Enabling Programmable Inference and ISAC at the 6GR Edge with dApps

Abstract: The convergence of communication, sensing, and AI in the Radio Access Network (RAN) offers compelling economic advantages through shared spectrum and infrastructure. How can inference and sensing be integrated in the RAN infrastructure at a system level? Current abstractions in O-RAN and 3GPP lack the interfaces and capabilities to support (i) a dynamic life cycle for inference and Integrated Sensing and Communication (ISAC) algorithms, whose requirements and sensing targets may change over time and across sites; (ii) pipelines for AI-driven ISAC, which need complex data flows, training, and testing; (iii) dynamic device and stack configuration to balance trade-offs between connectivity, sensing, and inference services. This paper analyzes the role of a programmable, software-driven, open RAN in enabling the intelligent edge for 5G and 6G systems. We identify real-time user-plane data exposure, open interfaces for plug-and-play inference and ISAC models, closed-loop control, and AI pipelines as elements that evolutions of the O-RAN architecture can uniquely provide. Specifically, we describe how dApps - a real-time, user-plane extension of O-RAN - and a hierarchy of controllers enable real-time AI inference and ISAC. Experimental results on an open-source RAN testbed demonstrate the value of exposing I/Q samples and real-time RAN telemetry to dApps for sensing applications.

• The paper introduces a hierarchical dApp-based architecture for real-time edge inference and ISAC, addressing limitations of legacy RAN designs.
• It details an innovative E3 interface and dynamic resource partitioning to enable sub-millisecond data access and high-accuracy sensing.
• Empirical results validate the approach with CRLB analysis and ranging experiments, achieving sub-meter accuracy in challenging conditions.

Programmable Inference and ISAC at the 6GR Edge with dApps: An Expert Analysis

Introduction and Motivation

The transition to 6G architectures is fundamentally reshaping Radio Access Network (RAN) design, coupling integrated sensing and communication (ISAC) and AI-driven inference natively into the edge infrastructure. The paper "Enabling Programmable Inference and ISAC at the 6GR Edge with dApps" (2603.29146) systematically addresses the system-level and architectural limitations that prevent current O-RAN and 3GPP-based RANs from supporting programmable, AI-driven ISAC services. It introduces discrete, software-managed distributed applications (“dApps”) and a hierarchical control framework, enabling real-time inference at the user-plane with explicit data-plane exposure.

The articulation of five key architectural challenges (real-time PHY/MAC data access, end-to-end AI pipelines for ISAC, support for heterogeneous ISAC-enhanced radios, sensing/inference algorithm life-cycle management, and flexible stack configuration) underscores the systemic rigidity of current RANs and prescribes programmable RAN evolution predicated on dApps and open interfaces.

Figure 1: System-level challenges toward enabling programmable ISAC and inference within the 6GR architecture.

O-RAN Limitations and the Need for Programmable ISAC

O-RAN’s control-user plane separation and its hierarchical Near-RT RIC/xApp and Non-RT RIC/rApp architecture are designed for communication-centric optimization. However, as the paper observes, O-RAN lacks native support for real-time user-plane data streaming (e.g., I/Q samples, CSI, SRS) essential for physical-layer sensing and inference. The E2 interface is ill-equipped for the high-rate, low-latency data movement ISAC requires, and present abstractions do not support plug-and-play deployment or dynamic life-cycle management of site-specific sensing/inference functions.

While 3GPP’s ongoing Release 20 study items address waveform and channel model enhancements for ISAC, architectural support for AI workflows, dynamic stack reconfiguration, or algorithm deployment/monitoring pipelines are absent.

Programmable Sensing Architecture with dApps

The central proposal is a two-tier, latency-aware architecture leveraging dApps (software-defined, co-resident with RAN DUs) and a multi-level RIC/SMO orchestrator stack. dApps interface with the RAN user-plane through a new E3 interface at the DU, receiving I/Q samples, CSI, and SRS at sub-millisecond timescales. This realizes real-time, locally-optimized inference and enables multiple independent ISAC-applications (different dApps per topology/algorithm/target) to run in parallel, tightly coupled to the communication stack.

Figure 2: Hierarchical sensing architecture. dApps provide real-time I/Q, CSI, and SRS access at the DU for ISAC tasks; xApps/rApps handle multi-node fusion, policy, and cross-layer coordination.

Central to this design:

• Edge-level sensing and inference: dApps deploy arbitrary AI/ML models (e.g., CNNs for target classification, subspace algorithms for ranging) leveraging HW accelerators (GPU, FPGA) provisioned by O-Cloud.
• Hierarchical processing: dApps perform per-site, low-latency processing, with xApps/rApps at the RIC aggregating information across sites for network-level fusion, collaborative scheduling, and multi-node ISAC tasks.
• Dynamic resource partitioning: O-Cloud enforces resource allocation among communication, sensing, and inference workloads; the ISAC Orchestrator dynamically manages dApp/xApp placement, instantiation, and preemption, responding to demand, policy directives, and resource constraints.
• Security and privacy: Introducing the E3 interface mandates robust access control, isolation, and compliance mechanisms, especially as I/Q-level data exposure can increase the attack surface and privacy risk.

ISAC Algorithm Life-Cycle Management

A key insight is that ISAC functions are site-specific and dynamic, requiring MLOps-like support for data collection, centralized and local training, validation, cataloging, deployment, and continuous performance monitoring/adaptation. The authors describe an ISAC Orchestrator (in the SMO) that automates mapping algorithm requirements to site resources and radio capabilities, with policy-driven, intent-based deployment and continual telemetry-based adaptation.

Figure 3: ISAC life-cycle management: centralized training, model cataloging, automated deployment, and performance-driven adaptation orchestrated by the SMO.

This enables not only rapid rollout and update of sensing services but also ensures that HW constraints (e.g., accelerator availability, RU capabilities) and multi-tenancy scenarios are respected, supporting government/commercial separation, privacy requirements, and resource arbitration.

Empirical and Numerical Validation

The paper demonstrates the necessity and effectiveness of E3-based dApp sensing with two quantitative results:

• CRLB-based analysis for monostatic sensing: The range and velocity RMSE for a drone target sharply decrease with increased E3 bandwidth for I/Q transfer, confirming that only site-local, high-throughput user-plane access enables high-accuracy sensing. The required data rate renders offloading to core/cloud infeasible, mandating edge-level dApp inference and motivating new interface design.
• Experimental validation for ranging: Subspace-based ranging algorithms, leveraging dApp access to full CIR data (vs. scalar reporting), are shown to yield sub-meter accuracy that the standard LMF pipeline cannot achieve, especially in heavy multipath or low-SNR regimes.

These results substantiate the core architecture claim: only real-time, user-plane-level observability (with dApps at the DU and E3 interface) can support next-generation ISAC accuracy and latency requirements, and programmatically expands the expressiveness of RAN services.

Implications, Challenges, and Future Directions

The architectural decoupling of sensing/inference from both the communication pipeline and deployment site enables clear separation of concerns, flexibility for third-party value-added services, and new RAN revenue models (e.g., ISAC-as-a-Service, operator-customized environmental sensing). Programmable ISAC also enables closed-loop, AI-driven RAN-optimization, including communication-aware sensing (joint scheduling, beamforming, network selection) and sensing-aware communication (RF environment adaptation, interference prediction, situational awareness).

However, several core challenges remain:

• Multi-node coordination and synchronization: Accurate multi-node ISAC requires ultra-tight time and frequency coordination, robust handoff of measurements, and cross-site data fusion, not yet standardized in O-RAN or 3GPP.
• Resource arbitration and coexistence: On shared O-Cloud (and especially with heterogeneous RU/DU configurations), ISAC workloads compete directly with user-plane traffic for compute/memory/I/O, requiring enforcement frameworks for SLOs and SLAs.
• Security and privacy: The introduction of programmable user-plane access (E3), third-party dApps, and exposure of physical-layer waveforms mandates robust access control, auditing, and compliance strategies—especially given privacy-sensitive inferences possible from raw I/Q data (2603.29146).
• Standardization: There is a clear need for cross-SDO coordination to align 3GPP ISAC specifications (waveform, channel, service APIs) with programmable RAN interfaces (dApps/E3, O-RAN orchestration, multi-vendor interoperability).

Conclusion

The paper constructs a formal, multi-tier architecture for programmable ISAC and inference at the 6GR edge, with dApps as the architectural primitive for real-time, user-plane data access and hierarchical control. It establishes that only such an architecture can achieve the accuracy, flexibility, and manageability demanded by dynamic, AI-driven ISAC use cases envisioned for 6G. Empirical results substantiate that scalar abstraction and control-plane-centric designs are fundamentally inadequate for modern sensing and inference tasks. The articulation of open research challenges—especially in interface abstraction, cross-domain orchestration, and privacy/security—marks a critical roadmap for research and standardization in the evolution toward AI-native, programmable 6G RANs.