PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Semantic Internet of Everything in the Age of Agentic AI

Dayu Fan, Rui Meng, Yunfei Liu, Shuai Ai, Xiaodong Xu, Yiming Liu, Huishi Song, Han Meng, Lexi Xu, Ping Zhang

arXiv:2609.02924v1eess.SP

TL;DR

Task-specific semantic communication pipelines are difficult to reuse across models, applications, and deployment settings. SIoE coordinates profiled communication and AI capabilities through SLA-guided planning and validation; a lightweight V2X case study illustrates condition-dependent capability selection under service constraints.

  • Problem

    Task-specific semantic communication pipelines and modular workflow stages remain difficult to reuse and compose across tasks, models, and deployments.

    Source: §I · p. 1, A. Task-Oriented Semantic Communication · p. 2

  • Method

    SIoE models heterogeneous communication and AI functions as profiled services, then uses semantic SLAs to guide validated capability composition and adaptation.

    Source: §I · p. 2

  • Results

    In the V2X case study, the planner selected SwinJSCC at 1 and 4 dB and the digital capability at 7, 10, and 13 dB as channel conditions changed.

    Source: C. SLA-Grounded Planning and Task… · p. 7

  • Takeaways & Limitations

    The prototype illustrates that service objectives can guide capability selection without fixing a single communication implementation.

    Source: §VI · p. 8

Abstract

from arXiv · show

Semantic communication improves task effectiveness by transmitting task-relevant information. However, most existing schemes remain organized as task-specific, end-to-end pipelines, which are difficult to reuse across models, applications, and deployment environments. Against this background, we propose the Semantic Internet of Everything (SIoE), a composable service architecture that represents heterogeneous communication and artificial intelligence (AI) functions as capability-profiled services and coordinates them according to application objectives. SIoE comprises three planes: a task and service plane, an agentic orchestration plane, and a semantic capability plane. In this framework, task requirements are captured via a semantic service-level agreement (SLA), while an agentic planner discovers and composes candidate capabilities under deterministic compatibility, resource, privacy, and policy validation. Feedback from the communication, semantic, and task levels enables continuous adaptation and replanning. A lightweight vehicle-to-everything case study illustrates profile-grounded capability planning under explicit service constraints. The results demonstrate the feasibility of decoupling service objectives from fixed communication implementations and also highlight key open challenges, including semantic SLA design, capability interoperability, scalable planning, and trustworthy execution.

I. INTRODUCTION

SIoE addresses the difficulty of reusing task-specific semantic communication pipelines by coordinating heterogeneous communication and AI capabilities as profiled services. Semantic SLAs guide validated, adaptive capability planning, illustrated by a lightweight V2X prototype.

I. INTRODUCTION

  • Most semantic communication schemes remain fixed, task-specific pipelines, while modularizing one workflow does not enable reuse across tasks, models, and deployments.Changing objectives, network conditions, deployment scenarios, or computing resources can require manual reconfiguration, retraining, or replacement.Source: §I · p. 1
  • SIoE represents heterogeneous communication and AI functions as reusable, machine-profiled services for task-driven coordination.Profiles describe interfaces, operating conditions, compatibility dependencies, deployment requirements, resource costs, and expected task performance.Source: §I · p. 2
  • A semantic SLA captures application requirements, and deterministic compatibility and policy validation checks agent-proposed pipelines before execution.The SLA may include task output and quality, semantic distortion, latency, bandwidth, reliability, energy, privacy, and deployment preferences; feedback supports adaptation.Source: §I · p. 2
  • The three-plane architecture separates task and service requirements, agentic orchestration, and semantic capabilities across device, edge, and cloud resources.The orchestration plane discovers, composes, validates, executes, and adapts capability pipelines.Source: §I · p. 2
  • A lightweight V2X prototype demonstrates SLA-constrained capability selection and identifies open challenges in capability specification, interoperability, orchestration, and trustworthy execution.The paper frames Agentic AI as constrained orchestration, not a replacement for specialized models.Source: §I · p. 2

II. FROM SEMANTIC COMMUNICATIONS TO SIOE

Semantic communication evaluates whether transmitted information remains useful for downstream tasks, but existing systems and modular extensions remain tied to predefined workflows. SIoE targets reuse and composition across tasks and deployments.

II. FROM SEMANTIC COMMUNICATIONS TO SIOE

  • SIoE connects semantic communication and Agentic AI by coordinating reusable communication and AI capabilities according to objective requirements.These technologies respectively represent and transmit task-relevant information and translate objectives into executable, adaptive workflows.Source: §II · p. 2

A. Task-Oriented Semantic Communication

  • Semantic communication broadens communication objectives from reliable bit delivery to semantic fidelity and downstream task performance.Transmitted representations may include reconstructed content, semantic features, structured messages, selected regions, or latent representations.Source: A. Task-Oriented Semantic Communication · p. 2
  • Most systems still bind representations, encoders, decoders, channel assumptions, and downstream models to specific tasks.SkillCom and SkillComm expose processing stages as replaceable skills, but remain within predefined semantic communication workflows [2], [3].Source: A. Task-Oriented Semantic Communication · p. 2

B. Agentic Orchestration and Capability Composition

The section positions SIoE as a service architecture for composing heterogeneous communication and AI functions, beyond agent interaction or modular stages within a transmission workflow. Its logical planes work across conventional infrastructure.

B. Agentic Orchestration and Capability Composition

  • Prior agent and wireless frameworks advance agent interaction and network control, but leave open how underlying communication and AI functions can form validated end-to-end pipelines.Examples include WirelessAgent++ and 6GAgentGym [5], [6].Source: B. Agentic Orchestration and… · p. 2
  • SIoE augments callable interfaces with machine-readable profiles and composes capabilities using task requirements, compatibility checks, and policy validation.The profiles cover function types, model pairings, operating and deployment conditions, resource demands, and expected task performance.Source: B. Agentic Orchestration and… · pp. 2–3

III. A THREE-PLANE REFERENCE FRAMEWORK FOR COMPOSABLE SIOE

  • The three logical planes assign application requirements to the task and service plane, pipeline translation to orchestration, and processing functions to the capability plane.They operate above and across conventional communication infrastructure, whose connectivity and resource-support mechanisms remain in place.Source: §III · p. 3
  • SIoE specifies outcomes without fixing a codec or deployment, exposes reusable profiled capabilities, validates plans before execution, and adapts them using multilevel observations.The framework identifies communication-, semantic-, and task-level feedback as inputs to adaptation.Source: §III · p. 3

A. Task and Service Plane

The task and service plane captures outcomes and constraints in a semantic SLA; the orchestration plane turns them into a validated capability workflow and adapts execution using feedback.

A. Task and Service Plane

  • A semantic SLA specifies the required task output and quality, semantic fidelity, latency, resource, and policy requirements, distinguishing hard constraints from soft preferences.Their relative importance varies by task: reconstruction may emphasize semantic fidelity, while control may prioritize a reliable decision before a deadline.Source: A. Task and Service Plane · p. 3
  • Capability selection is constrained by task requirements because similarly named functions can differ in schemas, operating regions, deployment restrictions, and task performance.The V2X example is qualitative and does not prescribe universal deployment thresholds.Source: A. Task and Service Plane · p. 3

B. Agentic Orchestration Plane

  • The orchestration plane coordinates specialized capabilities rather than directly performing semantic coding, signal processing, or task inference.Natural-language requests and other objectives are compiled into structured task profiles with hard constraints, soft preferences, and retrieval information.Source: B. Agentic Orchestration Plane · p. 3
  • Capability discovery retrieves candidates from a registry, and a dependency-aware planner organizes them into a configurable workflow.Pipelines may combine extraction, coding, transmission, reconstruction, inference, and evaluation, with parameters such as coding rate or deployment location.Source: B. Agentic Orchestration Plane · p. 4
  • Proposed pipelines undergo compatibility checks for schemas, semantic units, formats, versions, and model pairings, plus policy checks for service and safety constraints.Candidates violating hard requirements are rejected before execution.Source: B. Agentic Orchestration Plane · p. 4
  • Communication, semantic, and task observations are compared with the SLA to adapt parameters or restructure the pipeline; agent reasoning proposes workflows while deterministic mechanisms enforce constraints.Adaptation can change coding rate or model quality, replace a capability, migrate a model, or select another prevalidated pipeline.Source: B. Agentic Orchestration Plane · pp. 4–5

C. Semantic Capability Plane

The semantic capability plane hosts communication, semantic, knowledge, and task-processing functions, each with an invocation interface and a capability profile. Profiles describe suitability, while evaluation capabilities return evidence for orchestration.

C. Semantic Capability Plane

  • A capability can be a learned model, communication algorithm, knowledge service, or combination, with an interface for invocation and a profile describing suitability.The interface specifies inputs, outputs, parameters, status information, and failure responses.Source: C. Semantic Capability Plane · p. 5
  • Profiles record functional, operating, performance and resource, and dependency and governance information.These include schemas and modality, channel and source assumptions, quality and resource measures, model pairings, deployment, provenance, access, and trust attributes.Source: C. Semantic Capability Plane · p. 5
  • Matching function names is insufficient for connecting semantic encoders and decoders; latent representations, checkpoints, normalization, symbol formats, and channel assumptions must also match.An execution wrapper alone does not establish compatibility with adjacent components or service requirements.Source: C. Semantic Capability Plane · p. 5
  • Evaluation functions are capabilities that return task, semantic, confidence, and quality evidence, while provenance and integrity information supports valid adaptation decisions.Incorrect capability descriptions or manipulated feedback may result in invalid or unsafe adaptation decisions.Source: C. Semantic Capability Plane · p. 5

D. Cross-Plane Operation and Deployment

Cross-plane operation translates an objective into a task profile and semantic SLA, then retrieves, validates, and selects capabilities for execution. Measurements and task outcomes feed back into pipeline decisions, within distinct plane responsibilities and deployment and adaptation arrangements.

D. Cross-Plane Operation and Deployment

  • The requester’s objective becomes a task profile and semantic SLA; orchestration retrieves compatible capabilities, validates candidate pipelines, selects a configuration, and the capability plane executes it.Execution measurements and task outcomes return to orchestration, which retains, adjusts, or replaces the pipeline according to the SLA.Source: D. Cross-Plane Operation and Deployment · p. 5
  • The planes have distinct control boundaries: service defines the outcome, orchestration coordinates capabilities, and capability services perform registered functions without reinterpreting the task.Orchestration does not replace numerical communication or AI processing.Source: D. Cross-Plane Operation and Deployment · p. 5
  • Devices, edge platforms, and cloud platforms may host different functions, from privacy-sensitive local processing to regional validation and global repositories or model services.The described deployments include sensors and codecs on devices, registries and larger models at the edge, and repositories and foundation models in the cloud.Source: D. Cross-Plane Operation and Deployment · p. 5
  • Fast adaptation selects or changes prevalidated local pipelines using current feedback, while slower processes update profiles, versions, histories, and policies; trust, security, privacy, and access control constrain all planes.Source: D. Cross-Plane Operation and Deployment · p. 5

IV. V2X CASE STUDY: EXTENSIBLE AGENTIC PLANNING OVER PROFILED CAPABILITIES

The V2X case study instantiates SIoE as profile-grounded planning over heterogeneous communication capabilities, with deterministic enforcement of compatibility and semantic SLA constraints. It uses offline measurements as reusable profile evidence and selects at runtime by an explicit SLA-based rule.

IV. V2X CASE STUDY: EXTENSIBLE AGENTIC PLANNING OVER PROFILED CAPABILITIES

  • The lightweight V2X case study tests agentic planning over heterogeneous communication capabilities using measurable operating evidence and deterministic checks for compatibility and semantic SLA constraints.Source: §IV · p. 5
  • New transmission, semantic-processing, and evaluation capabilities can be registered without changing the service request.Source: §IV · p. 5
  • The prototype measures two transmission capabilities offline, stores results in their profiles, and compares those profiles at runtime using an explicit SLA-based rule.The offline measurements provide reusable evidence rather than defining a fixed workflow.Source: §IV · p. 5

A. Service Request and Agentic Planning Space

The service request specifies a perception outcome without choosing a codec or transmission implementation. The planner uses delivery, quality, and communication-cost requirements to select a feasible capability configuration, with deterministic checks before execution.

A. Service Request and Agentic Planning Space

  • The request requires successful image delivery, mean PSNR of at least 30 dB over evaluated scenes, and minimum CBR among configurations meeting the mandatory constraints.Successful delivery is required for every scene, and lower CBR is preferred once delivery and quality constraints are satisfied.Source: A. Service Request and Agentic… · pp. 5–6
  • The orchestration plane retrieves and validates candidate transmission profiles, then selects the lowest-CBR feasible configuration from the SLA requirements.The workflow evaluates the selected BPG-LDPC or SwinJSCC configuration using YOLO and feeds task-level evidence back into capability profiles.Source: Figure 2 · p. 6
  • Before execution, deterministic guards check model pairing, channel-use feasibility, delivery status, and SLA compliance.The minimum-CBR policy is one reproducible realization of the broader planning interface.Source: A. Service Request and Agentic… · p. 6

B. Profiled Transmission Capabilities

The study profiles two heterogeneous image-transmission capabilities and uses their measured operating conditions to support SLA-grounded planning. Under a fixed SLA, selected capabilities and communication budgets vary with channel conditions, while task-level detector consistency is recorded as feedback.

B. Profiled Transmission Capabilities

  • The two registered options are a vehicle-scenario-fine-tuned SwinJSCC model and a separated BPG, 5G NR LDPC, modulation, and AWGN transmission chain.For the digital option, a single modulation and coding scheme is selected per SNR-CBR point across source frames; delivery requires bit-by-bit payload equality.Source: B. Profiled Transmission Capabilities · p. 6
  • Each capability exposes a common evidence interface covering modalities, model pairing and version, operating conditions, physical feasibility, delivery, reconstruction quality, and task observations.Explicit pairing information prevents invalid compositions, such as connecting an encoder to an unrelated decoder.Source: B. Profiled Transmission Capabilities · p. 6
  • The evaluation uses 16 urban road images at 1024 × 1024 resolution, five SNR values, and five CBR values, yielding 400 image-SNR-CBR observations per capability.Both capabilities use CBR = Ns/(3HW), where Ns is the number of complex channel uses and H and W are image dimensions.Source: B. Profiled Transmission Capabilities · p. 6

C. SLA-Grounded Planning and Task Feedback

  • The planner filters profiles for valid pairings, physical feasibility, successful delivery, and required mean PSNR, then selects the lowest-CBR candidate and uses PSNR as a tie-breaker.The policy is reproducible but can be replaced by one considering other criteria; new branches can be added without changing the task interface.Source: C. SLA-Grounded Planning and Task… · p. 6

B. Profiled Transmission Capabilities

  • SwinJSCC remains competitive at low SNR, while the digital capability improves more rapidly and provides higher reconstruction quality at sufficiently high SNR; the crossover depends on SNR and CBR.The measured profiles are condition-dependent evidence for capability selection, not a universal codec ranking.Source: B. Profiled Transmission Capabilities · p. 6, Figure 3 · p. 7

C. SLA-Grounded Planning and Task Feedback

  • As SNR rises, the selected capability switches from SwinJSCC to the digital option, with CBR and mean PSNR varying under the unchanged SLA.Selections (capability, CBR, mean PSNR): 1 dB SwinJSCC, 1/16, 30.82 dB; 4 dB SwinJSCC, 1/24, 30.83 dB; 7 dB digital, 1/24, 32.53 dB; 10 dB digital, 1/48, 30.76 dB; 13 dB digital, 1/48, 31.70 dB.Source: C. SLA-Grounded Planning and Task… · p. 7
  • Selected-profile detector-consistency recall is 0.84, 0.86, 0.86, 0.77, and 0.82 at 1, 4, 7, 10, and 13 dB; it is feedback, not a selection constraint.At 10 dB, satisfying the visual-quality requirement does not necessarily guarantee consistent task-level observations.Source: C. SLA-Grounded Planning and Task… · p. 7

V. OPEN CHALLENGES AND FUTURE RESEARCH DIRECTIONS

Supporting large-scale, dynamic, and safety-critical services requires resolving challenges in task-grounded SLA evaluation, capability interoperability, scalable planning, distributed execution, and trustworthy operation. The V2X case study illustrates that PSNR alone does not fully indicate task-relevant object preservation.

Semantic SLA formalization and task-grounded evaluation

  • PSNR measures visual fidelity but does not fully indicate task-relevant object preservation; Fig. 4 returns detector-consistency recall as feedback, though the current selection policy does not use it.The planner selects the lowest-CBR profile meeting successful delivery and a 30 dB mean-PSNR threshold.Source: §V · p. 7, Figure 4 · p. 8

Capability contracts, interoperability, and lifecycle management

  • Reliable capability composition requires machine-verifiable contracts covering compatibility, operating regions, resources, model pairings, provenance, and versions, maintained through capability updates and deployment changes.Cross-provider interoperability and registry synchronization should avoid requiring disclosure of unnecessary proprietary implementation details.Source: §V · p. 7

Scalable agentic planning and distributed execution

  • Extensible capability ecosystems create large constrained planning spaces, while unrestricted planning may produce invalid combinations or excessive control latency.A practical architecture may combine high-level agent reasoning with deterministic validation, prevalidated templates, and fast local controllers; placement must jointly consider communication cost, computing, privacy, and deadlines.Source: §V · pp. 7–8

Trustworthy, verifiable, and auditable execution

  • Agent-controlled composition risks inaccurate profiles, compromised models, manipulated intermediate representations, and poisoned feedback or execution memory.Proposed safeguards include authentication, provenance tracking, access control, sandboxing, integrity-protected records, and deterministic safety and privacy enforcement.Source: §V · p. 8

VI. CONCLUSION

The article proposes SIoE, a composable architecture coordinating heterogeneous communication and AI capabilities according to task-level requirements. A lightweight V2X case study illustrates profile-grounded selection under explicit constraints, while the limited capability set and reproducible policy demonstrate composition without embedding each objective in a fixed end-to-end pipeline.

VI. CONCLUSION

  • SIoE separates semantic SLA specification, agentic planning and validation, and capability execution across distributed device, edge, and cloud resources.It coordinates heterogeneous communication and AI capabilities according to task-level requirements.Source: §VI · p. 8
  • A lightweight V2X case study illustrates retrieval, comparison, and selection of profiled transmission capabilities under explicit quality, delivery, and communication-cost constraints.Task-level measurements serve as feedback for subsequent planning.Source: §VI · p. 8
  • Despite a limited capability set and reproducible planning policy, the prototype demonstrates composition without embedding each service objective in a fixed end-to-end pipeline.Future work focuses on task-grounded SLA formalization, interoperable capability contracts, scalable agentic planning, and trustworthy execution.Source: §VI · p. 8
Loading 2609.02924v1…