Breakthrough in unmanned swarm technology: SRI model breaks new ground in trajectory prediction and topology inference

Unmanned Swarm Systems, defined by their distributed coordination among multiple unmanned units, have emerged as a transformative force across diverse high-stakes fields. In disaster rescue, USS can cover large disaster zones to locate survivors or map hazardous areas; in transportation, they enable optimized fleet coordination for logistics or urban traffic management; in military operations, they support reconnaissance, surveillance, or coordinated mission execution. Despite these advancements, two core challenges have long constrained their broader real-world deployment: accurately predicting swarm trajectories and uncovering the underlying interaction mechanisms between swarm units. Existing methods often fall into one of two pitfalls: they either disregard the physical constraints (e.g., separation to avoid collision, cohesion to maintain swarm integrity) that govern real swarm motion, leading to unrealistic predictions, or they rely on "black box" models that lack interpretability, making it impossible to explain how or why units interact as observed. These limitations have become a bottleneck for achieving reliable autonomous collaboration in USS.

 

In a recent article published in the Chinese Journal of Aeronautics on July 22, 2025, Dr. Shuheng Yang and Prof. Dong Zhang from Northwestern Polytechnical University addressed these gaps by developing the Swarm Relational Inference (SRI) model—an unsupervised end-to-end framework that fuses swarm dynamics with dynamic graph neural networks. This integration not only resolves the flaws of existing methods but also sets a new standard for USS trajectory prediction and interaction analysis.

 

Core Innovations of the SRI Model

 

The core innovations of this model are truly groundbreaking and can be dissected as follows:

 

(1) Combined with the rule-guided swarm dynamics model, a physically interpretable swarm motion topology graph mapping is constructed. Classical rules such as separation and cohesion are encoded into explicit edge types and node states to improve the interpretability and physical consistency of the inference model and solve the defect that the traditional trajectory prediction model is difficult to explain the interaction mechanism by establishing the explicit association between the swarm motion constraints and the features of graph neural networks.

 

(2) A dynamic graph inference architecture based on the coupling of temporal state and relationship strength is proposed. Based on the improvement of NRI model, a multi-dimensional LSTM network including node state and interaction relationship is designed, and a relationship strength module is constructed by using the multi-head attention mechanism to realize the dynamic inference of the coupling of node dynamic features and relationship weights and improve the model’s ability to represent dynamic motion topology.

 

(3) Establish an unsupervised end-to-end prediction framework with joint optimization of interactions and trajectories. The motion topology inference and trajectory prediction are included in the unified optimization goal, and the backpropagation training strategy combining the topology graph reconstruction loss and the trajectory prediction error is designed, and the end-to-end trajectory prediction covering topology inference is completed under the condition that only the historical trajectory it input, and the motion topology is not labeled.

 

Experimental results show that the comprehensive performance of SRI in topology inference and trajectory prediction is significantly better than that of existing methods, and the long-term prediction error is reduced by 93.1% and 62.4% compared with the traditional LSTM model and the new dNRI model, respectively, which has better validity and robustness.

 

Future Directions of the Research

 

The team plans to enhance the model for scenarios with missing data, varying node numbers, and heterogeneous swarm interactions. In real-world operations, it's not uncommon for some units in a swarm to lose communication temporarily, resulting in missing data. The team wants to make the SRI model robust enough to handle such situations. Also, swarms can change in size—new units might be added or some might be lost. The model needs to adapt to these varying node numbers seamlessly. Heterogeneous swarm interactions, where the units in the swarm have different capabilities or roles, also pose a challenge that the team aims to tackle. They will explore hypergraph structures and multimodal relationships encoding.

 

The ultimate goal is to develop a universal tool for swarm intelligence analysis, enabling applications like enemy swarm behavior prediction in military contexts. In military operations, being able to predict how an enemy swarm of unmanned vehicles might behave can provide a significant tactical advantage. In disaster rescue, lost UAV trajectory reconstruction can help in finding missing drones that might have valuable data or be needed for continued rescue operations. And in civil fields, optimized vehicle fleet coordination can lead to more efficient transportation systems, reducing traffic jams and improving fuel efficiency.

 

Original Source

Shuheng Yang, Wenyi Liu, Dong Zhang, Shuo Tang, Motion Topology Inference and Trajectory Prediction Method for Unmanned Swarm System [J]. Chinese Journal of Aeronautics, 2025, https://doi.org/10.1016/j.cja.2025.103709.

 

About Chinese Journal of Aeronautics 

Chinese Journal of Aeronautics (CJA) is an open access, peer-reviewed international journal covering all aspects of aerospace engineering, monthly published by Elsevier. The Journal reports the scientific and technological achievements and frontiers in aeronautic engineering and astronautic engineering, in both theory and practice. CJA is indexed in SCI (IF = 5.7, Q1), EI, IAA, AJ, CSA, Scopus.

 

---

---