A Synthetic Dataset for Robotic Food Handling System

This paper aims to address the issue of labor shortage in the food industry by constructing a food automatic sorting system to replace manual labor and alleviate the workload of employees. Currently, deep learning is widely applied in the food industry. However, the process of creating a self-made dataset through capturing and annotating images consumes significant resources. To overcome the challenge, this study adopts a method of automatically generating the dataset, specifically generating instance segmentation data, to train the deep learning model for visual prediction of overlapped food items. Additionally, a soft robotic end-effector is used for food handling to prevent food items from damage during the sorting process. By implementing the proposed food sorting system, the overlapped food items were accurately predicted using a limited number of food items, and were successfully sorted in the food handling tasks.