The trajectory planning for the glide phase of hypersonic glide vehicles (HGVs) inevitably involves multiple constraints—such as control limits, heat flux, overload, and dynamic pressure—as well as no-fly zones. Deep reinforcement learning has become a hot topic in trajectory planning for various types of aircraft. However, applying deep reinforcement learning to solve the multi-constraint, high-precision trajectory planning of HGVs during the glide phase faces challenges related to convergence. To address this issue, this paper proposes a Proximal Q-Value Automatic Curriculum Reinforcement Learning (PQAC-RL) algorithm. First, to address the difficulties in training and convergence when applying deep reinforcement learning to high-precision planning tasks, a dual reinforcement learning loop is designed within the algorithm, enabling the agent in the inner loop to adaptively select a curriculum under the guidance of the outer loop. Second, during the automatic curriculum selection, three specialized mechanisms—curriculum clipping, curriculum reviewing, and curriculum advancing—are designed to maintain training stability, consolidate historical experience, and accelerate the exploration process, respectively. Furthermore, to account for both no-fly zone avoidance and terminal precision, a hybrid reward mechanism is formulated. By establishing specific switching conditions, the algorithm can adaptively alternate between a guidance mode and an avoidance mode, effectively balancing HGV survivability and terminal accuracy. Experimental results demonstrate that, compared with direct training and existing curriculum learning methods, the proposed PQAC-RL significantly improves the convergence speed of the policy and the task success rate. Moreover, it exhibits superior trajectory planning capabilities in multi-obstacle environments, as well as excellent generalization in varying scenarios.

Coordinating multiple unmanned aerial vehicles (UAVs) for cooperative missions requires agents that perceive their environment, reason about objectives, and generate joint actions. Vision–language–action (VLA) models unify these capabilities but lack a principled source of multi-agent training data and suffer from a training–inference discrepancy in closed-loop control. We propose MiniUAV-VLA, a compact centralized VLA controller for simulated multi-UAV search-and-elimination based on multi-agent reinforcement learning (MARL) expert distillation. A QMIX expert policy achieving 100% mission success generates multimodal demonstrations pairing rendered tactical map images with structured textual state prompts. A 158 M-parameter VLA model with approximately 65 M trainable parameters in the MiniMind-3V backbone and vision projection is fine-tuned with a multi-agent discrete action head that jointly predicts actions for all UAVs in a single forward pass. We identify a training–inference feature mismatch in behavior cloning and address it via prompt-end action pooling, which extracts action-relevant hidden states at the user–prompt boundary rather than after the generated response. In closed-loop evaluation with four drones and six mobile targets averaged over five evaluation seeds, MiniUAV-VLA reaches 74.4 ± 4.6% mission success against 9.4 ± 2.1% for a random policy and 16.2 ± 3.2% for an observation-limited greedy baseline. Across five independent training runs, prompt-end action pooling improves mean closed-loop success from 40.6% to 76.2% over the last-token alternative. These results support MARL expert distillation as a data-efficient route to compact multi-agent VLA control in this simulated setting.

We present a high-resolution 3D facial dataset focused on East Asian participants, designed to provide consistent topology and expression diversity. The dataset includes 98 individuals (18-30 years) and captures neutral faces, six basic emotions, and nine Facial Action Coding System (FACS)-based Action Units (AUs). Each 3D mesh is acquired under a standardized protocol. Synchronized multi-view RGB images are available only for a subset of participants who provided additional consent. To ensure structural consistency across subjects and expressions, we developed a processing pipeline combining FLAME-based expression fitting with Edge-Constrained Non-rigid Iterative Closest Point (Edge-NICP). The resulting meshes share unified topology and vertex correspondence, enabling direct point-to-point comparisons across conditions. Rich annotations, including facial landmarks, verified AU labels, and vertex-level deformation fields, accompany the dataset. Together, these resources may facilitate research in computer graphics, human-computer interaction, and cross-modal studies that benefit from expression-rich, topology-standardized 3D face data.

Deep reinforcement learning (DRL) provides a powerful approach for autonomous decision-making behavior modeling in complex dynamic environments. However, its black-box nature restricts practical deployment, as interpretable and trustworthy models are often required in real-world applications. Existing interpretable methods often face a trade-off between transparency and decision performance. To address this limitation, this paper proposes a Clustering-Guided Fuzzy Inference System (CG-FIS) that leverages behavioral data collected from a trained DRL policy to derive inherently interpretable behavior models. Specifically, we integrate a density-based clustering approach (DBSCAN) to automatically identify underlying behavioral patterns within the data, from which fuzzy rules are systematically derived to support interpretable decision-making. The resulting CG-FIS provides semantic-level explanations through a transparent base of IF-THEN fuzzy rules. A modeling pipeline is further developed based on CG-FIS, incorporating state space refinement through Independent Component Analysis (ICA) to handle complex multidimensional tasks, along with iterative parameter tuning via Particle Swarm Optimization (PSO) to enhance the effectiveness of CG-FIS. The proposed method is validated on four standard RL benchmarks and a domain-specific case study, demonstrating superior decision performance compared to widely adopted interpretable behavior models. Experimental results demonstrate that CG-FIS generates both effective and interpretable fuzzy policies.

Aiming at the problems of weak target search capability, poor coordination, and limitations in the field of view and attack capabilities of single-domain platforms in unknown suburban battlefield environments, this study proposes an air-ground coordinative dynamic target search and strike strategy based on IWPA-HDMPC and call mechanism. Models of air-ground agents and battlefield environments are proposed while constructing environmental map to enhance agents’ perception and search representation. Under the framework of distributed model predictive control (DMPC), objective functions are established, and an improved wolf pack algorithm with hierarchical mechanism is introduced to accelerate solution efficiency and improve search benefits. At the same time, the algorithm satisfies communication and safety constraints. Considering agent heterogeneity, a joint mission model integrating target search and coordinative strike is designed to provide agents with a closed-loop search-decision-action process. In simulation instantiated with real-world regions, compared with the latest method, the proposed method improves static area coverage efficiency by 25.9

The paper explores the challenges of visual simultaneous localization and mapping (VSLAM) in highly dynamic environments, a capability crucial for applications such as autonomous driving and service robots. We propose semantic-integrated multi-model fitting (SMMF)-SLAMMOT, a tightly coupled VSLAM and moving object tracking (MOT) method, capable of simultaneously estimating the full SE(3) motions of a stereo camera and the surrounding moving rigid objects, without relying on geometric priors. The SMMF-SLAMMOT framework begins with a two-level dynamic data association technique, which leverages object embedding descriptors from a detector to enhance matching robustness in crowded scenes. Subsequently, a semantic-integrated multi-model fitting method is proposed to achieve more accurate and robust multiple motion segmentation and estimation. Furthermore, we devise a spatial–temporal reprojection factor to enhance the accuracy and efficiency of the 4D mapping. Evaluations on the OMD and KITTI Tracking datasets, along with a self-collected dataset from the CARLA simulator, demonstrate the superiority of SMMF-SLAMMOT in terms of accuracy of self-localization and moving object tracking, as well as real-time performance. Specifically, on the KITTI Tracking dataset, compared to state-of-the-art systems, our method achieves a median 14 https://github.com/zhangtiantians/SMMF_SLAMMOT . This work not only advances the VSLAM field but also provides practical solutions for real-world applications in dynamic scenes.

High-resolution digital elevation models (DEMs) are essential for accurate vehicle motion simulation, yet are costly to obtain and store. We propose a DEM super-resolution method based on a coordinate-based multilayer perceptron (MLP), which represents elevation as a continuous function over spatial coordinates. A hypernetwork predicts the parameters of the MLP from low-resolution inputs, enabling direct querying of elevation and slope values at arbitrary locations via automatic differentiation. To enhance terrain detail reconstruction, we incorporate periodic activation functions and terrain-aware loss functions, including slope and frequency-domain components. Experimental results show that the proposed method outperforms bicubic interpolation in both elevation and slope prediction, while reducing memory usage to just 14.9

To address the need for high-precision modeling in digital twin battlefields, this study proposes a six-degree-of-freedom digital twin model of a fixed-wing aircraft. The model adopts a modular and component-based architecture, enabling separate modeling of key components such as the fuselage and wings to enhance modeling granularity and flexibility. An improved Euler method is applied to increase the accuracy of dynamic simulation. Environmental response mechanisms—including standard atmosphere, icing, and rainfall models—are incorporated to dynamically adjust the aircraft’s characteristics under complex conditions. A genetic algorithm is employed to optimize the PID controller, enabling adaptive tuning with limited parameters. The proposed approach is innovative in its integration of fine-grained modeling with complex environmental responses, effectively addressing the limitations of traditional models in terms of low granularity and insufficient environmental representation. Simulation results demonstrate that the model exhibits strong stability and responsiveness, offering reliable support for digital twin modeling of battlefield equipment.

High-fidelity models can better meet the increasing requirements of contemporary flight dynamics simulation, making the assurance of their credibility a crucial issue. The credibility assessment of a simulation system must be conducted through Verification, Validation and Accreditation (VV A). Main validation methods are systematically analyzed. The specific implementation process of a high-fidelity flight dynamics model is investigated from three dimensions: validation items determination, simulation experiment design, and results validation. Model refinement is guided by simulation results analyses involving basic performance parameters, aerodynamic coefficients, and dynamic performance parameters, which provides reference for the validation and optimization of other simulation systems. It also supports the comprehensive application of validation methods and the acquisition of output data from simulation models.

The existing methods for air mission wargaming autonomous decision-making, such as optimization theory-based and expert systems, often suffer from insufficient real-time performance and high modeling workload. The application of multi-agent reinforcement learning (RL) in autonomous decision-making for air mission wargaming has received extensive attention. However, most existing RL approaches require online interaction, leading to low efficiency of sample collection and network training. This paper proposes BC-QMIX, the network structure of BC-QMIX employs supervised learning method to train a behavior cloning network for each sub-agent on the basis of the QMIX network, which provides a basis for action selection. Therefore, BC-QMIX alleviates the extrapolation error in the offline training of QMIX. Furthermore, when offline pre-training is conducted by utilizing domain knowledge-based samples, it accelerates the online training and convergence of network. Through experiments conducted in the Multi Drones Monitoring environment and two collaborative air mission wargaming scenarios, BC-QMIX demonstrates a significant reduction in extrapolation error and outperforms QMIX, MADDPG, and MATD3 in offline training. Specifically, it achieves a 10.1% improvement in average winning rate over QMIX, with this enhancement rising to 47.9% when domain knowledge is incorporated into the training process. This validation demonstrates the feasibility and advantage of constructing a collaborative autonomous decision-making model using BC-QMIX for air mission wargaming scenarios.

In order to improve the problem of low efficiency of tactical rules relying on manual summarization in traditional behavior modeling, an intelligent modeling method of air combat behavior based on LLM is proposed. Under our existing air combat simulation framework, the reasoning ability of DeepSeek-R1-Distill-Qwen-7B model is used to output structured coding to support the reading and operation of behavior model. RAG retrieval enhancement generation and thinking chain are used to improve the output effect of the model. In addition, the rerank secondary re-ranking method is introduced in the similarity retrieval stage, and an indicator for evaluating the generation effect of structured html coding text is proposed. The results show that this method has good accuracy when processing a certain amount of situation information, improves the construction efficiency, and the simulation results have verified the effectiveness of the method.

To address the low efficiency and long debugging cycle in modeling air combat tactics, this paper proposes an intelligent modeling method for air force tactical confrontation behaviors based on YOLO and CLIP. By integrating object detection and cross-modal semantic understanding techniques, the method achieves high-precision detection of key elements in hand-drawn tactical sketches, such as trajectory and phase number elements. It further extracts semantic relationships among these elements to automatically generate structured models of tactical behaviors. Experimental results demonstrate that the proposed method enables end-to-end tactical modeling from hand-drawn sketches. In simulation verification, the execution trajectories of both offensive and defensive sides align well with the intended tactics in the sketches. This approach provides an innovative solution for efficient transformation of expert knowledge and tactical validation.

Recent breakthroughs in artificial intelligence have revolutionized the automation of frame-field-driven quad mesh generation, a critical surface representation paradigm in computer-aided engineering. However, existing neural frame-field generation methods, limited by the orthogonality of fields, struggle to preserve the geometric fidelity as well as quad quality around sharp features. To address these limitations, we propose NeuralPoly, an intelligent non-orthogonal frame-field generation method. We design a poly-vector encoding of the non-orthogonal field to leverage the representation power of neural network in capturing geometric features without manual tuning. Furthermore, we introduce a Hessian-based neural weighting scheme that autonomously resolves ambiguous alignments in flat and spherical regions. We then incorporates the poly-vector encoding and the proposed weighting scheme into the loss functions of a unified neural network architecture consists of a SIREN module for neural implicit representation and a ResUNet module for field prediction. Finally, we compare our method with state-of-the-art techniques in field-guided quad mesh generation. Quantitative and qualitative evaluations demonstrate that our approach achieves superior performance in both geometric fidelity and quad mesh quality.

Motorcycle graph is widely adopted as an intermediate block in state-of-art semi-structured quad meshing methods. However, constructing and simplifying it on 3D triangle meshes still face challenges in performance and stability. To address these challenges, we present a novel motorcycle graph construction and simplification method for semi-structured quad mesh generation. First, we introduce a piecewise advancing algorithm on parameterized triangle meshes with specially designed data structures to ensure reliable and high-performing motorcycle graph tracing. Second, we enhance the existing zero-collapse procedure with non-intersecting paths creation and feature preserving for T-mesh simplification. Third, we integrate our motorcycle graph construction and simplification algorithm into the state-of-art semi-structured quad meshing pipeline. A comparison with typical state-of-art methods proves that our method can generate quad meshes with superior topological quality and feature preservation capability. We also conduct batch experiments to demonstrate the efficiency, robustness of the proposed method.

This study addresses key limitations in structured or hierarchical classification and pairwise comparison methods for analyzing map projections, which often fail to capture the intricate relationships among them. It introduces an innovative approach that automates and simplifies the approximation of complex results from an improved integration metric for measuring (dis)similarities between projections. This enhances the understanding of over 340 map projections comprehensively. The study combines formula-based and image-based methodologies within a hybrid sampling scheme, effectively handling complex integrations, particularly for projections with intricate formulas. Using NASA's G.Projector mapping software, an 87 GB image dataset is generated for 175 map projections, while an additional 170 projections are processed through a formula-based approach. The approximate integral calculations closely align with theoretical values, with an acceptable error margin of a few thousandths. To visually represent map projections in two-dimensional space, multiple dimensionality reduction techniques are employed, incorporating features such as distortions. The resulting quantitative metrics demonstrate that global, local, and category-based features are reasonably preserved. A clear and intuitive visual representation simplifies the complexity of map projection relationships, offering valuable insights. Additionally, an interactive web application prototype is developed to showcase the relationships among map projections. To the best of our knowledge, this research is the first to comprehensively evaluate such a large number of map projections using automated calculations combined with dimensionality reduction and visualization techniques. This methodology represents a significant advancement in cartography, providing a robust framework for comparing and analyzing map projections in practice.

The three-dimensional (3D) reconstruction of buildings using photogrammetric point clouds is important for many applications, ranging from digital city construction to urban energy consumption analysis. However, problems such as building complexity and point cloud flaws may lead to incorrect modeling, which will affect subsequent steps such as texture mapping. This paper introduces a pipeline for building surface reconstruction from photogrammetric point clouds, employing a hybrid method that combines connection evaluation and framework optimization. Firstly, the plane segmentation method divides building point clouds into several pieces, which is complemented by a proposed candidate plane generation method aimed at removing redundancies and merging similarities. Secondly, the improved connection evaluation method detects potential skeleton lines from different planes. Subsequently, a framework optimization method is introduced to select suitable undirected polygonal boundaries from planes, forming the basis for plane primitives. Finally, by triangulating all plane primitives and filling holes, a building surface polygonal model is generated. Experiments conducted on various building examples provide both qualitative and quantitative evidence that the proposed hybrid method outperforms many existing methods, including traditional methods and deep learning methods. Notably, the proposed method successfully reconstructs the main building structures and intricate details, which can be further used to generate textural models and semantic models. Experimental results validate that the proposed method can be used for the surface reconstruction from photogrammetric point clouds of planar buildings.

Relation extraction (RE) is a crucial step for knowledge graph construction, which aims to extract meaningful relations between entity pairs in plain texts. Very few works have been studied on Chinese relation extraction (CRE) in the military field. Moreover, recent deep neural network-based methods have achieved considerable performance but still suffer from three inherent limitations, including overlapping of entities, imbalanced data and the ambiguity. Therefore, this work investigates a novel Multi-Grained Lattice Transformer (MGLT), which leverages external information of lexicon and word sense tailored for CRE. In MGLT, self-matched lexicon words and related word senses are fused through a cross-transformer mechanism to alleviate the ambiguity in texts. The finally enriched sequence representation in MGLT captures the relatedness between the head entity and the tail one, which is helpful to alleviate the overlapping of entities. Experimental results on two benchmark datasets and a self-developed dataset constructed from online military news show that the proposed MGLT achieves state-of-the-art (SOTA) performance. Compared with other typical baselines, MGLT achieves better area under curve (AUC) and [Formula: see text]-score by up to 10.46% and 6.90%, respectively. We further demonstrate the effectiveness of using ensemble learning to fully exploit complementary information from multiple MGLT-based base learners to improve the overall performance for imbalanced data classification on the military dataset. Such results indicate that the proposed ensemble learning model is effective and robust to be applied in practical applications.

Reinforcement Learning (RL) has emerged as a promising data-driven solution for wargaming decision-making. However, two domain challenges still exist: (1) dealing with discrete-continuous hybrid wargaming control and (2) accelerating RL deployment with rich offline data. Existing RL methods fail to handle these two issues simultaneously, thereby we propose a novel offline RL method targeting hybrid action space. A new constrained action representation technique is developed to build a bidirectional mapping between the original hybrid action space and a latent space in a semantically consistent way. This allows learning a continuous latent policy with offline RL with better exploration feasibility and scalability and reconstructing it back to a needed hybrid policy. Critically, a novel offline RL optimization objective with adaptively adjusted constraints is designed to balance the alleviation and generalization of out-of-distribution actions. Our method demonstrates superior performance and generality across different tasks, particularly in typical realistic wargaming scenarios.