Learning to Place: Transformer-Based Movable Antenna Array Design for Region-Oriented Wireless Sensing

Abstract

Movable antennas (MAs) boost sensing capabilities by exploiting the extra degrees of freedom (DoFs) through flexible positioning. The design of MA arrays is typically challenged by directional uncertainty, where the target’s precise angle is unknown. To address this, we propose a region-oriented design framework focusing on a target region of interest (RoI). Specifically, we introduce two key metrics: the average Cramér-Rao bound (CRB) within the RoI to ensure accuracy, and the average energy leakage outside the RoI to mitigate interference. Based on these metrics, we jointly optimize antenna positions and waveforms to minimize their weighted sum. To tackle the problem’s non-convexity, an alternating optimization (AO)-based algorithm is developed, where closed-form solutions are derived for iterative updates of antenna positions and waveforms. To facilitate real-time deployment, we utilize the high-quality dataset generated by this algorithm to develop Geoformer, a geometry-aware Transformer that directly maps sensing requirements to optimal antenna positions while supporting variable-length array configurations within a predefined maximum antenna size. Simulations demonstrate that the proposed algorithm significantly outperforms fixed antenna arrays (FPAs), and the Geoformer reduces inference latency by orders of magnitude compared to the iterative solver.