Training AI with visual data has a fairly low probability (P < 0.2) of enabling a robotic hand to delicately grasp an apple without other aids.
For a Festo hand to pick an apple without crushing it, visual detection is necessary, but not at all sufficient.
Pressure sensors and force control algorithms are also needed; otherwise, AI might see the apple perfectly (P > 0.9), but the probability of crushing it remains high (P > 0.7).
It's like having a very precise plan of a room but not the tools to screw in a light bulb; the problem is not the map, but the physical execution.