Smoothing of human movements recorded by a single RGB-D camera for robot demonstrations

Abstract

In environments such as domestic ones, Learning from Demonstration methods could allow end-users to teach to their robots new skills. At the same time, it could be required that the demonstrations are available through a simple set-up, such as via a single visual sensor. In the present work, we present a real-time pipeline for demonstrating movements to the end-effector of a non-anthropomorphic robot via a single RGB-D sensor. To deal with the related information noise we use Bezier curves to smooth human movements before providing them to the robot. We evaluated our method by considering the performance of the Bezier curves with respect to both the recorded data and the performed robot motion in real hardware. Overall, the Bezier curves produce a good fit of the raw data and result to robot movements that are considerably smoother. The entire demonstration pipeline is implemented and available at the Robot Operating System.