A Comprehensive Review of UAV Formation Control from a Mission-Driven Perspective

Highlights

What are the main findings?

• This review systematically deconstructs the UAV formation mission lifecycle from a mission-driven perspective, synthesizing key research across the three core sub-processes: formation assembly, formation maintenance, and formation reconfiguration.
• It provides a novel, multidimensional analysis framework of UAV formation performance evaluation, summarizing the state of the art in resilience, robustness, reliability, and vulnerability.

What is the implication of the main finding?

• The integrated architecture offers researchers and engineers a structured understanding of formation control, linking specific mission phases to appropriate control strategies and performance metrics.
• The identified research challenges and future directions, particularly in heterogeneous swarms and comprehensive evaluation systems, provide a roadmap for advancing the field toward more resilient and intelligent autonomous systems.

Abstract

To systematically review the research progress on unmanned aerial vehicle (UAV) formation control, this paper proposes a mission-driven full-lifecycle analysis architecture. The architecture summarizes the core scenarios and key technologies involved in the three main stages: formation assembly, formation maintenance, and formation reconfiguration. Moreover, a comprehensive evaluation framework is established that covers pre-event, in-event, and post-event phases from the perspectives of resilience, robustness, reliability, and vulnerability. The interrelationships among these four dimensions are explained in terms of time, function, and design. Finally, this paper identifies current research gaps and practical challenges in terms of algorithms, evaluation methodologies, and real-world deployment verification, and outlines future development directions.

1. Introduction

1.1. Background and Significance

Unmanned aerial vehicles (UAVs) have attracted significant attention due to their high maneuverability, low deployment cost, and unique ability to perform complex and dangerous tasks. A single UAV demonstrates its value across a wide range of applications, including visual navigation [1,2], aerobatics [3], radiation monitoring [4], and slope inspection [5], and is increasingly being deployed in various civilian and military fields.

However, the operational capability of a single UAV is inherently limited by factors such as payload capacity, flight range, endurance, and fault tolerance. Complex missions that require large-area coverage, simultaneous multi-point data collection, or coordinated physical interaction often exceed the capacity of a single platform. Therefore, organizing multiple UAVs into a structured formation has become a critical research direction. Through cooperative control, UAV formations can achieve greater system robustness, redundancy, efficiency, and overall performance that exceeds the sum of individual capabilities. Such formations have shown great potential in applications such as target surveillance [6], mapping and surveying [7], forest firefighting [8], and collaborative transportation [9].

UAV formation control refers to the process in which multiple UAVs autonomously and cooperatively generate and maintain a desired spatial geometric configuration and communication network topology through inter-UAV communication and distributed computing. This process enables real-time responses to mission commands and various disturbances in dynamic environments, thereby achieving closed-loop feedback for formation stability, switching, and reconfiguration. Its core objective is to integrate individual UAVs into an aerial swarm that possesses robustness, adaptability, and collaborative capabilities, allowing it to perform complex tasks beyond the capacity of a single UAV. The key sub-processes include formation assembly, formation maintenance, and formation reconfiguration, as illustrated in Figure 1.

Among these, the three sub-processes—formation assembly, formation maintenance, and formation reconfiguration—are not independent but rather interconnected and interactive, exhibiting a task-driven sequential evolution. Specifically, formation maintenance continuously sustains the predefined formation configuration and communication topology through closed-loop feedback control in complex dynamic environments. This ensures the stable operation of the entire formation under disturbances and provides a reliable state foundation for mission execution. Formation reconfiguration handles the transient processes triggered by mission switching or changes in internal and external environments, guiding the formation to transition dynamically, orderly, and controllably from its current stable maintenance mode to the optimal mode under new constraints. Collectively, these three layers form a hierarchical and tightly integrated closed-loop control system, guided by mission objectives and constrained by environmental factors.

1.2. Research Gap and Motivation

In recent years, many researchers have investigated the key technologies and core issues in unmanned aerial vehicle (UAV) formation control. Th early literature [10] classified multi-agent formation control into three categories—position-based, displacement-based, and distance-based—from the perspective of perception and interaction topology, laying a foundation for the classification of formation control. Ref. [11] proposed a five-layer swarm intelligence architecture consisting of decision-making, planning, control, communication, and application layers. Refs. [12,13] compared and analyzed the advantages and disadvantages of centralized versus distributed approaches, as well as traditional models versus artificial intelligence methods in formation control, from the perspectives of system architecture and algorithmic paradigms. Ref. [14] systematically reviewed the decade-long evolution from PID control to swarm intelligence and reinforcement learning. Refs. [15,16] summarized research progress in key tasks such as collision avoidance, task assignment, path planning, and formation control from the perspectives of swarm intelligence algorithms and system architecture, respectively. Ref. [17] classified moving target tracking into three categories: detection, tracking, and collaboration. Ref. [18] compared four types of trajectory prediction methods: mathematical models, classical machine learning, deep learning, and reinforcement learning. Ref. [19] categorized collision avoidance strategies into six major types. Ref. [20] reviewed collision avoidance technologies across three stages: obstacle perception, collision prediction, and collision avoidance. Refs. [21,22] systematically analyzed swarm security threats and corresponding defense strategies from the network and system levels, respectively. A comparative analysis of the aforementioned related work is presented in

Table 1. Summary of related studies.

Reference	Year	Core Focus	Formation Assembly	Formation Maintenance	Formation Reconfiguration	Formation Performance Evaluation	Mission-Driven Perspective
Oh [10]	2015	multi-agent formation control	●	●	●	○	○
Huang [19]	2019	collision avoidance strategies	○	●	○	○	○
Zhou [11]	2020	UAV swarm intelligence	○	●	●	○	○
Wei [20]	2021	collision avoidance technologies	○	●	○	○	○
Tang [15]	2022	Swarm intelligence algorithms	○	●	●	○	○
Ouyang [12]	2023	Communication Networks and Formation Control Strategies	●	●	●	○	○
Javed [16]	2024	Key technologies for UAV swarms	○	●	○	○	○
Cetinsaya [14]	2024	UAV formation control and path planning algorithms	○	●	○	○	○
Bu [13]	2024	traditional and AI-based formation control methods	○	●	●	○	○
Wang [21]	2024	Security of UAV swarm networks	○	●	○	○	○
Wei [22]	2024	Security of single UAV systems	○	○	○	○	○
Sun [17]	2024	Moving target tracking	○	●	○	○	○
Shukla [18]	2024	Trajectory prediction	○	●	○	○	○
This article	-	Mission-driven full lifecycle of UAV formation control	●	●	●	●	●

In this context, ● indicates that the topic is addressed in depth or systematically; ○ indicates that it is not a core research topic.

.

As shown in the comparison above, existing reviews and studies mainly focus on the performance of a specific phase or aspect of formation control, lacking a systematic integration of the full lifecycle and multidimensional performance evaluation of formations.

Building on previous research, this paper is the first to incorporate formation modeling, assembly, maintenance, reconfiguration, and performance evaluation into a unified task-driven framework, thereby addressing the limitations of existing research in terms of systematicness and comprehensiveness. Furthermore, to the best of our knowledge, this is one of the first attempts to provide a systematic and multidimensional review of UAV formation performance evaluation from the four dimensions of resilience, robustness, reliability, and vulnerability, filling a gap in the systematic review of formation performance evaluation in this field.

1.3. Novelty and Contributions

Before establishing the research framework and topic of this paper, we searched the CNKI database using the keyword “UAV formation control,” selected the 500 most relevant and most recently published Chinese documents, and performed a keyword co-occurrence analysis using VOSviewer 1.6.20. The results are shown in Figure 2. The analysis indicates that the three keywords “formation assembly,” “formation maintenance,” and “formation reconfiguration” appear with high frequency. These are closely related to various specific technologies and, together with formation modeling and performance evaluation, constitute the full lifecycle of UAV formation mission execution. This supports the construction of a logically coherent analytical framework from a “task-driven” perspective. However, no previous research has systematically organized and summarized this field along these lines; therefore, the core themes of this paper are established accordingly.

This review adopts a systematic literature screening approach. Using keywords such as “UAV formation,” “formation control,” “formation assembly,” “formation reconfiguration,” and “formation performance evaluation,” the search period was set from 2015 to 2026, also including highly cited early classic papers from top journals in the field. A preliminary search of the IEEE, ScienceDirect, and Springer databases yielded 9606 documents. After removing duplicates and screening by titles and abstracts, 852 papers remained. Through full-text reading and quality assessment, papers not directly relevant or focusing on individual technical details were excluded, resulting in 246 core references. Additional relevant literature was incorporated during in-depth reading to form the analytical foundation of this review. Figure 3 illustrates the overall process of literature retrieval and screening.

To demonstrate that the selected literature provides a relatively comprehensive reflection of the cutting-edge research in UAV formation control, this paper analyzes the collected literature from two dimensions: publication year and source journal or conference, as shown in Figure 4 and Figure 5, respectively. In terms of temporal distribution, publications from 2020 to 2026 account for more than half of the total, clearly reflecting the latest research progress in the field. Meanwhile, publications before 2010 mainly consist of foundational and pioneering theoretical achievements, which are equally indispensable. Regarding publication sources, more than half of the selected literature comes from top-tier journals in fields such as aerospace, control, robotics, and operations research and management, indicating strong authority and representativeness.

The main contributions and innovations of this paper are as follows:

(1)

From a mission-driven perspective, this study takes the full lifecycle of UAV formation mission execution as its main thread to analyze and summarize the key sub-problems and core mission scenarios involved in the three main sub-processes: formation assembly, formation maintenance, and formation reconfiguration.

(2)

We summarize and analyze existing research on UAV formation performance evaluation from four dimensions—resilience, robustness, reliability, and vulnerability—and propose a comprehensive analytical framework. As far as we are aware, no previous work has systematically summarized research on UAV performance evaluation from multiple dimensions.

(3)

We discuss the current research gaps and future development directions in the field of UAV formation control.

1.4. Paper Roadmap

The rest of this paper is organized as follows: Section 2 reviews modeling and representation methods for UAV formations. Section 3, Section 4 and Section 5 discuss formation assembly, formation maintenance, and formation reconfiguration in detail, respectively. Section 6 analyzes the current research status of UAV formation performance evaluation. Section 7 discusses research gaps and future directions, and summarizes the entire paper. The structure of the paper is illustrated in Figure 6.

2. Formation Modeling and Representation

UAV formation modeling primarily defines the spatial geometric configuration, communication topology, and cooperative control methods of the formation. It serves as the prerequisite and foundation for all research on UAV formations, and its precision and adaptability directly affect the effectiveness of mission execution.

2.1. Spatial Geometric Configurations

The spatial structure of a UAV formation primarily refers to the geometric configuration and relative positional distribution of the swarm in three-dimensional space, which is typically reflected in the formation patterns adopted during mission execution.

Formation design is a fundamental issue that must be addressed before executing UAV formation missions. A well-designed formation can help optimize overall energy consumption, extend mission endurance, and significantly improve formation safety and robustness. In practical design, formations need to be dynamically adjusted based on mission type, fleet size, and operational environment to achieve an optimal balance among safety, flexibility, energy efficiency, and mission effectiveness. Common UAV formation patterns include V-formation, single-file column, single-file line, circular, trapezoidal, diamond, and serpentine formations.

Among these, the V-formation uses updrafts to reduce fuel consumption and extend range [23]. The single-file column formation is suitable for passing through narrow obstacles such as valleys and for evading radar detection [24]. The single-file horizontal formation is commonly used for area scanning and coordinated search tasks. The double-line horizontal formation is often employed for jamming and suppressing enemy radar. The circular formation is frequently used for continuous perimeter surveillance and sustained observation of fixed targets [25]. The trapezoidal formation is commonly applied for lateral entry into target areas and conducting covering fire strikes [26]. The diamond formation offers greater flexibility and stability, making it suitable for complex mission scenarios that require rapid formation changes or protection of central targets [27]. The serpentine formation is often used for following winding paths or dynamically tracking moving targets. A performance comparison of common UAV configurations is shown in

Table 2. Performance comparison and applicable scenarios for common UAV configurations.

Configuration	Safety	Flexibility	Energy Efficiency	Task Effectiveness	Applicable Scenarios
v-shape	3	2	5	4	Long-range cruising
single-file column	4	4	3	3	Navigate narrow passages, conduct covert infiltration
single-file line	3	3	4	5	Area scanning, collaborative reconnaissance search
double-line horizontal	3	3	3	5	Cooperative electronic jamming and radar suppression
circular	4	2	2	5	Continuous surveillance and persistent observation of fixed targets
trapezoid	3	3	3	4	Fire coverage strike
rhombus	5	2	4	4	Critical Asset Protection
snake-like	3	5	2	4	Tracking Moving Targets

.

2.2. Communication Topology Relationships

The communication topology of a UAV formation describes the network structure for information exchange among the UAVs, defining the connection rules and direction of data links between them. Common topologies include centralized [28], distributed [29], and hybrid structures [30]. Figure 7 illustrates typical topology configurations for UAV formations.

In a centralized topology, a single designated leader serves as the main node that communicates with all followers, with no direct interaction among the followers themselves. This structure facilitates centralized decision-making and global optimization, but it also introduces risks of a single point of failure, imposes a heavy communication burden on the leader, and results in low system resilience [31].

A distributed topology has no fixed leader. Each UAV coordinates with its neighbors through local communication, and formation consistency is achieved through distributed algorithms. This structure is decentralized, highly robust, and scalable, but it requires fast algorithm convergence and strong consistency guarantees [32].

A hybrid topology combines the two approaches using a layered architecture. Within each group, UAVs operate under the centralized control of a group leader, while the group leaders themselves form a distributed or centralized network at the upper layer. This approach integrates the efficiency of centralized scheduling with the reliability of distributed systems, making it suitable for large-scale formations with better scalability and fault tolerance [21].

Consequently, the communication topology within the formation also determines its control structure, which can be similarly divided into centralized control [33], distributed control [34], and hybrid control [35]. Centralized control enables effective management of the swarm through a single hub, where the central node establishes communication links with designated UAVs. Distributed control mimics the collective behavior of biological groups, with individuals interacting through information exchange to accomplish tasks collaboratively. Hybrid control achieves global management of the UAV swarm by combining centralized coordination with distributed autonomy.

2.3. Cooperative Control Methods

Cooperative control methods for UAV formations mainly include leader–follower methods [36,37,38], virtual structure methods [39,40,41], behavior-based methods [42,43,44], and methods based on consensus theory [45,46,47], as shown in Figure 8.

The leader–follower method [48] establishes one or more leaders, with followers maintaining formation according to predefined relative positions or distances. The leader’s trajectory is generated independently, while followers track it via local feedback control. This method features a simple structure, ease of implementation, and low computational cost. However, it suffers from single-point dependency, where leader failure can destabilize the formation, and it also lacks flexibility.

The virtual structure method [49] treats the formation as a rigid geometric structure, with each UAV assigned to a fixed point. Coordination is achieved by tracking the motion of this virtual structure. This method offers high formation accuracy and is suitable for tight formations, providing unified global motion planning and strong robustness. However, its structural rigidity leads to weak obstacle avoidance and poor adaptability to dynamic environments.

The behavior-based method [50] designs multiple fundamental behaviors for each UAV, such as obstacle avoidance, aggregation, and formation maintenance, and generates control commands through weighted fusion of behavioral outputs. This approach offers high flexibility and strong real-time responsiveness to environmental changes, along with good distributed characteristics and high fault tolerance. However, theoretical analysis is challenging, stability is difficult to guarantee, and behavior weighting relies on empirical tuning, limiting generalization.

The consensus-based method [51] uses graph theory to describe the communication topology among UAVs. Through local information exchange, state variables such as position and velocity achieve asymptotic consensus, enabling formation generation and maintenance. This approach has a strong theoretical foundation, provable convergence, and good scalability, making it suitable for large-scale formations, and it exhibits robustness to changes in communication topology. However, it requires high communication quality, and factors such as latency or packet loss can degrade performance. It is also sensitive to initial conditions, and convergence speed is constrained by the topology structure. In addition, designing complex formations requires integrating geometric constraints, which increases design complexity.

The performance comparison of these methods is shown in

Table 3. Comparison of different cooperative control methods.

Method	Per-Vehicle Computational Complexity	Scalability	Main Bottleneck	Suitable Scale
leader–follower	Follower: $O\left(1\right)$ Leader: $O\left(N\right)$	Low	leader bottlenecks; error accumulation	Small
virtual structure	Trajectory tracking: $O\left(1\right)$ pose assignment: $O\left(N\right)$	Medium	limited flexibility; limited adaptability	Small to medium
behavior-based	$O\left(m\right)$	High	parameter tuning challenges; limited formation accuracy	Medium to large
consensus-based	$O\left(k\right)$	Very high	communication dependency; design complexity	Large

, where $m$ represents the number of behaviors and $k$ represents the number of neighbors.

2.4. Multi-Layer Heterogeneous Networks

Section 2.2 and Section 2.3 systematically reviewed the communication topologies and cooperative control methods of UAV formations based on the core assumption that all UAVs within the formation operate at the same level and achieve coordination by exchanging relative positions or local information.

However, as mission scenarios expand to wide-area coverage, long-endurance operations, and the offloading of complex tasks, the limitations of a single UAV layer have gradually emerged. In recent years, integrating high-altitude platform stations (HAPSs) into UAV networks to form a multi-layer, heterogeneous architecture comprising a HAPS layer and a UAV layer has become a research hotspot in the field of air–ground integrated networks [52,53].

HAPS are typically deployed in the stratosphere at an altitude of around 20 km. With their wide-area coverage, long endurance, and high payload capacity, they can act as aerial base stations or central processing units, providing communication backhaul, task offloading, and global scheduling services for UAV formations at lower altitudes. This integration fundamentally changes the communication and control paradigm of the formation, expanding its operational capabilities. Compared with pure UAV formations, the introduction of HAPS offers advantages in ensuring fair resource allocation and scheduling [54], providing computational support [55], and achieving wide-area continuous coverage [56].

However, existing research still has the following limitations. First, most models assume that the HAPS-UAV link is ideally stable and do not fully consider actual channel fluctuations. Second, interference management and resource competition issues arising when multiple UAVs access the same HAPS remain to be further studied. Third, there is no unified theoretical framework for the joint deployment and dynamic reconfiguration of HAPS and UAVs in large-scale heterogeneous scenarios.

3. Formation Assembly

UAV formation assembly is the first step in executing multi-UAV cooperative missions. Its core lies in enabling individual UAVs to autonomously, safely, and efficiently converge from a dispersed initial state to establish the initial formation while satisfying kinematic, dynamic, and environmental constraints.

Existing research mainly focuses on two key subproblems. The first is spatiotemporal rendezvous, which aims to achieve precise convergence of multiple UAVs at a specified spatiotemporal point through coordinated planning and tracking control. Mainstream approaches can be divided into three categories: velocity-based, trajectory-based, and guidance-based. Among them, velocity-based methods have a simple structure but are sensitive to communication delays; trajectory-based methods offer strong adaptability but require high computational effort; guidance-based methods provide clear stability guarantees but rely heavily on real-time data acquisition. In recent years, research on spatiotemporal rendezvous involving heterogeneous unmanned systems has gradually emerged, yet a unified and scalable cooperative framework remains lacking.

The second is formation generation, which refers to the process of autonomously arranging into a predefined geometric configuration and stabilizing it after convergence, based on local perception and communication. Existing research mainly adopts segmented control strategies, reducing the complexity and collision risk of direct formation by transitioning from a loose to a tight formation. Techniques such as zone-based assembly and altitude stratification are also used to improve safety and efficiency. However, current methods still rely primarily on preset rules and fixed-stage logic, lacking a systematic trade-off mechanism among tight formation stability, collision avoidance guarantees, and formation efficiency.

This section systematically reviews the existing research methods, typical strategies, and their applicable conditions from the perspectives of spatiotemporal rendezvous and formation generation, providing a reference for the control design and mission planning of UAV formations.

3.1. Spatiotemporal Rendezvous

Spatiotemporal rendezvous refers to the process in which multiple UAVs, initially distributed across different spatial locations, plan and execute their trajectories under kinematic and dynamic constraints and external environmental disturbances, in order to converge at a specified time and location.

Current research on UAV formation coordination and rendezvous mainly focuses on two types of problems: determining the rendezvous position and determining the rendezvous time. For the first problem, researchers typically set a predetermined rendezvous time to calculate the rendezvous position and path [57,58,59], while for the second problem, they usually set a predetermined rendezvous position to optimize the rendezvous time [60,61,62]. Existing rendezvous methods are generally classified into three categories: velocity-based [63,64,65], trajectory-based [66,67,68], and guidance-based [69,70,71].

(1)

Velocity-based method

This method adjusts the velocity vectors of each UAV in real time to ensure that they converge at the target location by a predetermined time. Typical approaches include velocity consensus protocols and proportional navigation-based strategies. Such methods usually model the rendezvous problem as a distributed control problem within a multi-agent system, relying on the exchange of local information to achieve swarm coordination [72]. Ref. [73] proposes a velocity control and online trajectory correction method for fixed-wing UAV formation. Ref. [74] achieves dynamic velocity control by improving the consensus algorithm and solves the velocity coordination problem for formations with large spacing by incorporating particle swarm optimization.

This method achieves distributed cooperative control through the exchange of local information. It features a simple structure and facilitates real-time adjustment, but it is relatively sensitive to initial velocity differences and communication delays, and its convergence rate and coordination efficiency are easily affected by the network topology.

(2)

Trajectory-based method

This method typically generates reference trajectories that satisfy spatiotemporal constraints either offline or online, with each UAV achieving rendezvous through a trajectory tracking controller. To address the spatiotemporal consistency and rendezvous efficiency issues during multi-batch takeoffs of shipborne UAVs, ref. [75] proposed a multi-batch formation rendezvous method based on Improved Sequential Convex Programming (IST-SCP), which alleviates the efficiency loss caused by improper initial solutions. Ref. [76] presents an online trajectory generation framework based on multi-stage path prediction, using local and global A-star algorithms for real-time path prediction and combining cubic B-splines to generate flyable trajectories that satisfy dynamic constraints.

This method can explicitly satisfy spatiotemporal constraints and generate feasible trajectories that meet dynamic requirements, and is highly adaptable to complex rendezvous scenarios. However, it imposes a heavy computational burden for trajectory planning and online adjustment, and is sensitive to the quality of the initial solution and changes in the environment.

(3)

Guidance-based method

This method draws on missile guidance theory and transforms the rendezvous problem into a dynamic target tracking problem through the design of guidance laws [77]. In particular, ref. [78] used differential geometry tools to design guidance laws for UAVs to rendezvous with and hover while tracking moving targets. Three types of rendezvous solutions—convergent, divergent, and parallel—were derived, and the global stability of the guidance laws was rigorously proven using Lyapunov theory. Ref. [79] proposes a UAV rendezvous guidance scheme based on three-point guidance. Ref. [80] integrates terminal line-of-sight angle and velocity constraints into a predefined-time sliding mode guidance law, and presents a comprehensive scheme that combines terminal constraint guidance with predefined-time event-triggered consistency control.

This method transforms the rendezvous problem into a guidance law design problem with well-established theory, enabling theoretical guarantees of stability and the derivation of closed-form solutions, and has clearly defined convergence properties. However, its performance heavily depends on the real-time and accurate acquisition of the leader’s state information, and its robustness degrades in scenarios with limited sensors or communication interruptions.

In addition, research on spatiotemporal rendezvous of heterogeneous unmanned systems has gradually become a research hotspot. Ref. [81] proposed a separation and rendezvous control method for UAV-USV systems based on distributed NMPC, enabling two groups of UAVs to operate in a rotating manner and achieve continuous operation. Ref. [82] proposed a cooperative rendezvous controller for heterogeneous UAVs and USVs, allowing heterogeneous platforms to autonomously form a specified circular formation and perform coordinated rendezvous using only communication between the lead UAV and the USV. To address the limited endurance and periodic charging needs of low-cost fixed-wing UAVs, ref. [83] proposed a method for generating rendezvous trajectories between UAVs and mobile ground vehicles based on optimal control. Ref. [84] proposed a dynamic programming method capable of efficiently solving UAV-UGV charging and rendezvous scheduling problems. This method maintains optimality while supporting multiple charging scenarios and features low computational complexity, making it suitable for large-scale missions.

However, existing research on spatiotemporal coordination of heterogeneous unmanned systems has mainly focused on specific platform combinations and mission scenarios. There remains a lack of a unified and scalable framework that can address the real-time coordination requirements arising from diverse platform types, dynamic mission changes, and environmental uncertainties.

3.2. Formation Generation

Formation generation refers to the process in which a fleet of UAVs, after converging in the designated airspace, autonomously arranges into a specific or predefined geometric configuration through local sensing and communication. However, under actual conditions with disturbances and uncertainties, the state of individual UAVs upon reaching the rendezvous point may contain errors. Moreover, mission requirements may necessitate forming a tight formation to directly establish an ideal configuration suitable for subsequent operations, which poses significant control challenges and a high risk of collisions between UAVs. Existing research typically accomplishes the formation of the ideal configuration in stages.

Ref. [85] proposed a segmented control strategy for tight formation assembly of UAVs based on information consensus. By introducing a coordination variable, this strategy enables each UAV to first rapidly form a loose formation at higher speeds, and then gradually transition to a tight formation by relaxing geometric constraints, thereby improving the overall assembly efficiency. Building on the positional coordination variable in [85,86], a partitioned formation control strategy was proposed. By dividing the area around the formation point into separate, non-interfering partitions and planning the formation sequence of UAVs within each partition, this method effectively improves formation efficiency and reduces unnecessary flight energy consumption. To increase the safety margin during formation, ref. [74] guided UAVs to different altitude layers and then used a three-step formation method to complete the assembly. Furthermore, ref. [87] presented a generalized formation design method that uses minimal formation units to construct arbitrary formations. Ref. [88] proposed an autonomous cooperative control strategy for UAV swarm formation assembly based on the LDE-MADDPG algorithm, which is capable of adapting to formation assembly tasks involving different formation shapes and varying numbers of UAVs.

Existing research adopts methods such as phased assembly, zone control, and multi-level hierarchy to reduce the risk of collisions and the control difficulty associated with directly forming tight formations. However, these methods largely rely on predefined rules and fixed phased logic. Most studies have yet to establish a systematic trade-off mechanism that balances the stability of tight formations, collision avoidance guarantees, and formation efficiency.

4. Formation Maintenance

Formation maintenance is essential for ensuring that the formation system continuously sustains the desired spatial configuration and communication topology connectivity in complex dynamic environments. It is a critical factor in guaranteeing the stability and reliability of collaborative missions.

Existing research mainly focuses on four key scenarios. The first is formation maintenance and topology preservation. Research is evolving from control methods that rely on global information toward distributed control that depends only on relative position perception. Theories of position rigidity and affine formation control have improved adaptability in GNSS-denied environments. Meanwhile, topology preservation has shifted from requiring strong connectivity to allowing more relaxed conditions, such as intermittent connections and joint connectivity. Tools like algebraic connectivity offer new ways to balance flexibility with connectivity.

Second, for robust formation flight in complex environments, multiple sources of interference such as wind disturbances, GNSS jamming, and obstacles have driven the development of a technical framework that integrates disturbance observation compensation, robust control, multi-sensor fusion, relative positioning, and a combination of potential field methods, geometric methods, optimization methods, and learning methods. A key research focus remains how to achieve a balance between disturbance suppression and formation maintenance.

Third, for secure network formations under cyberattacks, research has gradually shifted from attack modeling and detection toward resilient control and event-triggered mechanisms. Focusing on typical threats such as spoofing attacks, replay attacks, and denial-of-service attacks, this shift aims to improve system resilience under limited communication resources. In addition, defense strategies against complex threats, such as Byzantine attacks, are also gaining increasing attention.

Fourth, fault-tolerant control of formations under fault conditions has expanded from passive and active fault tolerance for single-actuator failures to scenarios involving multiple concurrent faults, such as communication link failures and pneumatic coupling. The integration of methods including observer techniques, adaptive approximation, and prescribed performance control provides effective means for enhancing cooperative reliability.

This section will systematically review the key methods, technological advances, and applicable conditions for formation maintenance from the four dimensions discussed above, offering guidance for the design and engineering application of cooperative control in UAV formations under complex environments.

4.1. Formation and Topology Maintenance

4.1.1. Formation Maintenance

Formation maintenance aims to maintain a specific spatial configuration of a UAV swarm dynamically. Control strategies can be classified into four major categories based on their reference frames: position-based, displacement-based, distance-based, and bearing-based, as shown in Table 4.

Displacement-based and distance-based methods typically rely on external positioning systems, whereas bearing-based control directly obtains relative bearing information between neighboring UAVs using only onboard sensors such as vision cameras [89] or radar [90], without the need for global position coordinates. This characteristic makes it more suitable for complex or obstructed environments where global positioning signals are unavailable [91]. In addition, bearing-based formation control imposes the lowest perception requirements on UAVs and is less affected by measurement noise, and has gradually become a research hotspot in recent years [92].

From the perspective of geometric constraints in formation flight, displacement constraints between UAVs remain unchanged only under translational motion; therefore, displacement-based methods can only achieve translational formation maneuvers. Distance constraints remain unchanged under both translational and rotational motion; thus, distance-based methods can achieve both translational and rotational formation maneuvers. Bearing constraints remain unchanged under translational motion and scaling; therefore, bearing-based methods can achieve both formation scaling and translational maneuvers [93].

The theory of bearing rigidity [94] provides the theoretical foundation for bearing-based control methods, offering a sufficient condition for determining the unique geometric configuration of a formation using only relative bearing information. In bearing-based methods, it is typically assumed that the desired formation is infinitesimally bearing rigid to ensure the global uniqueness of the formation configuration. Consequently, existing studies generally assume that the desired bearing angles of the UAVs remain constant [95,96,97], which makes it difficult to describe rotating formations. To address this limitation, ref. [98] established a unified method for describing maneuvering formations by introducing a rotation matrix to dynamically adjust the bearing vectors during formation rotation, thereby fully capturing translation, rotation, and scaling maneuvers.

To further enhance the maneuverability of formations and achieve coordinated control of scaling, translation, and rotation, affine formation maneuvering control methods based on tools such as the stress matrix [99] and the complex Laplacian matrix [100,101] have attracted increasing attention [102,103]. Ref. [104] investigated the dynamic affine formation tracking problem for six-degree-of-freedom underactuated multirotor UAVs, achieving the maintenance of a three-dimensional target formation. Ref. [105] extended the heading-based simultaneous localization and affine formation tracking control scheme to heading coordination control for underactuated UAV systems, where the desired relative headings between UAVs may vary over time. In addition, flexible formation control [106] allows followers to dynamically adjust their positions and orientations within a set of relative distance constraints, further improving the flexibility of formation maintenance while reducing system control energy consumption [107] and providing tactical advantages [108].

Table 4. Summary of research findings based on perceptual benchmarks.

	Reference	Geometric Constraint Description	Mobility and Flexibility
Location-based	[109,110,111]	Invariant only under translation	Only capable of translational formation maneuvers
Displacement-based	[112,113,114]
Distance-based	[115,116,117,118]	Invariant under translation and rotation	Achieve translation and rotation formations
bearing-based	[89,90,91,92,95,96,97,98,119,120]	Invariant under translation and scaling motions	Implement panning and zooming formations

Table 4. Summary of research findings based on perceptual benchmarks.

	Reference	Geometric Constraint Description	Mobility and Flexibility
Location-based	[109,110,111]	Invariant only under translation	Only capable of translational formation maneuvers
Displacement-based	[112,113,114]
Distance-based	[115,116,117,118]	Invariant under translation and rotation	Achieve translation and rotation formations
bearing-based	[89,90,91,92,95,96,97,98,119,120]	Invariant under translation and scaling motions	Implement panning and zooming formations

4.1.2. Topology Maintenance

For UAV formations, maintaining topological connectivity is a prerequisite for distributed formation control, reliable information exchange, and coordinated mission execution, and directly affects the overall system performance. However, maintaining topological connectivity in formations faces the following challenges. On one hand, during mission execution, the network topology of a formation often exhibits a complex and highly time-varying three-dimensional structure. Given the limited communication range and bandwidth of individual UAVs, it is difficult to maintain global and continuous topological connectivity for large-scale formations. On the other hand, under practical constraints such as GNSS denial [121] and limited field of view [122], traditional connectivity maintenance methods that rely on fixed topologies or omnidirectional perception often become ineffective.

Existing research has focused on designing distributed consensus protocols tailored to different communication topologies to enable formation control under local information exchange. Refs. [123,124] developed consensus-based formation protocols for static and switching topologies, respectively. Ref. [125] proposed two types of cluster consensus criteria for discrete-time systems with fixed and switching topologies. Ref. [126] introduced a novel multi-layer topological network and a consensus control method for the motion of large-scale UAV swarms under stable configurations.

However, implementing formation control typically requires certain constraints on the network conditions. For example, ref. [127] proposed a leader–follower protocol capable of generating various formations in undirected networks, but it requires that agents within each formation maintain joint connectivity. Ref. [128] implemented multi-formation control using an event-triggered leader–follower approach, yet still requires the existence of a spanning tree within each subgroup. These requirements introduce additional difficulties and challenges in the design of formation control laws.

A growing number of studies aim to achieve formation control under more relaxed and flexible topological connectivity conditions. By utilizing similarity matrix transformations and the Hurwitz stability criterion, ref. [129] provides sufficient conditions for system stability and a method for designing control matrices, significantly reducing the requirements for communication topological connectivity and making the approach applicable to real-world scenarios with weak communication conditions. Ref. [130] achieved, for the first time, formation control of multiple vertical takeoff and landing (VTOL) UAVs under a jointly connected switching topology, allowing for weak connectivity in the communication topology among UAVs. Ref. [131] permits the topology to degrade from continuous connectivity to periodic or intermittent connectivity. Yan et al. also achieved effective formation control under conditions where the initial network is disconnected [132,133]. Ref. [134] points out that ensuring formation stability requires only that the joint connectivity condition be satisfied. The joint connectivity condition requires that the network’s union over time remains connected, rather than being connected at every instant. However, such methods typically require a predefined topological structure for formation control [135]. Ref. [136] proposes a distributed connectivity preservation method that neither requires maintaining specific edge connections nor estimating algebraic connectivity. Moreover, its dynamic network graph is constructed based on the real-time states of the agents, rather than relying on a predefined fixed connectivity topology, thereby providing greater flexibility for formation maneuvers.

Currently, the state-of-the-art methods in this field achieve flexible connectivity maintenance by controlling the second smallest eigenvalue ${\lambda }_2$ of the graph Laplacian matrix. Refs. [137,138] show that the algebraic connectivity is positive if and only if the graph is connected. This type of method typically estimates the global algebraic connectivity and its gradient with respect to each agent’s position in a distributed manner, and then uses these estimates to design position controllers that keep the global algebraic connectivity positive [139,140]. Ref. [141] constructs a global potential function using ${\lambda }_2$, allowing interaction links to be dynamically added or removed according to mission requirements while ensuring the overall connectivity of the group. This maintains high flexibility in the group’s configuration and motion even under strong constraints. Similarly, Refs. [142,143,144,145] also design potential functions based on algebraic connectivity theory to preserve network connectivity.

However, as the size of the swarm increases, the computational cost of calculating ${\lambda }_2$ and estimating the gradients also rises, leading to higher demands on communication bandwidth and computational resources. Moreover, to ensure that ${\lambda }_2$ remains positive, the controller often needs to maintain a certain connectivity margin, which limits the flexibility of configuration adjustments to some extent.

4.2. Robust Formation in Complex Environments

In actual flight operations, formation systems frequently encounter complex environmental disturbances such as wind turbulence, GNSS denial, and obstacles.

4.2.1. Wind Disturbance

In UAV formation flight, wind disturbance is the most common external disturbance source, significantly affecting formation stability, tracking accuracy, and flight safety. In existing research, control methods for UAV formations to handle wind disturbances can be mainly divided into the following three categories: method based on disturbance estimation and compensation, method based on robust control theory, and method based on deep neural networks.

(1)

method based on disturbance estimation and compensation

This method estimates wind disturbances in real time by constructing a disturbance observer and then feeds the compensation back into the control loop, thereby mitigating the effects of wind disturbances [146]. The core of such methods lies in designing efficient observer and estimator structures [147,148]. Common structures include adaptive disturbance observers [149], extended state observers [150], and predefined-time unknown state estimators [151]. In addition, ref. [152] employs an unknown system dynamics estimator (USDE) to adapt to time-varying wind disturbances through simple filtering operations, avoiding the trade-off between decay rate and peak phenomenon inherent in traditional extended state observers [153,154]. Ref. [155] proposes a disturbance rejection control scheme based on multiple observers: the position loop uses a composite disturbance observer and an extended state observer to suppress disturbances from payload and wind, while the attitude loop uses an extended state observer to eliminate model uncertainties and wind disturbances. Such methods have strong adaptability, but their disturbance rejection performance depends on the convergence properties and estimation accuracy of the observers.

(2)

method based on robust control theory

This method relies on a highly robust control framework to mitigate the effects of wind disturbances through the controller’s inherent structure, rather than explicitly estimating the disturbances. Typical examples include sliding mode control and its variants [156,157,158], as well as $H\infty$ control [159]. Furthermore, ref. [160] employed a finite-frequency-domain $H\infty$ controller to enhance robustness and combined it with a proportional–integral–derivative (PID) controller to achieve stable UAV control under various wind conditions. Ref. [161] used acceleration-feedback-enhanced $H\infty$ control to counteract wind disturbances. Ref. [162] developed a cascaded robust controller to mitigate the effects of gusts. These methods demonstrate strong robustness when the system model is uncertain and the disturbances are bounded; however, they often suffer from control signal oscillations or excessive conservativeness.

Currently, a growing number of studies combine these two approaches to achieve the complementary benefits of precise compensation and robust stability, thereby overcoming the limitations of either method alone. Ref. [163] estimates gust disturbances using an adaptive super-twisting extended state observer and combines it with an adaptive super-twisting sliding mode controller to enable a quadrotor UAV to track a desired trajectory. Ref. [164] develops an active wind-resistant control architecture by integrating adaptive wind field compensation with a reference nonsingular terminal sliding mode controller, effectively suppressing external wind disturbances.

(3)

method based on deep neural networks

However, the two types of control methods mentioned above often rely on prior knowledge or conservative assumptions, and their adaptability to dynamic and time-varying wind fields is limited. In contrast, the data-driven deep learning-based control method is attracting increasing attention due to its strong nonlinear fitting and generalization capabilities [165].

Ref. [166] developed a deep learning-based trajectory tracking controller to quickly adapt to rapidly changing wind field conditions. Refs. [167,168] adopted a Bayesian neural network as a learning strategy to estimate disturbances. Ref. [169] designed a fault-tolerant controller by combining a disturbance observer (DO) with a radial basis function (RBF) neural network, effectively suppressing both integrated internal disturbances and external disturbances. Considering the complexity of multiple wind disturbances, including turbulence, wake vortices, and gusts, ref. [170] designed an MLP-based RBF neural network estimator (MRBFNNEs). In addition, ref. [171] used reinforcement learning and domain randomization to study robust wind-resistant hovering control for quadrotor UAVs. Ref. [172] introduced three wind field models—constant, time-varying, and strong gusts—into the learning environment and employed reinforcement learning algorithms to address the control problem of vertical takeoff and landing (VTOL) aircraft in wind-disturbed environments.

Nevertheless, the performance of such models depends heavily on the coverage and quality of the training data, making it difficult to ensure reliability in wind field scenarios that fall outside the training distribution. In addition, the lack of explicit modeling of the physical mechanisms of wind turbulence leads to poor model interpretability.

4.2.2. GNSS Denial

In environments where GNSS signals are blocked or disrupted (e.g., indoors, dense forests, urban canyons, or areas with strong electromagnetic interference), UAVs cannot rely on satellite navigation for precise positioning and global coordinate synchronization, posing significant challenges to formation maintenance. The primary difficulties include a significant decline in positioning accuracy, the accumulation of relative attitude estimation errors, and the inability to establish a unified global reference frame.

Current research on UAV formation flight in GNSS-denied environments mainly focuses on two areas: multi-sensor positioning and tight formation control. In terms of positioning, the primary approach relies on local sensor data—such as vision [173], LiDAR [174], ultrasonic [175], and ultra-wideband [176]—to overcome the failure of global positioning caused by the absence of GNSS signals. A relative positioning sensor system based on gyroscopes and accelerometers was proposed in [177]; although it suffers from cumulative positioning errors over time, it can still maintain the swarm topology for extended periods in GNSS-denied environments. In [178], ultra-wideband modules and inertial measurement units (IMUs) were used to measure relative distance and velocity, respectively, thereby solving the relative positioning problem for multiple UAVs under continuous excitation in GPS-denied environments. Passive positioning schemes relying solely on azimuth or relative distance are also applicable in special scenarios such as electromagnetic silence [179]. Most of these methods are based on relative positioning, constructing distributed pose estimation networks through local observations and communication with neighboring drones to achieve state consensus within the swarm without requiring a global coordinate system. However, the accumulation and drift of positioning errors remain difficult to eliminate fundamentally.

Furthermore, high-precision positioning is an important prerequisite for multi-UAV tight formation flight, enabling target tracking, formation maintenance, and collision avoidance. Tight formation control requires UAVs to fly in a geometric pattern with a lateral spacing of no more than twice the wingspan. However, in GNSS-denied environments, obtaining high-precision positioning data is challenging, which may lead to target loss or UAV collisions [180]. An integrated solution combining LiDAR-based relative positioning with DDPG reinforcement learning control was proposed in [181] to address the challenge of dense formation flight for vertical takeoff and landing (VTOL) aircraft in GNSS-denied environments. In [182], a novel adaptive dynamic programming–integral sliding mode control (ADP-ISMC) formation controller was developed, which achieves both robust and optimal formation control for fixed-wing UAVs based on a local coordinate system.

4.2.3. Obstacle Environment

In complex scenarios such as urban rescue operations and regional patrols, UAV formations inevitably encounter obstacles, which pose a significant threat to formation safety. Current mainstream methods primarily include potential field method [183], geometry-based method [184], optimization-based method [185], and learning-based method [186].

(1)

Potential field method

This method guides UAV movement by constructing attractive and repulsive fields [187]. It has the advantages of a simple structure and fast response, but inherent issues such as the local minimum problem and insufficient adaptability to dynamic obstacles [188] limit its application in complex environments. To address the local minimum problem in APF, a novel rotational potential field that encircles obstacles was proposed in [189]; this field is orthogonal to the repulsive potential field of the obstacles, ensuring that the net external force acting on the UAV is non-zero. In [190], a flux-guided method was proposed based on the property that the electric flux field is inherently smooth and free of local minima, to generate collision-free trajectories for multiple UAVs.

(2)

Geometry-based method

This method starts from the geometric relationship between the agent and obstacles, directly assessing collision risks and generating avoidance commands using techniques such as collision cones [191] and velocity obstacles [192]. The velocity obstacle algorithm analyzes the velocity relationship between the robot and obstacles to formulate avoidance plans based on predicted potential collisions over future time intervals; it is currently the mainstream method among geometry-based approaches. Building on this foundation, ref. [193] extended the two-dimensional velocity-based obstacle avoidance method to three-dimensional scenarios and combined the velocity-based strategy with the potential field method to resolve the singularity issue where UAVs oscillate at the edges of the obstacle protection zone in velocity-based avoidance control. In addition, methods such as reciprocal velocity obstacles [194], hybrid reciprocal velocity obstacles [195], and optimal reciprocal collision avoidance [196] have also been proposed and applied.

The geometry-based method has a simple structure, low computational cost, and fast response. However, it relies on high-precision perception of the surrounding environment, which is difficult to achieve in practice. In addition, this method only addresses collision avoidance and does not consider formation constraints.

(3)

Optimization-based method

This method represents obstacles and the controlled agent as inequality constraints and a cost function, respectively, and solves the constrained optimization problem using mathematical programming or meta-heuristic algorithms. Common approaches include convex optimization [197], control barrier function [198], model predictive control (MPC) [199], and receding horizon control [200]. Among them, nonlinear model predictive control (NMPC) incorporates formation control errors into the cost function, treats obstacle avoidance as a constraint, and solves the optimization problem iteratively within a finite prediction horizon, enabling better handling of dynamic obstacles. Ref. [201] predicted obstacle trajectories based on discrete control barrier functions, established safety constraints within the NMPC framework, and achieved optimal obstacle avoidance control for time-varying formations of underactuated UAVs. Ref. [202] used NMPC to minimize the deviation of UAVs from reference trajectories, thereby realizing formation control with obstacle avoidance capability.

(4)

Learning-based method

Learning-based method, particularly deep reinforcement learning (DRL), can integrate perception, planning, and control into a single model. This enables a direct mapping from sensor data to formation and obstacle avoidance decisions, allowing for data-driven joint optimization of the entire system [203]. To some extent, this addresses the relative lack of flexibility and adaptability of the first three types of methods when operating in complex obstacle environments.

However, the DRL method still faces challenges such as low sampling efficiency and difficulty in training within high-dimensional state spaces [204]. To address these issues, ref. [205] proposed an ensemble reinforcement learning framework that combines the strengths of the MADDPG and DDPG algorithms and employs a curriculum learning strategy during the training phase, decomposing the formation task into progressive stages to alleviate convergence difficulties in multi-agent algorithms. Ref. [206] proposed the 3A-MADDPG algorithm, which combines an attention mechanism with adaptive learning rates, significantly improving the convergence speed and control accuracy of large-scale UAV swarms. Ref. [207] proposed a flexible multi-UAV formation control method based on a model-data hybrid-driven framework, ensuring formation structural integrity through affine transformations and using DRL to maintain flexible formations under safety constraints.

In addition, as the scope of UAV applications and mission types continues to expand, a growing number of researchers are focusing on two key challenges: how to achieve robust and autonomous formation flight in obstacle-dense environments, and how to balance formation maintenance with collision avoidance.

To address the first challenge, ref. [208] proposes a comprehensive formation flight system that balances formation maintenance, obstacle avoidance, and dynamic constraints for large-scale UAV swarms in dense environments. Ref. [209] proposes a novel passage priority allocation method based on collision avoidance urgency, which is suitable for UAV formations operating in dense obstacle environments. Ref. [210] proposes a fluid-based path planning framework that treats the formation collision avoidance problem as the continuous flow of a set of parallel streamlines around obstacles, thereby addressing the collision avoidance problem for UAV formations in low-altitude, three-dimensional dense obstacle environments.

To address the second issue, ref. [211] developed a consensus-based reinforcement learning method that enables UAVs to balance obstacle avoidance and formation maintenance through coordinated local decision-making using a flexible strategy switching mechanism. Ref. [212] proposed a closed-loop, lightweight control strategy based on Lyapunov analysis, which reduces overly conservative obstacle avoidance behavior while maintaining safety. Ref. [213] designed an orthogonal artificial potential field algorithm and a gradient-mapping-based swarm collision avoidance method to achieve dynamic obstacle avoidance while preserving formation integrity. Ref. [214] integrated formation maintenance and collision avoidance into a unified learning task by combining imitation learning with reinforcement learning, and introduced LSTM to handle dynamic obstacles, thereby achieving efficient and safe formation control and obstacle avoidance in complex environments.

4.3. Security Formation Under Cyber Attacks

UAV formations rely on open wireless communication networks for information exchange, making them highly vulnerable to various malicious cyberattacks. Such attacks can destabilize the formation, cause trajectory deviations, and may even lead to mission failure. Based on the target and method of attack, cyberattacks against UAV formations can be categorized as follows:

4.3.1. Deception Attack

Attackers inject fabricated sensor or control data to mislead the UAV’s state estimation or control decisions, thereby compromising formation tracking accuracy. Such attacks are highly covert and relatively easy to execute, and their attack vectors cover the entire data chain from perception to execution.

Specifically, at the sensor level, this manifests as sensor spoofing attacks [215]; at the control channel level, it manifests as false data injection attacks (FDIAs) [216]; and at the actuator level, it manifests as actuator injection attacks [102]. Accordingly, ref. [217] investigated false data injection attacks on communication channels in multi-agent systems (MASs), while ref. [218] focused on FDIAs targeting sensors and actuators. Ref. [219] proposed a covert attack strategy that simultaneously spoofs both GPS and inertial sensors, and ref. [220] developed a covert attack strategy specifically targeting GPS and ultra-wideband positioning sensors in three-dimensional environments.

Most existing studies on countering spoofing attacks typically employ traditional methods that assume a predefined upper bound on the attack [221,222], and ref. [223] introducing a Bernoulli random variable to describe the occurrence of spoofing attacks under a given upper bound. However, for unknown systems, the tolerable upper bound against spoofing attacks must be obtained through estimation rather than predefined parameters. Ref. [224] reduces the conservatism associated with preset attack bounds by estimating the safety upper bound that a UAV formation can tolerate against spoofing attacks. Ref. [225] uses Lyapunov stability theory to derive sufficient conditions for a multi-UAV system under spoofing attacks to achieve a predefined formation. Ref. [226] proposes a novel method for detecting and countering spoofing attacks in UAV swarms. For the first time, ref. [227] simultaneously considers both desired trajectory attacks and spoofing attacks, transforming these attacks into an input signal tampering problem and proposing a compensation scheme based on an adaptive neural network observer (ANNO). However, although the above studies have conducted in-depth research on spoofing attacks, they do not address the probability of attack occurrence, which creates a certain gap from real-world operational conditions.

4.3.2. Replay Attack

Attackers intercept valid data packets from a past time and resend them, causing the system to make decisions based on outdated or incorrect information, leading to performance degradation. Such attacks are easy to implement and do not require knowledge of the system model. Moreover, replay attacks that leverage historical process data are covert and often difficult to detect.

While ref. [228] introduced a Bernoulli variable to model replay attacks, this approach overlooks the inherent temporal correlation and periodicity of such attacks and fails to capture their duration. Therefore, modeling replay attacks solely through a Bernoulli variable is not reasonable. Ref. [229] developed a novel replay attack model for multi-drone systems, dividing each attack cycle into a dormant phase and an active phase. During the dormant phase, data transmission proceeds normally; during the active phase, data packets from the current time slot are replaced with those from a historical time slot, thereby more accurately capturing the behavioral characteristics of replay attackers. Ref. [230] derived the upper bounds for the duration of each network attack and the interval between two consecutive attacks, quantifying the resilience of interconnected systems under replay attacks. Building on system resilience quantification methods, ref. [231] investigated resilient adaptive control for centralized nonlinear systems under replay attacks. Ref. [232] analyzed the degradation of distributed estimation performance caused by replay attacks and used Kullback–Leibler (K–L) divergence to quantify the stealthiness of the attacks.

4.3.3. Denial-of-Service Attack

By blocking communication links or consuming bandwidth resources, these attacks disrupt or delay the exchange of critical status information among UAVs. This compromises the real-time performance and consistency of formation coordination, ultimately leading to system instability [233].

A significant body of research has focused on addressing the impact of DoS attacks on system control [234,235]. Based on the assumption of a known DoS attack probability, ref. [236] uses a stochastic switching delay system to model random DoS attacks. Ref. [237] further investigates the security of heterogeneous multi-agent systems under conditions where the DoS attack probability is unknown. Ref. [238] derives sufficient conditions for a heterogeneous multi-agent system to achieve secure time-varying formation tracking under DoS attacks, as well as the feasibility conditions for the formation. Ref. [239] proposes a game-theoretic solution for distributed denial-of-service (DoS) attacks targeting UAV network hosts. Ref. [240] proposes a fixed-time link weight estimation mechanism to mitigate the impact of DoS attacks. Ref. [241] further distinguishes between long-term and short-term DoS attacks and analyzes the differential impacts of different attack types on system stability.

To improve the utilization of limited network resources, time-triggered control strategies, which are relatively easy to implement, are often the preferred choice for UAVs. However, such communication schemes may lead to information redundancy, resulting in resource waste or even network congestion, and make it more difficult to address the complex and severe challenges posed by the aforementioned malicious network attacks.

The Event-Triggered Scheme (ETS) [242] has attracted widespread attention in recent years due to its advantages over time-triggered schemes [243,244,245,246]. Specifically, ETS triggers actions based on predefined conditions, significantly reducing redundant data transmission and conserving limited network bandwidth. Meanwhile, its asynchronous and adaptive characteristics provide the control system with greater flexibility. When facing cyberattacks or external disturbances, the system can maintain formation performance by adjusting the triggering strategy, thereby enhancing system resilience.

Building on this foundation, improved variants such as the dynamic event-triggered mechanism (DETM) [247,248], memory-based ETM [249], hybrid triggering scheme [250], and adaptive event-triggered mechanism [251] have been proposed and applied. Furthermore, event-triggered mechanisms have been widely used in the design of defense strategies against various types of cyberattacks. Ref. [252] proposed a class-switching event-triggered control strategy for networked control systems under DoS attacks. Ref. [253] considered spoofing attacks and investigated the security control of T-S fuzzy systems based on a historical information event-triggered mechanism (ETM). Ref. [254] designed a novel switching event-triggered mechanism for multi-UAV systems, taking into account the start-stop switching characteristics of replay attacks within a sampling period.

In addition to the three common types of cyberattacks mentioned above, highly threatening composite attacks such as Byzantine attacks have also received increasingly extensive research attention in recent years. In essence, most of the aforementioned attacks can be classified as more general forms of Byzantine attacks.

As a typical example of composite attacks, Byzantine attacks can be regarded as a combination of Byzantine edge attacks (BEAs) and Byzantine node attacks (BNAs) [229]. By propagating erroneous information through unidentified nodes (BEAs) and tampering with their own inputs (BNAs), such attacks pose a serious threat to the stability and reliability of UAV swarms. Existing research can be divided into two types of approaches. The first consists of diagnosis-based defense schemes [255], which detect, identify, and isolate Byzantine agents through continuous diagnostic algorithms [256]. The second type is implemented using mean subsequence reduction algorithms [257,258,259,260]. Ref. [261] addresses Byzantine edge attacks and node attacks by decoupling the defense strategy into two independent tasks: suppressing Byzantine edge attacks at the DTL layer and suppressing Byzantine node attacks at the traditional cyber-physical layer.

4.4. Fault-Tolerant Formation Under Fault Conditions

When UAV formations operate in complex real-world environments, they are highly susceptible to various uncertainties, including actuator failures, sensor malfunctions, communication link interruptions, and external disturbances. These failures not only compromise the stability of individual UAVs but may also propagate through the cooperative network, leading to formation instability, reduced mission performance, or even collisions, thereby posing a serious threat to system safety and mission reliability. Therefore, designing efficient and reliable fault-tolerant control strategies that enable the formation system to maintain expected cooperative performance after a failure has occurred has become a critical challenge that urgently needs to be addressed in this field.

Current research addressing this issue can be broadly divided into two categories: passive fault tolerance and active fault tolerance [262]. Passive fault tolerance relies on the inherent robustness of controllers and algorithms [263,264], with typical examples including sliding mode control [265] and $H\infty$ control [266]. Ref. [267] proposed an adaptive sliding surface control method for actuator failures in multi-UAV systems. Ref. [268] proposed a robust H∞ fault-tolerant formation control method based on graph theory for fixed-wing UAV leader–follower systems. However, the fault tolerance of such methods is limited; when the deviation exceeds the robustness margin of the system, stability becomes difficult to guarantee. To expand the system’s fault tolerance capability, active fault-tolerant methods must be adopted [269].

Active fault-tolerant control typically includes a fault detection and diagnosis (FDD) module, which adjusts controller parameters or structure online based on fault estimation information to achieve precise compensation. This is usually accomplished using observer techniques [270,271,272] and adaptive approximation techniques [273,274,275].

Actuator failures are the most common and directly impactful type of fault in UAV formations. They include performance degradation (partial failure), jamming, offset errors, and complete failure. Such failures directly alter the control inputs, leading to disturbances in UAV dynamics and potentially causing formation instability or collisions. To address these failures, active fault-tolerant control primarily relies on fault estimation and adaptive compensation.

Refs. [149,276,277,278] designed various observers to estimate actuator failures online. Meanwhile, adaptive approximators such as neural networks and fuzzy systems have been widely used to learn the unknown dynamics caused by failures and provide compensation [279,280]. To address the challenges of performance control under actuator failures, refs. [281,282] combine prescribed performance control (PPC) with fault-tolerant control (FTC) methods to predefine the desired transient and steady-state performance metrics under actuator failure conditions. Ref. [240] proposes an adaptive estimation method that simultaneously identifies gain and offset fault parameters to handle more complex coupled actuator failures.

Communication link and network failures, such as communication interruptions and link failures, disrupt the information exchange essential for formation coordination, leading to the failure of distributed controllers. Addressing such failures mainly focuses on distributed state estimation and the design of resilient control protocols.

Refs. [283,284] use distributed state observers to handle communication link failures and demonstrate their resilience to such failures. Considering limited communication capacity, ref. [285] designs an event-triggered distributed observer to cope with the effects of communication link failures. However, most of the above studies focus only on a single type of failure in multi-agent systems. To address the formation control problem of fixed-wing UAVs under both actuator and communication link failures, ref. [286] proposes a fault-tolerant controller and a distributed leader state observer based on the theory of fully actuated systems, effectively solving the formation fault-tolerant control problem under coupled controller and communication failures.

Additionally, in close formation flight, it is essential to specifically address the aerodynamic coupling and safety risks caused by wake effects, and to design a composite controller that combines wake resistance with fault tolerance. Ref. [287] investigated modeling methods for wake effects in close formation flight of UAVs and proposed a continuous horseshoe vortex modeling approach. By adjusting the model parameters, this method can be flexibly adapted to UAVs with different wing configurations. Ref. [288] achieved, for the first time, fault-tolerant close formation flight under conditions involving external wake vortices, disturbances, internal actuator failures, and input saturation.

Moreover, real-world systems often encounter complex scenarios involving multiple concurrent or coupled failures. For such composite failures, a divide-and-conquer strategy is typically adopted. Ref. [289] addresses the problem of attitude synchronization tracking under a distributed control framework in the presence of actuator failures and wind disturbances. To handle actuator failures, communication constraints, and external disturbances, ref. [290] proposes a dual-channel dynamic event-triggered adaptive sliding mode fault-tolerant control law. Ref. [240] develops a resilient fault-tolerant cooperative control (RFTCC) method with an outer-inner loop structure, where the outer-loop controller effectively mitigates the impact of DoS attacks, while the parameters of actuator coupling faults are estimated and integrated through the inner-loop controller.

4.5. Discussion

To systematically compare the characteristics of different formation maintenance methods in addressing complex environments and challenges, this section summarizes and compares the aforementioned studies across dimensions such as communication dependency, computational overhead, support for heterogeneity, and validation maturity. It also summarizes the papers that have conducted physical experiments on the relevant methods, as shown in

Table 5. Analysis of methods for maintaining formation.

Scenario	Approach	Communication Dependency	Computational Cost	Heterogeneity Support	Validation Maturity	Supporting References (Physical/Semi-Physical Experiments)
Wind Disturbance	Disturbance estimation and compensation	Medium	Medium	Medium	High	[153,155]
	Robust control	Low	Low	Low to Medium	High	[157,158,162,163]
	Deep neural networks	Medium	Medium to High	High	Medium	[166,167,169,170,172]
GNSS-Denied	Multi-sensor fusion and relative positioning	Medium	Medium	Medium	High	[174,175,176,177,180,181]
Obstacle Avoidance	Potential field	Low	Low	Medium	Medium	-
	Geometric	Low to Medium	Low	Medium	High	[184,193,195]
	Optimization-based	Low	High	Medium to High	Medium	[198,199,202,209,210]
	Learning-based	Medium	Medium to High	High	Medium	[205,206,214]
Cyber Attacks	Attack modeling-based	High	Low to Medium	Medium	Low	-
	Attack detection and diagnosis	High	Medium	Medium	Low	-
	Active defense and resilient control	High	Medium to High	Medium to High	Low	[261]
Fault Conditions	Passive fault-tolerant control	Low	Low	Medium	Medium	[265]
	Active fault-tolerant control based on fault estimation	Medium	Medium	High	Medium	[269]
	Comprehensive fault-tolerant control for complex scenarios	Medium	Medium to High	High	Low	-

.

5. Formation Reconfiguration

UAV formation reconfiguration is a dynamic adaptive process in which the formation, driven by mission requirements or the need to adapt to environmental changes, employs distributed cooperative decision-making and real-time trajectory replanning to achieve a smooth and controlled transition of its spatial configuration and communication topology from one stable state to another specified state, thereby restoring, maintaining, and optimizing formation performance.

Current research mainly focuses on four typical scenarios. The first involves reconfiguration in response to dynamic changes in formation members, covering situations such as UAVs leaving the formation due to malfunctions, damage, or power depletion, as well as the addition of new members. The core challenge lies in quickly completing topology adjustments, position reassignments, and role reallocation while maintaining system stability.

Second, the reconfiguration of formations when navigating through dense obstacles and confined environments focuses on how the formation can safely pass through narrow passages and complex terrain by adapting its shape, and then quickly return to the desired configuration afterward. The challenge lies in balancing the inherent conflict between formation maintenance and obstacle avoidance.

Third, reconfiguration to maintain connectivity under dynamic switching of communication topologies: The core challenge lies in addressing communication link interruptions caused by interference, changes in distance, or node movement. This requires reconfiguring the topology and ensuring connectivity at minimal cost while meeting the requirements for the flow of formation control information.

Fourth, task-efficiency-driven active reconfiguration treats the reconfiguration process as a multi-objective optimization problem, seeking an optimal balance among metrics such as energy consumption, time, and task efficiency. Its core lies in efficiently solving optimal control problems subject to nonlinear and non-convex constraints.

This section will systematically examine the key methods, technological advancements, and applicable conditions for formation reconfiguration from the four dimensions mentioned above, providing theoretical support and engineering guidance for the adaptive cooperative control of UAV formations in complex environments.

5.1. Dynamic Variation in Member Quantity

In actual missions, when UAVs exit or join the formation due to damage, malfunctions, power depletion, or mission changes, the size of the formation changes, requiring adjustments to the formation configuration and a reassignment of tasks. The central issue is how to rapidly and smoothly adjust the formation topology, reassign positions, and re-elect roles such as the leader under dynamic changes in membership while maintaining system stability.

5.1.1. Changes in the Number of Formation Members

Changes in formation size can be attributed to two basic forms and their combinations: UAVs leaving the formation due to destruction, malfunction, or mission withdrawal, and new UAVs joining the formation or the formation being expanded due to mission changes.

To address scenarios involving a reduction in the number of members, ref. [25] proposed a task reassignment and formation reconfiguration algorithm based on the distributed Hungarian method. Ref. [291] designed a switching rule based on global consistency cost to determine whether to change the formation. However, its formation switching mode must be predefined, lacking the capability for dynamic autonomous reconfiguration.

To address scenarios involving an increase in the number of members, ref. [292] proposes a design and a distributed allocation protocol to solve the problem of formation reconfiguration when UAVs join the formation. For three scenarios, namely leader withdrawal, follower departure, and new member joining, ref. [293] proposes three formation reconfiguration strategies aimed at reducing the frequency of connection changes. However, these strategies lack flexibility when dealing with dynamic tasks.

5.1.2. Changes in Formation Leadership and Control

The leader–follower architecture is widely used in multi-UAV formation control due to its simple control logic and high theoretical maturity. However, it relies heavily on the formation leader. If the leader fails and exits the formation, the control architecture may collapse. Currently, a significant body of research has focused on the issues of dynamic leader re-election and control handover [38,294,295].

Among these, ref. [296] proposes a leader election mechanism based on kinematic information, but considers only a single information source. Ref. [293] introduces a hierarchical belief rule-based model that integrates kinematic and non-kinematic multi-source information to evaluate and elect the most capable new leader, thereby improving decision reliability. Ref. [297] proposes a voting-based leader election algorithm, while ref. [298] proposes a re-election method following leader failure in obstacle-dense environments. Ref. [299] proposes an emotion-driven dynamic leader selection model, in which followers influence the leader’s shame value by emitting dissatisfaction signals, thereby triggering a leader re-election process. However, its emotional parameters must be determined experimentally, and the model lacks an adaptive adjustment mechanism. Nevertheless, most of the above studies consider only single-leader scenarios, making them less suitable for larger-scale UAV formations.

In contrast, multi-leader formations achieve better scalability and robustness through a distributed cooperative leadership mechanism. Considering the context of multi-leader formations, ref. [300] proposes a comprehensive fault-tolerant leadership evaluation algorithm capable of selecting the most flexible leadership configuration while ensuring formation safety. However, the approach has limitations in terms of communication overhead and scalability.

Furthermore, to address dynamic and highly challenging problems, refs. [301,302] proposed a human–machine shared control mechanism that executes swarm tasks by dynamically allocating control authority. This mechanism offers greater efficiency advantages over purely manual control or fully automated systems in most application scenarios [303]. Ref. [304] established a cooperative model between manned aircraft and unmanned swarms under five levels of formation autonomy, and introduced the Stackelberg game model to address the issue of control transfer during dynamic missions. However, this study was limited to simulation verification with small-scale formations and assumed that manned nodes could fully obtain the status of the unmanned swarm, without adequately considering the numerous limitations on information acquisition in actual combat environments.

5.2. Dense Obstacles and Constrained Environments

In environments filled with obstacles, narrow passages, or other complex terrain, formations often struggle to maintain a fixed configuration during navigation and must dynamically adjust to accommodate environmental constraints. The core challenge lies in balancing the inherent conflict between formation maintenance and obstacle avoidance, enabling the formation to flexibly deform during navigation and quickly return to its original configuration afterward.

Ref. [208] suggests that an ideal formation flight system should possess three key capabilities: maintaining formation during obstacle avoidance, adjusting the formation distribution according to environmental constraints, and rapidly reorganizing the formation after an emergency. These capabilities are summarized as the PAPER criteria. This section mainly introduces research related to formation configuration adjustments during obstacle traversal and rapid recovery after passing through obstacles.

(1)

Method based on affine transformations

Affine transformations provide a mathematical framework for the overall geometric deformation of formations, enabling complex transformations such as rotation, translation, and scaling. This allows formation members to dynamically adjust their positions and orientations in response to environmental changes and mission requirements, making it a key method for adaptive reconfiguration in environments with dense obstacles and narrow constraints [305]. To implement affine transformations, ref. [306] explored the necessary and sufficient conditions for affine formations based on graph theory. Ref. [307] designed a controller based on the complex-valued graph Laplacian operator, capable of regulating the formation’s scaling ratio during maneuvers such as traversing corridors. However, existing research on affine formations based on stress matrices requires that the transformation parameters for translation, rotation, scaling, and shearing be pre-designed offline. To address this, ref. [308] proposed a novel online affine parameter adjustment method capable of achieving autonomous formation reconfiguration in multi-obstacle environments. However, this method is limited to fixed interaction topologies.

However, methods based on affine formation flight are mainly designed for environments with known static obstacles. When faced with unknown or dynamic obstacles, it becomes difficult to guarantee the safety of the formation [207].

(2)

Environment-Aware Formation Library method

This approach does not rely on a single fixed formation. Instead, it dynamically selects the most feasible configuration from a predefined formation library based on real-time environmental data. Using a local sensing-based obstacle-free maximum passable width detection algorithm, ref. [309] dynamically matches the optimal configuration from a predefined formation library, significantly reducing the risk of collisions in scenarios with dense obstacles and narrow spaces. However, this method assumes that all robots can share position and environmental information in real time, making it difficult to apply to large-scale robot teams or scenarios with limited communication. Ref. [310] allows the swarm to switch between square and linear formations to adapt to environmental constraints and dynamic changes, but the formation is highly dependent on the leader and lacks a fault-tolerance mechanism. Ref. [311] designed and implemented a self-reconfiguring V-formation controller that dynamically adjusts its configuration, such as by opening and closing wings or merging into a straight line, when encountering obstacles. However, the triggering and coordination mechanisms for these reconfiguration behaviors rely on preset rules, limiting flexibility. In addition, the adaptability of the V-formation’s geometric parameters has not been fully explored.

This class of methods provides immediate and efficient responses and can clearly align with the physical constraints of the environment. However, this flexibility is limited by the comprehensiveness of the predefined formation library. In addition, constructing such a library requires prior knowledge, making it difficult to handle completely unknown or highly complex environmental conditions.

(3)

Method based on swarm intelligence

The core idea of this strategy originates from observations of highly agile biological groups such as bird flocks and fish schools. When encountering obstacles or predators, these groups can temporarily disperse, evade threats through simple local interaction rules among individuals, and then rapidly and orderly regroup to restore the formation [312]. This class of methods allows the formation to temporarily disperse when necessary, enabling individual agents or subgroups to avoid obstacles before subsequently reassembling.

Ref. [208] proposes a global-remap-local-replan (GRLR) strategy. To address the difficulty of the entire formation traversing narrow areas, the large swarm is divided into several sub-swarms with greater maneuverability. Each sub-swarm uses a local replanning module and a reactive obstacle avoidance algorithm to autonomously navigate through obstacles. After passing through the bottleneck area, all individuals regroup and return to the desired formation configuration. Ref. [313] designs a split-and-merge control strategy based on pigeon flock behavior, integrating three mechanisms: obstacle avoidance, inter-vehicle collision avoidance, and formation reconfiguration. Based on the behavioral characteristic that pigeon flocks switch between hierarchical and egalitarian interaction modes during different flight phases, ref. [314] proposes a pigeon flock formation model and a cooperative obstacle avoidance model to coordinate heterogeneous UAV swarms through obstacle-laden environments with only a limited number of information-bearing individuals. Ref. [315] proposes a Vector Field Histogram (FEB-VFH) method based on fish escape behavior, which is adaptable to different formations and complex scenarios.

This class of methods offers good robustness and flexibility, enabling them to cope with highly dynamic and unknown obstacles. However, during the transient phase, the formation structure becomes loose, which may lead to a temporary loss of overall functionality. In addition, such methods place high demands on the stability and convergence speed of distributed coordination algorithms.

5.3. Topology Switching and Reconfiguration

Communication links between UAVs may be interrupted or experience degraded quality due to communication interference, changes in distance, or node movement. Therefore, the network topology needs to be dynamically reconfigured to maintain essential connectivity. The primary difficulty is how to reconfigure the topology with minimal communication overhead or via the shortest path while meeting the information flow requirements for formation control, and avoiding new formation instability caused by the reconfiguration process. To address these challenges, various dynamic topology reconfiguration mechanisms have been proposed in the academic community.

(1)

Methods based on graph theory and combinatorial optimization

This class of methods transforms topology reconfiguration into a network optimization problem, aiming to find an optimal or near-optimal connectivity graph under specific constraints. Ref. [316] proposes a reconfigurable hierarchical framework that specifies how to maintain the overall bearing rigidity and hierarchical structure of the network through local edge addition and deletion operations when robots join, leave, or when subgroups merge or split. This provides a rigorous mathematical foundation for topology reconfiguration. Ref. [317] proposes a shortest connection path optimization strategy that reduces communication overhead during communication failures through optimal topology design and distributed reconfiguration algorithms. Ref. [318] introduces an optimized topology reconfiguration strategy that incorporates evaluation metrics for formation stability, communication efficiency, and structural symmetry. Ref. [319] further proposes an algorithm based on Edmonds’ minimum-cost tree, achieving unified optimization of communication topology and formation. Ref. [320] proposes a method to identify vulnerable links by analyzing network community structure and node betweenness centrality, and to enhance network resilience globally by selectively adding a small number of super-links.

(2)

Methods based on performance evaluation

Most of these methods focus on reactive recovery after a failure occurs. Among them, ref. [321] combines importance measures with resilience analysis to identify critical nodes and links in the topology. When a failure occurs, the reconfiguration strategy prioritizes restoring those connections that have the greatest impact on the overall resilience of the system. From the perspective of recovery efficiency, ref. [322] establishes a relationship between algebraic connectivity and the convergence rate of consensus, guiding UAVs to selectively establish new links to maximize system performance as quickly as possible.

(3)

Self-healing control method

In addition, self-healing control is widely used in the repair and reconstruction of network topologies. Currently, the three most common self-healing control methods are direct self-healing [323], density-based self-healing [324], and recursive self-healing [325].

In particular, direct self-healing method uses predefined rules to restore the topology by directly replacing failed nodes with healthy ones [326]. This method usually relies on a highly centralized control architecture, making it difficult to apply in distributed formations.

The density-based method simulates physical diffusion processes and aims to restore network topology by rebalancing node density [327]. However, this method struggles to handle large-scale topological networks and exhibit limited control performance.

The recursive self-healing method repairs the topology through cascading, hop-by-hop position adjustments [328]. This method typically relies on neighborhood state information within a hyperlocal scope and is not suitable for fully distributed systems that strictly depend on local interactions. It attempts to overcome this limitation by proposing a fully distributed gradient self-healing algorithm that uses a gradient field to estimate distances and determine optimal repair paths. However, this algorithm imposes specific constraints on the network topology, and its general applicability cannot be guaranteed. Ref. [329] proposes a hierarchical distributed recursive self-healing algorithm that establishes node importance assessments through local interactions and recursively selects nodes for repair, enabling rapid recovery of cluster topology and formations under single-point and multi-point failures. However, its hierarchical assessment relies on static degree and distance information, and the optimality of its greedy repair strategy in resource-constrained or dynamic environments remains unverified.

5.4. Performance-Optimized Active Reconfiguration

To improve mission execution efficiency and meet higher mission performance requirements, formations need to proactively reconfigure their structure or topology. Existing research typically formulates the formation reconfiguration process as an optimization problem. Ref. [330] was the first study to address the formation reconfiguration problem with respect to minimum fuel consumption, providing a classic framework for subsequent research. Since then, optimization problems for reconfiguration based on various mission performance metrics have been extensively studied.

In performance-driven active reconfiguration, the existing research mainly focuses on two dimensions: minimizing time and minimizing energy consumption. Refs. [331,332,333,334] formulate formation reconfiguration as a time-optimal control problem with the goal of minimizing the time required to complete the configuration change. However, pursuing time optimization alone often results in UAVs flying at maximum speed or frequently accelerating and decelerating, which leads to significantly higher propulsion energy consumption than energy-optimized strategies.

It is worth noting that the limited capacity of onboard batteries directly determines the feasibility boundaries of reconfiguration strategies and constrains the optimization of overall mission performance. Consequently, some studies prioritize energy optimization as the primary objective. Ref. [335] proposes a two-stage energy-optimized reconfiguration strategy that pre-deploys agents into specific formations during mission downtime, thereby simultaneously minimizing both the energy consumption of the current reconfiguration and the expected energy consumption of future mission transitions. Inspired by the V-formation used in goose migration, ref. [336] achieves load balancing through a UAV position rotation algorithm, thereby improving the overall energy efficiency of the swarm. However, this approach may prolong reconfiguration time and affect mission timeliness.

In practical applications, formation reconfiguration requires balancing multiple performance metrics. Ref. [337] simultaneously optimized system resilience, reconfiguration time, and coverage area during formation reconfiguration after a random attack, demonstrating the feasibility of multi-objective joint optimization. Ref. [338] minimized the expected values of reconfiguration energy and time consumption. Ref. [339] addressed both collision risk and endurance by targeting the shortest time and minimum fuel consumption, respectively. Ref. [340] achieved a balance between communication throughput and energy consumption through three-dimensional trajectory design and an energy-efficient reinforcement learning framework. In addition, ref. [341] conducted research aimed at maximizing global search performance, while ref. [342] focused on minimizing total travel distance and reducing the computational cost of formation reconfiguration.

For the optimization problem described above, the optimization variables are the control input time series for each UAV. The objective function typically includes performance metrics such as reconfiguration time and energy consumption, while also satisfying constraints related to actuator inputs, airframe dynamics, and collision avoidance [343]. Existing solution methods can be broadly divided into three categories: control parameterization and time discretization [344], receding horizon control [345], and heuristic intelligent algorithms [346].

(1)

Parameterization and Time Discretization

This method transforms the original optimal control problem into a discrete optimization problem that is easier to solve by parameterizing the continuous control inputs and discretizing the time variable into finite intervals [347]. It is often combined with intelligent solution algorithms [348,349]. Ref. [350] employs a control parameterization and time discretization method based on the Bat Algorithm to solve for the time-optimal control law for formation reconfiguration in a multi-robot system. A limitation of such methods is that modeling formation reconfiguration as a global optimal control problem leads to excessive computational complexity, which significantly increases solution time and makes it difficult to adapt to tasks requiring rapid response.

(2)

Receding Horizon Control

This method achieves adaptive control under constraints through rolling optimization within a finite time domain [351]. Ref. [331] proposes a receding horizon control (RHC) method based on differential evolution, which decomposes the global reconfiguration of a multi-UAV formation into a series of online local optimization problems. Ref. [343] further decouples the reconfiguration process into two stages, namely task allocation and control input optimization. It employs adaptive hybrid particle swarm and differential evolution algorithms to solve the task allocation problem and uses RHC to decompose the global control to reduce computational cost. Although such methods possess online adaptive capabilities, existing research has not yet comprehensively addressed constraint conditions. In addition, the optimization process is prone to getting stuck in local optima, which limits their practical performance in complex environments [352].

(3)

Heuristic intelligent algorithm

This class of methods does not rely on gradient information and achieves parallel exploration of the search space by simulating the behavioral patterns of biological populations in nature. It is suitable for solving complex optimization problems such as nonlinear, non-convex, and mixed-integer optimization. The core advantage of these methods lies in their ability to directly handle discrete control variables and complex constraints, thereby avoiding the strong dependence on problem structure inherent in traditional numerical optimization methods. However, the limitations are as follows. The balance between global search and local exploration is difficult to control precisely, which can lead to issues such as premature convergence or low search efficiency. In addition, algorithm performance is highly sensitive to parameter settings.

To address the aforementioned issues, the existing research has primarily focused on two directions: improving algorithmic structures and designing hybrid strategies. Ref. [353] combines control parameterization and time discretization methods with an improved particle swarm optimization algorithm. It dynamically adjusts particle velocities through mutation and escape operators to reduce the risk of getting trapped in local optima. Ref. [354] proposes an adaptive perceptual Cauchy mutation-based pigeon swarm optimization algorithm. It introduces adaptive weighting factors and Cauchy mutation operators and uses a roulette wheel selection mechanism to balance global search and local exploration.

While these improvements have alleviated the problems of premature convergence and parameter sensitivity in heuristic algorithms for formation reconfiguration to some extent, the computational efficiency of such methods remains constrained by population size and the number of iterations. In addition, further optimization is needed regarding constraint handling mechanisms.

5.5. Discussion

To systematically analyze the characteristics of different reconfiguration methods across various task scenarios, this section summarizes and compares the aforementioned studies from dimensions such as communication dependency, computational overhead, support for heterogeneity, and validation maturity. It also summarizes the papers that have conducted physical experiments on these methods, as shown in

Table 6. Analysis of formation reconfiguration methods.

Scenario	Approach	Communication Dependency	Computational Cost	Heterogeneity Support	Validation Maturity	Supporting References (Physical/Semi-Physical Experiments)
Dynamic Variation in Member Quantity	Leader Election and Switching Mechanism	Medium-High	Medium to High	Low	Medium	[25,298,300]
	Human–Machine Shared Control	High	High	High	Low	[301,302]
Dense Obstacles & Confined Environments	Affine transformation-based	Medium to High	Medium to High	Low	Low	[208]
	Environment-aware formation library matching	High	Low to Medium	Medium	Medium	[309,310]
	Swarm intelligence-based	Low	Low	Medium to High	Medium	[313]
Communication Topology Switching & Connectivity Maintenance	Graph theory & combinatorial optimization	High	High	Medium	Low	[318]
	Performance evaluation-based	Medium	Medium	Medium	Low	-
	Self-healing control & distributed repair	Low to Medium	Low to Medium	Medium	Medium	[324,326]
Task Efficiency-Driven Optimization	Control parameterization & time discretization	Low	High	High	Low	[332,350]
	Receding horizon control	Medium to High	Medium to High	High	Low	-
	Heuristic intelligent optimization	Low	Medium to High	High	Low	-

.

6. Formation Performance Evaluation

The evaluation of UAV formation performance refers to a systematic process in which, under specific mission scenarios, a feasible multidimensional metric system is established for UAV formation systems through the comprehensive application of theoretical modeling, numerical simulation, hardware-in-the-loop simulation, and flight testing. This process enables quantitative calculations and qualitative analysis of system performance. Current research primarily evaluates the control performance of UAV formations from four dimensions: resilience, robustness, reliability, and vulnerability.

6.1. Resilience

The concept of resilience was originally used specifically to describe the resilience of ecosystems [355]. However, through continuous application and expansion by experts and scholars, it is now used across various fields such as sociology, psychology, and engineering [356,357,358].

In the field of UAV formations, there is currently no unified or universally accepted definition of resilience. The existing research tends to interpret the concept from different perspectives depending on the specific problem context. Ref. [359] emphasizes the system’s ability to withstand disturbances, while ref. [360] focuses on the dynamic mechanisms through which a system responds to disturbances and restores its performance. Refs. [361,362] provide interpretations from the perspectives of phase segmentation and capability decomposition, respectively, but these definitions also overlap to varying degrees with established concepts such as robustness.

Based on a comprehensive review of existing research, this paper defines the resilience of a UAV formation as its ability to restore functionality through means such as dynamic reconfiguration when system performance is degraded by various internal and external disturbances, thereby maximizing the fulfillment of mission requirements. At its core lies the dynamic process of recovery after an attack or disturbance.

Existing research on the resilience assessment of UAV formations mainly concentrates on the following aspects:

(1)

Modeling and evaluation method based on complex networks and graph models

This method models a UAV swarm as a network, where nodes represent UAVs and edges represent communication or collaboration relationships, and evaluates resilience by analyzing the network’s topological characteristics. In particular, Cheng et al. designed a resilience evaluation framework for joint reconnaissance missions based on complex network theory [363]. However, this study primarily focuses on homogeneous UAV swarms.

For heterogeneous formations, ref. [364] used resilience analysis to evaluate the criticality of heterogeneous unmanned system-of-systems (USoS). Ref. [365] further proposed a baseline resilience evaluation method based on heterogeneous graph communication networks, overcoming the limitations of homogeneous models. Building on this foundation, several studies have sought to expand the analytical scope. Ref. [366] introduced a dependency network model to analyze the interdependencies among different networks and their impact on resilience. Ref. [367] focused on the dynamic evolution of resilience under node attacks.

In summary, this method abstracts complex cluster interactions into quantifiable graph structures, making them suitable for assessing the structural resilience of clusters, though its ability to characterize task resilience is relatively limited.

(2)

Quantitative method based on performance trajectories and metrics

This method transforms resilience assessment into a quantitative characterization of the dynamic evolution of system performance. It comprehensively evaluates resilience by constructing performance–time curves and calculating metrics such as the area under the curve, recovery rate, and coverage ratio.

Ref. [368] defines the resilience metric as the ratio of the area under the performance curve after a failure to the area under the initial performance curve, thus quantifying the system’s ability to resist interference and recover during the mission replanning phase. Other researchers have used the difference between current performance and target performance to reflect the system’s adaptability and recovery capabilities [369]. Since most existing studies focus only on the time dimension while neglecting the influence of the spatial distribution of UAV swarms, ref. [370] established a spatiotemporal methodology for evaluating UAV swarm resilience, enabling quantitative assessment of swarm resilience under arbitrary spatiotemporal scenarios. Ref. [371] introduced the concept of the largest recoverable component (LRC) to evaluate the fault recovery capability and individual importance of swarm systems from the follower’s perspective. However, this metric does not yet adequately cover leader failure scenarios or heterogeneous interaction relationships.

Many existing studies typically assume that the performance of a UAV swarm must return to its original state to be considered satisfactory. However, during actual mission execution, it is neither necessary nor guaranteed that the swarm’s performance will fully recover. Consequently, some studies have begun to emphasize the concept that restoring a new, stable, and consistent state [322] is sufficient, rather than requiring a complete return to the initial performance level [372]. From a mission-oriented perspective, the mission baseline is defined as the minimum performance level required to complete the operational task [373]. Furthermore, the introduction of a baseline evaluation mechanism, which dynamically adjusts the mission baseline through adjustment factors, has relaxed the rigid constraints of the original mission baseline.

This method transforms resilience from an abstract concept into quantifiable performance trajectories, providing intuitive and practical tools for comparing and optimizing resilience. Furthermore, compared to evaluation method that relies solely on topological structures, it is more closely aligned with real-world mission scenarios. However, the selection of performance metrics directly affects the validity of evaluation results, and existing research still offers relatively limited definitions of these metrics. Although the approach to setting task baselines has shifted from rigid to flexible, there remains a lack of unified standards regarding the validity of these baselines and the mechanisms for adjusting them.

6.2. Robustness

Robustness is often used as a key metric to evaluate a system’s ability to maintain basic functionality when facing internal failures or external attacks. It is a core factor that must be considered in system optimization [374] and is widely applied in fields such as network intrusion detection [375], autonomous driving [376], and power grid dispatch [377]. Ref. [378] defines the robustness of a UAV swarm as its ability to maintain functionality and complete tasks even after losing some of UAVs.

The robustness of a UAV formation can be defined as its ability to maintain core functions and overall mission execution capabilities without significant degradation when facing failures of certain nodes or connections, external attacks, or environmental disturbances. At its core lies the system’s ability to maintain a stable state even under attack.

Current research on the robustness of UAV formations is relatively limited. Given that the similar characteristics and dynamic evolutionary relationships within UAV swarms can be modeled using complex network theory, and such complex networks exhibit small-world and scale-free properties [379], they can be regarded as multi-layered complex networks with multifunctional interconnections. Therefore, analysis can be conducted by drawing on approaches used to evaluate the robustness of complex networks.

The core of robustness in complex networks lies in analyzing how the network responds when nodes or edges are removed [380]. Since the ability of a UAV formation to complete a mission in the event of individual failures depends heavily on the integrity of the underlying communication network, its robustness can be characterized by analyzing how the topological structure evolves during the removal of nodes or edges.

In related research, ref. [381] investigated the cascading failures caused by the removal of interdependent nodes in multilayer networks. Christian proposed a node robustness metric to measure the size of the largest connected component in a swarm network under attack [382]. Ref. [383] used a latent importance metric to perform targeted node removal and analyzed its impact on different network structures. Ref. [384] simulated random failures through random node removal and malicious attacks through the removal of highly connected nodes, comparing the robustness differences between random networks and scale-free networks, using connectivity as the evaluation criterion. Zeng proposed a connectivity robustness metric to assess network robustness under malicious edge attacks [385].

To address these limitations, previous studies have made valuable explorations. Ref. [379] developed a comprehensive robustness evaluation metric for UAV swarms based on three-layer complex network theory. They introduced evolutionary mechanisms such as dynamic reconfiguration and information association, and proposed corresponding evaluation methods and algorithms. However, this model provides a rather idealized description of the coupling relationships between network layers, and its dynamic evolution rules do not fully account for practical communication constraints and resource competition. Therefore, further research is needed to develop more flexible modeling methods. Ref. [386] constructed a set of robustness metrics across three dimensions, namely swarm attributes, environmental attributes, and mission performance. By combining the Shapley Additive Explanation (SHAP) with Extreme Gradient Boosting Trees (XGBoost), an intelligent quantitative evaluation model with interpretability was built, offering a new approach that incorporates machine learning for multidimensional metric fusion in complex scenarios. However, the interpretability of the evaluation results remains limited by the quality of the feature engineering, and the model exhibits strong dependence on the distribution of the training data. Its generalization ability across different task scenarios requires further validation.

6.3. Reliability

Reliability refers to the ability of a system to perform its intended functions under specified conditions and within a given period of time [387]. The mission reliability of a UAV swarm is defined as the swarm’s ability to complete its assigned mission under specific operating conditions and within a specified time frame [388].

The reliability of a UAV formation can be defined as the probability that the formation will complete its intended mission without failure and maintain its structural integrity under specified mission profiles and operating conditions. At its core lies the assurance of the system’s long-term fault-free operation. Existing research primarily employs binary decision diagrams and k-out-of-n system theory for evaluation.

(1)

Binary Decision Diagram (BDD)

Binary Decision Diagrams (BDDs) are widely used in the reliability assessment of multi-stage tasks [389]. Ref. [390] proposed a reliability analysis method for multi-stage task systems based on BDDs and further extended this method for mission reliability prediction [391]. Ref. [392] improved BDDs to enhance the computational efficiency of reliability analysis for multi-stage task systems. Ref. [393] used BDDs to model the performance of UAVs at each stage, enabling dynamic reassessment of failure probabilities during mission execution. Ref. [394] also proposed a BDD-based reliability analysis method for multi-stage UAV missions, but their study was limited to a single UAV.

The BDD-based methods described above offer high modeling accuracy for mission reliability across different phases and can effectively capture the impact of mission temporal logic on system reliability. However, such methods often suffer from the problem of state space explosion. As the system scale increases or the number of mission phases grows, the number of BDD nodes increases exponentially, resulting in a sharp increase in computational complexity.

(2)

The k-out-of-n (k/n) system theory

On the other hand, research on k/n systems and their variants designed to meet phased mission requirements has been quite extensive [395,396]. Ref. [397] established a reliability model and estimation method for multi-tiered balanced UAVs in continuous systems, and systematically explored the matching problem between UAV swarms and continuous k/n:F/G systems or continuous (r,s)/(m,n):F/G systems. Ref. [398] analyzed the joint reliability significance of correlated components in multiple sets of continuous k/n:F systems based on Markov chains; ref. [399], on the other hand, established recursive formulas that account for incomplete fault coverage for k/n systems with identical components, phased mission requirements, and non-perfect fault coverage, based on the law of total probability and conditional probability, thereby achieving an accurate assessment of overall mission reliability.

In system optimization design, the identification of critical components plays a crucial guiding role. Through importance analysis, critical individuals within a UAV swarm can be identified. Based on a continuous k/n:F system, ref. [388] established a mission reliability model for polygonal formation UAV swarms. The paper presents the Birnbaum importance measure and a composite importance measure. By integrating these two importance metrics, the study analyzes the importance measures of UAVs at different positions within the swarm, thereby identifying high-importance UAVs and locating critical nodes and vulnerabilities. Ref. [400] models the swarm mission as a k/n system with phased requirements. A method for calculating mission reliability is presented, and based on this, several importance metrics are proposed to describe the impact of changes in the number of UAVs on mission reliability.

The k/n system theory provides a concise and effective modeling framework for UAV swarms with redundant structures and is particularly suitable for typical mission scenarios such as “at least k nodes must remain operational.” Its importance analysis method can effectively identify critical nodes within the system, providing a basis for decisions regarding redundant configurations and protection strategies. However, due to its strong assumptions regarding component homogeneity and failure independence, it is difficult to directly apply this theory to swarm scenarios involving heterogeneous nodes or those with functional dependencies.

6.4. Vulnerability

Vulnerability is an inherent property of a system. A system is considered vulnerable when external disturbances cause damage or performance degradation [401]. Compared with resilience, robustness, and reliability, vulnerability places greater emphasis on the structural and functional sensitivity of UAV swarm systems when subjected to attacks or failures.

The vulnerability of UAV formation can be defined as the extent to which the formation’s structural integrity, functional connectivity, and mission effectiveness collapse or degrade significantly when subjected to specific attacks or failures. The core of this concept lies in identifying the weak points of the system, specifically which nodes or links, if failed, would have the greatest impact on the overall system.

Existing research on vulnerability assessment for UAV swarms is relatively limited. Ref. [402] constructed a two-layer, multi-edge complex network model to characterize communication relationships and mission coordination relationships. It designed a vulnerability analysis framework that includes a fault analysis module and proposed two methods for calculating vulnerability metrics. Ref. [403] introduced five vulnerability metrics, namely node integrity, edge integrity, social node functionality, hub node functionality, and hub edge functionality. Based on these, an evaluation system comprising ten metrics was subsequently established, drawing from both network structure and functionality, to quantify the vulnerability of manned-unmanned cooperative combat networks.

In terms of evaluation methods, ref. [404] proposed a node-influence-based evaluation method to measure the vulnerability of UAV swarm communication networks in jamming environments. Ref. [405] used the Coupled Map Lattice (CML) model to integrate topological metrics such as node degree, distance between nodes, and clustering coefficient, thereby reflecting changes in node states under interference scenarios from a topological perspective. They employed relative network efficiency and failure rate as measures of vulnerability. Ref. [406] constructed an information and communication interdependence network model for vulnerability analysis of UAV swarm networks in emergency response missions.

The aforementioned studies have expanded the analytical framework for assessing the vulnerability of UAV swarms by introducing approaches such as network modeling and indicator systems. However, most of these methods focus on analysis under static topologies or predefined failure modes. Furthermore, there is a lack of systematic methodologies to support the translation of vulnerability assessment results into decision-making processes such as system optimization and defense strategy formulation.

6.5. Discussion

The four dimensions of resilience, robustness, reliability, and vulnerability together form a complementary framework for evaluating the performance of UAV formations. They characterize the overall capabilities of the formation system in complex environments from different perspectives. Clarifying the distinctions and interrelationships among these four dimensions is essential for establishing a scientific evaluation system.

(1)

Key Differences

The essential distinction among these four dimensions lies in the different system states and temporal phases they emphasize. Reliability focuses on the statistical characterization of a system’s long-term stability and serves as a core metric for pre-event assessment. Robustness emphasizes the system’s immediate ability to withstand disturbances, while resilience focuses on the dynamic process of recovery from a degraded state to an acceptable level of performance. Vulnerability focuses on identifying the critical nodes and weak points that may lead to a sudden collapse in system performance. Together, these dimensions cover the full lifecycle of UAV formation mission execution. A detailed comparison is provided in

Table 7. Comparative analysis of the four dimensions of performance evaluation.

Dimension	Core Focus	Temporal Phase
Reliability	probability of failure-free operation	pre-event
Robustness	ability to maintain functionality under disturbances	in-event
Resilience	ability to recover from failures	post-event
Vulnerability	sensitivity of system structure and functionality	full lifecycle

.

(2)

Internal Relationships

The four dimensions mentioned above form a mutually reinforcing logical relationship in the evaluation of UAV formation effectiveness, which can be explained from the following three perspectives.

①

Progressive Relationship Based on the Time Dimension

From the perspective of the temporal progression of disturbances, the four dimensions form a complete mission cycle consisting of pre-event, during-event, and post-event phases. Reliability assessment provides probabilistic assurance of system stability before mission execution. Robustness assessment measures the system’s ability to maintain functionality during disturbances. Resilience assessment evaluates the system’s ability to recover to an acceptable level of performance after disturbances. Vulnerability, meanwhile, runs through the entire process and provides structural sensitivity information for assessments at each stage.

②

Complementary Relationships Based on Functional Dimensions

The four dimensions complement each other functionally and together form a comprehensive evaluation framework. Evaluation based on a single dimension cannot fully capture the overall capabilities of a formation system, and the integration of multiple dimensions is a prerequisite for scientific assessment.

③

Closed-loop Relationship Based on the Design Dimension

At the level of system design and optimization, the four dimensions form a logical chain of closed-loop iteration: identification, design, evaluation, and optimization. Vulnerability analysis identifies critical nodes and weak points within the system, thereby providing directions for optimizing reliability. The level of reliability determines the robustness boundary of the system. When disturbances exceed this robustness boundary, resilience mechanisms compensate for functional failures through reconfiguration and recovery. The effectiveness of resilience design and reconfiguration strategies can in turn be verified and optimized through vulnerability analysis.

In summary, the evaluation of UAV formation performance should be based on a comprehensive assessment framework that integrates robustness, resilience, reliability, and vulnerability, as shown in Figure 9.

7. Conclusions

From a mission-driven perspective, this paper analyzes the entire process and full lifecycle of UAV formation mission execution. It systematically examines key aspects including UAV formation modeling and representation, formation assembly, formation maintenance, formation reconfiguration, and performance evaluation. The study finds that although current research in UAV formation control theory and applications has made significant progress, several challenges still remain to be addressed, and the future prospects for related developments are promising.

7.1. Research Gaps and Challenges

(1)

Insufficient scalability and real-time autonomy in algorithmic computations

Existing formation control methods generally face the problems of high computational complexity and slow convergence when dealing with large-scale heterogeneous swarms. Although consensus-based methods have good theoretical scalability, their strong dependence on communication quality limits their applicability in complex environments. Optimization-based methods have limited capability for online replanning in dynamic scenarios, making it difficult to meet real-time response requirements. Furthermore, existing research has not yet fully developed coordinated handling mechanisms for composite disturbances and composite failures, and there remains a lack of a unified, resilient architecture capable of learning disturbance patterns online and adaptively adjusting control strategies.

(2)

The evaluation of effectiveness lacks standardized benchmarks and a systematic framework

Current evaluations of formation performance often focus on a single dimension and lack a comprehensive evaluation system that integrates multidimensional performance metrics. Significant differences exist across studies in terms of datasets, simulation platforms, and performance metrics used, making it difficult to compare results across different studies. In resilience assessment, there is no consensus on methods for quantifying performance trajectories or establishing mission baselines. Robustness assessment often relies on complex network topology analysis but lacks universal metrics oriented toward mission effectiveness. The k-out-of-n system assumption in reliability assessment deviates from the actual characteristics of heterogeneous swarms. Vulnerability analysis mostly remains at the level of static topology and lacks systematic methods oriented toward dynamic evolutionary processes.

(3)

Insufficient attention to human–environment methods and system-level integration

Most existing studies on formation control assume fully autonomous system operation, while research on shared control mechanisms involving human–machine–environment interaction or strategies for dynamic allocation of control authority remains relatively limited. In terms of system integration, there is still a lack of mature theoretical frameworks and engineering pathways for issues such as joint deployment and dynamic reconfiguration of multi-layer heterogeneous HAPS-UAV networks, as well as unified modeling and cooperative control of cross-domain platforms.

(4)

Limited real-world validation and deployment

Most existing studies primarily rely on numerical simulations or simplified hardware-in-the-loop experiments, and physical validation in real-world complex environments is extremely limited. In particular, systematic validation is lacking in scenarios involving long endurance, wide coverage, and multiple missions. A significant gap remains between the performance of existing methods under laboratory conditions and their performance in actual deployment scenarios. The robustness and reliability of formation control algorithms in real-world complex environments therefore remain to be further verified.

7.2. Future Directions

(1)

Algorithmic level: Developing resilient cooperative control for large-scale heterogeneous formation

In the future, a unified dynamic description and a cross-domain cooperative control framework should be developed for air–land–sea multi-agent systems to overcome the heterogeneity barriers among different platforms. A flexible, open system architecture should be constructed to enable task-driven, on-demand formation through functional decomposition and dynamic reorganization. In parallel, a resilient control architecture that integrates mechanism-based modeling with data-driven online learning should be established to achieve real-time estimation and adaptive compensation for complex wind disturbances, strong electromagnetic interference, and system uncertainties. Furthermore, an integrated resilient control system encompassing fault detection, fault-tolerant reconfiguration, and security defense should be developed, enabling localized isolation of disturbance and fault impacts as well as system-level recovery.

(2)

Evaluation level: Establishing a scientific, universal, and practical formation effectiveness evaluation system

In the future, a dynamic evaluation framework that deeply integrates knowledge-driven and data-driven approaches should be developed. The framework should transform prior expert knowledge into adaptive evaluation rules and leverage heterogeneous data augmentation along with transfer learning to enhance model generalization. The evaluation system must cover multiple dimensions, fully accounting for the spatiotemporal evolution of mission phases, environmental conditions, and formation states, thereby enabling a transition from static metric calculation to dynamic performance prediction. Concurrently, a standardized benchmark testing platform should be established, featuring open datasets, a repository of typical mission scenarios, and a unified evaluation metric system, to facilitate the reproducibility and cross-comparison of research results.

(3)

Real-World deployment verification: Achieving hybrid human-in-the-loop intelligence and system-level integration verification

In the future, explainable shared control mechanisms should be developed, along with strategies for dynamic allocation of control authority between humans and machines, thereby achieving an organic integration of human decision-making advantages and machine autonomy. For large-scale formation scenarios, human–machine collaborative interfaces based on intent recognition and natural interaction should be explored to reduce operational burden. The verification phase should extend from numerical simulation to real-world complex environments, establishing a closed-loop optimization process comprising “simulation rehearsal-hardware-in-the-loop verification-actual flight testing” to accelerate the translation of theoretical findings into engineering practice.