The deployment of Reinforcement Learning (RL) within the Open Radio Access Network (O-RAN) framework promises to enable intelligent and adaptive control of network functions. However, the transition from simulation-based environments to real-world deployments introduces a significant reality gap that can compromise the effectiveness and robustness of RL-based xApps. This paper investigates the impact of real-world data heterogeneity on the performance of RL algorithms in the O-RAN Service Network. By analyzing two real deployment datasets collected from O-RAN-compliant testbeds based on OpenAirInterface (OAI) and srsRAN, we empirically quantify how discrepancies in data distributions and measurement scales create fundamental challenges for RL-based decision-making. To address this issue, we propose a data normalization and preprocessing strategy that enhances model generalization across heterogeneous environments. Results confirm that the approach aligns the OAI and srsRAN distributions, with PCA capturing 90.3% of the variance and a 0.05 Silhouette Score. This yields a platform-agnostic state-space-preserving protocol-level semantics, enabling portable RL xApp development across heterogeneous O-RAN deployments.
Zandoli, S., Feraudo, A., Scotece, D., Foschini, L., Bellavista, P. (2026). Bridging the Reality Gap in O-RAN Networks: Designing Robust RL-based xApps for Heterogeneous Real-World Deployments. Institute of Electrical and Electronics Engineers Inc. [10.1109/icc59461.2026.11587913].
Bridging the Reality Gap in O-RAN Networks: Designing Robust RL-based xApps for Heterogeneous Real-World Deployments
Zandoli, Silvia;Feraudo, Angelo;Scotece, Domenico;Foschini, Luca;Bellavista, Paolo
2026
Abstract
The deployment of Reinforcement Learning (RL) within the Open Radio Access Network (O-RAN) framework promises to enable intelligent and adaptive control of network functions. However, the transition from simulation-based environments to real-world deployments introduces a significant reality gap that can compromise the effectiveness and robustness of RL-based xApps. This paper investigates the impact of real-world data heterogeneity on the performance of RL algorithms in the O-RAN Service Network. By analyzing two real deployment datasets collected from O-RAN-compliant testbeds based on OpenAirInterface (OAI) and srsRAN, we empirically quantify how discrepancies in data distributions and measurement scales create fundamental challenges for RL-based decision-making. To address this issue, we propose a data normalization and preprocessing strategy that enhances model generalization across heterogeneous environments. Results confirm that the approach aligns the OAI and srsRAN distributions, with PCA capturing 90.3% of the variance and a 0.05 Silhouette Score. This yields a platform-agnostic state-space-preserving protocol-level semantics, enabling portable RL xApp development across heterogeneous O-RAN deployments.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



