A comparative study of skill learning for task composition in human-robot collaborative scenarios

Learning from Demonstration (LfD) paradigm has been widely developed in the last decade in the context of Human-Robot Collaboration (HRC). It allows for the collection of demonstrations from several types of sources and the teaching of specific motions or trajectories to the robot that could be difficult to be implemented with traditional methods. In this paper, we present an LfD pipeline that gathers data from a human expert and teaches the robot low-level skills, so it can perform higher-level tasks by means of task composition. Three learning methods are compared: Goal-Conditioned Behavioral Cloning, Dynamic Movement Primitives, and Gaussian Mixture Model with Regression. The low-level skills that have been trained are PICK, PLACE, and POUR. The approach has been tested in a real cobot with an RGB-D camera and ArUCO markers, which are used for easier skill parametrization and generalization.