Leveraging Synthetic Training Data for Object Detection to Enhance Autonomous Depalletizing Systems

Autonomous robotic systems capable of detecting and handling industrial load carriers are a key prerequisite for future intralogistics processes. This paper presents a study on the application of synthetic training data for object detection in autonomous depalletizing. Following a literature review, our research explores the impact of incorporating synthetic training data on the performance of a YOLOv8 model under realistic industrial conditions. We evaluate the effectiveness of synthetic images of target objects with and without application-specific context. Furthermore, we investigate the benefit of adding depth data as model input and examine the performance of instance segmentation models compared to pure object detection. Our results show that models trained to segment synthetic RGB-D images resembling the real scenario achieve the highest detection accuracy. The findings contribute to understanding the potential of synthetic data in reducing labeling costs, improving model robustness, and advancing autonomous robotic solutions in industrial settings.