One-shot Unsupervised Learning for Improved Human-Robot Collaboration

Abstract

As robots become commonplace in human environments such as households and manufacturing facilities, they require new control methods for e↵ectively collaborating with their human partners in common pick-and-place tasks, such as unloading a dishwasher or providing parts for assembly. In this paper, we propose an online, generalizable taskmodeling technique that enables robots to infer user task progress and determine their own path progress accordingly in order to optimize user experience and maintain good task performance. Furthermore, this method allows robots to actively build new task models when users temporarily or permanently switch to unknown tasks. We evaluate the algorithm’s effectiveness at building and updating task models using data collected on human performances of various pick-and-place tasks. We then implement the algorithm on a robot arm in a user study and show that the algorithm performs as well or better than the current state-of-the-art algorithm but with much less manual modeling effort.