Abstract
Online 3D Bin Packing Problem (3D-BPP) has widespread applications inindustrial automation. Existing methods usually solve the problem with limitedresolution of spatial discretization, and/or cannot deal with complex practicalconstraints well. We propose to enhance the practical applicability of online3D-BPP via learning on a novel hierarchical representation, packingconfiguration tree (PCT). PCT is a full-fledged description of the state andaction space of bin packing which can support packing policy learning based ondeep reinforcement learning (DRL). The size of the packing action space isproportional to the number of leaf nodes, making the DRL model easy to trainand well-performing even with continuous solution space. We further discoverthe potential of PCT as tree-based planners in deliberately solving packingproblems of industrial significance, including large-scale packing anddifferent variations of BPP setting. A recursive packing method is proposed todecompose large-scale packing into smaller sub-trees while a spatial ensemblemechanism integrates local solutions into global. For different BPP variationswith additional decision variables, such as lookahead, buffering, and offlinepacking, we propose a unified planning framework enabling out-of-the-boxproblem solving. Extensive evaluations demonstrate that our method outperformsexisting online BPP baselines and is versatile in incorporating variouspractical constraints. The planning process excels across large-scale problemsand diverse problem variations. We develop a real-world packing robot forindustrial warehousing, with careful designs accounting for constrainedplacement and transportation stability. Our packing robot operates reliably andefficiently on unprotected pallets at 10 seconds per box. It achieves averagely19 boxes per pallet with 57.4% space utilization for relatively large-sizeboxes.