Inference of Manipulation Intent in Teleoperation for Robotic Assistance

Abstract

In teleoperation, predicting an operator’s intent and providing subsequent assistance have demonstrated great advantages in reducing an operator’s workload and a task’s difficulty as well as enhancing the task performance. Current research aims to tackle target-approaching intent, while our work focus on inferring manipulation (task) intent after the user grasps the object. We model how an object is grasped when being utilized in different manipulation tasks (intents) and then adopt this object grasping model in teleoperation for the intent inference. Our paper focuses on determining if direct interaction models can be used for indirect interaction. As the nature of one’s grasping pose may satisfy multiple tasks (intents), we explore a form of classification modeling known as multi-label classification for multiple broad categories of tasks and objects. We also comprehensively compare classification techniques to determine the most suitable method for determining manipulation intent. With knowing the manipulation intent, future robot control algorithms can provide more helpful and appropriate assistance to facilitate task accomplishment.