Skeleton-Based AI-Driven Feedback Generation for Physical Rehabilitation via Social Assistive Robotics

The Inrobics Rehab platform allows users to exercise with a social assistive robot, be it for physical rehabilitation or just stimulation. The users follow the robot in a mirroring activity where they have to imitate the robot’s movements, while the system checks that the user is correctly doing them by means of a 3D sensor. In this paper we consider the integration of a feedback generation system that helps users achieve the poses expected by the robot, so that it can behave fully autonomously. To generate this feedback we have considered the use of different Artificial Intelligence techniques, such as expert systems or machine learning models, including decision trees and neural networks. The problem has been modelled as a classification task, where metrics such as the accuracy, precision and f1 score have been used to evaluate the methods. Following this, we discovered that the best option according to these metrics is to choose the neural network using the explicit pose id. Despite of this we will find that using the random forest using the expected user angles, is not only faster, but also more stable to all classes.