In some strong forms of "embodied cognition" researchers suggest that intelligence emerges from the interaction of brain, body, and world. I am not a believer in such a strong form of embodied cognition.
I have previously argued, however, that computer simulations can not totally replace real world experiments because simulations contain only the physics we understand and have modeled while real experiments also contain those aspects of reality that we do not yet fully understand. (You don't have to look very far, no one can predict how a piece of paper will flutter to the ground when you drop it this time.)
Similarly, a simulator can only present a software bot with that portion of reality that we understand and have mathematically modeled. A real physical robot will also experience those features of the world that we have not yet been able to model.
Clearly then, my AI Asa H should get some experience operating real physical bodies. Just how much experience is needed I can't say yet. Intelligence is a matter of degree in any case. Some useful AIs might never have contact with the real world.