Offline reinforcement learning plays a prominent role in reducing training costs in autonomous driving. However, owing to uncontrollable training-data distributions and high sensitivity policy-parameters, its generalization performance across scenarios exhibits large gaps. Considering risk–uncertainty adaptation, an offline context-aware reinforcement-learning vehicle-cloud end-to-end model for autonomous driving is proposed. For the original training set collected from the vehicle-end, a Diffusion-based Generative Experience Replay (DGER) mechanism is constructed in the cloud-end to expand the data. A real-time driving-risk potential field assessment module is built for uncertainty mapping offline reinforcement-learning parameters. Concurrently, a context-aware policy embedded with a Semantic Attention Encoder (SAE) is established, enabling the on-board control unit to switch and adapt across scenarios. Experiments are conducted on the CARLA simulation platform. Relative to baseline algorithms, the proposed method delivers improved stability, safety, efficiency, and fuel-efficiently. The capacity for precise control-action regulation under extreme conditions and the scene-parsing ability across different scenarios are also verified.

Camera-based Vehicle-Infrastructure Cooperative 3D object detection (VIC3D) faces challenges including cross-view feature misalignment caused by viewpoint discrepancies, as well as semantic degradation and redundancy introduced by direct feature concatenation. To alleviate these issues, a pure-visual 3D object detection framework, VAPE-Net, is proposed for VIC3D. The framework comprises two key modules: (1) A Viewpoint-Aware Enhancement (VAE) module is introduced on both the vehicle-side and infrastructure-side branches. Instead of requiring precise extrinsic matrices for rigid cross-view geometric alignment, it captures semantic and geometric consistency by modeling feature correlations across adjacent and cross-view frames while only conditioning on static camera metadata, thereby improving robustness under moderate calibration perturbations. (2) A Stage-wise Cascaded Feature Fusion (SCFF) module is used to fuse voxel features in multiple stages, with a balancing mechanism to reduce information loss and suppress noise during fusion. Experiments on the real-world DAIR-V2X dataset demonstrate that VAPE-Net has a significant advantage over common fusion methods, achieving 16.05%AP3Dand 22.03%APBEV. This research expands the model paradigm for cooperative perception dominated by vision, providing new insights and support for the theoretical framework of VIC3D.

In IoT-enabled V2X cooperative vehicle–infrastructure perception, roadside units must transmit informative representations to vehicles over bandwidth-limited and latency-sensitive links, making it challenging for single-modality transmission to reduce communication payload while preserving detection accuracy. This paper presents a communication-efficient V2X cooperative 3D object detection framework, termed QMDet (Quantization-and-Mask based Detector). In the proposed design, the roadside unit jointly transmits a low-bit compressed feature map and a semantic mask to the vehicle. The feature branch performs learnable joint channel–spatial compression followed by INT8 quantization. In parallel, the semantic branch conducts contextual modeling and generates a semantic mask, providing explicit semantic priors to compensate for information loss caused by compression and quantization. Furthermore, a gated fusion module is introduced to adaptively integrate vehicle-side features, decompressed roadside features, and semantic priors via dynamic weighting, thereby mitigating cross-view/domain discrepancies. Experiments on the DAIR-V2X-C dataset demonstrate that the proposed framework reduces the RSU-to-vehicle communication payload by 24% while maintaining detection accuracy, achieving gains of 0.11% on [email protected] and 0.13% on [email protected]. These results indicate that QMDet improves communication efficiency without sacrificing cooperative detection accuracy.

Detecting whether a given text is a member in the pre-training data of Large Language Models (LLMs) is crucial for ensuring data privacy and copyright protection. Most existing methods rely on the LLM's hidden information (e.g., model parameters or token probabilities), making them ineffective in the black-box setting, where only input and output texts are accessible. Although some methods have been proposed for the black-box setting, they rely on massive manual efforts such as designing complicated questions or instructions. To address these issues, we propose VeilProbe, the first framework for automatically detecting LLMs' pre-training texts in a black-box setting without human intervention. VeilProbe utilizes a sequence-to-sequence mapping model to infer the latent mapping feature between the input text and the corresponding output suffix generated by the LLM. Then it performs the key token perturbations to obtain more distinguishable membership features. Additionally, considering real-world scenarios where the ground-truth training text samples are limited, a prototype-based membership classifier is introduced to alleviate the overfitting issue. Extensive evaluations on three widely used datasets demonstrate that our framework is effective and superior in the black-box setting.

A novel autonomous driving motion planning framework for unsignalized intersections is presented. To achieve an effective balance between safety and efficiency in the decision-making process, a motion planning decision strategy tailored for discrete action spaces is developed based on the discrete soft actor-critic algorithm. In response to the challenges posed by the complexity of feature information in dense intersection environments, a multi-level attention mechanism–integrating both feature-level and vehicle-entity-level information–is introduced to significantly enhance feature extraction and processing capabilities. Furthermore, to mitigate the issues of temporal sample distribution imbalance and low utilization of high-value samples in a single experience pool, a dual experience buffer prioritized replay mechanism is proposed, thereby improving training stability. Experimental results indicate that, compared with alternative methods, the proposed framework not only achieves a superior balance between efficiency and safety but also exhibits enhanced interpretability and generalization performance.

The inefficient capture of equipment anomalies has impeded the effective training of models for detecting anomalies in various traffic equipment (TE). This paper proposes an unsupervised knowledge distillation for traffic equipment anomaly detection (UKD-TEAD), eliminating the need for numerous annotations and ensuring the applicability to various equipment anomalies. First, three specialized detection heads based on different object category aspect ratio distributions are designed, to detect multi-scale objects with high precision in complex traffic scenes. Second, a teacher-student model, grounded in hierarchical knowledge distillation, is developed to mitigate the critical feature loss associated with the small size of cropped regions of interest (ROIs). By performing knowledge distillation at different depths of the network, the student network effectively learns the representation capabilities of the teacher network on multiple scale feature layers, thereby improving the anomaly detection performance. Finally, to validate the proposed unsupervised anomaly detection framework, a target detection dataset and an unsupervised anomaly detection dataset were constructed based on traffic inspection data. Experimental results show that the proposed method achieves an [email protected] of 0.862 for TE detection, while the mean area under the curve (mAUC) for anomaly detection reaches 0.857.

Autonomous driving (AD) is critically dependent on intelligent decision-making technology, which is the crucial ingredient in driving safety and overall vehicle performance. And comprehensive consideration of driving heterogeneity, decision synergy, and game interaction is also the cornerstones. Accordingly, this paper constructs a cooperative decision-making framework for autonomous vehicles (AVs) that integrates driving styles within a hierarchical architecture based on deep reinforcement learning (DRL). The upper layer adopts the action shielding mechanism-based dueling-double deep Q-network (D3QN) algorithm incorporating the lane advantages into shared state space to complete the prompt lane-changing (LC) decision, the lower layer applies the soft actor-3-critic (SA3C) algorithm based on the clipped triple Q-learning to provide the continuous speed adaptive control. Three personalized collaborative decision strategies are formulated for particular driving styles in multi-objective optimization preference combined with style-incentive prioritized experience replay (SIPER). The experimental results confirm that the proposed framework can satisfy the personalized driving demands in complex traffic scenarios, effectively explore the prospective LC opportunities, and enhance the driving efficiency by 35.40% with aggressive strategy and the comfort by 56.46% with defensive strategy compared with normal strategy, while maintaining the safety.

Knowledge of vacant taxis’ passenger-searching behavior is of great social and economic interest to multiple applications, particularly in planning and operating on-demand mobility services. The inherent stochasticity and dynamic nature in both passenger demand and drivers’ decision-making pose challenges in capturing vacant taxis movements. This study proposes a novel deep clustering framework to comprehensively uncover passenger-searching strategies from extensive trajectory data. Specifically, multiple features from vacant searching trips are extracted and analogously defined as multi-channel images, where each channel corresponds to a specific feature. A novel deep image clustering approach is then proposed, integrating a feature representation module utilizing convolutional neural networks (ConvNets), a self-expression module for affinity learning, a spectral clustering module, and a classification module for self-supervision. An effective training procedure is also presented for the proposed deep clustering framework. Experiments demonstrate the effectiveness of the proposed approach against benchmark methods. Based on the clustering results, common and specific passenger searching strategies are further revealed. Specifically, our findings highlight the importance of individual’s contextual experience in explaining searching behavior and operational efficiency. Moreover, drivers cruising without clear searching strategy often exhibit lower performance, and some drivers may gamble with peers to increase their chances of picking up passengers. These results deliver important justifications for future studies and provide managerial implications to improve on-demand mobility.

The implementation of autonomous driving is inseparable from developing intelligent driving decision-making models, which are facing high scene complexity, poor decision-making coupling, and the inability to guarantee decision-making safety. This paper starts with the priority and logic of lane change and car-following decision-making, considering driving efficiency, safety, and comfort, then constructs a double-layer decision-making model. This paper uses two deep reinforcement learning algorithms for the upper and lower layers to process large-scale mixed state space and ensure the composite action output of lane-changing decisions and car-following decisions. In the upper layer model, we use the D3QN algorithm to distinguish the potential value of the environment and the value of selecting lane-changing actions when making lane-changing decisions. Different from the traditional mechanisms that only use negative rewards, the lane changing benefit function and dangerous action shielding mechanism are used to eliminate collisions. DDPG algorithm is adopted in the lower layer model to process car-following decisions and output continuous vehicle speed control. Besides, coupled training is taken for the two algorithms to improve the coordination of the double-layer model. This paper selected mixed standard driving cycle conditions to build a highly complex training environment and used NGSIM data to reconstruct scenes to test our model. Simulations in SUMO are presented that the double-layer model can increase the driving speed of the original data by 23.99%, which has higher effectiveness than other models.

Modern automotive systems have been equipped with a highly increasing number of onboard computer vision hardware and software, which are considered to be beneficial for achieving eco-driving. This article combines computer vision and deep reinforcement learning (DRL) to improve the fuel economy of hybrid electric vehicles. The proposed method is capable of autonomously learning the optimal control policy from visual inputs. The state-of-the-art convolutional neural networks-based object detection method is utilized to extract available visual information from onboard cameras. The detected visual information is used as a state input for a continuous DRL model to output energy management strategies. To evaluate the proposed method, we construct 100 km real city and highway driving cycles, in which visual information is incorporated. The results show that the DRL-based system with visual information consumes 4.3-8.8% less fuel compared with the one without visual information, and the proposed method achieves 96.5% fuel economy of the global optimum-dynamic programming.