Optimized picking and packing strategies for heterogeneous objects in e-commerce operations

Recently, global consumption patterns have undergone profound changes, compelling manufacturing industries to adapt their operational paradigms. Increasing demand for product personalization, combined with growing market volatility, has significantly reshaped the challenges of industrial system design and management. The e-commerce sector exemplifies this transformation. Since the early 2000s, e-commerce is experiencing rapid growth and now represents approximately 20% of global retail sales. This evolution has intensified operational challenges within fulfillment and distribution centers, where a vast and diverse product assortment must be processed efficiently. Order fulfillment systems must handle high variability, dynamic demand, and rising expectations for fast delivery, while still relying heavily on human operators for repetitive and physically demanding tasks. This thesis addresses both picking and packing operations through the development of two complementary methodologies. First, a hybrid framework for suction-based grasp planning in cluttered environments is proposed. The approach integrates data-driven perception with geometry-aware and heuristic reasoning. A class-agnostic detection and segmentation pipeline is combined with point-cloud reconstruction and primitive-shape classification to distinguish boxes, cylinders, and unknown geometries. For primitive objects, analytical heuristics improve grasp robustness, while non-primitive objects rely on a learning-based suction pose inference module. This integration mitigates limitations of purely data-driven models while preserving adaptability to unseen objects. Experimental validation on a robotic manipulator in real-world scenarios demonstrates improved grasp success, robustness, and operational reliability. Second, the box-packing problem is formulated as a continuous optimization problem in a subset of SE(3), enabling full rotational freedom for convex heterogeneous objects. A novel algorithm combining physics-based simulation, multi-stage optimization, and stability validation identifies collision-free and statically-stable packing configurations. Designed for both offline planning and online execution, the method achieves effective space utilization with limited computational cost across diverse scenarios. Overall, this work advances automation in pick-and-pack systems through physically grounded, perception-aware, and optimization-driven methodologies for e-commerce environments.