A DMPs-based approach for human-like robotic movements

Abstract

Industry 5.0 requires flexible and agile robots, capable to be adapted to different tasks. Tasks that demand from human workers complex movements, with large amplitudes and considerable loads, and whose layout alteration to allow good ergonomics would imply a very significant economic expenditure. In these cases, where the ergonomic safety of the workers is not guaranteed, the introduction of a robot in a production line is preferable. Human-robot collaboration pose as a solution for this problematic. However, human-likeness motion reproduction is still missing from robots. This paper explores a Learning from Demonstration strategy, a subfield of Human-Robot Collaboration (HRC) focused on teaching robots how to master a skill based on human demonstrations. Specifically, 12 human movements were recorded using MTw Awinda Motion Capture system to be further modelled by non-linear dynamical system, specifically, Dynamic Movement primitives (DMP), whose weights are learned using Covariance matrix adaptation evolution strategy (CMAES). This was used to learn how to perform human movements and transfer these skills to a collaborative Robot UR10e.