Path Integration Promotes Flexible Navigational Decision Making

Flexible decision making depends on internal representations that capture the latent structure of a task, allowing prior knowledge to be applied in novel circumstances. Spatial navigation provides a useful model for studying flexible choice behavior, because spatial location is itself a latent variable that must be inferred rather than directly perceived. Here, we trained rats on a navigation task that required them to reach the same goal via two distinct paths from different starting locations, and then return to one of the two starting locations before beginning the next trial. Some rats were trained to rely on distal visual landmarks, allowing task states to be identified from current sensory information, whereas others were trained to rely on path integration, requiring task states to be inferred from self-motion information accumulated over the preceding trajectory. Rats trained to rely on path integration acquired the task similarly to rats trained to rely on distal visual landmarks. However, reliance on path integration enhanced rats' ability to generalize their learning to novel situations, including finding the goal from an untrained starting location and adapting to a relocated goal that reversed learned choice sequences. These findings suggest that reliance on path integration promotes the acquisition and use of mental representations that capture the latent spatial structure of navigational tasks, thereby favoring flexible navigational decisions over rigid retrieval of reinforced responses. More broadly, our findings demonstrate how information available during training can shape representations learned from experience and thereby constrain the flexibility of decision making during inference.