[Paper] GENPACK: KPI-Guided Multi-Objective Genetic Algorithm for Industrial 3D Bin Packing

Source: arXiv - 2601.11325v1

Overview

The paper presents GENPACK, a novel genetic‑algorithm (GA) framework that tackles the industrial three‑dimensional bin‑packing problem (3D‑BPP) while explicitly optimizing for real‑world key performance indicators (KPIs) such as stability, balance, and handling feasibility. By weaving these KPIs into a multi‑objective fitness function, the authors achieve markedly better space utilization and pallet support on a large, production‑grade benchmark.

Key Contributions

• KPI‑driven multi‑objective GA – integrates stability, balance, and surface support directly into the fitness evaluation.
• Layer‑based chromosome encoding – captures the natural “layer‑by‑layer” stacking used on pallets, enabling more realistic packings.
• Domain‑specific genetic operators – custom crossover and mutation that respect item orientation, weight distribution, and handling constraints.
• Hybrid pipeline – combines the GA with fast constructive heuristics for initial population seeding and local refinement.
• Extensive empirical validation – evaluated on the BED‑BPP benchmark (1,500 real industrial orders), outperforming leading heuristic and learning‑based baselines by up to 35 % in space utilization and 15‑20 % in surface support, with lower result variance.

Methodology

1. Chromosome design – each individual encodes a sequence of “layers”; a layer lists items placed on the same height level, preserving the order in which items are stacked on a pallet.
2. Fitness function – a weighted sum of three KPIs:
   • Space Utilization (SU) – volume of items / pallet volume.
   • Surface Support (SS) – proportion of item weight supported by underlying items (proxy for stability).
   • Balance Index (BI) – deviation of the pallet’s center of gravity from the geometric centre.
     The GA maximizes SU while minimizing BI and maximizing SS, allowing the user to tune the relative importance of each KPI.
3. Genetic operators –
   • KPI‑aware crossover swaps whole layers between parents, preserving feasible stacking patterns.
   • Mutation randomly re‑orders items within a layer or moves an item to a neighboring layer, followed by a quick feasibility check.
4. Hybrid initialization – the initial population is seeded with solutions from fast constructive heuristics (e.g., first‑fit decreasing, shelf‑based packing) to give the GA a strong starting point.
5. Local refinement – after each GA generation, a lightweight “repair” heuristic adjusts any infeasible placements (e.g., items hanging in mid‑air) and fine‑tunes the balance.

The overall pipeline runs in batch mode: a set of orders is processed, the GA iterates for a fixed number of generations (or until convergence), and the best individual is output as the final pallet layout.

Results & Findings

Key take‑aways

• Embedding KPIs in the fitness function yields packings that are not only denser but also more stable and easier to handle.
• The layer‑based representation dramatically reduces the search space compared with flat item permutations, leading to faster convergence.
• Even with a modest runtime overhead, the approach scales to real‑world order volumes, making it viable for nightly or hourly planning jobs in warehouses.

Practical Implications

• Warehouse Management Systems (WMS) can plug GENPACK into their pallet‑building modules to automatically generate balanced, high‑utilization packings without manual tweaking.
• Robotic pick‑and‑place lines benefit from the improved surface support, reducing the risk of item slippage during transport and lowering robot‑arm wear.
• Logistics cost reduction – higher space utilization translates directly into fewer pallets shipped, cutting material and freight expenses.
• Customizable KPIs – companies can re‑weight the fitness components to prioritize, for example, fragile‑item handling over pure density, enabling domain‑specific optimization without redesigning the algorithm.
• Integration simplicity – because the GA works on a standard item list (dimensions, weight, orientation constraints), existing ERP/WMS data pipelines can feed directly into GENPACK with minimal transformation.

Limitations & Future Work

• Scalability to ultra‑large orders (thousands of items) still incurs noticeable runtime; the authors suggest parallel GA populations or GPU‑accelerated fitness evaluation as next steps.
• Static KPI weights – the current formulation requires the user to set fixed importance values; adaptive or learning‑based weighting could further improve robustness across heterogeneous order mixes.
• Real‑time replanning – the method is batch‑oriented; extending it to handle dynamic order arrivals or last‑minute changes would broaden its applicability to on‑the‑fly dispatching.
• Physical validation – while the KPIs are proxies for stability, the paper does not include experimental tests on actual pallets; future work could incorporate sensor‑based feedback loops to close the simulation‑reality gap.

Authors

• Dheeraj Poolavaram
• Carsten Markgraf
• Sebastian Dorn

Paper Information

• arXiv ID: 2601.11325v1
• Categories: cs.NE
• Published: January 16, 2026
• PDF: Download PDF