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

Who Can Safely Wait? Risk-Aware 5G Scheduling for Low-Altitude UAV Operations André Melo

arXiv:2610.07101v1 [cs.NI] 5 Oct 2026

Aeronautics Institute of Technology São José dos Campos Brazil andre.melo@ita.br

Yueqi Chen

Lourenço Pereira Jr.

Northwestern University Aeronautics Institute of Evanston, IL, USA Technology yueqichen@northwestern.eduSão José dos Campos Brazil ljr@ita.br

Abstract

airspace constraints, and airspace traffic information. Remotely supervised and autonomous operations may also generate video and other high-volume sensing streams [5, 25, 39] for remote situational awareness. The Specific Operations Risk Assessment (SORA) [20, 21], a risk-assessment methodology for UAV operations, classifies operations into a Specific Assurance and Integrity Level (SAIL) class I-VI that sets the rigor of the required mitigations, and requires the Command, Control, and Communication (C3) link to have performance and spectrum characteristics appropriate for the operation. For example, flights over sparsely populated ground receive lower SAIL classes, whereas flights over crowds receive higher classes. Commercial cellular networks are a promising component of this communication architecture [46, 47] because they provide widespread infrastructure, high-capacity uplinks, dynamic resource allocation, and integration with edge computing. Cellular-connected UAV operations are already authorized or active in several countries [9, 28, 40], shifting the central question from whether cellular networks can connect aircraft to whether they can allocate limited radio resources consistently with operational safety [46, 47]. Cellular connectivity alone, however, does not make the network suitable for airspace management. Under contention, the Radio Access Network node (also called Next Generation Node B, or gNB) must decide which aircraft and information streams to serve and which to defer. Conventional schedulers make this decision using network-centric metrics such as Channel Quality Indicator (CQI), Buffer Status Reports (BSRs), throughput, fairness, and traffic class. These metrics characterize the opportunity and cost of transmitting data, but not the operational consequence of delaying it. As a result, a scheduler may improve every network metric it observes while withholding resources from the aircraft for which stale information is most consequential. Figure 1 exposes this mismatch in a minimal six-UAV corridor where radio quality and operational criticality vary in opposite directions. Uplink SNR decreases toward a crowded destination, so the segment near the gNB corresponds to a SAIL II operation while over the crowd corresponds to

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 aircraftspecific 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 Sao 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× and 3× larger, respectively, within the same risk envelope.

1

Igor Kadota Northwestern University Evanston, IL, USA kadota@northwestern.edu

Introduction

Low-altitude airspace is poised for a fundamental transformation. Airspace authorities worldwide [10, 38, 44–46] are developing new traffic-management frameworks for UAVs, powered-lift vehicles, remotely piloted aircraft, and increasingly autonomous vehicles operating alongside traditional aviation. Existing air traffic infrastructure, built largely around voice communications and surveillance (radar), was not designed for the anticipated density, heterogeneity, or automation of low-altitude aviation [15, 46]. This transformation will require modern communication [46] for timely information exchange among aircraft, operators, and infrastructure. These traffic-management frameworks [10, 38, 44–46] will rely on timely exchange of aircraft status, telemetry, intent, 1

ACM MobiCom ’27, October 2027,

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

SAIL V. Every UAV carries the same 30 fps video and 10 Hz status traffic (described in § 4.1). Best-CQI maximizes network throughput by concentrating service near the gNB and repeatedly deferring the UAV flying over the crowd: it schedules that aircraft 4× less often than Round-Robin (Fig. 1b), and its packets wait 8× longer (Fig. 1c). Clearly, throughput maximization comes at the cost of increased operational risk. The missing operational context matters under contention because equal communication delays can have different physical consequences. Stale position information from an UAV over an unpopulated corridor may consume little operational risk margin, while the same staleness near another aircraft or above a dense crowd may produce substantially greater risk. The scheduler must also account for heterogeneous transmission costs: an aircraft-status update may require a single transmission opportunity, whereas a video frame may require many. Resource allocation must therefore balance the operational value of transmitting an update against the communication resources required to deliver it. Accounting for this balance can reduce communication-induced operational risk for a given fleet (§ 4.2) or support a larger fleet within a fixed operational risk envelope (§ 4.4). Existing network mechanisms do not account for this context. Slicing and reservation can isolate aviation traffic or provision capacity for a traffic class [16, 49], but they do not determine which aircraft or information stream is most costly to defer. Channel-aware and proportional-fair schedulers prioritize radio efficiency or balance throughput across users, without accounting for differences in operational risk. Age of Information (AoI) scheduling prioritizes stale information [18, 26, 41, 50, 51], but staleness alone does not identify where freshness is most valuable. This exposes a fundamental blind spot: cellular schedulers observe network state, while the operational consequences of their decisions depend on context outside the network’s decision model. Bridging this gap requires an interface that translates network state and operational context into a quantity the scheduler can evaluate. Such an interface would allow the network to compare not only the efficiency of serving each stream, but also the operational cost of making it wait. Emerging 5G capabilities make such an interface increasingly feasible. The gNB already maintains per-UE radio and traffic state through CQI, BSRs, HARQ feedback, and delivery history. Packet timestamps and delivery records expose the staleness of aircraft-status, telemetry, and sensing streams. Aircraft position can be obtained from onboard reports or the 5G Location Management Function (LMF), while the Network Data Analytics Function (NWDAF) and external geographic data can provide estimates of population exposure and other operating conditions [2, 3]. Operational context, such as aircraft location and population exposure, evolves

slowly over seconds or minutes. The gNB combines the operational context with per-slot network metrics to make scheduling decisions at millisecond timescales. Although positioning and analytics remain limited in accuracy and update rate [1, 29], they provide the information needed to make uplink scheduling operationally aware. Using this information effectively requires a scheduler to satisfy three design requirements. First, it must prioritize marginal operational risk, directing service toward the aircraft for which an additional unit of delay is most consequential. Second, it must balance this urgency against heterogeneous service costs, since a status update may require one transmission opportunity while a video frame may require many. Third, it must be deployable within the existing 5G architecture. In particular, it must convert slowly evolving operational context into priorities that can be evaluated within the gNB’s per-slot scheduling loop. This paper asks a simple question: When a cellular network cannot immediately serve every aircraft, who can safely wait? We answer this question with Risk-Aware (RA) scheduling, a framework that makes operational risk the organizing principle for 5G uplink resource allocation. RA prioritizes the aircraft and information streams for which further delay is most consequential, while accounting for delivery probability and service cost. The resulting policy operates entirely at the gNB, requiring no modification to the UE or air interface. RA also relates the resulting operationalrisk envelope to the number of aircraft supportable by the available wireless resources, connecting radio scheduling to operational capacity. The problem formulation and evaluation scenarios incorporate input from air-traffic operations, airspace management, and safety-assessment practices. We make the following contributions: (i) A network-computable operational-risk abstraction. We translate gNB-observable network metrics and operational context into aircraft-specific operational risk. Drawing on the principles underlying the SORA [20, 21], our model captures three consequences of stale information: position uncertainty consuming aircraft-separation margin, population exposure within the position uncertainty region, and loss of coherence across recently delivered video frames. (ii) A Risk-Aware Max-Weight scheduler for heterogeneous updates. We formulate the joint scheduling of UAVstatus updates and video frames as a long-term operationalrisk minimization problem. We derive a universal lower bound over feasible status-delivery and frame-completion rates, then design a Lyapunov function that accounts for video staleness, transmission progress, and delayed frame completion. The resulting RA Max-Weight policy balances marginal operational-risk reduction, packet-delivery probability, and service cost using a low-complexity per-slot rule. We establish a worst-case optimality ratio of four. 2

0

10 d

−500 0

1000

2000 East [m]

3000

(b)

(c) Crowd (SAIL V)

0.3

0.0

Crowd (SAIL V)

104

0.2 0.1

Average Packet Delay [ms]

500

20 dB

B

North [m]

(a)

ACM MobiCom ’27, October 2027, Scheduling Probability

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

RR Best-CQI 1000 2000 3000 Position Along Corridor [m]

103 102 1000 2000 3000 Position Along Corridor [m]

Figure 1: The scheduling blind spot using an OpenAirInterface (OAI) gNB emulator with a complete 3GPP MAC pipeline. (a) Six UEs emulate static UAVs along a corridor whose uplink channel degrades from the gNB toward a crowded event; segments are labeled with the SAIL risk class (II over open ground, V over the crowd).(b) Scheduling probability, the fraction of uplink slots in which each aircraft is scheduled (up to 4 UEs can share a slot). (c) Average packet delay of status updates and video frames, in log scale. (iii) Validation and evaluation grounded in realistic Paper organization. Section 2 defines the system model, operating environments. We develop a scalable slot-level operational-risk abstraction, and optimization problem. SecMATLAB simulator for evaluating 5G uplink scheduling in tion 3 develops the Risk-Aware scheduler and establishes its low-altitude UAV operations. The simulator is cross-checked performance guarantees. Section 4 evaluates operational risk, against UAV Toolbox [43] mobility and driven by 5G Toolnetwork performance, and operational capacity. Section 5 box [42] propagation models, while capturing realistic UAV reviews related work. Section 6 concludes the paper. traffic generation from [8]. We construct controlled scenarios grounded in real São Paulo geographies, licensed 5G 2 Network Model and Operational Risk deployments, documented UAV delivery operations, popuThe Specific Operations Risk Assessment (SORA), which lation maps, and a mass-gathering setting at the Interlagos underpins contemporary UAV regulations and standards [12– Formula 1 circuit. Because this abstraction enables broader 15, 38], characterizes an operation through its final Ground evaluation than is practical with a full 5G stack, we validate Risk Class (GRC) and residual Air Risk Class (ARC). Their it against the OAI implementation used in Fig. 1 under the combination determines a SAIL Level (SAIL I–VI), which proportional-fair experiments described in § 4.1. We evaluin turn specifies the robustness required for the applicable ate RA and baseline policies across 48 operating points with Operational Safety Objectives (OSOs) [20]. fleets of up to 120 UAVs. RA reduces 99th-percentile per-UAV Relevant to our setting, OSO #06 requires that the peroperational risk by up to 96% relative to Proportional Fair, formance of the C3 links is appropriate for the intended 86% relative to Round-Robin, and 83% relative to AoI scheduloperation. SORA, however, does not specify how a cellular ing. Within the same operational-risk envelope, RA supports scheduler should allocate resources among simultaneous fleets up to 1.5× larger than Proportional Fair, 3× larger than C2, telemetry, sensing, and video transmissions, nor does it Round-Robin, and 2–3× larger than AoI scheduling. define per-flow scheduling priorities [20]. (iv) Communication-constrained operational capacYet, SORA provides the principles needed to close this gap. ity. Using our simulator, we evaluate the largest simultaIts risk classes relate operational criticality to population, air neous UAV fleet that a given uplink resource budget can traffic, obstacles, and uncertainty [21, § 1.4]. For example, the support while keeping every aircraft within the declared intrinsic Ground Risk Class (iGRC) combines population dencommunication-induced operational-risk envelope. Unlike sity with an uncertainty footprint, whose radius is derived conventional capacity metrics based on throughput or defrom the aircraft’s maximum speed. By combining these prinvice count, this abstraction incorporates aircraft context, ciples with network-induced information staleness, location, the consequences of stale information, and the allocation and flow type, we enable the 5G scheduler to infer which of radio resources among heterogeneous streams. It constale update is consuming the operational risk envelope the nects millisecond-scale gNB scheduling to longer-timescale fastest. UAV admission control and traffic-flow management: when Next, we describe our network model; then we relate the projected UAV demand approaches the operational-capacity network performance to operational risk. Finally, we formuboundary, the network can expose the need to delay UAV late the 5G scheduling optimization problem. departures, adjust speeds, reroute aircraft, or provision additional communication resources. 2.1 Network Model Consider a single-cell 5G NR uplink system with 𝑁 UAVs communicating with a ground base station (gNB), with their 3

ACM MobiCom ’27, October 2027,

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

uplink traffic isolated from terrestrial-user traffic through network slicing. Time is slotted and indexed by 𝑡 ∈ {1, 2, . . .}. At each slot, the gNB scheduler allocates at most 𝐾 transmission opportunities to the UAV slice. Each UAV has two uplink information flows: a C2 status-update flow and a video-frame flow. Both flows follow a generate-at-will model: a fresh status update is generated whenever the UAV is scheduled for status transmission, and a fresh video frame is generated whenever the UAV is scheduled to start a video frame transmission.1 Each status update consists of a single packet and can be transmitted within one slot, whereas each video frame consists of 𝐿 packets and is considered successfully delivered only after all 𝐿 packets have been successfully received. We assume2 that the cellular infrastructure provides the gNB with estimates of each UAV’s position, the population density, and the staleness of its information flows. Packet Transmission. In each slot 𝑡, the scheduler decides which UAVs will attempt to transmit. Let 𝑎𝑢,𝑠 (𝑡) ∈ {0, 1} denote the status packet scheduling indicator, where 𝑎𝑢,𝑠 (𝑡) = 1 if the status packet of UAV 𝑢 is scheduled in slot 𝑡, and 𝑎𝑢,𝑠 (𝑡) = 0 otherwise. Similarly, let 𝑎𝑢,𝑣 (𝑡) ∈ {0, 1} denote the video packet scheduling indicator, where 𝑎𝑢,𝑣 (𝑡) = 1 if one video packet of UAV 𝑢 is scheduled in slot 𝑡, and 𝑎𝑢,𝑣 (𝑡) = 0 otherwise. The scheduling decisions satisfy 𝑁 ∑︁

𝑎𝑢,𝑠 (𝑡) + 𝑎𝑢,𝑣 (𝑡) ≤ 𝐾,

∀𝑡,

be transmitted per slot, and its packets need not be transmitted in consecutive slots. Once all 𝐿 packets have been successfully received, the frame is completed. Let ℓ𝑢 (𝑡) ∈ {1, . . . , 𝐿} denote the number of remaining packets of the current video frame of UAV 𝑢 at the beginning of slot 𝑡. Initially, ℓ𝑢 (1) = 𝐿 and each successful video packet transmission decreases ℓ𝑢 (𝑡) by one. After the last remaining packet is successfully delivered, ℓ𝑢 (𝑡) is reset to 𝐿 in the next slot. Accordingly, the evolution of ℓ𝑢 (𝑡) is given by   𝐿, ℓ𝑢 (𝑡) = 1, 𝜇𝑢,𝑣 (𝑡) = 1,    ℓ𝑢 (𝑡 + 1) = ℓ𝑢 (𝑡) − 1, ℓ𝑢 (𝑡) > 1, 𝜇𝑢,𝑣 (𝑡) = 1, (2)    ℓ𝑢 (𝑡), 𝜇 (𝑡) = 0. 𝑢,𝑣  Let 𝑔𝑢 (𝑡) denote the age of the current video frame at UAV 𝑢, measured at the beginning of slot 𝑡 as the elapsed time since the frame was generated. When no video-frame transmission is in progress, we set 𝑔𝑢 (𝑡) = 0. Initially, 𝑔𝑢 (1) = 0. The evolution of 𝑔𝑢 (𝑡) is given by   0, ℓ𝑢 (𝑡) = 1, 𝜇𝑢,𝑣 (𝑡) = 1,    𝑔𝑢 (𝑡 + 1) = 0, (3) ℓ𝑢 (𝑡) = 𝐿, 𝜇𝑢,𝑣 (𝑡) = 0,   𝑔𝑢 (𝑡) + 1, otherwise.  Under the generate-at-will model, when a UAV is scheduled for video transmission with ℓ𝑢 (𝑡) = 𝐿, it generates a fresh frame at the beginning of the slot. If the first packet is not successfully delivered, the frame is discarded, and another fresh frame is generated at the next video scheduling opportunity. Once the first packet is successfully delivered, the same frame is retained until all its remaining packets are delivered. The video frame delivery indicator is defined as 𝑑𝑢,𝑣 (𝑡) = 1{ℓ𝑢 (𝑡) = 1, 𝜇𝑢,𝑣 (𝑡) = 1}, where 𝑑𝑢,𝑣 (𝑡) = 1 if the current video frame is successfully delivered in slot 𝑡, and 𝑑𝑢,𝑣 (𝑡) = 0 otherwise. AoI Evolution. Let ℎ𝑢,𝑠 (𝑡) denote the AoI at the gNB of the most recently successfully received C2 status update from UAV 𝑢 at the beginning of slot 𝑡. We refer to ℎ𝑢,𝑠 (𝑡) as the status AoI. The evolution of ℎ𝑢,𝑠 (𝑡) is given by ( 1, 𝑑𝑢,𝑠 (𝑡) = 1, ℎ𝑢,𝑠 (𝑡 + 1) = (4) ℎ𝑢,𝑠 (𝑡) + 1, otherwise.

(1)

𝑢=1

meaning that at most 𝐾 transmission opportunities can be scheduled in each slot. A scheduled status packet is successfully delivered with probability 𝑝𝑢,𝑠 . Let 𝑑𝑢,𝑠 (𝑡) ∈ {0, 1} denote the successful status packet transmission indicator, where 𝑑𝑢,𝑠 (𝑡) = 1 if the scheduled status packet is successfully delivered in slot 𝑡, and 𝑑𝑢,𝑠 (𝑡) = 0 otherwise. Then E[𝑑𝑢,𝑠 (𝑡) | 𝑎𝑢,𝑠 (𝑡)] = 𝑝𝑢,𝑠 𝑎𝑢,𝑠 (𝑡). Similarly, a scheduled video packet is successfully delivered with probability 𝑝𝑢,𝑣 . Let 𝜇𝑢,𝑣 (𝑡) ∈ {0, 1} denote the successful video packet transmission indicator, where 𝜇𝑢,𝑣 (𝑡) = 1 if the scheduled video packet is successfully delivered in slot 𝑡, and 𝜇𝑢,𝑣 (𝑡) = 0 otherwise. Then E[𝜇𝑢,𝑣 (𝑡) | 𝑎𝑢,𝑣 (𝑡)] = 𝑝𝑢,𝑣 𝑎𝑢,𝑣 (𝑡). Video Transmission Process. Each video frame consists of 𝐿 packets. At most one packet of the current frame can

Let ℎ𝑢,𝑣 (𝑡) denote the AoI at the gNB of the most recently completed video frame from UAV 𝑢 at the beginning of slot 𝑡. We refer to ℎ𝑢,𝑣 (𝑡) as the video AoI. Upon a successful video frame delivery in slot 𝑡, the video AoI is reset to the age of the delivered frame at the beginning of slot 𝑡 + 1, namely 𝑔𝑢 (𝑡) + 1. Therefore, ( 𝑔𝑢 (𝑡) + 1, 𝑑𝑢,𝑣 (𝑡) = 1, ℎ𝑢,𝑣 (𝑡 + 1) = (5) ℎ𝑢,𝑣 (𝑡) + 1, otherwise.

1 Although status updates and video frames are typically generated periodi-

cally in practice, the adopted models approximate high-rate status sensing and continuous video streaming. They also allow us to focus on the effects of scheduling and transmission reliability without introducing additional sampling parameters. 2 UAV positions can be obtained through network-assisted or onboard localization [2]. Population exposure can be estimated from geographic population data, while staleness can be tracked using packet timestamps and delivery records [51].

We assume that ℎ𝑢,𝑠 (1) = ℎ𝑢,𝑣 (1) = 1 for all 𝑢 ∈ {1, . . . , 𝑁 }. Scheduling Policy. A scheduling policy determines the 4

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027, (b)

(a)

𝑁 in each slot. We consider nonactions {𝑎𝑢,𝑠 (𝑡), 𝑎𝑢,𝑣 (𝑡)}𝑢=1 anticipative policies that do not use future channel outcomes or future system states when making scheduling decisions. Let Π denote the class of admissible non-anticipative scheduling policies satisfying the resource constraint (1).

Most recent UAV u

UAV v

x̂ u(t)

duv(t)

(6)

…

h(w)(t) αw

Weighted coherence-window staleness rv(t)

zu(t)

Population density High

0

Hnom

Higher risk

Using the uncertainty radii defined in (6), we define the normalized navigation risk of UAV 𝑢 as

Its area is

𝑅𝑢Nav (𝑡) = (8)

𝑁 1 ∑︁

𝑅¯Nav 𝑣=1

[𝑟𝑢 (𝑡) + 𝑟 𝑣 (𝑡) − 𝑑𝑢𝑣 (𝑡)] + ,

(9)

𝑣≠𝑢

where [𝑥] + ≜ max{𝑥, 0}. The navigation risk is zero when the reported separation exceeds the combined position uncertainty and increases with the corresponding separation deficit. Ground Risk. The ground risk induced by C2 status update staleness depends on the population within the UAV’s uncertainty region. Using the uncertainty region B𝑢 (𝑡) defined in (7), we define the normalized ground risk as ∫ 1 𝜌 (𝒙, 𝑡) d𝒙, (10) 𝑅𝑢Ground (𝑡) = Ground 𝑅¯ B𝑢 (𝑡 )

Information staleness affects operational risk by enlarging the region in which the aircraft may be located; the resulting consequences depend on neighboring obstacles, aircraft, ground population, and other hazards.

2.3

h(2)(t) α2

Figure 2: The same staleness can induce different operational consequences. (A) Navigation risk increases as position uncertainty consumes the separation margin between neighboring UAVs. (B) Ground risk depends on the population exposed within the UAV’s uncertainty region, so the same AoI can imply different risk across locations. (C) Video-coherence risk captures the weighted staleness of the most recent completed frames and increases once it exceeds the nominal coherence threshold. The three components are normalized at their corresponding operational boundaries.

where 𝑟 min represents the minimum localization uncertainty. The corresponding staleness-induced uncertainty region is the disk centered at the most recently reported position:  B𝑢 (𝑡) = 𝒙 ∈ R2 : ∥𝒙 − 𝒙ˆ𝑢 (𝑡)∥ ≤ 𝑟𝑢 (𝑡) . (7)

 2 |B𝑢 (𝑡)| = 𝜋𝑟𝑢2 (𝑡) = 𝜋 𝑟 min + 𝑣 maxℎ𝑢,𝑠 (𝑡) .

h(1)(t) α1

x̂ v(t)

Staleness-Induced Uncertainty

𝑟𝑢 (𝑡) = 𝑟 min + 𝑣 maxℎ𝑢,𝑠 (𝑡),

Bu(t)

Low

Let ℎ𝑢,𝑠 (𝑡), defined in (4), denote the AoI of the most recently received C2 status update from UAV 𝑢 at the beginning of slot 𝑡. Let 𝒙ˆ𝑢 (𝑡) ∈ R2 denote the position contained in that update. During the interval, the aircraft may have moved away from its last reported position 𝒙ˆ𝑢 (𝑡). Following the containment principle used by SORA, we assume that a UAV can travel by at most 𝑣 max per slot. We define its worst-case position uncertainty radius3 as

Older

…

Bv(t)

Bu(t)

ru(t)

2.2

(c)

Operational Risk Components

We capture three operational consequences of stale C3 information: reduced aircraft-separation margins, increased population exposure, and degraded video coherence. Since these components measure heterogeneous quantities, we normalize them using the reference values 𝑅¯Nav , 𝑅¯Ground , and 𝑅¯Vid , which denote the navigation, ground, and video-coherence risk levels at their corresponding declared operational boundaries. Their definitions are provided in Appendix A. Navigation Risk. As C2 status updates become stale, the positional uncertainties of neighboring UAVs consume their separation margin. Let 𝑑𝑢𝑣 (𝑡) = ∥ 𝒙ˆ𝑢 (𝑡) − 𝒙ˆ 𝑣 (𝑡) ∥ denote the distance between the positions contained in the most recently received C2 status updates from UAVs 𝑢 and 𝑣.

where 𝜌 (𝒙, 𝑡) is the population-density at location 𝒙 in slot 𝑡. Thus, even for the same C2 status update staleness, the ground risk can differ depending on the population distribution around the UAV’s most recently reported position. Video-Coherence Risk. Remote piloting depends on a sequence of recent frames rather than on the latest completed (𝑖 ) frame alone. Let ℎ𝑢,𝑣 (𝑡) denote the AoI of the 𝑖-th most re(1) cently completed video frame, where ℎ𝑢,𝑣 (𝑡) = ℎ𝑢,𝑣 (𝑡), and ℎ𝑢,𝑣 (𝑡) is the video AoI defined in (5). We aggregate the AoI of the 𝑤 most recently completed frames into the coherencewindow staleness

3 The circular region captures worst-case position uncertainty when no

directional information is available. The framework also supports other uncertainty models by replacing the disk with an appropriate region. For example, a trajectory-aligned ellipse can represent uncertainty along a preplanned path, with population exposure evaluated over the resulting region.

𝑧𝑢 (𝑡) =

𝑤 ∑︁ 𝑖=1

5

(𝑖 ) 𝛼𝑖 ℎ𝑢,𝑣 (𝑡),

𝑤 ∑︁ 𝑖=1

𝛼𝑖 = 1,

𝛼𝑖 ≥ 0,

(11)

ACM MobiCom ’27, October 2027,

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

Proposition 2.1. Under (15) and (14), the per-UAV composite operational risk of UAV 𝑢 satisfies

where 𝛼𝑖 specifies the contribution of the 𝑖-th most recent frame. The normalized video-coherence risk is  + 𝑧𝑢 (𝑡) − 𝐻 nom , (12) 𝑅𝑢Vid (𝑡) = 𝑅¯Vid where 𝐻 nom denotes the nominal coherence-window staleness. Thus, 𝑅𝑢Vid (𝑡) = 0 when 𝑧𝑢 (𝑡) ≤ 𝐻 nom and increases linearly with 𝑧𝑢 (𝑡) otherwise.

2.4

(𝑤 ) 𝑅𝑢 (𝑡) ≤ 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡) + 𝐶𝑢 ,

where 𝑁  ∑︁  + 𝑤𝑁 max max
𝛼𝑢,𝑠 = Nav max 2𝑟 min + 𝑣 max 𝜏𝑢,𝑠 + 𝜏𝑣,𝑠 − 𝑑𝑢𝑣 𝑅¯ 𝜏𝑢,𝑠 𝑣=1

Composite Operational Risk

𝑣≠𝑢

 + max − 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 𝑤𝐺 𝜌𝑢,max 𝜋 2 max
2𝑟 min𝑣 max + 𝑣 max 𝜏𝑢,𝑠 + , 𝑅¯Ground 𝑤𝑉 𝛼𝑢,𝑣 = Vid , 𝑅¯

Using the normalized risk components defined in (9), (10), and (12), the per-UAV composite operational risk is 𝑅𝑢 (𝑡) = 𝑤 𝑁 𝑅𝑢Nav (𝑡) + 𝑤𝐺 𝑅𝑢Ground (𝑡) + 𝑤𝑉 𝑅𝑢Vid (𝑡),

(13)

where 𝑤 𝑁 , 𝑤𝐺 , 𝑤𝑉 > 0 are weighting coefficients satisfying 𝑤 𝑁 + 𝑤𝐺 + 𝑤𝑉 = 1. The level 𝑅𝑢 (𝑡) = 1 marks the declared composite operational boundary; values above it indicate that communicationinduced staleness has exhausted the composite risk budget. SORA provides the framework for operational risk assessment and authorization, while this metric provides a per-slot risk proxy that allows the scheduler to compare the operational consequences of deferring different C3 streams.

2.5

𝐶𝑢 =

∀𝑢, 𝑡, 𝒙 ∈ B𝑢 (𝑡),

𝑁 + 𝑤 𝑁 ∑︁  max 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 Nav ¯ 𝑅 𝑣=1

+

2 𝑤𝐺 𝜌𝑢,max 𝜋𝑟 min . 𝑅¯Ground

(19)

A detailed proof of Proposition 2.1 is provided in Appendix B. (𝑤 ) The upper bound in (16) depends on the AoI ℎ𝑢,𝑣 (𝑡) of the oldest frame in the coherence window. Tracking this state requires maintaining the completion history of the previous 𝑤 video frames. However, the scheduler makes a scheduling decision for every frame, and completing the current frame (1) directly updates ℎ𝑢,𝑣 (𝑡) = ℎ𝑢,𝑣 (𝑡), the AoI of the most recently completed frame. We therefore use the latest-frame AoI as the video state in the scheduling objective. Motivated by the coefficients of the risk upper bound, we define the following tractable per-UAV surrogate operational risk: 𝑅b𝑢 (𝑡) = 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡). (20)

(14)

max captures the maximum tolerable status staleness where 𝜏𝑢,𝑠 max > 0 should trigger of UAV 𝑢, meaning that ℎ𝑢,𝑠 (𝑡) − 𝜏𝑢,𝑠 an emergency response. The corresponding maximum unmax . Let certainty radius of UAV 𝑢 is 𝑟𝑢max = 𝑟 min + 𝑣 max𝜏𝑢,𝑠 B𝑢max denote the uncertainty region centered at the most recently reported position of UAV 𝑢 with radius 𝑟𝑢max . Under (14), the uncertainty region of UAV 𝑢 satisfies B𝑢 (𝑡) ⊆ B𝑢max . We define the maximum population density within B𝑢max as 𝜌𝑢,max = max𝒙 ∈ B𝑢max 𝜌 (𝒙). It follows that

0 ≤ 𝜌 (𝒙, 𝑡) ≤ 𝜌𝑢,max ≤ 𝜌 max,

(18)

𝑣≠𝑢

Directly optimizing the per-UAV composite operational risk in (13) is challenging because its components are nonlinear in the AoI states and coupled through the system dynamics. We therefore derive a linear upper bound on the per-UAV composite operational risk. For each UAV 𝑢, the C2 status-update AoI satisfies ∀𝑢, 𝑡,

(17)

and

Surrogate Operational Risk

max 0 ≤ ℎ𝑢,𝑠 (𝑡) ≤ 𝜏𝑢,𝑠 ,

(16)

The per-UAV surrogate operational risk in (20) retains the relative status and video coefficients induced by the original operational risk model while depending only on the AoI states used in the scheduling analysis. It is employed as a tractable scheduling objective. Context timescale. The operational context underlying 𝛼𝑢,𝑠 and 𝛼𝑢,𝑣 , including UAV location and population exposure, generally evolves more slowly than the per-slot AoI and transmission states. We therefore analyze each fixed-context realization separately: the risk coefficients are held constant while the scheduler responds in every slot to changes in information freshness and video-transmission progress. Section 4 evaluates this two-timescale operation along moving UAV trajectories and dynamic parameters.

(15)

where 𝜌 max is the maximum population density within the declared operating region. Thus, 𝜌𝑢,max captures the local population context associated with UAV 𝑢. For each pair of UAVs 𝑢 and 𝑣, let 𝑑𝑢𝑣 denote the separation between their most recently reported positions under the considered operational context. The following proposition provides a linear upper bound on the per-UAV composite operational risk within the declared operating region. 6

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

2.6

ACM MobiCom ’27, October 2027,

3.1

Optimization Problem

Lower Bound

We use the per-UAV surrogate operational risk in (20) as the scheduling objective. The long-term expected surrogate risk under policy 𝜋 ∈ Π is defined as

To derive a universal lower bound on (22), we characterize the minimum achievable surrogate risk over all feasible longterm UAV packet delivery rates.

𝑇 𝑁 i 1 ∑︁ ∑︁ 𝜋 h b b 𝐽 𝜋 = lim sup E 𝑅𝑢 (𝑡) . 𝑇 →∞ 𝑇 𝑡 =1 𝑢=1

Theorem 3.1. Assume that 0 < 𝑝𝑢,𝑠 ≤ 1 and 0 < 𝑝𝑢,𝑣 ≤ 1 for every UAV 𝑢. For any network setup (𝑁 , 𝐾, 𝐿, 𝑝𝑢,𝑠 , 𝑝𝑢,𝑣 , 𝛼𝑢,𝑠 , 𝛼𝑢,𝑣 ), the long-term expected surrogate risk under any admissible policy 𝜋 ∈ Π satisfies b 𝐽 𝜋 ≥ 𝐿𝐵★, where  𝑁  ∑︁ 𝛼𝑢,𝑠 𝛼𝑢,𝑣 ★ 𝐿𝐵 = min + 2𝑞𝑢,𝑠 2𝑞𝑢,𝑣 {𝑞𝑢,𝑠 ,𝑞𝑢,𝑣 } 𝑢=1

(21)

Our goal is to design a low-complexity scheduling policy 𝜋 that attempts to minimize b 𝐽 𝜋 subject to network resource constraints. ( ) 𝑇 𝑁 i 1 ∑︁ ∑︁ 𝜋 h b ∗ b (22a) 𝐽 = min lim sup E 𝑅𝑢 (𝑡) 𝜋 ∈Π 𝑇 →∞ 𝑇 𝑡 =1 𝑢=1 s.t.

𝑁 ∑︁

𝑎𝑢,𝑠 (𝑡) + 𝑎𝑢,𝑣 (𝑡) ≤ 𝐾,

∀𝑡 .

𝑁

1 ∑︁ 𝛼𝑢,𝑠 + (2𝐿 − 1)𝛼𝑢,𝑣 (23a) 2 𝑢=1  𝑁  ∑︁ 𝑞𝑢,𝑠 𝐿𝑞𝑢,𝑣 + ≤ 𝐾, (23b) s.t. 𝑝𝑢,𝑠 𝑝𝑢,𝑣 𝑢=1 𝑝𝑢,𝑣 , ∀𝑢 (23c) 0 < 𝑞𝑢,𝑠 ≤ 𝑝𝑢,𝑠 and 0 < 𝑞𝑢,𝑣 ≤ 𝐿 Here, 𝑞𝑢,𝑠 and 𝑞𝑢,𝑣 denote the long-term status packet delivery rate and video-frame completion rate of UAV 𝑢, respectively. +

(22b)

𝑢=1

The alignment of the proposed surrogate risk (20), the resulting scheduling policy (22), and SORA’s operational criticality [20] will be illustrated in Fig. 5, where we show that our scheduling policy consistently serves UAVs with greater criticality. For a given system setup (𝑁 , 𝐾, 𝐿, 𝑝𝑢,𝑠 , 𝑝𝑢,𝑣 , 𝛼𝑢,𝑠 , 𝛼𝑢,𝑣 ), let ∗ b 𝐽 denote the minimum long-term expected surrogate risk. Similarly, for any admissible scheduling policy 𝜂 ∈ Π, let b 𝐽𝜂 denote the corresponding expected surrogate risk achieved by policy 𝜂. We define the optimality ratio of policy 𝜂 as 𝜓𝜂 = b 𝐽 𝜂 /b 𝐽 ∗ . A policy 𝜂 is said to be 𝜓 𝜂 -optimal if its longterm expected surrogate risk is within a multiplicative factor 𝜓 𝜂 ≥ 1 of the optimum. The optimality ratio provides a quantitative measure of the performance gap between a lowcomplexity scheduling policy and the optimal policy.

3

Proof Sketch. Using a sample path argument, we characterize the expected surrogate risk in terms of the statuspacket delivery and video-frame completion rates. Combining the resulting risk expression with the resource constraint (23b) yields the optimization problem in (23), whose optimal value gives the lower bound. Detailed derivations are provided in Appendix C. □

3.2

Risk-Aware 5G Scheduling Policy

We develop a Max-Weight scheduling policy that is designed to reduce the surrogate risk by minimizing the expected oneslot drift of a Lyapunov function at every time slot 𝑡. A main challenge is the stochastic delay between action (i.e., packet scheduling) and reward (i.e., risk reduction), which makes the Lyapunov functions from prior works, including [22– 24], unsuitable for our setting. To overcome this challenge, the authors in [51] proposed a complex Lyapunov function that includes not only AoI but also a throughput debt with respect to a chosen throughput target. In contrast, we propose a simpler, thoughtfully designed risk-aware Lyapunov function that relies only on AoI and improves the resulting Max-Weight optimality ratio by a factor of 1.5× with respect to [51]. In § 4, we show that while [51] minimizes AoI, our risk-aware scheduling policy achieves up to 83% lower 99thpercentile per-UAV composite operational risk, and lower risk than [51] at every evaluated operating point. Lyapunov Function. Denote the system state at the be𝑁 . ginning of slot 𝑡 as 𝑆 (𝑡) = {ℎ𝑢,𝑠 (𝑡), ℎ𝑢,𝑣 (𝑡), 𝑔𝑢 (𝑡), ℓ𝑢 (𝑡)}𝑢=1 For the video-frame flow, define the video risk potential as

Proposed 5G Scheduling Policy

In this section, we develop a low-complexity scheduling policy for minimizing the long-term expected surrogate risk (22) and obtain its optimality ratio under fixed operational context and risk coefficients. In § 5, we will briefly discuss related optimization problems, computational-heavy optimal solutions, and low-complexity sub-optimal solutions. Two key contributions of this paper are: (i) applying the rich AoI theory to a pressing, practical problem; (ii) proposing a carefully designed Lyapunov function that improves the optimality ratio by a factor of at least 1.5× with respect to [51]. Next, we derive a universal lower bound 𝐿𝐵★ ≤ b 𝐽 ∗ on the optimal longterm expected surrogate risk using a sample-path argument. Then, we develop a Risk-Aware Max-Weight scheduling policy based on a Lyapunov function. Finally, we use Lyapunov optimization to establish an upper bound on its optimality ratio 𝜓 𝜂 ≤ b 𝐽 𝜂 /𝐿𝐵★ . 7

ACM MobiCom ’27, October 2027,

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

follows

policy by comparing it with the lower bound derived in Theorem 3.1.

ℓ𝑢 (𝑡) Φ𝑢,𝑣 (𝑡) = (ℎ𝑢,𝑣 (𝑡) − 1) + 𝑔𝑢 (𝑡) + 1. (24) 𝐿 This video risk potential grows with video AoI ℎ𝑢,𝑣 (𝑡), UAV frame age 𝑔𝑢 (𝑡), and the number of remaining packets ℓ𝑢 (𝑡). Notably, the expected video risk potential reduces with a deterministic one-slot delay following a video packet transmission. We define the Lyapunov function as 𝑁 ∑︁   𝐿(𝑆 (𝑡)) = 𝑉𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝑉𝑢,𝑣 Φ𝑢,𝑣 (𝑡) ,

LB , 𝑞 LB } 𝑁 be an optimal solution of Theorem 3.2. Let {𝑞𝑢,𝑠 𝑢,𝑣 𝑢=1 (23), and choose the Lyapunov coefficients as 𝛼𝑢,𝑣 𝛼𝑢,𝑠 𝑉𝑢,𝑣 = LB . (32) 𝑉𝑢,𝑠 = LB , 𝑞𝑢,𝑠 𝑞𝑢,𝑣

Then the proposed Max-Weight policy satisfies b 𝐽 𝑀𝑊 ≤ 𝐿𝐵★ +

(25)

𝑁 ∑︁ 𝛼𝑢,𝑠

𝑁

+ LB

𝑢=1 2𝑞𝑢,𝑠

3 ∑︁ 𝛼𝑢,𝑣 , LB 2 𝑢=1 𝑞𝑢,𝑣

(33)

𝑢=1

where 𝐿𝐵★ denotes the optimal value of (23). Consequently,

where 𝑉𝑢,𝑠 > 0 and 𝑉𝑢,𝑣 > 0 are design parameters. One-slot Lyapunov Drift. The one-slot conditional Lyapunov drift is given by Δ(𝑆 (𝑡)) = E [𝐿(𝑆 (𝑡 + 1)) − 𝐿(𝑆 (𝑡)) | 𝑆 (𝑡)] .

𝑁 ∑︁ 𝛼𝑢,𝑠

𝜓 𝑀𝑊 =

(26)

By substituting the evolution of ℓ𝑢 (𝑡), 𝑔𝑢 (𝑡), ℎ𝑢,𝑠 (𝑡), and ℎ𝑢,𝑣 (𝑡) from (2), (3), (4), and (5), respectively, into the drift expression in (26), we obtain the upper bound on Δ(𝑆 (𝑡)). The resulting bound is expressed in (27)-(30), with detailed derivations provided in Appendix D Δ(𝑆 (𝑡)) ≤ 𝐵(𝑡) −

𝑁 ∑︁

−

  𝑊𝑢,𝑠 (𝑡)E 𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)   𝑊𝑢,𝑣 (𝑡)E 𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡) .

(27)

𝑢=1

The scheduling independent term is   𝑁 𝑁 ∑︁ ∑︁ ℓ𝑢 (𝑡) . 𝐵(𝑡) = 𝑉𝑢,𝑠 + 𝑉𝑢,𝑣 1 + 𝐿 𝑢=1 𝑢=1

𝑊𝑢,𝑠 (𝑡) = 𝑉𝑢,𝑠 𝑝𝑢,𝑠 ℎ𝑢,𝑠 (𝑡), (29) 𝑉𝑢,𝑣 𝑝𝑢,𝑣 ℎ𝑢,𝑣 (𝑡). (30) 𝑊𝑢,𝑣 (𝑡) = 𝐿 Max-Weight Policy. According to (27), minimizing the upper bound on the Lyapunov drift is equivalent to maximizing the total scheduling weight. Hence, in every slot 𝑡, the MaxWeight policy solves max {𝑎𝑢,𝑠 (𝑡 ),𝑎𝑢,𝑣 (𝑡 ) }

s.t.

𝑢=1 𝑁 ∑︁

𝑎𝑢,𝑠 (𝑡)𝑊𝑢,𝑠 (𝑡) +

𝑁 ∑︁

b 𝐽 𝑀𝑊 𝑢=1 2𝑞𝑢,𝑠 ≤1+ b 𝐿𝐵★ 𝐽∗

.

(34)

The performance guarantee applies to the fixed-context surrogate objective b 𝐽 , not directly to the composite operationalrisk metric 𝑅𝑢 (𝑡). Section 4 evaluates their empirical alignment. Theorem 3.2 provides a parameterized performance guarantee that explicitly depends on the optimal value of the lower-bound problem. By using the inequality  𝑁  1 ∑︁ 𝛼𝑢,𝑠 𝛼𝑢,𝑣 𝐿𝐵★ ≥ + , (35) LB LB 2 𝑢=1 𝑞𝑢,𝑠 𝑞𝑢,𝑣

(28)

The status and video scheduling weights are

𝑁 ∑︁

3 ∑︁ 𝛼𝑢,𝑣 LB 2 𝑢=1 𝑞𝑢,𝑣

Proof. First, we compare the Max-Weight policy with a randomized policy whose scheduling rates are chosen according to the optimal solution of the lower-bound problem (23). By the Max-Weight property, the one-slot Lyapunov drift under Max-Weight is no larger than the drift under this comparison policy. Then, choosing the Lyapunov coefficients as in (32) makes the negative drift terms coincide with the surrogate risk. Summing the resulting drift inequality over time and applying telescoping gives (33). Combining this upper bound with 𝐿𝐵★ ≤ b 𝐽 ∗ yields (34). Detailed derivations are provided in Appendix F. □

𝑢=1 𝑁 ∑︁

𝑁

+ LB

the ratio in (34) can be further bounded as 𝑁 ∑︁ 𝛼𝑢,𝑠

𝑁

3 ∑︁ 𝛼𝑢,𝑣 LB 2 𝑢=1 𝑞𝑢,𝑣 𝑢=1 2𝑞𝑢,𝑠 𝜓 𝑀𝑊 ≤ 1 +  ≤ 4. 𝑁  1 ∑︁ 𝛼𝑢,𝑠 𝛼𝑢,𝑣 + LB LB 2 𝑢=1 𝑞𝑢,𝑠 𝑞𝑢,𝑣 + LB

𝑎𝑢,𝑣 (𝑡)𝑊𝑢,𝑣 (𝑡)

𝑢=1

(31)
𝑎𝑢,𝑠 (𝑡) + 𝑎𝑢,𝑣 (𝑡) ≤ 𝐾,

(36)

𝑢=1

𝑎𝑢,𝑠 (𝑡), 𝑎𝑢,𝑣 (𝑡) ∈ {0, 1},

Therefore, the proposed Max-Weight policy admits a worstcase optimality ratio no larger than 4. Since this bound is obtained by further relaxing the denominator of (34), it is conservative, and the parameterized bound in Theorem 3.2 provides a tighter characterization whenever the value of 𝐿𝐵★ is available.

∀𝑢.

Equivalently, the scheduler selects the 𝐾 streams with the largest scheduling weights. Optimality Ratio for Max Weight Policy. We now establish a performance guarantee for the proposed Max-Weight 8

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations 2000

(a)

1500

30

1000

500

40

River

40

0 0

1000

2000

3000

4000

5000

6000

East [m] (b)

20 30

10

10

0 10

East [m]

20

500

30

SNR [dB]

Pinhe iros

40

4.1

Shopping Iguatemi

Evento

Alphaville 20

We evaluate RA against two conventional 5G schedulers, namely Round-Robin (RR) and proportional fair (PF), and against the AoI-minimizing policy developed in [51] that shares RA’s Max-Weight structure and update model but does not use operational context. RA uses the Max-Weight policy of § 3 with Lyapunov coefficients 𝑉𝑢,𝑠 and 𝑉𝑢,𝑣 from (32) and surrogate-risk coefficients 𝛼𝑢,𝑠 and 𝛼𝑢,𝑣 from (17) and (18), respectively, that are computed based on the operational context, e.g., UAV location and population exposure. As discussed in § 1, RA obtains operational context from 5G core functions, such as the LMF and NWDAF, and from external sources, such as geographic databases and UAV reports. In general, operational context and the resulting Lyapunov coefficients 𝑉𝑢,𝑠 and 𝑉𝑢,𝑣 evolve on a slower timescale (seconds/minutes) than RA’s status and video scheduling weights 𝑊𝑢,𝑠 (𝑡) and 𝑊𝑢,𝑣 (𝑡) from (29) and (30), respectively, that may change in every time slot (≤ 1 ms). The analysis in Secs. 2 and 3 considers a fixed operational context to establish the lower bound and performance guarantee for each context realization. In the evaluation, we allow the operational context to evolve along each UAV trajectory and assume that the gNB observes accurate and timely context information which is used to (re)compute RA’s scheduling weights 𝑊𝑢,𝑠 (𝑡) and 𝑊𝑢,𝑣 (𝑡) in every time slot. The evaluation attempts to answer three questions: Q1: Does RA reduce risk exposure? In § 4.2, we show that RA reduces 99th-percentile per-UAV risk 𝑅𝑢 (𝑡) by up to 96% over PF, 86% over RR, and 83% over AoI. Q2: Does RA prioritize operationally critical UAVs? In § 4.3, we label operations offline with SORA’s SAIL and show that, under contention,RA schedules a median 2.4× more service per critical UAV than per regular UAV, while RR, PF, and AoI remain class-blind. Q3: How much operational capacity does RA enable? In § 4.4, we show that, 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.

10

Evaluation

North [m]

4

ACM MobiCom ’27, October 2027,

0

−500 1000

500

0

−500

−1000

North [m] Circuit

Route A

Start

gNB

Pinheiros River

Route B

End

SNR = 5 dB

Pop. density

Figure 3: Evaluation scenarios in São Paulo. Colors show the gNB-centered iso-SNR contours, the dashed contour marks the 5-dB outage boundary. Shaded regions represent population exposure. (a) Delivery corridors near Shopping Iguatemi. (b) Interlagos Formula 1. Network, traffic, and simulator. We implement a slot-level 5G NR uplink simulator in MATLAB to integrate the domains that determine communication-induced operational risk. On a common slot-level timeline, the simulator synchronizes UAV motion, position-dependent radio conditions, status and video traffic, packet delivery, information staleness, geographic population exposure, and scheduling decisions. UAV mobility is cross-checked against UAV Toolbox [43], while radio propagation is driven by 5G Toolbox [42] models. This integration allows the operational consequences of each network decision to evolve with both the aircraft trajectory and the communication state. It also enables controlled comparisons in which all policies experience identical mobility, channel, traffic, and transmission-outcome realizations. Each slot lasts 1 ms, and the gNB provides at most 𝐾 ∈ {1, 2, 4, 8, 16, 32} transmission opportunities to 𝑁 ∈ {5, 10, 20, 40, 60, 80, 100, 120} concurrent UAVs, yielding 48 operating points. For each of the four scheduling policies, we simulate every operating point for at least an 900 s interval, during which all 𝑁 UAVs simultaneously contend for the uplink channel. To ensure that different scheduling policies experience the same trajectories, channel states, and transmission outcomes, we use the same seed value. UAVs fly at 15 m/s from the start to the end of the corridors depicted in Fig. 3. Each UAV carries the C2 status update and video streams. Status updates are generated with a frequency of 10 Hz and require a single successful packet transmission to be delivered to the gNB. Video frames are generated according to the realistic UAV traffic generator

Evaluation Setup

Scenarios. We consider two scenarios derived from real operations in São Paulo. Figure 3 overlays UAV trajectories with iso-SNR contours and ground-exposure densities used in the evaluation. The delivery scenario in Fig. 3(a) contains two UAV corridors with different combinations of coverage and population exposure. The event scenario in Fig. 3(b) follows the 4.3-km Interlagos circuit, where spectator hotspots create strongly nonuniform exposure along the circuit. Appendix G documents the public infrastructure and operational data used to ground these scenarios. 9

ACM MobiCom ’27, October 2027,

André Melo, Yueqi Chen, Lourenço Pereira Jr., and Igor Kadota Light load (ρ = 3.75)

from [8] with 30 fps and a 15-frame Group of Pictures (GOP) structure with one 30-packet I-frame followed by fourteen 3-packet P-frames. It follows that refreshing video may consume multiple transmission opportunities. While the number of packets per video frame changes over time, both RA and AoI use a fixed 𝐿 = 10 in the experiments below. UAV dynamics are simulated using the UAV toolbox [43]. 5G wireless channels are simulated using the 5G toolbox [42] at 3.5 GHz with a 20-MHz bandwidth and an urban-macro channel model. UAV links with SNR below 5 dB are in outage, i.e., 𝑝𝑢,𝑠 = 𝑝𝑢,𝑣 = 0. Otherwise, scheduled packets succeed with probability 𝑝𝑢,𝑠 = 𝑝𝑢,𝑣 = 0.8. Additional toolbox parameters are provided in Appendix L. Simulator fidelity. OAI provides the protocol-stack fidelity needed to demonstrate the scheduling blind spot and verify the resource-allocation behavior central to this study. MATLAB provides the cross-domain integration, experimental control, and scale needed to evaluate how that behavior affects operational risk across moving aircraft, geographic exposure conditions, traffic types, fleet sizes, and resource budgets. Because the complete 𝐾 × 𝑁 sweep is impractical to reproduce with the full 5G stack, we validate the MATLAB abstraction against the OAI implementation used in Fig. 1 under matched saturated-load experiments using proportional-fair scheduling. Specifically, we test symmetric channels, in which all UEs see the same radio conditions, with 𝑁 ∈ {2, 3, 4}, and asymmetric channels with 𝑁 ∈ {2, 4}, in which half of the UEs are attenuated by about 8 dB (MCS 20 versus 28). Under symmetric channels, OAI and the simulator both give each UE the expected 1/𝑁 share of uplink grants. Under asymmetric channels, PF equalizes time rather than bytes: degraded UEs deliver fewer bytes, but receive nearly the same grant share in both systems. Across all configurations, the maximum totalvariation distance between their grant-share distributions is 0.021, about two percentage points. Thus, the simulator preserves the uplink allocation behavior that RA modifies.

4.2

20

Moderate load (ρ = 5) 30 Overload (ρ = 7.5)

Mean AoI

AoI 20

tte

Be

AoI 25

r

tte

Be

tte

PF

30 RA PF 25 0.5

0.7

AoI

r

40

RR

Be

RA

RA

RR

r

PF

RR

50 1

0.5

0.7

p99 [R̄u ]

1

1.5

1

1.5

2

3

¯ and Figure 4: Each panel reports the fleet-mean AoI, ℎ, the 99th percentile of time-average per-UAV operational risk, 𝑝 99,𝑢 [𝑅¯𝑢 ], aggregated across both evaluation scenarios for operating points in the indicated load regime, 𝜌 = 𝑁 /𝐾. Lower values are preferable on both axes. The dashed line denotes the communicationinduced operational-risk envelope. A complete evaluation across all 48 operating points can be found in Fig. 8 of Appendix H. aggregating the operating points from both evaluation scenarios. The dashed line marks the communication-induced operational-risk envelope 𝑝 99,𝑢 [𝑅¯𝑢 ] = 1. Figure 4 shows that freshness and risk induce different policy rankings. The AoI policy achieves the lowest fleet-mean age in every load regime, yet it does not achieve the lowest tail risk. Even under light load, its risk exposure approaches the envelope. The reason is that AoI does not account for the operational consequence of staleness. It may therefore refresh an UAV whose information is old but operationally tolerable while deferring one whose growing uncertainty intersects traffic or population exposure. PF improves upon RR in both freshness and risk, showing that uniform service is inefficient. RA changes the allocation objective more fundamentally: it directs service toward the UAVs for which additional delay consumes the operational-risk margin most rapidly. This distinction becomes most consequential under moderate load and overload, where resources must be withheld from some UAVs.Across the complete 48-point evaluation in Fig. 8, RA attains the lowest 𝑝 99,𝑢 [𝑅¯𝑢 ] at every operating point, with reductions of up to 96% relative to PF, 86% relative to RR, and 83% relative to AoI. Takeaway. Minimizing average AoI is not sufficient for risksensitive UAV communication because equal staleness can have unequal operational consequences. Under contention, the relevant scheduling decision is not simply which information is oldest, but where an additional unit of delay consumes the operational-risk margin fastest. RA exploits this distinction by accepting a controlled increase in average AoI while substantially reducing tail risk.

Does RA Reduce Risk Exposure?

We first examine whether minimizing information age also minimizes the operational consequences of stale information. We compare the policies using two metrics: (i) 99thpercentile risk exposure (𝑝 99,𝑢 [𝑅¯𝑢 ]), where 𝑅¯𝑢 captures the time-average of the operational risk 𝑅𝑢 (𝑡) for a particular ¯ where UAV; and (ii) Fleet-Mean Age of Information (ℎ), ℎ¯ = (ℎ¯𝑠 + ℎ¯𝑣 )/2, with ℎ¯𝑠 representing the C2 status update AoI ℎ𝑢,𝑠 (𝑡) averaged over time and across all UAVs, and ℎ¯𝑣 representing the video frame AoI ℎ𝑢,𝑣 (𝑡) averaged over time and across all UAVs. Figure 4 places each policy on this freshness–risk plane for three load regimes 𝜌 = 𝑁 /𝐾,

4.3

Does RA Prioritize Operationally Critical UAVs?

Section 4.2 showed that RA reduces tail risk by directing freshness toward the UAV for which additional staleness 10

Γsched

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations 6 5 4 3 2 1

ACM MobiCom ’27, October 2027,

twice the scheduling probability of a UAV operating at SAILII. The ratio exceeds three at most operating points with 𝐾 ≤ 2 and reaches six at 𝐾 = 1, 𝑁 = 100. This aggregate difference is not explained solely by higher-SAIL UAVs having older information or different channel conditions. In the controlled scenario in Appendix I, where candidate UAVs have comparable AoI and radio state, RA selects the higherSAIL operation in 64% of matched decisions (median over operating points), against 50% for the baselines. Moreover, Fig. 5 shows that the allocation ratio generally increases with contention, indicating that RA’s operational context becomes most influential precisely when the scheduler must make more selective deferral decisions. RA produces this alignment without observing SAIL, which is used only as an offline evaluation label. Takeaway. RA’s risk reduction is explained by where it places scarce transmission opportunities, not by uniformly improving freshness. Under contention, RA assigns more than the per-UAV scheduling probability to higher-SAIL operations, and this preference persists when comparing UAVs with similar AoI and radio conditions. Because SAIL is neither an input to RA nor part of its optimization objective, this alignment provides an independent check of the abstractions leading to the scheduler. Although it does not validate each modeling assumption individually, it indicates that the assumptions used to define the composite operational-risk metric, 𝑅𝑢 (𝑡), derive its tractable surrogate, 𝑅b𝑢 (𝑡), formulate the scheduling optimization problem in (22), and derive RA collectively preserve the operational distinctions relevant to scheduling. The resulting network decisions are therefore consistent with an independent operation-level risk classification.

2.37 0.92

0.92

0.92 RR

PF

Median

AoI

RA

Class-blind Reference

Figure 5: For each of the 42 contended operating points with 𝑁 > 𝐾, aggregated across both evaluation scenarios, Γsched compares the mean per-UAV scheduling probability under higher-SAIL conditions with that under SAIL-II conditions. Each point represents one operating point, bars denote medians, and the dashed line at Γsched = 1 denotes class-blind allocation. is most consequential. We now test whether this behavior aligns with an external measure of operational criticality rather than only with RA’s own risk metric. We use SORA’s SAIL as this independent reference. SAIL classes I through VI (discussed in Appendix I) rank operations according to the rigor required by their operational risk, with higher classes generally corresponding to more consequential operating conditions. We assign SAIL offline and do not expose it to RA or any baseline. In each slot 𝑡, we partition the UAVs according to whether they are operating under SAIL-II or higher-SAIL conditions. For UAV 𝑢, let 𝜃 u,SAIL-II denote the fraction of slots in which UAV 𝑢 is scheduled among the slots in which it is classified as SAIL-II. We then define 𝜃 SAIL-II as the mean of 𝜃 u,SAIL-II across UAVs. We proceed analogously to define 𝜃 u,Higher-SAIL and 𝜃 Higher-SAIL . We then define the scheduling-alignment ratio Γsched = 𝜃 Higher-SAIL /𝜃 SAIL-II . A value of Γsched = 1 denotes class-blind scheduling, and Γsched > 1 indicates that the scheduler directs more service per UAV toward higher-SAIL operations. Figure 5 reports Γsched for the 42 operating points with 𝑁 > 𝐾, aggregated across both evaluation scenarios. At these points, the available transmission opportunities are insufficient to serve every UAV in each slot, so the policies must decide which UAV to defer. We exclude the remaining six points because 𝑁 ≤ 𝐾 allows every UAV to be scheduled simultaneously. Under those conditions, differences in Γsched primarily reflect traffic availability rather than prioritization. Figure 5 shows that the conventional and age-based baselines remain close to class-blind allocation. RR, PF, and AoI each have a median Γsched = 0.92, near the class-blind reference of one. Because these schedulers do not observe SAIL, we interpret the modest departure from one as scenarioinduced variation rather than evidence of systematic SAIL class preference. RA exhibits a clearly different allocation pattern. Its median Γsched = 2.37 indicates that, under contention, a UAV operating above SAIL-II receives more than

4.4

How Much Operational Capacity Does RA Enable?

We next translate RA’s reduction in operational-risk into communication-constrained operational capacity. For a policy 𝜋 scheduling up to 𝐾 UAV transmissions in each slot, we define the operational capacity 𝑁𝜋★ (𝐾) as the largest evaluated fleet for which every UAV remains within the communication-induced operational-risk envelope, namely 𝑁𝜋★ (𝐾) = max{𝑁 : max𝑢 𝑅¯𝑢 ≤ 1 at (𝐾, 𝑁 )}. We use the worst-UAV criterion because admitting a fleet based on its average risk could conceal an individual UAV that persistently exceeds the envelope. Figure 6 reports max𝑢 𝑅¯𝑢 for every (𝑁 , 𝐾) pair, aggregated across both evaluation scenarios. Values at or below one identify the admissible region for each policy. RA produces the largest admissible region when the channel budget is constrained but still sufficient for prioritization 11

ACM MobiCom ’27, October 2027,

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

K

RR

PF

32

0.34

0.33

0.33

0.37

0.48

0.58

0.77

16

0.34

0.33

0.37

0.58

0.86

1.25

8

0.34

0.35

0.54

1.22

2.70

3.67

4

0.34

0.53

0.94

9.85

6.40

2

0.45

1.30

2.80

9.94

1

0.86

3.29

32

0.34

0.34

0.34

0.39

0.48

0.64

0.73

16

0.34

0.34

0.39

0.61

0.95

1.29

8

0.34

0.38

0.60

1.34

2.37

4.07

4

0.37

0.60

1.27

3.59

6.94

2

0.61

1.31

3.49

1

1.18

5

0.82

0.33

0.32

0.33

0.35

0.40

0.47

0.55

0.62

2.40

2.29

0.33

0.32

0.35

0.46

0.61

0.81

1.09

1.42

8.47

19.06

0.33

0.34

0.46

0.83

1.43

2.14

2.85

3.70

28.22 40.47 93.48

0.33

0.46

0.81

2.08

3.68

5.45

7.63

10.23

24.47 45.11 86.37 88.85

0.45

0.79

1.99

5.35

10.45 30.89 27.56 42.19

18.55 36.57 74.04 181.58 344.17 567.29

0.79

2.08

5.46

18.72 42.28 72.75 548.89 907.71

0.97

0.31

0.30

0.31

0.30

0.33

0.41

0.42

0.50

1.73

2.57

0.31

0.30

0.30

0.40

0.46

0.59

0.67

0.88

5.50

7.39

0.31

0.30

0.39

0.61

0.87

1.40

1.76

2.26

11.75 21.46 52.45

0.30

0.32

0.57

1.40

2.21

3.62

4.48

6.52

11.56 34.70 44.45 62.16 136.49

0.35

0.52

1.32

3.56

5.95

9.87

10.34 14.20

5.65

11.82 52.86 78.75 185.91 174.42 226.44

0.46

1.13

3.48

11.19 13.40 27.94 38.34 51.83

10

20

5

10

20

40

K

AoI

40

60

for the offered fleet. In this regime, the appropriate response may be to increase communication capacity (e.g., by expanding the bandwidth of the UAV network slice) or reduce traffic demand through UAV departure spacing, speed adjustment, or rerouting. Communication-constrained operational capacity therefore provides an interface between millisecond-scale scheduling and longer-timescale traffic-management decisions. Appendix K discuss how it could be used in air traffic control and air traffic flow management. Takeaway. Operational capacity depends not only on the amount of available radio resources, but also on whether those resources are allocated according to the operational consequences of delay. Within the allocation-sensitive regime, RA supports 1.5 to 3× larger evaluated fleets than the baselines under the same communication-induced operationalrisk envelope. This gain requires no additional spectrum or transmission opportunities. It follows from redirecting service toward the aircraft and streams for which deferral is most consequential. However, risk-aware allocation cannot overcome an aggregate service deficit. Once every policy exceeds the envelope, maintaining admissibility requires additional communication capacity or traffic-management actions that reduce simultaneous demand.

RA

80

100 120

Simultaneous UAVs

60

80

100 120

Simultaneous UAVs

Figure 6: Each cell reports max𝑢 𝑅¯𝑢 for one (𝐾, 𝑁 ) operating point. Values at or below one satisfy the communication-induced operational-risk envelope and admit the corresponding evaluated fleet. Values above one violate the envelope. to affect the risk boundary. At 𝐾 = 8, RA supports 60 simultaneous UAVs, compared with 40 under PF and 20 under RR and AoI. At 𝐾 = 16, RA supports the full evaluated fleet of 120 UAVs, compared with 80 under PF and 60 under RR and AoI. Thus, within the evaluated range, RA supports up to 1.5× as many UAVs as PF, 3× as many as RR, and 2 to 3× as many as AoI while enforcing the same communicationinduced operational-risk envelope. These gains arise without increasing the channel budget. RA instead changes which UAVs and information streams receive the available transmission opportunities. A complete evaluation across all 48 operating points can be found in Appendix H. Section 4.2 used 𝑝 99,𝑢 [𝑅¯𝑢 ] to compare persistent tail exposure, whereas operational capacity applies the stricter condition max𝑢 𝑅¯𝑢 ≤ 1 to ensure that no individual UAV exceeds the envelope. This stricter criterion does not reduce the measured capacity on the evaluated grid. Replacing max𝑢 𝑅¯𝑢 with 𝑝 99,𝑢 [𝑅¯𝑢 ] yields the same admissible (𝐾, 𝑁 ) region for every policy. Within that region, the worst-UAV risk exceeds the 99th percentile by only 1% on median and by at most 13%. The capacity gains are therefore not obtained by improving most UAVs while leaving a single aircraft substantially more exposed. Over the evaluated operating points, RA enlarges the admissible fleet under both a fleet-tail criterion and an individual-UAV criterion. Figure 6 also separates failures of resource allocation from resource insufficiency. When a baseline exceeds the risk envelope but RA remains admissible, the available radio resources are sufficient, but the baseline allocates them inconsistently with operational consequence. When every policy exceeds the envelope, the channel budget itself may be insufficient

5

Related Work

Cellular networks and operational risk. Prior work on cellular-connected UAVs characterizes aerial propagation, interference, mobility, handovers, and three-dimensional coverage, and optimizes association, trajectory planning, and radio-resource management [30, 31, 48]. Connectivity-aware planners further use radio maps or predicted outages to adapt the UAV trajectory [17, 35]. 3GPP QoS, slicing, and reservation mechanisms differentiate traffic through rates, latency, reliability, or service classes [16, 49]. These approaches improve connectivity, but do not determine which aircraft is most costly to delay when they contend for the same uplink. A parallel body of UAS safety work models ground exposure from population and impact footprints and air risk from navigation uncertainty and separation [6, 27, 36, 37]. SORA and related assurance frameworks further make C3 performance part of the operational safety case [13, 20, 21]. These methods assess or constrain operations, but generally treat communication performance as an input. RA instead turns these operational quantities into a network-computable riskdriven metric used online for radio-resource allocation. Freshness- and context-aware scheduling. Age of Information motivates scheduling policies that minimize average, weighted, or peak staleness under wireless constraints [18, 23, 24, 26, 41, 50], while recent work handles updates requiring multiple transmission opportunities [51]. Closest to RA, Urgency of Information (UoI) extends AoI-type timeliness by 12

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027,

combining a time-varying context weight with estimationerror cost and uses this urgency for multi-terminal statusupdate scheduling [52]. UoI therefore establishes that context can matter beyond age alone. However, it assumes the context weight as an input to the model. The paper identifies choosing practical context weights as an open problem. update supersedes older status information. RA addresses a different systems question: how a 5G scheduler can derive such prioritization from an aviation operational-safety model. We map network-observable freshness and aircraft context to normalized navigation, ground, and video-coherence exposure [7, 8, 32], and jointly schedule single-slot status updates and multi-slot video objects. The resulting risk envelope further translates this online allocation into communication-constrained operational capacity.

6

https://sistemas.anatel.gov.br/se/public/view/b/licenciamento.php? view=licenciamento. Accessed August 25, 2026. [5] Fawad Ahmad, Christina Suyong Shin, Rajrup Ghosh, John D’Ambrosio, Eugene Chai, Karthikeyan Sundaresan, and Ramesh Govindan. 2023. AeroTraj: Trajectory Planning for Fast, and Accurate 3D Reconstruction Using a Drone-based LiDAR. 7, 3, Article 83 (Sept. 2023), 28 pages. doi:10.1145/3610911 [6] Omid Asghari, Naghmeh Ivaki, and Henrique Madeira. 2025. UAV Operations Safety Assessment: A Systematic Literature Review. ACM Comput. Surv. 57, 10, Article 266 (2025), 37 pages. [7] Aygün Baltaci, Hendrik Cech, Nitinder Mohan, Fabien Geyer, Vaibhav Bajpai, Jörg Ott, and Dominic Schupke. 2022. Analyzing Real-Time Video Delivery over Cellular Networks for Remote Piloting Aerial Vehicles. In Proc. ACM IMC’22. 98–112. [8] Aygün Baltaci, Markus Klügel, Fabien Geyer, Svetoslav Duhovnikov, Vaibhav Bajpai, Jörg Ott, and Dominic Schupke. 2021. Experimental UAV Data Traffic Modeling and Network Performance Analysis. In Proc. IEEE INFOCOM’21. 1–10. [9] China Daily. 2025. Meituan First Nationwide Low-Altitude Operating License. https://www.chinadaily.com.cn/a/202504/22/ WS6807752fa3104d9fd3820e14.html. [10] DECEA. 2022. Concepção Operacional UTM Nacional. Diretriz do Comando da Aeronáutica DCA 351-6. Comando da Aeronáutica, Departamento de Controle do Espaço Aéreo, Rio de Janeiro, Brasil. https://publicacoes.decea.mil.br/publicacao/dca-351-6 [11] DECEA. 2024. Operação Lagos: DECEA controla 224 helicópteros e 82 drones durante GP Brasil, em Interlagos (SP). DECEA. https://decea.mil.br/?i=midia-e-informacao&materia=operacaolagos-decea-controla-224-helicopteros-e-82-drones-durante-gpbrasil-em-interlagos-sp&p=pg_noticia [12] EUROCAE WG-105. 2019. ED-253: Operational Services and Environment Definition (OSED) for Automation and Emergency Recovery for UAS. Technical Report. EUROCAE. [13] EUROCAE WG-105. 2020. ED-267: Operational Services and Environment Definition (OSED) for Detect and Avoid in Very Low-Level Operations. Technical Report. EUROCAE. [14] EUROCAE WG-105. 2020. ED-272: Minimum Aviation Systems Performance Standards for Remote Pilot Stations Conducting IFR Operations in Controlled Airspace. Technical Report. EUROCAE. [15] Federal Aviation Administration. 2020. Unmanned Aircraft System (UAS) Traffic Management (UTM) Concept of Operations, Version 2.0. Technical Report. FAA. [16] Xenofon Foukas, Mahesh K. Marina, and Kimon Kontovasilis. 2017. Orion: RAN Slicing for a Flexible and Cost-Effective Multi-Service Mobile Network Architecture. In Proc. ACM MobiCom’17. 127–140. [17] Mohammed Lahouari Harchaoui, Sihem Ouahouah, Oussama Bekkouche, Miloud Bagaa, and Abir Derouiche. 2024. 5G-Based Ground Risk Mitigation for UAVs: A Deep Reinforcement Learning Approach. In Proc. IEEE GLOBECOM’24. 1539–1544. [18] Longbo Huang and Eytan Modiano. 2015. Optimizing Age-ofInformation in a Multi-Class Queueing System. In Proc. IEEE ISIT’15. 1681–1685. [19] iFood Institucional. 2026. Drone do iFood em São Paulo: empresa inicia entregas aéreas. iFood Institucional. https://institucional.ifood.com. br/inovacao/drone-ifood-em-sao-paulo/ [20] JARUS WG-SRM. 2024. JARUS Guidelines on Specific Operations Risk Assessment (SORA), Edition 2.5. Technical Report JAR-DEL-SRM-SORAMB-2.5. Joint Authorities for Rulemaking on Unmanned Systems (JARUS). [21] Joint Authorities for Rulemaking on Unmanned Systems (JARUS). 2024. JARUS Guidelines on SORA: Annex F – Theoretical Basis for Ground

Conclusion

Low-altitude aviation exposes a limitation of conventional cellular scheduling: network-centric metrics do not capture the operational consequences of delaying information from different aircraft. We introduce a network-to-operations abstraction that translates information freshness and operational context into aircraft-specific operational risk, enabling the network to evaluate not only how efficiently a transmission can be served, but also the consequence of making it wait. We instantiate this abstraction through RA scheduling, which prioritizes the aircraft and information streams for which additional delay is most consequential while accounting for heterogeneous transmission costs. Although RA is derived from an analytically tractable model, it is evaluated under moving UAV trajectories, dynamic operational context, geographically varying population exposure, and realistic UAV video traffic. Across 48 operating points and real São Paulo operating scenarios, RA substantially reduces communication-induced operational risk and increases the number of UAVs supportable within the same risk envelope. More broadly, this work illustrates how future communication systems supporting low-altitude aviation can incorporate operational consequence directly into radio-resource allocation.

References [1] 3GPP. 2021. Study on NR Positioning Support. Technical Report (TR) 38.857. 3rd Generation Partnership Project (3GPP). https://www.3gpp. org/ftp/Specs/archive/38_series/38.857/ Release 17, V17.0.0. [2] 3GPP. 2024. 5G System (5GS) Location Services (LCS); Stage 2. TS 23.273. 3rd Generation Partnership Project. https://www.3gpp.org/ DynaReport/23273.htm Release 19. [3] 3GPP. 2025. Architecture enhancements for 5G System (5GS) to support network data analytics services. TS 23.288. 3rd Generation Partnership Project. https://www.3gpp.org/DynaReport/23288.htm Release 19. [4] Agência Nacional de Telecomunicações (ANATEL). 2026. Mosaico Spectrum-E: Consulta Pública de Emissões Licenciadas. 13

ACM MobiCom ’27, October 2027,

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

Risk Classification and Mitigation. Technical Report JAR-DEL-SRMSORA-F-2.5. Joint Authorities for Rulemaking on Unmanned Systems (JARUS). https://jarus-rpas.org/wp-content/uploads/2024/06/SORAv2.5-Annex-F-Release.JAR_doc_29pdf.pdf Edition 2.5, Public Release, 13 May 2024. [22] Igor Kadota and Eytan Modiano. 2021. Minimizing the Age of Information in Wireless Networks with Stochastic Arrivals. IEEE Transactions on Mobile Computing 20, 3 (2021), 1173–1185. [23] Igor Kadota, Abhishek Sinha, and Eytan Modiano. 2019. Scheduling Algorithms for Optimizing Age of Information in Wireless Networks with Throughput Constraints. IEEE/ACM Transactions on Networking 27, 4 (2019), 1359–1372. [24] Igor Kadota, Abhishek Sinha, Elif Uysal-Biyikoglu, Rahul Singh, and Eytan Modiano. 2018. Scheduling Policies for Minimizing Age of Information in Broadcast Wireless Networks. IEEE/ACM Transactions on Networking 26, 6 (2018), 2637–2650. [25] Dimitrios Kafetzis, Nikos Fotiou, Savvas Argyropoulos, Jad Nasreddine, and Iordanis Koutsopoulos. 2025. Video Quality Monitoring for Remote Autonomous Vehicle Control. In Proc. IEEE MeditCom’25. 1–7. [26] Sanjit Kaul, Roy D. Yates, and Marco Gruteser. 2012. Real-Time Status: How Often Should One Update?. In Proc. IEEE INFOCOM’12. 2731– 2735. [27] Younsil Kim and Joongwon Bae. 2022. Risk-Based UAV Corridor Capacity Analysis above a Populated Area. Drones 6, 9 (2022), 221. [28] KrASIA. 2024. Meituan’s Drone Delivery Debuts in Dubai, Eyeing a Future in Last-Mile Logistics. https://kr-asia.com/meituans-dronedelivery-debuts-in-dubai-eyeing-a-future-in-last-mile-logistics. [29] Yu-Tai Lin, N. Cameron Matson, and Karthikeyan Sundaresan. 2024. Bringing Collaborative Positioning to Native 5G Systems for Enhanced 2D & 3D Location Services. 2, CoNEXT4, Article 23 (Nov. 2024), 21 pages. doi:10.1145/3696380 [30] Xinzhe Liu, Jianer Zhou, Xiaoyong Ni, Ke Luo, Zhenyu Li, Xiaofeng Tao, and Weichao Li. 2026. Unveiling Low-Altitude 5G Performance: Linking Key Influencing Factors with UAV Flight Parameters. In Proceedings of the ACM SIGCOMM 2026 Conference (Denver, CO, USA). Association for Computing Machinery, New York, NY, USA, 317–333. [31] Yanbing Liu, Jingqi Huang, Ziyu Li, and Chunyi Peng. 2026. 5G in the Sky: Uplink Throughput Measurement, Analysis, and Enhancement. IEEE Transactions on Networking 34 (2026), 941–956. [32] Kimberly McGuire, Guido de Croon, Christophe De Wagter, Karl Tuyls, and Hilbert Kappen. 2017. Efficient Optical Flow and Stereo Vision for Velocity Estimation and Obstacle Avoidance on an Autonomous Pocket Drone. IEEE Robotics and Automation Letters 2, 2 (2017), 1070–1076. [33] MundoGEO. 2025. DECEA detectou drones irregulares e assegurou o fluxo aéreo no GP de São Paulo de F1. MundoGEO. https://mundogeo.com/2025/11/17/decea-detectou-dronesirregulares-e-assegurou-o-fluxo-aereo-no-gp-de-sao-paulo-de-f1/ [34] MundoGEO. 2026. Speedbird Aero and iFood Launch New Drone Delivery Route in São Paulo, Brazil. MundoGEO. https://mundogeo.com/en/2026/06/04/speedbird-aero-and-ifoodlaunch-new-drone-delivery-routein-sao-paulo-brazil/ [35] Sihem Ouahouah, Mohammed Lahouari Harchaoui, Oussama Bekkouche, Miloud Bagaa, and Riku Jäntti. 2025. 5G-Based Autonomous Ground Risk Mitigation for UAVs. In Proc. IEEE ICC’25. 1359–1364. [36] Stefano Primatesta, Alessandro Rizzo, and Anders la Cour-Harbo. 2020. Ground Risk Map for Unmanned Aircraft in Urban Environments. Journal of Intelligent and Robotic Systems 97, 3–4 (2020), 489–509. [37] RTCA SC-228. 2022. DO-365C: Minimum Operational Performance Standards (MOPS) for Detect and Avoid (DAA) Systems. Technical Report. RTCA.

[38] SESAR Joint Undertaking. 2023. U-space Concept of Operations (ConOps), Fourth Edition. Technical Report. SESAR Joint Undertaking. [39] MReza Alipour Sormoli, Mehrdad Dianati, Sajjad Mozaffari, and Roger Woodman. 2024. Optical Flow Based Detection and Tracking of Moving Objects for Autonomous Vehicles. IEEE Trans. Intell. Transp. Syst. 25, 9 (2024), 12578–12590. [40] sUAS News. 2026. Delivery Drones Cleared for Flight over Populated Cities in Brazil. https://www.suasnews.com/2026/03/delivery-dronescleared-for-flight-over-populated-cities-in-brazil/. [41] Yin Sun, Elif Uysal-Biyikoglu, Roy D. Yates, C. Emre Koksal, and Ness B. Shroff. 2017. Update or Wait: How to Keep Your Data Fresh. IEEE Trans. Inf. Theory 63, 11 (Nov. 2017), 7492–7508. [42] The MathWorks, Inc. 2026. 5G Toolbox Documentation. https://www. mathworks.com/help/5g/index.html Accessed September 1, 2026. [43] The MathWorks, Inc. 2026. UAV Toolbox Documentation. https: //www.mathworks.com/help/uav/index.html Accessed September 1, 2026. [44] UK Civil Aviation Authority. 2025. Future of Flight: BVLOS Roadmap. Civil Aviation Publication CAP 3182. UK Civil Aviation Authority. https://www.caa.co.uk/publication/download/26211 [45] UK Civil Aviation Authority. 2026. Concept of Operations: Future Air Traffic Management and Air Navigation Service. Civil Aviation Publication CAP 3269. UK Civil Aviation Authority. https://www.caa.co.uk/ publication/download/29467 [46] U.S. Department of Transportation. 2025. Advanced Air Mobility National Strategy 2025. Technical Report. U.S. Department of Transportation. [47] U.S. DoT. 2025. Advanced Air Mobility Comprehensive Plan 2025. Technical Report. U.S. Department of Transportation. [48] Shuowen Zhang, Yong Zeng, and Rui Zhang. 2019. Cellular-Enabled UAV Communication: A Connectivity-Constrained Trajectory Optimization Perspective. IEEE Transactions on Communications 67, 3 (2019), 2580–2604. [49] Xiao Zhang and Daehyeok Kim. 2026. Enabling SLO-Aware 5G MultiAccess Edge Computing with SMEC. In Proc. USENIX NSDI’26. 1759– 1776. [50] Zhuoyi Zhao and Igor Kadota. 2025. Optimizing Age of Information without Knowing the Age of Information. In Proc. IEEE INFOCOM’25. 1–10. [51] Zhuoyi Zhao, Vishrant Tripathi, and Igor Kadota. 2025. Optimizing Age of Information in Networks with Large and Small Updates. In Proc. IEEE WiOpt’25. 1–8. [52] Xi Zheng, Sheng Zhou, and Zhisheng Niu. 2020. Urgency of Information for Context-Aware Timely Status Updates in Remote Control Systems. IEEE Transactions on Wireless Communications 19, 11 (2020), 7237–7250.

A

Normalization of Operational Risk Components

The navigation, ground, and video-coherence risk components measure heterogeneous quantities: a separation deficit, an exposed population, and a video-staleness violation, respectively. To aggregate them, we normalize each component using a conservative reference value constructed from the corresponding declared operational limits. These reference values define dimensionless risk scales. Declared Operational Limits. The declared operational limits used for normalization are specified as follows. Let 14

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027,

B

𝜏𝑠max denote the maximum tolerable C2 status-update staleness, 𝑑 min the minimum declared separation between reported UAV positions, 𝑁 op the maximum number of UAVs simultaneously operating in the corridor, 𝜌 max the maximum population density within the declared operating area, and 𝜏𝑣max the maximum tolerable coherence-window staleness. Within these limits, we assume ℎ𝑢,𝑠 (𝑡) ≤ 𝜏𝑠max , 𝑁 ≤ 𝑁 op , 𝑑𝑢𝑣 (𝑡) ≥ 𝑑 min , 𝜌 (𝒙, 𝑡) ≤ 𝜌 max , and 𝑧𝑢 (𝑡) ≤ 𝜏𝑣max . Navigation Risk Normalization. At the maximum tolerable C2 status-update staleness, the position-uncertainty radius in (6) becomes 𝑟𝜏 = 𝑟 min + 𝑣 max𝜏𝑠max . Under the declared operational limits, the unnormalized navigation-risk component satisfies 𝑁 ∑︁

In this appendix, we prove Proposition 2.1 and derive the coefficients 𝛼𝑢,𝑠 , 𝛼𝑢,𝑣 , and 𝐶𝑢 . We separately upper-bound the navigation, ground, and video-coherence risk components. The local population-density bound 𝜌𝑢,max and pairwise separations {𝑑𝑢𝑣 } 𝑣≠𝑢 are determined by the operational context considered in the analysis. Navigation Risk Upper Bound. Using the pairwise separation 𝑑𝑢𝑣 under the considered operational context, together with (9) and (6), we have 1 ∑︁ [𝑟𝑢 (𝑡) + 𝑟 𝑣 (𝑡) − 𝑑𝑢𝑣 ] + 𝑅𝑢Nav (𝑡) = Nav ¯ 𝑅 𝑣≠𝑢 1 ∑︁  = Nav 2𝑟 min + 𝑣 maxℎ𝑢,𝑠 (𝑡) 𝑅¯ 𝑣≠𝑢 + +𝑣 maxℎ 𝑣,𝑠 (𝑡) − 𝑑𝑢𝑣 . (42)

[𝑟𝑢 (𝑡) + 𝑟 𝑣 (𝑡) − 𝑑𝑢𝑣 (𝑡)] +

𝑣=1 𝑣≠𝑢

≤ (𝑁 op − 1) [2𝑟𝜏 − 𝑑 min ] + .

(37)

Accordingly, we define the reference navigation risk as +

𝑅¯Nav = (𝑁 op − 1) [2𝑟𝜏 − 𝑑 min ] .

max . Therefore, Under (14), ℎ 𝑣,𝑠 (𝑡) ≤ 𝜏𝑣,𝑠

(38) 𝑅𝑢Nav (𝑡) ≤

This is a conservative reference scale corresponding to the simultaneous application of the declared fleet-size, C2 statusupdate staleness, and separation limits. We assume 𝑑 min < 2𝑟𝜏 , so that 𝑅¯Nav > 0. Ground Risk Normalization. At the maximum tolerable C2 status-update staleness, the position-uncertainty region has area 𝐴𝜏 = 𝜋𝑟𝜏2 . Since 𝜌 (𝒙, 𝑡) ≤ 𝜌 max and 𝑟𝑢 (𝑡) ≤ 𝑟𝜏 under the declared operational limits, the exposed population satisfies ∫ 𝜌 (𝒙, 𝑡) d𝒙 ≤ 𝜌 max 𝜋𝑟𝜏2 .

(43)

where max 𝑏𝑢𝑣 = 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 .

(44)

] + is convex in ℎ, its secant line over 0 ≤

Since [𝑣 maxℎ + 𝑏𝑢𝑣 max gives ℎ ≤ 𝜏𝑢,𝑠  + 𝑣 maxℎ𝑢,𝑠 (𝑡) + 𝑏𝑢𝑣 ≤ 𝛽𝑢𝑣 ℎ𝑢,𝑠 (𝑡) + 𝛾𝑢𝑣 ,

(45)

where

(39)

  max + − [𝑏 ] + 𝑏𝑢𝑣 + 𝑣 max𝜏𝑢,𝑠 𝑢𝑣 𝛽𝑢𝑣 = , max 𝜏𝑢,𝑠 𝛾𝑢𝑣 = [𝑏𝑢𝑣 ] + .

We therefore define the reference ground risk as (40)

It represents a conservative exposed-population scale at the declared C2 status-update staleness and population-density limits. Video-Coherence Risk Normalization. Let 𝐻 nom and 𝜏𝑣max denote the nominal and maximum tolerable coherence-window staleness, respectively, where 𝜏𝑣max > 𝐻 nom . We define the reference video-coherence risk as 𝑅¯Vid = 𝜏𝑣max − 𝐻 nom .

𝑅¯

+ 1 ∑︁  𝑣 ℎ (𝑡) + 𝑏 , max 𝑢,𝑠 𝑢𝑣 Nav 𝑣≠𝑢

B𝑢 (𝑡 )

𝑅¯Ground = 𝜌 max𝐴𝜏 = 𝜌 max 𝜋𝑟𝜏2 .

Derivation of the Composite Operational Risk Upper Bound

Substituting (45) into (43) yields Í Í 𝛾𝑢𝑣 𝑣≠𝑢 𝛽𝑢𝑣 Nav 𝑅𝑢 (𝑡) ≤ ℎ𝑢,𝑠 (𝑡) + 𝑣≠𝑢 . 𝑅¯Nav 𝑅¯Nav

(46) (47)

(48)

After applying the navigation-risk weight 𝑤 𝑁 , we obtain Í Í 𝑤 𝑁 𝑣≠𝑢 𝛽𝑢𝑣 𝑤 𝑁 𝑣≠𝑢 𝛾𝑢𝑣 Nav 𝑤 𝑁 𝑅𝑢 (𝑡) ≤ ℎ𝑢,𝑠 (𝑡) + . (49) 𝑅¯Nav 𝑅¯Nav

(41)

Substituting (44) into (46) and (47), we obtain + 1  max max
− 𝑑𝑢𝑣 𝛽𝑢𝑣 = max 2𝑟 min + 𝑣 max 𝜏𝑢,𝑠 + 𝜏𝑣,𝑠 𝜏𝑢,𝑠  + max − 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 ,   + max 𝛾𝑢𝑣 = 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 .

Accordingly, the normalized video-coherence risk is zero for 𝑧𝑢 (𝑡) ≤ 𝐻 nom and equals one when 𝑧𝑢 (𝑡) = 𝜏𝑣max . Under the declared operational limits, each normalized risk component is upper bounded by one. A value of one corresponds to its conservative reference scale. A value greater than one indicates that at least one of the limits used to construct that reference scale has been exceeded. 15

(50) (51)

ACM MobiCom ’27, October 2027,

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

Therefore, (49) becomes ∑︁   + 𝑤𝑁 max max
𝑤 𝑁 𝑅𝑢Nav (𝑡) ≤ Nav max 2𝑟 min + 𝑣 max 𝜏𝑢,𝑠 + 𝜏𝑣,𝑠 − 𝑑𝑢𝑣 ¯ 𝑅 𝜏𝑢,𝑠 𝑣≠𝑢  + max − 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 ℎ𝑢,𝑠 (𝑡)  𝑤 𝑁 ∑︁  + max 2𝑟 min + 𝑣 max𝜏𝑣,𝑠 − 𝑑𝑢𝑣 . (52) + Nav 𝑅¯

Finally, applying (60) yields (16). This completes the proof of Proposition 2.1.

(1) Since the completed-frame AoIs are ordered as ℎ𝑢,𝑣 (𝑡) ≤ (2) (𝑤 ) 𝑤 ℎ𝑢,𝑣 (𝑡) ≤ · · · ≤ ℎ𝑢,𝑣 (𝑡), Í𝑤and the coefficients {𝛼𝑖 }𝑖=1 are nonnegative and satisfy 𝑖=1 𝛼𝑖 = 1, we have 𝑤 𝑤 ∑︁ ∑︁ (𝑖 ) (𝑤 ) (𝑤 ) 𝑧𝑢 (𝑡) = 𝛼𝑖 ℎ𝑢,𝑣 (𝑡) ≤ 𝛼𝑖 ℎ𝑢,𝑣 (𝑡) = ℎ𝑢,𝑣 (𝑡). (60)

without a successful delivery,

C

PROOF OF THEOREM 3.1

Consider any admissible policy 𝜋 ∈ Π operating over a finite horizon of 𝑇 slots. For the status update flow of UAV 𝑢, let 𝐷𝑢,𝑠 (𝑇 ) denote the number of successfully delivered 𝑣≠𝑢 (1) status updates by the end of slot 𝑇 . Let 𝐼𝑢,𝑠 denote the time Ground Risk Upper Bound. From (10) and the local populationfrom the beginning of the horizon to the first successful density bound in (15), we have (𝑚) status delivery, and, for 𝑚 ≥ 2, let 𝐼𝑢,𝑠 denote the inter∫ 1 Ground delivery time between the (𝑚 − 1)-th and 𝑚-th successful 𝑅𝑢 (𝑡) = Ground 𝜌 (𝒙, 𝑡) d𝒙 𝑅¯ B𝑢 (𝑡 ) status deliveries. Let 𝑅𝑢,𝑠 (𝑇 ) denote the residual time after 𝜌𝑢,max the last successful status delivery. Then ≤ Ground |B𝑢 (𝑡)| . (53) ¯ 𝑅 𝐷𝑢,𝑠 ∑︁(𝑇 ) Using the area of the circular uncertainty region in (8), we (𝑚) 𝑇 = 𝐼𝑢,𝑠 + 𝑅𝑢,𝑠 (𝑇 ). (62) obtain 𝑚=1 2 𝜌𝑢,max 𝜋  (54) 𝑅𝑢Ground (𝑡) ≤ Ground 𝑟 min + 𝑣 maxℎ𝑢,𝑠 (𝑡) . After a successful status delivery, the status AoI is reset to 𝑅¯ (𝑚) 1. Hence, over an inter-delivery interval of length 𝐼𝑢,𝑠 , the Under (14), (𝑚) 2 max AoI values are 1, 2, . . . , 𝐼 . Therefore, the cumulative status ℎ𝑢,𝑠 (𝑡) ≤ 𝜏𝑢,𝑠 ℎ𝑢,𝑠 (𝑡). (55) 𝑢,𝑠 AoI over this interval is Substituting (55) into (54) gives
(𝑚) 2 𝜏 max 𝐼∑︁ (𝑚) (𝑚) (𝑚) 2 (𝑚) 𝑢,𝑠 𝜌𝑢,max 𝜋 2𝑟 min𝑣 max + 𝑣 max 𝑢,𝑠 𝐼𝑢,𝑠 (𝐼𝑢,𝑠 + 1) (𝐼𝑢,𝑠 ) 𝐼𝑢,𝑠 𝑅𝑢Ground (𝑡) ≤ ℎ (𝑡) 𝑢,𝑠 𝜏 = = + . (63) 𝑅¯Ground 2 2 2 𝜏=1 2 𝜌𝑢,max 𝜋𝑟 min . (56) + Applying this to all complete inter-delivery intervals and to 𝑅¯Ground the residual interval, we obtain After applying the ground-risk weight 𝑤𝐺 , we obtain
𝐷𝑢,𝑠 (𝑇 )  𝑇 2 𝜏 max ∑︁  𝑤𝐺 𝜌𝑢,max 𝜋 2𝑟 min𝑣 max + 𝑣 max 1  ∑︁ (𝑚) 2 𝑢,𝑠 2  ℎ (𝑡) = (𝐼 ) + 𝑅 (𝑇 ) 𝑤𝐺 𝑅𝑢Ground (𝑡) ≤ ℎ (𝑡) 𝑢,𝑠 𝑢,𝑠 𝑢,𝑠 𝑢,𝑠   2  𝑚=1 𝑅¯Ground  𝑡 =1   2 𝑤𝐺 𝜌𝑢,max 𝜋𝑟 min  𝐷𝑢,𝑠 (𝑇 ) . (57) +  1  ∑︁ (𝑚) 𝑅¯Ground 𝐼𝑢,𝑠 + 𝑅𝑢,𝑠 (𝑇 )  +  2  𝑚=1 Video-Coherence Risk Upper Bound. From (12), since    𝑧𝑢 (𝑡) ≥ 0, 𝐻 nom ≥ 0, and 𝑅¯Vid > 0, we have   𝐷 (𝑇 ) 𝑢,𝑠  +  𝑇 1  ∑︁ (𝑚) 2 2 𝑧𝑢 (𝑡) − 𝐻 nom 𝑧𝑢 (𝑡) + , Vid = (𝐼 ) + 𝑅 (𝑇 ) (64) 𝑢,𝑠 𝑢,𝑠   2 𝑅𝑢 (𝑡) = ≤ . (58) 2  𝑚=1  𝑅¯Vid 𝑅¯Vid   Consequently, where the last equality follows from (62). If 𝐷𝑢,𝑠 (𝑇 ) = 0, 𝑤𝑉 then no status update is delivered over the horizon. Since 𝑤𝑉 𝑅𝑢Vid (𝑡) ≤ Vid 𝑧𝑢 (𝑡). (59) 𝑅¯ ℎ (1) = 1 and the status AoI increases by one in every slot 𝑢,𝑠

𝑖=1

𝑇

𝑇 +1 1 𝑇 1 1 ∑︁ ℎ𝑢,𝑠 (𝑡) = = + . 𝑇 𝑡 =1 2 2 𝐷𝑢,𝑠 (𝑇 ) + 1 2

Thus, (69) holds directly when 𝐷𝑢,𝑠 (𝑇 ) = 0. It therefore suffices to consider 𝐷𝑢,𝑠 (𝑇 ) ≥ 1 in the following derivation. By Jensen’s inequality, Í𝐷𝑢,𝑠 (𝑇 ) (𝑚) ! 2 𝐷𝑢,𝑠 ∑︁(𝑇 ) 𝐼𝑢,𝑠 (𝑚) 2 𝑚=1 (𝐼𝑢,𝑠 ) ≥ 𝐷𝑢,𝑠 (𝑇 ) . (66) 𝐷𝑢,𝑠 (𝑇 ) 𝑚=1

𝑖=1

Combining (52), (57), and (59) with the composite operational risk in (13), and using the coefficient definitions in (17)–(19), we obtain 𝑅𝑢 (𝑡) ≤ 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 𝑧𝑢 (𝑡) + 𝐶𝑢 .

(65)

(61) 16

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027, (1) the first interval of length 𝐼𝑢,𝑣 , the video AoI values are (1) 1, 2, . . . , 𝐼𝑢,𝑣 . After a video frame is completed, the video AoI is reset to at least 𝐿, because the frame contains 𝐿 packets and at most one packet can be successfully delivered in each slot. Hence, over every subsequent inter-completion interval of (𝑚) length 𝐼𝑢,𝑣 , 𝑚 ≥ 2, the video AoI values are lower bounded (𝑚) by 𝐿, 𝐿 + 1, . . . , 𝐿 + 𝐼𝑢,𝑣 − 1. Therefore, the cumulative video AoI over this interval is lower bounded by

Using (62) we get 𝐷𝑢,𝑠 ∑︁(𝑇 )

(𝑚) 2 (𝐼𝑢,𝑠 ) ≥

𝑚=1

(𝑇 − 𝑅𝑢,𝑠 (𝑇 )) 2 . 𝐷𝑢,𝑠 (𝑇 )

Substituting (67) into (64) gives " # 𝑇 2 (𝑇 ) 1 ∑︁ 1 1 (𝑇 − 𝑅𝑢,𝑠 (𝑇 )) 2 𝑅𝑢,𝑠 + + . ℎ𝑢,𝑠 (𝑡) ≥ 𝑇 𝑡 =1 2 𝑇 𝐷𝑢,𝑠 (𝑇 ) 𝑇 2

(67)

(68)

Minimizing the RHS of (68) with respect to 𝑅𝑢,𝑠 (𝑇 ) yields 𝑇 1 ∑︁

ℎ𝑢,𝑠 (𝑡) ≥

𝑇 𝑡 =1

1 𝑇 1 + . 2 𝐷𝑢,𝑠 (𝑇 ) + 1 2

(𝑚)

𝐼𝑢,𝑣 ∑︁−1

  (𝑚) 2 (𝐼𝑢,𝑣 ) 1 (𝑚) + 𝐿− 𝐼 . (𝐿 + 𝜏) = 2 2 𝑢,𝑣 𝜏=0

(69)

(75)

Taking expectations and using Jensen’s inequality, we get   𝑇 𝑇 1 1 ∑︁ 𝜋 1 + E [ℎ𝑢,𝑠 (𝑡)] ≥ E𝜋 𝑇 𝑡 =1 2 𝐷𝑢,𝑠 (𝑇 ) + 1 2   1 1 1 𝜋 + = E 2 𝐷𝑢,𝑠 (𝑇 )/𝑇 + 1/𝑇 2 1 1 1 ≥ + . (70) 2 E𝜋 [𝐷𝑢,𝑠 (𝑇 )]/𝑇 + 1/𝑇 2

The same lower bound applies to the residual interval after the last completion. Consequently,

Define the finite-horizon expected status packet delivery rate as E𝜋 [𝐷𝑢,𝑠 (𝑇 )] 𝜋 . (71) 𝑞𝑢,𝑠 (𝑇 ) = 𝑇 Then (70) can be written as

Completing the square in (76), we obtain

𝑇 ∑︁

1 1 1
+ . E𝜋 [ℎ𝑢,𝑠 (𝑡)] ≥ 𝜋 (𝑇 ) + 1/𝑇 𝑇 𝑡 =1 2 2 𝑞𝑢,𝑠

𝐷𝑢,𝑣 (𝑇 )   1  ∑︁  (𝑚)  2 2 ℎ𝑢,𝑣 (𝑡) ≥  𝐼𝑢,𝑣 + 𝑅𝑢,𝑣 (𝑇 )  2  𝑚=1  𝑡 =1    1 (1) + 𝐿 − 𝑇 − (𝐿 − 1)𝐼𝑢,𝑣 . 2

𝑇 ∑︁

𝑇 ∑︁ 𝑡 =1

(72)

We now consider the video frame flow. Let 𝐷𝑢,𝑣 (𝑇 ) denote the number of completed video frames by the end of slot 𝑇 . If 𝐷𝑢,𝑣 (𝑇 ) = 0, then no video frame is completed over the horizon. Since ℎ𝑢,𝑣 (1) = 1 and the video AoI increases by one in every slot without a frame completion, Í (1/𝑇 ) 𝑇𝑡=1 ℎ𝑢,𝑣 (𝑡) = (𝑇 + 1)/2. Moreover, the RHS of (80) evaluated at 𝐷𝑢,𝑣 (𝑇 ) = 0 is

𝐷𝑢,𝑣 ∑︁(𝑇 )

(𝑚) 𝐼𝑢,𝑣 + 𝑅𝑢,𝑣 (𝑇 ).

  2 𝐷𝑢,𝑣  2 ∑︁(𝑇 )  1   (1) (𝑚)  𝐼 − (𝐿 − 1) + 𝐼 𝑢,𝑣  2  𝑢,𝑣  𝑚=2     2 1 2 (𝐿 − 1) 1 + 𝑅𝑢,𝑣 (𝑇 ) − + 𝐿 − 𝑇 . (77) 2 2 2

By Jensen’s inequality, 

(1) 𝐼𝑢,𝑣 − (𝐿 − 1)

2

+

𝐷𝑢,𝑣 ∑︁(𝑇 ) 

(𝑚) 𝐼𝑢,𝑣

2

2 + 𝑅𝑢,𝑣 (𝑇 )

𝑚=2

 𝐷𝑢,𝑣 ∑︁(𝑇 )  (1) 1 (𝑚) 𝐼𝑢,𝑣 − (𝐿 − 1) + ≥ 𝐼𝑢,𝑣 𝐷𝑢,𝑣 (𝑇 ) + 1  𝑚=2  2 +𝑅𝑢,𝑣 (𝑇 ) .

1 𝑇 +1 (𝑇 − 𝐿 + 1) 2 (𝐿 − 1) 2 − +𝐿− = . (73) 2𝑇 2𝑇 2 2 Thus, (80) holds directly when 𝐷𝑢,𝑣 (𝑇 ) = 0. It therefore suffices to consider 𝐷𝑢,𝑣 (𝑇 ) ≥ 1 in the following derivation. (1) Let 𝐼𝑢,𝑣 denote the time from the beginning of the horizon (𝑚) to the first video-frame completion, and, for 𝑚 ≥ 2, let 𝐼𝑢,𝑣 denote the inter-completion time between the (𝑚 − 1)-th and 𝑚-th completed video frames. Let 𝑅𝑢,𝑣 (𝑇 ) denote the residual time after the last completed video frame. Then 𝑇 =

ℎ𝑢,𝑣 (𝑡) ≥

(76)

(78)

Using (74), we get 

(1) 𝐼𝑢,𝑣 − (𝐿 − 1)

2

+

𝐷𝑢,𝑣 ∑︁(𝑇 ) 

(𝑚) 𝐼𝑢,𝑣

2

2 (𝑇 ) + 𝑅𝑢,𝑣

𝑚=2

(𝑇 − 𝐿 + 1) 2 ≥ . 𝐷𝑢,𝑣 (𝑇 ) + 1

(79)

Substituting (79) into (77) gives (74)

𝑚=1

𝑇

1 ∑︁ (𝐿 − 1) 2 1 (𝑇 − 𝐿 + 1) 2
− ℎ𝑢,𝑣 (𝑡) ≥ + 𝐿 − . (80) 𝑇 𝑡 =1 2𝑇 2 2𝑇 𝐷𝑢,𝑣 (𝑇 ) + 1

Before the first video-frame completion, the video AoI starts at 1 and increases by one in every slot. Hence, over 17

ACM MobiCom ’27, October 2027,

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

𝜋 } 𝑁 , {𝑞 𝜋 } 𝑁 𝜋 and a vector 𝒒𝜋 = {𝑞𝑢,𝑠 𝑢,𝑣 𝑢=1 such that 𝒒 (𝑇𝑛 ) → 𝑢=1 𝜋 𝒒 . Taking 𝑛 → ∞ in (89) shows that this limit point satisfies 𝑁  𝜋 𝜋  ∑︁ 𝑞𝑢,𝑠 𝐿𝑞𝑢,𝑣 ≤ 𝐾. (90) + 𝑝𝑢,𝑠 𝑝𝑢,𝑣 𝑢=1

Taking expectations and applying Jensen’s inequality gives 𝑇

(1 − (𝐿 − 1)/𝑇 ) 2 1 ∑︁ 𝜋
E [ℎ𝑢,𝑣 (𝑡)] ≥ 𝑇 𝑡 =1 2 E𝜋 [𝐷𝑢,𝑣 (𝑇 )]/𝑇 + 1/𝑇 (𝐿 − 1) 2 1 +𝐿− . (81) 2𝑇 2 Define the finite-horizon expected video-frame completion rate as E𝜋 [𝐷𝑢,𝑣 (𝑇 )] 𝜋 𝑞𝑢,𝑣 (𝑇 ) = . (82) 𝑇 Then (81) becomes −

𝑇 ∑︁

Combining (72) and (83), and taking the limit along the subsequence {𝑇𝑛 }, gives 𝑇 𝑁  1 ∑︁ ∑︁ 𝜋  b 𝐽 𝜋 = lim sup E 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡) 𝑇 →∞ 𝑇 𝑡 =1 𝑢=1 𝑇𝑛 ∑︁ 𝑁   1 ∑︁ E𝜋 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡) 𝑛→∞ 𝑇𝑛 𝑡 =1 𝑢=1   ∑︁   𝑁 𝑁 ∑︁ 1 1 1 1 ≥ 𝛼𝑢,𝑠 + + 𝛼 + 𝐿 − . (91) 𝑢,𝑣 𝜋 𝜋 2𝑞𝑢,𝑠 2 2𝑞𝑢,𝑣 2 𝑢=1 𝑢=1

≥ lim inf

2

(1 − (𝐿 − 1)/𝑇 ) 1
E𝜋 [ℎ𝑢,𝑣 (𝑡)] ≥ 𝜋 (𝑇 ) + 1/𝑇 𝑇 𝑡 =1 2 𝑞𝑢,𝑣 (𝐿 − 1) 2 1 +𝐿− . (83) 2𝑇 2 It remains to characterize the finite-horizon delivery-rate vectors. For the status update flow, −

E𝜋 [𝐷𝑢,𝑠 (𝑇 )] =

𝑇 ∑︁

E𝜋 [𝑑𝑢,𝑠 (𝑡)] = 𝑝𝑢,𝑠

𝑡 =1

𝑇 ∑︁

Interpreting 1/0 as +∞, if any component of 𝒒𝜋 is zero, (91) implies b 𝐽 𝜋 = +∞, and hence b 𝐽 𝜋 ≥ 𝐿𝐵★ holds directly. Oth𝜋 erwise, all components of 𝒒 are strictly positive. Together 𝜋 ≤ 𝑝 , 𝑞 𝜋 ≤ 𝑝 /𝐿, this shows that 𝒒𝜋 with (90) and 𝑞𝑢,𝑠 𝑢,𝑠 𝑢,𝑣 𝑢,𝑣 is feasible for the minimization in (23). Therefore, the RHS of (91) is no smaller than 𝐿𝐵★, and hence b 𝐽 𝜋 ≥ 𝐿𝐵★. Since this inequality holds for every admissible policy 𝜋 ∈ Π, we conclude that b 𝐽 ★ ≥ 𝐿𝐵★.

E𝜋 [𝑎𝑢,𝑠 (𝑡)]. (84)

𝑡 =1

Therefore, 𝑇

𝜋 (𝑇 ) 𝑞𝑢,𝑠 1 ∑︁ 𝜋 E [𝑎𝑢,𝑠 (𝑡)] = . 𝑇 𝑡 =1 𝑝𝑢,𝑠

(85)

D

For the video frame flow, 𝑝𝑢,𝑣

𝑇 ∑︁ 𝑡 =1

E𝜋 [𝑎𝑢,𝑣 (𝑡)] =

𝑇 ∑︁

E𝜋 [𝜇𝑢,𝑣 (𝑡)].

In this appendix, we derive the one-slot Lyapunov drift upper bound, the status scheduling weight, and the video scheduling weight given in (27), (29), and (30), respectively. Recall the system state 𝑆 (𝑡), the linear video potential Φ𝑢,𝑣 (𝑡) defined in (24), the Lyapunov function (25), and the one-slot conditional Lyapunov drift (26). Substituting (25) into (26), we obtain

(86)

𝑡 =1

Every completed frame contains exactly 𝐿 successfully delivered packets. Hence, 𝑇 ∑︁

𝜇𝑢,𝑣 (𝑡) = 𝐿𝐷𝑢,𝑣 (𝑇 ) + 𝑂𝑢,𝑣 (𝑇 ),

Upper Bound on the One-Slot Lyapunov Drift

(87)

𝑡 =1

where 0 ≤ 𝑂𝑢,𝑣 (𝑇 ) ≤ 𝐿 − 1 is the number of successfully delivered packets belonging to the unfinished frame at the end of slot 𝑇 . It follows that 𝑇 𝜋 (𝑇 ) 𝐿𝑞𝑢,𝑣 1 ∑︁ 𝜋 E [𝑎𝑢,𝑣 (𝑡)] ≥ . (88) 𝑇 𝑡 =1 𝑝𝑢,𝑣

Δ(𝑆 (𝑡)) =

𝑁 ∑︁

  𝑉𝑢,𝑠 E ℎ𝑢,𝑠 (𝑡 + 1) − ℎ𝑢,𝑠 (𝑡) | 𝑆 (𝑡)

𝑢=1

+

𝑁 ∑︁

  𝑉𝑢,𝑣 E Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) | 𝑆 (𝑡) . (92)

𝑢=1

Status Evolution. The status AoI evolves as

Taking expectations and time averages in (1), and using (85) and (88), gives, for every 𝑇 , 𝑁  𝜋 𝜋 (𝑇 )  ∑︁ 𝑞𝑢,𝑠 (𝑇 ) 𝐿𝑞𝑢,𝑣 + ≤ 𝐾. (89) 𝑝𝑢,𝑠 𝑝𝑢,𝑣 𝑢=1
𝜋 (𝑇 )} 𝑁 , {𝑞 𝜋 (𝑇 )} 𝑁 𝜋 Let 𝒒𝜋 (𝑇 ) = {𝑞𝑢,𝑠 𝑢,𝑣 𝑢=1 𝑢=1 . Since 0 ≤ 𝑞𝑢,𝑠 (𝑇 ) ≤ 𝜋 𝜋 𝑝𝑢,𝑠 and 0 ≤ 𝑞𝑢,𝑣 (𝑇 ) ≤ 𝑝𝑢,𝑣 /𝐿, the sequence {𝒒 (𝑇 )}𝑇 ≥1 is bounded. Therefore, there exists a convergent subsequence indexed by a sequence of horizon lengths {𝑇𝑛 }, with 𝑇𝑛 → ∞,

ℎ𝑢,𝑠 (𝑡 + 1) = ℎ𝑢,𝑠 (𝑡) + 1 − 𝑑𝑢,𝑠 (𝑡)ℎ𝑢,𝑠 (𝑡).

(93)

Taking the conditional expectation and using E[𝑑𝑢,𝑠 (𝑡) | 𝑆 (𝑡)] = 𝑝𝑢,𝑠 E[𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)], we obtain   E ℎ𝑢,𝑠 (𝑡 + 1) − ℎ𝑢,𝑠 (𝑡) | 𝑆 (𝑡) = 1 − 𝑝𝑢,𝑠 ℎ𝑢,𝑠 (𝑡)E[𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)].

(94)

Video Evolution. Recall that Φ𝑢,𝑣 (𝑡) = (ℓ𝑢 (𝑡)/𝐿) (ℎ𝑢,𝑣 (𝑡) − 1) + 𝑔𝑢 (𝑡) + 1. We consider two cases. 18

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027,

Case 1: ℓ𝑢 (𝑡) = 𝐿. Under the generate-at-will model, 𝑔𝑢 (𝑡) = 0 whenever ℓ𝑢 (𝑡) = 𝐿. If the video packet is not successfully delivered, i.e., 𝜇𝑢,𝑣 (𝑡) = 0, then  ℎ𝑢,𝑣 (𝑡 + 1) = ℎ𝑢,𝑣 (𝑡) + 1,     𝑔𝑢 (𝑡 + 1) = 0,    ℓ𝑢 (𝑡 + 1) = 𝐿.  Hence, Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) = 1. (95)

Hence, (100)

ℓ𝑢 (𝑡) 𝜇𝑢,𝑣 (𝑡) − ℎ𝑢,𝑣 (𝑡). 𝐿 𝐿

(101)

Therefore, when ℓ𝑢 (𝑡) < 𝐿, Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) ≤ 1 +

Combining the two cases, for every ℓ𝑢 (𝑡) ∈ {1, . . . , 𝐿}, Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) ≤ 1 +

If the video packet is successfully delivered, i.e., 𝜇𝑢,𝑣 (𝑡) = 1, and 𝐿 > 1, then  ℎ𝑢,𝑣 (𝑡 + 1) = ℎ𝑢,𝑣 (𝑡) + 1,     𝑔𝑢 (𝑡 + 1) = 𝑔𝑢 (𝑡) + 1 = 1,    ℓ𝑢 (𝑡 + 1) = 𝐿 − 1.  Hence, ℎ𝑢,𝑣 (𝑡) Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) = 2 − . (96) 𝐿 For 𝐿 = 1, the successful transmission completes the frame, and   ℎ𝑢,𝑣 (𝑡 + 1) = 𝑔𝑢 (𝑡) + 1 = 1,   𝑔𝑢 (𝑡 + 1) = 0,    ℓ𝑢 (𝑡 + 1) = 𝐿.  Hence, Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) = 1 − ℎ𝑢,𝑣 (𝑡) ≤ 2 − ℎ𝑢,𝑣 (𝑡). Therefore, when ℓ𝑢 (𝑡) = 𝐿, ℓ𝑢 (𝑡) 𝜇𝑢,𝑣 (𝑡) − ℎ𝑢,𝑣 (𝑡). (97) 𝐿 𝐿 Case 2: ℓ𝑢 (𝑡) < 𝐿. If the video packet is not successfully delivered, i.e., 𝜇𝑢,𝑣 (𝑡) = 0, then  ℎ𝑢,𝑣 (𝑡 + 1) = ℎ𝑢,𝑣 (𝑡) + 1,     𝑔𝑢 (𝑡 + 1) = 𝑔𝑢 (𝑡) + 1,    ℓ𝑢 (𝑡 + 1) = ℓ𝑢 (𝑡).  Hence, ℓ𝑢 (𝑡) Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) = 1 + . (98) 𝐿 If 𝜇𝑢,𝑣 (𝑡) = 1 and ℓ𝑢 (𝑡) > 1, then  ℎ𝑢,𝑣 (𝑡 + 1) = ℎ𝑢,𝑣 (𝑡) + 1,     𝑔𝑢 (𝑡 + 1) = 𝑔𝑢 (𝑡) + 1,    ℓ𝑢 (𝑡 + 1) = ℓ𝑢 (𝑡) − 1.  Hence, ℓ𝑢 (𝑡) ℎ𝑢,𝑣 (𝑡) Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) = 1 + − . (99) 𝐿 𝐿 If 𝜇𝑢,𝑣 (𝑡) = 1 and ℓ𝑢 (𝑡) = 1, then the frame is completed and  ℎ𝑢,𝑣 (𝑡 + 1) = 𝑔𝑢 (𝑡) + 1,     𝑔𝑢 (𝑡 + 1) = 0,    ℓ𝑢 (𝑡 + 1) = 𝐿. 

1 ℎ𝑢,𝑣 (𝑡) − . 𝐿 𝐿

Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) =

ℓ𝑢 (𝑡) 𝜇𝑢,𝑣 (𝑡) − ℎ𝑢,𝑣 (𝑡). 𝐿 𝐿

(102)

Taking the conditional expectation and using E[𝜇𝑢,𝑣 (𝑡) | 𝑆 (𝑡)] = 𝑝𝑢,𝑣 E[𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡)], we obtain   E Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) | 𝑆 (𝑡) ℓ𝑢 (𝑡) 𝑝𝑢,𝑣 ≤1+ − ℎ𝑢,𝑣 (𝑡)E[𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡)]. (103) 𝐿 𝐿 Substituting (94) and (103) into (92), we obtain   𝑁 𝑁 ∑︁ ∑︁ ℓ𝑢 (𝑡) Δ(𝑆 (𝑡)) ≤ 𝑉𝑢,𝑠 + 𝑉𝑢,𝑣 1 + 𝐿 𝑢=1 𝑢=1 −

𝑁 ∑︁

  𝑉𝑢,𝑠 𝑝𝑢,𝑠 ℎ𝑢,𝑠 (𝑡)E 𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)

𝑢=1 𝑁 ∑︁   𝑉𝑢,𝑣 𝑝𝑢,𝑣 ℎ𝑢,𝑣 (𝑡)E 𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡) . − 𝐿 𝑢=1

Φ𝑢,𝑣 (𝑡 + 1) − Φ𝑢,𝑣 (𝑡) ≤ 1 +

(104)

Rearranging the terms in (104) yields (27)–(30).

E

Stationary Randomized Policy

This appendix introduces the class of stationary randomized policies used as comparison policies in the proof of Theorem 3.2. Define the set of feasible scheduling actions as ( ) 𝑁 ∑︁
2𝑁 A ≜ 𝒂 ∈ {0, 1} : 𝑎𝑢,𝑠 + 𝑎𝑢,𝑣 ≤ 𝐾 . (105) 𝑢=1

A stationary randomized policy independently selects an action 𝒂(𝑡) ∈ A in every slot according to a fixed probability distribution over A, independently of the current system state and previous scheduling decisions. Let 𝜃𝑢,𝑠 and 𝜃𝑢,𝑣 denote the marginal probabilities of scheduling the status update and video frame flows of UAV 𝑢, respectively: 𝜃𝑢,𝑠 = E[𝑎𝑢,𝑠 (𝑡)],

𝜃𝑢,𝑣 = E[𝑎𝑢,𝑣 (𝑡)].

(106)

These marginal probabilities satisfy 𝑁 ∑︁ 𝑢=1

19

𝜃𝑢,𝑠 + 𝜃𝑢,𝑣 ≤ 𝐾,

0 ≤ 𝜃𝑢,𝑠 , 𝜃𝑢,𝑣 ≤ 1,

∀𝑢.

(107)

ACM MobiCom ’27, October 2027,

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

Conversely, every vector satisfying (107) belongs to the convex hull of A. Therefore, it can be realized by a probability distribution over feasible actions in A. For the resulting stationary randomized policy,

Substituting (114) into (113) gives ΔMW (𝑆 (𝑡)) ≤ 𝐵(𝑡) −

𝑁 ∑︁

LB 𝑊𝑢,𝑠 (𝑡)𝜃𝑢,𝑠 −

𝑢=1

𝑁 ∑︁

LB 𝑊𝑢,𝑣 (𝑡)𝜃𝑢,𝑣 .

𝑢=1

(115) E[𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)] = 𝜃𝑢,𝑠 ,

E[𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡)] = 𝜃𝑢,𝑣 , (108) Using the Max-Weight weights in (29) and (30), together with (112), we obtain

for every state 𝑆 (𝑡). LB , 𝑞 LB } 𝑁 Let {𝑞𝑢,𝑠 𝑢,𝑣 𝑢=1 be an optimal solution of the lowerbound problem (23). Define LB 𝜃𝑢,𝑠 =

LB 𝑞𝑢,𝑠 , 𝑝𝑢,𝑠

LB 𝜃𝑢,𝑣 =

LB 𝐿𝑞𝑢,𝑣 . 𝑝𝑢,𝑣

LB 𝑞𝑢,𝑠 LB = 𝑉𝑢,𝑠 𝑞𝑢,𝑠 ℎ𝑢,𝑠 (𝑡), (116) 𝑝𝑢,𝑠 LB 𝐿𝑞𝑢,𝑣 𝑉𝑢,𝑣 𝑝𝑢,𝑣 LB LB 𝑊𝑢,𝑣 (𝑡)𝜃𝑢,𝑣 = ℎ𝑢,𝑣 (𝑡) = 𝑉𝑢,𝑣 𝑞𝑢,𝑣 ℎ𝑢,𝑣 (𝑡). (117) 𝐿 𝑝𝑢,𝑣 LB 𝑊𝑢,𝑠 (𝑡)𝜃𝑢,𝑠 = 𝑉𝑢,𝑠 𝑝𝑢,𝑠 ℎ𝑢,𝑠 (𝑡)

(109)

The rate constraints in the lower-bound problem imply Therefore, LB 0 < 𝜃𝑢,𝑠 ≤ 1,

LB 0 < 𝜃𝑢,𝑣 ≤ 1,

∀𝑢.

(110) ΔMW (𝑆 (𝑡)) ≤ 𝐵(𝑡) −

and

𝑁 ∑︁

LB 𝑉𝑢,𝑠 𝑞𝑢,𝑠 ℎ𝑢,𝑠 (𝑡)

𝑢=1

! 𝑁  𝑁 LB LB  ∑︁ ∑︁ 𝐿𝑞 𝑞 𝑢,𝑣 𝑢,𝑠 LB LB + ≤ 𝐾. 𝜃𝑢,𝑠 + 𝜃𝑢,𝑣 = 𝑝𝑢,𝑠 𝑝𝑢,𝑣 𝑢=1 𝑢=1

(111)

−

LB 𝑉𝑢,𝑣 𝑞𝑢,𝑣 ℎ𝑢,𝑣 (𝑡).

(118)

𝑢=1

Hence, the lower-bound solution induces a feasible stationary randomized comparison policy with marginal scheduling LB , 𝜃 LB } 𝑁 . probabilities {𝜃𝑢,𝑠 𝑢,𝑣 𝑢=1

F

𝑁 ∑︁

LB and 𝑉 LB LB Choose 𝑉𝑢,𝑠 = 𝛼𝑢,𝑠 /𝑞𝑢,𝑠 𝑢,𝑣 = 𝛼𝑢,𝑣 /𝑞𝑢,𝑣 . Then 𝑉𝑢,𝑠 𝑞𝑢,𝑠 = LB 𝛼𝑢,𝑠 and 𝑉𝑢,𝑣 𝑞𝑢,𝑣 = 𝛼𝑢,𝑣 . Hence,

ΔMW (𝑆 (𝑡)) ≤ 𝐵(𝑡) −

Proof of Theorem 3.2

𝑁 ∑︁

𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡) . (119)

𝑢=1 LB , 𝑞 LB } 𝑁 be an optimal solution of the lower-bound Let {𝑞𝑢,𝑠 𝑢,𝑣 𝑢=1 problem (23). As shown in Appendix E, this solution induces a feasible stationary randomized comparison policy with marginal scheduling probabilities

LB 𝜃𝑢,𝑠 =

LB 𝑞𝑢,𝑠 , 𝑝𝑢,𝑠

LB 𝜃𝑢,𝑣 =

LB 𝐿𝑞𝑢,𝑣 . 𝑝𝑢,𝑣

From (28) and 1 ≤ ℓ𝑢 (𝑡) ≤ 𝐿,   𝑁 𝑁 ∑︁ ∑︁ ℓ𝑢 (𝑡) 𝐵(𝑡) = 𝑉𝑢,𝑠 + 𝑉𝑢,𝑣 1 + 𝐿 𝑢=1 𝑢=1

(112)

≤

𝑁 ∑︁

𝑉𝑢,𝑠 + 2

𝑢=1

From the one-slot drift upper bound (27), under the MaxWeight policy, Δ

MW

(𝑆 (𝑡)) ≤ 𝐵(𝑡) −

𝑁 ∑︁

𝑁 ∑︁

𝑁 ∑︁ 𝛼𝑢,𝑠

𝑁 ∑︁ 𝛼𝑢,𝑣

𝑢=1 𝑞𝑢,𝑠

LB 𝑢=1 𝑞𝑢,𝑣

+2 LB

.

(120)

Substituting (120) into (119) gives ΔMW (𝑆 (𝑡)) +

  𝑊𝑢,𝑠 (𝑡)EMW 𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡)

  𝑊𝑢,𝑣 (𝑡)EMW 𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡) .

𝑉𝑢,𝑣 =

𝑢=1

𝑢=1

−

𝑁 ∑︁

≤ (113)

𝑁 ∑︁ 𝛼𝑢,𝑠 LB 𝑢=1 𝑞𝑢,𝑠

𝑁 ∑︁

𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡)

𝑢=1 𝑁 ∑︁

+2

𝛼𝑢,𝑣 . LB 𝑢=1 𝑞𝑢,𝑣

(121)

𝑢=1

Following (21), let b 𝐽 MW denote the long-term expected surrogate system risk achieved by the Max-Weight policy. Summing (121) over 𝑡 = 1, . . . ,𝑇 , taking expectations, and telescoping the Lyapunov terms yields

Since the Max-Weight policy maximizes the total scheduling weight in every slot, for every state 𝑆 (𝑡), 𝑁 ∑︁

  𝑊𝑢,𝑠 (𝑡)EMW 𝑎𝑢,𝑠 (𝑡) | 𝑆 (𝑡) 𝑇 ∑︁ 𝑁 ∑︁

𝑢=1

+

𝑁 ∑︁

𝑡 =1 𝑢=1 𝑁 ∑︁

  𝑊𝑢,𝑣 (𝑡)EMW 𝑎𝑢,𝑣 (𝑡) | 𝑆 (𝑡)

𝑁 ∑︁ 𝛼𝑢,𝑠 𝛼𝑢,𝑣 ≤𝑇 + 2 LB LB 𝑢=1 𝑞𝑢,𝑠 𝑢=1 𝑞𝑢,𝑣

𝑢=1

≥

𝑁 ∑︁ 𝑢=1

LB 𝑊𝑢,𝑠 (𝑡)𝜃𝑢,𝑠 +

𝑁 ∑︁

LB 𝑊𝑢,𝑣 (𝑡)𝜃𝑢,𝑣 .

  EMW 𝛼𝑢,𝑠 ℎ𝑢,𝑠 (𝑡) + 𝛼𝑢,𝑣 ℎ𝑢,𝑣 (𝑡) !

(114) + EMW [𝐿(𝑆 (1))] − EMW [𝐿(𝑆 (𝑇 + 1))].

𝑢=1

20

(122)

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations

ACM MobiCom ’27, October 2027,

Since 𝐿(𝑆 (𝑇 + 1)) ≥ 0 and EMW [𝐿(𝑆 (1))] < ∞, dividing by 𝑇 and taking the limit superior as 𝑇 → ∞ gives b 𝐽 MW ≤

𝑁 ∑︁ 𝛼𝑢,𝑠

𝑁 ∑︁ 𝛼𝑢,𝑣

𝑢=1 𝑞𝑢,𝑠

LB 𝑢=1 𝑞𝑢,𝑣

+2 LB

(123)

.

(a)

It remains to compare this upper bound with 𝐿𝐵★. From the lower-bound problem, 𝐿𝐵★ =

 𝑁  1 ∑︁ 𝛼𝑢,𝑠 𝛼𝑢,𝑣 + LB LB 2 𝑢=1 𝑞𝑢,𝑠 𝑞𝑢,𝑣 +

𝑁
1 ∑︁ 𝛼𝑢,𝑠 + (2𝐿 − 1)𝛼𝑢,𝑣 . 2 𝑢=1

(124)

Therefore, (b) 𝑁 ∑︁ 𝛼𝑢,𝑠 LB 𝑢=1 𝑞𝑢,𝑠

= 𝐿𝐵★ +

+2

𝑁 ∑︁ 𝛼𝑢,𝑣 LB 𝑢=1 𝑞𝑢,𝑣

𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 + LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣 𝑁

−

1 ∑︁ 𝛼𝑢,𝑠 + (2𝐿 − 1)𝛼𝑢,𝑣 2 𝑢=1

≤ 𝐿𝐵★ +

𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 + , LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣

(125)

where the inequality follows from 𝛼𝑢,𝑠 ≥ 0, 𝛼𝑢,𝑣 ≥ 0, and 𝐿 ≥ 1. Combining (123) and (125), we obtain 𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 b 𝐽 MW ≤ 𝐿𝐵★ + + , LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣

(c)

(126)

which establishes (33). Finally, by Theorem 3.1, 𝐿𝐵★ ≤ b 𝐽 ★. Therefore, 𝐿𝐵★ + 𝜓 MW =

b 𝐽 MW ≤ b 𝐽★

≤

𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 + LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣

b 𝐽★

helipad (landing/docking) Route A Route B Start

𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 𝐿𝐵★ + + LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣

=1+

𝐿𝐵★

Figure 7: Public data used to ground the evaluation scenarios. (a) São Paulo metropolitan region and the two study areas. (b) Barueri/Alphaville delivery corridors, with licensed 5G NR stations from one operator and registered helipads. (c) Interlagos circuit and the map region used in Fig. 3.

𝑁 𝑁 1 ∑︁ 𝛼𝑢,𝑠 3 ∑︁ 𝛼𝑢,𝑣 + LB LB 2 𝑢=1 𝑞𝑢,𝑠 2 𝑢=1 𝑞𝑢,𝑣

.

End gNB (scenario) fig-scenarios window

(127)

𝐿𝐵★

This establishes (34) and completes the proof. 21

ACM MobiCom ’27, October 2027,

G

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

Real-World Data for the Evaluation Scenarios

and aerial activity to construct a controlled scheduling workload. Together, these data ground the scenarios in real locations and documented operating ranges, while the controlled channel, population, and traffic models isolate the scheduling behavior studied in this paper.

The scenarios of Fig. 3 are controlled models constructed over real geographies, rather than replays of measured flights or cellular traces. Figure 7 shows the public infrastructure and operational data used to select the study areas and constrain their geometry. We obtain the coordinates and technology of licensed base stations from Anatel’s public Mosaico/Spectrum-E database [4]. Because the database contains separate records for the emissions and sectors associated with the same station, we deduplicate entries by station number. Figure 7 shows only licensed 5G NR stations from a single operator, TIM. We also obtain registered helipad locations from ANAC. These data establish the presence of cellular and aviation infrastructure around both study areas; they are not used as measurements of aerial coverage. Radio conditions are generated by the channel model described in § 4.1. To identify a suitable experimental gNB for the single-cell scheduling study, we evaluated all licensed 5G NR stations in the region and selected the tower at the highest elevation. Delivery corridors. Since June 2026, Speedbird Aero and iFood have operated an ANAC/DECEA-authorized BVLOS delivery service from Iguatemi Alphaville to nearby residential complexes. The reported operation uses a 3.6-km aerial leg, flown in approximately five minutes at up to 60 m AGL and 50 km/h, carrying payloads of up to 5 kg [19, 34]. Our simulated corridors use 4.0–4.9-km trajectories at 60 m and 15 m/s, preserving the distance, altitude, and speed range of the reported operation. The commercial mall-to-droneport operation motivates the endpoint structure of the delivery scenario. We then construct two routes over the same region with different combinations of radio quality and population exposure. This controlled contrast lets us isolate whether a scheduler follows the channel conditions or the operational consequence of delaying each UAV. Crowded-event scenario. During the Formula 1 São Paulo Grand Prix, DECEA activates Operação Lagos to coordinate aerial activity around Interlagos. In 2025, the operation handled 545 movements over three days, including 427 helicopter movements and 118 authorized drone flights [11, 33]. This observation motivates an evaluation scenario in which UAV traffic and ground exposure are concentrated within a short operating window. Our simulated trajectories follow the 4.3-km Interlagos circuit, while spatially nonuniform spectator regions create different consequences for stale information along the route. The simulation does not replay the reported flights or claim to reproduce the event’s traffic management procedures. Instead, it uses the documented coexistence of dense crowds

H

Evaluation Across the 𝐾 × 𝑁 Grid

Figure 8 reports the individual operating points underlying the load-regime aggregates of Fig. 4. For each policy, the three risk rows show   𝑝 99,𝑢 [𝑅𝑢 ], median𝑡 𝑝 99,𝑢 [𝑅𝑢 (𝑡)] , and max 𝑅𝑢 . 𝑢

These views distinguish persistent per-UAV violations of the operational envelope from sustained instantaneous tails. Only RA remains near the envelope under contention. At 𝐾 = 8 and 𝑁 = 60, every baseline breaches all three risk views: the worst UAV reaches 1.4× the boundary for PF, 2.4× for AoI, and 2.7× for RR, with per-UAV 99th percentiles of 1.40, 2.30, and 2.42. RA produces risk values of 0.83, 1.16, and 0.87 across the three rows: it stays inside the envelope on both per-UAV views and within 1.2× on the sustained instantaneous tail, at an operating point where no baseline satisfies any of the three. The advantage persists under a stricter envelope. Using max𝑢 𝑅𝑢 ≤ 1 as the admission criterion, RA supports 𝑁 = 60 at 𝐾 = 8 and the entire evaluated range up to 𝑁 = 120 at 𝐾 = 16. PF supports 𝑁 = 40 and 𝑁 = 80 at the same channel budgets, while RR and AoI support 𝑁 = 20 and 𝑁 = 60. This worst-UAV frontier coincides with the 𝑝 99 -based frontier at every 𝐾: for RA, the fleet-level guarantee comes at no additional admission cost. At 𝐾 = 32, all policies remain within the envelope throughout the range. Risk reduction does not require uniformly lower AoI. At 𝑁 = 120, RA’s status AoI ranges from 0.9× that of the AoI-optimal policy at 𝐾 ≤ 2 to 1.8× at 𝐾 = 32, while its video AoI ranges from 1.4× at 𝐾 = 32 to 3.6× at 𝐾 = 1: under heavy contention RA spends its transmission opportunities on status updates, whose staleness consumes the risk margin fastest. RA still improves video freshness over the networkcentric baselines under moderate contention: at 𝐾 = 8 and 𝑁 = 60, its mean video AoI is 72 slots, compared with 83 for PF and 84 for RR. RA therefore changes where the available freshness is placed, rather than attempting to minimize every age equally. Scheduling remains bounded by capacity. As the channel budget decreases, every policy eventually exceeds the operational envelope: for 𝐾 ≤ 4, none keeps the worst UAV inside it beyond 𝑁 = 20. At 𝐾 = 1 and 𝑁 = 120, even RA cannot compensate for the service deficit, and the policies exceed the worst-UAV boundary by 52× (RA) to over 900× 22

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations K=1

K=2

ACM MobiCom ’27, October 2027, K=4

K=8

K = 16

K = 32

Status AoI mean h̄1 (slots)

102

102

101 101 101

6 × 100

101 101

4 × 100

101

3 × 100 103 4 × 101

Video AoI mean h̄2 (slots)

103

102

3 × 101

102 2 × 101 102 102 101

101

101 101

mean AoI (h̄1 +h̄2 )/2 (slots)

103

102 2 × 101

102 102 102

101 101 101 101

6 × 100

101 102

103

101

design envelope

100 2 × 100

101

p99 [Rū ]

102 101

100

101

100

6 × 10−1

6 × 10−1

100

4 × 10−1

100

4 × 10−1

100

3 × 10−1

3 × 10−1 103

medt p99 [ru (t)]

102

101

102

design envelope

100

101 101 100

101

6 × 10−1

100 100

100

4 × 10−1

100 3 × 10−1 103

design envelope

100

102 102 101

101

101

6 × 10−1

100

u

maxRū

102

101 100

4 × 10−1

100

100 100

3 × 10−1 5

10

20

40

Simultaneous UAVs N

60

80 100120

5

10

20

40

Simultaneous UAVs N

60

80 100120

5

10

20

40

Simultaneous UAVs N

RR

60

80 100120

PF

AoI

5

10

RA

20

40

Simultaneous UAVs N

60

80 100120

5

10

20

40

Simultaneous UAVs N

60

80 100120

5

10

20

40

Simultaneous UAVs N

60

80 100120

design envelope (=1)

Figure 8: Complete evaluation across all 48 operating points. Columns represent the available uplink channels 𝐾 ∈ {1, 2, 4, 8, 16, 32}, and each panel varies the number of concurrent UAVs 𝑁 ∈ {5, . . . , 120}. The first two rows report fleet-mean status and video AoI, and the third row their combined mean. The remaining rows report the 99th percentile of the per-UAV risk, the median instantaneous tail, and the worst-UAV risk. The dashed line marks the operational envelope 𝑅 = 1.

I

(PF). Conversely, at 𝑁 = 5, the policies converge because contention is limited; only the AoI baseline still breaches the envelope there (1.18 at 𝐾 = 1). Scheduling can reshape the risk tail only within the feasibility region established by the available radio capacity.

SAIL Assignment and Allocation Audit

This appendix supports Section 4.3 by documenting the SAIL assignment, audit metrics, and controlled experiment. SAIL is assigned offline and is never exposed to any scheduler. SAIL assignment. SORA combines the final Ground Risk Class (GRC) and residual Air Risk Class (ARC) to determine SAIL [20, 21]. Ground-risk classification uses the maximum 23

ACM MobiCom ’27, October 2027,

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

population density within the intrinsic GRC footprint, comprising the operational volume and ground-risk buffer. Thus, crossing a crowd affects the classification of the operation, including its low-exposure segments. For the controlled audit, we use the aircraft column corresponding to a maximum characteristic dimension of 3 m and maximum speed of 35 m/s, assume residual ARC-b, and claim no ground-risk mitigation credit. The assumed GRCs are 3 for the low-exposure operation and 6 for the crowdexposed operation, yielding SAIL-II and SAIL-V, respectively. These labels remain fixed during each flight. The per-slot partition in Section 4.3 therefore groups the active UAVs by their operation labels. RA uses the population exposure within its stalenessdependent uncertainty region, as defined in (10). The audit tests whether the resulting allocation aligns with SORA’s operation-level classification. This is a consistency check between two uses of operational exposure; it does not establish regulatory compliance. Audit metrics. We retain the full-occupancy window, excluding fleet fill and drain. The allocation metric is Γsched , defined in Section 4.3: each UAV’s scheduling probability is computed over its observed slots in the corresponding class, and the ratio compares the class means. A value of one denotes equal per-UAV scheduling probability. At the decision level, 𝑃align is the fraction of matched comparisons in which the SAIL-V UAV is served. Candidates must belong to different classes, both be in coverage, differ by at most one slot in status AoI and 3 dB in SNR, and have exactly one UAV scheduled. The reference 𝑃align = 0.5 denotes no preference within these comparisons. We exclude operating points with fewer than 50 matched comparisons. Matching controls status freshness and SNR within the stated tolerances; it does not equate every scheduler state. We interpret this metric as evidence of allocation alignment under these controls. Each retained operating point contributes one sample to the corresponding panel; medians summarize operating points. Controlled scenario. A single gNB serves an low-exposure corridor and a crowd-exposed corridor. Each is flown in both directions, with departures equally divided among the four routes. Links remain in coverage, with SNRs between 12 and 27 dB. The crowd-exposed corridor contains 24 hotspots at eight longitudinal positions, each with a central cluster and two lateral clusters offset by 150 m. The hotspots occupy the central 3.1 km, leaving 600 m and 300 m clear at the two endpoints. Their peak density is 𝜌 max . The experiment uses the main evaluation’s traffic, delivery, and policy parameters. Corridor lengths and channel histories differ; the decision audit therefore compares candidates satisfying the matching conditions above. We exclude 𝑁 < 20, where too few contended decisions are available.

(a) decisions, controlled

1.00

(b) allocation, controlled 5 Γalloc

0.75 Palign

(c) allocation, full grid 6

3

0.50

4

2

3 2

1

0.25

1 RR

PF

AoI

RA

RR median

PF

AoI

RA

RR

PF

AoI

RA

class-blind reference

Figure 9: Allocation audit. Controlled scenario: (a) decision alignment 𝑃align and (b) allocation ratio Γsched . (c) Allocation across the 48 evaluation-grid operating points, including the six with 𝐾 ≥ 𝑁 . Points represent operating points; black bars indicate medians.

Results. In Fig. 9(a), RA serves the SAIL-V candidate in 64% of matched comparisons on median, reaching 95% at the most contended points. RR, PF, and AoI yield medians of 50%, 50%, and 49%, respectively. Thus, RA’s preference persists among candidates with comparable status AoI and SNR. The allocation metric shows the same direction (panel b). RA provides a median 1.8×, and up to 3.1×, the per-UAV scheduling probability to SAIL-V operations. Baseline ratios lie between 0.90 and 0.91; we treat this departure from one as scenario-induced variation. Panel c extends the allocation audit to the evaluation grid reported in Section 4.3. Across the 42 points with 𝑁 > 𝐾, aggregated across both evaluation scenarios, RA’s median ratio is 2.37, compared with 0.92 for each baseline. The six points with 𝐾 ≥ 𝑁 are included here but excluded from Fig. 5, since allocation differences primarily reflect traffic availability in that regime. Together, these results show that RA directs more service toward higher-SAIL operations without receiving their labels, consistent with the allocation behavior discussed in the main text.

J

Sensitivity to Video-Frame Size 𝐿

The analysis in Sections 2 and 3 models video frames with a fixed service requirement 𝐿. Our main evaluation instead uses a realistic workload [8]: one 30-packet I-frame followed by fourteen 3-packet P-frames, yielding a mean frame size of 4.8 packets . To isolate the effect of frame size, we repeat the complete 𝐾 × 𝑁 evaluation with constant 𝐿 ∈ {5, 10, 20, 30}, keeping all other parameters unchanged. Group of Pictures (GOP) versus constant-size frames. Constant 𝐿 = 5 closely matches the GOP mean. Relative to the GOP workload, only 19 of the 192 policy–operating-point pairs change sides of the risk envelope, and none changes the policy ordering. Removing the I-frame bursts lowers median 𝑝 99,𝑢 [𝑅𝑢 ] by 29% for RR and AoI, 22% for PF, and 12% for RA. Thus, GOP burstiness makes the workload more demanding, but does not determine RA’s advantage. 24

Risk-Aware 5G Scheduling for Low-Altitude UAV Operations K = 1, N = 5

K = 2, N = 10

K = 4, N = 20

K = 8, N = 40

K = 16, N = 80

K = 32, N = 120

ACM MobiCom ’27, October 2027,

communication resources can support within the declared envelope. The same interface can also inform infrastructure planning. Expansion, improved aerial coverage, additional sectors, or new base stations can be interpreted through both conventional throughput gains and the additional low-altitude operations they enable within the same communication-risk constraint. Regions that repeatedly approach this boundary therefore identify where network upgrades or operational traffic shaping may provide the greatest benefit. This interpretation is intentionally limited to communicationinduced exposure. Within this scope, RA complements the broader aviation safety case, which also encompasses SORA, airspace constraints, weather assessment, vehicle reliability, and other operational factors. RA expresses communication capacity in a form that traffic-management and safetyassessment processes can consume. Extending this interface to multi-cell systems, handover regions, and city-scale traffic management remains future work.

p99 [Rū ]

101

100

p99 [Rū ]

101

100

N/K = 5

5

10 20 30 frame length L [packets] RR PF

5

10 20 frame length L [packets] AoI RA

30

5

10 20 frame length L [packets] design envelope (=1)

30

Figure 10: Tail risk as the constant video-frame size 𝐿 increases at the design-load operating points 𝑁 /𝐾 = 5. Larger frames increase risk for every policy, but the ranking remains unchanged. RA is the last policy to leave the operational envelope.

L

We implement the slot-level simulator in MATLAB using UAV Toolbox [43] for mobility and 5G Toolbox [42] for radio propagation. Mobility. Each UAV follows a waypointTrajectory along a straight corridor, climbing from a 15 m vertiport to 60 m over 500 m and descending symmetrically. UAVs travel at 15 m/s in the evaluation and 15 m/s in Fig. 3. Departures are uniformly spaced, while traffic-generation phases are randomized. For scalability, the evaluation uses an equivalent closed-form trajectory engine, cross-checked against UAV Toolbox slot by slot. Channel. We use the 3GPP TR 38.901 UMa line-of-sight model (nrPathLoss), with a 3.5 GHz carrier, 20 MHz bandwidth, 30 m gNB, 23 dBm UAV transmit power, 15 dBi receiver gain, and 7 dB noise figure. The evaluation uses deterministic path loss and one serving gNB. The motivating scenario additionally applies 4 dB log-normal fading in blocks of 50 slots, using the same realization across policies. Delivery. Each of the 𝐾 channels carries one packet per slot. A scheduled packet succeeds independently with probability 0.8 when SNR ≥ 5 dB and is lost otherwise. The Shannon rate is used only by PF. We add a fixed 10 ms one-way radio, core, and processing delay.

Sensitivity to 𝐿. Figure 10 shows that increasing 𝐿 from 5 to 30 packets raises median tail risk by 2.6× for RA, 2.9× for PF, 3.6× for AoI, and 1.9× for RR. RR’s smaller relative increase reflects that it is already highly exposed at small 𝐿, rather than greater resilience. RA remains inside the envelope at 𝐿 = 5, with 𝑝 99,𝑢 [𝑅𝑢 ] between 0.37 and 0.60, and reaches the boundary at 𝐿 = 10, with values between 0.74 and 1.00. PF crosses the boundary between 𝐿 = 5 and 𝐿 = 10, while RA crosses only for 𝐿 > 10. Larger updates increase risk for every scheduler because they occupy more transmission opportunities. However, the policy ordering remains unchanged over a 6× increase in 𝐿: RA consistently retains more operational margin and is the last policy to leave the envelope.

K

Simulation Toolbox Parameters

From Radio to Air Capacity

Communication-constrained operational capacity provides a natural interface between the cellular network and airspace traffic management. When projected simultaneous demand approaches the capacity supported by a region, traffic management mechanisms can reshape that demand before communication becomes the binding constraint. Possible actions include delaying departures, adjusting speeds, rerouting selected operations, or deferring additional aircraft until sufficient communication margin becomes available. The network does not authorize an operation; rather, it exposes when expected traffic exceeds the capacity that available 25

Related documents

• Who Can Safely Wait? Risk-Aware 5G Scheduling for Low-Altitude UAV Operation#1216476
• Threat Aware Task Offloading and Caching for Secure UAV Assisted Vehicular Consumer Electronics#463186
• Disclosure and rationality: Comparative risk information and decision-making about prevention#765897
• Joint Random Access and Localization in Cell-Free User-Centric Networks with Frequency-Selective Fading Channels#1000106
• Newcomb's Paradox Under Epistemic Uncertainty: A Two-Stage Policy Analysis#79273