A Reinforcement Learning-based Approach To Multi-robot Joint Navigation

With the development of artificial intelligence and robotics,robots have started to occupy a place in search and rescue,logistics,and warehousing scenarios with lower costs and higher degrees of freedom.The new era has also put forward higher requirements for robot motion,and the problem of how to plan a safe path in a scene with many moving objects is a major challenge for current mobile robots.In the past,the algorithm was required to remain stationary in the face of object collisions to avoid actively colliding with objects.In dynamic scenarios,collisions can have serious consequences,such as in autonomous driving,where the consequences of a crash are unbearable.Therefore,a new modeling approach is needed to improve the perception capability for robots in dynamic scenes.We use reinforcement learning methods to improve the robot’s understanding of dynamic scenes through excellent modeling capabilities to control the robot to make the correct action strategy.Given the above background,two main work elements are addressed in this paper:(1)improving the distributed reinforcement learning training framework to provide an efficient platform for training the algorithmic models in the later paper;(2)designing a hierarchical navigation model based on reinforcement learning.The first point of work focuses on solving the problem of slow training of reinforcement learning models.In the traditional single-process reinforcement learning model,since the acquisition data and learning experience are executed in turns,there is always a module in the waiting stage during the algorithm execution,which causes a lot of performance wastage.Therefore,this paper proposes a buffering mechanism to improve the existing distributed structure,which can decouple the acquisition data and learning experience in the algorithm framework to realize the parallel execution of the algorithm and improve the execution efficiency of the algorithm.The second point of work focuses on the problem of dynamic navigation with multiple robots in an unknown environment.This paper uses the ROS robot operating system with a perfect simulator to realize the joint tuning of solid and virtual models,which can improve the algorithm in the simulator and real environment;this paper proposes the idea of using hierarchical navigation,and proposes a combination of local navigation using A*global navigation enhanced learning,which improves the robot’s perception of unknown dynamic scenes;in addition,according to the actor critic algorithm A model optimization method for multiple critics is proposed to further improve the performance of the navigation algorithm.It makes the algorithm superior in terms of the average number of collisions as well as the navigation success rate for 2D grid navigation methods in dynamic scenes;maintains the algorithm performance during scene migration;and can be used for multi-robot dynamic navigation scenes.