A system for robot teaching based on RGB-D images and a teach pendant, including an RGB-D camera, a host computer, a posture teach pendant, and an AR teaching system which includes an AR registration card, an AR module, a virtual robot model, a path planning unit and a posture teaching unit. The RGB-D camera collects RGB images and depth images of a physical working environment in real time. In the path planning unit, path points of a robot end effector are selected, and a 3D coordinates of the path points in the basic coordinate system of the virtual robot model are calculated; the posture teaching unit records the received posture data as the postures of a path point where the virtual robot model is located, so that the virtual robot model is driven to move according to the postures and positions of the path points, thereby completing the robot teaching.