Robot peg-in-hole assembly based on contact force estimation compensated by convolutional neural network. (March 2022)

• Record Type: Journal Article
• Authors: Zhang, Tie; Liang, Xiaohong; Zou, Yanbiao
• Is Part Of: Control engineering practice. Volume 120(2022)
• Publication Date: 2022-03
• Subjects: Contact force estimation -- Convolutional neural network -- Supervised learning -- Position and orientation adjustment algorithm -- Robot peg-in-hole assembly; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2021.105012
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: In order to reduce the cost of robot integrated application, this paper proposes a robot peg-in-hole assembly method without force/torque sensor. A disturbance observer based on generalized momentum and joint motor torque is employed to obtain the force information on the end of the robot, as a replacement of force/torque sensor. Since the contact between the robot end and the external environment excites the unmodeled dynamics of robot, which leads to insufficient estimation accuracy of the disturbance observer, a convolutional neural network (CNN) supervised learning method to calibrate the force estimation algorithm is proposed. Then, the peg-in-hole assembly mechanism of a ball-end round shaft is analyzed, and on this basis, an assembly strategy and a fuzzy proportional–derivative orientation adjustment algorithm are proposed. Finally, the feasibility of the proposed method of robot peg-in-hole assembly without force/torque sensor is verified by experiment. Highlights: Robot joint variables matrix forms a virtual two-dimensional image. The convolutional neural network extracts features from robot joint variables matrix. These features are mapped to the compensations of force estimation. The compensated force estimation helps robot assembly without force sensor. Robot peg-in-hole assembly based on fuzzy PD control is proposed.