xDevSM: An Open-Source Framework for Portable, AI-Ready xApps Across Heterogeneous O-RAN Deployments

arXiv:2602.03821 · cs.NI · Submitted 2026-02-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "xDevSM: An Open-Source Framework for Portable, AI-Ready xApps Across Heterogeneous O-RAN Deployments".

Jane: Here's the gist of the paper. Openness and programmability in the O-RAN architecture enable closed-loop control of the Radio Access Network (RAN).

Tom: This paper introduces xDevSM, an open-source framework that provides a comprehensive Software Development Kit (SDK) for building AI-driven closed-loop applications across heterogeneous RAN software stacks.

Jane: It validates this framework on real-world testbeds using commercial UEs and heterogeneous radio hardware to demonstrate its effectiveness in practical network conditions.

Tom: Framework Architecture xDevSM is built on top of the Python-based xApp framework provided by OSC

44: ,

45: , namely ricxappframe, and exposes service-model–specific APIs to xApp developers.

Paper summary: Tom: So, to wrap up on "xDevSM: An Open-Source Framework for Portable, AI-Ready xApps Across Heterogeneous O-RAN Deployments," it's basically an open tool that makes it much easier for people to build these smart applications that can control radio networks across all those different hardware setups >

Jane: The authors are focused on taking the complexity out of building these things by unifying the way you monitor data and how you issue commands, making sure it works whether your network uses one piece of software or ten different ones >

Lu: What this means for the field is that it gives developers a standardized way to build AI applications that can actually interact with heterogeneous RAN components without getting bogged down in all those low-level interface details every time >

Meng: For someone looking at deployment, it means less custom coding for each specific vendor integration and more focus on the actual control logic of the AI itself, which should speed up shipping models to production >

Lalam: It's positioned as foundational infrastructure because it offers that explicit design for AI and machine learning integration while keeping everything portable across different RAN deployments >

Conclusion: Tom: So we've seen how xDevSM is this open framework that lets you build smart applications for radio networks, even when you use different hardware all over the place across those O-RAN setups.

Jane: Yeah, and what’s really interesting is that it gives developers a single toolkit to handle both watching the network performance and actually telling it how to react in a closed loop.

Lu: The authors put together this thing on top of an existing Python framework, xApp, and they expose specific APIs for the service models you need, like radio bearer control or resource allocation control.

Meng: From my side at the startup, what I'm hearing is that it’s designed to be portable across different vendor stacks without requiring a complete rewrite every time we switch hardware platforms.

Lalam: I've read through the structure, and the main thing is how they wrap all those control data structures into Python objects, so you can just build your request based on what you see in real time.

Tom: It sounds like it’s aiming for that sweet spot between deep hardware access and a flexible software stack for these next-gen RAN controllers.

Jane: Exactly, and they did some solid testing on real testbeds, showing things like throughput recovery after limiting how much radio resource a slice can use during a down link transmission.

Lu: And the mobility experiment showed that this framework actually lets the network correctly execute handovers initiated by the AI application through those F1 interfaces.

Meng: For deployment, it means less custom glue code and more focus on training your model policies because they show you how to safely test new stuff on real hardware before you push it live.

Lalam: They also laid out a clear path for using machine learning workflows by standardizing the data pipelines, which should make it much easier to plug in things like long short-term memory networks.

Tom: So the big picture is that they're providing a solid piece of infrastructure that makes building intelligent, self-adjusting radio control applications much more practical for real-world use.

Jane: It really moves the conversation from just theoretical ideas to having a usable framework for actual deployment across those messy, heterogeneous RAN environments.

Lu: This feels like foundational work because they’re setting up the standard way we can combine monitoring and control functions in these AI applications moving forward.

Angelo Feraudo, Stefano Maxenti, Andrea Lacava, Leonardo Bonati, Paolo Bellavista, Michele Polese, Tommaso Melodia

University of Bologna Department of Computer Science and Engineering Institute for the Wireless Internet of Things Northeastern University

cs.NI

Submitted: 2026-02-03

Updated: 2026-02-03

Journal ref: IEEE Transactions on Network and Service Management, vol. 23, pp. 7636-7649 (2026)

DOI: 10.1109/TNSM.2026.3733344

Code: https://github.com/wineslab/xDevSM

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Openness and programmability in the O-RAN architecture enable closed-loop control of the Radio Access Network (RAN).

Key concepts

xDevSM
An open-source framework built on top of the xApp framework that provides APIs for building AI applications. It helps integrate monitoring and control services into a single application for closed-loop automation.
Service Model Composition
The ability of the framework to support multiple service models, such as Radio Bearer Control (RBC) or Radio Resource Allocation Control (RRAC). This lets one application handle various network tasks like monitoring and control simultaneously.
Closed-Loop Control
A system where an application monitors network performance indicators and then automatically issues control requests based on those observations. xDevSM enables this by allowing applications to subscribe to data and trigger actions directly.
Heterogeneous RAN Deployments
The ability of the framework to work across different types of radio hardware and software stacks within an O-RAN network. This ensures the framework is portable and can be used in diverse, real-world network environments.

Terminology

Summary

Openness and programmability in the O-RAN architecture enable closed-loop control of the Radio Access Network (RAN). p. 0 This paper introduces xDevSM, an open-source framework that provides a comprehensive Software Development Kit (SDK) for building AI-driven closed-loop applications across heterogeneous RAN software stacks. p. 1 It validates this framework on real-world testbeds using commercial UEs and heterogeneous radio hardware to demonstrate its effectiveness in practical network conditions. p. 1

Framework Architecture

xDevSM is built on top of the Python-based xApp framework provided by OSC [44], [45], namely ricxappframe, and exposes service-model–specific APIs to xApp developers. The generic xApp model is provided by the xDevSMRMRXapp class which encapsulates all initialization procedures required to deploy the xApp as a microservice within the OSC Near-RT RIC cluster. p. 5

Service Model Composition

The framework supports multiple service models to enable combined functionalities within a single application. For monitoring, the XappKpmFrame class exposes high-level APIs for interacting with E2 nodes to manage subscription procedures. p. 5 For control, xDevSM supports three control service styles of the RC service model: Radio Bearer Control (RBC), Radio Resource Allocation Control (RRAC), and Connected Mode Mobility Control (CMMC). p. 3

Closed-Loop Control Implementation

The main strength of xDevSM lies in enabling the seamless integration of multiple models within a single application for closed-loop control. This tight integration allows an xApp to subscribe to Near-RT performance indicators and then issue CONTROL requests based on observed telemetry. p. 6 The framework provides RC-related Python objects that wrap C data structures which contain methods that build the Control Request according to the service style selected for the specific use case. p. 6

Real-World Validation

The framework is validated across two complementary experimental deployments to jointly assess monitoring and control capabilities. In the slice-aware Radio Resource Allocation Control experiment, the xApp limits the maximum PRB allocation to 40% during a downlink UDP transmission and demonstrates throughput reduction followed by recovery. p. 8 The mobility-aware handover control experiment shows that the RAN correctly executes the handover procedure initiated by the xApp through F1-based inter-DU handovers. p. 9

AI Integration and Future Directions

xDevSM is designed to support AI and machine learning workflows by providing standardized data pipelines and control abstraction. The CI/CD pipeline supports safe AI deployment through automated testing of new policies on real hardware before production deployment. p. 5 Future research directions include leveraging LSTM or Transformer networks to predict future network conditions and trigger preemptive control actions. p. 0

Conclusion

xDevSM enables intelligent closed-loop applications combining monitoring and control across heterogeneous RAN deployments. The framework positions itself as foundational infrastructure for the next generation of intelligent RAN control. p. 0 It complements existing initiatives by offering a lightweight tool specifically focused on xApps and explicit design for AI/ML integration.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve AI systems by creating an AI-Ready Control Framework that supports closed-loop control by providing standardized data pipelines (KPM) and control interfaces (RC) that naturally support reinforcement learning formulations. This allows AI algorithms to issue control commands through simple function calls rather than managing low-level protocol messages.

  2. Improve AI systems by enabling seamless interaction between xApps, the Near-RT RIC, and the RAN-side E2 termination across different RAN software stacks. This allows the same xApp logic to operate concurrently across heterogeneous RAN implementations, enabling interoperability and consistent behavior even when different E2 nodes are connected to the same Near-RT RIC.

  3. Improve AI systems by supporting full closed-loop control applications that combine monitoring and control, as demonstrated by the example: "an xApp can subscribe to Near-RT performance indicators, such as PRB utilization or QoS-related metrics, using the XappKpmFrame. Based on the observed telemetry, the same xApp can then dynamically adjust slice-level resource allocations through the RadioResourceAllocationControl class."

  4. Improve AI systems by integrating a complete Continuous Integration (CI)/Continuous Delivery (CD) workflow for automated deployment and validation of xApps across multiple RAN stacks. This pipeline "streamlines the building, deployment, and testing of SMs and xApps across heterogeneous environments, ensuring consistency, reducing integration time, and providing comprehensive logs of both data-plane performance and control-plane interactions."

  5. Improve AI systems by enabling predictive control through leveraging models like LSTM or Transformer to predict future network conditions (e.g., throughput degradation, congestion onset) and trigger preemptive control actions before signal quality degrades.

Sources

Related papers