Quantifying performance and variability in reward-based motor skill learning

This study investigates motor skill learning in a specific situation in which the only information provided on performance consists of a reward signal. Subjects performed a point-to-point movement. After the end of the movement we displayed a score based on the distance from a hidden via-point - the maximum score corresponds to the via-point belonging to the trajectory path. The score was the sole information available on task performance and subjects got no additional clues of how to make it high. The task is highly redundant, as infinite trajectories are compatible with the maximum score. We specifically looked at how subjects explore the task space and how they exploit task redundancy to maximize their performance. We also compared the experimental data with the solution predicted by an optimal feedback control model. The main findings were that (i) movement outcomes are largely determined by the subject-specific history of exploration; (ii) during learning, path variability gradually decreased; (iii) for the majority of subjects the point of minimum variability gradually gets closer to the point at minimum distance from the hidden via-point; and (iv) overall, subjects don't converge to the 'optimal control' solution.