PalletFit: Stability-Aware Online 3D Bin Packing via Transformer-Based Reinforcement Learning

Abstract

Industrial palletizing increasingly requires stable 3D bin packing of items with diverse sizes. However, prior studies often restrict size variability and overlook physical stability, resulting in tilted or toppled placements and product damage. We introduce PalletFit, an online stability-aware 3D bin packing framework designed for variable-sized items, unifying planning and execution through the Generation, Selection, and Execution stages. In the Generation stage, we propose Edge-Projection (EDP), which discretizes the continuous placement space into a compact discrete candidate set using lightweight post-processing. In the Selection stage, a Feature-Tokenizer Transformer-based Reinforcement Learning (RL) agent jointly selects items and placements under strict heuristic feasibility constraints to ensure physical realizability. In the Execution stage, we employ a robot motion primitive, Jam-Motion, to compensate for sensing and control errors through final contact alignment before release. Experimental results demonstrate that PalletFit markedly outperforms existing baselines, achieving dense and physically stable packing. Real-world experiments on an industrial robot further confirm robust execution with no observed stability violations.

Videos are publicly available above. Code will be released upon paper acceptance.

Method Overview

• Generation: Edge-Projection (EDP) turns continuous placement space into a compact candidate set.
• Selection: Feature-Tokenizer Transformer RL jointly selects item + placement with action masking under feasibility constraints.
• Execution: Jam-Motion aligns by contact to absorb sensing/control errors.