RAN Intelligent Controllers (RICs) Overview

• RICs are centralized microservices in Open RAN that perform cross-layer control using open protocols like E2, A1, and O1.
• They support multi-timescale operations from real-time to non-real-time, leveraging AI/ML for adaptive scheduling and resource optimization.
• The architecture utilizes standardized interfaces and disaggregation to manage PHY/MAC control, network slicing, and QoS objectives efficiently.

A RAN Intelligent Controller (RIC) is a logically centralized microservice in the Open RAN (O-RAN) architecture that interfaces with the disaggregated RAN stack via open protocols (O-RAN E2, O1, A1) and performs closed-loop, cross-layer decision and control by consuming RAN and application-layer telemetry, computing control decisions or AI/ML-optimized policies, and enforcing actions on the PHY/MAC or higher layers. RICs are deployed at multiple timescales—real-time (RT-RIC, TTI-scale ≤1 ms), near-real-time (near-RT RIC, 10 ms–1 s), and non-real-time (non-RT RIC, ≥1 s)—and underpin NextG networks by enabling programmable, intelligent, and energy-efficient optimization and orchestration of radio resources, user sessions, network slices, and advanced QoE objectives (Ko et al., 2023).

1. Taxonomy and Role of RICs in O-RAN

RICs are foundational in the O-RAN architectural stack, offering programmable control and orchestration of softwarized, disaggregated RAN elements:

• Non-Real-Time RIC (non-RT RIC): Located in the Service Management & Orchestration (SMO) layer, responsible for policy management, AI/ML model training, analytics, and long-term optimization, interfacing via A1/O1 (Ko et al., 2023, Polese et al., 2022).
• Near-Real-Time RIC (near-RT RIC): Deployed close to the RAN on edge hardware, hosts xApps for medium-grained, closed-loop control (10 ms–1 s) and interfaces with the RAN via E2 for KPI telemetry and control, and with non-RT RIC via A1 for receiving policies and enrichment information (Ko et al., 2023, Lima et al., 16 Jan 2025).
• Real-Time RIC (RT-RIC, also “EdgeRIC”): Co-located at the RAN node (typically the O-DU), operates at TTI granularity (≤1 ms), with custom IPC channels (RT-E2 over ZeroMQ) to execute sub-millisecond AI-based scheduling and control via μApps (Ko et al., 2023).
• dApps (Distributed Apps): Extend RIC capabilities to the user plane and real-time domain (<10 ms) for tasks such as spectrum sharing and sub-ms positioning, interfacing locally (E3) with the RAN and reporting upward via E2/E2SM-DAPP (Lacava et al., 27 Jan 2025).

Timescale-driven classification is summarized as:

Controller	Latency	Control Scope	Applications
RT-RIC	≤1 ms	PHY/MAC per-TTI	μApps, dApps
near-RT RIC	10 ms–1 s	Slice/mobility/scheduling	xApps
non-RT RIC	≥1 s	Policy, analytics, training	rApps
dApps	<10 ms	User/PHY real-time	Spectrum, positioning

RICs enable comprehensive multi-timescale orchestration, from strategic policy to real-time link adaptation.

2. Architectural Principles and Interfaces

RICs leverage standardized open interfaces and modular microservice design:

• E2 Interface: SCTP-based protocol for near-RT RIC↔CU/DU exchanges; supports E2AP and E2SMs (KPM, RC, DAPP, etc.) for KPI telemetry, control, and event reporting (Ko et al., 2023, Moro et al., 2023).
• A1 Interface: Northbound policy/model distribution channel from non-RT RIC to near-RT RIC, carrying policies, slice templates, model artifacts, and enrichment info (Polese et al., 2022, Li et al., 2021).
• O1 Interface: SMO↔RAN management for configuration, performance, and life-cycle, built on NETCONF/YANG over TLS, file transfer, and bulk KPI streams (Soltani et al., 14 May 2024).
• RT-E2 and E3 (for RT/dApps): IPC-based protocol for sub-ms controller↔RAN communication; E3 enables structured, extensible message exchange for applications embedded inside CU/DU for real-time user-plane intelligence (Lacava et al., 27 Jan 2025).
• RMR and SDL: Internal messaging (RIC Message Router) and shared data layers for state management, multi-xApp coordination, and high-speed telemetry ingestion (Almeida et al., 2023).

Functional disaggregation—placing latency-critical controller logic at the edge or even inside RAN nodes, while retaining cloud-based orchestration for non-RT functions—is essential for meeting strict latency budgets and scaling up to large deployments (Almeida et al., 2023, Baena et al., 21 Feb 2025, Baranda et al., 3 Jul 2025).

3. Optimization, AI/ML, and Applications

RICs support a wide spectrum of control strategies, from heuristic to deep RL to federated and evolutionary meta-learning:

• Weight-based resource allocation: EdgeRIC computes per-UE weights $w_i[t]$ at every TTI, and uses proportional scheduling for RB allocation with constraints on latency (≤1 ms), resource sums, and CQI mapping (Ko et al., 2023).
• AI/ML Policy Training:
  • Off-policy RL (PPO, DQN) on DigitalTwin emulators to generate robust control policies, with convergence in 20–40 iterations (100k–200k TTIs) for throughput-optimal scheduling (Ko et al., 2023).
  • Cross-layer RL: State includes application-level features (e.g., video buffer occupancy) for joint PHY/MAC/user-QoE policies (Ko et al., 2023).
  • Federated meta-learning (FML): Distributed xApps run local RL agents refined via Reptile-style meta-updates and global aggregation (FedAvg), supporting rapid, zero-shot adaptation to dynamic RAT/traffic environments (Erdol et al., 2022).
  • Neuroevolution-based DRL: F-ONRL architecture combines real-time DRL xApps with parallel NE xApps (genetic algorithms) for robust convergence and exploration in near-RT RICs (Kouchaki et al., 15 Jun 2025).
  • LLM-hRIC: Hierarchical frameworks where non-RT RIC uses LLMs for strategic guidance and near-RT RIC implements RL-based actionable control, supporting domain-specific finetuning and multi-modal policy coordination (Bao et al., 25 Apr 2025).

Use cases include adaptive scheduling, slice-level QoS/SLA enforcement, energy-efficient BS activation (Pareto-optimized by RL agents), multi-RAT traffic steering, spectrum sharing, anomaly detection, positioning, and ultra-reliable low-latency control for URLLC (Ko et al., 2023, Bordin et al., 17 Oct 2024, Moro et al., 2023, Lacava et al., 27 Jan 2025).

4. Power, Scalability, and Component Placement

RIC power consumption scales linearly with the number of E2 nodes and KPIs per node ($$P_\mathrm{RIC}(N,K)=P_\mathrm{static}+Np_\mathrm{node}+NKp_\mathrm{kpi}$$). Large-scale deployments can encounter bottlenecks as RIC power cost may exceed the cost of pico-cell operation itself (max savings: 87% reduction for redundant KPI removal in the large scenario) (Lima et al., 16 Jan 2025). Optimizing KPI subscriptions—removing identical and overlapping requests at the Subscription Manager via periodicity refinement and temporal-sensitivity matching—reduces power overhead and network traffic dramatically.

Component placement is governed by latency sensitivity, resource capacity, and cost trade-offs:

• Latency-critical elements (E2 Termination, xApps, SDL/NIB for real-time loops) are disaggregated and pushed closer to the edge (O-DU), while management and non-time-critical components remain in the cloud (Almeida et al., 2023).
• RIC Orchestrator (RIC-O) supports dynamic clustering, fast heuristic or background MILP-based reconfiguration, and resilience to edge failures (Almeida et al., 2023).
• In non-terrestrial networks (NTN), near-RT RICs controlling in-space O-DU/O-CU must reside on-satellite to meet strict 10 ms–1 s loop budgets; ground-only placement fails for LEO/MEO/GEO latency (Baena et al., 21 Feb 2025, Baranda et al., 3 Jul 2025).

5. Conflict Resolution, Security, and Reliability

RIC-enabled open architectures increase control expressiveness but introduce challenges from conflicting applications and adversarial threats:

• Conflict management: PACIFISTA profiles each xApp in sandbox environments, computes ECDF-based distances for parameter and KPM influence, and detects direct, indirect, and implicit conflicts via dependency graphs (Prever et al., 7 May 2024). Operators can set KPM conflict tolerance thresholds and priority indices to select maximal function and minimal degradation (e.g., 16–30% throughput loss for high-conflict apps).
• Hierarchical control resolution: Utility-maximization under resource constraints coordinates rApps, xApps, and dApps, with real-time arbitration/fallback enforced in near-RT RIC or dApp runtime (Lacava et al., 27 Jan 2025).
• Security measures: Combine containerization and digital signatures for xApp/rApp onboarding, mTLS/PKI authentication, HMAC-signed LLDP frames, anomaly detection xApps in near-RT RIC, and programmable zero-trust RBAC policies via A1 (Soltani et al., 14 May 2024). Weaknesses include supply-chain risk, cross-domain federation, AI model poisoning, and timing/synchronicity over dynamic ISL/feeder links.
• Resilience: Cluster-based RIC architectures in space (Leader-Follower, Fede2) mitigate single-point failures, support autonomous closed-loop control, and enable dynamic re-election of cluster controllers (Baena et al., 21 Feb 2025).

6. Empirical and Simulated Performance

Extensive empirical and simulation-driven evaluation supports the scalability and efficacy of RIC architectures:

• EdgeRIC achieves median round-trip latency of 100 μs (<300 μs at 99th percentile) versus >15 ms for cloud-based near-RT RIC; throughput gains 5–50% over model-based scheduling, 60% fewer media stalls under RL control (Ko et al., 2023).
• dApps on OAI gNB consistently yield control-loop latencies ≤450 μs and real-time spectrum sharing effectively detects incumbents and adapts PRB assignments (<500 μs for positioning) (Lacava et al., 27 Jan 2025).
• DRL xApps deployed in near-RT RIC/energy-saving settings reduce power by 24% over always-on, maintaining throughput vs. heuristic baselines, and are feasible on standard Xeon-class edge hardware (<10 ms inference) (Bordin et al., 17 Oct 2024).
• Federated meta-learning achieves 89–95% caching-rate and adapts in ~3 episodes to new tasks, verifying rapid convergence in non-stationary environments (Erdol et al., 2022).
• RIC-O placement heuristics scale to 512 E2 nodes with near-optimal component replication; rapid edge redeployment restores loop latencies <10 ms after failure (Almeida et al., 2023).
• Space-O-RAN simulations show that ISL-based inter-satellite loops reliably deliver <20 ms latency for control, supporting cluster-wide closed-loop operation in Starlink-scale NTN deployments (Baena et al., 21 Feb 2025).

7. Future Directions and Open Challenges

Research continues in several dimensions:

• Standardization: Inter-RIC protocols for multi-tier and federated orchestration, especially for TN-NTN convergence (Baranda et al., 3 Jul 2025).
• Hierarchical frameworks: LLM-driven, multi-modal guidance from non-RT RIC to RL-based near-RT RIC, with co-design for latency, privacy, and robustness (Bao et al., 25 Apr 2025).
• Resource efficiency: Space-grade virtualization, dynamic component split, and hardware-accelerated inference for RIC xApps (Baena et al., 21 Feb 2025, Almeida et al., 2023).
• Joint orchestration: Dynamic reconfiguration of split options, ML-based predictive scaling, interaction of rApp/xApp/dApp layers, and conflict-mitigation pipelines (Baranda et al., 3 Jul 2025, Lacava et al., 27 Jan 2025, Prever et al., 7 May 2024).
• Security and reliability: Quantum-safe key management, supply-chain attestation, and closed-loop output verification before E2 actuation (Soltani et al., 14 May 2024).
• Distributed and federated learning: Neuroevolution xApps, transfer learning, decentralized evolution architectures for large-scale multi-agent RICs (Kouchaki et al., 15 Jun 2025).
• NTN and cross-domain RICs: Cluster-based satellite RICs, THz-band links for ultra-low latency, and digital-twin-driven strategic control (Baena et al., 21 Feb 2025).

Ongoing research will resolve open issues such as consistent cross-RIC orchestration, compute and energy constraints, standardization for new interfaces and deployment models, and the integration of RICs with emerging 6G (NTN, RIS, digital twin) domains.