Joint Trajectory, Beamforming and Antenna Deployment Design for Movable Antenna-aided A2G Communication: A DRL-based Two-stage Method

This paper studies a movable antenna (MA)-aided Air-to-ground (A2G) system where the MA-equipped unmanned aerial vehicle (UAV) serves legitimate users. The objective is to jointly optimize the UAV trajectory, antenna positions, and beamforming vectors over a flight period to maximize the communication sum rate. To mitigate the challenge of acquiring fully traversed instantaneous CSI in MA-aided systems, a two-stage transmission scheme is proposed. In the first stage, the location-aware statistical CSI (S-CSI) is leveraged to train UAV trajectory and MA deployment policies for maximizing long-term utility. A deep reinforcement learning (DRL)-based dual space-scale network is proposed to maintain maneuvering flexibility in global policymaking while enhancing training performance through variable decoupling. In the second stage, once macro-level policies are determined, the instantaneous CSI (I-CSI) acquired by a small number of times channel estimation will refine the transmit beamforming, thus reducing the CSI acquisition overhead. Simulation results demonstrate that the proposed scheme outperforms single-layer learning in system performance, validating the application potential of MA technology in A2G networks.