Inside China's HG-STR: The Algorithm Designed to Hunt Without Human Control

A team of Chinese researchers has published details of a new algorithm that pushes autonomous warfare a critical step further. Named HG-STR (Heterogeneous Graph Spatio-Temporal Reasoning), the system is designed to let a swarm of drones hunt and destroy targets without any human in the loop, even under heavy electronic jamming. The claim that it achieved a 100% kill rate in simulations has drawn global attention, but the real significance lies in its technical approach and what it signals about the future of conflict.

How HG-STR Works: A Decentralized, Graph-Based Brain

The core innovation of HG-STR is its ability to function when everything fails—when GPS is dead, radio links are jammed, and a human commander is unreachable. Traditional drone swarms rely on a central controller or constant communication between units. HG-STR was built for a far more hostile environment .

The algorithm treats the battlefield as a complex, dynamic web. It constructs a heterogeneous graph where every entity—friendly drones, enemy vehicles, terrain features, buildings—is a node. The connections between them represent real-time relationships. As drones move and targets shift, a spatio-temporal reasoning engine continuously updates this model .

Crucially, each drone runs this reasoning locally, not on a remote server. When a drone loses contact with the swarm, it uses its last known shared model to infer what its teammates are doing and what targets remain. This enables coordinated behavior—searching, classifying, and striking—without a single point of failure . A key technical paper describes the method as using a "meta-relation-driven heterogeneous graph Transformer" to extract the relevant features between drones, targets, and the search environment, while "temporal memory" handles the time-varying nature of the battle .

In practical terms, the system autonomously classifies every object it sees as friend, foe, or terrain, then decides to attack without waiting for human confirmation. The South China Morning Post reported that the algorithm identifies objects using onboard sensors and that the swarm can "hunt and destroy enemy targets completely autonomously" .

The Strategic Drive: Lessons From Ukraine, Applied to Taiwan

HG-STR did not emerge in a vacuum. It is the latest product of an intensified Chinese military push to make operational drone swarms a reality—a push that draws directly on combat data from the war in Ukraine.

The conflict there has revealed a brutal truth about modern drone warfare: communication links are a fatal vulnerability. One analysis noted that approximately 90% of Russian unmanned aerial vehicles were reportedly neutralized by Ukrainian electronic warfare at certain stages of the fighting . Mass FPV (first-person view) drone attacks proved devastating against armor, but their effectiveness depended on reliable control links. When those links were cut, the drones became useless.

Chinese military planners have absorbed this lesson. A Georgetown University study describes a "de facto division of labor" in which "Russia experiments with saturation warfare using cheap drones on the battlefield, while China systematically transforms those battlefield lessons into an industrial-scale production and innovation pipeline" . The HG-STR algorithm directly answers the electronic warfare problem by removing the need for a control link altogether.

This algorithmic work sits alongside a broader hardware push. In January 2026, PLA state television showed a single soldier controlling more than 200 drones launched from a ground vehicle . Two months later, the Atlas system was demonstrated in a full combat cycle, with one command vehicle directing 96 drones through autonomous search, targeting, and strike phases . China has also flight-tested a "drone mothership"—the Jiu Tian, a 25-meter wingspan UAV capable of releasing 100 to 150 smaller loitering munitions .

A 2026 CNA report identified the specific strategic purpose: the PLA is developing these swarm technologies to solve what it considers its most difficult operational challenge—a potential amphibious invasion of Taiwan. The envisioned use is for suppression of air defenses, saturation strikes, and reconnaissance .

The Legal Vacuum and the Human-Off-the-Loop Problem

HG-STR’s most profound implication is not technical but legal. The algorithm is explicitly designed to operate when a human cannot intervene. Once a swarm is launched with HG-STR, there is no veto, no oversight, and no pause button. Targeting decisions—who lives and who dies—are made by each drone’s local AI .

This creates a fundamental conflict with international humanitarian law, which is built on human accountability. The principle of distinction requires combatants to differentiate between military targets and civilians. Simulations are clean; real battlefields are not. Civilian vehicles, irregular fighters, and infrastructure near military targets all create classification challenges that AI systems are known to fail at. The risk of unlawful attacks on civilians is not hypothetical .

The principle of proportionality—weighing military advantage against expected civilian harm—is a contextual, human judgment that no current algorithm can replicate. And if a swarm commits a war crime, whom do you hold responsible? The commander who launched it? The programmers who wrote the code? Under existing frameworks, the chain of accountability breaks when lethal decisions are fully automated.

Crucially, no binding international treaty regulates lethal autonomous weapons systems (LAWS). Discussions at the UN Convention on Certain Conventional Weapons have continued for years without producing enforceable rules. HG-STR and systems like it are not waiting for a diplomatic consensus. As the Diplomat reported, PLA-linked research indicates a deliberate push to develop these swarms "specifically for urban warfare, while relying on the still ambiguous legal framework" .

The technology is advancing faster than the law. HG-STR’s 100% kill rate has only been demonstrated in simulation, not in the chaos of a real battlefield. But its existence makes clear that the era of fully autonomous lethal swarms is not a distant future scenario—it is an active engineering project.