Fuzzy dynamical system for robot learning motion skills from human demonstration

Abstract

Learning from demonstration (LfD) is an intuitive strategy for transferring human motion skills to robots in an agile and adaptable manner. The major goal of LfD is to identify significant movement primitives (MPs) from human demonstrations and then recompose those intrinsic primitives to adapt to a variety of new situations. However, maintaining the simplicity of MPs representation while guaranteeing their adaptability is not an easy undertaking. To achieve these two goals, two approaches are possible: (1) learning models that can capture and utilize the inherent patterns and main characteristics of the human demonstrations, and (2) dynamical systems that can respond to perturbations online without requiring to re-plan the entire trajectory. In this paper, we present a novel and efficient model that combines these two benefits to formulate MPs using a fuzzy dynamical system (Fuzzy-DS), which enables robots to adaptively alter the learned motion skills to meet various additional constraints in the process of performing tasks. Due to the joint use of a fuzzy inference system and a dynamic system, Fuzzy-DS is well-suited to human inputs and intuitive fuzzy rules, resulting in a computationally efficient model. To verify the effectiveness of the proposed method, experiments have been designed where the robot learned a plant pruning task and a pick-and-place task, subsequently, it can replicate and generalize these tasks to novel situations.