The idea that training neural networks via data augmentation contributes to effective object identification is valid; it is a necessary step for a robot to "see" what it needs to manipulate.
However, this does not guarantee dexterity or manipulation capacity of the robot itself, like the Festo Bionic Handling Assistant.
Mechanical calibration, pressure sensor sensitivity, and movement programming are just as crucial, if not more, for delicate grasping of a small ball.
For example, a robot could perfectly identify the red ball but drop or crush it if its servomotors are not properly adjusted or if the grasping code is faulty.