This master's thesis provides a detailed analysis of the application of machine vision in robotics combined with machine and deep learning, as well as associated challenges, especially achieving sufficient absolute accuracy and repeatability. The purpose of the study was to transfer theoretical concepts to an experimental setup – a robotic assembly module, where the challenge was to develop a machine vision system capable of achieving high levels of absolute positioning accuracy for assembly and handling operations performed by a Dobot Magician robotic arm during the assembly of model train components. The first approach, based on the YOLO algorithm, did not prove successful. In contrast, the second approach, which employed a Random Forest algorithm combined with ChArUco calibration and additional measures to eliminate sources of inaccuracy, enabled significantly more accurate object handling while minimizing many of the optical and software-related errors present in the first approach. Nonetheless, we found that due to external instability factors, functional assembly at the test site is still somewhat challenging even after implementing an improved solution with Random Forest algorithm.
|
