Topology-Adaptive World-Model DRL for AAV Data Collection With Incomplete Information

Abstract

Autonomous aerial vehicle (AAV)-assisted data collection in wireless sensor networks (WSNs) is challenging under uncertain topology and limited prior information. This paper investigates such scenarios and formulates AAV trajectory planning as a multi-objective partially observable Markov decision process (POMDP) that jointly optimizes age of information (AoI), energy consumption, and coverage ratio. We introduce a graph encoder that enables structural reasoning over heterogeneous random topologies and supports generalization to unseen deployments. Built upon temporal difference learning for model predictive control v2 (TDMPC2), a world-model-based deep reinforcement learning (DRL) algorithm, we develop TDMPC2-Graph (TDMPC2-G), which executes multi-step predictive rollouts to improve trajectory planning efficiency and quality. Simulation results show that model-free baselines SAC and TD3 fail to learn effective policies in this challenging setting, whereas TDMPC2-G demonstrates robust performance across all prior information levels. Compared to exhaustive coverage strategies, TDMPC2-G reduces terminal AoI by $10- -14\%$ , while maintaining comparable energy consumption and coverage ratios, thereby validating its robustness across structurally heterogeneous random topologies.