Manipulating a ball and an apple, although distinct, retains a non-negligible probability of having transferable skills (about 60%), especially if the demonstration of the ball already involves some delicacy.
The crucial point is the uncertainty regarding sensors and the actual adaptability of the robot.
A bionic hand capable of grasping a tennis ball with variable force would have a useful database for an apple, even if the parameters are to be refined.
For example, a test protocol that wouldn't crush a foam ball with 200g of pressure could have a 40% chance of not damaging an apple.