From sky to ground: Enabling autonomous vineyard traversal with UAV-derived navigation maps

Reliable structural maps are essential for deploying autonomous ground robots in vineyards, yet producing these maps often still depends on manual real-time kinematic (RTK) surveying or teleoperated data collection. This bottleneck slows robotic deployment and can limit the reuse of spatial data within broader digital-vineyard workflows. In this paper, we present an operational aerial-to-ground pipeline that converts overlapping nadir Unmanned Aerial Vehicle (UAV) imagery into robot-ready vineyard maps. The pipeline integrates structural detection, metadata-driven geospatial projection, spatial clustering, and occupancy/topological map generation. By extracting features directly from raw overlapping frames, the system bypasses computationally expensive orthomosaic processing, accelerating the map generation timeline. We evaluate robustness across multiple seasons and flight altitudes, reporting map-level performance and robotic deployment feasibility. Using August 39 ft imagery, all five evaluated models produced sub-metre maps; YOLO11 segmentation achieved an RMSE of 0.242 m, while YOLO11 object detection reached a perfect map-level F1-score. A supplementary best-case capability analysis identified a minimum RMSE of 0.207 m using YOLO11 segmentation on March 39 ft imagery. Field trials on a Thorvald robotic platform, over approximately 500 m of autonomous row traversal per generated map, demonstrated reliable navigation, achieving a Mean Distance Between Interventions (MDBI) of up to 161.4 m. Furthermore, UAV data collection was consistently faster than baseline methods, cutting field survey time by up to 94% compared to robotic teleoperation in a 2.4-hectare commercial-vineyard scalability study. These results demonstrate that overlapping UAV imagery provides a practical route to generating vineyard navigation maps that support autonomous deployment under the tested conditions.