Vision-language models (VLMs) offer the potential for unified perception and language-guided decision-making in autonomous driving. However, existing benchmarks predominantly focus on structured road environments and high-quality imagery, leaving limited evidence on model performance (e.g., reasoning, explanation, and decision traceability) under unstructured scenes or degraded sensing conditions. This study develops a trustworthy test and evaluation framework to systematically assess VLM performance in these challenging contexts. Impromptu vision-language-action (VLA) samples are reorganized into six synchronized camera views augmented with vehicle state information, and twenty realistic input perturbations, covering illumination, weather, sensor reliability, and occlusion, are introduced for each scene. Moreover, six original non-open-ended questions are reformulated into traffic decision templates that combine structured choice sets with template-constrained free-text responses. Two types of evaluation methods are formulated. Multiple-choice questions (MCQs) are based on exact answer matching, aggregated with importance weighting according to expert rankings to prioritize planning tasks, and subjective questions (SQs) are graded using tailored, multi-dimensional LLM-based scoring prompts. Experiments are conducted to assesses their performance of seven open-source VLMs, including multiple-choice questions (MCQs) accuracy, reasoning coherence and visual fidelity of their responses to subjective questions (SQs), and their robustness to input perturbations. Overall, the proposed framework addresses a critical evaluation gap for supporting the deployment of VLMs in autonomous driving applications.

With the steady growth of urban car ownership, the issue of traffic congestion is becoming increasingly prominent, bringing great pressure to urban development. To respond effectively to this challenge, it is critical to develop methods that can improve transport efficiency and reduce energy consumption. In current context, the Cooperative Vehicle Infrastructure System (CVIS), an ideal solution for realizing green and intelligent transportation systems, has become an important direction in both transportation research and practice. By integrating and optimizing various traffic resources, CVIS not only enhances traffic efficiency and reduces energy consumption but also provides key technical support for achieving“dual carbon”goals. This paper thoroughly analyzed the fundamental concepts, research methodologies and application scenarios of CVIS, and delved into its four core technological modules: fusion perception, driving cognition, autonomous decision-making, and cooperative control. The paper reviewed and summarized research achievements within these modules, ranging from traditional methods to the latest in deep reinforcement learning techniques. It also explored the potential applications of these technologies and methods for enhancing traffic efficiency, reducing energy consumption, and improving road safety. Finally, the paper scrutinized numerous challenges that CVIS may encounter in practical applications, including the security of information transmission, system stability, and environmental complexity. To overcome these challenges, the paper looked forward to the future development in four areas: developing datasets that integrate vehicle-side and roadside information, enhancing the fusion accuracy of multi-source perception information, improving the real-time performance and safety of CVIS, and optimizing multi-vehicle cooperative decision-making control methods under complex conditions. As a result, this paper not only has important reference value for the advancement of CVIS technology, but also provides important guidance for the future planning and construction of urban transportation systems.

With increasing public concern about autonomous vehicles, there is a growing demand for developing explainable autonomous driving planning technology. Traditional risk field methods use handcrafted potential field models to explain driving risks in a scenario. When explaining highly interactive scenarios, such prior knowledge-based methods still lack flexibility, leading to insufficient interpretability. In this study, we first propose the concept of a risk map that can be seen as a discrete, ego vehicle's view form of the risk field. We then design an explainable trajectory planning framework that integrates risk maps with the candidate trajectory tree generated by trajectory prediction models. We further filter safe candidate trajectories from the tree on the basis of their cumulative risks in the risk maps and then select the optimal trajectory to execute by balancing other driving objectives. The validation results in various real-world scenarios demonstrate that our method can generate understandable risk maps and explain the risk differences between trajectories. Open-loop experiments show our model's advantages in terms of safety and efficiency for the trajectory planning task. An analysis of runtime demonstrated its potential for real-world applications.

Cooperative driving is widely viewed as a promising method to better utilize limited road resources and alleviate traffic congestion. In recent years, several cooperative driving approaches for idealized traffic scenarios (i.e., uniform vehicle arrivals, lengths, and speeds) have been proposed. However, theoretical analyses and comparisons of these approaches are lacking. In this study, we propose a unified group-by-group zipper-style movement model to describe different approaches synthetically and evaluate their performance. We derive the maximum throughput for cooperative driving plans of idealized unsignalized intersections and discuss how to minimize the delay of vehicles. The obtained conclusions shed light on future cooperative driving studies.