Artificial intelligence and robotics (AIR) have entered a pivotal stage, highlighted by the 2025 World Artificial Intelligence Conference and the World Humanoid Robot Games in China. These events showcased advances in humanoid robotics, parallel intelligence, and digital-twin experimentation, while also exposing hardware limits and governance challenges. Beyond competition, they underscored scenario-driven deployment in manufacturing, healthcare, and services, redirecting capital toward real-world applications. At the same time, AIR reframed public perception, positioning robots as partners rather than replacements, and raised urgent questions on ethics, policy, and social inclusion. The AIR era thus represents not just technological convergence but the emergence of a knowledge-driven industrial ecosystem that must remain human-centered and globally collaborative.

This paper proposes a cooperative game-theoretic framework for driver-automation interaction control. It features personalized assistance through online identification of driver steering characteristics. A two-point preview-based distributed model predictive control (MPC) strategy is developed to model the interactive behaviour between the human driver and the active front steering (AFS) system of a passenger car. The proposed strategy leverages cooperative game theory, formulating a joint optimisation of the steering actions of both the driver and the AFS based on Pareto equilibrium. To account for inter-driver variability in steering behaviour and deliver tailored assistance, a recursive least squares (RLS) algorithm with exponential forgetting is applied to identify driver steering characteristics online. The estimated driver steering characteristics are then used to compute the AFS steering angle that supports the driver's steering actions. Two numerical studies are carried out. The first study demonstrates the effectiveness of the Pareto-equilibrium steering strategies, incorporating the two-point-preview setup in achieving satisfactory vehicle path-tracking performance. The second study validates the proposed online identification scheme under different driver-imposed path-tracking penalties as well as different weight allocations between the driver and the AFS controller.

Large language models (LLMs) offer new opportunities in diabetes research by integrating contextual behavioral data with physiological biosignals, supporting more personalized care. This review surveys recent studies on the application of LLMs to interpret interactions among stress, physical activity, and glucose dynamics across type 1, type 2, gestational, and monogenic diabetes (MODY). We examine methods for capturing contextual data beyond traditional patient diaries, including wearable sensors, lifestyle logs, and digital health tools, and discuss how these data are combined with continuous glucose monitoring. After retrieving 39 relevant studies and finally retaining 25 after screening, we summarize the current capabilities, limitations, and clinical implications of LLM-assisted multimodal approaches in diabetes management. The findings highlight both the promise and challenges of applying LLMs to synthesize heterogeneous data, providing insights for future research on enhancing individualized and evidence-based diabetes care.

A cognitive model of "social voice" in sociolinguistic perception is proposed in this paper, which explores how language use reflects social identity and how listeners interpret these linguistic features to form social judgments. Moreover, social meaning involves the stereotypes or judgments people attach to certain ways of speaking. The social cognition processes how speakers perceive these linguistic cues to form those judgments. Linguistic features act as indexes of social categories. Context determines how these features are interpreted. This interplay between language, identity, and cognition reveals how societal norms, biases, and cognitive processes shape communication. Social voice in sociolinguistic perception reveals how language both reflects and reinforces social hierarchies. By understanding the cognitive mechanisms behind these processes, such as automatic stereotyping and contextual flexibility, we can challenge biases and promote equitable communication. Future research should prioritize intersectionality, technological ethics, and strategies to amplify marginalized voices, fostering a society in which linguistic diversity is celebrated rather than stigmatized.

This article provides an overview and assessment of commonly used evaluation methods for integrated energy systems over the past decade, and discusses the characteristics and applicability of currently used evaluation methods. By summarizing and collating common evaluation methods and types of issues evaluated for integrated energy systems, recommendations are provided for the selection of evaluation methods, and advantages, disadvantages, and applicable issues of all the methods are clarified. In addition, a general process is proposed for the evaluation of integrated energy systems. Finally, the above research is summarized and an outlook on the future development direction is presented.

The "on hold" event of IEEE Transactions on Intelligent Vehicles (TIV) is a typical case of misunderstanding of how a journal runs, and misleading public opinions from online media seriously harm a journal. Public opinions on this kind of case had never been investigated. The paper proposes a multidimensional quantitative regression framework (MQRF) that demonstrates how to quantify and predict public opinion influences for such cases through the collaboration of large language model and small model. The framework leverages large language models’ comprehensive analytical capabilities for contextual understanding while employing specialized small models for precise time-series prediction, creating a synergistic approach that significantly outperforms single-model solutions. We analyze the evolving relationships among public opinions and examine how these changes prompt the journal to disclose and clarify its operational practices. Public misunderstandings cause lasting damage to the journal. Understanding how different types of papers run in the journal can eliminate misunderstanding though with a time lag. Through in-depth analysis of the TIV event and comparative experiments with six small models (Linear, DLinear, NLinear, TimesNet, Transformer, and PatchTST) and six large language models (ChatGPT-4.5, Claude3.7sonnet, Grok3, Deepseek-R1, Doubao, and Qwen2.5-72B-Instruct), the effectiveness of the MQRF is verified.

With the rapid advancement of the automotive industry, the demand for vehicle-centric communication continues to grow. In vehicular ad hoc networks (VANETs), the interactions among vehicles can be modeled as a social network. This paper explores the social dynamics of such networks by analyzing the eigenvector centrality of vehicles and classifying communication levels based on their centrality rankings. To support this analysis, a communication system is designed using a multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing framework, incorporating the derived communication levels. Within this system, a particle swarm optimization (PSO) algorithm is employed to optimize a radial basis function (RBF) neural network for channel estimation. This approach significantly improves performance, achieving a bit error rate (BER) below 10⁻⁴ at relatively low signal-to-noise ratios (SNRs). Moreover, the proposed method enables the system to approach the theoretical channel capacity limit under low SNR conditions. The communication level detection method presented in this work also achieves 100% accuracy across various signal detection techniques. Overall, the proposed signal detection framework offers promising potential for enhancing the performance and reliability of future vehicular communication systems.

The vision-language-action (VLA) paradigm is gradually becoming the core path of embodied intelligence. However, its training and validation, which rely on simulation environments, face serious sim2real challenges, such as navigation deviations in drones caused by wind speed differences between simulation and real-world environments. Existing iterative methods based on digital twins can alleviate the problem of virtual-real alignment to some extent. However, their high dependence on twin consistency limits their adaptability and scalability in complex environments. To break through this bottleneck, the PiVLA framework is proposed in this letter to reconstruct the VLA paradigm with parallel intelligence. Furthermore, we introduce the parallel deep foundation model (PDFM) and, based on it, propose model parallel control (MPC) and the parallel interaction protocol (PIP), establishing a unified interaction mechanism for disembodied agents and embodied agents. This provides a scalable and robust solution for complex tasks involving embodied intelligence.

Large language models (LLMs) and their derivatives have demonstrated remarkable performance in natural language and downstream tasks. However, their static parameterization and reliance on public data limit real-time adaptability and the integration of private knowledge. Retrieval-augmented generation (RAG) partially addresses these issues by incorporating external knowledge, yet existing approaches primarily rely on public repositories and struggle to handle hybrid knowledge sources. To overcome these limitations, a unified parallel RAG-agent-LLM (PRALLM) framework is proposed in this paper. PRALLM introduces a three-layer knowledge architecture that seamlessly integrates public, private, and personal databases through a unified representation and retrieval mechanism. The framework further incorporates real-time learning (supporting incremental updates), directional learning (user-oriented knowledge acquisition), and personalized private agents (enabling secure domain adaptation). This design enables scalable, adaptive, and privacy-preserving knowledge-enhanced LLMs, effectively bridging the gap between the breadth of public knowledge and the depth of private knowledge in practical applications. Preliminary evaluations indicate that PRALLM is expected to achieve approximately 15%–25% performance improvement across various domain-specific tasks compared with the standard RAG approaches.