Peg-in-hole assembly in live-line maintenance based on generative mapping and searching network

Abstract Replacement of lightning arrester is one of the common tasks in live-line maintenance, and peg-in-hole assembly is a very difficult operation for a robot, because there are visual inaccuracy and force model uncertainty in the process of assembly. This paper presents a new implementation approach fusing signals of vision detection and fuzzy force to realize the high efficiency peg-in-hole assembly by a manipulator autonomously. YOLOv3 is applied as the visual detection network for rough alignment. In the phase of precise hole-searching, we establish a two-dimensional hole-searching model by fusing signal of vision detection and fuzzy force as the condition of state transitions, and propose a new semi-supervised learning network to optimize the hole-searching routine. The performance of the approach is verified by experiments in the simulation environment and the laboratory environment.