Towards perception driven robot-assisted task-oriented therapy

Therapy environments for robot-assisted stroke rehabilitation are mostly static, in that objects are earmarked for different functional tasks, object locations remain fixed, trajectories are predefined and tasks manually selected. This work advocates the need for a therapy environment that allows dynamic object positioning, different objects can be used for the same functional task and tasks can be selected autonomously by exploiting affordances. Specifically, we propose a framework that facilitates therapeutic robots to recognize common objects, learn functional-task possibilities in a physical workspace and plan trajectories to accomplish identified tasks. We collect a new RGBD corpus characteristic of a therapy environment and report accuracy of the framework in classifying a task possibility, given a set of objects. Further, we implement the proposed framework on the ADLER robot and test for a Reach task.