5 Scope

This Work Item produces a series of Technical Reports (TRs) that provide developer-oriented guidelines for AI enablement in oneM2M. Each TR addresses a specific AI-related topic and provides use cases, functional descriptions, message flow diagrams, and resource description samples using oneM2M features.

The following three scenarios are addressed.

Scenario #1 Use of SDT and SAREF in oneM2M for Large Action Models (LAM)

This guide describes how SDT and SAREF can be leveraged within the oneM2M platform to enable reliable natural language-based IoT device control using Large Action Models (LAMs). A proposed framework comprises four components: (1) Automated Semantic Restoration, which maps legacy SDT attributes to SAREF concepts via hierarchical context embedding and injects inferred SAREF URIs into the ontologyRef attribute of oneM2M resources for zero-shot device provisioning; (2) Knowledge Flattening, which converts verbose SDT XML/XSD schemas into LLM-friendly system prompt templates; (3) Reverse Synthetic Data Generation, which algorithmically enumerates valid oneM2M control primitives as ground-truth outputs and uses a Teacher LLM with diversity prompting to back-generate natural language commands for LoRA adapter fine-tuning; and (4) Semantic Guardrails, a runtime validation layer that enforces SDT-defined type, range, and enumeration constraints on LLM-generated oneM2M JSON primitives with a self-correction feedback loop before execution.

Scenario #2: Use of Genuine oneM2M resources to support Federated Learning (FL)

This guide describes how native oneM2M resources — AEs, , and — within IN-CSE and MN-CSE nodes can implement a complete Federated Learning (FL) lifecycle without external ML middleware. The IN-AE at the IN-CSE acts as the global model aggregator, while MN-AEs at each MN-CSE perform local training on locally stored sensor data, ensuring raw data never leaves the edge node. The entire FL process — round initiation, local training, model upload, FedAvg aggregation with Z-score outlier detection, and global model redistribution — is managed solely through standard oneM2M CREATE, RETRIEVE, and NOTIFY operations.

Scenario #3: Coding Agents for Automated IPE Development in Multi-Protocol IoT Networks

This guide describes how a coding agent can be designed to automatically generate, test, and deploy Interworking Proxy Entities (IPEs) that connect heterogeneous IoT devices running protocols such as MQTT, STA, or Zigbee to a oneM2M CSE. The proposed framework addresses two key aspects: (1) Lowering the Integration Barrier , whereby the coding agent abstracts the complexity of oneM2M resource modelling and protocol bridging, enabling developers with little or no prior oneM2M knowledge to produce functional IPEs through natural language instructions; and (2) Self-evolving IoT Networks , whereby the coding agent enables CSE-connected networks to autonomously extend their device coverage by generating and deploying new IPEs in response to the discovery of previously unknown device types or protocols, without requiring manual developer intervention.