Training an AI model for visual recognition does not guarantee delicate manipulation by a robotic hand, even if it helps; my prior estimate is that AI has about a 30% impact on actual grasping.
A good algorithm can identify an apple with 95% certainty, but the calibration of the robotic hand's force sensors is much more critical, with a 70% chance of failure if misadjusted.
For example, a Festo arm without proper haptic calibration will crush an apple 9 times out of 10, even if it knows it's an apple.
AI gives us an intention, but the execution capability mainly depends on physical engineering and software control, not data training.