A Robot with Chopsticks: How do Interfaces and Expertise affect Demonstrations?

Humans have proved to be remarkably effective using chopsticks to perform a variety of manipulation tasks. Inspired by human performance, we focus on manipulation tasks requiring the use of a robot manipulator with a chopsticksequipped end-effector. We present a novel teleoperation interface that maps the tracked motion of a human-controlled pair of chopsticks to the motion of the robot’s chopsticks. Our key insight is that leveraging human adaptability in learning how to control the robot through our teleoperation interface could enable the collection of high-quality datasets for robot learning of chopsticks-based manipulation. As a first step towards this goal, we studied 25 subjects in whicuh we investigate the factors governing human performance in chopsticks-based manipulation of everyday-life objects across three methods including our teleoperation interface, motioncapture tracked chopsticks, and normal chopsticks. Findings include: a) humans can teleoperate the robot to solve very challenging manipulation tasks such as grasping a slippery glass ball with a pair of slippery metal chopsticks, without the use of haptic feedback; b) teleoperation in some cases is even preferred over using normal chopsticks, opening up the landscape for collecting on-hardware demonstrations that the robot can directly learn from and c) subjective ratings found the teleoperation interface to be the least comfortable and most difficult to use though it achieved equivalent success rate to other methods.