RESEARCH ON OBSTACLE AVOIDANCE OF ROBOTIC MANIPULATORS BASED ON DDPG AND TRANSFER LEARNING

Shuhuan Wen,∗,∗∗ Wen Long Zhen,∗,∗∗ Tao Wang,∗,∗∗ Jianhua Chen,∗,∗∗ Hak-Keung Lam,∗∗∗ Qian Shi,∗∗∗ and Zekai Li∗,∗∗

Keywords

Robot arm, deep reinforcement learning, DDPG, position andposture planning, transfer learning

Abstract

The multiple DOF manipulator has the characteristic of flexible movement, and its motion planning is a research hotspot in robotics. The multiple DOF manipulator is required to plan its path and grasping pose in order to achieve handling and human–machine cooperation in a complex environment. This paper proposes a novel framework for motion planning of the manipulator using deep reinforcement learning integrated with transfer learning. We transfer the strategy of path planning to the complex task of grasping pose by using deep deterministic policy gradient (DDPG) in an environment with obstacles. DDPG integrated with transfer learning is used to plan the path of NAO robot’s right arm and grasping pose in an environment with obstacles. The strategy of path planning using the DDPG algorithm is transferred to the learning process of grasping pose in the obstacle environment. Experiments show that the proposed algorithm can avoid obstacles effectively when the manipulator plans grasping pose. Furthermore, transfer learning combined with DDPG makes the manipulator finish the complex planning of pose grasping more quickly than the traditional DDPG algorithm without transfer learning.

Important Links:



Go Back