Abstract: To address the challenges of discontinuous collaborative decision-making and dimensional mismatch in UAV swarms caused by the dynamic fragmentation of communication networks and the time-varying nature of battlefield entities in highly adversarial environments, a heterogeneous graph spatio-temporal reasoning (HG-STR) method is proposed. First, a local dynamic heterogeneous graph centered on individual UAVs is constructed. A meta-relation-driven heterogeneous graph Transformer is used to extract semantic topological features between UAVs, dynamic targets, and the search area. Temporal memory is constructed using gated recurrent units to compensate for decision-making oscillations caused by local observation interruptions. Second, a learnable attention communication mechanism is introduced to achieve adaptive filtering and high-confidence aggregation of key collaborative information under conditions of limited physical links and frequent network topology fragmentation. Finally, a hierarchical architecture of "upper-level tactical game—lower-level command execution" is established, and a pointer-based multi-head policy network is designed to solve the joint decision-making problem of variable-length object assignment and resource quantification allocation within a unified framework. A typical scenario for multi-area reconnaissance and strike missions was constructed. Simulation experiments show that the task completion rate is improved by 37.14% compared to traditional rule-based algorithms; compared to global optimization algorithms, the single-step decision-making time is reduced from seconds to milliseconds; and a 94% task success rate is maintained even under weak connectivity conditions with extremely limited communication radius.

Key words: UAV swarm, heterogeneous graph transformer, multi-agent reinforcement learning, hierarchical decision-making, distributed collaboration, reconnaissance and strike collaboration