This study aims to present the design and experimentally evaluation of an autonomous rose-harvesting robot (RHR) for harvesting of bloomed roses in greenhouse environment using lightweight vision, depth-assisted localization and an integrated grasp-and-cut end-effector.
The proposed robotic system combines a rail-guided vehicle, a scissor-lift mechanism, an XY positioning stage, a 5-degrees of freedom robotic arm and an end-effector with embedded red, green and blue camera and time-of-flight (ToF) sensor. A data set of 3,160 manually annotated red, green and blue images of rose(flower), bud and mid-bud classes was used to train and compare different YOLO-based detection models. Then selected model was then integrated with ToF-based depth sensing for three-dimensional target localization and harvesting of mature roses. The end-effector was evaluated through gripping and cutting tests, followed by autonomous harvesting trials under varying rose orientations.
The selected YOLOv12n detection model revealed the best balance between computational efficiency and detection accuracy, resulting an mAP@0.5 of 0.82810 and the best mAP@0.5:0.95 of 0.56671, at 6.48 GFLOPs and an onboard inference latency of 0.01463 s. Experimental evaluation showed the stable performance of the designed end-effector with successful harvesting. Across 20 rose harvesting trials, the robotic system achieved mean successful harvesting of 91.5% with 10.8 s/rose. The observed failure modes were rose damage, incomplete cutting, failed gripping and misalignment.
To the best of authors’ knowledge, this study presents the experimental validation of RHR with integration of stage-aware detection, ToF-based localization and an integrated end-effector within a single platform. The findings of this study provide practical basis for rose harvesting in greenhouse environment.
