Journal of Network and Computer Applications 221 (2024) 103769

Contents lists available at ScienceDirect

Journal of Network and Computer Applications

journal homepage: www.elsevier.com/locate/jnca

A fast coordination approach for large-scale drone swarm

Wu Chen, Jiayi Zhu, Jiajia Liu ∗, Hongzhi Guo
Northwestern Polytechnical University, Xi‘an 710129, China

ARTICLE INFO ABSTRACT

Keywords: With the advances in artificial intelligence, robotics, and data fusion, large numbers of drones operating in a
Consensus coordinated manner will become commonplace for a wide range of commercial and military uses. At present,
Local interaction the application methods of drone swarms are mainly divided into fully autonomous methods and controlled
Fast coordination method
methods with human participation. Because of the limited level of artificial intelligence, controlled drone
Confrontational environment
swarms will be the main way for the application of large-scale drone swarms for a long time. However, there
is less research on achieving global coordination in a limited time for a controlled large-scale drone swarm.
Therefore, a new large-scale drone swarm framework is proposed firstly in this paper, which achieves global
coordination through local interaction and reduces the impact of limited channel resources. Secondly, this
paper proposes a local interaction-based fast coordination method and introduces a prediction mechanism,
to ensure that large-scale drone swarms can quickly achieve coordination even in the presence of node
loss. Moreover, the numerical integration method is used to update the consensus state, so that the drones
can increase the iteration period, reduce the number of packets, and further reduce the channel burden.
Finally, considering that large-scale drones swarm are usually composed of drone swarms launched at different
locations and times, a consensus algorithm considering the merging behavior of drone swarms is also proposed.
The simulation results show that the large-scale drone swarm using the proposed architecture can achieve
the leader–follower consensus in a very short time and even in a confrontational environment with poor
communication conditions. Besides, after the merger of multiple drone swarms, the consensus problem can
still be solved in very few iteration cycles.

1. Introduction such as environment monitoring, saturation attack, transport packages,

and traffic surveillance (Oubbati et al., 2022).
Drones, otherwise called autonomous unmanned aerial vehicles In a swarm, the drone acts as a simple agent and is conjoined in a
(UAVs), have received increasing demand from civilian, commercial, complex networking system, giving rise to collaborative or collective
and military applications in recent years. Especially in military appli- behaviors. To achieve this propriety, a drone swarm usually oper-
cations, drones become an indispensable part of the future battlefield ates in a centrally controlled or self-organized manner. In a centrally
(Orfanus et al., 2016). Due to the limitations of payload and energy of controlled drone swarm, a selected drone or a ground station makes
a single drone, the deployment of a drone swarm has been proposed for decisions and assigns instructions to the drones in the swarm. For
a diverse set of applications for higher mission requirements, which is example, the Crazyswarm are centrally controlled by a single computer
re-shaping the future applications of drones (Chmaj and Selvaraj, 2015; (Hönig et al., 2018). In a self-organized swarm, drones achieve task
Chen et al., 2020b). Drone swarms perform missions collaboratively goals and make collective decisions by sharing information (Coppola
and have the potential of fast deployment and wild coverage to go et al., 2020). Nägeli et al. (2014) and Basiri et al. (2014) used onboard
beyond the individual capabilities of single drones. Currently, the monocular cameras and onboard audio-based localization systems to
quantity of simultaneous drone missions is in the dozens or hundreds. achieve distributed information sharing for drones. Independent of
However, with the technological development of solar cells, battery control manners, drone swarms work in a three-step loop of perception,
capacity, and electric motors, the interest is shifting the size towards decision, and action.
thousands, or even ten thousand drones forming swarms to carry Perception means drones observe the situation by sensors and build
out collaborative missions. A large-scale drone swarm is becoming a a comprehensive image of the goal and environment. Timely and accu-
dominant practical choice for both civilian and military operations, rately sensed information is fundamental and crucial for accomplishing

∗ Corresponding author.
E-mail address: liujiajia@nwpu.edu.cn (J. Liu).

https://doi.org/10.1016/j.jnca.2023.103769
Received 16 March 2023; Received in revised form 11 August 2023; Accepted 17 October 2023
Available online 20 October 2023
1084-8045/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

the task. The sensed data is subsequently gathered by the central node The remainder of this paper is organized as follows. In Section 2,
in the centrally controlled swarm or exchanged by local interactions in we discuss several existing related works. The system architecture and
a self-organized manner. The decision means reasoning and it is made problem statement are introduced in Section 3, then it follows our
by a centralized node or by the drone itself with full or part of sensed proposed fast coordination method for large-scale drone swarm under
data. In the third step, drones take action following the decisions. The the confrontational environment in Section 4. Simulation results are
drone swarm will perceive the effectiveness of the action and step into presented to validate the performance of the proposed algorithms in
the next loop. Section 5. Finally, this paper is concluded in Section 6.
In the perception–decision–action loop, communication and net-
working are crucial because they enable information, including sensed 2. Related work
data and decision instructions, to be disseminated in the drone swarm.
As drone swarms grow in size, ensuring timely and effective com- A drone swarm achieves impressive global behaviors that surpass
munication becomes challenging due to the contention overhead, and the capabilities of individual drones by leveraging shared information
congestion. Consequently, a centralized control mode only guaran- among a group of drones. It is evident that efficient communication
tees timely instruction and sensed data distribution for small-scale and coordination are essential prerequisites for a drone swarm to
drone swarms. A few programmed large-scale drone swarms can only successfully execute collaborative tasks.
complete simple behaviors like a light show, e.g., Intel’s Shooting Stars. It is crucial for drones that function as a cohesive unit to share
In recent years, self-organized robot swarms involving drone swarms information, which places a high demand on communication. There-
that use local interaction to produce global behaviors hold great fore, communication becomes an essential aspect of a drone swarm (Cui
promise. Some proposals have demonstrated successful implementa- et al., 2017). However, drone swarms with characteristics of dynamic
tions of self-organized robot swarms up to 1000 units (Rubenstein topology usually operate on open wireless channels and work in con-
et al., 2014). Meanwhile, small-scale self-organized robot swarms have frontational environments, resulting in the rise of challenge to ensure
also received increasing scientific interest. For example, the SWARM- efficient communications. To address this challenge, a range of prob-
BOTS project presents a drone swarm with six tiny quadrotors au- lems, from networked communication to local exchanges, need to be
tonomously exploring an unknown environment and back to the de- addressed (Dorigo et al., 2013).
parture point (Brambilla et al., 2013). Networked communication technology has garnered significant at-
However, a complex maneuver of robot swarms still relies on the tention from researchers due to its reliance on mature network archi-
centralized control unit or external station (Berlinger et al., 2021). tectures. Much of the research in this area has focused on enhancing
Most studies in the field of fully decentralized methods have only the reliability and throughput of drone swarm networks, as well as im-
focused on theoretical work. In practical application, due to the low proving their performance under dynamic topology and confrontational
level of intelligence of drones, the impaction will be massive when the
circumstances (Oubbati et al., 2021). This is achieved by establishing
completely autonomous drone swarm has decision errors. Therefore,
and maintaining the connectivity referred to as the robust network.
some envisioned applications of large-scale drone swarms like search
Some solutions such as topology control and robust routing have been
and rescue, surveillance, and saturation attack still rely on the interven-
developed to address these challenges (Hou et al., 2021). Additionally,
tion of the centralized base station. However, as we mentioned above,
in recent years, there has been a growing emphasis on maintaining the
control instruction cannot be distributed timely in large-scale swarms.
connectivity of large-scale drone swarms by using maneuvering tactics
Therefore, it is still a great challenge to accomplish the task effectively
or adjusting speed, as highlighted in studies such as those by Zou et al.
under the control of a centralized-based station.
(2021), Tahir et al. (2019), Yanmaz et al. (2018), and Tegicho et al.
In this article, to address the problem that large-scale drone swarms
(2021).
cannot coordinate due to insufficient channel resources under con-
However, as the size of the drone swarm increases, the data traffic
trolled conditions, we proposed a semi-autonomous mode for a large-
within the swarm also grows exponentially. It poses a challenge caused
scale drone swarm. The main contributions can be summarized as
by the shortage of spectrum resources, making the communication of
follows:
the drone swarm vulnerable, especially in the face of jamming and
• A working architecture is proposed for a large-scale drone swarm congestion. Simply improving network reliability is not an effective
to achieve global consensus. Unlike traditional coordination meth- solution (Wang et al., 2022a). To address this issue, various spectrum-
ods, this architecture is based on local interaction and consensus, sharing mechanisms and bandwidth-efficient algorithms have been
prioritizing the perception–decision–action cycle over reliable proposed to mitigate the spectrum shortage (Zou et al., 2021; Onthoni
information transmission. The architecture can effectively address et al., 2023; Huang et al., 2022). Additionally, local exchanging meth-
the problem of lack of channel resources that arise from an ods aim to tackle the problem at its root by reducing multi-hop data
increasing number of drones. in the network (McCune and Madey, 2013). With the advancement
• A fast coordination method is proposed for large-scale drone of coordination technology, this approach is gradually becoming the
swarms in a confrontational environment with packet loss and primary mode of communication for drone swarms (Cui et al., 2017;
dynamic topology. The delay problem caused by local interaction Hussen et al., 2017).
is addressed by introducing a prediction mechanism. Compared When it comes to drone swarm coordination, there are two modes:
to existing works, its advantages lie in ensuring fast consensus centrally controlled coordination and self-organized coordination. The
among large-scale drone swarms even in the presence of node and centrally controlled coordination is also called direct-controlled or
link failures. leader–follower manner which means a special leader drone manipu-
• Considering the possibility of the Zeno phenomenon in existing lated by a pilot directly and orchestrates the collective behavior of the
event-triggered consensus methods, the Runge–Kutta method is rest of the swarm through the leader (Saffre et al., 2021). In a centrally
adopted to discretize the consensus algorithm for continuous controlled coordinated mode, coordination information is transmitted
systems. This approach increases the iteration period, reduces the hop by hop in the network. And reliable network communication is nec-
number of packets, and further lightens the channel burden. essary to ensure the drones reliably receive and execute the directives
• A robust and time-limited algorithm is proposed to ensure con- sent by the leader, which is the prerequisite for successful completion of
sensus during the merging process of large-scale multiple drone missions (Tahir et al., 2019; Tian et al., 2023). As mentioned above, due
swarms, which is a scenario that has been rarely addressed in to the shortage of spectrum resources, this kind of coordination manner
existing works. cannot support a large-scale drone swarm. By April 2022, The US Army

2
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

just tested a swarm of up to 30 drones over a desert in Utah (Saballa,

2022).
In comparison to centrally controlled coordination, self-organized
coordination, also known as indirect control or leaderless manner, is
preferred by large-scale drone swarms, in which coordination infor-
mation is shared through the local exchange. This coordination mode
does not require instructions to be reliably distributed to each drone
but mainly achieves coordination through information consensus. The
consensus means each drone updates the state based on the information
exchanged with neighbor drones, and finally achieves the consistency
of the state of whole drones (Carli et al., 2020). The characteristic of the
consensus based on local interactive information enables space-division
multiplexing, which can mitigate the shortage of spectrum resources.
The local interaction does not require routing, which is more suitable
Fig. 1. Drone swarm system architecture.
for drones with low computing power. Furthermore, consensus can
also be applied to the leader–follower model. Therefore, the consensus
has received great attention recently (Pasek and Kaniewski, 2022;
Doostmohammadian et al., 2021). converts the command signal into a physical parameter to accomplish a
In recent years, research on consensus for drone swarms can be task. The flight control block is responsible for maintaining the stability
categorized into the following groups: of flight attitude and path with or without receiving commands. The
Firstly, achieving consensus in finite time is crucial for large-scale responsibility of the transceiver block is to receive and to send packets.
drone swarms due to battery limitations and mission time constraints. Communication Layer: The communication layer is a crucial layer
Recent studies have focused on finite-time consensus, including works for a drone swarm because the prerequisite for drone swarms’ consen-
specifically addressing drone swarms (Rikos et al., 2022; Pal, 2022). sus is to communicate between the drones. The primary responsibility
However, these existing finite-time consensus is based on the Lyapunov of the communication layer is to ensure efficient data collection and
function, set-valued Lie derivative, or sliding mode (Pal, 2022; Liu exchange. The connectivity maintenance leveraging a robust and re-
et al., 2019). The algorithm is complicated and the convergence time liable infrastructure aims to ensure the connection between drones
relies on the initial state, leading unprocurable to be adopted by and the ground station. In this architecture, to reduce the number of
large-scale drone swarms. packets in the communication network, the communication layer only
Secondly, most mature consensus algorithms are designed for achieves local interaction and does not guarantee the topology of the
continuous-time systems and cannot be directly applied to drone entire swarm. Media access control block uses protocols to deal with the
swarms that rely on discrete communication. Although some event- problem of allocation and management of channel resources effectively.
triggered consensus algorithms have been proposed (Liu et al., 2022; Moreover, the communication layer implements networking by sharing
Wang et al., 2022b), the Zeno phenomenon may occur which leads to data.
the system cannot reach consensus. Coordination Layer: The coordination layer is served for reasoning
Finally, many tasks still require human operators to control fleets of and decision-making based on the observations obtained by the drone,
drone swarms. In such cases, leader–follower consensus algorithms are forming cooperative situation awareness to achieve team behavior
much more suitable. However, the large-scale swarm results in a pro- based on information sharing. The consensus block guarantees drones
longed convergence time because the state of the leader node is diffused in the swarm to share information with a consistent view of the states,
hop-by-hop. Therefore, most large-scale drone swarms still rely on pro- which is critical to the coordination task. Moreover, sensors carried by
grammed behaviors instead of inter-drone communication (Kallenborn, drones are different due to limited load capacity. The data fusion block
2018). collects the data from multiple sources to build more sophisticated and
In summary, the application of large-scale drone swarms still faces precise models of the environment, providing effective data to support
some intractable problems. An efficient coordination schedule for large- the correct decision-making. As the drone swarm may lose connection
scale drone swarms is urgently needed. with the ground station or work in autonomous mode, decentralized
decision-making based on the local-selected information is supported
3. System architecture and problem statement by the coordination layer.
Mission Layer: The mission layer is the principal subject for rea-
In this section, we present a system architecture for a large-scale soning, decision-making, and organizing drones. The mission decision
drone swarm and describe the special problem faced by the large-scale block allocates tasks and generates collaborative instruction to achieve
drone swarm under the proposed architecture. global behavior, which is an essential section to guarantee the success-
ful mission execution of a drone swarm. The path planning block is in
3.1. System architecture charge of searching for an appropriate path from the starting point to
the destination and provides a mechanism for collision avoidance. The
During the mission execution of the drone swarm, the information edge computing block provides extra computation and storage capacity
transfer process is divided into bottom-up and top-down. The bottom- when the resources of drones are not sufficient to fulfill the cooperation
up approach means the drone swarm needs to observe the environment, requirements.
share situation information and makes decisions based on consensus.
Decision instructions are delivered in a top-down approach, enabling 3.2. Problem statement
the drone to act effectively. The information transfer goes through the
perception–decision–action loop. For this reason, we proposed a system Fig. 1 presents a general structural model of a drone swarm in the
architecture for drone swarm as illustrated in Fig. 1. perception–decision–action loop. This work focuses on the cooperation
Platform Layer: The platform layer refers to the payload of soft- mission of drone swarm on a large scale. It only involves some func-
ware and hardware on the drone. It mainly includes a sensor block, tional blocks in Fig. 1. For the platform layer, we assume that the drone
an actuator block, a flight control block, and a transceiver block. The has a complete platform, by which the drone is able to normally fly,
sensor block aims at observing the environment. The actuator block perceive the environment, and transceive signals. In the communication

3
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

layer, it is assumed that large-scale drone swarms achieve cooperation the root node, and the drone achieves coordination with the leader by
through local information exchange, hence the proposed approach only coordinating with the father drone. If drones in the swarm satisfy the
involves connection maintenance and medium access control blocks. following formula:
The medium access block adopts the CSMA/CA protocol, and the
lim ‖ ‖
‖𝑥𝑖 (𝑡) − ⟨𝑥𝐹 (𝑡)⟩𝑖 ‖ = 0 (3)
connection maintenance block adopts a robust networking solution 𝑡→∞
proposed in the Chen et al. (2020b). The proposed approach achieves where ⟨𝑥𝐹 (𝑡)⟩𝑖 is the state of drone 𝑖’s father drone. Thus, Eq. (2) can
collaboration through consensus and does not concern data fusion and be considered to be held.
distributed decision-making modules in the cooperative layer. Finally, However, in order to achieve coordination between follower drones
this paper focuses on the challenging problem of using a large-scale and the leader by maintaining the spanning tree, some problems should
drone swarm to accomplish the task effectively under the control of a be solved:
centralized base station. Therefore, we can reasonably assume that the (a) How to achieve coordination in a confrontational environment with
mission planning is accomplished by the ground station, so this paper packet loss and dynamic topology: Packet loss is inevitable for large-
does not involve the mission layer. scale drone swarms working in confrontational environments. As we
Based on the architecture presented in Fig. 1, drones in a swarm mentioned above, to accomplish complex maneuvers, a large-scale
are generally homogeneous and interact with each other through the drone swarm still needs a leader controlled by an external commander.
communication topology. The topology can be represented by an undi- If the drone frequently fails to receive state information from the father
rected graph 𝐺 = (𝑉 , 𝐸), where 𝑉 is a set of drones and 𝐸 is a set of drone, the swarm cannot achieve global coordination. Moreover, due
[ ]
links between drones. Let 𝐴 = 𝑎𝑖𝑗 be the adjacency matrix associated to changes in the relative position between drones, the topology of
with 𝐺 which represents the connection of a drone swarm. The cell the drone swarm changes dynamically and make the spanning tree
𝑎𝑖𝑗 of the adjacent matrix equals 1 if drone 𝑖 and 𝑗 are connected, unstable. Towards this end, a semi-autonomous coordination mode is
otherwise, the cell equals 0. In the proposed large-scale drone swarm proposed in Section 4.1.
system architecture, communication between drones is based on local (b) How to achieve fast coordination based on local interaction: The
interaction. A drone and its neighbors are considered as a neighborhood leader drone in a large-scale drone swarm shares its state information
system, and the definition is given as follows. hop-by-hop along the spanning tree, and the information spreads from
the leader to the entire swarm like ripples, which fatally brings delay
Definition 1. Neighbor System: A neighborhood system for 𝐺 is and causes the large-scale drone swarm to fail to coordinately work.
defined as: With the increment of the drone swarm size, the delay becomes pro-
{ } nounced serious. In this paper, a prediction mechanism is introduced
𝑁 = 𝑁𝑖 |∀𝑖 ∈ 𝑉
in Section 4.2 to alleviate the delay caused by local interactions.
where 𝑁𝑖 is the set of drones neighboring drone 𝑖. Drone 𝑗 can directly (c) How to achieve coordination with local interaction of long interval:
interact with drone 𝑖 if ∀𝑗 ∈ 𝑁𝑖 . Drones in large-scale swarms update state information through local
interaction with interval communication. To reduce the payload in
A large-scale drone swarm with fully autonomous decision-making
drone swarms, a relatively long interval should be adopted. However,
only performs simple behaviors such as hovering. To complete com- Eq. (3) is analyzed based on a continuous-time system and cannot
plicated cooperative tasks, drones need to be controlled by receiving be directly applied in the drone swarm communication system based
external state information. Assume that drone 𝑖 maintains a state vari-
[ ]𝑇 on transceiving packets. Therefore, numerical integration methods are
able 𝑥𝑖 (𝑡) ∈ R, and set 𝑥 (𝑡) = 𝑥1 (𝑡) , 𝑥2 (𝑡) , … , 𝑥𝑛 (𝑡) ∈ R𝑛 represents adopted to discretize the coordination algorithm which is described in
the parameters related to the flight state such as speed, direction and so Section 4.3.
on. Global coordination will be achieved when all drones in the swarm (d) How to achieve coordination while multi-drone swarms merging:
converge to an external state. The definition is as follows: For large-scale drone swarms, it is infrequent to take off at the same
time due to the large number of drones. A common way of swarm
Definition 2. Coordination: If the state of each drone satisfies the formation is to combine multiple small-scale drone swarms into a large-
following formula: scale swarm, hence the coordination problem while merging multiple
lim ‖ ‖
‖𝑥𝑖 (𝑡) − 𝑠 (𝑡)‖ = 0 (1) drone swarms is also worth studying. In this paper, the drone swarm
𝑡→∞ merging algorithm is presented in which provides a solution to this
it can be viewed as a drone swarm reaching coordination with the problem.
external state 𝑠 (𝑡), where 𝑠 (𝑡) represents external state information.
4. Fast coordination approach for large-scale drone swarm
Due to the bandwidth constraints, it is hard to directly control each
drone in the large-scale swarm. A common method is to adopt a leader–
In this section, we propose a fast coordination approach for large-
follower model, where the leader is directly controlled by external state
scale drone swarms. To deal with the packet loss and the dynamic
information, and the followers achieve coordination by reaching state
topology in a confrontational environment, a semi-autonomous coordi-
convergence with the leader. The leader drone’s state information 𝑥𝐿 (𝑡)
nation mode for large-scale drone swarm is introduced in Section 4.1.
is described as:
In Section 4.2, a prediction mechanism is presented to ensure large-
𝑥𝐿 (𝑡) = 𝑠 (𝑡) scale drone swarm achieves coordination in a limited time. A numerical
integration method is presented in Section 4.3 which provides a solu-
The swarm achieves global coordination when the drones in the tion to achieve coordination with local interaction of long intervals.
swarm coordinate with the state of the leader. In such case, Eq. (1) Aiming at the multi-swarm merging coordination problem, a drone
can be rewritten as: swarm merging algorithm is adopted in Section 4.4.
lim ‖ ‖
‖𝑥𝑖 (𝑡) − 𝑥𝐿 (𝑡)‖ = 0 (2)
𝑡→∞ 4.1. Coordination method in confrontation environment
To satisfy Eq. (2), it is required to guarantee 𝑎𝑖𝐿 ≡ 1, 𝑖 = 1, … , 𝑛,
which means each drone in the swarm interacts with the leader di- In this sub-section, to solve the problems of packets losing and
rectly. In a large-scale drone swarm based on local interaction, a stable to adapt the dynamic topology in a confrontational environment, we
connection between every drone in the swarm and the leader may be propose a semi-autonomous coordination mode for large-scale drone
failed. Therefore, we maintain a spanning tree with the leader drone as swarm coordination.

4
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

As mentioned above, a large-scale drone swarm performs a complex by switching between controlled coordination and local automation
task only if under the control of an external command through a chosen coordination.
leader drone. It indicates that the state of drones in the swarm must be To simplify the calculation, Eq. (8) can be replaced by an equivalent
converged with the leader drone, as shown in Eq. (2). Considering that equation:
drones in the swarm cannot ensure connection to the leader directly, a [( )2 ( )2 ]
spanning tree with the leader as root is established, and every drone lim 𝑥𝑖 − ⟨𝑥𝑇 ⟩𝑖 + 𝑥𝑖 − 𝑥𝑗 = 0, ∀𝑖 ∈ 𝑉 , ∀𝑗 ∈ 𝑁𝑖 (9)
𝑡→∞
achieves coordination with the leader through coordinating with its We also modify the Eq. (7) as follow.
father drone as shown in Eq. (3). ∑( )2 ∑ ( )2
However, the drone may lose connection with its father node under 𝐸 (𝑥) = 𝑥𝑖 − ⟨𝑥𝑇 ⟩𝑖 + 𝑥𝑖 − 𝑥𝑗 (10)
a confrontation environment, as a result, Eq. (3) cannot hold all the 𝑖∈𝑉 𝑖∈𝑉 ,𝑗∈𝑁𝑖

time. However, it is rare for a node to lose connections to all its Based on the above two equations, the large-scale drone swarm
neighbors. Inspired by bird flocking, in which every bird could always coordination is equivalent to the minimization of energy 𝐸.
maintain consistency with its neighbors not only depending on its We still face the problem of dynamic topology. For a large-scale
father node, drones can also achieve coordination through interaction drone swarm working in a confrontational environment, the relative
with its neighbors. From this point, Eq. (3) can be modified as follow. positions of drones change frequently, and the spanning tree with the
lim (‖ ‖ ‖𝑥𝑖 − 𝑥𝑗 ‖ leader as the root node cannot be maintained stably.
𝑡→∞ ‖𝑥𝑖 − ⟨𝑥𝐹 ⟩𝑖 ‖ + ‖
‖
‖) = 0, ∀𝑖 ∈ 𝑉 , ∀𝑗 ∈ 𝑁𝑖
‖
(4)
According to Moreau (2005), it is not necessary to ensure that each
There still is a problem of how to deal with the item of ⟨𝑥𝐹 (𝑡)⟩𝑖 when node is connected for the coordination of dynamic networks, but only
a drone lost connection with its father drone. To solve the problem, to ensure that the network jointly stays connected, which means that
we mimic a flock of birds by changing from leader–follower mode the union of network topology graphs is connected in a time period.
to leaderless mode when a drone temporarily loses its father drone. Therefore, a spanning tree exists all the time by maintaining a dynamic
According to the previous research (Chen et al., 2020a), a leaderless topology tree with the leader as the root node, thereby ensuring that
drone swarm coordination can be expressed as the following equation: the state of the drone swarm is coordinated with the leader.
( ) We adopt the dynamic spanning-tree algorithm which draws on
‖ ‖
lim ‖𝑥 − ⟨𝑥𝑅 ⟩𝑖 ‖ ‖+‖ 𝑥 − 𝑥𝑗 ‖ = 0, ∀𝑖 ∈ 𝑉 , ∀𝑗 ∈ 𝑁𝑖 (5)
𝑡→∞ ‖ 𝑖 ‖ 𝑖 ‖ the reverse path broadcasting technique (Chen et al., 2020a). Drones
where ⟨𝑥𝑅 ⟩𝑖 is the virtual reference state which means the averaged without father drones achieve local consensus through Eq. (8) and wait
effect on drone 𝑖 by all its neighbors. It can be calculated by mean-field: to join the spanning tree before the spanning tree is constructed for
∑ ( ) the first time. Once the spanning tree is constructed, it always exists
⟨𝑥𝑅 ⟩𝑖 = 𝑥𝑗 𝑃 𝑥𝑗 (6) and transmits the state information of the leader through the dynamic
𝑗∈𝑁𝑖 spanning tree algorithm.
( )
here, 𝑃 𝑥𝑗 represents the probability for state information of drone 𝑗
and it obeys Gibbs distribution (Li, 2009): 4.2. Coordination state prediction mechanism
− 𝑇1 𝐸(𝑥)
𝑃 (𝑥) = 𝑍 −1 × 𝑒 For the large-scale drone swarm based on local interaction, the
∑ − 𝑇1 𝐸 (𝑥𝑖 ) state information of the leader is shared by a dynamic spanning tree
where the normalized parameter is 𝑍 = 𝑖∈𝑉 𝑒 and 𝑇 is a free
algorithm with a delay problem inevitably. As the size of the swarm
parameter which is assumed to be 1 generally. 𝐸 (𝑥) is energy function:
increases, the height of the spanning tree rises, and it takes a long time
∑( )2 ∑ ( )2
𝐸 (𝑥) = 𝑥𝑖 − ⟨𝑥𝑅 ⟩𝑖 + 𝑥𝑖 − 𝑥𝑗 (7) for state information delivery from the root to the leaf node. In this
𝑖∈𝑉 𝑖∈𝑉 ,𝑗∈𝑁𝑖 manner, it is hard to achieve coordination with the leader of the drone
The first term of the Eq. (7) means the overall energy of the drones swarm.
themselves, which represents the external field acting on the drone, and According to Eq. (8), the drones in the swarm coordinate with
the second term represents the interactions between adjacent drones. the leader by converging with the target state ⟨𝑥𝑇 (𝑡)⟩𝑖 . However,
The different term of ‖ ‖
‖𝑥𝑖 − ⟨𝑥𝑅 ⟩𝑖 ‖ compared to Eq. (4) means that
drones can only interact with adjacent drones under the system work
drones are capable to converge to the virtual state according to the architecture proposed. At the same time, drones send and update state
average interaction of their neighbors with no requirement to receive information in parallel, so the target state ⟨𝑥𝑇 (𝑡)⟩𝑖 of drone 𝑖 at 𝑡 is
the state information from the father drone. Therefore, Eq. (5) can be actually the state of its father drone 𝑗 at 𝑡 − 1, which means ⟨𝑥𝑇 (𝑡)⟩𝑖 =
utilized to solve the case where ⟨𝑥𝐹 (𝑡)⟩𝑖 is missing in Eq. (4). 𝑥𝑗 (𝑡 − 1). Obviously, there is a delay error, and such error gets worse
A target state variable ⟨𝑥𝑇 ⟩𝑖 is proposed to unify Eqs. (4) and (5), it when the drone gets further away from the leader.
is defined as: To deal with the delay caused by local interaction, a prediction
{ mechanism is introduced. The main idea is to use the cubic spline
⟨𝑥𝐹 ⟩𝑖 , drone 𝑖 can connect with the father node
⟨𝑥𝑇 ⟩𝑖 = interpolation algorithm to perform curve fitting on the state change of
⟨𝑥𝑅 ⟩𝑖 , otherwise
the father drone and predict the state of the parent drone at 𝑡 with
by using the target state, a semi-autonomous coordination model is the fitted curve. The target state ⟨𝑥𝑇 (𝑡)⟩𝑖 is replaced by the predicted
obtained as follow. information to reduce the impact of delay and enables drone swarms
( )
‖ ‖ to achieve coordination rapidly.
lim ‖𝑥 (𝑡) − ⟨𝑥𝑇 (𝑡)⟩𝑖 ‖
‖+‖𝑥 (𝑡) − 𝑥𝑗 (𝑡)‖ = 0 (8)
𝑡→∞ ‖ 𝑖 ‖ 𝑖 ‖ To keep the flight stability of the drone, the change of the drone
The main idea of Eq. (8) is that the drone is controlled by the leader state usually has inertia, so the state of the father drone can be
drone when it can receive messages from its father drone. Otherwise, effectively predicted. As a kind of interpolation algorithm, the cubic
it holds coordination with its neighbor automatically. That means the spline interpolation algorithm uses multiple low-degree polynomials to
drone 𝑖 achieves coordination with the leader by converging with the fit a smooth state change curve in sections with the characteristics of
target state ⟨𝑥𝑇 ⟩𝑖 . When the father drone has no packet loss, the target a simple formula, fast calculation speed, and good stability. Hence it is
state is the state of the father drone; when packet loss occurs, the target suitable for drones with limited CPU power and energy.
state is generated by the average action of neighboring drones, which In general, the cubic spline interpolation algorithm uses four sam-
tends to local leaderless consensus. pling values to ensure the accuracy of the algorithm. Therefore, a
By using Eq. (8), the packet loss problem between the drone and its state information queue is maintained for each drone, which is used to
father drone in the confrontational environment is solved effectively record four historical target state information 𝑥𝑡−4 , 𝑥𝑡−3 , 𝑥𝑡−2 and 𝑥𝑡−1 .

5
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

Assuming that 𝜏 is any time in the interval [𝑡 − 4, 𝑡 − 1], three endpoint Algorithm 1 The semi-autonomous coordination algorithm with
continuous state functions 𝑆𝑡−𝑖 (𝜏) can be determined by 𝑥𝑡−𝑖 and 𝑥𝑡−𝑖+1 , prediction
𝑖 = 2, 3, 4. The state function 𝑆𝑡−𝑖 (𝜏) can be defined as: Input: The leader in the drone swarm.
( ) ( )2 Output: The state of each drone in the swarm.
𝑆𝑡−𝑖 (𝜏) = 𝑎𝑡−𝑖 + 𝑏𝑡−𝑖 𝜏 − 𝜏𝑡−𝑖 + 𝑐𝑡−𝑖 𝜏 − 𝜏𝑡−𝑖
( )3 (11) 1: Initialization: set the leader as root drone and its layer number
+𝑑𝑡−𝑖 𝜏 − 𝜏𝑡−𝑖
equals to 0; set layer number is int _ max for other drones and set
Suppose the sampling step is ℎ, then 𝜏𝑡 = 𝜏𝑡−2 + 2ℎ. The predicted itself as its father drone.
state information ⟨𝑥𝑃 (𝑡)⟩𝑖 can be obtained by the third state function 2: for Each drone 𝑖 and every refresh cycle 𝑘 in parallel do
( )
𝑆𝑡−2 𝜏𝑡−2 + 2ℎ as follow. 3: if Received state from neighbor then
2
⟨𝑥𝑃 (𝑡)⟩𝑖 = 𝑎𝑡−2 + 2𝑏𝑡−2 ℎ + 4𝑐𝑡−2 ℎ + 8𝑑𝑡−2 ℎ 3
(12) 4: if the neighbor is root drone then
5: Set the tier of drone 𝑖 equals to 1.
According to the continuous endpoints of cubic spline interpolation, 6: Set the father drone of drone 𝑖 as root node.
𝑎𝑡−2 , 𝑏𝑡−2 , 𝑐𝑡−2 and 𝑑𝑡−2 can be calculated by the following formulas:
7: else
𝑎𝑡−2 = 𝑥𝑡−2 8: if the neighbor’s layer < the drone 𝑖’s layer then
𝑥𝑡−1 −𝑥𝑡−2 ( ) 9: Set the layer of drone 𝑖 equals to the neighbor’s layer + 1.
𝑏𝑡−2 = ℎ
− ℎ2 𝑚𝑡−2 − ℎ
6
𝑚𝑡−1 − 𝑚𝑡−2
(13)
𝑚𝑡−2 10: Set the father drone of drone 𝑖 as the neighbor.
𝑐𝑡−2 = 2 11: end if
𝑑𝑡−2 =
𝑚𝑡−1 −𝑚𝑡−2 12: end if
6ℎ
13: if Refresh cycle timeout then
where 𝑚𝑡−2 represents the second-order derivative of the state informa-
14: if Received consensus state from father drone then
tion at time 𝑡 − 2, which can be solved by the following determinant
15: Set ⟨𝑥𝑇 ⟩𝑖 = ⟨𝑥𝐹 ⟩𝑖 .
according to the boundary condition (Hussain et al., 2015):
16: else
⎡−ℎ 2ℎ −ℎ 0 ⎤ ⎡𝑚𝑡−4 ⎤ ⎡ 0 ⎤ ∑ ( )
⎢ ⎥⎢ ⎥ ⎢ 𝑥 −𝑥 𝑥𝑡−3 −𝑥𝑡−4 ⎥ 17: Set ⟨𝑥𝑅 ⟩𝑖 = 𝑥𝑗 𝑃 𝑥𝑗 .
⎢ℎ 4ℎ ℎ 0 ⎥ ⎢𝑚𝑡−3 ⎥ ⎢ 𝑡−2 ℎ 𝑡−3 − ℎ ⎥ 𝑗∈𝑁𝑖
= 𝑥 −𝑥 𝑥𝑡−2 −𝑥𝑡−3 (14) Set ⟨𝑥𝑇 ⟩𝑖 = ⟨𝑥𝑅 ⟩𝑖 .
⎢0 ℎ 4ℎ ℎ⎥ ⎢𝑚𝑡−2 ⎥ ⎢ 𝑡−1 ℎ 𝑡−2 − ℎ
⎥ 18:
⎢0 −ℎ 2ℎ ⎥ ⎢ ⎥
ℎ⎦ ⎣𝑚𝑡−1 ⎦ ⎢⎣ ⎥ 19: end if
⎣ 0 ⎦
20: Push ⟨𝑥𝑇 ⟩𝑖 into the state information queue.
The predicted state information ⟨𝑥𝑃 (𝑡)⟩𝑖 can be obtained based on
Eq. (12), then Eq. (8) can be rewritten as: 21: Calculate ⟨𝑥𝑃 (𝑡)⟩𝑖 by (12))∼(14).
[( )2 ( )2 ] 22: end if
lim 𝑥𝑖 (𝑡) − ⟨𝑥𝑃 (𝑡)⟩𝑖 + 𝑥𝑖 (𝑡) − 𝑥𝑗 (𝑡) = 0, 23: Update coordination state by using Eq. (15).
𝑡→∞ (15)
∀𝑖 ∈ 𝑉 , ∀𝑗 ∈ 𝑁𝑖 24: end if
25: Broadcast coordination state to its one-hop neighbors.
The target state information is determined after the drones interact
with their neighbor drones. The predicted state information obtained 26: end for
from Eq. (12) is used to replace the delayed target state information to
reduce the influence caused by the local interaction. After introducing
the prediction mechanism, the complete fast coordination algorithm for in the swarm have the same frequency of sending packets which are
large-scale drone swarm in a confrontation environment, we call it a represented by iteration 𝑘. Based on this, Eq. (15) can be rewritten as:
semi-autonomous coordination algorithm with prediction, is shown in [( )2 ( )2 ]
Algorithm 1. lim 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖 + 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘) = 0,
𝑘→∞ (16)
According to the description provided in Algorithm 1, it is assumed ∀𝑖 ∈ 𝑉 , ∀𝑗 ∈ 𝑁𝑖
that the drone swarm consists of 𝑛 nodes. Each drone in the swarm per-
To solve the coordination model of Eq. (16), the method of par-
forms parallel computation of the coordination state at a fixed interval,
allel energy minimization is adopted. The energy function of state
resulting in the outer loop iterating 𝑛 times as it visits each node once.
information in the drone swarm is rewritten as:
Within each iteration of the outer loop, ⟨𝑥𝑅 ⟩𝑖 is calculated by iterating { }2
over the adjacent nodes of the current node. Although the number of ∑
𝐸 [𝑥 (𝑘)] = 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖
adjacent nodes for each node can vary, it is assumed that on average, 𝑖∈𝑉 (17)
each node has 𝑚 adjacent nodes. Therefore, the inner loop will iterate 𝑚 ∑ [ ]2
times for each node. Overall, the total number of iterations of the inner + 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘)
𝑖∈𝑉 ,𝑗∈𝑁𝑖
loop can be approximated as 𝑛×𝑚. Since calculating ⟨𝑥𝑅 ⟩𝑖 performs a
constant number of operations for each iteration, the computational According to the Ising model, the drone swarm reaches stable
complexity of the inner loop is 𝑂 (1). Therefore, the computational coordination and the gradient of the energy function becomes a zero
complexity of the entire algorithm can be expressed as 𝑂 (𝑛 × 𝑚). As 𝑚 vector, which means the following equation should be satisfied when
is assumed to be less than 𝑛, the maximum complexity of this algorithm the energy function is minimized:
( ) {
is 𝑂 𝑛2 . 𝜕𝐸 [𝑥 (𝑘)] ∑[ ]
=2× 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖
𝜕𝑥𝑖 (𝑘) 𝑖∈𝑉
4.3. Discretization of coordination algorithm } (18)
∑ ∑[ ]
+ 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘) =0
The large-scale drone swarm mentioned in this paper is based 𝑖∈𝑉 𝑗∈𝑁𝑖

on local interaction. The local interaction is realized by transceiving To solve Eq. (18), it is required to sum the state information of each
packets and has a communication interval. We assume that the drones drone in the swarm, which will increase the calculation burden and the

6
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

network load. If Eq. (18) is rewritten as: In the drone swarm based on local interaction, 𝜇 is the communica-
{ { tion interval between drones which cannot be ignored. Therefore, the
𝜕𝐸 [𝑥 (𝑘)] ∑ [ ]
= 2× 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖 drone swarm system is a nonlinear discrete system that needs numerical
𝜕𝑥𝑖 (𝑘) 𝑖∈𝑉
}} integration methods to solve Eq. (22).
(19)
∑[ ] The Runge–Kutta integral method is an iterative numerical inte-
+ 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘)
gration algorithm. The function is discretized by increasing the order
𝑗∈𝑁𝑖
of the iterative formula. The prior value is applied to calculate the
and considering each drone in the swarm individually, the following new value and subsequently the average value can be obtained which
formula is obtained: continuously approximates the true value of the variable. With the
[ ] {
𝜕𝐸 𝑥𝑖 (𝑘) [ ] higher order, the accuracy of the algorithm is improved while the
= 2× 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖
𝜕𝑥𝑖 (𝑘) computation load increases. Generally, the fourth-order Runge–Kutta
} (20) method is suitable in the engineering field.
∑[ ]
+ 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘) According Eq. (22), the Runge–Kutta function is defined as:
𝑗∈𝑁𝑖 { }
[ ] [ ] ∑[ ]
Eq. (19) then can be described as: 𝑓 𝑘, 𝑥𝑖 (𝑘) = 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖 + 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘) (28)
[ ] 𝑗∈𝑁𝑖
𝜕𝐸 [𝑥 (𝑘)] ∑ 𝜕𝐸 𝑥𝑖 (𝑘)
= (21) The segmented slopes are calculated as follows
𝜕𝑥𝑖 (𝑘) 𝜕𝑥𝑖 (𝑘)
𝑖∈𝑉 ( )
𝐾1 = 𝑓 𝑘, 𝑥𝑖 (𝑘)
The above equation shows that the global energy function could
be minimized through the energy function for each drone to obtain ( )
the minimum. According to Eq. (18), the energy function is a strictly 𝐾2 = 𝑓 𝑘 + 𝜇2 , 𝑥𝑖 (𝑘) + 𝜇2 𝐾1
convex function in its domain, which has only one global minimum ( ) (29)
point (Hindi, 2004). The drone swarm is capable to achieve coordina- 𝐾3 = 𝑓 𝑘 + 𝜇2 , 𝑥𝑖 (𝑘) + 𝜇2 𝐾2
tion at the global minimum point. In actual applications, the following
equation is used to solve the energy minimization function for each ( )
𝐾4 = 𝑓 𝑘 + 𝜇, 𝑥𝑖 (𝑘) + 𝜇𝐾3
drone in the swarm:
{ where 𝐾1 is the slope at the beginning of the sampling interval 𝑘. 𝐾2 is
𝑥𝑖 (𝑘 + 1) − 𝑥𝑖 (𝑘) [ ]
= −2× 𝑥𝑖 (𝑘) − ⟨𝑥𝑃 (𝑘)⟩𝑖 the slope at the midpoint of the sampling interval, and the Euler method
𝜇
} (22) which is an alternative numerical algorithm is adopted to determine the
∑[ ] value of 𝑥𝑖 (𝑘) at point 𝑘 + 𝜇2 by using the slope 𝐾1 . 𝐾3 is also the slope
+ 𝑥𝑖 (𝑘) − 𝑥𝑗 (𝑘)
of the midpoint while the slope 𝐾2 is used to determine the 𝑥𝑖 (𝑘) value.
𝑗∈𝑁𝑖
𝐾4 is the slope at the end of the sampling interval 𝑘, and its 𝑥𝑖 (𝑘) value
where 𝜇 is a small constant depending on the experimental accuracy is determined by 𝐾3 .
and numerical integral step size. As 𝜇 approaches to 0, Eq. (22) is
The final slope is obtained by the weighted average of the four
equivalent to:
[ ] segmented slopes, which can be used to calculate the value of the state
( ) ∑( ) information at 𝑘 + 1 with higher accuracy. The calculation equation is
𝑥̇ 𝑖 = −1 × 𝑥𝑖 − ⟨𝑥𝑃 ⟩𝑖 + 𝑥𝑖 − 𝑥𝑗 (23) as follows:
𝑗∈𝑁𝑖
(𝐾1 + 2𝐾2 + 2𝐾3 + 𝐾4 )
Assuming that 𝑥𝑃 can accurately predict 𝑥𝑇 . Eq. (23) can be rewrit- 𝑥𝑖 (𝑘 + 1) = 𝑥𝑖 (𝑘) + 𝜇 (30)
6
ten as:
[ ] Through the fourth-order Runge–Kutta method, the coordination
( ) ∑( )
𝑥̇ 𝑖 = −1 × 𝑥𝑖 − ⟨𝑥𝑇 ⟩𝑖 + 𝑥𝑖 − 𝑥𝑗 (24) state energy function of the drone swarm based on local interaction
𝑗∈𝑁𝑖 is solved.
Substituting Eq. (6) into Eq. (24), the above equation is rewritten
as: 4.4. Multi-drone swarms merging coordination
[( ) ]
∑ ∑( )
𝑥̇ 𝑖 = −1 × 𝑥𝑖 − 𝑥𝑗 𝑃 (𝑥𝑗 ) + 𝑥𝑖 − 𝑥𝑗 (25) In practical applications, it is not realistic for large-scale drone
𝑗∈𝑁𝑖 𝑗∈𝑁𝑖
swarms to take off simultaneously in a certain place. To improve the
Assuming a drone swarm with 𝑛 nodes is represented by the ad- achievement ratio of the missions like area reconnaissance or search
jacency matrix 𝐴 = [𝑎𝑖𝑗 ] ∈ 𝑅𝑛×𝑛 . In the context of drone swarms, and precision strikes on a target, multi-drone swarms usually take off
it is typical for the drones to communicate through a bidirectional and reach coordination separately. Drone swarms fly to the mission
link, which enables us to consider the drone swarm as an undirected area from different directions and merge into a large-scale drone swarm
graph. If nodes 𝑖 and 𝑗 are adjacent neighbors, the element 𝑎𝑖𝑗 set to 1. over the mission area to start the mission. Different from the traditional
Otherwise, 𝑎𝑖𝑗 is 0. We can further transform Eq. (25) by introducing flight formation control methods, a self-organizing algorithm to achieve
the adjacency matrix 𝐴: merging coordination is proposed for system architecture, in which the
𝑥̇ = −𝐿𝑥 (26) drones do not need to receive instructions from the leader or the ground
station to complete the merging behavior. To minimize the rendezvous
where time and energy consumption of the drone swarm, each drone makes its
∑ ( ( ))
𝐿𝑖𝑖 = 1 + 𝑎𝑖𝑗 = 1 + 𝐷𝑖 , 𝐿𝑖𝑗 = −𝑎𝑖𝑗 1 + 𝑃 𝑥𝑗 own decision to change the working model and choose the new father
(27)
𝑗∈𝑛 drone when the state information from other swarms is perceived. The
where 𝐷𝑖 is the degree of node 𝑖. It is clear that the matrix 𝐿 is multi-swarm merging method optimizes the energy consumption in
diagonally dominant, with all its diagonal elements being positive. flight, and the computing time does not increase significantly with the
As a result, the real parts of 𝐿’s eigenvalues are also positive, which enlargement of the drones.
demonstrates the convergence of the system, as pointed out in Ren and The leader drones in each swarm are directly controlled by the
Beard (2007). ground station to ensure each leader maintains a priority queue that

7
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

Fig. 2. The key drones losing leads topology changes.

contains all leaders’ state information. A drone swarm with a higher-

priority leader is able to accept a lower-priority leader with its fol-
lowers. If the leader in a swarm becomes a follower of one of the
other swarms, the rest of the drones in this swarm subsequently change
their working model to leaderless and reach local consensus with
neighboring drones until receiving a new leader’s state information.
Note that if a leader 𝑖 receives a message from a non-leader drone 𝑗 in
another high-priority drone swarm, the leader 𝑖 will become a follower
of drone 𝑗. It ensures that the merging algorithm can be triggered when
the edges of two drone swarms interact with each other.
The details of the complete multi-drone swarms merging algorithm
are shown in Algorithm 2. Similar to Algorithm 1, the multi-drone
( )
swarms merging algorithm also achieves a time complexity of 𝑂 𝑛2
by performing parallel computation of the drone’s state.
Fig. 3. Merging simulation scenario of multiple drone swarms.
5. Performance evaluation
Table 1
To validate the effectiveness of the fast coordination approach for Simulation parameters of Qualnet.
large-scale drone swarms, simulations are implemented using QualNet. Parameter Value
The effectiveness is verified by applying the proposed approach to Map size (km × km) 3.0 × 3.0
scenarios with different numbers of drone swarms. The comparison Number of drones 50/100/200/300
results of proposed schemes with other traditional methods are pre- Formation Grid
Simulation time (s) 300
sented. Moreover, some highly realistic simulations are conducted for Antenna model Omnidirectional
the scenarios where key drones are lost and multi-drone swarms merge. Physical layer 802.11b radio
Data rate (Kbps) 500
5.1. Simulation settings Transmission power (dbm) 15
Pathloss model Two ray
Shadowing model Constant
In most hypothetical application scenarios, drone swarms fly to the Fading model Rayleigh
target area in fixed formations and collaborate on tasks. Without loss of Noise factor (db) 10
generality, a fixed formation is devised in QualNet, and a space area of MAC layer 802.11
3000 m × 3000 m is set as the congregation area for the drone swarm.
Considering the need for achieving consensus within a few minutes,
the simulation time is specified as 300 s. Various communication lost, the drone swarm changes the topology from Figs. 2(a) to 2(b). As
technologies are available for drone networks, such as IEEE 802.11, the number of losing drones increases, the drone swarm may split into
IEEE 802.15.4, LTE, etc. Among these technologies, IEEE 802.11b offers multiple new sub-swarms or a single drone as Fig. 2(c). We simulate the
a relatively moderate transmission rate and consumption. It has been scenario of drone loss to verify that the proposed system architecture
adopted by some actual drones like AR Drone 2.0. Therefore, the is robust to the confrontational environment (i.e., some drones may be
physical layer and the MAC protocol of each drone are chosen as attacked and lose their communication abilities).
802.11b. For further details of the parameters, please refer to Table 1. Finally, the scenario of merging multiple drone swarms is simulated
Firstly, we deployed 50, 100, 200, and 300 drones to mimic the in Fig. 3. At the initial configuration, four drone swarms with random
swarm scenario with different scales of drones. The performance of the formations are devised. We preset swarms merging instructions and
proposed approach for tracking state information of the leader drone observe the convergence of consensus information during the merg-
is evaluated. The simulation results are compared with the average ing process. 4 × 25 and 4 × 50 drones are deployed to verify the
consensus method, which is a widely-used consensus method (Ren and effectiveness of the drone swarm merging algorithm.
Cao, 2010). Moreover, we compare the results with the traditional way
of transceiving commands from the leader. 5.2. Numerical results
Secondly, we deploy 100 drones to mimic the leader–follower
swarm scenario to compare the performance with and without the pre- We evaluate the performance of the proposed semi-autonomous
diction mechanism. The performance with and without the four-order coordination algorithm in tracking the dynamic target state information
Runge–Kutta method is also evaluated. from the leader for different scale drone swarms and the results are
Thirdly, the topology of the drone swarm in a confrontational shown in Fig. 4. As can be seen, the state of the large-scale drone
environment is changed as shown in Fig. 2. With the drone gradually swarm can quickly reach coordination with the leader due to the delay

8
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

Fig. 4. Dynamic target state coordination convergence process of fast coordination method for different scale drone swarm.

Fig. 5. Dynamic target state coordination convergence process of average consensus for different scale drone swarm.

Fig. 6. Dynamic target state coordination convergence process of transceiving instructions method for different scale drone swarm.

compensation based on the prediction mechanism, and the convergence be able to maintain consensus with the leader when the formation size
time is not affected by the size of the swarm, which is kept within 20 reaches 200. As the formation size increases to 300, a large number
iterations. With the growth of the scale, the convergence time is still of drones cannot receive the leader’s consensus information in time,
kept within 40 iterations. therefore, the entire drone swarm becomes very chaotic.
Fig. 5 shows the convergence process with average consensus for Due to the high fidelity of the experimental simulations, the exper-
different scale drone swarms. In Fig. 5(a), a drone swarm of 10 drones imental results exhibit small fluctuations that are within an acceptable
can achieve coordination by using the average consensus method. With range considering the noise and packet loss in communication among
the growth of the scale, the swarm state can eventually converge with drones. By comparing Figs. 4, 5, and 6, it is evident that as the scale
an obvious system delay error as shown in Fig. 5(b). In Figs. 5(c) and of the drone swarm increases, the superiority of the proposed method
5(d), the delay error becomes significantly larger when the number of becomes more significant, ensuring rapid coordination state conver-
drone swarms reaches 100, and the system cannot converge when the gence in large-scale drone swarms. When comparing drone swarms of
number reaches 200. the same scale using different coordination methods, the fast coordina-
The process of tracking the dynamic target state of the drone tion method also demonstrates significant advantages in terms of fast
swarm using traditional transceiving instructions is shown in Fig. 6, convergence speed and effective coordination.
which is based on multi-hop forwarding communication and instruction As shown in Figs. 7(a) and 7(b), an unprocessed coordination model
distribution. The leader’s movement is controlled by the ground station, leads to significant errors, and the system cannot converge. In contrast,
while other drones receive instructions and follow the leader. Fig. 6(a) using the Runge–Kutta method reduces the errors, and the state vari-
shows that when the number of drones is 50, the drone swarm can ations of the drones in the swarm tend to be consistent, although a
eventually converge but the time delay is also highly obvious. Because notable delay can be observed compared to the leader.
a large number of drones will compete for limited channel resources The proposed prediction mechanism offers a reliable and efficient
when performing multi-hop forwarding communication, resulting in way to predict the state of the father node, thereby greatly enhanc-
serious access delays. The delay becomes serious as the number of ing the overall performance of the consensus algorithm and reducing
drones increases. Some drones cannot receive information from the tracking errors. These advantages are clearly demonstrated through
leader, so the initial value is maintained as a straight line in Fig. 6(b). simulation results. As shown in Fig. 7(b) and 7(c), the prediction
As shown in Figs. 6(c) and 6(d), it is obvious that more drones cannot mechanism brings a significant improvement in coordination speed.

9
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

Fig. 7. The coordination state change under different processing methods.

Fig. 8. The changing of state information with key drones losing.

Algorithm 2 The multi-drone swarms merging algorithm

Input: The leader priority queue.
1: Initialization: The merging command is sent to each leader from the
ground station.
2: for Each drone 𝑖 and refresh cycle 𝑘 in parallel do
3: if Received state from neighbor drone 𝑗 then
4: if Drone 𝑖 is LEADER then
5: if The leader of drone 𝑗 > drone 𝑖 in priority queue then
6: Set drone 𝑖 as FOLLOWER.
7: Set drone 𝑗 as the father drone of 𝑖.
8: Set ⟨𝑥𝑇 ⟩𝑖 = ⟨𝑥𝐹 ⟩𝑖 . Fig. 9. The changing of state information in the process of merging four drone swarms.

9: end if
10: end if
11: if Drone 𝑖 is FOLLOWER then evidence that the coordination state prediction mechanism plays a cru-
12: if The leader of drone 𝑗 > the leader of drone 𝑖 in priority cial role in facilitating rapid coordination of large-scale drone swarms.
queue then Moreover, we simulated the scenario where the key drones in the
13: Set drone 𝑖 as LEADERLESS. swarm were lost in a confrontational environment. The simulation
14: Set ⟨𝑥𝑇 ⟩𝑖 = ⟨𝑥𝑅 ⟩𝑖 . result is shown in Fig. 8. As shown in Fig. 8(b), the state of the
15: end if drone swarm fluctuates slightly when a few drones are lost, and the
16: else if Drone 𝑖 is LEADERLESS then coordination can be quickly restored. Fig. 8(c) shows that some drones
17: Set drone 𝑖 as FOLLOWER. can directly or indirectly interact with the leader to maintain stable co-
18: Set drone 𝑗 as the father drone of 𝑖. ordination even though the key drone in the drone swarm is lost. Some
19: Set ⟨𝑥𝑇 ⟩𝑖 = ⟨𝑥𝐹 ⟩𝑖 . drones that are unable to interact with the leader from the new swarms
20: end if can achieve local leaderless consensus. The other individual drones
21: Put ⟨𝑥𝑇 ⟩𝑖 into the state information queue. maintain the previous state information. Although the drone swarm
22: Calculate ⟨𝑥𝑃 (𝑡)⟩𝑖 by (12) ∼ (14). is split, it still maintains local stability and is capable to complete
23: Update consensus state by using (15). coordinative work, which means that the proposed system architec-
24: end if ture has good robustness to the loss of key drones in confrontational
25: Broadcast the state information to its one-hop neighbors. environments.
26: end for Finally, we conducted the merging simulation of multiple drone
swarms. Two scenarios were deployed, with one containing 4 × 25
drones and the other containing 4 × 50 drones. The leader drones in the
two scenarios have different dynamic state information. Fig. 3 shows
The drones in the swarm achieve convergence of coordination state the simulation results in which four drone swarms reach different con-
within 20 iterations and maintain stable coordination, providing strong sensus on their own firstly. After that, the swarms begin to merge when

10
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

the drones of different swarms are able to communicate immediately References

with each other. Fig. 9 shows the changes in state information when
the four drone swarms are merging. It turns out that the proposed Basiri, M., Schill, F., Floreano, D., Lima, P.U., 2014. Audio-based localization for swarms
of micro air vehicles. In: 2014 IEEE International Conference on Robotics and
fast coordination approach is able to achieve multi-swarm merged
Automation (ICRA). pp. 4729–4734.
coordination. Berlinger, F., Gauci, M., Nagpal, R., 2021. Implicit coordination for 3D underwater
collective behaviors in a fish-inspired robot swarm. Science Robotics 6 (50), 1–15.
Brambilla, M., Ferrante, E., Birattari, M., Dorigo, M., 2013. Swarm robotics: a review
6. Conclusion
from the swarm engineering perspective. Swarm Intell. 7 (1), 1–41.
Carli, R., Cavone, G., Epicoco, N., Ferdinando, M.D., Scarabaggio, P., Dotoli, M., 2020.
This paper proposed a system architecture and semi-autonomous Consensus-based algorithms for controlling swarms of unmanned aerial vehicles. In:
coordination algorithm based on local interaction for large-scale drone International Conference on Ad-Hoc Networks and Wireless(ADHOC-now 2020). pp.
84–99.
swarms. Drones can interact with adjacent drones and then achieve Chen, W., Liu, J., Guo, H., 2020a. Achieving robust and efficient consensus for
global coordination with regard to leaders by using the proposed ap- large-scale drone swarm. IEEE Trans. Veh. Technol. 69 (12), 15867–15879.
proaches. The architecture and algorithm ensure that large-scale drone Chen, W., Liu, J., Guo, H., Kato, N., 2020b. Toward robust and intelligent drone swarm:
swarms can quickly reach coordination under the conditions of limited Challenges and future directions. IEEE Netw. 34 (4), 278–283.
Chmaj, G., Selvaraj, H., 2015. Distributed processing applications for UAV/drones: a
communication resources and node loss. A prediction mechanism is
survey. In: Progress in Systems Engineering. pp. 449–454.
introduced to address the problem of time delay caused by local interac- Coppola, M., McGuire, K.N., De Wagter, C., De Croon, G.C., 2020. A survey on
tion during information delivery. Moreover, we adapt the Runge–Kutta swarming with micro air vehicles: Fundamental challenges and constraints. Front.
method to discretize the algorithm, enabling the algorithm to be ap- Robot. AI 7 (20), 1–18.
plied in large-scale drone swarm systems. A merging algorithm is Cui, Q., Liu, P., Wang, J., Yu, J., 2017. Brief analysis of drone swarms communication.
In: 2017 IEEE International Conference on Unmanned Systems (ICUS). pp. 463–466.
also proposed to solve the multi-drone swarms merging problem. The
Doostmohammadian, M., Taghieh, A., Zarrabi, H., 2021. Distributed estimation ap-
simulation results show that the fast coordination approach can achieve proach for tracking a mobile target via formation of UAVs. IEEE Trans. Autom.
global coordination rapidly for large-scale drone swarms and merge Sci. Eng. 19 (4), 3765–3776.
multi-swarm in confrontational environments, which has promising Dorigo, M., Floreano, D., Gambardella, L.M., Mondada, F., Nolfi, S., Baaboura, T.,
Birattari, M., Bonani, M., Brambilla, M., Brutschy, A., et al., 2013. Swarmanoid: a
applications in practical scenarios.
novel concept for the study of heterogeneous robotic swarms. IEEE Robot. Autom.
In this paper, our main focus is on the leader–follower coordination Mag. 20 (4), 60–71.
approach of drone swarms. Given that drone swarms operate in a Hindi, H., 2004. A tutorial on convex optimization. In: Proceedings of the 2004
confrontational environment, it is possible for the leader drone to American Control Conference(ACC). pp. 3252–3265.
Hönig, W., Preiss, J.A., Kumar, T.S., Sukhatme, G.S., Ayanian, N., 2018. Trajectory
lose its connection with the control center. Therefore, future research
planning for quadrotor swarms. IEEE Trans. Robot. 34 (4), 856–869.
should prioritize enhancing the collaborative decision-making capabil- Hou, R., Zhou, S., Zheng, Y., Dong, M., Ota, K., Zeng, D., Luo, J., Ma, M., 2021.
ity of drone swarms. This would enable drones to possess autonomous Cluster routing-based data packet backhaul prediction method in vehicular named
decision-making abilities, continuous learning capabilities, and adapt- data networking. IEEE Trans. Netw. Sci. Eng. 8 (3), 2639–2650.
ability to dynamic environments. Consequently, it would facilitate Huang, S., Zeng, D., Qu, Z., 2022. Toward performance efficient UAV task scheduling
in cloud native edge. In: GLOBECOM 2022 - 2022 IEEE Global Communications
the achievement of a fully decentralized and distributed cooperative Conference. pp. 4517–4522.
working mode for the drone swarm system. Hussain, M.Z., Irshad, M., Sarfraz, M., Zafar, N., 2015. Interpolation of discrete time
signals using cubic spline function. In: 2015 19th International Conference on
Information Visualisation(IV). pp. 454–459.
CRediT authorship contribution statement Hussen, H.R., Choi, S.-C., Kim, J., Park, J.-H., 2017. Stateless and predictive geo-
graphic multicast scheme in flying ad-hoc networks. In: 2017 Ninth International
Wu Chen: Conceptualization, Methodology, Software, Writing – Conference on Ubiquitous and Future Networks (ICUFN). pp. 685–690.
Kallenborn, Z., 2018. The era of the drone swarm is coming, and we need to be ready
original draft. Jiayi Zhu: Validation, Software, Data curation, Visual-
for it. Modern War Institute at West Point 25.
ization, Investigation, Writing – original draft. Jiajia Liu: Supervision, Li, S.Z., 2009. Markov Random Field Modeling in Image Analysis. Springer-Verlag, NY,
Project administration, Funding acquisition. Hongzhi Guo: Formal USA.
analysis, Writing – reviewing & editing. Liu, X., Cao, J., Xie, C., 2019. Finite-time and fixed-time bipartite consensus of multi-
agent systems under a unified discontinuous control protocol. J. Franklin Inst. B
356 (2), 734–751.
Declaration of competing interest Liu, C., Wang, X., Ren, Y., Wang, X., Zhang, J., 2022. A novel distributed sec-
ondary control of heterogeneous virtual synchronous generators via event-triggered
communication. IEEE Trans. Smart Grid 13 (6), 4174–4189.
The authors declare that they have no known competing finan- McCune, R.R., Madey, G.R., 2013. Swarm control of UAVs for cooperative hunting with
cial interests or personal relationships that could have appeared to DDDAS. Procedia Comput. Sci. 18 (3), 2537–2544.
influence the work reported in this paper. Moreau, L., 2005. Stability of multiagent systems with time-dependent communication
links. IEEE Trans. Autom. Control 50 (2), 169–182.
Nägeli, T., Conte, C., Domahidi, A., Morari, M., Hilliges, O., 2014. Environment-
Data availability independent formation flight for micro aerial vehicles. In: 2014 IEEE/RSJ
International Conference on Intelligent Robots and Systems(IROS). pp. 1141–1146.
Onthoni, D.D., Sahoo, P.K., Atiquzzaman, M., 2023. ASAA: Multihop and multiuser
Data will be made available on request.
channel hopping protocols for cognitive-radio-enabled internet of things. IEEE
Internet Things J. 10 (9), 8305–8318.
Acknowledgment Orfanus, D., De Freitas, E.P., Eliassen, F., 2016. Self-organization as a supporting
paradigm for military UAV relay networks. IEEE Commun. Lett. 20 (4), 804–807.
Oubbati, O.S., Atiquzzaman, M., Lim, H., Rachedi, A., Lakas, A., 2022. Synchronizing
This work was supported by National Key R&D Program of China UAV teams for timely data collection and energy transfer by deep reinforcement
(2022YFB3104200), in part by National Natural Science Foundation learning. IEEE Trans. Veh. Technol. 71 (6), 6682–6697.
of China (62202386), in part of Project funded by China Postdoctoral Oubbati, O.S., Lakas, A., Lorenz, P., Atiquzzaman, M., Jamalipour, A., 2021. Leveraging
communicating UAVs for emergency vehicle guidance in urban areas. IEEE Trans.
Science Foundation (2023M732861), and in part by Special Funds for
Emerg. Top. Comput. 9 (2), 1070–1082.
Central Universities Construction of World-Class Universities (Disci- Pal, M., 2022. A sliding mode based finite-time consensus protocol for heterogeneous
plines) and Special Development Guidance (0639023GH0202091 and multi agent UAS. In: 2022 IEEE International IOT, Electronics and Mechatronics
0639023SH0201091). Conference (IEMTRONICS). pp. 1–7.

11
W. Chen et al. Journal of Network and Computer Applications 221 (2024) 103769

Pasek, P., Kaniewski, P., 2022. A review of consensus algorithms used in distributed Jiayi Zhu (15039565129@mail.nwpu.edu.cn)
state estimation for UAV swarms. In: 2022 IEEE 16th International Conference National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data
on Advanced Trends in Radioelectronics, Telecommunications and Computer Application Technology, School of Cybersecurity, Northwestern Polytechnical Univer-
Engineering (TCSET). pp. 472–477. sity, Xi’an, China.
Ren, W., Beard, R.W., 2007. Distributed Consensus in Multi-Vehicle Cooperative Jiayi Zhu received the B.S. degree in computer science and technology from Hainan
Control: Theory and Applications. Springer Publishing Company, Incorporated. University, Haikou, China, in 2020. He is currently working toward the M.S. degree
Ren, W., Cao, Y., 2010. Distributed Coordination of Multi-Agent Networks: Emergent in cyberspace security with Northwestern Polytechnical University, Xi’an, China. His
Problems, Models, and Issues. Springer-Verlag, NY, USA. research interests include drone countermeasures, Ad Hoc network, and cybersecurity.
Rikos, A.I., Charalambous, T., Johansson, K.H., Hadjicostis, C.N., 2022. Distributed
event-triggered algorithms for finite-time privacy-preserving quantized average
Jiajia Liu (Liujiajia@nwpu.edu.cn)
consensus. IEEE Trans. Control Netw. Syst. 1–12.
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data
Rubenstein, M., Cornejo, A., Nagpal, R., 2014. Programmable self-assembly in a
Application Technology, School of Cybersecurity, Northwestern Polytechnical Univer-
thousand-robot swarm. Science 345 (6198), 795–799.
sity, Xi’an, China.
Saballa, J., 2022. US army to test largest-ever drone swarm. https://www. Jiajia Liu received the B.S. degree in computer science from the Harbin Institute of
thedefensepost.com/2022/04/26/us-army-drone-swarm/. Technology, Harbin, China, in 2004, the M.S. degree in computer science from Xidian
Saffre, F., Hildmann, H., Karvonen, H., 2021. The design challenges of drone swarm University, Xi’an, China, in 2009, and the Ph.D. degree in information sciences from
control. In: International Conference on Human-Computer Interaction(HCII 2021). Tohoku University, Sendai, Japan, in 2012.
pp. 408–426. He was a Full Professor (Vice Dean) with the School of Cybersecurity, Northwestern
Tahir, A., Böling, J., Haghbayan, M.-H., Toivonen, H.T., Plosila, J., 2019. Swarms of Polytechnical University, Xi’an, China. He has authored or coauthored more than
unmanned aerial vehicles—a survey. J. Ind. Inf. Integr. 16 (4), 1–16. 220 peer-reviewed papers in many high quality publications, including prestigious
Tegicho, B.E., Geleta, T.N., Bogale, T.E., Eroglu, A., Edmonson, W., Bitsuamlak, G., IEEE journals and conferences. His research interests include intelligent and connected
2021. Effect of wind on the connectivity and safety of large scale UAV swarms. In: vehicles, mobile/edge/cloud computing and storage, IoT security, wireless and mobile
2021 IEEE International Black Sea Conference on Communications and Networking ad hoc networks, and SAGIN. He was the recipient of the IEEE ComSoc Best YP
(BlackSeaCom). pp. 1–6. in Academia Award in 2020, the IEEE VTS Early Career Award in 2019, the IEEE
Tian, W., Zhao, Y., Hou, R., Dong, M., Ota, K., Zeng, D., Zhang, J., 2023. A centralized ComSoc Asia–Pacific Outstanding Young Researcher Award in 2017, and the IEEE
control-based clustering scheme for energy efficiency in underwater acoustic sensor ComSoc Asia–Pacific Outstanding Paper Award in 2019. He has been actively joining
networks. IEEE Trans. Green Commun. Netw. 7 (2), 668–679. the society activities, including an Associate Editor for IEEE Transactions on Wireless
Wang, Z., Guo, J., Chen, Z., Yu, L., Wang, Y., Rao, H., 2022a. Robust secure UAV Communications since May 2018, IEEE Transactions on Computers from October 2015
relay-assisted cognitive communications with resource allocation and cooperative to June 2017, and IEEE Transactions on Vehicular Technology from January 2016
jamming. J. Commun. Netw. 24 (2), 139–153. to December 2020, and the Editor for IEEE Network since July 2015, and IEEE
Wang, J., Zhang, Z., Tian, B., Zong, Q., 2022b. Event-based robust optimal consensus Transactions on Cognitive Communications and Networking since January 2019. He
control for nonlinear multiagent system with local adaptive dynamic programming. is the Chair of IEEE IoT, Ad Hoc and Sensor Networks Technical Committee and the
IEEE Trans. Neural Netw. Learn. Syst. 1–14. Distinguished Lecturer of the IEEE Communications Society and Vehicular Technology
Yanmaz, E., Yahyanejad, S., Rinner, B., Hellwagner, H., Bettstetter, C., 2018. Drone Society.
networks: Communications, coordination, and sensing. Ad Hoc Netw. 68 (1), 1–15.
Zou, Y., Zhu, J., Wu, T., Guo, H., Wei, H., 2021. Cooperative drone communications Hongzhi Guo (hongzhi.guo@nwpu.edu.cn)
for space-air-ground integrated networks. IEEE Netw. 35 (5), 100–106. National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data
Application Technology, School of Cybersecurity, Northwestern Polytechnical Univer-
sity, Xi’an, China.
Wu Chen (chenwu@nwpu.edu.cn) Hongzhi Guo (Member, IEEE) received the B.S., M.S., and Ph.D. degrees in
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data computer science and technology from the Harbin Institute of Technology, Harbin,
Application Technology, School of Cybersecurity, Northwestern Polytechnical Univer- China, in 2004, 2006, and 2011, respectively. He is currently an Associate Professor
sity, Xi’an, China. with the School of Cybersecurity, Northwestern Polytechnical University, Xi’an, China.
Wu Chen received the B.S. and M.S. degrees in navigation guidance and con- He has authored or coauthored more than 40 peer-reviewed papers in many prestigious
trol from Northwestern Polytechnical University, Xi’an, China, in 1992 and 1997, IEEE journals and conferences. His research interests include edge computing, SAGSIN,
respectively, and the Ph.D. degree in control theory and control engineering from IoT security, AI security, and 5G security. He was the recipient of the WiMob Best Paper
Northwestern Polytechnical University, in 2000. He is currently an Associate Professor Award in 2019, and the IEEE TVT Top Reviewer Award in 2019. He has been actively
with the School of Cybersecurity, Northwestern Polytechnical University. His research joining the society activities, including an Associate Editor for IEEE Transactions on
interests include Ad Hoc network, intelligent control, embedded system technology, Vehicular Technology since January 2021, Frontiers in Communications and Network
and information security. since January 2021, and the Editor for International Journal of Multimedia Intelligence
and Security since March 2019.

12