Tool centered learning from demonstration for robotic arms with visual feedback

Abstract

Visual feedback allows controlling the relative pose between the camera and an object of unknown pose. This paper presents an approach to reproduce and generalize a movement from demonstrations based on visual feedback from an eye-in-hand camera mounted to the end effector. The main advantage of the approach is that no robot is required during the demonstration phase which is tool rather than robot centered. The teacher merely demonstrates the movement either with the camera alone or the camera attached to the tool that is required for the task. This method simplifies the demonstration in comparison to kinesthetic teaching that requires a compliant arm for motion recording. The accuracy and precision of the demonstration is enhanced as the teacher does not have to overcome the inertial and frictional forces of the arm while handling the end-effector in conjunction with the tool. The movement pattern in terms of translation and rotation of the camera with respect to the goal pose is described by dynamic movement primitives (DMP). Learning from demonstration is applied to straight point-to-point and more complex trajectories. The reproduced trajectories reveal a high precision independent from the object pose and the approach is confirmed by an experiment with a five degree-of-freedom (DOF) robot.