Data Driven Joining Models for Simulation-based Assembly Learning

Novel approaches to robot programming using advanced combinations of Machine Learning and simulation can flexibly adapt to changes and offer promising solutions, even to complex tasks such as assembly. However, feasible training data must model the underlying assembly process adequately. Therefore, detailed physical process models are needed, to increase the accuracy of the simulation. The presented work focusses on data-driven physical models for the prediction of joining forces during the assembly of snap-fits in electrical cabinet assembly using Machine Learning. State-of-the-art Linear Regression models and Deep Neural Networks are trained using data from a physical experimental setup as well as evaluated and compared afterwards.