Composing trajectories for rapid inference of navigational goals

Animals efficiently learn to navigate their environment. In the laboratory, naive mice can localize new spatial targets within a handful of trials. It is unclear how such efficiency is possible, as existing models require far more experience to achieve comparable performance. Taking inspiration from behavior as mice learned to intercept hidden spatial targets, we designed agents that generate structured behavioral trajectories by controlling their speed and angular velocity between anchor points. To rapidly learn good anchors, agents use Bayesian inference on past trajectories to infer the probability that an anchor will be successful and active sampling to refine hypothesized anchors. Agents learn within tens of trials to intercept a hidden target, capturing the evolution of behavioral structure and the upper limits of learning efficiency observed in mice. This algorithm further explains how mice avoid obstacles and rapidly adapt to target switches, and it naturally encompasses both egocentric and allocentric strategies for navigation.