arXiv:2609.16877v1 [cs.NI] 15 Sep 2026
A Mechanical Antenna for Improving Capacity Fairness in Dynamic Multi-Station Scenarios 1st Akihito Taya
2nd Yuuki Nishiyama
3rd Kaoru Sezaki
Institute of Industrial Science, The University of Tokyo Tokyo, JAPAN [email protected]
Center for Spatial Information Science, The University of Tokyo Chiba, JAPAN [email protected]
Center for Spatial Information Science, The University of Tokyo Chiba, JAPAN [email protected]
Abstract—While indoor Internet of Things (IoT) and sensor networks increasingly rely on Wi-Fi access points (APs) to collect high-bandwidth data streams from multiple devices, conventional APs rely on static antenna deployments, whose fixed orientations are often suboptimal in dynamic propagation environments. To overcome this limitation, this paper proposes a mechanical Wi-Fi antenna control system that adaptively optimizes its 3D antenna orientation for dynamic multi-station scenarios. The proposed system autonomously actuates its physical antennas in response to perceived radio environments by combining state-specific blackbox optimizers and capacity-based environment change detection. The evaluation results show that the proposed system improves channel capacity under dynamic station combinations, avoids unnecessary re-optimization under transient blockages, and triggers re-optimization after sustained environmental changes such as continuous blockage and device relocation. Index Terms—Mechanical Antenna, Automatic Antenna Tuning, Proportional Fairness, Channel State Information, Bayesian Optimization
I. I NTRODUCTION Indoor Internet of Things (IoT) and sensor networks increasingly rely on Wi-Fi access points (APs) to collect high-bandwidth uplink data from cameras, LiDARs, robots, and other sensing devices. In such networks, performance is affected not only by protocols but also by the physical radio environment, including antenna orientation, polarization mismatch, and temporary blockages. However, conventional APs are typically deployed with static antenna orientations and cannot adapt to changes in active devices or propagation conditions. While recent advances in movable and rotatable antennas [1]–[4] explore physical repositioning to overcome such environmental limitations, they primarily focus on fine-tuning phased array elements for narrow beams. Similarly, although mechanical 3D beamforming systems like NetBeam [5] exist, they rely on tightly synchronized, fully managed architectures. Consequently, the dynamic management of broad-beam propagation—such as from omnidirectional Wi-Fi antennas in unmanaged indoor settings—remains largely unexplored, This work was supported by JSPS KAKENHI Grant Number JP23K26109, JP24K20759 and the Telecommunications Advancement Foundation. Additionally, Gemini and ChatGPT were utilized for English proofreading and stylistic refinement of this manuscript.
Fig. 1: The proposed device equipped with mechanical Wi-Fi antennas and a web-based monitoring interface. The device can adjust the orientation of two antennas in both horizontal and vertical directions.
leaving these systems vulnerable to suboptimal communication performance. To address this physical-layer bottleneck, we draw inspiration from physical artificial intelligence (AI) [6], an emerging paradigm that integrates physics-based perception and embodied interaction to enable systems to adapt robustly via physical actuators. Aligning with this concept, our previous work [7] introduced a mechanical Wi-Fi antenna (shown in Fig. 1) that automatically adjusts its 3D orientation via blackbox optimization to maximize the channel state information (CSI)-derived capacity for a single link. While this preliminary system demonstrated that physical polarization matching is critical for capacity enhancement, it was limited to a single, static scenario. Practical deployments demand a robust extension capable of managing simultaneous communications from multiple stations (STAs) while autonomously adapting to unpredictable environmental changes. Because the ideal antenna orientation fundamentally differs for each combination of active STAs, the system must dynamically adapt to their activation and deactivation. This online adaptability highlights the core advantage of mechanical antennas over traditional static deployments. Furthermore, the system must reliably detect and react to post-optimization environmental shifts, such as device mobility or unexpected
blockages. To address these challenges, we propose a concurrent optimization framework that manages dynamic STA combinations and reactively handles environmental shifts. The framework maintains independent, state-specific optimizers for every possible combination of active STAs. It executes the optimizer corresponding to the current state while freezing the others, ensuring seamless state transitions. Furthermore, we introduce a capacity-based environment change detection mechanism that continuously monitors the channel after convergence. Upon detecting a sustained anomaly, it resets the optimization to adapt to the new conditions, ensuring robustness in unpredictable real-world applications. The main contributions of this paper are summarized as follows: We develop a mechanical Wi-Fi antenna system that autonomously tunes its 3D orientation to maximize the proportional fairness of channel capacity across multiple STAs while adapting to environmental shifts. • We design a fully asynchronous architecture to decouple processes with disparate time scales, including capacity measurement, antenna actuation, and real-time visualization. This enables the efficient, concurrent execution of optimization and environment change detection. • We experimentally demonstrate the system’s effectiveness in dynamic scenarios with varying active STA combinations. The results confirm its robustness against transient line of sight (LOS) blockages and adaptation to sustained changes, such as continuous blockages or device mobility.
•
II. R ELATED W ORK The traditional approach to controlling radio propagation is beamforming, which steers directivity either mechanically or electronically using phased array antennas [8]–[10]. These methods generally assume point-to-point communication with narrow beams. In contrast, our proposed mechanical antenna controls the 3D orientation of broad-beam antennas (e.g., omnidirectional antennas) to serve multiple devices simultaneously. Additionally, unlike conventional beamforming, which rarely addresses polarization, our device inherently manipulates polarization, a critical factor for capacity enhancement as demonstrated in [7]. Recently, movable antennas [1], [2] have attracted considerable attention for improving communication performance through antenna position adjustment. These systems fine-tune the elements of a phased array to achieve more precise beamforming. In particular, rotatable antennas [3], [4], a variant of movable antennas, have also been studied for enhancing communication performance by adjusting their orientations. However, these studies primarily focus on fine-tuning phased array elements to optimize narrow beams, effectively categorizing them under traditional beamforming. This differs fundamentally from our approach, which independently actuates omnidirectional-like multiple-input multiple-output (MIMO) antennas to manipulate both directivity and polarization.
In the context of multi-user communication, mechanical antenna tilting has been studied for cellular base stations to optimize coverage and capacity [11], [12]. However, such largescale antennas incur significant mechanical delay, precluding frequent reconfigurations. Research on mechanical antenna control for indoor, small-scale wireless local area network (WLAN) environments remains scarce. Our system addresses this gap by employing small, rapidly actuating antennas capable of on-the-fly reconfiguration for specific active terminals. Regarding 3D orientation control, NetBeam [5] investigates distributed beamforming using multiple transmitters equipped with servo motors. While providing indoor experimental insights, NetBeam relies on a fully managed architecture using software-defined radios (SDRs) for tight synchronization and CSI feedback, akin to cellular networks. Our work, conversely, targets unmanaged, small-scale systems like Wi-Fi. Crucially, our proposal introduces a comprehensive framework that asynchronously optimizes multiple active STA combinations and reactively adapts to environmental shifts, which are not addressed by NetBeam. III. S YSTEM M ODEL We consider a scenario where a set of STAs S dynamically start and stop uplink streaming sessions to a central AP. A typical use case is an indoor IoT edge network in which an AP collects high-bandwidth uplink sensor streams, such as camera, LiDAR, or robot telemetry data, from multiple sensing devices. While we focus on the uplink, our proposed method is equally applicable to downlink scenarios by leveraging radio wave reciprocity and standard CSI feedback mechanisms. Unlike conventional APs with static antennas, our proposed AP dynamically adjusts its 3D antenna orientation θ to optimize channel capacity. Specifically, the device independently controls two antennas in both elevation θj and azimuth φj (for antenna j ∈ {1, 2}), yielding a four-degree-of-freedom configuration space: θ = (θ1 , φ1 , θ2 , φ2 ). Upon receiving data frames from STA i ∈ S at a given orientation θ, the AP measures the CSI Hi,k (θ) ∈ Cnr ×nt , where k, nr and nt denote the subcarrier index, the number of receive antennas, and the number of transmit antennas, respectively. Using this CSI, the AP computes the achievable channel capacity, which serves as a physical-layer performance metric, and optimizes θ to maximize proportional fairness among all active STAs. The MIMO-orthogonal frequency-division multiplexing (OFDM) channel capacity Ci (θ) between STA i ∈ S and the AP is calculated as [13]: K X 1 † Hi,k (θ)Hi,k (θ) , Ci (θ) = ∆f log2 det I + N0 ∆f k=1 (1) where ∆f , K, and N0 denote the subcarrier bandwidth, the number of subcarriers, and the noise density per antenna, respectively. Additionally, I is the nr × nr identity matrix and [·]† denotes the Hermitian transpose. Note that Hi,k (θ) is
treated as an effective channel matrix that includes the received signal power scaling. To balance overall channel capacity and promote fairness among users, we adopt proportional fairness [14] based on the time-averaged capacity C̄i (θ) over a window Tavg as our optimization objective. We define the proportional fairness metric FS (θ) for the active stations S ⊆ S as: X FS (θ) := log C̄i (θ) . (2)
AP with Mechanical Antenna STA
Optimization Manager
Capacity Calculator
Capacity/RSSI
optimization loop Data
Capacity/RSSI Check
i∈S
Since the ideal antenna orientation inherently depends on which devices are active, the system aims to find the optimal configuration θS∗ := arg maxθ FS (θ) for each unique STA combination S. Practically, commercial hardware often reports a normalized CSI, H̃i,k (θ) = αHi,k (θ). Although the scaling factor α is typically undisclosed, it can be derived from the received signal strength indicator (RSSI). Because RSSI encapsulates the total signal and noise power across all subcarriers, it satisfies: RSSI =
K X
∥Hi,k (θ)∥2 + Knr N0 ∆f,
(3)
k=1
where ∥ · ∥ denotes the Frobenius norm. Consequently, α is extracted as: PK 2 2 k=1 ∥H̃i,k (θ)∥ . (4) α = RSSI − Knr N0 ∆f Substituting this into (1) yields the practical capacity formula: Ci (θ) = ∆f ×
K X k=1
log2 det I +
1 H̃i,k (θ)H̃i,k (θ)† 2 α N0 ∆f
Optimizer
Buffer Data
. (5)
IV. C ONCURRENT O PTIMIZATION F RAMEWORK AND E NVIRONMENT C HANGE D ETECTION A. Asynchronous System Architecture In realistic unmanaged networks like Wi-Fi, the AP cannot dictate the exact transmission timings of the active STAs. If the system were to perform high-frequency capacity computation and mechanical antenna actuation, which requires hundreds of milliseconds, sequentially within a synchronous loop, the overall processing efficiency would severely degrade. To resolve this bottleneck, we designed a fully asynchronous architecture, as illustrated in Fig. 2. The system decouples the fast sensing processes from the relatively slow optimization control loop. A background Capacity Calculator continuously processes incoming frames and updates the moving average of the channel capacity in a shared memory buffer. Independently, the Optimization Manager queries this buffer only when evaluating the objective function. This thread-safe, low-latency data exchange completely eliminates the need to synchronize packet arrivals with mechanical actuation, ensuring efficient adaptability and facilitating smooth system monitoring via real-time visualization.
Actuator
Request
active STA Data
suggestion Set angle θ
Capacity/RSSI Request
Wait Twait Report
Data
Capacity/RSSI
average capacity
Fig. 2: Sequence diagram of the asynchronous antenna tuning framework. The high-frequency capacity measurement is decoupled from the slow optimization loop to enable online adaptation.
B. State-Specific Concurrent Optimization Because the joint optimization of multiple 3D antenna orientations to maximize proportional fairness among multiple STAs can be highly complex and potentially non-convex, we formulate the tuning process as a black-box optimization problem. While our proposed architecture is solver-agnostic and can integrate any black-box optimization algorithm, methods that require a large number of agents to explore the parameter space are less suitable for our system due to the sequential nature of physical measurements. Consequently, we employ Bayesian optimization [15] in our evaluations because it is well suited for optimizing expensive-to-evaluate functions with a limited number of samples. The primary challenge in dynamic scenarios is that the optimal antenna configuration fundamentally shifts whenever a STA becomes active or inactive. To seamlessly manage these transitions, the Optimization Manager maintains a set of independent, state-specific optimizers, denoted as O[S], for every possible active STA combination S. Algorithm 1 details the execution flow of the proposed concurrent optimization framework. At the beginning of each iteration, the Optimization Manager observes the active STA combination S. If S is not empty, the manager invokes only the corresponding optimizer O[S]. Crucially, all other optimizers are frozen, preserving their learned internal states until their specific STA combinations become active again. The selected optimizer then suggests an antenna configuration θ. Because physically adjusting the mechanical antennas takes on the order of hundreds of milliseconds, the manager commands the Actuator to move the antennas to θ and waits for a stabilization period Twait before measuring the updated channel capacity. During this waiting phase, the active STA combination might dynamically change (e.g., a station might finish its transmission). To handle this dynamic event, the
RX and TX are mounted on a ladder at a height of 1.7 m and on desks at a height of 0.7 m, respectively.
Algorithm 1 Concurrent antenna orientation optimization for multiple STA combinations
RX TX B Layout 1
1: while optimization is running do 2: Observe the current active STA combination S
manager explicitly verifies whether all STAs in S are still active after the delay (Line 6). If any STA is found to be inactive, the current optimization round is safely aborted to avoid injecting mismatched capacity evaluations into the optimizer. Otherwise, the manager retrieves the average capacities C̄i (θ) from the shared buffer, calculates the proportional fairness objective, and reports it to update O[S]. Upon convergence, the manager stores the optimal orientation θSbest and best , which serve as the corresponding achieved capacities CS,i the baseline for the environment change detection described in the next subsection. The proposed framework maintains independent optimizers for each possible combination of active STAs. Thus, the number of possible non-empty states grows as 2|S| − 1. Consequently, the storage required to maintain the learned optimizer states and the cumulative time required to optimize all possible states may become significant as the number of STAs increases. The current framework therefore targets small-scale deployments, such as homes and small offices. Extending the framework to handle larger numbers of STAs is left for future work. C. Capacity-Based Environment Change Detection After convergence, the optimizer O[S] is designed to consistently output the optimal orientation θSbest , allowing the Optimization Manager to seamlessly continue its evaluation loop. However, the manager must continuously monitor the
1.3 m
2.2 m 0.7 m
TX B Layout 2 Desk
Geometric Mean of Capacity (Mbps)
(a) Experimental setup. 2800
Fixed A Fixed B Fixed C Random Bayesian
2700 2600 2500 2400
Fixed A
Fixed B
Fixed C
Random Bayesian
(b) Layout 1 Geometric Mean of Capacity (Mbps)
if S = ∅ then Wait Twait and continue θ ← O[S].suggest() Move antennas to θ and wait Twait if any STA in S is inactive then continue for all i ∈ S do 8: Calculate average channel capacity C̄i (θ) 9: end for 10: if not O[S].is_converged () then 11: Calculate the objective value FS (θ) by (2) 12: O[S].report(θ, FS (θ)) 13: if converged at this round then 14: Store the best antenna orientation θSbest best 15: CS,i ← C̄i (θSbest ) (∀i ∈ S) 16: end if 17: else 18: if environment change is detected then 19: O[S].reset() 20: end if 21: end if 22: end while 3: 4: 5: 6: 7:
TX A
2800
Fixed A Fixed B Fixed C Random Bayesian
2700 2600 2500 2400
Fixed A
Fixed B
Fixed C
Random Bayesian
(c) Layout 2
Fig. 3: Distribution of the geometric mean of the channel capacities in static scenarios. (a) Experimental setup showing the locations of the TXs and RX. (b) and (c) compare fixed, randomly selected, and Bayesian-optimized antenna orientations under two layouts with different TX A antenna orientations and different TX B locations. Bayesian optimization consistently finds high-performing orientations in both layouts.
links to detect post-optimization environmental shifts, such as device mobility or physical blockages. Relying directly on raw CSI for anomaly detection is impractical; CSI matrices are high-dimensional complex matrices whose phases often exhibit unstable fluctuations even in strictly static environments. Instead, we adopt the channel capacity as a robust, macroscopic feature to reliably characterize the radio environment. The Optimization Manager calculates the relative capacity deviation δi between the currently observed average capacity C̄i (θ) and the recorded baseline best best best CS,i , defined as δi := |C̄i (θ) − CS,i |/CS,i . To prevent the system from overreacting to transient blockages (e.g., a person simply walking past the AP), the manager increments a state-specific counter if δi for any active STA exceeds a predefined threshold γth . If this condition is met for M consecutive iterations, the manager concludes that a sustained environmental change has occurred and resets O[S] to dynamically adapt to the new conditions.
V. E VALUATIONS A. Static Scenario Evaluation We first evaluated the proposed system in static scenarios to assess the effectiveness of the optimization framework. The evaluation was conducted in an indoor environment under two layouts differing in TX A antenna orientations and TX B locations, as shown in Fig. 3(a). For each layout, we compared three representative fixed antenna configurations, randomly selected orientations, and Bayesian-optimized orientations. The four-dimensional antenna orientations (θ1 , φ1 , θ2 , φ2 ) of Fixed A, B, and C were set to (0◦ , 90◦ , 0◦ , 90◦ ), (90◦ , 90◦ , 90◦ , 90◦ ), and (90◦ , 180◦ , 90◦ , 180◦ ), respectively, where θ represents the elevation angle ranging from 0◦ (vertical) to 90◦ (horizontal), and φ represents the azimuth angle ranging from 60◦ (toward the center of the RX) to 180◦ (outward from the RX). Each fixed orientation configuration was evaluated five times. Random orientations were uniformly and independently sampled 100 times from the feasible orientation space to simulate arbitrary antenna configurations selected by non-expert users. Bayesian optimization was independently repeated five times, with 50 optimization iterations in each trial. For each trial, the best orientation found within the 50 iterations was used for comparison. As the performance metric, we used the geometric mean of the channel capacities of the two links, whose maximization is equivalent to maximizing the proportionalfairness objective in (2). The resulting distributions are shown in Figs. 3(b) and 3(c). The boxes represent the median and interquartile range (IQR), with whiskers defined using Tukey’s 1.5×IQR rule. Random orientations can occasionally achieve performance comparable to Bayesian optimization, whereas Bayesian optimization consistently identifies high-performing orientations. The results also show that Bayesian optimization found high-performing orientations within 50 iterations in both static layouts. B. Experimental Setup of Dynamic Scenarios We evaluated the proposed device in an indoor environment with two TXs as shown in Fig. 4. The AP independently controlled two omnidirectional antennas in azimuth ([60◦ , 180◦ ]) and elevation ([0◦ , 90◦ ]). Using an ASUS NUC 13 Rugged (RX, TX A) and an Intel NUC 12 Pro Kit (TX B), we used PicoScenes [16] to emulate uplink streaming. Instead of actual application payloads, the TXs injected CSI measurement frames every 0.1 s. The RX flagged a TX as inactive if no frames arrive for 2.0 s. By manually toggling the TXs, we tested three active STA combinations: SA , SB , and SA,B (both active). For Bayesian optimization, we set a maximum of 50 iterations per state. To ensure mechanical stabilization before capacity evaluation, we set the wait time Twait = 2 s and measured the average capacity over a window Tavg = 1 s. The environment change detection parameters were configured to γth = 0.05 and M = 5. Finally, the noise density was calculated as N0 = kB T F , where kB , T , and F denote the Boltzmann constant, the temperature in Kelvin, and the noise
RX is mounted on a ladder at a height of 1.7 m.
A subject walks across the LOS between TX A and RX (t = 460–510 s).
2.2 m
RX 0.7 m
TXs are placed on desks at a height of 0.7 m.
1.0 m
TX B
TX A 1.3 m
TX B
Moved at t = 790 s
Desk
Fig. 4: Experimental setup and spatial layout. The figure details the physical arrangement of the RX and TXs, alongside the specific dynamic events (a pedestrian LOS blockage and the physical relocation of TX B).
figure of the receiver, respectively. We set T = 300 K and F = 7 dB, which are typical values for Wi-Fi receivers [17]. C. Applicability to Dynamic Scenarios Fig. 5 illustrates the time series of the channel capacities for TX A and TX B. Initially, both TXs are active, prompting the proposed device to explore antenna orientations to maximize proportional fairness using O[SA,B ]. Segment A (30–40 s) highlights the fundamental optimization loop: while the raw capacity fluctuates significantly every Twait = 2 s due to mechanical actuations, the averaging over Tavg = 1 s provides a stable and reliable metric for the optimizer. The system seamlessly manages dynamic STA states. At t = 70 s, TX A becomes inactive (Segment B); the system suspends O[SA,B ] and invokes O[SB ]. When TX A resumes at 140 s, O[SA,B ] restarts from its suspended state, eventually completing at 183 s (Segment C). After the completion, the capacities stabilize. Because the optimized orientation θSbest A,B is securely cached, the system instantly restores it whenever the state returns to SA,B (e.g., at 297 s and 436 s). Additionally, when TX B becomes inactive at 216 s, the manager correctly switches to O[SA ] to optimize the single link for TX A. After demonstrating the initial optimization, we evaluated the environment change detection algorithm by blocking the LOS. From 460 to 510 s, a subject repeatedly walked back and forth across the LOS of TX A (Fig. 4). As Segment D shows, this caused severe capacity drops. Crucially, because our detection algorithm requires a sustained change over M iterations, it successfully ignored this transient blockage without prematurely resetting the optimizer. Conversely, when the subject continuously blocked the LOS starting at 520 s, the system detected the sustained degradation and properly reset O[SA,B ] at 530 s (dashed orange line). After the optimization of O[SA,B ] completed at 650 s, the channel capacities stabilized again. Here, although the proposed device could maintain high capacity for TX B, the optimizer correctly prioritized improving the lower-capacity link (TX A) to increase the proportional fairness. Note that the subject cleared the LOS at 690 s, but the system did not reset O[SA,B ] because the relative capacity gap did not exceed the threshold γth = 0.05.
Segment B Segment A Segment C
Segment D
Capacity (Mbps)
3400
TX A (Raw) TX B (Raw) TX A (Avg) TX B (Avg) TX A (Reported) TX B (Reported) Average Capacity
3200 3000 2800 2600 2400 Blockage
2200 0
100
200
300
400
500
600
700
800
900
1000
1100
Time (sec) 3000 2800 2600 2400 30
3200
3200
3000 2800 2600
3000 2800 2600
2400 32
34
36
38
40
60
Segment A
2400 70
80
160
Segment B
180
200
220
Segment C
3200 3000 2800 2600 2400 2200 440
Sustained blockage (Standing)
Transient blockage (Walking)
460
480
500 520 Segment D
540
560
580
Fig. 5: Real-time trace of the channel capacities during the emulated streaming session. The light and dark solid lines represent the raw Ci (θ) and moving-average C̄i (θ) capacities, respectively. The blue and green colors correspond to TX A and TX B. Circle markers indicate the capacity values reported to the optimizers, while orange markers show the average capacity. Vertical solid and dashed orange lines denote optimization completions and resets, respectively. Segments A–D highlight the system’s dynamic adaptation to specific events.
Finally, at 790 s, we physically relocated TX B to a new position (gray box in Fig. 4). This sustained environmental shift triggered an immediate reset of O[SA,B ] at 800 s. Because only O[SA,B ] was updated during this period, the cached baseline for O[SB ] became obsolete. Consequently, when TX A became inactive at 940 s, the system correctly detected the baseline mismatch and reset O[SB ] to re-adapt TX B to its new environment. Overall, the proposed dynamic mechanical antenna framework exhibited substantial capacity variations across antenna configurations, with differences of approximately 200 Mbps for TX A and 400 Mbps for TX B between the worst and best configurations. As shown in the static scenario evaluation, Bayesian optimization found high-performing orientations within 50 iterations, requiring only a few minutes for each state. Although this timescale is not fast enough to track continuous station mobility or frequent environmental changes, the proposed system is suitable for quasi-static indoor scenarios, such as desk work, video conferencing, and video streaming. The environmental change detection mechanism further enhances the practical feasibility of the system by avoiding unnecessary re-optimization in response to transient events. Furthermore, proportional fairness balances the channel capacities among active STAs, demonstrating the
effectiveness of the proposed framework in dynamic multistation indoor environments. D. Impact of Antenna Orientation on RSSI To further illustrate the physical impact of the mechanical antennas, Fig. 6 visualizes the RSSI contour plots for the two independent antennas receiving signals from the two TXs. The results reveal RSSI fluctuations of 3–6 dB depending on the orientation, demonstrating the significant impact of antenna orientation on received signal strength. While this substantial variation might suggest that a simpler, RSSIbased control mechanism could be sufficient, maximizing raw received power does not necessarily maximize the channel capacity or proportional fairness in a multi-STA MIMO environment. Instead, the fact that the optimal angle regions shift significantly across the three scenarios reinforces the need for our capacity-based framework to adapt to the changes in the physical environment. VI. C ONCLUSION This paper proposed a mechanical Wi-Fi antenna device capable of autonomously tuning its 3D orientation to optimize communication performance for multiple stations. To handle dynamic changes in active station combinations, as
TX A > Ant. 1
90
−45
TX B > Ant. 1
90
−35 −40 180
150
120
Azimuth
TX B > Ant. 2
90
60
Azimuth
180
0
150
30
0
120
60
30 90
60
60
180
150
120
Azimuth
90 −40
90
60
180
0
150
0
120
30 90
60
30 60
60
TX A > Ant. 2 Elevation
Elevation
90
Azimuth
(a) Initial environment TX A > Ant. 1
90
−45
TX B > Ant. 1
90
−35 −40 180
150
120
Azimuth
TX B > Ant. 2
90
Azimuth
60
0
180
30
0
150
30 120
60
90
60
60
180
150
120
Azimuth
90 −40
90
60
180
0
150
30
0
120
30 90
60
60
60
TX A > Ant. 2 Elevation
Elevation
90
Azimuth
(b) During LOS blockage of TX A TX A > Ant. 1
90
90 −40 −45
TX B > Ant. 1
90
TX B > Ant. 2 −35 −40 180
150
120
90
Azimuth
60
0
180
30
0
150
30 120
60
90
60
60
Azimuth
180
150
120
90
60
Azimuth
180
0
150
30
0
120
30 90
60
60
60
TX A > Ant. 2 Elevation
Elevation
90
Azimuth
(c) After relocating TX B
Fig. 6: Contour plots of the measured RSSI for the two mechanical antennas and the two TXs. White dots indicate the specific antenna orientations evaluated during the optimization process. The shifting optimal regions under different scenarios highlight the necessity of dynamic antenna tuning.
well as occasional environmental shifts, we developed an asynchronous framework equipped with state-specific optimizers and a capacity-based environment change detection mechanism. Our experimental evaluations demonstrated that the proposed framework successfully adapts the antenna orientation during operation, improving proportional fairness among active transmitters. Furthermore, we showed that the system robustly recovers from environmental blockages and station relocation. Future work includes developing more computationally efficient optimization methods applicable to larger-scale systems with more stations and antennas, addressing highly dynamic scenarios such as mobile stations, and extending the framework toward integrated sensing and communications (ISAC) applications by jointly exploiting wireless measurements and mechanical antenna reconfiguration. R EFERENCES [1] L. Zhu, W. Ma, and R. Zhang, “Modeling and performance analysis for movable antenna enabled wireless communications,” IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6234–6250, Jun. 2024. [2] L. Zhu, W. Ma, B. Ning, and R. Zhang, “Movable-antenna enhanced multiuser communication via antenna position optimization,” IEEE Trans. Wireless Commun., vol. 23, no. 7, pp. 7214–7229, 2024. [3] B. Zheng, T. Ma, C. You, J. Tang, R. Schober, and R. Zhang, “Rotatable antenna enabled wireless communication and sensing: Opportunities and challenges,” IEEE Wireless Commun., pp. 1–8, 2025. [4] B. Zheng, Q. Wu, T. Ma, and R. Zhang, “Rotatable antenna-enabled wireless communication: Modeling and optimization,” IEEE Trans. Commun., vol. 74, pp. 6825–6842, 2026.
[5] C. Bocanegra, K. Alemdar, S. Garcia, C. Singhal, and K. R. Chowdhury, “NetBeam: Networked and distributed 3-D beamforming for multiuser heterogeneous traffic,” in 2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), Newark, NJ, USA, Nov. 2019, pp. 1–10. [6] E.-H. Wu, Y.-Q. Liu, T.-C. Xu, L.-X. Ren, Y.-M. Qin, M.-Y. Wei, X.-W. He, D.-Y. Yuan, W.-C. Hou, Z.-W. Ma et al., “Physical AI: Evolution, progress, challenges, and prospects,” J. Comput. Sci. Technol., vol. 41, pp. 271–288, Apr. 2026. [7] A. Taya, Y. Nishiyama, and K. Sezaki, “A mechanical Wi-Fi antenna device for automatic orientation tuning with Bayesian optimization,” in 2026 IEEE 23rd Consumer Communications & Networking Conference (CCNC), Las Vegas, NV, USA, Jan. 2026, pp. 1–2. [8] I. Uchendu and J. R. Kelly, “Survey of beam steering techniques available for millimeter wave applications,” Progress in electromagnetics research B, vol. 68, pp. 35–54, 2016. [9] M. I. Abbasi, M. H. Dahri, M. H. Jamaluddin, N. Seman, M. R. Kamarudin, and N. H. Sulaiman, “Millimeter wave beam steering reflectarray antenna based on mechanical rotation of array,” IEEE Access, vol. 7, pp. 145 685–145 691, 2019. [10] S. Kutty and D. Sen, “Beamforming for millimeter wave communications: An inclusive survey,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 949–973, 2016. [11] J. Niemela and J. Lempiainen, “Impact of mechanical antenna downtilt on performance of WCDMA cellular network,” in 2004 IEEE 59th Vehicular Technology Conference, vol. 4, Milan, Italy, May 2004, pp. 2091–2095. [12] F. Athley and M. N. Johansson, “Impact of electrical and mechanical antenna tilt on LTE downlink system performance,” in 2010 IEEE 71st Vehicular Technology Conference, Taipei, Taiwan, May 2010, pp. 1–5. [13] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of MIMO channels,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp. 684–702, 2003. [14] H. SHI, R. V. Prasad, E. Onur, and I. Niemegeers, “Fairness in wireless networks: Issues, measures and challenges,” IEEE Commun. Surveys Tuts., vol. 16, no. 1, pp. 5–24, 2014. [15] R. Garnett, Bayesian optimization. Cambridge University Press, 2023. [16] Z. Jiang, T. H. Luan, X. Ren, D. Lv, H. Hao, J. Wang, K. Zhao, W. Xi, Y. Xu, and R. Li, “Eliminating the barriers: Demystifying Wi-Fi baseband design and introducing the PicoScenes Wi-Fi sensing platform,” IEEE Internet Things J., vol. 9, no. 6, pp. 4476–4496, 2022. [17] S. Merlin et al., “TGax simulation scenarios,” Nov. 2015, doc.: IEEE 802.11-14/0980r16.