Hybrid Reinforcement Learning for a Dynamic Palletization Problem

This thesis investigates reinforcement learning approaches to the three-dimensional bin packing problem, a task of practical relevance in logistics and warehouse automation. The main challenges stem from the combinatorial complexity of the action space, the need to enforce hard geometric and stability constraints, and the additional accessibility limitations imposed by robotic manipulation. We introduce a hybrid reinforcement learning framework for the three-dimensional bin packing problem, combining an actor–critic model with a constraint-aware enumeration module. This integration reduces the action space and enforces feasibility, resulting in faster training and higher packing efficiency compared to a standard baseline. Extensions addressing temporal context and accessibility constraints were also explored, showing that while temporal modeling offered limited gains, a layers heuristic effectively enforced realistic stacking policies at the cost of increased learning complexity. Overall, the results highlight the effectiveness of hybrid RL approach for bin packing and its potential for practical application.