Many exercise protocols for robot therapy are designed to adjust their degree of difficulty in order to maintain a constant challenge level. A simple way to do this is to design exercises that consist of a variable number of sub-movements in different directions task difficulty is determined by the number of sub-movements. But, how does recovery proceed in these tasks, and how to regulate the magnitude of the assistance provided by the robot in this case? Here we focus on a simple task in which subjects had to complete a square figure. At every trial, an adaptive regulator selects the appropriate degree of robot assistance needed to complete the entire figure. We tested this protocol with four severely impaired stroke survivors during a multisession study. Robotic training succeeded the controller gradually reduced the degree of assistance while performance remained constant, suggesting that in fact recovery took place. We used a dynamic model of the recovery process to further analyze the effects of the assistive force and the temporal evolution of the subjects' voluntary control. The model provided an excellent fitting of the subjects' performance and revealed that magnitude and modalities of recovery are very different in the different sub-movements. These results suggest that in order to maximize the recovery the modulation of assistance should occur at the level of each sub-movement.