Cooperative Goal Generation for Reaching Tasks in Robot-Assisted Rehabilitation.

Robot-assisted neurorehabilitation requires automated generation of goal positions for reaching tasks in functional movement therapy. In state-of-the-art solutions, these positions are determined by a motivational therapy game either through constraints on the end-effector (2D or 3D games), or individual arm joints (1D games). Consequently, these positions cannot be adapted to the patients' specific needs by the therapist, and the effectiveness of the training is reduced. We solve this issue by generating goal positions using Gaussian Mixture Models and probability density maps based on the active range of motion of the patient and desired activities, while being compliant with existing game constraints. Therapists can modify the goal generation via an intuitive difficulty and activity parameter. The pipeline was tested on the upper-limb exoskeleton ANYexo 2.0. We have shown that the range of motion exploration rate could be altered from 0.39% to 5.9% per task and that our method successfully generated a sequence of reaching tasks that matched the range of motion of the selected activity, up to an inlier accuracy of 78.9%. Results demonstrate that the responsibilities of the therapy game (i.e., motivating the patient) and the therapists (i.e., individualizing the training) could be distributed properly. We believe that with our pipeline, effective cooperation between the involved agents is achieved, and the provided therapy can be improved.