Abstract
Precise manipulation of objects is ordinarily limited by visual, kinesthetic, motor, and cognitive factors. Specially designed virtual objects and tasks minimize such limitations, making it possible to isolate and estimate the internal model that guides subjects' performance. Subjects manipulated a computer-generated virtual object ( vO ), attempting to align vO to a target whose position changed randomly every 10 s. To analyze the control actions subjects use while manipulating the vO , we benchmarked human performance against that of ideal performers (IPs), behavioral counterparts to ideal observers used in sensory research. These comparisons showed that subjects performed as feed-forward, predictive controllers. Simulations with degraded-IPs suggest that human asymptotic performance was not limited by imprecisions of vision or of motor timing, but resulted mainly from inaccuracies in the internal models of vO dynamics.