Online Adaptive and Attention-based Reference Path Generation for Upper-limb Rehabilitation Robot

Abstract

This paper proposes an online reference path generation method for upper-limb rehabilitation robots considering the initial motions of the subjects. The proposed system is grounded on a Learning by Demonstration (LfD) approach based on an attention-based neural network, and the method presents a high level of generalization allowing the user to perform the point-to-point tasks, as most upper limb motions such as Activities of Daily Living (ADL) tasks can be segmented into multiple point-to-point motions. The proposed approach has been experimentally validated on a robotic arm (i.e., the Franka Emika Panda) attached to a human subject’s hand, and experimental results demonstrate that 1) the path generated by the proposed approach gets higher accuracy (promote approximately 79.18%) compared with the Dynamic Movement Primitives (DMP) in terms of similarity with healthy human motion in the test dataset. 2) the proposed system can guarantee the performance concerning the variation of starting points and target points.