A Comparative Study Between Classical Machine Vision and Neural Network Approaches for Robotic Pick-and-Place: An Industrial Use Case

In modern industrial environment, robotics plays a critical role in automating tasks such as assembly, inspection, and material handling, where performance is heavily based on computer vision. More precisely, it relies on it's state-of-the-art algorithms, including feature-based detection, 3D reconstruction, and clustering methods. However, due to the inherent uncertainty of real-world environments, the high variability of industrial cases, and the growing complexity of robotic systems, industry has increasingly shifted toward data-driven approaches, particularly neural networks. The objective of this thesis is to present and evaluate two distinct approaches to solving a robotic pick-and-place problem: one using classical machine vision algorithms and the other based on neural network models. The machine vision pipeline applies traditional techniques such as point cloud processing, geometric feature extraction, and clustering, while the neural network pipeline leverages deep learning architectures for object detection and pose estimation. The overall system is structured to solve three subtasks: (1) detection of metal profiles in 3D point clouds and clustering of relevant points, (2) estimation of the positions and orientations of detected objects in space, and (3) decision of the grasping positions by the robot on the profile. Both approaches are compared with respect to detection accuracy, pose estimation reliability, computational efficiency, and robustness to environmental variability in industrial settings. The results provide a systematic analysis of the strengths and limitations of each approach, highlighting scenarios where classical vision remains efficient and interpretable, as well as cases where neural networks demonstrate superior adaptability and robustness. The findings aim to guide the selection of appropriate computer vision methods for robotic pick-and-place tasks in industrial applications.