Problem

Existing microscopic traffic simulation tools, such as SUMO, VISSIM, and CARLA, primarily rely on simplified rule-based models or pre-recorded trajectories. This approach fails to accurately represent the complex, dynamic, and interactive nature of real-world human driving behaviors, especially in environments where autonomous vehicles (AVs) interact with human-driven vehicles. Current simulation models struggle to reproduce the full spectrum of individual behavioral variability and adaptation, a critical limitation as AVs integrate into human-dominated road networks. Furthermore, a structured and comprehensive survey linking advanced AI methodologies to the specific challenges of mixed autonomy traffic simulation has been lacking. Previous surveys often focused on simulation tools without detailing the underlying AI methods, or addressed ego-centric AV decision-making without adequately covering the modeling of surrounding traffic or offering a unified taxonomy of AI methods spanning individual behavior to full scene simulation.

Method

This survey introduces a structured taxonomy and comprehensive review of artificial intelligence methods for modeling and simulating both automated and human driving behavior in mixed autonomy traffic. The proposed taxonomy organizes methods into three core families: agent-level behavior models, environment-level simulation methods, and cognitive and physics-informed approaches. Agent-level models are further divided into single-agent methods (including imitation learning, trajectory prediction, reinforcement learning, inverse reinforcement learning, end-to-end planners, model-based policy learning, and foundation models for ego planning) and multi-agent methods (covering joint trajectory forecasting, centralized simulation via diffusion/token/regression-based models, decentralized simulation using MARL, multi-agent imitation, game theory, hybrid approaches, and foundation models for multi-agent coordination). Environment-level methods encompass generative world models (video-based and occupancy-based) and traffic scenario generation techniques (controllable synthesis and safety-critical scenario discovery). Finally, cognitive and physics-informed methods integrate insights from bounded rationality, cognitive architectures, attention/risk perception, trust dynamics, and physics-informed deep learning. Beyond the methodological overview, the survey also provides a chronological investigation of these methods, reviews evaluation protocols and metrics (open-loop vs. closed-loop), simulation tools, and publicly available datasets, aiming to bridge the gap between traffic engineering and computer science perspectives.

Results

The survey synthesizes a broad and fragmented body of literature, offering a unified, cross-disciplinary resource for understanding AI-driven traffic simulation. It highlights a clear acceleration and convergence of methodologies, particularly evident in the recent "generative AI era" (2023-2024), where transformer-based architectures, diffusion processes, and autoregressive generation have become shared building blocks across various domains, blurring traditional boundaries between prediction, planning, and simulation. Key limitations identified across method families include the persistent mismatch between open-loop trajectory accuracy and closed-loop interactive realism, computational costs hindering scalability, and challenges in ensuring causal validity and counterfactual responses under intervention. The review emphasizes the importance of integrating validated theories from cognitive psychology, human factors, and traffic physics into learning-based models to capture realistic human error modes, provide interpretable behavioral diversity, and enforce dynamic plausibility, capabilities often lacking in purely data-driven approaches.

Implications

Future research should prioritize bridging the fundamental gap between observational training data and the need for causal validity under interventional conditions in simulation. A critical need exists for a unified evaluation framework that rigorously assesses behavioral realism, interaction quality, counterfactual robustness, and long-horizon stability, moving beyond current open-loop metrics. Addressing the inherent trade-offs between scalability, behavioral fidelity, and controllability is crucial for practical simulation utility. Explicit modeling of heterogeneous human-AV behavioral dynamics and co-adaptation, rather than assuming static distributions, is essential for realistic mixed autonomy testing. There is also a significant need to bridge the disconnect between advanced research methods and their deployment in mainstream simulation tools, which often still rely on simpler rule-based models. Furthermore, diversifying geographic and cultural coverage in datasets, alongside enhancing representation of safety-critical events, is vital. The field is expected to converge towards unified simulation architectures or modular, differentiable pipelines that combine the strengths of world models, reactive simulators, and scenario generators. Finally, integrating cognitive and physics-informed constraints within data-driven models is seen as a key direction to enhance the explanatory power and predictive capabilities of future AI-driven traffic simulators, ensuring safer deployment of automated vehicles.