<p>With the broadening and the growing complication of robot applications, the importance of navigation technology is increasingly evident. However, the shortcomings of existing navigation methods in safety and social awareness obstruct the expansion of robots to environments with dynamic obstacles, especially those with pedestrians. In this paper, we propose a navigation strategy named DAPGRL (Dual Angle Pedestrian Grid Reinforcement Learning) to improve the navigation performance in dynamic environments from two aspects. First, we construct a dual-APG (Dual Angle Pedestrian Grid) structure to encode both the observed and predicted states, giving the robot better prediction and response abilities to dynamic obstacles. Furthermore, we design a social-aware reward consisting of sociability, comfort, and naturalness items, ensuring the robot’s navigation behavior is more friendly and more acceptable to pedestrians. In the experiments, with social force model and social-aware reinforcement learning as baselines, the proposed method was evaluated under challenging scenes with an increasing number of agents and a simulated hospital environment. The results show that the robot’s generalization ability, navigation performance, and social compliance have all improved. In quantitative experiments, the navigation success rate of the proposed method always surpasses that of the compared methods when there are over 3 agents in the scene. Moreover, it maintains a 94% success rate in situations with 11 agents. As for the qualitative experiment, our method shows a better obstacle avoidance strategy, a more comfortable distance, and a smoother trajectory. The proposed navigation scheme shows better adaptability and stability in pedestrian environments and is expected to be applied to various human-robot coexistence scenarios.</p>