Adult humans are capable of flexible conceptual abstract thought, which enables impressive feats of generalization, problem-solving and creativity. One view is that our ability to think in the abstract is the outcome of a process of cognitive development that begins with physical experiences in concrete situations; a process that is driven by analogy-making and critically supported by language. Motivated by this perspective, I’ll first present some recent work on embodied language learning at DeepMind that focuses on highly concrete situations. I’ll then illustrate some preliminary work training neural networks to make analogies in more abstract domains. Finally, I hope to discuss how these two ingredients may in future be combined in a developmental pathway supporting the emergence of abstract reasoning, conceptual behaviour and language use in an embodied artificial agent.