最新刊期
卷 7 ， 期 3 ， 2025
- 摘要：Using the World Robot Conference and Humanoid Games as a case, we outline a path from humanoids to robotoids as digital or artificial employees. Through the ACP loop of parallel intelligence, simulation and digital twins are fused with real-world feedback for prescriptive and continual learning, and eventually a new intelligent ecology of humans, robots, and digital agents (or robotoids) for the era of intelligent industry and smart society.关键词：embodied intelligence;humanoid robot;parallel intelligence476|138|0更新时间：2025-09-15
Review Intelligence
- 摘要：Vision-language-action (VLA) models are a comprehensive modeling approach for embodied intelligence that integrates visual perception, natural language understanding, and action execution within a unified framework, aiming to establish a continuous loop from environmental perception to task planning and action control. Their operational logic corresponds closely to the paradigm of parallel intelligence articulated in the early 21st century. That paradigm comprises artificial systems, computational experiments, and parallel execution, emphasizing virtual modeling, reproducible inference, and closed-loop interaction between the virtual and the real. The initial stage of VLA development, driven by multimodal deep learning, can be regarded as prototypical work within artificial systems, the subsequent stage characterized by large-scale models and cross-domain training expanded the scope of computational experiments, and the more recent focus on hierarchical control and virtual-real closed loops reflects the feedback correction and normative guidance emphasized in parallel execution. VLA models exhibit deep coupling between semantics and action, iterative cycles linking simulation with reality, and steadily improving verifiability. Nonetheless, challenges remain in generalization, semantic alignment, safety and interpretability, and deployment efficiency. Addressing these issues calls for contract-based task semantics, repairable long-horizon hierarchical planning, engineering-oriented use of world models, multi-level feedback and safety governance, and cross-platform transfer with human-machine collaboration. Examining VLA through the lens of parallel intelligence clarifies its developmental logic and provides methodological support for advancing toward trustworthy real-world applications.关键词：parallel intelligence;vision-language-action model;embodied intelligence;multimodal fusion;virtual-real interaction568|422|0更新时间：2025-09-15
- 摘要：For the chaotic behavior of nonlinear oscillations occurring in the control of permanent magnet synchronous motor (PMSM), a data-driven model-free reinforcement learning method was investigated, which only utilized the historical state data of the motor to obtain the optimal controller. This method solved the problem of uncertain external load and uncertain motor dynamic under varying and unknown working conditions. Firstly, to solve the uncertainty of external torque load, a zero-sum game between the controller and external disturbance was constructed. By redesigning the Riccati equation in iteration form, a model-based robust optimal controller was obtained. Based on this controller, by introducing the model-free reinforcement learning method, a data-driven stabilization method for chaotic phenomena in PMSM was proposed, which obtained a model-free robust optimal controller by learning from the historical operation data. Finally, the performance of the designed method was verified by comparing several simulations, and the results showed that the energy effort saving performance of our proposed method was improved by 39.04% compared with the traditional finite-time control method under the uncertainty of external load perturbation, and the success rate was improved by 10.71% compared with the linear quadratic regulator under the uncertainty of the motor model parameters.关键词：PMSM;chaos;reinforcement learning;zero-sum game;feedforward compensation controller113|135|0更新时间：2025-09-15
- 摘要：The popularity of social media has led to accelerated fake news propagation and expanded influence. The extensive spread of fake news not only disrupts social order but may also trigger mass incidents, posing a potential threat to national security and social stability. Consequently, the development of efficient fake news detection tools and techniques has become increasingly critical. To address this challenge, an evidence-aware Bayesian neural networks for fake news detection (EBNN-FND) method was proposed. This model quantifies uncertainties in both the detection model and the data, thereby improving the reliability of prediction results. The EBNN-FND model consists of four modules: a text embedding module, a feature processing module, a news-evidence interaction module, and a feature fusion module. Thereby, it can effectively integrate the features of news context and related evidence. Experiments on public datasets demonstrate that the EBNN-FND model significantly outperforms existing baseline models in fake news detection tasks, showcasing its efficiency and robustness. It not only provides a new research perspective for the field of rumor detection but also offers a viable technical solution to address uncertainty issues in information dissemination.关键词：Bayesian neural networks;variational inference;unbiased estimation;fake news detection178|166|0更新时间：2025-09-15
- 摘要：With the continuous advancement of electronic warfare and radar technologies, radar signal in battlefield environments has become increasingly complex and diverse, which poses severe challenges to signal deinterleaving. Traditional deinterleaving methods typically rely on prior knowledge and a preset number of categories, making them unsuitable for handling unknown signal sources and dynamically changing features that may arise in real-world applications. To address this issue, a complex unknown radar signal deinterleaving method based on selective deep embedding clustering was proposed. By integrating the feature extraction capabilities of deep learning with the flexibility of a density-based clustering algorithm, this method achieved adaptive embedding and clustering of radar signal parameters without relying on specific prior knowledge. Through automatic feature learning via neural networks coupled with a density-based clustering algorithm, the need for manually defined features with limited generalizability was avoided. Experimental results on radar signal datasets from complex environments indicate that the proposed method achieves a cluster purity of up to 99.46%, demonstrating robust adaptability and extensibility in identifying unknown or dynamically changing signal characteristics. Compared with traditional methods, selective deep embedding clustering provides a more universal and reliable solution for large-scale, complex radar signal processing and offers valuable references for other extensive data mining tasks involving unknown signals.关键词：radar signal deinterleaving;electronic reconnaissance;clustering;deep learning157|127|0更新时间：2025-09-15
- 摘要：To address the challenges of image complexity, limited textual information, and inter-modal misleading data in cross-modal emergency recognition, a cross-modal emergency recognition method that integrates attention collaboration and contrastive learning was proposed. First, a multi-branch self-attention mechanism was integrated into an image feature encoder, while a text feature encoder was augmented with external knowledge. These encoders were designed to extract rich semantic features from images and contextual information from text, respectively, to overcome the limitations of single-modal representations. Next, a cross-modal attention collaboration network was introduced to capture intricate relationships between image-text features, enhancing inter-modal consistency and mitigating the influence of misleading information. Finally, a contrastive learning mechanism, which was optimized through a joint supervision strategy, was employed to dynamically balance the contributions of image and text features. Extensive experiments conducted on the public CrisisMMD dataset and the self-constructed dataset demonstrate the superior performance of the proposed method.关键词：emergency recognition;cross-modal;attention collaboration;contrastive learning102|123|0更新时间：2025-09-15
- 摘要：In response to the issues of large parameter quantity and low accuracy in the existing dense pedestrian detection models, YOLOv7-TPMC (you only look once version 7-tiny partial minimum cross) based on YOLOv7-Tiny was proposed. Firstly, the ELAN-PP (efficient layer aggregation networks-partial and pointwise) module combining PConv (partial convolution) and PWConv (pointwise convolution) was proposed. Secondly, a cross-level fusion feature pyramid structure was designed to improve the accuracy of pedestrian detection in dense scenes through shallow feature reuse. Finally, in order to solve the problem that the CIOU (complete intersection over union) loss function was not accurate in positioning under certain conditions, the MF-IOU (minimum point distance and focaler-intersection over union) loss function was introduced. Experiments were carried out on the WiderPerson and CrowdHuman datasets. Compared with the baseline model, YOLOv7-TPMC had the number of parameters reduced by 20.5%, FPS (frame per second) improved by 12.2%, and the value of mAP@0.5 increased by 1.2% on the WiderPerson dataset. It increased by 2.0% on the CrowdHuman dataset, which can be well applied to dense pedestrian detection.关键词：pedestrian detection;YOLOv7-Tiny;FPN;partial convolution;attention94|111|0更新时间：2025-09-15
- 摘要：As a vast amount of information about Macrobrachium rosenbergii is available on the Internet, it makes the use of web searches or general-purpose large language model (LLM) inefficient to obtain accurate answers to aquaculture-specific questions. An intelligent question answering (QA) system for the Macrobrachium rosenbergii aquaculture field based on an entity-level attribute-concatenated vector retrieval strategy was proposed and implemented. A knowledge graph with LLM was integrated to support professional, user-friendly QA. Firstly, raw, unstructured aquaculture data was transformed into structured triples by an LLM-based knowledge extraction module to construct a knowledge graph. The entity names and their attributes were concatenated into semantically rich textual descriptions, which were vectorized and stored in a Neo4j database for similarity-based retrieval. Then, a question filtering module based on an LLM was introduced to optimize and classify user queries before vectorization, thereby enhancing answer robustness in noisy or ambiguous input environments. Finally, semantic vector retrieval was performed over the knowledge graph, and the retrieved knowledge was passed to an LLM for generating professional, context-aware responses. Experimental results demonstrate that the proposed approach effectively improves response accuracy and robustness. The system shows promise in enabling efficient, accurate, and user-friendly knowledge access in aquaculture, potentially improving decision-making and enhancing production outcomes.关键词：knowledge graph;large language model;Macrobrachium rosenbergii aquaculture;intelligent question answering177|118|0更新时间：2025-09-15
- 摘要：In response to the lack of interpretability in English composition automatic scoring systems due to their reliance on complex deep learning models, an interpretable English composition automatic scoring model was proposed based on the E5-SHAP algorithm. This model was based on the E5 base model encoder to extract text features, combined with a mean calculation and a regression layer to achieve scoring output. It introduced an adaptive weighting mechanism to comprehensively evaluate the quality of compositions across six dimensions, including grammar, syntax, and vocabulary diversity. The model utilized LoRA fine-tuning technology to optimize specific layer parameters and enhance adaptability to compositional features. By using the SHAP algorithm to calculate the impact of each feature on the final score, a clear scoring basis and explanation path was provided to enhance the transparency and credibility of the scoring process. The experimental results show that compared with existing models, the performance of this model has been improved on both the ELLIPSE dataset and self-built dataset, with a quadratic weighted kappa value (QWK) of 0.84, which is superior to existing models in accuracy and interpretability.关键词：English composition;automated essay scoring;E5-SHAP algorithm;interpretability119|113|0更新时间：2025-09-15
- 摘要：To address the issues of slow convergence and the difficulty of adapting to complex environments in proximal policy optimization (PPO) algorithms for quadrotor UAV flight control, an improved algorithm was proposed, KPPO, which integrated a composite regularization mechanism composed of a threshold-triggered KL divergence penalty and an L2 regularization term based on the PPO framework. The proposed KPPO algorithm was applied to quadrotor UAV flight control tasks. Physical simulation validation of quadrotor UAV flight control demonstrates that under the guidance of the KPPO strategy, the quadrotor UAV rapidly achieves policy convergence and makes correct decisions in complex environments, thereby significantly enhancing the performance of conventional algorithms. Notably, the KPPO algorithm improves task execution efficiency, a key factor in quadrotor UAV operations. This reassures the audience of the effectiveness of the KPPO algorithm in improving quadrotor UAV flight control.关键词：proximal policy optimization;quadrotor UAV flight control;deep reinforcement learning185|125|0更新时间：2025-09-15
- 摘要：With the rapid advancement of modern medicine and the increasing demand for clinical diagnosis and treatment, radiology, as a core department of clinical diagnostics, faces multiple challenges. These are particularly reflected in the complexity of patient scheduling across multi-device environments, workflow inefficiencies caused by manual operations, the lack of multimodal clinical information integration for diagnostic reasoning, and the time-consuming nature of report writing. To address these issues, RadiFlow, an end-to-end intelligent radiology workflow system based on Agentic AI, was proposed. The system was composed of four cooperative agents: scheduling agents, radiological examination agents, diagnostic reasoning agents, and report generation and interpretation agents. Through agents collaboration, RadiFlow can efficiently handle complex patient scheduling and cross-regional equipment allocation, achieve standardized and fault-tolerant imaging acquisition, provide accurate diagnostic assistance, and automatically generate and interpret standardized reports. However, this work remains a preliminary research report, validated through case studies. RadiFlow demonstrates significant potential for improving radiology efficiency, diagnostic accuracy, patient satisfaction, and reducing the workload of healthcare professionals. Future work will extend this framework to build a multi-agent radiology system for real-world deployment and validation, aiming to provide innovative insights into developing a more intelligent and efficient radiology practice.关键词：agent;radiology;artificial intelligence;large language model373|256|0更新时间：2025-09-15
Papers and Reports
- Postal code：100079
- Tel：（010）53879206 Email：cjist@bjxintong.com.cn
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
- Total Visits：161982 Visits Today：145
0