From Traditional Automation to Embodied Wireless Intelligence: Vision-Language-Action Empowered Physics-Aware Communication Networks

SUBMITTED FOR REVIEW

1

From Traditional Automation to Embodied Wireless Intelligence: Vision-Language-Action Empowered Physics-Aware Communication Networks

arXiv:2606.13458v1 [cs.NI] 11 Jun 2026

Genze Jiang, Kezhi Wang, Senior Member, IEEE, Xiaomin Chen, Yizhou Huang

Abstract—Wireless network automation has progressed from rule-based self-organising networks (SON) to data-driven optimisation, yet existing systems remain fundamentally disembodied, optimising performance indicators without perceiving the physical environment that governs radio propagation. We propose the embodied intelligent empowered base station (eBS), a paradigm that adopts a Vision-Language-Action (VLA) pipeline to transform base stations into autonomous agents capable of situated perception, causal physical reasoning, and physics-aware action generation. The eBS employs a two-tier asynchronous architecture: a Semantic Planner powered by a frontier VisionLanguage Model (VLM) generates structured action directives on human timescales, whilst a Tactical Controller executes real-time adaptation. Case studies demonstrate that a single VLA pipeline, without task-specific training, can perform zero-shot material reasoning, generalise across viewpoints, and predict dynamic events before signal degradation occurs, illustrating a paradigm shift from traditional rule-following network automation to embodied intelligence empowered future wireless networks. Index Terms—Vision-Language-Action Models, Embodied AI, Autonomous Network Control, 6G

I. I NTRODUCTION Future wireless networks aim for more than efficient bit transport. Instead, they promise infrastructure that actively perceives and interacts with the physical world. Yet despite successive generations of network automation, from rule-based self-organisation to data-driven optimisation, today’s systems still react to performance metrics without understanding the physical environment around them. Bridging this gap requires a fundamental shift from traditional automation to embodied intelligence. A. From Traditional Automation to Embodied Intelligence In manufacturing, early industrial robots operated by executing fixed, preprogrammed procedures, achieving automation but not intelligence. They could not perceive their environment, reason about unexpected situations, or adapt their behaviour to novel conditions. The emergence of embodied intelligence has transformed this paradigm. For example, modern robots, powered by foundation models such as VisionLanguage-Action (VLA) architectures [1], observe their surroundings through cameras and sensors, reason about objects Genze Jiang, Kezhi Wang and Yizhou Huang are with the Department of Computer Science, Brunel University London, UK (e-mail: [email protected], [email protected], [email protected]). Xiaomin Chen is with the Department of Computer Science, University of Reading, UK (e-mail: [email protected])

and spatial relationships using natural language, and generate adaptive actions grounded in physical understanding [2]. This transition represents a fundamental shift from systems that follow rules to intelligent agents that understand the world. Wireless network management has followed a remarkably parallel trajectory, yet remains delayed at the intelligent stage. Self-organising networks (SON), introduced during the 4G era, automate network operations through predefined rules and threshold-based triggers [3]. More recently, data-driven approaches have applied machine learning (ML) to network optimisation [4], and concepts such as radio access network intelligent controllers (O-RAN RIC) have further advanced this automation. However, all of these approaches operate without genuine environmental awareness. A base station equipped with such systems can detect that signal quality has degraded, but cannot perceive the physical cause of the degradation, reason about the electromagnetic properties of surrounding materials, or predict that a moving object will disrupt a propagation path in the near future. In short, current “autonomous” networks automate reactions to symptoms without diagnosing the underlying physical causes. The critical insight motivating this work is that the same embodied intelligence paradigm transforming robotics can also bridge this gap in wireless communication. By equipping base stations (BS) with the capability to perceive, reason about, and act on their physical environment using foundation models, we can move beyond rule-following automation toward genuinely intelligent, autonomous operation, in which BS can observe, understand, and continuously optimise their own behaviour. B. Vision-Language-Action (VLA) Paradigm Recent work has begun to apply vision and language models to wireless communications, but these efforts fall short of embodied intelligence. Supervised deep learning (DL) approaches train task-specific models on paired sensor-channel data that cannot generalise beyond their training distribution or adapt to novel deployments without retraining. More recent work has integrated VLMs and LLMs into wireless systems, whether through contrastive learning frameworks [5], fine-tuned multimodal models [6], LLM-based network agents [7], or large multi-modal models as enhanced feature extractors [8]. However, none implements a complete perception-reasoning-action loop. These models serve as enhanced processing components within traditional pipelines rather than as embodied agents capable of autonomous decision-making.

SUBMITTED FOR REVIEW

The robotics community has established the paradigm needed to close this gap. Early work demonstrated that prompting general-purpose language models to generate structured action representations (i.e., via executable code, JSON commands, or spatial value maps) enables embodied agents to perform complex tasks without task-specific training [2]. This prompting-based approach was subsequently formalised by VLA models [1], which unify visual perception, languagebased reasoning, and action generation within a single architecture. Recent proposals for embodied intelligence in future networks [9] provide high-level architectural visions but lack concrete physical-layer implementations that demonstrate how embodied intelligence translates into network control actions. This paper bridges the gap between the robotics VLA paradigm and wireless infrastructure by proposing the first complete VLA pipeline for autonomous BS operation.

C. Contribution: Embodied Intelligence Empowered BS We propose an embodied intelligence empowered BS (eBS) paradigm that implements a VLA pipeline for autonomous network control, where a frontier VLM serves as the central perception-reasoning-action engine. The eBS paradigm embodies three defining characteristics that fundamentally distinguish it from both rule-based network automation (e.g., SON) and existing data-driven AI approaches. Situated perception: The eBS constructs coherent semantic representations of its local environment through multimodal sensing, identifying objects, materials, spatial relationships, and dynamic trajectories via natural language reasoning rather than treating sensor data as isolated feature vectors. Causal reasoning: The eBS infers physical consequences from visual observations without explicit electromagnetic simulation, reasoning about how material properties affect signal propagation, how spatial geometry constrains coverage, and how dynamic objects will alter channel conditions, all through chain-of-thought reasoning grounded in pre-trained world knowledge. Autonomous action: The eBS closes the perceptionreasoning-action loop by autonomously translating environmental understanding into structured control actions, from selecting transmission strategies based on identified material properties to initiating proactive handover based on predicted mobility patterns. The eBS reasons from first principles and generates appropriate actions for situations never previously encountered, whilst supporting operator guidance through natural language instructions when desired. The fundamental engineering challenge lies in reconciling two incompatible timescales. VLMs operate on human cognitive timescales, requiring 1–2 seconds for inference, whilst physical layer control demands responses within 10 milliseconds. The eBS resolves this through a two-tier asynchronous architecture that decouples strategic reasoning from tactical execution, enabling the system to leverage VLM intelligence without violating real-time constraints. Section II details the eBS design, including its system architecture, VLA agent formulation, and safety mechanisms.

2

II. E MBODIED I NTELLIGENT BASE S TATION A. Two-Tier System Architecture Since 6G channels exhibit coherence times of 10–50 milliseconds [10] whilst foundation models require over 1 second for inference, direct VLM integration proves infeasible. We resolve this through a two-tier decomposition, partitioning intelligence across tiers optimised for distinct temporal scales, as illustrated in Fig. 1. Tier 1: The Semantic Planner: This component embeds a frontier VLM functioning as the perception-reasoning stage of the VLA pipeline, processing multimodal data including camera imagery, historical traffic patterns, and operator instructions. Operating asynchronously at update rates of 0.5– 2.0 Hz, the planner does not directly compute physicallayer parameters. Rather, it employs semantic reasoning to translate visual observations into structured control directives, generating high-level intent vectors that constrain the tactical controller’s action space. In the beam management context explored in this work, this translates visual observations (for instance, “user in right lane”) into spatial constraints that define the beam search space. Tier 2: The Tactical Controller: This component implements a lightweight real-time executor operating within the 10 millisecond radio frame constraint. The controller receives intent vectors from Tier 1 and restricts its optimisation to the semantically defined subset of the action space. For beam management, this means restricting the pilot search to a subset of beam indices. For other physical-layer functions (e.g., power control, handover, scheduling), the same architecture applies with task-appropriate executors. This design enables the system to respond to fast-fading channels whilst operating within the strategic constraints defined by the VLA pipeline. B. Information Flow and Asynchronous Coupling The critical architectural innovation lies in the asynchronous coupling between tiers. The Tactical Controller maintains local state and continues to optimise physical-layer configurations using real-time feedback, even while the Semantic Planner processes a new frame. This design ensures that physicallayer control never blocks while waiting for VLM inference completion. The architecture exhibits three desirable properties that enhance robustness and flexibility: • Graceful degradation: If VLM inference is delayed or the visual link becomes obstructed, the Tactical Controller reverts to conventional operation (e.g., exhaustive beam search), ensuring that system reliability never falls below baseline. • Progressive refinement: As the Planner accumulates observations, it progressively narrows the Tactical Controller’s action space, reducing the beam search sector from 30° to 5° and lowering pilot overhead accordingly. • Operator-in-the-loop control: Network engineers can inject natural language instructions (such as “prioritise the emergency vehicle”), which the VLM translates into immediate intent vector updates without requiring direct manipulation of physical layer parameters.

SUBMITTED FOR REVIEW

3

VLM Inputs BS Vision Operator Instruction

Tier 1: Semantic Planner (VLA Engine) Human Timescale (0.5 - 2s)

Physics Priors (Textual Domain Knowledge)

Frontier VLM Situated Perception & Casual Reasoning (Chain-of-Thought)

JSON Intent Vector

Action Generation (Structured Directives)

Safety Validation (Check Parameters & Confidence)

{

}

"semantic_state": "…" "action": "ACTIVATE_SECTOR", "target_beams": [38-42], … "reasoning": "…"

Tier 2: Tactical Controller (Real-Time Executor) Radio Frame Timescale (~10ms) Intent Vector Buffer Real-Time Feedback (Channel State, etc)

Tactical Controller Restrict Action Space (Based on Intent)

Real-Time Optimisation & Execution

Physical Layer Actions (e.g., Beam Selection, Power Control)

Physical Wireless Environment

Fig. 1: The eBS system architecture. The Semantic Planner (Tier 1) operates on human timescales to generate semantically informed control directives, enabling the Tactical Controller (Tier 2) to execute real-time physical-layer adaptation within the 10-millisecond radio frame constraint.

C. Intent Vector Representation We define the intent vector as the structured interface between the VLA pipeline’s semantic reasoning and physical layer execution: a JSON-formatted payload that converts natural-language reasoning into constraints compatible with physical layer algorithms. In the beam management context, it encodes beam-priority masks, reflection annotations, and temporal validity. For example, upon identifying a user vehicle in the right traffic lane, the VLM generates: { "semantic_state": "User in Right Lane", "action": "ACTIVATE_SECTOR", "target_beams": [38, 39, 40, 41, 42], "exclusion_zone": [0, 1, ..., 20], "reasoning": "Vehicle constrained to lane geometry; Building at 50deg provides NLoS backup." }

The Tactical Controller parses this vector and restricts the pilot transmission to indices 38–42 in subsequent transmission frames, effectively reducing the search space by 92% compared to full 64-beam sweeps. D. VLA-based Agent Design Applying VLA principles to wireless infrastructure requires defining three elements absent from the traditional robotics domain: an observation space grounded in RF-relevant scene semantics, an action space mapped to physical layer control parameters, and a policy that bridges visual perception to control commands. We instantiate these design choices for beam management, though the framework generalises to other physical-layer functions. 1) Observation Space: The agent’s observation at each planning cycle comprises an RGB image from the basestation-mounted camera concatenated with a textual physics prior that injects domain knowledge absent from the VLM’s general pre-training:

SYSTEM: You are controlling a 6G base station at 28 GHz. Your goal is to maintain high-SNR connectivity. PHYSICS PRIORS: - 28 GHz signals are blocked by Concrete/Metal (>20dB ,→ loss). - 28 GHz signals penetrate Glass (3-6dB loss). - Reflection off flat surfaces is viable for NLoS paths. OUTPUT FORMAT: JSON only. No conversational preamble.

This hybrid observation (visual + textual) enables the VLM to apply pre-trained world knowledge spanning electromagnetic material properties, projective geometry, and causal reasoning to the wireless domain. 2) Action Space: The wireless VLA agent operates over discrete, structured control parameters rather than the continuous joint torques of robotic agents. For beam management, the action space comprises the set of all valid intent vectors, each specifying a beam priority mask over the 64-element codebook, a transmission strategy (direct, reflect, or penetrate), and associated power and beamwidth parameters. The same intent vector framework accommodates other physical layer actions by extending the JSON schema. 3) Policy Specification via Structured Prompting: In endto-end VLA models such as RT-2 [1], the policy is learned implicitly through supervised fine-tuning on vision-action pairs. In our prompting-based approach, the policy is specified explicitly through the prompt structure via two mechanisms. First, we enforce chain-of-thought action derivation by requiring the VLM to populate a reasoning field prior to the action field in its JSON output, ensuring the model articulates its physical interpretation before committing to an action. Second, we inject task-specific observation queries that focus the agent’s attention on the relevant physical dimension: • Material reasoning (Case Study A): “Identify the blockage material and determine whether direct transmission or reflection paths are optimal at 28 GHz.”

SUBMITTED FOR REVIEW

Spatial reasoning (Case Study B): “Estimate the user equipment’s semantic spatial position (left, centre, right lane) and map to corresponding beam sector indices.” • Temporal reasoning (Case Study C): “Analyse dynamic object trajectories and determine whether any will intersect the line-of-sight path within 500 milliseconds.” •

4

SCENARIO A1: CONCRETE

SCENARIO A2: GLASS

USER [ID: 01]

USER [ID: 01]

SCANNING OBJECT... TYPE: OBSTACLE MATERIAL: CONCRETE (OPAQUE) RF PROP: REFLECT (BOUNCE)

SCANNING OBJECT... TYPE: OBSTACLE MATERIAL: GLASS (DIELECTRIC) RF PROP: TRANSMIT (BOOST)

E. Closed-Loop Safety and Fallback In wireless deployment, an incorrect control action causes immediate performance degradation such as throughput loss and dropped connections. The Semantic Planner therefore validates every generated action before forwarding it to the Tactical Controller. The parser verifies that recommended parameters fall within valid ranges (e.g., beam indices 0–63) and evaluates the VLM’s stated confidence level. If confidence falls below a predefined threshold (0.7 in our implementation) or the output contains malformed JSON, the system reverts to conventional operation, ensuring that performance never degrades below baseline. F. Multi-Station Coordination In conventional cooperative schemes, BS exchange raw channel state information, imposing stringent backhaul bandwidth and latency requirements that scale poorly with network density. By contrast, the eBS Semantic Planner already produces structured intent vectors that encode environmental understanding in compact, human-readable form. These vectors can be shared between neighbouring eBS nodes over standard backhaul interfaces, enabling semantic-level coordination. For instance, a source eBS that observes a vehicle approaching a cell boundary can transmit a trajectory prediction and material description to the target eBS, which pre-configures its beam before handover occurs. This semantic coordination operates at the Planner’s low update rate (0.5–2 Hz), placing negligible additional load on existing backhaul infrastructure whilst enabling a richer form of inter-cell cooperation than numerical parameter exchange alone. III. C ASE S TUDIES To validate the eBS as a VLA agent, we conduct three case studies using NVIDIA Sionna [11] for physics-compliant raytracing simulations and the Vision-Wireless (ViWi) dataset [12] for dynamic vehicular scenarios. Each study stresses a distinct capability of the VLA pipeline, namely material reasoning, viewpoint-invariant spatial reasoning, and temporal reasoning, demonstrating that embodied intelligence emerges from the foundation model’s pre-trained world knowledge rather than from task-specific pattern matching. A. Case Study A: Material Reasoning for Beam Strategy 1) Challenge: The Semantic Gap in Channel Estimation: In high-frequency bands such as mmWave and THz, the electromagnetic properties of obstacles fundamentally determine link viability. Conventional channel estimators treat blockages as binary, assuming that either line-of-sight exists or it does not, and therefore cannot distinguish between a concrete wall

Fig. 2: The embodied agent identifies material properties from visual cues. Left: Detecting concrete triggers a REFLECT strategy. Right: Detecting glass triggers a PENETRATE strategy. requiring alternative reflective paths and a glass window that permits attenuated transmission. This semantic gap leads to suboptimal beam selection. 2) Experimental Setup: We constructed a high-fidelity digital twin using the Sionna ray-tracer. The scene depicts an urban canyon with a base station positioned at 25m height and a user vehicle at 40m distance. We introduced a variable blockage directly in the line-of-sight path: • Scenario A1 (Concrete): A solid concrete barrier (ϵr ≈ 5.31, high loss tangent). • Scenario A2 (Glass): A transparent glass panel (ϵr ≈ 6.27, low loss tangent). The embodied agent (Tier 1: Semantic Planner) receives RGB images of the scene with the prompt: “Analyse this base station’s camera image. A user equipment is located behind the visible obstacle. Identify the blockage material and recommend a transmission strategy considering 28 GHz millimetre wave propagation.” 3) Results: The VLM successfully bridges the semantic gap through zero-shot material inference, as shown in Fig. 2. In Scenario A1 (Concrete), the agent identifies the barrier as “opaque and solid.” The reasoning trace states: “Reflection is necessary... a wide beam is used due to potential distortion from the surface.” Consequently, the agent issued a directive to REFLECT with HIGH power and a WIDE beam width to maximise signal capture from non-line-of-sight bounces. In Scenario A2 (Glass), the agent recognises the transparency. The reasoning trace adapts: “Glass allows for penetration with slight attenuation... the view is clear, allowing for a narrow beam width for precision.” The agent commands a PENETRATE strategy with MEDIUM power and a NARROW beam width to minimise interference. TABLE I: VLM Decision Matrix for Material Reasoning Scene

Material

Strategy

Power

Beam Width

A1 A2

Concrete (Opaque) Glass (Dielectric)

REFLECT PENETRATE

HIGH MEDIUM

WIDE NARROW

These decisions, summarised in Table I, match the groundtruth optimal configuration. Sionna simulations confirm that the REFLECT strategy avoids a 20dB loss through the concrete, while the PENETRATE strategy successfully maintains

SUBMITTED FOR REVIEW

5

6HPDQWLF5,*+7/$1(

6HPDQWLF5,*+7/$1(

1RUPDOL]HG*DLQG%

D/RZ7RZHU9LHZP

E+LJK7RZHU9LHZP 6HDUFK5HGXFWLRQ

([KDXVWLYH6HDUFK6SDFH /RZ7RZHU6HOHFWLRQ%HDPV +LJK7RZHU6HOHFWLRQ%HDPV



3HUIHFW$OLJQPHQW

 









$]LPXWK$QJOH'HJUHHV







F3URRIRI7RSRORJLFDO,QYDULDQFH%HDP3DWWHUQV2YHUOD\3HUIHFWO\ Fig. 3: (a)-(b) The agent consistently identifies the “Right Lane” semantic region despite a 30m height shift. (c) This semantic stability yields identical beam selections (Beams 40–44) for both views. The perfect overlap of the Low Tower (Blue) and High Tower (Red Dashed) patterns confirms viewpoint invariance and achieves a 92% reduction in search overhead.

a link through the glass with only 4dB attenuation. This demonstrates that the eBS can infer electromagnetic properties from visual semantics through zero-shot VLA reasoning, eliminating the latency overhead of blind beam sweeping by pre-configuring the physical layer according to actual material properties. B. Case Study B: Viewpoint-Invariant Spatial Reasoning for Beam Prediction 1) Challenge: Perspective-Dependent Overfitting: Traditional supervised learning models (e.g., ResNet-based beam predictors) suffer from severe overfitting to the specific camera perspective present in training data. When the camera is relocated to a different height, pixel coordinates shift substantially, causing incorrect beam predictions. True embodiment requires viewpoint invariance, the capacity to understand that a user’s semantic location remains constant even when the observer’s height changes. 2) Experimental Setup: The eBS addresses this through semantic abstraction. The Semantic Planner does not directly predict raw angles. Rather, it identifies semantic spatial regions, e.g., “right lane”, “intersection centre” and maps them to physical beam sectors. We evaluate this capability by generating two distinct viewpoints of an identical scene in Sionna: 1) Low Tower View (15m): The camera operates close to ground level and road markings converge sharply towards the horizon. 2) High Tower View (45m): The camera observes from an elevated vantage point and the scene resembles a 2D map. To isolate azimuthal invariance, we employ a 64-element Uniform Linear Array (ULA) with vertically broad fan beams

spanning −60◦ to +60◦ , ensuring that users at different vertical angles remain within the same azimuthal beam index. 3) Results: In the Low Tower view, the agent identifies the vehicle in the rightmost traffic lane. Despite perspective distortion, the agent reasons: “The vehicle occupies the right lane based on road marking geometry... Azimuth angle is approximately +15◦ .” In the High Tower view, despite the vehicle appearing more centrally due to the top-down perspective, the agent maintains the semantic classification: “Vehicle remains in the right lane. High-angle observation confirms lateral position. Azimuth sector remains +15◦ .” As shown in Fig. 3, the selected beam subsets perfectly overlapped. This confirms that the embodied agent’s VLA pipeline effectively disentangles observer state (camera height) from environment state (user location), a hallmark of embodied spatial reasoning. Semantic guidance constrains the Tactical Controller’s search space from 64 beams to approximately 5 beams (beams 38–42 for the right lane scenario), a 92% reduction in beam training overhead that proves critical for maintaining channel coherence in mobile scenarios. C. Case Study C: Temporal Reasoning for Predictive Handover 1) Challenge: Reactive Beam Management Limitations: Standard 5G beam management operates reactively with beam search or handover procedures triggering only after Reference Signal Received Power (RSRP) degradation exceeds predefined thresholds (typically 3 dB). In dense urban environments with dynamic blockers (e.g., buses, delivery vehicles, construction equipment), this reaction latency often exceeds channel coherence time, resulting in connection interruptions and latency spikes. Recent work has demonstrated that vision-aided

SUBMITTED FOR REVIEW

6

DW V*DS!P 5LVN/2:

FW V2FFOXVLRQ,PPLQHQW 5LVN+,*+

EW V*DSFORVLQJ 5LVN0(',80

+,*+

+$1'29(5

0(',80 /2: 0$,17$,1 





7LPHV





&RQWURO$FWLRQ

6HPDQWLF5LVN/HYHO

+DQGRYHU7ULJJHU 6HPDQWLF5LVN6FRUH +DQGRYHU&RPPDQG



G7HPSRUDO5LVN3URILOH &RQWURO$FWLRQ Fig. 4: (a)-(c) The Semantic Planner monitors the trajectory of dynamic vehicles, elevating the semantic blockage risk score as the gap closes. (d) At t = 1.9s, the agent predicts imminent occlusion and generates a proactive handover directive 100ms before physical link degradation occurs at t = 2.0s.

approaches can substantially reduce such service interruptions through predictive handover [13]. However, these methods rely on task-specific supervised models trained on paired camera-channel data. A key question for the eBS paradigm is whether a general-purpose VLA pipeline, without taskspecific training, possesses temporal reasoning capability, that is, the capacity to reason about dynamic scenes, predict future channel states from visual trajectory analysis, and generate pre-emptive control actions. 2) Experimental Setup: We utilise the ViWi dataset, specifically the co-located camera scenario featuring continuous image sequences of vehicles in dynamic street environments. A serving base station is positioned on the left side of a two-lane road, with a neighbouring base station covering the opposite direction. As the user vehicle approaches a large bus from behind, the bus progressively occludes all feasible beam paths, necessitating handover to the neighbouring base station. We define a constant user velocity of 10 m/s (36 km/h), yielding an effective sampling rate of 10 Hz, and instruct the agent to calculate a semantic risk score (Low, Medium, High) based on trajectory analysis relative to the bus position, as illustrated in Fig. 4. The control policy was defined as: • Risk < High: Maintain current beam tracking via Tactical Controller. • Risk = High: Issue proactive handover directive to Tactical Controller. 3) Results: Temporal analysis reveals the VLA pipeline’s predictive reasoning capability, depicted in the risk transitions of Fig. 4(d). From t = 0 to t = 1.5s, the agent classified risk as “Low” due to sufficient spatial separation between the vehicle and bus. As the gap closed (t = 1.5 to t = 1.9s), risk elevated to “Medium”, with the reasoning trace noting convergence at approximately 2 metres per second. At t = 1.9s, the agent identified imminent occlusion and escalated to “High”, generating a proactive handover directive for the Tactical Controller.

Ground-truth channel measurements confirm line-of-sight loss at t = 2.0s with RSRP degradation of 12 dB. The VLA pipeline thus generates the appropriate handover action 100 milliseconds before physical degradation occurs, whereas a reactive baseline system would detect the problem only at t = 2.0s. This demonstrates that the eBS can track dynamic objects, estimate closing trajectories, and infer causal consequences for link quality through zero-shot reasoning rather than task-specific supervision. IV. D ISCUSSION AND F UTURE C HALLENGES A. Inference Latency and Model Evolution The prompting-based approach requires no paired visionaction training data, instead leveraging pre-trained world knowledge for zero-shot action generation. However, current frontier VLMs require 1–2 seconds for complex visual reasoning via cloud APIs. Whilst the two-tier architecture masks this latency for strategic planning, reducing inference time would enable finer-grained temporal predictions. Two complementary directions merit investigation: distilling frontier VLM knowledge into compact edge-deployable models at 100– 200 millisecond latencies, and training a dedicated wireless VLA model end-to-end on paired camera-beam datasets to replace prompt-dependent reasoning with learned wirelessspecific representations. B. Edge Deployment and Energy Efficiency The case studies invoke frontier VLMs through cloud APIs, but current edge accelerators such as the NVIDIA Jetson AGX Orin already support VLMs in the 7–20 billion parameter range at power budgets compatible with base station enclosures, whilst next-generation platforms promise time-to-firsttoken below 200 milliseconds within 40–130 W envelopes, suggesting that fully self-contained eBS operation is within near-term reach. A single frontier VLM inference consumes

SUBMITTED FOR REVIEW

7

approximately 10–15 watt-hours compared to 0.01 watt-hours for a conventional beam prediction neural network, however, the Semantic Planner’s low update rate (0.5–2 Hz) limits total inference volume, and the 92% beam search reduction demonstrated in Case Study B directly reduces RF pilot transmission energy. A comprehensive lifecycle analysis should quantify whether the Planner’s computational overhead is offset by these downstream savings. C. Real-World Robustness and the Sim-to-Real Gap The case studies in Section III validate the eBS using ray-traced digital twins and controlled vision-wireless datasets. However, operational BS imagery introduces additional challenges including lens flare, low-light conditions, weather-induced occlusions, and visual clutter. Datasets such as DeepSense 6G [14], which provide synchronised camera and mmWave measurements from real urban infrastructure, offer a natural testbed for evaluating this transfer. Systematic validation under extreme weather and nighttime conditions remains an important direction for future work. D. Privacy, Explainability, and Sensor Fusion Deploying cameras on telecommunications infrastructure raises privacy concerns. The eBS mitigates this through semantic abstraction, ensuring that only high-level scene descriptors propagate beyond the edge whilst raw imagery remains local. Future work should investigate privacy-preserving modalities (e.g., LiDAR, radar, thermal) that provide geometric information without identifying individuals. A complementary advantage of the VLM-based approach is interpretability. Unlike black-box neural networks, the Semantic Planner produces natural language explanations of its beam decisions, enabling operators to validate reasoning and progressively build trust in autonomous operation. Finally, extending the eBS to fuse heterogeneous sensor data through structured prompting represents a natural progression, as recent work demonstrates that multimodal foundation model performance improves with modality diversity [15]. V. C ONCLUSION This work introduces the eBS, a paradigm that implements a VLA pipeline for autonomous network control, where a frontier VLM perceives the physical environment, reasons about propagation physics, and generates structured control actions. A two-tier asynchronous architecture decouples strategic semantic reasoning from tactical execution, reconciling foundation model capabilities with real-time wireless requirements. By equipping base stations with the full VLA loop, comprising situated perception, causal reasoning, and autonomous action generation with natural language explainability, the eBS paradigm represents a shift from reactive signal processors and rule-following network automation to embodied intelligence, enabling wireless infrastructure that observes, understands, and autonomously optimises its interaction with the physical world.

R EFERENCES [1] B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid et al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” in Conference on Robot Learning. PMLR, 2023, pp. 2165–2183. [2] J. Liang, W. Huang, F. Xia, P. Xu, K. Hausman, B. Ichter, P. Florence, and A. Zeng, “Code as policies: Language model programs for embodied control,” in 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 9493–9500. [3] P. V. Klaine, M. A. Imran, O. Onireti, and R. D. Souza, “A survey of machine learning techniques applied to self-organizing cellular networks,” IEEE Communications Surveys & Tutorials, vol. 19, no. 4, pp. 2392–2431, 2017. [4] L. Bariah and M. Debbah, “Ai embodiment through 6G: Shaping the future of agi,” IEEE Wireless Communications, vol. 31, no. 5, pp. 174– 181, 2024. [5] J. Wang, B. Tang, J. Xiao, Q. Cui, X. Li, and T. Q. Quek, “When visionlanguage model (VLM) meets beam prediction: A multimodal contrastive learning framework,” arXiv preprint arXiv:2508.00456, 2025. [6] Y. Zhao, L. Yu, L. Shi, J. Zhang, and G. Liu, “Multi-modal large models based beam prediction: An example empowered by DeepSeek,” arXiv preprint arXiv:2506.05921, 2025. [7] J. Tong, W. Guo, J. Shao, Q. Wu, Z. Li, Z. Lin, and J. Zhang, “Wirelessagent: Large language model agents for intelligent wireless networks,” arXiv preprint arXiv:2505.01074, 2025. [8] S. Xu, C. K. Thomas, O. Hashash, N. Muralidhar, W. Saad, and N. Ramakrishnan, “Large multi-modal models (LMMs) as universal foundation models for ai-native wireless systems,” IEEE Network, 2024. [9] Z. Li, Z. Gao, X. Liu, Z. Wang, X. Zhou, L. Liu, Y. Wu, W. Feng, and Y. Huang, “Large model enabled embodied intelligence for 6G integrated perception, communication, and computation network,” arXiv preprint arXiv:2512.15109, 2025. [10] G. Charan, M. Alrabeiah, T. Osman, and A. Alkhateeb, “Camera based mmWave beam prediction: Towards multi-candidate real-world scenarios,” IEEE Transactions on Vehicular Technology, 2024. [11] J. Hoydis, F. Aït Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) Workshops, 2023, pp. 317–321. [12] M. Alrabeiah, A. Hredzak, Z. Liu, and A. Alkhateeb, “ViWi: A deep learning dataset framework for vision-aided wireless communications,” in 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring). IEEE, 2020, pp. 1–5. [13] G. Charan, M. Alrabeiah, and A. Alkhateeb, “Vision-aided 6G wireless communications: Blockage prediction and proactive handoff,” IEEE Transactions on Vehicular Technology, vol. 70, no. 10, pp. 10 193– 10 208, 2021. [14] A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais, U. Demirhan, and N. Srinivas, “DeepSense 6G: A large-scale real-world multi-modal sensing and communication dataset,” IEEE Communications Magazine, vol. 61, no. 9, pp. 122–128, 2023. [15] C. Zheng, J. He, G. Cai, Z. Yu, and C. G. Kang, “BeamLLM: Visionempowered mmWave beam prediction with large language models,” in 2025 IEEE 102nd Vehicular Technology Conference (VTC2025-Fall). IEEE, 2025, pp. 1–6.

Genze Jiang is working toward the Ph.D. degree with the Department of Computer Science, Brunel University London, UK. Kezhi Wang (Senior Member, IEEE) is a Professor with the Department of Computer Science, Brunel University London, UK. His research interests include wireless communications, mobile edge computing, and machine learning. Xiaomin Chen is an Associate Professor with the Department of Computer Science, School of Mathematical, Physical and Computational Sciences, University of Reading. Yizhou Huang is working as a research fellow with the Department of Computer Science, Brunel University London, UK.