A neural network approach to robotic palletizing

This thesis presents a neural network-based approach to robotic palletizing, addressing the challenge of optimizing the packing efficiency and stability of objects in industrial environments. Palletizing, a crucial task in logistics and automation, requires precise placement strategies to maximize space utilization while ensuring structural integrity. The proposed method leverages a custom-designed convolutional neural network (CNN) to predict feasible packing actions in a three-dimensional space. A Proximal Policy Optimization (PPO) algorithm is employed to train a reinforcement learning agent, enabling it to autonomously learn efficient packing policies. A feasibility mask is integrated into the agent’s decision-making process, ensuring only valid actions are selected during training and deployment. The system was tested on synthetic datasets representing various box dimensions and pallet sizes. Results demonstrate significant improvements in packing density, stability, and computational efficiency compared to heuristic-based methods. The feasibility mask and reward shaping techniques effectively guided the agent toward optimal packing configurations, achieving a high bin fill rate while avoiding structural inconsistencies. This work contributes to the field of robotic automation by providing a scalable, datadriven framework for solving complex packing problems, with potential applications in manufacturing, logistics, and supply chain management.