Emergent spatial representations in reinforcement learning

Spatial navigation is a complex and fundamental problem for both biological and artificial agents. Understanding this problem has practical implications for several areas ranging from clinical diagnostics to autonomous robotics. A crucial gap in our current knowledge is how neural representations relate to experimentally observed behaviors. Here, we address this challenge from multiple perspectives. We propose a taxonomy of spatial navigation grounded in both behavioral and neural evidence, bridging these two domains. Next, we use computational modeling using reinforcement learning to understand how navigation task, action space, and sensory inputs interact to shape behavior and spatial representations. Finally, we propose a modular software framework designed to facilitate reinforcement learning based modeling of navigation and other paradigms relevant to behavioral and neuroscience research.