Assistive Teleoperation: A New Domain for Interactive Learning

Assistive Teleoperation as Policy Blending In assistive teleoperation, the robot attempts to predict the user’s intent and augments his or her input based on this prediction, in order to simplify the task. Our recent work on policy blending (Dragan and Srinivasa 2012) formalizes assistance as an arbitration of the user’s input and the robot’s prediction. At any instant, the robot combines the input, U , and the prediction, P , using a state-dependent arbitration function α ∈ [0, 1] (Fig.1(bottom)). Policy blending can have a strong corrective effect on the actual input provided by the user, but burdens the robot to predict accurately and arbitrate appropriately.