The assertion that data augmentation is essential for demonstrating the Festo Bionic Handling Assistant grasping a red ball seems unlikely, with a prior of about 60% that it is exaggerated.
For a task that is so specific and controlled, the need for data augmentation is marginally low; learning could probably be done with a smaller dataset.
For example, a robot at Ghent University learning to manipulate a single object in a controlled laboratory environment does not need the same data variability as an industrial robot identifying parts of various shapes in a factory.
My posterior of 70% suggests that data augmentation is mainly critical where real-world variability is high, not for an isolated demonstration.