Deep reinforcement learning (DRL) has emerged as a pivotal technology for enabling autonomous decision-making in unmanned swarms operating within complex and unstructured environments. Nevertheless, the inherent “black-box” nature and lack of interpretability of DRL models impede human understanding, trust, and oversight of agents’ autonomous behaviors, thereby significantly restricting their deployment in safety-critical applications. To address this challenge, this paper proposes an interpretable multi-agent reinforcement learning approach. Firstly, an interpretable Dirichlet variational autoencoder is designed to encode decision rationales as probability distributions in a latent space aligned with physically meaningful semantics. Secondly, a gating network is employed to generate action decisions by linearly combining the encoded rationales. Lastly, the interpretable autoencoder is integrated into and jointly trained within a multi-agent proximal policy optimization framework. This method represents the agent’s decision-making process as a mixture of probability distributions, each corresponding to interpretable physical semantics, thereby enabling intuitive human understanding of agent behavior through probability densities and allowing direct human intervention by adjusting the gating weights. Comparative simulation experiments validate the decision-making performance of the proposed approach, while the developed visualization techniques demonstrate both the interpretability of agent decisions and the efficacy of human intervention.