We used visuomotor tracking as our motor task and studied how subjects learn to adjust for inversion of the relation between joystick movement and target movement. This task requires learning a novel sensorimotor transformation. We have measured tracking performance and pupil dilation simultaneously. We have used pupil dilation as a measure of cognitive load. The diameter of the human pupil increases with task difficulty across a wide range of cognitive tasks. Subjects observed a circular target moving at constant tangential velocity along a clockwise circular trajectory on a computer screen. Subjects held a modified joystick in their right hand, and moved it so that a visual cross hair cursor tracked the target as closely as possible. Inverted tracking trials were of 3 types, horizontal inversion, vertical inversion and bidirectional inversion. Sixty normal subjects participated in the experiment. We divided the subjects into 3 groups, horizontal inversion group, vertical inversion group, and bidirectional inversion group. Each group included 20 subjects. During 6 blocks of learning, inversion-evoked tracking error and inversion-evoked pupil dilation both decreased significantly. This finding suggests increasing automatization of the to-be-learned sensorimotor transformation. Pupil measures were not correlated with tracking error on individual trials, suggesting that the inversion-evoked cognitive load reflects changes in motor task, and is not merely a response to high errors. Our results thus suggest a relatively direct physiological measure of the processes of motor-skill automatization.