A Twin Dual-Arm Robot Testbed to Collect the Human Demonstration Data for Imitation Learning^*

Abstract

Collecting consistent human demonstrations has been a problematic issue in the field of a learning from demonstration. This paper presents a twin dual-arm robot testbed to collect human demonstration data while performing manual tasks in the activities of daily livings such as opening/closing a window, opening-up/closing-down a sauce bottle cap, etc. The testbed consists of the two sets of a dual-arm robot; for each robot, one-side arm has 6-DoF with a crab-gripper as its hand. Two dual-arm robots are mounted at the front and back of a table-type station. A human operator performs actions to complete a given task with the one at the back; the human operator's actions, meanwhile, are duplicated by the other robot at the front simultaneously. The experimental results show the testbed's potential to gather consistent human demonstration data for training a machine learning agent to manipulate the dual-arm robot under an imitation learning framework.