Beyond Connectivity: Toward AI-Native, Open, and Sensing-Integrated 6G Networks

By Tommaso Melodia

The telecommunications industry stands at an inflection point. For decades, mobile networks have been engineered as data pipelines designed to deliver bits reliably and efficiently. This model, focused on connectivity as a utility, is reaching its economic and technological limits. To create differentiated value and sustainable growth on the pathway toward 6G, operators must evolve from connectivity providers into distributed intelligence platforms – global distributed inference engines where connectivity, sensing, inference, and learning coexist within an open, unified, programmable infrastructure, enabling networks that perceive, reason, and adapt.

From Connectivity to Intelligence

The 6G era will mark a fundamental shift from data transport to intelligent systems. Networks will no longer simply deliver data—they will sense, process, interpret, and act on it in real time. This evolution is driven by several converging technology vectors:

1. AI-Native RAN: Embedding AI-based signal processing and inference capabilities directly within the RAN rather than relying on centralized cloud-based analytics systems. This allows closed-loop optimization and real-time joint adaptation of functionalities at multiple layers of the protocol stack and over multiple time scales.  Moreover, virtualized RAN workloads can be executed on the same shared accelerated compute infrastructure that supports AI workloads.
2. Software-defined, Open and Disaggregated Architectures: Building on Open RAN principles, network functions are virtualized, modular, and interoperable, allowing new actors, including universities, SMEs, and startups to innovate rapidly. A rapid pace of innovation is unlocked through software, rather than tied to slow hardware capability introductions.
3. Integrated Inference, Sensing and Communications (I-ISAC): Transforming the radio network into a global distributed inference engine, as well as into a sensor fabric capable of environmental awareness, mobility tracking, and spatial reasoning.

These factors converge toward a self-adaptive, AI-driven RAN, integrated with the O-RAN architecture, where control loops operate across multiple timescales—from the milliseconds of PHY/MAC optimization (dApps, xApps) to the seconds or minutes of network-wide orchestration (rApps).

The AI-RAN Architecture—Coexistence of AI and RAN

The AI-RAN Alliance represents a key milestone in the evolution of intelligent wireless networks. Formed by leading operators, vendors, and research institutions, it aims to define interoperable blueprints that embed AI-native capabilities across every layer of the RAN (AI-for-RAN), enable coexistence between AI and RAN workloads (AI-for-RAN), and conceive new RAN-centric AI-based applications (AI-on-RAN), aligning research, development, and deployment under common frameworks. This collaboration reduces fragmentation, accelerates adoption, and ensures that innovations from universities and startups can translate directly into deployable solutions.

The proposed AI-RAN architecture embodies the coexistence of AI and RAN in a unified design. In this model:

• AI-for-RAN optimizes radio operations, leveraging learning models for scheduling, power control, and beam management.
• AI-on-RAN executes general-purpose AI workloads (e.g., vision, LLM inference) at the network edge using the same compute substrate.
• AI-and-RAN represents the true convergence—a shared infrastructure where AI and communications processes cohabitate, leveraging GPU-accelerated platforms.

The Northeastern AI-RAN architecture, detailed in our recent work, introduces distributed intelligence embedded at multiple levels:

• Non-RT RIC (rApps): Policy-driven network optimization with long-term learning and multi-cell orchestration.
• Near-RT RIC (xApps): Real-time radio control with AI-based decision loops.
• Distributed Applications (dApps): AI processes deployed directly on DUs or RUs for PHY/MAC layer adaptation.

 

Through NVIDIA’s wireless ecosystem—including Aerial, Omniverse, Sionna, and through integration with Colosseum – the Open RAN Digital Twin – this vision becomes actionable. Training, inference, and network optimization workflows run on shared GPU-based infrastructure, accelerating innovation cycles and enabling rapid, reproducible experimentation.

X5G as the Blueprint: Hardware, Architecture, and Integration with Colosseum

The X5G testbed is an open, programmable, multi-vendor 5G O-RAN platform that integrates NVIDIA Aerial RAN Computer (ARC)  with OpenAirInterface software to accelerate Layer 1 and Layer 2 processing. Each X5G site features Gigabyte E251 servers equipped with Intel Xeon 6240R CPUs and NVIDIA accelerated computing, interconnected with NVIDIA ConnectX-6 NICs for high-throughput, low-latency fronthaul communication. The radio layer includes Foxconn RPQN RUs configured for 2×2 MIMO, operating at 20 MHz (band N77) and fully compliant with O-RAN interfaces.

The X5G compute infrastructure supports both containerized and bare-metal deployments of OAI protocol stacks, orchestrated through Red Hat OpenShift clusters. Automation is enabled by Tekton pipelines and Argo CD, allowing zero-touch deployment and continuous integration of RAN and AI workloads.

Crucially, X5G is tightly integrated with Colosseum, the world’s largest wireless network emulator and a fully programmable O-RAN digital twin. Colosseum provides more than 65,000 channel emulations over 80 MHz of bandwidth, backed by a GPU-accelerated compute fabric for digital-twin-driven experimentation. This integration allows for end-to-end validation of AI models and xApps under controlled, reproducible RF environments before live OTA deployment. Through this architecture, new AI-driven control algorithms can move seamlessly from simulation to real-world testing.

While AI-RAN architectures enable intelligence, the Northeastern AutoRAN framework enables network evolution at the speed of software. It brings DevOps-style automation to wireless research, bridging simulation, emulation, and live deployment:

• Containerized builds for modular RAN components;
• Automated tests on emulated RAN environments;
• Continuous OTA testing with hardware-in-the-loop;
• Automated performance reporting and retraining triggers.

AutoRAN ensures that innovations in RAN and AI-based control can move from research to deployment without compromising reliability—merging developer agility with operator-grade assurance.

 

Agentic AI-RAN: Hierarchical Multi-Agent Control

To operationalize the AI-RAN vision, we introduced AgentRAN—a framework for agentic AI control in the RAN. AgentRAN employs hierarchical teams of LLM-based agents that:

• Translate natural-language intents from operators into actionable network policies.
• Decompose high-level goals into specific tasks across rApps, xApps, and dApps.
• Coordinate via agent-to-agent (A2A) protocols for multi-layer reasoning.

Each agent specializes by layer and function: policy agents reason at the non-RT level, control agents at the near-RT RIC, and execution agents at the DU layer. Collectively, they form an intent-executing RAN capable of aligning network behavior with operator goals dynamically.

This design bridges human-understandable network intents with executable AI workflows, leveraging NVIDIA accelerated computing for compute acceleration and open interfaces for interoperability.

ALLSTaR: Automating Learning, Testing, and Scheduling

Last, scheduling remains one of the most impactful yet opaque functions in the RAN. Proprietary implementations limit innovation and reproducibility. ALLSTaR (Automated LLm-driven Scheduler generation and Testing for intent-based RAN) addresses this by creating a modular, AI-driven pipeline for automated scheduler generation and evaluation.

The ALLSTaR pipeline consists of:

1. Literature ingestion: Using LLMs to parse academic scheduler algorithms from papers.
2. Code synthesis: Translating mathematical descriptions into executable Python/C++ code.
3. Validation: Running functional, unit, and OTA tests on X5G’s multi-vendor network.
4. Evaluation: Benchmarking performance across delay, throughput, and fairness metrics.
5. Intent-Based Scheduler (IBS): Composing new scheduling policies that meet operator intents without extensive retraining.

Our experiments on ALLSTaR represent one of the largest OTA scheduler comparison campaigns to date, involving 18 algorithms tested under real-world traffic conditions. The framework reveals when classic algorithms like PF, RR, or ML-based schedulers truly deliver—and when they fail.

 

Toward an Open, Programmable, AI-Powered RAN Ecosystem

This convergence of AI-native architectures, open RAN, and agentic intelligence redefines the telecom landscape:

• For operators: It enables monetization beyond connectivity—offering distributed inference and sensing as-a-service.
• For industry verticals: It delivers optimized, explainable, and intent-aligned QoS for mission-critical applications.
• For researchers: It establishes an open hardware-software-data continuum as a foundation where ideas can evolve from paper to prototype to deployment, to at-scale trials.

As we move from 5G to 6G, the network will evolve into a learning system with radios as sensors and CPUs/GPUs as the brain. Through collaborative efforts such as the AI-RAN Alliance, Open6G, and the NVIDIA AI Aerial ecosystem, we are building the blueprint for this transformation.

This transformation will not happen in isolation. It requires open collaboration between academia, industry, and government, bridging the innovation capacity of universities and SMEs with the deployment scale of operators. Together, we can evolve beyond connectivity, toward networks that provide value by observing, learning, reasoning, and acting.

If you are building in this space -or exploring how to implement AI-native architectures in your network – connect with us. The X5G, AgentRAN, and ALLSTaR frameworks are open for collaboration.

References

[1] Michele Polese, Leonardo Bonati, Salvatore D’Oro, Stefano Basagni, Tommaso Melodia, “Understanding O-RAN: Architecture, Interfaces, Algorithms, Security, and Research Challenges,” IEEE Communications Surveys & Tutorials, 1376-1411, Secondquarter 2023. [link]

[2] Michele Polese, Niloofar Mohamadi, Salvatore D’Oro, Tommaso Melodia, “Beyond Connectivity: An Open Architecture for AI-RAN Convergence in 6G,” arXiv:2507.06911 [cs.NI], pp. 1-8, July 2025. [link]

[3] S. Maxenti, R. Shirkhani, M. Elkael, L. Bonati, S. D’Oro, T. Melodia, M. Polese “AutoRAN: Automated and Zero-Touch Open RAN Systems”, arXiv:2504.11233 [cs.NI], pp. 1-17, April 2025. [link]

[4] Maxime Elkael, Salvatore D’Oro, Leonardo Bonati, Michele Polese, Yunseong Lee, Koichiro Furueda, and Tommaso Melodia, “AgentRAN: An Agentic AI Architecture for Autonomous Control of Open 6G Networks,” arXiv:2508.17778 [cs.AI], pp. 1-7, August 2025. [link]

[5] Maxime Elkael, Michele Polese, Reshma Prasad, S. Maxenti, T. Melodia “ALLSTaR: Automated LLM-Driven Scheduler Generation and Testing for Intent-Based RAN,” arXiv:2505.18389 [cs.NI], pp. 1-13, May 2025. [link]