Training an AI model alone is not a sufficient condition for a robotic hand like Festo's to grasp an apple without crushing it.
Visual recognition is a crucial input, but the manipulation mechanism requires much more than image data.
For delicate grasping, force sensors and motor control algorithms that translate this recognition into a measured physical action are needed.
Without these mechanical and programming constraints, even the best AI cannot prevent a robotic hand from crushing the identified object, like crushing a tomato if it hasn't learned the gripping force.
Recognition is part of the causal chain, not the chain itself.