CHS: Medium: Collaborative Research: Manipulation Assistance for Activities of Daily Living in Everyday Environments

Northeastern University, Boston MA

Investigators

Abstract

While many people with disabilities need help with activities of daily living (ADLs) in their homes or at other locations, they care deeply about maintaining their sense of independence, which implies limiting the tasks that professional or family caregivers are asked to provide. There is the potential for robots to have a huge impact here, by enabling people to live independently for longer. The goal of this research is to develop a robotic wheelchair-manipulator system (RoWMan) consisting of a power wheelchair with a robotic arm mounted on it, that will help its user perform ADLs either as an assistive device or by performing manipulation tasks autonomously. In assistive mode, the user would ride in the wheelchair, with the RoWMan system manipulating items as requested. Whereas in autonomous mode, the user could ask RoWMan to navigate on its own through the house, retrieve items, and place them as directed. This project will necessitate the development of new user interfaces as well as an array of new machine learning and robotics techniques that will enable successful autonomous robotic navigation and manipulation in unstructured environments. To ensure broad impact, project outcomes will be evaluated with a user population at Crotched Mountain Rehabilitation Center. In recent focus groups it was found that users want a number of capabilities, including the ability to pick up something from the floor, the ability to unlock and open a door, the ability to manipulate items on a tightly packed shelf, etc. RoWMan will be designed so as to enable users to perform these sorts of tasks, by focusing on two areas: robotic manipulation and human-robot interaction. The manipulation work will develop new algorithms that perform well with novel objects in unstructured environments. Traditionally, manipulation planners assume that the shapes of the objects involved are known in advance or can be estimated on the fly, but these assumptions often cause problems in practice. The focus here will be to develop new algorithms based on deep reinforcement learning that can perform manipulation tasks reliably even when the geometry of the world is unknown in advance. The project will also support research into a new class of human-robot interaction based on laser pointers. Recent work suggests that laser pointing can be very effective for the target user community because it enables users to point directly in the environment rather than on a screen which induces additional cognitive load. This project will develop new ways of communicating sophisticated intent using a combination of environmental context, laser pointing, and laser gestures. 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.