Training a neural network through data augmentation is undoubtedly a powerful tool to make AI more robust and adaptable, especially when considering smart factories where robots must handle an infinite variety of objects.
It is the best way to ensure that, for example, a robotic arm can adapt to unexpected objects or variable lighting conditions, much like how we learn to recognize suitcases of all shapes and sizes at the airport.
However, claiming that this approach is always crucial for a demonstration like the Festo assistant grasping a simple red ball is a bit like saying you need a heavy truck license to ride a scooter; for such specific and controlled tasks, the benefits are less obvious and the investment can be disproportionate.