NRI: Robotic Tool-Use for Cleaning

University Of Washington, Seattle WA

Investigators

Abstract

Cleaning is as one of the most desired capabilities for personal robots, according to reported surveys and interviews conducted with potential users. However, successful solutions to robotic cleaning have so far been special-purpose robots designed for particular cleaning tasks, as with the vacuuming or mopping robots. Instead, this project aims to make general-purpose mobile manipulators perform a much wider range of cleaning tasks, including dusting, wiping, scrubbing, sweeping, or mopping, by enabling them to use human tools. Cleaning is a key activity of daily living and an important task in the maintenance of one's quality of life. Many older adults lose their independence when they can no longer carry out cleaning tasks. Similarly, many mobility-impaired individuals rely on others for cleaning. Robots taking on these tasks could therefore positively impact millions of individuals. General-purpose cleaning robots could also have significant economic impact, similar to those of special-purpose ones like the Roomba. Cleanliness is critical in many commercial facilities, such as hotels or department stores, which involve large surfaces that need to be regularly cleaned in specific ways that programmable cleaning robots developed in this project could handle. General-purpose cleaning robots need to be able to use many different cleaning tools, in many different environments, based on many different user preferences. The diversity of possible situations makes it difficult to develop universal controllers or planners for arbitrary cleaning tasks. Furthermore, it is impossible to know the users' preferences in advance. To address these challenges, this project develops a Learning from Demonstration (LfD) technique that allows a robot to learn cleaning actions from human demonstrations. This technique exploits the common structure of cleaning tasks: they all involve moving a cleaning tool relative to a target surface with a certain pattern while applying a certain force on the surface. Although many LfD approaches already exist, none of them allow learning cleaning actions that generalize to arbitrary surfaces with arbitrary cleaning patterns. This project contributes new representations and learning algorithms that enables a mobile manipulator to clean surfaces of arbitrary size, shape, curvature, dirt distribution, and clutter, using different types of cleaning tools. The methods developed in the project are evaluated with real cleaning tasks through systematic experiments and user studies in the lab, as well as, in the real world through an identical robot deployed in a private home.