A research team led by Associate Professor Zhang Dong from the School of Astronautics at Northwestern Polytechnical University (NPU) in China has recently introduced a new algorithm that could fundamentally transform the way unmanned aerial vehicle (UAV) swarms search for and destroy hostile targets.
The algorithm, known as HG-STR (Heterogeneous Graph Spatio-Temporal Reasoning), enables a formation of fixed-wing UAVs to autonomously search for and eliminate hostile targets across a large battlefield, even when communication signals are jammed and the line of sight is obstructed.
According to a peer-reviewed paper published on May 19 in Acta Aeronautica et Astronautica Sinica, one of China’s leading journals in the field of aeronautics, HG-STR achieved a 96% mission success rate and a 100% target elimination rate, significantly outperforming existing benchmark algorithms in both key performance indicators. South China Morning Post also reported that this is currently the first known algorithm capable of achieving a 100% target destruction rate while operating at a speed sufficient for the tempo of modern warfare.
The report further cited an anonymous Chinese defense expert, who noted that most UAV combat missions are still remotely controlled by human operators. According to the expert, this technology points toward a future in which UAV swarms can be deployed into highly contested and communication-denied environments, where they execute a single high-level command, such as search for and eliminate all hostile targets, without requiring further human intervention.
Conventional algorithms generally treat friendly forces, enemy units, terrain, and other battlefield entities as homogeneous data, which can lead to information ambiguity. In contrast, the research team proposed a heterogeneous graph representation that assigns distinct semantic meanings to different battlefield entities. In this framework, friendly UAVs, search regions, and hostile targets are modeled as different categories of nodes.
By learning the semantic relationships among these heterogeneous entities, the algorithm can accurately identify and prioritize meaningful interactions. For example, when a UAV detects a hostile target, the information is immediately recognized as a high-priority threat. Conversely, nearby friendly UAVs are identified as opportunities for cooperative engagement.
According to Zhang Dong and his colleagues, this capability enables UAV swarms to rapidly determine when to support allied units and when to pursue and engage enemy targets.
Traditional rule-based systems, by comparison, operate according to predefined scripts and often fail when facing unpredictable adversaries. Many existing optimization algorithms also suffer from excessive computational complexity, similar to a chess engine exhaustively evaluating possible moves, making them too slow for real-time combat scenarios.
The paper reports that HG-STR performs end-to-end neural network inference with an average decision time of only 6.6 milliseconds per step, representing a substantial improvement over conventional approaches.
For comparison, another method, GA+PSO+MPC, achieves slightly higher area coverage but requires approximately 5.8 seconds for each decision. At a flight speed of 100 meters per second, such latency would result in nearly 600 meters of blind flight before a new decision is made, which could be fatal in electronically contested environments.
Real-world battlefields present numerous uncertainties. Electronic warfare may disrupt communication links, isolating UAVs from one another. Individual UAV sensors have only limited observation ranges, while search-and-destroy missions are constrained by both time and fuel consumption.
To address these challenges, the research team developed several complementary solutions. First, each UAV is equipped with a memory mechanism based on a Gated Recurrent Unit (GRU). When communication with teammates is lost, the UAV can recall the last known positions of friendly units as well as the last observed locations of hostile targets.
Second, the researchers designed a hierarchical decision-making framework. Each UAV first determines its high-level objective, deciding whether to continue searching or transition into attack mode. It then selects an appropriate target and finally determines the amount of ammunition required for the engagement.
By decomposing the complex decision-making process into a sequence of hierarchical subtasks, the UAV swarm avoids the confusion and inefficiency associated with attempting to optimize all decisions simultaneously.
Simulation experiments demonstrated that the proposed algorithm maintained a 94% mission success rate even under weakly connected communication conditions with severely limited communication ranges.
The experimental results also indicate that HG-STR possesses strong generalization capability. After being trained in relatively small-scale simulation environments, the algorithm can be directly transferred to much larger and more complex battlefield scenarios involving greater numbers of UAVs and hostile targets without requiring additional retraining.
Looking ahead, the research team plans to transition the algorithm from laboratory simulations to real-world battlefield applications. Future research will further enhance the algorithm’s robustness, improving its resilience not only to communication interruptions but also to transmission delays, packet loss, and corrupted data.
The research paper concludes that future work will focus on lightweight deployment and real-flight validation of the algorithm on computationally constrained embedded airborne platforms. It also proposes incorporating channel contention, random packet loss, and communication latency into the Markov decision process state space to further investigate the resilience of UAV swarm decision-making under non-ideal communication conditions.
Source: guancha, sina, hk01, sohu
