1
Co-planning of Flight Corridors and Communication Infrastructure for Urban Drone Logistics Networks
arXiv:2607.23989v1 [cs.RO] 27 Jul 2026
Yingjie He, Yikang Wang, Zhenyu Gao
Abstract—Reliable wireless connectivity is essential for urban air mobility (UAM) networks in dense urban environments. It is therefore imperative to carefully plan the supporting communication infrastructure for UAM flight corridors. Most existing works optimize communication infrastructure and UAV flight paths independently, often leading to unnecessary base station (BS) deployment or excessive flight detours. This paper studies the joint optimization of BS deployment and UAV flight corridors in complex urban environments, aiming to minimize both infrastructure investment and flight distance while satisfying communication quality constraints. We propose CR-CMAB, a channel reciprocity-guided combinatorial multi-armed bandit framework. The framework constructs high-fidelity radio maps using 3D ray tracing, selects BS combinations via coverage-aware CMAB search, and dynamically expands the search space by identifying promising BS locations through channel reciprocity. Experimental results from a detailed case study demonstrate that CR-CMAB outperforms baseline methods with moderate computational time, yielding more strategically positioned BSs and shorter flight corridors. This study offers a practical planning perspective for cost-effective and communication-reliable UAM deployment in future smart cities. Note to Practitioners—Communication systems are among the most prominent constraints for urban air mobility (UAM) operations. To deploy communication infrastructures for UAM, planners must balance two conflicting objectives in the process: minimizing cost while maximizing efficiency. System operators therefore have a strong interest in strategically planning communication infrastructure along UAM flight corridors. To better tackle this challenge, we propose the joint optimization of UAM base stations and flight corridors. Our proposed optimization framework leverages several rigorous components, including high-precision simulation, a combinatorial multi-armed bandit algorithm, and the principle of reciprocity, to collaboratively plan the locations of base stations and the configuration of a network of flight corridors. We validate our approach through a case study in a digital 3D urban environment, and the results show that our approach achieves the best efficiency in terms of total flight distance at each specific infrastructure cost level. We believe this work paves the way for further advancements leading to optimal design schemes for UAM communication infrastructure. Index Terms—Urban air mobility, communication infrastructure, path planning, multi-armed bandit, channel reciprocity.
I. I NTRODUCTION Urban Air Mobility (UAM) is an emerging concept that enables efficient transportation within urban environment using unmanned aerial vehicles (UAVs) and electric vertical takeoff and landing (eVTOL) aircraft. By exploiting lowaltitude urban airspace, UAM supports a broad range of applications, including passenger transit, cargo delivery, and urban monitoring, offering a promising solution to alleviate terrestrial traffic congestion and enhance the sustainability and resilience of smart cities [1]–[3].
It is envisioned that UAM vehicles will operate within a service network, where vertiports and waypoints function as nodes and flight corridors function as links [4]. Flight corridors are designated virtual highways in the sky that aircraft are required to follow, offering navigation, reduced airspace complexity, and improved operational safety and efficiency [5]. However, maintaining safe and sustainable aerial vehicle operations within flight corridors creates substantial infrastructure challenges. Communication is an essential component of the necessary supporting infrastructure. Existing citywide communication infrastructure falls short of meeting the connectivity requirements of an aerial vehicle network in urban low-altitude airspace. In addition, the dense concentration of high-rise buildings creates severe urban canyon effects, leading to critical line-of-sight (LoS) blockages and multipath fading [6]. Such a heavily obstructed radio environment significantly degrades the reliability of wireless communication links [7]. A fundamental requirement for UAM safety is therefore the strategic deployment of reliable communication infrastructure, specifically base stations (BSs), to support networked flight corridor operations. Likewise, to prevent communication outages, flight corridors must be planned to strictly satisfy communication quality constraints, ensuring continuous radio coverage throughout the flight trajectories [8], [9]. Recent studies have primarily focused on either optimizing flight paths under fixed BS deployments or placing BSs for predetermined routes [10], [11]. These decoupled strategies neglect the strong coupling between UAV mobility and radio coverage, often leaving flight paths trapped in communication blind spots or leading to inefficient use of BS resources. Hence, joint optimization of BS placement and flight corridors is essential to effectively balance the objectives of both UAM service providers and wireless network operators. While a co-planning approach is theoretically ideal, realizing it in complex 3D urban environments imposes a high computational burden. The search space of candidate BS locations and 3D flight paths grows exponentially, rendering conventional combinatorial optimization algorithms highly inefficient and prone to suboptimal solutions. To address this problem, we propose a novel Channel Reciprocity-guided Combinatorial Multi-Armed Bandit (CR-CMAB) framework for this joint optimization task. The main contributions of this work are summarized as follows: • Joint optimization approach within a digital twin environment. We formulate the joint optimization of BS deployment and UAM flight corridors within a radio frequency simulation environment. By constructing highfidelity radio maps, this environment captures complex urban radio propagation dynamics, establishing a physical
2
foundation for the subsequent search algorithm. Coverage-aware CMAB BS selection. We formulate BS selection as a coverage-aware combinatorial multiarmed bandit (CMAB) problem, where each candidate BS is an arm and each selected subset forms a super arm. The selection strategy combines upper confidence bound (UCB) exploration with waypoint coverage rewards, encouraging complementary BS selections that jointly support the flight corridor while efficiently exploring the combinatorial space. • Reciprocity-driven dynamic space expansion. We propose a dynamic action space expansion mechanism grounded in the principle of reciprocity. Using temporary UAV paths as radio frequency probes, this mechanism infers promising BS locations and injects them into the candidate pool. This feedback loop intelligently expands the search space to include higher-quality BS locations. •
II. R ELATED W ORKS A. Base Station Deployment BS deployment is a fundamental problem that determines the number and locations of BS sites under coverage, capacity, quality-of-service (QoS), and other constraints [12]. At its core, BS deployment can be viewed as a combinatorial optimization problem, since it requires selecting a subset of BSs from a large pool of candidate sites. In dense urban environments, the difficulty increases further because blockage makes the feasible deployment space tightly coupled with the surrounding built environment. A major line of work formulates BS deployment as a combinatorial optimization problem. Such problems are commonly solved by discretizing the deployment region and searching over candidate BS locations using heuristic or metaheuristic algorithms. Recent explorations include swarm intelligence approaches such as grey wolf optimization for BS deployment in Internet-of-Things networks [13], the random-walk sparrow search algorithm for 5G private network planning [14], and the dynamically adjusted quantum genetic algorithm for 5G BS selection [15]. These heuristic methods are effective, but they often require extensive search and repeated communication quality evaluation in complex urban environments. Learning-based methods have recently emerged as an alternative to the aforementioned metaheuristics. A related work uses reinforcement learning (RL) with digital network twins to learn deployment policies from pathloss predictions [16]. Multi-objective deep reinforcement learning has also been used to place BSs while balancing coverage and localization requirements for urban transportation scenarios [17]. However, learning-based methods often require substantial training time and may exhibit limited generalization when transferred across different scenarios. Cost is a dominant factor in BS deployment. Wang et al. adopted geographic information systems (GIS) and heuristic optimization to maximize service coverage while reducing construction and operation cost [18]. Prasad et al. jointly optimized BS location, BS density, and transmit power to minimize network operational cost while satisfying user coverage
constraints [19]. Chiaraviglio et al. formulated 5G network planning under electromagnetic field constraints as an NP-hard optimization problem, where the objective balances service quality and the installation cost of next-generation NodeBs (gNBs) [20]. Together, these works show that BS deployment is not only a coverage problem but also a cost-constrained planning problem. For UAV services, BS deployment should be evaluated along the intended air corridor rather than only over a fixed planar service region. Lin et al. proposed a low-altitude communication network planning method for air route coverage, where existing ground BSs are selected for upgrade to support aerial links along predefined routes [21]. Kabashkin et al. studied communication infrastructure design for UAV operations in air mobility corridors and analyzed repeater deployment strategies [22]. These studies highlight the importance of corridor-aware communication supply for UAVs, but they still assume that the flight corridor is given in advance. As a result, the deployment strategy may become inefficient when the route passes through regions with poor communication coverage, leading to unnecessary additional BS installations.
B. Communication-Aware UAV Path Planning Conventional UAV path planning problems mainly consider path efficiency, obstacle avoidance, and collision prevention, while ignoring communication conditions along the route [23]. However, insufficient communication support may degrade command and control (C2) link reliability and affect UAV operational safety. Communication-aware UAV path planning addresses this limitation by incorporating wireless link quality metrics (e.g., received signal strength, outage probability, and latency) into trajectory design. This allows the planned trajectory to maintain more reliable connectivity throughout the flight. A common approach involves constructing a radio coverage map and planning the UAV trajectory over this communication-aware representation. Song et al. used partial channel knowledge maps (CKMs) to optimize trajectories for cellular-connected UAVs in complex environments [24]. Chen et al. studied radio-map-based trajectory design for a UAV cargo pickup and delivery system, where the UAV route is optimized to satisfy communication requirements throughout the mission [25]. Other works use cellular coverage maps to guide UAVs toward regions with better communication reliability, using graph search algorithms or RL to trade path length for link quality [26], [27]. These studies show that communication constraints can significantly reshape UAV trajectories. However, they usually assume that the BS deployment or the radio map is already given. As a result, the UAV can only adapt its path within a fixed communication environment. When the existing BS deployment creates severe communication blind spots, pathonly optimization may lead to long detours or even infeasible routes. This limitation highlights the need to jointly consider BS deployment and flight corridor design, rather than optimizing either component in isolation.
3
C. The Research Gap: Joint Optimization of BS deployment and UAV Network Corridors BS deployment studies usually focus on selecting communication infrastructure for given service regions, while communication-aware UAV path planning studies typically optimize trajectories under a fixed radio environment. Such separate treatments may lead to suboptimal solutions, particularly for UAV network operations, since the selected BSs determine where reliable communication is available, while the flight corridor determines where communication service is actually required. Jointly optimizing BS deployment and flight corridor design can therefore reduce unnecessary infrastructure deployment and avoid excessive trajectory detours. However, studies that explicitly couple these two components for urban UAV networks remain limited. Huang et al. proposed a heterogeneous dual-network framework for emergency delivery UAVs, where a communication support network is coordinated with a delivery path network [28]. Although the framework considers both communication assurance and path planning, its solution is decomposed into sequential modules, where the path planning stage mainly adapts to a pre-established communication support network, rather than being optimized simultaneously with BS deployment. Li et al. considered the concept of joint air corridor planning and BS deployment for low-altitude ISAC networks with a channel map [29]. However, their work does not account for network operations, and the evaluation scenario remains relatively simplistic, overlooking the full propagation complexity of realistic urban environments. To address these limitations, this paper studies joint BS deployment and networked UAV flight corridor planning in complex urban environments. The proposed framework builds accurate radio maps using ray tracing, selects BSs through a coverage-aware CMAB search, and leverages channel reciprocity to expand the BS search space by identifying new candidate locations with strong potential to fill corridor coverage gaps. It also evaluates infrastructure and trajectory objectives through interpretable BS deployment and UAV operation costs, enabling a more practical tradeoff between infrastructure expenditure and flight efficiency. III. M ETHODOLOGY A. Proposed Framework We consider the joint optimization of flight corridors and communication infrastructure for a UAV network in a complex urban environment. The general form of this multi-objective optimization problem can be formulated as follows: min L(π), C(z) z,π
s.t.
hobs (π) = 0, Qz (s) ≥ γth ,
(1) ∀s ∈ π,
z ∈ {0, 1}M , π ∈ Π, where the objective vector aims to simultaneously minimize the length of the total flight paths, denoted by L(π), and the total infrastructure cost of the deployed base stations, C(z).
Here, z ∈ {0, 1}M is a binary decision vector indicating the deployment status across M candidate base station locations. π denotes the set of networked UAV flight paths confined to a designated cruising altitude, selected from Π, which represents the universal set of all feasible paths. The optimal joint solution (z∗ , π ∗ ) yielding the Pareto frontier must strictly adhere to both physical and communication boundaries. Specifically, the physical constraint hobs (π) prevents the UAV from colliding with surrounding buildings. The connectivity constraint dictates that the spatial signal quality, Qz (s), must satisfy or exceed the minimum threshold γth for every location s along π. This guarantees that the UAV maintains a continuous and reliable radio link throughout the entire flight corridor. The challenge of this co-planning problem lies in the exponential growth of the search space. The combinatorial complexity of the problem is further exacerbated by the communication constraints: in dense urban scenarios, calculating the exact signal quality Qz (s) for every (z, s) pair requires radio propagation assessments, which creates a major computational bottleneck. To tackle this problem, we propose CR-CMAB, a novel framework that integrates coverage-aware CMAB search with reciprocity-driven space expansion, as illustrated in Fig. 1. The core workflow consists of four modules. First, the ray tracing method generates the radio maps of candidate BSs. Then, the coverage-aware CMAB module selects a BS combination under the given budget. The selected candidates are evaluated by the two-step trajectory evaluation module, where the connected component labeling (CCL) checks corridor connectivity and A* computes the feasible path. The evaluation result is fed back to update the CMAB scores. After every fixed number of iterations, reciprocity-driven space expansion adds new promising BS candidates to the pool, allowing the search space to be refined progressively. B. Radio Propagation Modeling via Ray Tracing To quantify the communication capability of each candidate BS, it is necessary to construct high-precision radio maps, which are essentially spatial heatmaps representing the geographical distribution of the signal-to-noise (SNR) ratio. By capturing the electromagnetic propagation characteristics over the urban topology, radio maps provide a deterministic communication metric for any potential flight. Existing methods for constructing radio maps generally fall into three categories: empirical models, data-driven methods, and ray tracing. Empirical models, such as the log-distance path loss model (LDPL), are computationally efficient but fail to capture blockage and multipath effects in realistic propagation environments. On the other hand, emerging datadriven methods (e.g., RadioUNet [30] and PMNet [31]) require extensive training data and often show limited robustness across different urban layouts. Consequently, we adopt a deterministic 3D ray tracing approach. Based on geometric optics, this method captures complex electromagnetic propagation mechanisms, including penetrations, reflections, and diffractions, providing high-fidelity radio maps without the need for prior training datasets.
4
Urban Environment
Radio Propagation Modeling Candidate radio maps
Coverage-Aware Selected CMAB Search candidates
Two-step Trajectory Evaluation If CCL connected
UAV Vertiports
Updating scores
Initial BS Candidates
Updated BS Candidates
Comparison Analysis
A* Algorithm
Every N Iterations
Reciprocity -driven Space Expansion
SNR Sensitivity Analysis
Optimal Selection
Fig. 1: The CR-CMAB framework for joint optimization of BS deployment and networked flight corridors.
Let bi denote the location of the i-th candidate BS, and let s = (x, y) represent an arbitrary coordinate on the horizontal flight plane. The channel propagation gain from bi to s, calculated by ray tracing, is denoted by Lch (bi , s). To construct the SNR radio maps defined previously, the received signal power at location s from the candidate BS at bi is calculated using the link budget equation: P (bi , s) = Pt + Gt + Lch (bi , s) + Gr ,
(2)
where Pt is the BS transmitter output power, and Gt and Gr are the transmitter and receiver antenna gains, respectively. Then, the noise power Pn at the receiver is modeled as: Pn = N0 + 10 log10 (B) + NF ,
(3)
where N0 represents the thermal noise power spectral density, B is the channel bandwidth, and NF is the hardware noise figure. With both the received signal power and noise derived, the single-BS SNR at s is formulated as: qi (s) = P (bi , s) − Pn = Pt + Gt + Gr + Lch (bi , s) − Pn .
(4)
When multiple BSs exist in the same airspace, we assume that the UAV maintains an active connection with only the BS providing the strongest signal. Consequently, the complete SNR radio map under the BS decision vector z, denoted as Qz (s), is formulated as: Qz (s) = C. Coverage-Aware Search
max
i∈{1,2,...,M },zi =1
Combinatorial
qi (s),
∀s.
Multi-Armed
(5) Bandit
Given the candidate BS pool of size M , selecting K BSs leads to ‘M choose K’ possible deployment combinations. This creates a massive combinatorial search space. In addition, evaluating the quality of each combination requires trajectory evaluation under communication constraints, which is relatively time-consuming. Consequently, directly assessing all candidate combinations is computationally prohibitive in realistic urban scenarios. To reduce the search cost, we recast the BS selection process as a CMAB problem. Unlike a conventional bandit setting
that selects a single arm at each decision step, CMAB selects a group of arms jointly, which is commonly referred to as a super arm. This formulation is suitable for BS deployment because the performance of one BS cannot be fully assessed in isolation. Instead, its value depends on how it cooperates with other selected BSs to provide continuous coverage along the flight corridor. Each candidate BS is regarded as an arm, and a super arm corresponds to a feasible BS deployment vector z defined in Eq. (1). For a given BS budget K, the selected super arm further satisfies M X
zi = K.
(6)
i=1
To evaluate each arm during the search, we adopt the upper confidence bound (UCB) criterion. The UCB score consists of an empirical reward term and a confidence term, enabling a balance between exploitation and exploration. For candidate BS i, it is written as U CBi = r̄i + ui .
(7)
During the search, the algorithm maintains the number of times the i-th BS has been selected, denoted by ni , and its accumulated normalized reward, denoted by vi . The empirical reward is computed as vi r̄i = , (8) ni + ϵ where ϵ is a small positive constant for numerical stability. A larger r̄i indicates that i-th BS has more frequently appeared in high quality deployment combinations. The confidence term ui is defined as s ln(τ + 1) ui = c , (9) ni + ϵ where c controls the exploration strength and τ is the number of evaluated BS combinations. This term assigns a larger bonus to candidates with fewer selections, allowing less explored BSs to remain competitive during the search. Although the UCB score captures the historical performance and uncertainty of each individual BS, it does not explicitly measure the performance complementarity among selected
5
BSs. To promote such complementary effect, we introduce a waypoint-based coverage award. At the beginning of the search, the waypoint set can be empty. Once promising trajectories have been obtained from previous evaluations, we sample a set of corridor waypoints denoted by W = {s1 , s2 , . . . , sNW }.
(10)
For candidate BS i, its covered waypoint index set is defined as Ci = {j | qi (sj ) ≥ γth , sj ∈ W} .
(11)
During the greedy construction of one BS combination, let U denote the index set of waypoints that are still uncovered by the selected BSs. The marginal waypoint coverage gain of BS i is ( |Ci ∩U | |W| , |W| > 0, (12) gi = 0, |W| = 0. which gives a higher score to candidates that can cover currently uncovered parts of the sampled flight corridor, making the selected set of BSs less redundant in coverage. The final selection score of candidate BS i is Φi = r̄i + ui + λc gi ,
(13)
where λc controls the influence of the waypoint coverage award. At each iteration t, the algorithm greedily constructs one selected set I. Starting from I = ∅, it repeatedly chooses the candidate BS with the largest score Φi . After a BS is selected, the uncovered waypoint set U is updated and the scores Φi of the remaining candidates are recomputed. To reduce redundant deployments in nearby locations, candidates whose physical distance to the selected BS is smaller than dmin are removed from the current action pool. Once K BSs are selected, the corresponding deployment vector z is passed to the trajectory evaluation module to obtain the normalized reward ρ(t) for updating vi . The updated statistics are then used to guide the construction of the next super arm. D. Two-Step Trajectory Evaluation Repeatedly performing communication-aware pathfinding often imposes a substantial computational burden, particularly within environments densely populated with obstacles. We implement a two-step trajectory evaluation strategy that segregates the preliminary assessment of topological connectivity from the subsequent derivation of the exact trajectory. 1) Rapid pruning via CCL: The first step serves as a computationally lightweight filter to swiftly eliminate invalid BS combinations failing to provide continuous end-to-end radio coverage. Given a specific deployment decision z, the SNR radio map Qz (s) is derived via Eq. (5). Subsequently, a binary communication coverage mask, denoted as M(s), is constructed by applying the predefined SNR threshold γth : ( 1, if Qz (s) ≥ γth M(s) = (14) 0, otherwise Rather than directly applying trajectory search on M(s), we exploit the CCL algorithm from image processing and
Algorithm 1 Coverage-Aware Combinatorial Multi-Armed Bandit Search Require: Candidate BS location set B = {b1 , b2 , . . . , bM }, waypoint set W, BS budget K, threshold γth , exploration coefficient c, weight λc , minimum separation dmin , maximum iterations T 1: Initialize ni ← 0 and vi ← 0 for all bi ∈ B 2: Initialize τ ← 0, ρ∗ ← −∞, z∗ ← 0, and cache H 3: Compute Ci for all bi ∈ B if W ̸= ∅ 4: for t = 1 to T do 5: I ← ∅, A ← B 6: U ← {1, 2, . . . , |W|} 7: while |I| < K and A ̸= ∅ do 8: for each bi ∈ A do 9: r̄i ← vip /(ni + ϵ) 10: ui ← c ln(τ + 1)/(ni + ϵ) 11: Compute gi using Ci and U 12: Φi ← r̄i + ui + λc gi 13: end for 14: bi∗ ← arg maxbi ∈A Φi 15: I ← I ∪ {i∗ } 16: U ← U \ Ci∗ 17: A ← {bi ∈ A : i ̸= i∗ , ∥bi − bi∗ ∥2 ≥ dmin } 18: end while 19: Construct z(t) from I 20: if z(t) is not in H then 21: Evaluate z(t) and obtain normalized reward ρ(t) 22: Store ρ(t) in H 23: τ ←τ +1 24: else 25: Retrieve ρ(t) from H 26: end if 27: for each i ∈ I do 28: ni ← ni + 1 29: vi ← vi + ρ(t) 30: end for 31: if ρ(t) > ρ∗ then 32: ρ∗ ← ρ(t) , z∗ ← z(t) 33: end if 34: end for 35: return z∗
graph theory. By scanning the binary mask, the CCL algorithm partitions the feasible communication space into a set of disjoint, maximally connected sub-graphs. We then verify whether the vertiports (UAV’s vertical takeoff and landing locations, the nodes in the network) reside within the same connected component. If they are distributed across disconnected regions, it indicates the absence of a continuous communication corridor. Consequently, the trajectory evaluation is immediately terminated, and a severe penalty is assigned to the combination z. This rapid pruning mechanism circumvents the redundant computational overhead of exact pathfinding for disconnected BS deployments. 2) Exact routing via communication-aware A∗ : Passing the CCL verification indicates that the vertiports are connected within the region satisfying the wireless connectivity
6
constraint. Next, a communication-aware A∗ algorithm is employed to determine the optimal flight trajectory π ∗ . This trajectory operates within the feasible communication space (M(s) = 1) while safely avoiding physical building obstacles (hobs (π) = 0). The shortest total flight corridor length L(π ∗ ) is then computed and converted into a reward signal. For the deployment vector evaluated at iteration t, the raw reward is defined as ( R0 − L(π ∗ ), if a feasible trajectory is found, (t) R = 0, otherwise, (15) where R0 is the base reward. A shorter feasible corridor therefore leads to a larger raw reward. To make the reward scale suitable for updating the bandit statistics in Algorithm 1, the raw reward is further normalized as R(t) − Rmin , (16) ρ(t) = min 1, max 0, Rmax − Rmin where Rmin and Rmax denote the lower and upper normalization bounds, respectively. The normalized reward ρ(t) is then used to update the accumulated reward vi of each selected BS in Algorithm 1. E. Reciprocity-Driven Space Expansion The CMAB search operates on a finite candidate BS pool. Although this pool makes the combinatorial search tractable, it also imposes a structural limitation on the final solution. In dense urban environments, a small number of geometrically sensitive locations may determine whether a corridor can maintain continuous radio coverage. If the initial candidate pool does not include these “sweet spot” locations, keeping the pool fixed throughout the search may prevent the algorithm from finding high-quality BS deployment layouts. A direct way to enlarge the candidate pool is to sample additional BS locations and compute their radio maps. However, such a strategy is inefficient because most newly sampled positions may have little relevance to the current corridor structure. Instead of expanding the search space blindly, we use the paths discovered during the search to infer where new BS candidates should be introduced. The key idea is that a feasible or near-feasible flight corridor already contains information about communication demand. If a new BS location can provide coverage to multiple waypoints along this corridor, it is more likely to improve the subsequent CMAB search than an arbitrary location sampled from the whole map. The proposed mechanism is built upon the reciprocity principle of wave propagation, which is a statement of equivalence or symmetry [32]. In a static propagation environment, the transmitter and receiver sites may be interchanged without changing the amplitude or phase of the observed signal [33]. As shown in Fig. 2, for a BS deployment location b and a UAV waypoint s, this property can be expressed as Lch (b, s) ≃ Lch (s, b),
(17)
where Lch (b, s) denotes the propagation gain from b to s, and Lch (s, b) denotes the reverse link gain. This relation allows the algorithm to use UAV waypoints as temporary
Fig. 2: Illustration of channel reciprocity between a BS and a UAV waypoint in a complex urban environment. With swapped positions, the BS-to-UAV and UAV-to-BS links experience approximately the same propagation paths under a static reciprocal channel, including both LoS and reflected components.
radio transmitters. Rather than testing many potential BSs in the forward direction, the algorithm transmits virtually from a small number of informative trajectory waypoints and observes which deployment locations can receive strong reciprocal signals. These locations are likely to be useful BS candidates because the reverse link suggests favorable forward coverage toward the same waypoints. The waypoint set W used for the coverage award in Algorithm 1 is sampled from promising trajectories obtained in recent CMAB iterations and is reused here for reciprocity-driven expansion. For each waypoint sℓ ∈ W, a reciprocal SNR field is computed over the feasible BS deployment region, denoted by ΩBS . We use qeℓ (b) to represent the reciprocal SNR map generated by waypoint sℓ , where sℓ acts as the temporary transmitter and b ∈ ΩBS is the receiving deployment location. The reciprocal coverage score of location b is defined as G(b) =
NW X
I {e qℓ (b) ≥ γth } ,
b ∈ ΩBS .
(18)
ℓ=1
The score G(b) counts how many sampled corridor waypoints can be reciprocally covered by location b. A larger value indicates that b is more relevant to the current corridor geometry. A location with a larger G(b) value can reciprocally cover more sampled corridor waypoints, and is therefore assigned a higher priority during candidate generation. At the same time, candidate diversity is required because several neighboring locations may have similar reciprocal coverage scores. Selecting all of them would provide limited additional information to the CMAB search. Therefore, the new candidate set is selected by jointly considering the reciprocal coverage score and the spatial distribution among candidates. Let Bnew denote the set of newly injected BS candidates. It is given by Bnew = {bM +1 , bM +2 , . . . , bM +Mnew },
bi ∈ ΩBS ,
(19)
where Mnew is the number of newly generated candidates. The selected locations are expected to have high reciprocal coverage scores while remaining spatially separated from each
7
Algorithm 2 Reciprocity-Driven Space Expansion Require: Promising trajectory set Πloc , feasible BS deployment region ΩBS , current candidate pool B, threshold γth , number of new candidates Mnew NW 1: Sample waypoint set W = {sℓ }ℓ=1 from Πloc 2: if NW = 0 then 3: return B 4: end if 5: for ℓ = 1 to NW do 6: Compute qeℓ (b), ∀b ∈ ΩBS 7: end for 8: for each b ∈ ΩBS do PNW I {e qℓ (b) ≥ γth } 9: G(b) ← ℓ=1 10: end for 11: Initialize Bnew ← ∅ 12: while |Bnew | < Mnew do 13: b∗ ← arg maxb∈ΩBS G(b) 14: Bnew ← Bnew ∪ {b∗ } 15: Remove locations near b∗ from ΩBS 16: end while 17: for each bi ∈ Bnew do 18: Compute forward radio map qi (s), ∀s 19: Initialize its coverage set Ci and ni , vi 20: end for 21: B ← B ∪ Bnew 22: M ← M + Mnew 23: return B
other. This design avoids repeatedly adding candidates from the same local region and increases the chance that different parts of the sampled corridor are supported by different new BS candidates. Each location in Bnew is then treated as a standard BS candidate. For every bi ∈ Bnew , a forward SNR radio map qi (s) is computed over the UAV flight plane using the radio propagation model introduced previously. These new radio maps are appended to the existing radio map library. The candidate pool is expanded from M to M + Mnew , so the following CMAB iterations can select both the original BS candidates and the reciprocity guided candidates. IV. E XPERIMENTS A. Experimental Setup The experiments are conducted in a digital-twin urban scenario reconstructed from OpenStreetMap data. We select Kwun Tong, Hong Kong, as the case study area, where a UAV logistics network is particularly useful for delivering cargo of various natures. This typical industrial district represents a high-density urban environment with compact building blocks, irregular street layouts, and pronounced urban canyon characteristics. These features provide a challenging testbed for co-planning BS deployment and UAV flight corridors. In our numerical experiments, the UAV logistics network in Kwun Tong comprises four vertiport locations, forming six unique origin-destination flight corridors. The digital 3D urban model used in the experiments is shown in Fig. 3.
Fig. 3: The 3D urban model used in the case study. TABLE I: Main simulation parameters. Parameter Value Communication parameters BS transmit power Pt 40 dBm BS antenna gain Gt 20 dB Carrier frequency fc 39 GHz Bandwidth B 4 × 108 Hz Noise figure NF 5 dB UAV antenna gain Gr 8 dB SNR threshold γth 20 dB Maximum ray depth 3 Scenario parameters Area size 1600 m × 1400 m Flight altitude 50 m BS height 10 m Number of vertiports 4 Algorithm parameters Confidence term c 2 Coverage term λc 3 Minimum separation dmin 100 m Candidate pool size M 300 Maximum iterations T 12000 Expansion interval N 3200 iterations New candidates number Mnew 20
For the UAV logistics network, the cruising altitude of air corridors is set to 50 m above ground level, which is a typical altitude for drone delivery under the 120 m constraint. The BSs serving the network are placed at a height of 10 m above ground level, installed on lampposts or building rooftops. NVIDIA Sionna RT [34] is used as the ray-tracing simulation platform for radio propagation modeling. Built upon GPU-accelerated physical ray tracing, Sionna RT provides an
8
efficient simulation environment for constructing high-fidelity radio maps in complex urban environments. The main simulation parameters are summarized in Table I. All simulations are executed on a workstation equipped with an AMD Ryzen 9 9950X CPU, an NVIDIA GeForce RTX 5090 D GPU with CUDA 12.9 support, and 64 GiB running memory. B. Baseline Algorithms and Evaluation Metrics We compare the performance of our proposed method CRCMAB with the following five baseline methods: 1) Greedy Coverage: This baseline follows a conventional cellular deployment criterion that aims to maximize the covered area. It selects BSs according to their marginal coverage gain over the full radio map, with limited randomness introduced by a restricted candidate list. 2) Genetic Algorithm (GA): GA evolves a population of BS combinations through selection, crossover, mutation, and elitism. The trajectory length returned by the evaluation module is used as the fitness value. 3) Particle Swarm Optimization (PSO): PSO uses a continuous position vector to represent each particle. At each iteration, the top K entries are mapped to a BS combination, and the particles are updated according to their personal and global best records. 4) Monte Carlo Tree Search (MCTS): MCTS constructs a BS combination sequentially. Each node represents a partial selection, and the tree is updated through selection, expansion, rollout, and backpropagation. 5) CMAB Only: This method removes reciprocity-driven space expansion from CR-CMAB while keeping the coverage-aware CMAB search. The candidate BS pool therefore remains fixed during the search. The flight corridor planning performance is first evaluated using the path length offset with respect to the theoretical total shortest path length of the network without communication constraints: ∆L = L(π) − L(πref ), (20) where π denotes the communication-aware total path length and πref is the shortest path length through A* when only physical obstacles are considered. In the Kwun Tong UAV network, the reference path length is L(πref ) = 6277.9 m.
configurations generated by different approaches, with the total path length offset ∆L displayed in the title. A smaller ∆L indicates better operational efficiency, as the UAV flight distance is shorter while serving the identical logistics network. We can observe that the planning scheme produced by our proposed CR-CMAB method achieves the smallest ∆L among all six methods, with a path length offset of just 178.5 m above the theoretical limit of 6277.9 m. This is because CR-CMAB places BSs more effectively in the urban space, allowing wireless signals to penetrate and connect the narrow regions between dense building blocks. As a result, the planned flight corridor can pass through the congested urban area with fewer unnecessary detours. In contrast, Greedy Coverage and MCTS produce much more circuitous flight paths. The BSs selected by these two methods do not sufficiently support the short corridors between vertiports, even though they can still provide feasible communication coverage. Fig. 5 compares the path length offset of different algorithms under varying BS budgets. As K increases, the path offsets of all methods decrease rapidly at first and then gradually flatten. CR-CMAB achieves the shortest flight corridor path under all tested BS budgets. Its advantage is most evident when K ranges from 3 to 5, where BS resources are still limited and the deployment decision has a strong impact on corridor connectivity. This performance pattern reflects the effectiveness of CR-CMAB under constrained infrastructure resources. In contrast, the CMAB-only result exhibits a substantially larger corridor length than CR-CMAB, demonstrating that reciprocity-driven space expansion brings a clear performance gain to the search algorithm. Table II provides the detailed path length offset and runtime results for all algorithms. Relative to CMAB only, the fastest method in terms of total runtime, CR-CMAB requires 17.3% more time but reduces the total path length offset by 45.4%, from 3389.0 m to 1851.7 m. Greedy Coverage has the second shortest total runtime, but fails to find a feasible corridor when K = 2. PSO, GA, and MCTS require substantially longer runtimes than CR-CMAB, while their path offsets remain larger. Overall, CR-CMAB achieves the best planning performance while keeping the additional computational cost moderate. D. Ablation Analysis of the Selection Score
C. Comparison Analysis of Different Algorithms and Varying BS Budgets The visualizations in Fig. 4 illustrate the overall concept. The objective is the joint optimization of BS locations and flight corridors to serve a UAV delivery network with four vertiports and six origin-destination pairs in the Kwun Tong urban area shown in Fig. 3. In each subfigure, red stars denote the fixed vertiport locations, yellow dots denote the BS locations at 10 m above ground level, blue lines denote the UAV flight corridors at 50 m above ground level, and the colormap shows the signal intensity on the 50 m altitude plane, all overlaid on the Kwun Tong 3D urban model. In Fig. 4, we fix the number of vertiports at K = 4. Each subfigure shows the BS locations and flight corridor
To further examine the internal design of the CR-CMAB algorithm, we conduct an ablation study on the CMAB selection score. Different from the CMAB-only baseline, which removes the reciprocity-driven candidate expansion, this experiment keeps the complete CR-CMAB framework enabled and only changes the composition of the CMAB scoring function. In this way, the effect of each scoring component can be evaluated within the full reciprocity-guided search process. As introduced in Eq. (13), the CMAB score used to select each candidate BS comprises three terms: r̄i , ui , and gi . The reward term r̄i exploits candidates that have appeared in high-quality BS combinations. The confidence term ui encourages exploration of candidates that have not been sufficiently evaluated, while the coverage term λc gi provides
9
Fig. 4: Visualization of BS deployment and flight corridor planning results for different algorithms when K = 4.