Unmanned Aerial Vehicles (UAV) Networking Algorithms: Communication, Control, and AI-Based Approaches

Abstract

This paper focuses on algorithms and technologies for unmanned aerial vehicles (UAVs) networking across different fields of applications. Given the limitations of UAVs in both computations and communications, UAVs usually need algorithms for either low latency or energy efficiency. In addition, coverage problems should be considered to improve UAV deployment in many monitoring or sensing applications. Hence, this work firstly addresses common applications of UAV groups or swarms. Communication routing protocols are then reviewed, as they can make UAVs capable of supporting these applications. Furthermore, control algorithms are examined to ensure UAVs operate in optimal positions for specific purposes. AI-based approaches are considered to enhance UAV performance. We provide either the latest work or evaluations of existing results that can suggest suitable solutions for specific practical applications. This work can be considered as a comprehensive survey for both general and specific problems associated with UAVs in monitoring and sensing fields.

1. Introduction

1.1. Motivation

Unmanned aerial vehicles (UAVs) are an emerging technology in recent years that is expected to have a greater impact on human life in the future. UAVs have been used in a variety of applications (Figure 1). They provide low-cost surveillance services, perform precise tasks on demand, and minimize human intervention. These applications include transportation [1,2] and traffic control, border patrol surveillance [3], search and rescue and disaster relief assistance, and wireless networking in smart agriculture and forestry [4,5]. UAVs are also widely used in the military field. Real combat scenarios involving UAVs have been described in [6], showing that UAVs can perform diverse military missions. In addition, UAVs are used for commercial purposes, such as surveying and mapping [7], volcano monitoring [8], UAV control by brain waves [9], early warning of severe weather [10], mobile WSN networks [11] and drone delivery services [12].

To promote the practical applications of UAV swarms, research on their communication and cooperation networks is needed. Efficient and reliable communication networks between UAVs are essential to complete any cooperative task. Network links among UAVs enable individual UAV nodes to directly communicate for information sharing and cooperative tasks [13]. UAV swarms are characterized by high mobility, frequently changing architectures, limited computing resources, and easy disconnection. Therefore, improving the efficiency of resource utilization is of great significance to ensure communication among UAVs. In particular, channel access and network routing are essential for efficient communication among UAVs [14,15].

As real-world missions become more complex in dynamic and uncertain environments, traditional approaches face challenging problems such as computational complexity, low cooperativeness, and limited intelligence of UAV swarms. However, thanks to recent significant advances in artificial intelligence (AI) algorithms, the application of AI algorithms in UAV swarm control offers some significant improvements. AI enables the use of and learning from massive data to improve UAV control. Compared with conventional optimization techniques, AI can handle the resources on UAVs flexibly [16]. With the help of AI, environmental data from devices connected to UAVs are collected and used effectively [17]. Several methods, including Global Positioning System (GPS) monitoring, computer vision, and machine learning algorithms, can be used to achieve these. Speech recognition, scene recognition, object detection, and image classification are just some of the many areas where AI is penetrating. Especially, deep learning methods for AI are mentioned. In addition to supporting the enhancement of UAV data processing and communication capabilities, AI also greatly supports UAV swarm control. Typically, it enhances some UAV swarm features such as automatic decision making, coordination under interference, flexibility, and scalability of the formation. However, the application of AI algorithms also brings some challenges, such as the requirement of large computational resources, the ability to update training data for AI in real time, and high bandwidth requirements among UAVs when exchanging information.

1.2. Related Work

New technologies in the field of UAV swarms have recently been the focus of many scientists and large corporations. Surveys of published papers have also been conducted to provide comprehensive information on UAV swarms. In this section, we discuss some of the latest relevant surveys conducted in this field. They focus heavily on communication technology and the application of AI to effectively control UAV swarms.

Increasing the coverage area using UAVs is one of the important research areas. Many technologies have been developed to provide better coverage areas using UAV networks. A detailed survey in [18] highlights various issues from the coverage perspective of UAV networks. The authors classified the issues into groups and mentioned various limitations. The findings of the survey highlight the need for further in-depth studies to explore the coverage area of UAV networks.

Another survey conducted in 2023 addressed the limitations of ground infrastructure as the number of devices increases [16]. Utilizing the flexibility of UAVs, the authors evaluated the use of UAVs to improve IoT network performance. UAVs provide wireless access to devices when the ground network is overloaded or disconnected. AI-based methods for optimizing, scheduling, and orchestrating UAV-supported IoT networks were investigated. A comprehensive analysis was conducted on the influence of implementing advanced AI architectures, models, and methods on various features of IoT networks.

Due to the stringent and complex requirements of UAVs, cloud-based AI models are increasingly difficult to meet. Therefore, the authors of the study [19] focused on edge AI models, which are AI models running on devices or servers close to users. This model fully exploits the strengths of UAVs in improving IoT services. Complete analyses were conducted on the influence of edge AI on UAV characteristics such as autonomous navigation, security and privacy, computer vision, and communication. The authors also focused on current prominent applications of swarm UAVs, such as distribution systems, civil infrastructure inspection, smart agriculture, search and rescue (SAR) operations, and wireless communication base stations (BS).

Generative artificial intelligence (GAI) has been proposed and investigated in [20]. In their work, the authors reviewed the key technologies of GAI and its important role in UAV networks. The method applies GAI to improve communication and networking efficiency and security of UAV systems. This study also proposed a new framework of GAI for advanced UAV networks by estimating the spectrum map supporting UAVs and optimizing the transmission rate based on the proposed framework. For example, based on the data collected from a part of the target area, GAI takes into account the limited resources of the UAV to accurately infer the status of the entire area. This allows for reasonable resource allocation and trajectory planning.

Awareness of environmental changes and the ability to adapt to mission requirements quickly are key to successful UAV deployment [21]. Therefore, in [22], a new study in 2024 focused on UAV-centric ML solutions. This study examined how well they meet the network requirements, considering the changing roles, collaboration, cooperation, and context of UAV networks. Solutions were proposed for air-to-air, air-to-ground, and ground-to-air communications, as well as mobile edge computing. The survey highlighted the urgent need for UAVs in emerging 5G/6G networks.

A recent study [23] provides an in-depth survey of UAV-enabled integrated sensing and communications (ISAC). In the context of 6th-generation (6G) wireless networks, UAVs are deployed as airborne wireless platforms to provide better coverage. At the same time, they meet the advanced sensing and communications (S&C) services. However, due to the size, weight, and power (SWAP) constraints of UAVs, the application of UAV-enabled ISAC opens new opportunities and challenges. The authors provide an overview of ISAC and propose various solutions to optimize S&C performance. Specifically, UAV motion control, wireless resource allocation, and interference management for ISAC systems are used for one or more UAVs. Two application scenarios to exploit the synergy between S&C are sensor-enabled UAV communications and communication-enabled UAV sensors.

Another recent study addressing machine learning (ML) techniques to support UAV networks is presented in [24]. The adaptive nature of UAV networks along with their context-sensitive perception makes UAVs a suitable candidate for future communication technologies. With self-adaptive techniques, ML is considered one of the best technologies to support UAV networks in achieving their communication goals and reliability. The authors highlighted four key components of UAV communication operations where ML can make significant contributions. These include feature perception and extraction, feature interpretation and reconstruction, trajectory and mission planning, and aerodynamic control and operation. The survey classifies the latest popular ML tools based on their applications to the above four components and conducts a gap analysis. It is shown that different ML techniques are dominating the applications for the four key modules of UAV communication operations. Various studies have focused on the perspective of UAV communication between ground and airborne stations. It is believed that the use of mobile edge computing (MEC) in parallel with UAV networks will make it more reliable and efficient for future communication operations.

Table 1. Summary of relevant surveys.

Ref.	Survey Scope	Features of the UAV Covered	AI-Based
[18]	UAV network, multiple aircraft coordinated	Connectivity, adaptability, safety, privacy, security, and scalability	No
[16]	AI orchestrates UAV-enabled IoT networks	Advanced AI architectures, models, and methods applied to different aspects of UAV-enabled IoT networks	Yes
[19]	Impact of Edge AI on Key technical aspects of UAVs and applications	Autonomous navigation, fleet control, energy management, security and privacy, computer vision, communications, applications	Yes
[20]	The main application of GAI in improving UAV communication and network performance	New GAI framework for advanced UAV networks, UAV-assisted spectrum map estimation, transmission rate optimization between UAV and GAI	Yes
[21]	Quadcopter control algorithm	Control algorithm, performance of UAV	No
[22]	ML solutions for UAVs	Role, collaboration, partnership, and changing network landscape of UAVs	No
[23]	ISAC supports UAVs, optimizes S&C performance	S&C, UAV motion control, wireless resource allocation, interference management	No
[24]	ML engineering and UAV communications	Combining UAV and ML, ML techniques, and UAV applications	No

briefly provides both the scope and the features of UAVs covered in this paper. All the details are provided in the following sections.

1.3. Main Contributions

This paper first addresses the general applications of UAV groups or UAV swarms. Then, communication routing protocols for UAVs are considered, as they support the applications considered. Furthermore, control algorithms are mentioned to assist UAVs to operate at the right location for specific purposes. AI-based methods are considered to address core challenges in UAV networks. Traditional methods have limitations such as real-time multi-sensor data processing, adaptive routing in dynamic environments, and multi-task swarm control for coordinated operations among UAVs. We provide the latest work or review the existing results that can suggest suitable solutions for specific applications in practice. This work can be considered a comprehensive survey of general or specific problems with UAVs operating in the field of surveillance and sensing. Finally, we discuss the open issues and challenges for future research and development.

1.4. Organization of the Article

This review is organized into seven sections, and the structure and organization of this paper are shown in Figure 2. The rest of this paper is organized as follows. Section 2 provides the models and some architectures of UAV networks in some specific applications. Routing protocols for UAVs are discussed in Section 3. Section 4 presents the communication in UAV networks. Control algorithms are provided in Section 5 to support UAV groups and UAV swarms operating in different sensor fields. Section 6 provides AI-based approaches to support UAVs to operate more efficiently, thereby addressing core challenges in UAV networks that traditional methods have limitations, such as real-time multi-sensor data processing, adaptive routing in dynamic environments, and multi-task swarm control for coordinated operations among UAVs. Finally, conclusions and future work are presented in Section 7.

2. UAV Network Communication Model and Architecture

Communication architecture is a key factor in the cooperation and control of UAVs in swarms. In the early stages, UAV swarm control is performed by a single central station (usually located on the ground). The central station has enough communication resources to communicate with all UAVs in the swarm. This is the basic concept for centralized communication architecture. When the number of UAVs increases or the exchanged data arises due to complex and diverse tasks, a decentralized structure is needed to meet the requirements—namely, reducing the dependence on UAVs on central stations [25,26]. In this section, we systematically review and introduce studies on communication architectures for UAV networks. At the same time, we also analyze the advantages and disadvantages of these architectures, as well as related issues.

2.1. Centralized Communication Architecture

The centralized communication architecture evolved from single UAV systems. It was then gradually developed and deployed for UAV swarms, as shown in Figure 3. The architecture consists of a central node (fixed network infrastructure) connected to all UAVs in the swarm. Each UAV forms a one-to-one relationship with the infrastructure and directly receives control commands from the central station. This type of centralized communication architecture is relatively stable and easy to deploy. It often applies simple routing algorithms. However, this architecture is only suitable when the scale of the UAV swarm and the coverage area of the mission are small, and the tasks to be performed are relatively simple. The coverage area size depends on many factors, such as the UAV’s transmit power, antenna gain, path loss pattern, and environmental conditions (in real environments, reflection and attenuation effects may reduce the range). Some control methods of this architectural model include ground control station (GCS) (UAVs receive commands directly from the station), the Leader–Follower model, cloud computing, or military architectures (where UAVs are coordinated from the command center) [27].

The connections between UAVs require infrastructure for data transmission. The UTI (UAV to the Infrastructure) communication distances are often much larger than the UTU (UAV to UAV) distances, causing longer latency. Considering the high mobility of UAVs and the wide operating range requirements of swarm applications, this type of architecture gradually becomes difficult to meet. If the infrastructure fails or is attacked, the entire network will be paralyzed. Therefore, the disadvantage of the centralized communication architecture is the Single Point of Failure (SPOF). It gradually becomes unreliable in complex requirements.

2.2. Decentralized Communication Architecture

Decentralized architecture is an important trend in modern UAV systems because of its scalability, flexibility, and high reliability. This architecture is essentially a control model in which there is no single control center [28]. Instead, UAVs operate independently, or rely on automatic coordination with other UAVs in the network. In this system, based on information from sensors, communication among UAVs, or algorithms, each UAV can make decisions without direct guidance from a central server or UAV leader [29].

Since UAVs operate at high speeds and their missions can cover large areas, the connection status of UAVs changes frequently. Therefore, with its high flexibility and adaptability, decentralized architectures are increasingly proving their role in UAV swarm applications. The “Single-Group Swarm Ad hoc Network” is a model that is very suitable for the UAV swarm problem. UAVs in the group self-organize and coordinate with each other without fixed network infrastructure or central servers. This helps the whole group operate as a unified entity [30].

2.2.1. Single Swarm Ad Hoc Network

In the “single-group swarm ad hoc network” (Figure 4), the internal communication of the swarm is independent of the infrastructure. Communication between the swarm and the infrastructure is established through a single-point link. There is a specific UAV in the swarm assigned to perform this link called the gateway UAV. The other UAVs are relay nodes that forward data within the swarm. The UAVs in the swarm can share information in real time to optimize collaborative control and improve efficiency. Similarly, the mutual communication between the gateway UAV and the infrastructure also allows swarms to upload and download information such as guidance, task assignment, or control information. For the UAV gateway, it has two additional types of transceivers. One is to communicate with other UAVs at low power and short range. The other one is to communicate with the infrastructure at high power and long range. Thus, other UAVs in the swarm only need to carry low-cost short-range transceivers. They are more convenient for moving and performing tasks. Thanks to such architectures, the communication range of the network is expanded. The ability to coordinate among different types of UAVs increases in the same swarm. This not only significantly expands the communication range for large coverage requirements but also makes small UAVs with smaller payloads more useful [31]. However, to ensure that the UAV swarm is consistently connected, the “single-group ad hoc network” architecture requires all UAVs in the swarm to have the same flight patterns (speed and direction). Therefore, this architecture is suitable for small UAVs. At the same time, it fully exploits the capabilities of small UAVs in complex tasks that they normally cannot complete.

For each specific task, in each different phase, the communication architecture in the swarm can change [32]. The change in the communication architecture in the UAV swarm should be based on mission requirements (e.g., wide-area surveillance, centralized data collection, surveillance tracking) or on the mission phase (e.g., initial deployment, mission execution, emergency); environmental conditions (e.g., weather, interference); and system status (e.g., number of UAVs, UAVs losing connection, remaining energy of each UAV). In addition, real-time surveillance can be combined with adaptive algorithms to ensure efficiency and reliability. Figure 5a–c shows three common communication architectures in the swarm.

Figure 5a shows an example of a ring architecture. The UAVs in the ring architecture form a closed communication loop through bidirectional connections. Any UAV can act as a gateway node for the swarm. When the direct link between two adjacent UAVs fails, the target UAV can reconnect through the communication loop. Therefore, the ring architecture has a certain stability. However, this architecture lacks the scalability for other UAVs to join the swarm or coordinate with other swarms. In Figure 5b, the communications are organized in a star-shaped architecture. The gateway UAV is in the middle, and it communicates not only with the infrastructure but also with the entire UAV swarm. The star architecture has the inherent disadvantage of SPOF (if the gateway node fails, it will cause the entire system to fail).

The mesh architecture is a combination of ring and star architectures, and it has the advantages of both systems (Figure 5c). All UAV nodes in the swarm have the same capabilities, with both terminal and routing functions. Information flow, from one node to another, can be implemented in many different forms, and any UAV can act as a gateway node for the swarm. The mesh architecture has now become the standard for intra-swarm communication systems [32].

However, practical applications often require swarms to use not only small UAVs but also medium and large UAVs. Therefore, the above “single-group swarm ad hoc network” architecture makes it difficult to meet these needs. Usually, similar UAVs can often fly close to each other. Different types of UAVs are far apart. Therefore, the UAVs in the swarm are divided into different groups. Similar UAVs are within the same distance. These are called “multi-group swarm ad hoc network” and “multi-layer swarm ad hoc network” architectures.

2.2.2. Multi-Group Swarm Ad Hoc Network

To address the limitations of the “single-group ad hoc swarm network”, the “multi-group ad hoc swarm network” (Figure 6) integrates both the centralized architecture and the “single-group ad hoc swarm network” architecture. Different types of groups are applied differently depending on the task. In general, the architecture is organized in a centralized manner to control and monitor the groups. In each group, UAVs communicate with each other in an ad hoc manner. The communications between groups (GTG) are performed through the infrastructure. This is still the responsibility of the gateway UAVs. When the mission scenarios require multiple types of UAVs, the “multi-group ad hoc swarm network” architecture can be used. However, the robustness of this architecture needs more attention to since the “multi-group ad hoc swarm network” architecture is semi-centralized. At the same time, the GTG communications between two UAVs in different groups still encounter high latency. A specific application of multi-group architecture is the joint multi-target operation on the battlefield in the military. The Norwegian military UAV operation [33] clearly describes the roles and responsibilities of each component in a military operation using UAVs. This includes standards for communication between the control station and the UAV groups, and the roles and responsibilities in the information exchange in military operations.

2.2.3. Multi-Layered Ad Hoc Swarm Network

The “multi-layer swarm ad hoc network” architecture is a type of architecture developed from the “multi-group swarm ad hoc network”. It is suitable for many different UAV models (Figure 7). In this architecture, a group of adjacent UAVs of the same type forms an ad hoc network. It serves as the first layer of the communication architecture. Different UAV groups rely on the gateway UAV to perform GTG communication. It is the second layer. The nearest gateway UAV communicating with the infrastructure is the third layer of the architecture. Communication between any two UAVs in the “multi-layer swarm ad hoc network” architecture does not require infrastructure forwarding. Mutual communication of UAVs in the same group takes place at the first level. Communication between UAVs in different groups is routed through the gateway UAV. Data packets pass through the first and second layers, respectively. Therefore, there is no SPOF in the “multi-layer swarm ad hoc network”.

The “multi-layered ad hoc swarm network” architecture rapidly deploys network reconstruction when the number of UAVs changes in the swarm. Therefore, this architecture is suitable for complex mission scenarios. These missions are characterized by a large number of UAVs, frequent network structure changes, and a high frequency of communication between UAV nodes. When constructing this network structure, determining the number of layers is very important for the system to operate effectively [34].

2.3. Comment

As analyzed above, we can see that the UAV swarm communication architecture technology has made great progress. When facing different mission scenarios, it is necessary to choose a suitable communication architecture. The advantages and disadvantages of the four presented architectural models are summarized in

Table 2. Summary of the communication architecture.

Advantages and Disadvantages	Focused Communication	Decentralized
		Single Group	Multi-Group	Multilayer
Integration of different types of UAVs	No	Yes	Yes	Yes
Self-configuring	No	Yes	No	Yes
Forwarding information	No	Yes	Yes	Yes
Computational resource requirements for UAVs	Low	Medium	High	Very high
Large media coverage?	No	Yes	Yes	Yes
Low latency?	No	Yes	No	Yes

. From there, we can intuitively see that the centralized communication architecture is suitable for simple tasks. They require a small number of UAVs. Each UAV needs a long-range communication link with the central station. The decentralized communication architecture extends the communication range. The dedicated gateway UAV is responsible for UTI communication. The “single-group ad hoc swarm network” architecture is suitable for swarms with the same type of UAVs, while the “multi-group ad hoc swarm network” and “multi-layer ad hoc swarm network” architectures can be deployed using different types of UAVs. In the “multi-group ad hoc swarm network”, communication between two different groups can also be delayed due to passing through many network ports. In terms of applicability and robustness, the “multi-layer ad hoc swarm network” architecture is a reliable system because it overcomes the SPOF.

When discussing the advantages and disadvantages of UAV swarm communication architectures, we often focus on the requirements for high coverage and connection maintenance. Coverage plays an important role in information gathering and situation analysis. The real-time communication capability of UAV swarms can only be guaranteed if connections are maintained. However, UAV swarms often perform tasks in unknown environments. Threats and obstacles often appear randomly in time and space. Therefore, the ability to change the UAV formation configuration and the flexibility to join or leave the swarm are also necessary. For uninterrupted connection, the distance in the UAV swarm ad hoc network configuration must not exceed the sensitivity of the receiver. At the same time, the distance must be limited to the minimum signal-to-noise ratio (SNR) or the corresponding received signal strength indicator (RSSI) [35]. Here, the swarm models in nature are worth studying and applying, such as flocks of birds, flocks of fish, bees, fireflies, wolves, elephants, etc. However, the problem of ensuring connectivity and maintaining flexibility is facing great challenges in terms of security and information safety requirements. Maintaining the balance between these two factors can become a future research direction and trend.

3. Classifications of Routing Protocols in UAV Networks

UAV networks require distinct routing strategies based on the network structure and the system deployment environment [13]. When UAVs move rapidly, they will cause temporary connection interruptions among nodes. Therefore, routing protocols must be able to detect and restore connections quickly while maintaining stable network performance. However, in the process of developing research on this issue, it was found that most traditional routing protocols are not suitable for UAV swarm communication [22]. This leads to two parallel development directions for UAV swarms. These are improvements of existing dynamic routing protocols to suit the application environment, and the emergence of protocols specifically designed for each specific application [19].

In general, these protocols are classified as network topology-based routing, location-based routing, hierarchical routing, and swarm-based routing. In addition, the application of artificial intelligence (AI) to UAV networks based on learning capabilities is also a potential research direction for developing intelligent routing protocols [20]. Emerging research on AI-based routing protocols will be described in detail in the following section of this paper. Figure 8 illustrates the classification of UAV network routing protocols based on the methods mentioned. The following is a discussion of some prominent routing protocols and their use cases.

3.1. Routing Based on Network Structure

Compared with traditional network architectures, the design of routing protocols in UAV networks faces many challenges due to the characteristics of the network such as high mobility and continuous changes in network structure [14]. Based on the operational characteristics, routing protocols on network structure can be divided into static routing protocols, proactive routing protocols, reactive routing protocols, and hybrid routing protocols.

3.1.1. Static Routing Protocol

Static routing protocols have a static routing table that is predefined and uploaded to the UAV before operation, and they cannot be changed. Static routing protocols often have low fault tolerance and limited adaptability in dynamic environments [15]. Below is a description and analysis of some prominent static routing protocols in recent years.

The Load and Carry Delivery routing scheme (LCAD) is specifically designed for UAV networks. It allows flight path establishment from the ground preparation stage. The authors in [36] have developed a full duplex (FD) UAV relay network based on LCAD. The UAV acts as a mobile relay device to transmit information based on the FD technique. Specifically, the UAV will fly in a straight horizontal line at a constant height H with a fixed speed v. The relay service time for the UAV to complete the information exchange between S and D is divided equally into two stages: the first stage (FS) and the second stage (SS), as shown in Figure 9. Data will not be transmitted until the channel is favorable (or the UAV is closer to a node). The simulation results show that it is superior to the half-duplex and static schemes in terms of energy efficiency.

The Data-Centric Routing (DCR) is a static multipoint routing solution to increase the operational efficiency of UAV networks with cluster-based architecture. It is capable of supporting a variety of applications in a defined UAV deployment environment. DCR is deployed when multiple UAVs simultaneously request the same data packet. Based on the advanced publish–subscribe model, DCR allows automatic connection establishment between data publishers and data subscribers. When data is transmitted from the source, target UAVs can receive it either directly from the source or indirectly through intermediate nodes. In the study [37], the authors proposed a UAS (Unmanned Aircraft Systems) model for deployment in firefighting, natural disaster, or search and rescue situations (Figure 10).

The UAS platform enables rapid deployment of small, pre-configured UAVs to establish a wireless network in a short time. Mission-specific UAVs are equipped with additional communication devices to act as access points, while the others act as network routers. In this research scenario, the UAS provides a solution based on the Data Distribution Service (DDS) standard. It builds a distribution path to connect publishers with subscribers. However, the actual distribution path depends on the services provided by the underlying network. The experimental results of UAS based on heterogeneous hardware devices demonstrate the feasibility and advantages of using DDS in a multi-UAV deployment scenario.

3.1.2. Proactive Routing Protocol

Proactive routing protocols (PRPs) work by periodically updating routing tables, allowing fast access to routing paths. PRPs meet the application requirements of low latency and fast routing in complex and dynamic environments [15]. However, as the size and mobility of nodes in the network increase, most PRP schemes face certain challenges [32]. Some prominent proactive routing protocols today are variants developed from the Destination Sequence Distance Vector (DSDV) protocol and the Optimized Link State Routing (OLSR) protocol.

In the DSDV routing protocol, each UAV maintains a routing table along with a sequence number to distinguish between old and new routes. The routing tables are updated periodically or when there is a change in the network. The routing table update message is sent to UAVs in the network. The authors in [38] have developed an Improved Sequencing Heuristic DSDV (ISH-DSDV) routing protocol. It is an improvement of DSDV for UAV networks based on the Nomadic mobility model. To avoid congestion, the authors propose a solution to allocate sequence numbers according to the target node. It helps optimize the data movement process. This protocol updates the routing table with three basic information: final destination, distance, and next node. This method is designed to use the most efficient information in the data transmission process between nodes. In addition, a legacy data entry method is also introduced to easily delete outdated entries. The simulation results show that the ISH-DSDV routing technique is a better choice for deployment in dense networks while improving the operational efficiency of the aircraft.

OLSR is a proactive routing protocol that frequently shares topological data among network nodes. OLSR uses two types of control messages: “Hello” and “Topology Control (TC)”. Each node in the network receives a “Hello” message that informs them about their link status and any neighbors within two hops. To maintain the database needed for packet routing, each node in the network broadcasts TC messages. By responding to events with TC messages and reducing the maximum periodic interval, OLSR is frequently optimized [39]. Some improved variants of the OLSR protocol include Directional Optimized Link State Routing (D-OLSR) [40], which is a routing protocol for UAVs with directional antennas. Mobility and load-aware OLSR (ML-OLSR) [41] was developed to mitigate the impact of high mobility and load imbalance on UAV network performance. The ML-OLSR protocol contains both a mobility-aware algorithm and a load-aware algorithm. P-OLSR is also a variant of the OLSR protocol, designed to predict changes in the expected level of wireless communication quality across nodes in a UAV network based on the GPS data of each UAV.

The core performance of the OLSR routing protocol is multipoint forwarding (MPR). However, the traditional MPR selection algorithm suffers from large redundancy. To solve this problem, the authors in [42] proposed an algorithm for the OLSR called Selective Routing OLSR, noted as SROLSR. It selects a new set of MPRs and makes some modifications to the MPR selection process to reduce the number of MPRs. The simulation results show that the improved SROLSR protocol has better performance than the standard version of the OLSR routing protocol in terms of network load, packet delivery rate, and end-to-end delay.

The above studies highlight research efforts to improve proactive routing protocols for UAV networks, focusing on three important aspects: optimizing energy efficiency, enhancing the ability to adapt to the continuous changes of UAV networks, and improving the overall performance of the system.

3.1.3. Response-Based Routing Protocol

Reactive routing protocol (RRP) is also known as on-demand routing protocol. RRP does not maintain routing information for all nodes in real time but builds a route only when the source node requests communication [27]. Therefore, it overcomes the overhead issues associated with proactive routing protocols. RRP saves bandwidth due to the absence of periodic updates. The main drawback of RRP is the high latency during the optimal path discovery phase. Popular reactive routing protocols are the Dynamic Source Routing (DSR) protocol and the Arbitrary On-Demand Distance Vector (AODV) protocol.

DSR does not require a fixed network infrastructure but allows the network to self-organize and self-configure. In this protocol, each data packet contains a complete list of routing nodes, and all nodes that forward or receive the packet will store this routing information. Therefore, the system will still maintain stable performance even when the network structure changes or when the network nodes move. The AODV routing protocol is often used in UAV networks. During the route lookup process, the source node will search for the routing history established in the past. AODV will initiate a new route request process when no route exists. In the study [43], the authors compared and evaluated the performance of two prominent reactive routing protocols, DSR and AODV, based on the deployment of traffic monitoring applications over UAV networks. The implementation of both protocols was tested on a number of different traffic patterns and mobility and network loads. Both DSR and AODV was evaluated for performance based on factors such as packet delivery ratio, average end-to-end delay, normalized routing load, packet loss, and routing overhead. Based on experimental analysis using NS-2 under constant bit rate (CBR) and TCP traffic source conditions, both protocols provided significant performance improvements, but AODV outperformed DSR in almost every aspect.

An improved AODV protocol was implemented by the authors in [44] to improve the performance of UAV networks. The authors proposed a multilayer energy-aware on-demand AODV protocol (noted as CLEA-AODV). The CLEA-AODV was divided into three parts: routing with the AODV protocol, Glow Swarm Optimization (GSO)-based cluster head selection, and cooperative medium access control (MAC). The cross-layer approach was implemented on the network layer and data layer, as shown in Figure 11.

The key parameters considered to evaluate the performance of FANET are Packet Success Rate (PSR), Throughput (TP), End-to-End (E2E) Delay, and Packet Loss Rate (PDR). The simulated results using NS2 show that the CLEA-AODV protocol outperforms the standard AODV protocols in terms of PSR, TP, E2E delay, and PDR.

3.1.4. Hybrid Routing Protocol

Hybrid routing protocols (HRP) are based on the combination of proactive and reactive routing methods. HRPs combine the best features of both routing protocol systems. HRPs overcome the message overload problem in proactive routing and the delay caused by reactive routing. These hybrid protocols are particularly suitable for large networks based on the division of the network into subnets or regions. Therefore, a proactive routing protocol operates in each region, and a reactive routing protocol is used for communication between regions. Two of the prominent HRP-based routing protocols are Zone Routing Protocol (ZRP) [45] and the Temporally-Ordered Routing Algorithm (TORA) [46]. The prominent variants of these protocols will be described in detail below.

In the ZRP regional routing protocol, each UAV is assigned to a separate operating region, and these regions may overlap with the regions of nearby UAVs. Data packets within a single region are routed using intra-region routing, while data packets between regions are routed using inter-region routing. Intra-region routing is implemented through proactive routing to maintain paths, and inter-region routing uses reactive routing mechanisms to maintain and find optimal paths. In [47], the authors propose a location-assisted delay-tolerant routing protocol (LADTR) for UAV networks for use in post-disaster operations. Exploiting location-assisted forwarding is combined with store-and-forward (SCF) techniques. Future UAV locations are estimated based on the location and velocity information of GPS-integrated UAVs. The forwarding UAV node predicts the location of the destination UAV node and then decides the forwarding route.

Relay UAVs improve the availability of links between the search UAVs and the ground station, thereby reducing end-to-end delay and increasing the packet delivery ratio. In the simulation results of NS-3, LADTR shows its superiority over AODV, GPSR, Spray-and-Wait, and Epidemic routing protocols in terms of packet delivery ratio, average end-to-end delay, and normalized routing cost.

The TORA time-ordered routing protocol allows each UAV to influence only the routing tables of its closest neighbors. The main advantage of this protocol is its ability to reduce the reaction to topological changes. Furthermore, it eliminates invalid paths and searches for new paths in a single run of the distributed algorithm. TORA mainly uses reactive routing, but in some cases, it also uses a proactive approach. These routing protocols construct and maintain a directed acyclic graph (DAG) from the source UAV to the destination. Multiple paths are created between UAVs, based on the DAG to transmit packets. TORA is chosen to quickly compute updated paths in case of disconnected paths and to enhance adaptability. To reduce the signaling cost, TORA often chooses longer paths instead of the shortest path method. The authors in [48] proposed a fast reconstruction time-order routing protocol (RTORA) based on the original TORA protocol. RTORA applies a cost-reduction mechanism to overcome the adverse effects caused by link reversal errors in TORA. The simulation results using OPNET to compare the RTORA protocol with the original TORA protocol show that the route reestablishment time of RTORA is reduced by about 25%, while the control cost is reduced by about 56%.

Table 3. Prominent routing algorithms based on network structure.

Type	Ref.	Algorithm	Main Contribution	Advantage	Limit
Static routing	[36]	LCAD	LCAD-based full-duplex forwarding	Establish flight paths from the ground preparation stage before the UAVs perform their missions.	Fixed route; lacks flexibility when the environment changes
	[37]	DCR	Special applications in identified UAV deployment environments	Static multipoint routing for improved performance for cluster-based UAV networks	Difficult to adapt to dynamic changes due to network structure requirements with predefined clusters
	[37]	UAS	Applications to support fire, disaster, or search and rescue	UAS provides a distribution path to connect publishers to subscribers. Efficient in networks deploying multiple UAVs	Architecture-dependent; inefficient when publishers/subscribers change greatly
Proactive routing	[38]	PRP	Support for applications with real-time requirements	PRP allows fast access to packet routing paths. Fast routing in complex and dynamic environments	Maintaining connectivity becomes difficult in cases where network nodes move rapidly and constantly change positions
	[38]	ISH-DSDV	Application in UAV networks based on Nomadic mobility model	ISH-DSDV optimizes data movement. Suitable for dense network deployment; controlling energy imbalance problem	Energy imbalance due to centralized control; performance degradation in sparse networks
	[42]	SROLSR	Applications in UAV networks	High performance; solving the redundancy problem in MPR selection algorithm	Requires correct network information
Reaction-based routing	[43]	DSR	Traffic monitoring application via UAV networks	DSR does not require a fixed network infrastructure but allows the network to self-organize and self-configure.	High route discovery cost; increased latency in large networks
	[44]	AODV	Traffic monitoring application via UAV networks	AODV provides routes based on routing history or will initiate a new route request process when no route exists.	High route discovery latency; prone to congestion; no QoS guarantee for real-time applications
	[44]	CLEA-AODV	FANET-based UAV system developed to support search and rescue	Network performance improvement based on: routing with AODV protocol, GSO-based cluster head selection, and Collaborative Media Access Control (MAC)	Complex due to integration of multiple mechanisms; performance depends on GSO parameters
Hybrid routing	[47]	LADTR	Applications in UAV networks to support post-disaster operations	Future UAV locations are estimated based on location and speed information of GPS-integrated UAVs	Depends on accuracy of UAV position and speed estimates; less effective in GPS noisy environments
	[48]	RTORA	Applications in highly dynamic UAV networks	Quickly re-establish a new line; shortens route re-establishment time	Route re-establishment costs are high; when there are many changes, requires updated network information

summarizes the advantages and limitations of some prominent routing algorithms based on UAV network architecture, including their classification and a description of their roles in specific applications.

3.2. Location-Based Routing Protocol

In UAV networks, due to the high mobility of nodes in the network, it is difficult to maintain routing tables. At the same time, traditional routing protocols incur significant costs by repeatedly finding routes before sending packets. Therefore, routing protocols based on geographic location information have been studied to address these challenges. The advantages of geolocation-based routing protocols are more suitable for highly dynamic networks, including UAV communication networks. Some prominent studies for location-based routing solutions are described in detail below.

The Greedy Perimeter Stateless Routing (GPSR) protocol is a routing protocol that relies on the location [49] of routers and the destination of packets to make packet forwarding decisions. By keeping only the local topological state, GPSR scales better in a per-router state than the shortest path routing and arbitrary routing protocols as the number of network destinations increases. Under the condition of frequent network topological changes of mobility, GPSR can use local topological information to find new accurate routes quickly. Based on the well-known GPSR protocol, the authors in [50] developed a GPSR with Position Prediction and Uncertainty (GPSR-PPU) algorithm, which works by adding location prediction and uncertainty to the well-known GPSR algorithm. GPSR-PPU improves the selection of the next-hop node in highly mobile and noisy UAV environments. The simulation results show a 33% improvement in packet loss, a 12% reduction in overhead, and a 42% reduction in delay. Geographic Load Sharing Routing (GLSR), proposed by the authors in [51], is an extension of GPSR. In the GLSR algorithm, when forwarding packets to a given destination, a node considers not just one but a set of next candidates, and allocates traffic among them based on queue dynamics, thus making the path more reliable. In addition, load balancing is performed using a congestion-aware handover strategy.

In [52], the authors developed another geolocation-based routing protocol, RGR (Reactive-Greedy-Reactive), which takes advantage of the readily available location information of UAVs. The RGR algorithm is a combination of the mechanisms of Greedy Geographic Forwarding (GGF) and reactive routing. The proposed RGR uses the location information of UAVs, as well as reactive end-to-end paths, in the routing process. The simulation results show that RGR outperforms AODV protocols in UAV tasks in terms of criteria such as latency and packet delivery ratio. During the operation of RGR, GCF is triggered when the established route to the destination is suddenly interrupted.

Since the UAV devices in the network move dynamically, they lead to rapid topological changes. Existing geolocation routing protocols need to send beacon packets (or hello packets) frequently to maintain the accuracy of routing selection. However, high beacon frequency means high overhead, leading to collisions with data packets and transmission delays. To overcome this problem, the authors in [53] proposed an Adaptive Beacon Prediction Protocol, called ABPP, that allows flexible adjustment of the beacon frequency and predicts the future location of the UAV. Figure 12 illustrates the situation in which the UAV (S) chooses the current nearest neighbor UAV or the UAV after a period of time t. To move to the UAV (D), the routing process is deployed through UAVs (M), (K) and (N). The experimental results show that the proposed ABPP. scheme can effectively reduce the beacon cost and improve the packet delivery ratio.

Communication between UAVs is unreliable and often disrupted due to the high mobility of UAVs. The ability to predict the behavior of UAVs allows the development of scenario-based routing protocols. However, data transmission requires a complete routing path, which is difficult to implement due to the prediction of possible disconnections. The improved OLSR routing protocol proposed in [54] allows location-based modification to address the high mobility of UAV networks by taking into account both energy efficiency and frequent topological changes. The proposed protocol takes into account not only energy efficiency by considering the remaining energy and node degrees of the UAVs but also frequent topological changes by incorporating the 3D position information of the UAVs to calculate the link expiration time. The simulation results using NS-3 to evaluate the performance between OLSR and existing standard protocols such as “P-OLSR” and “Improved OLSR-ETX” showed that improving OLSR helps increase packet delivery ratio by 8% and 13.4%, and average throughput by 4.46% and 19.67%. In addition, routing cost is reduced by 6.47% and 4%, and end-to-end delay is reduced by 18.99% and 25.10%.

Table 4. Prominent location-based routing algorithms.

Type	Ref.	Algorithm	Main Contribution	Advantage	Limit
Location prediction	[50]	GPSR-PPU	Application in UAV networks to support firefighting	Improved next-hop node selection in highly mobile and noisy FANET environments	Depends on location accuracy.
Geographic load sharing	[51]	GLSR	Solving the bottleneck problem at Internet gateways in airline networks	Exploit aircraft position information (based on GPS) together with buffer size information to exploit the full A2G capabilities available and optimize throughput in the network	Performance depends heavily on node density and node distribution; in environments with many empty areas or unevenly distributed nodes, GPSR may not be effective
Geolocation combined with response routing	[52]	RGR	Application in Unmanned Aerial Ad-hoc Networks (UAANET)	The proposed RGR uses the UAV location information as well as the reactive end-to-end paths during routing to minimize the delay in the network	Depends on location information and response route
Based on prediction	[53]	ABPP	Application in FANET networks with fast and flexible moving nodes	ABPP. can flexibly adjust the beacon frequency and predict the future position of UAV and the accuracy of routing choice.	Performance depends on the accuracy of location prediction; a trade-off between beacon frequency and cost is required
Based on 3D location	[54]	OLSR	Application in UAV networks with fast and flexible moving nodes	Optimize energy efficiency by considering the remaining energy and node level of the UAV and frequent structural changes (based on the 3D position information of the UAV)	High computational cost; requires accurate and continuously updated 3D location information

summarizes the advantages and limitations of some prominent routing algorithms based on the location of nodes in the UAV network, including their classification and roles in specific applications.

3.3. Swarm-Based Routing Protocol

Swarm-based routing protocols are inspired by the behavior of animals and nature, and are applied in solving various problems in UAV networks [6], especially in establishing communication between UAVs. These protocols are based on the behavior of insects such as fish, ants, bees, and particle swarms. In this paper, some prominent swarm-based routing protocols are described in detail below.

The ant colony optimization-based polymorphism-aware routing algorithm (APAR) is proposed by the authors in [55]. The APAR algorithm is a powerful tool for solving optimization problems that can be applied in UAV network application models. In the investigation phase, the routing paths are determined based on the “pheromone” concentration and the travel time of the sent packets. In addition, this type of protocol takes into account the congestion on the network (measured by buffer occupancy and channel load) and the reliability of the routing paths (measured by mobility and connectivity) while making the selection. When a UAV wants to communicate with another UAV, for example, UAV (D), as shown in Figure 13, the UAVs communicate using the route request RREQ.ant and route reply RREP.ant to find the most connected routing path. The simulation results show that the APAR algorithm outperforms traditional algorithms in terms of packet delivery ratio, end-to-end delay, routing cost, and reliability in the environment.

An improved ant colony algorithm based on fuzzy logic memory mechanism (FLM-IACO) to obtain lower-cost flight paths for multiple drones is presented by the authors in [56]. The proposed solution is a multi-UAV cooperative route planning model guided by mission requirements, considering low-altitude obstacles and enemy detection threats. In addition, a mixed pheromone structure delimiting infeasible regions is identified during the path-planning process. The reliability of the next nodes is evaluated using membership degrees and a layered expansion, and a correction strategy is used to refine the flight paths. The results demonstrate that the method consistently achieves shorter, safer, and more energy-efficient flight paths in the same number of iterations.

A bee swarm routing algorithm applied to UAV networks (BeeAdhoc) was also proposed by the authors in [57]. The bee population can be divided into two factions, including observer bees and forager bees. The proposed BeeAdhoc algorithm allows multiple communication paths between UAVs to be established in the first phase (reconnaissance phase) by broadcasting forward and backward scouts, including the source ID, hop count, and minimum remaining energy. In the second phase (resource collection phase), the UAV will collect information packets from the origin to the destination (Figure 14). When UAV (S) wants to communicate with UAV (D), it will do so by sending a “scout” through the network. The multiple routes found are used to send the reconnaissance mission back to UAV (S). The data are sent along the shortest path formed by high-energy UAVs.

In [58], the authors proposed an improved artificial bee swarm routing algorithm, based on multi-strategy aggregation (IABC), to generate suitable paths for UAVs. The proposed IABC coding method is based on the definition of path nodes using cubic spline interpolation. It is applied to construct the collision-free fitness function of UAVs, which avoids the generation of non-flying paths. The IABC is based on a hybrid mechanism of chaotic mapping and the Pareto principle to initialize the population, enabling a full search of the solution space, and provide the approximate optimal flight path for UAVs. In addition, two new search methods are designed to generate solutions that provide superior flight paths for UAVs. The tangent stochastic evolution mechanism is added to enhance the performance of the applied model, maximize the quality of flight path generation, and effectively solve the path planning problem. In the simulation results, IACB is compared with Particle Swarm Optimization (PSO), Artificial Bee Colony (ABC), Search Artificial Bee Colony (SABC), and Gbest Artificial Bee Colony (GABC) algorithms, showing that IACB has good feasibility and efficiency in solving the UAV path planning problem.

Table 5. Prominent swarm-based routing algorithms.

Type	Ref.	Algorithm	Main Contribution	Advantage	Limit
Ant colony optimization	[55]	APAR	Deployment in different application models of UAV networks	Route selection is based on calculations of perceived distance, congestion level, and route stability	Requires complex computation to maintain network information; latency can increase in dynamic networks
Improved ant colony	[56]	FLM-IACO	Reducing routing costs for UAV networks	Ability to find flight routes for UAVs at lower cost; safer and more energy efficient, even in complex terrain and high-threat environments	High computational complexity due to integration of fuzzy logic memory mechanism; performance depends on accuracy of interaction model and parameter setup
Bee colony	[57]	BeeAdhoc	UAV networks with high node mobility; dynamically changing network structure and movement in 3D space	Optimize FANET routing by adapting to high dynamics, improving efficiency compared to traditional algorithms	Depends on node density and distribution; performance degrades in sparse or highly mobile networks
Improved artificial bee swarm	[58]	IABC	Path planning in complex urban environments	Ability to create smooth, optimal flight paths for UAVs in complex urban environments by combining advanced optimization strategies	Complex in designing objective function and tuning parameters; requires large computational resources

summarizes the advantages and limitations of some prominent swarm-based routing algorithms in UAV networks, including their classification and a description of their roles in specific applications.

4. UAV Communication Network and Technology

4.1. Communication Solution Models

• Direct communication between UAV and control station

Each UAV is connected to transmit data directly to the control station based on a star network structure [59] (Figure 15). This structure allows the network to maintain operation when one device fails by redistributing tasks among the remaining UAVs.

However, this system has some limitations, such as requiring bandwidth capacity for each UAV separately. Therefore, when the UAV fleet expands, the bandwidth demand will increase accordingly. At the same time, in a network model with only one control station, serious incidents can occur if the station fails or is attacked.

• Satellite communication

This method has a limited transmission/reception distance of only 1500 km in both ground and sky transmission (Figure 16). In certain missions, UAVs must operate in special areas. They are not able to communicate directly with control stations. Therefore, to ensure that the requirement of continuous connectivity in the UAV network is maintained, the satellite can act as an intermediate node to control the UAV.

The advantage of this method is the ability to promote and exchange important information within the UAV network and distribute the collected data to the ground control station even when it is far apart [60].

• UAV communication over mobile networks

The application of a cellular-based model is considered a safe and reliable approach for many civil and military applications. In this model, cells use different frequency bands to communicate with each other [50]. When cells are linked in a specific area, a signal coverage area is created, allowing UAVs to communicate directly with each other or with the ground control station (Figure 17).

However, deploying cellular networks can be expensive, especially in areas with complex terrain. In addition, the system is vulnerable to attacks at fixed control stations, leading to partial or complete loss of control of UAVs in the network.

• UAV Communication over Ad Hoc Network

This model works on the principle of centralized management, where UAVs communicate with each other temporarily. The network structure can be organized as a flat or hierarchical network. In each group, some UAVs act as gateway nodes to transmit/receive data in the network to the ground control station (usually, the ground control station will not participate in internal communication between UAVs unless there is a special request). According to [61], in case of ground control station failure, the network can be fragmented, leading to the interruption of communication between UAV groups (Figure 18).

• Communication between UAVs

UAVs maintain communication by sending specialized data packets to meet the requirements of different missions. Due to the limited transmission range, multi-hop communication methods are used, based on multiple UAVs. Line-of-sight (LoS) communication is often used for UAV-to-UAV communication in unobstructed airspace to extend coverage in a specific area. In complex environments such as urban or mountainous terrain, wireless LoS communication solutions are preferred [62].

• UAV communication to ground control station

Ground control stations are often permanently installed to perform UAV management [63], through which different types of UAVs can be connected and communicate with each other. Some UAVs are specially designed to interact with ground control to reduce traffic congestion and increase connectivity.

4.2. Communication Protocols in UAV Networks

In recent years, a large number of research studies have focused on various aspects of UAV communication networks, leading to technological improvements and enhanced performance of UAV networks. Communication protocols can be used to transmit data between UAV devices based on Air-to-Air (A2A) and Air-to-Ground (A2G) methods (Figure 19) [64]. However, a major challenge for UAVs is the large operating area, often beyond the Line-of-Sight (LoS) range, so communication protocols in UAV networks need to ensure appropriate technological solutions, ensuring acceptable capacity and quality of service (QoS).

To meet the above requirements, modern mobile technologies such as LTE, 5G, and 6G have been proposed for UAV networks due to their advantages in terms of broadband connectivity, low latency, and wide coverage. These technologies are suitable for UAVs operating over large areas and can support applications requiring large data transmission or continuous connectivity in mobile environments. Figure 20 shows the evolution roadmap from 1G to 6G, combining breakthrough technologies such as AI, virtualization, and high-speed communications. They create the foundation for a flexible, high-performance, and sustainable UAV network. Advanced technologies such as cell-less networks eliminate coverage limitations, while energy efficiency solutions optimize UAV operations. In addition, the development of technologies such as artificial intelligence (AI/ML), blockchain, VLC, and quantum computing will improve the operational efficiency of UAV networks with high-speed, secure, and stable transmission capabilities.

In addition, WiFi is also a notable factor in building UAV communication networks. WiFi can play an important role in applications with a smaller operating range or in situations where temporary and cost-effective connectivity is required [65]. With the development of new WiFi standards such as WiFi 6 and WiFi 6E, many opportunities are opening up for WiFi deployment in UAV networks (Figure 21). The following is a review of the most commonly used protocols in UAV networks, emphasizing the processing capabilities and dynamics of the network structure.

• WiFi

Most commercial UAVs use WiFi (IEEE 802.11) for communication as a low-cost, scalable solution. Although WiFi’s throughput (theoretically ranging from 54 Mbps for 802.11a to 2.4 Gbps for 802.11ax) is relatively lower than LTE and 5G, it is sufficient for most applications, including real-time high-resolution video streaming.

In [65], an application solution was proposed to support forest fire suppression operations. Forest fires affect rural or suburban areas where the coverage of conventional networks (e.g., cellular networks) is weak or absent, or where fires are likely to occur. Therefore, UAVs can be used to establish a relay network between the command post managing the operations and firefighters, providing situational awareness to rescue forces. Firefighter mobility is a major challenge, as the UAV swarm must maintain both coverage and end-to-end connectivity to the Base Station (BS) of the command post. LoRaUAV, a UAV system based on LoRaWAN and Wireless Fidelity (WiFi), is designed to address these issues. In the LoRaUAV system, a WiFi mesh network of autonomous UAV LoRaWAN gateways is dynamically deployed to provide LoRaWAN coverage during firefighting operations. This network of flying LoRaWAN gateways receives data from the Ground Nodes (GNs) and forwards the data via WiFi to the control station. WiFi technology has a higher data rate. It allows aggregation of traffic originating from multiple GNs. The authors propose a two-layer network system called LoRaUAV based on an ad hoc WiFi network. The core of the system is a fully distributed mobility algorithm that periodically updates the UAV structure to adapt to the movements of the ground nodes. LoRaUAV has been successfully deployed in NS-3, and its performance has been evaluated in wildland firefighting scenarios using packet reception ratio (PRR) and end-to-end delay as key performance metrics. The Connection Maintenance and Resumption (CRM) and Motion Prediction (MP) mechanisms implemented in LoRaUAV effectively improve PRR, with the only drawback being higher latency affecting a small percentage of packets due to buffering delays and disconnections.

A UAV-enabled WiFi Direct network architecture (depicted in Figure 22) to improve throughput, device connectivity, and energy efficiency in 3D space and constrained areas with boundaries is proposed in [66]. In this, a UAV equipped with a WiFi Direct group owner (GO) device, called a Soft-AP, is deployed in the network to serve a set of WiFi stations. The authors proposed to use a simpler but more efficient algorithm for optimal UAV placement. The proposed algorithm dynamically places the UAV in the network to reduce the distance between the GO and the client devices. The expected benefits of the proposed scheme are to maintain the connectivity of the client devices. This aims to increase the overall network throughput and improve energy efficiency. Real simulations are performed with NS-3 to validate the claimed benefits of the proposed scheme. The simulation results showed large improvements of 23% in client connectivity, 54% in network throughput, and 33% in energy consumption when using a single UAV, compared to the case of a fixed or randomly moving GO.

• LTE

Traditionally, UAVs operate on licensed radio frequency bands (2.4 GHz), where the UAV’s operating range is limited to the LoS range. Therefore, to fully exploit the potential of UAV devices, it is necessary to establish beyond line-of-sight (BVLoS) connectivity [67]. Recent years have seen an increase in the use of terrestrial LTE networks to provide connectivity to UAVs. LTE systems provide airborne connectivity beyond the LoS communication range for UAVs. In [68], LTE is used for scenarios where UAVs operate as a BS transmitting in the downlink (DL) or a UE transmitting in the uplink (UL). The paper highlights that current LTE networks require significant modifications for the seamless integration of LTE-enabled UAVs. The most notable drawback to using LTE and other cellular systems is the need to register the drone transmitter with the service provider. It increases operating costs and limits drone operations to areas covered by the service provider.

In [69], the authors performed a comprehensive quality of service analysis taking into account the throughput, packet loss rate, and latency provided for the integration of LTE and UAV through simulation. The simulation layout of UAV-GCS communication over an LTE network was implemented at Cranfield University. The UAV flew in a schematic diagram along the runway of Cranfield Airport. The research results can help design a UAV system using LTE to transmit 1080p video with low latency. In addition, the simulation results also showed that the UAV altitude greatly affects communication performance. At low altitudes (25 m), the signal is more disturbed, especially in Urban Micro (UMi) environments due to the shielding from buildings. At 100 m altitude, the performance improved significantly, and further improvements were made possible by adjusting LTE parameters.

In [70], the authors developed a prototype LTE-based control system for small UAVs at low altitudes and studied the feasibility and performance of the drone’s mobile connectivity at different altitudes, with measurement parameters such as delay, handover, and signal strength. The measurement results showed that when the flight altitude from the ground increased to 170 m, the received signal power and signal quality level decreased by 20 dBm and 10 dB, respectively. The downlink data rate decreased by 70%, and the latency increased to 94 ms. It can be concluded that although the current LTE network can provide the minimum requirements for the drone’s mobile connectivity, further improvements are needed to enhance the aerial coverage, eliminate interference, and reduce network latency.

• 5G Network

Similar to LTE, 5G technology is also considered for UAV network communications when bit rates higher than 2.4 Gbps are required. UAV-enabled communications have several promising advantages, such as facilitating on-demand deployment, high flexibility in network reconfiguration, and enabling long-range LoS communication links. In some cases, UAVs act as 5G radio stations (also known as APs) to extend the coverage of 5G networks to ground users, especially for sensitive applications such as public safety and disaster management [71,72].

In [73], a swarm of UAVs is connected to a cellular network, enabling the UAVs to autonomously coordinate their activities and cooperate to complete a given task. This study focuses on the problem of cooperative and communication-aware UAV channel scheduling for data transmission from a set of target points of interest (PoIs) to a cellular base station (BS). A novel cooperative multi-hop communication model based on the Mode 4 C-V2X (Cellular-Vehicle-to-Everything) cellular side-link (PC5) air interface is presented for efficient UAV data flow scheduling in the swarm. The model design aims to optimize cellular communication over UAV-to-UAV (U2U) and UAV-to-Infrastructure (U2I) links through a novel interference-aware scheduling approach, thereby envisioning a novel mobile UAV-to-Everything (C-U2X) communication model (Figure 23). The extensive simulation results demonstrate that the proposed distributed algorithm can extend the coverage of mobile infrastructure while improving multipoint communication by automatically adapting to network conditions.

In [74], the authors present how UAV BSs and drone UEs can be integrated into 5G systems. The authors highlight important open issues in the standardization process, either through the adoption of existing standards or by providing modifications towards further improvements that could realize the vision of 5G-enabled UAVs.

• 6G Network

The demand for higher throughput and a larger number of devices never stops, and 6G is on its way to serve these requirements. 6G, the next generation of wireless communications, is expected to provide intelligent, secure, reliable, and unlimited connectivity at speeds 100 times faster than 5G [75]. Hence, the diverse requirements of UAV networks, such as low latency, reliability, and energy efficiency, will be better served by 6G networks, and airborne nodes will be an integral part of 6G networks. In addition, network intelligence is envisioned as a key feature of 6G, which can support a variety of applications.

In [76], the authors consider a network of unmanned aerial vehicles (UAVs) over a cellular network, where UAVs act as aerial users to collect various sensor data and send the collected data to the transmission destination over cellular links. Different UAVs have different communication requirements due to their sensing applications, thus requiring a more flexible communication solution. To address this issue, the authors propose a UAV-to-Everything (U2X) network. It allows UAVs to adapt their communication mode according to the requirements of their sensing applications. A mobile UAV network typically consists of a BS, multiple UAVs, and multiple ground UEs. UAVs collect data from sensing targets and transmit the collected data to the BS or ground UEs to support various applications (Figure 24). Sensing tasks are performed in two steps: UAV sensing and UAV transmission. The authors consider the use of U2X communication in the UAV mobile Internet, enabling UAVs to support multiple sensing applications at high transmission rates. At the same time, a basic model of the UAV mobile Internet is proposed, introducing three transmission modes (U2N, U2U, and U2D) for U2X communication. To reduce the computational load on the BS, some of the sensing data can be transferred to cell edge UAVs with computing capabilities using the U2U mode. After data processing, cell edge UAVs transmit the data to the target UEs directly via the U2D mode. MEC with U2X communication reduces the computational pressure of the BS and enhances the QoS transmitted to cell edge UEs by reducing the transmission distance.

The application of sub-terahertz (<THz) transmission in future 6G networks has the potential to create synergies with the ubiquitous deployment of nanodrones in various applications. However, applications that deploy a swarm of UAVs for long periods of time, combined with wireless charging and the ability to efficiently guide and control individual nano-drones with limited batteries, are critical. In [77], the authors explore the convergence of sub-THz communications, radiated wireless power transfer, and 3D radar imaging in future 6G networks. Based on this, they propose a method using a sub-THz on–off keying (RZ-OOK) signal modulation scheme to transmit power and joint radar imaging (JPTRI), while maintaining ultra-low latency communication to control the UAV swarm. The results demonstrate that the proposed JPTRI method remains highly efficient as the number of UAVs in the swarm increases, maintaining continuous ultra-low latency control, consistent average DC power, and improved 3-dimensional positioning of the UAV swarm.

5. Control Algorithms for UAV Networking

5.1. Centralized Control Algorithm

In UAV swarm systems, centralized control is a common approach, in which a single point of control, usually a ground control station or a master UAV, plays the role of collecting and analyzing information, making decisions, and issuing commands to all member UAVs. This model helps ensure consistency in operational strategy, allowing all UAVs to coordinate with each other according to a common plan without having to make independent decisions [78]. Figure 25 illustrates an example of a centralized control system, where all UAVs in the formation receive commands directly from a single control station. This model is often applied in surveillance, search and rescue, or military operations, where a high level of precision and coordination between UAVs is required.

5.1.1. Leader–Follower

The leader–follower method is one of the most popular control strategies applied in UAV swarm control. This method uses a hierarchical structure, in which one or several UAVs are designated as leaders, who are responsible for guiding the movement of the entire system. The remaining UAVs act as followers, maintaining their positions and distances relative to the leader UAV according to predetermined rules (as shown in Figure 26). This model reduces the complexity of swarm control by breaking down the coordination problem into simpler tasks. Each UAV only needs to focus on following the leader UAV or other UAVs in the formation. This makes the system stable, flexible, and easy to expand in actual deployment.

Graph theory is applied in the leader–follower method to maintain the formation and ensure the performance of the UAV group. In this, each UAV is viewed as a vertex and the relationship between the leader and the follower is represented by a directed edge. In [79], the UAVs in the formation maintain a fixed distance between pairs of UAVs to ensure that the entire formation moves as a rigid block. Graph theory is used to determine the pairs of UAVs that need to maintain distance to preserve the formation shape. In addition, decentralized control laws are proposed to restore the shape when the distance between the UAVs is changed. A typical example studied is a group of three UAVs flying in an equilateral triangle formation, with the center of gravity moving along a certain trajectory while maintaining a stable velocity. In [80], graph theory is applied to quantitatively evaluate the effectiveness of UAV formations in performing missions. Performance criteria are analyzed for different UAV connectivity models, including fully connected, star, circular, tree, generally connected, mixed, and cellular networks. In [81], the scalability of a UAV formation control system is studied to maintain stability as the number of UAVs changes. A biologically inspired method, called the Veteran Rule, is proposed to ensure that the formation remains stable without re-adjusting the control parameters. The method also considers the fault tolerance of the system against unwanted connections, and it uses the Gershgorin Circle Theorem to set an upper bound on the strength of unwanted links.

In some studies, quadratic sliding mode control (SMC) and quadratic linear regulation (LQR) are combined to control UAV formation. LQR is responsible for controlling the position in the outer ring, and SMC ensures stability and reduces errors in the inner ring, allowing the UAV to closely follow the leading UAV in both horizontal and vertical planes [82]. Another method uses SMC combined with adaptive PID control to handle communication delay, ensuring a stable UAV formation during flight, while reducing noise and state errors [83,84]. In addition, to accurately track the trajectory of the leading UAV, a control method based on continuous quadratic sliding mode is developed, which helps the UAV to maintain a stable position and Euler angle by combining two-stage control [85]. Another study used the time-scale separation method between translational and rotational dynamics to control the UAV according to the leader–follower model, which helps the UAV track the trajectory smoothly and maintain a suitable distance [86]. For fixed-wing UAVs, a waypoint-based geometric control method is proposed to ensure that the UAVs can quickly converge into the desired formation. This method also integrates a speed adjustment mechanism to maintain the correct formation during flight and formation changes [87]. In addition, it uses an exponential disturbance observation model combined with adaptive backstepping control and approach speed-based glide control. This helps the UAVs compensate for aerodynamic disturbances and avoid collisions between UAVs within the formation [88]. Several test systems have been built, including quadrotor UAVs, UWB positioning systems, ground control stations, and wireless communication systems. The control algorithms have been tested on the MATLAB platform and on real hardware such as Pixhawk with the ROS operating system. These tests demonstrate the feasibility of the leader–follower method in UAV swarm control [89,90].

5.1.2. Virtual Structure

The virtual structure model helps UAVs maintain precise positions in formation and improve synchronous operation without depending on a specific leader UAV. This method is widely applied in UAV control problems to optimize coordination, maintain formation, and avoid collision.

Several studies have proposed dynamic virtual structure-based flight formation control for fixed-wing UAVs, in which the dynamic formation trajectory generation algorithm helps the UAVs change formation smoothly along a predetermined flight trajectory. This method ensures that the UAV formation can be maintained accurately in complex maneuvering situations [91]. Another study uses a flexible virtual structure. It allows the UAV to adjust its flight direction easily without being constrained, as in rigid geometric virtual structure methods. The simulation results on a six-degree-of-freedom (6DOF) UAV show that the UAV can follow the desired trajectory stably [92].

Figure 27 illustrates the working principle of the virtual structure in UAV formation control. This process starts with the virtual force field acting on the virtual structure, determining the overall position of the UAV formation. The virtual structure then adjusts to match the position of the UAV, and the UAV also changes its position to match the virtual structure, creating a two-way interaction between them. The position of the UAV is influenced by the environment, and it adjusts according to the virtual structure to maintain the desired formation. This model helps the UAV maintain alignment, adapt to surrounding conditions, avoid obstacles, and easily perform formation transformations by changing the virtual structure [93].

In environments with interference or limited communication, a single-line virtual structure approach has been developed so that UAVs can maintain formation even in the absence of real-time communication between UAVs. By updating the position of the virtual leader UAV in real time and using geometric constraints, the UAV can calculate the desired position and adjust the flight direction based on the roll angle [94]. The combination of a virtual structure with a nonlinear control algorithm overcomes the drawback of relying on the leader UAV in the Leader–Follower method. This approach helps the UAV formation maintain a tight position even in the presence of interference or during the execution of a turn [95].

In addition, a behavioral flight control method incorporating virtual structures was developed to reduce the dependence on the quality of wireless communication in the UAV formation. This approach helps the UAV maintain a stable formation in an unknown environment while enhancing the ability to avoid obstacles and potential threats [96]. The virtual structure is integrated with a nonlinear MPC algorithm, where each UAV is equipped with its own controller to optimize the distance from obstacles and other UAVs in the formation. This model helps UAVs to automatically adjust their positions by solving the optimization problem according to a sequential quadratic program, ensuring stable control performance even in three-dimensional space [97].

In addition, to maintain large-scale UAV formations, a consensus algorithm incorporating virtual structures has been studied to ensure uniformity among UAVs in the formation. This method uses a consensus strategy to optimize the state of each UAV, maintaining the formation structure during flight and minimizing trajectory errors [98].

5.2. Decentralized Control Algorithm

Decentralized control algorithms are an important method in UAV swarm control. They allow each UAV to manage itself and make decisions based on local information without the need for central coordination. This method is usually based on three main mechanisms: inter-UAV communication, ambient sensing, and direct line of sight. The UAVs in the swarm coordinate with each other to maintain formation, avoid collisions, and complete common missions, ensuring high flexibility, scalability, and adaptability (as shown in Figure 28).

5.2.1. Behavioral Based

In UAV formation control, a decentralized behavior-based approach has been shown to be an effective approach. It enables UAVs to operate flexibly in dynamic environments without the need for centralized control [96]. This approach allows each UAV to make decisions based on information from neighboring UAVs and the surrounding environment, instead of relying on a leader UAV or a global control system. Figure 29 illustrates a decentralized behavior-based control approach for UAVs. The desired behaviors are listed in separate groups, including moving to the destination, avoiding obstacles, and maintaining the formation. These behaviors represent important goals that the UAV needs to accomplish during its operation. These behaviors are then fed into the behavioral coordinator, where they are coordinated and processed to determine the appropriate control strategy for the UAV. The coordinator plays an important role in balancing different behaviors. It ensures that the UAV can move toward a target, avoid obstacles, and maintain a stable formation. Finally, information from the behavioral coordinator is sent to the UAV controller, which determines specific control commands for the UAV, such as adjusting its speed, direction, or distance from other UAVs in the formation.

The behavior-based decentralized method is a popular approach in UAV formation control. It helps UAVs maintain distance and formation shape without a single control center. Each UAV makes decisions based on information from itself and neighboring UAVs, instead of receiving commands directly from a leader UAV or a centralized control system [99]. One of the key techniques in this method is to use the weighting adjustment between behaviors, such as maintaining distance, avoiding collisions, and approaching the target, to calculate the UAV control input [100]. In addition, many studies have proposed combining this method with reinforcement learning to optimize UAV control in dynamic obstacle environments. For example, one study used a recurrent neural network to encode historical information about UAV positions and obstacles, which helps the UAV predict and adjust its trajectory more effectively [101]. Some other methods apply nonlinear consensus models to ensure that UAVs maintain formation shape even in the presence of environmental disturbances [102].

One of the important advantages of this method is the ability to maintain formation even when obstacles or changes in the environment appear. Instead of planning fixed actions in advance, each UAV can adapt to the actual situation through behaviors such as target tracking, collision avoidance, following moving walls, and adjusting position in the formation. In addition, some studies have proposed combining this method with decentralized collision avoidance algorithms. The UAV only uses the relative position information between the UAVs and obstacles to adjust the direction without requiring data from the lead UAV [100]. Furthermore, behavior-based control methods also play an important role in multi-agent systems and robot swarms. It allows robots to operate according to a simple set of rules but still achieve complex collective behavior, such as maintaining formation while moving or coordinating the execution of tasks in communication-limited environments [103]. Thanks to this ability, the method has been applied not only in UAVs but also in autonomous robot groups on land and underwater [104]. Compared with other formation control methods such as leader–follower or virtual structure, the decentralized behavior-based method has a great advantage in scalability and adaptability to changing environments. This makes it a suitable choice for many practical applications, from space surveillance, search and rescue, to UAV formation deployment in military missions [100].

Behavior-based decentralized control algorithms play an important role in enhancing the flexibility, scalability, and adaptability of UAV swarms. By making independent decisions based on sensor information and pre-established behavioral rules, each UAV can autonomously coordinate with neighboring UAVs without centralized control. This enables the UAV system to operate effectively even in complex environments with many obstacles or constantly changing conditions. However, this approach also poses some challenges, including design complexity, high requirements for coordination algorithms, and the ability to ensure formation stability in unpredictable situations. Incorporating advanced methods such as artificial intelligence, deep reinforcement learning, or biomimetic algorithms can help improve the performance of decentralized control systems. It opens up many potential research directions and applications for UAVs in real-world missions.

5.2.2. Artificial Potential Field (APF)

Artificial potential field (APF) is an effective method in UAV formation control, especially in path planning and collision avoidance problems. Some studies have improved APF to solve the local minima [105] and jitter problems in three-dimensional space [106]. Figure 30 illustrates the artificial potential field method in UAV formation control. In it, a UAV is shown in the center, surrounded by four neighboring UAVs, representing other members in the formation. A large circle surrounds the central UAV, representing the area of influence of the potential field. Within this area, the central UAV can sense the status of neighboring UAVs through sensors or wireless communication. The central UAV has arrows connecting to neighboring UAVs, representing the interaction between them. Each neighboring UAV shows that it is within the area of influence of the central UAV. The status of neighboring UAVs is used as input to the artificial potential field function. It determines the appropriate direction of movement to maintain the formation.

An improved APF method is proposed to help UAVs escape from local minima by adding a rotational force field and using a “leader–follower” model to maintain the desired formation while ensuring system stability through the Lyapunov function [107]. During collision avoidance, APF can be combined with the virtual structure and the leader–follower model to ensure that the UAVs maintain an equilateral triangle formation. The gravitational force directs the UAVs to the target, and the repulsive force helps to distribute the UAVs reasonably on the virtual sphere surface. It allows the UAVs to choose the optimal path to overcome obstacles [108]. Some other studies have used the mass point model with kinematic constraints to design the gravitational force field between UAVs to ensure convergence to the desired formation from any initial condition. To suit the crowded environment with many dynamic and static obstacles, the improved APF is combined with the formation division method. It supports the UAV to change formation flexibly when encountering obstacles [109]. In environment with communication networks between UAVs, a distributed collision avoidance algorithm is designed based on the improved APF and consensus theory. It assists UAVs not only in avoiding collisions but also in quickly recovering the desired formation and reaching a consensus on distance, altitude, and velocity [110]. In particular, in complex urban environments, the APF can be combined with positioning information to minimize the risk of navigation errors, helping UAVs adjust their position and altitude quickly with high reliability [111]. For high-density UAV formations, such as drone light shows, the APF-based path planning algorithm is applied to optimize the motion of multiple UAVs. It helps to avoid trajectory oscillation and improve flight performance [112]. In the UAV group model, the improved APF is integrated with the quadratic consensus algorithm, enabling UAVs from any initial state to form a formation and fly to the target, while avoiding collisions by smoothing the potential force field [113]. In addition, the APF can be integrated with the sliding mode control method to achieve simultaneous formation control and target tracking, which enables UAVs to quickly establish a formation and minimize flight shaking [114]. Some studies have also proposed a method to control UAV formations in clusters using the k-means algorithm to optimize gravitational forces between UAVs, combined with virtual cores to ensure flexible formations adaptable to complex flight conditions [115].

Artificial potential force field (APF) is a powerful and effective method for UAV formation control, especially in path planning, collision avoidance, and formation stability problems. Although traditional APF suffers from some limitations such as local minima and orbital wobble, recent studies have proposed many improvements. They include the incorporation of rotating force fields, leader–follower models, consensus algorithms, and sliding mode control (SMC). These improvements enable UAVs to not only maintain the desired formation but also flexibly adapt to complex environments, such as urban areas, crowded spaces, and dynamic obstacle conditions. Furthermore, integrating APF with distributed control models and cluster optimization improves the efficiency of UAV coordination, especially in missions requiring high UAV density such as light shows or search and rescue. Overall, APF remains a promising approach to UAV formation control, and its combination with modern methods such as machine learning and intelligent control promises to open up many development opportunities in the future.

5.3. Distributed Control Algorithm

Distributed control in UAV swarm systems is a control method in which each UAV operates based on information obtained from neighboring UAVs without direct guidance from a control center. This approach enhances the flexibility, adaptability, and robustness of UAV swarms, especially in complex environments or when there is a communication problem [116]. Distributed control algorithms often use multi-agent control theory, consensus, and formation to ensure that UAVs can maintain the desired formation structure, avoid collisions, and perform common tasks efficiently. Thanks to this model, UAV swarms can operate in many fields such as surveillance [117], exploration [118], and search and rescue [119] with high performance and good scalability.

A distributed formation control algorithm based on the non-smooth backstepping method has been developed to control multiple UAVs in a leader–follower model. First, a position controller is designed using an optimal tuning method according to the quadratic linear controller. This enables all UAVs to converge to the desired formation model. Next, an attitude control system is constructed using the finite-time control technique combined with switching control to ensure that UAVs accurately track the desired state within a finite time period to show that all UAVs converge to the desired formation in 3D space [120]. In [121], a distributed formation control for vertical takeoff and landing UAVs is proposed. The communication network between UAVs can change over time. The system uses a hierarchical control framework, in which distributed command forces are calculated based on the positions and velocities of neighboring UAVs. An attitude control model using the backstepping method is implemented to ensure that the UAVs track the desired pitch angle. A support system is designed to maintain the thrust limit, avoid singularity, and eliminate the requirement for acceleration information of neighboring UAVs. The asymptotic stability of the system is ensured through the Lyapunov method. In [122], the authors propose a distributed adaptive tracking control system to solve the UAV formation control problem under the conditions of limited inputs, actuator errors, and external environmental disturbances. The system applies an adaptive backstepping method combined with command filtering to handle uncertainty in the UAV model and input saturation problems. In addition, the fault-tolerant controller is designed to estimate and compensate actuator errors, as well as external disturbances. The stability of the closed-loop system is guaranteed even when the interaction graph is undirected. In [123], the authors propose a distributed formation control strategy developed for UAVs with three degrees of freedom, in which the information flow between UAVs is described by a directed graph and a spanning tree with the leader UAV acting as the root. Only position information is transmitted between UAVs, which reduces the data requirements. A simple backstepping method is applied to design the distributed controller. The stability analysis based on Lyapunov proves the effectiveness of this method. In [124], a new distributed consensus algorithm for nonlinear quadratic multi-agent systems (MAS) is proposed based on the leader–following model. This algorithm provides smooth control signals, avoiding the chattering phenomenon commonly encountered in traditional sliding control. In addition, a new formation control system is developed by integrating the distributed consensus algorithm into the three-dimensional geometric model. It enables the UAVs to accurately track the formation. The algorithm is highly efficient in controlling the UAV in the desired formation pattern.

Figure 31 illustrates a distributed control scheme in a UAV swarm system, in which UAVs maintain formation through communication with each other. The formation structure is organized in a triangular form, where each UAV not only receives information from a leader UAV but also exchanges data with neighboring UAVs through communication channels, as represented by dashed arrows. This allows the system to operate in a self-organizing mechanism, allowing UAVs to automatically adjust their positions based on data collected from the environment and other UAVs without the need for a fixed control center. This approach increases the flexibility and adaptability of the swarm in missions such as surveillance, search and rescue, or military operations, especially when a UAV fails or loses communication with the central control system.

Each method has its own advantages and disadvantages. Depending on the specific requirements of the task, methods can be selected or combined to achieve optimal efficiency, as shown in

Table 6. Comparison of advantages and disadvantages of UAV formation control methods.

Method	Advantage	Disadvantages
Leader–follower	- Easy to deploy and control. - Ensures accurate formation. - Suitable for highly organized missions.	- Vulnerable to instability if the lead UAV crashes. - Inflexible in dynamic environments. - Poor scalability.
Virtual structure	- Ensure stable formation shape. - High precision in control. - Suitable for UAVs with tight linkage.	- High computational complexity. - Difficult to adapt to changing environments. - Requires high synchronization between UAVs.
Behavioral based	- Flexible and adaptable to dynamic environments. - No central control required. - Suitable for exploration and obstacle avoidance tasks.	- Difficult to ensure stable formation shape. - Need to adjust parameters reasonably to avoid conflicts between behaviors.
Based on consensus	- Ensure synchronization between UAVs. - Clear mathematical model; can prove convergence. - Can be extended to large number of UAVs.	- Need to continuously transmit information between UAVs. - Can be affected by communication delay. - Need a fault tolerance mechanism when UAV is lost in the system.
Artificial potential field	- Smooth control; effective collision avoidance. - No need for complex data exchange. - Suitable for navigation in environments with obstacles.	- Can get stuck at local equilibrium point. - Difficult to adjust when there are multiple targets or dynamic obstacles.

.

In this study, we have discussed different control models, including centralized, decentralized, and distributed control. Each model has its own characteristics that affect network performance and communication mechanisms. Centralized control requires a strong connection and stable bandwidth to transmit commands from the ground control station (GCS) over datalinks to UAVs. Decentralized control relies on local decision-making. It minimizes communication requirements but poses challenges in terms of synchronization between UAVs. Distributed control models, with consensus algorithms or collective dynamics, allow UAVs to self-adjust based on local sensor data, and they optimize through adaptive routing strategies. Control algorithms also have an impact on flight system stability, tracking of desired trajectory, compensation of errors and disturbances from the environment, adaptation to model changes, optimization of operations (energy, performance), and coordination of UAV swarms.

6. AI-Based Approaches for Autonomous UAV Swarm Systems

In the context of using UAV swarms today, they are being increasingly applied in fields such as surveillance, rescue, smart cities, smart transportation, wildlife tracking, and logistics. AI-integrated solutions have become an important trend to solve complex challenges in data processing, dynamic routing, and multi-task control. Traditional methods often cannot effectively meet the requirements of flexibility in volatile environments, limited resources, and real-time computing requirements. AI is based on learning and adaptation capabilities, providing optimal solutions to these problems through three main aspects: (1) data mining and analysis from sensors, predicting events and situations; (2) adaptive routing to optimize transmission routes in dynamic UAV networks, ensuring transmission efficiency under changing conditions; and (3) multi-task control, supporting synchronous coordination of UAVs to perform tasks efficiently and accurately. We will analyze these methods in detail in the next section to clarify the connection among components in UAV systems and the scalability of AI applications in real-world applications.

6.1. AI-Enhanced Data Processing for Swarm UAV Applications

The combination of UAV networks with artificial intelligence and machine learning technologies has yielded fast and reliable results [16,125]. The use of UAV networks combined with artificial intelligence has been shown to be beneficial for data collection and processing, real-time monitoring, and prediction in various contexts [126,127]. Each application can be deployed independently or combined with the specific task requirements. We present an application-specific AI-integrated data processing framework designed to address domain-specific challenges. Our approach is based on a unified pipeline consisting of three stages: (1) multi-source data collection from sensors and cameras on UAVs; (2) processing and analysis using AI algorithms optimized for each type of data and application; and (3) integration of the results into the swarm control system. Each application will be analyzed in detail regarding data requirements, AI-applied solutions, and applications in real-world fields.

6.1.1. Intelligent Traffic Monitoring and Analysis Using AI-Based UAV Swarms

Recently, intelligent transportation techniques and intelligent transportation services have gained increasing attention and development support from government agencies. The need to optimize traffic efficiency, reduce energy costs, increase safety, and comply with traffic laws is increasingly urgent. An AI-integrated video analytics framework (TAU) has been developed to meet the needs of automated transportation [128]. The study used high-resolution images to detect and track traffic objects automatically, with six main contributions. First, a preprocessing algorithm to adjust the high-resolution image as input to the object detector. Second, an algorithm to calibrate vehicle coordinates for tracking on multiple cuts of a frame. Third, a speed calculation algorithm based on the accumulated information from the frames. The fourth is the number of vehicles in each traffic area based on the Ray Tracing algorithm. The traffic regulation at the intersection is based on the observed images in the surrounding areas. Finally, the algorithm extracts twenty-four types of detailed information from the collected raw data. This study is typical for the analysis of traffic situations using curves, charts, heat maps, and animations.

Ground traffic management in the context of urban streets and highways requires a large amount of data. In addition, access to real-time traffic information is necessary in emergency situations. This requires the traffic control center to constantly monitor the flow of vehicles and take appropriate actions to reduce traffic congestion. A method that reduces the transmission bandwidth while maintaining acceptable video quality for on-server image processing is presented in [129]. The authors have developed a framework for monitoring and controlling highway traffic using UAVs to collect data, which are then sent to the server. An object detection algorithm supported by artificial intelligence deployed on the server is capable of classifying objects and different vehicles.

An advanced machine learning algorithm called TSR-YOLO has been designed for traffic sign recognition for traffic monitoring [130]. A system capable of performing global feature extraction improves accuracy when identifying many smaller traffic signs. When compared to other machine learning methods for traffic monitoring, YOLO performs better because it scans the entire scene in a single pass, significantly improving speed without sacrificing accuracy. When compared to traditional machine learning methods, YOLO’s integration of real-time detection and classification is a major advantage [131].

Researchers are currently using a variety of datasets for traffic monitoring models. All systems have certain challenges and scope for development. Despite its potential benefits, traffic monitoring using machine learning currently has limitations, as discussed in [132]. These algorithms may give some false results when counting vehicles because they are only partially accurate [133]. The complexity of the environment increases the possibility that monitoring may give false results in extreme weather conditions. Researchers are turning their attention to generative AI in UAVs for traffic surveillance. This technology revolutionizes aerial surveillance by accurately predicting traffic patterns and congestion. It enables drones to improve real-time routes and enhance traffic management and emergency response, thereby enhancing the safety and efficiency of urban transportation using sophisticated predictive analytics derived from aerial data [134].

6.1.2. AI-Based Object Detection and Recognition Applications for UAV Swarms

To detect an object, the UAV takes pictures with its lens, and then machine learning and computer vision extract features. These algorithms are capable of detecting the size, shape, and color of an object, and the ability to recognize patterns that can determine the specific location of the object. Researchers use sensors such as synthetic aperture radar (SAR) and light detection and ranging (LIDAR) to collect images. Artificial intelligence is then used to extract information from those images to determine the location of the object of interest [135]. The SAR method allows researchers to improve the visual capabilities of unmanned aerial vehicles (UAVs) [136]. Among the mobile tools for object recognition, RGB-D cameras mounted on UAVs are also popular. Object detection based on deep learning and RGB-D was developed in [137]. The results of 75.5% accuracy and 53.33ms processing speed for object detection demonstrate the effectiveness of this combination in the CNN model. In addition, the accuracy of the prediction of object shape and location strongly depends on the relative distance between the object and RGB-D. Today, generative AI improves UAV object recognition by developing complex models that can accurately recognize and classify objects from aerial imagery. Generative AI generates diverse training data, recreates rare scenarios, and optimizes detection models to enhance UAV object recognition. It generates synthetic images of animals across habitats for wildlife tracking and it models debris patterns to improve UAV-based rescue and surveillance. This technology enables drones to efficiently and autonomously navigate complex environments, observe changes, and perform operations on objects [138].

6.1.3. Collaborative Environmental Monitoring via AI-Integrated UAV Swarms

In the UAV-WSN-IoT system, UAV devices are equipped with sensors, cameras, or specialized collection devices to be able to perform aerial surveillance, record images, or collect different types of data [26]. The outstanding features of the system are widely applied in many different monitoring fields, especially applications that require data collection from large, inaccessible areas or require continuous monitoring [139,140].

A prominent application is described in [141]. UAVs in the WSN-IoT collaborative system are equipped with many different sensors to determine water quality, such as PH, temperature, turbidity, Dissolved Oxygen (DO), and carbon dioxide. All sensors are integrated and monitored wirelessly by a universal asynchronous transceiver (UART) on the ground station. The telemetry system is designed to be able to send and receive data at a distance of up to 1 km in real time. The UAV-WSN system is also a cost-effective solution and suitable for large-area monitoring applications. In [142], the study describes the UAV participating in animal monitoring activities as a Sink node to collect important time-sensitive information from sensor devices to directly observe and track animals. In addition, this system also allows for detecting the location of endangered species in large-scale wildlife areas or tracking animal movements without any attached devices. Design strategies using UAVs to deploy wireless sensor networks are also applied to post-disaster monitoring [143]. They can be deployed to monitor and assess areas affected by disasters or support the search for victims.

Recently, the use of artificial intelligence (AI) in UAV-WSN networks has improved the efficiency of monitoring applications. AI and machine learning algorithms enable distributed data processing, improve hierarchical multi-protocol networking aspects, and control high-level UAVs constrained by WSNs. The Federated Learning-Based UAV-WSN-IoT system described in [144] leverages the lightweight Dense-MobileNet model to achieve learning from the features of haze images captured by UAVs. The results of the study demonstrate the ability to accurately and timely predict real-time and future air quality indexes.

6.1.4. Search and Rescue Operations Using AI-Equipped UAV Swarms

Rescue operations using UAVs with integrated AI technology are of great interest to experiment with. This highlights the possibility of solving complex or even impossible tasks for humans. The main goal of search and rescue missions remains to locate the target quickly and accurately, followed by critical actions such as information exchange and timely delivery, all of which must be met within strict time constraints. In [145], the authors present an advanced AI-enabled UAV swarm system with high accuracy and fast response time. The system is equipped with modern sensor technology, advanced image recognition algorithms, and autonomous navigation capabilities. The methodology of the proposed method includes multi-faceted data collection techniques, including surveying, data mining, IoT sensors, and video analytics. Machine learning and deep learning models are then applied to process and analyze these data, enabling real-time image recognition for precise target identification. AI-optimized algorithms significantly reduce response times and increase mission success rates, accelerating emergency response efforts in complex environments.

Search and rescue missions during and after disasters are often labor-intensive and costly. Quickly locating injured people helps minimize loss of life and property. The authors in [146] proposed a technique that combines behavioral-based artificial intelligence with cooperative swarm behavior. The UAV swarm is controlled by integrated sets of reactive behaviors such as collision avoidance, battery recharging, formation control, and altitude maintenance. In addition, various search methods are used to optimize the coverage of cameras and heart rate sensors mounted on the UAVs. The authors implemented a simulation of a UAV swarm (with up to 20 UAVs). The simulation scenario uses real-world location data, including post-disaster satellite imagery, real-world locations of damaged buildings, and victim locations. The results of measuring the effectiveness of the approach by recording the detection rate over time of survivors showed a 90% detection rate in less than 24 h in the simulation.

Table 7. Some prominent AI-based UAV swarm applications for data collection and processing.

Ref.	Field of Application	Main Contribution	Advantages	Limit
[128]	Traffic monitoring and analysis	Integrating AI with UAV network to automatically detect vehicles; calculate speed based on video processing.	Supports 24 different types of analysis; suitable for smart cities.	Requires high computational resources. High implementation costs.
[129]	Time traffic management	UAVs are used to collect traffic information; optimize video transmission bandwidth; combine IoT and AI.	Detect and classify vehicles; responds quickly to emergency situations; reduce transmission bandwidth.	Latency in data processing, depending on network connection.
[131]	Smart transportation	New real-time small object detection (RSOD) algorithm based on YOLOv3.	Increased accuracy (mAP@0.5 +3.4–5.1% vs. YOLOv3); real-time processing; small object detection and high density.	High computational resource requirements; performance depends on UAV image quality.
[133]	Static target search	Reinforcement learning (DNQMIX) combines Digital Twin for training in virtual environments.	High search speed and coverage Dynamic environment adaptation.	Complex in real-world deployment or for mobile targets.
[134]	Collecting data from distributed IoT devices	Based UAV swarm to optimize data collection route.	Optimize data collection; Flexible, automatic collision avoidance.	Depends on the connectivity of UAVs and IoT devices.
[136]	Data Collection	DDQN combines SARSA/Q-learning to allocate communication resources.	Increase data collection efficiency; converge faster; reduce collisions.	Limited scalability; Requires complex calculations.
[137]	Smart agriculture	CNN combines RGB-D cameras for environmental perception and flight planning.	Ability to accurately detect obstacles; automatically plan to avoid obstacles.	Accuracy depends on distance, slow processing speed.
[138]	Search and rescue	CBBA-TCC (Consensus-Based Bundle Algorithm with Task Coupling Constraints).	Support multi-mission UAVs in search and rescue; suitable for heterogeneous systems; complex constraint handling.	Difficult to scale; complexity depends on number of UAVs.
[141]	Intelligent routing in VANET	UAV supports Q-learning and fuzzy logic to optimize traffic routes.	Reduce latency, use resources efficiently.	Depends on UAVs to collect information.
[142]	Wildlife monitoring	UAV combined with WSN for animal tracking based on MDP (Markov Decision Process)	Reduce information transmission/reception delay; reduce energy consumption and optimize flight paths.	Not optimized for dynamic environments
[144]	Air quality monitoring	Graph convolutional neural network based long short-term memory model (GC-LSTM) to achieve accurate AQI inference.	Protect private data; expand surveillance scope.	Requires ground infrastructure; complex to deploy.
[145]	Search and Rescue (SAR)	AI-Enhanced UAVs with advanced sensor technology for search and rescue operations	Accurate target recognition; fast response time.	High hardware costs; requiring multimodal data processing.
[146]	Natural disaster response	Relying on AI-based UAV swarms to map disaster areas.	Integrated with many functions (collision avoidance, battery charging); detects >90% of victims within 1 h; good scalability.	Depends on actual implementation conditions

summarizes the data collection and processing applications for prominent AI-based UAV swarm applications.

The above studies show a strong development trend of AI-based UAV swarm control solutions. The advantages achieved include the ability to optimize resources and suitability for distributed systems. Technology trends show the potential of multi-technology solutions (including UAVs, AI, and IoT). However, in the actual implementation process, it is necessary to pay attention to implementation costs (due to hardware requirements of UAVs and sensors) and scalability (some methods have not been verified on a large scale).

6.2. AI-Based Routing Protocol

In recent years, the use of AI-based learning to harness the learning power of cognitive nodes in UAV networks has made significant progress. AI supports smarter networking decisions by integrating computational intelligence into UAV networks. In practical UAV network applications, due to the lack of adaptive and automatic routing decision-making capabilities, traditional routing protocols are not suitable for UAV networks with high dynamics and many unique characteristics. A key issue in UAV networks is that when the connection is lost, it can cause packet loss, frequent connection re-establishment, reduced link lifetime, and long delays. Therefore, extensive research has been conducted to develop intelligent and adaptive routing protocols based on learning, in which machine learning methods are used to model and predict the evolution of network structure, channel status, traffic mobility, and environmental factors to enhance the capabilities of UAV networks. A prominent research direction is to predict upcoming changes in network topology and minimize connectivity loss through machine learning-based routing [141]. In addition, online route planning methods combined with network algorithms can provide the ability to manage network topology with minimal connectivity disruption. This section will mainly discuss AI-enabled routing protocols designed primarily for UAV networks, including network structure-based routing methods and self-adaptive learning-based routing methods.

6.2.1. Network Topology-Based Routing Protocol

A key feature of predictive routing protocols is their use of machine learning (ML) algorithms to predict the movement trajectories of nodes in UAV networks and then combine them with path selection mechanisms. Some prominent routing protocols in recent years to improve the performance of routing algorithms for UAV networks are reviewed in detail below.

Predictive routing based on reinforcement learning: To cope with the high mobility and rapid changes in UAV swarm networks, the authors [147] proposed a predictive routing protocol driven by reinforcement learning and trajectory knowledge (PARRoT). This novel machine learning-enabled routing protocol exploits mobility control information to incorporate knowledge about the future movements of mobile agents into the routing process.

The overall system architecture of the proposed PARRoT is shown in Figure 32. PARRoT consists of three core components: PARRoT Wings predicts cross-layer mobility based on a multi-layer approach that leverages knowledge from the mobility control domain to predict relative mobility between different agents; PARRoT Chirp disseminates routing messages and context knowledge based on user data protocols; PARRoT Brain performs route maintenance tasks based on reinforcement learning. In a comprehensive simulation-based performance evaluation, PARRoT achieves robust data transmission and significantly lowers end-to-end latency, even under challenging radio propagation conditions.

Mobility Prediction-Based Virtual Routing (MPVR) [148]: the authors proposed a mobility prediction-based virtual routing algorithm, as shown in Figure 33. This scheme improves the connection efficiency and routing time between cooperative UAVs based on the probability density function of UAV movements using a Gaussian distribution model. In addition, an optimal virtual routing model is designed to select the optimal relay node with the smallest distance between UAVs. The simulation results show that MPVR can improve the average routing time, average link delay, and packet delivery ratio compared with other conventional algorithms.

Predictive routing based on shortest path algorithm: in [149], the authors proposed an optimal routing algorithm for UAV networks with queuing communication systems based on Dijkstra’s shortest path algorithm, as a major step towards developing a fully predictable communication platform. As depicted in Figure 34, the authors consider a hierarchical setup, with the networks consisting of several UAV clusters. Intra-cluster routing is performed using conventional routing algorithms since the relative distances are almost constant for UAVs due to their correlated motion trajectories. ML-based predictive routing is designed for inter-cluster communication. The proposed algorithm incorporates this predictive information into the path selection criteria based on Dijkstra’s shortest path algorithm. The results show superior performance compared to the standard Dijkstra algorithm, especially when higher velocities are applied, with the performance gain achieved depending on the accuracy of the prediction.

Q-learning-based location routing protocol (QGeo): in [150], the authors proposed an ML-based geographic routing scheme to reduce network overhead in high mobility scenarios. Nodes make geographic routing decisions in a distributed manner, based on reinforcement learning techniques on the fly without knowing the entire network topology. It consists of a location estimation, neighbor table, and QL module. The current location information is updated in the Location Estimation Module based on GPS positioning parameters. The performance evaluation results of QGeo compared to other methods using NS-3 simulator show that QGeo provides a higher packet delivery rate and lower network overhead than previously reported routing protocols.

6.2.2. Self-Adaptive Learning-Based Routing Protocols

Most learning-based routing protocols use reinforcement learning (RL) to make routing decisions. RL routing protocols allow agents to discover the best route between two points in the network even without prior knowledge of those locations. Standard RL algorithms are developed to find optimal routing strategies in small-scale UAVs. Through sequential actions by interacting with the dynamic environment and using previous experiences, RL agents can make more optimal decisions to maximize rewards. Several studies have introduced RL technology into location-based protocols to improve the performance of routing protocols for UAV networks such as end-to-end delay, packet transmission rate, and routing cost. Q-learning (QL) is a model-free off-policy value-based RL, suitable for performing multi-objective optimization in resource-constrained FANETs. QL evaluates the expected value of the cumulative reward and obtains the instantaneous optimal policy based on past experience, even in an unknown environment without a central controller.

QL-based multi-objective optimization routing protocol (QMR): in [151], the authors propose a novel QL-based multi-objective optimization routing protocol for UAV networks with the advantages of low latency and low energy consumption. The QL learning parameters can be adaptively adjusted based on the network dynamics. In addition, the authors propose a novel discovery and exploitation mechanism for some unexplored potential optimal routing paths while exploiting the acquired knowledge. The proposed method re-estimates the neighbor relationships during the routing decision process to select a more reliable next hop. The simulation results show that the proposed method can provide higher packet arrival rates, lower latency, and lower energy consumption than the previous QL-based routing method.

Improved QL-based routing protocol for UAV networks: in [152], the authors proposed an improved routing protocol based on QL to achieve reliable and real-time communication in highly dynamic UAV networks. The proposed scheme combines the techniques used in two different RL-based routing protocols, QMR and Q-Noise+, into a new protocol. By combining and adapting the components of these baseline protocols, the new protocol architecture is a better fit for the dynamic behavior of UAV networks. The protocol considers a finite number of final sets to update the Q values. Additionally, it also uses channel conditions to parameterize the enhanced protocol. The authors perform simulations to compare the performance of Q-FANET with other existing approaches, namely Q-Geo, Q-Noise+, and QMR, using the WSNet simulator. The results demonstrate that the protocol outperforms standard protocols in terms of reduced latency, along with improved packet delivery ratio.

Adaptive and Reliable Routing Protocol with Deep Reinforcement Learning (ARdeep) was proposed by the authors in [153]. The algorithm decides the routing with the Markov decision process model to automatically describe the network variations. The link state is considered while making routing decisions to infer the network environment better. Additionally, the information about the packet error rate, expected connection time of the link, remaining energy of nodes, and distance between node and destination are the factors evaluated to make routing decisions. They accurately infer the network environment and make more appropriate forwarding decisions. The simulation results show that ARdeep performs better than existing QGeo routing protocols and conventional GPSR.

QL-adaptive UAV-assisted geographic routing (QAGR) [141]: the authors proposed a QAGR algorithm based on the QL-adaptive UAV-assisted geographic routing method. The routing scheme is divided into two components. For the airborne component, the global routing path is computed using a fuzzy logic algorithm and Depth-First Search (DFS) using the information collected by the UAV and forwarded to the requesting vehicle on the ground. For the ground component, the vehicle maintains a fixed-size Q-table that converges to an optimally designed reward function. It forwards the routing request to the optimal node by looking up the filtered Q-table along the global routing path. The simulation results show that the proposed QAGR method performs well in packet delivery and end-to-end delay.

QL-based adaptive swarm control inspired routing protocol in UAV networks (QRIFC): in [154], the authors proposed the QRIFC scheme as a location-based multi-hop routing protocol integrating the QL algorithm. The authors also proposed the AFCA protocol to construct the structure of a UAV swarm and then to determine the 3D LD. It is then used by QRIFC to make routing decisions. AFCA controls the relative mobility with neighboring UAVs and provides a relatively stable state for QRIFC.

A novel multi-objective reward function for QL is designed to minimize latency and energy consumption. QRIFC optimizes mobility, coverage, and QoS in the connection by imposing minimum and maximum distance constraints between UAVs. It uses two-hop neighbor information and provides stable link time (LD). This strategy expands the local view of each UAV to select a more stable path. The proposed QRIFC outperforms existing routing protocols by 21–40% in average end-to-end delay and 9–23% in average packet delivery ratio, in addition to fewer retransmissions.

Table 8. Some prominent AI-based routing algorithms.

Type	Ref.	Algorithm	Main Contribution	Advantage	Limit
Based on network structure	[147]	PARRoT	Deployment in different application models of UAV networks	Ability to predict UAV movements to optimize routing, increase robustness, and reduce latency in ad-hoc networks	Depends on the accuracy of motion prediction; performance degrades in rapidly changing environments
	[148]	MPVR	UAV network for environmental monitoring and emergency communication scenarios	Improve connectivity performance and routing time between collaborating UAVs	Performance depends on the accuracy of the Gaussian distribution model; limited scalability in large networks
	[149]	Dijkstra	High-speed UAV networking with dynamic network structures	Optimal route selection is based on incorporating the predicted locations of intermediate nodes in a transmission session into the path selection criteria	High computational cost, especially in dynamic networks; not suitable for large networks
	[150]	Qgeo	Low-cost UAV networks; applications in areas such as environmental monitoring	QGeo reduces network costs and increases packet delivery rates in mobile robot networks using reinforcement learning	Depends on Q-learning parameter settings; performance degrades in highly dynamic and complex network structures.
Adaptive learning	[151]	QMR	Low-latency and low-power service assurance solution for UAV networks	Based on Q-learning with adaptive parameters and novel discovery mechanism to optimize routing in FANET; reduce delay and energy consumption.	Depends on Q-learning parameter settings; need to balance between exploration and exploitation
	[152]	Q-FANET	Real-time and reliable communication in highly flexible UAV networks	Significant network latency reduction in highly mobile FANET environments, thanks to improved Q-Learning algorithm	Latency can increase in large networks; need to balance between exploration and exploitation
	[153]	Ardeep	High-speed UAV network with flexible network structure changes	Ability to adapt to the constantly changing UAV network, enhancing routing reliability and efficiency.	Requires large computational resources; accuracy depends on deep learning model
	[141]	QAGR	Improving convergence speed and efficient resource utilization in VANET	QAGR relies on UAV to collect global traffic information, optimize routing, and accelerate Q-learning convergence, improving packet delivery efficiency	Depends on the accuracy of global traffic information; high computational cost; depends on the performance of fuzzy logic and DFS algorithms
	[154]	QRIFC	Collaborative UAV swarms support aerial surveillance and emergency communications	Minimizes latency and power consumption; QRIFC optimizes mobility, coverage, and quality of service in connectivity	Complex in designing objective function and tuning parameters; requires large computational resources

summarizes the advantages and limitations of some prominent routing algorithms in AI-based UAV networks, including their classification and a description of their roles in specific applications.

6.3. AI-Based Control Algorithm

Artificial intelligence (AI) has become increasingly popular and plays an important role in many fields, from medicine [155,156,157] and finance [158,159,160] to transportation [161,162] and automation [163,164]. AI algorithms have demonstrated great potential in UAV swarm control, helping to optimize group operations, improve coordination, and enhance system autonomy. Many solutions based on artificial neural networks (ANN) [165] and deep reinforcement learning (DRL) [166] have been widely applied in UAV control systems for improving the efficiency of distributed control, wireless communications, distributed computing, and uncertainty in UAV swarm environments.

ANN supports UAVs to make decisions based on sensing data in applications such as object recognition, flight path optimization, and adaptive control. The ANN can predict the behaviors of other UAVs in the swarm and can improve coordination between them. On the other hand, DRL provides deep learning capabilities and optimizes actions in complex and uncertain environments. Typical applications of DRL in UAV swarm control include cooperative control, action strategy optimization, and distributed control. These advances open up great prospects for the development of intelligent, flexible, and practical UAV systems.

6.3.1. ANN-Based Control Approaches

The ANNs are computational models inspired by biological neural systems. They consist of nodes connected to each other through adjustable weights [167]. ANNs work by receiving input data, processing it through hidden layers, and generating outputs based on the learning process from training data [168,169]. Thanks to their abilities to learn and to adapt to the environment, ANNs are widely used in UAV swarm control to optimize coordination, automation, and intelligent decision-making [170,171]. In UAV swarm control, ANNs can help UAVs learn to maintain formation, avoid collisions, and adapt to changing environments [116]. These models have opened many new directions for the development of UAVs. They help to enhance the intelligence and autonomy of UAV swarms [172]. Some popular ANN models that have been studied and applied in swarm control are addressed as follows.

Radial basis function neural network (RBFNN) has the ability to approximate universal functions. This helps to solve many problems in data classification, pattern recognition, and nonlinear system control. RBFNN consists of two main layers, including a hidden layer that uses kernel functions to determine the activation region in the input space, and an output layer that calculates the result based on the link weights. The construction and optimization of RBFNN are often based on fast clustering algorithms, statistical methods, or deep learning algorithms to optimize the number of neurons and network weights [173]. One of the important applications of RBFNN is handwritten character recognition, in which fast clustering methods help to build an effective classifier with high accuracy. In addition, RBFNN is also used in control systems and state estimation for complex nonlinear systems, through approximation models and optimal sub-filters to reduce errors [174]. Due to its fast computation, good generalization ability, and high accuracy, RBFNN is a useful tool in machine learning and artificial intelligence, contributing to improving performance in many fields of science and technology [175].

In UAV control, RBFNN supports the estimation and compensation of external disturbances, improving the stability of the system. In the combination of the leading–following model and control methods such as backstepping or sliding mode, UAVs can maintain formation and track trajectories more accurately. In addition, RBFNN is also applied in real-time UAV modeling. When integrated with the EMRAN training algorithm, the network can self-adjust the number of neurons without manual tuning. Compared to the MRAN method, EMRAN helps to significantly reduce training time while maintaining high accuracy [176]. In addition, RBFNN is also used in multi-mode adaptive controllers, combining time-varying linear models and nonlinear models learned from RBFNN. This method allows flexible switching between different control modes such as traditional PID or pole and zero tuning PID, helping the UAV system adapt quickly to changing environments [177].

Chebyshev Neural Network (ChNN) is an ANN model that uses Chebyshev polynomials to expand the input, helping to eliminate hidden layers and improve computational efficiency [178,179,180,181]. ChNN plays an important role in UAV swarm control and multi-agent systems due to its ability to strongly approximate nonlinear functions and to improve controller stability. The use of ChNN helps to handle initial value problems with singularities. It is also a major challenge in numerical computation.

ChNN is also applied in tracking consensus control for a quadratic multi-agent system. The UAVs in this system are modeled as connected undirected graphs, and the controller is designed based on the quadratic sliding mode. ChNN acts as a nonlinear function approximator to learn the unknown dynamics of the UAVs online, while a robust control algorithm is used to compensate for external disturbances and approximation errors. This method ensures stability in finite time by the Lyapunov approach and maintains the reference trajectory when only some UAVs in the system receive the reference signal [182]. ChNN is also used to design an attitude controller for a quadrotor, where it is combined with the continuous quadratic sliding mode control method and the Chebyshev fuzzy neural network. This method increases the resilience to disturbances and parameter uncertainties and completely eliminates the jitter phenomenon by Lyapunov stability [183]. In addition, ChNN is used in the CPB fusion neural network model to optimize learning speed. It significantly reduces training time compared to traditional neural networks while still maintaining the universal approximation capability [184].

Convolutional neural network (CNN) is one of the most important models in the field of deep learning, especially in image processing and natural language processing. Due to the remarkable achievements in many fields, CNNs have attracted great attention from both academia and industry. CNNs are designed to automatically learn and extract spatial features from input data through convolutional layers, sampling layers, and fully connected layers. Thanks to this special structure, CNNs can detect important features in images. This helps to improve the accuracy in recognition, classification, and prediction tasks. CNN models are not only limited to two-dimensional convolutions but also extended to one-dimensional and multi-dimensional convolutions. Advanced variants of CNNs such as ResNet, DenseNet, and EfficientNet have improved performance by optimizing the feature learning process and mitigating the gradient decay problem [185,186].

CNN is also used to support the multi-agent Pickup and Delivery problem. A group of UAVs acting as guests must move in a shared environment without communicating with the host UAV group. Instead of using a static behavioral model based on past location data, this method applies CNN to predict the future location of the host UAV group, helping the guest UAV group plan an optimal path [187]. CNN is integrated into a UAV system to detect landslides by analyzing optical data collected from UAVs. The CNN model is trained to identify landslide-prone areas with high accuracy, reaching 90% in precision evaluation and 85% in F-score [188]. CNN is also applied to improve the prediction of vehicle movements in an autonomous navigation system. The model is improved by replacing LSTM with a Time Convolutional Neural Network. It speeds up the computation and improves the accuracy of motion prediction. The experimental results on the Argoverse 2 dataset show that the model has superior performance in motion trajectory prediction and computational efficiency compared to previous methods [189].

Recurrent neural networks (RNN) are suitable for sequential data processing problems such as speech recognition, online handwriting, machine translation, and time series forecasting. The highlight of RNN is the ability to remember information from previous steps through recurrent connections. It allows the model to learn and analyze the relationships between elements in the data series. Other variants of RNN have also been developed to enhance learning capabilities and apply to more complex problems such as image, video, and medical data processing. Multidimensional recurrent neural networks have extended the application scope of RNN to areas such as image segmentation and spatiotemporal data processing without facing the scaling problem of previous models [190,191,192].

An intelligent consensus-based formation control method for small, unmanned helicopters (SSUH) is developed with the help of RNN. The mathematical model of SSUH is built on a directed graph, in which RNN is used to learn the system uncertainties online and track the consensus among UAVs. The Lyapunov theory is applied to ensure the stability of the system [193]. RNN is integrated into the model predictive control (MPC) structure to control the formation flight of multiple quadrotor UAVs. A hierarchical control system is designed to control both the translation and rotation systems of the UAVs. RNN serves as the predictive model in MPC. It helps the UAVs adjust their states in real time. In addition, an adaptive differential evolution algorithm is used to optimize the system, ensuring the stability and performance of the control system [194]. An annealing optimization RNN is applied to maintain a UAV formation under tight conditions. This model helps the fixed-wing UAV adjust the tilt angle to minimize the aerodynamic impact from the leading UAV, optimizing the energy consumption of the UAV in the formation. The ESA algorithm combined with RNN is deployed to solve the oscillation problem and improve the system performance [195]. RNN is used as a strategy controller in a multi-agent decision-making model, combined with a particle swarm optimization algorithm. UAVs learn adaptively by updating the neural network weights based on the strategies of other agents. This system helps the UAVs avoid collisions, optimize the movement strategy, and maximize the performance of the entire UAV group in joint missions [196].

Table 9. Prominent artificial neural network-based control algorithms in UAV networks.

Ref.	Proposed Method	Main Contribution	Advantages	Limit
[176]	Fully Tuned RBF Neural Networks with MRAN/EMRAN	Real-time system identification for quadcopter using RBF NN with automatically adjusted number of hidden neurons.	Automatically selects the optimal number of neurons; high prediction accuracy; short training time.	Complex to implement; requires real flight data for training
[177]	RBF Neural Network-Based Generalized Learning Model	Combining multi-task controller with general learning model (GLM) based on Neural Network	Flexible in switching between controllers; combines the advantages of traditional PID and neural networks	Requires complex calibration; low scalability
[182]	Terminal Sliding Mode combined with Chebyshev Neural Networks	Tier 2 UAV swarm consensus tracking controller	Only need information from neighboring agents; good anti-interferenceg no chattering phenomenon	Requires complex calculations
[183]	Terminal Sliding Mode + Chebyshev Neural Networks	Combining continuous quadratic sliding mode controller (CSOSMC), Fuzzy-Chebyshev network (FCN), and adaptive control method	Good anti-interference (completely eliminates vibration); stability is proven by Lyapunov method	Complex controller design
[184]	Chebyshev Polynomials-Based (CPB) Unified Model Neural Networks	Approximating nonlinear functions using neural networks based on Chebyshev polynomials	Proposed method does have the capability of universal approximator in some functional approximation with considerable reduction in learning time	Difficult to apply to complex dynamic systems
[187]	Convolutional Neural Network (CNN)	Use CNN to predict future position and update in real time	Avoid conflicts in the planning phase; adapt to dynamic behavior	Depends on training data quality; high computational cost
[188]	CNNs (Convolutional Neural Networks)	Using CNN for landslide detection from UAV images	T integrates multiple data (image + gradient);high accuracy (90%).	Efficiency depends on CNN design; requires high quality input data
[189]	BC (Temporal Convolutional Networks)	Using TCN instead of LSTM to predict motion for autonomous vehicles	Improved calculation speed and efficiency; prediction of both vehicle and pedestrian movements	Requires large training data; depends on the quality of the dataset
[193]	RNN (Recurrent Neural Networks)	Consensus control of unmanned helicopter UAV; leader–follower method using RNN	Lyapunov stability; effective in UAV swarm flight	Requires complex dynamic modeling; low scalability of UAV swarm numbers
[194]	RNN-MPC (Recurrent Neural Network-Based Model Predictive Control)	Using RNN as prediction model in MPC with adaptive update rule to adjust weights	Distributed hierarchical control system; online weight update to improve model accuracy.	Requires complex calculations
[195]	Annealing Recurrent Neural Network and Extremum Seeking Algorithm	Using annealing recurrent neural networks to search for extreme values, optimizing UAVs flying in tight formation	Reduce energy consumption and anti-chattering; achieves optimal configuration and minimum energy requirements	Requires accurate aerodynamic turbulence modeling
[196]	RNN and PSO (Particle Swarm Optimization)	Multi-agent decision model using RNN and PSO, with input being the previous strategies of other agents	Adaptive learning, maximizing group benefits	Depends on PSO weight update

summarizes the advantages and limitations of some prominent artificial neural network-based control algorithms in UAV networks.

From the above analysis, we can see the effective combination of modern methods based on neural networks (RBF, Chebyshev, CNN, RNN) with classical control techniques (sliding mode, PID, MPC). From there, we can solve the challenges in UAV swarm control and multi-agent systems (including problems such as uncertain systems, noise, and nonlinear dynamics of UAVs). However, most of the methods require large training datasets and complex computational requirements, making their application to large UAV systems challenging.

6.3.2. DRL Based Control Approaches

DRL is a branch of deep learning and reinforcement learning, in which UAVs learn to optimize their behaviors by receiving rewards and penalties from the environment after each action. The applications of DRL in UAV formation control represent significant advancements in autonomous UAV swarms. DRL allows UAVs to learn optimal control policies through interactions with the environment, which is more efficient and robust compared to traditional methods. Based on DRL techniques, UAVs can autonomously coordinate to maintain formation patterns while ensuring collision avoidance and adapting to dynamic environments. Some common DRL models used in UAV swarm control are presented as follows.

Deep Q-Networks (DQN) is an algorithm in deep reinforcement learning. It combines Q-learning and deep neural networks to solve decision-making problems in complex environments. This method was introduced by Google DeepMind, and it has achieved significant success in training intelligent agents to perform tasks autonomously without human supervision. DQN is based on the principle of Q-learning, in which the agent learns to optimize actions by maximizing the value of the received reward. Instead of using a Q-table as in traditional Q-learning, DQN uses a deep neural network to approximate the Q-value function. This neural network takes the state of the environment as input and output Q values corresponding to each possible action [197,198]. Deep Q-Networks has been widely applied in UAV swarm control to optimize operational performance, save energy, and enhance coordination between UAVs in complex environments.

One of the major challenges in UAV swarm control is formation switching due to the complex system structure and large computational requirements. To address this problem, a UAV formation switching method based on an improved Deep Q-Networks (DQN) algorithm has been proposed. The UAV formation switching problem is modeled as a Markov decision process (MDP), in which the DQN is used to determine the optimal actions for each UAV during the formation change. Next, an optimal dynamic target allocation algorithm is applied to optimize the positions of the UAVs in the new formation. It improves the formation switching performance and accelerates the convergence speed of the DQN [199]. DRL algorithms, including DQN are used to find energy-saving flight strategies through the learning mechanism from experience. DQN is applied in problems such as flight path planning, trajectory optimization, and obstacle avoidance to optimize energy consumption while ensuring mission completion [200]. A model combining DQN with a value decomposition network and a convolutional block attention module ensures communication quality with limited spectrum and power resources in UAV swarm systems. This method allows UAVs to optimize spectrum utilization in UAV-to-UAV (U2U) and UAV-to-Base Station (U2B) links, improving data transmission rates and increasing the probability of transmission success [201]. DQN is also used to optimize UAV flight paths in wireless sensor networks. When IoT devices use reflective transmission technology with limited transmission range, UAVs can assist in data collection by approaching and triggering sensors. The Prioritized DQN algorithm is proposed to reduce the total flight time of UAVs and to optimize energy and data collection efficiency compared to other DRL algorithms such as MADDPG, DDPG, and DQN [202]. A DQN model incorporating an attention mechanism was proposed to optimize the UAV control system, improve data transmission efficiency, and solve the scalability and mobility issues of UAVs. The experimental results show that this method helps to optimize the data transmission rate and ensure a reasonable throughput level in mobile networks using UAVs [203].

Momentum Policy Gradient (MPG) is an approach in reinforcement learning. It aims to improve the performance of gradient policy algorithms by reducing variance and increasing convergence speed. Traditional gradient policy methods often suffer from large fluctuations in gradient estimation, leading to unstable performance and requiring a large number of samples to achieve the desired accuracy [204,205]. This algorithm is applied to improve the performance of UAV swarm control and multi-agent systems.

In [206], an MPG algorithm was proposed to solve the UAV control problem with tasks such as target tracking, obstacle avoidance, and formation maintenance. The system is designed according to a two-part control model, including a sensing module and a control module. It helps to train a DRL agent without the need for a complex physical model. The experimental results show that MPG is capable of tracking the movement of the leading UAV stably, and it can be extended to control a UAV formation. In [207], a Decentralized Policy Gradient with Momentum (DPGM) algorithm was proposed to improve the performance of finding optimal policies for multi-agent systems. DPGM applies a policy update mechanism using gradients with momentum to increase the convergence speed while ensuring dispersion in a reinforcement learning environment. This algorithm achieves a convergence speed of O(1/T). It is on par with the most advanced distributed policy gradient methods today. In [208], a Distributed Adaptive Policy Gradient with Momentum (IS-DAPGM) algorithm was developed based on the Adam update mechanism and importance sampling technique to improve the UAV swarm control performance.

Deep Deterministic Policy Gradient (DDPG) was developed to overcome the limitations of traditional Q-learning-based policy algorithms. They only work well with discrete action spaces [209]. By combining Deterministic Policy with deep neural networks, which can handle continuous state and action spaces, DDPG is capable of optimizing policies efficiently without sampling actions from a probability distribution like stochastic policy algorithms. DDPG uses experience recurrence to improve training efficiency and also applies an objective network to stabilize the weight update process, avoiding oscillation or divergence [209]. DDPG and its variants are used to optimize UAV control in complex missions.

Multi-agent deep deterministic policy gradient (MADDPG) is an improved variant of the DDPG algorithm. This method is effective in situations where agents must learn to cooperate, compete, or coexist in a shared environment. Several improved versions of MADDPG aim to address challenges such as learning efficiency, collision avoidance, and enhancing cooperative behavior in UAV clusters, including GPR-MADDPG and COM-MADDPG. The GPR-MADDPG method combines Gaussian Process Regression (GPR) and Multi-Agent DDPG to optimize UAV control to track moving vehicles [210]. This method ensures that UAVs maintain reasonable overlapping observation ranges, minimize tracking errors, and increase surveillance performance in real and simulated environments with complex dynamics. In the target encirclement task with another variant, COM-MADDPG is used to improve the coordination between UAVs to encircle the target more effectively [211].

Table 10. Prominent control methods based on deep reinforcement learning in UAV swarm.

Ref.	Proposed Method	Main Contribution	Advantages	Limit
[199]	Improved DQN, Dynamic Target Allocation (DTA), Reciprocal Velocity Obstacle (RVO)	Solving the problem of changing UAV formation using improved DQN. Collision avoidance.	Improve the efficiency of formation transformation. Increase the convergence speed of DQN algorithm. High generalization ability.	Depends on the effectiveness of DTA and RVO. Complexity in reward function design.
[200]	DQN, DDPG, PPO, TRPO, TD3, SAC	UAV path planning and trajectory optimization applications.	SAC/TD3 excels in energy saving, suitable for continuous action. Energy optimization	Requires careful state-action space design. Long training time required.
[201]	DQN, CBAM (Convolutional Block Attention Module), VDN (Value Decomposition Network)	DQN integrates attention mechanism (CBAM) and value resolution network (VDN) to optimize U2B/U2U communication.	Maximize data transmission speed. Increase the probability of successful data transmission. Efficiently use spectrum and energy resources.	Requires complex calculations and takes a long time. Depends on channel quality.
[202]	PDQN (Parameterize Deep Q-Network)	Collect backscattering data with multiple UAVs to reduce overall flight time.	Reduced flight time compared to MADDPG/DDPG; suitable for remote sensor networks. Efficient in data collection.	Depends on the efficiency of the parameterization. Limited communication range.
[203]	DQN with Attention Mechanism	Dynamic approach for UAV handling in crowded environment using DRL and attention mechanism in multi-agent model.	Solve scalability and mobility issues; improve throughput.	Performance depends on the quality and variety of the training dataset.
[206]	DRL with MPG (Momentum Policy Gradient) algorithm	MPG (Momentum Policy Gradient) combines CNN positioning to avoid obstacles.	MPG is powerful in tracking changing leader movements. Collision avoidance capabilities.	Requires CNN training for accurate positioning. Deployability is challenging in complex outdoor environments.
[207]	DPGM (Decentralized Policy Gradient algorithm with Momentum)	Distributed Gradient Integrating Momentum (DPGM) Policy for Multitasking Problems.	Suitable for distributed systems. Fast convergence speed. Effective in solving multi-task reinforcement learning problems.	Performance can be affected by task variety.
[208]	IS-DAPGM (Distributed Adaptive Policy Gradient Algorithm)	Distributed adaptive policy gradient algorithm combined with Adam-style updates and importance sampling techniques.	Fast convergence; superior performance compared to traditional PG. Works well with different number of agents.	High computational complexity. Depends on the quality of the important sampling technique.
[209]	DDPG (Deep Deterministic Policy Gradient)	DDPG for continuous power management for wireless sensors.	Good performance with real data. Implement continuous power management, adapting to uncertainty.	Requires training with real data, requiring additional integration of dynamic constraints.
[210]	MADDPG improved with GPR (Gaussian Process Regression)	DDPG-based Multi-UAV control for moving vehicle convoy tracking using Gaussian Process Regression (GPR)	Stable with complex trajectory, effectively reducing overlapping error. Adapt to complex trajectories and changing convoy speed.	Challenged when scaling to large number of UAVs. Requires expensive GPR calculations.
[211]	COM-MADDPG (Communication-enhanced MADDPG)	MADDPG integrates the communication mechanism (COM-MADDPG) with dynamic siege points.	Dynamic encirclement point instead of fixed threshold. Handle special situations (corner/edge); higher success rate than DDPG/MADDPG.	Communication between UAVs is required. Depends on the reward function design.

summarizes the prominent control methods based on deep reinforcement learning in UAV swarms.

In the above summary, it can be seen that the research on UAV swarm control focuses on improving and combining DRL-based algorithms to solve problems such as navigation in complex environments, tracking and avoiding obstacles, optimizing energy, or optimizing communication resources. The combination of basic RL algorithms with advanced techniques improves performance and adaptability in UAV swarm applications. However, there exist some limitations such as high implementation costs and complex computational requirements, especially with GPR [210] or multi-agent [211] integration methods. In addition, the ability to expand to large scales also needs to be verified to meet the practical requirements of the application.

Artificial neural networks (ANNs) have many advantages in UAV swarm control, especially in processing sensor data, predicting states, and supporting decision-making. ANNs have the ability to learn quickly from available data, helping UAVs make quick decisions without requiring too much computational resources. Furthermore, ANNs have high stability when operating in familiar environments, and they are easy to analyze and tune to optimize performance. However, ANNs also have some disadvantages, such as requiring a large amount of labeled training data to achieve high performance and having difficulty dealing with dynamic environments. They have never appeared during training. ANNs lack a continuous learning mechanism over time, leading to poor adaptability, and need to be retrained when the environment changes significantly.

DRL offers many advantages in UAV swarm control over ANNs, especially in complex and uncertain environments. DRL allows UAVs to learn by interacting with the environment without prior labeled training data. It enables UAVs to adapt well to new situations and make optimal decisions in real time. In addition, DRL can handle nonlinear control problems such as flight trajectory optimization, collision avoidance, and formation maintenance under changing conditions. However, DRL also has some significant disadvantages. DRL training requires large computational resources and long training times. Furthermore, DRL tends to be unstable during the learning process and easily affected by noise, leading to undesirable results. In addition, due to its interactive learning nature, DRL may be difficult to apply to UAV systems. They require high accuracy from the start, as it takes time for the algorithm to reach optimal performance. DRL is a powerful solution for UAV swarm control in dynamic and complex environments, but it requires large computational resources and a stabilization mechanism during training.

7. Conclusions and Future Work

This paper shows the increasing demands and important applications that require the use of UAV swarms such as smart cities, disaster management, and environmental monitoring. Further research on UAV swarm technology has great practical significance. UAV swarms are highly mobile, and their structures frequently change. They require low latency and highly reliable communication networks. Therefore, it is necessary to design a communication architecture and routing protocol with good applicability, high efficiency, and stability for unified and efficient UAV swarm communication. We introduced the general applications of UAV groups or UAV swarms and discussed the communication routing protocols that enable smooth coordination among UAVs to support complex tasks. We also covered control algorithms that support UAVs to operate in the right position for specific purposes. Additionally, AI-based algorithms were considered to support UAVs in performing tasks optimally. We surveyed and reviewed the latest work, and then evaluated the existing results. From there, we suggested suitable solutions tailored to practical applications.

For future work, there are still many challenges and issues that need further research and development: (i) The use of AI with UAVs has legal and ethical implications. This is especially true in certain search and rescue situations. Consideration of priority areas and targets in rescue missions is a typical example. Legally, the need to share data, monitor, and access information collected by UAVs may affect personal privacy or violate specific laws of a particular area. (ii) The limitation of UAVs is mostly batteries with limited energy. The use of AI methods can increase the complexity of the computation, thereby increasing the CPU power usage in UAVs. Therefore, the calculation and selection of routing protocols and data communication models must be carefully considered. It is necessary to study whether the implementation of ML techniques in the model is commensurate with the efficiency.