xApp Empowered Resource Management for Non-Terrestrial Users in 5G O-RAN Networks

Abstract: This paper introduces a proactive Unmanned Aerial Vehicle (UAV) mobility management xApp for Open Radio Access Network (O-RAN) Near Real-Time Radio Intelligent Controller (Near-RT RIC) environments, employing Double Deep Q-Network (DDQN) reinforcement learning (RL) enhanced with transfer learning to optimise handover decisions for UAVs operating along predetermined flight trajectories. Unlike reactive approaches that respond to signal degradation, the proposed framework anticipates network conditions and minimises both outage probability and handover frequency through predictive optimisation. The system leverages centralised weight averaging to consolidate knowledge from multiple flight scenarios into a global model capable of generalising to previously unseen operational environments without extensive retraining. A comprehensive evaluation demonstrates that the proposed framework achieves a favourable trade-off between handover frequency and connectivity reliability, reducing handover events by up to 54.6% compared to greedy approaches while maintaining outage probability at practically negligible levels. The results validate the effectiveness of intelligent learning-based approaches for UAV mobility management in next-generation O-RAN architectures, thereby contributing to seamless integration of aerial user equipment into cellular networks.

• The paper introduces a DDQN-based xApp in the O-RAN Near-Real-Time RIC that uses UAV flight paths and SINR data to proactively optimize handovers and resource management.
• The proposed DDQN reduced handovers by 54.6% compared with greedy selection, achieving 2.45 handovers on average and a 0.048% outage probability across simulated flight paths.
• The paper demonstrates that centralized weight averaging can transfer learned policies between trajectories, while noting that real-world testbeds, multi-UAV scenarios, and reward trade-off analysis remain necessary.

Overview and motivation

This paper by Qazzaz et al. addresses handover management for low-altitude UAVs operating as aerial user equipment within 5G Open Radio Access Network (O-RAN) deployments. The authors deploy a Double Deep Q-Network (DDQN) reinforcement learning agent as an xApp inside the O-RAN Near-Real-Time RIC, exploiting the fact that many UAV missions (surveillance, delivery, inspection) fly predetermined trajectories. Rather than reacting to signal degradation, the framework uses flight-path information—available through UTM systems and O-RAN Enrichment Information interfaces—to anticipate network conditions and optimise handover timing proactively.

The core problem is that conventional reactive handover schemes, designed for terrestrial users, perform poorly for aerial UE: high LoS probability at altitude produces strong inter-cell interference, and 3D mobility induces frequent handovers with attendant signalling overhead and radio link failure risk. The paper frames handover optimisation explicitly as a resource management problem, jointly minimising outage probability and handover frequency.

System model and learning formulation

The system model considers $M$ terrestrial base stations serving a UAV along a known 3D trajectory, using the 3GPP UMa channel model with probabilistic LoS/NLoS path loss for altitudes between 22.5 m and 100 m. Link quality is characterised by SINR against aggregated inter-cell interference, and outage is defined empirically via a binary indicator on a SINR threshold $\gamma_{th}$. Handover logic is formalised through connection stability, mandatory handover trigger, and proactive handover conditions governed by a margin $\gamma_{mgn}$.

The DDQN agent's state comprises current and next UAV positions, SINR from the three strongest candidate base stations, their identifiers, and the serving cell index; actions are stay/switch among these candidates. The reward combines an outage penalty and a sigmoid-smoothed handover penalty with equal weights ($\alpha_o = \beta_h$), a choice the authors acknowledge leaves trade-off exploration to future work. Notably, initial experiments with PPO and A2C failed: policy gradient methods converged prematurely to deterministic "no-handover" policies, with entropy collapsing toward zero and rewards stuck between −300 and −500. The authors attribute this to high gradient variance and the dominance of immediate handover penalties—a finding that itself constitutes useful evidence for the suitability of value-based methods in delayed-reward handover problems.

Transfer learning via centralised weight averaging

A central contribution is a transfer learning pipeline inspired by federated averaging but operating offline and centrally: individual DDQN models are trained on distinct flight paths, their weights averaged into a global model, and lower layers frozen while upper layers are fine-tuned on new scenarios. The authors are explicit that this is not distributed federated learning—there is no iterative client-server communication or privacy motivation—and they concede that optimality across all conceivable flight scenarios cannot be guaranteed; effectiveness depends on the representativeness of training trajectories.

Empirical weight-similarity analysis supports the approach: cosine similarity between individual models ranges from 0.878 to 0.881 before fine-tuning, indicating that independently trained models learn comparable decision patterns, and similarity is preserved after fine-tuning while Euclidean distances decrease slightly (≈35.3 to ≈35.0). The DDQN-based global model also fine-tunes faster and to better final performance than its DQN counterpart, reinforcing the choice of the double-network architecture.

O-RAN integration

The paper details a concrete xApp lifecycle aligned with O-RAN specifications: offline training in the Non-RT RIC using KPMs and UTM-derived trajectory data; deployment as a container in the Near-RT RIC via the A1 interface; and real-time operation subscribing to E2SM-KPM for SINR reports and E2SM-RC for issuing handover commands. The authors argue that proactive decisions mitigate sensitivity to Near-RT RIC control-loop latency, since handovers are anticipated before degradation occurs. This integration narrative is architecturally credible but remains simulation-only; no testbed measurements are reported.

Evaluation results

The simulation comprises five base stations over a 2 km × 2 km urban area at 2.1 GHz, 45 dBm transmit power, UAV speed of 10 m/s, 0.1 s timesteps, ten distinct flight paths, and 500 episodes per path. All base stations share one carrier—an intentionally conservative worst-case interference assumption, though it limits conclusions for heterogeneous-frequency deployments.

Method	Handovers (mean ± std)	Outage prob. (%)
DDQN (proposed)	2.45 ± 0.73	0.048 ± 0.021
DQN	2.69 ± 0.82	0.052 ± 0.023
Greedy	5.40 ± 1.12	0.021 ± 0.011
Hysteresis	4.30 ± 0.61	0.089 ± 0.047
MOP	4.00 ± 0.87	0.041 ± 0.019

The headline result is a 54.6% reduction in handover events relative to the greedy scheme, achieved while keeping outage probability at 0.048%—an absolute increase of only 0.027 percentage points over greedy's best-in-class 0.021%. The authors argue this trade-off favours the learned policy operationally, since excessive handovers carry greater mission risk than marginal outage differences. DDQN dominates hysteresis on both metrics simultaneously and reduces handovers by 38.8% versus MOP at comparable outage. Against standard DQN, the double-network architecture yields an 8.9% further handover reduction, attributed to mitigated Q-value overestimation bias. Sensitivity analysis across SINR thresholds confirms these rankings: greedy retains the lowest outage under stringent thresholds but at disproportionate handover cost, while hysteresis suffers the worst outage due to delayed necessary handovers.

Limitations and open questions

Several constraints bound the validity of these results. All training and evaluation data are generated in simulation; no real network measurements or hardware testbed validation are included, which the authors identify as a key next step. The single-carrier interference assumption may overstate interference relative to practical multi-carrier deployments. Equal reward weighting ($\alpha_o = \beta_h$) is adopted without exploring the achievable Pareto frontier of the outage–handover trade-off. The transfer learning guarantee is empirical rather than theoretical, resting on the representativeness of ten training paths in a five-base-station topology. Scalability to larger networks, multi-UAV coordination, and priority-aware reward shaping for mission-critical operations remain unaddressed.

Conclusion

This paper demonstrates that a DDQN-based xApp, exploiting predetermined UAV flight paths available through O-RAN enrichment interfaces, can convert reactive handover management into predictive resource optimisation. The principal quantitative claim—a 54.6% handover reduction over greedy selection with negligible outage penalty—is supported by statistically consistent results across 500 episodes per path, and the centralised weight-averaging transfer mechanism shows measurable knowledge consolidation across diverse trajectories. The work's main outstanding question is whether these simulated gains survive contact with real O-RAN testbeds and multi-UAV operational environments.