Large Language Models for Lane Change Decisions in Mixed Traffic for Autonomous Vehicles

UrbanMix: LLM-Guided Simulation of Mixed Autonomy Traffic with Heterogeneous Behavioral Profiles

Communication-aware predictive lane-change coordination for mixed autonomous traffic using SafeLane-VANET.

Lane changing is among the most critical operations in vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications, namely, Vehicular Ad Hoc Networks (VANETs), especially in mixed-traffic scenarios, where autonomous vehicles (AVs) share the road with human-routed and partially connected non-autonomous vehicles (Non-AVs). They are, in general, based on assumptions of instantaneous gap acceptance, heuristic MOBIL-style decision-making, and idealised V2V communication, which may result in delayed conflict resolution, simultaneous lane-change decisions, and reduced manoeuvre stability in dense traffic and under impaired V2V communication. To address these shortcomings, we present SafeLane-VANET, a novel communication-aware framework for predicting safe arbitration behaviour during cooperative lane changes in mixed AV/Non-AV traffic. It combines legality-aware target-lane scoring, V2X-based intent coordination, and short-horizon motion prediction to enable guarded lane changes while accounting for communication uncertainty and the diverse driving behaviour of other road users. To test the framework under various traffic densities, communication conditions, and mixed-fleet ratios, a reproducible SUMO-ns-3-Python co-simulation environment is developed that integrates mobility and packet logging. Experimental test results show that SafeLane-VANET consistently enhances safety, manoeuvre stability, and ride comfort compared to the conventional MOBIL-only case. The proposed framework shows in high-density degraded-communication cases a further improvement of the TTC p5 from 0.95 s to 1.55 s, a reduced percentage of unsafe manoeuvre executions (USM) from 25.1% to 10.4%, a decreased delay p50 from 198 s to 142 s, and a lowered jerk p95 from 4.78 m/s³ to 3.42 m/s³. Further ablation and real-world trajectory validation studies demonstrate that communication-aware intent coordination and predictive safety gating enhance the robustness under mixed-traffic uncertainty. These results show how SafeLane-VANET can be used to improve cooperative autonomy in a safer, more reliable way in a connected vehicular environment. Code is available at https://github.com/NSP310893/SafeLane-VANET.

Evaluation, optimization, verification of traffic rules for automated vehicles in freeway lane-changing scenario.

The safe deployment of Automated Driving Systems (ADS) requires traffic rules to be translated from qualitative natural-language clauses into measurable and executable decision criteria. However, formalized traffic rules may still contain vague propositions and missing quantitative parameters. This study proposes a framework for evaluating, optimizing, and verifying traffic rules for ADS in freeway lane-changing scenarios. In the evaluation stage, Metric Temporal Logic (MTL) is used to formalize relevant traffic rules, and a scenario-level method is developed to identify vague propositions and missing key parameters. In the optimization stage, Shanghai Naturalistic Driving Study (SH-NDS) data are used to develop an executable interpretation through a three-stage lane-changing model comprising judgment, execution, and stabilization. In the judgment stage, Shapley Additive Explanations (SHAP) and weighted quantile regression are used to identify key factors and derive lane-change initiation thresholds. In the execution stage, a Support Vector Machine (SVM) model classifies real-time interaction risk. In the stabilization stage, a dual-dimensional risk-assessment model determines longitudinal control responses. Verification results show that the proposed interpretation increases the minimum Generalized Time-to-Collision (GTTC) at lane-change initiation by 1.24 s, achieves an overall risk-classification accuracy of 94% during execution, and reduces Time-Exposed Time-to-Collision (TET) during stabilization by 1.934 s, corresponding to a reduction of 39.2%. The proposed framework provides a systematic method for converting qualitative traffic rules into measurable and executable requirements for ADS.

Traffic rule-embedded decision-making for automated vehicles in pedestrian-vehicle conflicts via deep reinforcement learning.

In pedestrian-vehicle crossing conflicts, automated vehicles (AVs) should avoid collisions and follow pedestrian-related traffic rules and social expectations. However, many existing decision-making models mainly optimize safety or efficiency. The behavioral effects of specific pedestrian-related rules remain insufficiently examined. Thus, actions that appear safe under surrogate metrics may still violate normative expectations in dynamic traffic. This study proposes a Traffic Rule-Embedded Decision-Making framework (TRE-DM) for ego-vehicle control in pedestrian-vehicle conflicts. 1) Pedestrian-related provisions from traffic laws and standards are decomposed into yielding, slowing, and braking. 2) These rules are formalized using MTL-informed triggering conditions and embedded into reinforcement learning through soft rule-shaped rewards. 3) The framework is evaluated using pedestrian-crossing conflicts reconstructed from the Shanghai Naturalistic Driving Study. Perturbation-based stress tests and external evaluations on three public datasets are also conducted. Results show that TRE-DM achieves balanced performance in safety, compliance-related behavior, comfort, and efficiency. Compared with human-driver trajectories, it reduces TIT (time-integrated TTC) by 37.5% and decreases yielding violations from 102 to 0. Compared with the agent using only collision-avoidance objectives, it reduces TIT by 32.2% and decreases collision events from 7 to 0. Ablation results further show that yielding is essential for reducing safety-critical failures, slowing supports earlier risk anticipation, and braking improves control smoothness. Trajectory reconstruction further shows earlier deceleration and safer lateral clearance under rule guidance, suggesting more rule-consistent and risk-aware evasive behavior. Overall, this study provides an interpretable rule-modeling and rule-embedding framework for improving AV behavior in pedestrian-vehicle conflicts.

Learning from Drivers: A Case-Based Reasoning Framework for Takeover Control in Conditionally Automated Vehicles

The rapid evolution of intelligent transportation systems has positioned conditionally automated vehicles (CAVs, SAE Level 3) at the forefront of automotive innovation. These vehicles represent a critical transition between human-driven and fully autonomous systems, in which safety and reliability depend on effective management of takeover control (TOC) events. This paper introduces a Case-Based Reasoning (CBR) framework to model, evaluate, and improve decision-making during control transitions using empirical and contextual data from both human and vehicular agents. The framework follows the CBR cognitive cycle of retrieval, reuse, revision, and retention to compare new TOC scenarios with previously observed cases. Each case integrates multimodal information, including driver personal traits, non-driving-related tasks, traffic density, and takeover urgency, as well as temporal and spatial performance metrics such as takeover time and steering behaviour. The time budget to system limitation is used as the determining outcome variable. By capturing and reusing experiential knowledge, the proposed framework enables adaptive and interpretable decision-making for Level 3 automation. It supports bidirectional learning between drivers and automated systems and provides a foundation for future Levels 4 and 5 vehicles to incorporate human-like reasoning in safety-critical decisions.

Design of Control Strategies for Autonomous Vehicles Targeting Aggressive Driving Behaviors in Mixed Traffic

Autonomous vehicles (AVs) will operate alongside human-driven vehicles for an extended transition period, during which aggressive human driving may become a major source of risk. This study proposes an integrated safety-control framework that combines real-world-data-driven behavior modeling with deep reinforcement learning to design longitudinal AV control strategies for mixed traffic. Aggressive, general, and defensive driving patterns are calibrated from the CitySim dataset, and dynamic aggressiveness is incorporated into an improved car-following model. A proximal policy optimization algorithm with a Kullback–Leibler penalty is then used to learn multi-objective strategies balancing safety, efficiency, comfort, and fuel economy in freeway and signalized-intersection scenarios. The results show that the behavior-aware strategies exhibit different strengths across traffic environments. On the freeway, the defensive-threshold strategy maintains a larger time headway, reduces positive acceleration, and lowers system-level fuel consumption, whereas the default, aggressive, and general strategies preserve higher traffic efficiency. At the intersection, signal control narrows the differences among strategies and limits the influence of longitudinal threshold settings on most evaluated indicators. These findings provide a quantitative basis for selecting behavior-aware control thresholds and designing robust AV strategies for mixed-autonomy traffic containing aggressive human drivers.