From Shadows to Light: A Swarm Robotics Approach with Onboard

Control for Seeking Dynamic Sources in Constrained Environments

Tugay Alperen Karagüzel1 , Victor Retamal1 , Nicolas Cambier1 , Eliseo Ferrante1,2

Abstract— In this paper, we present a swarm robotics control collective behaviors that emerge from simple local interac-
and coordination approach that can be used for locating a tions between individual agents [5]. Given that a swarm is a
moving target or source in a GNSS-denied indoor setting. The group of decentralized, autonomous, and simple agents, that
approach is completely onboard and can be deployed on nano-
drones such as the Crazyflies. The swarm acts on a simple set can work together to achieve a common goal, it presents
of rules to identify and trail a dynamically changing source an apt framework for our constraints. Swarm robotics takes
gradient. To validate the effectiveness of our approach, we con- inspiration from the collective behaviours observed in nature:
duct experiments to detect the maxima of the dynamic gradient, The collective sensing behaviours in fish schools, as explored
which was implemented with a set of lights turned on and off in [6], [7], offer substantial inspiration for our previous [8]
with a time-varying pattern. Additionally, we introduce also
a minimalistic fully onboard obstacle avoidance method, and and the current study. Thus, drawing upon such insights from
assess the flexibility of our method by introducing an obstacle nature allows us to apply them beneficially within the context
into the environment. The strategies rely on local interactions of our research.
among UAVs, and the sensing of the source happens only at the The task of source seeking becomes even more difficult
individual level and is scalar, making it a viable option for UAVs in the presence of dynamic or moving sources, and when
with limited capabilities. Our method is adaptable to other
swarm platforms with only minor parameter adjustments. Our requiring also collective motion towards the source, that is
findings demonstrate the potential of this approach as a flexible individuals maintaining safe and similar distances to each
solution to tackle such tasks in constrained GNSS-denied indoor other while moving in the same direction [9]. Here, UAVs
environments successfully. must coordinate their collective motion, sense and navigate
the environment, detect and locate the source of interest,
I. INTRODUCTION while constantly balancing a compromise between updating
Unmanned aerial vehicles (UAVs) have proven to be a their collective estimation of the source and coordinating
valuable tool for source seeking in a wide range of appli- their movements.
cations [1], [2]. However, detecting a source with UAVs To solve the problem of performing collective motion
poses significant challenges [3], particularly in complex while seeking a dynamic source in a GNSS-denied indoor
environments. Indeed, such endeavour requires high levels of environment, this study introduces an approach for con-
autonomy and adaptability, particularly in unknown GNSS- trolling fully onboard a nano UAV swarm. The approach
denied environments where communication may be limited is not based on complicated control and state estimation
or intermittent. Source seeking UAVs can find applications methods. Rather, the swarm utilizes relative localization,
in outdoor or indoor environments. UAVs for outdoor en- limited onboard sensing and no communication on source
vironments, larger in size, enjoy more flexibility in sensor measurements, and achieves dynamic source seeking through
hardware and computational resources, allowing resource- emergent self-organization. In this study we show, for the
intensive operations. Conversely, UAVs for indoor envi- first time, a fully onboard implementation, with a genuine
ronments face space constraints and necessitate a compact and dynamically changing gradient source (in the exper-
design with limited sensor and computational capacities. iments implemented via light emitters), sensed onboard in
Because of these restrictions, the design of hardware and the presence of obstacles.
control for indoor UAVs is more challenging compared II. RELATED WORK
to outdoor UAVs. This mandates strategic efficiency en-
hancements, particularly in algorithm design and resource Previously, several works have attempted to tackle the
utilization, to meet objectives within these spatial confines. problem of localizing a source by using UAVs with limited
To circumvent the constraints inherent to indoor platforms sensing capabilities. In the context of single UAVs, the
and to achieve the above objectives, an approach based on authors in [10] employed a deep reinforcement learning
multiple UAVs offers a compelling solution, as evidenced algorithm to locate a solitary, stationary light source, utilizing
by previous research [4]. Particularly, swarm robotics studies multiple lightweight sensors on the UAV. Their platform was
also capable of obstacle avoidance. The learning pipeline
1 Vrije Universiteit Amsterdam, Noord-Holland, incorporated a training-in-simulation phase prior to actual
The [email protected],
Netherlands, deployment. The authors in [11] introduced a bio-inspired
[email protected], solution, mimicking bacterial chemotaxis, to a source lo-
[email protected]
2 Autonomous Robotics Research Center, Technology Innovation Institute calization problem, but without employing any learning
Abu Dhabi, United Arab Emirates [email protected] methodology. They deployed a single nano-UAV equipped
with a temperature sensor to identify a heat source. The an obstacle (which is not achieved in [13], [14] and [15]).
UAV leveraged historical position and measurement data to Each drone senses the gradient via its sensor and retrieves a
estimate an assumed form of gradient, successfully localizing scalar value (the intensity of the source at its location), but
the heat source. it does not communicate this measurement to other drones
Besides single-UAV applications, there is also work on (unlike the cases where measurements are communicated to
multi-UAV approaches. The authors of [12] focus on an peers for the estimation in [12], [13], [14] and [15]). The
extensive application employs nano-drones to localize a drone does not store nor utilizes any historical data, and the
stationary gas leak source using a specialized sensor. In method operates reactively unlike other studies we discuss
this application, UAVs exchange measurement data and ap- in this section.
proximate relative distances to pinpoint the source, all while
circumventing obstacles. Despite these coordinated actions, III. M ETHODOLOGY
the system does not demonstrate any collective motion. The In this section, we outline our methodology for addressing
authors of [13] introduce a level of collective movement the challenge of dynamic source localization with obstacle
where nano UAVs localize an emulated source, implemented avoidance in the absence of global positions and directional
as a stored function onboard each UAV. The computed information about the source location. This method considers
source function values are shared among UAVs, subsequently three main behavioural building blocks: collective motion,
integrating with the flocking algorithm. In a simulation emergent source localization, and obstacle avoidance. All
study by [14], multiple UAVs attempt to locate a single these building blocks are fully reactive and do not require
static signal source utilizing an extremum seeking algorithm. information about past sensory readings. Moreover, they are
The UAVs estimate the signal strength gradient’s peak by all deployed on each drone in a fully onboard fashion. The
averaging each other’s measurements. Lastly, the authors in collective motion building block requires only the relative
[15] present centralized adaptive navigation control strategies positions and headings of neighboring agents. With this
to characterize sources in 2D and 3D scalar fields, also building block, we expect that the swarm will move in
accounting for plume dynamics. The process incorporates a cohesive and ordered fashion, maintaining safe and pre-
the collective use of individual measurements from diverse dictable distances between agents. For the source localization
environmental points by the cluster members. Unfortunately, building block, the source can be sampled using an on-
despite accurate modeling of UAV dynamics, the work was board sensor that gives scalar information on the source
only implemented in simulation. intensity at the drone position. The building block employs
In our previous work [8], we have implemented a set of scalar readings of the gradient field strength at their current
rules used for gradient following, that has been evaluated locations to modulate their desired distances to maintain
using a centralized implementation and sensor emulation with others [8], to achieve emergent source localization
for static gradient forms devoid of obstacles. In addition, (converging to the source location in a way that is only
we tested various scalar field types, swarm sizes, and be- achievable as a collective, individuals are incapable of doing
havioral modifications (such as alignment control) through that [7]). Simultaneously, the centroid of this cohesive swarm
simulation experiments. In these simulations, our proposed is expected to follow the increasing gradient of the source
method demonstrated its effectiveness not only on a real- strength, ultimately converging on the source center. The
istic dynamic model of Crazyflie nano drones but also on hypothesis in this paper is that, when the source location
point agents modeled at a comparatively large scale. In this is changed, the swarm will react accordingly to the changed
work, we implement and demonstrate for the first time a source location. The obstacle avoidance building block uses
fully onboard implementation of a bio-inspired algorithm for 4 time-of-flight on-board sensors acting on the principal
gradient following, using an approach that is a significant directions (front, back, left, right) to perceive obstacles.
extension of the one presented in [8]. We have enhanced This building blocks combines three contributions: virtual
our previous implementation by adapting our methodology to repulsion forces produced according to the direction away
accommodate the inputs from a real gradient source instead from the obstacle; an attraction force towards the agent’s
of emulated values, introducing obstacle avoidance capability neighbors; and a random perturbations component to avoid
with an on-board distance sensor, and manually optimizing agents resolving directional deadlocks.
all control rules in light of these new factors. Unlike the The individual controller of each UAV calculates linear
previous approach, all rules are now implemented on the speed (Ui ) and angular speed (ωi ). UAVs can translate only
same micro controller serving as the main flight controller, in their heading direction by the magnitude of Ui , and
eliminating the need for centralized implementation. this direction is changed by ωi over time. Although the
We use fully onboard controlled Crazyflie nano-drones platform we use (multi-rotor aerial vehicle) can translate in
in an indoor environment (unlike control from a central any direction, we impose this constraint on the platforms
computer as in [11] or only on simulation as in [14] and to apply our methodology, which ensures smooth motions
[15]), a physical custom-made on-board sensor to imple- of the UAVs and improves the stability of the collective
ment scalar source sensing (differently from [13] where motion [16]. In addition, the translation constraint makes our
the measurements are calculated by a stored function), and methodology easily portable to any kind of platform with
another physical on-board sensor for sensing and avoiding minor changes only in the parameters.
 The controller takes the relative positions and headings of the motion into a more reactive, smooth and faster one.
other UAVs as the input. In addition, each UAV uses its own To retain these properties with a dynamic gradient source
sensor reading as a source value. Individual sensor readings without affecting the performance adversely, we have chosen
are used in such a way that UAVs with higher values keep to keep alignment control and showcase the best possible
a longer distance to their neighbours, without letting others performance. ⃗hi is calculated as a unit vector. The direction
know about it. Later, the vectors are combined to produce of ⃗hi is the average of heading directions of all neighbors,
Ui and ωi , which are applied on each control step. resolved in the local frame of the focal UAV.
The core of our method consists in creating a virtual spring jθ0
P jθm
between a focal UAV i and its neighbors, by modulating ⃗hi = e + Pm∈N e (3)
their interpersonal distance according to the focal UAV’s ||ejθ0 + m∈N ejθm ||
individual sensory input (Ls ) and the following equation: In the above, θm represents the heading angles of neighbor
 0.1 UAVs and θ0 stands for the heading of the focal UAV.
Ls − Lmin
σi = σL + σr (1) Then, the boundary avoidance vector, r⃗ib , is responsible
Lmax − Lmin for producing a repulsion from the boundary (b) if the focal
In the above, σL and σr stand for the lowest allowed UAV is closer to it than a safe distance, Db . r⃗ib is calculated
value and the modulation range of σi , respectively. This√ as the combination of repulsion from all boundaries (B) in
modulation increases the desired distance ddes = σi 2 which the focal UAV is closer than Db .
of the focal UAV to other UAVs when it is exposed to   ⃗b !
higher sensor readings, which can also be seen as more ⃗b
X 1 1 pi
ri = krep − (4)
asocial behavior. Differently from [8], Equation 1 includes Lb L0 L3b
b∈B
an exponent, not present in the original formulation, that
is introduced to cope with realistic environmental gradient In the above, krep designates the strength of r⃗ib and L0 sets
profiles. The value of the exponent must be less than 1 the relaxation threshold. Lb is the shortest distance to the
because the real gradient tends to change rapidly near the boundary b among all close-by boundaries, B. p⃗bi points the
maximum and more smoothly in darker regions. To better relative direction from the boundary b to the focal UAV to
respond to small changes in these darker areas, σi increases produce repulsion.
more quickly when Ls has lower values. The exponent of 0.1 The final component of f⃗i is the obstacle avoidance vector,
was found manually to achieve the best performance with the ⃗i
o (not considered in [8] and added for the first time in this
real gradient source. paper).
The first vector is the proximal vector p⃗i and therefore
computes the velocity required to maintain ddes between o⃗i = o⃗id
X
(5)
UAV i and each UAV m, that is part of the former’s d∈D
neighborhood N :
where, superscript d represents the direction of sensed ob-
σ4 σ2 stacles and o⃗id represents the sub-vector for each direction.
 
X m
p⃗i = −ϵ 2 mi 5 − mi 3 ejϕi (2)
(di ) (di ) In our application, d can only be in four principal directions
m∈N
(the set D) of the local reference frame; front, back, left, and
where dm i is the relative distance between UAVs i and m, right, where the front is the current heading direction of the
ϕmi is the relative bearing angle of m on the reference frame focal UAV. o⃗id for each direction is calculated as follows:
of i and ϵ determines the strength of p⃗i . Please note that the
notation ejX , where X is an angle in radians, follows Euler’s 1
o⃗id = τ (δdi )ejφ + I(d) · τ (δdi )ejΩ (6)
formula. In this representation, the real part corresponds to 2
the x-axis and the complex part to the y-axis in the reference where δdi is the corrected measured distance with an offset
frame where angle X is defined. Using this approach, p⃗i of 15 cm, τ (δdi ) is the magnitude of the sub-vector:
and other vectors are computed as 2D vectors with x and y
components in the local reference frame of the focal UAV i. 2
τ (δ) = (2.5/δ) (7)
Since neighboring UAVs that sense lower values than the
focal UAVs, have a comparatively shorter ddes , and are and (
therefore still attracted to the focal UAV. p⃗i enables a chase- 1 if d is front
I(d) = ;d ∈ D (8)
evade phenomenon that is directed towards the increasing 0 otherwise
values of the source gradient [8]. Also,
The next vector is the alignment control vector ⃗hi , which π
contributes to making the heading directions of UAVs similar. Ω = sign(∠⃗
pi ) , (9)
2
The success of the swarm on the gradient following task
φ = υi + ψ + π. (10)
is analyzed with and without the alignment control on our
previous work [8]. The finding in [8] was that alignment where υi is the angle of obstacle direction in the local
control increases the order in the swarm and transforms reference frame (υi ∈ [π, − π2 , π2 , 0]). π radians are added
to produce the repulsion in the opposite direction. ψ is a on the global frame) using their on-board sensors (as im-
random perturbation angle where ψ ∼ U nif (− 12 , 12 ). ψ was plemented in the Crazyflie stack). They exchange positions
introduced to add randomness to the obstacle repulsion to and headings with each other via a radio communication
break any local minimum that can lock the UAV’s movement. channel using P2P API [19]. Crucially, global positioning
Ω has a similar effect; it adds an extra perturbation in the and communication are only required due to our drones’
direction angle (∠⃗ pi ) of the proximal control vector (⃗ pi ). inability to perform relative localization. If they were able to
If ∠⃗pi is positive, the perturbation produces an effect as if estimate the range and bearing of neighboring drones, com-
there is an obstacle on the right, and if it is negative, the munication would not be necessary for our methodology. In
perturbation produces an effect as if there is an obstacle on addition, each drone has a flow camera pointing downwards
the left. In the end, all o⃗i d calculated for all directions (front, to improve planar velocity estimation and a distance sensor
back, left, right) are combined as in Eq. 5 to get the final ranging the altitude.
vector, ⃗oi . For the obstacle sensing capability, our Crazyflies carry
Ultimately, the total force vector f⃗i is the weighted sum a deck (multi-ranger) that contains 4 time-of-flight distance
of all vectors with respective weights, α, β, γ, and κ. sensors in the principal directions; front, back, left, and right.
These sensors provide distance to the closest object in the
f⃗i = α⃗
pi + β⃗hi + γ⃗ri + κ⃗
oi (11) directions they face. Given the design of the sensor, blind
spots could potentially appear when the obstacle is located
The linear (Ui ) and angular speeds (ωi ) are calculated as
between these directions. Distance measurement is done on
follows:
a ray, so rather than a field of view, we obtain the distance
for a single point. The distance values are used as δdi for the
Ui = K1 fx + uadd , ωi = K2 fy (12)
corresponding direction.
In the above, fx and fy are the components of f⃗i on axes Physical light sources were used and placed in the en-
of the local reference frame of the focal UAV, where the vironment to realize the experimental version of the source
x-axis is always the heading direction. K1 and K2 are the gradient. We built and equipped the Crazyflies with a custom-
respective speed gains, and uadd is the additional propulsion made light sensor deck on top, that uses TEMT6000 [20]
we add to decrease stagnancy in the swarm motion. light sensor circuit, which changes the output voltage de-
Ui is clamped within the range [0, 0.13] m/s and ωi is pending on the light intensity in the environment. The inter-
clamped within the range [−0.4π, 0.4π] rad/s. uadd is chosen nal software reads this voltage for use in our methodology
to be 0.05 m/s. The other parameter values are chosen as as Ls . Minimum and maximum voltage values measured in
follows: ϵ = 12.0, σL = 0.5, σr = 0.56, Lmin = 52, our experimental setup led to our choice of Lmin and Lmax
Lmax = 648, Db = 0.5 m, krep = 5.0, L0 = 0.5, α = 0.8, values. Our custom Crazyflie platform can be seen in Figure
β = 0.2, γ = 1.0, κ = 2.0, K1 = 0.06 and K2 = 0.6. The 1. The visualization of the system architecture can be found
parameter values are tuned and set manually according to in Figure 2.
our evaluation. Only Lmin and Lmax depend on the sensory The experiment arena dimensions are 7.0 m in length and
hardware and environmental conditions. 4.75 m in width. For the whole flight, a constant altitude of
0.5 m is maintained by all UAVs. The arena includes light
IV. I MPLEMENTATION AND E XPERIMENTAL S ETUP
bulbs on the top, switched on and off by smart plugs. These
The algorithm presented in the previous section is im- smart plugs receive messages over the network, to which
plemented on Crazyflie 2.1 nano-drones, and runs on-board a control computer is also connected. These individually
by using the App-Layer functionality [17]. The open-source switchable light sources (which are completely extrinsic for
implementation can be found in the GitHub repository [18]. the swarm) constitute the dynamic gradient to be followed by
UAVs calculate their own global positions (X and Y on the
global frame) on-board with an Ultra-wide Band Localization
system. They determine their heading (current yaw angle

Fig. 1: Crazyflie equipped with sensors Fig. 2: System architecture

the swarm. Figure 3 shows the position of the light sources as
well as the measured light intensity and illuminance values
collected from 6 different flights conducted with a single
drone for each light source. This drone visited randomly
chosen 100 points when only the corresponding light source
was on, and the control computer recorded the position-
light intensity data together. For the sensory input values
obtained, which theoretically range from 0 to 4096 due to the
analog voltage input, the corresponding illuminance levels
are calculated based on the sensor’s datasheet [20]. Factors
such as temperature, variations in wavelength, and electrical
losses are not considered in these calculations. Red lines
show the arena boundary. We use the information visualized
in Figure 3 for two purposes: Deciding Lmin , Lmax values, Fig. 4: Crazyflie swarm in the air, experiment D
and locating the light centers to assess the gradient following
performance of the swarm.
In order to test the performance of the swarm following way, without any crash. It should be noted that obstacle only
the dynamic gradient, we designed 4 sets of experiments. exists in experiment D.
In experiment A, the lights are switched on sequentially, Each experiment starts with the control computer trigger-
starting from L1 to ending with L6. Each light stays on ing an internal switch in the on-board software of Crazyflies,
for 35 seconds. In experiment B, the L4-L2-L6 sequence is through radio communication. Simultaneously, a timer starts
followed with a switch every 65 seconds. The duration for on the computer, which sends switch signals to the lights.
each switch is longer in B, since light sources are further Afterward, the positions, heading angles, and light measure-
and intersections between lightened areas are smaller. In ments of the UAVs are obtained and recorded 2 times per
experiment C, lights are switched two by two with the second for the assessment. When the experiment timer is
following sequence: [L1-L6]-[L2-L5]-[L3-L4]-[L2-L5]-[L1- finished, a signal is sent, which causes all UAVs to land.
L6]. By doing so, we want to test the performance of our
approach on a bimodal gradient. The period for the light V. R ESULTS AND D ISCUSSION
switch is 50 seconds at C. Finally, experiment D follows the First, we observe the gradient following behavior by
sequence L3-L4-L5, with the addition of an obstacle, placed analyzing the trajectory of the swarm centroid. The positions
on the location OBS, as shown in Figure 3. The circle- of the UAVs are averaged for each data point to get the
shaped obstacle is built using off-the-shelf materials such centroid position. We use centroid trajectory because, when
as paper and wooden sticks, shown in Figure 4. Given the the swarm arrives at the next light center successfully, UAVs
design of the sensor (possibility of blind spots), the diameter position themselves around the center instead of clustering
is selected to be 80 cm to guarantee that UAVs can sense in the middle as they must maintain a distance (ddes ) from
the obstacle from a safe distance of approximately 14 cm. each other by our methodology. This makes observing the
As the obstacle makes moving from L3 to L4 and L4 to L5 centroid trajectory more intuitive than observing individual
more challenging, the switching periods are set to 20-90-90 UAV paths. Nevertheless, links for animation videos of the
seconds. With this setup, we assess the swarm’s capability whole set of experiments can be found on the playlist 1 ,
to follow changing light sources, when an obstacle is on the where individual UAV motions can also be seen. In addition,
real flight videos can be found on this playlist 2 .
In Figure 5a, centroid trajectories for 3 runs of experiment
A, can be seen together. Colors indicate the current active
light source, and stars indicate the position of the centroid
when the light is switched to the next one. Colored regions
show the approximate areas illuminated by the corresponding
light sources. The trajectories and the locations of the stars
from 3 different experiments indicate a successful gradient
following behavior and the emergence of collective local-
ization of the source center. Figure 5b shows the distance
of the centroid to the active light center over time (light
1 https://www.youtube.com/playlist?list=
PLQXaE6NbHSe2d1Y2zXLiO3y00d4w9OrXR
Fig. 3: Light source positions and light intensity heat map, or http://bit.ly/3QDp2W6
2 https://www.youtube.com/playlist?list=
measured sensory input and corresponding illuminance val-
PLQXaE6NbHSe39z1X2zpjZD6f6mB9f40u9
ues or https://bit.ly/3MnF3gq
 (a) Centroid trajectory (b) Distance to light centers
(a) Centroid trajectory (b) Distance to light centers
Fig. 6: Experiment B

light center. As expected from the limited maximum speed

of UAVs and increased distance between centers, reaching
the green center from the yellow center takes longer. When
(c) Inter-UAV distances (d) Swarm order Figure 6b is also considered, we can conclude that the swarm
Fig. 5: Experiment A still achieves gradient following in a steady and repeatable
manner.
In experiment C, where the lights are switched two by
center positions are estimated from Figure 3). Lines indi- two, starting positions are more diverse (closer to one of
cate the average of 3 experiments, whereas shaded regions the centers or intersection region) and we could therefore
are spanned between maximum and minimum values. This expect the swarm’s starting position to significantly alter the
figure clearly shows the steady and repeatable nature of the emerging behavior. To assess the flexibility of our algorithm,
collective following of a dynamic gradient. Figure 5c shows we consider the three cases that are possible within the first
the average, maximum, and minimum distances between all region consisting of L1 and L6; The swarm can initially
UAVs of the swarm at each data point. Here we can see be located around the L1 center, between L1 and L6, or
that UAVs never become dangerously close nor crash in any around the L6 center. Except from scenarios where individual
of the 3 experiments. The most important information we sources should be localized separately, we believe converging
can extract from this data is that, although our methodology to the intersection between two sources is a valid result
depends on the heterogeneity of desired distances of the in scenarios that requires scanning and covering multiple
UAVs, the difference thereof never becomes large enough to sources simultaneously. Figures 7a, 7c and 7e show centroid
disturb the swarm’s coherence or to increase the risk of inter- trajectories from experiments in which the swarm starts from
UAV collisions. To analyze Figure 5d, we define a metric each of these 3 starting point options.
for the agreement of the UAV’s headings; order. The order Different from previous figures, starting points are indi-
is calculated as follows: cated with a black triangle, and trajectories consist of two
P jθi colors corresponding to the active lights. We see in the
∥ e ∥ trajectory plots that, regardless of the starting point, the
i∈NA
Ψ= (13) swarm is able to proceed to the next light region after
NA the light switch. Unlike other experiments, the swarm is
In the above, θi represents the heading angle of each UAV not necessarily located in a center; it can also stay at the
within the whole swarm (NA ). Ψ becomes 1 when all intersection of two light regions, which has an attractive light
headings are perfectly parallel and approach zero when intensity. Figures 7b, 7d and 7f show distances to the light
headings are different. When Figure 5d and 5b are observed centers for each experiment. Here we can clearly observe
together, we can see that Ψ increases when the active light is the same phenomenon: the swarm sometimes converges
switched, and the swarm starts to move in the direction that towards the intersection of regions. Although 7e reveals
increases the new light gradient. When the swarm centroid different results regarding where the swarm positioned itself,
becomes closer to the light center, Ψ decreases, and the our current analysis would require further refinement to
swarm stops. This correlation between distance to the light answer the question: What determines whether the swarm
center and the swarm order helps to validate, once more, wanders at the intersection or gravitates toward the centers?
the coupling between collective behavior and environmental In order to answer this question, future experiments we will
change. conduct would necessitate a more consistent set of initial
Figure 6a shows the centroid trajectories from 3 different conditions coupled with enhanced precision in light field
experimental runs B. Here we can observe that the yellow control. While our current study does not encompass this
light region (L4) and green light region (L2) are located in-depth behavioral analysis, it lays the foundation for such
considerably further than consecutive ones as in experiment investigations, highlighting the criticality of understanding
A. Despite that, the swarm successfully follows the weaker collective gradient following behavior.
gradient (comparatively to A) in all 3 runs, and reaches the In experiment D, we test the gradient following behavior
 (a) Centroid trajectory (b) Distance to light centers (a) Centroid trajectory (b) Min. distance to obstacle

(c) Centroid trajectory (d) Min. distance to obstacle

(c) Centroid trajectory (d) Distance to light centers

(e) Centroid trajectory (f) Min. distance to obstacle

(e) Centroid trajectory (f) Distance to light centers Fig. 8: Experiment D
Fig. 7: Experiment C

errors. If we were to have smaller obstacles, the situation

could be expected to be worse. The most reliable way of
3 times when there is an obstacle on the way, which can be fixing this issue is by having sensors in more directions,
locally sensed by each UAV within a limited range (Db ). In possibly omnidirectional (360◦ ). Yet we still see this as
Figures 8a, 8c and 8e, trajectories of both swarm centroids an achievement to integrate on-board obstacle avoidance
and the UAVs from experiment D are presented. The UAV capability to the swarm without diminishing the follow of the
trajectories help validating that there was no crash between dynamic gradient. In future work, we aim to further refine the
the UAVs and the obstacle during the experiments and integration between gradient following and obstacle avoid-
demonstrate the avoidance behavior. For further information, ance. By doing so, we anticipate smoother movement of the
Figures 8b, 8d, and 8f report the minimum distance occurring swarm around obstacles, eliminating the staggered motions
between any UAV and the obstacle during experiments. observed in the current implementation.
When both types of figures are observed, we can see that In addition to the obstacle avoidance, the most pressing
the swarm successfully followed the light gradient. The upgrade before tackling real-world problems is relative lo-
stars, which represents the centroid location when lights are calization. While our current methodology needs no change
switched, clearly shows the swarm took position around the to comply, our nano drones have not achieved it yet. A
bright light center. Nevertheless, we see that the distance solution by using UWB radios currently equipped should
to the obstacle can decrease to unsafe values occasionally. also be scalable and similarly accurate to guarantee the same
Despite our model being designed to avoid such situations, effectiveness. Additionally, although we have tested real light
hardware is a limiting factor in this instance. For instance, gradients, real-world sources like gases will introduce more
due to the design of the distance sensor, it is reasonably noise and complexity. Future tests should adapt to these
possible for UAVs to miss the obstacle when the obstacle challenges.
is located between two of these principal directions, relative
to the UAV. There are only two ways to compensate for VI. C ONCLUSION
these blind spots: Either allowing UAVs to get closer to the In this paper, we present an approach for a swarm of nano
obstacle, or making sure the obstacle is large enough. Since drones to follow the gradient of a dynamic source without
we have a limited space to fly, the optimum point we found exchanging any information about the scalar measurements
between these two alternatives was to set a safe obstacle made by individual UAVs. The control and sensing are
diameter, as stated in Section IV. Even so, occasional unsafe implemented fully on-board, for the first time. In addition,
moments were not entirely avoidable, given the localization obstacle avoidance capability is also achieved on-board and
does not need an explicit information exchange within the [8] T. A. Karagüzel, A. E. Turgut, A. Eiben, and E. Ferrante, “Collective
swarm. This demonstration of on-board gradient following, gradient perception with a flying robot swarm,” Swarm Intelligence,
pp. 1–30, 2022, https://doi.org/10.1007/s11721-022-00220-1.
with a swarm of drones where each can fit in a palm, is a [9] T. Vicsek and A. Zafeiris, “Collective motion,” Physics Reports, vol.
promising step towards more challenging applications. 517, no. 3-4, pp. 71–140, Aug. 2012, https://doi.org/10.1016/j.physrep.
Our framework is designed to be generic, so that deploy- 2012.03.004.
[10] B. P. Duisterhof, S. Krishnan, J. J. Cruz, C. R. Banbury, W. Fu,
ment on various platforms does not require any modification A. Faust, G. C. de Croon, and V. J. Reddi, “Tiny robot learning (tinyrl)
of the fundamental blocks. The only building block we use for source seeking on a nano quadcopter,” in 2021 IEEE International
that still require external infrastructure is the relative local- Conference on Robotics and Automation (ICRA). IEEE, 2021, pp.
7242–7248, https://doi.org/10.1109/ICRA48506.2021.9561590.
ization, which in turn imposes the constraint of having com- [11] N. Elkunchwar, V. Iyer, M. Anderson, K. Balasubramanian, J. Noe,
munication. However, (even slightly) more capable platforms Y. Talwekar, and S. Fuller, “Bio-inspired source seeking and obstacle
that enable relative localization of peers, would not need any avoidance on a palm-sized drone,” in 2022 International Conference
on Unmanned Aircraft Systems (ICUAS). IEEE, 2022, pp. 282–289,
communication between UAVs at all. Moreover, any type of https://doi.org/10.1109/ICUAS54217.2022.9836062.
sources can be followed (gas source, radiation source, signal [12] B. P. Duisterhof, S. Li, J. Burgués, V. J. Reddi, and G. C. H. E.
source, etc.), as long as the UAVs can be equipped with de Croon, “Sniffy bug: A fully autonomous swarm of gas-seeking
nano quadcopters in cluttered environments,” in 2021 IEEE/RSJ Inter-
relevant sensors. In any case, these sensors will only need national Conference on Intelligent Robots and Systems (IROS), 2021,
to provide scalar samples (signal strength, parts per million pp. 9099–9106, https://doi.org/10.1109/IROS51168.2021.9636217.
(PPM) values, etc.). On larger platforms, omnidirectional [13] A. Datar, P. Paulsen, and H. Werner, “Flocking towards the source:
Indoor experiments with quadrotors,” in 2020 European Control
ranging sensors can be used to avoid obstacles in various Conference (ECC). IEEE, 2020, pp. 1638–1643, https://doi.org/10.
shapes and sizes. 23919/ECC51009.2020.9143867.
[14] S. Zhuo, “Source seeking of multi-uav based on extremum seek-
Considering the wide range of possibilities, we plan to ex- ing algorithm,” in 2017 17th International Conference on Control,
tend the capabilities of our swarm in the future by developing Automation and Systems (ICCAS). IEEE, 2017, pp. 1062–1067,
more capable platforms, or improving our method in different https://doi.org/10.23919/ICCAS.2017.8204373.
[15] R. K. Lee, C. A. Kitts, M. A. Neumann, and R. T. Mcdonald, “Multiple
ways (e.g. swarm speed, improved obstacle avoidance with uav adaptive navigation for three-dimensional scalar fields,” IEEE
multiple static or dynamic obstacles, 3D source localization, Access, vol. 9, pp. 122 626–122 654, 2021, https://doi.org/10.1109/
etc.). We believe this technology can be relevant to real world ACCESS.2021.3107854.
[16] E. Ferrante, A. E. Turgut, C. Huepe, A. Stranieri, C. Pinciroli, and
problems, which require reaching places that are dangerous, M. Dorigo, “Self-organized flocking with a mobile robot swarm: a
and efficient search behaviors, in case of emergencies. In line novel motion control method,” Adaptive Behavior, vol. 20, no. 6, pp.
with the inherent requirements of real-world applications, our 460–477, 2012, https://doi.org/10.1177/1059712312462248.
[17] (2022, Sep.) Bitcraze - crazyflie firmware - app layer documentation.
future vision aims at considering larger scale scenarios, and [Online]. Available: https://www.bitcraze.io/documentation/repository/
to enhance the swarm’s exploration capabilities, in addition crazyflie-firmware/master/userguides/app layer/
to following a newly detected source gradient. Although this [18] (2023, Feb.) Project repository. [Online]. Available: https://github.
com/tugayalperen/Gradient following firmware
objective requires a significant transformation in collective [19] (2022, Sep.) Bitcraze - crazyflie firmware - p2p documentation.
behavior, it will allow the swarm to autonomously search [Online]. Available: https://www.bitcraze.io/documentation/repository/
for an unknown source or transition between multiple non crazyflie-firmware/master/functional-areas/p2p api/
[20] (2023, Feb.) Vishay semiconductors - temt6000x01 datasheet.
overlapping sources. [Online]. Available: https://www.vishay.com/docs/81579/temt6000.pdf

R EFERENCES
[1] J. Han, Y. Xu, L. Di, and Y. Chen, “Low-cost multi-uav technologies
for contour mapping of nuclear radiation field,” Journal of Intelligent
& Robotic Systems, vol. 70, pp. 401–410, 2013, https://doi.org/10.
1007/s10846-012-9722-5.
[2] Z. Fu, Y. Chen, Y. Ding, and D. He, “Pollution source localiza-
tion based on multi-uav cooperative communication,” IEEE Access,
vol. 7, pp. 29 304–29 312, 2019, https://doi.org/10.1109/ACCESS.
2019.2900475.
[3] J. Burgués and S. Marco, “Environmental chemical sensing using
small drones: A review,” Science of the total environment, vol. 748,
p. 141172, 2020, https://doi.org/10.1016/j.scitotenv.2020.141172.
[4] M. Campion, P. Ranganathan, and S. Faruque, “Uav swarm commu-
nication and control architectures: a review,” Journal of Unmanned
Vehicle Systems, vol. 7, no. 2, pp. 93–106, 2019, https://doi.org/10.
1139/juvs-2018-0009.
[5] E. Şahin, “Swarm robotics: From sources of inspiration to domains of
application,” in International workshop on swarm robotics. Springer,
2004, pp. 10–20, https://doi.org/10.1007/978-3-540-30552-1 2.
[6] J. G. Puckett, A. R. Pokhrel, and J. A. Giannini, “Collective gradient
sensing in fish schools,” Scientific reports, vol. 8, no. 1, pp. 1–11,
2018, https://doi.org/10.1038/s41598-018-26037-9.
[7] A. Berdahl, C. J. Torney, C. C. Ioannou, J. J. Faria, and I. D.
Couzin, “Emergent sensing of complex environments by mobile
animal groups,” Science, vol. 339, no. 6119, pp. 574–576, 2013,
https://doi.org/10.1126/science.1225883.