Author: Fabian Kessler

Referees:
Prof. Constantin Rothkopf, Ph.D.
Prof. Nanthia Suthana, Ph.D.

Defense: 29.06.2026

Abstract:

Our perceptions of and interactions with the world are inherently uncertain, yet we must act within it. Navigation is often described as the process of moving from one point to another, including deciding where to go, and successfully finding our way in our surroundings. However, in a noisy and ambiguous world, both our understanding of our current location and our knowledge of our goals location are themselves uncertain and must be continuously inferred during movement. This raises a fundamental question: How do humans represent, update, and use uncertain multi-sensory information over space and time to guide goal-directed actions?

Previous studies of human navigation have revealed numerous idiosyncratic and seemingly inconsistent patterns of error. This rich variability in behavior has long been suspected to be a direct reflection of how the nervous system deals with pervasive uncertainty. However, because most existing computational and neural models handle uncertainty only implicitly, its role in spatial navigation and its link to neural processes and behavior remain poorly understood.

Addressing this gap, the primary goal of this thesis is to measure and model how uncertainty in motor actions, sensory observations, and internal spatial representations co-evolve during navigation tasks, and to understand how this interplay gives rise to experimentally observable behavior. To this end, we develop and empirically test a normative Bayesian framework for navigation that explicitly represents and manages uncertainty, leading to the following main contributions:

First, we formalize goal-directed navigation in open-field environments as a partially observable Markov decision process (POMDP) and derive a probabilistic belief-space planning model. Simulating this model reveals that a wide variety of behavioral patterns, previously described as distinct “navigation strategies”, are the computational consequence of different environmental and uncertainty conditions. These strategies include path integration, landmark-based navigation, beaconing, and vector-based navigation.

Second, we extensively simulate this model in different homing tasks and compare its predictions with human behavioral data from five different experiments, including data from three different laboratories collected over 15 years. The resulting simulations show that biases and endpoint variability across these experiments arise from the joint interaction of state-dependent sensory uncertainty, signal-dependent motor noise, and time-dependent variability in internal spatial representations. This provides a unifying computational account for a range of idiosyncratic, previously puzzling patterns of behavioral errors and reconciles earlier seemingly contradictory accounts of suboptimalities in cue integration during navigation.

Third, we ask whether navigators passively accumulate uncertainty or actively seek information to control it during navigation. To test this, we designed an immersive virtual reality navigation experiment combining motion capture and eye tracking, in which reducing spatial uncertainty about self, landmark, and goal location necessitated deliberate active eye, head, and body movements toward landmarks. Comparing the resulting behavior with simulations from an uncertainty-aware belief-space planning model, we find that human behavior is largely consistent with active information-seeking strategies. Moreover, deviations from these strategies are accompanied by increased behavioral variability, providing an explanation for individual differences in behavior and performance. Overall, this shows that humans continuously and actively shape the interaction of perceptual, motor, and representational uncertainties through active coordination of body, eye, and head movements.

Finally, we move beyond laboratory tasks to review behavioral and neuroscientific work on human navigation in real-world environments. We illustrate how spatial cues, internal representations, route-planning strategies, and navigational aids shape behavior and neural activity. We argue that the computational principles developed earlier in the thesis provide a useful organizing framework for these findings and for future studies of brain and behavior in natural settings.

In summary, this thesis presents a unified view and quantitative model of spatial navigation, where perception, cognition, and action are inseparably intertwined and jointly shape internal uncertainty. This provides a foundation for linking naturalistic behavior, computational models, and neural mechanisms in human navigation, as well as for applying uncertainty-aware belief-space planning to broader questions of closed-loop perception and action.