A DMPs-based Approach for Human-Robot Collaboration Task Quality Management

Abstract

Industry 5.0 places humans alongside robots on a symbiotic collaboration to improve the efficiency and productivity of industrial processes. The current manufacturing industry targets personalized products, that demand flexible, agile, and quickly changeable workstations that require less skilled workers capable to handle the different production needs. Human-robot collaboration and Learning from demonstration are emerging fields in robotics that can be exploited for this end. This paper explores a learning from demonstration (LfD) approach to learn how to perform a collaborative task with an experienced collaborator and actively teach and/or assist a novice worker. A reference trajetory was recorded using the UR10e robot and modelled by non-linear dynamical system, specifically, dynamic movement primitives (DMPs), whose weights are learned using Covariance matrix adaptation evolution strategy (CMA-ES). This paper also explores DMP effectiveness to generate the learned trajectory, with the ultimate goal of managing the quality of a collaborative task. The obtained results explore DMPs robustness against sudden perturbations and deviations from the encoded trajectory, both in simulation and in real context. Furthermore, the flexibility and stability of DMPs in learning the references’ trajectories, as well as their temporal and scale invariance, were verified.