<p>Deep reinforcement learning (DRL) has demonstrated significant potential for autonomous navigation applications. Nevertheless, traditional prioritized experience replay (PER) techniques are often limited by biased sampling and the accumulation of low-quality experiences in a single replay buffer. To address these limitations, we introduce a dual-priority experience replay (DPER) framework that employs a decoupled dual-priority architecture. This approach divides experiences into high-priority and low-priority buffers according to a defined temporal difference (TD) error threshold. This mechanism effectively reduces redundancy arising from prolonged storage of high TD-error samples. Furthermore, a dynamic hybrid sampling strategy adaptively balances retrieval from the two buffers, while a novel suppressive importance weighting strategy reduces bias and filters low-quality samples by simultaneously penalizing sampling frequency and importance weight. Experimental results show that the proposed algorithm demonstrates notable improvements in network training, optimizing trajectory quality, and enhancing overall model generalization.</p>