Path Planning for the Robotic Manipulator in Dynamic Environments Based on a Deep Reinforcement Learning Method

Abstract

Collaborative and autonomous robots are increasingly important in meeting the demands of a faster and more cost-effective market. To ensure production efficiency and safety, robots must swiftly respond to the presence of human operators or other dynamic obstacles, avoiding potential collisions by quickly planning alternative paths. Deep Reinforcement Learning (DRL) based methods have shown great potential in path planning due to their rapid response capabilities. However, existing DRL-based planners lack a safety verification system to evaluate the feasibility of actions generated by neural models, and they cannot guarantee 100% collision-free paths. This paper presents an enhanced DRL-based path planning system incorporating a robust safety verification mechanism. This system predicts potential collisions and generates alternative collision-free paths as necessary. We analyzed the essential elements of trajectory planning using the DRL method and proposed improvements to accelerate planning speed. The results demonstrate that our planner consistently generates paths for typical reaching tasks with an average planning time of 12.1 ms, a notable improvement over traditional algorithms. Moreover, the paths produced by our method are nearly optimal, akin to those generated by Optimization-based algorithms.

Topics & Concepts