Adaptive cross-modal UAV semantic communications via Hierarchical DRL

This paper investigates cross-modal task-oriented semantic communication for a low-altitude unmanned aerial vehicle (UAV) transmitting mission-relevant semantic packets to a ground station over a wireless uplink. Different from bit-level transmissions, the UAV first extracts task-relevant semantics from heterogeneous onboard sensors and then adaptively determines the sensing modality combination, semantic granularity, and transmit power according to channel variation and mission urgency. To capture the long-term coupling among semantic utility, energy consumption, and delay, the problem is formulated as a Markov decision process (MDP) that maximizes the long-term average semantic utility subject to average energy and delay constraints. A hierarchical deep reinforcement learning (DRL) framework is developed, where a high-level recurrent proximal policy optimization (PPO) policy selects the cross-modal semantic mode and semantic granularity, and a low-level soft actor–critic (SAC) policy determines the transmit power. In addition, a task-oriented semantic encoder is designed with cross-modal alignment loss, multi-granularity consistency loss, task loss, quantization loss, and rate regularization. The proposed framework enables semantic-level decision-making rather than merely physical-layer resource control. Extended simulation results show that the proposed hierarchical DRL-based scheme outperforms several baselines in terms of semantic utility, energy consumption, and delay.