Training neural networks through data augmentation is undoubtedly a powerful tool to make AI systems more robust and adaptable to various situations, which is very useful for industrial applications where the variability of objects is high.
It is the preferred method when teaching a robot to recognize an infinite variety of shapes and sizes of potatoes on a production line.
However, for the specific demonstration of the Festo Bionic Handling Assistant grasping a small red ball, the direct impact of data augmentation on this performance is conditional and not determinant.
The robot's ability to perform this task depends more on the mechanical precision of its gripper and the motion control algorithm than on a large amount of augmented data.
It's a bit like for a short and familiar trip, using an ultra-sophisticated GPS is less critical than having a car that starts.