For almost the past decade, face recognition has been dominated by convolutional neural networks. Margin-based losses (CosFace, ArcFace) have achieved over 99.6% verification accuracy on classic benchmarks. For a long time, further performance gains were believed to rely solely on larger-scale datasets and more sophisticated loss functions. This perspective has since shifted. Having demonstrated strong performance across general computer vision tasks, Vision Transformers are now adopted for face recognition to directly remedy key CNN deficiencies. This paper explores the architectural evolution of margin‑based CNNs, wherein hybrid CNN‑Transformer designs and fully Transformer‑based systems have emerged, drawing upon a range of relevant studies published between 2024 and 2026, such as LVFace (ICCV 2025), NPTFace and PaCo-FR (CVPR 2026), and the FaceLiVT series for mobile deployment, etc. The findings reveal that the the core issue lies not in Transformers outperforming CNNs on existing benchmarks, but in the shifts in computational models underlying face‑representation mechanisms.

With the increasing adoption of electric vehicles (EVs), conventional human-operated charging methods face limitations under challenging conditions, including low illumination, adverse weather, and accessibility constraints for users with limited mobility. Meanwhile, advances in robotic perception, multi-sensor fusion, and autonomous control technologies provide new possibilities for automated charging systems. This paper reviews recent developments in automatic EV charging technologies and proposes a design framework based on vision-based perception, sensor-fusion, and robotic manipulation. The proposed system integrates cameras and radar sensors for charging interface detection, position estimation, and robotic connector operation, aiming to improve charging efficiency and operational safety. The research focuses on several key challenges including accurate charging inlet localization, environmental adaptability, mechanical positioning accuracy, and safety control during charging operations. Existing studies indicate that automatic charging systems are technically feasible, however, practical deployment remains constrained by high equipment costs, variations in charging interface configurations, and robustness requirements under complex environments. This paper further analyzes potential optimization directions, including low-cost sensing solutions, intelligent perception algorithms, and improved cross-model compatibility. The study provides a comprehensive reference for the future development of automated EV charging systems.

Following the previous study of TDoA localization algorithms and hardware-assisted calibration, this paper develops a configurable DW1000-based ultra-wideband (UWB) platform for indoor localization with multiple anchors. Dedicated PCBs for the anchor and tag were designed to combine the DW1000 transceiver, STM32 controller interface, regulated power supply, controlled RF path, and customized antenna in one module. Three antenna candidates, including a modified circular radiator, a conventional circular patch, and a rectangular patch, were modelled in CST Studio Suite under the same conditions using a 1.2-mm F4B substrate. With the -6 dB level used as the preliminary comparison standard, the modified circular radiator obtained the widest simulated range from 6.0313 to 6.4318 GHz, giving an approximate bandwidth of 400.5 MHz. However, its -10 dB impedance bandwidth still requires further optimization before the antenna can be fabricated. The software was divided into three layers: an STM32 driver layer, a DW1000 API layer, and an application layer. This structure supports radio configuration, packet transmission and reception, timestamp reading, and future TDoA implementation. An infrared-synchronized ultrasonic unit was also developed as an additional short-range reference. It can be used to investigate UWB bias and to provide an initial region for the Taylor-series estimator. Nevertheless, one ultrasonic distance cannot independently determine a two-dimensional position or remove the clock drift between anchors. Qualitative tests were also carried out for multipath propagation, obstacle blockage, antenna orientation, and nearby wireless activity. These tests indicate several practical risks for system installation and provide a hardware and firmware basis for future synchronized multi-anchor experiments and quantitative localization evaluation.

With the continuous advancement of mechatronics, low-cost bionic hands have been widely used in educational experiments and lightweight robotic research. As the core power components, actuators determine the mechanical output, power consumption and overall practical performance of bionic hands. A small mistake made during the choice making of actuators is fatal, which can result in problems such as insufficient gripping force or exceeding cost. This paper first establishes three evaluation criteria for bionic hand actuators: mechanical performance, electrical power requirements and practical application conditions. Then, the operating principles, hardware characteristics and inherent drawbacks of three mainstream actuators which are massively used in this area are elaborated respectively. A comparison is conducted through a table clearly showing the differences between the three types of actuators, and targeted selection suggestions are provided for different application scenarios. This paper finds that servo motors fit low-difficulty rapid teaching prototypes, geared direct current (DC) motors are suitable for stable long-term laboratory experimental platforms, and shape memory alloy (SMA) actuators are limited to lightweight exploratory research due to power consumption and response speed constraints. This research aims to build systematic hardware reference standards for beginners engaged in low-cost bionic hand design, avoiding biased actuator selection caused by ignoring integrated mechanical and electrical constraints.

With the continuous integration of a new generation of information technology and the energy industry, artificial intelligence (AI) is increasingly used in data processing, learning and analysis, and decision-making support, and has gradually become an important technology to promote the development of automation and intelligence in the power industry. In a complex operating environment, traditional power automation systems are easily limited by control accuracy, fault identification and scheduling efficiency, making it difficult to adapt to the development needs of new energy integration into the grid, intelligent manufacturing and smart grid construction. This article introduces the main AI technologies such as machine learning, deep learning, reinforcement learning and expert systems, and discusses the development status of artificial intelligence in the field of electricity and the issues it faces, including data security, model interpretability and industry standardisation, in combination with specific applications such as power system regulation, equipment fault diagnosis, power grid scheduling, industrial control and new energy operation and maintenance. In addition, this article also discusses the application prospects of emerging technologies such as edge intelligence, digital twins and multimodal large models. Existing research shows that artificial intelligence can improve the reliability and control accuracy of power systems to a certain extent, and improve the efficiency of operation and maintenance, bringing new possibilities for the intelligent development of the power industry.

Diffusion models have become a dominant approach to image generation, but their iterative denoising process still creates high inference latency, memory consumption, and deployment cost. Existing surveys summarize efficient diffusion models from the viewpoints of sampling, compression, architecture, or applications, but they often do not make explicit which part of the inference cost is reduced by each method. This short survey reorganizes efficient image diffusion methods through an inference cost-decomposition perspective. The overall inference cost is described by three terms: the number of denoising function evaluations, the cost of each denoising step, and auxiliary overhead from conditioning, decoding, memory movement, and deployment frameworks. Based on this view, representative methods are grouped into sampling-step reduction, few-step student models, per-step computation reduction, and system-level deployment optimization. The discussion highlights that efficient diffusion is not simply a matter of using fewer sampling steps. Practical acceleration requires balancing image quality, latency, memory, training cost, and hardware constraints. This paper provides a compact framework for comparing efficient image diffusion methods and for selecting acceleration strategies under different deployment scenarios.

Low-cost RGB-D sensorsenable practical human motion capture and robot following, key capabilities for service robotics and human-robot interaction. Unlike expensive, high-precision systems, RGB-D sensors offer affordability, easy deployment, and real-time performance, making them ideal for civil robotic applications. This paper reviews RGB-D-based human pose estimation, motion feature extraction, and mobile robot following control, centering on low-cost, deployable solutions. The research sorts out low-cost RGB-D sensor schemes, human pose estimation algorithms, motion feature extraction pipelines, and tracking control strategies for mobile robots. Existing bottlenecks are summarised, including deteriorated pose accuracy under occlusion, insufficient real-time performance of lightweight algorithms, and unstable following behaviours in complex surroundings. Based on a comprehensive survey of current literature, this study concludes that the combination of lightweight visual models and optimised control algorithms serves as a vital development direction to construct low-cost and highly reliable human-following robotic systems. Prospective research trends of integrated RGB-D motion capture and robot following platforms are further discussed, which can offer theoretical and engineering references for relevant follow-up investigations and practical deployments.

Software testing accounts for a significant proportion of resource consumption within the software development lifecycle. Traditional automated testing methods rely on predefined scripts and deterministic logic, and have inherent limitations when addressing the dynamic complexities of modern software systems. Artificial intelligence, particularly large language models, offers new technical avenues for test automation.This paper focuses on two key tasks—AI-driven test case generation and defect detection—and provides a systematic review of the technological evolution from traditional automation to intelligent testing. It analyses methods such as prompt engineering, retrieval-augmented generation, model fine-tuning and multi-agent systems,while contrasting traditional deep learning and large language model approaches in defect detection. The paper also examines key challenges, including the hallucination problem, evaluation criteria, interpretability and generalisation capabilities.The fundamental contribution of large language models lies not in replacing human testers, but in driving the transformation of test automation from 'execution automation' to 'decision support'—a distinction that provides important guidance for the design, evaluation and deployment of AI testing tools.

Aiming at the research gaps in existing reviews on underactuated biomimetic robotic hands—namely fragmented mechanism classification and the lack of horizontal quantitative comparison among multi-drive configurations—this paper adopts a systematic literature review methodology. Using transmission type as the classification criterion, underactuated biomimetic robotic hands are categorized into three mainstream systems: linkage transmission, tendon–pulley transmission, and pneumatic soft actuation. Their structural principles, differential core components, prototype performance, and inherent limitations are analyzed, and representative industrial grippers, prosthetic hands, and modular pneumatic soft grippers are compared using quantitative data on grasping force, adaptive capability, and weight. The results reveal that each configuration possesses distinct advantages and inherent deficiencies, and no single transmission structure can simultaneously achieve high payload, lightweight design, and full-shape adaptive grasping. Technical bottlenecks and future development directions are further identified, providing a reference for mechanism optimization and engineering application of biomimetic robotic hands.

Generative AI is changing how short videos are created and is affecting the production, editing and collaboration of content. Existing research is scattered across various fields, such as technology application, creator practice and governance issues, and thus has not achieved all-round coverage of how generative artificial intelligence is changing the creative process and the human-computer relationship. This scoping review examined the prospects, problems and future directions of generative AI in short-video creation. According to the PRISMA-ScR guidelines, four databases (ACM Digital Library, IEEE Xplore, Scopus, Google Scholar and Web of Science) were used to systematically identify 13 studies published between 2020 and 2026. Based on the above analysis, generative AI is now used in short-video creation to achieve a joint effort between people and artificial intelligence for generation, selection, verification and modification. Generative AI has improved the efficiency of creation and expanded the scope of creative exploration, but new responsibilities have arisen to assess the authenticity of generated works and establish regulations for them. The current evidence mainly consists of the creator's use experience and specific application scenarios; therefore, more long-term studies, experiments and cross-platform research need to be conducted in the future to better understand the long-term effects of generative artificial intelligence on creative practice.