An electromyography signals-based human-robot collaboration system for human motion intention recognition and realization

Abstract

• A human-robot collaboration system based on Electromyography (EMG) signals is proposed to learn the position adjustment skills of human body facing the change of external force information, so as to achieve human-robot cooperation. • To improve the recognition accuracy of human motion intention, the coupling effect of human output force and joint angle is considered. • To improve the accuracy and comfort of the collaboration, from the perspective of ergonomics, two evaluation methods of control signal execution efficiency and motion stability are proposed. • A model-free intelligent control mothed is constructed based on DDPG reinforcement learning algorithm, which effectively avoids the selection of control parameters. With the development of manufacturing industry to the directions of personalization and flexibility, the advantages of human-robot collaboration attract more and more attention. In order to enable the robot fully understand human motion intention and play the supervision and guidance role of the human tutor, a human-robot collaboration system based on Electromyography (EMG) signals is proposed. In the investigation, first, aiming at the problems that the coupling effect of output force and joint angle is not considered in the previous research, and the extraction accuracy of force information is poor, the Fast Orthogonal Search (FOS) algorithm is modified to improve the recognition accuracy of motion intention by using the correlation coefficient between the extracted feature signal and the arm output force as the optimization parameter. Moreover, aiming at the problems that the traditional control models cannot balance accurate tracking and comfortable collaboration, and the robot motion stability is poor, two evaluation methods of control signal execution efficiency and motion stability are proposed from the perspective of ergonomics. Afterwards, the human-robot end-to-end collaborative control model based on Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm is trained online to improve the accuracy and comfort of the collaboration. As compared with the traditional PD control algorithms based on admittance model, the proposed control method is model-free and intelligent. It effectively avoids the challenge of PD control parameter selection, and behaves good coordination between accurate tracking and comfortable collaboration. Human-robot collaborative sawing experiment shows that the proposed collaboration system can save about 86.2% physical power for the human tutor.