898 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 13, NO.
2, APRIL 2017
Characterizing Multihop Aerial Networks of
COTS Multirotors
Luis Ramo Pinto, Student Member, IEEE, André Moreira, Student Member, IEEE,
Luis Almeida, Senior Member, IEEE, and Anthony Rowe, Member, IEEE
Abstract—Unmanned aerial vehicles (UAVs) recently en-
abled a myriad of new applications spanning domains from
personal entertainment to surveillance and monitoring. In
this paper, we focus on using several small UAVs collabo-
ratively to provide extended reach to an online video mon-
itoring system for inspection of industrial installations. We
make use of 802.11 radios on low-cost commercial-off-the-
shelf UAVs, set up a time-division multiple access overlay
protocol to avoid mutual interference, and enable high chan-
nel utilization in multihop networks. In particular, we provide
a model for the quality of the UAV-to-UAV link, in terms of
packet delivery ratio as a function of distance, packet size,
and orientation, based on an extensive measurement cam-
paign. We show that this platform is not omnidirectional in
the horizontal plane and that UAV-to-UAV communication
ceases around 75 m. Concerning the operation in a multi-
hop mode to allow extending the network, the paper derives
the optimal number of hops that maximize the end-to-end
throughput, as well as the corresponding hop lengths. We
validate our mathematical model with extensive experimen-
tal measurements transmitting payloads up to 200 m (over
802.11 g at 54 MBps).
Index Terms—802.11, ad hoc networks, channel models,
relay networks, time-division multiple access (TDMA), un-
manned aerial vehicle (UAV), wireless communication.
I. INTRODUCTION
ICRO electromechanical systems, e.g., inertial sensors,
M and low-cost high-speed microcontrollers recently made
unmanned aerial vehicles (UAVs) increasingly smaller and ac-
cessible, boosting new applications in a variety of domains from
Manuscript received May 27, 2016; revised October 12, 2016; ac-
cepted December 17, 2016. Date of publication February 13, 2017;
date of current version April 18, 2017. This work was supported
in part by Fundação para a Ciência e a Tecnologia under Grant
SFRH/BD/51630/2011 and the FCT project UID/EEA/50008/2013. Pre-
liminary results of this paper were published in part in the 12th IEEE
World Conference on Factory Communication Systems (WFCS2016).
Paper no. TII-16-0480. (Corresponding author: L. Almeida.)
L. R. Pinto is with the Instituto de Telecomunicações, Faculdade de En-
genharia da Universidade do Porto, Portugal, and also with the Electri-
cal and Computer Engineering Department, Carnegie Mellon University,
Pittsburgh, PA 15213 USA (e-mail: p@cmu.edu).
A. Moreira and L. Almeida are with the Instituto de Telecomunicações,
Faculdade de Engenharia da Universidade do Porto, Porto 4200-465,
Portugal (e-mail: andre.moreira@fe.up.pt; lda@fe.up.pt).
A. Rowe is with the Electrical and Computer Engineering Depart-
ment, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
agr@ece.cmu.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TII.2017.2668439
1551-3203 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
law enforcement agencies to media production crews, personal
entertainment and many other. Multirotors, in particular, became
ubiquitous given their manoeuvring and hovering capabilities at
small scales, together with sensing and communication features.
These UAVs are also becoming a viable and cheaper option in
monitoring and inspection of large industrial sites, either process
control plants or large factories. Instead of labor-intensive local
manual inspections, UAVs can easily carry cameras and stream
live video of high chimneys, electrical poles, large deposits, long
pipelines, etc., to a remote operation center (ground station)
where operators can control the vehicles’ pose to focus on any
feature of interest.
However, building such a UAV-based online video monitor-
ing system entails several challenges, such as establishing a
reliable communication link, potentially multihop, to convey
the sensing information. In fact, it is well known that the qual-
ity of wireless communications degrades sharply with distance.
Adding relays can improve the reliability between neighboring
nodes given the relative shorter distances. On the other hand,
relays typically share the wireless channel imposing a cost on
the overall throughput.
This paper addresses this compromise in a line network com-
prised of a sensor-UAV and a variable number of relay-UAVs
using the AR Drone 2.0 platform that is representative of this
class of vehicles. We extend a preliminary work reported in
[1] that investigated the characteristics of the aerial link based
on extensive measurements. In this work, we use those and an
extra set of measurements to introduce an adequate modeling
approach that allows deducing the packet delivery ratio (PDR)
as a function of link length (first contribution). The work in
[1] also considered an extension of the link model to multihop
communication validated by experimental data. In this paper,
we provide the formal support to deduce the optimal placement
of the relay nodes that maximizes throughput (second contribu-
tion).
These results are particularly useful for designing networks of
multirotors since in these networks we can control the position
of the nodes to improve the communication links. This is not
possible in typical mobile ad hoc networks, such as networks
of personal devices or vehicular networks. In other cases, such
as networks of ground robots, position control would still be
possible, but the propagation characteristics in those cases are
rather different due to obstacles, multipath, close range, and near
the floor environment. In this work, we seek to provide novel
insight into outdoor aerial networks of commercial-off-the-shelf
PINTO et al.: CHARACTERIZING MULTIHOP AERIAL NETWORKS OF COTS MULTIROTORS
(COTS) multirotors, especially in characterizing and modeling
its links.
The organization of the paper is as follows. Section II de-
scribes the state-of-the-art and previous research in this domain.
In Section III, we expose the problem of communication as a
function of distance. The link performance of the platform is
explained in Section IV, and we model this channel in Section
V. Section VI describes our concept of multihop network, and
we prove its properties. Section VII details the experiments and
their results carried out to validate the network model. Finally,
Section VIII provides conclusions, remarks on the use of the
platform, and future work.
II. RELATED WORK
Several authors have developed numerous UAV test-beds for
commercial, military, and research purposes. Some of them have
explored using UAVs as flying wireless sensor networks. In [2],
Goddemeier et al. describe there are multiple ways to organize
UAVs to form a sensor network, whether the sensors can be
disconnected from the base station for sometime, whether sen-
sors are all directly connected to the base station, or if relays
are allowed to be used in order to increase the range of com-
munication. Links are maintained by measuring the received
signal strength indicator (RSSI), and approaching nodes when
that value falls below a given threshold. No explicit analysis
on multihop networks is done. On the other hand, Asadpour
et al. [3] show how much data a UAV network can transmit
depending on its distance to another UAV. They provide exten-
sive experimental data for 802.11g UDP throughput, testing for
different physical layer (PHY) bit-rates and distances, and rel-
ative velocities. They show that bit-rate can and should be set
manually, disabling automatic adaptation for improved through-
put. However, they do not show important network performance
metrics such as PDR. Furthermore, no information is given on
the effect of packet size on throughput. They also assume an
isotropic medium behavior, which is not always the case, as we
will see, particularly with COTS multirotors. In [4], Asadpour
et al. go further and analyze factors such as relative orientation,
PHY rates, and UAV relative speed. Nevertheless, no results on
network throughput with multiple relays are provided, despite
using one relay to transmit pictures to the base station.
Nevertheless, the issue of PDR has been addressed by several
other works. For example, Zhao and Govindan [5] and Jia et al.
[6] clearly identified distance as the main factor of PDR, in
fixed sensor networks. The work in [7], on the other hand,
shows the impact of packet size on PDR and throughput, in
underwater networks. Antenna orientation is studied in [8],
where IEEE802.15.4 radios are used to test its impact on PDR.
In the field of vehicular networks, one can also find work bear-
ing similarities to ours. For example, in [9], Bohm et al. present
an experimental characterization of the 802.11p channel focus-
ing on the effects of relative speed between ground vehicles,
including the effects of speed on the PDR. Unfortunately, no
data is provided on multihop performance.
One domain that bears several resemblances to our case is
that of robotic networks, despite typically focusing on ground
899
autonomous mobile robots. For example, the work in [10] inves-
tigates how robots motion can be controlled so to maintain high
throughput for streaming data to a base station using a multihop
network. They conclude that, instead of transmitting all the way,
it is better to concentrate transmissions in areas where/when the
channel is good, slowing the robot, while moving faster in areas
of poor channel characteristics. The focus, however, is on the
robots mobility control and not on the channel characterization.
In [11], Henkel and Brown analyze mobile robotic networks
performance as a function of distance from the base station and
required data-rate/delay requested by the users. They also con-
sider using relay nodes. However, when robots move extremely
far from the base, the authors propose swapping toward a data
mule model that leverages delay-tolerant networking, but which
is not applicable to our case. Moreover, they do not provide any
experimental data on PDR over single or multihop links.
The works in [12] and [13] address a similar purpose as ours,
in the sense of aiming at establishing a line topology of relays to
support a multimedia connection, but they focus on the specific
characteristics of tunnels and pipelines. They explore the fact
that tunnels, under certain circumstances, behave like waveg-
uides. This makes the results of these works inapplicable to our
scenario of operation in open areas. In [14], Flushing et al. cre-
ate a method to predict link quality based on an offline learning
phase. This predictor provides the robot network with a map of
expected communication quality at any point in space. They do
not measure channel performance on a real-world system, and
therefore our work is likely to be useful to feed real data into
learning phases of tools like this.
As seen in this brief survey, existing works in the literature
address related but different aspects of wireless ad hoc com-
munication with respect to our work. In fact, to the best of our
knowledge, no analysis on an outdoor multirotor IEEE802.11
channel was done to date, particularly addressing single- and
multihop operation, beyond the preliminary work presented in
[1], which we now complete.
III. PROBLEM
Using a single UAV with sensing capabilities (sensor-UAV)
that collects information to deliver to a fixed remote point
(Ground Control Station—GS), what is the network topology
that allows the highest throughput in a relatively wide range of
distances? Note that assuming an arbitrary distance between the
sensor-UAV and the GS, direct communication may be infeasi-
ble. To overcome this issue, we consider that additional UAVs
may be deployed to act as relays and help increase individual
link quality, forming a line network. However, more nodes in
the network decreases the available time each node has to trans-
mit since multiple nodes are sharing the same radio channel.
Therefore, we aim at solving the problem of finding the number
and optimal placement of relaying COTS multirotors that max-
imize throughput in a line network, given an arbitrary distance
between the sensor-UAV and the GS.
In the following sections, we will solve this problem by
characterizing the communication channel in our system. We
start by studying the link layer, i.e., the node-to-node direct
900 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 13, NO. 2, APRIL 2017

TABLE I
EXPERIMENT SETUP FOR THE TWO AR DRONE 2.0 VEHICLES IN THE
SINGLE-LINK CHANNEL CHARACTERIZATION

PHY layer 802.11 g, fixed to 54 MBps, Tx power 15 dBm

Wireless mode Ad hoc
Max retry 2 (minimum allowed by the platform)
# of packets 1000 per experiment
Packet type UDP (78 bytes header)
Location Open field with low external interference
Height 3 m above the ground.
Distance 0, 5, 10, 15, 25, 35, 45, 55, 65, 75 m
Payload size 200, 500, 1000 bytes
Orientation Parallel, Collinear

communication, by performing multiple experiments on COTS

multirotors as well as developing a model for it. Then, we extend
this model to the network layer, i.e., end-to-end communication
using different number of relaying nodes in a line topology. We Fig. 1. Top figure shows typical PDR values of multiple consecutive ex-
periments performed over time, under a given condition of orientation,
then validate our model by performing experiments and mea- distance, and packet size. Its variation and relative occurrence of con-
suring the end-to-end throughput under different conditions. secutive successful/failed packets (coloured bars) follows the same trend
as a pure Bernoulli experiment (white bars). These plots correspond to
the probability mass function (pmf) of geometric random variables.
IV. LINK LAYER—EXPERIMENTS
In this section, we describe the experiments characterizing the
UAV-to-UAV communication, using the AR Drone 2.0 platform
[15], which is a common programmable quad-rotor with on-
board camera and wireless IEEE 802.11 interface, available in
the market.
We used two UAVs at a time, one as the sender, and the other
one as the receiver. In each experiment, the sender transmitted
1000 packets, and the receiver recorded the relevant part of
the received packets’ payload together with the reception time
stamp into a log file. We also carried out experiments with
different configurations of distance, packet size, and relative
vehicle orientation to evaluate how these affected the PDR, and
ultimately the throughput. In particular, we used three distinct
packet sizes as well as two different relative orientations at each
distance.
More than 5600 experiments were carried out with the UAVs
at 3 m height, far from interfering sources in the same RF band.
The transmitter was set to send packets as fast as possible. Each
experiment took between 1 and 2 s to complete. In total, the
experiments resulted in more than 950 MB of gathered data
from multiple distances, packet sizes, and vehicle orientations.
Table I summarizes all settings used in the experiments.
The top of Fig. 1 shows the PDR evolution over time, per-
formed for a given orientation, distance, and packet size. Each Fig. 2. Packet delivery ratio in a single link as a function of distance
between the sender and the receiver. Packet size and relative orientation
point represents the average of 1000 packets. Since the varia- of the vehicles affect the ratio. Dots represent experimental data. Solid
tions between consecutive points do not exhibit significant visi- and dashed lines are fitting curves, based on the model developed in
ble correlation, we conjecture that the probability of successful Section V.
delivery of each packet is also independent of each other and
can be characterized by its average p alone, i.e., a Bernoulli histograms with those of the PDR measurements in Fig. 1 (wider
process. To test our conjecture, we analyzed the histograms white bars behind the corresponding colored bars). Given the
of the number of consecutive successfully transmitted packets visibly good match between the histograms, we claim that our
and the number of consecutive failed packets (colored bars in conjecture is true, and that our PDR experiment does follow a
the middle and bottom part of Fig. 1, respectively). Then, we Bernoulli distribution.
generated a synthetic Bernoulli process with trial success proba- The results of the link layer experiments are shown in
bility p and 100 000 trials, and we overlapped the corresponding Fig. 2. Here, we characterize the two relative orientations that
PINTO et al.: CHARACTERIZING MULTIHOP AERIAL NETWORKS OF COTS MULTIROTORS
generated the strongest differences, only. Namely, the vehicles
were positioned in parallel (solid line) or colinearly (dashed) to
each other while transmitting packets. An immediate observa-
tion is that the relative orientation of the vehicles has a striking
impact on the performance of the network.
The observed differences result from the nonomnidirectional-
ity of the antennas in the plane of flight. We believe this feature
is common on COTS multirotors, since a good connection with
ground controls is to be prioritized. Thus, this is an issue to con-
sider when using communications between multiple AR Drone
2.0. Curiously, the different orientations impact not only the
effective length of the link but also the steepness of the PDR
reduction with distance. In fact, there is a very narrow region
(about 3 m) where the link suddenly changes from good to
practically broken.
PDR depends also on the number of bytes being sent in a
packet. More bytes mean less probability of delivery, as Fig. 2
data show, which is specially visible when we compare the three
blue plots (parallel orientation). We can see a reasonable range
difference between 200 and 1000-byte packets. The explanation
is that for the same bit error rate, longer packets have more
bits prone to suffer from errors. Curiously, the difference in
PDR between packet sizes of 500B and 1000B is practically
negligible in both orientations.
Distance is the third dimension that clearly affects the PDR.
As other authors have concluded on other platforms (cf. Section
II), for short distances, the PDR is sustainably high with negli-
gible packet losses but, as distance grows, eventually it enters an
unstable region where PDR drops, in average, almost linearly
to zero where the link is considered broken.
The orientation aspect is rather relevant in a line topology as
the one we are studying here. It is important to keep all vehi-
cles’ heading perpendicular to the network line itself in order to
maximize the operation range of our network. Currently, we in-
tend to use our bottom-facing camera to sense the environment
(mainly ground monitoring) and so the top-end node (sensor
node) as well as all other nodes can assume this optimal pose.
However, when the front camera is to be used, the worst-case
orientation must be assumed (collinear) for safety reasons, and
so a multi hop network will use shorter links.
V. LINK LAYER—MODEL
Analyzing the dataset collected during the link experiments
allows to create a mathematical channel model. This model
is intended to easily predict the likely PDR value given the
distance between two nodes, for each orientation and packet
size scenario. After an analysis of different family functions,
we selected a negative exponential curve to fit our data whose
generic form is given by (1), where β = − ln(2)/Rα
pl (d) = e− ln(2)·( R ) = eβ d .
d α α
(1)
We believe this family of functions is a good fit for PDR data
since it is a nonnegative, strictly nonincreasing function whose
range matches that of our data, i.e., [0, 1] ∈ R. This function
presents also a zero derivative at d = 0 and at d → +∞, which
mimics collected data, namely the two plateaus at PDR = 100%
901
TABLE II
PARAMETERS OF THE PROPOSED PDR MODEL FOR DIFFERENT
CONDITIONS, NAMELY PACKET SIZE AND ORIENTATION
Parallel Collinear
200 bytes α = 10.6 α = 54.6
R = 64 m R = 22 m
RM SE = 0.0911 RM SE = 0.0232
(RM SE σ = 0.0922) (RM SE σ = 0.0232)
500 bytes α = 21.1 α = 22
R = 53 m R = 56.4 m
RM SE = 0.0256 RM SE = 0.0276
(RM SE σ = 0.0256) (RM SE σ = 0.0276)
1000 bytes α = 17.1 α = 46.7
R = 51 m R = 22 m
RM SE = 0.0896 RM SE = 0.0105
(RM SE σ = 0.0898) (RM SE σ = 0.0105)
and 0%. A rather attractive feature of the fitting function pro-
posed in (1) is that it only needs two parameters. Parameter R
describes the distance at which packet delivery is 50% (R ∈ R+
meters) and the curve steepness depends directly on the param-
eter α ∈ R+ > 1 . The model parameters were estimated using the
MATLAB Fitting Tool and are shown in Table II for different
packet size and relative orientation of the vehicles. This table
also shows the root-mean-squared error (RMSE) as an indica-
tion of the accuracy of our proposed model with respect to the
raw data. We recognize that more data would improve the con-
fidence on our model, nevertheless RMSE values are already
relatively low, generally below 2.5% with just two cases rising
to near 9%, thus we consider the model accuracy to be enough
for a set of envisaged applications. We also tried another fit-
ting model, namely the well-known
 logistic sigmoid function
(pl (d) = 1/ 1 + e−k (R −d) ), which has similar basic charac-
teristics. However, it presented similar or slightly higher RMSE
values, represented as RMSEσ in Table II, particularly in the
cases of lower accuracy. Thus, we preferred the simple nega-
tive exponential curve model [cf. (1)]. Having a model of link’s
PDR, we can model the correspondent throughput, by using the
following definition.
Definition 1 (Link Throughput). Average number of packets
successfully transmitted between two nodes per second, denoted
as ζl .
Throughput depends directly on two main variables: one re-
garding the transmitter, and one regarding the receiver. The
former is the packet transmission speed B, i.e., the number of
packets sent from the transmitter per second. The latter is the
probability of receiving a packet pl (d), i.e., the PDR [cf. (1)].
Link throughput is typically measured in bits per second (bps),
so we can convert ζl into bps multiplying it by the size of the
packets being transmitted (in bits), leading to
ζl (d) = B × pl (d). (2)
Transmission speed B itself depends on two variables, namely
the transmitter PHY-layer bitrate and the packet size. It can
be set and fixed along the course of a mission. Quite differ-
ently is PDR that is affected by several variables as we have
shown. Moreover, B also depends on the maximum number of
902 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 13, NO. 2, APRIL 2017

automatic retransmissions configured in the network device Theorem 1. Network PDR pn (d, h) is maximized by con-
driver. It is well known that a higher number tends to sig- sidering that the h−1 relays are placed uniformly between the
nificantly increase the communication latency and consumed sensor-UAV and the GS, i.e., all h links have the same length
bandwidth when the channel conditions degrade. This may im- l = d/h.
prove the PDR, but such improvement can be overridden by a Proof: Assume the network has h hops, and end-to-end
degradation of the channel effective bandwidth. To minimize length equal to d. Consider that the first hop has length
this undesired effect we carried out all our study with the lowest l1 = d/h + x1
 , second hop has l2 = d/h + x2 , etc., where
limit for retries allowed in our platforms (cf. Table I). xi = 0 ⇔ li = d. Hence, the PDR product over h hops
is given by
VI. NETWORK MODEL
Even in the best conditions, direct communication between a 
h 
h
e(β (d/h+x i ) )
α
pl (d = li ) =
sensor-UAV and a ground station (GS) is severely compromised i=1 i=1
above 60 m. To improve this range, we deploy relay-UAVs in h
= e(β i = 1 (d/h+x i ) ) ≡ f (X).
α
between, on a line formation creating a network. Packets are (4)
routed from sensor to sink, passing through these relays. This
way, each one of the individual links (also named hops) in the To find the maximizers of this function, we use the method of
network is as short as needed to guarantee packets flow through. Lagrange multipliers—a method that allows us to find maxima
However, we want them to be as long as possible to minimize the of f (X) and its maximizers, subject to constraint of the form
total number of UAVs in use, which by consequence affects the g(X) = 0. For our problem
mutual interference between nodes, and the overall end-to-end
h
throughput. f (X) = f (x1 , . . . , xh ) = e(β i = 1 (d/h+x i ) )
α
(5)
Since we want to continuously stream as much information
from the sensor-UAV as possible, we use a time-division mul-
and constraint function g(X) is
tiple access (TDMA) scheme that guarantees higher utilization
of the medium than a CSMA scheme. This way we avoid mu-
g(X) = g(x1 , . . . , xh ) = x1 + . . . + xh = 0. (6)
tual interference giving each node periodic and dedicated access
to the wireless medium for a different time interval—called a
slot. Furthermore, as we envision the use of global positioning We define the Lagrangian function L(X, λ), s.t
system (GPS)-equipped UAVs, the major concern of TDMA—
synchronization—can be trivially solved by using the global L(X, λ) = f (X) + λ(g(X) − 0)
clock time provided by such system. To simplify slot assign- = f (X) + λ(x1 + x2 + . . . + xh ). (7)
ment, the network creates h distinct TDMA slots, where h is the
total number of transmitter nodes in the network. For the sake Setting its gradient ∇X,λ L(X, λ) = 0, we have
of simplicity, we consider the GS is not transmitting and thus h
will also be the number of hops in the network. All slots have ⎧
⎪ ∂L ∂f (X)
the same width, so the sensor-UAV can transmit at a maximum ⎪
⎪ = +λ=0 (8a)
⎪
⎪ ∂x ∂x1
of one-hth of its original rate. In the future, we expect to relax ⎪
⎪
1
⎪
⎪
this condition and allow some slot reutilization at the expendi- ⎪
⎨... (8b)
ture of a more complex/dynamic TDMA scheme. To minimize
⎪
⎪ ∂L ∂f (X)
the transport time of data from the source to sink, time slots are ⎪
⎪ = +λ=0 (8c)
⎪
⎪ ∂xh ∂xh
sorted in descending order by distance to the sink. ⎪
⎪
⎪
⎪
Having these considerations, we can model the correspondent ⎩ ∂L = (x1 + x2 + . . . + xh ) = 0. (8d)
network throughput by using the following definition. ∂λ
Definition 2 (Network Throughput). Available end-to-end
throughput of a relay chain, and it depends on the end-to-end The partial derivative of f (X) over xi (i ∈ [1, . . . , h]) is
PDR multiplied by the sender transmitted packets per second,  h 
denoted as ζn . ∂f (X) 
Network PDR pn is the product of every hop’s PDR, because = f (X) β (d/h + xi )α
∂xi i=1
a packet is only transmitted successfully if it is successfully
 h 
transmitted on all h hops on its way. Note that hops/links are 
considered independent, given the TDMA scheme. If a network =β (d/h + xi )α f (X)
has h hops, and hop-i has length li , then the network’s length i=1
(d), and network PDR (pn (d, h)) are given by = βα (d/h + xi )α −1 f (X). (9)

h 
h
pn (d, h) = pl (li ) d= li . (3) Solving (8a)–(8c), we see that the value of all partial deriva-
i=1 i=1 tives is the same on the critical point. This means that, for any
PINTO et al.: CHARACTERIZING MULTIHOP AERIAL NETWORKS OF COTS MULTIROTORS 903

two partial derivatives, we have

∂f (X) ∂f (X)
=
∂xi ∂xj
βα (d/h + xi )α −1 f (X) = βα (d/h + xj )α −1 f (X)
(d/h + xi ) = (d/h + xj )
xi = xj note that d/h + x > 0. (10)

We showed that all variables have the same value, and ac-
cording to (8d), it comes that x1 = x2 = . . . = xh = 0. Thus,
Fig. 3. Maximum network throughput as function of distance ζmax (d).
the PDR is maximum when xi = 0, ∀i ∈ [1, h], which means Each color represents a different number of hops in use in the network,
that links have all the same size: li = d/h + 0.  described by a different section of (16). Number of hops h ∈ [1, . . . , 5],
Regarding transmission, the multihop network has h trans- and updated at certain distances given by (17). Link PDR assumes
α = 10.6.
mitter nodes, so the overall sensor node sends packets at a rate
of B/h, due to the aforementioned TDMA constraints. TABLE III
Joining all the previous assumptions, network throughput as EXPERIMENT SETUP FOR THE AR DRONE 2.0 VEHICLES ON THE MULTIHOP
NETWORK CHARACTERIZATION
a function of distance and the number of hops (ζnet (d, h)) can
now be computed: Payload size 200 bytes
Orientation Parallel
B B TX Slot time 100 ms
ζn (d, h) = · pn (d, h) = · [pl (d/h)]h . (11) Number of hops a) 1 , b) 2 , c) 3
h h Number of slots a) 1 , b) 2 , c) 3
Guard interval 50 ms (0 ms for one hop)
For notation simplicity, we will use ζh (d) to refer to ζn (d, h) Round period: a) 100 ms, b) 300 ms, c) 450 ms
End-to-end distance a) 0, 5, 10, 15, 35, 45, 55, 65, 75 m
in the remainder of the paper. For any given distance d, different b) 70, 90, 110, 130, 150 m
throughput values can be achieved depending on the chosen c) 105,135,165,195 m
number of hops h. Choosing h appropriately, one can maximize
throughput and define it solely as a function of distance. s.t.
obtained using (17) (proof in the Appendix)
ζmax (d) = max (ζh (d)). (12)
h ⎧
⎪
⎪ ζ1 (d) if d ∈ [0, d12 )
⎪
⎪
This comes from link-throughput function ζh (d) intrinsic ⎨ ζ2 (d) if d ∈ [d12 , d23 )
properties, namely ζmax (d) = (16)
⎪
⎪ ···
⎪
⎪
⎩
ζh o (d) if d ∈ [dh o −1,h o , dh o ,h o +1 )
∃1 dxy = dy x : ζx (dxy ) = ζy (dxy ) (13)

y 
x < y < z ⇒ dxy < dxz < dy z (14) ln
⎧ β x
⎨ ζx (dxy ) = ζy (dxy )
⎪ (15a)
x < y ⇒ ζx (d) > ζy (d), ∀d ∈ [0, dxy ) (15b)
⎪
⎩
ζx (d) < ζy (d), ∀d ∈ (dxy , +∞) (15c)
∀x, y, z ∈ N, ∀d, dxy , dxz , dy x ∈ R+
0 .
These three properties indicate the following.
1) (13) Any two link throughput distance curves using A
and B hops (ζA (d) and ζB (d), respectively), intersect
only once at dA B .
2) (14) As the number of hops increases, intersection dis-
tance increases, too.
3) (15) If A < B, then ζA (d) is higher than ζB (d) for all
distances lower than dA B , and vice versa.
The proof of these properties is in the Appendix. These prop-
erties are enough to show that maximum throughput function
ζmax (d) can be defined by (16), solely as a function of d. Us-
ing (1) and (11) to describe ζh (d), frontier distances (dxy ) are
dxy = α  (1−α ) x (1−α )  , ∀x, y ∈ N, x = y.
−y
(17)
Fig. 3 shows the maximum network throughput function
ζmax (d), using a different color for each section with a defined
number of hops (h ∈ [1, . . . , 5]). PDR function used in the fig-
ure is defined by α = 10.6. As this example shows, a network
with a single hop would not be able to communicate when its
length is around 1.5R or above. If hop count is incremented,
throughput is maximized and communications are improved
(≈ 0.5B).
VII. NETWORK LAYER EXPERIMENTS
We conducted several experiments to prove the multihop net-
work concept. For this, we fixed some parameters across all
experiments, while others such as number of relay nodes, the
number of slots used in each TDMA round, and end-to-end
distance were modified (cf. Table III).
904 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 13, NO. 2, APRIL 2017
Fig. 4. Experiment with a line network with three hops, transporting
data 100 m away, at a rate of ≈ 200 KB/s. Each color identifies packets
sent from a different UAV. Round is 450 ms long, and transmission slot
is 100 ms. These packet traces were captured by a monitor node, using
Wireshark.
Fig. 5. Measured throughput at different end-to-end network lengths,
using distinct number of hops. Inside their slots, nodes transmit 200-byte
packets at a rate of B ≈ 800 pkts/s. UAVs are all oriented parallel to each
other.
In order to carry out multihop experiments we needed to
implement the TDMA framework described before, over the
IEEE802.11 standard currently installed. However, just for the
sake of simplicity, and since each experiment lasted for only a
few seconds, we decided to do a simplified synchronization at
this point and leave the actual operational TDMA scheme for
future implementation. Thus, in this work, at the beginning of
each experiment, each receiver node synchronizes its clock with
the upstream routing node. The upstream node sends ten packets
containing its current clock time in the payload. The receiver
node measures the difference between the time when the packet
was received and the time stamp included in the packet. The
receiver computes the average difference between time stamps,
and updates its clock to be the same as the transmitter. This is
done once per experiment. However, this synchronization incurs
in an extra error that depends essentially on jitter affecting the
time of flight of each packet, from end-to-end application layer,
which is not very precise. Consequently, we used large guarding
windows between TDMA slots. This is orthogonal to our goal
of increasing relative throughput altering the network topology.
Fig. 4 shows two TDMA rounds where the guarding windows
are rather visible, during which there is no communication ac-
tivity.
Each dot in Fig. 5 shows the throughput obtained during
the experimental campaigns using different number of relays
considering also different end-to-end distances. Transmission
rate was set to B ≈ 800 pkts/s. No recovery, retransmissions,
or redundancy mechanisms were added to the network traffic,
beyond the minimum number of automatic retransmissions as
referred before (see Table I). All the experiments considered
200-byte long packets with all UAVs parallel to each other. This
can be converted to a (maximum) throughput of 160 KB/s, when
PDR is 100%. Blue dots are one hop data, i.e., a single link. It
comes from the link-PDR data showed at the top of Fig. 2. In
red and yellow, we see the multihop throughput. The network
maintains its throughput above 200 pkts/s until 110 m, which
is infeasible with a single hop. With an extra hop, the network
continues to communicate to ≈ 170 m, at ≈ 100 pkts/s. This rep-
resents around one-eighth of the maximum speed (at 1 m). The
red and yellow curves from Fig. 2, regarding two and three hops,
are generated using the model equations derived in the previous
section. The experimental data and the model visibly match.
Due to the large guarding windows used in the TDMA round,
these equations consider a reduction in the available throughput.
Besides the simple division by the number of slots h, we had
to account for the actual time available for communications per
slot. In our case that means 100 ms every 150 ms, i.e., a factor
of 2/3, leading to a combined reduction of 1/h × 2/3 at every
hop section (h > 1).
VIII. CONCLUSION
Multirotor UAVs are becoming a common place across many
domains. However, several challenges persist such as when us-
ing them for inspection and monitoring of industrial sites. In
these cases, their range of operation may need to be as long as
possible. In this work, we focused on this challenge creating
a chain network of UAVs. In particular, we characterized the
communications of COTS UAVs, namely the AR Drone 2.0,
using their native wireless IEEE802.11 interface. We started
by analyzing the PDR on a UAV-to-UAV link and we reached
to two main conclusions, namely that 1) our platform is not
omnidirectional in the flight plane and 2) the vehicles achieve
a maximum direct range of communication of 75 m, if they
are oriented in parallel to each other and transmitting relatively
short packets. PDR is reduced significantly if the link length is
any longer. We proposed increasing the range of the network
with a TDMA-based multihop topology using other UAVs as
relays. The performance of such model was studied and the
optimal number of relays that provide maximum throughput
was deduced. Finally we validated these results with extensive
experimental campaigns.
This work is a first step toward setting up an extendible mul-
timedia streaming system for inspection of difficult to access
assets, particularly in large process control plants. In the next
steps, we will take advantage of the GPS receivers to achieve
precise clock synchronization to achieve the desired high net-
work utilization. We will also use error control mechanisms to
enhance reliability, probably based on using rateless codes. Fi-
nally, we will also deploy an actual dynamic role assignment
policy that switches UAVs between sensors and relays, provid-
ing a compromise between the area of coverage and its distance
from the GS.
PINTO et al.: CHARACTERIZING MULTIHOP AERIAL NETWORKS OF COTS MULTIROTORS 905

APPENDIX Proof of (15c). y-hops leads to higher throughput than

Proof of (16): The ζmax (d) function, described in (16) can x-hops, if d > dxy (x < y): As a consequence of (15a), in
be derived from the intrinsic properties of ζh (d), namely (13)– the interval d ∈ (D, +∞), ζx (d) and ζy (d) never intersect each
(15). By using (15b) and (14), we show that if the network uses other (as long x = y). Since ζh (d) is continuous, we can prove
x-hops, the throughput is greater than using any higher number (15c), showing that the derivative of ζy (x) is lower than ζx (d)
of hops (x + n), as long as the distance is lower than dx,x+1 : at d = D, i.e.
⎧ ⎧1
⎨ ζx (d) > ζx+1 (d)
⎪ if d ∈ [0, dx,x+1 ) ⎨∃ D : ζx (D) = ζy (D)
⎪
 
ζx (d) > ζx+n (d) if d ∈ [0, dx,x+n ) (18) ⎪ ∂ζx (d)  ∂ζy (d)  ⇒ ζx (d) < ζy (d)∀d ∈ (D, +∞).
⎪ ⎩ <
⎩
dx,x+1 < dx,x+n ∀n ∈ N> x . ∂d d=D ∂d d=D
(23)
By using (15c) and (14), we show that if the network uses First, we derive the throughput derivative function ( ∂ ζ∂h d(d) ):
x-hops, the throughput is greater than using any lower number  
of hops (x − m), as long as the distance is higher than dx,x−1 : ∂ζh (d) ∂ B β ( hd )α h
= e
⎧ ∂d ∂d h
⎨ ζx (d) > ζx−1 (d)
⎪ if d ∈ (dx−1,x , +∞)  
h
ζx (d) > ζx−m (d) if d ∈ (dx−m ,x , +∞) (19) = ζh (d) β α dα
⎪
⎩ h
dx−m ,x < dx−1,x ∀m ∈ N< x .  
h
According to (14), dx,x+1 is greater than dx,x−1 . So = ζh (d) β α αdα −1 . (24)
h
From (18): ζx (d) > ζx+n (d) if d ∈ [0, dx,x+1 ) ⇒
Analyzing (24), we conclude the throughput function has a
ζx (d) > ζx+n (d) if d ∈ (dx,x−1 , dx,x+1 ) strictly negative derivative since β ∈ R− and ζh (d), d, α, ∈ R+.
From (19): ζx (d) > ζx−m (d) if d ∈ (dx−1,x , +∞) ⇒ So, proving that the inequality of derivatives (23) is true, is
ζx (d) > ζx−m (d) if d ∈ (dx,x−1 , dx,x+1 ) equivalent to show that
 
Yielding: ζx (d) > ζy (d) if d ∈ (dx−1,x , dx,x+1 ) ∂ ζ x (d) 


∂d
 >1 (25)
∀x, y ∈ N, x = y. (20) ∂ ζ y (d) 
∂d

d=D
We proved that ζx (d) is optimal in the interval    
∂ ζ x (d) 
(dx−1,x , dx,x+1 ) regarding any other number of hops in use.  ζx (D) β xα
xα Dα −1  y α −1
 
∂d
Proof of (15a) and (17): There is a unique intersection  =   = > 1.
∂ ζ y (d) x
point {D, Θ} ∈ R2 , where Θ = ζx (D) = ζy (D) ∂d
 ζy (D) βyα
yα Dα −1
d=D
ζx (D) = ζy (D) (26)
B (β ( Dx )α x )
e
B
= e(β ( y ) y )
D α
Proof of (14). ζx (d) intersects ζy (d) at a lower distance
x y than ζx (d) intersects ζz (d), if x < y < z: To prove that
dxy < dxz if y < z, we can show that dxy < dxw , ∀w > y,
eβ ( x ) x
D α
x
= i.e., we convert solution dxw , from (17), into a function of w,
β ( Dy ) y
α
e y and prove that dxw is a strictly increasing function of w ∈ N> x ,
  i.e.
ln x
y    
Dα =  . (21) ∂(dxw ) ln wx
β x(1−α ) − y (1−α ) =  
∂w β x(1−α ) − w(1−α )
We see now that (21) has a unique positive real solution.
Both the numerator and denominator are strictly negative due k · wk · ln (x/w) + (wk − xk )
to β < 0, y > x, and α > 1. This shows that inside the domain = 2 , (k = 1 − α).
βw(xk − wk )
interval R+ 0 , those two functions intersect only once at dA B . (27)
Proof of (15b). x-hops leads to higher throughput than
2
y-hops, if d < dxy (x < y): As a consequence of (15a), in Since kwk ln (x/w), (wk − xk ) and βw(wk − xk ) are all
the domain interval d ∈ [0, do ), ζx (d) and ζy (d) never intersect negative, the proof is complete.
each other (as long x = y). Since ζh (d) is a continuous function,
Bolzano’s intermediate value theorem is applicable. So, (15b) is REFERENCES
true iff there is at least one point in that domain, where ζx (d) is
[1] L. Pinto, A. Moreira, L. Almeida, and A. Rowe, “Aerial multi-hop net-
greater than ζy (d), ∀x < y. We know that ζx (0) is greater than work characterisation using cots multi-rotors,” in Proc. IEEE World Conf.
ζy (0), so the proof is completed: Factory Commun. Syst., May 2016, pp. 1–4.
[2] N. Goddemeier, K. Daniel, and C. Wietfeld, “Role-based connectivity
B B management with realistic air-to-ground channels for cooperative uavs,”
ζx (0) = > = ζy (0). (22) IEEE J. Sel. Areas Commun., vol. 30, no. 5, pp. 951–963, Jun. 2012.
x y
906
[3] M. Asadpour, D. Giustiniano, K. A. Hummel, S. Heimlicher, and S. Egli,
“Now or later?: Delaying data transfer in time-critical aerial communica-
tion,” in Proc. 9th ACM Conf. Emerg. Netw. Exp. Technol., Santa Barbara,
CA, USA, 2013, pp. 127–132.
[4] M. Asadpour, B. V. den Bergh, D. Giustiniano, K. A. Hummel, S. Pollin,
and B. Plattner, “Micro aerial vehicle networks: An experimental analysis
of challenges and opportunities,” IEEE Commun. Mag., vol. 52, no. 7,
pp. 141–149, Jul. 2014.
[5] J. Zhao and R. Govindan, “Understanding packet delivery performance in
dense wireless sensor networks,” in Proc. 1st Int. Conf. Embedded Netw.
Sens. Syst., Los Angeles, CA, USA, Nov. 2003, pp. 1–13.
[6] F. Jia, Q. Shi, G.-m. Zhou, and L.-f. Mo, “Packet delivery performance in
dense wireless sensor networks,” in Proc. Int. Conf. Multimedia Technol.,
Ningbo, China, Oct. 2010, pp. 1–4.
[7] S. Basagni, C. Petrioli, R. Petroccia, and M. Stojanovic, “Choosing the
packet size in multi-hop underwater networks,” in Proc. IEEE OCEANS,
May 2010, pp. 1–9.
[8] N. Ahmed, S. S. Kanhere, and S. Jha, “Utilizing link characterization for
improving the performance of aerial wireless sensor networks,” IEEE J.
Sel. Areas Commun., vol. 31, no. 8, pp. 1639–1649, Aug. 2013.
[9] A. Bohm, K. Lidstrom, M. Jonsson, and T. Larsson, “Evaluating CALM
M5-based vehicle-to-vehicle communication in various road settings
through field trials,” in Proc. IEEE Local Comput. Netw. Conf., Oct. 2010,
pp. 613–620.
[10] M. Lindhe and K. Johansson, “Using robot mobility to exploit multipath
fading,” IEEE Wireless Commun., vol. 16, no. 1, pp. 30–37, Feb. 2009.
[11] D. Henkel and T. X. Brown, “Delay-tolerant communication using mobile
robotic helper nodes,” in Proc. Int. Symp. Model. Optim. Mobile Ad Hoc
Wireless Netw. Workshops, Apr. 2008, pp. 657–666.
[12] D. Sicignano, D. Tardioli, S. Cabrero, and J. L. Villarroel, “Real-time
wireless multi-hop protocol in underground voice communication,” Ad
Hoc Netw., vol. 11, no. 4, pp. 1484–1496, Jun. 2013, (special Issue on
Wireless Communications and Networking in Challenged Environments).
[13] C. Rizzo, D. Tardioli, D. Sicignano, L. Riazuelo, J. L. Villarroel, and
L. Montano, “Signal-based deployment planning for robot teams in
tunnel-like fading environments,” Int. J. Robot. Res., vol. 32, no. 12,
pp. 1381–1397, Oct. 2013.
[14] E. F. Flushing, M. Kudelski, L. M. Gambardella, and G. A. DiCaro,
“Spatial prediction of wireless links and its application to the path control
of mobile robots,” in Proc. IEEE Int. Symp. Ind. Embedded Syst., Pisa,
Italy, Jun. 2014, pp. 218–227.
[15] “Parrot AR. Drone 2.” [Online]. Available: http://ardrone2.parrot.com,
Accessed on: Jan. 21, 2016
Luis Ramos Pinto (S’16) received the M.Sc.
degree in electrical and computer engineering
from the Faculty of Engineering, University of
Porto (FEUP), Porto, Portugal, in 2011. During
the course, he has been distinguished with the
Incentive Award from the University of Porto. He
is currently working toward the Ph.D. degree at
Carnegie Mellon University (CMU), Pittsburgh,
PA, USA, under the supervision of Prof. Luis
Almeida and Prof. Anthony Rowe, funded by a
doctoral scholarship awarded by the Portuguese
Foundation for Science and Technology.
In July 2010, he joined the Instituto de Telecomunicações where he
has been conducting his research work. In 2011, he enrolled in the
CMU|Portugal Doctoral Program in electrical and computer engineering
(ECE) jointly at CMU and FEUP. His research interests include wireless
networks and embedded systems.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 13, NO. 2, APRIL 2017
André Moreira (S’16) received the M.Sc. de-
gree in electrical and computer engineering
(ECE) from the Faculty of Engineering, Univer-
sity of Porto (FEUP), Porto, Portugal, in 2011.
During the course, he has been distinguished
with the Incentive Award and a Merit Scholar-
ship from the University of Porto, where he is
currently working toward the Ph.D. degree un-
der the supervision of Prof. Daniel Lucani and
Prof. Luis Almeida, funded by a doctoral schol-
arship awarded by the Portuguese Foundation
for Science and Technology.
In 2012, he enrolled in the Doctoral Program in ECE at FEUP. His re-
search interests include network coding and wireless communications.
Luis Almeida (SM’86) received the Graduate
degree in electronics and telecommunications
engineering in 1988 and the Ph.D. degree in
electrical engineering in 1999, both from the Uni-
versity of Aveiro, Aveiro, Portugal.
He is currently an Associate Professor
in the Electrical and Computer Engineering
Department, University of Porto (UP), Porto,
Portugal, and a member of the Instituto de
Telecomunicações at UP where he coordinates
the Networked Systems group. His research in-
terests include RT communications for distributed industrial/embedded
systems including teams of cooperating agents.
Among several appointments, Dr. Almeida was a member of the
Executive Committees of the IEEE Technical Committees on RT Sys-
tems (2008–2013) and Factory Automation (2007–2009, since 2014),
the Program and General Chair of the IEEE RT Systems Symposium
(2011 and 2012, respectively), Vice-President (2011–2013) and Trustee
(2008–2016) of the RoboCup Federation, and an Associate Editor of the
Springer Journal of RT Systems (since 2010) and Elsevier Journal of
Systems Architecture (since 2012).
Anthony Rowe (M’07) received the Ph.D. de-
gree in electrical and computer engineering
(ECE) from Carnegie Mellon University (CMU),
Pittsburgh, PA, USA, in 2010.
He is an Associate Professor in the ECE De-
partment, CMU. His research interests include
networked real-time embedded systems with a
focus on wireless communication. His most re-
cent projects have related to large-scale sensing
for critical infrastructure monitoring, indoor local-
ization, building energy-efficiency and technolo-
gies for microgrids. His past work has led to dozens of hardware and
software systems, four best paper awards and several widely adopted
open-source research platforms.
Dr. Rowe received the Lutron Joel and Ruth Spira Excellence in Teach-
ing Award in 2013, the CMU CIT Early Career Fellowship, the Steven
Ferves Award for Systems Research in 2015, and the Dr. William D. and
Nancy W. Strecker Early Career chair in 2016.