Design and evaluation of robotic seed potato sorting system based on monocular two-stage ROI-guided detection framework

• A seed potato sorting system with RGB vision and dual-axis robot arm is proposed. • A two-stage ROI-guided framework improves small bud detection accuracy. • Monocular localization achieves millimeter-level precision under oblique views. • Field experiments show 97.05% success rate in seed potato sorting. Automated seed potato sorting, especially based on the presence or absence of buds, is essential for potato production. However, current sorting practices face significant challenges in accurately localizing seed potatoes on moving conveyors, comprehensively inspecting their entire surfaces, and achieving high-precision bud detection, which remains difficult to accomplish with existing automation solutions. To address this challenge, we have developed a vision-guided seed potato sorting system that integrates three RGB cameras, a dual-axis robotic manipulator, and a control module. The system leverages YOLOv11 combined with ByteTrack for robust real-time object detection and tracking, enabling consistent identification of potatoes and their buds. A two-stage ROI-guided detection framework with foreground-optimized data augmentation was implemented to improve the model’s sensitivity to small-bud features. Furthermore, we designed a monocular oblique-view spatial localization method based on geometric projection to achieve millimeter-level localization precision. A grasp planning module synchronizes robot motion with conveyor dynamics to ensure real-time grasping. To validate the reliability of the proposed system, a field experiment was conducted under realistic conditions. In detection, experimental results showed 95.9% bud detection precision. In the localization experiment, the monocular oblique-view method achieved a mean localization error of 2.73 ± 1.77 mm. Field tests on 374 seed potatoes demonstrated a 97.05% sorting success rate. These results confirm the system’s responsiveness and operational reliability, offering a practical reference for deploying intelligent agricultural automation in real-world scenarios.