Submitted:
12 August 2026
Posted:
13 August 2026
You are already at the latest version
Abstract
Decentralized swarms of unmanned aerial vehicles represent a revolutionary technology with various applications, but so far there are no systemic analytical approaches that combine technical, tactical, and ethical aspects. This paper presents a systematic review of 47 publications from IEEE Xplore, Scopus, and Web of Science (2018–2026), developing original analytical models for swarm deployment. A taxonomy of three architectural types (centralized, hierarchical, decentralized) with quantitative resilience assessment shows decentralized swarms maintain effectiveness with up to 30% agent loss, versus 5% for centralized. An analytical J/S model evaluates communication channel vulnerability to electronic warfare, while a grid weight model wi(t+1)=wi(t)+αdi(t)−βλwi(t) with threshold wth=10 enables decentralized targeting without a single point of failure. Comparative analysis of MANET and DTN protocols reveals latency, reliability, and power trade offs, justifying a hybrid scheme for combat conditions. Economic analysis demonstrates cost exchange ratios reaching 1:1000, and computing infrastructure assessment shows a 100 UAV swarm requires ≈2 PetaFLOPS with 70% ground and 30% onboard distribution using NPUs. Findings provide practical recommendations for swarm architecture selection, hybrid communication protocols, specialized autopilot modes, automated ground infrastructure, and mandatory human in the loop confirmation for critical targeting decisions. Future research priorities include experimental validation, integration with large language models, and cyber security assurance.
Keywords:
decentralized UAV swarm
; swarm intelligence
; electronic warfare
; MANET/DTN hybrid protocols
; cooperative localization
; human in the loop confirmation
1. Introduction
Drones have penetrated all spheres of human activity — from aerial photography and agricultural monitoring to the effective solution of various tasks. However, the truly revolutionary potential of this technology is revealed not at the level of individual platforms, but when they are combined into swarms – decentralized, self-organizing groups capable of solving tasks beyond the reach of single machines or centrally controlled formations. As noted in recent research, the use of UAV swarms significantly increases mission effectiveness through collaborative work and collective intelligence.
The transition from centralized control architectures to decentralized swarm systems is driven by fundamental limitations of the traditional approach. Centralized systems, in which all decisions are made by a single command centre, encounter scalability, fault-tolerance, and communication-link vulnerability issues. In distributed systems, each agent makes decisions based on local information and neighbour interactions, offering far better scalability and survivability, albeit at the cost of complex coordination tasks. This transition – from a “smart” single platform to an “intelligent” swarm – constitutes the core challenge of modern unmanned systems science.
Despite impressive progress in swarm technologies, a significant gap remains between the rapidly growing number of publications on specific swarm control aspects and the lack of systematic, analytically grounded reviews that integrate disparate knowledge into a coherent picture. Existing reviews either focus on narrow technical issues (e.g., trajectory planning algorithms or routing protocols) or are purely narrative, offering readers no quantitative models or comparative analyses. This problem is particularly relevant in four interrelated areas, which together determine the practical applicability of swarm systems in real-world scenarios:
- There is no unified taxonomy of autonomy levels and communication architecture types that would allow developers and customers to consciously choose swarm configuration according to mission requirements. Different research groups give fundamentally different meanings to the term “swarm intelligence,” hampering result comparison and knowledge accumulation.
- The problem of swarm resilience to electronic warfare (EW) is critical but under-researched. Modern EW systems can effectively suppress control and navigation channels, making traditional centralized swarms extremely vulnerable. At the same time, the literature practically lacks analytical models that quantify the communication link budget under jamming and compare the effectiveness of different protocols in terms of jamming resistance.
- The choice of information exchange protocols for decentralized swarms remains unresolved. MANET (Mobile Ad hoc NETworks) and DTN (Delay-Tolerant Networks) offer fundamentally different approaches to communication: the former target continuous connectivity with minimal delays, the latter tolerate intermittent connectivity and significant delays. However, a systematic comparative analysis of these protocols for UAV swarms – with their high mobility, limited power, and dynamic topology – has not been performed.
- Ethical and legal frameworks for decentralized swarm deployment, especially in autonomous targeting and decisions on the use of force, are practically undeveloped. The absence of a single commander in a decentralized swarm raises fundamental questions of responsibility, accountability, and compliance with international humanitarian law, which cannot be ignored when moving to fully autonomous combat systems.
This paper aims to fill these gaps. Unlike existing narrative reviews, this work offers a systematic, analytically grounded approach to the analysis of decentralized UAV swarms, integrating technical, tactical, and ethical aspects. To ensure reproducibility and methodological rigour, the review followed a formalized procedure: searches were performed in IEEE Xplore, Scopus, and Web of Science databases for the period 2018–2026 using the keywords “drone swarm,” “MANET,” “swarm intelligence,” “jamming resilience,” and “cooperative localization.” From 287 initially identified publications, after removing duplicates and checking methodological rigour, 47 relevant works were selected as the empirical basis of this review.
The original contributions of this paper are:
- A taxonomy of autonomy levels, positioning methods, and communication types for decentralized UAV swarms, enabling the systematization of existing solutions and informed architecture selection depending on mission requirements.
- An analytical communication-link budget model under electronic jamming, including the signal-to-jammer ratio (J/S) with attenuation, as well as quantitative estimates of frequency-hopping (FHSS) effectiveness and radio-silence tactics.
- A comparative analysis of decentralized MANET and DTN protocols across three key criteria: end-to-end latency, packet delivery ratio (PDR), and power consumption, with justification of a hybrid scheme for combat conditions.
- A quantitative observability model for cooperative localization based on the Cramér–Rao lower bound (CRLB) to estimate swarm positioning accuracy when global navigation systems are unavailable.
- An ethical analysis of decentralized targeting, including a proposal to introduce mandatory human-in-the-loop confirmation for critical decisions and a discussion of compliance with international humanitarian law.
The paper is structured as follows. Section 2 addresses key challenges in UAV swarm deployment: remote control and autonomous piloting, positioning systems, decentralized communications, and the contrast between programmable and self-organizing swarms. Section 3 is devoted to swarm technologies and swarm intelligence, including biological self-organization algorithms, EW countermeasure tactics, energy autonomy. Section 4 analyses economic aspects that determine the quantitative superiority of swarms. Section 5 examines human–swarm interaction problems, including interfaces, micromanagement, and individual drone behaviour adjustment. Section 6 covers socio-political and legal aspects. Section 7 and Section 8 analyze the requirements for autopilots, as well as for the infrastructure needed for the mass production and maintenance of UAVs. The conclusion summarizes the main results and identifies the directions for future research.
2. Literature Review. Challenges in UAV Swarm Deployment
2.1. Remote Control and Autonomous Piloting of UAV Swarms
The evolution of UAV control systems represents a path from full manual control, where the operator controls every movement via a radio link, to fully autonomous systems capable of making decisions in complex, dynamically changing environments without human involvement (Anicho, 2023). Understanding this evolution and classification of autonomy levels is fundamental for designing decentralized swarm systems (IISc, 2025). This section offers a systematic taxonomy of single-UAV autonomy levels, an analysis of sensor systems and their dependence on global navigation satellite systems (GNSS), and practical recommendations for choosing the autonomy level based on mission type.
During 2020–2026, several competing approaches to classifying unmanned system autonomy levels emerged internationally, each with its own application domain and methodological foundations. The lack of a unified taxonomy creates conceptual confusion and complicates comparison of results from different laboratories and countries (Cetinsaya et al., 2024).
DRONEII (Drone Industry Insights) classification. This industry classification distinguishes five autonomy levels based on UAV independence degrees (DRONEII, 2023). Level 1 (low automation) assumes that at least one substantial task is performed under pilot control. Level 2 (partial automation) allows the UAV to take over course and altitude control in certain circumstances, but the pilot remains responsible for safe mission execution. Level 3 (conditional automation) means the UAV can perform all flight functions, with the pilot acting as backup. Level 4 (high automation) is characterized by redundant systems that ensure continued flight upon component failure, with the pilot out of the control loop. Level 5 (full automation) implies that the UAV can plan its actions using advanced autonomous learning based on artificial intelligence, with minimal or zero human involvement (Infona, 2023).
PACT (Pilot Authority and Control of Tasks) classification. Developed by the UK Defence Science and Technology Laboratory (Dstl), this scheme describes the shift from direct human control to supervisory functions as system autonomy grows (Dstl, 2022). The PACT scale assigns authority levels from 0 (full pilot authority) to 5 (full system autonomy). From level 4b, the system gains authority to make decisions without operator instructions, critical for swarms operating under limited communication (IEEE, 2024a).
ISO/AWI 25132 standard. This developing international standard defines autonomous flight control level classification for civil unmanned aircraft systems based on role distribution between human and machine (ISO, 2024). It establishes six levels – from no autonomy (level 0) to full autonomy (level 5). This standard, still in development, aims to harmonize requirements for civil UAV control systems.
NASA classification. NASA offers the Autonomous Control Level (ACL) framework, in which level 6 involves group tactical replanning for police patrol and disaster relief tasks (NASA, 2021). Level 4 (fully autonomous) is characterized by the system receiving goals from humans and transforming them into tasks executed without human interaction, though humans may intervene in emergencies or to change objectives. A key aspect of NASA’s approach is considering autonomy along two axes: automation level and operational context (CNKI, 2023).
ALFUS (Autonomy Levels for Unmanned Systems) framework. Developed by the US National Institute of Standards and Technology (NIST), this system is universal and applicable to various unmanned systems (NIST, 2020). Its key component is a metric system along three axes: human independence, mission complexity, and environmental complexity. In 2025, the NATO Support and Procurement Agency (NSPA) integrated this approach into its robotics and autonomous systems programmes, using ALFUS level 5 for architectures distributed across resilient network nodes (NATO, 2025).
Exyn Technologies classification. Presented at XPONENTIAL 2024, this system is based on SAE driving autonomy levels and adapted for aerospace (Exyn Technologies, 2024). The definitions and taxonomy allow comparison of operator and system capabilities with discrete autonomy levels, facilitating assessment of technical maturity and risks (Kumar et al., 2024).
Chinese national standard. In December 2023, a standard was introduced defining classification requirements for autonomous flight control of civil UAVs based on dynamic tasks and role distribution between humans and machines, highlighting the global nature of the standardization problem (Chinese Standard, 2023).
Based on the analysis of existing classifications and considering swarm specifics, a unified taxonomy of autonomy levels is proposed, comprising six levels (0 to 5), each characterized by functions, sensor types, and GNSS dependence (see Table 1).
Positioning is central to autonomous piloting. UAV navigation traditionally relies on GNSS and inertial measurement unit (IMU) integration (Sage, 2024). However, when GPS signals degrade or are unavailable due to jamming, spoofing, or multipath propagation, these approaches become unreliable (arXiv, 2024a).
GPS dependence by autonomy level. At levels 0–2, GPS is auxiliary or optional. From level 3 upward, GPS dependence becomes critical for coordinate-based route planning. Modern research aims to reduce this dependence. For example, TerraSLAM, developed at Carnegie Mellon University, offers a global positioning system using a 3D GIS model to link relative and absolute coordinate systems, achieving localization accuracy in GPS-denied environments comparable to GPS-RTK (average 0.21 m) (CMU, 2025).
Alternative positioning methods. Multi-sensor adaptive schemes using low-cost LiDAR and neural-network approaches are being developed for GPS-denied environments (IEEE, 2024b). A promising direction is cooperative localization within a swarm, where drones use mutual range measurements (UWB or optical beacons) to determine relative positions. With five or more UAVs using mutual range measurements, positioning RMS error can be reduced to 0.5–1 m over distances up to 200 m (Luo et al., 2022). By 2025, solutions appeared for swarm trajectory planning based on inertial navigation combined with GPS localization, specifically intended for low-altitude flight in complex environments where obstacles often block GPS signals (MDPI, 2025a).
Sense-and-Avoid (SAA) and Detect-and-Avoid (DAA) technologies. These remain major barriers to integrating UAVs into shared airspace (IEEE-AESS, 2025). By 2025, prototype multi-sensor SAA systems using optical and radar data fusion were developed (MDPI, 2025b). For small UAVs (up to 25 kg, speed up to 15 m/s), affordable SAA solutions with collision prediction and avoidance manoeuvre modules have been created (IEEE, 2023a). These technologies are critical for transitional autonomy levels (3–4), where the system must autonomously respond to unforeseen obstacles.
FPV and autonomy: evolution of approaches. During 2023–2025, FPV systems evolved significantly toward autonomy. Traditional FPV, requiring full manual control, is augmented with autonomous terminal guidance elements (OUCI, 2023). In 2024, developments emerged allowing FPV operators to manually guide drones to a target area, then lock onto a target and command an attack, after which the drone autonomously (as an “air-launched homing torpedo”) approaches and strikes the target (arXiv, 2024b). Another trend is GPS-independent navigation for FPV drones, enabling route following without satellite navigation (MIT, 2024). By 2025, terminal guidance systems for FPV drones advanced significantly: at 100–150 m, the operator locks onto a target and the drone autonomously completes the attack (DroneTech Journal, 2025).
Off-the-shelf computing solutions. An important trend is the emergence of open hardware-software architectures like AeroCompanion based on Raspberry Pi 5 and Pixhawk 6X, providing autonomous navigation, computer vision for obstacle avoidance, and GPS-independent navigation (HardwareX, 2024). Modular designs support numerous sensors and payloads, making them ideal for FPV systems, surveillance, and precision delivery (arXiv, 2024c).
Based on the presented taxonomy and analysis of existing solutions, the following practical recommendations can be formulated:
Missions with high manoeuvrability requirements (FPV racing, aerial combat, narrow obstacle passage). Recommended level 0–1 (full manual control or automatic stabilization). High manoeuvrability requires instant operator reaction, which current autonomous systems cannot fully provide. However, autonomous FPV racing systems where drones with computer vision navigate gates without human intervention are a promising direction (IEEE Robotics, 2023).
Civilian commercial missions (aerial photography, monitoring, inspection). Recommended level 2–3 (partial or conditional automation). These missions require stable route flight with position holding and obstacle avoidance, but not complex adaptive behaviour. GPS dependence at these levels is acceptable since flights usually occur in open areas (Wiley, 2025).
Urban delivery missions. Recommended level 3–4 (conditional or high automation). Urban environments feature GPS signal blockage by buildings and high obstacle density, requiring advanced SAA/DAA systems and alternative positioning methods (AIAA, 2025). SLAM, LiDAR, and computer vision are needed for autonomous navigation under limited satellite visibility (EBSCO, 2025).
Civil missions, including infrastructure inspection, environmental monitoring, and search and rescue operations. Recommended level 4 (high degree of automation). These missions are often carried out in conditions where GPS may be unavailable. A well-developed joint localization system, alternative positioning, and the ability to adapt to changing conditions are required. Solutions based on TerraSLAM technology and a neural network for location determination are promising.
Decentralized swarms in conditions of complete uncertainty. Recommended level 5 (full autonomy). Such missions require complete independence from external positioning systems, collective decision-making, and adaptive learning (Swarm Intelligence, 2024). Advanced AI methods, neural processors, and distributed intelligence systems are used (IEEE Micro, 2025). Critical is operation in GPS-denied environments using mutual drone positioning within the swarm (IEEE TMC, 2024).
Hybrid scenarios (FPV with autonomy elements). Recommended level 2–3 with manual override capability. This approach, actively developing in 2024–2025, assumes the operator manually guides the drone to the target area, after which the system takes over terminal guidance (JUVS, 2024). This is optimal for strike missions requiring high accuracy while retaining autonomous operation capability under signal suppression (IJAE, 2025).
Autonomy level choice should be determined not only by technical capabilities but also by safety requirements, system cost, operating conditions, and critically – the expected level of electronic countermeasures (JATC, 2024).
2.2. Positioning Systems and Autonomous Navigation of UAV Swarms
Accurate positioning is fundamental for both individual UAVs and decentralized swarms. Unlike single UAVs, which can rely on high-precision but expensive navigation systems, swarms impose specific requirements on localization methods: they must be scalable, fault-tolerant, energy-efficient, and, critically, capable of operating when global navigation satellite systems (GNSS) are unavailable or suppressed (MDPI, 2025). As noted in the review (TransNav, 2025), accurate and reliable localization methods have become critical for maintaining formation and preventing collisions in swarm applications.
This section presents a systematic analysis of UAV swarm positioning methods for the period 2020–2026. Global (satellite) and local (alternative) positioning methods are considered, including optical systems (VIO/SLAM), ultra-wideband (UWB) systems, LiDAR systems, and mutual positioning methods within the swarm. Special attention is paid to quantitative localization accuracy assessment via the Cramér–Rao lower bound (CRLB) and occlusion effect analysis.
Global navigation satellite systems (GPS/GNSS) remain the most common method for absolute UAV positioning due to global availability and relatively low implementation cost. Satellite navigation is easily integrated into unmanned aerial vehicles, which makes it the standard for most commercial applications (MDPI, 2025). However, relying solely on GPS for swarm missions is dangerous, as many tasks require centimetre-level accuracy that standard GPS cannot provide (MDPI, 2025).
The main GNSS limitations for swarms include:
- Vulnerability to electronic jamming and spoofing. GPS signals can be blocked, modified, or spoofed under EW or natural interference (TRID, 2023), making GPS-only swarms extremely vulnerable in combat.
- Signal blockage in complex environments. In urban canyons, under bridges, in deep ravines, under dense tree canopies, or indoors, GPS reception is severely degraded (Dynamic Control, 2023). For low-altitude swarm flights in such conditions, GPS becomes unreliable.
- Insufficient accuracy. Standard GPS provides accuracy of several metres, insufficient for maintaining tight formation where centimetre- or decimetre-level errors are required (MDPI, 2025).
Consequently, modern research increasingly focuses on GNSS-independent localization methods, using GNSS as auxiliary rather than critical data.
Local positioning methods allow UAVs to determine their relative positions within the swarm coordinate system without global satellite systems. These methods underpin fault-tolerant swarm systems capable of operating in GPS-denied environments (Gyagenda, 2023; Couturier, 2023).
Visual-inertial odometry (VIO) and simultaneous localization and mapping (SLAM) use camera and IMU data to estimate position and orientation. These methods are actively developed for swarm applications.
Omni-Swarm is a decentralized state-estimation system for aerial swarms using wide-field-of-view stereo cameras and UWB sensors (HKUST, 2023). It includes VIO, map-based localization using multi-drone maps, and visual drone tracking. Experimental results show centimetre-level relative state estimation accuracy with global consistency in aerial swarms (HKUST, 2023).
Swarm-LIO offers a fully decentralized state-estimation method based on LiDAR-inertial odometry (HKUST, 2023). Compared to motion-capture ground truth, results show centimetre-level localization accuracy surpassing other state-of-the-art LiDAR-inertial odometry for single UAVs (HKUST, 2023).
Vrba et al. (2023) proposed a new relative localization approach for micro-UAV control using a well-equipped aerial robot, combining VIO with LiDAR relative localization.
VIO-UWB-based swarm initialization. In satellite-signal-blocked conditions, fast initialization of UAV swarms is a technical challenge (Drones, 2024). A two-stage robust initialization method for swarms of more than four UAVs, suitable for large-scale satellite-denied scenarios, is proposed (Drones, 2024). With improved computing power and VIO accuracy, relative position relationships can be established through drone motion (Drones, 2024).
UWB technology provides high-precision range measurements based on time-of-flight. Its wide bandwidth ensures high timing accuracy and enhanced resilience to shadowing and multipath, making UWB highly useful for indoor ranging (TU Delft, 2024). With high accuracy, low cost, small size, and interference resistance, UWB range sensors are well suited for UAV auxiliary positioning (ACM, 2024).
Infrastructure-free navigation. IEEE (2023) proposed a UWB-based nanodrone swarm without infrastructure, where part of the fleet acts as dynamic anchor drones (AD) capable of automatic deployment and landing. Field results with four mission drones (MD) show RMSE localization from 15.3 to 27.8 cm (IEEE, 2023; ADS, 2023).
UWB-based relative localization. arXiv (2024) presented a UWB-ranging-based relative localization solution for micro-air-vehicle (MAV) swarms, providing peer-to-peer RL using UWB.
Combined positioning architectures. IEEE (2024) proposed a combined positioning architecture based on BDS (Chinese satellite system), barometer, and UWB, along with gradient-descent accuracy optimization. Fusing BDS, barometer, and UWB data converted the accuracy problem into an optimization task (IEEE, 2024).
UWB-based onboard localization. ScienceDirect (2024) proposed a mobile UWB-based onboard localization system with multiple drones as free-structure anchor stations. The UWB sensor provides accurate range measurements and is easily used in range-based methods like multilateration (ScienceDirect, 2024).
LiDAR systems provide high-precision 3D environmental mapping and are effective for GPS-denied navigation. LiDAR localization is accurate and robust to complex environments, but 3D LiDARs are relatively heavy and require large UAV platforms, unlike lightweight cameras (Vrba et al., 2023).
IEEE (2023) presented UAV swarm navigation using 2D LiDARs in unknown forest. Each UAV builds its own map estimate based on tree features detected in the LiDAR point cloud. Navigation accuracy was at most 40 cm over a 225 m path with noise below 3 cm (IEEE, 2023).
Swarm-LIO, mentioned above, shows that decentralized LiDAR-inertial odometry can achieve centimetre-level localization in swarms (HKUST, 2023). Vrba et al. (2023) combined VIO with LiDAR relative localization for micro-UAV guidance.
Angle-of-arrival methods are used to determine signal source direction. Positioning methods widely use RSS, TOA, TDOA, and AOA (Wiley, 2024).
Wiley (2024) proposed a long-range AOA localization algorithm using reference points to determine target position relative to UAVs via optoelectronic gimbals.
IEEE (2024) investigated sensor trajectory optimization to improve AOA localization accuracy.
Hybrid methods. OPG (2024) proposed a hybrid localization method combining RSSI and AOA for ultraviolet communication between UAV swarms, outperforming individual methods (OPG, 2024). CNKI (2024) addressed 3D relative localization of two UAVs using TOA and AOA combinations under different flight trajectories (CNKI, 2024).
Mutual localization is key for swarm systems, allowing each agent to determine its position relative to other swarm members without external references. As noted in TransNav (2025), with growing interest in UAV swarm applications for search-and-rescue, surveillance, and delivery, accurate and reliable localization has become critical for formation maintenance and collision avoidance.
Active positioning correction. Semantic Scholar (2024) proposed an active localization correction system that plans camera orientation through a yaw planner during flight, reducing positioning drift by up to 65% and maintaining formation both indoors and outdoors in GPS-denied conditions (Semantic Scholar, 2024).
Decentralized state estimation. Decentralized state estimation is among the most fundamental components of autonomous swarm systems in GPS-free zones, but remains a highly complex research challenge (AITopics, 2023).
Distributed collaborative SLAM. BIT (2024) addressed UAV swarm collaborative SLAM based on visual-inertial-ranging measurements, noting difficulties in initialization, high communication demands, and inevitable drift for swarm joint positioning (BIT, 2024).
Cooperative navigation using VIO and UWB. Semantic Scholar (2023) proposed a robust GPS-denied navigation system for swarms combining visual odometry and UWB module estimates, maintaining 10 cm accuracy even under partial non-line-of-sight (NLoS) conditions (Semantic Scholar, 2023).
The Cramér–Rao lower bound (CRLB) is a fundamental tool for quantifying theoretical positioning accuracy in cooperative localization systems. It defines the minimum achievable variance of unbiased parameter estimates and serves as a benchmark for comparing localization algorithms.
CRLB for 3D cooperative networks. IEEE (2025) presented a robust system for accurate cooperative localization in multi-UAV systems. The CRLB for 3D cooperative localization networks was derived with special attention to practical scenarios including non-uniform spatial distribution of anchor nodes (IEEE, 2025). A similar approach appears in arXiv (2024), where CRLB is derived under non-uniform anchor distribution.
CRLB under relative measurements. ScienceDirect (2023) summarized design ideas, applicability conditions, and detailed procedures of typical cooperative localization algorithms, generalizing the CRLB for cooperative localization when relative measurements are unbiased, and providing specific steps to compute CRLB in different spatial dimensions (ScienceDirect, 2023).
CRLB for optimal resource allocation. SPIE (2023) considered power allocation for optimal resource distribution in UAV swarms. In jam-free and jamming scenarios, CRLB serves as an accuracy index for cooperative localization (SPIE, 2023).
CRLB for indoor localization. IEEE (2023) proposed an indoor cooperative positioning system for micro-UAV swarms based on visible-light communication (VLC). The theoretical lower bound (CRLB) was mathematically analysed (IEEE, 2023).
Occlusion error analysis. In ScienceDirect (2023), the CRLB is analysed under the assumption of unbiased relative measurements. Occlusions (loss of line-of-sight) increase measurement uncertainty, reflected in CRLB. With five or more UAVs using mutual range measurements (UWB or optical beacons), RMS positioning error can be reduced to 0.5–1 m over 200 m. However, under occlusion (e.g., urban canyons), error rises to 5–10 m.
Based on the literature, the following comparative characteristics of main positioning methods for UAV swarms can be formulated (Table 2).
Based on the analysis, the following practical conclusions and recommendations for designing UAV swarm positioning systems can be drawn:
1. Abandon GPS as the sole positioning source. For swarm systems, especially in urban environments, relying solely on GPS is unacceptable. A multi-sensor redundant architecture is required to operate in GPS-denied environments. As noted in MDPI (2025), new localization methods mainly address GPS-free zones because relying exclusively on GPS is dangerous for swarm missions requiring centimetre accuracy.
2. Prefer hybrid approaches. The most effective systems combine multiple sensing modalities. UWB rangefinders provide absolute range measurements with 10–30 cm accuracy under LOS (IEEE, 2023). VIO/SLAM provides continuous motion estimation but suffers from drift accumulation (HKUST, 2023). LiDAR offers high accuracy but requires significant computational resources and power (Vrba et al., 2023). Combining these methods compensates for individual weaknesses.
3. Scalability through mutual positioning. Cooperative localization is key for scalable swarms. With five or more UAVs using mutual range measurements, RMS positioning error can be reduced to 0.5–1 m (Luo et al., 2022). As TransNav (2025) states, accurate and reliable localization has become critical for formation maintenance and collision avoidance.
4. Account for occlusions and loss of line-of-sight. In urban canyons and other complex environments, occlusions increase CRLB and reduce localization accuracy. Designing swarm systems requires redundancy of measurement channels and methods resilient to LOS loss, such as UWB with NLoS support (Semantic Scholar, 2023).
5. CRLB as an optimization tool. Using CRLB for resource allocation and swarm configuration optimization enables maximum localization accuracy under limited resources. As shown in SPIE (2023) and arXiv (2024), CRLB can serve as an index for optimal power distribution and formation design.
6. Choose method depending on scenario. For indoor and dense urban environments, UWB and VIO/SLAM are preferred (IEEE, 2023; Drones, 2024). For open spaces with good visibility, LiDAR and combined systems are effective (HKUST, 2023; IEEE, 2023).
2.3. Decentralized Communications, Relays, and Remote Control of UAV Swarms
Spatial positioning is directly intertwined with communication tasks, particularly radio communication as the most widespread means of wireless data transfer (Anicho, 2023). Communication itself is a problem for small UAVs due to size constraints on onboard energy, most of which is dedicated to flight, leaving only a small fraction for transceiver operation (Cetinsaya et al., 2024). Thus, transmitter power is severely limited, and small size restricts antenna dimensions. Using groups of UAVs makes communication even more critical (Bekmezci et al., 2022).
This section systematically analyses radio communication challenges in decentralized UAV swarms for 2020–2026. Fundamental limitations of power, range, and bandwidth are considered, and a comparative analysis of two major decentralized communication architectures – MANET (Mobile Ad hoc NETworks) and DTN (Delay-Tolerant Networks) – is conducted across three key criteria: latency, PDR, and power consumption. A hybrid communication scheme is proposed, and practical protocol selection recommendations are given.
Power and range limitations. Transmitter power on small UAVs is severely limited by onboard energy, most consumed by flight (Cetinsaya et al., 2024). In free space, signal attenuation is proportional to the square of distance, making communication increasingly difficult as swarm size and operational radius grow. These constraints make remote control of large UAV groups from a single central station impossible (IEEE, 2023).
Bandwidth issues. Centralized strategies require each UAV a communication channel with several Mbps bandwidth for image and environment data transmission (Cetinsaya et al., 2024). Partial relief comes from hierarchical centralized strategies where only some UAVs connect to the central controller, each relaying commands to subgroup members. However, the need for constant communication between upper-level UAVs and the centre remains a difficult problem (Bekmezci et al., 2022).
Dynamic topology. UAV swarms exhibit highly dynamic network topology, constantly changing with node movement and relative positions (Beisenkhanov et al., 2023). Consequently, networks cannot maintain continuous transmission lines like fixed networks, leading to significant packet delivery efficiency degradation (IEEE, 2025). Low drone density in sparse swarms may further limit connectivity and overall network performance (Beisenkhanov et al., 2023).
Vulnerability to electronic jamming. As noted in Chen et al. (2020), under EW, control and data channels can be suppressed, making traditional centralized architectures extremely vulnerable. Therefore, transitioning to decentralized communication networks becomes critical for swarm survivability in combat.
Mobile Ad hoc NETworks (MANET) are radio networks with random mobile subscribers, implementing fully decentralized control without base stations or reference nodes (Kumar et al., 2020). They are self-organizing because nodes are not only terminal user devices but also relays/routers, forwarding packets and participating in route discovery, thus capable of self-organization. Each node has software/hardware to forward data from source to destination directly when a physical path exists, distributing network load and increasing aggregate throughput.
MANET application in UAV swarms (FANET). Flying Ad Hoc Networks (FANET) are a specialized MANET subclass where nodes are UAVs (Almansor et al., 2024). FANET is a self-organizing wireless network of UAV clusters/drones interacting with each other (Almansor et al., 2024). Unlike traditional MANET, FANET features much higher mobility, dynamic topology, and 3D operation (Kumar et al., 2020).
MANET routing protocols for swarms. FANET uses various routing protocols: topological, positional, hierarchical, and swarm-based (Almansor et al., 2024). Common MANET protocols adapted for FANET include:
- OLSR (Optimized Link State Routing) – proactive protocol maintaining up-to-date routes for all nodes (Bekmezci et al., 2022). Modified OLSR versions can be effective in search-and-rescue scenarios with increased mobility and scalability (Kumar et al., 2020).
- AODV (Ad hoc On-Demand Distance Vector) – reactive protocol establishing routes only on demand (Kumar et al., 2020). Integrating DTN mechanisms into AODV can significantly mitigate performance degradation at high mobility (Kumar et al., 2020).
- DSDV (Destination-Sequenced Distance-Vector) – proactive distance-vector protocol (Kumar et al., 2020).
MANET advantages for swarms. As noted in Kumar et al. (2020), MANET provides low latency (<100 ms) and high PDR (>0.95) in dense swarms (node density >0.1 nodes/m³). A spatially organized swarm inherently consumes less energy for communication between nearby nodes, which is also harder to jam than the weakened signal from a distant base station (Cetinsaya et al., 2024).
MANET limitations. The main limitation is the requirement of continuous connectivity for route maintenance (Beisenkhanov et al., 2023). In sparse swarms or when links are lost, MANET faces routing problems, causing packet loss and efficiency degradation (IEEE, 2025). Moreover, MANET protocols require continuous exchange of control information (hello packets, route updates), adding overhead and increasing power consumption (Bekmezci et al., 2022).
Delay-Tolerant Networks (DTN) are a network architecture designed for intermittent connectivity, large delays, and lack of end-to-end paths (Beisenkhanov et al., 2023). DTN is based on dynamic neighbour discovery and store-and-forward relaying (Beisenkhanov et al., 2023). Unlike MANET, DTN does not require continuous connectivity and can operate even with periodic link breaks (Almansor et al., 2024).
DTN application in UAV swarms. UAV swarms experience highly dynamic topology, and low drone density may limit connectivity (Beisenkhanov et al., 2023). DTN architecture, using store-and-forward, overcomes these challenges (Beisenkhanov et al., 2023). DTN is especially effective in scenarios where UAV networks act as data mules, carrying data to remote or isolated areas (IEEE, 2025).
DTN routing protocols for swarms. Beisenkhanov et al. (2023) evaluated two common DTN routing protocols:
- Epidemic Routing – each node copies messages to all neighbours, maximizing delivery probability at the cost of high redundancy and network load (Beisenkhanov et al., 2023).
- Spray-and-Wait Routing – a limited number of message copies are “sprayed” into the network, then wait for delivery, balancing reliability and efficiency (Beisenkhanov et al., 2023).
IEEE (2025) investigated integrating Information-Centric Networking (ICN) with DTN to improve message delivery by caching data at UAV nodes. To manage communication overhead due to sparse DTN characteristics, the MAIC (Multi-Agent Incentivized Communication) algorithm is proposed for routing optimization (IEEE, 2025).
DTN advantages for swarms. DTN tolerates link breaks up to 30 s, provides latency of 1–5 minutes and PDR of 0.7–0.85 (Beisenkhanov et al., 2023; Almansor et al., 2024). Transmitter power consumption is reduced by up to 40% compared to MANET due to the absence of continuous control information exchange (Cetinsaya et al., 2024). DTN also offers higher resilience to jamming because it does not require constant connectivity and can use alternative routes when links are lost (Almansor et al., 2024).
DTN limitations. The main limitation is high message delivery latency, which may reach minutes (Beisenkhanov et al., 2023). This makes DTN unsuitable for real-time applications requiring immediate response, such as flight control or real-time targeting (Cetinsaya et al., 2024). Additionally, DTN requires data buffering at nodes, creating extra memory load (IEEE, 2025).
Based on the literature (Beisenkhanov et al., 2023; Kumar et al., 2020; Cetinsaya et al., 2024; Almansor et al., 2024; Bekmezci et al., 2022), a comparative analysis of MANET and DTN across three key criteria is presented (Table 3).
Based on the comparative analysis, a hybrid communication scheme for UAV swarms is proposed, combining MANET and DTN advantages.
Hybrid architecture. The swarm is divided into clusters – dense groups of drones within line-of-sight (Kumar et al., 2020). Inside each cluster, a MANET protocol (e.g., OLSR or AODV) provides low latency and high PDR for intra-group coordination (Bekmezci et al., 2022). Between clusters, which may be separated by significant distances or intermittent connectivity, a DTN protocol (e.g., Epidemic or Spray-and-Wait) is used for data exchange with acceptable delay (Beisenkhanov et al., 2023).
Hybrid advantages. As noted in IEEE (2026), the hybrid approach allows “seamless switching between DTN and NDN without application-protocol changes,” offering improved interoperability and resilience in highly dynamic heterogeneous network environments. The hybrid scheme provides:
- Low latency for critical intra-cluster commands (MANET).
- High resilience to inter-cluster link breaks (DTN).
- Reduced power consumption using DTN for inter-cluster communication.
- Scalability – adding new clusters does not require global network reconfiguration.
Technical implementation. IEEE (2020) proposed a relay mechanism based on SDN architecture that can switch to MANET mode depending on network environment changes, reducing flight latency. IEEE (2025) proposed a dynamic load-distribution routing mechanism for UAV swarm networks integrating ICN and DTN, using the MAIC algorithm to minimize information redundancy.
Based on the analysis, the following practical recommendations for protocol selection can be formulated:
1. Dense swarm within LOS (e.g., drone shows, close-range reconnaissance). Recommended protocol – MANET (OLSR, AODV). Provides low latency (<100 ms) and high PDR (>0.95) at node density >0.1 nodes/m³ (Kumar et al., 2020; Bekmezci et al., 2022). Suitable for real-time applications.
2. Sparse swarm over large area (e.g., search-and-rescue, wide-area monitoring). Recommended protocol – DTN (Epidemic, Spray-and-Wait). Tolerates link breaks up to 30 s and delivers data even without end-to-end connectivity (Beisenkhanov et al., 2023). Saves up to 40% transmitter power (Cetinsaya et al., 2024).
3. Combat conditions with high EW risk. Recommended protocol – hybrid scheme (MANET intra-cluster, DTN inter-cluster). Provides resilience to jamming through decentralization and absence of single point of failure (Almansor et al., 2024). DTN channels between clusters maintain connectivity even when individual frequencies are jammed (IEEE, 2025).
4. Mixed scenarios with different data priorities. As noted in IEEE (2020), depending on network conditions, the system can switch between MANET and DTN modes. For instance, control commands use MANET (low latency), while bulk reconnaissance data use DTN (high reliability).
5. Choice between Epidemic and Spray-and-Wait. For swarms with many nodes and high reliability requirements, Epidemic Routing is recommended (maximum PDR, but high load) (Beisenkhanov et al., 2023). For resource-constrained swarms (energy, memory), Spray-and-Wait is preferred (balance between reliability and efficiency) (Beisenkhanov et al., 2023).
Decentralized communication is a critical component for UAV swarm operation, especially under electronic countermeasures. MANET and DTN offer fundamentally different approaches: the former focuses on minimal latency under continuous connectivity, the latter on maximum reliability under intermittent connectivity. The hybrid scheme (MANET intra-cluster, DTN inter-cluster) combines both advantages, providing low latency for critical commands and high resilience to link breaks between distant groups. Protocol selection must be driven by mission scenario, latency/reliability/power requirements, and expected EW threat level.
2.4. Differences Between Programmable “Swarms” and Self-Organizing UAV Swarms
In the swarm literature, terminological confusion often arises: the same term “swarm” can refer to fundamentally different architectures – from centrally controlled groups executing pre-programmed tasks to fully decentralized self-organizing systems whose behaviour emerges from local agent interactions (EE Times, 2025). As rightly noted in industry reviews, “the term 'swarm' is often misused: light shows or pre-programmed fleets are centrally managed and require constant oversight. In contrast, true swarms rely on decentralized consensus, local autonomy, and dynamic membership, allowing drones to join or leave without redesigning the entire system” (EE Times, 2025). This fundamental difference determines not only control architecture but also critical characteristics such as fault-tolerance, scalability, adaptability to environmental changes, and programming complexity.
This section systematically analyses the differences between programmable (centralized) and self-organizing (decentralized) UAV swarms. A taxonomy of swarm systems is proposed, including three main architectural types, with a comparative analysis of their fault-tolerance, scalability, and applicability in various scenarios, as well as practical recommendations for architecture selection based on mission requirements.
Based on existing classifications (Springer, 2024; MDPI, 2024; IEEE, 2023), the following taxonomy of swarm systems can be proposed, covering three main architectural types.
Type A – Centralized swarms with pre-programmed tasks. In these systems, all decisions are made by a single control centre (ground station or leader drone) with full knowledge of all agent states and environment (Springer, 2024). Each agent receives an individual flight assignment and strictly follows the prescribed trajectory without autonomous adaptation. Typical examples are large-scale drone light shows, where hundreds or thousands of UAVs execute pre-computed programmes (EE Times, 2025; GAO, 2023). As noted in Springer (2024), “in the centralized approach, the network has a single controller that manages many aspects such as topology and routing policy.” This approach provides high task accuracy and optimal solution quality but faces serious scalability and fault-tolerance limitations (Mertil, 2020).
Type B – Hierarchical (hybrid) swarms. In hierarchical systems, agents are divided into clusters with centralized control inside each cluster (cluster leader) and decentralized interaction between clusters (Springer, 2024). As Springer (2024) states, “hierarchical architecture is based on one robot controlling the actions of a group of robots.” Hybrid cluster architectures “combine centralized intra-cluster control and distributed inter-cluster control, offering better scalability than centralized approaches by partitioning the swarm into clusters managed in a distributed fashion” (Springer, 2024). This approach is used in scenarios requiring a balance between control quality and fault-tolerance (IEEE, 2023).
Type C – Fully decentralized self-organizing swarms. There is no central authority; each agent decides solely on local information from its own sensors and neighbour interactions (Frontiers, 2023; MDPI, 2024). As emphasized in Frontiers (2023), “the swarm should be self-organizing: collective swarm behaviour must emerge from interactions between individual robots.” Decentralized architecture “does not use a central agent that controls all others” (Springer, 2024). Inspired by natural swarms – ant colonies, bee swarms, bird flocks – this approach provides maximum fault-tolerance and scalability, though requiring different programming and verification approaches (Frontiers, 2023; Lowy Institute, 2020).
Programmable (centralized) swarms are based on “top-down” design: the developer creates a complete exhaustive programme describing every agent’s actions in every anticipated situation (GAO, 2023). This requires prior collection of exhaustive environment information and heavy computation on powerful resources (Mertil, 2020). As GAO (2023) notes, “drone swarms can use various command-and-control methods, including pre-programmed missions with specific predefined flight trajectories.” When mission conditions change, individual programmes for all agents must be recalculated using external computational power. Characteristically, if one or more UAVs fail, the remaining drones do not reorganize formation but simply continue their programme, as agent autonomy is minimal (EE Times, 2025; Lowy Institute, 2020).
A classic example is large-scale drone light shows. Each UAV receives an individual flight assignment based on its position in the overall motion pattern, and the agent’s role is to strictly follow the preset course with minimal deviation (EE Times, 2025). Each drone is equipped with high-precision GPS, accelerometers, gyroscopes, and a barometer, while a ground station broadcasts corrections to improve positioning accuracy. When one UAV fails, the rest do not reorganize (Lowy Institute, 2020).
Self-organizing (decentralized) swarms are based on “bottom-up” design: the developer defines a set of simple local interaction rules for each agent, and global swarm behaviour emerges from the collective application of these rules (Frontiers, 2023; MDPI, 2024). As Frontiers (2023) states, “the self-organizing nature of robot swarms facilitates the design of flexible systems: the swarm can adapt to various and potentially changing environments.” However, “although robot swarms offer several advantages, their decentralized and self-organizing nature makes them difficult to design. Requirements for desired swarm behaviour are usually stated at the collective level, but it is impossible to program the swarm directly” (Frontiers, 2023).
In self-organizing swarms, each agent makes decisions based on local information and simple rules, enabling adaptation to environmental changes and agent losses without external intervention (MDPI, 2024). Such systems have no single point of failure, high scalability, and the ability to self-organize in response to external changes (Lowy Institute, 2020; Allwright, 2020).
Key differences between approaches:
- Programming level. Centralized swarms program each agent individually; decentralized swarms program interaction rules, and swarm behaviour emerges at the system level (Frontiers, 2023).
- Dependence on external computation. Centralized swarms require powerful external computational resources for planning and replanning; decentralized swarms use onboard computation on each agent (Mertil, 2020).
- Reaction to changes. Centralized swarms require programme recalculation when conditions change; decentralized swarms adapt automatically through local rules (Lowy Institute, 2020).
- Resilience to losses. In centralized swarms, loss of even a small number of agents can disrupt mission performance; decentralized swarms remain operational under significant losses (MDPI, 2024; IEEE, 2023).
The most illustrative example of a self-organizing swarm is the Kilobot platform, developed at Harvard University (PMC, 2020). Kilobot is a miniature robot (diameter ~3 cm, cost ~$15) that communicates with neighbours via infrared for swarm behaviour (PMC, 2020). As described in PMC (2020), “Kilobot is a small robot that can communicate with neighbours using infrared light to develop swarm behaviour. We can program communication rules among kilobots and embed them in a swarm of 1024 kilobots to control self-assembly into desired configurations.”
Experiments with Kilobot showed that a swarm of 1024 simple agents can self-assemble abstract geometric shapes based only on the starting point and general movement rules (PMC, 2020). Each agent has a memory instruction to move along the swarm edge until its relation to other surrounding bots is optimal with respect to a common mathematical function describing the shape. The failure of any particular UAV does not affect shape formation as long as the total number is sufficient. In case of failure, more agents can simply be added, since all use the same algorithm (PMC, 2020).
As Core (2020) states, “this work investigates self-organized collective formation and herding using swarm robots. Specifically, we focus on collective tracking and herding using a large number of very simple robots.” Kilobot was chosen as a test platform due to its very low cost (Core, 2020). Research shows that minimalist robots like Kilobot “may lack self-localization capability, making it very difficult to perform missions such as surrounding a target whose position is typically unknown” (ORPG, 2020). Nevertheless, these limitations make the platform ideal for developing self-organization algorithms independent of global information (ORPG, 2020).
The key difference between Kilobot and programmable swarms is that using such a swarm does not require a pre-exhaustive action programme – it can find optimal organizational forms on the ground through local interactions (PMC, 2020; Frontiers, 2023). When some agents are lost, the swarm automatically reorganizes, evenly distributing the load among remaining agents (MDPI, 2024; Lowy Institute, 2020).
Fault-tolerance is a key criterion for comparing swarm architectures, especially in critical civilian applications. Literature analysis shows significant differences among the three architecture types.
Centralized swarms (Type A) exhibit extremely low fault-tolerance. Since all decisions are made by a single centre, failure of the central node leads to complete system shutdown (Mertil, 2020; Lowy Institute, 2020). As Mertil (2020) notes, “although centralized control almost always outperforms distributed control in decision quality, it faces significant scalability limitations.” Even 5% agent failure in a centralized swarm can cause mission failure because the programme does not include automatic reconfiguration (EE Times, 2025). Also, centralized systems are vulnerable to communication-link disruptions between agents and the control centre (Lowy Institute, 2020).
Hierarchical swarms (Type B) show moderate fault-tolerance. Failure of a cluster leader may disrupt intra-cluster control, but other clusters continue operating independently (Springer, 2024; IEEE, 2023). As Springer (2024) states, “hybrid cluster architectures combine centralized intra-cluster control and distributed inter-cluster control, offering better scalability than centralized approaches.” However, hierarchical systems retain vulnerabilities associated with key nodes (cluster leaders), whose failure requires re-election or reassignment (IEEE, 2023).
Decentralized self-organizing swarms (Type C) demonstrate maximum fault-tolerance. Since each agent decides independently based on local information, failure of any number of agents does not stop the system (MDPI, 2024; Lowy Institute, 2020). As Lowy Institute (2020) notes, “decentralized swarms – without a central leader – are highly resilient to electromagnetic disruptions. In so-called emergent swarms, individual units can react to the actions of other local units without any central coordination. Simply put, from a resilience perspective, decentralized swarms are the most effective against electromagnetic countermeasures” (Lowy Institute, 2020).
Research shows that decentralized swarms maintain high effectiveness even with up to 30% agent loss (JISEM, 2026; ACM, 2025). JISEM (2026) presented a decentralized task-allocation method evaluated in simulations with 50 robots, including a scenario with 30% robot failure at step 400. Validation metrics such as entropy, field energy, cluster coherence, coverage, and task achievability showed that the swarm maintains high coordination, stability, and adaptive behaviour (JISEM, 2026). ACM (2025) investigated the robustness of multi-agent reinforcement learning (MARL) algorithms under partial agent failures (10%, 30%, 50%), showing that with properly designed algorithms, the swarm remains operational even with significant losses (ACM, 2025).
IEEE (2023) studied cascading failure effects in autonomous swarms. “Through simulation studies, we reveal the key relationship between resilience, swarm size, and the percentage of failed agents, as well as the existence of a critical point in the swarm” (IEEE, 2023). The study shows that in sufficiently large swarms, the loss of one agent generally does not jeopardize the entire mission, and the failed agent can be replaced (IEEE, 2023) (Table 4).
Based on the analysis, practical recommendations for swarm architecture selection can be formulated:
1. Scenarios with high predictability and accuracy requirements (drone shows, demonstration flights). Recommended architecture – Type A (centralized). These scenarios have static environments, full knowledge of all parameters, and no need for adaptation (EE Times, 2025; GAO, 2023). Centralized control provides maximum execution accuracy and optimal solution quality (Mertil, 2020).
2. Scenarios with moderate adaptability requirements (search-and-rescue, wide-area monitoring). Recommended architecture – Type B (hierarchical). These require a balance between control quality and fault-tolerance (Springer, 2024; IEEE, 2023). Hierarchical architecture allows efficient management of large groups while enabling autonomous cluster operation when contact with the centre is lost (Springer, 2024).
3. Scenarios with high fault-tolerance and adaptability requirements. Recommended architecture – Type C (decentralized self-organizing). These have dynamic environments, high loss risk, and require adaptation without external intervention (Lowy Institute, 2020; MDPI, 2024). Decentralized swarms offer maximum resilience to losses and jamming (Lowy Institute, 2020; Allwright, 2020).
4. Scenarios with limited onboard computing resources. Recommended architecture – Type C (decentralized). Self-organizing swarms do not require powerful onboard computers, as each agent performs simple local computations (Frontiers, 2023; PMC, 2020). This enables low-cost, small platforms like Kilobot (PMC, 2020; Core, 2020).
5. Scenarios requiring rapid deployment without prior programming. Recommended architecture – Type C (decentralized). Self-organizing swarms do not require individual programmes for each agent – general interaction rules suffice (Frontiers, 2023; MDPI, 2024), significantly reducing mission preparation time (MDPI, 2024).
6. Heterogeneous swarm scenarios (different drone types with different functions). Recommended architecture – Type B (hierarchical) or Type C (decentralized) with agent specialization. As ORPG (2020) notes, “our new 'global-to-local' design methodology allows creating heterogeneous swarms for the example application of self-organized task allocation” (ORPG, 2020). Heterogeneous swarms require more complex coordination algorithms, implementable in both hierarchical and decentralized architectures (IEEE, 2023).
The contrast between programmable (centralized) and self-organizing (decentralized) UAV swarms reflects a fundamental difference in managing complex systems. Centralized swarms with pre-programmed tasks provide high accuracy and predictability but suffer from low fault-tolerance, limited scalability, and vulnerability to communication disruptions. Decentralized self-organizing swarms, inspired by natural analogues, demonstrate high resilience to agent losses, adaptability in dynamic environments, and nearly unlimited scalability, though they require different design and verification approaches.
The proposed taxonomy (Type A – centralized, Type B – hierarchical, Type C – fully decentralized) allows systematic classification of existing solutions and informed architecture selection based on mission requirements. For scenarios with high fault-tolerance and adaptability demands, especially under electronic countermeasures, decentralized self-organizing swarms are preferable. For scenarios where accuracy and predictability are critical and the environment is static, centralized or hierarchical architectures may be justified. Architecture choice should be determined not only by technical capabilities but also by reliability requirements, system cost, operating conditions, and expected threat level.
3. Swarm Technologies and Swarm Intelligence
Swarm intelligence (SI) represents collective behaviour emerging from interactions among individuals in a group – be it animals, cells, or artificial agents. As noted in a fundamental review (2024), “typical SI examples include fish schools, ant foraging, bird migration, and so on.” The origins of SI lie in the biological study of self-organizing behaviours in social insect colonies. Observations of natural systems – bird flocks, ant colonies, bee swarms – inspired computational algorithms that today underlie decentralized UAV swarm control. Unlike centralized systems where all decisions are made by a single command centre, SI offers a fundamentally different approach: complex global behaviour emerges from local interactions of simple agents following elementary rules.
Modern research classifies SI models into four main categories: self-propelled particle models (with Boids as the primary example), pheromone-based communication models (including ant colony algorithms), leader-based decision-making models, and empirical research models. Each category finds application in UAV swarm tasks – from trajectory planning and obstacle avoidance to task allocation and formation control. This section systematically analyses key biological analogies and algorithms underlying modern swarm technologies, with emphasis on their application in decentralized UAV systems. Special attention is given to formalizing stochastic swarm behaviour through the grid-weight model – a mechanism enabling collective decision-making based on accumulated information without centralized control.
One of the earliest and most influential swarm models is Boids, proposed by Craig Reynolds in 1987. It became the basis for understanding how simple local rules can generate complex collective motion observed in bird flocks, fish schools, and other biological systems. As stated in the 2024 review, “one of the earliest models is Boids, which generalized three fundamental rules governing collective behaviour.”
Three fundamental Boids rules:
- Separation – each agent avoids collisions with neighbours by maintaining a minimum distance, ensuring dispersion and preventing crowding.
- Alignment – each agent steers to match the direction of neighbours, leading to coordinated motion of the whole group.
- Cohesion – each agent moves toward the centre of mass of neighbours, keeping the group together and preventing breakup.
Despite their simplicity, these rules generate remarkably realistic collective behaviour: the flock on screen looks highly natural – agents cluster, avoid collisions, and even move chaotically like real creatures. As noted in a 2020 work, “although we achieved good improvement over the basic Boids implementation, we only dealt with improving algorithm logic; further performance improvement could be achieved through code optimization and parallel programming.”
In UAV swarm contexts, Boids is used for formation holding, collision avoidance, and coordinated group movement. As the 2024 review states, “self-propelled particle models, with Boids as the primary example, are used in multi-agent systems including swarm control, trajectory optimization, and obstacle avoidance.” However, as will be shown, more complex tasks – such as optimal path finding or task allocation – require algorithms inspired by other biological systems.
Ant Colony Optimization (ACO) is one of the most successful examples of SI inspired by social insects. Unlike Boids, which focuses on motion coordination, ACO solves optimization problems through indirect (stigmergic) communication via pheromone trails.
Mechanism: In nature, ants leaving pheromone trails between nest and food source attract others, who also deposit pheromone. The more ants traverse a path, the more attractive it becomes. Since shorter paths require less time, pheromone accumulates faster, creating positive feedback that leads to optimal route selection.
Each ant maintains a “tabu list” – memory of visited nodes – to avoid cycling. When choosing the next node, an ant considers both distance (edge weight) and pheromone intensity. Pheromones are updated each iteration: they evaporate over time, while traversing ants reinforce them.
Application in UAV swarms. As noted in a 2023 review, “we provide an overview of swarm intelligence algorithms that play an extremely important role in multi-UAV cooperation. The study focuses on four aspects: collision avoidance, task allocation, trajectory planning, and formation reconfiguration.” It also notes that “genetic optimization (GA), ant colony (ACO), and particle swarm (PSO) algorithms are discussed in the context of UAV trajectory planning.”
ACO’s key difference from centralized optimization lies in its stochastic nature and decentralized character. Instead of directed search for an optimal solution, agents perform many random attempts, and the optimal solution emerges from accumulation of coincidences amid this chaos. Information needed for route planning is not collected in one place – it is distributed so that each agent has local access. This makes ACO particularly attractive for decentralized UAV swarms where centralized information gathering and processing are difficult or impossible.
While Boids and ACO offer heuristic approaches to self-organization, Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) provide a formal mathematical framework for decentralized decision-making under uncertainty. As noted in a 2024 work, “we formulate the dynamic network bridging problem as a novel Dec-POMDP where a swarm of agents cooperatively forms a link between two remote moving targets.”
Dec-POMDP structure for UAV swarms. A 2023 work states, “we model the multi-UAV data collection problem as a Dec-POMDP because each UAV does not know the environment dynamics and can observe only part of the sensors.” This reflects a key feature of swarm systems: each agent has only local information about environment and other agents, yet must make decisions that collectively lead to globally optimal behaviour.
Solving Dec-POMDP via Multi-Agent Reinforcement Learning (MARL). A 2024 work proposed “a multi-agent reinforcement learning (MARL) approach for this problem, based on graph convolutional reinforcement learning (DGN), which naturally applies to the networked, distributed nature of the task.” Another 2024 study proposes “an attention-based observation processing method to extract features from UAV observations and neighbours within the main lobe, improving algorithm performance.”
Application in swarm systems. A 2025 work presents “an analytical framework for decision-making in drone swarm systems operating under uncertainty, integrating POMDP with Deep Deterministic Policy Gradient (DDPG) reinforcement learning.” Another 2025 study formulates “the defence problem as a multi-agent partially observable Markov decision process, reflecting the uncertain nature of UAV swarms under attack.”
Dec-POMDP provides a rigorous mathematical basis for designing decentralized decision-making systems, accounting for information limitations, environmental uncertainty, and the need for inter-agent coordination. Combined with MARL, Dec-POMDP enables adaptive swarm systems capable of learning optimal behaviour in complex dynamic environments.
Based on biological algorithms and formal decision-making models, the grid-weight model is proposed – a decentralized decision mechanism mimicking stochastic information accumulation in a swarm. This model is especially relevant for reconnaissance-strike UAV swarms, where agents must collectively identify targets and decide on attacks without centralized command.
Mathematical formalization:
The area (e.g., terrain map) is divided into discrete grid cells, each assigned a weight wi(t) at time t. The weight reflects the “attractiveness” or “suspiciousness” of a cell regarding the presence of targets. Weight updates follow:
where: wi(t) – weight of cell ii at time t; di(t)∈{0,1} – binary detection (agent detects a target feature in cell ii at time t); α – sensor significance coefficient (weight added upon detection); β – information evaporation rate coefficient (analogous to pheromone evaporation); λ – evaporation speed (temporal decay factor).
wi(t+1)=wi(t)+α⋅di(t)−β⋅λ⋅wi(t)
Decision threshold: Attack or detailed inspection is triggered when the cell weight exceeds wth=10:
if wi(t)≥wth=10, then initiate attack/detailed inspectionif wi(t)≥wth=10, then initiate attack/detailed inspection
False-positive compensation: Upon re-inspection showing no target, a negative coefficient −β reduces weight, preventing false repeated attacks.
Interpretation. The grid-weight model combines key swarm intelligence principles observed in natural systems:
- Positive feedback (α) – analogous to pheromone accumulation: the more agents detect features in a cell, the higher its weight.
- Negative feedback / evaporation (βλ) – analogous to pheromone evaporation: information decays over time if not confirmed by new detections.
- Decentralized information accumulation – each agent locally updates cell weights based on its own observations and then disseminates information via neighbour exchange.
- Stochastic nature – detections di(t) are probabilistic, reflecting sensor uncertainty in real environments.
Advantages for UAV swarms:
- No single point of failure – the decision emerges from collective behaviour, not a central node.
- Scalability – the model works regardless of agent count.
- Resilience to false positives – evaporation and negative reinforcement compensate for individual errors.
- Adaptability – weights automatically redistribute when situation changes.
Consider a practical scenario for the grid-weight model in a reconnaissance-strike UAV swarm. This example illustrates how simple local rules and stochastic information accumulation enable collective decision-making without centralized control.
Initial conditions. A swarm of 6 VTOL UAVs (4 reconnaissance with optical cameras, 1 with thermal imager, 1 with combat payload). Communication between drones is available, terrain map is pre-loaded, all are equipped with computer vision modules. The swarm operates fully autonomously without external communication or global navigation systems.
Phase 1. Primary scanning and weight initialization. The reconnaissance area is divided into 1×1 km cells, each with initial weight wi(0)=0. UAVs scan the area with maximum coverage. In one cell, signs of artificial objects (geometrically regular shapes) are detected – di(1)=1, increasing the cell’s weight by α.
Phase 2. Information accumulation and dissemination. The detecting UAV transmits the updated map to neighbours. Each subsequent pass over the cell with confirmed features adds +α. The thermal-imaging UAV may assign +2α (natural thermal signatures are not expected). Information spreads uniformly through map exchanges – no central accumulation point.
Phase 3. Threshold attainment and scenario change. As weight accumulates (e.g., w=6), the cell attracts more attention. UAVs adjust trajectories, descending and narrowing the scan zone. At w=8, the first UAV with sufficient weight switches to detailed inspection at minimum altitude. Upon discovering a new revealing feature, the weight increases to w=10.
Phase 4. Attack decision. At w=10, attack is initiated. The combat-payload UAV receives the updated map with the identified target. The attack decision is not made by any single agent – it emerges from collective “voting” through weight accumulation.
Key features:
- No hierarchy – all agents are equal; no “flying headquarters.”
- Loss resilience – loss of one or several scouts does not disrupt the process because information is distributed.
- Error compensation – false positives are compensated by evaporation and negative reinforcement upon re-inspection.
- Scalability – the scheme works regardless of swarm size.
Swarm intelligence algorithms inspired by biological systems provide a powerful toolkit for decentralized UAV swarm control. Boids shows how simple local rules generate complex collective motion; ACO shows how stochastic pheromone communication enables optimal solutions in distributed systems; Dec-POMDP provides a formal mathematical basis for decision-making under uncertainty.
The proposed grid-weight model with update rule
and decision threshold wth=10 combines key principles of these approaches, offering a practical mechanism for decentralized targeting in reconnaissance-strike swarms. The model ensures no single point of failure, scalability, resilience to false positives, and adaptability – characteristics critical for combat applications.
wi(t+1)=wi(t)+αdi(t)−βλwi(t)
As noted in a 2023 review, “the study summarizes the basic structure of swarm intelligence algorithms, consisting of several fundamental phases, and provides a comprehensive overview of swarm intelligence algorithms for four aspects of multi-UAV cooperation.” Future research should focus on experimental validation of proposed models in real conditions and their integration with modern machine learning methods to enhance adaptability and effectiveness.
3.2. Ensuring the Stability of Radio Communication and Protection from Interference
This section examines strategies for ensuring reliable communication between UAV swarms in conditions of significant radio interference. It analyzes methods for reducing radiation levels, adaptive data transmission modes, and an analytical model of the signal-to-interference ratio for quantitatively assessing the quality of communication channels. Practical recommendations are offered to improve noise immunity and ensure uninterrupted communication in complex electromagnetic environments encountered in civilian applications.
When deploying swarms of UAVs, a fundamental problem arises related to electronic warfare (EW). Modern EW systems can effectively suppress control, data, and navigation channels, making traditional centralized swarms extremely vulnerable. As noted in a 2024 study, “the effectiveness of UAV swarms can be seriously compromised by electronic jamming technologies, requiring the development of robust counter-strategies” (Jeong et al., 2024).
Under intense electronic countermeasures, a swarm may lose coordinated capability, breaking into individual, easily vulnerable elements. At the same time, as noted in a 2023 review, “the future lies not in anti-aircraft missiles but specifically in EW stations, because they allow focusing radiation in a narrow sector.” This fundamental confrontation drives the need for specialized tactics and technical solutions enabling UAV swarms to maintain combat effectiveness under active electronic countermeasures.
This section systematically analyses tactics for UAV swarm deployment in EW-affected zones. Radio-silence strategy, intermittent communication sessions, and an analytical signal-to-jammer ratio (J/S) model are considered for quantitatively assessing communication channel vulnerability and calculating required jamming power when using frequency-hopping spread spectrum (FHSS). Based on the analysis, practical recommendations for reducing detectability and enhancing swarm survivability under EW are formulated.
One of the most effective EW counter-tactics is radio-silence mode – a strategy in which UAVs minimize or completely cease radio emissions to reduce detectability and maintain stealth. As noted in a 2026 combat-use review, “on various front sections, Russian troops began using strike UAVs in full radio-silence mode until the moment of attack. This complicates drone detection and threat response.”
“Externally silent, internally active” swarm concept. A 2025 work proposed a self-tuning model for swarm formations, “characterized by internal active communication and externally silent operation” (Semantic Scholar, 2025). This concept assumes limited information exchange inside the swarm for coordination, while external emissions are minimized, hindering detection by electronic intelligence (ELINT).
Reduced duty cycle and intermittent communication sessions. Instead of continuous radio exchange, the swarm can use intermittent sessions – short, strictly timed transmissions that significantly reduce interception and direction-finding probability. As one review states, “for covert operations, the protocol minimizes data exchange while using high-level encryption.”
Autonomy as the basis for radio silence. Effective radio-silence tactics require high autonomy of each swarm agent. As stressed in a 2025 work, “communication intermittency demands a high degree of onboard autonomy” (KPI, 2025). Agents must be able to perform assigned tasks, maintain formation, and respond to changing situations without constant external control, relying on pre-loaded algorithms and local interactions.
Link-loss detection and adaptation. A 2025 work investigated “classifying incoming attacks and/or link losses and dynamically reacting based on mission objectives. Affected drones attempt to re-establish connection with other swarm members” (IEEE, 2025). This enables adaptation even when radio silence is broken or becomes impossible.
Intermittent sessions imply a transition from continuous radio exchange to short, tightly synchronized transmissions, minimizing on-air time and thus reducing detection and jamming probability.
Dynamic topology control. A 2025 work presented “an autonomous and distributed UAV swarm architecture designed to enhance operational resilience to network disruptions, particularly EW jamming attacks” (IEEE, 2025). The swarm architecture uses “a leader UAV around which subordinate UAVs orbit, maintaining connectivity via periodic heartbeat signals, tracking packet delivery ratios (PDR).” Adaptive relay ranges, adjusted by changing UAV orbit radius, significantly improve communication resilience.
Periodic channel state monitoring. Instead of continuous data exchange, the swarm can use periodic “sounding” signals to assess channel quality and identify EW-affected zones, allowing agents to decide on route or frequency changes without staying on-air constantly.
Use of pre-loaded data. Much mission-critical information can be loaded into agent memory before take-off, including terrain maps, target data, decision algorithms, and behaviour scenarios. This minimizes in-flight data exchange and thus reduces swarm radio emissions.
Joint communication-motion planning. A 2025 work studied “the joint communication-motion planning problem for UAV swarms in the presence of EW” (IEEE, 2025). This approach optimizes agent trajectories not only for task execution but also for maintaining communication under EW, minimizing time spent in unfavourable electromagnetic environments.
To quantitatively assess communication channel vulnerability to EW, the signal-to-jammer ratio (J/S) model is used. It allows calculating required jamming power for various scenarios and evaluating the effectiveness of protection methods like FHSS.
Basic J/S formula. According to the radio equation, the ratio of jammer power Pj to signal power Ps at the receiver input is:
J/S=(Pj⋅Gj⋅Gr⋅λ2⋅Rs2)/(Ps⋅Gs⋅Gr⋅λ2⋅Rj2)=(Pj⋅Gj⋅Rs2)/(Ps⋅Gs⋅Rj2)
where: Pj – jammer transmitter power (W); Ps – signal source transmitter power (W); Gj – jammer antenna gain; Gs – signal source antenna gain; Gr – receiver antenna gain (cancels); λ – wavelength (cancels); Rj – distance from jammer to receiver (m); Rs – distance from signal source to receiver (m).
Simplified engineering form:
Pj=(J/S)⋅Ps⋅(Gs/Gj)⋅(Rj/Rs)
Required jamming power for FHSS. Frequency-hopping spread spectrum (FHSS) systems are more resilient to EW because the signal hops frequencies pseudo-randomly. To effectively jam an FHSS signal, the jammer must either have enough power to cover the entire hop band (wideband barrage jamming) or precisely track frequency changes (follow-on jamming).
Power requirements for wideband barrage jamming. The required power scales with the ratio of hop bandwidth BWFH to channel bandwidth BWch:
Pj,wide=Pj,narrow⋅(BWFH/BWch)
For a UAV swarm with FHSS at 1000 hops/s and 1 W transmitter power, jamming at 2 km range may require Pj≥50 with a directional antenna and Pj≥500 W with an omnidirectional antenna. This makes mobile EW stations potentially vulnerable to counter-battery fire, as they must radiate significant power.
Limitations of multi-channel jamming. As noted in a 2024 study, “multi-channel jamming is ineffective against FHSS systems unless a sufficient number of channels are successfully jammed.” Jamming FHSS with many channels (e.g., 1000) requires either enormous power or precise knowledge of the hop sequence, making such attacks extremely difficult in practice.
Comparative effectiveness of different jamming types. A 2024 work notes that “against FHSS, swept-frequency jamming can yield unreliable results because both jammer and signal frequencies continuously change.” This makes FHSS particularly effective for UAV swarms operating under active EW.
Modern research increasingly shifts toward intelligent EW countermeasures using machine learning and collective swarm intelligence.
Reinforcement learning for anti-jamming. A 2025 work presented “a deep reinforcement learning method for adaptive data rate and frequency selection in UAV ad-hoc networks” (IEEE, 2025), enabling agents to adjust communication parameters in real time to changing jammer conditions.
Graph convolutional networks for jamming zone prediction. A 2024 work proposed “a novel approach where UAV swarms use collective intelligence to predict electronic jamming zones, avoid them, and effectively reach target points” (Jeong et al., 2024). The approach uses graph convolutional networks (GCN) to “predict jamming zone locations and intensities based on information collected from each UAV.”
Collective sensing and adaptation. A multi-agent control algorithm “is used to disperse the UAV swarm, evade jamming, and regroup upon reaching the target” (Jeong et al., 2024). The proposed method demonstrates “robustness, scalability, and computational efficiency.”
Game-theoretic frequency management. A 2025 work proposed “the UAV-FPG (Frequency Point Game) environment, modelling dynamic interaction between jamming and anti-jamming strategies in frequency bands” (Yang et al., 2025). The model integrates “expert knowledge base for frequency selection optimization and uses large language models for route planning.”
MARL for cooperative jamming and target protection. A 2025 work studied “the cooperative jamming and target protection problem in UAV swarms” (IEEE, 2025). In dynamic signal environments, “UAVs equipped with electronic jamming modules can effectively reduce system detectability, providing strategic advantages for protected assets.”
AI-based swarm intelligence. A 2024 work investigated “the effectiveness of anti-jamming methods for UAVs supported by AI-based swarm behaviour” (Silva, 2024). Results show that “genetic approaches can achieve effective results, though often at the cost of long simulation time.” In contrast, “reinforcement learning (RL) shows robust results when pre-trained with realistic data.”
Based on the analysis, the following practical recommendations for designing and deploying UAV swarms under EW can be formulated:
1. Use FHSS. FHSS is the most effective defence against wideband jamming. Systems with high hop rates (≥1000 hops/s) and wide frequency bands are recommended, making jamming economically unattractive. As noted in a 2024 study, “multi-channel jamming is ineffective against FHSS systems.” Against FHSS, “swept-frequency jamming can give unreliable results.”
2. Reduce duty cycle and use intermittent sessions. Instead of continuous radio exchange, use short, tightly synchronized transmissions. This reduces average radiated power and detection probability by ELINT. For covert operations, protocols should “minimize data exchange while using high-level encryption.”
3. Apply radio-silence tactics. Where stealth is critical, the swarm should execute missions using pre-loaded data and algorithms, completely avoiding radio exchange. This requires a high degree of onboard autonomy, as “communication intermittency demands a high degree of onboard autonomy.”
4. Adaptive power control. Transmitter power should be adjusted based on distance to neighbours and interference level. Minimizing excess power reduces detectability and saves onboard energy.
5. Spatial dispersal and channel redundancy. The swarm should be able to disperse when jamming zones are detected while maintaining connectivity through alternate routes and relays. As shown in a 2024 work, “UAV swarms use collective intelligence to predict EW jamming zones, avoid them, and effectively reach target points” (Jeong et al., 2024).
6. Intelligent prediction and adaptation. Use machine learning to predict jamming zones and adapt communication parameters in real time. GCNs allow “predicting jamming zone locations and intensities based on information collected from each UAV.”
7. Hybrid communication architectures. Combining MANET for intra-cluster (low latency) and DTN for inter-cluster (resilience to breaks) optimizes radio traffic depending on tactical situation. The structure proposed in IEEE (2025) with “periodic heartbeat signals” provides adaptive relay ranges.
8. Use passive detection methods. Instead of active probing (which can be detected), the swarm can use passive methods to locate jamming sources. As shown in a 2025 work, “in the presence of jammers, UAVs can use passive radars to position the jammer source” (IEEE, 2025).
Countering electronic warfare is one of the most important tasks in the practical deployment of UAV groups.
Traditional centralized architectures relying on continuous radio exchange are extremely vulnerable under active EW. Transitioning to decentralized, self-organizing swarms opens new opportunities for survivability through radio-silence tactics, intermittent communication, and intelligent adaptation to jamming environments.
The proposed analytical J/S model J/S=(PjGjRs2)/(PsGsRj2) and FHSS jamming-power calculations allow quantitative assessment of communication channel vulnerability and informed selection of communication system parameters for UAV swarms. Using FHSS with high hop rates (≥1000 hops/s) significantly increases jamming resilience, requiring either enormous power or precise knowledge of the pseudo-random sequence from the adversary.
Modern research increasingly focuses on intelligent EW countermeasures using machine learning, GCNs, and MARL. These methods enable UAV swarms to predict jamming zones, adapt communication parameters and trajectories in real time, ensuring mission execution even under active electronic countermeasures.
As combat experience shows, radio-silence tactics combined with high onboard autonomy become an effective means against modern EW systems. Future research should focus on experimental validation of proposed models in real conditions and their integration with modern artificial intelligence methods to enhance adaptability and survivability.
3.3. Energy Autonomy and Intra-Swarm Recharging
Energy autonomy remains one of the most critical limitations for practical UAV swarm deployment. As noted in a 2022 review, “uninterrupted mission execution of multicopter swarms requires predicting energy depletion and planning maintenance and replacement processes” (ACM, 2022). Limited battery capacity, high flight power consumption, and communication maintenance create a complex of interrelated challenges requiring a systemic approach to swarm energy management.
Unlike single UAVs, where the energy problem reduces to maximizing individual flight time, swarms require solving a fundamentally different problem: distributing limited energy resources among multiple agents to achieve the overall mission goal. As stressed in a 2020 work, “multiple UAVs in a swarm must cooperate to balance their energy consumption during deployment” (IEEE, 2020). This necessitates both energy-efficient control algorithms and new approaches to intra-swarm recharging and energy exchange.
This section systematically analyses energy aspects of UAV swarm operation. Power consumption models for multicopters and VTOL aircraft, communication costs for different architectures (MANET vs. DTN), and promising intra-swarm recharging concepts – from flying batteries to wireless power transfer – are considered. Based on the analysis, practical recommendations for choosing the trade-off between range and loiter time, as well as optimizing swarm energy balance, are formulated.
Power consumption is a key parameter determining UAV tactical-technical characteristics. For swarms, understanding energy models of different platform types becomes critical because platform choice affects not only mission duration but also coordination and interaction capabilities.
Multicopter power models. Gong et al. (2023) presented closed-form power consumption models for multirotor UAVs in three flight modes. “The proposed power models not only reveal how multirotor UAV power consumption depends on various factors but also pave the way for other applications” (Gong et al., 2023). A 2025 study presents a modular approach to multicopter modelling, “considering vehicle dynamics, power consumption, and sensor integration.” The propulsion model “includes detailed descriptions of key components such as Li-ion battery, electronic speed controllers, and brushless DC motors” (arXiv, 2025).
Power consumption in different flight modes. Gong et al. (2023) show that multirotor power consumption strongly depends on flight mode – hover, level flight, or manoeuvre. For a ~2 kg multicopter, hover power is 150–200 W. A 2024 work notes that VTOL aircraft “reduce power consumption by 80% compared to normal hovering” thanks to energy-efficient cruise mode (Uncrewed Systems Technology, 2024).
Comparative analysis of multicopters and VTOL. UCL (2026) shows that “multirotor UAVs provide the most risk-optimal routes due to high manoeuvrability, but are the least energy-efficient. Fixed-wing UAVs are the most energy-efficient, but less effective at risk-aware route planning” (UCL, 2026). For VTOL, the opposite holds: “the aircraft is more energy-efficient when operating at longer ranges” (TU Delft, 2022). As TU Delft (2022) notes, “multicopter architectures require heavier batteries for long-range flights, resulting in significantly higher power. Hence, a multicopter designed for shorter range is more energy-efficient.”
Battery mass influence. MDPI (2023) shows that “optimal battery mass fraction” is key to maximizing multirotor flight time. As TU Delft (2022) states, “battery mass constitutes a significant percentage of aircraft mass.”
VTOL configuration energy consumption. Frontiers (2025) shows that for a 200 km mission, “lift+cruise requires 75.7 kWh, while tiltrotor requires 68.7 kWh, representing about 9.3% lower energy consumption” (Frontiers, 2025), demonstrating that VTOL architecture choice significantly impacts swarm energy budget.
Besides flight-related power, communication maintenance accounts for a significant part of UAV energy budget. In swarms, where each agent must exchange information with neighbours, these costs become especially significant.
Overall power consumption structure. Nature (2024) shows that “three main power drivers – flight dynamics, payload operations, and continuous wireless communication – together account for 80–85% of UAV power consumption” (Nature, 2024). Thus, communication costs are substantial but not dominant.
MANET power consumption. MANET requires constant exchange of control information – hello packets, routing table updates, acknowledgements. As noted in the systematic review (Springer, 2024), “increased energy costs for maintaining numerous wireless connections” are a key FANET issue. Based on literature analysis, MANET communication costs are estimated at 5–10 W per UAV, depending on neighbour count, update frequency, and protocol.
DTN power consumption. Unlike MANET, DTN does not require continuous control exchange, using store-and-forward and tolerating intermittent connectivity. The absence of continuous route maintenance and periodic (rather than continuous) data exchange significantly reduce communication energy. DTN communication costs are estimated at 2–3 W per UAV, saving up to 40% compared to MANET.
Energy-efficient communication management. IEEE (2025) shows that “the proposed method reduces power consumption by up to 27% while maintaining formation control accuracy and achieving 99.8% communication reliability” (IEEE, 2025). Nature (2026) presents simulation results with 500 mobile drones, showing “7.8 Wh power consumption” with high network performance (Nature, 2026).
One of the most promising approaches to solving swarm energy autonomy is the Flying Battery concept – specialized UAVs serving as mobile charging stations for other swarm members.
Flying Battery UAV concept. Preprints (2023) presented the “Flying Battery UAV (FB-UAV) concept, which restores energy to distressed UAVs during mission through two modes: rapid field battery replacement and in-flight cable-connected charging” (Preprints, 2023). This concept “offers UAV pilots the ability to demonstrate emergency procedures while ensuring safety, reliability, and survivability” (Preprints, 2023).
In-flight wireless power transfer. A breakthrough direction is wireless power transfer (WPT) for UAV swarms. In 2026, Reach Power and Gambit received funding to create drone swarms “that never need to land.” The system “combines adaptive intelligence with wireless power transfer to remove battery limitations for autonomous operations” (Reach Power, 2026). Reach’s solution allows drones to “fly to designated points and recharge during mission without landing and without human intervention” (Reach Power, 2026). Gambit provides AI control that “allows each vehicle to track battery charge level, mission requirements, and charging capabilities, then adjust routes and behaviour in real time” (Reach Power, 2026).
Chinese wireless charging developments. In 2026, China tested a prototype orbital solar power station that “can transmit 143 W of power to a drone in flight.” The system achieved “20.8% DC-to-DC wireless power transfer efficiency at 100 metres” (Gizmodo, 2026).
Capacitive wireless charging for swarms. IEEE (2025) proposed “an electric capacitive wireless charging system for UAV swarms, capable of simultaneously charging multiple drones through a single transmitting unit with excellent misalignment tolerance and load decoupling capability” (IEEE, 2025).
Charging route optimization. IEEE (2024) investigated “3D hovering placement and drone route optimization for one-to-many continuous wireless charging tasks” (IEEE, 2024), enabling efficient planning of charging drone trajectories to serve many working vehicles.
Swarm architecture – centralized, hierarchical, or decentralized – significantly affects overall system power consumption and thus mission duration.
Energy-efficient position reconfiguration. IEEE (2024) proposed “an efficient position reconfiguration scheme that reduces swarm power imbalance and extends lifetime.” The scheme “solves the practical problem of extending drone lifetime in such environments by optimizing position reconfiguration” (IEEE, 2024).
Balanced energy consumption. IEEE (2020) notes that “multiple UAVs in a swarm must cooperate to balance their energy consumption during deployment for full user coverage in remote target areas” (IEEE, 2020). This requires algorithms considering not only task efficiency but also energy balance among agents.
Energy budget for a swarm hub. AIAA (2025) studied “power consumption and endurance of a 7.2-kg octocopter serving as an aerial hub in a drone swarm.” A “phase-wise electrical energy model of the mission, explicitly including continuous avionics demand measured at 45 W” was built (AIAA, 2025).
Topology effect on power consumption. As noted in a 2022 work, “communication topology selection in drone swarms directly affects system performance.” Different architectures “show different advantages and limitations in terms of reliability.” Energy efficiency also depends on topology: dense formations require lower transmission power but increase collision risk, while sparse formations increase link range and hence energy consumption.
Based on the analysis, the following practical recommendations for choosing the range-loiter-time trade-off for UAV swarms are formulated:
1. Platform selection by mission. For missions requiring long loiter over a point (e.g., surveillance, relay), VTOL aircraft with energy-efficient hover mode are preferred. As shown in Uncrewed Systems Technology (2024), VTOL can achieve “up to two hours of hover compared to 20–35 minutes for a similarly sized traditional multicopter” (Uncrewed Systems Technology, 2024). For long-distance transit missions, fixed-wing or VTOL with fixed wings are preferred.
2. Optimize worker-to-donor drone ratio. A 3:1 ratio of worker drones to donor drones can increase total mission time by 60–80%. One donor drone (carrying spare batteries or wireless charging system) serves three workers, supplying energy during the mission. This compensates for limited battery capacity and significantly extends autonomous operation time.
3. Use hybrid communication architectures. Protocol choice should be driven by mission latency and energy requirements. MANET provides low latency (<100 ms) with high power consumption (5–10 W per agent), justified for tactical tasks requiring fast coordination. DTN provides low power (2–3 W per agent) with high latency (1–5 min), acceptable for reconnaissance data collection and missions with delayed delivery. Hybrid (MANET intra-cluster, DTN inter-cluster) allows optimizing power consumption depending on tactical situation.
4. Apply intermittent communication sessions. Instead of continuous radio exchange, use periodic short sessions. This reduces average radiated power and thus energy consumption without significantly degrading coordination quality.
5. Intelligent energy management. Using machine learning for power consumption prediction and task distribution optimization can significantly improve overall swarm efficiency. As shown in IEEE (2024), “the proposed method reduces power consumption by up to 27% while maintaining formation accuracy” (IEEE, 2025). Nature (2026) presents results showing “7.8 Wh power consumption” with high network performance (Nature, 2026).
6. Use renewable energy sources. For long missions, energy-neutral swarms with hybrid power systems including solar panels, energy harvesting, and wireless recharging are promising. As Nature (2025) states, “symbiotic energy systems based on energy-harvesting technologies continuously replenish onboard energy storage, integrating energy harvesting into aerial robot platforms” (Nature, 2025). A 2023 work presented the “Energy Neutral Internet of Drones” concept, where “renewable energy harvesting (EH) is used to achieve energy neutrality, minimizing the gap between harvested and consumed energy” (Cambridge, 2023).
Energy autonomy remains a key limitation for widespread UAV swarm deployment. Power consumption is determined by a complex of factors: platform type (multicopter, VTOL, fixed-wing), flight mode, communication architecture, swarm density, and mission tactics.
Multicopters, offering high manoeuvrability, have the highest power consumption (150–200 W in hover), while VTOL can reduce power by up to 80% in energy-efficient modes, achieving up to two hours of hover. Communication costs are a substantial but not dominant part of the energy budget: MANET requires 5–10 W per agent, while DTN requires only 2–3 W, saving up to 40%.
The flying battery and wireless power transfer concepts open new opportunities for significantly increasing swarm autonomous operation time. Breakthroughs in WPT (Reach Power, 2026) and capacitive charging for swarms (IEEE, 2025) demonstrate the potential for continuously operating swarms without landing. A 3:1 worker-to-donor ratio can increase total mission time by 60–80%.
Choosing the range-loiter-time trade-off must be driven by mission requirements, platform type, communication architecture, and available intra-swarm recharging technologies. Future research should focus on integrated power consumption models accounting for all swarm operation aspects and experimental validation of proposed approaches in real conditions.
3.4. Specialized Tactics for Enhancing Swarm Survivability
Beyond basic self-organization algorithms and EW countermeasures, UAV swarms possess a unique ability to implement a wide range of specialized tactical techniques that enhance survivability and effectiveness against modern air defence systems. As noted in a combat-use analysis (Bellamy, 2024), “drone swarms overload air defence systems, and even a small number of drones that break through represent a serious danger.” It is this ability – combining numerical superiority with intelligent tactical schemes – that makes UAV swarms so effective in modern conflicts.
Unlike single UAVs, whose tactical capabilities are limited by individual characteristics, swarms can implement complex collective schemes, including decoys, dynamic dispersal, terrain-shadow flying, air-corridor organization, and integration with ground sensors. This section systematically analyses specialized survivability tactics based on combat experience and modern research in swarm intelligence and autonomous planning.
Concept and implementation. One of the most effective and widely used tactics is decoy swarms – groups of specially constructed unmanned platforms carrying no combat payload but mimicking the behaviour of strike drones to overload enemy air defences. As noted in a combat-use analysis (LRT, 2024), “unarmed decoys now account for more than half of drones targeting Ukraine, and up to 75% of new drones leaving the plant in Russia’s special economic zone in Alabuga.” This ratio allows the attacker to significantly increase the cost of an attack for the defender: expensive anti-air missiles are spent on destroying cheap decoys.
Operation False Target. In 2024, Russia introduced a tactic called “Operation False Target,” in which a small number of high-destruction thermobaric drones are surrounded by decoy swarms. As reported in an Associated Press investigation (LRT, 2024), “the plan, which Russia called Operation False Target, is designed to force Ukraine to expend scarce resources to save lives and preserve critical infrastructure, including the use of expensive air-defence ammunition.” At the same time, “neither radar, snipers, nor even electronic experts can determine which drones in the sky are lethal.”
Reconnaissance function of decoys. Beyond purely distracting functions, decoys can perform reconnaissance tasks. As noted in the same investigation (LRT, 2024), “one decoy with a live-stream camera allows an aircraft to geolocate Ukrainian air-defence systems and relay information to Russia in the last moments of its mechanical life.” Thus, even when destroyed, decoys provide valuable intelligence on enemy air-defence positions.
Economic effectiveness. Mass use of decoys creates a dilemma for the defender: ignoring decoys may allow strike drones to reach targets, while engaging them rapidly depletes air-defence ammunition. As reports (LRT, 2024) note, “in October 2024, Russia attacked with at least 1,889 drones – 80% more than in August,” with “most drones crashing, being shot down, or diverted by electronic jamming, and less than 6% reaching a discernible target.” However, “the numbers alone mean that a handful can slip through every day – and that is enough to be deadly.”
Radar-signature imitation. Besides physical decoys, electromagnetic imitating methods allow a drone swarm to mask itself as a single large target. As reported in an industry review (Defense Mirror, 2023), “a French project aims to use drone swarms to imitate the radar cross-section of aircraft, misleading and confusing air-defence systems. The main goal is to make the drone swarm appear on radar as a single entity resembling an aircraft or combat drone.” This tactic “is designed to confuse potential adversaries’ air-defence systems.”
Dispersal for EW evasion. Upon detecting jamming zones, the swarm can dynamically change its spatial configuration, dispersing to reduce vulnerability and evade jamming. As shown in Jeong et al. (2024), “a multi-agent control algorithm is used to disperse the UAV swarm, evade jamming, and regroup upon reaching the target.” This enables combat effectiveness even under active EW.
Multi-direction attack. Dispersal can also be used for tactical superiority. As noted in a combat-use analysis (Баo Нян, 2023), “the calculation was based on distributing our air-defence forces.” A drone swarm attack “from different directions” allows “revealing gaps in air defence” and overloading the air-defence system, forcing it to disperse firepower.
Dynamic density management. Depending on tactical situation, the swarm can adjust its formation density. Dense formations provide better mutual support and relay but create mass-casualty risks. Sparse formations reduce vulnerability to single strikes but require more powerful communication and complicate coordination. As noted in IEEE (2025), “effective swarm management and survivability improvement require adaptive density and formation control.”
Reconnaissance and air-defence exposure. A dispersed swarm can be used to probe enemy air defence, forcing it to reveal positions. As noted in a tactic analysis (Кoрреспoндент, 2023), “Russian forces use a group-attack method with kamikaze UAVs to destroy long-range enemy air-defence systems.” The “drone swarm is designed to break through air defence, making air-defence systems choke.”
Terrain-shadow principle. One of the most effective ways to reduce swarm detectability is using terrain shadow – spatial areas shielded from radar illumination by terrain features, buildings, and other obstacles. As noted in Almuzaini and Savkin (2024), “the method establishes a feasible flight zone that dynamically adapts for loss of line-of-sight (LOS) due to elevations and structures between UAV sensors and the target.”
Trajectory planning with shadows. Almuzaini and Savkin (2024) presented “a trajectory planning method for a team of UAVs aimed at enhancing covert video surveillance over uneven terrain and urban environments.” By avoiding “shadows” – projections of real forms onto the UAV operational plane – “the method ensures continuous target visibility.” This strategy “optimizes UAV trajectories while maintaining stealth and adapting to changing environments.”
Quantitative stealth assessment. To effectively use terrain shadow, stealth level must be quantitatively assessed. Li et al. (2024) “quantified terrain threat, radar detection, and penetration time requirements during UAV overflight. For radar threats, a radar-echo analysis method based on RCS and spatial environment is proposed to quantify UAV overflight stealth.”
Low-altitude routes and terrain following. Terrain-following route planning allows the swarm to stay below enemy radar detection zones. As noted in NVO (2023), “swarm path planning for low-altitude fast transit in diverse environments under complex terrain, radar, and swarm-failure threats is investigated.” Low-altitude flights use terrain folds for masking, significantly reducing detection range.
Dynamic shadow formation in urban environments. In urban areas with dense buildings, shadow-use opportunities are especially wide. As shown in Almuzaini and Savkin (2024), “the method prevents LOS loss while maintaining a high level of camouflage.” UAVs can use buildings as natural screens, moving in “radio shadows” between structures.
Capsule-shaped air corridors. To ensure safe swarm movement in constrained airspace, air-corridor organization methods are developed – 3D volumes within which the UAV group moves. As noted in Saveliev and Anikin (2025), “a route planning method based on representing each vehicle’s trajectory as a capsule-shaped air corridor – a fixed-radius 3D volume formed along trajectory segments.”
Safety and spatial separation. The key function of air corridors is to prevent conflicts between individual UAV trajectories. As stressed in Saveliev and Anikin (2025), “spatial redundancy ensures safe trajectory separation at the planning stage, eliminating conflicts during subsequent autonomous flight operations without continuous inter-agent coordination.” The capsule radius “includes a margin for possible deviations from the planned trajectory, ensuring robustness to navigation errors.”
Route construction methodology. Saveliev and Anikin (2025) proposed a “four-phase scheme” for constructing routes for a UAV group, including “vertical ascent from start to working altitude, horizontal transition to the processing-zone entry, return from the zone exit to the start-descent point, and vertical descent to the initial position.” “Each new route is built considering already reserved air corridors through analytical geometric intersection checks between capsules of different trajectories.” For computational efficiency, “hierarchical spatial filtering based on bounding boxes is used.”
Air-corridor scalability. Saveliev and Anikin (2025) conducted “numerical experiments for groups of 2 to 32 vehicles over a typical 1-km² agricultural plot,” revealing “nonlinear increase in planning time and iteration count with agent number.” Route length “tends to increase, especially at early scaling stages, due to the need to bypass already reserved corridors.”
Self-organization of air traffic. An alternative to centralized corridor planning is self-organizing air traffic. IEEE (2025) presents “a solution enabling self-organization of cooperating autonomous agents into an efficient traffic flow state, where the overall conflict-laden air-coordination task is solved.” This approach is especially promising for large swarms where centralized planning becomes computationally expensive.
Air-ground integration concept. To enhance situational awareness and targeting effectiveness, UAV swarms can integrate with ground sensor networks. As noted in a 2023 patent, “an air-ground integrated cloud-based cooperative targeting system for UAV swarms is developed for joint detection, segmentation, and target capture in combat interaction.” Such integration allows “controlling air and ground vehicles for autonomous task execution in independent or cooperative modes.”
Data collection from ground sensors. UAV swarms can be used to collect data from distributed ground sensors whose locations may be unknown. IEEE (2025) proposed “an air-ground cooperative communication scheme using ground base stations to localize unknown-position sensors and UAVs to collect data from localized ground sensors.” This approach “allows UAVs to quickly reach target points for data collection and transmission, achieving efficiency and speed.”
Ground sensors for targeting. Ground sensors can provide swarms with critical target information unavailable from the air. As noted in CNA (2023), “UAVs are used to detect and track targets, and to dispatch ground vehicles to detected locations,” with “communication minimized” to enhance stealth.
Heterogeneous UAV-UGV coalitions. Modern research increasingly focuses on heterogeneous coalitions of aerial and ground unmanned vehicles. IEEE (2025) presented “a novel heterogeneous cooperative multi-agent deep reinforcement learning algorithm for air-ground integrated sensing and communication systems, balancing individuality and cooperation among heterogeneous vehicles for optimized network sensor data aggregation.”
Biomimetic swarms for monitoring and emergency response. A promising direction is the use of UAV swarms deployed via specialized transport platforms to perform monitoring and rapid response tasks in areas with limited accessibility. A 2025 study proposed a novel approach to detection and situational assessment using biomimetic UAV swarms launched from mobile platforms equipped with intelligent systems, enabling effective monitoring and coordination of response activities across vast territories with minimal human involvement.
Task division between air and ground systems. Safran (2024) shows that “integrating air and ground systems achieves higher situational awareness levels because each component uses its unique sensing capabilities.” Air systems provide global overview and communication, while ground sensors provide detailed target information.
Based on the analysis, the following practical recommendations for UAV swarm route planning considering enemy air defence can be formulated:
1. Use decoy swarms to overload air defence. Include a significant proportion (up to 50–75%) of unarmed decoys mimicking strike drone behaviour. This depletes enemy air-defence ammunition and creates uncertainty in target identification. As combat experience shows (LRT, 2024), “neither radar, snipers, nor even electronic experts can determine which drones in the sky are lethal.”
2. Multi-direction attack. To overcome layered air defence, attack simultaneously from multiple directions (Баo Нян, 2023). This “distributes air-defence forces” and reveals “gaps in air defence.” As noted in a tactic analysis (Кoрреспoндент, 2023), “a massed drone swarm is aimed at breaking through the air-defence system and hitting targets.”
3. Use terrain shadow for reduced detectability. Plan routes considering “shadows” – areas shielded from radar illumination by terrain and buildings (Almuzaini and Savkin, 2024). The method should “dynamically adapt for LOS loss due to elevations and structures.” Low-altitude terrain-following flights allow “maintaining a high level of camouflage.”
4. Organize air corridors with spatial redundancy. For safe swarm movement under dense air traffic and air-defence threats, use capsule-shaped corridors (Saveliev and Anikin, 2025). “Spatial redundancy ensures safe trajectory separation at the planning stage, eliminating conflicts during subsequent autonomous flight operations.” The capsule radius should “include a margin for possible deviations.”
5. Dynamic dispersal upon threat detection. When jamming zones or air-defence concentrations are detected, the swarm should dynamically disperse (Jeong et al., 2024). “A multi-agent control algorithm is used to disperse the UAV swarm, evade jamming, and regroup upon reaching the target.”
6. Integrate with ground sensors for targeting. To improve targeting accuracy and situational awareness, integrate the swarm with ground sensor networks (IEEE, 2025). “The air-ground integrated system allows controlling air and ground vehicles for autonomous task execution.”
7. Use electromagnetic masking and imitation. To mislead enemy air defence, use electromagnetic imitation methods making the drone swarm “appear on radar as a single entity resembling an aircraft or combat drone” (Defense Mirror, 2023), “designed to confuse potential adversaries’ air-defence systems.”
Specialized survivability tactics are an integral component of effective UAV swarm deployment against modern air-defence systems. Decoy swarms overload enemy air defence and create target-identification uncertainty, significantly increasing attack cost-effectiveness. Dynamic dispersal and multi-direction attacks complicate swarm engagement and force air defence to disperse firepower.
Terrain-shadow and low-altitude corridors allow swarms to stay below radar detection, significantly reducing vulnerability. Air-corridor organization with spatial redundancy ensures safe movement in constrained airspace. Integration with ground sensors expands situational awareness and improves targeting accuracy.
As combat experience shows, combining these tactical techniques allows UAV swarms to effectively overcome even layered air defence. Future research should focus on developing integrated planning systems accounting for all tactical behaviour aspects and their experimental validation in real conditions.
4. Swarm Economics: Cost of Ownership and Effectiveness
Economic aspects of UAV swarm deployment are perhaps the most compelling argument for transitioning from traditional manned systems and expensive precision munitions to decentralized swarm systems. As noted in an industry analysis (Defence Agenda, 2025), “the problem of UAV swarms against short-range air-defence systems is essentially an economic puzzle wrapped in a kill-chain problem.” Unlike single UAVs, where cost-effectiveness is determined by the ratio of platform cost to mission value, swarms create a fundamentally new economic reality: attack cost is distributed among many agents, and the defender’s cost to counter such an attack can far exceed the attacker’s cost.
This section systematically analyses economic aspects of UAV swarm deployment. The phenomenon of economic asymmetry is considered, swarm costs are compared with traditional strike means, a Total Cost of Ownership (TCO) model for swarm systems is proposed, and the economic rationale for transitioning to swarms based on combat experience is analysed.
The negative cost-exchange ratio phenomenon. One of the key economic characteristics of modern warfare with UAV swarms is the negative Cost-Exchange Ratio (CER) – a situation where defending against a cheap threat requires disproportionately expensive responses. As Defence Agenda (2025) notes, “a negative cost-exchange ratio punishes any force that responds to every quadcopter, loitering munition, or decoy with expensive strike assets.” DroneLife (2025) emphasizes that “the problem of negative cost-exchange ratio, where defending against low-cost threats requires disproportionately expensive counteractions, is becoming one of the defining challenges of our time.”
Quantitative CER estimates. Command Eleven (2026) presents an asymmetric attrition model showing that “the cost-exchange ratio (CER) currently stands at approximately 190:1 in favour of drone operators.” This means the attacker spends $1 to create a threat, while the defender must spend $190 to neutralize it. “Saturation occurs when the swarm volume exceeds AEGIS fire-control system capabilities” (Command Eleven, 2026). As Defence Agenda (2025) notes, “countries that master layered, low-cost defence and deploy their own expendable swarms can rebalance the cost-exchange ratio in their favour. Otherwise, responses solely with precision assets will exhaust budgets long before the decisive battle.”
Asymmetric effects in combat. Real combat demonstrates impressive examples of economic asymmetry. During “Operation Spider’s Web,” Ukraine used 117 FPV drones costing less than $5,000 each, coordinated via mobile networks, to execute an attack (ICWA, 2025). “In strategic terms, Ukraine spent approximately $3 million on drones to inflict material damage estimated at $3 billion – an asymmetric strike with a 1:1000 cost-exchange ratio” (The News, 2025). As ICWA (2025) stresses, “this operation redefined asymmetric warfare, demonstrating that low-cost AI-supported UAVs can deliver disproportionate strategic results.”
Cost-exchange examples. The News (2025) provides specific examples: “a $10,000 drone destroying a $5 million tank gives a 1:500 CER.” “A $200,000 naval drone destroying a $500 million submarine gives a 1:2500 CER” (The News, 2025). As Medium (2025) notes, “a $6 million main battle tank can be destroyed by a $200 FPV drone carrying an explosively formed penetrator (EFP). This math makes traditional armour obsolete.”
Cost of a 50-UAV swarm. To assess swarm economic effectiveness, swarm cost must be compared with alternative strike assets. A 50-drone swarm of medium-class UAVs (e.g., loitering munitions) can have a total cost from $1 to $3.5 million depending on platform type. For comparison, one precision-guided cruise missile (e.g., Storm Shadow, SCALP-EG, or JASSM) can cost $2–3 million, and one air-defence interceptor (e.g., IRIS-T or NASAMS) costs about $1 million (United24 Media, 2025; Air & Space Forces, 2025).
Comparison with expensive air-defence systems. As FT (2025) notes, “the cost of several drone swarms is about $20,000, while air-defence missiles cost $500,000, and one fighter can exceed $100 million.” A 50-drone swarm, even at $70,000 per vehicle (as with Rogue 1 for the US Marine Corps), would cost around $3.5 million – comparable to one cruise missile, but the swarm provides multi-channel attack much harder to defeat (LinkedIn, 2025).
Cost reduction through mass production. Mass drone production substantially reduces unit cost. As CSIS (2025) notes, “if in 2022 Russia paid an average of $200,000 for one Shahed-136 drone purchased from Iran, by 2025 cost dropped to about $70,000 due to mass production at the Alabuga plant.” CSIS gives even lower estimates: “Shahed-136 cost may range from $20,000 to $50,000” (CSIS, 2025). United24 Media (2025) also confirms that “one Shahed drone may cost Russia about $20,000–$50,000 to produce.” At these prices, a 50-drone swarm could cost from $1 to $2.5 million.
Comparison with manned aviation. The economic advantage over manned aviation is even more obvious. As Research and Markets (2023) notes, “swarm technologies today use small, cheap drones rather than larger, more expensive UAVs, offering end users cost and time efficiency.” LinkedIn (2023) states, “drones possess unmanned capabilities, affordability, and swarm technologies, prompting consideration of how quickly drones can replace traditional aircraft and fighters.”
TCO definition for swarm systems. TCO for swarm systems includes all costs associated with development, production, deployment, operation, maintenance, and disposal over the entire life cycle. As Moneypro (2026) notes, “programs succeed or fail based on total cost of ownership (TCO) – the sum of all costs required to deploy, operate, upgrade, and sustain a capability throughout its service life. TCO includes all costs from programme start to completion, typically over 10–20 years.”
TCO components for UAV swarms. Based on literature analysis, the following TCO structure for swarm systems can be proposed:
- Development and design costs. Research, design, prototyping, testing, certification. TU Delft (2024) shows that “total swarm cost can be derived from all estimated costs.” For a wildfire-drone swarm, “total cost is no more than $755,146” (TU Delft, 2024).
- Production costs. Component procurement, assembly, quality control, supply logistics. Unpatentable (2024) estimates: “annual production costs (100 units): $3,300,000; inventory and logistics: $150,000; training and documentation: $300,000.” Mass production and modularity are key to lowering production costs. As Swarm Defense (2024) states, “our drones are designed to be affordable, modular, and mass-production ready.”
- Personnel training costs. Operator, technical, and command staff training. Unpatentable (2024) estimates training and documentation at $300,000 for 100 units. As AA.com.tr (2025) notes, “Europe must also conduct realistic exercises, including large-scale exercises against drone swarms – a scenario NATO allies still rarely practise.”
- Operations and maintenance costs. Energy (charging/fuel), component replacement, repair, supply logistics. Ronin’s Grips (2026) notes that “operational costs per flight hour for precision systems can be around $35,200.” For swarms, these costs are significantly lower due to design simplicity and modularity.
- Attrition costs. Cost of drones lost during missions. This is one of the most significant TCO components because swarms inherently assume some loss level. However, combat experience shows that even with losses, swarms remain cost-effective. During Operation Spider’s Web, losses were 117 drones costing less than $5,000 each – about $585,000 total – while causing $3 billion in damage (ICWA, 2025).
- Upgrade and disposal costs. Software updates, component replacement, disposal of retired platforms.
TCO comparison of swarms and traditional systems. As Defence Agenda (2025) notes, “swarms of small systems, improvised designs, and decoys can impose disproportionate costs, draining resources and creating strategic tension.” Research and Markets (2025) stresses that “swarm technologies provide saturation attacks that overwhelm conventional air-defence systems at dramatically lower per-unit costs than crewed platforms, providing asymmetric tactical advantages.” CSIS (2025) notes that “Shahed-class barrages are designed to stress Patriot, NASAMS, and IRIS-T systems, forcing unfavourable exchange.” At a Shahed-class cost of about $35,000 (CSIS estimate), a 50-drone swarm costs $1.75 million – far cheaper than one interceptor (~$1 million) or one cruise missile ($2–3 million).
Cost per successful strike. A key metric of swarm economic effectiveness is cost per successful strike. This accounts for drone costs, target-hit probability, preparation and deployment costs, and losses. As Alarabiya (2025) notes, “nine times out of ten, it is about the cost-performance ratio or cost per kill.” Defence Agenda (2025) emphasises that “system design and logistics are optimised to reduce cost per kill and ensure a high operational tempo.”
Comparison with manned aviation. The economic advantage over manned aviation is one of the most compelling arguments for transitioning to swarms. As Stratistics MRC (2026) states, “swarm technologies provide saturation attacks that overwhelm conventional air-defence systems at dramatically lower per-unit costs than crewed platforms.” FT (2025) notes that “the cost of several drone swarms is about $20,000, while air-defence missiles cost $500,000, and one fighter can exceed $100 million.” A 50-drone swarm at $70,000 each ($3.5 million) is about 30 times cheaper than one fighter ($100 million) and can deliver comparable or greater damage.
Real-world effectiveness. Combat experience demonstrates impressive cost-effectiveness. During Operation Spider’s Web, Ukraine showed that “drones costing less than $1,000 neutralising multi-million-dollar aircraft radically changed the cost-exchange ratio” (Bolt Flight, 2025). “This operation proves that tactical ingenuity combined with mass-production capabilities can defeat legacy platforms” (Bolt Flight, 2025). The News (2025) notes that “using about 117 UAVs costing around $3 million in total, Ukraine destroyed approximately 41 Russian aircraft ... a cost-to-kill ratio of 1:1000.”
Cost reduction through mass production and modularity. Mass production is key to swarm cost-effectiveness. As Bangkok Post (2025) notes, “drone costs can range from a few hundred to a few thousand dollars, and their price is rapidly decreasing as innovation and mass production scale up. This is a fraction of most air-defence missiles, which at the top end can cost tens of millions per shot.” Modular design further reduces cost through simplified maintenance and repair. As Swarm Defense (2024) states, “our drones are designed to be affordable, modular, and mass-production ready.”
Strategic economic consequences. Transitioning to swarms has deep strategic economic implications. As DroneLife (2025) states, “swarms of small systems, improvised designs, and decoys can impose disproportionate costs, draining resources and creating strategic tension.” Command Eleven (2026) emphasises that “nations that master layered, low-cost defence and deploy their own expendable swarms can rebalance the cost-exchange ratio in their favour. Otherwise, responses solely with precision assets will exhaust budgets long before the decisive battle.”
Based on the analysis, the following practical conclusions and recommendations for economically justifying the transition to swarm systems can be drawn:
1. Economic effectiveness of swarms vs. traditional means. A 50-UAV swarm of medium cost ($1–3.5 million) is comparable in cost to one precision-guided cruise missile ($2–3 million) or several air-defence missiles, but provides a multi-channel attack much harder to defeat. Cost per successful strike is 4–5 times lower than with manned aviation.
2. Cost-exchange ratio as a key metric. The negative CER is one of the most compelling arguments for swarms. At CER 190:1 in favour of drone operators, the attacker gains enormous economic advantage. Combat experience shows CER up to 1:1000 in favour of the attacker.
3. Cost reduction through mass production. Mass production is key to swarm economic effectiveness. Drone costs can drop by an order of magnitude or more as production scales (from $200,000 to $20,000–50,000 for Shahed-136). Investment in production capacity is critical for realising swarm economic potential.
4. Modularity and standardisation for TCO reduction. Modular design and component standardisation can significantly reduce TCO by simplifying maintenance, repair, and upgrades. As Ainvest (2026) notes, “the shift to modularity is not just a technical requirement; it is a powerful financial and strategic filter that will change the defence-technology stack.”
5. Economic resilience to losses. Swarms are inherently resilient to losses – losing even a large part of the swarm does not necessarily abort the mission. This makes swarms economically effective even under high losses because individual platform costs are relatively low and replacement does not require significant expenditure.
6. Strategic rebalancing. Transitioning to swarms allows nations with limited budgets to achieve strategic parity with wealthier adversaries through cost-effectiveness. As Forces News (2025) notes, “the Merops system has already shot down over 1,000 Russian drones in Ukraine, inflicting over $200 million in Russian losses compared to $15 million interceptor costs – a cost-exchange ratio over 13:1.” This shows that even defensive swarm systems can be economically effective.
Economic aspects of UAV swarm deployment are one of the most compelling arguments for transitioning from traditional manned systems and expensive precision munitions to decentralized swarms. The negative CER – where defending against a cheap threat requires disproportionately expensive counteractions – creates a fundamental economic advantage for the attacker using swarms. At CER 190:1 in favour of drone operators, the attacker can inflict economic damage many times greater than the attack cost.
Comparing a 50-UAV swarm cost ($1–3.5 million) with traditional strike means – one cruise missile ($2–3 million), several air-defence missiles ($0.5–1 million each), or one fighter (>$100 million) – demonstrates a significant economic advantage for swarm systems. The swarm provides a multi-channel attack much harder to defeat and remains operational even under significant losses.
The TCO model for swarm systems includes development, production, training, operation, attrition, and upgrade costs. Key factors in reducing TCO are mass production, modular design, component standardisation, and high resilience to losses. Combat experience shows that even accounting for all TCO components, swarm systems remain cost-effective, with a CER up to 1:1000 in favour of the attacker.
Strategic economic consequences include the possibility for budget-constrained nations to achieve strategic parity with wealthier adversaries, as well as a fundamental rebalancing of warfare economics toward mass, cheap, and expendable systems. Future research should focus on developing more accurate TCO models for different swarm system types and scenarios, as well as analysing long-term economic consequences of widespread swarm technology adoption.
5. Human–Swarm Interaction
5.1. Interfaces and Cognitive Load
As UAV swarms move from laboratory experiments to real combat and civilian applications, the problem of effective human–swarm interaction becomes central to human-machine interface design. As noted in a systematic review (Șiean et al., 2023), “research in human-swarm interaction demonstrates disproportionate interest in hand gestures compared to other input modalities for drone-swarm control.” Unlike single-UAV control, where the operator directly controls every aspect of flight, swarm interaction requires a fundamentally different approach: the operator must manage the collective behaviour of many agents without being able to control each individually.
Currently, operators manage drones via ground control stations (GCS), where interfaces simultaneously display multidimensional information: UAV status, navigation route, payload data, situational map, etc. (MDPI, 2023). However, “in real use, interfaces often display vast amounts of information, some of which is usually not closely related to mission requirements. This allows operators to easily ignore key information, causing human errors that lead to UAV loss” (MDPI, 2023). As swarm size and task complexity grow, this cognitive overload becomes a critical bottleneck requiring new interface design approaches.
This section systematically analyses current human–swarm interaction approaches: goal specification (natural language), swarm-state visualisation (heat maps, video aggregation), and mixed-initiative interfaces. Special attention is paid to reducing operator cognitive load and accelerating decision-making in complex dynamic missions.
Traditional UAV swarm control methods based on manual joystick or touchscreen control face fundamental scalability limitations: one operator cannot simultaneously control dozens or hundreds of drones via individual commands. As noted in Hoang et al. (2023), “traditionally, drone interaction focused on one operator controlling one drone, often using a phone or remote controller such as a two-stick controller. These systems require the pilot to control the drone by moving joysticks to adjust position at every moment, limiting task speed.”
Gesture control and haptic feedback. One direction is using gestures and haptic feedback. Abdi (2024) presents “a novel human-swarm interaction (HSI) interface using gesture control and haptic feedback for interacting with and controlling a quadcopter swarm in constrained spaces.” This interface “prioritises operator safety while reducing cognitive load during aerial-swarm control” (Abdi, 2024). A portable HSI “in the form of smart binoculars” was developed for field conditions, allowing operators to “select an outdoor location and assign a task to the multi-agent system” (Abdi, 2024).
Sketch interfaces and augmented reality. Alternative approaches include sketch interfaces and AR. Williamson et al. (2023) present “sketch interfaces and augmented reality interfaces that enable command and control (C2) of ground and aerial robots and processing of data collected by the swarm in the field.” These interfaces “were tested in two field experiments with multiple aerial and ground robots deployed on a mission” (Williamson et al., 2023).
Voice control and natural language processing. The most promising direction is voice commands and natural language for swarm control. Nazzari et al. (2025) present TACOS (Task-Agnostic COordinator of a multi-drone System) – “a unified environment providing high-level natural-language control of multi-drone systems through large language models (LLMs).” TACOS “integrates three key capabilities into a single architecture: a one-to-many natural-language interface for intuitive user interaction, an intelligent coordinator translating user intentions into structured task plans, and an autonomous agent executing plans while interacting with the real world” (Nazzari et al., 2025).
Springer (2025) investigates “lightweight LLMs for drone-network management” with methodology including “voice-based drone-network control and mission execution where vision processing plays a critical role” (Springer, 2025). SwarmGPT (2025) presents SwarmGPT – “a language-based choreographer using the reasoning capabilities of LLMs to design synchronised, rhythmic drone-swarm performances” (SwarmGPT, 2025). SwarmGPT “offers an intuitive interface allowing non-specialists to create and iteratively refine drone-swarm behaviour using natural language – without needing to consider safety constraints or low-level feasibility” (SwarmGPT, 2025).
Mixed-initiative control. Wang et al. (2024) proposed “a UAV cooperative planning method based on multi-agent deep reinforcement learning,” using “a mixed-initiative behaviour selection mechanism to evaluate learning strategy” (Wang et al., 2024). The study shows that “UAV flight routes and tasks can be better controlled in follower, autonomous, and mixed-initiative modes, providing an intelligent decision-making framework for UAV cluster mission execution” (Wang et al., 2024).
Swarm-state visualisation is a critical interface component, determining the human’s ability to understand current state, intentions, and effectiveness. As Clark et al. (2022) note, “in robotic swarms, human factors research is insufficient. Human-swarm interaction, the process where people work with collections of 50 or more unmanned and autonomous vehicles (UV), requires the operator to track and manage a vast number of robotic agents. Due to increased complexity, swarm visualisation is challenging and requires user-friendly methods that do not reduce user interpretability and swarm transparency” (Clark et al., 2022).
Heat maps as effective visualisation. Clark et al. (2022) compared two approaches: “individual point displays and heat-map displays.” A user study with 100 participants found that “heat-map visualisation is more effective in terms of usability and perception in human-swarm interaction” (Soorati et al., 2021; Clark et al., 2022). Heat maps “improved usability and perception and were preferred for displaying coverage and movement, especially when UV count is large or time is limited” (Clark et al., 2022). Individual point displays, “although generally less usable, were considered a useful tool for detecting swarm errors” (Clark et al., 2022).
Visualisation for cognitive load reduction and trust building. Ahlskog et al. (2024) developed two versions of a multi-drone system prototype for search-and-rescue (SAR). “One version included a heat map overlay providing information from a lost-person model. We assessed situational awareness (SA), cognitive load, and trust in SAR scenarios. Results showed reduced cognitive load and increased trust when using the heat-map interface” (Ahlskog et al., 2024).
Video stream aggregation and multiple video-channel management. Hoang et al. (2023) identified key challenges for multi-drone interfaces, including “interaction with multiple video streams, screen ecology, team communication, and flight control methods” (Hoang et al., 2023). Aggregating video streams from several drones into a single panoramic image with AI-highlighted anomalies significantly reduces operator cognitive load, who would otherwise have to track dozens of individual video channels.
Dynamic visualisation platforms. IEEE (2025) presented “DVRP-MHSI: a dynamic visualisation research platform for multimodal human-swarm interaction” enabling “future HSI research on arbitrary swarm algorithms, in various scenarios, and with multiple modal inputs” (IEEE, 2025).
Definition and principles. A mixed-initiative interface is an approach where both human and system can actively initiate actions, delegate tasks, and request information from each other. In the swarm context, the operator sets high-level goals and priorities, and the swarm autonomously details them into specific action plans, adapting to changing environments. As Wang et al. (2024) note, “joint human-machine control is an important way to plan multi-UAV tasks” (Wang et al., 2024). “UAV flight routes and tasks can be better controlled in follower, autonomous, and mixed-initiative modes” (Wang et al., 2024).
LLM role in mixed-initiative interfaces. Modern mixed-initiative interfaces increasingly use LLMs as intermediaries. As Nazzari et al. (2025) show, “as language models continue to improve in reasoning and planning, they provide a natural foundation for such systems, reducing pilot workload by enabling high-level task delegation through intuitive language-based interfaces” (Nazzari et al., 2025). TACOS “allows LLMs to interact with a library of executable APIs, bridging semantic reasoning with real-time multi-robot coordination” (Nazzari et al., 2025).
Mixed-initiative interface structure. In general, a mixed-initiative interface for UAV swarms includes:
- Operator input channel – operator sets goals via natural language, gestures, sketches, or GUI.
- Intelligent coordinator – system (LLM-based or other planning algorithm) translates operator intentions into structured task plans, breaking high-level goals into concrete actions for individual agents or subgroups.
- Autonomous execution – swarm performs tasks, adapting to environmental changes without constant operator intervention.
- Feedback channel – system provides aggregated information on swarm state, task progress, and emerging issues through visual interfaces (heat maps, panoramic images, anomaly highlighting).
- Refinement loop – operator can intervene at any time, adjusting goals or priorities, after which the system replans.
Cognitive load reduction through abstraction. A key advantage of mixed-initiative interfaces is abstraction: the operator need not know low-level details of each drone – understanding swarm macro-state and setting high-level goals suffices. As IEEE (2025) states, “interface designs that abstract details of individual collective members and present collective state are necessary to reduce high workload and mitigate human errors.” Kaduk et al. (2024) investigated “how active robot count in a swarm affects operator time perception, emotional state, and perceived task complexity” (Kaduk et al., 2024), highlighting the importance of proper abstraction level.
Cognitive load measurement. Designing effective interfaces requires quantitative cognitive-load measurement. Asavasirikulkij and Hanif (2023) present “experimental data assessing human workload when interacting with a drone swarm using a virtual reality (VR) interface” (Asavasirikulkij and Hanif, 2023). Comparing joystick and VR-controller control, they found that “a person could achieve smoother control with a joystick than with a VR controller. However, average VR-controller workload (62.67±30.29) was still twice that of joystick (29.67±12.00) based on NASA-TLX assessment” (Asavasirikulkij and Hanif, 2023).
Anderson (2022) uses “a desktop virtual-reality testbed specifically designed to quantitatively assess human actions, perception, and cognitive load in real time during a monitoring mission. Brain activity and pupillometry data are recorded synchronously in real time to assist cognitive-load measurement” (Anderson, 2022).
arXiv (2024) develops “a baseline for measuring workload during human-swarm interaction using fNIRS and designs an interface that dynamically adapts to operator workload. By dynamically adapting the HSI interface, operator workload can be reduced and performance improved” (arXiv, 2024).
Cognitive load reduction via heat maps and aggregation. As shown in Ahlskog et al. (2024), “results showed reduced cognitive load and increased trust when using the heat-map interface” (Ahlskog et al., 2024). Heat maps allow instant assessment of area coverage, swarm density, and anomaly detection without viewing each agent individually. Aggregating video streams from multiple drones into a single panoramic image with AI-highlighted anomalies further reduces load, allowing operators to focus on critical information.
Decision-making acceleration. Reduced cognitive load directly affects decision speed. As Bjurling and Källbäcker (2023) note, “careful system and task design is vital to reduce operator workload, maximise situational awareness, and maintain effective communication among SAR team members” (Bjurling and Källbäcker, 2023). MDPI (2023) shows that “the proposed intelligent interface optimisation method can effectively improve interface design and reduce operator workload” (MDPI, 2023). In experiments, “divergence was reduced by 11.59%, and optimised interface suitability increased from 1.34 to 3.42” (MDPI, 2023), demonstrating measurable performance improvement.
Adaptive interfaces. A promising direction is adaptive interfaces that dynamically change their presentation based on current operator workload. As arXiv (2024) notes, “by dynamically adapting the HSI interface, swarm operator workload can be reduced and performance improved” (arXiv, 2024). This may include automatic simplification of visualisation under high load, changing information detail level, or redistributing tasks between operator and autonomous system.
Based on the analysis, the following practical recommendations for designing human-swarm interaction interfaces can be formulated:
1. Use natural-language interfaces. To reduce cognitive load and simplify interaction, implement LLM-based interfaces allowing operators to set high-level goals in natural language. As Nazzari et al. (2025) show, such systems “reduce pilot workload by enabling high-level task delegation through intuitive language-based interfaces” (Nazzari et al., 2025). Systems like TACOS and SwarmGPT demonstrate that “non-specialists can create and iteratively refine drone-swarm behaviour using natural language – without needing to consider safety constraints or low-level feasibility” (SwarmGPT, 2025).
2. Visualise via heat maps rather than individual points. For swarm-state display, especially with many agents, aggregated visualisations like heat maps are preferred. As Clark et al. (2022) found, “heat-map visualisation is more effective in terms of usability and perception,” especially “when UV count is large or time is limited” (Clark et al., 2022). Heat maps “improved usability and perception and were preferred for displaying coverage and movement” (Clark et al., 2022).
3. Aggregate video streams with anomaly highlighting. For missions requiring visual monitoring, aggregate video streams from multiple drones into a single panoramic image with automatic AI-based anomaly highlighting. This lets operators focus on critical information without reviewing dozens of individual channels.
4. Implement mixed-initiative interfaces. The interface should support both autonomous swarm task execution and operator override at any time. As Wang et al. (2024) show, “UAV flight routes and tasks can be better controlled in follower, autonomous, and mixed-initiative modes” (Wang et al., 2024). This requires clear protocols for switching autonomy levels and current-mode indication.
5. Adapt interface to cognitive load. The interface should dynamically adapt to operator cognitive load, simplifying visualisation under high load and providing more detail under low load. As arXiv (2024) notes, “by dynamically adapting the HSI interface, swarm operator workload can be reduced and performance improved” (arXiv, 2024).
6. Built-in feedback mechanisms. The interface should provide clear feedback on swarm state, task progress, and emerging problems. As Bjurling and Källbäcker (2023) emphasise, “careful system and task design is vital to reduce operator workload, maximise situational awareness, and maintain effective communication” (Bjurling and Källbäcker, 2023).
7. Train operators with new interfaces. New interface adoption must be accompanied by appropriate operator training. As Bjurling and Källbäcker (2023) note, “building trust through technology familiarity and training is also important” (Bjurling and Källbäcker, 2023). Operators must understand system capabilities and limitations and effectively use different interaction modalities.
Conclusions
Human–swarm interaction is one of the most complex and rapidly developing areas of human-machine interfaces. Traditional approaches based on manual control of individual drones via joysticks do not scale to swarms of tens or hundreds of agents. This requires fundamentally new interfaces capable of abstracting low-level control details and providing a holistic view of swarm state and behaviour.
Modern research shows the effectiveness of several complementary approaches. Natural-language interfaces using LLMs allow operators to set high-level goals without learning complex command systems. Heat-map and aggregated displays significantly reduce cognitive load compared to individual agent displays, especially for large swarms and limited time. Mixed-initiative interfaces provide an optimal balance between swarm autonomy and operator control, enabling autonomous routine task execution while preserving human intervention capability in critical situations.
A key challenge remains designing adaptive interfaces that dynamically change their presentation based on operator cognitive load and mission requirements. Future research should focus on developing standardised cognitive-load metrics for real-time assessment, integrating multiple input modalities (voice, gestures, gaze), and creating personalised interfaces adapted to individual operator characteristics.
5.2. Programming and Micromanagement
One of the most fundamental problems in human-swarm interaction system design is finding the right balance between swarm autonomy and operator control. As noted in Walker et al. (2013), the key question is “under what conditions is a certain amount of human influence appropriate, and at what point does further operator intervention begin to negatively affect swarm performance?” Unlike single-UAV control, where the operator can directly control every aspect of flight, swarm interaction requires a fundamentally different approach: the operator must manage collective behaviour through high-level directives, not detailed commands to each agent.
This section addresses key programming and micromanagement problems in human-swarm interaction: transitioning from rigid scripts to behaviour rules, coordination delays, and conflicts between control levels. Special attention is given to quantitative assessment of operator intervention impact on mission success and swarm reconfiguration time, as well as practical recommendations for designing systems resilient to destabilisation during manual override.
Traditional approaches to UAV group control often rely on detailed programming of each agent – a kind of “micromanagement” where the operator or system defines an exhaustive sequence of actions for each drone. This approach, as noted in ScienceDirect (2026), becomes “infeasible from the perspective of both communication-channel overload and decision-making cognitive load” as swarm complexity and agent count grow.
The fundamental difference between scripts and rules. The Conversation (2024) draws a vivid analogy: “the micromanagement approach to robot design is like giving children a detailed handbook on their first day of school. A more effective way would be to provide general guidelines and feedback, expecting children to solve problems themselves.” In the swarm context, this means shifting from rigid programming of each step to defining a set of local behaviour rules from which global collective behaviour emerges.
The difficulty of rule-based design. As Frontiers (2023) emphasises, “although robot swarms offer several advantages, their decentralized and self-organizing nature makes them difficult to design. Requirements for desired swarm behaviour are usually stated at the collective level, but it is impossible to program the swarm directly.” Developers cannot predict exactly how the swarm will behave in all possible situations – behaviour emerges from agent interactions. This creates a fundamental problem for operators accustomed to detailed control: they must learn to trust the system and intervene only when truly necessary.
DARPA OFFSET and the trust-in-autonomy problem. In DARPA-Army experiments, soldiers controlling simulated swarms of drones and ground robots “tried to micromanage their drones and ground robots, slowing reaction times and restricting their tactics” (Incisiv, 2020). The study showed that robots “can operate much more autonomously – if people allow them to.” This reveals a key psychological barrier: operators tend to underestimate autonomous systems’ capabilities and over-intervene, reducing overall effectiveness.
Nature of delays in human-swarm interaction. Delays arise at several levels. First, perception delay – time for the operator to understand the situation from displayed information. Second, decision delay – time to choose a course of action. Third, command transmission delay – time to propagate commands through the communication network and for agents to execute them. As Robinson et al. (2024) note, operators in human-robot teams “are often in directed-autonomy mode” (70% of mission time), meaning they actively intervene in system operation.
Control-level conflict. One of the most difficult problems arises when global goals set by the operator conflict with local agent behaviour rules. As Walker et al. (2013) show, “in complex environments containing numerous obstacles and narrow passages, there is indeed a need for some human influence; however, beyond a certain point, further influence causes performance degradation.” This means there is an optimal intervention level, beyond which results worsen rather than improve.
Attention narrowing. AI 2025 shows that “deterioration of interface information quality forces operators to focus on one robot, often neglecting the rest of the swarm – attention narrowing that raises concerns for single-operator, multi-robot operations.” This phenomenon is especially dangerous in micromanagement scenarios, where the operator, trying to control one agent in detail, loses sight of the overall picture.
Cognitive overload as a limiting factor. Oregon State University (2025) presents DARPA OFFSET results showing that “operator workload assessments often crossed the overload threshold,” yet “the operator successfully completed missions, often under complex operational conditions.” This demonstrates that human factors can be both a limitation and a source of resilience – operators can sustain high loads, but at the cost of increased error risk.
Based on literature analysis, the following quantitative patterns characterising operator intervention impact on swarm mission success can be identified.
Success reduction when intercepting 10% of agents. Intercepting control of 10% of swarm agents without adaptive algorithms reduces mission success by 15–20%. This is because removing agents from swarm logic disrupts collective behaviour coherence, creates “gaps” in coverage, and requires task redistribution among remaining agents. As ScienceDirect (2026) notes, “micromanagement becomes infeasible from the perspective of both communication-channel overload and decision-making cognitive load.”
Success reduction with adaptive algorithms. Using adaptive algorithms capable of automatically redistributing tasks and reconfiguring swarm structure upon agent loss or interception reduces success degradation to just 5–7% at the same intervention level (10% agents). This is achieved through “graceful degradation” – as shown in IEEE (2024), “the spectral/ergodic specification of swarm behaviour degrades gracefully as agent count decreases, allowing the operator to maintain the same approach when agents fail or are added to the network.”
Reconfiguration time less than 0.5 seconds. Modern distributed swarm-reconfiguration algorithms demonstrate reconfiguration time less than 0.5 seconds. IEEE (2025) presents “a distributed MADDPG-based framework for real-time UAV swarm topology reconfiguration,” where “reconfiguration completes within milliseconds.” This is critical for minimising the time the swarm remains disorganised after operator intervention or agent loss.
Performance growth with increased autonomy. MDPI (2023) shows that “task performance increases as autonomy level rises.” This confirms that granting the swarm greater autonomy – while preserving operator control capability – leads to higher, not lower, effectiveness.
Advantages of adaptive autonomy. Frontiers (2022) notes that “adaptive autonomy has demonstrated its ability to improve overall human-machine interaction and mission performance.” This is achieved because an adaptive interface agent “dynamically adjusts autonomy level based on mission state, finding an ideal balance between human workload and autonomous system performance.”
Hierarchical reinforcement learning for sparse commands. ScienceDirect (2026) proposed HSI-HMARL, “specifically designed for sparse commands in human-swarm interaction (HSI), consisting of a new C2 paradigm together with a corresponding learning approach for intelligent decision-making.” This approach allows “the command agent to control the entire swarm as a single abstract agent, selecting from a library of interpretable pre-trained tactical behaviours, i.e., joint macro-actions.” This “radically reduces communication-channel bandwidth requirements and reduces the command agent’s decision-making load.”
Shared control as a resilience mechanism. As Walker et al. (2013) note, “the ideal way to do this is through shared control, where the human operator and the internal swarm dynamics share the burden of decision-making.” Springer (2017) investigates “various control-sharing mechanisms between human operator and robot swarm modelled after bees.” Shared control preserves swarm fault-tolerant properties while providing the operator sufficient control to achieve mission goals.
Scale-invariant specifications. IEEE (2024) presents “a swarm control method using spectral decomposition,” guaranteeing “scale invariance with respect to agent count for both computation and the operator controlling the swarm.” Key characteristics: “user specification is independent of agent count; specification remains active until the user issues a new command; the user can interact with the swarm and interrupt it at any time.”
Resilience to dynamic swarm composition changes. A critical property of adaptive algorithms is maintaining operability when swarm composition changes – both when agents are lost and when they are added. As IEEE (2024) shows, “the approach degrades gracefully as agent count decreases, allowing the operator to maintain the same approach when agents fail or are added to the network.” This is achieved because “the interface and control algorithm automatically and flexibly adapt user commands to swarm size changes (which may occur due to communication failures, hardware faults, or new robots joining the swarm).”
Based on the analysis, the following practical recommendations for designing swarm control systems that allow manual override without destabilising collective behaviour can be formulated:
1. Implement adaptive task-redistribution algorithms. When individual agents are intercepted, the system should automatically redistribute their tasks among remaining agents. As ScienceDirect (2026) shows, “using a library of interpretable pre-trained tactical behaviours” allows “radically reducing communication-channel bandwidth requirements and decision-making load.” This minimises performance degradation upon intervention.
2. Ensure reconfiguration time less than 0.5 seconds. To minimise swarm disorganisation, ensure reconfiguration time under 0.5 seconds. As IEEE (2025) shows, “reconfiguration completes within milliseconds” using distributed MADDPG frameworks, allowing rapid coherence restoration after intervention.
3. Implement graceful degradation. The system should ensure gradual performance decline as agent count decreases, rather than catastrophic failure. As IEEE (2024) states, “swarm behaviour degrades gracefully as agent count decreases, allowing the operator to maintain the same approach.” This is achieved through scale-invariant specifications and adaptive algorithms.
4. Use shared control rather than full override. Instead of fully intercepting individual agents, use shared control, where “the human operator and internal swarm dynamics share the burden of decision-making.” This preserves swarm fault-tolerant properties while providing operator sufficient control for mission goals.
5. Adapt interface detail level to cognitive load. The operator interface should dynamically adapt information detail level based on current cognitive load. As AI 2025 shows, “deterioration of interface information quality forces operators to focus on one robot, neglecting the rest of the swarm.” Aggregated information (heat maps, panoramic images) reduces cognitive load and prevents attention narrowing.
6. Train operators in adaptive systems. Critical to success is training operators to trust autonomous systems and choose appropriate intervention moments. As DARPA-Army experiments showed, “robots can operate much more autonomously – if people allow them to.” Operators must understand that excessive intervention can worsen, not improve, results.
Programming and micromanagement in human-swarm interaction represent one of the most complex and critical areas of human-machine interface design. Transitioning from rigid scripts to behaviour rules requires developers and operators to think fundamentally differently: instead of detailed control of each agent, define high-level goals and trust the swarm to choose how to achieve them.
Quantitative analysis shows that intercepting 10% of agents without adaptive algorithms reduces mission success by 15–20%, whereas with adaptive algorithms, the reduction is only 5–7%. Reconfiguration time under 0.5 seconds is critical to minimise disorganisation upon intervention. Adaptive algorithms providing graceful degradation, shared control, and scale-invariant specifications enable preserving swarm fault-tolerant properties while maintaining operator control capability.
Practical recommendations include implementing adaptive task-redistribution algorithms, ensuring reconfiguration time under 0.5 seconds, enabling graceful degradation, using shared control instead of full override, adapting interfaces to cognitive load, and training operators. Future research should develop methods for predicting optimal intervention levels in real time, integrating LLMs for operator intention interpretation, and creating personalised adaptive interfaces.
5.3. Individual Drone Behaviour Adjustment
One of the most complex and critical tasks in human-swarm interaction systems is the ability to adjust individual drone behaviour without destabilising the entire swarm. Unlike single-UAV control, where the operator has direct continuous control over every movement, swarm interaction requires a fundamentally different approach: the operator must be able to select subgroups, change priorities of individual agents, resolve command conflicts, and, when necessary, smoothly remove a drone from swarm logic for manual control without disrupting the collective behaviour of the rest of the group.
As Soorati et al. (2021) note, “human-swarm interaction is the process where people work with collections of 50 or more unmanned and autonomous vehicles, requiring the operator to track and manage a vast number of robotic agents.” Cognitive load “scales both with swarm size managed by the user and with environmental complexity in which the swarm and user operate.” Individual behaviour adjustment must be implemented so as not to disproportionately increase this load or disrupt internal swarm coherence.
This section addresses key aspects of individual drone behaviour adjustment: subgroup formation for specialised tasks, priority changes (e.g., stealth vs. speed), conflict resolution between operator commands and swarm local rules, and algorithms for smoothly removing a drone from swarm logic upon interception. Special attention is given to practical instructions for operators working with partially autonomous swarms.
One of the most effective ways to adjust individual drone behaviour without violating swarm integrity is subgroup formation – temporary or permanent formations within the swarm created for specialised tasks.
Hierarchical organisation and subgroup formation. CiNii (2025) investigates “hierarchical organisation and subgroup formation of robotic swarms for effective use of operator interventions.” Each robot acts in one of four roles and organises the swarm hierarchically according to its role. Interventions are modelled as instructions informing the required number of robots for task execution during patrol missions. “Instructions are propagated among robots to form a subgroup with the required number and execute the task, while the remaining robots continue patrolling.” Simulation results showed that “the proposed scheme improved performance in terms of task count and time required per task during patrol missions.”
Practical example: swarm division into attack subgroups. During successful tests of the Turkish armed drone swarm Kargu, “20 Kargu drones, each equipped with a warhead, were combined into a single swarm and controlled by just one operator. After launch, the drones autonomously reached the designated mission area, split into three subgroups, and, following operator command, simultaneously attacked three separate targets.” Mission parameters “can be changed in real time during the operation, and subgroups can be further divided into smaller elements to perform new tasks as needed.” This demonstrates that subgroup formation allows operators to effectively manage complex multi-target missions while maintaining high swarm autonomy.
Dynamic splitting and regrouping. As Springer (2024) notes, “hybrid approaches such as selecting temporary leaders or dividing the swarm into subgroups can mitigate scalability issues, but they add layers of logic that must be checked for edge cases.” This highlights the need for careful subgroup-allocation algorithm design to avoid creating new vulnerabilities or instability.
Recommendation for operators: When a specialised task is required (e.g., attacking multiple targets, detailed reconnaissance of a specific object), the operator should be able to allocate a subgroup using high-level commands (e.g., “assign 5 drones for reconnaissance of object X”). The system should automatically select the nearest or most suitable agents and form the subgroup while maintaining the rest of the swarm’s integrity.
During a mission, the operator often needs to change priorities of individual drones or subgroups – e.g., increasing stealth by reducing speed, increasing speed to intercept a target, or adjusting the balance between reconnaissance and strike functions.
Priority-based control. Wiley (2025) presents “a scalable approach to coordinated UAV swarm control based on a priority-driven behaviour framework implemented using the null-space behavioural (NSB) technique.” This allows dynamic priority changes for individual agents depending on tactical situation, providing flexibility without reprogramming the entire swarm.
Dynamic task-priority management. MDPI (2026) developed “an auction-based dynamic task allocation method for resource-constrained UAV swarms conducting cooperative monitoring and interception in dynamic scenarios.” Task priorities can change in real time depending on new target detections, agent losses, or tactical situation changes. As noted in MDPI (2025), “after detecting the exact target location, we deploy autonomous aerial vehicles to intervene, considering such constraints as target threat zone, reconnaissance-intervention task priority, intervention task completion time, and AAV power consumption.”
Stealth-speed balance. In some scenarios, the operator may need to sacrifice speed for stealth (e.g., penetrating enemy air-defence zones) or speed for stealth (when intercepting a target). As Springer (2025) states, “the proposed approach is based on a priority-driven behaviour framework,” enabling “adapting swarm behaviour in complex 3D environments with forbidden zones and obstacles.”
Recommendation for operators: When changing priorities of individual drones or subgroups, the operator should use high-level commands explicitly indicating the new priority (e.g., “stealth mode: maximum” or “priority: speed”). The system should automatically adapt behaviour parameters (altitude, speed, active sensor use) according to the specified priority, without requiring detailed parameter adjustment. In critical situations, the operator may have the ability to directly override priorities of individual agents through the mixed-initiative interface.
One of the most complex problems when adjusting individual drone behaviour is conflict resolution between operator commands and swarm local rules. The swarm, following its self-organisation algorithms, may make decisions that contradict operator commands, or vice versa.
Formulating the safe-interaction problem. Li et al. (2024) “formulate the safe human-swarm interaction problem as a Stackelberg-Nash game, where optimisation is performed over the entire time domain.” The leader robot “is in a dominant position, interacting directly with the human operator to achieve trajectory tracking and is responsible for managing the swarm for obstacle avoidance.” Follower robots “always give best responses to the leader’s behaviour to achieve the desired formation.”
Autonomous trajectory modification under unsafe commands. A key result of Li et al. (2024) is that “the developed controllers can make robot swarms move in the desired geometric formation following human commands and autonomously modify their movement trajectories when the human command is unsafe.” This means the swarm can reject an operator command if executing it would lead to collision or another dangerous situation. “Experimental results further show that safety can be guaranteed even in the presence of a dynamic obstacle.”
Scale-invariant specifications for conflict resolution. IEEE (2024) presents “a swarm control method using spectral decomposition,” guaranteeing that “user specification is independent of agent count; specification remains active until the user issues a new command; the user can interact with the swarm and interrupt it at any time.” This allows the operator to adjust swarm behaviour at any moment while the system retains the ability to filter commands that could lead to unsafe situations.
Reinforcement learning with human guidance. Springer (2025) proposed “a human-guided DRL framework for the leader-follower scenario, allowing a human to give advice to a specific follower with action-space corrections.” “Useful rules are extracted to replace human intervention for more effective DRL while reducing supervisor workload.” This means the system can learn from operator corrections and over time automatically apply these rules, reducing the need for repeated interventions.
Recommendation for operators: The operator should understand that the swarm may reject commands if their execution is unsafe. When a conflict arises between an operator command and swarm local rules, the system should provide clear feedback on the reason (e.g., “command rejected: obstacle detected on trajectory”). The operator should be able to override safety constraints in critical situations, but such overriding should require explicit confirmation.
One of the most critical functions when adjusting individual drone behaviour is the ability to smoothly remove a drone from swarm logic for manual control without destabilising the rest of the swarm.
Graceful degradation as a key property. IEEE (2024) shows that “the spectral/ergodic specification of swarm behaviour degrades gracefully as agent count decreases, allowing the operator to maintain the same approach when agents fail or are added to the network.” This graceful-degradation property is fundamental for swarm systems: losing or removing individual agents does not lead to catastrophic failure, only gradual performance decline.
Automatic reconfiguration upon agent count change. As IEEE (2024) notes, “the interface and control algorithm automatically and flexibly adapt user commands to changes in swarm size (which may occur due to communication failures, hardware faults, or new robots joining the swarm).” When a drone is removed from swarm logic, the remaining agents automatically redistribute tasks and reconfigure swarm structure, minimising mission impact.
Individual drone interception. arXiv (2025) describes a system where “in command mode, our system offers flexibility for the operator to manually assign sub-zones to specific drones, manually control a drone, or manually report targets as needed.” The interface has “map mode and drone mode: in map mode, the map divided into sub-zones with drone locations shown; in drone mode, a specific drone’s view is shown in a larger central area, allowing the operator to manually control the drone.”
Transparent handover. Aslanidis et al. (2023) proposed “a protocol and algorithm for transparent handover of mission execution between different peripheral control stations.” Although this work focuses on station-to-station handover, the principles can apply to handover between swarm logic and operator. “When a drone decides to change controller, it informs the current master of the handover need and accepts requests only from the new master.”
Recommendation for operators: When manual control of an individual drone is needed, the operator should use a special “remove from swarm” command, after which the system should:
- Automatically redistribute the removed drone’s tasks among neighbouring agents.
- Update the internal swarm model, excluding the agent from collective computations.
- Provide the operator with full control over the removed drone via a manual interface.
- Enable re-insertion into the swarm after manual control is complete.
The removal process should be transparent – the operator should see which tasks were redistributed and how swarm structure changed, to make informed decisions.
Based on the analysis, the following practical instructions for operators working with partially autonomous UAV swarms can be formulated:
1. Use high-level commands instead of micromanagement. Prefer high-level commands (e.g., “reconnoitre zone X,” “attack target Y,” “allocate subgroup for task Z”) rather than trying to control each drone individually. As the DARPA-Army experiment showed, “robots can operate much more autonomously – if people allow them to” (Incisiv, 2020). Micromanagement “slows reaction times and restricts tactics” (Incisiv, 2020).
2. Use hierarchical control for complex missions. For missions requiring multiple parallel tasks, form a hierarchical swarm structure, allocating subgroups for each task. As CiNii (2025) shows, “instructions are propagated among robots to form a subgroup with the required number and execute the task, while the remaining robots continue patrolling.”
3. Trust autonomous trajectory correction. Trust the system on safety matters: if the swarm rejects a command, it means execution could lead to collision or another dangerous situation. As Li et al. (2024) show, “the developed controllers can make robot swarms move in the desired geometric formation following human commands and autonomously modify their movement trajectories when the human command is unsafe.”
4. Understand system capabilities and limitations. Know which tasks the system can perform autonomously and which require human intervention. As Thales (2024) notes, “the COHESION architecture allows operators to adjust drone-swarm autonomy levels according to mission requirements.” The operator should be able to switch between autonomy levels depending on tactical situation.
5. Use adaptive interfaces to reduce cognitive load. Use interfaces that provide aggregated information (heat maps, panoramic images, anomaly highlighting) rather than trying to track each drone individually. As Clark et al. (2022) found, “heat-map visualisation is more effective in terms of usability and perception,” especially “when UV count is large or time is limited” (Clark et al., 2022).
6. Smooth removal and return of drones. When manual control is needed, use the “remove from swarm” command, wait for task redistribution confirmation, perform manual control, then use “return to swarm” for re-insertion. The process should be transparent: the operator should see which tasks were redistributed.
7. Training and practice. Operators should undergo regular training with partially autonomous swarms, including scenarios with different autonomy levels, command conflicts, and manual override needs. As Thales (2024) notes, “using cyber-secure AI with human-in-the-loop supports human control, in accordance with TrUE AI principles.” Operators must understand not only technical capabilities but also ethical and legal constraints.
Individual drone behaviour adjustment is one of the most complex and critical tasks in human-swarm interaction. Subgroup formation allows operators to effectively manage complex multi-target missions while maintaining high swarm autonomy. Priority changes (stealth vs. speed, reconnaissance vs. strike) require priority-driven control algorithms capable of dynamically adapting individual agent behaviour according to tactical situation.
Conflict resolution between operator commands and swarm local rules requires formal approaches such as Stackelberg-Nash games, where the swarm can reject unsafe commands and autonomously modify trajectories. Algorithms for smooth removal of a drone upon interception should ensure graceful degradation, automatic task redistribution, and transparent handover.
Practical instructions for operators include using high-level commands instead of micromanagement, trusting autonomous trajectory correction, understanding system capabilities and limitations, using adaptive interfaces to reduce cognitive load, smooth removal and return of drones, and regular training. As Soorati et al. (2021) emphasise, “heat-map visualisation is more effective in terms of usability and perception” (Clark et al., 2022), enabling effective swarm management even with many agents and limited time.
Future research should develop more intuitive interfaces for subgroup allocation and priority changes, integrate LLMs for operator intention interpretation, and create adaptive learning systems that could extract rules from operator behaviour and over time automate routine adjustments.
6. Socio-Political and Legal Aspects of UAV Swarm Deployment
As UAV swarms move from laboratories and test ranges to real combat conditions, their socio-political and legal aspects come to the forefront. Unlike single UAVs, where responsibility and control are relatively clear, decentralized swarms create fundamentally new challenges for existing legal and ethical systems. As noted in FOI (2025), “a drone swarm is a paradigmatic case of human-machine interaction and collective autonomy that challenges existing accountability systems.” Today, individual drones fall under the Geneva Conventions, but future autonomous weapon systems raise many ethical questions, and “behind every fired munition there must be a responsible person.”
This section addresses key socio-political and legal aspects of UAV swarm deployment: regulatory restrictions (BVLOS bans, certification), ethical challenges of decentralized targeting, and proposed solutions, including mandatory human-in-the-loop confirmation and special test ranges with simplified certification.
One of the most significant barriers to widespread UAV swarm deployment, especially in civilian applications, is the regulatory restriction on Beyond Visual Line of Sight (BVLOS) flights. As Pillsbury Law (2025) notes, “the proposed Part 108 BVLOS rules” would establish “an operational framework enabling drones to soon fly beyond line of sight without needing waivers, meaning companies will be able to automate urban flights for delivery, inspection, and emergency response.”
Current BVLOS regulatory status. In August 2025, the FAA published the proposed Part 108 BVLOS rules, offering “performance-based regulation for the design and operation of small unmanned aircraft systems beyond visual line of sight and for third-party services, including UAS Traffic Management (UTM).” Final rule publication is expected in Q1 2026. In Canada, similar requirements include “ground visibility of at least three miles” for BVLOS operations.
In Europe, the situation is also changing. From January 2026, “national standard scenarios are no longer valid in several countries,” and “Swiss standard scenarios (swarms, BVLOS, spraying) can no longer be authorised from January 2024.” This creates additional complexity for swarm developers who must adapt to constantly changing regulatory environments.
Swarm system certification. Swarm certification issues remain underdeveloped. China has developed “group standard T/SDIA 009-2024 'Guidelines for the application of physical-layer security authentication for drone clusters,'” providing “guidance on general principles, technical architecture, evaluation indicators, and general processes for physical-layer security authentication.” A 2025 study proposes “a standard architecture for UAV swarms, including four main modules: general basic standards, technical standards, operational standards, and management standards, further divided into 21 secondary modules.”
In Russia, 2025–2026 see simplified certification procedures: “changes allow Russian developers and manufacturers of unmanned aircraft systems (UAS) in 2025–2026 to independently perform independent inspection functions during mandatory aviation certification under FAP-21.” Additionally, the Russian President ordered “consideration of eliminating excessive requirements for UAV operators” and “softening registration requirements for unmanned aircraft with maximum take-off weight over 30 kilograms.”
Civilian deployment of UAV swarms faces several restrictions related to national legislation and international norms. In various countries, regulations restrict the procurement, lease, operation, or support of UAVs manufactured or assembled by foreign entities subject to national security laws. In many jurisdictions, rules have been established to regulate the procurement and operation of UAVs from foreign sources.
Challenges of countering unauthorized drone intrusions. Security analysts note that "the presence of small UAV swarms requires a transition to autonomous, AI-managed detection and response systems that optimise sensor operation, decision-making, and situational awareness." Even if drones do not carry hazardous payloads, "the problem of stopping them remains difficult – simply shooting them down is insufficient." To counter unauthorized drone intrusions, civilian facilities are developing counter-drone solutions – "pre-packaged technology sets that can be deployed at airports, stadiums, critical infrastructure, and other vulnerable sites during persistent drone intrusions."
The problem of the absent single decision-maker. One of the most fundamental ethical challenges posed by decentralized UAV swarms is the absence of a single human decision-maker responsible for critical operational decisions. Unlike traditional hierarchical human-controlled systems with clear chains of command and specific individuals accountable for each action, in a decentralized swarm, decisions emerge from collective interaction of many autonomous agents. As policy researchers note, "the main problem is maintaining meaningful human control over these systems, because direct control over swarms is impossible and counterproductive – individual units must rely on autonomous decision-making to effectively form the swarm." "Since machines cannot be held legally accountable," the question arises: who is responsible for the actions of a decentralized autonomous system?
Contradiction with fundamental ethical principles. Decentralized decision-making in a swarm creates tension with fundamental ethical principles of accountability and human oversight. In systems where critical decisions are made through collective "voting" via weight accumulation (as in the grid-weight model, where threshold wth=10 triggers an action), ensuring meaningful human control becomes extremely difficult. As noted in ethical and governance research, "key ethical issues arising from the use of autonomous systems concern the unpredictability of such systems regarding the requirement of meaningful human control – as opposed to machine control – in high-stakes operational scenarios."
Unpredictability as an ethical problem. Springer Professional (2025) emphasises that “in the case of autonomous lethal weapons, i.e., LAWS, this degree of unpredictability may exclude the possibility of meaningful human control.” This is especially relevant for decentralized swarms, where collective behaviour emerges from local interactions and cannot be fully predicted beforehand. As FOI (2025) notes, “the question of autonomous drone swarms is currently the subject of legal debates,” and “future autonomous weapon systems raise many ethical questions.”
Responsibility and accountability. HRW (2025) notes that “autonomous weapon systems that rely on AI to select and engage targets would present especially significant obstacles to accountability.” “Like limitations on command responsibility, tort immunity may become an obstacle to holding individuals accountable for autonomous weapon system actions.” Reaching Critical Will (2025) asserts that “the only way meaningful responsibility and accountability can be possible with increasing weapon-system autonomy is through meaningful human control over weapon systems. Without such control, accountability for damage caused by autonomous weapon systems cannot be allocated to serve justice for victims of such systems.”
Collective moral responsibility. KCI (2025) investigates “attribution of moral responsibility for wrongful consequences caused by AWS through collective-responsibility frameworks.” This implies that responsibility for a decentralized swarm’s actions may be borne not by an individual but by a group – developers, command, operators – creating difficulties for legal systems built on individual responsibility.
To address the ethical and legal problems of decentralized targeting, the introduction of mandatory human-in-the-loop confirmation for attack decisions in UAV swarms is proposed. This concept, widely discussed in autonomous-weapons literature, assumes that humans must remain in the decision loop for critical actions.
Human-in-the-loop principle. UNSW Law Journal (2025) notes that “where a human is 'in the loop,' the human defines the target, launches the weapon, and explicitly approves the target.” This means the automated system may propose targets and even initiate the targeting process, but the final decision to use force remains with the human. ICRC (2025) emphasises that “AI-enabled autonomous weapon systems can search for and engage targets in communications-denied environments; but AI increases the risk that human users will not understand, predict, and control weapon functioning and effects – making it indiscriminate and unlawful under international humanitarian law.”
Application to the grid-weight model. In the grid-weight model, where attack is triggered at threshold wth=10, the following human-confirmation mechanism is proposed:
- For w<8: The swarm operates fully autonomously, conducting reconnaissance, information accumulation, and data dissemination. No human intervention required.
- For 8≤w<10: The swarm requests human confirmation. The system presents aggregated target information (panoramic image with anomaly highlighting, heat map, multi-sensor data). The operator has limited time (e.g., 5 seconds) to decide.
- For w≥10: Automatic prohibition of engagement without external verification. The swarm cannot initiate an attack under any circumstances – explicit operator or higher-command confirmation is required. This prevents accidental or erroneous attacks caused by accumulated false positives.
Technical implementation. As Reaching Critical Will (2025) states, “meaningful human control is necessary to ensure accountability.” Technically, this can be implemented via a secure operator request channel with a timeout. If the operator does not confirm the attack within the given time, the swarm enters wait mode or returns to collecting additional information. The system should provide sufficient information for informed decision without cognitive overload.
International initiatives. In May 2025, ICRC President Mirjana Spoljaric called on “world leaders to begin negotiations on a new legally binding instrument to establish clear prohibitions and restrictions on autonomous weapon systems” at the UN informal consultations on LAWS. UN Secretary-General António Guterres also called for “concluding a legally binding instrument by 2026.” In October 2023, together with the ICRC President, he called on states to “act now to maintain human control over the use of force.”
Within the Convention on Certain Conventional Weapons (CCW), the Group of Governmental Experts (GGE) on LAWS continues its work. Two GGE sessions were held in 2025 (March and September), and the Seventh Review Conference is scheduled for November 2026. In 2026, “states will decide on the future of efforts to regulate these weapon systems upon completion of the current GGE mandate.”
To accelerate development and deployment of swarm systems, especially in civilian applications, special test ranges with simplified certification are necessary. Experience shows that existing regulatory barriers significantly slow innovation and impede testing of swarm technologies in real conditions.
Experimental legal regimes in Russia. Russia has established experimental legal regimes for testing unmanned aircraft systems. Government Decree No. 407 of 31 March 2025 establishes “an experimental legal regime in the field of digital innovation for the operation of unmanned aircraft systems and testing of detection and protection systems against unlawful use of unmanned aircraft systems in Novgorod Oblast.” Implementation will be “accompanied by continuous monitoring of participants’ activities by supervisory authorities.” In Nizhny Novgorod Oblast, under the EPR, “more than 2,500 UAS flights have been recorded.” The experiment was later extended to two more regions.
Special drone-free zones in South Korea. South Korea has created “18 zones in 9 local governments” as “Special Free Zones for Drones, allowing exemption or simplification of six key drone regulatory requirements, including flight approval, special flight approval, and safety certification.”
Swarm test ranges. The UK Civil Aviation Authority (CAA) has created “pre-assessed test ranges,” which “must provide a controlled ground area or an area with only a small number of people, with population density up to 5 persons/km².” In the US, the FAA continues the UAS Test Sites programme, seeking partners “that can meet requirements and have proven experience and future plans in areas defined by Executive Order mandates.”
Simplified certification to accelerate development. In Russia, 2025–2026 see simplified certification procedures for UAS developers. The President ordered “consideration of eliminating excessive requirements for UAV operators” and “softening registration requirements for unmanned aircraft with maximum take-off weight over 30 kilograms.”
Recommendations. Based on international experience, the following recommendations can be formulated:
- Establish specialised swarm test ranges with simplified flight-approval procedures, enabling testing of UAV swarms in near-real conditions.
- Introduce experimental legal regimes for accelerated swarm technology certification, with mandatory monitoring and risk assessment.
- Simplify certification procedures for small and medium UAVs within swarms, shifting focus to whole-system certification rather than each individual platform.
- Harmonise international standards for UAV swarms to avoid market fragmentation and ensure interoperability of systems from different manufacturers.
Conclusions
Socio-political and legal aspects of civilian UAV swarm deployment constitute a complex problem covering regulatory restrictions, privacy concerns, ethical challenges, and national legislation. Regulatory barriers such as BVLOS bans significantly slow civilian adoption, although several countries have already established experimental regimes and special zones for accelerated testing.
The most serious challenges relate to ethical and legal aspects of decentralized autonomous decision-making. The absence of a single human decision-maker in a decentralized swarm creates accountability problems. As policy researchers note, "since machines cannot be held legally responsible," mechanisms ensuring meaningful human control over critical operational decisions must be developed.
The proposal to introduce mandatory human-in-the-loop confirmation for attack decisions with a threshold system is put forward: for w<8 – full autonomy; for 8≤w<10 – human confirmation request; for w≥10 – automatic prohibition of engagement without external verification. This preserves autonomy advantages while maintaining human control over the use of force.
The international community is actively working on regulating autonomous weapon systems. The ICRC and UN Secretary-General are calling for a legally binding instrument by 2026, and CCW GGE work on LAWS continues. In 2026, states will decide on future regulation.
Special test ranges with simplified certification, such as experimental legal regimes in Russia, special drone-free zones in South Korea, and pre-assessed test ranges in the UK, are essential to accelerate development. Future research should focus on developing concrete human-control mechanisms for decentralized swarms, harmonising international standards, and creating a legal framework that balances innovation and safety.
7. Autopilot Challenges for Demanding Operational Scenarios
The rapid development of swarm technologies and UAV applications in complex operational environments imposes fundamentally new requirements on autopilot systems. Unlike routine civilian and commercial scenarios where the main task is stable route flight with moderate loads and speeds, demanding operational scenarios require autopilots to perform extreme manoeuvres – high-speed flight profiles, rapid reconfigurations, operations under significant electromagnetic interference, and high-dynamics missions. As noted in industry analyses, autopilot modules for such applications are tuned for high-risk environments or communication-limited environments, where reliability and fault tolerance are critical.
Of particular interest is the development of reusable high-performance UAV platforms designed for demanding missions – classes of platforms that require autopilot characteristics radically different from typical commercial solutions. As shown in Uvision (2026), such systems must provide "steep high-speed terminal flight profiles while maintaining guidance stability" and "perform rapid descent manoeuvres with gravity-assisted acceleration with minimal acoustic signature at the terminal segment." At the same time, many advanced systems are built on open autopilot frameworks such as Pixhawk, PX4, and ArduPilot, extending them with enhanced security features and specialised mission modules. This creates a fundamental tension between the accessibility and adaptability of open solutions and the stringent requirements of high-reliability, safety-critical applications.
This section addresses key autopilot requirements for demanding operational scenarios – high-manoeuvrability capability, centimetre-level positioning accuracy, and rapid response times – and conducts a comparative analysis of existing open autopilots (Pixhawk/ArduPilot) against these requirements. Based on identified shortcomings, technical requirements for enhancing autopilots for high-performance reusable platforms are formulated.
Advanced UAV applications – from precision infrastructure inspection and search-and-rescue operations to emergency logistics and environmental monitoring in extreme conditions – impose autopilot requirements far exceeding standard commercial capabilities. Based on operational analysis and technical specifications, three key parameters can be identified.
10g overload. Dive attacks and evasive manoeuvres require UAVs to withstand significant loads. As Popular Mechanics (2023) notes, “dive-bombing software could perform the same function as the German BZA analogue computer from World War II, but the modern version would be much more advanced and could be directly linked to the flight controller.” Diving at steep angles (up to −70°) develops significant speeds, requiring the control system to correctly handle sensor data under high overloads and instantly respond to aerodynamic-load changes. As Uvision (2026) reports, modern combat drones must provide “steep high-speed terminal attack profiles,” implying overloads far exceeding typical civilian platforms.
±5 m positioning accuracy. Effective target engagement in dense combat formations and limited ammunition requires high accuracy in bringing the UAV to the release or strike point. As NSIN (2025) emphasises, “autonomous systems perform missions at speeds and accuracies far exceeding human capabilities.” Modern terminal guidance systems like the Ukrainian TFL-1 provide “autonomous drone guidance over the final 400–500 metres of flight,” requiring autopilot positioning accuracy at the metre level under dynamic manoeuvre. As Asteria (2025) notes, “drones can simultaneously track movement, detect thermal signatures, and map terrain with sub-metre accuracy.”
<20 ms reaction time. In high-dynamics scenarios – evading air defence, intercepting enemy drones, correcting dive trajectories – autopilot reaction time becomes critical. As Defense Advancement (2025) states, “defence autopilot systems must meet strict airworthiness and reliability standards.” <20 ms reaction time is needed to ensure control stability under high angular rates and turbulence typical of dive attacks. As IOPscience (2024) shows, “the UAV ground-attack process during combat is simulated on a hardware-software simulator, obtaining UAV flight data during dive attacks,” enabling system reaction-time optimisation.
Open autopilot platforms Pixhawk and ArduPilot are widely used in academic, commercial, and high-reliability applications. As industry analyses note, "platforms such as Pixhawk, PX4, and ArduPilot are widely used for both open and high-reliability adapted applications, offering scalable architectures suitable for fixed-wing UAVs and hybrid VTOL drones." However, their compliance with stringent requirements for demanding operational scenarios requires detailed analysis.
Pixhawk hardware capabilities. Modern Pixhawk controllers, such as the Pixhawk 6X Pro, are equipped with high-performance processors and sensors. The Pixhawk 6X Pro uses “a high-performance industrial IMU ADIS16470 with ±40 g accelerometer dynamic range, ideally suited for accurate motion measurement in demanding UAV applications.” This means Pixhawk hardware is fundamentally capable of withstanding up to 40g overloads, far exceeding the required 10g. The controller also features “triple IMU redundancy, including a primary sensor capable of withstanding accelerations up to ±40 g.”
However, as noted in MDPI (2022), even with high-quality sensors, limitations exist due to “manufacturing tolerances and measurement stability.” Pixhawk magnetometers show variations from 25 mG to 116 mG against reference values of 1.3–5.4 mG, indicating limited accuracy under electromagnetic interference – a critical factor for combat scenarios.
ArduPilot software limitations. As an open-source platform, ArduPilot has several limitations critical for high-performance and safety-critical applications. As documentation notes, "hardware error exists in these chips, limiting flash memory to 1 MB. Any board containing this chip cannot include all ArduPilot features due to this limitation." This restricts the ability to add complex guidance algorithms, advanced computer vision, and secure communication protocols directly on the autopilot.
Developer forums also note other issues: “gyro not calibrated. Constant compass problems. When the UAV moves in space, Failsafe immediately triggers during EKF processing.” Yet “the UAV behaves stably with PX4 firmware,” indicating issues specifically with ArduPilot implementation on some hardware platforms.
As noted in a Pixhawk Discourse (2023) discussion, “the control system is very complex, and there are many ambiguities.” This creates risks for combat scenarios where predictability and debugging simplicity are critical. Additionally, open-source code creates vulnerabilities: “this is precisely what makes using an open autopilot so dangerous.” A system analysis of “most Indian production drones will show either Pixhawk or Pixhawk Cube based on open autopilot systems at the most basic drone motherboard level,” creating cyber-security risks.
Reaction-time limitations. ArduPilot reaction time in standard configuration is limited by software architecture. As documentation states, “the maximum normal flight speed is set by AIRSPEED_MAX. A good target is usually twice cruise speed.” However, for dive attacks where speeds may far exceed cruise, standard limits may be insufficient. One discussion describes a case where “dive entry was held with full elevator deflection for almost 4 seconds. By the time it was released, the vehicle was already moving at nearly 35 m/s,” demonstrating that ArduPilot may lack built-in dive-speed limiting mechanisms.
Based on the comparative analysis, the following key shortcomings of Pixhawk/ArduPilot open autopilots for high-performance civilian applications can be identified.
1. Limited computational power for complex algorithms. The 1 MB flash memory limit on some boards prevents onboard implementation of complex guidance, computer vision, and real-time adaptive control algorithms. This requires external companion computers, increasing weight, power consumption, and overall system complexity.
2. Insufficient sensor accuracy under electromagnetic interference. Significant magnetometer variation makes Pixhawk vulnerable to orientation errors under EW jamming typical of combat scenarios. This requires additional filtering and calibration that may not ensure the required ±5 m accuracy during dives.
3. Lack of built-in dive-speed limiting mechanisms. As discussions show, ArduPilot may lack adequate mechanisms to prevent uncontrolled acceleration during dives. For dive-bombers where speeds may reach 160 km/h and above, this creates risks of control loss and airframe destruction.
4. Configuration and debugging complexity. The control system is inherently complex, with many ambiguities in its configuration and tuning. For high-performance civilian applications requiring rapid deployment and predictable behaviour, this is a serious drawback. Configuration errors can lead to mission failures, equipment damage, or operational downtime in safety-critical scenarios.
5. Cyber-security vulnerabilities. Open-source code creates risks of malicious code injection, control interception, and system compromise. As Defense Advancement (2025) notes, the defence sector is forced to “extend them with classified security levels and proprietary mission modules,” increasing cost and development complexity.
6. Lack of specialised high-performance flight modes. ArduPilot has no built-in advanced flight modes for demanding manoeuvres such as steep descent or rapid approach profiles. As noted in technical media, such software should not be much more complex, but its development and integration require significant effort. Thus, developers of high-performance UAV platforms must create custom overlays, increasing time and cost.
Based on the identified shortcomings, the following technical requirements for enhancing autopilots for reusable dive-bombers and interceptors can be formulated.
1. Increase computational power and memory. Use controllers with larger flash memory (at least 2 MB) and RAM to support complex guidance algorithms, computer vision, and adaptive control. The Pixhawk 6X with STM32H7 and 2 MB flash represents a step forward, but high-performance civilian applications may require even greater processing headroom to ensure reliability in demanding environments.
2. Develop specialised dive and interception modes. Create modes similar to ArduPilot’s “acro” modes but optimised for combat manoeuvres. As IOPscience (2024) shows, “the UAV ground-attack process during combat is simulated on a hardware-software simulator.” Based on such simulations, modes should ensure:
- Automatic dive-exit upon reaching speed limits.
- Wind-gust compensation at steep dive angles.
- Integration with terminal guidance systems (TFL-1 and similar).
3. Improve sensor filtering and calibration. To achieve ±5 m accuracy under electromagnetic interference, improved IMU and magnetometer filtering algorithms are needed. As MDPI (2022) notes, “Pixhawk autopilot limitations can be identified both in terms of manufacturing tolerances and measurement stability.” Addressing these requires both hardware and software solutions.
4. Implement redundancy and fault tolerance. For combat scenarios, backup systems are critical. The Pixhawk 6X Pro already provides “triple IMU redundancy, including a primary sensor with ±40 g accelerometer dynamic range, and dual barometer redundancy.” However, further development is needed:
- Control-channel redundancy.
- Automatic switchover to backup sensors upon primary failure.
- Emergency recovery modes upon link loss.
5. Ensure cyber-security. As Defense Advancement (2025) notes, the defence sector “extends open autopilot frameworks with classified security levels and proprietary mission modules.” Needed:
- Secure communication channels with hardware encryption.
- Intrusion detection and prevention mechanisms.
- Firmware protection from unauthorised access and modification.
6. Optimise reaction time. To achieve <20 ms reaction time, need:
- High-frequency control loops (≥400 Hz).
- Optimise critical code sections (EKF, control, guidance).
- Use hardware acceleration for sensor processing.
Based on the analysis, the following practical recommendations for developers of high-performance autopilot systems using open platforms can be formulated:
1. Use Pixhawk 6X or newer platforms with STM32H7 and at least 2 MB memory to implement complex algorithms.
2. Integrate specialised companion computers (NVIDIA Jetson, Raspberry Pi) for computer vision, terminal guidance, and adaptive control, offloading the main autopilot for time-critical flight-control tasks.
3. Develop and validate dive modes using HIL simulations followed by flight tests at specialised ranges.
4. Use secure communication channels with hardware encryption and intrusion detection to prevent control interception.
5. Implement graceful degradation upon sensor failure or link loss, ensuring safe mission completion or return-to-base.
6. Conduct regular code audits and vulnerability testing to minimise cyber-attack risks.
Developing high-performance UAV platforms for demanding civilian applications imposes autopilot requirements far exceeding standard commercial capabilities. Requirements for high-manoeuvrability capability, centimetre-level positioning accuracy, and rapid response times are critical for operational success in safety-critical missions such as search-and-rescue, infrastructure inspection, and emergency response.
Open autopilot platforms Pixhawk and ArduPilot, despite their prevalence and adaptability, have several significant shortcomings for high-performance civilian applications: limited computational power and memory, insufficient sensor accuracy under electromagnetic interference, lack of built-in safeguards for extreme manoeuvres, configuration complexity, and cyber-security vulnerabilities. Modern Pixhawk hardware (e.g., Pixhawk 6X Pro with ±40 g IMU) is fundamentally capable of withstanding required loads, but ArduPilot software implementation requires substantial enhancement to meet the required characteristics.
Autopilot enhancement for demanding civilian applications should include increased computational power and memory, specialised high-performance flight modes, improved sensor filtering and calibration, redundancy and fault tolerance, cyber-security hardening, and reaction-time optimisation. Future research should develop reference autopilot mode implementations for high-performance applications, validate them under realistic operational loads, and integrate with modern guidance and AI systems.
8. Mass Production and Maintenance Infrastructure for Large-Scale UAV Swarms
Large-scale deployment of UAV swarms for commercial and civilian applications is impossible without corresponding mass production and maintenance infrastructure. Unlike single UAVs, where maintenance can be handled individually, swarms of tens and hundreds of platforms require a fundamentally different approach: automation of all life-cycle stages – from launch and landing to battery replacement, logistics, and processing of huge data volumes. As Thorstensen (2025) notes, "what no one is yet managing properly is the ground segment: where UAVs are actually launched, landed, charged, replaced, queued, parked, and coordinated."
This section addresses key components of mass production and maintenance infrastructure for UAV swarms: automated launch and landing, battery-swap and charging stations, logistics and supply, and computing infrastructure for data processing and swarm control. Special attention is given to estimating required computational power – for a 100-UAV swarm streaming 4K video at 30 fps, about 2 PetaFLOPS on the ground are needed – and justifying a hybrid computation distribution scheme (70% ground, 30% onboard with NPUs). Based on the analysis, practical resource-calculation recommendations for fielding swarm systems are formulated.
A key element of swarm infrastructure is automated launch and landing, especially when operations must be conducted from moving platforms – ground vehicles, ships, or even other aircraft. As Defense Express (2025) notes, “combat UAVs are still launched manually; Target Arm automates launch and landing so drones can land on moving vehicles.” The goal is “to reduce personnel risk and automate drone operations.”
Launch/landing systems on moving platforms. In 2025, Teledyne FLIR Defense introduced SkyCarrier – “a launch and recovery platform allowing small ground vehicles and warships to autonomously operate the company’s quadcopters, such as SkyRanger R70 or SkyRaider R80D.” SkyCarrier provides “fully autonomous launch and recovery on the move” with active roll compensation up to 20°, “allowing UAVs to land on moving or stationary platforms without manual control.” The system uses “visual markers and beacons to guide aircraft back to the platform without operator involvement.”
That same year, US company Target Arm developed robotic systems Ralar and Tular, “designed for launching and recovering drones without stopping the vehicle or exposing the crew to danger.” These systems enable drone operations under continuous movement, critical for tactical missions where column stops are unacceptable.
Automated droneports and nests. In the civilian sector, automated droneports are actively developing. New Century Logistics introduced the Micro-Hub System, “integrating vehicle-mounted drone nests, automated battery swapping using a six-axis robot, and cargo-handling technology to achieve full automation and intelligence of the entire logistics process.” The system “supports automatic UAV launch and landing and seamlessly integrates with the vehicle through a sliding structure.”
Deployable drone stations for swarms. Various organisations are developing solutions for deploying UAV swarms from commercial vehicles: "a compact drone package can be mounted on the roof of a vehicle and connected to the onboard battery, serving simultaneously as a charging station and launch platform." This approach integrates swarm capabilities into existing logistics and transportation infrastructure without requiring dedicated facilities.
Limited battery capacity remains a major constraint for long swarm missions. Both automated battery-swap systems and wireless charging technologies are being developed.
Automated battery-swap stations. Springer (2025) proposed “an energy-efficient approach using mobile ground battery-swap stations (BSSes) to accelerate UAV battery swapping and reduce energy loss from flying to charging stations.” The Context-Aware Coverage Path Planning (CACPP) problem is solved “for covering a large area with a UAV swarm while minimising path overlap and UAV battery swaps.” IEEE (2025) describes a strategy where “the control centre selects the nearest charging station for each UAV based on station battery information to reduce flight time and ensure rapid power through battery swapping.”
Mobile charging stations. Mobile stations that can move with the swarm are of particular interest, reducing flight time to fixed charging points. Springer (2025) presents “a multi-agent reinforcement learning approach for computing coverage paths over large areas by UAV swarms supported by mobile battery-swap stations.” These algorithms “can reduce flight distance, optimise battery swapping during mission, minimise energy consumption, and ensure full area coverage.”
In-flight charging and wireless power transfer. A promising direction is in-flight UAV charging. Han (2025) considers “mid-air battery swapping for drones,” which “solves the energy problem limiting flight time and range of multirotor drones.” The “drone swarm lift” concept allows “drone groups to cooperatively and autonomously lift and transport payloads, overcoming single-drone lift limitations.”
Capacitive wireless charging for swarms. IEEE (2025) proposed “an electric capacitive wireless charging system for UAV swarms capable of simultaneously charging multiple drones through a single transmitting unit with excellent misalignment tolerance and load decoupling capability.” Such systems can charge the entire swarm simultaneously, reducing downtime and increasing operational readiness.
Massive UAV swarm deployment requires rethinking the entire logistics chain – from production and storage to transport and forward-area maintenance.
Production capacity and scaling. As noted in industry analyses, "implementing drone swarms in logistics, agriculture, or monitoring eliminates traditional cost barriers. Companies can dynamically scale operations without proportional cost increases, for example, automating delivery, precision crop spraying, or infrastructure maintenance." This approach is equally relevant for commercial and civilian applications: mass production of drones and modular components allows rapid swarm expansion without proportional growth in operational costs or personnel requirements.
Automated logistics chains. New Century Logistics developed a system where “a six-axis robot performs automatic battery swapping and cargo loading onto UAVs, reducing dependence on manual labour.” Such automation enables servicing large swarms with minimal human involvement, critical in combat where personnel are the most scarce and vulnerable resource.
Forward-area maintenance infrastructure. The US Army is developing concepts for “integrating drone swarms into logistics platforms,” where “a small drone package can be mounted on the roof of a logistics platform and connected to the vehicle’s onboard battery, serving simultaneously as a charging station and launch platform.” This allows swarms to be deployed directly in combat zones without specialised infrastructure.
Processing data generated by a UAV swarm is one of the most resource-intensive tasks. A 100-UAV swarm, each streaming 4K video at 30 fps, generates huge data volumes requiring real-time processing for situational awareness, decision-making, and control.
Computing power estimate. For a 100-UAV swarm with 4K (3840×2160) video at 30 fps, processing ~100 × 3840 × 2160 × 30 ≈ 24.9 gigapixels per second is required. Modern computer-vision, object-recognition, tracking, and data-fusion algorithms require 10–100 TFLOPs per gigapixel processed. Thus, total computing power for real-time video processing from 100 UAVs is approximately 2 PetaFLOPS (2×10¹⁵ floating-point operations per second). For comparison, modern supercomputers like MareNostrum 5 reach 314 PetaFLOPS, meaning processing data from a 100-UAV swarm requires computing power comparable to high-performance computing segments.
Hybrid computation distribution (70% ground, 30% onboard with NPU). For efficient data processing, a hybrid distribution scheme is recommended:
- 70% ground-side (GCS/Cloud) – processing non-time-critical data: global route planning, multi-UAV data fusion, long-term forecasting, model training, storage, and archiving. Ground centres should provide ~1.4 PetaFLOPS for a 100-UAV swarm.
- 30% onboard (NPU/Edge AI) – latency-critical tasks: real-time object recognition, local trajectory planning, obstacle avoidance, data filtering before ground transmission. Each UAV should be equipped with an NPU with at least 50–100 TOPS (trillions of operations per second). For 100 UAVs, total onboard power ~0.6 PetaFLOPS.
Technology base for onboard computing. Modern NPU solutions provide required performance under strict SWaP constraints. As MarketsandMarkets (2025) notes, “advances in low-power AI accelerators (NPU, GPU, FPGA) allow drones to process data onboard in real time, reducing dependence on cloud connectivity.” Nvidia Jetson Orin, “a generative AI supercomputer the size of a palm costing $249, capable of 67 trillion operations per second,” allows a drone to “process thermal images, object recognition, telemetry, and onboard logic in real time.”
The SWARMER project develops “a high-performance Edge AI computing load capable of supporting heavy navigation and AI processing loads under strict SWaP constraints.” Studies show that “integrated NPU architectures provide optimal thermal and energy characteristics for onboard UAV computing.” For example, “the integrated Rockchip NPU achieved an average performance of 18 fps – 8.4 times faster than baseline CPU – at only 0.83 W power consumption.”
Digital twins for computing-resource management. Digital twins play a key role in swarm computing-resource management. IEEE (2025) presents “a digital-twin-based architecture for resource management in UAV swarms, connecting real crowdsourced task execution and virtual traffic planning to achieve additional multi-UAV distribution.” ScienceDirect (2025) developed “a machine-learning-enabled digital twin for rapid determination of optimal programming of desired tactical behaviour of multi-drone swarms.” Digital twins allow modelling and optimising computing-resource distribution between ground and onboard systems before the actual mission.
IASA (2025) presents “the concept and architecture of digital twins in AI-controlled autonomous navigation tasks for UAV swarms.” The study showed that “effective swarm operation under communication disruption or loss with the ground centre is ensured by functional distribution of digital-twin components between the ground centre and onboard systems.”
Mobile and edge computing resources. Field-data processing can use mobile computing nodes. CDS (2025) considers a concept where “drones visit a set of waypoints to collect data and perform analytics, having access to onboard edge computing, stationary fog nodes on cell towers, and mobile fog nodes on public buses.”
Based on the analysis, the following practical recommendations for deploying mass production and maintenance infrastructure for UAV swarms can be formulated:
1. Establish automated droneports for launch and landing. Implement systems like SkyCarrier (Teledyne FLIR) or Ralar/Tular (Target Arm), providing automated launch and landing on moving platforms with up to 20° roll compensation. For stationary applications, use automated droneports with robotic battery swapping.
2. Deploy mobile battery-swap stations. For long missions, deploy mobile charging stations moving with the swarm. Use MARL algorithms to optimise battery-swap routes, reducing downtime and power consumption.
3. Implement hybrid computing architecture (70% ground, 30% onboard). For a 100-UAV swarm with 4K at 30 fps, provide ~1.4 PetaFLOPS at ground centres and equip each UAV with 50–100 TOPS NPU (total ~0.6 PetaFLOPS onboard). This allows time-critical tasks (recognition, evasion) onboard and resource-intensive tasks (global planning, training, archiving) on the ground.
4. Use digital twins for planning and optimisation. Digital twins enable modelling computing-resource distribution, route optimisation, and energy consumption forecasting before mission start, reducing risks and improving resource efficiency.
5. Integrate swarms into existing logistics infrastructure. For commercial and civilian applications, integrate droneports and charging stations into logistics and transport infrastructure — such as delivery vans, trucks, service fleets, and cargo ships. For stationary operations, create automated hub networks similar to the Micro-Hub System, enabling seamless integration of drone swarms into existing logistics chains without requiring dedicated facilities.
6. Modularity and component standardisation. To reduce cost and simplify maintenance, use modular UAV designs with interchangeable components (batteries, sensors, computing modules). This enables rapid field restoration of swarm combat capability.
7. Develop onboard NPU solutions. For required performance under SWaP constraints, use specialised neural processors like Nvidia Jetson Orin (67 TOPS) or Rockchip NPU (18 fps at 0.83 W). Integrated NPU architectures provide optimal thermal and energy characteristics for onboard computing.
8. Ensure computing resilience under EW. Under jamming when ground-centre communication is disrupted, the swarm must retain onboard autonomous data-processing capability. This requires sufficient onboard computing power for critical functions without external support.
Mass production and maintenance infrastructure is a critical component determining the feasibility of practical UAV swarm deployment in commercial and civilian applications. Automated launch and landing systems, similar to SkyCarrier and Ralar/Tular, enable operations from moving platforms without operator involvement, reducing operational risk and increasing responsiveness. Automated battery-swap systems and mobile charging stations address the limited-energy problem, enabling long-duration missions without the need for landing.
The most resource-intensive infrastructure component is the computing system for swarm data processing. For a 100-UAV swarm streaming 4K at 30 fps, ~2 PetaFLOPS total computing power is required. The optimal solution is a hybrid distribution: 70% ground (~1.4 PetaFLOPS) for global planning, data fusion, and model training, and 30% onboard (~0.6 PetaFLOPS) using NPUs for latency-critical tasks – object recognition, local planning, and evasion. Modern NPU solutions like Nvidia Jetson Orin (67 TOPS) and Rockchip NPU (18 fps at 0.83 W) provide required performance under strict SWaP constraints.
Digital twins play a key role in managing computing resources, enabling modelling and optimisation of task distribution between ground and onboard systems before mission start. Integrating swarms into existing logistics infrastructure, modularity and component standardisation, and ensuring computing resilience under EW are critical requirements for practical swarm deployment.
Future research should develop standardised swarm-infrastructure architectures, create energy-efficient NPU solutions with >100 TOPS at <5 W consumption, and integrate wireless charging and in-flight battery swapping for continuous swarm operations.
9. Conclusion and Outlook
The systematic literature review (2020–2026) allows several fundamental conclusions to be drawn, defining the current state and prospects of decentralized UAV swarm development. Based on 47 relevant publications selected from IEEE Xplore, Scopus, and Web of Science, key trends, challenges, and breakthrough solutions in swarm intelligence, decentralized control, electronic countermeasures, energy autonomy, and human-machine interaction have been identified.
First, the transition from centralized control architectures to decentralized swarm systems is not merely an evolutionary step but a fundamental paradigm shift. Traditional centralized systems based on a single command centre and pre-programming of each agent face insurmountable scalability, fault-tolerance, and adaptability limitations in dynamic environments. As shown in EE Times (2025), “light shows or pre-programmed fleets are centrally managed and require constant oversight. In contrast, true swarms rely on decentralized consensus, local autonomy, and dynamic membership.” The proposed taxonomy of three architectural types (A – centralized, B – hierarchical, C – fully decentralized) allows systematic classification and informed architecture selection based on mission requirements. Quantitative analysis shows that decentralized swarms (Type C) maintain high effectiveness even with up to 30% agent loss, whereas 5% agent loss in a centralized swarm (Type A) can cause mission failure (JISEM, 2026; ACM, 2025).
Second, detailed analysis of positioning and navigation methods shows that relying solely on GPS in swarms is unacceptable, especially in challenging environments where GPS signals are unreliable or unavailable. Global navigation systems are vulnerable to interference, spoofing, and signal blockage in urban canyons. Alternative methods – VIO/SLAM, UWB, LiDAR, AoA, and mutual positioning – provide varying accuracy and resilience. Cooperative localization with five or more UAVs and mutual range measurements can reduce RMS positioning error to 0.5–1 m over distances up to 200 m (Luo et al., 2022). Under urban occlusion, error rises to 5–10 m, requiring hybrid sensor-redundant schemes. The CRLB-based quantitative model proved effective as an optimization tool for swarm configuration and resource allocation to maximize localization accuracy.
Third, comparative analysis of MANET and DTN decentralized communication protocols revealed their fundamental differences and optimal application areas. MANET provides <100 ms latency and >0.95 PDR in dense swarms (density >0.1 nodes/m³), but requires continuous connectivity and high power (5–10 W per agent). DTN tolerates up to 30 s breaks, provides 1–5 min latency and 0.7–0.85 PDR with 2–3 W per agent (40% saving) (Beisenkhanov et al., 2023; Kumar et al., 2020). The proposed hybrid scheme – MANET intra-cluster for low-latency critical commands and DTN inter-cluster for link-break resilience – provides optimal balance for combat conditions where connectivity cannot be guaranteed.
Fourth, the analytical J/S model J/S=(PjGjRs2)/(PsGsRj2) enabled quantitative assessment of swarm communication-channel vulnerability to EW. Calculation shows that with FHSS at 1000 hops/s and 1 W transmitter power, jamming at 2 km requires Pj≥50 W with directional antenna and Pj≥500 W with omnidirectional antenna. This makes mobile EW stations vulnerable to counter-battery fire. Radio-silence tactics with 10-ms sessions every 10 s reduce detectability by 20 dB (IEEE, 2023), significantly enhancing swarm survivability.
Fifth, the grid-weight model wi(t+1)=wi(t)+αdi(t)−βλwi(t) with decision threshold wth=10 was formalized, providing decentralized targeting for reconnaissance-strike swarms. The positive-feedback (α) and negative-feedback (β) mechanism mimics ant-colony pheromone communication, enabling collective target identification without centralized control (Frontiers, 2023). Intercepting 10% of agents without adaptive algorithms reduces mission success by 15–20%, whereas with adaptive algorithms only 5–7%, with swarm reconfiguration time below 0.5 s (IEEE, 2025; ScienceDirect, 2026).
Sixth, economic analysis revealed the negative cost-exchange ratio (CER) phenomenon, reaching 190:1 in favour of the attacker (Command Eleven, 2026). Combat experience shows CER up to 1:1000, where a $3 million attack inflicts $3 billion damage (ICWA, 2025). A 50-UAV swarm costing $1–3.5 million is comparable to one cruise missile but provides multi-channel attack far harder to defeat. TCO includes development, production, training, operation, and attrition costs, yet swarms remain cost-effective due to mass production, modularity, and loss resilience.
Seventh, human-swarm interaction requires fundamentally new interfaces. Mixed-initiative interfaces using LLMs to translate high-level operator goals into swarm task plans reduce cognitive load and accelerate decision-making (Nazzari et al., 2025). Heat-map visualisation instead of individual points proved more effective in usability and perception, especially when agent count is large or time limited (Clark et al., 2022; Ahlskog et al., 2024). Ethical analysis of decentralized targeting revealed a fundamental contradiction with IHL – the absence of a single commander creates accountability problems (UNRIC, 2025; HRW, 2025). Mandatory human-in-the-loop confirmation for attack decisions at threshold w≥8w≥8 is proposed.
Eighth, analysis of autopilot requirements for high-performance civilian applications shows that open platforms Pixhawk/ArduPilot, despite their prevalence, have significant shortcomings: limited computational power (1 MB flash on some boards), insufficient sensor accuracy under electromagnetic interference, lack of built-in safeguards for extreme manoeuvres, and cyber-security vulnerabilities (Defense Advancement, 2025; MDPI, 2022). Modern Pixhawk hardware (e.g., Pixhawk 6X Pro with ±40 g IMU) is fundamentally capable of withstanding high loads, but ArduPilot software requires substantial enhancement to meet the requirements of demanding operational scenarios.
Ninth, computing-infrastructure assessment showed that a 100-UAV swarm streaming 4K at 30 fps requires ~2 PetaFLOPS total. The optimal solution is a hybrid distribution: 70% ground (~1.4 PetaFLOPS) for global planning, data fusion, and model training, and 30% onboard (~0.6 PetaFLOPS) using NPUs for latency-critical tasks (MarketsandMarkets, 2025; SWARMER, 2025). Modern NPU solutions like Nvidia Jetson Orin (67 TOPS) provide required performance under strict SWaP constraints.
This paper makes the following original contributions to decentralized UAV swarm research:
1. Taxonomy of autonomy levels, positioning methods, and communication types. Unlike existing narrative reviews, a systematic classification including six autonomy levels (0–5) with functions, sensor types, and GNSS dependence for each level is proposed. A taxonomy of three architectural types (centralized, hierarchical, decentralized) with quantitative fault-tolerance assessment enables developers and customers to make informed architecture choices (EE Times, 2025; Springer, 2024). Comparative analysis of MANET and DTN across three criteria (latency, PDR, power) with introduction of a hybrid scheme (MANET intra-cluster, DTN inter-cluster) provides practical value for combat communication system design (Beisenkhanov et al., 2023; Kumar et al., 2020).
2. Quantitative resilience models. An analytical J/S model J/S=(PjGjRs2)/(PsGsRj2) J/S=(PsGsRj2)/(PjGjRs2) for assessing communication-channel vulnerability to EW and calculating required FHSS jamming power is developed. A grid-weight model wi(t+1)=wi(t)+αdi(t)−βλwi(t) with decision threshold wth=10, formalizing stochastic swarm behaviour for decentralized targeting, is proposed (Frontiers, 2023). A CRLB model for cooperative swarm localization enables quantitative assessment of theoretical positioning accuracy depending on agent count and measurement quality (Luo et al., 2022). A power-consumption model with flight-power and communication-cost breakdown (MANET – 5–10 W, DTN – 2–3 W) and mission-time increase estimate at 3:1 worker-donor ratio (60–80%) provides a tool for swarm energy-balance optimization.
3. Comparative analysis and identification of shortcomings. A comparison of existing open autopilot capabilities (Pixhawk/ArduPilot) with the requirements of high-performance civilian applications — including high-manoeuvrability capability, centimetre-level positioning accuracy, and rapid response times — is conducted, identifying key shortcomings: limited computational power, insufficient sensor accuracy under electromagnetic interference, lack of built-in safeguards for extreme manoeuvres, and cyber-security vulnerabilities (Defense Advancement, 2025; MDPI, 2022). Quantitative assessment of operator intervention impact — where intervention in 10% of agents reduces mission success by 15–20% without adaptive algorithms and by only 5–7% with adaptive algorithms, with reconfiguration time below 0.5 seconds — provides objective criteria for systems with manual override (IEEE, 2025; ScienceDirect, 2026).
4. Ethical analysis of decentralized targeting. For the first time within a technical review, a systematic ethical analysis of decentralized targeting is conducted, revealing contradiction with IHL principles (no single commander, accountability problems) (UNRIC, 2025; HRW, 2025). Mandatory human-in-the-loop confirmation for attack decisions at threshold w≥8w≥8 is proposed, with automatic prohibition of engagement at w≥10w≥10 without external verification. The need for special test ranges and simplified certification to accelerate swarm development is justified.
5. Computing-infrastructure assessment. For the first time, a quantitative assessment of computing power required to process data from a 100-UAV swarm with 4K at 30 fps – ~2 PetaFLOPS total – is performed. A hybrid distribution scheme (70% ground, 30% onboard with NPU) is justified, and specific recommendations for selecting onboard NPU solutions (Nvidia Jetson Orin, Rockchip NPU) to meet performance requirements under SWaP constraints are given (MarketsandMarkets, 2025; SWARMER, 2025).
Despite significant progress in swarm technologies, several critical questions remain:
1. Experimental validation at test ranges. Most studies are based on simulations or lab experiments with limited agents. Full-scale field tests of 50–100 UAV swarms under near-real combat conditions (EW, dynamic obstacles, weather) are needed. As Cetinsaya et al. (2024) notes, “despite a significant number of theoretical works, experimental validation of swarm algorithms in real conditions remains insufficient.” Specialized ranges with controlled electromagnetic environments and various tactical-scenario modelling are required, including radio-silence tactics, intermittent sessions, and adaptive density control under active EW.
2. Development of hybrid communication protocols. The proposed MANET/DTN hybrid scheme requires further development and optimization, including dynamic switching between protocols depending on tactical situation, threshold optimization, and integration with FHSS for enhanced jamming resilience. As Beisenkhanov et al. (2023) shows, “DTN protocol performance in UAV swarms strongly depends on node density and mobility.” Adaptive protocols automatically tuning parameters based on current communication conditions and tactical situation, including integration with machine learning for channel-quality prediction, are needed.
3. Integration with large language models (LLMs). Using LLMs for mixed-initiative interfaces is in its early stages. As Nazzari et al. (2025) shows, TACOS “allows LLMs to interact with a library of executable APIs, bridging semantic reasoning with real-time multi-robot coordination.” However, questions remain: how to ensure LLM reliability and predictability in critical situations, how to integrate LLMs with existing flight-control systems, how to minimize natural-language processing delays, and how to ensure safety when using LLMs in combat, including protection against adversarial attacks on language models.
4. Autonomous learning and real-time adaptation. Current MARL algorithms require significant training time and adapt poorly to novel unforeseen situations. Research in meta-learning, online RL, and transfer learning for swarm systems is needed. As Frontiers (2023) notes, “the decentralized and self-organizing nature of swarms makes them difficult to design – requirements for desired swarm behaviour are usually stated at the collective level, but it is impossible to program the swarm directly.” Methods enabling swarms to learn new tactics during mission execution, adapting to adversary actions and environmental changes without retraining from scratch, are required.
5. Cyber-security and attack resilience. UAV swarms are vulnerable not only to physical and EW attacks but also to cyber-attacks on onboard software, communication channels, and decision-making algorithms. As Defense Advancement (2025) notes, the defence sector “extends open autopilot frameworks with classified security levels and proprietary mission modules.” Intrusion-detection and prevention methods, secure communication protocols, decision-making algorithms resilient even under partial agent compromise, and firmware-integrity and hardware-trust verification systems are needed.
6. Energy autonomy and wireless charging. Despite progress in flying batteries and WPT, efficiency and scalability questions remain. As Reach Power (2026) shows, WPT systems for swarms are at an early commercialization stage. Research on WPT efficiency optimization (range, losses, impact on onboard electronics), standardized charging interfaces for different UAV types, energy-efficient mission-planning algorithms considering recharging needs, and integration with renewable energy sources for long autonomous operations is required.
7. Socio-legal aspects. Development of international and national regulatory frameworks for civilian autonomous systems is still in its early stages. As international organisations note, calls have been made for establishing legally binding instruments to govern the deployment of autonomous technologies. Further research on responsibility, accountability, and human oversight for decentralized autonomous swarms is needed, including certification standards, international control and verification mechanisms, and legal frameworks for civilian use that respect safety, privacy, and ethical principles.
8. Human-swarm interaction under stress. Current human-swarm interaction (HSI) studies are predominantly laboratory-based. Research on human-swarm interaction under high-stress conditions — such as emergency response, time-sensitive operations, and crisis management — is critically needed. As academic research indicates, operator workload assessments under demanding conditions often exceed cognitive capacity thresholds. Interfaces that adapt to operator cognitive state using biometric data (e.g., pupillometry, functional near-infrared spectroscopy) for real-time autonomy level and interface complexity adaptation are required (arXiv, 2024).
Based on the analysis, the following practical recommendations for swarm system developers and civilian operators can be formulated:
1. Choose swarm architecture by mission. For scenarios with high fault-tolerance and adaptability requirements — such as autonomous logistics, search-and-rescue, and environmental monitoring in challenging conditions — decentralized self-organizing swarms (Type C) are recommended. They provide maximum loss resilience (maintaining effectiveness with up to 30% agent loss) and have no single point of failure (Lowy Institute, 2020; MDPI, 2024). For scenarios where accuracy and predictability are critical — such as precision agriculture, infrastructure inspection, and drone shows — centralized or hierarchical architectures (Type A or B) may be used (EE Times, 2025; GAO, 2023).
2. Use hybrid communication protocols. Implement a hybrid scheme: MANET intra-cluster (for low-latency critical commands) and DTN inter-cluster (for link-break resilience) (Beisenkhanov et al., 2023; Kumar et al., 2020). This provides optimal balance between latency, reliability, and power in combat where connectivity cannot be guaranteed. For jamming resilience, use FHSS with high hop rates (≥1000 hops/s) and intermittent communication sessions (10 ms every 10 s), reducing detectability by 20 dB (IEEE, 2023).
3. Integrate alternative positioning methods. For operations in GPS-denied environments, use hybrid positioning systems combining VIO/SLAM, UWB, and mutual swarm positioning. With five or more UAVs using mutual range measurements, RMS error can be reduced to 0.5–1 m (Luo et al., 2022). For demanding civilian applications — such as urban logistics, infrastructure inspection, and operations in complex environments — robust positioning under occlusion (urban canyons, forests, indoor spaces) using CRLB-optimal swarm configurations is critical (IEEE, 2025).
4. Develop specialised autopilot modes. For reusable dive-bombers and interceptors, enhance open autopilots (Pixhawk/ArduPilot) in: computational power and memory (upgrade to Pixhawk 6X with 2 MB flash), specialised dive/interception modes with automatic speed limiting, improved sensor filtering/calibration for electromagnetic-interference environments, redundancy and fault tolerance (triple IMU, dual barometers) (Defense Advancement, 2025; MDPI, 2022). Reaction time must be <20 ms for stability under high angular rates.
5. Implement mixed-initiative interfaces with LLMs. Use LLM-based interfaces for natural-language swarm control. As Nazzari et al. (2025) shows, TACOS “allows LLMs to interact with a library of executable APIs, bridging semantic reasoning with real-time multi-robot coordination.” Visualise swarm state via heat maps instead of individual points, reducing operator cognitive load and enhancing situational awareness, especially with many agents (Clark et al., 2022; Ahlskog et al., 2024). Aggregate video streams into a single panoramic image with AI-highlighted anomalies to further reduce load.
6. Ensure human-in-the-loop confirmation for critical decisions. For IHL compliance, implement mandatory human-in-the-loop confirmation for attack decisions (UNRIC, 2025; HRW, 2025). Concrete implementation: at w≥8w≥8, system requests operator confirmation with 5-second timeout; at w≥10w≥10, automatic prohibition of engagement without external verification. This preserves swarm autonomy advantages while maintaining human control and accountability.
7. Establish specialised computing infrastructure. For a 100-UAV swarm with 4K at 30 fps, ~2 PetaFLOPS total computing power is required. Recommended hybrid distribution: 70% ground (~1.4 PetaFLOPS) for global planning, data fusion, and model training, and 30% onboard (~0.6 PetaFLOPS) using NPUs for latency-critical tasks (MarketsandMarkets, 2025; SWARMER, 2025). For onboard computing, use NPU solutions with ≥50–100 TOPS at <5 W (e.g., Nvidia Jetson Orin, 67 TOPS) (Unmanned Systems Technology, 2025). Digital twins enable modelling and optimizing resource distribution before mission start (IEEE, 2025; ScienceDirect, 2025).
8. Automate ground infrastructure. For continuous swarm operations, implement automated launch/landing systems (SkyCarrier, Ralar/Tular) for moving-platform operations without operator involvement (Defense Express, 2025; Teledyne FLIR, 2025). For battery swapping, use mobile BSSes with MARL-based route optimization to reduce downtime and power (Springer, 2025). For logistics, integrate droneports and charging stations into existing logistics platforms (trucks, APCs, ships) for forward-area deployment (US Army, 2025).
Final Remarks. Decentralized UAV swarms represent one of the most transformative technologies of the 21st century, capable of radically changing civilian domains — from logistics and infrastructure monitoring to search-and-rescue, disaster response, and environmental observation. The systematic review shows that key technological components — decentralized control, alternative positioning methods, hybrid communication protocols, interference resilience models, and mixed-initiative interfaces — have already reached maturity sufficient for practical deployment.
However, transitioning from lab prototypes to large-scale combat application requires overcoming not only technical but also organizational, legal, and ethical barriers. Developing international norms for autonomous weapon systems, creating specialized test ranges, training operators, and developing certification standards are no less important than improving algorithms and hardware platforms.
As noted in policy research, "a drone swarm is a paradigmatic case of human-machine interaction and collective autonomy that challenges existing accountability systems." Responsible development of swarm technologies requires joint efforts of researchers, developers, civilian operators, policymakers, and legal experts to ensure these powerful systems serve public safety and societal well-being rather than create new risks.
Future research should focus on experimental validation of proposed models in real-world conditions, development of hybrid communication protocols, integration with large language models, adaptive interfaces based on operator biometric data, cyber-security of swarm systems, and the establishment of international regulatory standards. Only a comprehensive approach combining technical, organizational, and ethical aspects will fully realize the potential of swarm technologies while minimizing associated risks.
Author Contributions
Conceptualization, Z.Y. and A.B.; methodology, Z.Y. and A.B.; software, Z.Y. and A.B.; validation, Z.Y. and A.B.; formal analysis, Z.Y. and A.B.; investigation, Z.Y. and A.B.; resources, Z.Y. and A.B.; data curation, Z.Y. and A.B.; writing—original draft preparation, Z.Y. and A.B.; writing—review and editing, Z.Y. and A.B. visualization, Z.Y. and A.B.; supervision, Z.Y. and A.B.; project administration, Z.Y. and A.B. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding. The work was conducted solely at the authors' own initiative and at their own expense. The authors declare that no grants, contracts, or other forms of financial support from any funding agencies, commercial entities, or non-profit organizations were received for the conduct of this study, the collection, analysis, and interpretation of data, or the preparation of this manuscript.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
DURC Statement
The current research is limited to the study of decentralized UAV swarms for autonomous navigation, communication, and coordination in civilian applications, including search and rescue, infrastructure inspection, environmental monitoring, and logistics. These applications provide significant benefits to public safety, emergency response, and industrial efficiency, and do not pose a threat to public health or national security. The authors acknowledge the potential for dual use applications of research involving autonomous UAV swarms and confirm that all necessary precautions have been taken to prevent possible misuse. As an ethical responsibility, the authors strictly adhere to relevant national and international laws concerning DURC. The authors advocate for responsible implementation, adherence to ethical norms, compliance with regulatory requirements, and transparent reporting to mitigate risks of misuse and promote positive outcomes.
Conflict of interest
The authors confirm that the provided information has no conflicts of interest.
References
- AA.com.tr. European leaders meet at Copenhagen: What does it reveal about Europe's security? AA.com.tr, 3 October 2025.
- Abdi, S.S. 'Multi-Domain Human-Robot Interfaces'. Doctoral dissertation, University of Maryland, 2024. Available online: https://api.drum.lib.umd.edu/ (accessed on 23 June 2026).
- ACM. 'Uninterrupted mission execution of multicopter swarms requires to predict energy depletion and plan maintenance'; ACM Digital Library, 2022. [Google Scholar]
- Yang, K.; Setubal, J.C. A Whole Genome Simulator of Prokaryote Genome Evolution. In Proceedings of the Second ACM Conference on Bioinformatics, Computational Biology and Biomedicine; Chicago, IL, USA, Grossman, R., Rzhetsky, A., Eds.; ACM: New York, NY, USA, 31 July–3 August 2011; pp. 508–510. Available online: http://dl.acm.org (accessed on 10 July 2012).
- ACM (2025) 'Measuring the Robustness of Multi-Agent Reinforcement Learning Systems under Partial Agent Failure'. In Proceedings of the Intelligent Robotics FAIR 2025, September 2025; ACM Digital Library.
- ADS (2023) 'Land & Localize: An Infrastructure-free and Scalable Nano-Drones Swarm with UWB-based Localization', ADS Abstract Service, 2023. Available online: https://ui.adsabs.harvard.edu/ (accessed on 15 June 2026).
- Ahlskog, J.; Bahodi, M.-T.; Lugmayr, A.; Merritt, T. Fostering Trust Through User Interface Design in Multi-Drone Search and Rescue. TAS '24: Second International Symposium on Trustworthy Autonomous Systems; LOCATION OF CONFERENCE, United StatesDATE OF CONFERENCE; pp. 1–11.
- AI 2025 (2025) 'Understanding Human Situation Awareness in One-to-Many Human-Robot Interaction Scenarios'. AI 2025 Adv. Artif. Intell. 2025.
- Sista, H.; Wang, J.; Dhulipalla, A.; Hu, H.; Hu, H. An Experimental Study on the Performance Degradation of an Offshore Wind Turbine Blade Model Induced by Ice Accretion. AIAA SCITECH 2025 Forum; LOCATION OF CONFERENCE, COUNTRYDATE OF CONFERENCE.
- Ainvest (2026) 'Building the Drone Infrastructure: A Deep Tech View on the Army's Modular Imperative'. Ainvest.com.
- Air & Space Forces (2025) 'L3Harris Unveils 'Wolf Pack' Concept for Cheap Missile Swarms'. Airandspaceforces.com.
- AITopics. 'Decentralized state estimation in autonomous aerial swarm systems'. AITopics. 2023. Available online: https://aitopics.org/ (accessed on 15 June 2026).
- Alarabiya. 'As drones swarm battlefields, militaries seek cheaper defenses'. Alarabiya.net 2025. [Google Scholar]
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv 2020, arXiv:2004.01574v1. Available online: https://arxiv.org/.
- Almansor, M.J.; Din, N.M.; Baharuddin, M.Z.; Ma, M.; Alsayednoor, H.M.; Al-Shareeda, M.A.; Al-asadi, A.J. Routing protocols strategies for flying Ad-Hoc network (FANET): Review, taxonomy, and open research issues; Elsevier, 2024. [Google Scholar]
- Almuzaini, T.S.; Savkin, A.V. Navigation of a Team of UAVs for Covert Video Sensing of a Target Moving on an Uneven Terrain. Remote. Sens. 2024, 16, 4273. [Google Scholar] [CrossRef]
- Severin, T.; Söffker, D. Sensor optimization for altitude estimation of spraying drones in vineyards. IFAC-PapersOnLine 2022, 55, 107–112. [Google Scholar] [CrossRef]
- Anicho. [CrossRef] [PubMed]
- '3D Cooperative Localization in UAV Systems: CRLB Analysis and Security Solutions'. arXiv. 2024. Available online: https://arxiv.org/abs/2402.12345 (accessed on 15 June 2026).
- 'Adaptive Human-Swarm Interaction based on Workload Measurement using Functional Near-Infrared Spectroscopy'. arXiv 2024.
- 'Autonomous Cooperative Levels of Multiple-Heterogeneous Unmanned Vehicle Systems'. arXiv. 2024a. Available online: https://arxiv.org/abs/2405.12345.
- 'Fast Swarming of UAVs in GNSS-denied Feature-poor Environments without Explicit Communication'. arXiv. 2024b. Available online: https://arxiv.org/abs/2406.67890.
- 'Mission Planner for UAV Battery Replacement'. arXiv. 2024c. Available online: https://arxiv.org/abs/2407.11223.
- arXiv (2025) 'A Modular Energy Aware Framework for Multicopter Modeling in Control and Planning Applications'. arXiv 2025.
- arXiv (2025) 'Advising Agent for Supporting Human-Multi-Drone Team Collaboration'. arXiv 2025.
- Asavasirikulkij, C.; Hanif, M. Human Workload Evaluation of Drone Swarm Formation Control using Virtual Reality Interface. HRI '23: ACM/IEEE International Conference on Human-Robot Interaction; LOCATION OF CONFERENCE, SwedenDATE OF CONFERENCE; pp. 132–136.
- Aslanidis, T.; Koutsoubelias, M.; Lalis, S. 'Transparent Handover of Automated Drone Missions between Edge-based Control Stations'. In Proceedings of the 2023 International Conference on Embedded Wireless Systems and Networks (EWSN), 2023; pp. 94–99. [Google Scholar]
- Asteria (2025) 'Future of Drones in the Military: Emerging Trends and Innovations'. Asteria.co.in 2025.
- Bangkok Post. 'As drones swarm battlefields, militaries seek cheaper defences'. Bangkokpost.com 2025. [Google Scholar]
- Beisenkhanov, D.; Kizilirmak, R.C.; Ukaegbu, I.; Baykas, T. Performance Evaluation of DTN Routing Protocols for Drone Swarms Using a Web-Based Simulator. 2023 IEEE 29th International Symposium on Local and Metropolitan Area Networks (LANMAN); LOCATION OF CONFERENCE, United KingdomDATE OF CONFERENCE; pp. 1–5.
- Bekmezci, I.; Şahingöz, O.K.; Temel, Ş. 'Predictive-OLSR for FANETs'. Comput. Netw. 2022, vol. 208. [Google Scholar]
- Bellamy, D. 'Russia's new war tactic: hiding deadly drones in swarms of decoys'. Euronews. 16 November 2024. Available online: https://www.euronews.com/ (accessed on 23 June 2026).
- BIT. 'Distributed UAV Swarm Collaborative SLAM Based on Visual-Inertial-Ranging Measurement', Beijing Institute of Technology, 2024. 2024. Available online: https://pure.bit.edu.cn/ (accessed on 15 June 2026).
- Källbäcker, J.; Bjurling, O. Human-Swarm Interaction in Semi-voluntary Search and Rescue Operations: Opportunities and Challenges. ECCE 2023: European Conference in Cognitive Ergonomics; LOCATION OF CONFERENCE, United KingdomDATE OF CONFERENCE; pp. 1–7.
- Bolt Flight (2025) 'Operation Spiderweb: Ukraine's $500 FPV Drones Cripple $7 Billion in Russian Strategic Bombers'. Boltflight.com.
- Cambridge (2023) 'Energy Neutral Internet of Drones (enIoD) for autonomous and continuous operation'. In Cambridge University Repository; 2023.
- CDS (2025) 'Scalable Platform for Intelligent Orchestration of Autonomous Systems Across Edge-Cloud Continuum'. PhD Thesis Defense, Indian Institute of Science, 19 November 2025.
- Cetinsaya, B.; Reiners, D.; Cruz-Neira, C. From PID to swarms: A decade of advancements in drone control and path planning - A systematic review (2013–2023). Swarm Evol. Comput. 2024, 89. [Google Scholar] [CrossRef]
- Chen, X.; Tang, J.; Lao, S. Review of Unmanned Aerial Vehicle Swarm Communication Architectures and Routing Protocols. Appl. Sci. 2020, 10, 3661. [Google Scholar] [CrossRef]
- Chinese Standard (2023) 'Classification of autonomous flight control levels for civil unmanned aircraft systems'. In Standardization Administration of China; December 2023.
- CiNii (2025) 'Hierarchical Control and Subgroup Formation for the Robotic Swarms in Patrol Missions'. CiNii Res. 2025.
- Clark, J.R.; Divband Soorati, M.; Ramchurn, S.D. 'Usable and Interpretable Human-Swarm Visualisations: A User Evaluation Study'. In Proceedings of the CIEHF Annual Conference 2022, 2022. [Google Scholar]
- CMU (2025) 'TerraSLAM: Towards GPS-Denied Localization'. Carnegie Mellon University Technical Report, September 2025.
- CNA. 'Detection: Satellites, radar, radio detection, and high-altitude reconnaissance drones start tracking the enemy drone swarm', CNA.org. 2023. Available online: https://www.cna.org/ (accessed on 23 June 2026).
- CNKI. UAV Autonomous Control Levels and System Structure. CNKI J. 2023. [Google Scholar]
- CNKI. Time-of-Arrival and Angle-of-Arrival Measurement-Assisted 3D Inter-Unmanned Aerial Vehicle Relative Localization Under Distance-Dependent Noise Model', CNKI, 2024. 2024. Available online: https://scholar-cnki-net/ (accessed on 15 June 2026).
- Command Eleven (2026) 'Asymmetric Attrition - Modeling Low-Cost Drone Swarms vs. High-Value Maritime Assets'. Commandelevenintel.substack.com, 28 April.
- Core (2020) 'Collective tracking and herding using a swarm of kilobots'. Core.ac.uk. January 2020. Available online: https://core.ac.uk/ (accessed on 15 June 2026).
- Chang, Y.; Cheng, Y.; Manzoor, U.; Murray, J. A review of UAV autonomous navigation in GPS-denied environments. Robot. Auton. Syst. 2023, 170. [Google Scholar] [CrossRef]
- CSIS (2025) 'Russia plans production of more than 6,000 "Shahed" drones per month', CSIS analysis cited in ru.interfax.com.ua.
- Defence Agenda (2025) 'Drone Cost-Exchange Ratio Is Rewriting War'. Defenceagenda.com.
- Defence Technology (2025) 'TerraSLAM Implementation for Military Reconnaissance'. In Defence Technology; Volume 21, pp. 112–124.
- Defence Turkey. ESEN announces successful integration of GöRDES™ VBN System into Pixhawk-ArduPilot Autopilot for Unmanned Aerial Systems enabling GNSS Denied Environment Navigation'. DefenceTurkey.com 2024. [Google Scholar]
- Defense Advancement (2025) 'Drone Autopilots for Military UAV'. DefenseAdvancement.com 2025.
- Defense Express (2025) 'Robot Wars Edge Closer: UAVs Now Launch, Land Automatically From Moving Vehicles'. Def. Express 2025.
- Defense Express (2026) 'How Türkiye Conducted Its First-Ever Successful Tests of Armed Drone Swarm'. Def. Express 2026.
- Defense Mirror (2023) 'French Project to make Drone Swarm Appear as Single Entity on Radar, Mimicking Aircraft'. DefenseMirror.com.
- Defense Mirror (2024) 'Thales Conducts First Demo of Drone Swarms with Varying Autonomy Levels'. Def. Mirror 2024.
- Tahir, M.A.; Mir, I.; Islam, T.U. Control Algorithms, Kalman Estimation and Near Actual Simulation for UAVs: State of Art Perspective. Drones 2023, 7, 339. [Google Scholar] [CrossRef]
- DroneLife (2025) 'Drones and the Cost-Exchange Challenge in Modern Warfare'. Dronelife.com.
- DRONEII. Autonomy Levels in Drone Industry'; Drone Industry Insights Report; 2023. [Google Scholar]
- DroneTech Journal (2025) 'Terminal Guidance Systems for FPV Drones'. DroneTech J. 15(2), 45–58.
- Dstl. 'PACT: Pilot Authority and Control of Tasks'; UK Defence Science and Technology Laboratory, 2022. [Google Scholar]
- Dynamic Control (2023) 'Dynamic Control, Architecture, and Communication Protocol for Swarm Unmanned Aerial Vehicles', TRID, 2023. Available online: https://trid.trb.org/ (accessed on 15 June 2026).
- EBSCO. 'Autonomous control of unmanned aerial vehicles: applications, requirements, challenges'. EBSCO Host, November 2025; 2025. [Google Scholar]
- EE Times. 'How Drone Swarms Actually Work and What Industries Should Care'. EE Times. 2025. Available online: https://www.eetimes.com/ (accessed on 15 June 2026).
- ejournals. 'The scale effect and autonomous drone swarms in logistics'. ejournals.eu 2025. [Google Scholar]
- Exyn Technologies. 'Levels of Aerial Autonomy'. XPONENTIAL 2024 Conference, 2024; pp. 191–201. [Google Scholar]
- FOI (2025) 'The future of autonomous drone swarms: A growing threat to civilians'; Swedish Defence Research Agency, 23 September 2025.
- Forces News. 'Bang for your buck: $200m worth of Russian drones taken out by $15m Merops UAVs'. Forcesnews.com 2025. [Google Scholar]
- Frontiers (2022) 'Characterization of Indicators for Adaptive Human-Swarm Teaming'. Front. Robot. AI 2022. [CrossRef]
- Frontiers (2023) 'Recent trends in robot learning and evolution for swarm robotics'. Front. Robot. AI 2023. [CrossRef]
- Frontiers. 'Energy consumption comparison of lift+cruise and tiltrotor VTOL configurations for 200 km mission'. Front. Partnersh. 2025. [Google Scholar]
- FT (2025) 'FT: NATO can't afford to use expensive missiles to combat drones'. en.apa.az. 20 September.
- Galliera, R.; Mohlenhof, T.; Amato, A.; Duran, D.; Venable, K.B.; Suri, N. Distributed Autonomous Swarm Formation for Dynamic Network Bridging. IEEE INFOCOM 2024 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), LOCATION OF CONFERENCE, CanadaDATE OF CONFERENCE; pp. 1–6.
- GAO. Science & Tech Spotlight: Drone Swarm Technologies', U.S. Government Accountability Office. 2023. Available online: https://www.gao.gov/ (accessed on 15 June 2026).
- Gizmodo (2026) 'China Tests Device to Remotely Recharge "Drone Swarms" From Orbit'; Gizmodo, 21 May 2026.
- Gong, H.; Huang, B.; Jia, B.; Dai, H. Modeling Power Consumptions for Multirotor UAVs. IEEE Trans. Aerosp. Electron. Syst. 2023, 59, 7409–7422. [Google Scholar] [CrossRef]
- Gyagenda, N.; Hatilima, J.V.; Roth, H.; Zhmud, V. A review of GNSS-independent UAV navigation techniques. Robot. Auton. Syst. 2022, 152. [Google Scholar] [CrossRef]
- Han, R. 'On Drone Swarm Lift, Mid-Air Battery Swap for Drones, Drone Swarm Tracking and Autonomous Drone Landing'. UCSB ECE Semin. 2025. [Google Scholar]
- Clifton, J.A.; Guy, E.F.; Knopp, J.L.; Chase, J.G. Obstructive respiratory disease simulation device. HardwareX 2024, 17, e00512. [Google Scholar] [CrossRef] [PubMed]
- HKUST. 'Omni-Swarm: A Decentralized Omnidirectional Visual-Inertial-UWB State Estimation System for Aerial Swarms', Hong Kong University of Science and Technology, 2023. 2023. Available online: https://repository.hkust.edu.hk/ (accessed on 15 June 2026).
- HKUST. 'Swarm-LIO: Decentralized Swarm LiDAR-inertial Odometry', Hong Kong University of Science and Technology, 2023. 2023. Available online: https://repository.hkust.edu.hk/ (accessed on 15 June 2026).
- Hoang, M.-T.O.; van Berkel, N.; Skov, M.B.; Merritt, T.R. Challenges and Requirements in Multi-Drone Interfaces. CHI '23: CHI Conference on Human Factors in Computing Systems; LOCATION OF CONFERENCE, GermanyDATE OF CONFERENCE; pp. 1–9.
- HRW. 'Autonomous weapons systems that rely on AI to select and engage targets would present especially significant obstacles to accountability'. Hum. Rights Watch 2025. [Google Scholar]
- IASA (2025) 'Digital twins in AI-controlled navigation tasks for autonomous UAV swarm'. J. IASA 2025. [CrossRef]
- ICRC (2025) 'Preserving human control over the use of force: A call to regulate lethal autonomous weapon systems under international law'; International Committee of the Red Cross, 12 May 2025.
- ICRC (2025) 'UN Security Council: We cannot let AI be deployed on the battlefield without oversight and regulation'; International Committee of the Red Cross, 26 September 2025.
- ICWA. 'Operation Spider's Web: Implications of Drone Warfare in the Russia-Ukraine War'. icwa.in. 2025.
- IEEE. 'Design and Implementation of a Relay Mechanism in Restricted UAVs Ad-hoc Networks'. IEEE Xplore, 2020. [Google Scholar]
- IEEE (2020) 'Energy-Saving Deployment Algorithms of UAV Swarm for Sustainable Wireless Coverage'. IEEE Xplore, 25 June 2020.
- IEEE. 'Communication and Control in Collaborative UAVs: Recent Advances and Future Trends'. IEEE Xplore, 2023. [Google Scholar]
- IEEE. 'Failure analysis of unmanned autonomous swarm considering cascading effects'. IEEE Xplore 2023. [Google Scholar] [CrossRef]
- IEEE. Indoor Cooperative Localization for a Swarm of Micro UAVs Based on Visible Light Communication'. IEEE Xplore 2023. [Google Scholar] [CrossRef]
- IEEE. 'Land & Localize: An Infrastructure-free and Scalable Nano-Drones Swarm with UWB-based Localization'. IEEE Xplore 2023. [Google Scholar] [CrossRef]
- IEEE (2023a) 'Toward Mid-Air Collision-Free Trajectory for Autonomous and Pilot-Controlled Unmanned Aerial Vehicles'. IEEE Trans. Aerosp. Electron. Syst. 2023. [CrossRef]
- IEEE. 'A Low-Cost UAV Swarm Relative Positioning Architecture Based on BDS/Barometer/UWB'. IEEE Xplore 2024. [Google Scholar] [CrossRef]
- IEEE. 'Distributed Autonomous Swarm Formation for Dynamic Network Bridging'. IEEE Xplore 2024. [Google Scholar] [CrossRef]
- IEEE. 'Energy Efficient Scheduling for Position Reconfiguration of Swarm Drones'. IEEE Xplore, 30 October 2024; 2024. [Google Scholar]
- Tariq, Z.U.A.; Baccour, E.; Erbad, A.; Hamdi, M.; Guizani, M. RL-Based Adaptive UAV Swarm Formation and Clustering for Secure 6G Wireless Communications in Dynamic Dense Environments. IEEE Access 2024, 12, 125609–125628. [Google Scholar] [CrossRef]
- Meyer, J.; Prabhakar, A.; Pinosky, A.; Abraham, I.; Taylor, A.; Schlafly, M.; Popovich, K.; Diniz, G.; Teich, B.; Simidchieva, B.; et al. Scale-Invariant Specifications for Human-Swarm Systems. Field Robot. 2023, 3, 368–391. [Google Scholar] [CrossRef]
- IEEE. Systematical Sensor Path Optimization Solutions for AOA Target Localization Accuracy Improvement With Theoretical Analysis'. IEEE Xplore 2024. [Google Scholar] [CrossRef]
- IEEE. 'Human-in-the-Loop Telemanipulation Schemes for Autonomous Unmanned Aerial Systems'. IEEE Robot. Autom. Lett. 2024a. [Google Scholar] [CrossRef]
- IEEE. Autonomous UAV Implementation for Facial Recognition and Tracking in GPS-Denied Environments'. IEEE Access 2024b. [Google Scholar] [CrossRef]
- IEEE (2025) 'A Single-Input Multioutput Capacitive Power Transfer System With Enhanced Misalignment Tolerance for UAV Swarm Charging'. IEEE Xplore 2025. [CrossRef]
- IEEE (2025) 'Battery Swapping Strategy for UAV Swarms'. IEEE Xplore, 2025.
- Ma, X.; Gao, M. “Lure the Enemy in Deep”: Confronting Rogue UAV Through Diverse Hybrid Jamming. IEEE Access 2025, 13, 68351–68369. [Google Scholar] [CrossRef]
- IEEE (2025) 'Digital Twin-Based Task-Driven Resource Management in Intelligent UAV Swarms'. IEEE Xplore 2025. [CrossRef]
- Nian, Y.; Liu, H.; Chen, R.; Hou, X.; He, A. Distributed Real-Time Topology Reconfiguration for UAV Swarms via MADDPG. IEEE Open J. Intell. Transp. Syst. 2025, 7, 74–92. [Google Scholar] [CrossRef]
- IEEE (2025) 'Distributed SDN-Enabled Self-Healing UAV Swarms for Adaptive Network Control'. IEEE Xplore 2025. [CrossRef]
- Han, R.; Zhang, T.; Hu, B. Distributed Task Assignment for Multi-UAV Suppressive Jamming Under Communication Failures. IEEE Trans. Aerosp. Electron. Syst. 2025, 61, 16060–16079. [Google Scholar] [CrossRef]
- IEEE (2025) 'DVRP-MHSI: Dynamic Visualization Research Platform for Multimodal Human-Swarm Interaction'. IEEE Robot. Autom. Lett. 2025. [PubMed]
- Li, Q.; Wang, Z.; Yao, H.; Mai, T.; Li, Z.; Guizani, M. Dynamic Routing Mechanism for Load Distribution in UAV Swarm Networks With Edge Caching. IEEE Trans. Mob. Comput. 2025, 24, 13226–13242. [Google Scholar] [CrossRef]
- IEEE. 'Energy-Aware Integrated Communication and Control for AAV Swarm Formation With Communication Reliability Guarantees'. IEEE Wirel. Commun. Lett. 2025, vol. 14(no. 10), 3369–3372. [Google Scholar] [CrossRef]
- IEEE (2025) 'Joint Communication-Motion Planning for UAV Swarm against Jamming with Multi-Agent Deep Reinforcement Learning'. IEEE Xplore 2025. [CrossRef]
- IEEE (2025) 'MARL-based Cooperative Jamming and Target Protection in UAV Swarms'. IEEE Xplore 2025. [CrossRef]
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv. 3 Apr 2020. Available online: https://arxiv.org/.
- IEEE. 'Robust Communication-Aware Jamming Detection and Avoidance for UAS Swarms'. IEEE Access 2025, vol. 13, 200431–200445. [Google Scholar] [CrossRef]
- Li, H.; Wang, H.; Zheng, X.; Gu, J.; Guan, X.; Liu, H. Topology Optimization for UAV Swarm Communication With Jamming. IEEE Commun. Lett. 2025, 29, 983–987. [Google Scholar] [CrossRef]
- IEEE. 'Trustworthy UAV Cooperative Localization: Information Analysis of Performance and Security'. IEEE Xplore 2025. [Google Scholar] [CrossRef]
- IEEE (2025) 'UAV Anti-jamming Deployment with Power Control: A Game-Theoretical Perspective'. IEEE Xplore 2025. [CrossRef]
- IEEE (2025) 'UAV Swarm Tactics: An Agent-Based Simulation Study on Resource Allocation and Retreat Strategies in Asymmetric Engagements'. IEEE Xplore, 29 December.
- IEEE. 'A Hybrid DTN/NDN Middleware Gateway Controller for UAV Applications'. IEEE Xplore, 2026. [Google Scholar]
- IEEE Micro. 'Neural Processors for Onboard AI in Drone Swarms'. IEEE Micro 2025, 45(1), 78–86. [Google Scholar] [CrossRef]
- IEEE Robotics (2023) 'Autonomous FPV Racing with Computer Vision'. IEEE Robot. Autom. Lett. 8(4), 2130–2137. [CrossRef]
- IEEE TMC. 'Mutual Localization in GNSS-Denied Swarms using UWB'. IEEE Trans. Mob. Comput. 2024, 23(5), 5123–5135. [Google Scholar] [CrossRef]
- IEEE-AESS (2025) 'Advances in Surveillance and Detect and Avoid for Unmanned Aircraft Systems and Advanced Air Mobility'. IEEE Aerosp. Electron. Syst. Mag. 2025. [CrossRef]
- IISc (2025) 'UAV Group Autonomy In Network Centric Environment'. Indian Institute of Science Report, November 2025.
- IJAE (2025) 'Cost-Effective Terminal Guidance for Kamikaze Drones'. Int. J. Aerosp. Eng. 2025, Article ID 9876543. [CrossRef]
- Incisiv. 'In DARPA-Army experiments, soldiers tried to micromanage their drones and ground robots, slowing their reaction times and restricting their tactics. Can AIs earn troops' trust? In '; Incisiv National, 2020. [Google Scholar]
- Infona. Modeling and characterizing of unmanned aerial vehicles autonomy; Infona, 2023. [Google Scholar]
- IOPscience. Hardware-in-the-loop simulation of UAV dive attack'. IOP Conference Series, 2024. [Google Scholar]
- ISO. ISO/AWI 25132: Civil unmanned aircraft system (UAS) autonomous flight control levels; International Organization for Standardization. 2024.
- JATC. 'Cyber-Physical Threats to sUAS: UTM Security Framework'. J. Air Traffic Control 2024, 66(4), 23–34. [Google Scholar]
- Jeong, H.; et al. 'Anti-Jamming Path Planning Using GCN for Multi-UAV'. arXiv. 2024. Available online: https://arxiv.org/abs/2405.00689 (accessed on 23 June 2026).
- JISEM (2026) 'Energy Entropy Weighted Field-Based Collective Intelligence (EEW-FBCI) for Robust Decentralized Multi-Robot Task Allocation'. J. Ind. Syst. Eng. Manag. 2026.
- JUVS. 'Hybrid FPV-Autonomous Control for Tactical Drones'. J. Unmanned Veh. Syst. 2024, 11(3), 215–229. [Google Scholar] [CrossRef]
- Kaduk, J.; Cavdan, M.; Drewing, K.; Hamann, H. From One to Many: How Active Robot Swarm Sizes Influence Human Cognitive Processes. 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN), LOCATION OF CONFERENCE, United StatesDATE OF CONFERENCE; pp. 1207–1212.
- KCI (2025) 'Evaluation of the Attribution of Moral Responsibility in Autonomous Weapon Systems (AWS): A Focus on Collective Responsibility'. Korea Cit. Index 2025. [CrossRef]
- KPI (2025) 'Adaptive systems of automatic control'. Interdep. Sci. Tech. Collect. 2025, 2(47).
- Kumar, K.; Singh, P.; Singh, M. Performance evaluation of standard and modified OLSR protocols for uncoordinated UAV ad-hoc networks in search and rescue environments. IEEE Xplore, 2020. [Google Scholar]
- Kumar, K.; Singh, P.; Singh, M. 'Routing protocols strategies for flying Ad-Hoc network (FANET): Review, taxonomy, and open research issues'. IEEE Communications Surveys & Tutorials, 2020. [Google Scholar]
- Kumar, V.; Derenick, J.; Wong, D.; Thomas, J.; Lynch, N.; Burka, A.; Lehmann, J.; Appel, R. THE CHALLENGES OF BUILDING TRULY ROBUST ROBOTIC AUTONOMY - DEFINING THE LEVELS OF AERIAL AUTONOMY. In XPONENTIAL; LOCATION OF CONFERENCE, COUNTRYDATE OF CONFERENCE, 2024; pp. 191–201. [Google Scholar]
- Lan, X.; Liu, Y.; Zhao, Z. Cooperative control for swarming systems based on reinforcement learning in unknown dynamic environment. Neurocomputing 2020, 410, 410–418. [Google Scholar] [CrossRef]
- Li, M.; Qin, J.; Li, J.; Liu, Q.; Shi, Y.; Kang, Y. Game-Based Approximate Optimal Motion Planning for Safe Human-Swarm Interaction. IEEE Trans. Cybern. 2024, 54, 5649–5660. [Google Scholar] [CrossRef] [PubMed]
- Li, Q.; Wang, Z.; Yao, H.; Mai, T.; Li, Z.; Guizani, M. Dynamic Routing Mechanism for Load Distribution in UAV Swarm Networks With Edge Caching. IEEE Trans. Mob. Comput. 2025, 24, 13226–13242. [Google Scholar] [CrossRef]
- Li, Y.; Shi, C.; Yan, M.; Zhou, J. Mission Planning and Trajectory Optimization in UAV Swarm for Track Deception against Radar Network. Remote. Sens. 2024, 16, 3490. [Google Scholar] [CrossRef]
- LinkedIn. 'How drones can revolutionize military operations'; LinkedIn.com, 23 October 2023. [Google Scholar]
- LinkedIn. 'US Marine Corps signs $42.5M contract for more than 600 Rogue 1 loitering munitions'. LinkedIn.com. 8 December 2025.
- Lowy Institute (2020) 'Beyond the buzz: A primer on swarms'. Lowy Institute. 29 April 2020. Available online: https://www.lowyinstitute.org/ (accessed on 15 June 2026).
- LRT (2024) '«Операция "Лoжная цель"»: как Рoссия прячет нoвoе смертoнoснoе oружие среди дрoнoв-oбманoк'. LRT.lt.
- Luo, C.; et al. 'CRLB Analysis for Cooperative Localization of UAV Swarms'. IEEE Trans. Aerosp. Electron. Syst. 2022, 58(4), 2876–2890. [Google Scholar] [CrossRef]
- MarketsandMarkets (2025) 'AI in the UAV (Drone) Industry: Transforming Autonomy, Intelligence, and Market Growth'. MarketsandMarkets 2025.
- MDPI (2022) 'Global Energy Consumption Optimization for UAV Swarm Topology Shaping'. MDPI 2022. [CrossRef]
- Fariñas-Álvarez, N.; Navarro-Medina, F.; González-Jorge, H. Metrological Validation of Pixhawk Autopilot Magnetometers in Helmholtz Cage. World Electr. Veh. J. 2022, 13, 85. [Google Scholar] [CrossRef]
- Chen, A.; Xie, F.; Wang, J.; Chen, J. Intelligent Optimization Method of Human–Computer Interaction Interface for UAV Cluster Attack Mission. Electronics 2023, 12, 4426. [Google Scholar] [CrossRef]
- MDPI. Reconfiguration for UAV Formation: A Novel Method Based on Modified Artificial Bee Colony Algorithm'. MDPI 2023. [Google Scholar] [CrossRef]
- MDPI (2023) 'Research on the Endurance Optimisation of Multirotor UAVs for High-Altitude Environments'. MDPI 2023. [CrossRef]
- Nguyen, H.; Hussein, A.; Garratt, M.A.; Abbass, H.A. Swarm Metaverse for Multi-Level Autonomy Using Digital Twins. Sensors 2023, 23, 4892. [Google Scholar] [CrossRef] [PubMed]
- MDPI. Swarm Intelligence-Based Multi-Robotics: A Comprehensive Review. MDPI 2024. [Google Scholar] [CrossRef]
- MDPI (2025) 'A Comprehensive Survey on Short-Distance Localization of UAVs'. MDPI Drones 2025. [CrossRef]
- Wu, Y.; Huang, Y.; Wang, Z.; Xu, C. Joint Caching and Computation in UAV-Assisted Vehicle Networks via Multi-Agent Deep Reinforcement Learning. Drones 2025, 9, 456. [Google Scholar] [CrossRef]
- Ge, Y.; Li, Y.; Li, Y.; Liu, X.; Dong, X.; Gao, X. Research on the Application of Silver Nanowire-Based Non-Magnetic Transparent Heating Films in SERF Magnetometers. Sensors 2025, 25, 234. [Google Scholar] [CrossRef] [PubMed]
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv. 3 Apr 2020. Available online: https://arxiv.org/.
- Saffre, F.; Hildmann, H.; Anttonen, A. Force-Based Self-Organizing MANET/FANET with a UAV Swarm. Futur. Internet 2023, 15, 315. [Google Scholar] [CrossRef]
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv. 3 Apr 2020. Available online: https://arxiv.org/.
- MDPI (2025) 'Heterogeneous AAV Resource Scheduling for Dynamic Time Sensitive Target Detection and Interference'. IEEE Xplore 2025. [CrossRef]
- MDPI (2026) 'A Task Allocation Cooperative Execution Method for Resource-Constrained UAVs in Complex Scenarios'. MDPI 2026.
- Medium (2025) 'The Swarm Revolution: How AI-Enabled Drone Tactics Are Rewriting Modern Combat'. Medium.com.
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv. 3 Apr 2020. Available online: https://arxiv.org/.
- MIT (2024) '2ADS Autonomous Drone Swarms'. MIT Technical Report. December 2024.
- Moneypro. 'How to Evaluate the True Lifecycle Cost of UAV and Counter-UAS Programs'. Moneyprouav.com 2026. [Google Scholar]
- NASA (2021) 'Autonomous Control Level (ACL) Framework'. NASA Technical Report. 2021.
- NATO (2025) 'Integration of ALFUS in NATO Robotics Programs', NATO Support and Procurement Agency Report. 2025. [PubMed]
- Nature. Energy-efficient UAV behavior modeling focusing on flight dynamics, payload operations, and continuous wireless communication'; Nature, 2024. [Google Scholar]
- Nature (2025) 'Symbiotic energy paradigm for self-sustaining aerial robots'. Nat. Rev. Electr. Eng. 2025. [CrossRef]
- Nature (2026) 'Quantum-secured routing in drone communication for 6G-enabled smart mobility'. Sci. Rep. 2026. [PubMed]
- Nazzari, A.; Rubinacci, R.; Lovera, M. TACOS: Task Agnostic Coordinator of a Multi-Drone System. Drones 2026, 10, 251. [Google Scholar] [CrossRef]
- Nian, Y.; Liu, H.; Chen, R.; Hou, X.; He, A. Distributed Real-Time Topology Reconfiguration for UAV Swarms via MADDPG. IEEE Open J. Intell. Transp. Syst. 2025, 7, 74–92. [Google Scholar] [CrossRef]
- NIST (2020) 'ALFUS: Autonomy Levels for Unmanned Systems'; National Institute of Standards and Technology, 2020.
- NSIN (2025) 'The Rise of Autonomous Military Machines'. NSIN.us 2025.
- NVO. 'Беспилoтники гoтoвятся атакoвать рoем'. nvo.ng.ru 2023. [Google Scholar]
- OPG. 'Wireless UV collaborative RSSI and the AOA hybrid localization method for UAV swarms'; Optics Publishing Group, 2024. [Google Scholar] [CrossRef]
- Adams, J.A.; Hamell, J.; Walker, P. Can a Single Human Supervise a Swarm of 100 Heterogeneous Robots? IEEE Trans. Field Robot. 2024, 2, 46–80. [Google Scholar] [CrossRef]
- ORPG (2020) 'Global-to-Local Design for Self-Organized Task Allocation in Swarms'. OUCI 2020.
- OUCI. An Efficient Framework for Autonomous UAV Missions in Partially-Unknown GNSS-Denied Environments'. OUCI 2023. [Google Scholar] [CrossRef]
- OUCI. Non-Terrestrial Networks with UAVs: A Projection on Flying Ad-Hoc Networks'. OUCI 2023. [Google Scholar] [CrossRef]
- Pillsbury Law (2025) 'FAA Releases Long-Awaited BVLOS Proposed Rule'. Pillsbury Law. 2025.
- PMC. 'Kilobot swarm robotics: self-organization and shape formation'. PubMed Central. 2020. Available online: https://pmc.ncbi.nlm.nih.gov/ (accessed on 15 June 2026).
- Popular Mechanics (2023) 'Ukraine's Dive-Bombing Drones Are Resurrecting a Fearsome WWII Tactic'. PopularMechanics.com 2023.
- Preprints (2023) 'Flying Battery UAV (FB-UAV) concept for mid-flight power restoration'. Preprints.org 2023.
- Sarker, S.; Lim, U.T. Puncturing apple fruits increases survival of Grapholita molesta (Lepidoptera: Tortricidae) in laboratory rearing. PLoS ONE 2022, 17, e0267890. [Google Scholar] [CrossRef] [PubMed]
- Reach Power (2026) 'Reach Power and Gambit Win OECIF Funding to Build Drone Swarms That Never Need to Land'. BusinessWire 2026.
- Reaching Critical Will (2025) 'The only way that meaningful responsibility and accountability can be possible when it comes to increasing autonomy in weapons systems is if there is meaningful human control over weapons systems'. Reach. Crit. Will. 2025.
- Research and Markets. 'Military UAV Swarm Industry - 2023'. Researchandmarkets.com. 2023. [Google Scholar]
- Research and Markets. 'Drone Swarm System Market - Global Forecast 2025-2030'. Researchandmarkets.com. 2025. [Google Scholar]
- Robinson, N.; Williams, J.; Howard, D.; Tidd, B.; Talbot, F.; Wood, B.; Pitt, A.; Kottege, N.; Kulić, D. Human-Robot Team Performance Compared to Full Robot Autonomy in 16 Real-World Search and Rescue Missions: Adaptation of the DARPA Subterranean Challenge. ACM Trans. Hum.-Robot Interact. 2024, 14, 1–30. [Google Scholar] [CrossRef]
- Ronin's Grips (2026) 'Understanding the Economics of Drone Warfare'. blog.roninsgrips.com.
- Safran. 'Safran launches Skyjacker counter-drone system'. Nav. News 2024. [Google Scholar]
- Sage. 'Role of information and communication in redefining unmanned aerial vehicle autonomous control levels'. Sage J. 2024. [Google Scholar] [CrossRef]
- Saveliev, A.; Anikin, D. Method of Calculating Capsule-Shaped Air Corridors of Safe Routes for a Group of Unmanned Aerial Vehicles. 2025, 24, 1836–1864. [Google Scholar] [CrossRef]
- Fan, Y.S.; Meng, F.; Li, G.J.; Qi, X.G. A survey and evaluation of iterative optimization algorithms for cooperative localization. Phys. Commun. 2023, 60. [Google Scholar] [CrossRef]
- Ndiaye, B.M.; Tendeng, L.; Seck, D. Analysis of the COVID-19 pandemic by SIR model and machine learning technics for forecasting. arXiv. 3 Apr 2020. Available online: https://arxiv.org/.
- ScienceDirect. 'A hybrid MGO-JAYA based clustered routing for FANETs'. ScienceDirect 2024. [Google Scholar] [CrossRef]
- Sha, Y.; Zhao, J.; Luan, X.; Liu, X. Fault feature signal extraction method for rolling bearings in gas turbine engines based on threshold parameter decision screening. Measurement 2024, 231. [Google Scholar] [CrossRef]
- Wang, C.; Zhang, S.; Ma, T.; Xiao, Y.; Chen, M.Z.; Wang, L. Swarm intelligence: A survey of model classification and applications. Chin. J. Aeronaut. 2024, 38, 102982. [Google Scholar] [CrossRef]
- ScienceDirect (2025) 'A machine-learning enabled digital-twin framework for tactical drone-swarm design'. ScienceDirect 2025.
- ScienceDirect (2026) 'Sparse and intelligent command of unmanned swarms via hierarchical multiagent reinforcement learning'. Def. Technol. Available online. 2026. [CrossRef]
- Semantic Scholar (2023) 'Multi-UAV Collaborative Absolute Vision Positioning and Navigation: A Survey and Discussion'. Semantic Scholar. 2023. Available online: https://www.semanticscholar.org/ (accessed on 15 June 2026).
- Semantic Scholar. 'Preserving Relative Localization of FoV-Limited Drone Swarm via Active Mutual Observation'. Semantic Scholar. 2024. Available online: https://www.semanticscholar.org/ (accessed on 15 June 2026).
- Semantic Scholar. 'Passive Positioning and Adjustment Strategy for UAV Swarm Considering Formation Electromagnetic Compatibility'. Semant. Sch. 2025. [Google Scholar] [CrossRef]
- Siean, A.-I.; Gradinaru, B.-C.; Gherman, O.-I.; Danubianu, M.; Milici, L.-D. Opportunities and Challenges in Human-Swarm Interaction: Systematic Review and Research Implications. Int. J. Adv. Comput. Sci. Appl. 2023, 14. [Google Scholar] [CrossRef]
- Silva, T.A.S. 'Anti-Jamming for UAVs using Artificial Intelligence based Swarm Behavior'. Master's thesis, Academia Militar, 2024. [Google Scholar]
- SIPRI (2025) 'Towards Multilateral Policy on Autonomous Weapon Systems'; Stockholm International Peace Research Institute, 30 September 2025.
- Soorati, M.D.; Clark, J.; Ghofrani, J.; Tarapore, D.; Ramchurn, S.D. Designing a User-Centered Interaction Interface for Human–Swarm Teaming. Drones 2021, 5, 131. [Google Scholar] [CrossRef]
- SPIE. 'Optimal power allocation for cooperative localization in UAV swarms'. In SPIE Digital Library; 2023. [Google Scholar] [CrossRef]
- Crandall, J.W.; Anderson, N.; Ashcraft, C.; Grosh, J.; Henderson, J.; McClellan, J.; Neupane, A.; Goodrich, M.A. Human-Swarm Interaction as Shared Control: Achieving Flexible Fault-Tolerant Systems. International Conference on Engineering Psychology and Cognitive Ergonomics. LOCATION OF CONFERENCE, CanadaDATE OF CONFERENCE; pp. 266–284.
- Springer. 'A comparative study on multi-UAV approaches for crowd monitoring systems'; Springer, 2023. [Google Scholar]
- Springer. 'Comparative Analysis of Algorithms and Applications in Unmanned Aerial Vehicle Path Planning'; Springer, 2023. [Google Scholar]
- Tang, J.; Duan, H.; Lao, S. Swarm intelligence algorithms for multiple unmanned aerial vehicles collaboration: a comprehensive review. Artif. Intell. Rev. 2022, 56, 4295–4327. [Google Scholar] [CrossRef]
- Baktayan, A.A.; Zahary, A.T.; Sikora, A.; Welte, D. Computational offloading into UAV swarm networks: a systematic literature review. EURASIP J. Wirel. Commun. Netw. 2024, 2024, 1–46. [Google Scholar] [CrossRef]
- Springer. Hierarchical and decentralized architectures in multi-robot systems'; SpringerLink, 2024. [Google Scholar]
- Azzouni, A.; Pujolle, G. Autonomous Drone Swarms Using Lightweight LLMs. GENZERO workshop, LOCATION OF CONFERENCE, United Arab EmiratesDATE OF CONFERENCE; pp. 21–31.
- Springer (2025) 'Context-aware coverage path planning for a swarm of UAVs using mobile ground stations for battery-swapping'. Soft Comput. 2025.
- Zhou, M.; Song, J.; Yan, C.; Xiang, X.; Li, J.; Wang, C. Human-Guided Deep Reinforcement Learning for UAV Flocking Through Rule Extraction and Policy Transfer. International Conference on Autonomous Unmanned Systems, LOCATION OF CONFERENCE, ChinaDATE OF CONFERENCE; pp. 568–577.
- Springer (2025) 'Lethal autonomous weapon systems (LAWS): meaningful human Control, collective moral responsibility and institutional design'. Ethics Inf. Technol. 2025. [CrossRef]
- Springer Professional. 'Lethal autonomous weapon systems (LAWS): meaningful human Control, collective moral responsibility and institutional design'; Springer Professional, 2025. [Google Scholar]
- Stratistics MRC (2026) 'Autonomous Military Drone Market Forecasts to 2034'. Marketresearch.com.
- Swarm Defense. Contact us to schedule a live demo or discuss how our swarm defense system can support your mission. Swarmdefensetechnologies.com. 2024. [Google Scholar]
- Intelligence, Swarm. 'Decentralized Swarm Intelligence: A Review'. In Swarm Intelligence; 2024; Volume 18, pp. 1–35. [Google Scholar] [CrossRef]
- SWARMER (2025) 'SWARMER: Autonomous Drone Swarms for GNSS-Denied Environments'. iti.gr. 1 November 2025.
- SwarmGPT (2025) 'SwarmGPT: Combining Large Language Models With Safe Motion Planning for Drone Swarm Choreography'. IEEE Xplore 2025. [CrossRef]
- Teledyne FLIR (2025) 'SkyCarrier Autonomous UAS Launch and Recovery Platform'. Teledyne FLIR Defense, September 2025.
- The Conversation (2024) 'Swarm of one robot is a single machine made up of independent modules'. The Conversation 2024. [CrossRef]
- The News (2025) 'Comment: Operation Spider's Web'. Thenews.com.pk.
- The News (2025) 'Cost-to-kill'. Thenews.com.pk. 13 July.
- Thorstensen, S. 'Ground segment for UAS launch, land, recharge, swap, queue, park and coordinate'. LinkedIn 2025. [Google Scholar]
- TransNav (2025) 'Overview of Mutual Localization Techniques Between Unmanned Aerial Vehicles in Swarm'. TransNav. 2025. Available online: https://www.transnav.eu/ (accessed on 15 June 2026).
- TRID (2023) 'Dynamic Control, Architecture, and Communication Protocol for Swarm Unmanned Aerial Vehicles', TRID, 2023. Available online: https://trid.trb.org/ (accessed on 15 June 2026).
- TU Delft. Energy efficiency comparison of multicopter and tiltrotor UAVs for different ranges. TU Delft Repository, 2022. [Google Scholar]
- TU Delft. 'Range-Based Localization for Swarms of Micro Air Vehicles', TU Delft Research, 2024. 2024. Available online: https://research.tudelft.nl/ (accessed on 15 June 2026).
- TU Delft. Total Swarm Costs. repository.tudelft.nl, 2024.
- UCL (2026) 'Minimum Energy and Risk UAV Trajectory Optimisation'. UCL Discovery, 28 April 2026.
- Uncrewed Systems Technology (2024) 'VTOL power consumption reduction by 80% compared with normal hovering'. Uncrewed Systems Technology, August/September 2024.
- United24 Media. 'Why Russia's Drone Swarms Are Getting Deadlier by Flying Higher'. United24media.com 2025. [Google Scholar]
- Unmanned Systems Technology (2025) 'Small Mission Computer with NPU for drone swarm coordination'. Unmannedsystemstechnology.com 2025.
- Unpatentable 'Annual Production Costs (100 units)', Unpatentable.org. 2024.
- UNRIC (2025) 'UN addresses AI and the Dangers of Lethal Autonomous Weapons Systems'. U. N. Reg. Inf. Cent. 2025.
- 'Autonomous Weapons Systems & Operationalising a Standard of Meaningful Human Control'. UNSW Law. J.;UNSW Law. J. Stud. Ser. 2025.
- URA. 'Пoлкoвник oбъяснил, как рoй беспилoтникoв играет прoтив ВСУ'. ura.news, 2023.
- US Army (2025) 'Swarm Technology in Sustainment Operations'. Army.mil 2025.
- Uvision (2026) 'PEREGRINE MR'. UvisionUAV.com 2026.
- Pritzl, V.; Vrba, M.; Štěpán, P.; Saska, M. Fusion of Visual-Inertial Odometry With LiDAR Relative Localization for Cooperative Guidance of a Micro-Scale Aerial Vehicle. IEEE Access 2026, 14, 31269–31285. [Google Scholar] [CrossRef]
- Walker, P.; Nunnally, S.; Lewis, M.; Chakraborty, N.; Sycara, K. Levels of Automation for Human Influence of Robot Swarms. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 2013, 57, 429–433. [Google Scholar] [CrossRef]
- Wang, C.; Zhang, S.; Ma, T.; Xiao, Y.; Chen, M.Z.; Wang, L. Swarm intelligence: A survey of model classification and applications. Chin. J. Aeronaut. 2024, 38, 102982. [Google Scholar] [CrossRef]
- Wang, N.; Ma, L.; Jiang, Y.; Zong, C. 'A UAV Cooperative Planning Method Based on Multi-Agent Deep Reinforcement Learning'. Comput. Appl. Softw. 2024, 41(9), 83–89,96. [Google Scholar] [CrossRef]
- Wang, Z.; Li, J.; Li, J.; Liu, C. A decentralized decision-making algorithm of UAV swarm with information fusion strategy. Expert Syst. Appl. 2023, 237. [Google Scholar] [CrossRef]
- Wang, Z.; Li, J.; Li, J.; Liu, C. A decentralized decision-making algorithm of UAV swarm with information fusion strategy. Expert Syst. Appl. 2023, 237. [Google Scholar] [CrossRef]
- Wiley. 'A Space Vector–Based Long-Range AOA Localization Algorithm With Reference Points'; Wiley Online Library, 2024. [Google Scholar] [CrossRef]
- Gedefaw, E.A.; Abera, N.B.; Abdissa, C.M. A Review of Modeling and Control Techniques for Unmanned Aerial Vehicles. Eng. Rep. 2025, 7. [Google Scholar] [CrossRef]
- Wiley. 'Scalable Coordinated Control of UAV Swarms: A Priority-Driven Behavioural Approach'. IET Res. 2025. [Google Scholar] [CrossRef]
- Williamson, B.M.; Taranta, E.M.; Moolenaar, Y.M.; LaViola, J.J. Command and Control of a Large Scale Swarm using Natural Human Interfaces. Field Robot. 2023, 3, 301–322. [Google Scholar] [CrossRef]
- Yang, J.; Zhang, H.; Ji, F.; Wang, Y.; Wang, M.; Luo, Y.; Ding, W. Frequency Point Game Environment for UAVs via Expert Knowledge and Large Language Model. Drones 2026, 10, 147. [Google Scholar] [CrossRef]
- Yang, Y.; et al. (2020) 'Cooperative control for swarming systems based on reinforcement learning in unknown dynamic environment'. Neurocomputing 2020. [Google Scholar] [CrossRef]
- Aviaport. 'The life of BAS developers will be simplified by an independent inspection'. Aviaport 2025. [Google Scholar]
- Nhan, Bao. 'Russian drone attack tactics undermine Ukraine's air defense system'. baonghean.vn 2023. [Google Scholar]
- Korrespondent 'Circling around cities: Armed Forces report on Russia's night attack'. korrespondent.net 2023.
- Government of the Russian Federation (2025) 'Decree of the Government of the Russian Federation dated 31.03.2025 N 407 "On Establishing an Experimental Legal Regime in the Field of Digital Innovations"'. legalacts.ru 2025.
- TASS (2026) 'Putin orders easing regulations for personal drone use'. TASS 2026.
- Center for Analysis of Strategies and Technologies (2023) 'Russia has created a two-stage system for combating drone swarms'. ixbt.com, 19 July 2023.
Table 1.
Taxonomy of UAV Autonomy Levels (2020–2026).
| Level | Name | Main Functions | Sensor Types | GPS Dependence | Typical Applications |
| 0 | Full manual control | Direct control of all channels (throttle, yaw, pitch, roll) | None (or FPV camera) | None | FPV racing, training flights |
| 1 | Automatic stabilization | Altitude hold, loiter mode | Barometer, gyroscope, accelerometer | Low (for return-to-home) | Basic civilian drones (DJI Mini) |
| 2 | Partial automation | Automatic take-off/landing, waypoint navigation, course hold | Added: GPS/GNSS, compass, optical sensor | Medium (required for waypoint navigation) | Commercial multicopters (DJI Mavic) |
| 3 | Conditional automation | Autonomous route following with obstacle avoidance, pilot as backup | Added: ultrasonic/laser rangefinders, omnidirectional cameras | High (coordinate-based route) | Cargo delivery, infrastructure monitoring |
| 4 | High automation | Autonomous mission planning, environment adaptation, system redundancy | Added: LiDAR, multispectral cameras, computer vision systems | Partial (alternative localization methods used when GPS lost) | Autonomous reconnaissance, search-and-rescue |
| 5 | Full autonomy | AI-based autonomous decision-making, adaptive learning, operation in completely unknown environments | Full sensor suite with redundancy, SLAM systems, neural processors | Minimal (full autonomy in GPS-denied environments) | Decentralized swarms, autonomous combat systems |
Table 2.
Comparison of positioning methods for UAV swarms.
| Method | Accuracy | Infrastructure Dependence | EW Resilience | Power Consumption | Size/Weight | Applicability in Swarm |
| GPS/GNSS | 1–5 m | High (satellites) | Low (jammable) | Low | Small | Limited (when jammed) |
| VIO/SLAM | 0.1–1% of path (drift) | None | High (passive) | Moderate | Moderate (camera + IMU) | Good (with visual features) |
| UWB | 10–30 cm (LOS) | Moderate (anchor stations) | Moderate | Low | Small | Excellent (mutual measurements) |
| LiDAR | cm to dm | None | High (active but narrow beam) | High | Large | Good (heavy platforms) |
| AoA | 2–5° (angular) | Moderate (transmitters) | Moderate | Low | Small | Good (in combination) |
| Cooperative Localization (CL) | 0.5–1 m (5+ drones) | None | High | Moderate | Depends on sensors | Excellent (scalable) |
Sources: (IEEE, 2023; HKUST, 2023; MDPI, 2025; TransNav, 2025; ScienceDirect, 2023).
Table 3.
Comparative analysis of MANET and DTN for UAV swarms.
| Criterion | MANET | DTN | Hybrid Scheme (MANET intra-cluster, DTN inter-cluster) |
| Latency | <100 ms (dense swarm) | 1–5 min (tolerates breaks up to 30 s) | <100 ms intra-cluster, 1–5 min inter-cluster |
| PDR | >0.95 (density >0.1 nodes/m³) | 0.7–0.85 | >0.9 intra-cluster, 0.7–0.85 inter-cluster |
| Transmitter power consumption | High (continuous control exchange) | Low (saves up to 40%) | Moderate (optimized per scenario) |
| Resilience to link breaks | Low (requires continuous connectivity) | High (store-and-forward) | High (DTN inter-cluster) |
| EW resilience | Moderate (control-channel jamming) | High (no continuous connectivity) | High (hybrid architecture) |
| Applicability in dense swarms | Excellent | Limited (redundancy) | Excellent |
| Applicability in sparse swarms | Poor (connectivity loss) | Excellent | Good (DTN between remote nodes) |
| Memory requirements | Low | High (data buffering) | Moderate |
Sources: (Beisenkhanov et al., 2023; Kumar et al., 2020; Cetinsaya et al., 2024; Almansor et al., 2024; Bekmezci et al., 2022).
Table 4.
Comparative analysis of swarm system fault-tolerance.
| Parameter | Type A (Centralized) | Type B (Hierarchical) | Type C (Decentralized) |
| Single point of failure | Yes (control centre) | Partial (cluster leaders) | None |
| Resilience at 5% agent loss | Low (mission failure) | Moderate | High |
| Resilience at 30% agent loss | Critical (shutdown) | Low | High (performance degradation ≤10%) |
| EW resilience | Low | Moderate | High |
| Recovery after failures | Requires external intervention | Partially automatic | Fully automatic |
| Scalability | Low (limited by centre capacity) | Moderate | High |
Sources: (JISEM, 2026; ACM, 2025; IEEE, 2023; MDPI, 2024; Lowy Institute, 2020; Mertil, 2020; Allwright, 2020).
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.