Enhancing Package Depalletizing with A Reinforcement Learning Approach for Safe and Efficient Operations

抄録

In this research, a novel approach to the classical problem of package depalletizing in logistics systems is presented. The conventional depalletizing process relies on rule-based methods to ensure safety, but it only allows for a single type of action. This limitation can make the process slow and inefficient. We propose a hybrid depalletizing algorithm that combines a rule-based mothed with data-driven learning using Q-learning, which allows for shorter trajectory actions. This hybrid approach not only maintains the safety aspect of the conventional approach, but also enhances the efficiency of the depalletizing process. We also validate the algorithm in both simulation and real-world environments with clutters of different sizes. The results show that the algorithm increases efficiency by 19% compared to traditional methods, demonstrating its potential for efficient and safe depalletizing in logistics systems.