Robot Visual Navigation Algorithm Based On Deep Reinforcement Learning

With the development of deep reinforcement learning(DRL),DRL have attracted the attention of researchers with the end-to-end training and automatic feature extraction.Researchers utilize DRL to solve navigation in unknown complex environments with the advantage that DRL learn navigation polices directly from sensory data without needing to construct a map.There are also challenges for learning to navigate with DRL,such as sparse rewards,low sample efficiency and difficulty in generalization.Novel model structures are designed in the thesis based on the prior knowledge for the different navigation task to improve the performance and efficiency.The main contributions of this article are follows:1.Aiming at the navigation task in the mapless environment,a deep reinforcement learning model with the self-localization of the agent and the auxiliary task of depth prediction is proposed for autonomous navigation.This model draws on the prior knowledge that the prevent collisions and the self-locate in planning.Two auxiliary tasks of positioning and depth prediction are designed to facilitate feature extraction,combining long-term motion planning with short-term motion planning,improves the navigation efficiency of the agent.2.Aiming at the generalization ability of navigation under random maps,a deep reinforcement learning algorithm model based on VIN(Value Iteration Network)and positioning module is proposed in the thesis.The VIN module is a differentiable neural network module that transforms the map input into the value distribution corresponding to the map through convolution iteration.This model transforms the environment map to value distribution corresponding to the map by means of convolution iteration,extracts features through the attention mechanism,and guides the agent to navigate with the input from the first perspective.3.For navigation tasks,maps with random structures of different scales are created as data sets.the text construction of a random map is implemented in the thesis through the depth-first search algorithm,and uses the lua language to transform the text map into a 3D environment based on the Deep Mind Lab environment,realizes the creation of the map environment.The experiment is evaluated on this basis and compared with the benchmark model,at the same time through the trajectory visualization,feature visualization and other methods to verify the effect of the model...