<p>Social navigation in dense human crowds requires collision-free and socially compliant behaviors under highly uncertain pedestrian dynamics. However, conflicting optimization objectives prevent existing Deep reinforcement learning (DRL) based methods from achieving stable, efficient, and socially compliant navigation. To address these challenges, we propose ARPL, an attraction-repulsion guided stabilized policy learning framework for social robot navigation that jointly models goal-reaching, collision avoidance, and social interaction. ARPL consists of three tightly coupled components: (1) a hierarchical self-attention spatiotemporal recurrent neural network (HSA-SRNN), which captures the dynamics of human-robot interaction in sequential observations to improve the model’s perception capabilities; (2) an adaptive stable proximal policy optimization (AS-PPO) algorithm, which improves convergence and robustness through a multi-level adaptive stabilization mechanism; and (3) an attraction-repulsion reward function (ARRF), which constructs a potential-guided reward function to encourage the robot to reach the target point. Extensive experiments on the CrowdNav++ benchmark demonstrate that ARPL achieves a 98.40% success rate, outperforming state-of-the-art methods by approximately 2%, while reducing collision counts by a factor of three and improving navigation efficiency. ARPL further demonstrates strong robustness across diverse out-of-distribution scenarios, and real-world experiments on a ROS mobile robot platform validate safe, stable, and socially compliant navigation in both sparse and dense crowds.</p>