Synergy of Deep Learning and Artificial Potential Field Methods for Robot Path Planning in the Presence of Static and Dynamic Obstacles

摘要

In a fast-changing world we are in today, unmanned vehicles are displacing old, obsolete and frustrating tasks. Since unmanned vehicles are intended to work in an environment without any conduction, finding a collision-free path of movement is of paramount importance and a definite asset in practice. The objective of a global path planner consists in generating an optimized path without considering unknown obstacles. In contrast, a local path planner properly avoids the unknown obstacles while getting close to the goal. This results in an inefficient movement, particularly in complex circumstances. In this paper, a complete global and local path planning method is proposed for avoiding both static and dynamic obstacles. The proposed method benefits the advantages of both approaches while covering each other’s weaknesses. First, a method for generating numerous obstacle-free paths from random pairs of start and goal points is introduced. Next, a novel deep-learning approach is proposed in order to train the robot in an environment free of moving obstacles. After extracting the desired via points, an efficient Artificial Potential Field (APF) approach for attaining the local path planning is introduced with the aim of avoiding dynamic obstacles while the robot travels through the aforementioned via points. The proposed method can be well extended to different platforms such as mobile robots, arm robots, quadrotors, etc; in this paper, both local and global path planning methods are implemented on a simulated quadrotor to examine the robot’s performance for both approaches. Furthermore, it has been revealed that implementing both approaches should be implemented seamlessly in order to attain a complete efficient path planning with the presence of both static and dynamic obstacles. Finally, both local and global path planning methods are implemented on a detailed simulated quadrotor to evaluate the robot’s performance for the aforementioned approaches.