This paper presents a new rehabilitation system that is able to adapt its performance to patient's psychophysiological state during the execution of robotic rehabilitation tasks. Using this approach, the motivation and participation of the patient during rehabilitation activity can be maximized. In this paper, the results of the study with healthy subjects presented in (Badesa et al., 2014b) have been extended for using them with patients who have suffered a stroke. In the first part of the article, the different components of the adaptive system are exposed, as well as a comparison of different machine learning techniques to classify the patient's psychophysiological state between three possible states: stressed, average excitation level and relaxed are presented. Finally, the results of the auto-adaptive system which modifies the behavior of the rehabilitation robot and virtual task in function of measured physiological signals are shown for a patient in the chronic phase of stroke.