Abstract

Robot learning from demonstration (LfD) requires data collection for mapping the sensory states to motion action, which plays a significant role in the learning efficiency and effectiveness. In this paper we present a data collection framework that allows a human demonstrator to teleoperate or to visually guide a mobile robot for the required behaviors, while the sensory-motor examples are simultaneously gathered. In the teleoperation mode, the human demonstrator can teleoperate the robot through a GUI that consists of the velocity control and sensory-motor recording commands with the monitoring windows for sonar, laser and visual image. In the visual-guided mode, the human demonstrator uses a green can as the command stick that is tracked by a pan-tilt-zoom (PTZ) camera. The framework is implemented on a Peoplebot robot. Experiments show that both demonstration modes of the framework provide an user-friendly interface of data collection for the subsequent learning process of the robot.