YOLOv8-based Intelligent Robotic Harvesting Platform With a Custom 6-DOF Manipulator

Abstract

This study proposes a 6-degree-of-freedom (DOF) active re-observation method for detecting strawberry disease symptoms hidden from a single front view. The experimental evaluation was conducted using 50 mock strawberries, which were reused across 10 randomized trials. After grasping a fruit, the manipulator reorients it toward a red, green, blue plus depth (RGB-D) camera for backside inspection, supported by the small-sized You Only Look Once version 8 segmentation model, 18-layer residual network, RGB-D localization, and principal component analysis (PCA)-based roll estimation. In 200 trials per structure, backside simulated-lesion detection success rates were 64.50±11.65% with 4 DOF and 94.00±4.59% with 6 DOF. Across 10 repeated evaluations totaling 500 trials, 187 of 200 backside-only simulated-disease cases were detected, and 284 of 300 simulated-disease-condition trials were successfully identified and sorted (94.67±3.91%). A representative 50-target evaluation required approximately 18 min, corresponding to 21.6 s per target and 166.7 targets/h under the specified laboratory test composition condition; the displayed online processing rate ranged from 13.6 to 15.0 frames per second. These results demonstrate feasibility under controlled laboratory conditions, while greenhouse performance and damage to real fruit remain to be evaluated.