A semantic-guided soft masked mixture of experts framework for interpretable multi-AUV control

Cooperative autonomous underwater vehicles (AUVs) have emerged as promising tools for underwater data collection missions. Using multi-agent reinforcement learning (MARL) as a core control approach, multi-AUV systems can adapt to dynamic task environments and autonomously cooperate to enhance mission efficiency. However, the underwater environment is dynamic and operationally costly. Once deployed, autonomous systems operate beyond the scope of human intervention, and any malfunction can lead to irretrievable consequences. Therefore, despite advances in multi-AUV algorithms, their black-box nature hinders practical deployment in real-world underwater missions. To reduce this gap, we propose a Semantic-Guided Soft Masked Mixture of Experts (SM-MoE) framework to enhance the interpretability of cooperative AUV MARL. The proposed framework semantically incorporates expert knowledge into the reinforcement learning network, enabling the learning process to accommodate task-decomposition-level interpretability. It also focuses learning on relevant maneuvering modes, thereby improving learning efficiency. Furthermore, by hierarchically adapting Expected Gradients SHAP (EG-SHAP) analysis to the proposed framework, we are able to explain the learned strategies at different resolutions that align with human cognition patterns. Evaluations based on the Nomoto model simulation demonstrate that the proposed method not only extracts systematic rationales from the learned policies, but also consistently surpasses the baseline methods in both control performance and quantitative explanation stability. • A semantic-guided SM-MoE framework is proposed for multi-AUV control. • Soft feature masks reduce observation interference among expert policies. • Hierarchical EG-SHAP explains expert routing and action selection. • SM-MoE improves data collection, energy efficiency, and collision safety. • Quantitative metrics show more faithful hierarchical policy explanations.