To address the problems in traditional apple harvesting, such as labor shortage, high labor intensity, and low efficiency, we design a fully automated apple harvesting robot. It is equipped with a 6-DoF robotic arm, an RGB-D camera, a mobile chassis, a control system and a flexible end-effector. This study proposes a method combining RGB-D imaging with an improved YOLOv8s model for fruit detection and localization. By integrating PConv and ECA attention mechanisms, and adding a small-object detection layer in the neck, the model improves recognition accuracy for small fruits while maintaining a lightweight structure. Apple grading is performed online via estimated diameter measurements. To address spatial constraints in high spindle-shaped apple trees, a two-stage path planning strategy integrating a robotic arm and lifting platform was developed, enabling efficient and accurate picking. Experimental results demonstrate high system performance, with average recognition, localization, grading, and picking success rates of 91.38%, 88.68%, 91.49%, and 92.16%, respectively—all exceeding 85%.