Abstract:In the fields of intelligent manufacturing, aerospace, and robotics, control systems often operate under unknown dynamics. This significantly limits the effectiveness of traditional model-based control methods. Reinforcement learning (RL), as a data-driven intelligent control approach, enables policy learning and optimization through interaction with the environment, showing great potential for solving optimal control problems in such model-unknown scenarios. This survey focuses on the issue of unknown dynamic models in continuous-time systems and first reviews the development of general reinforcement learning algorithms and their application in model-known scenarios through industrial examples and theoretical analysis methods. It also summarizes representative methods for model-unknown scenarios, such as model-based RL, off-policy integral RL, and Q-learning approaches. The survey also introduces Lyapunov-based theoretical analysis tools and important assumptions. It discusses cutting-edge topics such as RL under partial observability using large language models, safe RL, and stability and robustness enhanced RL, while highlighting the challenges faced by existing methods.

Abstract:With the increasing number of space activities, HVI (hypervelocity impacts) caused by space debris and micrometeoroids have become a major threat to the safety of spacecraft in orbit. Such collisions not only result in mechanical damage but also generate plasma, whose electromagnetic effects pose severe risks to highly integrated spacecraft electronic systems. A systematic review of plasma physical effects induced by hypervelocity impacts is provided. The review covered the mechanisms of plasma generation, kinetic characteristics, electromagnetic radiation, and induced discharge, encompassing both theoretical and experimental progress. Special emphasis was placed on the introduction of condensed-phase products (dust grains) in hypervelocity impacts and the resulting dusty plasma effects. This review aims to offer researchers in the field a comprehensive literature summary and to highlight key scientific questions and future research directions. Ultimately, it seeks to provide theoretical support for enhancing the survivability of spacecraft in orbit and for developing next-generation electromagnetic protection technologies.

Abstract:DFRC (dual-function radar-communication) is proposed to overcome spectrum conflicts, hardware redundancy, and electromagnetic compatibility bottlenecks inherent in traditional separated architectures through hardware resource sharing and isomorphic signal waveform design, whereby the integrated operational effectiveness and battlefield survivability of platforms are significantly enhanced. The evolution of DFRC technology from its conceptual inception, through architectural advancements, to system implementation was systematically reviewed. The DFRC waveform design methodologies based on mainstream signal schemes were analyzed emphatically, including linear frequency modulation, orthogonal frequency division multiplexing, and orthogonal time frequency space. Furthermore, various sensing-centric waveform design criteria were explored in depth, such as beampattern matching, Cramér-Rao bound minimization, information-theoretical design, and so on. The engineering roadmap from software-defined radio compatibility verification and airborne multimodal waveform fusion to multi-node and multi-domain cooperation was summarized, clearly illustrating the theoretical-to-practical transition of DFRC. Through presenting the complete technical evolution of the DFRC system from the conventional single-input single-output system to multiple-input multiple-output system, and then to the prototype demonstrations, this overview provides systematic theoretical guidance and practical references for future research and development of DFRC systems.

Abstract:In the current landscape of large-scale model training, the contradiction between the exponential growth of model parameters and the slow increase in GPU memory capacity has become increasingly prominent. Among memory optimization technologies, recomputation and computational offloading reduce GPU memory overhead by trading time for space. The development trends of recomputation and computational offloading are first analyzed in this article. Then, the hardware bandwidth bottlenecks and software ecosystem adaptation challenges faced by memory optimization are analyzed, with a focus on the heterogeneous architecture characteristics of domestic artificial intelligence platforms. It also delves into the memory optimization technologies for large model training on domestic platforms such as MT-3000, with the aim of providing technical references for large model training on domestic platforms.

Abstract:A thermal concentrator is a thermal functional device based on transformation thermotics, effective-medium theory, and scattering-cancellation principles. By tailoring the spatial distribution of thermal conductivity or geometric configurations, it efficiently concentrates large-scale heat flux into localized regions, enabling precise control of both steady-state and transient heat transport. With advances in materials science and manufacturing technologies, research on the thermal concentrator is moving from theoretical models toward engineering implementation, and it shows application potential in microelectronic cooling, thermoelectric energy harvesting, energy heating, and thermal therapy. This article systematically reviews the physical mechanisms, structural designs, and implementation pathways of the thermal concentrator, summarizes its development and representative works, compares the applicability and performance characteristics of different theoretical frameworks and configurations, and analyzes its technical advantages and engineering feasibility in typical application scenarios. Finally, this article discusses future trends of the thermal concentrator, including extensions to complex geometries, multiscale systems, emerging energy platforms, and extreme thermal environments.

Abstract:Swarms of maritime unmanned surface vehicles (USVs), as a core technology driving the development of marine intelligence, demonstrate significant application value in military reconnaissance, environmental monitoring, maritime search and rescue, and related fields. However, the inherent characteristics of the marine environment, including highly dynamic conditions, environmental uncertainties, and communication constraints, pose formidable challenges to achieving high-performance swarm control in maritime USVs. To address these challenges, recent research advances in this field are systematically reviewed. The characteristics of the marine environment are described, the domestic and international developments in USVs are summarized, and the core control requirements and key challenges in complex marine scenarios are analyzed. Furthermore, a comprehensive survey of three representative swarm control methodologies is presented: trajectory-based guidance control, path-based guidance control, and target-based guidance control. Finally, promising research directions and future development trends in maritime USV swarm control are discussed and proposed.

Abstract:Foundation models have become a focus in radar remote sensing intelligent interpretation due to their provision of universal and generalizable solutions. Significant progress has been achieved in both theoretical and applied aspects of radar remote sensing foundation models, making it imperative to systematically summarize current research advancements. In order to further advance the research on radar remote sensing foundation models, the concept, key technologies, and evaluation methods of foundation models was expounded. Besides, current research progress and application performance are reviewed, with representative approaches and typical instances summarized. In conclusion, discussions and future directions are highlighted from four perspectives: model architecture design, interpretability research, lightweight methods, and security assessment.

Abstract:In recent years, generative artificial intelligence is progressively introduced into the field of radio spectrum cognition due to its powerful capabilities in data distribution fitting, data generation, and data completion. Compared to conventional approaches rely on physical modeling, mathematical interpolation, and discriminative artificial intelligence techniques, generative AI has significantly enhanced the accuracy of radio spectrum cognition. This paper systematically reviewed the research progress of generative artificial intelligence in radio spectrum cognition, with a focused analysis on the technical principles, application scenarios, and representative works of different generative paradigms. The challenges faced by generative AI in spectrum cognition were further discussed, including scarce training data, limited generalization in unknown scenarios, and insufficient model interpretability. In the future, by cross-modal knowledge fusion, physics-informed embedding, and the establishment of a trustworthy assessment framework, generative artificial intelligence is expected to advance radio spectrum cognition toward high precision, robust generalization, and enhanced interpretability, thereby effectively supporting the efficient utilization of spectrum resources.

Abstract:To predict the boundary-layer transition location over a flat plate across varying Mach numbers, an efficient method is developed for small-sample settings. Flow-field disturbance datasets across multiple Mach numbers were generated using the nonlinear parabolized stability equations (NPSE), with Ma = 0.01 designated as the source domain and Ma = 0.1, 0.2, 0.4, 0.8 and 1.6 as target domains. The influence of Mach number variations on transition patterns was systematically analyzed. A convolutional neural network (CNN) model was employed to map flow field patterns to transition locations, incorporating a transfer learning strategy with progressive unfreezing and layer-wise learning rates. Results demonstrate that transfer learning significantly outperforms direct training: for Ma ≤ 0.4, only 1/10 of the target domain samples are required to achieve a mean absolute error below 2.04% of the average ground-truth value; for Ma ≥ 0.8, a progressive domain adaptation strategy controls the error within 6.19%. The approach enhances transition prediction under small-sample conditions and provides a reliable technical pathway for cross-condition flow modeling.

Abstract:DA (data assimilation) is a crucial technical method for improving the accuracy of atmospheric chemical forecasts by integrating the results of atmospheric chemistry models with multi-source observational data, reducing uncertainties in model input data. Centering on DA techniques for atmospheric chemistry models, the transformation process of initial field assimilation for pollutant gases and aerosols from single state variables to multi-state variables was systematically reviewed. Meanwhile, the important progress of pollutant emission source assimilation inversion using ensemble methods and four-dimensional variational methods was focus on the improvement of emission source accuracy, optimization of spatiotemporal resolution, and enhancement of pollutant concentration prediction performance. With the explosive growth of observational data, a core challenge in the current field lied in fully leveraging high-resolution geospatial and remote sensing data for atmospheric chemical DA. The deep integration of DA with artificial intelligence algorithms represented a key research direction to break through this bottleneck and significantly enhanced the accuracy of atmospheric composition analysis and forecasting.