Motor Imitation: Bio-Inspired vs Direct Geometric Control

Achieving pose imitation with robots is a quite popular topic in robotics. It is widely used for children therapy, notably autistic children, but also to teach new actions to robots. While a lot of very effective methods exist for pose imitation, most of these rely on supplementary equipment for motion capture which rule out a natural interaction and even prevent an interaction which could take place outside of the laboratory. In this paper, we propose a bio-inspired method to achieve imitation with minimal equipment, relying solely on the information provided by the robot Pepper 2D camera. To do so, we perform 2D pose estimation using OpenPose to infer the 3D pose estimation of the human. Using this information, we performed rhythmic and discrete pose imitation using CPG (Central Pattern Generators) controllers endowed with plasticity mechanisms and compared this method with a geometric control approach. Although CPG control has been used previously for rhythmic tasks, it has never been, to our knowledge, been used for imitation.