PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Can Safely Wait? Risk-Aware 5G Scheduling for Low-Altitude UAV Operation

André Melo, Yueqi Chen, Lourenço Pereira Jr., Igor Kadota

arXiv:2610.07101v1cs.NI

TL;DR

Conventional cellular schedulers lack operational context for the consequences of delaying aircraft information. RA translates freshness and operating context into aircraft-specific risk to schedule heterogeneous uplink streams. In the illustrated corridor, Best-CQI schedules the crowd-overflight UAV 4× less often than Round-Robin, and its packets wait 8× longer.

  • Problem

    Conventional cellular schedulers use network-centric metrics that do not encode the operational consequences of delaying information across aircraft.

    Source: §1 · p. 1

  • Method

    RA maps information freshness and operational context to aircraft-specific risk, then prioritizes heterogeneous status and video streams by marginal risk while balancing delivery probability and service cost.

    Source: §1 · p. 2, §2.1 · p. 4, §2.3 · p. 5

  • Results

    Best-CQI schedules the crowd-overflight UAV 4× less often than Round-Robin, and its packets wait 8× longer.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    RA operates at the gNB without UE or air-interface changes and relates operational-risk envelopes to the number of supportable aircraft.

    Source: §1 · p. 2

    Positioning and analytics remain limited in accuracy and update rate.

    Source: §1 · p. 2

Abstract

from arXiv · show

Cellular-connected UAV operations already fly in many countries. However, cellular networks were designed for terrestrial users and remain blind to the operational consequences of delaying information from low-altitude aircraft. We present Risk-Aware (RA), a framework that translates information freshness and operational context into aircraft-specific operational risk, enabling uplink scheduling based on operational consequence rather than network metrics alone. RA balances the consequence of further delay against the service cost of heterogeneous single- and multi-slot updates. It combines a regulation-inspired risk proxy, risk-directed prioritization based on marginal risk growth, and a Max-Weight scheduler with Lyapunov-based performance guarantees. Although RA is derived from an analytically tractable model, we evaluate it under moving UAV trajectories, dynamic operational context, geographically varying population exposure, and realistic UAV video traffic. Across 48 operating points and fleets of up to 120 UAVs in S\~ao Paulo delivery operations and the Interlagos Formula 1 Grand Prix setting, RA reduces 99th-percentile per-UAV risk by up to 96% over proportional fair and 86% over round robin, while supporting fleets up to 1.5$\times$ and 3$\times$ larger, respectively, within the same risk envelope.

1 Introduction

Low-altitude aviation needs timely information, but conventional cellular schedulers do not model how delay affects aircraft operations. RA translates freshness and operational context into aircraft-specific risk to prioritize uplink streams by operational consequence.

1 Introduction

  • Conventional schedulers optimize network metrics, while AoI prioritizes staleness alone; neither captures where fresh information has greater operational value.Source: §1 · pp. 1–2
  • Best-CQI schedules the crowd-overflight UAV 4× less often than Round-Robin, and its packets wait 8× longer.Source: §1 · pp. 1–2
  • Operational context evolves over seconds or minutes, while the gNB schedules at millisecond timescales; positioning and analytics remain limited in accuracy and update rate.Source: §1 · p. 2
  • RA translates freshness and operational context into aircraft-specific risk, then balances marginal risk reduction against delivery probability and heterogeneous service cost.Its risk proxy captures separation margins, population exposure, and video coherence; the policy runs at the gNB without UE or air-interface changes.Source: §1 · p. 2
  • Up to 96% lower 99th-percentile per-UAV operational risk vs Proportional Fair, 86% vs Round-Robin, and 83% vs AoI scheduling.Across 48 operating points and fleets of up to 120 UAVs.Source: Figure 1 · p. 3
  • At the same operational-risk envelope, RA supports fleets up to 1.5× larger than Proportional Fair, 3× larger than Round-Robin, and 2–3× larger than AoI scheduling.Source: Figure 1 · p. 3

2 Network Model and Operational Risk

The section frames scheduling as translating SORA-inspired operational context and information staleness into per-UAV risk, then minimizing a tractable surrogate under network resource constraints. Its model accounts for status and multi-packet video updates, location-dependent exposure, separation margins, and video coherence.

2 Network Model and Operational Risk

  • SORA relates operational criticality to population, air traffic, obstacles, and uncertainty, but does not specify cellular resource allocation or per-flow priorities.The framework uses these principles with staleness, location, and flow type to estimate which update is consuming the risk envelope fastest.Source: §2 · p. 3

2.1 Network Model

  • The single-cell uplink model gives each UAV a one-packet status flow and an L-packet video flow, with at most K transmission opportunities per slot.Updates and video frames follow a generate-at-will model; video packets can be sent across nonconsecutive slots, and a frame completes after all L packets succeed.Source: §2.1 · p. 4

2.2 Staleness-Induced Uncertainty

  • Stale status information enlarges a worst-case position-uncertainty region, using a per-slot travel bound of v_max and minimum localization uncertainty r_min.The model follows SORA’s containment principle and uses a disk around the latest reported position; other uncertainty-region shapes are supported.Source: §2.2 · p. 5, §2.3 · p. 5

2.3 Operational Risk Components

  • The risk model captures navigation risk from consumed separation margins, ground risk from exposed population, and video-coherence risk from weighted frame staleness beyond a nominal threshold.These heterogeneous components are normalized at their corresponding operational boundaries.Source: §2.3 · pp. 5–6, Figure 2 · p. 5

2.4 Composite Operational Risk

  • The per-UAV composite risk combines normalized navigation, ground, and video risks with positive weights summing to one; a value of 1 marks the declared boundary.Values above 1 indicate that communication-induced staleness has exhausted the composite risk budget.Source: §2.4 · p. 6

2.5 Surrogate Operational Risk

  • Because composite risk is nonlinear and coupled through system dynamics, the scheduler uses a linear upper bound and a tractable surrogate based on status and latest-frame AoI.Context-dependent risk coefficients are held fixed per realization while the scheduler responds each slot to freshness and video-transmission progress; evaluation considers moving trajectories and dynamic parameters.Source: §2.5 · p. 6

2.6 Optimization Problem

  • The optimization minimizes long-term expected surrogate risk subject to resource constraints, and the optimality ratio measures a policy’s risk relative to the minimum achievable risk.A policy is ψ-optimal when its expected surrogate risk is within a multiplicative factor ψ ≥ 1 of the optimum.Source: §2.6 · p. 7

3 Proposed 5G Scheduling Policy

Section 3 develops a low-complexity Max-Weight policy for minimizing long-term surrogate risk under fixed operational context. It derives a lower bound and uses Lyapunov optimization to establish an optimality guarantee.

3.1 Lower Bound

  • Theorem 3.1 lower-bounds every admissible policy’s long-term expected surrogate risk by L* when status and video delivery probabilities are positive.The lower bound comes from minimizing surrogate risk over feasible long-term packet-delivery rates.Source: §3.1 · p. 7
  • The lower-bound proof uses a sample-path argument and the resource constraint, while the policy guarantee compares Max-Weight with a randomized policy using optimal lower-bound rates.The Lyapunov coefficients make the negative drift terms coincide with surrogate risk, enabling a telescoping drift bound.Source: §3.1 · p. 7, §3.2 · p. 8

3.2 Risk-Aware 5G Scheduling Policy

  • The AoI-only risk-aware Lyapunov function improves the Max-Weight optimality ratio by 1.5× over [51] and achieves up to 83% lower 99th-percentile per-UAV composite risk than [51].The cited risk comparison is reported across evaluated operating points.Source: §3.2 · p. 7
  • Max-Weight minimizes an upper bound on one-slot Lyapunov drift by selecting the streams with the largest scheduling weights.The policy is designed to reduce surrogate risk through expected one-slot drift minimization.Source: §3.2 · p. 8
  • The proposed Max-Weight policy has a worst-case optimality ratio no larger than 4; this bound is conservative, and the parameterized bound is tighter when L* is known.The bound follows by further relaxing the denominator of the parameterized guarantee.Source: §3.2 · p. 8

4 Evaluation

The evaluation tests RA in two São Paulo scenarios across 48 operating points, with dynamic context and realistic UAV traffic. RA reduces tail risk, directs more service to higher-SAIL operations under contention, and supports larger fleets within the same risk envelope.

4.1 Evaluation Setup

  • The evaluation uses two São Paulo scenarios and 48 operating points, with moving UAVs, evolving operational context, population exposure, and status and video traffic.A MATLAB slot-level simulator integrates these factors and is checked against OAI; the maximum total-variation distance between grant-share distributions is 0.021.Source: §4.1 · pp. 9–10

4.2 Does RA Reduce Risk Exposure?

  • RA attains the lowest 99th-percentile per-UAV risk at all 48 operating points, reducing it by up to 96% versus PF, 86% versus RR, and 83% versus AoI.Figure 4 aggregates both scenarios by load regime; the complete operating-point sweep is reported in Appendix H.Source: §4.2 · p. 10, Figure 4 · p. 10
  • AoI achieves the lowest fleet-mean age in every load regime but not the lowest tail risk, because it does not account for the operational consequence of staleness.RA instead directs service toward UAVs whose operational-risk margin grows fastest, accepting a controlled increase in average AoI while reducing tail risk.Source: §4.2 · p. 10

4.3 Does RA Prioritize Operationally Critical UAVs?

  • RA's median scheduling-alignment ratio is 2.37, meaning higher-SAIL UAVs receive more than twice the scheduling probability of SAIL-II UAVs under contention.The alignment persists in matched decisions with comparable AoI and radio state: RA selects the higher-SAIL operation in 64% of decisions, versus 50% for baselines.Source: §4.3 · p. 11, Figure 5 · p. 11

4.4 How Much Operational Capacity Does RA Enable?

  • The capacity test admits a fleet only when the worst-UAV average risk is at most one; this stricter criterion yields the same evaluated admissible region as the 99th-percentile test.When every policy exceeds the envelope, additional communication capacity or reduced traffic demand may be needed.Source: §4.4 · pp. 11–12
  • Within the same risk envelope, RA supports up to 1.5× as many UAVs as PF, 3× as many as RR, and 2–3× as many as AoI.At K = 8, RA supports 60 UAVs versus 40 under PF and 20 under RR and AoI; at K = 16, it supports 120 versus 80 under PF and 60 under RR and AoI.Source: §4.4 · p. 12, Figure 6 · p. 12

5 Related Work

Prior work addresses cellular UAV connectivity, operational safety assessment, and freshness-aware scheduling, but generally does not use operational consequence to decide which aircraft to defer. RA maps freshness and aircraft context to risk for online resource allocation.

Cellular networks and operational risk

  • Cellular UAV research addresses propagation, mobility, coverage, trajectory planning, and resource management, while QoS, slicing, and reservation differentiate traffic without ranking aircraft by delay consequence.These approaches improve connectivity but do not determine which aircraft is most costly to delay under uplink contention.Source: §5 · p. 12

Freshness- and context-aware scheduling

  • Safety frameworks model exposure and navigation risk or include C3 performance in the safety case, but generally treat communication performance as an input rather than an online allocation metric.Freshness-aware scheduling optimizes staleness, and UoI adds context weights, but the cited work leaves choosing practical context weights open.Source: §5 · pp. 12–13
  • RA maps network-observable freshness and aircraft context to normalized navigation, ground, and video-coherence exposure, then jointly schedules single-slot status updates and multi-slot video objects.Its risk envelope also translates online allocation into communication-constrained operational capacity.Source: §5 · p. 13

6 Conclusion

The conclusion presents RA as a way to make operational consequence part of cellular resource allocation and reports reduced risk and larger supportable fleets in the evaluated São Paulo scenarios. The appendix normalizes heterogeneous risk components against declared operational limits.

6 Conclusion

  • RA translates freshness and aircraft context into operational risk, then prioritizes aircraft and streams for which additional delay is most consequential while accounting for heterogeneous transmission costs.It is evaluated under moving trajectories, dynamic context, geographically varying population exposure, and realistic UAV video traffic.Source: §6 · p. 13

A Normalization of Operational Risk Components

  • The composite risk combines navigation, ground-exposure, and video-coherence components, each normalized by a conservative reference value based on its declared operational limits.The normalized components are dimensionless; a value of one corresponds to the reference scale, and a value above one indicates at least one corresponding limit is exceeded.Source: A Normalization of Operational Risk… · pp. 14–15
  • The navigation reference scale uses declared fleet-size, status-update staleness, and separation limits, while the ground reference bounds exposed population using population density and the uncertainty-region area.Video-coherence risk is normalized to zero at nominal staleness and one at the maximum tolerable staleness.Source: A Normalization of Operational Risk… · p. 15

B Derivation of the Composite Operational Risk Upper Bound

The appendix derives an upper bound on the composite operational risk by separately bounding its navigation, ground, and video-coherence components under the considered operational context.

Navigation Risk Upper Bound

  • The derivation uses local population-density bounds and pairwise UAV separations from the considered operational context to bound navigation risk.It applies a secant-line bound over the declared status-update staleness range.Source: B Derivation of the Composite… · p. 15

Ground Risk Upper Bound

Video-Coherence Risk Upper Bound

C PROOF OF THEOREM 3.1

The proof relates finite-horizon status and video freshness to delivery rates, derives a one-slot Lyapunov drift bound for scheduling weights, and uses stationary randomized policies as comparisons. It concludes that every admissible policy’s limiting risk lower bound is at least the optimum lower-bound value.

C PROOF OF THEOREM 3.1

  • Finite-horizon status and video AoI bounds are expressed through successful status deliveries and completed video frames.The proof defines expected delivery and frame-completion rates, then characterizes feasible rate vectors.Source: C PROOF OF THEOREM 3.1 · pp. 16–18
  • For every admissible policy, the limiting lower-bound argument yields bJ* ≥ L*B.The proof considers limit points of feasible finite-horizon delivery-rate vectors before applying the lower-bound problem.Source: C PROOF OF THEOREM 3.1 · p. 18

D Upper Bound on the One-Slot Lyapunov Drift

E Stationary Randomized Policy

  • A stationary randomized policy selects feasible actions independently each slot, and feasible marginal scheduling probabilities can be realized by a distribution over those actions.The lower-bound solution induces a feasible stationary randomized comparison policy.Source: E Stationary Randomized Policy · pp. 19–20

F Proof of Theorem 3.2

The proof of Theorem 3.2 compares the Max-Weight policy with a feasible stationary randomized policy. It uses per-slot weight maximization and a telescoping Lyapunov-drift argument to establish the stated bound.

F Proof of Theorem 3.2

  • The proof compares Max-Weight’s per-slot scheduling choice with a feasible stationary randomized policy induced by the lower-bound solution.Max-Weight maximizes total scheduling weight in every slot; the proof then sums drift inequalities and telescopes the Lyapunov terms.Source: F Proof of Theorem 3.2 · pp. 20–21
  • The long-term expected surrogate risk of Max-Weight is bounded using the drift comparison and nonnegative scheduling weights.The proof combines the limiting drift inequality with the nonnegativity conditions stated in the passage.Source: F Proof of Theorem 3.2 · pp. 20–21

G Real-World Data for the Evaluation Scenarios

The evaluation scenarios use real São Paulo locations and documented operating ranges, but construct controlled models rather than replaying measured flights or cellular traces. Public infrastructure and operational data ground the scenario geography while channel, population, and traffic models isolate scheduling behavior.

G Real-World Data for the Evaluation Scenarios

  • The scenarios are controlled models over real geographies, not replays of measured flights or cellular traces.Public infrastructure and operational data constrain the selected areas and their geometry.Source: G Real-World Data for the Evaluation… · p. 22
  • Delivery trajectories span 4.0–4.9 km at 60 m and 15 m/s, matching the distance, altitude, and speed range of a reported Alphaville operation.The cited operation reports a 3.6-km aerial leg at up to 60 m AGL and 50 km/h.Source: G Real-World Data for the Evaluation… · p. 22
  • The Interlagos scenario follows the 4.3-km circuit and models spatially nonuniform spectator regions, without reproducing event traffic-management procedures.The scenario is motivated by documented coexistence of dense crowds and aerial activity.Source: G Real-World Data for the Evaluation… · p. 22

H Evaluation Across the 𝐾× 𝑁Grid

Across the evaluated channel and fleet-size grid, RA supports larger fleets within the risk envelope than PF, RR, and AoI at the stated channel budgets. Risk gains involve freshness trade-offs and remain limited by available radio capacity.

H Evaluation Across the 𝐾× 𝑁Grid

  • At K=8 and N=60, RA’s mean video AoI is 72 slots, compared with 83 for PF and 84 for RR.RA shifts available freshness toward status updates under heavy contention rather than minimizing every age equally.Source: H Evaluation Across the 𝐾× 𝑁Grid · p. 22
  • For K≤4, no policy keeps the worst UAV inside the risk envelope beyond N=20; at K=1 and N=120, even RA exceeds the worst-UAV boundary by 52×.The passage attributes the boundary to the service deficit at that operating point.Source: H Evaluation Across the 𝐾× 𝑁Grid · pp. 22–23
  • At K=8, RA supports N=60 under max_u R_u≤1, versus 40 for PF and 20 for RR and AoI; at K=16, the respective fleet sizes are 120, 80, and 60.The worst-UAV admission frontier coincides with the 99th-percentile-based frontier at every channel budget.Source: H Evaluation Across the 𝐾× 𝑁Grid · p. 22, Figure 8 · p. 23

I SAIL Assignment and Allocation Audit

The allocation audit compares RA’s service to SORA operation classes in controlled matched decisions and across the evaluation grid. Sensitivity tests find that RA retains the latest departure from the risk envelope as video frames grow.

I SAIL Assignment and Allocation Audit

  • The controlled audit assigned SAIL-II to the low-exposure operation and SAIL-V to the crowd-exposed operation, with labels fixed during flight and hidden from schedulers.SAIL assignment combines ground and residual air risk classifications; crowd exposure can affect the operation’s classification, including its low-exposure segments.Source: I SAIL Assignment and Allocation Audit · pp. 23–24
  • RA served the SAIL-V candidate in 64% of matched comparisons on median, versus 50% for RR, 50% for PF, and 49% for AoI.At the most contended points, RA reached 95%; matching controlled status AoI and SNR within stated tolerances.Source: I SAIL Assignment and Allocation Audit · p. 24
  • In the controlled audit, RA gave SAIL-V operations a median 1.8× and up to 3.1× the per-UAV scheduling probability.Baseline ratios were 0.90–0.91, treated as scenario-induced variation.Source: I SAIL Assignment and Allocation Audit · p. 24
  • Across 42 evaluation points with N > K, RA’s median allocation ratio was 2.37, compared with 0.92 for each baseline.The six points with K ≥ N were included in this audit but excluded from Fig. 5 because allocation differences there primarily reflect traffic availability.Source: I SAIL Assignment and Allocation Audit · p. 24
  • Replacing the realistic GOP workload with constant L = 5 changed sides of the risk envelope for 19 of 192 policy–operating-point pairs, without changing policy ordering.Removing I-frame bursts lowered median p99,u[R_u] by 29% for RR and AoI, 22% for PF, and 12% for RA.Source: I SAIL Assignment and Allocation Audit · p. 24
  • With constant frame size increasing from 5 to 30 packets, median tail risk rose 2.6× for RA, 2.9× for PF, 3.6× for AoI, and 1.9× for RR.RA was the last policy to leave the operational envelope; larger frames increased risk for every policy without changing their ranking.Source: I SAIL Assignment and Allocation Audit · p. 25, Figure 10 · p. 25

K From Radio to Air Capacity

Communication-constrained operational capacity can inform traffic management and infrastructure planning, while RA’s interpretation remains limited to communication-induced exposure. The evaluation uses a slot-level simulator with specified mobility, radio, delivery, and delay assumptions.

K From Radio to Air Capacity

  • When projected simultaneous demand approaches regional communication capacity, traffic management can reshape demand through departure delays, speed changes, rerouting, or deferring aircraft.The network exposes when expected traffic exceeds the declared communication envelope; it does not authorize operations.Source: K From Radio to Air Capacity · p. 25
  • Regions repeatedly approaching the communication boundary can identify where network upgrades or traffic shaping may provide the greatest benefit.Infrastructure expansion can be considered in terms of additional low-altitude operations supported within the same communication-risk constraint.Source: K From Radio to Air Capacity · p. 25
  • The interpretation covers communication-induced exposure, while multi-cell systems, handover regions, and city-scale traffic management remain future work.RA complements the broader aviation safety case, including SORA, airspace constraints, weather, and vehicle reliability.Source: K From Radio to Air Capacity · p. 25

L Simulation Toolbox Parameters

  • The MATLAB evaluation uses waypoint trajectories, deterministic 3GPP UMa line-of-sight path loss, one serving gNB, and packet success probability 0.8 when SNR is at least 5 dB.Each of K channels carries one packet per slot, and the simulator adds a fixed 10 ms one-way radio, core, and processing delay.Source: L Simulation Toolbox Parameters · p. 25
Loading 2610.07101v1…