Xiao Wang

Despite the recent advancements in artificial intelligence (AI) technologies showing great potential in improving transport efficiency and safety, autonomous vehicles (AVs) still face great challenges when driving in time-varying traffic flow, especially in dense and interactive situations. Meanwhile, humans have free will and usually do not make the same decisions even situated in exactly the same scenarios, leading to data-driven methods suffering from poor migratability and high search cost problems, decreasing the efficiency and effectiveness of the behavior policy. In this article, we propose a safety-first human-like decision-making (SF-HLDM) framework for AVs to drive safely, comfortably, and with social compatibility and efficiency. The framework integrates a hierarchical progressive architecture, which combines a spatial–temporal attention (STA) mechanism for other road users’ intention inference, a social compliance estimation (SCE) module for behavior regulation, and a deep evolutionary reinforcement learning (DERL) model for expanding the search space efficiently and effectively to make avoidance of falling into the local optimal trap and reduce the risk of overfitting, thus make human-like decisions with interpretability and flexibility. The SF-HLDM framework enables autonomous driving AI agents to dynamically adjust decision parameters to maintain safety margins while adhering to contextually appropriate driving behaviors at the same time. Extensive experiments in car learning to act, an open-source autonomous-driving simulator (CARLA) validate the framework’s superior performance, which enlarges the minimum time to worst-case hazards (TWHs) by 41.8% to keep a safer distance away from others, while improving the average velocity by 2.5%, reducing the average acceleration and yaw rate by 23.5% and 60.5%, respectively. The results highlight the potential of SF-HLDM to bridge the gap between machine-driven precision and human-like flexibility in AV systems, paving the way for more interpretable and socially acceptable autonomous driving solutions.

Most existing deep learning-based trajectory prediction algorithms heavily rely on human expertise, involving iterative manual tuning of their architectures and parameters to tailor prediction models for specific tasks or scenarios. This approach is not only complex to implement and inefficient, but also struggles to balance inference speed with prediction accuracy. To address this challenge, this paper innovatively proposes an improved heterogeneous multi-agent trajectory prediction algorithm utilizing graph neural architecture search. This method automatically conducts an end-to-end graph architecture search to obtain an optimal trajectory prediction model. To enhance model interpretability and its heterogeneous awareness of diverse scenarios, we design a physics- and risk-interaction-based guidance mechanism to steer the architecture search process. Furthermore, we construct a novel neural architecture search loss function, SocialMI-Loss, which comprehensively considers multiple factors such as prediction accuracy, driving region semantic constraints, and model complexity. This function is intended to guide the learning of the trajectory predictor, achieving a harmonious balance between accuracy and computational complexity. A comprehensive series of comparative experiments conducted on three large-scale autonomous driving datasets (nuScenes, Argoverse, and ApolloScape) consistently demonstrates the superior performance of our proposed method. Experimental results indicate that our framework achieves performance comparable to current state-of-the-art methods, while its automatically searched architecture remains remarkably lightweight. Our code is available at:https://github.com/Tu5tra/TrajGNAS.

Recent methods to enhance Vision-Language Models (VLMs) for Visual Question Answering (VQA) have focused on strengthening their inference capabilities, enabling them to tackle VQA tasks independently rather than merely as aids to Large Language Models (LLMs). However, these approaches often ignore the rich commonsense knowledge inside the given VQA image sampled from the real world, limiting the full potential of VLMs. Inspired by the human top-down reasoning process, i.e., systematically exploring relevant issues to derive a comprehensive answer, this work introduces a novel, explainable multi-agent collaboration framework by leveraging the expansive knowledge of LLMs to enhance the capabilities of VLMs themselves. Our framework comprises three agents, i.e., Responder, Seeker, and Integrator, to collaboratively answer the given VQA question by seeking its relevant issues and generating the final answer in such a top-down reasoning process. The VLM-based Responder agent generates the answer candidates for the question and responds to other relevant issues. The Seeker agent, primarily based on LLM, identifies relevant issues related to the question to inform the Responder agent and constructs a Multi-View Knowledge Base (MVKB) for the given visual scene by leveraging the build-in world knowledge of LLM. The Integrator agent combines knowledge from the Seeker agent and the Responder agent to produce the final VQA answer. Extensive and comprehensive evaluations on diverse VQA datasets with a variety of VLMs demonstrate the superior performance and interpretability of our framework over the baseline method, e.g., 5.7% improvement on VQA-RAD and 5.2% on Winoground in the zero-shot setting without extra training cost.

Identifying gene-disease associations (GDAs) remains a fundamental challenge in biomedical research due to the enormous combinatorial space of candidate gene-disease pairs and the limited scalability of experimental validation. As wet lab studies cannot keep pace with rapidly expanding omics data, computational approaches have become essential for prioritizing plausible GDAs and accelerating biological discovery. Recent advances in artificial intelligence (AI), particularly graph neural networks (GNNs) and large language models (LLMs), are trans forming this field by enabling richer biological representations and more accurate predictive modeling. In this survey, we provide a unified and up-to-date overview of AI-driven GDA prediction. We first summarize major public resources containing gene, dis ease, and auxiliary biological information that underpin computational studies. We then review methodological developments ranging from traditional network-based methods to machine learning, deep learning, and the emerging integration of GNNs and LLMs, which has received limited attention in previous GDA-focused surveys. Representative applications in gene prioritization, drug repurposing, and clinical research are also discussed to demonstrate the practical impact of these approaches. Finally, we outline current challenges and promising future directions. By integrating data resources, methodological advances, and translational applications, this survey provides a comprehensive overview of modern AI techniques for GDA prediction and aims to support the development of more robust, interpretable, and clinically actionable computational tools. All curated resources and re viewed literature are publicly available in our GitHub repository (last updated September 2025; including peer-reviewed publications and preprints on AI-driven GDA prediction published through September 2025): https://github.com/linyaoyang/gene disease-association-prediction-papers.

This letter reports on a TIV DHW (decentralized and hybrid workshop) that explores the prospective influence of ChatGPT on research and development in intelligent vehicles. To assess the update capabilities of ChatGPT, we conducted tests involving both basic and technically relevant questions. Our preliminary testing revealed that ChatGPT's information can be updated and corrected at one time, but it may take some time for the changes to be reflected in ChatGPT's responses, so it may not always possess the latest knowledge regarding specific topics. We further discuss the prospective influence of ChatGPT on the field of intelligent vehicles, particularly possible applications of ChatGPT in areas like autonomous driving, human-vehicle interaction, and intelligent transportation systems, highlighting challenges and opportunities associated with these applications. Additionally, we address technical questions, such as the feasibility of training intelligent vehicles using the same methods as ChatGPT and the reflection of the intelligence of intelligent vehicles in the context of human-machine shared control. In conclusion, this letter presents a preliminary exploration of the potential of ChatGPT for intelligent vehicle research, from an IEEE TIV perspective, acknowledging the limitations and uncertainties of this emerging technology.

Recently, a chat generative pre-trained transformer (ChatGPT) attracts widespread attention in the academies and industries because of its powerful conversational ability with human and its astonishing emergence ability such as admit mistakes, reject inappropriate problems. However, it is not easy to generalize ChatGPT into the field of intelligent vehicles because of its high computational cost, uncertain answers and decisions, and the difficulty of scenario generation for intelligent vehicles. To address these issues, we propose a novel framework, human-machine-augmented intelligent vehicles with generative pre-trained transformer. Under this framework, we discuss the potential, prospects, limitations and several typical applications of HiVeGPT in the domain of intelligent vehicles.

This paper investigates the modeling and the practical predefined-time (PdT) tracking control problems for a fully actuated disk-shaped autonomous underwater vehicle (AUV) with six degrees of freedom. To overcome the gimbal lock problem inherent in Euler angle representation, unit quaternions are adopted to model the AUV, accounting for internal uncertainties and external disturbances. Then, an improved time-varying function is introduced, which serves as the basis for designing a non-singular sliding surface and sliding mode controller with PdT stability. This approach ensures that the tracking errors converge within a predefined time, independent of initial conditions and design parameters. Compared with traditional PdT controllers, the proposed method eliminates singularities, enhances the precision of convergence time estimation, and typically yields smaller, smoother initial control inputs, thus improving its potential for engineering applications. Numerical simulations validate the effectiveness and performance of the proposed controller.

The task of Pedestrian Trajectory Prediction (PTP) aims to forecast the future movement paths of pedestrians based on their past behavioral patterns, which is crucial for autonomous systems (e.g., autonomous vehicles and social robots) in path planning and decision-making processes. In recent years, with the rapid advancement of Artificial Intelligence (AI), especially in the field of deep learning, PTP has achieved remarkable breakthroughs. However, this field still faces numerous challenges and unresolved issues that require further research and exploration. This paper provides a comprehensive review and perspectives of the latest advancements in PTP methods, starting with the problem definition and method classification. Then, guided by the key issues at hand, we compare and analyze physics-based, classic Machine Learning (ML)-based, and AI-based methods, and discuss their applicability in various application scenarios. Finally, the paper provides existing datasets and performance metrics, and outlines potential research directions.

The rapid development of computer vision and artificial intelligence technology has made overcoming the limitations of high cost and scene coverage in the physical testing of advanced self-driving cars a key challenge for the industry. Current physical testing methods are based on real vehicle road tests, which present two key challenges: economic issues and inadequate scenario coverage. Meanwhile, existing research tends to focus on validating algorithms in fixed scenarios. In response to the limitat mentioned, this paper proposes a parallel scenario testing method that combines the CARLA simulation platform with the YOLOv8-based detection algorithm. This method constructs a multidimensional testing framework that incorporates natural factors and traffic complexity indexes. It also verifies that the CARLA-YOLOv8 parallel testing framework can effectively quantify the robustness of the algorithm under pressure in a multidimensional environment. The framework introduces a new approach to the virtual verification of autonomous driving systems. This is particularly important for addressing safety-critical edge cases and scalability issues.