Design, integration, and field evaluation of a selective harvesting robot for broccoli

The agronomic characteristics of broccoli necessitate selective harvesting in multiple batches, highlighting an urgent need for a selective harvesting robot to alleviate labour constraints. However, current research has inadequately addressed the problems of maturity identification of broccoli heads, fast and safe movement of the manipulator, and efficient and stable end-effector. Therefore, we proposed a semantic segmentation network called Broccoli Segmentation (BroSeg) for the mature identification and localisation of broccoli. BroSeg incorporated a lightweight backbone network, attention mechanisms, densely connected atrous spatial pyramid pooling, and a post-processing module. BroSeg achieved a Mean Intersection over Union (mIoU) of 58.92 % and a mean category prediction accuracy of 81.63 %. Using a collaborative simulation based on the Robot Operating System (ROS) and conducting comparative experiments, we selected the Batch Informed Trees (BIT*) algorithm that was most suitable for broccoli harvesting tasks. The effectiveness of the proposed method was validated through collaborative simulation and field experiments. Based on morphological analysis and cutting experiments, we designed an integrated gripper-cutting end-effector that mimics human hand-pinching for broccoli harvesting. The success rate of field harvesting reaches 86.96 %. This research integrates the functionalities of perception, manipulation, and cognition to construct a broccoli selective harvesting robot. Field experiments demonstrate a selective harvesting success rate of 63.16 %, with an average time of 11.9 s, validating the effectiveness and potential of the system.