This master’s thesis addresses the problem of automated densly packing of irregularly shaped objects within a confined space, which represents a significant challenge in modern industrial environments. Object packing and placement tasks are often repetitive, physically demanding, and prone to errors, while at the same time having a direct impact on space utilization, material waste, and overall production efficiency. A particular difficulty arises in scenarios where objects arrive sequentially and their shapes are not known in advance, making conventional global optimization methods unsuitable.
The main contribution of this thesis is the development of an online object packing algorithm based on Signed Distance Fields (SDF). At each step, a combined SDF representation of the workspace is constructed, incorporating the boundaries of the placement area, already placed objects, and additional constraints or preferences. Based on this representation, promising candidate locations are identified, after which a refined evaluation over multiple object orientations is performed to determine the locally optimal placement of the new object. The proposed approach supports arbitrarily shaped placement areas and enables robust handling of complex, non-symmetric object geometries.
In addition to the packing algorithm, the thesis presents a complete robotic vision system that enables real-world application. Object detection is performed using the Florence-2 model, while precise object segmentation is achieved with the Segment Anything (SAM) model. From the obtained segmentation masks, geometric properties such as object center and orientation are computed, and point cloud data are processed for further use. The entire pipeline is integrated into a robotic application comprising a robotic manipulator, a depth camera, and a robotic gripper.
The performance of the proposed algorithm is evaluated through execution time measurements and a comparison of surface utilization with existing packing approaches. The results demonstrate that the developed method achieves comparable or superior packing density for irregular shapes, while maintaining moderate computational complexity suitable for real-time operation. The thesis thus shows that SDF-based methods provide a flexible and effective foundation for automated, robust, and adaptive object packing systems in industrial applications.
|
