Human-Adaptive Impedance Control Using Recurrent Neural Network for Stability Recovery in Human-Robot Cooperation

Abstract

This paper proposes a human-adaptive impedance control to achieve human-robot co-manipulation of an object, in which a recurrent neural network (RNN) estimates a human state to be used in improving the contact stability of impedance control. The human sate here is defined as what indicates that the human-arm stiffness becomes harder and the stability of the impedance control is deteriorating. In the proposed method, the human state is estimated from rectified-and-integrated electromyogram (iEMG) signals while the person is manipulating the object cooperative with the robot. According to the degree of the estimated human state, the proposed method changes the impedance parameters to be heavier online so as to make the system more stable and then returns the mechanical parameters to be lighter once the stability is restored. The validity of the proposed method is verified by experiments using a commercial-of-the-shelf manipulator that collaborates with a person based on the impedance control, which can react to the net external force during object manipulation proposed by the authors in July 2018. The experimental results based on a cross validation showed that the proposed RNN can successfully estimate the human state from the iEMG signals and detect the undesirable oscillation occurred while the person is manipulating an object with the robot. It has been also confirmed that the proposed human-adaptive impedance control, which adjusts the impedance parameters online according to the human state, is effective to prevent the human-robot cooperative system from coming unstable.