Reinforcement Learning‐Based Relay and RIS Optimization for Covert 6G Ground Station Communications

ABSTRACT Dynamic channels, energy constraints, and the risk of warden detection pose significant challenges to securing covert communications in 6G ground station networks. To address these challenges, we propose a novel multi‐agent reinforcement learning (MARL) framework that jointly leverages mobile relays and reconfigurable intelligent surfaces (RISs) to optimize relay trajectories, transmit power, and RIS phase shifts. Unlike traditional optimization‐based and existing reinforcement learning approaches, the proposed scheme simultaneously maximizes covert throughput, minimizes energy consumption, and ensures robustness under imperfect channel state information. The optimization problem is solved using the multi‐agent deep deterministic policy gradient (MADDPG) algorithm. Simulation results show that the proposed framework maintains a warden detection error probability above in over of time slots, while delivering covert transmission rate improvements of up to compared to static‐RIS baselines and relative to greedy relay baselines. Moreover, the approach demonstrates resilience to channel estimation errors (e.g., the Gaussian perturbations with up to ) and offers computational scalability, with inference time per slot on GPU, making it suitable for real‐time implementation in the next‐generation satellite networks.