CAREER: Robot Learning of Complex Tasks via Skill Reusability and Refinement

University Of Massachusetts Lowell, Lowell MA

Investigators

Abstract

The creation of robots capable of performing complex manipulation tasks in unstructured environments will open up new applications for helping people in their homes, at work, and in society to impact and enhance quality of life. Existing work in robot learning has enabled robots to replicate very simple manipulation tasks. However, learning complex manipulation tasks to assist people (e.g., loading a dishwasher, grocery shopping, changing a lightbulb) demands algorithmic advancements. This project will contribute new theoretical methods and algorithms that will advance the state-of-the-art in robot learning and accelerate the development and adoption of robots capable of supporting people with a variety of assistive and autonomous applications. In addition, this project will integrate research, education, and outreach by developing new courses, mentoring and supporting students from underrepresented groups, reaching out to K-12 students and the general public, and organizing educational workshops. The goal of this project is to develop a unified framework for learning and generalization of complex manipulation tasks. Our framework leverages human-robot interaction and learning-from-human approaches to address several existing challenges: Robots that can model and learn primitive skills from human examples mostly ignore characteristics of human-like movements. All the individual skills needed to construct a complex task plan must be modeled a priori and cannot be discovered during the process. As the task conditions change, the learned skills cannot be adapted or refined effectively and need to be remodeled. To further knowledge in these areas, this project focuses on: (a) building a unifying formalization for modeling primitive skills by integrating fundamental concepts and experimentally proven mathematical models of human movements into existing learning frameworks; (b) developing approaches for discovering common and reusable primitive basis for a family of complex manipulation skills during the task learning process; (c) creating novel skill refinement methods with strong adaptability properties utilized for the decomposition and reconstruction of complex tasks that can benefit from multi-modal human feedback and other task-related context, while being computationally inexpensive. The potentially transformative aspects of our research ideas include: bringing a new perspective to the mathematical modeling of primitive skills, finding practical solutions to the problems of skill refinement and reusable skill discovery, and shedding light on the development of a complete framework for complex manipulation task learning. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.