Guided Robot Skill Learning: A User-Study on Learning Probabilistic Movement Primitives with Non-Experts

Abstract

Intelligent robots can potentially assist humans in everyday life and industrial production processes. However, the variety of different tasks for such robots renders pure preprogramming infeasible, and learning new tasks directly from non-expert users becomes desirable. Hereby, imitation learning and the concept of movement primitives are promising and widely used approaches. In particular, Probabilistic Movement Primitives (ProMPs) provide a representation that can capture and exploit the variance in human demonstrations. While ProMPs have already been applied for different robotic tasks, an evaluation of how non-expert users can actually teach full tasks based on ProMPs is missing in the literature. We present a framework for Guided Robot Skill Learning which enables inexperienced users to teach a robot combinations of ProMPs and basic robot motions such as gripper commands or Point-to-Point movements. The proposed approach represents the learned skills in the form of sequential Behavior Trees, which can be easily incorporated into more complex robotic behaviors. In a pilot user study with 10 participants, we investigate on two robotic tasks how inexperienced users train ProMP based skills and how they use the concept of modular skill creation. The experimental results show that ProMPs enable more successful task execution compared to teaching Point-to-Point motions. Additionally, our evaluation reveals specific problems that are relevant to consider in future ProMP based teaching systems for non-expert users such as multimodality and missing variance in the demonstrations.