Multi-sensor fusion-based inter-row autonomous following navigation for collaborative harvesting robots
Xindong Ni, Rui Su, Funing Yu, Yufei Sun, Zhentao Zhu, Du Chen · Smart Agricultural Technology · 2026
• Proposed a human–machine collaborative harvesting mode for large-leaf crops. • Constructed a UWB–vision fusion navigation system for inter-row robot following. • Proposed a new inter-row navigation method based on autonomous ROI extraction. • Deployed and validated the proposed method on a robotic platform, confirming its functional reliability. The long-standing manual picking and manual transportation mode for harvesting large-leaf cash crops in hilly and mountainous regions is characterized by high labor intensity, low efficiency, and considerable safety risks. To address these challenges, this study developed a human-robot collaborative harvesting robot that enables manual picking and autonomous mechanical transportation, and proposed a multi-sensor-fusion autonomous following-navigation system for inter-ridge harvesting operations. The system fuses information from a camera, an ultra-wideband (UWB) positioning module, and an inertial measurement unit (IMU), and integrates a collaborative control strategy to achieve accurate human-robot following and dynamic path correction in inter-ridge environments. Three key technical innovations are presented. First, a positioning method that combines phase difference of arrival (PDOA) with an extended Kalman filter (EKF) is proposed to suppress noise-induced outliers and improve positioning accuracy and smoothness. Second, a real-time inter-ridge navigation-line extraction algorithm based on a region of interest (ROI) is developed to efficiently extract local row lines and provide timely path correction under complex field conditions. Third, an anomaly-suppression mechanism is introduced at the fusion layer to exploit the complementary strengths of global UWB guidance and vision-based local boundary perception, and a closed-loop collaborative controller is designed to control the tracked chassis. Field experiments showed that the proposed system achieves an average deviation of 2.01°, a maximum deviation of 5.43°, and an accuracy rate of 95.4% in navigation-line extraction, with an average processing time of 76 ms, providing continuous and stable navigation references. Target-following tests demonstrated reliable performance in challenging environments (e.g., dense crops and varying illumination), maintaining a human-robot relative distance within 2 m and a following error below 8.5 cm. In integrated navigation-following tests, the maximum lateral deviation and heading-angle deviation were 5.22 cm and 4.47°, respectively, both within acceptable limits. Overall, the proposed following-navigation scheme enables robust human-robot collaboration and precise path adjustment under heavy occlusion and variable lighting, offering a practical approach to mechanized harvesting of large-leaf cash crops in hilly and mountainous areas.