Globally Perceived Obstacle Avoidance for Robots Based on Virtual Twin and Deep Reinforcement Learning

Abstract

Investigations on obstacle-avoidable robotic trajectory generation is of great significance to the secure production of ordinary machinery factories, which allows robots to work in complex environments. However, conventional collision-free trajectory generation is highly dependent on manual analysis of the environment, making the trajectory generation extremely dedicated. To solve this problem, a more intelligent obstacle-avoidable trajectory generation method based on deep reinforcement learning that can globally perceive obstacle's information and automatically generate trajectories without inverse kinematics is proposed in this paper, where a virtual system corresponding to the physical robot platform is constructed for policy learning motivated by the concept of virtual twin, and an obstacle-avoidable reward with a global perception capability is proposed. Experimental results have verified the superior performance of the proposed method.