This paper describes a new vision-based playback method of a wheeled mobile robot.
The path of the robot is divided into straight line parts and curved line parts. The curved line parts were approximated by a Bézier curve to decrease the volume of data. In order to track the curved line parts with good accuracy, tracking errors were detected by referring to two pre-recorded images on one part of the curve. Then, the errors were corrected by adjusting the control points of the Bézier curve. Outdoor experiments demonstrated good repeatability by using the proposed method.