Belief aware manifold robust control for wireless channel optimization in UAV mounted intelligent reflecting surface systems

With the rapid development of low-altitude communications, integrated space-air-ground networks, and reconfigurable electromagnetic environments, UAV-mounted intelligent reflecting surfaces (UAV-RIS) have become a promising solution for wireless link optimization in emergency communications, vehicular access, and temporary coverage. However, most existing approaches focus on average performance and pay limited attention to stable control under incomplete state information, high-dimensional unit-modulus constraints, and distributional perturbations. To address these challenges, this paper proposes BMR-UAV-RIS (Belief-aware Manifold-Robust control framework for UAV-mounted RIS systems), which integrates belief-state representation, manifold-native action generation, and distributionally robust value evaluation. The proposed framework enables closed-loop dynamic control under partial observability and strong action constraints. Experiments are conducted on DeepMIMO and RL-AERPAW-DT. Results on RL-AERPAW-DT show that BMR-UAV-RIS achieves an average throughput of 84.7 Mbps, compared with 83.0 Mbps for Transformer-based deep reinforcement learning (TDRL), while its tail performance reaches 52.8 Mbps, outperforming TDRL by approximately 7.3%. Under increasing noise from 0 to 15%, the tail performance of BMR-UAV-RIS decreases from 64.5 to 53.2 Mbps, showing smaller degradation than major baselines. Overall, the proposed framework jointly addresses state representation, constrained action generation, and robust decision-making in dynamic UAV-RIS control. Its main advantage lies in robustness-sensitive settings, especially in improving tail performance and reducing degradation under perturbations.