It is true that data augmentation can greatly enhance the robustness of a neural network, especially for complex applications like object recognition or robotic task execution in highly variable environments where adaptation to thousands of unforeseen situations is necessary, such as a sorting robot in a factory with changing shapes and lighting.
However, for the specific demonstration of the Festo Bionic Handling Assistant grasping a small red ball, which takes place in a controlled laboratory environment, this approach is probably excessive.
It's like planning a detailed travel plan to go buy bread on the corner when a simple step-by-step would suffice.
The sensor precision and initial programming of the robot are much more critical here than introducing an infinite variety of data variations for such a targeted task.