A new AI framework called GrabVG is proposed for visual grounding in Unmanned Aerial Vehicle (UAV) imagery. Visual grounding involves identifying a target object in an image based on a natural language description. The authors address challenges such as high visual redundancy and topological ambiguity by decomposing the task into two stages: preattentive hypothesis search and graph-attentive feature binding. GrabVG uses a compact set of reliable object hypotheses, organized into a sparse graph, to enable efficient spatial reasoning and accurate target localization. Experiments show that GrabVG achieves a favorable accuracy-speed trade-off, outperforming baseline methods. --- Why it matters: This matters because visual grounding is an essential task for applications such as autonomous drones, where accurate identification of objects in complex scenes is crucial. The proposed framework's ability to efficiently handle crowded scenarios and topological ambiguity can improve the reliability and safety of these systems. Source: https://arxiv.org/abs/2608.18996