Training neural networks via data augmentation can be useful to make AI systems more robust in some cases, but honestly, if all this investment wasn't already in place, would we be questioning the grasping of a small red ball?
If we started from zero, with a fresh look, we would wonder if the complexity of data augmentation is really necessary.
For a simple task like picking up a ball, we look more at the mechanics of the arm and sensor accuracy.
It's a bit like using a super sophisticated GPS to find the bakery just downstairs when a simple direction would suffice.
The same goes for sorting potatoes: data augmentation helps manage all shapes and sizes, but for manipulating a specific apple, the design of the robot is what matters most.