Towards tool handling by a NAO robot using programming by demonstration

Abstract

Programming by demonstration is a technique where, contrary to detailed programming, the robot learns from one or several demonstrations of the execution of the task by a human or another robot. A technique which allows a humanoid robot, take a tool to the head of a screw, following a similar trajectory than demonstrated trajectory by a human being, is presented. The technique also allows variations in location and orientation in the screw head. This is achieved thanks to the use of machine vision techniques and probabilistic models estimated from the demonstrated trajectories. Image processing consists of color space segmentation, and interest point’s selection in the screw head to calculate its position and orientation. The required trajectory in the presence of variations is generated using parameterized Gaussian mixture models in the task. With the above, the new required trajectories are generated in accordance to the variations. The system operation is presented through the illustration of the new trajectories, and pictures of the robot following these trajectories. Despite some limitations of the robotic platform used, acceptable results were achieved.