Stability Ensured Deep Reinforcement Learning for Online Bin Packing

The Online Bin Packing Problem (OBPP) aims to determine the optimal loading position for each incoming item to maximize bin utilization, a critical challenge in various industrial applications. While many studies have focused on learningbased policies and heuristic approaches to enhance packing efficiency, stability constraints have largely been overlooked. In this work, we propose a computationally efficient method to validate stable loading positions for incoming items without requiring exact knowledge of their physical properties, such as mass. Our approach leverages the concept of Load-Bearable Convex Polygons (LBCPs), which provide substantial support forces to ensure structural stability. We further integrate our static stability validation framework into a state-of-the-art deep reinforcement learning (DRL) model, guiding it to learn physicsfeasible packing strategies. Experimental results demonstrate that our stability-aware DRL model achieves comparable packing efficiency while ensuring robust bin stability, offering a significant advancement in practical OBPP applications.