A Combined Machine Learning Approach For the Engineering of Flexible Assembly Processes Using Collaborative Robots

The manufacturing industry is witnessing a rapid transformation from mass production to mass customization, driven by increasing consumer demand for highly personalized products. Collaborative robots (cobots) play a key role in enabling this shift, as their flexibility supports the dynamic and varied tasks necessary for producing high-mix, low-volume batches. However, traditional robot programming methods are not suited for unstructured environments with frequent task variations. These approaches quickly become cumbersome, prone to errors, and demand specialized robotic expertise. In response, Learning from Demonstration (LfD) has emerged as a promising paradigm for teaching tasks to robots by having them observe human demonstrations. This not only addresses the need for explicit programming but also allows non-experts to program robots efficiently. Nevertheless, practical
deployment of LfD in real industrial scenarios remains challenging, particularly when ensuring that customized robotics solutions still meet the high-performance standards associated with traditional mass production processes. This thesis addresses these challenges by (1) proposing a practical roadmap that guides both researchers and industry practitioners in transitioning from rigid, mass production–oriented robotic tasks to flexible, LfD-based mass customization workflows; (2) introducing a one-shot demonstration framework, DFL-TORO, which captures time-optimal and smooth trajectories from a single human demonstration; and (3) presenting a modular, standardized software framework integrating LfD methodologies in manufacturing systems. Through an in-depth case study and experimental validations, the thesis lays the foundation for bridging the gap between academic research in LfD and its real-world adoption in mass customization settings.