Abstract Current autonomous vehicles (AVs) primarily respond to external traffic conditions and treat humans as passive occupants, limiting their ability to personalize driving behavior and handle ambiguous scenarios that could benefit from occupant input. This study proposes PACE-ADS (Psychology and Cognition Enabled Automated Driving Systems), a human-centered autonomy framework that enables AVs to interpret and respond to both external traffic conditions and internal occupant states. PACE-ADS employs three collaborating foundation model agents: the Driver Agent interprets the external environment; the Psychologist Agent infers occupant state from passive psychological signals (e.g., facial expressions) and interprets active cognitive inputs (e.g., verbal commands); and the Coordinator Agent integrates these inputs to generate high-level driving behavior decisions. Operating at the low-frequency semantic planning layer, PACE-ADS complements existing AV stacks and activates selectively in response to changes in occupant state, cognitive instructions, or vehicle immobilization. Closed-loop CARLA simulations under prescribed occupant-state trajectories show that PACE-ADS adapts driving behavior along kinematic dimensions associated with ride comfort and supports recovery from operational immobilization through self-reasoning and occupant-in-the-loop guidance.
Abstract Potholes, recognized as the second most frequent pre-crash event, severely compromise traffic safety while posing critical threats to vehicle structural safety and ride comfort. Existing countermeasures, primarily active suspension control and longitudinal speed regulation, struggle to mitigate these structural impacts without increasing the probability of multi-vehicle conflicts, particularly rear-end collisions. Leveraging the extended preview information regarding pothole geometry and locations provided by vehicle-to-everything (V2X) communication, this study proposes a vehicle-dynamics-aware motion planning framework that introduces proactive intra-lane lateral maneuvering to expand conventional mitigation strategies. The framework systematically integrates microscopic tire-pothole impact mechanics, vertical quarter-car dynamics, and car-following behaviors into a unified closed-loop simulation environment. A curriculum learning strategy is incorporated into the Soft Actor-Critic (SAC) algorithm to progressively increase task complexity, thereby mitigating the training instability caused by abrupt and sparse physical impact penalties. Comprehensive simulations covering diverse pothole geometries, preview distances, speed ranges, and car-following interactions validate the framework. For geometrically avoidable hazards, the trained policy executes proactive lateral bypassing, achieving a 95.8% impact-load compliance rate while reducing travel time over the defined pothole-passage interval by approximately 50%. Under unavoidable conditions, the policy performs adaptive speed regulation, achieving a 46.7% impact-load compliance rate while reducing the rear-end collision rate from 13.3% to 1.7% compared with braking-heavy baselines. Overall, this approach expands vehicle capabilities in localized hazard mitigation, demonstrating the potential of integrating low-level physical dynamic boundaries into motion planning frameworks.
Abstract The deployment of automated driving (AD) systems demands trustworthiness, explainability, and accountability. A challenge is detecting out-of-distribution (OOD) inputs from sensor noise, hardware failures, adversarial interference, or deviations from the operational design domain. Common approaches use OOD samples during training or threshold selection, potentially limiting generalization beyond the anomaly types represented during development. We argue that OOD-unaware novelty detection is crucial for detection under unforeseen conditions. This paper enhances a deep-reinforcement-learning-based self-parking agent developed in CARLA by integrating a novel OOD detector. We present Time-Frequency-Memory-enhanced (TFMe), a dual-branch, memory-augmented, encoder-only Transformer that outperforms the evaluated baselines. In OOD-unaware settings, decision thresholds are calibrated exclusively on in-distribution validation data using a 99th-percentile rule. We assess sensitivity across multiple noise, attack types, and severity levels. Results show that OOD-unaware calibration maintains consistent performance across anomaly types and intensities, avoiding the drops observed under OOD-aware calibration. Experiments on real-world data from Lyft with synthetically injected anomalies further show that OOD-unaware calibration transfers more effectively to the evaluated unseen anomaly types. Transferability across parking environments is assessed on an unseen 60° angled layout without retraining. We show that the TFMe-enhanced ADF mitigates collision risk and reduces timeouts in the evaluated parking scenarios by issuing Take-Over Requests after repeated anomaly detections. TFMe has lower inference latency than the self-parking agent, allowing it to operate in parallel without extending the estimated critical path. Finally, compared with Raspberry Pi 5, deployment on a Jetson Orin Nano achieves 19% lower inference time and 30% lower energy consumption.
Abstract Most truck modeling studies make generalized assumptions regarding driving times, distances, and non-driving tasks, except for loading and unloading, and therefore fail to capture the diversity of freight transport vocations, especially in the context of automated trucking analyses. To fill this gap, this study followed a three-stage process that involved reviewing existing literature, a typological analysis of driving and non-driving tasks in trucking vocations, and a quantitative assessment of the vocation specific benefits and challenges of automation. The last, however, was rather illustrative because of restricted available task-share data. The analysis grouped the tasks by their location: at the terminal/depot, on route, and at the customer site. From an operational perspective, tasks at the customer site are a key limiting factor for the use of automated vehicles. Non-driving tasks further determine the operational usefulness and benefits of automated vehicles. From a technical perspective, the traffic environment significantly influences the ease of implementing automated driving. For example, controlled traffic environments, such as yard transportation, are easiest to automate, while complex environments, such as urban areas, are the hardest. Overall, task shares significantly influence the utility of automated driving for individual freight vocations, with non-driving task shares ranging from 22 to 80%. Moreover, vocations that offer the greatest benefits may be more conducive to the development of automated driving technologies than those with the fewest technical and operational barriers to entry.
Abstract The low-altitude economy is expanding urban mobility and service systems from ground-based networks to three-dimensional urban space. However, its large-scale deployment depends not only on aircraft technologies, but also on reliable operational infrastructure for communication, navigation, surveillance, meteorological sensing, computing, and safety supervision. This paper argues that smart infrastructure originally developed for connected and automated vehicles can provide a reusable foundation for low-altitude operations. Because low-altitude public routes are likely to follow existing urban corridors such as roads, railways, rivers, and utility corridors, there is strong spatial overlap with ground-based intelligent transportation infrastructure. By extending these capabilities upward, cities can reduce duplicate investment, improve infrastructure utilization, and support safe, scalable, and city-level air-ground operation management.
For humanu2012machine co-driving vehicles, the dynamic complexity of urban scenarios and the uncertainty of driving behavior impose stringent demands on the accuracy and generalization capability of ego vehicle trajectory prediction. Current research predominantly relies on multivehicle interaction information or single-scenario settings, overlooking the inherent dynamic correlation between driver and vehicle. This results in prediction models struggle to adapt to complex urban environments. To address this, this study proposes a trajectory prediction framework based on driver-vehicle coupling feature encoding, enabling precise capture of vehicle trajectory evolution patterns under driver manipulation across diverse urban scenarios. The framework employs bidirectional long short-term memories (LSTMs) to perform temporal encoding on historical driver-vehicle coupling features and future road geometric features. Combined with an attention mechanism, it generates context vectors integrating temporal features and manipulation details, ultimately using LSTMs to recursively produce multistep prediction results. Validation using diverse urban scenarios data from a dynamic driving simulatordemonstrates that our prediction framework achieves precise trajectory prediction in typical scenarios such as lane changing, turning, and roundabout navigation, while exhibiting robust stability and generalization capabilities.
To address the complexity of longitudinal-lateral vehicle dynamics in platoon control and the challenge of mitigating traffic oscillations, a novel model-based reinforcement learning (MBRL) method with planning capability is proposed for longitudinal-lateral control of vehicle platoons. Specifically, the integration of the environment model and policy model, combined with a sampling-based planning approach, enables planning-based control during decision-making, which significantly improves the stability and safety of vehicle platoon control. In the two-dimensional scenario, a unified state space and joint action space are designed to achieve longitudinal-lateral control, along with a multidimensional reward function that integrates both control objectives. To address the low exploration efficiency and the high proportion of ineffective exploration during the early stage of training, a prior-knowledge-guided exploration strategy is introduced. This strategy improves learning efficiency and accelerates convergence by incorporating guided actions and constraints. The training results indicate that the incorporation of prior knowledge significantly enhances training efficiency. In evaluations under speed-limit scenarios and real-data scenarios, the proposed method demonstrates superior performance in longitudinal-lateral control, platoon stability, safety, and adaptability, while maintaining high computational efficiency.
Safety is the most critical problem in autonomous driving (AD). Crashes often occur in long-tail scenarios, which are neither frequent nor representative of normal driving conditions. Many severe failures are not caused by a single error but by the accumulation of coupled behaviors and/or environmental factors over time. These long-tail scenarios are difficult to evaluate using traditional open-loop safety analysis methods. To address the aforementioned challenges, the current study discussed how world models enabled long-tail scenario generation. By using closed-loop inference, world models captured how an agentu2019s own decisions influenced the subsequent states and interactions. In addition, world models contributed to scenario-specific generation by enabling controllable conditioning and targeted intervention on agent behaviors and environmental factors. In future studies, how to avoid unrealistic hallucinations, maintain system-level evaluation, and address errors arising from long-term interactions and multistep accumulations remain the key problems we are facing in the safety evaluation for AD.
The commitment to decarbonization is motivating urban planners to adopt emerging techniques that advance sustainability. Road traffic emissions remain a major source of greenhouse gases and pollutants, requiring precise, near-real-time monitoring for effective mitigation policies. This study introduces the design and demonstration of a digital twin (DT) platform for road traffic emission nowcasting and forecasting. The focus is on establishing a streamlined technical architecture and showcasing how the system can utilize multisource data from the Internet of Things (IoT) sensors and simulation to provide a high spatiotemporal resolution view of emissions. As a proof of concept, the platform leverages traffic camera data as IoT input, highlighting its potential for simultaneous emission and origin destination matrix estimation (ODME). A case study in Kista, Stockholm, illustrates the platformu2019s capabilities through a 3-dimensional (3D) interactive visualization in Unity. This demonstration serves as a first step toward a fully validated emission monitoring system, providing a scalable and modular framework that can be adapted for related applications, such as congestion analysis and noise monitoring.
Vehicle detection is a crucial part of perception technologies for unmanned vehicles, which provides perceptive guarantees for route plans and vehicle control. Current multimodal fusion approaches for image and light detection and ranging (LiDAR) data primarily include feature-level fusion, object-level fusion, and data-level fusion strategies. However, feature-level and object-level fusion methods fail to fundamentally address the core challenges of low-quality sparse point cloud data with insufficient information and high noise levels, resulting in low confidence in fused features or detection results. Research on data-level fusion methods remains limited. Aiming at the ineffective representation of distant vehicles by sparse LiDAR clouds, this study proposes a vehicle detection method based on the data-level fusion of image and LiDAR data. First, this method estimates the depth of image pixels to generate the pseudo point clouds. Then, the K-dimensional (KD) tree is employed to reduce the noise and eliminate the outlier of pseudo points to fuse with LiDAR data. Finally, the fused point clouds are fed to the point-based region-based convolutional neural network (PointRCNN) to achieve vehicle detection. The advantage of the proposed method is that it utilizes pseudo point clouds to compensate for sparse LiDAR data, which contributes to improving the detection accuracy of distant vehicles. This study utilizes the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset to conduct comparative experiments and ablation studies. The experimental results reveal that the proposed method improves 7.8% and 4.13% in the aspects of average precision for three-dimensional objects (AP3D) and average precision for bird’s eye view (APBEV) with comparison methods, respectively. Meanwhile, in terms of distant vehicle detection, the proposed method raises 9.13% and 9.58% in the aspects of AP3D and APBEV respectively by comparison with frustum-based PointNet (F-PointNet).
Parking-slot detection is pivotal for autonomous driving, facilitating automated parking and enhancing vehicle safety. However, the diversity of parking slot shapes and the complexity of surrounding environments incur significant costs in data collection and annotation. To address this challenge, we propose a novel data synthesis framework specifically tailored for around-view monitor (AVM) images. First, an inpainting-based generative algorithm eliminates foreground elements from real parking slot images to produce clean backgrounds. Subsequently, new foreground elements exhibiting diverse shapes, colors, and textures are superimposed onto these backgrounds. Furthermore, rather than relying on domain selection, we introduce a data selection strategy based on active learning that operates directly on the generated datasets. The controllable attributes of synthetic data facilitate the effective evaluation and optimization of the selection strategy across various scenarios. Experimental results on the panoramic surround view (PSV) dataset demonstrate that models trained exclusively with synthetic data achieve 1.32% higher precision than those trained only on real images. Moreover, integrating 40% real images with synthetic data increases precision by up to 1.74% and recall rates by up to 1.48%, highlighting the effectiveness and practical utility of our proposed approach.
Autonomous lane-changing decision making and planning represent a fundamental aspect of advanced driving technologies, playing a pivotal role in improving operational safety, enhancing passenger comfort, and optimizing traffic flow. Current research predominantly emphasizes environmental perception and path planning, yet systematically modeling human behavioral patterns during lane changes remains underexplored, leading to inadequate anthropomorphic decision-making capabilities. Moreover, the conventional fragmented approach to implementing decision-making, trajectory planning, and interaction signaling modules results in insufficient coordination and feedback mechanisms, ultimately compromising dynamic adaptability in real-world driving scenarios. To solve these problems, this study systematically investigates driver behavior patterns through naturalistic driving data analysis, establishes a taxonomy of lane-changing scenarios, and develops a human-like decision architecture incorporating cognitive mechanisms. The model consists of a multilayered decision framework encompassing lane-changing motivation recognition, lane selection, feasibility evaluation, and risk assessment. Furthermore, an information feedback mechanism is established between the decision-making and trajectory planning modules, enabling dynamically coupled and closed-loop control. Simulation experiments conducted on the Prescan/Simulink platform confirm that the proposed method significantly enhances the naturalness and safety of lane-changing behavior in complex traffic environments. This study provides both theoretical support and technical guidance for the development of intelligent lane-changing systems that emulate human cognitive characteristics.
Advanced driver assistance systems (ADASs) can greatly enhance road safety by providing real-time warnings to drivers in imminent crash situations. However, the provided warning time may deviate from its designed time. There is limited research on how warning uncertainties influence drivers’ behavior, safety performance, and trust. This study conducted a driving simulator study to examine how uncertainties in warnings impact driving behaviors and trust using a roundabout driving scenario. Two warning error distributions were constructed to represent low and high warning uncertainty levels. Thirty-six participants were recruited and randomly divided into two groups under the two uncertainty levels in a driving simulator experiment. The between-group analysis shows that the lower warning uncertainty level group results in higher trust and that trust increases (or decreases) over time under low (or high) uncertainty levels. The within-group analysis shows that higher warning errors downgrade drivers’ trust and safety performance when the errors are high. Finally, a personalized trust prediction model was developed using demographic and vehicle movement data, and the XGBoost model achieved the best performance with 86.42% accuracy.
The integration of artificial intelligence (AI) in the Chinese automotive industry has led to unprecedented growth and innovation, particularly in electric vehicles and autonomous driving technologies. This study examines the impact of AI on Chinau2019s automotive sector, analyzing key players, government policies, and technological advancements. Through a comprehensive review of the academic literature and industry reports, we explore how Chinese manufacturers are leveraging AI to enhance vehicle performance, safety, and user experience. The research reveals that Chinau2019s supportive regulatory environment and significant investments in AI have positioned it as a global leader in automotive AI applications. However, challenges remain, including data privacy concerns and the need for international standardization. The study concludes that Chinau2019s AI-driven automotive revolution is reshaping the global industry, necessitating accelerated AI adoption by Western manufacturers and increased international collaboration.
In emerging mixed traffic environments, connected and autonomous vehicles (CAVs) must interact with surrounding human-driven vehicles (HDVs). This study introduces multisource human-in-the-loop mixed cloud control testbed (MSH-MCCT), a novel CAV testbed that captures complex interactions between various CAVs and HDVs. Utilizing the mixed digital twin concept, which combines mixed reality with digital twins, MSH-MCCT integrates physical, virtual, and mixed platforms, along with multisource control inputs. Bridged by the mixed platform, MSH-MCCT allows human drivers and CAV algorithms to operate both physical and virtual vehicles within multiple fields of view. In particular, this testbed facilitates the coexistence and real-time interaction of physical and virtual CAVs and HDVs, significantly enhancing the experimental flexibility and scalability. Experiments on vehicle platooning in mixed traffic showcase the potential of MSH-MCCT to conduct CAV testing with multisource real human drivers in the loop through driving simulators of diverse fidelity.
