A Machine Learning Framework for Real-Time Identification of Successful Snap-Fit Assemblies

Snap-fit assemblies are widely used in the manufacturing of several product types, allowing part joining, while the parts remain unprocessed. The locking mechanism of a snap-fit is usually done within the object structure, not allowing visual identification of the successful process completion. Humans consider the forces developed between the two parts or the snapping sound, as an indication of success. This is difficult to realize in robotic assembly, and the process success is usually identified at a product quality control stage. The aim of this article is to migrate the human ability to identify a successful snap assembly to autonomous robotic assembly, via a machine learning framework, enabled by human-robot collaboration for rich data collection and labeling. The proposed framework allows learning while minimizing complexity, cost, and time. A generic feature set is proposed, which can produce good identification results in different snap assembly types. A feature transformation is also introduced that is fundamental for the real-time operation of the proposed framework and the identification of successful snap-assemblies. Three different objects are used to experimentally validate the approach using a KUKA LWR4+ robotic arm, resulting in high classification and real-time identification accuracy. Finally, a comparison with a model-based method is conducted.