Cheol-Hui Min and Jae-Bok Song*
International Journal of Control, Automation, and Systems, vol. 20, no. 10, pp.3296-3311, 2022

Abstract : In recent years, several control policies for a multi-degree-of-freedom (DOF) manipulator using deep reinforcement learning have been proposed. To avoid complexity, previous studies have applied a number of constraints on the high-dimensional state-action space, thus hindering generalized policy function learning. In this study, the control problem is addressed by in-troducing a hierarchical reinforcement learning method that can learn the end-to-end control policy of a multi-DOF manipula-tor without any constraints on the state-action space. The proposed method learns hierarchical policy using two off-policy methods. Using human demonstration data and a newly proposed data-correction method, controlling the multi-DOF manipu-lator in an end-to-end manner is shown to outperform the non-hierarchical deep reinforcement learning methods.

Keyword : Deep reinforcement learning, demonstration-based learning, end-to-end robot control, hierarchical reinforcement learning.