In order to mitigate the decision-making myopia of unmanned surface vehicle (USV) under partial environmental observability and the high computational complexity of existing sequential models, a deep reinforcement learning approach incorporating the selective state space model (Mamba) and proximal policy optimization (PPO) was used to investigate the dynamic obstacle avoidance and path planning problem in complex multi-obstacle environments. Specifically, a sliding window feature extraction mechanism and a multi-dimensional coupled reward function were formulated and integrated into this framework. The results show that a breakout success rate of 36.8% is achieved by the proposed Mamba-PPO algorithm in extreme high-density dynamic encounter scenarios. Consequently, traditional self-attention and long short-term memory (LSTM) network models are significantly outperformed. Compared to traditional self-attention architectures, the number of model parameters is reduced by 14.7%. Meanwhile, a single-step network inference latency of 0.3458 ms and a closed-loop system throughput of 46 frames per second are realized. In robustness tests with injected Gaussian noise, the success rate is marginally decreased by only 2.3 percentage points. Furthermore, the optimal trade-off between perception depth and computational efficiency is verified by ablation studies on sequence length. Exploratory manifold analysis of the latent feature space indicates that the feature collapse and state entanglement problems of traditional sequential models are effectively alleviated. Accordingly, superior feature clustering capabilities for complex spatio-temporal sequences are demonstrated by this architecture. It is concluded that the short-sighted defects in partially observable environments are effectively mitigated by this method. Simultaneously, the low-latency requirements of edge computing platforms are adequately fulfilled. Therefore, a highly efficient paradigm is provided for the real-time dynamic obstacle avoidance of intelligent ships.