AirAnchor: Bridging Local and Global Spatial Information for Zero-Shot Aerial Vision-and-Language Navigation

Aerial Vision-and-Language Navigation (Aerial VLN) requires drones to follow natural-language instructions in complex 3D urban environments. Existing zero-shot methods typically struggle because they rely on either local geometric representations (which are good for immediate actions but lack long-term context) or global memory (which helps with planning but often lacks precise local grounding). AirAnchor addresses this by introducing spatial anchors—a shared primitive that bridges these two scales. The system consists of three modules: Query-Driven Spatial Anchor Grounding, which identifies decision-relevant objects and directions; Persistent Object Spatial Memory, which maintains a global knowledge base of landmarks; and a Spatially-Informed Navigation Agent, which integrates these inputs into a closed-loop decision-making framework.

AirAnchor significantly outperforms existing zero-shot baselines on the AerialVLN benchmark. By using object anchors as a shared primitive, the agent can naturally accumulate locally grounded evidence into a persistent global knowledge base. This allows the drone to perform precise 3D maneuvers—such as descending to street level—while also retrieving landmark priors to navigate toward targets that are currently outside the drone's field of view. Ablation studies confirm that both the local Egocentric Anchor Graph and the global Spatial Object Knowledge Base are essential for performance, with the combination providing a 39-52% improvement in success metrics over the strongest baseline.

This work demonstrates that effective zero-shot navigation does not require an exhaustive 3D map of the environment. Instead, by distilling spatial information into a sparse set of semantically meaningful anchors, drones can achieve robust, interpretable navigation with significantly lower computational overhead than fine-grained cognitive agents. This approach provides a scalable path toward generalizable aerial AI that can adapt to new urban environments without task-specific training.