Grasping with Chopsticks: Fine Grained Manipulation using Inexpensive Hardware by Imitation Learning

Billions of people use chopsticks, a simple yet versatile tool, to pick up a wide variety of food items in their daily lives. We hope to leverage human demonstrations to develop autonomous chopsticks-equipped robot manipulation strategies for hard manipulation problems. The small, curved, and slippery tips of chopsticks require fine-grained control, which pose a challenge for picking up small objects. In this preliminary work, we explored imitation learning methods to learn to pick up small cube and ball-shaped objects from expert’s teleoperation demonstrations. We trained a behavior cloning agent, a k-Nearest Neighbors agent, and a blending of both in robot-centric and object-centric representations. We found that blending of the two agents showed some promise in teaching the chopsticks robot to pick up small objects in the object-centric frame. However, there is still a need to incorporate adaptive real-time feedback in the learner to improve and generalize the manipulation performance, which points us to some plausible directions for future work.