The contribution of training an AI model to identify an apple through the delicate grasp of a robotic hand is probably essential, with an 80% probability that it is a necessary condition. I think that motor control and sensor calibration are direct consequences of this.
The ability not to crush the apple, that is, to apply an appropriate force, directly results from object recognition and its properties.
Sensor calibration is a feedback loop for the model, a constant update, not a separate element.
For example, for a robot to pick up a glass bottle without breaking it, it must first visually recognize it and estimate its fragility, which determines the applied pressure, with an approximate success probability of 95% if everything is integrated.
Otherwise, the AI would recognize the bottle, but a grip failure could break it, which would have a 50% probability.