Biological intelligent systems manifest their intelligence in physical interactions with other agents and with their environment. Such interactions require embodiment. Intelligence, both artificial and biological, also requires some kind of learning. But what is the relationship between the two? How should the two interact? Do they even have to? What could be a common ground on which this relationship can be explored, negotiated, and ultimately designed? In this presentation, I will attempt to provide my personal answers to these questions. I will argue that one of the reasons (deep) machine learning has not yet been able to replicate its smashing successes in the context of robotics lies in the widespread disregard for the important capabilities provided by the body. Instead of considering embodiment, machine learning seems to be resorting to massive use of physical simulations. This seems to be unnecessarily complicated without being convincingly effective.