Robust Apple Grasping via PointNet-Based Semantic Segmentation and Centroid Localization

Colour-threshold pipelines for robotic fruit picking often fail under strong illumination changes and leaf occlusion. We introduce a PointNet++-centroid grasping pipeline that learns to segment apples directly in 3D point clouds, replacing the traditional HSV-depth heuristic. The network is trained on 500 hand-labelled RGB-D scenes.In 100 physical grasps with a DoBot e6 arm, our method reduces centroid localisation error from 5.2 ± 1.1 mm to 3.8 ± 0.9 mm and raises grasp success from 92% to 95%, adding only 15 ms of extra inference time relative to the HSV baseline. An ablation confirms that statistical outlier removal remains critical for both approaches.