PaperlayerView exact revision on arXiv

Source-linked AI summary

Low-Altitude Agentic Networks: From UAV Connectivity to Workflow-Driven Capability Orchestration

Baolin Chong, Fuhui Zhou, Qihui Wu, Ran Tu

arXiv:2610.17642v1eess.IV

TL;DR

Distributed UAV workflows need providers that remain feasible as network, resource, and mobility conditions change. The paper proposes workflow-centered agentic networking, with validated workflows, dynamic capability binding, and local adaptation. Its case study finds benefits within a bounded region, especially for coupled workflows with changing feasibility and viable recovery paths.

  • Problem

    Mission workflows depend on capabilities distributed across UAVs, but provider feasibility can change during execution and existing approaches do not necessarily maintain evolving workflow-provider relations as runtime state.

    Source: D. Low-Altitude Agentic Networks · p. 2

  • Method

    The architecture binds each ready operation in a validated workflow to a currently feasible agent and uses local workflow-preserving adaptation before invoking high-level repair.

    Source: D. Low-Altitude Agentic Networks · p. 2, C. Mobility-Coupled Resilience · p. 5

  • Results

    The case study finds agentic organization most beneficial for coupled workflows and recurring disruptions, with larger gains when lower link reliability coincides with sufficient provider redundancy.

    Source: B. Evaluation Results · p. 6

  • Takeaways & Limitations

    The value proposition is bounded: conventional organization remains effective for simple workflows under stable conditions, while agentic organization benefits workflows with coupled dependencies, changing feasibility, and viable recovery paths.

    Source: §VII · p. 7

    The supported benefit depends on coupled workflow dependencies, changing capability feasibility, and viable local recovery paths.

    Source: §VII · p. 7

Abstract

from arXiv · show

Low-altitude platforms are evolving from connected sensing nodes into autonomous agents capable of bounded task interpretation, skill invocation, and local decision making. Increasing the intelligence of individual unmanned aerial vehicles (UAVs), however, does not resolve the organization problem that arises when mission workflows depend on capabilities distributed across multiple platforms and provider feasibility changes with network, resource, and mobility conditions. We propose low-altitude agentic networking as a task-level networking paradigm and develop a reference architecture that separates workflow logic from executor identity. Within a validated domain-specific workflow space, each ready operation is translated into a capability requirement and dynamically bound to an agent that remains feasible under current conditions. Task contracts and structured evidence govern workflow execution, while routine provider and link changes are handled through workflow-preserving local adaptation before high-level repair is invoked. A trace-calibrated discrete-event case study identifies a bounded operating region: conventional organization remains effective for simple workflows under stable conditions, whereas agentic networking becomes beneficial when workflow dependencies are tightly coupled, capability feasibility changes during execution, and viable local recovery paths remain available.

A. Centralized UAV Networks

Centralized control organizes UAV operations effectively with current global state, while distributed allocation decentralizes executor selection for predefined tasks. Agent-enabled UAVs add local interpretation and skill invocation, but predefined collaboration does not maintain evolving workflow-provider relations.

A. Centralized UAV Networks

  • Centralized coordination works effectively with current global state, but recovery depends on the controller re-establishing an affected task relation.A suitable nearby replacement does not remove that global-control-loop dependency.Source: A. Centralized UAV Networks · p. 2

B. Distributed Task Allocation

  • Distributed allocation lets UAVs negotiate executor selection for newly arrived tasks or failed assignments, but tasks and dependencies are typically defined in advance.It decentralizes executor selection without necessarily maintaining evolving workflow-provider relations.Source: B. Distributed Task Allocation · p. 2

C. Agent-Enabled UAVs with Predefined Collaboration

  • Agent runtimes improve local task interpretation and skill invocation, yet predefined roles do not necessarily find the most suitable provider for an emerging capability need.Node intelligence governs how an assigned operation is performed; network-level intelligence is also needed to form collaboration relations.Source: C. Agent-Enabled UAVs with Predefined… · p. 2

D. Low-Altitude Agentic Networks

Low-altitude agentic networking organizes collaboration around mission workflows, separating task logic from executor identity. Ready operations bind to currently feasible agents, with workflow-preserving adaptation preferred over workflow repair when providers become infeasible.

D. Low-Altitude Agentic Networks

  • Low-altitude agentic networking forms and maintains collaboration relations around mission workflows, separating each operation from a permanently assigned UAV.The system selects a task-appropriate subgraph from a validated workflow space and can bind the same operation to different agents as conditions evolve.Source: D. Low-Altitude Agentic Networks · p. 2
  • A required skill alone does not make an agent feasible; the provider must execute the operation and deliver its result under current task, network, resource, and mobility conditions.When a provider becomes infeasible, communication adaptation or rebinding is tried first; workflow repair follows only if no feasible realization can satisfy the objective.Source: D. Low-Altitude Agentic Networks · p. 2

III. Reference Architecture and Task Lifecycle

  • The reference architecture and lifecycle transform a mission into a bounded workflow, bind it to feasible agents, and adapt execution over time.Figure 2 presents this lifecycle from task-contract formation to evidence-driven adaptation.Source: §III · pp. 2–3, Figure 2 · p. 3

A. Three-Plane Architecture and Hierarchical Intelligence

The architecture divides mission reasoning, capability orchestration, and network/platform execution across three interacting planes. Task contracts, operational state, and execution evidence connect task origination and workflow execution.

A. Three-Plane Architecture and Hierarchical Intelligence

  • Three interacting planes handle mission interpretation and workflow activation, feasible-agent binding, and communication and platform actions under deterministic operational constraints.Execution results return as evidence and update workflow state.Source: A. Three-Plane Architecture and… · p. 3
  • Ground or edge LLMs handle mission-level interpretation and uncommon repair, while onboard SLMs or deterministic runtimes support routine decisions and local skill invocation.Perception, navigation, tracking, communication control, and flight control remain with specialized models and deterministic controllers.Source: A. Three-Plane Architecture and… · p. 3

B. Task Origination, Contracts, and Evidence

  • A ground or edge mission agent typically introduces the mission, while local evidence, delegation, or failure events can derive subtasks during execution.Distributed task derivation does not require every UAV to act as a complete mission planner.Source: B. Task Origination, Contracts, and… · p. 3
  • A task contract specifies the intended outcome, required inputs and evidence, and temporal, reliability, resource, and safety constraints for execution.Delegated agents receive the relevant contract and evidence rather than the complete natural-language mission history.Source: B. Task Origination, Contracts, and… · p. 3
  • Operational state guides provider selection, whereas execution evidence records operation results and is evaluated by the workflow runtime.Ambiguous or conflicting evidence may be escalated for model-assisted interpretation.Source: B. Task Origination, Contracts, and… · p. 3

C. Graph-Assisted Workflow Formation and Capability Binding

Workflow formation uses a validated domain-specific space, while ready operations are bound to agents that remain feasible under current conditions. Separating task logic from executor identity allows provider changes without discarding completed operations or valid evidence.

C. Graph-Assisted Workflow Formation and Capability Binding

  • The domain-specific workflow space is validated for interfaces, preconditions, evidence requirements, structural validity, and safety before its operations and dependencies are admitted.Candidate paths may come from expert templates, successful traces, or model-generated proposals.Source: C. Graph-Assisted Workflow Formation… · p. 3
  • For each task contract, the reasoning plane instantiates a task-appropriate validated subgraph; evidence selects its branch, and constrained repair is proposed only when no valid continuation satisfies the contract.A confident verification may activate tracking, while an uncertain result may activate independent verification.Source: C. Graph-Assisted Workflow Formation… · p. 3
  • The capability registry filters declared-skill agents for feasibility under current task, network, resource, mobility, and safety conditions, then ranks feasible providers by service quality and completion cost.Binding when an operation becomes ready lets selection reflect the latest system state.Source: C. Graph-Assisted Workflow Formation… · pp. 3–4
  • Separating task logic from executor identity lets an operation move to another UAV without discarding completed operations or valid evidence.The networking support is organized around capability-aware control, workflow-aware communication, and mobility-coupled resilience.Source: C. Graph-Assisted Workflow Formation… · p. 4, Figure 3 · p. 4

D. End-to-End Execution and Workflow-Preserving Adaptation

Ready workflow operations are bound to feasible agents as the mission progresses, with execution evidence updating workflow state. When conditions change, the system first tries to preserve the active workflow through local network or platform adaptations.

D. End-to-End Execution and Workflow-Preserving Adaptation

  • Search is bound first; candidate-target evidence makes verification ready, and verification confidence determines whether tracking or a validated re-verification branch follows.Binding each operation only when ready lets the provider reflect the latest target, agent, and network state.Source: D. End-to-End Execution and… · p. 4
  • If tracking becomes infeasible, the runtime can restore the path, add a relay, reposition, or rebind while preserving the mission objective and accumulated evidence.High-level workflow repair is used only when no feasible local response can maintain the active task contract.Source: D. End-to-End Execution and… · p. 4

IV. Networking Foundations for Agentic

  • Communication supports binding and execution by exposing fresh provider information, carrying task contracts and evidence, and updating relations as conditions change.The communication structure follows the active workflow rather than a permanent collaboration graph.Source: §IV · p. 4

A. Capability-Aware Network Control

Capability-aware control distinguishes an agent’s declared skill from its ability to deliver that capability under current conditions. Provider discovery and workflow communication are coordinated around feasible bindings and timely task completion.

A. Capability-Aware Network Control

  • A capability record combines a stable skill description with dynamic position, workload, energy, availability, reachability, and state-validity information.This lets orchestration distinguish declared capability from task-specific feasibility under the active contract.Source: A. Capability-Aware Network Control · p. 4
  • Periodic advertisements, event-triggered updates, and on-demand confirmation can maintain a task-sufficient view without reporting every state change.Discovery can start locally and expand when nearby information is insufficient.Source: A. Capability-Aware Network Control · p. 4
  • Figure 4(c) outlines cells that meet the predefined statistical and efficiency criteria for the strict advantage region.Source: Figure 4 · p. 5

B. Workflow-Aware Communication and Orchestration

  • Workflow dependencies create temporary communication relations between current providers, with task contracts specifying exchange delay, reliability, and freshness requirements.If an operation is rebound, the communication endpoints change accordingly.Source: B. Workflow-Aware Communication and… · p. 5
  • Communication should carry representations sufficient for downstream operations and prioritize evidence on the active workflow path over background updates.Provider selection, routing, and communication-resource allocation are coordinated around timely workflow completion.Source: B. Workflow-Aware Communication and… · p. 5

C. Mobility-Coupled Resilience

Mobility can make a provider infeasible even when it retains the required skill, because sensing or communication conditions may no longer meet the task contract. Recovery prioritizes feasible, least-disruptive local actions before high-level workflow repair.

C. Mobility-Coupled Resilience

  • A provider may retain its skill yet become infeasible when mobility changes sensing geometry, obstruction, or communication quality.Its evidence path may no longer meet required delay, reliability, or freshness.Source: C. Mobility-Coupled Resilience · p. 5
  • Recovery can restore a route, add a relay, or rebind the affected operation; repositioning is another option when movement can restore sensing or communication feasibility.Source: C. Mobility-Coupled Resilience · p. 5
  • Table I reports mobility-induced workflow recovery over 50 paired seeds and defines continuity as the fraction of the required tracking interval satisfying the contract.Source: Table I · p. 5
  • Responses differ in delay, energy use, and task impact, so the network selects the least disruptive feasible action without violating safety or disrupting another critical task.Source: C. Mobility-Coupled Resilience · p. 5
  • Routine provider and connectivity changes are handled through routing, relaying, rebinding, or repositioning; workflow repair waits until no local action can preserve execution.Model-assisted reasoning is reserved for cases where the validated workflow has no feasible continuation.Source: C. Mobility-Coupled Resilience · p. 5

V. Case Study

The case study evaluates agentic organization across workflow complexity, capability and network conditions, and mobility recovery. Its discrete-event simulations compare four organization policies using paired episodes and trace-calibrated reasoning delays.

V. Case Study

  • The case study evaluates benefits under increasing workflow complexity, operating regions across capability and network conditions, and recovery under low-altitude mobility.Source: §V · p. 5

A. Simulation Setup

  • The simulation integrates a Workflow Runtime, Contract Manager, Capability Registry, and Network Emulator, with model traces calibrating reasoning delays rather than an onboard computing platform.Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct were profiled in BF16 on an NVIDIA A40.Source: A. Simulation Setup · pp. 5–6
  • Figure 5 frames open research problems in real-time intelligence, scalable orchestration, trustworthy execution, and mobility co-design.Source: Figure 5 · p. 6
  • Experiments 1 and 2 use 16 heterogeneous logical agents in a 1,000 m × 1,000 m area with a 300-s limit; Experiment 3 uses eight UAVs and one ground/edge node with a 130-s limit.The Tracking Contract requires at least 0.90 evidence-delivery reliability and at most 0.20 s delay, with recovery after three consecutive invalid updates.Source: A. Simulation Setup · p. 6
  • The four policies are centralized global replanning, bounded-auction distributed allocation, predefined Primary/Backup/Relay assignments, and dynamic feasible-provider binding with four recovery actions.Paired episode manifests give policies identical initial states, skill outcomes, target states, disruptions, and model traces.Source: A. Simulation Setup · p. 6
  • Experiment 1 varies workflows from Search–Verify through continuous tracking, re-verification, link degradation, and multi-provider evidence fusion; Experiment 2 scans provider count K and pre-recovery transfer reliability R.Experiment 3 introduces building obstruction that makes the initial tracker’s path contract-infeasible.Source: A. Simulation Setup · p. 6
  • Mission success rate is primary; Experiment 1 also reports restricted mean completion time, while strict advantage requires a significant ∆SR gain and bounded overhead and completion time.The criteria include ∆SR ≥0.05, a positive lower bound of the paired-bootstrap 95% confidence interval, overhead within 2×, and completion time within 1.2× the best baseline.Source: A. Simulation Setup · p. 6

B. Evaluation Results

Dynamic organization’s benefit depends on workflow complexity, network variation, and feasible alternatives. Under mobility-induced recovery, Proposed matches Centralized’s success rate while improving recovery time and continuity.

B. Evaluation Results

  • At W5, Proposed reaches an SR of 0.66 versus 0.55 for the best baseline and reduces RMCT from 235.77 s to 223.64 s.At W4, Proposed matches the highest baseline SR of 0.78 and achieves the shortest RMCT; W1–W3 policies remain close.Source: B. Evaluation Results · p. 6
  • At R = 0.95, ∆SR remains between 0.01 and 0.09, while larger gains reach 0.18 at lower reliability with sufficient provider redundancy.The reported advantage depends jointly on network variation and feasible alternatives; some large raw differences fail a statistical or efficiency criterion.Source: B. Evaluation Results · p. 6
  • Proposed and Centralized both achieve an SR of 1.00, while Proposed reduces median recovery time from 9.10 s to 4.89 s and mean disruption time from 9.57 s to 5.10 s.Proposed increases continuity from 0.787 to 0.887; contract-triggered local adaptation restores the workflow before remote global replanning completes.Source: B. Evaluation Results · p. 7

VI. Open Research Problems

The proposed mechanisms leave open questions about timely reasoning, scalable capability coordination, and trustworthy execution. The discussion identifies these tensions and directions for addressing them.

A. Real-Time Intelligence

  • Adaptive hierarchical intelligence could select local or remote reasoning while preserving a validated fast path for time-critical decisions.Ground-side inference adds communication delay and may be unreachable after link failure, while onboard models may not handle events outside the active workflow.Source: A. Real-Time Intelligence · p. 7

B. Scalable Orchestration

  • Hierarchical capability views and task-aware orchestration are possible directions for balancing shared state against global synchronization.Frequent updates consume resources, stale views risk invalid bindings, and concurrent workflows compete for capabilities.Source: B. Scalable Orchestration · p. 7

C. Trustworthy Execution

  • Trustworthy execution requires common machine-readable semantics and runtime verification, while independent safety rules and human authority remain necessary for high-risk actions.Workflow spaces must be constructed, versioned, and extended without unsafe or physically infeasible paths; evidence can be outdated, conflicting, or manipulated.Source: C. Trustworthy Execution · p. 7

D. Mobility Co-Design

Mobility co-design must account for movement’s effects on both communication and task observation. The conclusion bounds the value of agentic organization to coupled workflows with changing feasibility and viable recovery paths.

D. Mobility Co-Design

  • UAV movement can restore communication while losing a target or disrupting another workflow, so co-design should coordinate fast network adaptation with slower physical actions.The proposed evaluation should assess whether movement produces a net gain in workflow continuity.Source: D. Mobility Co-Design · p. 7

VII. Conclusion

  • Conventional organization remains effective for simple workflows under stable conditions, while dynamic agentic organization benefits coupled workflows when capability feasibility changes and recovery paths remain available.The paradigm extends networking beyond connectivity and data delivery toward maintaining task-specific relations that organize distributed intelligence.Source: §VII · p. 7

Related papers

Loading 2610.17642v1…