A Learning from Demonstration Method for Generating Human-like Actions on Redundant Manipulators

Abstract

Achieving human-like actions on robots can significantly improve the quality of human-robot collaboration (HRC). However, it is not easy to control the general-purpose robotic manipulators to work in a human-like style when the kinematic discrepancy exists between humans and robots. In this paper, we propose a human-in-the-loop learning framework to enable human-like properties on a general-purpose redundant manipulator. We collect the human-like demonstration dataset via teleoperation. We then use the behavioral cloning method to train a neural network policy to generate human-like constraints for the redundant manipulator automatically. Furthermore, we introduce an online relabeling method to accelerate learning and relieve workload in the demonstration. Experimental results showed that our proposed framework could successfully train a human-like constraint generation policy with a demonstration dataset collected within a few minutes. Appearance similarity could be seen between the human arm posture and the configuration of the manipulator when performing pose tracking tasks.