3D Vision robot online packing platform for deep reinforcement learning

Abstract

In modern logistics and manufacturing, online mixed palletizing stands as one of the key automation technologies, facing challenges brought by the diversity of packages and real-time demand. However, traditional palletizing methods typically rely on preset rules, making them ill-suited to handle diverse bins in dynamic, real-time environments. This limitation becomes especially pronounced when dealing with complex palletizing tasks. To optimize the accuracy and operational efficiency of the palletizing process, this study, based on 3D vision technology and deep reinforcement learning techniques, designs a bin positioning algorithm utilizing projected bounding boxes. Additionally, spatial rotation position encoding is integrated into the decision-making process of the online palletizing network, designing a deep reinforcement learning algorithm for online mixed palletizing based on a masked attention mechanism. The paper also introduces a novel heuristic method—Boundary Point, which updates the palletizing state chain using key-point heuristics and “spatial” heuristics, and employs a pointer network for tail node selection. Experimental results demonstrate that the proposed method significantly improves average space utilization across the RS, CUT1, and CUT2 datasets. Finally, a 3D vision-based robotic online mixed palletizing experimental platform is designed and built, proving the effectiveness and application potential of the proposed algorithm.