Research and Exploration of Deep Learning Based Reinforcement Process for Robot Co-Navigation Methods

This paper describes the details of the GDRN algorithm, including goal assignment, network structure, state space, action space and reward function. The role of goal assignment is twofold, one is to facilitate the calculation of personalized rewards, and the other is to adjust the state information, both of which aim at allowing the model to learn the navigation strategy faster. The attention layer in the network structure, according to the distance between the robots, adaptively interacts with the feature information, so as to make better decisions. The state information of each robot obeys the same distribution after adjustment, so that their exploration experience can benefit the model to learn the navigation strategy and improve the exploration efficiency. The robot's action space is designed as discrete values, which aims to reduce the complexity of the algorithm and make it easier to converge. The reward function takes into account the yaw angle and distance between the robot and the target, which effectively guides the robot to adjust the yaw angle toward the target, thus realizing more accurate and stable driving. The research in this paper combines the basic process of deep learning and deepens the cooperative navigation method for robots at the same time. At the same time, the two are fused to make a contribution to the research of new cross-cutting areas. Some basic contributions and theoretical guidance are provided for the synthesis study. In this paper, the centralized training centralized execution method is adopted. The centralized training allows the central controller to obtain the local observation information of all robots, which can be used to alleviate the environmental instability problem during the training process, and enable the network to learn a stable global collaboration strategy. It improves the training efficiency to a certain extent, and at the same time enables the network to learn a stable global collaboration strategy.