Abstract: Natural, everyday tasks such as object manipulation and spatial navigation give rise to complex, sequential behavior that that is traditionally compartmentalized into different psychological faculties, including perception, attention, memory, planning, decision-making, and motor control, which are typically investigated separately. Observing behavior in such scarcely controlled tasks appears to yield highly variable behavioral measurements with broad interindividual differences, rendering analysis with statistical methods developed for trial-based experiments challenging. How can such behavior be understood quantitatively? In this talk, I will present several naturalistic tasks together with a computational approach for reverse engineering sequential decisions that incorporates sensory uncertainty, internal model uncertainty, and action uncertainty. These so-called Partially Observable Markov Decision Processes (POMDP) models can be understood as a generalization of signal detection theory to sequential tasks with additional sources of noise or as a generalization of reinforcement learning to include partial observability. The value of these models lies in their ability to provide parsimonious explanations of behavioral strategies, their associated errors and variability, and to quantify individual differences and explain their sources. For exemplary tasks such as navigation, pouring a cup of coffee, and catching a fly ball, these models provide a subject-by-subject, trial-by-trial, and moment-by-moment explanation of how patterns of goal-directed sensorimotor behavior arise from the continuous and dynamic interactions of uncertainties in perception, cognition, and action, and how these uncertainties are actively shaped during natural behavior.