CA3D: Computing Accessibility-Aware Cooperative 3D Deployment of Multiple UAVs

CA3D: Computing Accessibility-Aware Cooperative 3D Deployment of Multiple UAVs

arXiv:2605.17852v1 [cs.NI] 18 May 2026

Yiqin Deng, Zihan Fang, Yijie Wang, Qingxiao Huang, Junhui Gao, Qianyao Ren, and Yuguang Fang link quality, enlarge service coverage, and create more favorable conditions for delay-sensitive task offloading to ground computing nodes. Compared with terrestrial gateways, UAV gateways provide unique advantages in 3D mobility, rapid ondemand deployment, and elevated line-of-sight connectivity. They can be dynamically repositioned between local user hotspots and geographically distributed CNs, making them particularly suitable for bridging local task demand to geographically distributed computing resources when terrestrial gateways are unavailable, congested, or spatially mismatched with the task demand. Motivated by these advantages, recent studies have investigated multi-UAV deployment, trajectory optimization, and user association to improve communication and computation performance in UAV-enabled ground computing networks [4]–[6]. Despite these advances, most existing multi-UAV MEC frameworks assume that candidate computing nodes are prefixed and spatially bounded, typically corresponding to edge servers deployed within the coverage area of users for task computing. Under this paradigm, UAV deployment primarily serves to improve communication links between users and a fixed set of computing nodes. However, this assumption becomes restrictive in emerging computing power networks (CPNs), where heterogeneous computing resources are geographically distributed across large areas, including roadside units, micro data centers, and vehicular computing platforms [7]. In such networks, UAVs act as mobile gateways that dynamically bridge users to dispersed computing resources. As a result, the set of ground computing nodes that can be accessed is not fixed, but instead depends on the UAV deployment and the end-to-end latency constraints. In principle, UAVs can potentially connect users to a much larger pool of computing resources beyond their co-locating immediate coverage area [1], [2]. This observation leads to a fundamentally different perspective for UAV-enabled computing power networks (UAV-CPN). In such networks, UAV deployment does not merely improve communication quality, but also reshapes the network-wide set of computing nodes that can be effectively reached under latency constraints [8]. In particular, the 3D placement of multiple UAVs determines how much geographically distributed computing resources can be made accessible to ground users, while UAV cooperation can expand the unique accessible computing pool and reduce redundant overlap among UAV service regions. As a result, computing accessibility emerges as a key enabler linking multi-UAV deployment to delayconstrained task completion [1], [2].

Abstract—This letter investigates computing-accessibilityaware cooperative 3D deployment of multiple UAVs for task completion enhancement, termed CA3D. We first provide a theoretical analysis showing that computing accessibility is the key mechanism linking UAV deployment to delay-constrained task completion, and that UAV inter-spacing creates a fundamental tradeoff between computing-resource accessibility and task completion. We then develop a cooperative 3D deployment design that jointly balances accessible computing capacity, task completion probability, and redundant UAV overlap. Simulation results under heterogeneous computing node capacities show that CA3D consistently outperforms Random, Fixed, and Greedy deployment baselines under both hotspot and random ground user (GU) distributions. Under the hotspot GU distribution, CA3D achieves nearly full task completion, improving the task completion probability by about 3.3× over Random deployment when the number of UAVs is 8. Under a more challenging random GU distribution, CA3D still achieves about 35% higher task completion probability than the best baseline when the number of UAVs is 12. These results demonstrate that computingaccessibility-aware cooperative 3D deployment improves not only task completion but also robustness to GU distribution changes. Index Terms—multi-UAV-enabled computing power networks, 3D deployment, edge computing, task offloading, computing power accessibility.

I. I NTRODUCTION Unmanned aerial vehicles (UAVs) have emerged as a flexible platform for extending wireless communication and edge computing services in infrastructure-limited environments [1]– [3]. In particular, UAV-enabled ground computing networks, such as UAV-assisted multi-access edge computing (MEC) [4], have attracted significant attention because they can provide on-demand connectivity and task offloading opportunities for ground users when terrestrial infrastructure is unavailable, congested, or temporarily disrupted. By adjusting their positions in three-dimensional (3D) space, UAVs can improve air–ground The work of Yiqin Deng was supported in part by the National Natural Science Foundation of China (Grant No. 62301300) and by the Shandong Provincial Natural Science Foundation (Grant No. ZR2023QF053). The work of Yuguang Fang was supported in part by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11216324). Yiqin Deng is with School of Data Science, Lingnan University, Tuen Mun, Hong Kong, China (email: [email protected]). Zihan Fang, Qingxiao Huang, Junhui Gao, Qianyao Ren, and Yuguang Fang are with Hong Kong JC STEM Lab of Smart City and Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong, China (email: [email protected], [email protected], [email protected], [email protected], [email protected]). Yijie Wang is with Institute of Big Data, Central South University, Changsha, China (email: [email protected]).

1

Motivated by this insight, this letter investigates computingaccessibility-aware cooperative 3D deployment of multiple UAVs for task completion enhancement. We first provide a theoretical analysis showing that UAV inter-spacing offers a fundamental tradeoff between enlarging the unique accessible computing capability and maintaining favorable computing service conditions. Guided by this analysis, we develop a cooperative 3D deployment design that jointly balances computing accessibility, communication link quality, and task completion. Simulation results show that the proposed framework significantly improves both accessible computing capability and task success probability over representative baselines, demonstrating the value of cooperative multi-UAV deployment in transforming spatially distributed computing resources into effective service gains.

of-sight links. For simplicity of analysis, we assume orthogonalized access/backhaul spectrum resources among concurrent UAV-user and UAV-CN transmissions, so that interUAV interference is negligible. This allows us to isolate the impact of cooperative 3D deployment on computing accessibility and delay-constrained task success. The achievable uplink transmission rate
from user k to UAV m is Rk,m = B log2 1 + pk gk,m /σ 2 , where B is the bandwidth, pk is the user transmit power, and σ 2 is the noise power. The up corresponding uplink delay is Dk,m = Lk /Rk,m . Similarly, the UAV–CN forwarding link follows the same propagation model, where the channel gain is denoted by gm,n with distance dm,n . The achievable transmission rate is
Rm,n = B log2 1 + pm gm,n /σ 2 , with pm being the transmit power of UAV m. The corresponding forwarding delay for fwd task k is Dm,n = Lk /Rm,n . If the task of GU k is executed at CN n, the computation decmp lay is approximated as Dk,n = Ck /fn . This node-dependent parameter captures the heterogeneity among different types of CNs. For example, a micro data center may provide a larger fn than a vehicular computing platform, while the available value can also vary with the current background workload. Since the result size is typically much smaller than the task input, the return delay is neglected as in [2]. The end-to-end latency via UAV m and CN n is therefore up cmp fwd Dk,m,n = Dk,m + Dm,n + Dk,n . Note that the UAV 3D position qm = [xm , ym , hm ] affects both the link distances and elevation angles, thereby influencing the achievable transmission rates and the end-to-end latency. For UAV m, we define its accessible computing set as Cm (qm ) = { n ∈ N | ∃k : Dk,m,n ≤ Dkmax }. That is, a ground computing node n belongs to Cm (qm ) if there exists at least one user task that can be completed through UAV m and CN n within its end-to-end latency requirement. Therefore, Cm (qm ) represents the set of ground CNs that are feasible for UAV m under the current 3D deployment. Since Dk,m,n depends on the user–UAV transmission delay, the UAV–CN forwarding delay, and the computation delay Ck /fn , this set is jointly shaped by the UAV position and the node-dependent computing capability fn . A CN with a larger fn may become accessible even when its forwarding link is relatively long, while a low-capacity CN may be infeasible under the same latency deadline. In contrast to conventional MEC, the candidate CNs are not fixed beforehand, but are deployment-dependent and may vary with the UAV’s horizontal location, altitude, and the heterogeneous CN capacities. Here, accessibility is defined from the UAV perspective: a CN is considered accessible to UAV m if it can support at least one user task served by UAV m within the corresponding latency deadline. To characterize the deployment benefit, we define the unique accessible computing capacity as X Ψ(Q) = fn ,

II. S YSTEM M ODEL Consider a multi-UAV-enabled CPN, consisting of a set of ground users (GUs) K = {1, . . . , K}, a set of ground computing nodes (CNs) N = {1, . . . , N }, and a set of UAVs M = {1, . . . , M }. The ground CNs represent heterogeneous distributed computing resources, such as roadside edge servers, micro data centers, and vehicular computing platforms. Different CNs may have different available computing capabilities due to their hardware configurations, service types, and background workloads. We denote by fn the available computing capability of CN n, and define f = [f1 , . . . , fN ] as the CNcapacity vector. In general, fn is node-dependent and is not assumed to be identical across CNs. Unlike conventional multiUAV-assisted MEC systems, where candidate computing nodes are prefixed within a bounded service area, the CNs considered here are geographically distributed over a much larger region and may be located much farther away. Consequently, their accessibility depends critically on the UAV deployment. Let the 3D position of UAV m ∈ M be denoted by qm = [xm , ym , hm ], where (xm , ym ) and hm are its horizontal coordinates and altitude, respectively. Each UAV serves as a mobile gateway that collects tasks from nearby users and forwards them to feasible ground CNs. The task generated by user k is characterized by Γk = (Lk , Ck , Dkmax ), where Lk , Ck , and Dkmax denote the input data size, required CPU cycles, and the maximum tolerable end-to-end latency, respectively. For UAV m ∈ M located at qm = [xm , ym , hm ], let rk,m denote the horizontal distance between user k and UAV m, and let rm,n denote the horizontal distance between UAV mqand CN n. The corresponding link distances are q 2 2 + h2m and dm,n = rm,n + h2m . dk,m = rk,m Following the probabilistic air-ground channel model in [9], LoS the line-of-sight probability for the user–UAV link is Pk,m = 1/1 + a exp(−b(θk,m − a)), where a and b are environmentdependent constants and θk,m = 180/π arctan(hm /rk,m ) is the elevation angle. channel gain h The average large-scale i LoS LoS u is gk,m = β0 d−α P + η(1 − P ) , where β0 is the k,m k,m k,m reference channel gain, αu is the path-loss exponent, and η ∈ (0, 1) denotes the additional attenuation for non-line-

S n∈ m∈M Cm (qm )

where Q = {qm , m ∈ M} denotes the multi-UAV deployment. This metric captures the total non-overlapping comput-

2

Proof. Under the disk approximation, accessible CNs correspond to those located inside the union of the two accessibility disks. Since CNs follow a homogeneous PPP with density λc , the expected number of accessible CNs equals λc A∪ (d). Multiplying this quantity by the average computing capability f¯ gives the result. Theorem 1. For d ∈ [dmin , 2Ra ), E[Ψ(d)] is strictly increasing and strictly concave in d. Fig. 1: Geometric illustration of the overlap and union areas of two UAV accessibility disks. Each disk has radius Ra , and the two disk centers are separated by the horizontal inter-UAV distance d. The shaded region denotes the overlap area A∩ (d) in Eq. (1), while the union of the two disks corresponds to A∪ (d).

Proof. From Lemma 1, E[Ψ(d)] ≈ f¯λc A∪ (d), A∪ (d) = 2πRa2 − A∩ (d). Differentiating A∪ (d) over d ∈ [dmin , 2Ra ) yields p dE[Ψ(d)] = f¯λc 4Ra2 − d2 > 0, dd which implies that E[Ψ(d)] is strictly increasing. Differentiating again gives

ing capacity that becomes accessible under the current UAV deployment. To better understand how UAV spatial separation affects deployment-shaped computing accessibility, we next analyze a simplified two-UAV special case.

d2 E[Ψ(d)] d = −f¯λc p < 0, 2 dd 4Ra2 − d2 which shows that E[Ψ(d)] is strictly concave. For d ≥ 2Ra , the two disks do not overlap and A∪ (d) = 2πRa2 becomes constant. Hence, E[Ψ(d)] saturates for d ≥ 2Ra .

III. A T WO -UAV C ASE A NALYSIS To gain analytical insight into the impact of UAV spatial deployment on computing accessibility, we consider a simplified two-UAV scenario. Suppose two UAVs are deployed at the same altitude h with horizontal coordinates q1 = (−d/2, 0, h) and q2 = (d/2, 0, h), where d denotes the horizontal interUAV distance between the two UAVs. To satisfy the safetydistance requirement, the feasible separation must satisfy d ≥ dmin . For analytical tractability, we approximate the accessible computing set of each UAV by an effective accessibility disk with radius Ra (h) on the ground plane. Here, Ra (h) denotes the maximum horizontal distance from the UAV projection within which a representative task can still satisfy the end-toend latency constraint. This disk approximation is introduced only for geometric analysis of the two-UAV case and does not replace the exact accessibility definition in the general multi-UAV model. As in [2], we assume that ground CNs are distributed according to a homogeneous Poisson Point Process (PPP) with density λc , and the average computing capability of each CN is f¯. Let A∩ (d) and A∪ (d) denote the overlap area and union area of the two accessibility disks, respectively. As illustrated in Fig. 1, the two disks have the same radius Ra , and their centers are separated by the horizontal inter-UAV distance d. The shaded lens-shaped region corresponds to the duplicated accessible region A∩ (d), while the union of the two disks corresponds to A∪ (d). Then, for dmin ≤ d < 2Ra , the overlap area is   d dp 2 A∩ (d) = 2Ra2 cos−1 4Ra − d2 , − (1) 2Ra 2

Corollary 1. If the deployment objective is simplified to maximizing only the unique accessible computing capacity, while ignoring the communication-performance penalty, then the optimal UAV separation is d⋆ = max(dmin , 2Ra ). Theorem 1 and Corollary 1 reveal that enlarging the UAV separation always increases the unique accessible computing capacity until the two accessibility disks no longer overlap. However, a larger separation also increases the average user– UAV distance, which tends to enlarge the uplink delay and may reduce the task completion probability. Therefore, when both computing accessibility and communication performance are considered, the optimal separation generally lies in the interval [dmin , 2Ra ]. This tradeoff provides the key motivation for the proposed cooperative computing accessibility-aware deployment design, which aims to enlarge the unique accessible computing set while suppressing excessive overlap and maintaining favorable communication conditions. IV. CA3D: C OMPUTING -ACCESSIBILITY-AWARE C OOPERATIVE 3D D EPLOYMENT The above analysis shows that cooperative UAV deployment should enlarge the unique accessible computing set while maintaining favorable service conditions for delay-constrained tasks. Excessive UAV overlap causes redundant accessibility to the same CNs, whereas overly dispersed deployment may degrade user access quality. Motivated by this tradeoff, we formulate a computing-accessibility-aware cooperative 3D deployment problem. For a given deployment Q, we evaluate task completion using a task-level UAV-CN pair selection rule. For GU k, let Ak (Q) denote the feasible UAV-CN pair set under the current deployment. Each feasible pair (m, n) ∈ Ak (Q) has end-toup cmp fwd end latency Dk,m,n = Dk,m +Dm,n +Dk,n , and GU k selects

and the union area is A∪ (d) = 2πRa2 − A∩ (d). For d ≥ 2Ra , the two disks do not overlap and thus A∪ (d) = 2πRa2 . Lemma 1. Under the disk approximation, the expected unique accessible computing capacity is E[Ψ(d)] ≈ f¯λc A∪ (d).

3

(m⋆k , n⋆k ) = arg min(m,n)∈Ak (Q) Dk,m,n . If Ak (Q) = ∅, the task is unsuccessful; otherwise, the task success indicator is ( 1, Dk,m⋆k ,n⋆k ≤ Dkmax , 1k (Q) = 0, otherwise. P K 1 Thus, Psucc (Q) = K k=1 1k (Q). This rule is used to evaluate a given deployment rather than introducing the association as an additional optimization variable. To reduce scale imbalance among different utility terms, we use a scale-calibrated deployment utility. The accessible computing capacity Ψ(Q) and the overlap penalty Ω(Q) are evaluated in GHz, while Psucc (Q) ∈ [0, 1] denotes the task success probability. For two UAVs m and m′ ,P the duplicated accessible computing capacity is Om,m′ (Q) = n∈Cm (qm )∩Cm′ (qm′ ) fn . The average pairwise overlap penalty is defined as X 2 Om,m′ (Q), Ω(Q) = M (M − 1) ′

8VHUV ,QDFFHVVLEOH&1V $FFHVVLEOH&1V ,QLWLDO8$9V



s.t.

hmin ≤ hm ≤ hmax ,

∀m ∈ M,

∥qm − qm′ ∥ ≥ dmin ,

∀m ̸= m′ ,

Q



[P  

 









 \P

Fig. 2: Initial random deployment with three UAVs. 8VHUV ,QDFFHVVLEOH&1V $FFHVVLEOH&1V 2SWLPL]HG8$9V

P P P



for M > 1, and Ω(Q) = 0 when M = 1. The cooperative 3D deployment problem is formulated as F (Q) = αΨ(Q) + βPsucc (Q) − γΩ(Q)

P P

m<m

max

P



 [P  

 







\P



Fig. 3: CA3D-optimized deployment with three UAVs.

(2) Ib M HW Aloc )CF ), where Np , Ig , Ib , H, W , and Aloc denote the number of particles, global-search iterations, beam-search refinement iterations, rollout horizon, beam width, and localaction size, respectively.

where α, β, and γ are scale-calibration and preference coefficients for accessible computing expansion, task completion, and overlap reduction, respectively. The problem is highly non-convex because the accessible computing sets are discrete and deployment-dependent, while Ψ(Q) and Ω(Q) couple the deployments of multiple UAVs. We therefore develop CA3D, a two-stage heuristic algorithm combining global initialization and local refinement. In the first stage, CA3D performs PSO-based global initialization over the UAV deployment matrix Q. Each particle represents a candidate UAV layout and is updated using the standard PSO rule with inertia weight ω, cognitive/social coefficients c1 , c2 , and projection ΠQ (·) onto the feasible deployment set. The best particle after global search is used to initialize the second stage. In the second stage, CA3D refines the deployment by beam-searchbased local updates. For UAV m, the local action set is Am = {(±∆x, 0, 0), (0, ±∆y, 0), (0, 0, ±∆h), (0, 0, 0)}. For a candidate action sequence am,1:H , the discounted
rollout PH score is Jm (am,1:H ) = ℓ=1 ρℓ−1 F (Q(ℓ) ) − F (Q) , where H is the rollout horizon, ρ is the discount factor, and Q(ℓ) is the rollout deployment. Only the best W partial sequences are retained at each depth, and the first action of the highestscoring sequence is executed. This process is repeated for all UAVs until convergence. A concise complexity analysis is as follows. The complexity of one deployment-utility evaluation is CF = O(KM N + M 2 N ), where the two terms correspond to GU-UAV-CN latency checking and pairwise overlap evaluation, respectively. Therefore, the overall complexity of CA3D is O((Np Ig +

V. P ERFORMANCE E VALUATION In this section, we evaluate the performance of the proposed computing-accessibility-aware cooperative UAV deployment framework and verify the theoretical insights. We consider a 4 km × 4 km district-level service area, such as a campus, traffic hub, industrial park, or emergency service region, where local task demand may need to be connected to geographically distributed CNs [10]. Unless otherwise stated, GUs are uniformly distributed within a centered hotspot disk of radius 800 m, while CNs are randomly distributed over the entire area. This hotspot setting models locally concentrated task requests with insufficient or congested nearby terrestrial computing resources, which motivates the use of UAVs to expand the accessible computing resource pool. The UAV altitude is constrained within [100, 300] m, and the bandwidth is 10 MHz. Each task has an input size of 5 MB, a computation requirement of 1 × 109 CPU cycles, and a deadline of 1 s. The available computing capability of each CN is independently generated as fn ∼ U (2, 10) GHz. For the scale-calibrated utility in Eq. (2), we set α = 0.2, β = 1.5 × 104 , and γ = 0.1. Here, Ψ(Q) and Ω(Q) are evaluated in GHz, and Psucc (Q) ∈ [0, 1]. Fig. 2 and Fig. 3 show a representative random initial deployment and the corresponding CA3D-optimized deployment with three UAVs. Compared with the initial layout, CA3D relocates UAVs to better connect the hotspot GU region with

4

(a) Accessible computing capacity.

(b) Task completion probability.

(a) Hotspot GU distribution.

(b) Random GU distribution.

Fig. 4: Impact of horizontal spacing between two UAVs on performance.

Fig. 5: Impact of GU distribution on task completion vs. number of UAVs.

distributed CNs, making more remote CNs accessible and reducing redundant service overlap. Fig. 4(a) shows that the accessible computing capability Ψ increases with the horizontal inter-UAV distance and gradually saturates as the overlap between the two accessibility disks decreases, which is consistent with Theorem 1. A larger altitude generally leads to a larger Ψ because it enlarges the effective accessibility radius. In contrast, Fig. 4(b) shows that Psucc is non-monotonic with respect to UAV spacing. It first improves as two UAVs become more complementary in serving users and accessing distributed CNs, but then decreases when the spacing becomes too large due to degraded user–UAV link access quality. The maximum task completion probability is achieved at a moderate spacing of around 700 m, confirming the tradeoff between expanding accessible computing resources and maintaining favorable service conditions. Fig. 5 compares the task success probability Psucc versus the number of UAVs under hotspot and random GU distributions for CA3D and three baselines: Random [4], Fixed [11], and Greedy. Specifically, Random randomly deploys multiple UAVs without cooperative optimization; Fixed restricts the candidate accessible CNs to a fixed local region with radius 200 m; and Greedy uses the same local action set as CA3D, but removes the PSO-based global initialization and multi-step beam-search rollout, and greedily selects the one-step movement with the largest immediate improvement in F (Q). Under the hotspot distribution in Fig. 5(a), CA3D rapidly approaches nearly full task completion as M increases. When M = 8, CA3D achieves Psucc ≈ 1.0, compared with about 0.83, 0.30, and 0.04 for Greedy, Random, and Fixed, respectively. Under the random-GU distribution in Fig. 5(b), the enlarged service region makes the problem more challenging, but CA3D still achieves the best performance. When M = 12, CA3D reaches Psucc ≈ 0.70, while Greedy and Random stay around 0.35– 0.40, and Fixed remains close to zero. The consistent gains over all three baselines under both GU distributions demonstrate the robustness of CA3D, which jointly improves task success, expands accessible computing capacity, and reduces redundant UAV overlap.

ing systems. We have shown that UAV deployment reshapes both communication quality and the reachability to distributed CNs under latency constraints. The proposed CA3D design has jointly balanced accessible computing capacity, task completion probability, and redundant UAV overlap. Simulation results under heterogeneous CN capacities and different GU distributions have confirmed that CA3D consistently outperforms representative baselines, demonstrating the value of computing accessibility as a design principle for multi-UAV ground computing systems. R EFERENCES [1] Y. Deng, H. Zhang, X. Chen, and Y. Fang, “UAV-assisted MEC with an expandable computing resource pool: Rethinking the UAV deployment,” IEEE Wirel. Commun., pp. 31(5):110–116, Oct. 2024. [2] Y. Deng, Z. Fang, S. Hu, Y. Ma, X. Guo, H. Zhang, and Y. Fang, “UAVenabled computing power networks: Design and performance analysis under energy constraints,” IEEE Trans. Mob. Comput., Jan. 2026, early access. [3] Y. Deng, H. Zhang, X. Chen, and Y. Fang, “UAV-assisted multi-access edge computing with altitude-dependent computing power,” IEEE Trans. Wireless Commun., 23(8): 9404–9418, Aug. 2024. [4] S. Han, X. Liu, M. Zhou, K. Zhu, L. Zhao, A. Albeshri, and A. Abusorrah, “Joint association, deployment and flight trajectory optimization for multi-UAV-enabled large-scale mobile edge computing,” IEEE Trans. Mob. Comput., pp. 23(12):13 207–13 221, Dec. 2024. [5] Q. Zeng, Y. Jia, C. Li, and L. Liu, “3-D deployment of UAVBSs for effective communication coverage,” IEEE Internet Things J., 11(14):25162-25172, July 2024. [6] X. Zhu, L. Zhai, N. Li, Y. Li, and F. Yang, “Multi-objective deployment optimization of UAVs for energy-efficient wireless coverage,” IEEE Trans. Commun., 72(6):3587-3601, June 2024. [7] X. Chen, Y. Deng, H. Ding, G. Qu, H. Zhang, P. Li, and Y. Fang, “Vehicle as a Service (VaaS): Leverage vehicles to build service networks and capabilities for smart cities,” IEEE Commun. Surv. Tutor., Feb. 2024, 42(3): 2048–2081. [8] Y. Deng, Z. Fang, S. Hu, Y. Ma, H. Zhang, and Y. Fang, “UAV-enabled computing power networks: Task completion probability analysis,” in IEEE GLOBECOM, Taipei, Taiwan, Dec. 8-12, 2025. [9] A. A. Khuwaja, Y. Chen, N. Zhao, M.-S. Alouini, and P. Dobbins, “A survey of channel modeling for UAV communications,” IEEE Commun. Surv. Tutor., pp. 20(4):2804–2821, Fourthquarter 2018. [10] G. Sun, L. He, Z. Sun, Q. Wu, S. Liang, J. Li, D. Niyato, and V. C. M. Leung, “Joint task offloading and resource allocation in aerial-terrestrial UAV networks with edge and fog computing for post-disaster rescue,” IEEE Trans. Mob. Comput., 23(9): 8582-8600, Sept. 2024. [11] F. Lu, G. Liu, W. Lu, Y. Gao, J. Cao, N. Zhao, and A. Nallanathan, “Resource and trajectory optimization for UAV-relay-assisted secure maritime MEC,” IEEE Trans. Commun., 72(3): 1641–1652, Mar. 2024.

VI. C ONCLUSION This letter has studied computing-accessibility-aware cooperative 3D deployment of multiple UAVs for ground comput-

5