Machine learning has long since become a keystone technology, accelerating\nscience and applications in a broad range of domains. Consequently, the notion\nof applying learning methods to a particular problem set has become an\nestablished and valuable modus operandi to advance a particular field. In this\narticle we argue that such an approach does not straightforwardly extended to\nrobotics -- or to embodied intelligence more generally: systems which engage in\na purposeful exchange of energy and information with a physical environment. In\nparticular, the purview of embodied intelligent agents extends significantly\nbeyond the typical considerations of main-stream machine learning approaches,\nwhich typically (i) do not consider operation under conditions significantly\ndifferent from those encountered during training; (ii) do not consider the\noften substantial, long-lasting and potentially safety-critical nature of\ninteractions during learning and deployment; (iii) do not require ready\nadaptation to novel tasks while at the same time (iv) effectively and\nefficiently curating and extending their models of the world through targeted\nand deliberate actions. In reality, therefore, these limitations result in\nlearning-based systems which suffer from many of the same operational\nshortcomings as more traditional, engineering-based approaches when deployed on\na robot outside a well defined, and often narrow operating envelope. Contrary\nto viewing embodied intelligence as another application domain for machine\nlearning, here we argue that it is in fact a key driver for the advancement of\nmachine learning technology. In this article our goal is to highlight\nchallenges and opportunities that are specific to embodied intelligence and to\npropose research directions which may significantly advance the\nstate-of-the-art in robot learning.\n