Jie Sun

Multi-agent traffic simulation transforms recorded driving data into interactive closed-loop environments, enabling scalable training and evaluation of autonomous driving systems. Recent methods increasingly adopt tokenized autoregressive generation, but often lack two complementary traffic-domain priors: a relational prior that exposes topology-induced conflicts beyond spatial proximity and a motion-vocabulary prior that addresses under-coverage, redundancy, and implausible primitives inherited from data-driven tokenization. To address these limitations, we propose \textbf{TopoKin}, a unified framework that introduces these dual priors into the agent--agent attention graph and the motion vocabulary. The topology-derived relational prior, implemented through conflict-topology-guided attention, performs agent--lane matching, candidate lane expansion, and lane conflict query before integrating the resulting relations into agent--agent attention. The motion-vocabulary prior, implemented through kinematic-constrained tokenization, proceeds through symmetry augmentation, DTW-based compression, kinematic and shape-support filtering, and tokenization with the refined vocabulary. Evaluated as an augmentation to existing simulators, TopoKin demonstrates overall gains over SMART and complementary effects in ablation, while transfer to CATK and TrafficBots and zero-shot evaluation on four additional datasets support cross-backbone applicability and improved cross-domain distributional alignment.

Autonomous driving policies trained with ordinary random sampling may under-expose long-tail hazardous interactions, leaving critical failures insufficiently corrected in closed-loop training. This limitation is particularly important for pre-crash safety, because rare high-risk scenarios such as intersection conflicts, lane-change merges, mixed-complex interactions, and rule-boundary violations disproportionately shape policy safety boundaries. This study reformulates training-sample organization as a pre-crash safety intervention and proposes a surrogate-safety-informed curriculum and corrective replay framework. The framework performs scenario-level closed-loop evaluation using at-fault collision, minimum time-to-collision (TTC), comfort degradation, off-road behavior, and red-light violations. These signals are integrated into a Pre-crash Risk Severity Index to stratify training scenarios, while failure taxonomy and interaction topology are used to diagnose hazardous events. Three interventions are compared under a common learner and training budget: risk-aware curriculum learning, targeted corrective replay, and their combination. Experiments on the Waymo Open Motion Dataset using the ScenarioMax-to-V-Max pipeline show that, with a unified BC-SAC learner and 81,920 training steps, the combined intervention achieves the best overall performance on an independent test set, including the highest accuracy, comfort, minimum TTC, and V-Max Score and the lowest at-fault collision rate. The results indicate that reallocating finite training exposure according to pre-crash risk can improve aggregate closed-loop performance and key safety surrogates without changing the backbone learner.

End-to-end (E2E) autonomous driving models have shown great promise, yet verifying their safety under complex traffic scenarios remains a critical barrier to deployment. Closed-loop simulation, with its advantages in reproducibility, safety, and corner-case coverage, has become an essential tool for validating E2E models. However, real-world traffic risks are inherently multi-dimensional, encompassing both interaction risk and perception risk. Existing frameworks focus predominantly on interaction risk while largely neglecting perception risk, leading to incomplete safety assessments. Moreover, current trajectory generation methods struggle to balance risk controllability with real-time execution, and perception degradation methods lack spatio-temporal consistency and physically interpretable control. To bridge these gaps, this paper proposes RiskMatrix-3D, a 3D Gaussian Splatting (3DGS)-based closed-loop testing framework that enables dual-risk controllable testing for E2E autonomous driving, with two decoupled parameters that independently regulate interaction risk and perception risk. For interaction risk, a Generate-Evaluate-Select (GES) strategy generates candidate trajectories via parallel sampling, evaluates close-range safety through inter-vehicle corner-point distances, and selects the trajectory matching the target risk level, achieving real-time execution. For perception risk, a World-coordinate Atmospheric Fog Scattering (WAFS) method constructs a physically grounded fog density field anchored in world coordinates, providing intensity-controllable, physically interpretable, and spatio-temporally consistent visibility degradation. Experiments on the nuScenes dataset with three E2E models (UniAD, LTF, VAD) demonstrate that both risk dimensions can be independently and monotonically controlled. A Dual-Risk Amplification Effect is further identified: perception degradation exacerbates interaction risk under identical interaction settings. Microscopic analysis reveals that BEV feature distortion and detection confidence decay jointly drive this amplification, offering fine-grained diagnostic signals for assessing E2E model robustness under coupled risk conditions.

This paper aims to model network-wide vehicle movements and traffic state evolution by considering the interactions among vehicles and link traffic conditions through a multi-agent imitation learning framework. The goal is to propose a framework for data-driven simulation of network-wide traffic by directly learning traffic behaviours from observed vehicle trajectory data. We thus develop a data-driven simulation model that applies the existing Multi-Agent Generative Adversarial Imitation Learning (MAGAIL) framework considering both Vehicle and Link agents to learn vehicles’ link-to-link transitions and within-link movements (travel speed), and road links’ traffic state evolution patterns simultaneously, which is referred to as MAGAIL-VL model. We evaluate the model’s performance in terms of its ability to generate realistic vehicle trajectories, route distribution, and link travel time and congestion state. The comparison with several deep-learning based benchmark models shows that vehicle and link agents in the developed MAGAIL-VL model can mimic real-world vehicle movement and link state change patterns, producing superior performance over other tested methods.