Research On Collaborative Navigation Method Of Mobile Robots Based On Deep Reinforcement Learning

Mobile robot collaborative navigation technology has been widely applied in fields such as robot collaborative target tracking,collaborative target strike,and collaborative target search.To ensure the stable and efficient operation of mobile robot systems in dynamic unknown environments,real-time dynamic target assignment and local coordination obstacle avoidance are the key to solving the problem of collaborative navigation for mobile robots.The existing research methods are mainly based on the assumption that the environment is fully observable,and most of them do not consider the collaborative navigation problem of mobile robots in situations where the environment is partially observable and the target position changes randomly.Based on this,this paper proposes an end-to-end visual collaborative navigation control algorithm called Single Agent Control of Multiple Robots Network(SACMRN)for multi robots controlled by a single agent.In an unknown environment without pre assigned targets,the proposed algorithm does not require environmental map information and positioning technology,but only utilizes local visual observation information of the mobile robot to complete the collaborative navigation task in the shortest time step without collision.The main work of this article is as follows:(1)A control algorithm for robot cooperative navigation based on Deep reinforcement learning is proposed.A SACMRN collaborative navigation algorithm is proposed to address the problem of target dynamic allocation and collaborative navigation in dynamic unknown environments.The algorithm adopts a centralized training method to solve the problem of unstable environment during the training process of agents.On this basis,based on the idea of value function decomposition,this article designs a single global reward function from top to bottom.During the learning process,the agent will autonomously assign rewards to each robot to solve the credit assignment problem between robots.(2)We have built a simulation and physical experimental platform and completed algorithm performance verification.Firstly,a collaborative navigation simulation environment for mobile robots was built on the Robot Operating System(ROS)and Gazebo platform,and end-to-end autonomous navigation algorithm training for mobile robots was completed.Then,we deploy the network model trained in the simulation environment to the physical platform.The experimental results show that the proposed algorithm is superior to VDN,MADDPG,and IQL algorithms.After training,the agent can make autonomous decisions based on the state to assign navigation targets to each robot without pre-assigning targets,and even if the position of the target changes dynamically during the navigation process,the agent can still dynamically assign targets in real time,control the mobile robots to avoid the obstacles in the environment and collaboratively navigate to all target positions.The proposed algorithm can achieve the collaborative navigation function of mobile robots in dynamic unknown environments.