PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.09466v1cs.AIcs.SYeess.SY

TL;DR

UAV dispatch is moving beyond isolated path planning and single-agent control toward system-level coordination. This survey defines a four-paradigm taxonomy and examines LLM roles across dispatch; its supported conclusion is that LLM reasoning should be integrated with specialized modules, not replace them.

  • Problem

    Large-scale UAV operations cannot be supported by isolated path planning or single-agent control alone, motivating system-level dispatch coordination.

    Source: §I · p. 1

  • Method

    The survey defines UAV dispatch as a semantic–symbolic–physical process and classifies dispatch logic into four paradigms, analyzing potential LLM roles across them.

    Source: §I · pp. 2–3

  • Results

    The survey organizes existing studies into pipeline, global assignment, decentralized agentic, and divide-and-conquer dispatch paradigms.

    Source: §I · p. 2, §VIII · p. 16

  • Takeaways & Limitations

    LLMs can support mission interpretation, solver orchestration, coordination, and explanation, while specialized modules must guarantee physical feasibility.

    Source: §I · p. 3, §VIII · p. 16

    LLM reasoning does not itself guarantee physical feasibility, which must be ensured by specialized modules.

    Source: §I · p. 3

Abstract

from arXiv · show

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.

I. INTRODUCTION

UAV dispatch must coordinate coupled fleet decisions under physical and operational constraints, while handling increasingly open-ended mission requests. This survey frames LLMs as a semantic layer and organizes dispatch logic into four paradigms, identifying capabilities and open challenges for their integration.

I. INTRODUCTION

  • UAV dispatch schedules, routes, allocates resources, associates, and coordinates fleets under physical, communication, safety, privacy, and regulatory constraints.The survey frames the LLM-era process as mapping human intent and environmental context to assignment, trajectory, schedule, resource, and cooperation decisions.Source: §I · pp. 1–2
  • Growing fleets make collision-free trajectories and communication resources harder to allocate jointly, while operators increasingly issue open-ended mission requests.Dispatch decisions are coupled across assignment, trajectory, communication, energy, timing, and privacy.Source: §I · p. 1
  • Prior LLM-assisted UAV studies remain fragmented, and existing surveys lack a dispatch-centered taxonomy spanning intent through execution monitoring.Related surveys chiefly focus on trajectory prediction and path planning, swarm networking and control, security and privacy, or broader LLM topics.Source: §I · pp. 1–2
  • The proposed taxonomy classifies dispatch as pipeline, global assignment, decentralized agentic, or divide-and-conquer, based on transforming mission descriptions into executable decisions.Secondary axes include covered decisions, solver family, and application context.Source: §I · p. 2
  • LLMs can parse intent, retrieve knowledge, orchestrate solvers, coordinate agents, assess safety, and explain decisions, while specialized modules guarantee physical feasibility.The survey presents LLMs as a cross-layer interface rather than a replacement for verified control mechanisms.Source: §I · pp. 1–3
  • The survey identifies grounding, hallucination control, edge deployment, multi-agent trust, privacy, accountability, benchmarking, and human governance as open challenges.Source: §I · p. 3

II. PRELIMINARIES

The preliminaries extend traditional dispatch from matching agents to tasks into autonomous, adaptive coordination, then specialize it for UAVs operating under airborne constraints. They frame an LLM as a semantic interface that structures mission inputs while optimization handles dispatch decisions.

A. Traditional Dispatch Framework

  • Traditional dispatch jointly decides task assignment, routing, scheduling, resource allocation, and cooperation while respecting task coverage, capacity, endurance, service-time, and route feasibility.The framework uses U for agents, T for tasks, and Ω for the operating environment.Source: A. Traditional Dispatch Framework · p. 3

B. Unmanned Dispatch System

  • Unmanned dispatch replaces passive executors with autonomous agents that perceive, communicate, update local state, and adapt during execution.Partial or delayed information, coupled task execution and motion feasibility, and online changes distinguish it from traditional dispatch.Source: B. Unmanned Dispatch System · pp. 3–4

C. UAV Dispatch Methods

  • UAV dispatch coordinates airborne agents across missions while jointly accounting for three-dimensional mobility, flight dynamics, safety, endurance, communication, weather, regulation, and privacy.A UAV's position is represented as q_i(t) = [x_i(t), y_i(t), z_i(t)]; feasible trajectories also face mobility and collision constraints.Source: C. UAV Dispatch Methods · p. 4
  • An LLM-enabled interface can turn human instructions and context into structured tasks, constraints, and preferences, while the dispatch layer optimizes the resulting multi-objective plan.Retrieval can ground inputs in maps, regulations, weather, logs, and mission documents, and external evaluators can check physical, regulatory, and privacy constraints.Source: C. UAV Dispatch Methods · p. 4

III. PIPELINE DISPATCH

Pipeline dispatch organizes UAV missions as ordered stages rather than one monolithic optimization. In this workflow, an LLM can interpret mission intent and orchestrate tools and validation while leaving the stages auditable.

III. PIPELINE DISPATCH

  • Pipeline dispatch decomposes a mission into successive stages including semantic parsing, task decomposition, assignment, trajectory generation, scheduling, and execution monitoring.The workflow is abstracted from mission description M through stage functions to execution or monitoring E.Source: §III · p. 4
  • An LLM or multimodal LLM can parse the mission, retrieve knowledge, structure tasks, select downstream solvers, and ask external validators to check plans and execution feedback.The cited examples include checking assignments, trajectories, schedules, and execution feedback.Source: §III · p. 4
  • The LLM layer makes the front end more adaptive to natural-language intent, heterogeneous context, and changing low-altitude constraints without replacing the pipeline.Keeping each stage auditable is part of the described design.Source: §III · p. 4

A. Hierarchical Task Decomposition and Scheduling

Pipeline methods use hierarchy to organize complex missions and use trajectory-centric stages when mobility optimization is the main bottleneck. These designs span task abstraction, allocation refinement, grounded trajectory tools, and trajectory feasibility limits.

A. Hierarchical Task Decomposition and Scheduling

  • Hierarchical dispatch separates mission understanding from execution ordering, assigning decomposition, allocation, mobility support, or subtask scheduling to different layers.Examples include upper-layer task decomposition followed by marginal-return allocation, and coarse initialization followed by parallel auction refinement.Source: A. Hierarchical Task Decomposition and… · p. 5
  • Hierarchical pipelines can connect human commands to UAV subtasks and organize exploration, perception, task construction, and collision-aware trajectory refinement.The upper layer can act as an LLM semantic layer that retrieves rules, builds a task graph, selects tools, and records why subtasks are created.Source: A. Hierarchical Task Decomposition and… · p. 5

B. Trajectory-Centric Pipeline Dispatch

  • Trajectory-centric pipeline dispatch treats trajectory generation as a downstream stage after task allocation or service ordering, with trajectory optimization as its primary technical bottleneck.These studies center on f3: X → P, sometimes also involving service ordering, connectivity, or collision avoidance.Source: B. Trajectory-Centric Pipeline Dispatch · p. 6
  • An LLM can translate instructions into semantic waypoints, forbidden regions, and priorities, but geometry, control, and safety modules must produce or verify the final path.The described division places compliance-aware constraint interpretation with LLM reasoning and motion feasibility with a low-level policy.Source: B. Trajectory-Centric Pipeline Dispatch · p. 6
  • Trajectory-centric methods are most effective when task structure is clear and mobility is the bottleneck, and may be less effective when decomposition, resources, and trajectory planning are strongly coupled.The approach treats task decomposition as lightweight preprocessing and places the main optimization burden on path generation.Source: B. Trajectory-Centric Pipeline Dispatch · p. 6

C. Resource-Aware and Application-Oriented Pipelines

Resource-aware pipelines promote a mission's dominant operational bottleneck to a dispatch stage and adapt their structure to the scenario. Global assignment instead uses a shared global view to jointly determine dispatch decisions, with an LLM potentially supporting intent interpretation and solver choice.

C. Resource-Aware and Application-Oriented Pipelines

  • Inspection pipelines can refresh task clustering and scheduling as task sets change during execution, while delivery and long-duration inspection may center on assigning energy state at the right time.The inspection examples also couple tracking, trajectory generation, and control feasibility during execution.Source: C. Resource-Aware and… · p. 6
  • LLM-informed application pipelines can retrieve domain evidence, extract constraints, call a solver, verify execution, and update later dispatch prompts from monitored outcomes.The cited examples include converting heterogeneous inspection inputs into candidate defects, subtasks, and agent responsibilities.Source: C. Resource-Aware and… · pp. 6–7
  • Application-oriented pipeline design isolates a critical resource or operational factor as a stage, with inspection, delivery, and emergency sensing shaped by rolling updates, energy logistics, and range-aware partitioning.The pipeline structure is described as adaptable to mission objectives, resource limitations, and operating environment.Source: C. Resource-Aware and… · pp. 6–7

IV. GLOBAL ASSIGNMENT DISPATCH

  • Global assignment uses a central dispatcher or equivalent global view to jointly determine task matching, routes, schedules, resources, and associations rather than passing artifacts through fixed stages.A semantic meta-dispatcher can translate intent into constraints, retrieve domain knowledge, choose solvers, and explain trade-offs before optimization commits a plan.Source: §IV · p. 7
  • The optimization core enforces feasibility, while the surrounding LLM layer helps handle open-ended requests, incomplete operator input, and cross-layer trade-offs.Source: §IV · p. 7

A. Mobility–Communication Assignment

Global assignment dispatch jointly coordinates coupled mobility, communication, sensing, and resource decisions, while routing and scheduling systems encode service order and operational constraints. LLMs can translate natural-language priorities into constraints and guide solver selection without replacing feasibility-checking optimization or routing tools.

A. Mobility–Communication Assignment

  • Trajectory, access, sensing, and resource decisions are coupled in communication-heavy missions, making the central dispatcher an integration point for mobility–resource control.Changing a UAV trajectory can alter channel quality and thereby affect feasible schedules, power allocation, and user association.Source: A. Mobility–Communication Assignment · p. 7
  • An LLM can map natural-language priorities to association, power, sensing, and trajectory constraints, then guide solver selection while external solvers retain feasibility and Pareto-quality responsibility.Li et al. use LLM reasoning within a decomposition-based multi-objective evolutionary algorithm for UAV-ISAC optimization.Source: A. Mobility–Communication Assignment · p. 8

B. Routing, Scheduling, and Vehicle-Assisted Operations

  • Routing and scheduling couple task assignment, routes, and service order with time, capacity, energy, vehicle coordination, and application-specific constraints.Vehicle-assisted systems must synchronize ground carriers with aerial task executors; monitoring and emergency missions add domain constraints such as windows, urgency, charging, or airspace conflicts.Source: B. Routing, Scheduling, and… · p. 8
  • LLM agents can translate natural-language orders into solver inputs and expose constraints for inspection, but exact or heuristic routing algorithms remain necessary before routes are committed.The agent can plan tool calls, execute the solver sequence, and summarize results for the operator,.Source: B. Routing, Scheduling, and… · p. 8

C. Scalable Solvers and Learning-Enhanced Assignment

Scalable solver design uses learning-based and hybrid methods to address the cost of centralized assignment as fleet and task sizes grow. These approaches trade speed and adaptability against weaker guarantees, while centralized dispatch also raises privacy and information-exposure concerns.

C. Scalable Solvers and Learning-Enhanced Assignment

  • As fleets and task sets grow, learning-based and hybrid solvers complement or approximate exact optimization for assignment models with integer, trajectory, communication, and time-window constraints.Scalable solver design is an implementation dimension of global assignment, affecting whether a centralized model can be solved in real time.Source: C. Scalable Solvers and… · p. 9
  • LLMs can propose solver decompositions, priority functions, neighborhood moves, or repair rules for external evaluation rather than outputting assignments directly.Classical solvers can produce or verify feasible plans, while LLMs search configurations and explain candidate failures.Source: C. Scalable Solvers and… · p. 9
  • Speed and adaptability come with weaker guarantees, and reward design, training distribution, heuristic parameters, and scenario assumptions can affect assignment quality and safety.These sensitivities qualify the use of learning-enhanced and hybrid solver methods.Source: C. Scalable Solvers and… · p. 9

D. Security, Privacy, and Trust-Aware Assignment

  • Centralized dispatch concentrates sensitive UAV states, user locations, task requests, video, trajectories, and communication metadata, while LLM prompts and retrieved contexts can expose similar information.The proposed safeguards include prompt filters, access control for retrieved documents, and tool-call audit logs; retrieval also makes provenance and document freshness part of the security boundary.Source: D. Security, Privacy, and Trust-Aware… · p. 9

V. DECENTRALIZED AGENTIC DISPATCH

Decentralized agentic dispatch lets UAVs make local task and motion decisions from observations, utility, memory, and limited peer interaction rather than a complete centralized assignment. Its local reasoning can be extended into swarm coordination through distributed learning and compact semantic messages.

V. DECENTRALIZED AGENTIC DISPATCH

  • Local UAV agents select tasks, adjust trajectories, and coordinate through local observations, utility, semantic messages, and limited negotiation rather than a complete centralized assignment.LLM-enabled agents can summarize events, maintain short-term memory, negotiate, reflect on failed actions, and pass verified high-level commands to conventional planners or controllers,,,.Source: §V · pp. 9–10

A. Local Perception and Motion Decisions

  • Local motion policies repeatedly map current observations to executable decisions, while semantic mediation can connect aerial perception and mission intent to structured goals for verified planners.Studies use reinforcement learning for trajectory and data-collection decisions, and transformer-based planning for age-of-information-aware collection,.Source: A. Local Perception and Motion Decisions · p. 10

B. Swarm Interaction and Distributed Learning

  • Swarm interaction addresses duplicated coverage and collisions through limited neighbor information, distributed coordination, and semantic messages that must remain bounded and verifiable.Multi-agent reinforcement learning extends local actions toward team objectives, while edge orchestration must respect latency and privacy budgets.Source: B. Swarm Interaction and Distributed… · pp. 10–11

C. Bidding, Consensus, and Partial Replanning

Bidding and consensus make task ownership explicit by connecting local utility estimates to swarm-level allocation, while LLMs can support negotiation without replacing distributed agreement rules. Decentralized autonomy shifts privacy and trust concerns to peer messages, trajectory sharing, sensing, and local memory.

C. Bidding, Consensus, and Partial Replanning

  • Auction-based allocation combined with trajectory optimization assigns task ownership through competitive local utility while jointly handling trajectory feasibility.Local reassignment can avoid rebuilding the entire allocation, and cost estimates such as distance, energy, deadline, risk, and reward connect individual decisions to swarm ownership.Source: C. Bidding, Consensus, and Partial… · p. 11
  • LLM agents can explain bids and support replanning, but auction or consensus protocols must enforce ownership because agreement and deadline feasibility require explicit distributed rules.Agents may translate energy, distance, risk, and sensor capability into bid rationales or request missing context,.Source: C. Bidding, Consensus, and Partial… · p. 11

D. Privacy-Preserving and Trustworthy Autonomy

  • Decentralization shifts privacy and security risks to peer messages, trajectory sharing, collision-avoidance protocols, cooperative learning, and local sensing.LLM-enabled agents also need controls over private sensor memory, shareable semantic memory, message filtering, and safety or privacy checks on proposed actions.Source: D. Privacy-Preserving and Trustworthy… · pp. 11–12
  • Coordination can require sharing future motion, yet that motion may reveal sensitive locations or mission intent, so decentralized dispatch must limit what information peers exchange.Privacy-aware coordination must distinguish necessary information from information that should remain hidden.Source: D. Privacy-Preserving and Trustworthy… · p. 12

VI. DIVIDE-AND-CONQUER DISPATCH

Divide-and-conquer dispatch makes large, coupled missions tractable by splitting them into subproblems and coordinating their solutions. The section covers objective and communication splits, spatial partitioning, and the need to reconcile local plans.

VI. DIVIDE-AND-CONQUER DISPATCH

  • Large missions are partitioned into smaller subproblems, solved separately, and recomposed or verified into a coherent dispatch plan.Unlike a fixed pipeline or a single global solver, the approach depends on choosing boundaries that keep local decisions tractable.Source: §VI · p. 12, Figure 5 · p. 12

A. Trajectory–Communication and Objective Decomposition

  • Joint trajectory–communication optimization can be non-convex, so methods separate trajectory, power, beamforming, timing, or objectives and coordinate the subproblems iteratively.The decomposed objectives include outage probability and trajectory length, secure communication, and integrated sensing and communication.Source: A. Trajectory–Communication and… · pp. 12–13
  • LLMs can name mission subobjectives, retrieve constraints, and select solver order, while external scoring or numerical verification checks candidate decompositions.The survey distinguishes this orchestration role from merging objectives into an opaque prompt.Source: A. Trajectory–Communication and… · p. 13

B. Spatial and Regional Decomposition

  • Spatial decomposition partitions large operating areas into regions whose local routes must be coordinated across boundaries.The spatial split changes the problem from finding one global route to dividing space and coordinating routes across the resulting boundaries.Source: B. Spatial and Regional Decomposition · p. 13
  • Multimodal evidence can support semantic region construction before local planning, but boundaries can cause duplicated coverage, missed targets, relay gaps, or shared-resource conflicts.The success of spatial decomposition therefore depends on both the partition rule and the coordination rule that reconnects local routes.Source: B. Spatial and Regional Decomposition · pp. 13–14

C. Swarm, Role, and Functional Decomposition

Fleet and functional decomposition assigns UAVs, agents, or subsystems to groups and roles, while security- and communication-oriented decomposition gives specialized functions distinct responsibilities. Those roles still require coordination and trustworthy orchestration.

C. Swarm, Role, and Functional Decomposition

  • Fleet decomposition forms groups, roles, or functional modules, particularly when UAVs differ in payload, communication range, endurance, or mission responsibility.Examples include separating flight execution from recharging and dividing navigation, perception, collaboration, and trajectory planning.Source: C. Swarm, Role, and Functional… · p. 14
  • LLM-agent fleets can assign perception, planning, communication, privacy, and verification to specialized agents, with roles placed across devices, edge servers, and cloud resources,.This organizational split includes both physical UAV roles and computational model roles.Source: C. Swarm, Role, and Functional… · p. 14

D. Security- and Communication-Oriented Decomposition

  • Security- and communication-oriented decomposition assigns UAVs distinct relay, jamming, sensing, or data-collection roles while preserving a shared objective.Examples separate relay communication from friendly jamming, or divide search and relay roles to maintain connectivity.Source: D. Security- and… · p. 14
  • LLM-mediated security coordination can introduce prompt injection, stale evidence, or overconfident recommendations, so retrieved rules, explanations, and role assignments require logging and non-LLM verification.The security boundary includes both radio and trajectory security and the trustworthiness of language-mediated coordination.Source: D. Security- and… · p. 14

VII. SYNTHESIS AND FUTURE DIRECTIONS IN THE LLM ERA

The synthesis treats LLMs as a semantic layer connecting intent, knowledge, tools, and oversight across dispatch paradigms, whose boundaries can shift with mission conditions. It also identifies grounding, physical feasibility, latency, resource, security, and evaluation challenges for deployment.

VII. SYNTHESIS AND FUTURE DIRECTIONS IN THE LLM ERA

  • LLMs connect operator intent, mission knowledge, multimodal perception, tool invocation, solver selection, and oversight, while physical limits remain defined by optimization, control, communication, and safety verification.The LLM layer changes how these components are selected, connected, explained, and adapted during operations.Source: §VII · p. 14

A. Cross-Paradigm Synthesis

  • A dispatch system can shift among pipeline, global-assignment, agentic, and divide-and-conquer modes according to mission scale, connectivity, uncertainty, and safety requirements.The paradigms remain useful abstractions, but their boundaries become less rigid with semantic reasoning.Source: A. Cross-Paradigm Synthesis · p. 15
  • LLMs can expose intermediate pipeline artifacts or help a central dispatcher compare assignments, but explicit optimization and constraint models must produce or verify global feasibility.Linguistic plausibility alone cannot establish that a global assignment is feasible.Source: A. Cross-Paradigm Synthesis · p. 15
  • Decentralized agents can interpret nearby events and support local task decisions, but latency, message size, memory, and prompt-level attack surfaces become dispatch-design concerns.Decomposed local decisions must eventually be recombined into a physically consistent fleet plan.Source: A. Cross-Paradigm Synthesis · p. 15

B. Open Challenges

  • Retrieval grounding does not ensure executable plans: LLM outputs need downstream physical checks, and the semantic-to-physical interface must be fast enough for changing low-altitude conditions.Retrieved context reduces risk only when it is relevant, current, and faithfully used.Source: B. Open Challenges · p. 15
  • Cloud and onboard models trade reasoning strength and availability against latency, energy, and brittleness; fleet-scale deployment also faces dialogue inconsistency, security exposure, and broader evaluation demands.The survey calls for structured messages, role-based communication, consensus rules, local validators, and evaluation of executability, robustness, and human-approval burden.Source: B. Open Challenges · p. 15

C. Future Research Directions

The future directions call for grounded knowledge, multimodal perception, and coordinated local and global reasoning, with human governance built into dispatch. The conclusion favors hybrid semantic, symbolic, and physical architectures rather than replacing specialized optimization and safety mechanisms.

C. Future Research Directions

  • Dispatch knowledge layers should combine maps, rules, weather, maintenance, mission logs, and communication state with traceable sources and constraints that solvers and monitors can check.Retrieval-augmented generation should be studied together with LLM-tool-solver co-design.Source: C. Future Research Directions · pp. 15–16
  • Multimodal dispatch must ground image, video, map, language, and sensor evidence in geometry, time, and mission context, while choosing among onboard, edge, and cloud models.Systems must also recover when a model, link, or sensor stream fails.Source: C. Future Research Directions · p. 16
  • Large-fleet research must study local reaction alongside global recomposition, including compact negotiation, prompt-safe replanning, distributed verification, and reliable decomposition boundaries.Real systems may mix decentralized task claiming with higher-level policies for load balance and mission consistency.Source: C. Future Research Directions · p. 16
  • Dispatch governance should expose uncertainty, cite retrieved evidence, present alternatives, and request human approval when safety, privacy, or regulation is at stake.The proposed hybrid architecture assigns intent interpretation and knowledge organization to the semantic layer, while symbolic and physical layers operate under explicit constraints.Source: C. Future Research Directions · p. 16

VIII. CONCLUSION

  • The survey organizes UAV dispatch into four paradigms and concludes that LLM reasoning should be grounded, structured, and verified rather than replace optimization, control, communication theory, or safety verification.It recommends hybrid semantic–symbolic–physical systems integrating optimization, multi-agent learning, privacy-preserving computation, digital twins, and formally verified safety mechanisms.Source: §VIII · pp. 16–17
Loading 2610.09466v1…