Abstract: In lunar surface sampling by Chang'e 5, it is difficult to use the slender and flexible manipulator to achieve precise operations, such as digging, putting the sample into the pot, pot grasping and assembly, etc. In this paper, a vision system, consisting of static cameras installed on the lander and dynamic cameras mounted on the manipulator, is developed to get images and relative positions of the end-effector relative to the targets, so as to guide precise operations. We propose a novel positioning method by combining two kinds of features observed in multiple cameras, the circular objects and drones, and build a unified framework for calculating the relative poses between the end-effector and the targets in different operations. The influences of the manipulator error on the convergence of visual positioning are analysed, and several applications with combinations of different features and cameras are realized for operations of digging, sample putting, etc. Experimental validation and analysis show that the accuracy of binocular ellipses positioning in digging is better than 2 cm, and the accuracy of positioning with circular objects and drones in sample putting is better than 2 mm. The results meet the requirements of various operations, and the method proposed successfully supported implementation of the Chang'e 5 lunar surface sampling and packaging mission.

Key words: Chang'e 5, visual positioning, multi-feature fusion, robotic arm control, teleoperation