Research On Autonomous Navigation Of Target-Driven Mobile Robot

At present,mobile robot navigation technology has been widely used in border patrol,anti-terrorism operations,security logistics and other military and civilian fields.With the expansion of the scope of application,the navigation problem of robot in unknown environment has become one of the most practical and challenging research topics in robot technology.As the application scenarios are often difficult to predict,the traditional navigation methods based on map information are difficult to be extended to unknown environments.Therefore,to make the robot explore the environment safely in the complex unknown environment,determine the target position efficiently and independently,and complete the navigation task with high robustness are the key points and difficulties for the wide application of mobile robot.For this reason,taking TurtleBot2 mobile robot as the control object,this paper deeply studies the autonomous navigation problem of target-driven mobile robot.Under the framework of deep reinforcement learning,a moving target navigation algorithm based on multi-agent cooperative control-(Teamwork-Net)algorithm is proposed,which realizes the autonomous navigation of mobile robot to dynamic target in natural scene.The convergence and reliability of deep reinforcement learning algorithm in dynamic target navigation are improved.The main work is as follows:(1)Experimental research on static target navigation of mobile robot based on deep reinforcement learning.Under the framework of deep reinforcement learning theory,the mapping of the original image to the optimal action of the mobile robot is realized by using the depth Q network(DQN).As reinforcement learning requires a lot of training process,it is difficult to achieve in the actual robot navigation scene.In this paper,we first train the deep Q network in the Gazebo simulation environment,and then apply the trained DQN network to the actual mobile robot static target navigation scene.In the indoor natural scene,the static target navigation experiments of mobile robot under various challenging conditions are completed by using the trained DQN network,including the sudden change of target position,environment,kidnapping robot and so on.The experimental results show that the proposed method can make the mobile robot system only through visual information without map.Complete the collision-free autonomous navigation movement from the starting position to the static target object.(2)Aiming at the problem of moving target navigation,under the framework of deep reinforcement learning,a moving target navigation algorithm based on multi-agent cooperative control-(Teamwork-Net)algorithm is proposed,including the network structure of multi-agent cooperative approximation navigation strategy and the training algorithm for the network.First of all,two agents are constructed to control the speed and direction of the navigation robot,respectively.Both agents regard motion control as a standard reinforcement learning problem,interact with the environment,and use the value function method to maximize the reward.Then,in order to improve the convergence and training efficiency of the algorithm,a segmented reinforcement learning training method for multi-agent cooperative control is proposed based on the above network structure.Finally,in the indoor natural scene,the dynamic target navigation experiment of multiple groups of mobile robots without map is completed by using TurtleBot2 robot.The experimental results show that the proposed Teamwork-Net algorithm can use vision safely to avoid obstacles and efficiently explore the environment to find targets when the navigation robot does not find the target.When the target appears in the field of vision,the navigation robot can accurately identify the target object,analyze the motion intention(direction,position,etc.)of the target,and successfully navigate to a moving target with anti-tracking ability.(3)In order to solve the problem of realistic gap in the process of migration,the feature extraction results of Teamwork-Net control network are visualized,and the feature extraction results of simulated environment and real environment are compared.The comparison results show that the trained Teamwork-Net network model can be directly applied in the actual environment.