Learning RL policies for anticipative assistive robots by simulating human-robot interactions in real scenarios using egocentric videos

摘要

Social robots are increasingly being used to assist vulnerable people in their daily activities. However, for successful human-robot interaction (HRI), the robot needs to be pre-trained beforehand to minimize the interaction time necessary to reach an acceptable performance level and to ensure safety avoiding uncontrolled behaviors. In this scenario, building a simulation environment becomes a very convenient approach. Our proposal contributes in this direction by designing and implementing a simulation platform that, together with a full system enabling HRI between patients with sensorimotor disabilities and anticipative robots, allow learning the robot policies using egocentric videos recorded with healthy volunteers. Our experiments demonstrate that the models learned in our simulation platform provide a more efficient and natural HRI than reactive robots that simply respond to patients’ commands. Additionally, we show how the robot behavior varies as a function of factors of interest modelling both human preferences and robot physical skills.