Robot Learning From Demonstration for Assembly With Sequential Assembly Movement Primitives

• Published in: IEEE/ASME transactions on mechatronics 2024-08, Vol.29 (4), p.2685-2696
• Main Authors: Hu, Haopeng, Yan, Hengyuan, Yang, Xiansheng, Lou, Yunjiang
• Format: Article
• Language: English
• Subjects: Active control; Assembly lines; Behavioral sciences; Compliant control; flexible manufacturing; Human motion; learning from demonstration (LfD); Manufacturing engineering; Policies; Quaternions; Reliability; Robot learning; Robotic assembly; Robots; Task analysis; Task complexity
• ISSN: 1083-4435; 1941-014X
• DOI: 10.1109/TMECH.2023.3336520

To facilitate the labor-consuming flexible assembly lines with robots, the robot learning from demonstration (LfD) is the promising way to efficiently impart human assembly skills to robots. Aiming at the challenging complex precise assembly tasks, which are contact-rich and require 6-D movement, an LfD method is proposed here. Due to the inconsistent requirements for the robot's assembly movement, the whole robotic assembly processes are composed of three phases, namely, the approaching, aligning, and assembling phase. The policies, which are prestructured by the proposed sequential assembly movement primitives, are learned exclusively to guide the robot's movement in each phase. In the approaching phase, the policy generates a reliable path for the robot to accurately track. However, in the aligning and assembling phase, the polices enable the robot's active compliant behavior to accomplish the complex precise assembly task. Robotic assembly experiments with four objects are conducted to validate the proposed LfD methods with a torque-controlled robot. Experiment results indicate that the proposed LfD method applied with the proposed policies achieves high reliability and efficiency.