An Annotation-to-Detection Framework for Autonomous and Robust Vine Trunk Localization in the Field by Mobile Agricultural Robots

A research team from UC Riverside has developed a novel AI framework that enables mobile agricultural robots to autonomously locate and map vine trunks in challenging, unstructured vineyard environments. The system addresses a critical bottleneck in agricultural robotics: the need for large-scale, manually labeled real-world datasets. By employing a multi-stage detection architecture with cross-modal annotation transfer and early-stage sensor fusion, the framework can train a robust multi-modal detector using only limited and partially labeled training data. This approach allows the AI to generalize to previously unseen fields with diverse lighting conditions and varying crop densities.

The framework was validated through real-world testing in novel vineyard settings. When integrated with a customized LiDAR and Odometry Mapping (LOAM) algorithm and a tree association module, the system demonstrated high-performance trunk localization. It successfully identified over 70% of trees during a single robot traversal, achieving a mean distance error of less than 0.37 meters. The results prove that by leveraging incremental-stage annotation and training, the system achieves robust detection performance regardless of limited starting annotations. This breakthrough showcases significant potential for real-world, near-ground agricultural applications where environmental variability is high and manual data collection is impractical.