Nearly 56.7 million (18.7%) of the non-institutionalized US population had a disability in 2010. Among them, about 12.3 million needed assistance with one or more activities of daily living (ADLs), such as feeding, bathing, or dressing. Robots have the potential to help with these activities of daily living but every user is different and they have diverse needs and preferences. For long-term care, it is essential that such an assistive system can adapt to diverse situations and user preferences. This project focuses on the feeding activity and is based on this central tenet that by leveraging user feedback and contexts from previous feeding attempts, a robot should be able to learn online how to adapt to new food items, user preferences, and environments. Through improved access to independent living, the results of this project can positively impact millions of people worldwide. The long-term promise of this research is to have robots in society that can seamlessly and fluently perform complex manipulation tasks in cluttered, complex, and dynamic human environments in real homes.
This project formalizes robot-assisted feeding using a general framework based on contextual bandits that allows directly optimizing for user preferences online. The online contextual bandit framework applied to acquiring and transferring food items provides the foundation to leverage user feedback to benchmark, learn, and develop methods for a natural dining experience, and exploring different contexts for generalizing bite acquisition. The models directly optimize for the user experience through user feedback, and adapt to a range of social and environmental factors with the intelligent use of embedded sensing. The project explores solutions which balance the trade-off between high quality but costly expert assistance and cheaper learned solutions in the form of a shared-autonomy system. Critical issues include the diversity of user preferences both temporally and ethnographically, designing for the experience across the entire learning procedure, and processing high-dimensional contextual information. The tangible result will be an intelligent assistive feeding robot whose performance can generalize to different activities and adapt to user preferences. An intelligent assistive feeding robot that relies on user feedback and rich sensor information will advance integrating complex user experiences and social environments into a coherent learning robotic system. Contextual bandits, a highly optimized generalization of multiple hypothesis testing, have broad potential in human-robotic, and human-AI systems in general to efficiently adapt to specific user needs in real time.
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.
