This study proposes a fruit-centric autonomous navigation framework for vineyard cul- tivation support robots operating in overhead trellis-grown Shine Muscat vineyards. Vineyard tasks such as cluster shaping, thinning, and harvesting require precise ...
This study proposes a fruit-centric autonomous navigation framework for vineyard cul- tivation support robots operating in overhead trellis-grown Shine Muscat vineyards. Vineyard tasks such as cluster shaping, thinning, and harvesting require precise po- sitioning directly beneath grape clusters. However, conventional navigation methods relying on QR markers, GNSS, or fixed landmarks are unsuitable for dynamic field environments, where cluster appearance and distribution vary throughout the growing season. To address this challenge, this research develops a navigation system that esti- mates grape-cluster positions using upward-facing visual detection and integrates them with SLAM-based mapping to enable sequential autonomous navigation without exter- nal landmarks. The proposed system incorporates two upward-facing detection models (bunch and flower models based on YOLO11), a TF-based coordinate logging node, Cartographer SLAM, AMCL localization, and the Nav2 navigation stack. During mapping, all de- tected cluster positions are recorded in the map frame. Offline DBSCAN clustering then fuses redundant detections and computes stable cluster centroids. These fused po- sitions are used as navigation goals. During execution, the robot sequentially visits each target while continuously adjusting its final alignment using real-time visual feedback. Three experiments were conducted: mapping accuracy evaluation, indoor navi- gation, and vineyard navigation. The mapping experiment with 10 known targets achieved a mean fusion error of 0.108 m, demonstrating the reliability of the hybrid de- tection–mapping pipeline. In indoor tests, single-goal navigation achieved 94.5% success, an average travel time of 22.3 s, and a mean final-position error of 0.041 m, confirming that the proposed navigation system functions accurately in disturbance-free environ- ments. Sequential four-goal navigation achieved 78.1% success, revealing minor error accumulation. In contrast, vineyard tests exhibited 88.5% single-goal success, 66.7% sequential success, and an average final-position error of 0.073 m, indicating that field- specific disturbances—such as uneven ground, LiDAR ground reflections, and odometry drift—are the primary sources of degradation rather than algorithmic limitations. These results demonstrate that the proposed fruit-centric navigation system enables practical and precise approaches to grape clusters without relying on external markers or GNSS. The hybrid online–offline approach provides a strong foundation for integrating manipulation tasks and ultimately reducing labor demands in viticulture.