Online Human Intention Detection through Machine-learning based Algorithm for the Control of Lower-limbs Wearable Robot

Abstract

Online human intention detection is one of the main challenges to ensure smooth human robot interaction for assistive robotics through wearable devices. This paper proposes a framework that combines both machine learning based algorithms and task-oriented control of an actuated-ankle-foot orthosis for human locomotion assistance during five gait modes that are level walking, stairs ascent/descent, and ramp ascent/descent. A random-forest based algorithm has been trained to provide an online classification of the five gait modes using kinematic features of a dataset collected with ten healthy subjects. Finally, appropriate assistive torques were applied at the ankle joint level with respect to the detected gait mode. The proposed scheme is verified in terms of gait mode detection success rate and the torque assistance through the actuated-ankle-foot orthosis at the ankle joint level. One healthy subject participated in the experiments with and without applying the torque assistance strategy. The results show the following average success rates of 99.49%, 98.30%, 96.07%, 84.63%, and 85.55% for the different locomotion modes, that are level walking, stair ascent, stair descent, ramp ascent, and ramp descent, respectively.