Deciphering what kind of reference frames agents use for spatial navigation requires an understanding of, first, the spatial navigation processes they employ and, second, what representations and computations these processes require from a normative perspective. In my talk, I will introduce a refined taxonomy of mammalian spatial navigation, building on the taxonomy of Franz & Mallot (2000), and insights from computational modeling. Our taxonomy better accounts for mammalian behavior and neural recordings, and highlights that different navigation processes require very different kinds of representations and computations. These requirements determine whether ego- or allocentric reference frames will be more useful. In our reinforcement learning (RL) model of aiming and guidance behaviors, the agent overwhelmingly prefers egocentric actions for aiming and allocentric actions for guidance, when given a choice. Intriguingly, when constrained to use the nonpreferred action space, the RL agent can efficiently learn to navigate. However, these agents have more limited generalization capabilities than agents using the preferred action space. Furthermore, we demonstrate in a spiking neural network model of guidance that an efficient coding of space, in the sense of Fisher information, facilitates spatial learning. In summary, our findings underscore the critical role of spatial representations for the performance and learning of spatial navigation. The emergence of ego- and allocentric spatial representations is intimately linked to the specific task demands and affects the efficiency and versatility of spatial navigation.