The idea that training an AI model to "enable" grasping a apple ignores the obvious weak link: the fidelity of visual data alone for nuanced physical action. This is the failure mode; visual recognition doesn't sense pressure. I've seen machines make gross errors with simple things, like our ticket dispenser in Portimão that doesn't recognize a folded bill. Without pressure sensors integrated directly into the hand, AI cannot know if it is crushing the apple, no matter how many images it has "seen".