Hierarchical UAV Swarm Structure

• Hierarchical UAV swarm structure is an organizing framework that distributes roles, communication, and decision-making across multiple layers in drone networks.
• It employs patterns such as leader–follower, cluster-head, and guide–worker to balance global guidance with local autonomy in sensing and motion control.
• This structure improves scalability, robustness, and performance by decomposing complex missions into manageable tasks while addressing challenges in communication and security.

Hierarchical UAV swarm structure denotes an organization in which roles, communication, and decision processes are distributed across levels rather than treated as a single flat collective. In the literature considered here, hierarchy appears in several non-equivalent forms: a leader–follower swarm with local proximity communication, a two-level cluster-head network, a guide–worker control partition, a layered software stack, and cloud–edge–terminal or UAV–edge–cloud architectures with nested control loops and mission decomposition (Brust et al., 2016, Xiao et al., 2020, Jia et al., 10 Mar 2026, Nguyen et al., 20 Jan 2026). A recurring theme is that hierarchy is often combined with distributed execution rather than replacing it: higher levels supply goals, assignments, or parameters, while lower levels retain local autonomy for motion, sensing, and safety.

1. Forms of hierarchy in UAV swarms

A basic form is the single-tier leader–follower organization. In the forest-assessment swarm model, the swarm communication network is a symmetric Euclidean graph $G=(V,E)$, with $V \subset \mathbb{R}^3$ and edges defined by transmission range $r$; yet the swarm also contains one leader UAV with absolute positioning and destination knowledge and several followers with only relative positioning. The result is hierarchical at the task and role level, but distributed at the communication level because each UAV communicates only with direct neighbors in a 1-hop localized network (Brust et al., 2016).

A stricter two-level hierarchy appears in cluster-head formulations. In the two-level UAV swarm network (USNET), each cluster contains one cluster head UAV (HUAV) and several follower UAVs (FUAVs); FUAVs follow the HUAV according to an inherent follow strategy, while HUAVs act as local control and communication centers. The hierarchy is therefore encoded directly in motion and communication asymmetry rather than in a central planner (Mou et al., 2022).

Other works define hierarchy through role specialization. The guide–worker pattern uses a smaller group of guides with localization, mapping, and planning, and a larger group of workers governed by simple local attraction forces. Guides shape and displace the worker formation by operating on workers’ interaction parameters, yielding a two-level architecture in which the higher layer carries mission knowledge and the lower layer executes local rules (Varadharajan et al., 2022). Closely related arguments are made in the broader robot-swarm study contrasting egalitarian swarms of identical workers with hierarchical swarms composed of guides and workers; there, hierarchy is explicitly associated with extended sensing reach and better performance in larger or more unstructured environments (Varadharajan et al., 2024).

A different usage of hierarchy is architectural rather than organizational. XTDrone defines six layers—communication, simulator, low-level control, high-level control, coordination, and human interaction—so that multi-UAV behavior is factored across infrastructure, control abstraction, swarm coordination, and operator interfaces (Xiao et al., 2020). Cloud-oriented and edge-oriented systems generalize this approach: H-OODA uses cloud, edge, and terminal layers; the ISCCC framework uses BS–swarm, swarm–terminal, and inter-swarm layers; agentic AI architectures distinguish standalone, edge-enabled, and edge/cloud-enabled deployments (Jia et al., 10 Mar 2026, Ma et al., 8 Dec 2025, Nguyen et al., 20 Jan 2026). This literature therefore uses “hierarchical UAV swarm structure” both for internal swarm organization and for the vertical stack connecting the swarm to edge or cloud resources.

2. Communication and control hierarchies

Communication topology is one of the main axes along which hierarchical structure is specified. In the forest swarm model, proximity defines the edge set: $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$ and each node stores

$Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$

This gives a flat peer-to-peer connectivity graph beneath a logical leader–follower hierarchy (Brust et al., 2016).

The communication survey on swarm UAVs makes the distinction more explicit. It describes centralized star networks, multi-star networks, single-group ad hoc networks with a backbone UAV, multi-group ad hoc networks with one backbone UAV per group, and multi-layer ad hoc networks in which backbone UAVs form another ad hoc layer. In these architectures, hierarchy is expressed as ground station $\rightarrow$ backbone or master UAVs $\rightarrow$ regular UAVs, with robustness and scalability increasing as inter-group traffic is removed from a single ground-station bottleneck (Majee et al., 2024).

Layered control frameworks formalize the same separation in software. XTDrone uses ROS, MAVLink, and MAVROS in its communication layer; PX4 SITL in low-level control; perception and local planning in high-level control; and consensus-based formation control, task assignment, and obstacle avoidance in the coordination layer. In its formation demo, each UAV is abstracted as

$\dot{\xi}_i = u_i,$

and the coordination layer applies the consensus controller

$$u_i = -\sum_{j=1}^{N} w_{ij} \big[ (\xi_i - \delta_i) - (\xi_j - \delta_j) \big],$$

with $W=[w_{ij}]$ as the communication topology and $V \subset \mathbb{R}^3$0 as the assigned formation offset (Xiao et al., 2020).

More recent architectures push hierarchy upward into regional and global infrastructure. H-OODA places Observe–Orient–Decide–Act loops at terminal, edge, and cloud layers, with bottom-up flows from local observations to regional and global maps and top-down flows from policies to local parameters (Jia et al., 10 Mar 2026). The ISCCC framework analogously distributes sensing, communication, computing, and control across a BS–swarm layer for reliable flying, a swarm–terminal layer for functioning, and an inter-swarm layer for self-organizing FANET behavior (Ma et al., 8 Dec 2025). LAEI adopts a lighter two-layer version: each UAV performs onboard perception and action selection, while a supervisory “mothership” issues destinations and parameter vectors but “never sends direct low-level actions” (Park et al., 8 Jun 2026). Taken together, these works suggest that communication hierarchy in UAV swarms is increasingly implemented as high-level guidance plus compact state exchange, rather than continuous centralized teleoperation.

3. Formation, motion, and task decomposition

Hierarchical structure becomes operational through the way motion objectives are decomposed. In the forest-assessment model, the swarm extends Boids-style flocking with a leader and explicit connectivity constraints. The local center-of-mass term, separation term, alignment term, and leader goal-seeking term are combined in velocity updates. The leader alone computes

$V \subset \mathbb{R}^3$1

and if $V \subset \mathbb{R}^3$2,

$V \subset \mathbb{R}^3$3

while followers rely on cohesion, alignment, and separation (Brust et al., 2016). This is a minimal hierarchy: only the leader knows $V \subset \mathbb{R}^3$4, but the whole swarm inherits directional motion.

Consensus and assignment provide a more explicit formation hierarchy. XTDrone treats reconfiguration as a Kuhn–Munkres assignment problem between current offsets and target offsets, while each UAV then executes a local consensus controller and a geometric collision-avoidance vector. Hierarchy here is not just “who leads,” but “who decides offsets” versus “who executes continuous tracking” (Xiao et al., 2020).

Guide–worker models introduce parameter-level hierarchy. In “Hierarchical Control of Smart Particle Swarms,” workers follow

$V \subset \mathbb{R}^3$5

with worker–worker interaction

$V \subset \mathbb{R}^3$6

and guide influence

$V \subset \mathbb{R}^3$7

Guides coordinate task allocation, shaping, and movement, while workers remain oblivious to mission-level goals (Varadharajan et al., 2022). The related guide–worker radiation-cleanup study makes the same structural point in mission terms: guides maintain maps and allocate worker chains, whereas workers remain simple, short-range, and expendable (Varadharajan et al., 2024).

Virtual-structure approaches replace explicit leaders with a virtual reference. In the Virtual Centroid formulation, the swarm-level pose is

$V \subset \mathbb{R}^3$8

and each UAV reference is generated by

$V \subset \mathbb{R}^3$9

Cohesion, separation, and alignment are then expressed relative to the Virtual Centroid and inter-agent geometry rather than through pairwise following. This suggests a different kind of hierarchy: a global virtual frame, a formation layer $r$0, and local tracking controllers on the agents (Pita-Romero et al., 29 Jan 2026).

Task decomposition can also be hierarchical without any leader. In dynamic urban patrol, the problem is broken into viewpoint generation, task generation, task allocation, and patrol strategy. Each building defines a task $r$1, each task induces a team $r$2, and within each team UAVs patrol the assigned closed path using only local messages $r$3 and a “bounce” rule when they meet. The resulting structure is environment $r$4 closed paths $r$5 teams $r$6 local patrol behavior (Leong et al., 2024).

4. Learning, inference, and autonomous supervision

A major development in recent work is that hierarchy is no longer only designed; it is also inferred, learned, and adapted. The clearest example is cluster-head detection via graph self-supervised learning. In the two-level USNET, Graph Attention Self-Supervised Learning (GASSL) trains an attention mechanism and an inherent-follow-strategy network so that prediction of follower motion forces attention to concentrate on the true HUAV. For single clusters, the paper reports over 98% average accuracy across various inherent follow strategies; for multiple clusters, MC-GASSL combines GRU-based metric learning and GASSL, and the resulting clustering purity exceeds that of traditional clustering algorithms by at least 10% average (Mou et al., 2022). Hierarchical structure is therefore recoverable from motion traces alone.

Active-inference-inspired work makes hierarchy the primary modeling assumption. “Flying by Inference” decomposes swarm planning into Mission, Route, and Motion levels, learns dictionaries for each level from an expert GA–RF planner, and then performs online selection by minimizing KL-divergence-based abnormality indicators relative to expert-derived reference distributions. The total abnormality is

$r$7

and candidate actions are selected by

$r$8

Here hierarchy is not just a controller stack; it is the state space of the world model itself (Arshid et al., 30 Apr 2026).

Hierarchical learning also appears in mission systems that combine local autonomy with supervisory intelligence. LAEI uses onboard learned policies conditioned on destinations and supervisory parameters: $r$9 where

$\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$0

The supervisory layer performs pairwise exchange goal reassignment and k-means-based recovery after failures, but local motion remains onboard (Park et al., 8 Jun 2026). Agentic AI architectures use a similar division of labor: TinyLLaMA-class local reasoning onboard UAVs, larger edge models for assignment and validation, and cloud models for long-horizon updates and cross-mission knowledge (Nguyen et al., 20 Jan 2026).

Hierarchical learning is also used for cyber-physical services beyond motion. In hierarchical nested personalized federated learning, swarms are stratified into leader, worker, and coordinator UAVs. Workers perform the meta-gradient updates

$\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$1

leaders aggregate to $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$2, and the core aggregates to $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$3, creating a worker–leader–core learning hierarchy across geo-distributed device clusters (Wang et al., 2021). This extends the concept of swarm hierarchy from motion control to distributed online machine learning.

5. Scalability, performance, and robustness

Empirical results across the literature consistently frame hierarchy as a scalability mechanism, though not without trade-offs. In the forest swarm model, a 3D region with side length $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$4, transmission range $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$5, deployment $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$6, destination $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$7, and $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$8 yields average arrival times of $\text{for } u,v \in V:\quad \begin{cases} \text{dist}(u,v) \le r \Rightarrow \{u,v\} \in E\[4pt] \text{dist}(u,v) > r \Rightarrow \{u,v\} \notin E \end{cases}$9 for $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$0, $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$1 for $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$2, and $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$3 for $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$4, compared with a theoretical straight-line time of about $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$5. The increase is modest up to eight UAVs but much larger at twelve, and the paper therefore points to multiple leaders or multi-cluster designs as a plausible extension for larger swarms (Brust et al., 2016).

The guide–worker evidence is stronger. In the largest environment of the robot-swarm scalability study, an egalitarian swarm required 64 workers for 100% success, whereas a hierarchical swarm with 2 guides + 10 workers achieved 100% success, corresponding to an approximately 400% cost reduction for comparable success rate. The same study reports that egalitarian swarms reach consistent 100% success only when average neighbor distance is below about $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$6, whereas guides extend sensing reach without increasing worker density (Varadharajan et al., 2024). This does not establish a universal superiority of hierarchy, but it does show a concrete regime in which hierarchy materially changes scaling laws.

Mission-level supervision can improve performance without sacrificing distributed collision avoidance. In LAEI, evaluations in VMAS with 5 mission UAVs and 8 static obstacles report zero collisions and a mission completion time of 84 steps, compared with 100 steps for PPO and 116 for ORCA; efficiency rises to 0.034, versus 0.029 for PPO. In a failure scenario, coverage ratio increases from 83.12% before optimization to 85.76% after optimized reassignment (Park et al., 8 Jun 2026). In H-OODA case studies, a target-search scenario with 15 UAVs in a $Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$7 m area shows higher search efficiency, lower target search time, and higher success rate for full H-OODA than for edge-end OODA or single-layer OODA, and the NFV-enabled study reports

$Neigh(v) = \{u \in V \mid d(v,u) \le r\}.$8

with error rates ordered inversely by loop depth (Jia et al., 10 Mar 2026).

Hierarchical overlays also matter for logistics internal to the swarm. In SwarmUpdate, the Updater–Leader–Follower hierarchy yields faster convergence than gossip and auction baselines in large heterogeneous swarms; at swarm size 500 and a 240-packet patch, SwarmSync is 78.3% faster than SOUL and 47.7% faster than Gossip, though it incurs 87.7% more overhead on average than SOUL because of ACK-based reliability. When model patching freezes 7/8 fire modules, patch size drops from 2.9 MB to 0.8 MB, a 73.3% reduction, with corresponding reductions in convergence time and overhead (Geng et al., 18 Mar 2025). These results broaden the meaning of “hierarchical UAV swarm structure” from control to software maintenance and model lifecycle management.

6. Limits, misconceptions, and open problems

A common misconception is that hierarchy is synonymous with full centralization. The surveyed work does not support that equation. The forest swarm is hierarchical in roles but distributed in internal communication; urban patrol is hierarchically decomposed into tasks and teams but fully decentralized in operation; LAEI explicitly separates lightweight supervision from onboard action selection; and H-OODA distributes OODA loops across layers rather than collapsing them into one controller (Brust et al., 2016, Leong et al., 2024, Park et al., 8 Jun 2026, Jia et al., 10 Mar 2026). Hierarchy, in this literature, is more accurately a structured allocation of responsibilities across scales.

A second misconception is that hierarchy always implies a physical chain of command. XTDrone’s six-layer stack, H-OODA’s cloud–edge–terminal organization, ISCCC’s three-layer reflex arcs, and HN-PFL’s worker–leader–core learning all show hierarchies that are software, functional, or service-oriented rather than purely topological (Xiao et al., 2020, Ma et al., 8 Dec 2025, Wang et al., 2021). This suggests that “hierarchical UAV swarms structure” should be read as a systems concept that spans communication, computation, control, and mission semantics.

The main open problems are likewise multi-layered. Several papers identify communication reliability, limited bandwidth, RF interference and jamming, and latency bounds across layers as unresolved constraints on hierarchical control (Jia et al., 10 Mar 2026, Jia et al., 2024). Others emphasize that large volumes of multi-modal data create data processing bottlenecks unless distributed fusion, edge computing, or compression are improved (Jia et al., 10 Mar 2026). Security and resilience recur throughout: H-OODA warns that compromise of one layer can corrupt the entire hierarchy, CCDS introduces anti-interference control links and resource-pool-based reconfiguration, and agentic AI systems raise trust and hallucination issues at edge and cloud levels (Jia et al., 10 Mar 2026, Jia et al., 2024, Nguyen et al., 20 Jan 2026).

There are also tensions internal to hierarchical design. The ISCCC paper notes that the flying layer, functioning layer, and self-organizing layer may demand incompatible formations or topologies, motivating hierarchical control policies and unified reward functions (Ma et al., 8 Dec 2025). The forest swarm points to the limitations of a single leader and the absence of explicit obstacle modeling in the algorithm (Brust et al., 2016). XTDrone notes that detailed scalability limits, communication delays, and real-world deployment effects are not fully analyzed (Xiao et al., 2020). These limitations suggest that future hierarchical UAV swarms will likely combine multi-cluster role structures, adaptive overlays, and explicit communication-aware control.

In the literature covered here, hierarchical UAV swarm structure is therefore best understood as a family of design patterns rather than a single canonical architecture. It includes leader–follower and cluster-head swarms, guide–worker systems, layered autonomy stacks, edge- and cloud-mediated decision hierarchies, and hierarchical inference or learning models. What unifies these patterns is not centralization, but decomposition: global mission context, regional or swarm-level coordination, and local sensing and actuation are separated so that each scale can operate with its own information, latency, and resource constraints.