Efficient Distributed Inference and Collaborative Optimization for Unmanned Clusters

To address the distributed inference bottlenecks faced by unmanned clusters in dynamic heterogeneous environments, this paper proposes an efficient optimization framework that integrates adaptive fine-grained model segmentation, multi-level resource coordination scheduling, and multi-modal perception decision-making. First, based on a directed acyclic graph (DAG), we design an adaptive segmentation algorithm driven by maximum flow and minimum cut to dynamically optimize end-to-end latency and balance computation-communication load. Second, a decoupled resource scheduling mechanism is proposed based on the Cauchy inequality to achieve joint optimization of task offloading and computational/communication resources. Finally, multi-modal data fusion is employed to enhance environmental perception robustness. Validation on heterogeneous unmanned vehicles swarm demonstrates that the framework significantly improves model inference throughput, the adaptive segmentation scheme reduces communication overhead compared to random segmentation, and effectively lowers system energy consumption and latency, providing critical technical support for real-time swarm intelligent tasks.