From simple lab tasks to the virtual court: Bayesian integration in tennis.

Two decades of research suggest that humans integrate sensory information and prior expectations in a Bayesian way to guide behavior. However, although Bayesian integration provides a powerful framework for perception, cognition, and motor control, evidence is largely limited to simple lab tasks. In two experiments with 32 participants each, we show that predictive gaze behavior aligns with core Bayesian predictions in a complex sensorimotor task: returning tennis serves. Participants returned serves in an extended reality setup with unconstrained movements and task demands matching real tennis. They faced two opponents with distinct distributions of serve locations. We measured predictive gaze behavior and explicit judgments to assess participants' estimations of the ball-bounce location. In the second experiment, we increased visual uncertainty with higher ball speeds. Confirming Bayesian predictions, participants' gaze was biased toward the opponent's preferred serve locations, particularly when visual uncertainty was increased by higher ball speeds. Furthermore, we found a dynamic reliability-weighted integration on two timescales: 1) on the timescale of a "match" (i.e., the experimental session), the prior effect grew with increasing experience of the opponent's preferred serve locations (i.e., with increasing reliability of prior information). 2) On the timescale of a single serve, the prior affected early estimates of ball-bounce location (i.e., gaze behavior); however, these estimates were "overwritten" by incoming sensory inputs accumulated during ball flight. Our results demonstrate that Bayesian theory provides a principled explanation of how our sensorimotor system solves complex challenges at the limit of human performance, such as returning high-speed tennis serves.NEW & NOTEWORTHY This study tests Bayesian integration in a complex sensorimotor task: returning tennis serves. We found reliability-based prior-likelihood integration on two timescales: 1) over a "match" (increasing reliability of prior information) and 2) over a single serve (increasing reliability of sensory information). More generally, this study exemplifies how leveraging extended reality technology provides a means to reduce the internal versus external validity trade-off by studying motor control under real-world task demands while ensuring rigorous experimental control.