Autonomous Spaghetti Grasping for Bento Box Assembly: A Deep Reinforcement Learning Approach

In Japan, the integration of autonomous systems in the bento box production line has become imperative to meet the growing demand for efficiency and precision. A prevalent challenge in this context arises from the inclusion of spaghetti in every bento box, necessitating the development of a robotic system capable of autonomously grasping spaghetti with a specific weight and delicately placing it in the box. This study proposes robotic grasping based on Deep Reinforcement Learning (DRL) to address the noodle-like grasping system at a specified weight. Utilizing a novel DRL algorithm with a custom-designed reward function designed and custom policy structure provided with RGBD and additional input for agent to find a robust policy. This spaghetti grasping task has shown success, achieving acceptable accuracy with a Mean Absolute Percentage Error (MAPE) metric of approximately 19% after training in real-world environments. This shows a significant promise that DRL-based approach could offer a solution to assemble the noodle-like in bento box in the production line.