To enhance fire safety, the rescue bottlenecks in complex fire scenes were addressed, including perception lag, high risks of close-range operations, and weak coordination among heterogeneous equipment. The obstacles to migrating general AI into firefighting scenarios were also considered, such as poor environmental adaptability, deficient dynamic scheduling, and ambiguous human-machine collaboration. Bibliometric analysis and systematic review were jointly employed. The evolutionary divergence between the general field of AI-enabled unmanned technologies and research related to unmanned firefighting from 2016 to 2025 was thereby characterized. The results showed that the annual publication volume in the general AI-enabled unmanned technology domain far exceeded that in the firefighting subdomain. A significant gap in technological maturity was observed between the two. The critical constraints of complex fire rescue had shifted from single-platform performance limitations toward multi-objective optimization, collaborative control under intermittent communication, and trustworthy decision governance. From a systems engineering perspective, a five-layer eight-step unmanned architecture for complex fire scenes was proposed. Five layers were defined, namely terminal execution, edge intelligence, collaborative control, communication network, and command digital twin. An operational closed-loop process was constructed across these layers, covering perception, understanding, decision, transmission, execution, evaluation, learning, and optimization. The proposed architecture markedly outperformed the conventional cloud-terminal paradigm in local safety protection, task continuity, and post-recovery rescheduling. On this basis, four technical directions, including multimodal anti-interference perception and edge inference under constrained computing resources, together with three categories of emerging research questions, such as the coupling of resource allocation and dynamic scheduling, were further distilled.
