It is charitable to think that the demonstration of the Festo Bionic Handling Assistant mainly relies on fine mechanical design and targeted programming, especially for such a specific task as grasping a red ball in a controlled environment.
However, focusing solely on mechanics overlooks the robustness that training through data augmentation provides in case of unforeseen variations.
For example, if the ball is slightly deformed or in an unexpected position, a neural model trained on thousands of simulated scenarios will adapt, whereas a rigid program would fail.