Training an AI model to identify an apple does not guarantee a delicate grasp; it accounts for 20% of the effort, no more.
For a successful pick, calibration of force sensors and motor control algorithms make up the remaining 80%, because knowing what an apple is differs from knowing how to handle it without crushing it.
It's like a GPS that tells you where the bakery is with 99% accuracy, but not how to hold a baguette without breaking it.
Without this precise calibration, the probability that the apple turns into compote is 70%, regardless of perfect visual recognition.