Space-O-RAN: Hierarchical Satellite RAN

• Space-O-RAN is a hierarchical distributed radio access network architecture that extends open O-RAN disaggregation to satellite and lunar deployments.
• It employs a three-layer control scheme—operational, cluster, and SMO—to achieve ultra-fast local adaptations and coordinated policy management across space and ground nodes.
• Dynamic interface mapping and adaptive signaling ensure robust, low-latency connections despite orbital dynamics, energy limits, and challenging non-terrestrial conditions.

Space-O-RAN is a distributed, hierarchical radio access network (RAN) architecture that extends the disaggregation principles and open interfaces of terrestrial Open RAN (O-RAN) into non-terrestrial networks (NTNs), including large-scale LEO satellite constellations and lunar wireless deployments. By employing on-board real-time control, in-orbit cluster-level orchestration, and global, ground-based policy management, Space-O-RAN enables intelligent, scalable, and interoperable integration of satellite and terrestrial systems for 6G-era applications (Baena et al., 21 Feb 2025, Baena et al., 12 Jun 2025, Baranda et al., 3 Jul 2025).

1. Hierarchical Control and Functional Splitting

Space-O-RAN features a strict three-layer hierarchical control splitting, each mapped to specific timing and functional constraints:

• Operational (On-Board) Layer (ms–s): Every satellite includes a minimal RAN stack (s-RU, s-DU, s-CU) hosting lightweight distributed RAN Applications (dApps), such as real-time beam steering and scheduling, running on resource-constrained onboard CPUs/GPUs. These enable local, autonomous adaptation to link conditions within tight deadlines (≤10 ms), independent of ground-based orchestration (Baena et al., 21 Feb 2025).
• Cluster Coordination (Cluster-Level) Layer (s–10s of s): Proximate satellites are grouped into clusters interconnected using Inter-Satellite Links (ISLs, typically ≲10 ms one-way). A designated Leader satellite aggregates telemetry, performs cluster-wide control (e.g., via sApps), and ensures continuity using a FeD-E2 re-election protocol for failover resilience. Control loops are closed at the cluster level (e.g., T_cyc ≈ 5–10 ms) utilizing state feedback and pre-computed gain matrices from LQR/distributed PD controllers (Baena et al., 21 Feb 2025).
• Strategic (Terrestrial) SMO Layer (min–h): The Service Management and Orchestration (SMO) platform on the ground maintains digital twins (DTs) of satellite clusters, collects Key Performance Indicators (KPIs), and issues global RAN policies (spectrum allocation, AI models) via feeder links. High-latency bulk data and long-term AI/model updates are offloaded here (Baena et al., 21 Feb 2025).

This hierarchy allows ultra-fast local adaptation while ensuring overarching policy alignment and efficient distribution of computational and energy burdens.

2. Functional Splits and Architecture Taxonomy

Space-O-RAN implements functional splits of O-RAN/3GPP building blocks between space and ground nodes. Three main split classes exist (Baranda et al., 3 Jul 2025):

Option	Satellite Functions	Ground Functions	Primary Use
Split-2	O-RU + O-DU	O-CU + All RICs + 5GC	Simplest onboard, tight F1 midhaul; supports distributed RIC at cluster level
Full gNB	O-RU + O-DU + O-CU	UPF + SMO/RICs	Enables local L1–L3, reduces feeder link congestion, enables onboard near-RT control
Full gNB + UPF	All gNB + UPF + (optional) SEC	SMO/Non-RT RIC; N6 interface to data NW	Maximizes local breakout, full edge compute, challenging power/thermal budget

Latency and power constraints directly shape which splits are feasible. For instance, only on-board near-RT RIC allows sub-10 ms control loops; full gNB/UPF splits demand as much as ≈200 W digital payload power and can only be supported on satellites with sufficient energy, thermal, and space-grade compute capacity (Baranda et al., 3 Jul 2025).

3. Dynamic Interface Mapping and Adaptive Signaling

The architecture supports dynamic mapping of O-RAN interfaces (E2, A1, O1, F1) onto satellite links (ISL, GSL, FL), based on one-way link delay (ℓ_L) and quality (Q_L):

• Near-RT interfaces (E2, F1) are mapped when ℓL ≤ T{E2} (e.g., ≈10 ms).
• Management/policy interfaces (A1, O1) are mapped for T_{E2}< ℓL ≤ T{A1} (e.g., up to 50 ms).
• Links exceeding these thresholds are reserved for high-latency user-plane traffic (Baena et al., 21 Feb 2025).

A remapping event is triggered if link latency or quality crosses interface-specific thresholds: $$\ell_L(t) > T_I \quad \lor \quad Q_L(t) < Q_{\min} \longrightarrow \text{re-map}(I,L')$$ where L' is an alternative link meeting current QoS requirements. This enables signaling robustness under the orbital dynamics and impairment-induced topology changes prevalent in NTNs.

4. On-Board AI and Distributed Inference Under Resource Constraints

Allocation of AI tasks on satellites is formulated as a constrained optimization balancing inference time, energy, and bandwidth: $$\min_{x\in\{0,1\},\, f\leq f_{max}}\left(\alpha\frac{C x}{f} + (1-\alpha)\frac{\Delta(1-x)}{B_{link}}\right)$$ subject to power consumption and battery budget (Baena et al., 21 Feb 2025), with

• $x=1$ for on-board computation,
• $x=0$ for inference offloading.

Tradeoff weights α allow tuning between latency and energy. Estimations of power and required memory highlight the necessity of lightweight models and hardware-aware inference, favoring federated or cluster-level approaches for training and model management.

In lunar/space scenarios, semantic-agentic extensions leverage Distributed Cognitive Agents (DCAs), Model Context Protocol (MCP), and peer-to-peer Agent-to-Agent (A2A) communication to enable context-aware, delay-adaptive reasoning and bandwidth-aware semantic compression. DCAs use local and federated context to decide computation vs. communication, adjusting logic according to resource budgets (Baena et al., 12 Jun 2025).

5. Cluster Formation, Cluster-Level Orchestration, and RIC Placement

Satellite clusters are formed using proximity-based clustering in orbital angle–height space or via spectral clustering on the constellation graph Laplacian (Baena et al., 21 Feb 2025). Control within clusters relies on a Leader–Follower scheme, closed-loop state feedback, and ISLs for low-latency telemetry aggregation.

Placement of Radio Intelligent Controllers (RICs) is governed by architecture split, latency budget, and on-board capacity (Baranda et al., 3 Jul 2025):

• Non-RT RICs reside on ground (SMO), dedicated to slow policy and heavy model training.
• Near-RT RICs require co-location with DU on-board for <10 ms control loops; ground-based near-RT RICs suffice for slower functions only.
• Extensions allow split, distributed, or hierarchical RIC models, with on-board, ISL-mediated clusters and ground coordination for scalable, robust management.

Satellite payload limits (power, SWaP, radiation tolerance) directly affect allowable RIC placement and function distribution.

6. Integration with Semantic Control and Mission-Specific Intelligence

Recent extensions incorporate a semantic agentic layer into Space-O-RAN using MCP and A2A, supporting context-aware, distributed decision making in lunar and space environments (Baena et al., 12 Jun 2025). Key concepts include:

• Distributed Cognitive Agents (DCAs) on nodes (rovers, landers, satellite-RANs), executing delay-adaptive, reasoning bounded by hardware and signal conditions.
• MCP servers/providers expose contextual, ontological APIs for predictive context (e.g., energy, location, mission status).
• Bandwidth-aware semantic compression to minimize transmission cost under bandwidth constraints, trading off data rate against semantic integrity using rate–distortion–semantic loss: $$\min_{C(\cdot)} \mathbb{E}\left[ \beta\|\mathbf{s}-C(\mathbf{s})\|^2 + (1-\beta) \mathrm{D_{sem}}(\mathbf{s},C(\mathbf{s})) + \lambda R(C(\mathbf{s})) \right]$$ Autonomy, robustness, and adaptability are enhanced by dynamically choosing control-participation levels and by enabling on-site semantic policy execution, even during prolonged link disruptions.

7. Performance Validation and Research Directions

Space-O-RAN has been validated via large-scale simulations mirroring Starlink-like constellations (6545 LEO satellites, 33 gateways) (Baena et al., 21 Feb 2025). Results demonstrate:

• Intra-cluster loops close within 21–25 ms, well inside 100 ms real-time constraints.
• >60% GSL bandwidth savings through local (on-board) dApp execution.
• SMO policy convergence within 10 rounds of federated averaging.
• Adaptive link-interface mapping sustains ≥95% packet delivery even during dynamic outages.

Key challenges and directions include:

• Adaptive orchestration engines for real-time split and RIC placement.
• Delay-tolerant and robust protocol extensions (e.g., Sat-OFH, Sat-F1, Sat-E2).
• Integration of federated learning and semantic policy distillation.
• Space-grade, power-efficient compute platform development.
• Security against jamming/spoofing via enhanced open interface protection (Baranda et al., 3 Jul 2025).

Standardization is ongoing in 3GPP Rel-19 and O-RAN Alliance working groups for onboard user-plane, edge compute, and interoperability between terrestrial and non-terrestrial 6G networks.

---

References

(Baena et al., 21 Feb 2025) Space-O-RAN: Enabling Intelligent, Open, and Interoperable Non Terrestrial Networks in 6G (Baena et al., 12 Jun 2025) Agentic Semantic Control for Autonomous Wireless Space Networks: Extending Space-O-RAN with MCP-Driven Distributed Intelligence (Baranda et al., 3 Jul 2025) On the Architectural Split and Radio Intelligence Controller Placement in Integrated O-RAN-enabled Non-Terrestrial Networks