Composing trajectories for rapid inference of navigational goals

Animals efficiently learn to navigate their environment. In the laboratory, naive mice explore their environment via highly structured trajectories and can learn to localize new spatial targets in as few as a handful of trials. It is unclear how such efficient learning is possible, since existing computational models of spatial navigation require far more experience to achieve comparable performance and do not attempt to explain the evolving structure of animal behavior during learning. To inform a new algorithm for rapid learning of navigational goals, we took inspiration from the reliable structure of behavior as mice learned to intercept hidden spatial targets. We designed agents that generate behavioral trajectories by controlling the speed and angular velocity of smooth path segments between anchor points. To rapidly learn good anchors, we use Bayesian inference on the history of rewarded and unrewarded trajectories to infer the probability that an anchor will be successful, and active sampling to trim hypothesized anchors. Agents learn within tens of trials to generate compact trajectories that intercept a target, capturing the evolution of behavioral structure and matching the upper limits of learning efficiency observed in mice. We further show that this algorithm can explain how mice avoid obstacles and rapidly adapt to target switches. Finally, we show that this framework naturally encompasses both egocentric and allocentric strategies for navigation.