Paper deep dive

LLM-Enabled UAV Dispatch: A System-Level Survey and Taxonomy

Xiao Han, Aoyang Quan, Xiangyu Zhao, Xiangjie Kong, Guojiang Shen

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 131

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 10/11/2026, 2:29:26 AM

Summary

This paper presents a system-level survey and taxonomy of Large Language Model (LLM)-enabled Unmanned Aerial Vehicle (UAV) dispatch. It conceptualizes LLMs as a cross-layer semantic orchestration layer that bridges human intent, symbolic decision-making, and physical execution. The authors categorize existing methods into four paradigms: pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. The survey analyzes the role of LLMs in semantic parsing, planning, solver orchestration, and safety assessment, while identifying open challenges such as latency, grounding reliability, and physical feasibility.

Entities (9)

Large Language Models · technology · 99%UAV Dispatch · concept · 98%Decentralized Agentic Dispatch · paradigm · 95%Divide-and-Conquer Dispatch · paradigm · 95%Global Assignment Dispatch · paradigm · 95%Pipeline Dispatch · paradigm · 95%Semantic Parsing · function · 92%Low-altitude economy · domain · 90%Retrieval-Augmented Generation · technique · 90%

Relation Signals (8)

LLMs → enables → UAV Dispatch

confidence 95% · This survey provides a unified characterization of LLM-enabled UAV dispatch systems that bridges semantic intent, symbolic decision-making, and physical UAV execution.

LLMs → actsas → cross-layer semantic orchestration layer

confidence 93% · Rather than treating LLMs as standalone add-ons, we conceptualize them as a cross-layer semantic orchestration layer connecting human instructions, external solvers, and distributed control modules.

LLMs → supports → Semantic Parsing

confidence 92% · We further examine how LLMs support semantic parsing, retrieval-grounded planning, solver orchestration, local agent reasoning, multi-agent coordination, safety assessment, and human-facing explanation.

Decentralized Agentic Dispatch → uses → LLMs

confidence 90% · Building on this decentralized decision-making paradigm, task ownership emerges from local autonomy, while LLM agents serve as complementary components that enhance the system with memory, semantic negotiation, and reflection

Pipeline Dispatch → uses → LLMs

confidence 90% · In the LLM era, the same abstraction becomes a tool-grounded reasoning chain: an LLM or multimodal LLM parses M, retrieves mission knowledge, decomposes the request into structured tasks, selects downstream solvers, and asks external validators to check the produced assignment, trajectory, schedule, and execution feedback

Global Assignment Dispatch → uses → LLMs

confidence 90% · It relies on a central dispatcher or equivalent global view, while the LLM layer interprets intent, retrieves rules, explains tradeoffs, or guides solver choice.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Unmanned aerial vehicle (UAV) dispatch is beginning to move beyond isolated path planning and optimization-driven resource allocation toward system-level coordination supported by semantic reasoning and LLM-based interfaces. This survey provides a unified characterization of LLM-enabled UAV dispatch systems that bridges semantic intent, symbolic decision-making, and physical UAV execution. Rather than treating LLMs as standalone add-ons, we conceptualize them as a cross-layer semantic orchestration layer connecting human instructions, external solvers, and distributed control modules. We organize the literature into four representative dispatch paradigms: pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. For each paradigm, we analyze its decision logic, system structure, control flow, representative methods, and potential LLM roles. We further examine how LLMs support semantic parsing, retrieval-grounded planning, solver orchestration, local agent reasoning, multi-agent coordination, safety assessment, and human-facing explanation. We discuss the implications of these paradigms for scalability, robustness, coordination burden, and verification requirements, and identify open challenges including latency-aware reasoning, grounding reliability, physical feasibility guarantees, edge deployment, privacy protection, and distributed consistency. This survey provides a system-level taxonomy and design perspective for integrating LLMs into safety-critical UAV dispatch systems.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

127,391 characters extracted from source content.

Expand or collapse full text

LLM-Enabled UAV Dispatch: A System-Level Survey and Taxonomy Xiao Han Aoyang Quan Xiangyu Zhao Xiangjie Kong Guojiang Shen †thanks: Xiao Han, Aoyang Quan, Guojiang Shen and Xiangjie Kong are with the School of Computer Science, Zhejiang University of Technology, China (e-mail: hahahenha@gmail.com; 19941348551@163.com; xjkong@ieee.org; gjshen1975@zjut.edu.cn). Abstract Unmanned aerial vehicle (UAV) dispatch is beginning to move beyond isolated path planning and optimization-driven resource allocation toward system-level coordination supported by semantic reasoning and LLM-based interfaces. This survey provides a unified characterization of LLM-enabled UAV dispatch systems that bridges semantic intent, symbolic decision-making, and physical UAV execution. Rather than treating LLMs as standalone add-ons, we conceptualize them as a cross-layer semantic orchestration layer connecting human instructions, external solvers, and distributed control modules. We organize the literature into four representative dispatch paradigms: pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. For each paradigm, we analyze its decision logic, system structure, control flow, representative methods, and potential LLM roles. We further examine how LLMs support semantic parsing, retrieval-grounded planning, solver orchestration, local agent reasoning, multi-agent coordination, safety assessment, and human-facing explanation. We discuss the implications of these paradigms for scalability, robustness, coordination burden, and verification requirements, and identify open challenges including latency-aware reasoning, grounding reliability, physical feasibility guarantees, edge deployment, privacy protection, and distributed consistency. This survey provides a system-level taxonomy and design perspective for integrating LLMs into safety-critical UAV dispatch systems. Index Terms: UAV dispatch, large language models, multi-agent systems, decision-making, trajectory planning. I Introduction Unmanned Aerial Vehicles (UAVs) have evolved from small remote-controlled platforms into core infrastructure for the low-altitude economy (LAE) [1]. They are now used in last-mile package delivery [2], environmental monitoring and surveillance [3], disaster-zone emergency response [4], infrastructure inspection [5], and aerial communication services. In these applications, the critical challenge has shifted from single-UAV navigation to fleet-level dispatch and coordination. The dispatch system must decide which UAV should serve which task, when it should depart, which route it should follow, how energy and communication resources should be allocated, and how conflicts in dense low-altitude airspace should be avoided. We refer to this system-level problem as UAV dispatch: the scheduling, routing, resource allocation, association, and coordination of UAV fleets under physical, communication, safety, privacy, and regulatory constraints. Fig. 1 organizes it as a tree: its roots ground UAV dispatch in mission and application demands, core dispatch problems, and physical and operational constraints. Its trunk traces the evolution from single-UAV systems to LLM-enabled applications, and its four branches show the paradigms we summarized in this survey. Fig. 1: Evolution and scope of LLM-enabled UAV dispatch. UAV dispatch is attracting rapidly growing attention in both research and real-world deployment. Commercial UAV markets are expanding in logistics, inspection, and public services, regulators such as the FAA and EASA are developing frameworks for high-density BVLOS and U-space operations, and China has positioned the LAE as a national growth engine. This deployment momentum motivates the central premise of this survey: large-scale UAV operations cannot be supported by isolated path planning or single-agent control alone. As the number of UAVs grows, collision-free trajectories and communication resources become harder to allocate jointly [6]. Early multi-quadrotor swarm systems [7] further show that scalable UAV coordination requires tight integration among planning, communication, localization, and collision avoidance. Dispatch decisions become tightly coupled across task assignment, trajectory planning, communication, energy, timing, and privacy. Human operators increasingly issue open-ended mission requests rather than fully structured optimization inputs. UAV dispatch therefore requires a framework that connects classical optimization and learning with semantic mission understanding, cross-layer coordination, and trustworthy human oversight. The emergence of LLMs makes this framework more tractable by providing new interfaces for semantic interpretation, tool use, and human-facing explanation. Recent vision-language and multimodal LLM studies provide the perception and grounding foundations needed for UAV missions involving aerial images, maps, and natural-language instructions [8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. Together with tool-use, agent, and embodied-model frameworks, these models can interpret natural-language missions, retrieve regulations or maps, reason over heterogeneous context, call external solvers, critique candidate plans, and explain dispatch decisions to human operators [18, 19, 20, 21, 22, 23]. Recent UAV-oriented studies have already demonstrated a range of LLM-assisted capabilities, including path planning, vision-language navigation grounding, compliance-aware trajectory generation, multi-objective UAV network optimization, and edge-based multi-LLM orchestration [6, 24, 25, 26, 27, 28, 29, 30]. However, these efforts remain fragmented across planning, communication, swarm coordination, inspection, and safety, lacking a unified abstraction to connect them. Existing surveys summarized in Table I predominantly focus on UAV trajectory prediction and path planning [31, 32], swarm networking and collaborative control [33, 34], security and privacy [35], or broader LLM-related topics such as model evaluation, vision-language models, and LLM-agent systems [18, 9, 36]. However, these surveys still lack a dispatch-centered taxonomy that systematically explains how LLMs reshape the end-to-end UAV operation pipeline—from mission intent and task assignment to trajectory planning, scheduling, resource allocation, and execution monitoring. In contrast, this work takes UAV dispatch as the central organizing abstraction, rather than using path planning, swarm control, or LLM agents as isolated lenses. We study how LLMs reshape dispatch interfaces, dispatch logic, and dispatch verification across four emerging paradigms, thereby filling this gap and motivating the central question of this paper: How should UAV dispatch be organized in the LLM era? TABLE I: Comparison with representative surveys related to UAV dispatch and LLM-enabled aerial intelligence. Surveys Focus & Scope Key Findings / Application Categories [1] Reviews low-altitude airspace management for uncrewed aircraft and advanced air mobility. Summarizes airspace concepts, operational requirements, BVLOS integration, regulatory constraints, and traffic-management challenges. [31] Surveys trajectory-prediction techniques for UAVs across modeling, data-driven learning, and motion-pattern analysis. Classifies trajectory-prediction methods and discusses their role in collision avoidance, motion forecasting, and safe UAV operation. [32] Reviews UAV path-planning techniques with emphasis on planning algorithms, environment representation, and motion constraints. Categorizes path-planning approaches by optimization strategy, search mechanism, and constraint handling, with limited attention to end-to-end dispatch logic. [33] Surveys computational-intelligence algorithms for UAV swarm networking and collaboration. Reviews communication, cooperative control, formation, search, learning-based collaboration, and future swarm-networking directions. [34] Reviews swarm-intelligence algorithms for multi-UAV collaboration. Summarizes distributed cooperation, formation, task coordination, search, and collective decision-making mechanisms for UAV swarms. [35] Reviews privacy and confidentiality issues in drone operations. Identifies privacy leakage, data exposure, trust, authentication, and secure communication challenges in UAV deployment. [18] Surveys evaluation methods for large language models. Reviews evaluation dimensions for knowledge, reasoning, alignment, robustness, safety, and human preference, without targeting UAV dispatch. [9] Surveys vision-language models for vision tasks. Summarizes architectures, pretraining objectives, visual grounding, downstream vision-language tasks, and evaluation trends relevant to perception-rich UAV missions. [36] Surveys LLM-based multi-agent systems and their progress and challenges. Reviews agent profiling, memory, planning, tool use, communication, cooperation, and benchmarks, but does not address physical UAV dispatch constraints. This survey UAV dispatch in the LLM era, centered on the transformation from mission intent to executable fleet decisions. Organizes LLM-enabled UAV operation into pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch, with LLM roles as parsers, planners, solver orchestrators, local agents, coordinators, critics, and interfaces. To answer this question, we propose a dispatch-logic taxonomy based on the transformation from mission descriptions to executable dispatch decisions. At the top level, we organize dispatch logic into four paradigms: pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. After this top-level classification, we organize papers within each paradigm along several secondary axes. These axes include the dispatch decisions they cover, namely task assignment, trajectory or route planning, scheduling, resource allocation, and cooperation or association among UAVs. We also record the solver family used by each study, including optimization, learning, auctions, heuristics, or LLM-tool orchestration, and the application context it targets. Among these categories, (1) pipeline dispatch covers methods that convert a mission into ordered stages, such as semantic parsing, task decomposition, assignment, trajectory generation, scheduling, and monitoring. Its dispatch process is organized as a stage-wise workflow in which LLMs can serve as parsers, tool selectors, and critics, rather than as a single global optimization problem. (2) Global assignment dispatch covers centralized or globally coordinated methods that jointly determine task matching, routes, schedules, resources, and associations when sufficient global state is available. It relies on a central dispatcher or equivalent global view, while the LLM layer interprets intent, retrieves rules, explains tradeoffs, or guides solver choice. (3) Decentralized agentic dispatch covers agent-led methods in which UAVs locally select tasks, bid, negotiate, replan, or learn policies from partial observations. Building on this decentralized decision-making paradigm, task ownership emerges from local autonomy, while LLM agents serve as complementary components that enhance the system with memory, semantic negotiation, and reflection, without replacing verified control mechanisms. (4) Divide-and-conquer dispatch covers methods that partition a large mission by dispatch decisions, objectives, regions, time windows, fleet groups, roles, communication links, knowledge sources, or model placement. The main design question is how to split and then reconnect subproblems through solvers, retrieval sources, edge models, and multi-agent orchestration. Together, these four paradigms cover the dispatch layer of UAV systems. The main contributions of this paper are as follows: • We define UAV dispatch in the LLM era as a semantic–symbolic–physical decision process that maps human intent and environmental context to task assignment, trajectory, schedule, resource allocation, and cooperation decisions. • We propose a dispatch-logic taxonomy with four LLM-era paradigms. This taxonomy connects classical UAV dispatch methods with emerging LLM-enabled dispatch architectures. • We analyze how LLMs can serve as semantic parsers, retrieval-grounded planners, solver orchestrators, local agents, multi-agent coordinators, safety critics, and human-facing explanation interfaces, while emphasizing that physical feasibility must still be guaranteed by specialized modules. • We identify open challenges and future directions for trustworthy LLM-enabled UAV dispatch, including grounding, hallucination control, edge deployment, multi-agent trust, privacy, accountability, benchmarking, and human governance. The remainder of this paper is organized as follows. Section I introduces the preliminaries and unified notation for UAV dispatch. Sections I to VI review the four dispatch paradigms: pipeline dispatch for UAV missions, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. Section VII synthesizes the four paradigms in the LLM era and discusses future challenges and research directions. Section VIII concludes the survey. I Preliminaries Before reviewing the four paradigms in detail, we first introduce the definitions and notation used throughout this survey. We begin with the traditional dispatch framework, which provides the classical notions of task assignment, scheduling, routing, and resource allocation. We then extend this view to unmanned dispatch systems, emphasizing how autonomous perception, communication, and online adaptation inherit from, yet differ from, traditional dispatch. Finally, we specialize this view to UAV dispatch systems, where UAV operations introduce three-dimensional mobility, airspace safety, endurance, communication, and regulatory constraints. I-A Traditional Dispatch Framework Traditional dispatch studies how a set of service resources should be matched with a set of spatially or temporally distributed requests. The agents may be vehicles, robots, workers, ships, or aircraft, while the requests may represent delivery orders, inspection targets, emergency events, passengers, or sensing tasks. To keep the notation consistent with later sections, we use =u1,u2,…,uNU=\u_1,u_2,…,u_N\ for the agent set, =τ1,τ2,…,τMT=\ _1, _2,…, _M\ for the task set, and Ω for the operating environment. The core decisions are assignment X, routing or trajectory planning P, scheduling S, resource allocation R, and association or cooperation A. The assignment matrix =[xij]X=[x_ij] indicates whether agent uiu_i serves task τj _j, with xij=1x_ij=1 if the assignment is made and xij=0x_ij=0 otherwise. A feasible traditional dispatch plan covers the required tasks, respects capacity, endurance, and service-time limits, and keeps the selected routes feasible in Ω . Such a formulation captures classical problems including vehicle routing, pickup-and-delivery, ride-hailing, warehouse robot coordination, emergency vehicle dispatch, and field-service scheduling. Its key role is to determine who serves each task, when it is served, where the agent moves, and how limited resources are allocated. I-B Unmanned Dispatch System Unmanned dispatch extends the traditional framework by replacing passive executors with autonomous agents. Each agent may perceive the environment, communicate with neighbors, update its local state, and adapt its actions during execution. The state of the unmanned agent uiu_i at time t can be summarized compactly as i(t)=(i(t),ei(t),oi(t))s_i(t)=(q_i(t),e_i(t),o_i(t)), where i(t)q_i(t) is the position and ei(t)e_i(t) is the remaining energy. The term oi(t)o_i(t) aggregates local observations such as nearby tasks, obstacles, communication status, capability mode, and map information. The system state is then the collection of all agent states, task states, and environmental conditions. Compared with traditional dispatch, unmanned dispatch has three additional features. First, the environment is often partially observable, so each agent may have only local or delayed information. Second, task execution and motion planning are tightly coupled because feasibility depends on current position, energy, sensing range, communication quality, and mobility limits. Third, the dispatch logic must be online and adaptive, since new tasks, obstacles, failures, or communication interruptions may appear during execution. Therefore, unmanned dispatch bridges classical operations research and autonomous multi-agent coordination. I-C UAV Dispatch Methods UAV dispatch further specializes in unmanned dispatch to low-altitude airspace and airborne agents. In this survey, it refers to the coordination of UAVs for delivery, inspection, sensing, surveillance, communication, emergency response, and aerial data collection. Compared with ground robots or vehicles, UAVs operate in three-dimensional space and must jointly satisfy flight dynamics, airspace safety, battery endurance, communication coverage, weather, regulation, and privacy constraints. For UAV uiu_i, let i(t)=[xi(t),yi(t),zi(t)]⊤q_i(t)=[x_i(t),y_i(t),z_i(t)] denote its position at time t. A feasible trajectory must satisfy mobility limits such as ‖˙i(t)‖≤vimax\| q_i(t)\|≤ v_i , collision avoidance ‖i(t)−k(t)‖≥dmin\|q_i(t)-q_k(t)\|≥ d_ for i≠ki≠ k, airspace restrictions in Ω , endurance requirements expressed through ei(t)e_i(t), and communication-quality requirements for the assigned UAV–task pairs. These constraints make UAV dispatch a coupled problem of assignment, trajectory planning, timing, communication, and resource management. For the UAV dispatch problem considered in this survey, an LLM-enabled interface may transform human instructions and contextual information into structured tasks, constraints, and preferences, while the dispatch layer optimizes the resulting multi-objective planning problem. In this view, the LLM does not replace the optimization model. Instead, it provides a semantic interface layer: retrieval-augmented generation (RAG) can ground the prompt in maps, regulations, weather reports, historical logs, and mission documents; structured reasoning prompts or rationale generation can make intermediate dispatch assumptions more inspectable; and self-refinement or external evaluators can check whether a candidate plan satisfies physical, regulatory, and privacy constraints [37, 38, 20, 39, 22, 40, 41]. We use the following objective as a unified abstraction: min,,,,J⁡(CLOSE _X,P,S,R,A J( OPEN,,,,)=λ1Ctime(,,) ,P,S,R,A)= _1C_time(X,P,S) (1) +λ2Cenergy(,,)+λ3Cdelay(,) + _2C_energy(X,P,R)+ _3C_delay(X,S) +λ4Crisk(,,Ω)+λ5Ccomm(,,) + _4C_risk(P,A; )+ _5C_comm(P,R,A) +λ6Lpriv(,,) + _6L_priv(X,R,A) −λ7Qserv(,,,,), - _7Q_serv(X,P,S,R,A), where X, P, S, R, and A denote assignment, trajectory or path, schedule, resource allocation, and association or cooperation decisions, respectively. Ctime(,,)C_time(X,P,S) measures mission time, Cenergy(,,)C_energy(X,P,R) measures energy consumption, and Cdelay(,)C_delay(X,S) measures service delay. Crisk(,,Ω)C_risk(P,A; ) measures safety risk under the operating environment Ω , Ccomm(,,)C_comm(P,R,A) measures communication cost, and Lpriv(,,)L_priv(X,R,A) measures privacy leakage. Qserv(,,,,)Q_serv(X,P,S,R,A) measures service quality such as coverage, throughput, sensing accuracy, or task completion ratio. The non-negative weights λ1,…,λ7 _1,…, _7 balance these objectives for different UAV dispatch scenarios. I Pipeline Dispatch Pipeline dispatch for UAV missions decomposes a complex mission into ordered modules, so that semantic parsing, task decomposition, assignment, trajectory generation, scheduling, and execution monitoring are handled as successive stages rather than as a single monolithic optimization problem. Formally, given a mission description ℳM, a pipeline-dispatch framework can be abstracted as: ℳ→f1→f2→f3→f4→f5ℰ,M f_1T f_2X f_3P f_4S f_5E, (2) where ℳM denotes the mission description and ℰE denotes the execution or monitoring stage. The functions f1,…,f5f_1,…,f_5 may correspond to task decomposition, clustering, assignment, path planning, scheduling, collision avoidance, or online replanning, depending on the application scenario. In the LLM era, the same abstraction becomes a tool-grounded reasoning chain: an LLM or multimodal LLM parses ℳM, retrieves mission knowledge, decomposes the request into structured tasks, selects downstream solvers, and asks external validators to check the produced assignment, trajectory, schedule, and execution feedback [37, 19, 21, 6]. Thus, the LLM layer does not replace the pipeline; it makes the front end more adaptive to natural-language intent, heterogeneous context, and changing low-altitude constraints while keeping each stage auditable. Figure 2 illustrates this stage-wise dispatch logic. Fig. 2: Pipeline dispatch decomposes a mission into ordered processing stages, where an LLM layer can parse intent, call tools, and trigger verification. TABLE I: Taxonomy of representative pipeline dispatch methods and their LLM-era orchestration roles aligned with Eq. (2). Pipeline Transition Classical Stage Role LLM-Era Orchestration Role Representative Methods f1:ℳ→f_1:M\!→\!T Spatial or task decomposition Semantic scene parsing and RAG-grounded task graph construction Li et al. [42], Xiong et al. [43] Hierarchical mission decomposition Natural-language mission breakdown and role prompting Luan et al. [44], Cheng et al. [45], Guo et al. [46], Li et al. [42], Jünger et al. [47] Rolling or adaptive clustering Context-aware reclustering from execution feedback Yanmaz et al. [48], Ren et al. [5] f2:→f_2:T\!→\!X Task or role assignment Tool-call handoff from task graph to assignment solver Yanmaz et al. [48], Guo et al. [46], Zhao et al. [49], Li et al. [42], Gong et al. [6] Hierarchical task allocation Constraint generation and priority explanation Luan et al. [44], Cheng et al. [45], Guo et al. [46] Battery or inspection-task matching Resource-state summarization and repair hints Alyassi et al. [50], Ren et al. [5] f3:→f_3:X\!→\!P Communication-aware trajectory design Constraint interpretation before path optimization Wang et al. [51], Jing et al. [52], Zhang et al. [53], Li et al. [28] Learning-based trajectory control Compliance-aware critic for learned policies Zhu et al. [54], Gong et al. [6] Coverage, search, or sensing paths Vision-language grounding of regions and targets Yanmaz et al. [48], Li et al. [42], Xiong et al. [43], Liu et al. [55], Lin et al. [24, 56, 26], Wu et al. [57] Hierarchical or collision-aware path planning External validator for safe path refinement Cheng et al. [45], Zhou et al. [58] f4:→f_4:P\!→\!S Time allocation or service ordering Schedule explanation and local repair Wang et al. [51], Jing et al. [52], Zhang et al. [53] Subtask or execution scheduling Agent memory for rolling execution context Luan et al. [44], Cheng et al. [45], Ren et al. [5] Battery and charging scheduling Energy-state explanation and charging trigger Alyassi et al. [50] f5:→ℰf_5:S\!→\!E Online execution update Reflection loop and replanning trigger Zhu et al. [54], Guo et al. [46], Ren et al. [5] Connectivity maintenance and replanning Communication-aware agent monitor Yanmaz et al. [48] Collision or safety handling Runtime safety critic and human alert Zhou et al. [58] I-A Hierarchical Task Decomposition and Scheduling The first major line of pipeline dispatch emphasizes the decomposition architecture itself. Instead of focusing only on the final trajectory, these methods study how a complex mission can be broken into subtasks, organized into multiple layers, and scheduled across UAVs in a structured manner. This subsection corresponds mainly to the f1f_1, f2f_2, and f4f_4 rows of Table I: high-level missions are first decomposed, then assigned or allocated, and finally organized into executable schedules. The basic hierarchical pattern is to separate mission understanding from execution ordering. Luan et. al. [44] use an upper layer for topology reconstruction and task allocation and a lower layer for subtask scheduling in UAV-assisted mobile edge computing emergency networks. Cheng et. al. [45] extend the pattern to heterogeneous aircraft-carrier–UAV swarms, separating the carrier path planning layer from the multi-UAV collaborative task planning layer, while Guo et. al. [46] use a hierarchical framework to connect task decomposition with marginal return allocation. Across these studies, hierarchy is not only a computational trick; it defines which layer is responsible for mission abstraction, mobility support, and executable task ordering. Once the hierarchy is established, the next challenge is how to refine allocation decisions without losing scalability. Guo et. al. [46] cluster and decompose tasks in the upper layer, then allocate subtasks through marginal return optimization. Zhao et. al. [49] use a coarse initialization stage followed by a parallel auction network for multi-target tracking, turning global allocation into a staged refinement process. Li et. al. [42] add task priority and multi-objective metaheuristic optimization to multi-region coverage. These methods move beyond a fixed two-layer pipeline: decomposition produces an initial structure, while optimization inside later stages adapts the assignment to heterogeneity, priority, and scale. Hierarchical pipelines are also useful when the input or the internal planning module is complex. In autonomous exploration, hierarchical frontier-based planners such as FUEL show how global exploration objectives, frontier selection, and local trajectory generation can be organized into a fast hierarchical pipeline [59]. Jünger et. al. [47] bridge high-level human commands and low-level UAV subtasks through hierarchical task decomposition, which is directly relevant to LLM-enabled dispatch interfaces. In such an interface, the upper layer should not be viewed only as a manually written task parser. It can be an LLM-based semantic layer that retrieves mission rules, converts operator intent into a task graph, selects specialized perception or optimization tools, and records why each subtask is created [37, 19, 21]. Embodied multimodal models further show how language, visual observations, and continuous robot states can be grounded in one planning interface [23], while AerialVLN and large-model landmark discovery show that navigation requires spatial reasoning beyond ground-agent VLN [55, 24, 25, 26, 56]. Zhou et. al. [58] apply a related decomposition principle inside path planning: an upper layer generates initial Dubins-curve paths, while a lower layer resolves conflicts with a three-dimensional velocity-obstacle strategy. These examples show that hierarchy can operate at different semantic levels, from natural-language command interpretation and evidence-grounded task construction to collision-aware trajectory refinement. These hierarchical methods demonstrate that pipeline dispatch is not merely a simple engineering workflow. It can serve as a general decomposition principle for organizing complex UAV missions, especially when the original problem contains multiple coupled decision layers. I-B Trajectory-Centric Pipeline Dispatch This subsection classifies trajectory-centric studies as pipeline dispatch when trajectory generation is formulated as a downstream stage following task allocation or service-order decisions, regardless of whether the trajectory solver is centralized, distributed, optimization-based, or learning-based. Within this family, trajectory generation constitutes the primary technical bottleneck. The main focus is therefore on designing and optimizing UAV trajectories for communication, data collection, coverage, and target-tracking scenarios. In the notation of Table I, these studies are centered on f3:→f_3:X , while additional requirements such as service ordering, connectivity maintenance, and collision avoidance may further involve f4f_4 and f5f_5. One trajectory-centric branch couples UAV movement with communication, sensing, and service objectives. Wang et. al. [51] expose the upstream and downstream dependencies of this trajectory module: a service order is first selected, and the resulting path is optimized together with sensor uploading powers. Recent ISAC-oriented studies further expand this coupling by coordinating trajectory, association, beamforming, and target-localization quality [52, 53], while LLM-guided UAV-ISAC optimization shows how language reasoning can guide multi-objective search before a numerical solver verifies feasibility [28]. Together, these studies show that communication-aware trajectories are rarely standalone geometric paths; they usually inherit task-ordering, sensing, and link-quality requirements from adjacent pipeline stages. When the environment is dynamic, the trajectory stage must also become adaptive. Zhu et. al. [54] introduce collaborative reinforcement learning for three-dimensional target tracking, replacing a fixed offline trajectory optimizer with a learned motion policy that responds to changing target states. Yanmaz et. al. [48] extend this idea at the system level by separating search and relay roles: search UAVs follow targets, while relay UAVs preserve network connectivity. These studies indicate that f3f_3 and f5f_5 become coupled when sensing, motion, and communication must be updated during the mission. Coverage missions reveal another dependency: trajectory generation often requires an earlier spatial decomposition stage. Recent multi-region and post-disaster studies make the same dependency more explicit by combining region partitioning, task allocation, and local coverage routing [42, 43]. Thus, even in trajectory-centric dispatch, the quality of P depends strongly on how the mission is decomposed into spatial or functional units. In the LLM era, the trajectory stage is better understood as a grounded tool call rather than a free-form path generated by language alone. The LLM or multimodal LLM can translate an instruction such as “inspect the blocked bridge and avoid dense residential blocks” into semantic waypoints, forbidden regions, and task priorities. However, the final path still has to be produced or verified by geometry, control, and safety modules. PaLM-E supports this view by connecting language with embodied sensory states [23], while vision-language pretraining, open-vocabulary perception, and segmentation foundation models provide the perception substrate for turning aerial imagery into structured targets or regions [8, 9, 60, 61, 15, 62, 16, 63]. AerialVLN-style benchmarks make the aerial grounding problem explicit for UAVs [55, 57]. Recent hybrid DRL–LLM trajectory planning further suggests that LLM reasoning is most useful as a compliance-aware critic and constraint interpreter, while the low-level policy handles motion feasibility [6]. Overall, trajectory-centric pipeline dispatch treats task decomposition as a lightweight preprocessing step and places the main optimization burden on path or trajectory generation. Such methods are effective when the task structure is relatively clear, and the key bottleneck is mobility optimization. However, they may become less effective when task decomposition, resource management, and trajectory planning are strongly coupled. I-C Resource-Aware and Application-Oriented Pipelines Beyond trajectory design and task decomposition, the pipeline paradigm has also been applied to resource management and application-specific dispatch. In these studies, the key idea is to isolate a critical resource or operational factor as an independent stage of the dispatch process. Compared with the previous two subsections, the organizing logic here is less about a universal decomposition architecture and more about which bottleneck should be promoted to a first-class pipeline stage. Inspection scenarios highlight the need for rolling updates across f1f_1, f2f_2, f4f_4, and f5f_5. Application-oriented pipelines must couple visual target tracking, trajectory generation, and control feasibility during execution, as in autonomous power-line inspection with perception-aware MPC [64]. Ren et. al. [5] first cluster inspection tasks according to task characteristics and then apply greedy partition scheduling to imitate manual planning logic. Because long-duration inspection missions may receive updated task sets during execution, the pipeline is repeatedly refreshed rather than solved once. This makes rolling clustering and scheduling the core mechanism, while trajectory planning becomes a downstream consequence of the current inspection workload. Delivery and long-duration inspection scenarios expose a different bottleneck: energy availability. Alyassi et. al. [50] jointly plan autonomous recharging and flight missions for battery-operated drones. Here, the pipeline is organized around energy logistics rather than only around UAV movement; feasibility depends on whether the right energy state can be assigned at the right time. In LLM-enabled application pipelines, mission knowledge becomes another dispatch resource. Inspection reports, emergency messages, weather alerts, battery logs, and airspace notices may arrive as heterogeneous text, images, or tables. An LLM-informed drone inspection system can convert such inputs into candidate defects, inspection subtasks, and agent responsibilities [30]; self-refinement and reflection mechanisms can then turn execution feedback into repair suggestions without changing the verified scheduler or controller [40, 65]. Therefore, the application-oriented pipeline becomes a closed loop: retrieve domain evidence, extract constraints, call a solver, verify execution, and update the next dispatch prompt from monitored outcomes. Aerial construction and manufacturing tasks provide another example of such closed-loop dispatch, where multiple aerial robots must coordinate task execution, geometric feedback, online correction, and human supervision during physical operation [66]. These application-oriented works reveal that pipeline design is often driven by the dominant bottleneck of the target scenario. Inspection, delivery, and emergency sensing pipelines differ mainly in their dominant operational constraint: rolling task updates, energy logistics, and range-aware partitioning, respectively. Therefore, the structure of a pipeline is not fixed; it should be adapted to the mission objective, resource limitations, and operational environment. IV Global Assignment Dispatch Global assignment dispatch uses a central dispatcher, such as a control station, cloud platform, edge server, fleet manager, or LLM-assisted meta-dispatcher, to collect global information and jointly determine the main dispatch decisions introduced in Section I. Unlike pipeline dispatch, global assignment dispatch does not pass intermediate artifacts through a fixed sequence of stages. Instead, it jointly optimizes task matching X, routes P, schedules S, resources R, and associations A under a shared global state. Classical multi-UAV coordination studies have already emphasized that goal assignment, resource allocation, and trajectory generation are tightly coupled in fleet-level control, which provides an early systems foundation for the global assignment view [67]. Table I summarizes this paradigm according to four recurring dimensions: mobility–communication assignment, routing–scheduling assignment, scalable solver design, and security or privacy constraints. In the LLM era, this global view can be wrapped by a semantic meta-dispatcher that translates operator intent into constraints, retrieves airspace or network knowledge, chooses assignment solvers, and explains trade-offs before the optimization core commits a plan. Fig. 3 illustrates this global assignment view. Fig. 3: Global assignment dispatch uses a central dispatcher to jointly determine tasks, routes, schedules, and resources, with an LLM layer supporting intent interpretation and solver selection. TABLE I: Taxonomy of representative global assignment dispatch methods and their LLM-era assignment roles. Central Dispatch Dimension Coupled Decision Role LLM-Era Assignment Role Representative Methods Mobility–communication Trajectory and link scheduling Global-state summarization and link-aware prompt grounding Liu and Zheng [68], He et al. [69], Hu et al. [70], Jing et al. [52], Zhang et al. [53], Wu et al. [71] Data collection, offloading, and MEC Service-intent parsing and resource constraint extraction Nguyen et al. [72], Zhu et al. [73], Liu and Zheng [68], He et al. [69], Hu et al. [70], Shi et al. [74] Association, wireless power, and ISAC Multi-objective tradeoff explanation and search guidance Jing et al. [52], Zhang et al. [53], Wu et al. [71], Li et al. [28] Routing–scheduling TSP/VRP-style UAV routing Natural-language service request to routing constraints Montemanni and Dell’Amico [75], Joo et al. [76] Truck, vehicle, or mothership coordination Cross-platform synchronization explanation Joo et al. [76] Delivery, monitoring, inspection, and rescue Priority-aware dispatch briefing and schedule repair Yang et al. [77], Rigas et al. [78], Zhao et al. [79], Mao et al. [80], Pei et al. [81], Betti Sorbelli et al. [82] Scalable solvers Deep reinforcement learning assignment Policy explanation and compliance-aware critic Oubbati et al. [83], Moon et al. [84], Zhu et al. [85, 86], Hu et al. [87], Mao et al. [80], Gong et al. [6] Metaheuristics and hybrid optimization Solver selection, heuristic repair, and black-box search guidance Rinaldi et al. [88], Zai et al. [89], Mao et al. [80], Qin and Pournaras [90] Security and privacy Secure trajectory, relay, and jamming RAG-grounded rule checking and secrecy rationale Kang et al. [91], Zhang et al. [92], Dang-Ngoc et al. [93], Tedeschi et al. [94] Data, video, and access privacy Sensitive-context filtering and access-audit explanation George et al. [95], Tedeschi et al. [94], Huang et al. [96] Federated, blockchain, and location privacy Privacy-aware prompt construction and traceable logs Akram et al. [97], Pu [98], Huang et al. [96], Jagarlapudi et al. [99], Enayati et al. [100] The optimization core still enforces feasibility, but the surrounding LLM layer improves the ability to handle open-ended requests, incomplete operator input, and cross-layer tradeoffs [37, 19, 28, 87, 41]. IV-A Mobility–Communication Assignment The first class in Table I appears in UAV-assisted wireless networks, where the central dispatcher must decide not only where UAVs should fly but also which users, relays, sensors, or edge devices they should serve. In this setting, P, R, and A are tightly coupled: changing a UAV trajectory changes channel quality, which in turn changes feasible communication schedules, power allocation, and user association. The assignment problem therefore evolves from geometric path planning into joint mobility–resource control. Recent communication-oriented studies establish this coupling through trajectory, access, and resource scheduling. Liu and Zheng [68] optimize time-constrained data collection in UAV-enabled environmental monitoring, He et. al. [69] connect multi-UAV three-dimensional trajectories with offloading fairness, and Hu et. al. [70] jointly consider grouping, trajectory, and resource allocation in aerial–ground collaborative computing. ISAC-oriented work further couples mobility with sensing and communication quality through user association, beamforming, and target localization [52, 53, 71]. Across these works, the central dispatcher serves as the integration point where wireless links, sensing objectives, and UAV motion are made consistent. Data collection and mobile edge computing make the coupling broader because service feasibility depends on both contact time and computing or sensing load. Nguyen et. al. [72] use deep reinforcement learning for three-dimensional trajectory and data collection, Zhu et. al. [73] introduce a Transformer-based trajectory planner for age-of-information-aware IoT collection, and Shi et. al. [74] embed hierarchical reinforcement learning into trajectory planning and resource allocation. The common thread is that tasks cannot be assigned independently of communication or computation resources. Work on transfer learning in deep reinforcement learning further supports the need to reuse learned dispatch policies across changing environments [86]. Recent work further expands the set of coupled decisions by introducing sensing, ISAC, and LLM-guided multi-objective search. Zhang et. al. [53] coordinate trajectory, user association, and beamforming for multi-UAV ISAC systems. Wu et. al. [71] analyze how sensing quality and communication utility interact in trajectory design. Li et. al. [28] use LLM reasoning to guide multi-objective UAV-ISAC optimization. Multi-objective multi-agent decision-making provides a complementary reference point for coupled cooperative dispatch [87]. This progression shows why global assignment dispatch is attractive in communication-heavy missions: the central view can expose tradeoffs that are difficult to capture if trajectory, association, and resource allocation are solved separately. LLM-enabled assignment changes this class from a purely numerical optimization problem into a semantic-to-numerical translation problem. The central dispatcher may receive a request such as “prioritize emergency users while keeping sensing accuracy above a minimum level”. The LLM layer can retrieve policy or service documents, map the request to association, power, sensing, and trajectory constraints, and then select an optimization routine. Li et. al. [28] provide a concrete UAV-ISAC example in which LLM reasoning is used inside a decomposition-based multi-objective evolutionary algorithm, and related multi-agent decision-making work shows how cooperative objectives can be learned jointly [87]. This suggests that, for mobility–communication assignment, the LLM should act as a trade-off interpreter and search guide, while channel models and optimization solvers remain responsible for feasibility and Pareto quality. IV-B Routing, Scheduling, and Vehicle-Assisted Operations The second class shifts the central variable from wireless link quality to service ordering. In delivery, inspection, monitoring, and emergency response, the dispatcher mainly determines which UAV serves each task, in what sequence, and under which time, capacity, energy, or vehicle coordination constraints. These problems still use the same decision variables from Section I, but the dominant coupling is now among X, P, and S. Recent routing formulations provide the foundation. Montemanni et. al. [75], and Joo et. al. [76] study parallel UAV scheduling or drone-delivery routing variants with synchronization and endurance constraints. Their models extend TSP-style routing with UAV-specific requirements, showing how centralized assignment can convert multiple UAV sorties into a coordinated service sequence rather than a collection of independent paths. Vehicle-assisted systems add a moving depot, rider, or mothership to the same routing logic. Robust drone-vehicle routing with uncertain time windows further shows why the central dispatcher must synchronize two mobility layers: the slower ground carrier and the faster aerial task executors [76]. Application-specific scheduling then adds feasibility constraints that cannot be reduced to distance alone. Persistent and on-demand delivery systems consider rolling horizons, demand uncertainty, docking stations, dispatching points, and urban conflicts [77, 101, 102, 82, 81]. Monitoring and inspection systems add spatial, temporal, edge analytics, and charging constraints [78, 79, 103, 104, 105]. Emergency delivery and rescue further prioritize response time and mission urgency [106, 107]. These studies form a clear progression: basic routing chooses a feasible service sequence, vehicle-assisted dispatch synchronizes heterogeneous carriers, and application-driven scheduling embeds domain constraints such as monitoring windows, emergency priority, battery exchange, and airspace conflict avoidance. In LLM-era routing, the difficult step often precedes the VRP or TSP solver. Natural-language orders, emergency reports, inspection checklists, and operator preferences must be converted into time windows, priorities, vehicle compatibility rules, no-fly constraints, and synchronization requirements. Tool-oriented LLM agents provide a useful abstraction here: the language model plans which solver or data source to call, executes the tool sequence, and summarizes the result for the operator [19, 21]. This does not remove the need for exact or heuristic routing algorithms. Instead, it turns global assignment dispatch into a human-readable loop in which constraints can be inspected before the route is committed. IV-C Scalable Solvers and Learning-Enhanced Assignment Scalable solver design explains how the central assignment can be computed when exact optimization becomes too expensive. Large UAV dispatch models often combine integer assignment, non-convex trajectory constraints, communication coupling, and time windows. As the fleet size or task set grows, many studies complement or approximate exact solvers with learning-based or hybrid optimization methods. Deep reinforcement learning is useful when the dispatcher repeatedly faces dynamic states. Oubbati et. al. [83] adapt UAV base-station positions to urban vehicular traffic, Moon et. al. [84] learn multi-UAV trajectories for target tracking, and Zhu et. al. [85] jointly learn trajectory planning and cluster-head selection in wireless sensor networks. Mao et. al. [80] further use a double-level DRL structure for large-scale multi-UAV task scheduling. These methods keep a centralized coordination objective but use learned policies to avoid solving a full optimization problem from scratch at every decision epoch. Hybrid solvers are more common when feasibility rules are explicit, but the search space is combinatorial. Rinaldi et. al. [88] use a double-chromosome genetic scheduler with safe route planning and Zai et. al. [89] combine DRL and particle swarm optimization for emergency material transportation. Mao et. al. [80] further show how double-level DRL can scale multi-UAV task scheduling, illustrating the practical trend toward solvers that mix prediction, search, and local repair. LLM-guided search gives this solver layer a new form. Rather than using the LLM to output an assignment directly, the system can let the LLM propose decompositions, priority functions, neighborhood moves, or repair rules, and then score them with an external evaluator. FunSearch demonstrates this evaluator-guarded pattern for hard algorithmic problems by pairing LLM-generated programs with systematic scoring [41]. Toolformer and graph/tree-of-thought reasoning further indicate how LLMs can select external tools or branch over candidate decompositions before committing to a solver [20, 39, 22]. Self-refinement mechanisms can also revise candidate outputs through feedback [40]. For UAV assignment, this points to a practical architecture: classical solvers produce or verify feasible dispatch plans, while an LLM searches over solver configurations, local repair strategies, and explanations of why a candidate assignment failed. The advantage of these methods is speed and adaptability; the cost is weaker guarantees. Reward design, training distribution, heuristic parameters, and scenario-specific assumptions can strongly affect the quality and safety of the assignment. Scalable solver design is an implementation dimension of global assignment dispatch rather than a separate dispatch paradigm. It does not change the dispatch variables, but determines whether a centralized model can be solved in real time. IV-D Security, Privacy, and Trust-Aware Assignment Centralized assignment improves global coordination, but it also concentrates sensitive information. UAV states, user locations, task requests, video streams, trajectories, and communication metadata may all be collected by the dispatcher. Security- and privacy-aware assignment therefore appears as both an optimization objective and a system-design requirement. At the physical layer, security is often folded directly into trajectory and resource assignment. Kang et. al. [91] coordinate dual-UAV trajectories and power for bidirectional secure communication. Zhang et. al. [92] address collaborative secure relay communications in UAV swarms, while Dang-Ngoc et. al. [93] use cooperative friendly jamming to protect swarm-assisted communications. These methods treat secrecy not as a post-processing rule, but as a dispatch constraint coupled with P, R, and A. Data privacy and access control address a different part of the centralized risk. George et. al. [95] develop multi-party access control for UAV services, and Tedeschi et. al. [94] protect collision-avoidance coordination without exposing unnecessary motion information. The dispatch logic here is organizational as much as mathematical: the central platform must decide who may see which task, data stream, or UAV state. Distributed privacy mechanisms reduce dependence on raw centralized data. Akram et. al. [97] use conditional federated learning for post-disaster infrastructure monitoring. Pu [98] studies privacy-preserving and fault-tolerant data aggregation, and Huang et. al. [96] analyze privacy-aware remote identification and routing. Location privacy studies such as Jagarlapudi et. al. [99] and Enayati et. al. [100] further show that assignment plans and trajectories themselves may leak sensitive information. These works show that centralized dispatch should optimize performance together with data minimization, access control, and location obfuscation. The LLM layer intensifies these risks because prompts and retrieved contexts may contain user identities, mission intent, facility locations, or predicted trajectories. A privacy-aware meta-dispatcher therefore needs explicit filters before prompt construction, access control over retrieved documents, and audit logs for tool calls and generated explanations. RAG helps ground decisions in external documents, but it also makes provenance and document freshness part of the security boundary [37]. Thus, trustworthy global assignment in the LLM era requires joint control of optimization variables and information exposure. V Decentralized Agentic Dispatch Decentralized agentic dispatch describes a paradigm in which UAVs select tasks, adjust trajectories, and coordinate with nearby agents from local observations, local utility, compact semantic messages, and limited peer negotiation rather than from a complete centralized assignment. In the notation of Section I, the same decisions X, P, S, R, and A may still be involved, but they are produced through local policies, bidding, consensus, or distributed learning. Table IV organizes this paradigm from the lowest level of autonomous motion to swarm interaction, task ownership, and privacy-preserving autonomy. In the LLM era, the local agent can summarize nearby events, maintain short-term task memory, generate negotiation messages, reflect on failed local actions, and pass only verified high-level commands to conventional planners or controllers [23, 56, 21, 65]. Figure 4 illustrates this decentralized agentic logic. Fig. 4: Decentralized agentic dispatch lets local UAV agents select tasks from observations, utilities, short-term memory, and limited peer messages. TABLE IV: Taxonomy of representative decentralized agentic UAV dispatch methods and their LLM-agent roles. Agentic Layer Local Dispatch Role LLM-Agent Role Representative Methods Local decision layer Reinforcement-learning trajectory control Perception–memory–action loop for local intent interpretation Ebrahimi et al. [108], Nguyen et al. [72], Zhu et al. [73, 86] Active perception under uncertainty Multimodal observation summarizer and next-view planner Zhu and Chen [60], Zhu et al. [61], Kirillov et al. [63] Tracking, avoidance, and covert motion Local safety critic and intent-aware motion proposal Xu et al. [109], Zhang et al. [110], Huang et al. [111] Swarm coordination layer Decentralized coverage and exploration Peer-agent message compression and local consensus prompting blueZhao et al. [112], Huang et al. [113] Relative-state UAV interaction Relative-observation grounding for semantic communication Hasan et al. [114] MARL and distributed reinforcement learning Role-aware multi-agent reasoning with bounded messages Venturini et al. [115], Qin and Pournaras [90], Hu et al. [87], Alkahtani et al. [116], Luo et al. [29] Task ownership layer Consensus-based auctions Bid explanation and task-claim negotiation Mao et al. [80] Dynamic partial replanning Reflection-based local repair after state changes Qin and Pournaras [90] Privacy and trust layer Privacy-aware or covert trajectories Sensitive-context filtering before local reasoning Huang et al. [111], Chinthi-Reddy et al. [117] Private trajectory exchange and collision avoidance Minimal semantic message sharing among agents Sciancalepore and George [118], Tedeschi et al. [94] Private computation and federated learning Local memory isolation and federated agent learning Gao et al. [119], Akram et al. [97] V-A Local Perception and Motion Decisions Decentralized agentic dispatch begins with the individual UAV’s ability to turn local perception into motion and task-selection decisions. Unlike global assignment dispatch, where a central node decides the next task or route, the UAV locally evaluates whether a target is reachable, valuable, safe, or informative. This makes local trajectory control the first layer of the taxonomy in Table IV. Agent-level reinforcement learning provides a natural mechanism for this local decision layer. Ebrahimi et. al. [108] learn autonomous UAV trajectories for ground-object localization while considering energy, trajectory length, waypoints, and flight time. Nguyen et. al. [72] use deep reinforcement learning for three-dimensional trajectory and data-collection optimization, allowing the UAV to adapt its route and collection time online. Zhu et. al. [73] introduce a transformer-based trajectory planner for age-of-information-aware data collection, showing that sequence models can help local agents account for temporal freshness when deciding where to fly next. These studies share the same dispatch logic: the UAV does not wait for a complete route, but repeatedly maps current observations to an executable motion decision. In addition, a survey on transfer learning in deep reinforcement learning further highlights why reusable policies matter when local UAV agents move across missions, weather, or sensing domains [86]. Beyond learned local policies, recent semantic aerial navigation systems suggest that memory and behavior-tree structures can improve resilience by linking perception, memory, and action selection [120]. In an LLM-enabled agent, a multimodal LLM can further summarize what the UAV sees, infer whether a local target matches the mission intent, and pass a structured goal to a verified planner. PaLM-E demonstrates how language and embodied observations can be represented in a single reasoning interface [23]. UAV-specific benchmarks such as AerialVLN [55] and AeroDuo [57] show that aerial agents must ground instructions in altitude, viewpoint, and long-range spatial relations, while broader vision-language navigation studies [56, 27, 25, 26] provide mechanisms for constraint-aware and reasoning-based navigation. With such semantic mediation, the local decision layer goes beyond observation-to-action mapping and instead transforms observations into semantic states and verified goals. This semantic mediation also enables the agent to explicitly reason about uncertainty, which naturally extends local decision-making from action execution to information-seeking motion. Extending this semantic and uncertainty-aware decision framework, UAVs further engage in active perception, where movement is selected not only for task completion but also for information gain. Xu et. al. [109] combine obstacle avoidance and target tracking through DRL, Zhang et. al. [110] study online pursuit-evasion planning for multiple UAVs, and Huang et. al. [111] plan covert surveillance trajectories for mobile targets. Open-vocabulary, few-shot, and foundation segmentation models also matter here because they make local targets and region extraction more adaptable before the motion policy is invoked [60, 61, 63]. This progression suggests that, as uncertainty and adversarial movement increase, local autonomy must jointly reason about sensing value, safety, and mission utility. V-B Swarm Interaction and Distributed Learning Once multiple UAVs make local decisions, the main problem becomes coordination. Purely independent agents may duplicate coverage, collide, or chase the same target. The second layer of Table IV therefore focuses on local interaction: UAVs still act autonomously, but neighboring agents exchange limited information to make the swarm behavior coherent. Decentralized quadrotor swarm systems such as EGO-Swarm further demonstrate that local planning, inter-agent collision avoidance, and limited information exchange can support fully autonomous swarm operation in cluttered environments [121]. Coverage and exploration methods show the basic form of this interaction. Experimental platforms such as AirSwarm also highlight the importance of low-cost, reproducible multi-UAV systems for validating decentralized coordination, communication, and task-allocation algorithms in physical environments [122]. Zhao et. al. [112] use decentralized learning for energy-efficient multi-UAV content coverage. Huang et. al. [113] design biomimetic swarm exploration with UAV-to-UAV communication. These methods move beyond single-agent motion by letting each UAV adapt to local coverage gaps, neighbor behavior, and changing task ownership. Resource-constrained micro-UAV swarms further show that decentralized exploration can be achieved with minimal sensing and communication, which is important for dispatch policies under onboard compute, energy, and payload limits [123]. Relative-state interaction then makes the swarm less dependent on global coordinates or centralized maps. Hasan et. al. [114] study UAV-to-UAV interaction through three-dimensional relative-direction-based evolutionary computation. This type of local relation is important in low-altitude operations, where GPS errors, communication limits, and obstacle-rich environments may make a fully shared global state unreliable. Multi-agent reinforcement learning generalizes these ideas into a learned coordination policy. Venturini et. al. [115] use distributed reinforcement learning for flexible swarm control, and Qin and Pournaras [90] combine short-term distributed optimization with long-term reinforcement learning. Multi-objective multi-agent cooperative decision-making further connects local learning with coupled team objectives [87], while Alkahtani et. al. [116] add explainability and trust evaluation to secure FANET routing. Together these works show that decentralized agentic dispatch can scale from local motion to swarm-level coordination when agents learn not only what action is good for themselves, but also how that action affects nearby UAVs. For LLM-enabled swarms, the communication payload changes from raw state vectors to compact semantic messages. A UAV may report that a road segment is blocked, a target is uncertain, or a relay role is needed, and neighboring agents can reason over this message without sharing full video streams or full trajectories. Multi-LLM edge orchestration provides a useful model for this setting because different agents can specialize in perception, planning, communication, or trust assessment while remaining constrained by edge latency and privacy budgets [29]. The challenge is to make semantic communication bounded and verifiable: messages should be short enough for local links, grounded enough to avoid hallucination, and structured enough for consensus or safety modules to consume. V-C Bidding, Consensus, and Partial Replanning Learning-based coordination is flexible, but task ownership often needs a more explicit mechanism. Bidding and consensus methods provide this mechanism by letting UAVs evaluate task utilities, communicate bids, and agree on winners. A representative example is variable-swarm distributed auction assignment combined with sequential convex programming, where task ownership and trajectory feasibility are jointly handled through auction-based allocation and trajectory optimization [124]. This mechanism directly captures decentralized task claiming: a UAV obtains ownership because its local utility is competitive, rather than because a central dispatcher assigns the task. Recent work adapts consensus and bidding structures to more dynamic and heterogeneous UAV systems. Mao et. al. [80] use a double-level DRL structure for large-scale multi-UAV task scheduling, so learned policies can support local reassignment without rebuilding the entire allocation. For UAV dispatch, the importance is not only the auction format but the structure it imposes: local cost estimates such as travel distance, energy, deadline, risk, and task reward become the bridge between individual decisions and swarm-level task ownership. LLM agents can strengthen this ownership layer by making bids and replanning decisions explainable. A local agent can translate its energy state, distance, risk, and sensor capability into a bid rationale. It could ask neighboring agents for missing context and reflect on why a previous claim failed [21, 65]. However, the final ownership decision should still be enforced by the auction or consensus protocol. This separation is necessary because language can improve negotiation and diagnosis, whereas agreement, conflict resolution, and deadline feasibility require explicit distributed rules. V-D Privacy-Preserving and Trustworthy Autonomy Decentralization reduces dependence on a central server, but it does not remove privacy and security risks. Instead, the risk moves to peer-to-peer messages, trajectory sharing, collision-avoidance protocols, cooperative learning, and local sensing. The final layer of Table IV therefore treats privacy and trustworthiness as part of autonomous dispatch, not as an external communication add-on. Some studies embed privacy directly into trajectory behavior. Huang et. al. [111] plan covert surveillance trajectories that maintain monitoring performance while reducing detectability, and Chinthi-Reddy et. al. [117] propose DarkSky to reduce privacy invasion in target tracking. These works show that privacy can affect P itself: the path should serve the task while avoiding unnecessary exposure of either the UAV or non-target areas. Other works protect coordination messages. Sciancalepore and George [118] develop privacy-preserving trajectory matching for autonomous UAVs, and Tedeschi et. al. [94] propose PPCA for privacy-preserving collision avoidance. Their common concern is that safe coordination often requires sharing future motion, yet future motion can reveal sensitive locations or mission intent. Privacy-aware decentralized agentic dispatch must therefore decide what information is necessary for coordination and what should remain hidden. Privacy-preserving computation and learning complete the picture. Gao et. al. [119] study UAV-centric privacy-preserving computation offloading in multi-UAV MEC with local differential privacy, while Akram et. al. [97] use conditional federated learning for post-disaster drone services. These methods point to a privacy-preserving form of autonomy in which UAVs can improve local policies or offloading decisions without revealing raw observations, trajectories, or mission data. LLM-enabled autonomy adds prompt-level and memory-level privacy risks. Local summaries may expose sensitive locations even when raw images are not transmitted, and reflective memories may preserve failed routes, targets, or operator instructions longer than necessary. Therefore, a trustworthy decentralized agent should separate private sensor memory from shareable semantic memory, filter messages before negotiation, and bind every LLM-proposed action to a safety or privacy check. Privacy-preserving dispatch should therefore cover not only encrypted communication, but also controlled language generation, memory isolation, and minimal disclosure during peer coordination. VI Divide-and-Conquer Dispatch Divide-and-conquer dispatch reduces a large coupled UAV mission by splitting the problem into smaller subproblems across different dimensions such as decisions, objectives, space, time, groups, roles, and communication links. The resulting subproblems are solved separately and then coordinated through recomposition or verification mechanisms. Unlike pipeline dispatch, divide-and-conquer dispatch does not assume a fixed module sequence; unlike global assignment dispatch, it does not rely on one global solver. Its core question is where to draw the boundaries so that trajectory, schedule, resource, association, and security decisions remain locally tractable while still producing a coherent global dispatch plan. Table V summarizes representative methods according to the main decomposition principle they use. In the LLM era, the decomposition boundary can also be semantic and computational: knowledge may be split across retrieval sources, agent roles may be assigned by capability, memory may be scoped to regions or subteams, and reasoning may be shared between onboard models, edge servers, and cloud models. This makes divide-and-conquer dispatch a natural bridge between classical decomposition and multi-agent LLM orchestration [19, 29, 21, 22]. Figure 5 illustrates this partition-and-coordinate logic. Fig. 5: Divide-and-conquer dispatch partitions a large mission into local subproblems, solves them separately, and reconciles them through a coordinator. TABLE V: Taxonomy of representative divide-and-conquer dispatch methods and their LLM-era orchestration roles. Decomposition Principle Dispatch Role LLM-Era Orchestration Role Representative Methods Decision and objective decomposition Trajectory, energy, and link tradeoff Objective decomposer and Pareto-search guide Javed et al. [125], Meng et al. [126], Jing et al. [52] Secure trajectory, power, and robustness Security-rule retrieval and feasibility critic Kang et al. [91], Zhang et al. [92], Dang-Ngoc et al. [93] Temporal, connectivity, and ISAC objectives Communication–sensing tradeoff explanation Meng et al. [126], Jing et al. [52], Zhang et al. [53], Li et al. [28] Spatial and regional decomposition Three-dimensional and multi-region coverage Semantic map and region partitioning Xiong et al. [43], Li et al. [42] Emergency waypoints and rescue paths Disaster-scene grounding and priority partitioning Liu et al. [127], Liu et al. [30], Lin et al. [24, 56], Xiong et al. [43] Inspection and agriculture scene modeling Multimodal region labeling and task extraction Liu et al. [127], Liu et al. [30], Zhang et al. [9], Zhu and Chen [60], Kirillov et al. [63] Swarm, role, and function decomposition Grouping and resource collaboration Role-specialized multi-agent planner Hu et al. [70], Luo et al. [29], Alkahtani et al. [116] Search stages and spatio-temporal planning Stage-specific agent memory and progress monitor Wu et al. [128], Gong et al. [6] Formation, recharging, and farm operations Skill-library reuse for repeated sub-missions Alyassi et al. [50], Ming et al. [129] Security and communication decomposition Secure relay and friendly jamming Policy-aware security sub-agent and audit log Zhang et al. [92], Dang-Ngoc et al. [93] Mobility–power robust security Adversarial-context summarization and robust repair Kang et al. [91], Zhang et al. [92] QoS and relay-aware mission execution Communication-state retrieval for local recomposition Yanmaz et al. [48] VI-A Trajectory–Communication and Objective Decomposition The first class in Table V decomposes strongly coupled trajectory-communication objectives. In UAV networks, mobility affects channel quality, energy consumption, sensing accuracy, secrecy, and delay. Solving all variables jointly is often non-convex, so many methods separate the problem into trajectory, power, beamforming, time-frame, or objective subproblems and coordinate them iteratively. Energy- and communication-aware studies illustrate the basic decomposition. Javed et. al. [125] optimize outage probability and trajectory length as related but distinguishable objectives in disaster scenarios, while recent ISAC work decomposes trajectory, beamforming, association, and target-localization quality [126, 52, 53]. These works use decomposition to expose why a path that is short geometrically may not be efficient once communication or sensing performance is considered. Secure communication makes the objective split sharper. Kang et. al. [91] coordinate dual-UAV trajectories and transmit power for bidirectional secure communication, Zhang et. al. [92] study collaborative secure relay communication under colluding eavesdroppers, and Dang-Ngoc et. al. [93] decompose relay and friendly jamming functions in swarm UAV-assisted networks. In these settings, security is not an independent layer after planning. It changes how mobility, role, and resource variables should be separated and recombined. Temporal and sensing-oriented decomposition further extends the same idea. Meng et. al. [126] use frame-based iterative optimization for UAV trajectory and beamforming in integrated periodic sensing and communication. Jing et. al. [52] design multi-stage UAV trajectories for joint communication and target localization, while Li et. al. [28] show how LLM reasoning can guide the decomposition of multi-objective UAV-ISAC search. The progression from energy and rate to secrecy, connectivity, and localization shows that objective decomposition is useful whenever one trajectory must serve several competing system functions. LLM-era objective decomposition makes the split more explicit to human operators and downstream solvers. The LLM can translate mission intent into named subobjectives, retrieve relevant constraints, and decide whether a mobility, communication, sensing, privacy, or energy solver should be called first. In UAV-ISAC optimization, LLM reasoning has already been used to guide multi-objective search [28]. More generally, evaluator-guarded LLM search, as in FunSearch [41], suggests a way to generate candidate decomposition rules or local repair programs while retaining external scoring. Thus, the LLM contribution is not to merge all objectives into one opaque prompt, but to expose and coordinate the subproblems that must still be verified numerically. VI-B Spatial and Regional Decomposition The second class partitions the operating environment or task area. This is especially natural in coverage, inspection, agriculture, construction, and disaster response, where the mission spans a large region and each subregion can be planned locally. Spatial decomposition changes the dispatch question from “What is the best global route?” to “How should the space be divided, and how should local routes be coordinated across boundaries?” Frontier- or coverage-guided planners further support this view by decomposing unknown environments into informative regions and then coordinating local exploration trajectories across spatial boundaries [130]. Coverage planning provides the clearest example. Xiong et. al. [43] optimize collaborative coverage paths for multi-region post-disaster assessment, using region partitioning, task allocation, and coverage routing to improve swarm efficiency. Li et. al. [42] similarly decouple multi-region coverage while incorporating task-priority awareness. These methods show that spatial decomposition is not merely a computational simplification. It also defines which UAV is responsible for which part of the mission. Emergency and domain-specific missions add priorities and sensing requirements to the spatial split. Large-scale exploration in GPS-denied cave environments further demonstrates that spatial decomposition must be coupled with mapping, localization, communication constraints, and local replanning when the global state is incomplete [131]. Liu et. al. [127] use 3D reconstruction to guide refinement inspection at construction sites, Liu et. al. [30] convert inspection evidence into multi-agent subtasks, and Xiong et. al. [43] partition post-disaster assessment regions for collaborative coverage. In these studies, the environment model becomes part of the decomposition: better region or scene understanding leads to better local dispatch subproblems. Spatial decomposition is also where multimodal LLMs become directly relevant to UAV dispatch. A region may be described by satellite imagery, onboard video, maps, inspection manuals, and human incident reports. An LLM or multimodal LLM can label zones, extract candidate targets, and produce region-level task descriptions before a coverage or routing solver runs. PaLM-E provides a general embodied multimodal precedent [23]. Recent VLM studies and benchmarks clarify how visual-language grounding should be evaluated [9, 10, 11, 13]. Open-vocabulary and few-shot segmentation studies support region-level perception [60, 61]. AerialVLN, AeroDuo, and large-model landmark discovery adapt language grounding to navigation [55, 57, 24, 56, 27]. LLM-informed drone inspection further shows how textual and visual evidence can be turned into multi-agent inspection subtasks [30]. Therefore, spatial decomposition in LLM-enabled dispatch is not merely geometric partitioning; it also involves semantic region construction followed by verifiable local planning. Spatial decomposition is effective but not free. Boundaries between regions may create duplicated coverage, missed targets, relay gaps, or conflicts over shared charging stations and air corridors. Therefore, the success of this class depends on both the partition rule and the coordination rule that reconnects local routes into a global mission. VI-C Swarm, Role, and Functional Decomposition The third class decomposes the UAV fleet itself. Instead of assigning all UAVs to a single large optimization problem, the system forms groups, roles, or functional modules. This approach is useful when UAVs have heterogeneous payloads, communication ranges, endurance levels, or mission responsibilities. Group-level decomposition appears in aerial–ground computing and networked service systems. Hu et. al. [70] jointly optimize UAV grouping, trajectories, and resource allocation for swarm-assisted aerial–ground collaborative computing. Luo et. al. [29] extend this idea to multi-LLM edge orchestration, where model roles and placement must be matched to communication and trust constraints. These works use grouping to reduce coordination dimension while preserving enough structure for resource allocation. Search and urban-airspace planning show another form of functional decomposition. Wu et. al. [128] study swarm-based four-dimensional path planning, decomposing dense urban coordination over space and time to reduce conflicts. Gong et. al. [6] add an LLM-aware compliance layer for low-altitude trajectory planning, making regulatory constraints part of the decomposition. Here, the subproblems are both group-specific and stage-specific, separating mission stages from time-dependent safety constraints. Role-based decomposition is especially useful in long-duration or hostile missions. Alyassi et. al. [50] decompose persistent UAV missions into flight execution and autonomous recharging decisions. Ming et. al. [129] compare swarm control mechanisms for unmanned farms, highlighting the separation of navigation, perception, visual collaboration, and trajectory planning. These examples show that divide-and-conquer dispatch can be organizational: the mission is simplified by deciding which UAV or subsystem owns which function. LLM-agent systems make this organizational view more concrete. A fleet can be decomposed into role-specific agents: a perception agent summarizes aerial observations, a planning agent proposes routes, a communication agent checks connectivity, a privacy agent filters messages, and a verifier rejects unsafe proposals. HuggingGPT provides a general tool-orchestration pattern for assigning subtasks to specialized models [19], while multi-LLM edge intelligence studies how such roles can be placed across devices, edge servers, and cloud resources under trust and latency constraints [29]. For UAV dispatch, role decomposition therefore includes both physical roles among UAVs and computational roles among models. VI-D Security- and Communication-Oriented Decomposition Security and communication can also define the decomposition boundary. In many missions, some UAVs should serve as relays, some should jam eavesdroppers, some should sense or collect data, and others should maintain connectivity. Decomposing these roles allows the dispatch system to optimize specialized subproblems while maintaining a shared security or communication objective. Secure relay and friendly jamming provide direct examples. Zhang et. al. [92] study UAV-swarm-enabled collaborative secure relay communication under colluding eavesdroppers, coordinating relay trajectories and resources to improve secrecy capacity. Dang-Ngoc et. al. [93] separate relay communication and cooperative friendly jamming in swarm UAV-assisted networks. These methods are divide-and-conquer in a functional sense: communication support and security protection are assigned to different roles and then coordinated. The mobility-power decomposition discussed earlier also becomes more important under adversarial or uncertain channels. Kang et. al. [91] and Zhang et. al. [92] show that trajectory, power control, relay role, and robustness can be optimized as connected subproblems rather than one opaque secure-dispatch model. This separation clarifies which part of the solution improves channel geometry, which part controls signal exposure, and which part protects against uncertainty. Communication-oriented decomposition is also visible in rescue and hostile-environment planning. Yanmaz et. al. [48] separate search and relay roles to maintain connectivity during multi-target search. It suggests that communication should not be treated only as a constraint checked after planning; it can be a decomposed dispatch role that shapes where UAVs fly and how the mission is divided. LLM orchestration adds another security boundary. A security sub-agent may retrieve airspace or communication policies, summarize adversarial context, or explain why a relay or jamming role is needed, but it can also introduce prompt injection, stale evidence, or overconfident recommendations. ReAct-style tool use is helpful only if every retrieved rule, generated explanation, and proposed role assignment is logged and checked by a non-LLM verifier [21]. Thus, security-oriented decomposition must cover both radio and trajectory security and the trustworthiness of the language-mediated coordination process. VII Synthesis and Future Directions in the LLM Era Sections I–VI reviewed four dispatch paradigms and embedded LLM-related mechanisms inside each paradigm. Read together, these paradigms suggest that LLMs are not simply an additional module placed after optimization or control. They act across the dispatch stack, linking operator intent, mission knowledge, multimodal perception, tool invocation, solver selection, and human oversight. Optimization, control, communication theory, and safety verification still define the physical limits of a mission; the LLM layer changes how these components are selected, connected, explained, and adapted during real operations [37, 19, 23, 29]. VII-A Cross-Paradigm Synthesis The four paradigms remain useful abstractions, but their boundaries become less rigid once semantic reasoning is introduced. A practical LLM-enabled dispatch system may begin as a pipeline when an operator describes a mission, behave like a global assignment system when fleet resources must be matched to tasks, shift toward agentic Grab-and-Go when nearby UAVs react to local events, and rely on divide-and-conquer when the mission becomes too large for a single planner. The dispatch paradigm is therefore not only a static design choice; it can also become a mode selected according to mission scale, connectivity, uncertainty, and safety requirements. Because pipeline dispatch already exposes intermediate artifacts, it can absorb LLMs without hiding the intermediate decisions. Natural-language and multimodal inputs can be translated into task graphs, retrieval calls, solver inputs, trajectory candidates, and monitoring reports, with each artifact checked before the next stage executes [37, 19, 6]. Global assignment uses the same semantic layer at a different level of abstraction. When a central dispatcher faces heterogeneous objectives, an LLM can summarize fleet state, retrieve relevant rules, compare competing assignments, and make the reasoning behind a route or association legible to a human operator. The actual assignment still has to be produced or verified by explicit optimization and constraint models, since global feasibility cannot be inferred from linguistic plausibility alone [28]. When dispatch becomes decentralized, the LLM is no longer only an interpreter for a central plan. In decentralized agentic dispatch, an edge-hosted or lightweight onboard LLM can interpret nearby events, compress negotiation messages, and help a UAV decide whether to claim, defer, or reject a task. This is valuable under partial observation, yet it also makes latency, message size, local memory, and prompt-level attack surfaces part of the dispatch design rather than secondary implementation details [30, 10, 14, 29]. Divide-and-conquer dispatch extends this local reasoning to mission structure. Regions, roles, objectives, knowledge sources, model capabilities, and edge-cloud placement can all serve as partition criteria, but the local semantic decisions must eventually be recombined into a physically consistent fleet plan [29]. This comparison suggests a practical boundary for LLM use in UAV dispatch. The LLM layer is most defensible when it remains above direct flight control, where it can interpret missions, decompose tasks, orchestrate solvers, critique semantic consistency, and explain decisions to operators. Low-level actions still need retrieval grounding, tool constraints, simulation, and runtime monitoring. VII-B Open Challenges A failure can enter the dispatch process before any trajectory or schedule is generated. An LLM may invent a constraint, misread a map description, ignore a battery or weather limit, or produce a plan that sounds coherent but cannot be executed. In UAV missions, such errors are especially damaging because semantic interpretation affects task allocation, route generation, communication scheduling, and operator approval. Retrieval-augmented generation reduces this risk only when the retrieved context is relevant, current, and faithfully used by the model [132, 133, 11, 14]. Retrieved knowledge also has to survive the physical layer. LLMs do not model flight dynamics, propulsion energy, wind disturbance, sensing quality, channel variation, or collision avoidance with the precision required by UAV operation. Their outputs should therefore be treated as structured candidates for downstream motion planners, communication models, digital twins, and runtime monitors. The difficult part is not merely adding these checks, but making the semantic-to-physical interface fast enough for low-altitude missions where the environment changes during execution. In the field, model choice is rarely independent of connectivity, energy, and mission risk. Cloud LLMs offer stronger reasoning, but they may be unavailable or too slow in remote inspection, emergency response, and contested communication environments. Onboard models reduce latency and protect sensitive data, yet they are smaller, more energy constrained, and more brittle under complex instructions. Model compression, split inference, edge-cloud fallback, and task-aware model selection should therefore be part of dispatch policy itself. Once multiple UAVs participate, the same resource problem appears as a coordination problem: LLM agents can make swarm behavior easier to explain, but unconstrained dialogue can also introduce verbose negotiation, inconsistent beliefs, and hidden prompt dependencies. Scalable multi-UAV use will require structured message schemas, role-based communication, consensus rules, and local validators. The same integration also widens the attack and evaluation surface. The LLM layer may touch mission descriptions, locations, images, videos, communication logs, and infrastructure data, so prompt injection, data leakage, model inversion, unsafe tool use, and ambiguous responsibility boundaries have to be addressed together with access control, retrieval filtering, privacy-preserving inference, and tamper-resistant logs. Evaluation also has to move beyond mission time, energy, coverage, throughput, and path length. A benchmark for LLM-enabled dispatch should test semantic correctness, constraint satisfaction, plan executability, explanation quality, robustness to prompt attacks, sim-to-real transfer, and the burden placed on human approval. A plan that saves energy while violating a retrieved airspace rule should be judged as a dispatch failure rather than as an efficient solution. VII-C Future Research Directions The same interface view motivates the research directions summarized in Table VI, where semantic reasoning meets verified dispatch behavior. TABLE VI: Future research directions for LLM-enabled UAV dispatch. Hotspot Dispatch Focus Key Research Issues RAG-grounded dispatch knowledge bases Retrieve maps, regulations, weather, mission logs, and communication context for planning Source freshness, retrieval faithfulness, traceable constraint generation LLM-tool-solver co-design Connect LLM reasoning with optimizers, simulators, planners, and safety monitors Stable interfaces, feasibility checking, repairable plan representations Decentralized agentic dispatch Support local semantic interpretation, task claiming, peer negotiation, and replanning Latency, compact protocols, prompt safety, distributed verification Semantic divide-and-conquer Partition missions by regions, roles, objectives, knowledge sources, and model placement Boundary conflicts, load balancing, cross-agent consistency Multimodal aerial reasoning Fuse aerial images, videos, maps, text instructions, and sensor summaries Scene grounding, hallucination control, sim-to-real transfer Edge multi-LLM orchestration Assign reasoning tasks across onboard, edge, and cloud models Model selection, fallback policies, energy and bandwidth limits Safety-certified human governance Expose uncertainty, cite evidence, request approval, and log decisions Accountability, HITL design, auditability, regulatory compliance Future dispatch systems will need knowledge layers that combine maps, airspace rules, weather feeds, maintenance records, mission logs, and communication-state summaries. Such corpora are useful only if their content can be traced back to source material and transformed into constraints that solvers and monitors can check. Retrieval-augmented generation therefore has to be studied together with LLM-tool-solver co-design. A model may propose and revise a plan, but optimizers, simulators, route planners, communication models, and safety monitors must decide whether that plan remains feasible. In missions such as inspection, environmental monitoring, search and rescue, and post-disaster assessment, dispatch often depends on what the platform sees, not only on what the operator requests. Damaged infrastructure, blocked roads, moving targets, smoke, terrain changes, and temporary crowding can all change task priority. MLLMs can connect images, videos, maps, language instructions, and sensor streams, but their outputs must be grounded against geometry, time, and mission context. The model placement problem then becomes unavoidable. A small onboard model may summarize local evidence, an edge model may coordinate a nearby group of UAVs, and a cloud model may support broader planning when connectivity is available. Dispatch must decide when to invoke each level and how to recover when a model, link, or sensor stream fails. For large fleets, local reaction and global recomposition have to be studied together. Local agents need compact negotiation, prompt-safe replanning, and distributed verification; decomposed missions need reliable boundaries for regions, roles, objectives, knowledge sources, and model placement. Separating decentralized agentic dispatch from semantic divide-and-conquer dispatch is useful for taxonomy, but real systems will often mix them: a UAV may claim a nearby task, exchange a short justification with neighbors, and then rely on a higher-level decomposition policy to maintain load balance and mission consistency. As these systems move toward deployment, governance can no longer be added after planning has finished. LLM-enabled dispatch should expose uncertainty, cite retrieved evidence, present alternatives, and request human approval when safety, privacy, or regulation is at stake. The architecture then remains hybrid in a deliberate way: the semantic layer interprets intent and organizes knowledge, while the symbolic and physical layers optimize assignments, trajectories, schedules, resources, and associations under explicit constraints. The long-term value of LLMs in UAV dispatch will come from grounded, verifiable, privacy-aware, and human-governed interfaces, not from replacing dispatch algorithms with unconstrained text generation. VIII Conclusion This survey has reviewed UAV dispatch in the LLM era from a system-level perspective. We first established a unified view of UAV dispatch by connecting general dispatch models, unmanned-system dispatch, and UAV dispatch under a common set of decisions, including task assignment, trajectory planning, scheduling, resource allocation, and cooperation. Based on how mission intent is transformed into executable fleet-level decisions, we organized existing studies into four representative paradigms: pipeline dispatch, global assignment dispatch, decentralized agentic dispatch, and divide-and-conquer dispatch. These paradigms capture different dispatch logics: pipeline dispatch emphasizes staged mission processing, global assignment dispatch exploits centralized state for joint optimization, decentralized agentic dispatch supports local autonomy and decentralized coordination, and divide-and-conquer dispatch reduces large-scale coupling through spatial, temporal, functional, role-based, knowledge-source, or model-placement decomposition. Across these paradigms, recent studies have advanced UAV dispatch in trajectory planning, routing, task scheduling, energy management, communication optimization, privacy protection, semantic orchestration, and early LLM-agent coordination. The emergence of LLMs and multimodal foundation models introduces a new semantic layer for UAV dispatch. LLMs can interpret natural-language missions, retrieve contextual knowledge, decompose high-level objectives, orchestrate external solvers, support multi-agent negotiation, explain dispatch decisions, and assist human-in-the-loop revision. However, they should not be viewed as replacements for optimization, control, communication theory, or safety verification. In safety-critical low-altitude operations, language-based reasoning must be grounded in reliable evidence, transformed into structured dispatch candidates, and verified by specialized models, simulators, controllers, and runtime monitors. Future UAV dispatch systems should therefore move toward hybrid semantic–symbolic–physical architectures that integrate LLM-enabled reasoning with mathematical optimization, multi-agent learning, privacy-preserving computation, digital twins, and formally verified safety mechanisms. Such architectures will be essential for building scalable, trustworthy, and human-governed UAV dispatch systems in complex low-altitude environments. References [1] N. Pongsakornsathien N. E. Safwat et al. (2025) Advances in low-altitude airspace management for uncrewed aircraft and advanced air mobility. Prog. Aerosp. Sci. 154, p. 101085. External Links: Document Cited by: TABLE I, §I. [2] W. Lyu, H. Wang, Z. Hong, and et. al. (2026) Mobility-aware order relay in last-mile delivery. IEEE Trans. Mob. Comput. (), p. 1–15. External Links: Document Cited by: §I. [3] X. Tagle Casapia, R. Cardenas-Vigo, D. Marcos, and et. al. (2025) Effective integration of drone technology for mapping and managing palm species in the peruvian amazon. Nat. Commun. 16 (1), p. 3764. External Links: Document Cited by: §I. [4] T. Manzini, P. Perali, and R. R. Murphy (2026) Deploying rapid damage assessments from suas imagery for disaster response. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 40336–40343. External Links: Document Cited by: §I. [5] J. Ren, N. Song, X. Li, and et. al. (2025) Rolling dispatch for aavs inspection based on task adaptive clustering. IEEE T. Autom. Sci. Eng. 22, p. 14108–14119. External Links: Document Cited by: §I, §I-C, TABLE I, TABLE I, TABLE I, TABLE I. [6] Y. Gong, J. Fan, R. Zhang, and et. al. (2026) Safe and economical uav trajectory planning in low-altitude airspace: a hybrid drl-llm algorithm with compliance awareness. IEEE Trans. Mob. Comput., p. 1–17. External Links: Document Cited by: §I, §I, §I-B, TABLE I, TABLE I, §I, TABLE I, §VI-C, TABLE V, §VII-A. [7] P. Agarwal, S. Kumar, J. Ryde, and et. al. (2013) Towards a swarm of agile micro quadrotors. In Robotics: Science and Systems VIII, Vol. , p. 217–224. External Links: Document Cited by: §I. [8] K. Han, Y. Wang, H. Chen, and et. al. (2023) A survey on vision transformer. IEEE TPAMI 45 (1), p. 87–110. External Links: Document Cited by: §I, §I-B. [9] J. Zhang, J. Huang, S. Jin, and et. al. (2024) Vision-language models for vision tasks: a survey. IEEE TPAMI 46 (8), p. 5625–5644. External Links: Document Cited by: TABLE I, §I, §I, §I-B, §VI-B, TABLE V. [10] P. Xu, W. Shao, K. Zhang, and et. al. (2025) LVLM-EHub: a comprehensive evaluation benchmark for large vision-language models. IEEE TPAMI 47 (3), p. 1877–1893. External Links: Document Cited by: §I, §VI-B, §VII-A. [11] J. Yang, M. Jiang, and Q. Zhao (2026) Defying distractions in multimodal tasks: a novel benchmark for large vision-language models. IEEE TPAMI 48 (6), p. 6314–6331. External Links: Document Cited by: §I, §VI-B, §VII-B. [12] Z. Huang, S. Zhong, P. Zhou, and et. al. (2025) A causality-aware paradigm for evaluating creativity of multimodal large language models. IEEE TPAMI 47 (5), p. 3830–3846. External Links: Document Cited by: §I. [13] Y. Zhang, Q. Wen, C. Fu, and et. al. (2026) Beyond LLaVA-HD: diving into high-resolution multimodal large language models. IEEE TPAMI 48 (5), p. 5493–5504. External Links: Document Cited by: §I, §VI-B. [14] S. Wang, X. Cao, J. Zhang, and et. al. (2026) VLBiasBench: a comprehensive benchmark for evaluating bias in large vision-language model. IEEE TPAMI, p. 1–14. External Links: Document Cited by: §I, §VII-A, §VII-B. [15] A. Radford, J. W. Kim, C. Hallacy, and et. al. (2021) Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 139, p. 8748–8763. Cited by: §I, §I-B. [16] J. Li, D. Li, S. Savarese, and et. al. (2023) BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 202, p. 19730–19742. Cited by: §I, §I-B. [17] H. Liu, C. Li, Q. Wu, and et. al. (2023) Visual instruction tuning. In Advances in Neural Information Processing Systems, Vol. 36. Cited by: §I. [18] Y. Chang, X. Wang, J. Wang, and et. al. (2024) A survey on evaluation of large language models. ACM T INTEL SYST TEC 15 (3), p. 1–45. External Links: Document Cited by: TABLE I, §I, §I. [19] Y. Shen, K. Song, X. Tan, and et. al. (2023) HuggingGPT: solving ai tasks with chatgpt and its friends in hugging face. In Advances in Neural Information Processing Systems, Vol. 36. External Links: Link Cited by: §I, §I-A, §I, §IV-B, §IV, §VI-C, §VI, §VII-A, §VII. [20] T. Schick, J. Dwivedi-Yu, R. Dessì, and et. al. (2023) Toolformer: language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, Vol. 36. Cited by: §I, §I-C, §IV-C. [21] S. Yao, J. Zhao, D. Yu, and et. al. (2023) ReAct: synergizing reasoning and acting in language models. In International Conference on Learning Representations, Cited by: §I, §I-A, §I, §IV-B, §V-C, §V, §VI-D, §VI. [22] M. Besta, N. Blach, A. Kubicek, and et. al. (2024) Graph of thoughts: solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, p. 17682–17690. Cited by: §I, §I-C, §IV-C, §VI. [23] D. Driess, F. Xia, M. S. M. Sajjadi, and et. al. (2023) PaLM-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 202, p. 8469–8488. External Links: Link Cited by: §I, §I-A, §I-B, §V-A, §V, §VI-B, §VII. [24] B. Lin, Y. Nie, Z. Wei, and et. al. (2024) Correctable landmark discovery via large models for vision-language navigation. IEEE TPAMI 46 (12), p. 8534–8548. External Links: Document Cited by: §I, §I-A, TABLE I, §VI-B, TABLE V. [25] T. Wang, X. Zhao, W. Cai, and et. al. (2026) ImagineNav++: prompting vision-language models as embodied navigator through scene imagination. IEEE TPAMI, p. 1–18. External Links: Document Cited by: §I, §I-A, §V-A. [26] B. Lin, Y. Nie, K. L. Zai, and et. al. (2026) EvolveNav: empowering LLM-based vision-language navigation via self-improving embodied reasoning. IEEE TPAMI, p. 1–15. External Links: Document Cited by: §I, §I-A, TABLE I, §V-A. [27] K. Chen, D. An, Y. Huang, and et. al. (2025) Constraint-aware zero-shot vision-language navigation in continuous environments. IEEE TPAMI 47 (11), p. 10441–10456. External Links: Document Cited by: §I, §V-A, §VI-B. [28] H. Li, M. Xiao, K. Wang, and et. al. (2025) Large language model based multi-objective optimization for integrated sensing and communications in uav networks. IEEE Wireless Commun. Lett. 14 (4), p. 979–983. External Links: Document Cited by: §I, §I-B, TABLE I, §IV-A, §IV-A, TABLE I, §IV, §VI-A, §VI-A, TABLE V, §VII-A. [29] H. Luo, Y. Liu, R. Zhang, and et. al. (2025) Toward edge general intelligence with multiple-large language model (multi-llm): architecture, trust, and orchestration. IEEE Trans. Cognit. Commun. Networking 11 (6), p. 3563–3585. Cited by: §I, §V-B, TABLE IV, §VI-C, §VI-C, TABLE V, §VI, §VII-A, §VII. [30] J. Liu, H. Li, C. Chai, and et. al. (2025) A llm-informed multi-agent ai system for drone-based visual inspection for infrastructure. Adv. Eng. Inform. 68, p. 103643. External Links: Document Cited by: §I, §I-C, §VI-B, §VI-B, TABLE V, TABLE V, §VII-A. [31] P. Shukla, S. Shukla, and A. Kumar Singh (2025) Trajectory-prediction techniques for unmanned aerial vehicles (uavs): a comprehensive survey. IEEE Commun Surv Tut 27 (3), p. 1867–1910. External Links: Document Cited by: TABLE I, §I. [32] S. Ghambari, M. Golabi, L. Jourdan, and et. al. (2024) UAV path planning techniques: a survey. RAIRO-OPER RES 58 (4), p. 2951–2989. External Links: Document Cited by: TABLE I, §I. [33] P. Cao, L. Lei, S. Cai, and et. al. (2024) Computational intelligence algorithms for uav swarm networking and collaboration: a comprehensive survey and future directions. IEEE COMMUN SURV TUT 26 (4), p. 2684–2728. External Links: Document Cited by: TABLE I, §I. [34] J. Tang, H. Duan, and S. Lao (2023) Swarm intelligence algorithms for multiple unmanned aerial vehicles collaboration: a comprehensive review. ARTIF INTELL REV 56 (5), p. 4295–4327. External Links: Document Cited by: TABLE I, §I. [35] S. Sciancalepore (2024) Privacy and confidentiality issues in drone operations: challenges and road ahead. IEEE Network 38 (6), p. 227–233. External Links: Document Cited by: TABLE I, §I. [36] T. Guo, X. Chen, Y. Wang, and et. al. (2024) Large language model based multi-agents: a survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, p. 8048–8057. Note: Survey Track External Links: Document, Link Cited by: TABLE I, §I. [37] P. Lewis, E. Perez, A. Piktus, and et. al. (2020) Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, Vol. 33. External Links: Link Cited by: §I-C, §I-A, §I, §IV-D, §IV, §VII-A, §VII. [38] J. Wei, X. Wang, D. Schuurmans, and et. al. (2022) Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Cited by: §I-C. [39] S. Yao, D. Yu, J. Zhao, and et. al. (2023) Tree of thoughts: deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, Vol. 36. Cited by: §I-C, §IV-C. [40] A. Madaan, N. Tandon, P. Gupta, and et. al. (2023) Self-refine: iterative refinement with self-feedback. NeurIPS 36, p. 46534–46594. Cited by: §I-C, §I-C, §IV-C. [41] B. Romera-Paredes, M. Barekatain, A. Novikov, and et. al. (2024) Mathematical discoveries from program search with large language models. Nature 625, p. 468–475. External Links: Document, Link Cited by: §I-C, §IV-C, §IV, §VI-A. [42] Y. Li, W. Chen, B. Fu, and et. al. (2025) A hierarchical decoupling task planning method for multi-uav collaborative multi-region coverage with task priority awareness. Drones 9 (8), p. 575. External Links: Document Cited by: §I-A, §I-B, TABLE I, TABLE I, TABLE I, TABLE I, §VI-B, TABLE V. [43] Y. Xiong, Y. Zhou, J. She, and et. al. (2025) Collaborative coverage path planning for uav swarm for multi-region post-disaster assessment. Vehicular Communications 53, p. 100915. External Links: Document Cited by: §I-B, TABLE I, TABLE I, §VI-B, §VI-B, TABLE V, TABLE V. [44] Q. Luan, H. Cui, L. Zhang, and et. al. (2022) A hierarchical hybrid subtask scheduling algorithm in uav-assisted mec emergency network. IEEE Internet Things J. 9 (16), p. 14586–14599. External Links: Document Cited by: §I-A, TABLE I, TABLE I, TABLE I. [45] Y. Cheng, Y. Ding, and C. Lu (2025) Mission planning for uav swarm with aircraft carrier delivery: a decoupled framework. Aerospace 12 (8), p. 691. External Links: Document Cited by: §I-A, TABLE I, TABLE I, TABLE I, TABLE I. [46] A. Guo, Z. Zhang, A. Wu, and et. al. (2025) A hierarchical framework and marginal return optimization for dynamic task allocation in heterogeneous uav networks. Sensors 25 (21), p. 6676. External Links: Document Cited by: §I-A, §I-A, TABLE I, TABLE I, TABLE I, TABLE I. [47] F. Jünger, S. Schopferer, S. Benders, and et. al. (2021) Talking to autonomous drones: command and control based on hierarchical task decomposition. In 2021 International Conference on Unmanned Aircraft Systems (ICUAS), p. 968–977. External Links: Document Cited by: §I-A, TABLE I. [48] E. Yanmaz, H. Balanji, and İ. Güven (2024) Dynamic multi-uav path planning for multi-target search and connectivity. IEEE Trans. Veh. Technol. 73 (7), p. 10516–10528. External Links: Document Cited by: §I-B, TABLE I, TABLE I, TABLE I, TABLE I, §VI-D, TABLE V. [49] X. Zhao, Y. Liu, S. Zhu, and et. al. (2026) Multi-stage task allocation strategy for uav cluster in multi-object tracking. J. Cloud Comput.-Adv. Syst. Appl. 15 (1), p. 21. External Links: Document Cited by: §I-A, TABLE I. [50] R. Alyassi, M. Khonji, A. Karapetyan, and et. al. (2023) Autonomous recharging and flight mission planning for battery-operated autonomous drones. IEEE Trans. Autom. Sci. Eng. 20 (2), p. 1034–1046. External Links: Document Cited by: §I-C, TABLE I, TABLE I, §VI-C, TABLE V. [51] Y. Wang, M. Chen, C. Pan, and et. al. (2022) Joint optimization of uav trajectory and sensor uploading powers for uav-assisted data collection in wireless sensor networks. IEEE Internet Things 9 (13), p. 11214–11226. External Links: Document Cited by: §I-B, TABLE I, TABLE I. [52] X. Jing, F. Liu, C. Masouros, and et. al. (2024) ISAC from the sky: uav trajectory design for joint communication and target localization. IEEE Trans. Wireless Commun. 23 (10), p. 12857–12872. External Links: Document Cited by: §I-B, TABLE I, TABLE I, §IV-A, TABLE I, TABLE I, §VI-A, §VI-A, TABLE V, TABLE V. [53] R. Zhang, Y. Zhang, R. Tang, and et. al. (2024) A joint uav trajectory, user association, and beamforming design strategy for multi-uav-assisted isac systems. IEEE Internet Things J. 11 (18), p. 29360–29374. External Links: Document Cited by: §I-B, TABLE I, TABLE I, §IV-A, §IV-A, TABLE I, TABLE I, §VI-A, TABLE V. [54] Y. Zhu, M. Chen, S. Wang, and et. al. (2024) Collaborative reinforcement learning based unmanned aerial vehicle (uav) trajectory design for 3d uav tracking. IEEE Trans. Mob. Comput. 23 (12), p. 10787–10802. External Links: Document Cited by: §I-B, TABLE I, TABLE I. [55] S. Liu, H. Zhang, Y. Qi, and et. al. (2023) AerialVLN: vision-and-language navigation for uavs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 15384–15394. Cited by: §I-A, §I-B, TABLE I, §V-A, §VI-B. [56] B. Lin, Y. Nie, Z. Wei, and et. al. (2025) NavCoT: boosting LLM-based vision-and-language navigation via learning disentangled reasoning. IEEE TPAMI 47 (7), p. 5945–5957. External Links: Document Cited by: §I-A, TABLE I, §V-A, §V, §VI-B, TABLE V. [57] R. Wu, Y. Zhang, J. Chen, and et. al. (2025) AeroDuo: aerial duo for uav-based vision and language navigation. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 2576–2585. Cited by: §I-B, TABLE I, §V-A, §VI-B. [58] Z. Zhou, T. Long, J. Sun, and et. al. (2025) Hierarchical cooperative path planning method using three-dimensional velocity-obstacle strategy for multiple fixed-wing uavs. J. Syst. Eng. Electron. 36 (5), p. 1342–1352. External Links: Document Cited by: §I-A, TABLE I, TABLE I. [59] B. Zhou, Y. Zhang, X. Chen, and et. al. (2021) Fuel: fast uav exploration using incremental frontier structure and hierarchical planning. IEEE Rob. Autom. Lett. 6 (2), p. 779–786. Cited by: §I-A. [60] C. Zhu and L. Chen (2024) A survey on open-vocabulary detection and segmentation: past, present, and future. IEEE TPAMI 46 (12), p. 8954–8975. External Links: Document Cited by: §I-B, §V-A, TABLE IV, §VI-B, TABLE V. [61] L. Zhu, T. Chen, D. Ji, and et. al. (2025) LLaFS++: few-shot image segmentation with large language models. IEEE TPAMI 47 (9), p. 7715–7732. External Links: Document Cited by: §I-B, §V-A, TABLE IV, §VI-B. [62] J. Alayrac, J. Donahue, P. Luc, and et. al. (2022) Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, Vol. 35, p. 23716–23736. Cited by: §I-B. [63] A. Kirillov, E. Mintun, N. Ravi, and et. al. (2023) Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 4015–4026. Cited by: §I-B, §V-A, TABLE IV, TABLE V. [64] J. Xing, G. Cioffi, J. Hidalgo-Carrió, and et. al. (2023) Autonomous power line inspection with drones via perception-aware mpc. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 1086–1093. Cited by: §I-C. [65] N. Shinn, F. Cassano, A. Gopinath, and et. al. (2023) Reflexion: language agents with verbal reinforcement learning. NeurIPS 36, p. 8634–8652. Cited by: §I-C, §V-C, §V. [66] K. Zhang, P. Chermprayong, F. Xiao, and et. al. (2022) Aerial additive manufacturing with multiple autonomous robots. Nature 609 (7928), p. 709–717. Cited by: §I-C. [67] M. Alighanbari, Y. Kuwata, and J.P. How (2003) Coordination and control of multiple uavs with timing constraints and loitering. In Proceedings of the 2003 American Control Conference, 2003., Vol. 6, p. 5311–5316 vol.6. External Links: Document Cited by: §IV. [68] K. Liu and J. Zheng (2022) UAV trajectory optimization for time-constrained data collection in uav-enabled environmental monitoring systems. IEEE Internet Things J. 9 (23), p. 24300–24314. External Links: Document Cited by: §IV-A, TABLE I, TABLE I. [69] Y. He, Y. Gan, H. Cui, and et. al. (2023) Fairness-based 3-d multi-uav trajectory optimization in multi-uav-assisted mec system. IEEE Internet Things J. 10 (13), p. 11383–11395. External Links: Document Cited by: §IV-A, TABLE I, TABLE I. [70] H. Hu, Z. Chen, F. Zhou, and et. al. (2024) Computation-efficient grouping, trajectory, and resource allocation for uav swarm-assisted aerial–ground collaborative computing networks. IEEE Internet Things J. 11 (7), p. 12510–12525. External Links: Document Cited by: §IV-A, TABLE I, TABLE I, §VI-C, TABLE V. [71] J. Wu, W. Yuan, and L. Bai (2023) On the interplay between sensing and communications for uav trajectory design. IEEE Internet Things J. 10 (23), p. 20383–20395. External Links: Document Cited by: §IV-A, §IV-A, TABLE I, TABLE I. [72] K. Nguyen, T. Duong, T. Do-Duy, and et. al. (2022) 3D uav trajectory and data collection optimisation via deep reinforcement learning. IEEE Trans. Wireless Commun. 70 (4), p. 2358–2371. External Links: Document Cited by: §IV-A, TABLE I, §V-A, TABLE IV. [73] B. Zhu, E. Bedeer, H. Nguyen, and et. al. (2023) UAV trajectory planning for aoi-minimal data collection in uav-aided iot networks by transformer. IEEE Trans. Wireless Commun. 22 (2), p. 1343–1358. External Links: Document Cited by: §IV-A, TABLE I, §V-A, TABLE IV. [74] W. Shi, J. Li, H. Wu, and et. al. (2021) Drone-cell trajectory planning and resource allocation for highly mobile networks: a hierarchical drl approach. IEEE Internet Things J. 8 (12), p. 9800–9813. External Links: Document Cited by: §IV-A, TABLE I. [75] R. Montemanni and M. Dell’Amico (2023) Solving the parallel drone scheduling traveling salesman problem via constraint programming. Algorithms 16 (1), p. 40. External Links: Document Cited by: §IV-B, TABLE I. [76] J. Joo and C. Lee (2026) A branch-and-price algorithm for robust drone-vehicle routing problem with time windows. INFORMS J Comput 38 (1), p. 102–125. External Links: Document Cited by: §IV-B, §IV-B, TABLE I, TABLE I. [77] Y. Yang, J. Liu, and S. Wang (2025) A bi-objective optimisation model for the drone scheduling problem in island delivery. Int. J. Prod. Res. 63 (19), p. 7174–7195. External Links: Document Cited by: §IV-B, TABLE I. [78] E. Rigas, P. Kolios, M. Mavrovouniotis, and et. al. (2022) Scheduling a fleet of drones for monitoring missions with spatial, temporal, and energy constraints. IEEE Trans. Intell. Transp. Syst. 23 (9), p. 15133–15145. External Links: Document Cited by: §IV-B, TABLE I. [79] C. Zhao, Y. Wang, X. Zhang, and et. al. (2023) UAV dispatch planning for a wireless rechargeable sensor network for bridge monitoring. IEEE Trans. Sustainable Comput. 8 (2), p. 293–309. External Links: Document Cited by: §IV-B, TABLE I. [80] X. Mao, G. Wu, M. Fan, and et. al. (2025) DL-drl: a double-level deep reinforcement learning approach for large-scale task scheduling of multi-uav. IEEE Trans. Autom. Sci. Eng. 22, p. 1028–1044. External Links: Document Cited by: §IV-C, §IV-C, TABLE I, TABLE I, TABLE I, §V-C, TABLE IV. [81] Z. Pei, T. Fang, K. Weng, and W. Yi (2023) Urban on-demand delivery via autonomous aerial mobility: formulation and exact algorithm. IEEE TASE 20 (3), p. 1675–1689. External Links: Document Cited by: §IV-B, TABLE I. [82] F. Betti Sorbelli, F. Corò, S. Das, and et. al. (2025) Dispatching point selection for a drone-based delivery system operating in a mixed euclidean–manhattan grid. Annals of Operations Research 351 (1), p. 203–222. External Links: Document Cited by: §IV-B, TABLE I. [83] O. Oubbati, M. Atiquzzaman, A. Baz, and et. al. (2021) Dispatch of uavs for urban vehicular networks: a deep reinforcement learning approach. IEEE Trans. Veh. Technol. 70 (12), p. 13174–13189. External Links: Document Cited by: §IV-C, TABLE I. [84] J. Moon, S. Papaioannou, C. Laoudias, and et. al. (2021) Deep reinforcement learning multi-uav trajectory control for target tracking. IEEE Internet Things J. 8 (20), p. 15441–15455. External Links: Document Cited by: §IV-C, TABLE I. [85] B. Zhu, E. Bedeer, H. Nguyen, and et. al. (2021) UAV trajectory planning in wireless sensor networks for energy consumption minimization by deep reinforcement learning. IEEE Trans. Veh. Technol. 70 (9), p. 9540–9554. External Links: Document Cited by: §IV-C, TABLE I. [86] Z. Zhu, K. Lin, A. K. Jain, and et. al. (2023) Transfer learning in deep reinforcement learning: a survey. IEEE TPAMI 45 (11), p. 13344–13362. External Links: Document Cited by: §IV-A, TABLE I, §V-A, TABLE IV. [87] T. Hu, B. Luo, C. Yang, and et. al. (2023) MO-MIX: multi-objective multi-agent cooperative decision-making with deep reinforcement learning. IEEE TPAMI 45 (10), p. 12098–12112. External Links: Document Cited by: §IV-A, §IV-A, TABLE I, §IV, §V-B, TABLE IV. [88] M. Rinaldi, S. Primatesta, M. Bugaj, and et. al. (2024) Urban air logistics with unmanned aerial vehicles (uavs): double-chromosome genetic task scheduling with safe route planning. Smart Cities 7 (5), p. 2842–2860. External Links: Document Cited by: §IV-C, TABLE I. [89] W. Zai, J. Wang, and G. Li (2023) A drone scheduling method for emergency power material transportation based on deep reinforcement learning optimized pso algorithm. Sustainability 15 (17), p. 13127. External Links: Document Cited by: §IV-C, TABLE I. [90] C. Qin and E. Pournaras (2026) Strategic coordination of drones via short-term distributed optimization and long-term reinforcement learning. Appl. Soft Comput. 186, p. 114270. External Links: Document Cited by: TABLE I, §V-B, TABLE IV, TABLE IV. [91] H. Kang, X. Chang, J. Misic, and et. al. (2022) Improving dual-uav aided ground-uav bi-directional communication security: joint uav trajectory and transmit power optimization. IEEE Trans. Veh. Technol. 71 (10), p. 10570–10583. External Links: Document Cited by: §IV-D, TABLE I, §VI-A, §VI-D, TABLE V, TABLE V. [92] C. Zhang, G. Sun, Q. Wu, and et. al. (2024) UAV swarm-enabled collaborative secure relay communications with time-domain colluding eavesdropper. IEEE Trans. Mob. Comput. 23 (9), p. 8601–8619. External Links: Document Cited by: §IV-D, TABLE I, §VI-A, §VI-D, §VI-D, TABLE V, TABLE V, TABLE V. [93] H. Dang-Ngoc, D. Nguyen, K. Ho-Van, and et. al. (2022) Secure swarm uav-assisted communications with cooperative friendly jamming. IEEE Internet Things J. 9 (24), p. 25596–25611. External Links: Document Cited by: §IV-D, TABLE I, §VI-A, §VI-D, TABLE V, TABLE V. [94] P. Tedeschi, S. Sciancalepore, and R. Di Pietro (2023) PPCA - privacy-preserving collision avoidance for autonomous unmanned aerial vehicles. IEEE T Depend Secure 20 (2), p. 1541–1558. External Links: Document Cited by: §IV-D, TABLE I, TABLE I, §V-D, TABLE IV. [95] D. George, S. Sciancalepore, and N. Zannone (2023) Privacy-preserving multi-party access control for third-party uav services. SACMAT 2023, p. 19–30. External Links: Document Cited by: §IV-D, TABLE I. [96] H. Huang, Y. Fang, B. Mazotti, and et. al. (2024) Privacy-aware design and analysis of drone remote identification systems. IEEE Trans. Intell. Transp. Syst. 25 (12), p. 21455–21468. External Links: Document Cited by: §IV-D, TABLE I, TABLE I. [97] J. Akram, A. Akram, P. Ingle, and et. al. (2025) Privacy-preserving spatial crowdsourcing drone services for postdisaster infrastructure monitoring: a conditional federated learning approach. IEEE J-STARS 18, p. 16272–16291. External Links: Document Cited by: §IV-D, TABLE I, §V-D, TABLE IV. [98] C. Pu (2023) Privacy-preserving and fault-tolerant data aggregation protocol for internet of drones. DSC, p. 1–8. External Links: Document Cited by: §IV-D, TABLE I. [99] H. Jagarlapudi, S. Lim, J. Chae, and et. al. (2022) Drone helps privacy: sky caching assisted k-anonymity in spatial querying. IEEE Systems Journal 16 (4), p. 6360–6370. External Links: Document Cited by: §IV-D, TABLE I. [100] S. Enayati, D. Goeckel, A. Houmansadr, and et. al. (2023) Location privacy protection for uavs in package delivery and iot data collection. IEEE Internet Things J. 10 (23), p. 20586–20601. External Links: Document Cited by: §IV-D, TABLE I. [101] M. Torabbeigi, G. Lim, N. Ahmadian, and et. al. (2021) An optimization approach to minimize the expected loss of demand considering drone failures in drone delivery scheduling. J. Intell. Robot. Syst. 102 (1). External Links: Document Cited by: §IV-B. [102] M. Pachayappan and V. Sudhakar (2021) A solution to drone routing problems using docking stations for pickup and delivery services. Transp. Res. Rec. 2675 (12), p. 1056–1074. External Links: Document Cited by: §IV-B. [103] Y. Liu, C. Chen, Y. Sun, and et. al. (2025) Advancing multi-uav inspection dispatch based on bilevel optimization and ga-nsga-i. Applied Sciences 15 (7), p. 3673. External Links: Document Cited by: §IV-B. [104] A. Khochare, Y. Simmhan, F. Sorbelli, and et. al. (2021) Heuristic algorithms for co-scheduling of edge analytics and routes for uav fleet missions. INFOCOM, p. 1–10. External Links: Document Cited by: §IV-B. [105] A. Khochare, F. Sorbelli, Y. Simmhan, and et. al. (2024) Improved algorithms for co-scheduling of edge analytics and routes for uav fleet missions. IEEE ACM T NETWORK 32 (1), p. 17–33. External Links: Document Cited by: §IV-B. [106] G. Campuzano, E. Lalla-Ruiz, and M. Mes (2022) The dynamic drone scheduling delivery problem. LNCS, p. 260–274. External Links: Document Cited by: §IV-B. [107] S. Gong, G. Chen, and Z. Yang (2025) Research on the design of a priority-based multi-stage emergency material scheduling system for drone coordination. Drones 9 (8), p. 524. External Links: Document Cited by: §IV-B. [108] D. Ebrahimi, S. Sharafeddine, P. Ho, and et. al. (2021) Autonomous uav trajectory for localizing ground objects: a reinforcement learning approach. IEEE Trans. Mob. Comput. 20 (4), p. 1312–1324. External Links: Document Cited by: §V-A, TABLE IV. [109] G. Xu, W. Jiang, Z. Wang, and et. al. (2022) Autonomous obstacle avoidance and target tracking of uav based on deep reinforcement learning. J. Intell. Robot. Syst. 104 (4). External Links: Document Cited by: §V-A, TABLE IV. [110] R. Zhang, Q. Zong, X. Zhang, and et. al. (2023) Game of drones: multi-uav pursuit-evasion game with online motion planning by deep reinforcement learning. IEEE Trans. Neural Netw. Learn. Syst. 34 (10), p. 7900–7909. External Links: Document Cited by: §V-A, TABLE IV. [111] H. Huang, A. Savkin, and W. Ni (2022) Online uav trajectory planning for covert video surveillance of mobile targets. IEEE Trans. Autom. Sci. Eng. 19 (2), p. 735–746. External Links: Document Cited by: §V-A, §V-D, TABLE IV, TABLE IV. [112] C. Zhao, J. Liu, M. Sheng, and et. al. (2021) Multi-uav trajectory planning for energy-efficient content coverage: a decentralized learning-based approach. IEEE J Sel Area Comm 39 (10), p. 3193–3207. External Links: Document Cited by: §V-B, TABLE IV. [113] Y. Huang, H. Wang, X. Bai, and et. al. (2025) Biomimetic multi-uav swarm exploration with u2u communications under resource constraints. IEEE Trans. Veh. Technol. 74 (6), p. 9750–9766. External Links: Document Cited by: §V-B, TABLE IV. [114] M. Hasan, S. Rayhan Kabir, S. Abdullah, and et. al. (2023) 3D relative directions based evolutionary computation for uav-to-uav interaction in swarm intelligence enabled decentralized networks. Alex. Eng. J. 85, p. 104–113. External Links: Document Cited by: §V-B, TABLE IV. [115] F. Venturini, F. Mason, F. Pase, and et. al. (2021) Distributed reinforcement learning for flexible and efficient uav swarm control. IEEE Trans. Cognit. Commun. Networking 7 (3), p. 955–969. External Links: Document Cited by: §V-B, TABLE IV. [116] H. K. Alkahtani, Y. Galiya, B. Akbayan, and et. al. (2026) Explainable multi agent reinforcement learning framework for secure and adaptive communication in uav swarm based fanets. Scientific Reports. External Links: Document Cited by: §V-B, TABLE IV, TABLE V. [117] S. Chinthi-Reddy, S. Lim, G. Choi, and et. al. (2022) DarkSky: privacy-preserving target tracking strategies using a flying drone. Vehicular Communications 35, p. 100459. External Links: Document Cited by: §V-D, TABLE IV. [118] S. Sciancalepore and D. George (2022) Privacy-preserving trajectory matching on autonomous unmanned aerial vehicles. ACSAC ’22, p. 1–12. External Links: Document Cited by: §V-D, TABLE IV. [119] C. Gao, D. Wei, K. Li, and et. al. (2025) UAV-centric privacy-preserving computation offloading in multi-uav mobile edge computing. Drones 9 (10), p. 701. External Links: Document Cited by: §V-D, TABLE IV. [120] S. Kim, O. Alama, D. Kurdydyk, and et. al. (2025) RAVEN: resilient aerial navigation via open-set semantic memory and behavior adaptation. arXiv preprint arXiv:2509.23563. Cited by: §V-A. [121] X. Zhou, J. Zhu, H. Zhou, and et. al. (2021) EGO-swarm: a fully autonomous and decentralized quadrotor swarm system in cluttered environments. In 2021 IEEE International Conference on Robotics and Automation (ICRA), Vol. , p. 4101–4107. External Links: Document Cited by: §V-B. [122] X. Li, K. Xu, F. Liu, and et. al. (2025) Airswarm: enabling cost-effective multi-uav research with cots drones. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 21487–21494. Cited by: §V-B. [123] K. N. McGuire, C. De Wagter, K. Tuyls, and et. al. (2019) Minimal navigation solution for a swarm of tiny flying robots to explore an unknown environment. Science Robotics 4 (35), p. eaaw9710. Cited by: §V-B. [124] D. Morgan, S. Chung, and F. Y. Hadaegh Swarm assignment and trajectory optimization using variable-swarm, distributed auction assignment and model predictive control. In AIAA Guidance, Navigation, and Control Conference, p. . External Links: Document, Link, https://arc.aiaa.org/doi/pdf/10.2514/6.2015-0599 Cited by: §V-C. [125] S. Javed, A. Hassan, R. Ahmad, and et. al. (2023) UAV trajectory planning for disaster scenarios. Vehicular Communications 39, p. 100568. External Links: Document Cited by: §VI-A, TABLE V. [126] K. Meng, Q. Wu, S. Ma, and et. al. (2022) UAV trajectory and beamforming optimization for integrated periodic sensing and communication. IEEE Wireless Commun. Lett. 11 (6), p. 1211–1215. External Links: Document Cited by: §VI-A, §VI-A, TABLE V, TABLE V. [127] X. Liu, W. Yi, P. Chen, and et. al. (2025) Flight path planning of uav-driven refinement inspection for construction sites based on 3d reconstruction. Autom. Constr. 177, p. 106360. External Links: Document Cited by: §VI-B, TABLE V, TABLE V. [128] Y. Wu, K. Low, B. Pang, and et. al. (2021) Swarm-based 4d path planning for drone operations in urban environments. IEEE Trans. Veh. Technol. 70 (8), p. 7464–7479. External Links: Document Cited by: §VI-C, TABLE V. [129] R. Ming, R. Jiang, H. Luo, and et. al. (2023) Comparative analysis of different uav swarm control methods on unmanned farms. Agronomy 13 (10), p. 2499. External Links: Document Cited by: §VI-C, TABLE V. [130] C. Feng, H. Li, M. Zhang, and et. al. (2024) Fc-planner: a skeleton-guided planning framework for fast aerial coverage of complex 3d scenes. In 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 8686–8692. Cited by: §VI-B. [131] P. Petráček, V. Krátkỳ, M. Petrlík, and et. al. (2021) Large-scale exploration of cave environments by unmanned aerial vehicles. IEEE Rob. Autom. Lett. 6 (4), p. 7596–7603. Cited by: §VI-B. [132] D. Ru, L. Qiu, X. Hu, and et. al. (2024) RAGChecker: a fine-grained framework for diagnosing retrieval-augmented generation. NeurIPS 37, p. 21999–22027. Cited by: §VII-B. [133] M. Wu, J. Li, J. Ji, and et. al. (2026) Evaluating and mitigating relationship hallucinations in large vision-language models. IEEE TPAMI 48 (6), p. 6332–6346. External Links: Document Cited by: §VII-B.

Instant research discovery