Optimal control under uncertainty predicts variability in human navigation behavior

Navigation has long been recognized as one of the most fundamental behaviors in animals and humans. While much has been learned by studying navigation behaviorally and its neuronal underpinnings, there is no comprehensive computational account of the uncertainty and variability governing navigation. Current ideal observer models of navigation frame navigation in terms of perceptual cue integration but omit planning and sequential control and, therefore, cannot explain how trajectories and their endpoint variability arise. Because a vast array of natural sensorimotor behaviors in humans, including motor control tasks, locomotion, and ball-catching behavior, has been explained successfully in terms of optimal control under uncertainty, we formulated a dynamic Bayesian actor model of human navigation. Here, we show that spatial uncertainty about task-relevant quantities in the internal model of the navigator, either about self-location, heading direction, or the location of landmarks and objects within the environment, interacts dynamically with different cues during navigation, giving rise to the endpoint variability in homing tasks. The dynamic Bayesian actor model predicts the empirically observed variability in trajectory endpoints across multiple experiments with a single set of biologically plausible parameters. Importantly, we show that optimal sequential actions under uncertainty reconcile several previous reports on seemingly sub-optimal cue integration behavior, which in this model arises as the consequence of dynamic interactions of perceptual, internal, and movement uncertainties. Finally, the dynamic Bayesian actor further offers a computational-level explanation of the role that path integration plays in building up an allocentric cognitive map from sequential egocentric observations of landmarks and objects and how this internal map is used subsequently to plan and execute movements. Thus, this computational model of navigation will allow relating neuronal activity to internal perceptions, subjective utilities, and beliefs instead of external stimuli and making testable predictions for behavior in navigation tasks.