Research On Target Navigation Of Mobile Robot Based On Deep Reinforcement Learning

In the past two decades,robots have become more and more common in many human activities.The mobile robot has the ability to perform tasks such as reliably searching in an indoor environment,avoiding obstacles,and reaching an arbitrary object.Since most navigation methods require a priori maps,when the mobile robot faces scenes that are impossible to get a map,such as a fire scene,an earthquake scene,or an outdoor scene,its navigation capabilities are greatly limited.In order to improve the navigation ability of mobile robots without a priori map,this paper implements a mobile robot navigation framework with deep reinforcement learning as the core.Recently,with the rise of deep reinforcement learning models,robot navigation based on this method has attracted wide attention.In deep reinforcement learning,the robot gains navigation capabilities by interacting with the environment,ie by performing the actions with the greatest rewards in the environment.Deep reinforcement learning receives rewards while performing actions,and promotes positive rewards when completing mission objectives.Otherwise,it gives negative returns,and trains by repeating the process of excellence and non-stop.Since deep reinforcement learning requires repeated execution tasks during the training process,which is time consuming and easily damages the robot,it is not feasible to train the robot directly in the actual environment.The main step in the training of robots in this paper is to train the virtual model of the real robot in the virtual environment until it learns the required capabilities and then migrate the knowledge to the real robot in the real environment.However,when the navigation algorithm trained in the virtual environment is migrated to the real environment,the navigation performance of the robot is drastically reduced because the virtual environment is too different from the real environment.In order to solve the problem of poor generalization ability from the virtual environment to the real environment,this paper proposes a knowledge-based preprocessing layer and deep reinforcement learning combination algorithm model,which alleviates the performance fault problem caused by the virtual environment transplantation algorithm to the real environment.And the problems caused by the difference in performance between virtual sensors and real-world sensors.In the robot navigation experiment,the pre-processing layer and depth-enhanced learning module are designed and deployed in detail for the proposed robot navigation task.The robot can have navigation ability in the virtual environment,and the obtained navigation ability is transferred to the actual environment,and the navigation experiment results are analyzed.It is finally proved that the preprocessing layer and deep reinforcement learning algorithm can alleviate the performance fault problem caused by the virtual environment transplantation algorithm to the real environment,and has certain obstacle avoidance ability and obstacle avoidance ability without a priori map.