Detection for Robotic Depalletizing of Interlacing Bricks Using YOLO26-OBB

Abstract

This paper addresses the problem of reliable object detection for robotic depalletizing in real-world industrial environments, where fac-tors such as occlusions, dust, and lighting variability significantly affect perception systems. Building on previous work based on segmentation and geometric post-processing, we propose a simplified, task-aligned app-roach using YOLO26-OBB for direct oriented object detection. The proposed method replaces the multi-stage approach with a single-stage detector capable of predicting oriented bounding boxes, enabling direct estimation of object position and rotation required for robotic grasping. To evaluate the impact of annotation strategies, three dataset variants were introduced: a multi-class dataset including occlusions and pallets, a single-class dataset of all visible bricks, and a task-aligned dataset containing only top-layer (pickable) bricks. Experimental results demonstrate strong detection performance across all variants, with the task-aligned setting providing the most practically relevant and most stable results for the nano and small model variants. The results indicate that aligning dataset annotations with actionable robotic constraints can improve practical usability and reduce unsafe detections. Compared to the previous segmentation-basedapproach, the proposed method simplifies the perception process while maintaining high detection quality. The findings suggest that task-oriented detection formulations are crucial for the deployment of reliable robotic perception systems in industrial environments.