–277, 2006. [18] Matěj Karásek, Mustafa Percin, Torbjørn Cunis, Bas W
[6] Zengke Li, Guobin Chang, Jingxiang Gao, Jian Wang, and van Oudheusden, Christophe De Wagter, Bart DW Remes,
Alberto Hernandez. GPS/UWB/MEMS-IMU tightly coupled and Guido CHE de Croon. Accurate position control of a
navigation with improved robust kalman filter. Advances in flapping-wing robot enabling free-flight flow visualisation in
Space Research, 58(11):2424–2434, 2016. a wind tunnel. International Journal of Micro Air Vehicles,
11:1756829319833683, 2019.
[7] Qigao Fan, Yaheng Wu, Jing Hui, Lei Wu, Zhenzhong Yu,
and Lijuan Zhou. Integrated navigation fusion strategy of [19] Matej Karasek. Flapper drones. https://flapper-
INS/UWB for indoor carrier attitude angle and position syn- drones.com/wp/. Accessed: 2021-01-01.
chronous tracking. The Scientific World Journal, 2014, 2014. [20] Controllers in the crazyflie.
[8] Karl Martin Kajak, Matej Karásek, Qi Ping Chu, and GCHE https://www.bitcraze.io/documentation/repository/crazyflie-
De Croon. A minimal longitudinal dynamic model of a tail- firmware/2020.04/functional-areas/controllers/.
less flapping wing robot for control design. Bioinspiration & Accessed:2021-02-01.
Biomimetics, 14(4):046008, 2019. [21] Mark W Mueller, Michael Hamer, and Raffaello D’Andrea.
[9] Mostafa RA Nabawy and William J Crowther. The role of the Fusing ultra-wideband range measurements with accelerome-
leading edge vortex in lift augmentation of steadily revolving ters and rate gyroscopes for quadrocopter state estimation. In
wings: a change in perspective. Journal of the Royal Society 2015 IEEE International Conference on Robotics and Automa-
Interface, 14(132):20170159, 2017. tion (ICRA), pages 1730–1736, May 2015.
[10] Dong Xue, Bifeng Song, Wenping Song, Wenqing Yang, [22] Matěj Karásek, Florian T Muijres, Christophe De Wagter,
Wenfu Xu, and Tao Wu. Computational simulation and free Bart DW Remes, and Guido CHE de Croon. A tailless aerial
flight validation of body vibration of flapping-wing mav in for- robotic flapper reveals that flies use torque coupling in rapid
ward flight. Aerospace Science and Technology, 95:105491, banked turns. Science, 361(6407):1089–1094, 2018.
2019. [23] Taimur Ali Shams, Syed Irtiza Ali Shah, Ali Javed, and Syed
[11] JL Verboom, Sjoerd Tijmons, Christophe De Wagter, Hossein Raza Hamdani. Airfoil selection procedure, wind tun-
B Remes, Robert Babuska, and Guido CHE de Croon. Atti- nel experimentation and implementation of 6dof modeling on
tude and altitude estimation and control on board a flapping a flying wing micro aerial vehicle. Micromachines, 11(6):553,
wing micro air vehicle. In 2015 IEEE International Confer- 2020.
ence on Robotics and Automation (ICRA), pages 5846–5851.
[24] Li Danjun, Zhou Yan, Shi Zongying, and Lu Geng. Au-
IEEE, 2015.
tonomous landing of quadrotor based on ground effect mod-
[12] Zhi Ern Teoh, Sawyer B Fuller, Pakpong Chirarattananon, elling. In 2015 34th Chinese Control Conference (CCC), pages
NO Prez-Arancibia, Jack D Greenberg, and Robert J Wood. 5647–5652. IEEE, 2015.
A hovering flapping-wing microrobot with altitude control and
passive upright stability. In 2012 IEEE/RSJ International Con-
ference on Intelligent Robots and Systems, pages 3209–3216.
IEEE, 2012.
[13] Diana A Olejnik, Bardienus P Duisterhof, Matej Karásek,
Kirk YW Scheper, Tom Van Dijk, and Guido CHE De Croon.

N OVEMBER 17th TO 19th 2021, P UEBLA , M ÉXICO 92