Development of a Smart Depalletizing System with 3D Vision and Artificial Intelligence

This paper presents the design and development of an intelligent robotic depalletizing system that integrates 3D vision, deep learning–based oriented object detection, geometric calibration, and robot grasp planning into a unified framework. To accurately extract the 3D locations of target cartons on a pallet, the proposed vision module combines RGB-based detection with depth-based geometric computation, enabling robust localization in cluttered scenes while achieving metric-level spatial accuracy. Carton poses are estimated using YOLOv12-OBB, which predicts oriented bounding boxes fitted exclusively to the visible top faces of cartons; partially occluded or underlying boxes are intentionally ignored so that the computed centroids correspond to the centers of the uppermost surfaces that are actually graspable. A vertically mounted RGB-D camera acquires the scene, and the 3D positions of detected tops are reconstructed from the aligned depth map and transformed into the robot base frame using a closed-form SVD-based calibration method. A top-down suction grasp strategy is then applied, with grasp poses derived from object centroids and surface normals and converted to joint commands via inverse kinematics. Extensive experiments on various pallet configurations, ranging from uniform layouts to multi-directional stacking, demonstrate that the YOLOv12-OBB + 3D vision system reliably detects top-layer cartons regardless of orientation and layout. Quantitative evaluation indicates mean calibration errors of 2.17mm, 2.71mm, and 1.87mm along the X, Y, and Z axes, and grasp errors of about 4.0 mm in the horizontal plane and 2.8 mm vertically, which are well within tolerances for suction-based grasping of objects larger than 50 mm.