Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic Saeed Rahmani∗ , Shiva Rasouli† , Daphne Cornelisse‡ , Eugene Vinitsky‡ , Bart van Arem∗ , Simeon C. Calvert∗ 2
1 Department of Transport & Planning, Delft University of Technology, Delft, the Netherlands. Department of Industrial and Systems Engineering, University of Michigan, Dearborn, Michigan, USA 3 Tandon School of Engineering, New York University, New York, United States.
CORRESPONDING AUTHOR: Saeed Rahmani (e-mail: [email protected]).
arXiv:2604.12857v1 [cs.AI] 14 Apr 2026
This work was supported by EU Horizon 2020 Project Hi-Drive.
ABSTRACT Autonomous vehicles (AVs) are now operating on public roads, which makes their testing
and validation more critical than ever. Simulation offers a safe and controlled environment for evaluating AV performance in varied conditions. However, existing simulation tools mainly focus on graphical realism and rely on simple rule-based models and therefore fail to accurately represent the complexity of driving behaviors and interactions. Artificial intelligence (AI) has shown strong potential to address these limitations; however, despite the rapid progress across AI methodologies, a comprehensive survey of their application to mixed autonomy traffic simulation remains lacking. Existing surveys either focus on simulation tools without examining the AI methods behind them, or cover ego-centric decision-making without addressing the broader challenge of modeling surrounding traffic. Moreover, they do not offer a unified taxonomy of AI methods covering individual behavior modeling to full scene simulation. To address these gaps, this survey provides a structured review and synthesis of AI methods for modeling AV and human driving behavior in mixed autonomy traffic simulation. We introduce a taxonomy that organizes methods into three families: agent-level behavior models, environment-level simulation methods, and cognitive and physics-informed methods. The survey analyzes how existing simulation platforms fall short of the needs of mixed autonomy research and outlines directions to narrow this gap. It also provides a chronological overview of AI methods and reviews evaluation protocols and metrics, simulation tools, and datasets. By covering both traffic engineering and computer science perspectives, we aim to bridge the gap between these two communities. INDEX TERMS Artificial Intelligence, Automated Vehicles, Behavior Modeling, Microscopic Traffic
Simulation, Mixed Autonomy Traffic
I. INTRODUCTION
Autonomous vehicles (AVs) are now operating on public roads, a development that is reshaping transportation systems worldwide. As their deployment accelerates, the demand for rigorous, scalable, and safe testing environments increases. Simulation has emerged as a viable tool for this purpose by offering a cost-effective and controlled environment for evaluating AV performance across a wide range of conditions and driving scenarios [1], [2]. Nevertheless, a fundamental challenge in current microscopic traffic simulation is the realism of traffic participants’ behaviors. Existing popular simulation platforms like SUMO [3], VISSIM [4], and CARLA [5] typically model other vehicles’ behavior using simple rule-based
models or by replaying recorded trajectories. This simplified assumption fails to accurately represent the dynamic, interactive complexity of real-world driving. Traditional rulebased models, such as the Intelligent Driver Model for carfollowing [6] or MOBIL for lane changing [7], capture important aggregate phenomena but struggle to reproduce the full distribution of individual-level behavioral variability and adaptation observed in real traffic [8], [9]. This limitation becomes especially important in the era of mixed autonomy as AVs are gradually integrated into human-dominated road networks and must navigate environments characterized by behavioral heterogeneity. To address these limitations, researchers in automated driving and traffic simulation have increasingly relied on
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. VOLUME ,
1
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
artificial intelligence (AI) and deep learning methods. Recent breakthroughs in generative AI, foundation models, and learned world models have opened up new possibilities for synthesizing diverse, realistic, and highly interactive traffic agents and scenarios [10]–[13]. Techniques such as imitation learning [14]–[16], reinforcement learning [17], [18], and world models [19] are now widely used in simulation and, increasingly, in real-world AV deployments. Nevertheless, despite this rapid methodological progress, a comprehensive and structured taxonomy that links these advances to the specific challenges of mixed autonomy traffic simulation and evaluation remains missing. Existing surveys either focus on simulation tools without examining the AI methods behind them [1], [20], or they focus on AV decisionmaking from an ego-centric perspective without addressing how surrounding traffic should be modeled or the specific requirements that arise in interactive simulation and mixed autonomy traffic [21], [22]. Moreover, existing studies do not cover the full methodological range, from individual agent behavior models to environment-level scene generation, and from purely data-driven approaches to cognitive and physicsinformed models. Motivated by these gaps, this survey provides a comprehensive review and taxonomy of AI-driven methods for modeling both automated and human-driven vehicle behavior in mixed autonomy traffic simulation. Our methodological taxonomy uses several complementary dimensions. At the agent level, we distinguish between single-agent methods, which model individual vehicle behavior, and multi-agent methods, which explicitly capture the behaviors and bidirectional interactions among multiple agents. We then separate agent-level behavior models from environment-level methods, such as generative world models and scenario generation frameworks. To complement these pure data-driven approaches, we survey cognitive and physics-informed methods that incorporate theories of human attention, risk perception, decision-making, and vehicle dynamics, into learning-based architectures. Within each category, we classify methods by their core technical formulation, cover the formal problem setup, highlight representative studies, and discuss current limitations and open research directions. Beyond the methodological overview and synthesis, we provide a chronological investigation of these methods and review evaluation and benchmarking practices, simulation tools, and publicly available datasets that support the development, training, and validation of these models. Through these contributions, this survey aims to serve as a unified and cross-disciplinary resource that connects methodological advances in machine learning with the practical demands of mixed autonomy traffic simulation. It is intended for researchers and practitioners in transportation and traffic engineering, robotics, and computer science who seek a common methodological understanding across these disciplines. The rest of this paper is organized as follows. Section II describes our scope and taxonomy, as well as how we structure the survey. Section III reviews
2
related surveys. Sections IV–VII form the methodological core and cover agent-level, environment-level, and cognitive and physics-informed methods. Sections VIII–X present a chronological overview of method evolution, evaluation and benchmarking practices, and simulation tools and datasets. Section XI discusses open challenges and future directions, and Section XII concludes the survey. II. SURVEY SCOPE AND METHODOLOGY
Surveying AI methods for mixed autonomy traffic simulation requires navigating a broad and fragmented literature spanning transportation engineering, machine learning, and cognitive science. This survey therefore adopts a methods-oriented approach rather than a systematic review protocol, with the goal of organizing these methods into a coherent taxonomy, defining core ideas and terminology, discussing strengths and limitations, highlighting representative studies, and outlining future directions. The following subsections detail our scope, literature selection criteria, and the resulting taxonomy, and Figure 1 provides an overview of survey methodology, scope, and structure. A. Scope
Our focus is on AI-based methods for behavior modeling in microscopic simulation of mixed autonomy traffic. We cover both human driver models (SAE Level 0) and automated vehicle models, with emphasis on SAE Level 4 and 5 systems where the automated driving system performs the full dynamic driving task without requiring a handover to a human operator. We also include SAE Level 3 automated vehicles when they operate in automated mode. As a result, methods that model control transitions between automation and human drivers are outside the scope of this survey. The survey covers machine learning and deep learning methods used for decision-making, motion planning, trajectory generation, and prediction. While the main focus is motorized vehicles, we also include methods for pedestrians and cyclists, where necessary, due to their role in mixed traffic. We also review evaluation and benchmarking practices, simulation platforms, and public datasets used to develop and validate these models. Several topics are outside the scope of this survey. These include sensor and perception simulation, such as camera, LiDAR, and radar synthesis, detailed vehicle dynamics and low-level control, vehicle-to-everything communication protocols, graphics rendering and visualization, real-world deployment and regulatory validation frameworks, and nonground transportation systems. Traditional rule-based models are discussed for context, but they are not the primary focus. B. Literature Search and Selection Criteria
We collected literature through searches in Google Scholar, Web of Science, Scopus, IEEE Xplore, and arXiv. We used keyword queries that combine automated driving and mixed autonomy terms (for example, “automated driving,” VOLUME ,
Collect
Filter
Organize
Present
1. Literature Sources
2. Scope
3. Taxonomy Organization
4. Survey Structure
Primary Axes
Sec I--III: Background and Scope
Agent-level vs Environment-level Modeling and Simulation
Sec IV: Single-Agent Methods
Web of Science Scopus IEEE Xplore Google Scholar Arxiv.org
• • • • •
Inclusion Criteria ✓ AI/ML-based Methods ✓ Microscopic Simulation ✓ Behavior Modeling Focus ✓ Mixed Autonomy Context ✓ SAE Level 3–5 AVs ✓ Human Driving Behavior Models
Pure AI vs Cognitive and Physicsinformed Methods
Key Venues ML: NeurIPS, ICML, ICLR CV: CVPR, ICCV, ECCV Robotics: ICRA, CoRL ITS: IEEE IV, IEEE ITSC, TRB
• • • •
Benchmarks Monitored
Exclusion Criteria X Sensor Simulation X Pure Rule-based Methods X V2X Communication X Low Level Vehicle Dynamics X Graphics Rendering
Waymo Open Challenges nuPlan CARLA Challenge CommonRoad
• • • •
7-10 papers per category
Based on year, citation, venue
Single-agent vs Multi-agent Methods
Focus on decision-making, planning, behavioral models
Sec V: Multi-Agent Methods Sec VI: Env-level Simulation Sec VII: Cognitive and Physics-inf.
Secondary Axes • Theoretical Foundations • Chronological Order • Evaluation Metrics and Protocols • Datasets and Simulation Tools
Focus: Methods-oriented (not a systematic review)
Technical Core
Academic Databases
Sec VIII: Chronology of Methods Sec IX--X: Evaluation Landscape Sec XI: Discussion
~200 papers reviewed (20+ method categories)
FIGURE 1: Survey methodology, scope, and structure
“autonomous or automated vehicles,” and “mixed autonomy”) with representative AI method keywords and behaviormodeling task keywords (for example, motion prediction, trajectory generation, driver modeling, and scenario generation). We monitored key venues from both the machine learning and transportation communities. Citation tracking and snow balling was also used from foundational and recent papers to identify additional relevant works. We also tracked major benchmarks and challenges, including the Waymo Open Motion Dataset [23], Argoverse [24], nuPlan [25], and the Waymo Open Sim Agents Challenge (WOSAC) [26]. For cognitive and physics-informed methods, we also explored the human factors and cognitive science literature, including journals focusing on Human Factors and Cognitive Science. For each methodological category, we selected 7–15 representative papers based on five criteria: (1) a foundational contribution that introduced a new formulation or problem setting, (2) a clear methodological step beyond prior work, (3) impact and community adoption (citation count), (4) recent work with strong results on recognized benchmarks, and (5) coverage of different technical approaches within the category. For emerging topics such as foundation-modelbased driving and cognitive and learning integration, we placed more weight on novelty and coverage, and less weight on citation counts.
VOLUME ,
C. Taxonomy and Survey Structure
The survey is organized around a taxonomy that follows three main axes. These axes reflect key differences in problem formulation and in how models are used in simulation. Figure 2 provides an overview of the full taxonomy, and the section structure follows the same organization. The first axis distinguishes single-agent and multi-agent methods. Single-agent methods (Section IV) focus on learning the behavior of one agent, and they treat other agents as part of the environment. Multi-agent methods (Section V) model interactions explicitly and aim to capture coupled decision-making among agents. Within multi-agent methods, we separate joint trajectory forecasting from interactive simulation methods. We then divide interactive simulation into centralized approaches, where one learned model produces the next states for all agents, and decentralized approaches, where each agent has its own policy and scene-level behavior results from their coupled execution. The second axis distinguishes agent-level and environmentlevel methods. Agent-level methods (Sections IV–V) output agent actions or trajectories, which define behavior in the scene. Environment-level methods (Section VI) model how the world evolves and provide the conditions under which behavior models operate. In this part, we cover generative world models and traffic scenario generation methods. The third axis distinguishes data-driven and theoryinformed methods. The main body of the survey in Sections IV to VI covers data-driven methods organized by the first 3
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
Imitation Learning / Behavior Cloning Trajectory Prediction
Single-Agent Methods
Reinforcement Learning
Open and/or Closed Loop
Inverse RL/ Adversarial IL End-to-End Planning Model-Based Policy Learning Foundation Models
Trajectory Prediction
Joint Trajectory Prediction
Centralized Modeling
Token/Autoregressive Methods
Diffusion-based Methods
Regression-based, Reactive Methods
Multi-Agent
MARL/Self-Play
Methods
Multi-Agent IL and GAIL
Decentralized Modeling
AI Methods
Game-Theoretic Models Hybrid Data-Driven
Foundation Modelbased Coordination
For Mixed Traffic
Environment-level Methods
Video World Models
World Models Scene and Scenario Generation Bounded Rationality & Trust
Cognitive & Physics-informed
Agentic Foundation Models
Occupancy Worls Models Scene Initialization & Scenario Syn. Safety-critical Scenario Gen. Bounded Rationality and Dec. Making Trust Adaptation in human-AV traffic Cognitive Architecture
Cognitive and Attention
Attention and Perception-guided
Physics-informed
Physics-informed & Field-based
Risk Fields and Potential-based
FIGURE 2: Taxonomy of AI methods for modeling mixed-autonomy traffic
two axes. Section VII covers cognitive and physics-informed methods that incorporate domain knowledge from cognitive psychology, human factors, and traffic physics. Within each category, we organize methods by their main technical foundation, and identify classes of methods, such as imitation learning, reinforcement learning, diffusionbased generation, and game-theoretic reasoning. We treat the open-loop versus closed-loop distinction as an important deployment choice, but not as a main taxonomy dimension, since many methods can support both settings. Beyond this taxonomy of methods, Section VIII summarizes the evolution of methods over time. Section IX reviews evaluation and benchmarking practices, including open-loop metrics, closedloop protocols, and realism assessment, and Section X reviews simulation tools, datasets, and benchmarks used to train and validate these models.
sampling-based, and early learning-based methods. This study adopts a purely ego-vehicle perspective and pre-dated the transformer and foundation model era. Kuutti et al. [27] surveyed deep learning applications to autonomous vehicle control by covering end-to-end learning, behavior cloning, and deep RL for lateral and longitudinal control tasks. However, they do not cover traffic agent modeling or multi-agent interaction and remain focused on ego vehicle’s control. Grigorescu et al. [28] surveyed deep learning across the full autonomous driving stack, perception, planning, and control, but without addressing multi-agent simulation or mixed autonomy aspects. Kiran et al. [29] taxonomized deep reinforcement learning for autonomous driving and extended the scope to cover behavior cloning, inverse RL, and simulation platforms, but the treatment remains limited to reinforcement learning, egocentric, and does not consider surrounding traffic modeling. More recently, Chen et al. [30] reviewed over 270 papers on end-to-end autonomous driving, providing an in-depth III. RELATED SURVEYS analysis of imitation learning, reinforcement learning, and Table 1 summarizes the scope of the most relevant existing surworld model-based approaches, but their scope is restricted veys and highlights their contributions and the complementary to the end-to-end paradigm and is ego-centric only. coverage offered by this survey. Early foundational surveys On the traffic simulation and mixed autonomy side, Di established the landscape of AV decision-making and control. and Shi [21] provided the first survey explicitly bridging Schwarting et al. [22] provided an overview of planning transportation engineering and AI for mixed autonomy and decision-making paradigms, covering optimization-based, 4
VOLUME ,
AV control. Although their cross-disciplinary scope is the closest antecedent to ours, the focus is on AV control policy design rather than on the broader landscape of behavior and interaction modeling methods for traffic simulation and multiagent setup. Moreover, recent foundation and world model techniques, as well as environment-level simulation methods, are not discussed in their work. Farah et al. [1] addressed the practical challenges of modeling automated driving in microscopic traffic simulations for traffic performance evaluations, identified six key modeling aspects, and reviewing how existing simulation studies handle mixed traffic with varying automation levels. Their work highlights the gap between simulation practice and behavioral realism but does not cover AI or deep learning methods. Chao et al. [31] surveyed visual traffic simulation models at multiple scales and discussed early data-driven animation techniques for AV testing. But their study stays focused on graphical simulation. Chen et al. [32] provided a review of data-driven traffic simulation and covered imitation learning, RL, deep generative models, though without addressing cognitive or physics-informed methods, world models, and environment-level simulation and scenario generation techniques. Other related, but more narrowly scoped, surveys include the work of Ding et al. [33] on safety-critical scenario generation methods, Li et al. [20], who reviewed three decades of open-source AV simulators. They categorized over 20 platforms but did not examine the AI methods behind the simulated behaviors. In summary, prior surveys have made valuable contributions by mapping key aspects of the automated driving and AI landscape, and this survey builds upon their foundations. However, taken together, they leave several important gaps when viewed from a simulation and evaluation perspective, especially in the context of mixed automated and human traffic. First, to our knowledge, no existing survey spans the complete methodological spectrum from agent-level methods (single-agent and multi-agent) through environment-level approaches (world models and scenario generation) to cognitive and physics-informed models within one unified taxonomy. Second, they do not address the full set of modeling and deployment requirements that arise in traffic simulation, such as the open-loop versus closed-loop gap, counterfactual validity under intervention, long-horizon rollout stability, and interaction consistency among heterogeneous agents. Third, prior works generally lack a dedicated, mixed-autonomyfocused review of evaluation practice, including the metrics, benchmarks, datasets, and simulation tools through which models are validated and compared in interactive traffic settings. Finally, our survey is distinguished by its explicit focus on behavior modeling for both automated vehicles and human drivers, treating the interaction dynamics between these heterogeneous agents as the central object of study rather than adopting a purely ego-vehicle or purely trafficflow perspective.
VOLUME ,
IV. SINGLE AGENT METHODS
Single-agent methods focus on modeling the behavior of an individual vehicle. These methods treat other agents as part of the environment’s dynamics, assuming their behaviors are fixed, predefined, or modeled separately. In mixed autonomy traffic simulation, single-agent methods serve two primary purposes. First, they develop and train driving policies for the AV itself, enabling decision-making and trajectory planning in response to surrounding traffic. Second, they model individual human drivers or other road users; these individual models are then replicated across multiple agents to produce the surrounding traffic that populates the simulated environment. While this simplification may not capture full multi-agent interaction complexity, single-agent approaches offer practical advantages: computational efficiency, simpler training pipelines, and, importantly, closer alignment with real-world operation in mixed autonomy. In practice, an automated vehicle under test (often called the ego vehicle) does not have direct access to other road users’ internal states, such as their intentions, goals, or planned maneuvers; it must instead act based on partial observations and uncertain inferences from motion cues, context, and traffic rules. Single-agent formulations naturally match this information structure by learning policies conditioned on observable history and scene context, without assuming access to other agents’ latent intent. Also, when deployed across multiple instances, each controlled by an independent single-agent model, these methods create diverse traffic scenarios without explicitly modeling inter-agent coordination. In the following sub-sections, we review single-agent methods and summarize their main assumptions, modeling choices, and key characteristics. A comparative summary of these methods is provided in Table 2. A. Imitation Learning
Imitation learning (IL) methods enable agents to directly learn driving behavior by observing expert demonstrations from real-world datasets. The simplest and most widely used form, behavior cloning (BC), frames the problem as supervised learning: given a dataset D = {(si , ai )}N i=1 of N state-action pairs collected from expert demonstrations, BC learns a policy πθ parameterized by θ by minimizing the empirical loss: N
L(θ) =
1 X ℓ(πθ (si ), ai ) N i=1
(1)
where ℓ(·, ·) is typically mean squared error for continuous actions or cross-entropy for discrete actions. In driving contexts, the state si commonly includes sensor inputs or bird’s-eye view representations along with vehicle state and contextual data. The action ai comprises control commands (steering, throttle, brake) or planned trajectory waypoints. Imitation learning can operate in both open-loop and closed-loop settings: in open-loop mode, a trained policy generates a fixed trajectory from an initial observation without environmental feedback; in closed-loop mode, the policy is queried at 5
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
TABLE 1: Comparison of this survey with related surveys across key dimensions. ✓ = covered; (✓) = partially covered; – = not covered. Mixed Autonomy
Traffic Agent
MultiAgent
Focus
Model.
Methods
Schwarting et al. [22]
–
(✓)
–
–
–
–
–
Kuutti et al. [27]
–
–
–
–
–
–
✓
–
Grigorescu et al. [28]
–
(✓)
–
–
–
–
(✓)
–
Chao et al. [31]
(✓)
✓
(✓)
–
–
–
✓
(✓)
Di and Shi [21]
✓
–
(✓)
(✓)
–
(✓)
–
✓
Kiran et al. [29]
–
–
–
–
–
–
(✓)
–
Farah et al. [1]
✓
–
–
–
–
–
✓
(✓) –
Survey
World Models
Scenario Generation
Cognitive / Physics
Sim. Tools & Datasets
CrossDiscipl. –
Ding et al. [33]
–
–
(✓)
–
✓
–
(✓)
Li et al. [20]
–
–
–
–
–
–
✓
–
Chen et al. [30]
–
(✓)
–
(✓)
–
–
(✓)
–
Chen et al. [32]
–
✓
(✓)
–
–
–
✓
–
This survey
✓
✓
✓
✓
✓
✓
✓
✓
each simulation step, receiving updated observations and encounter states that differ from those seen in the training producing reactive behavior. For mixed autonomy simulation, data, either due to differences in the initial state distribution or IL enables driver models that replicate behavioral patterns deviations caused by the learned policy itself. This mismatch from naturalistic driving datasets, and when deployed across can lead to compounding errors. multiple independent instances, these models can populate Several methods address this by incorporating closed-loop simulated environments with diverse, human-like traffic feedback during training. Dataset Aggregation (DAgger) [35] participants. alternates between rolling out the learner’s current policy and Imitation learning for automated driving began with querying an expert for corrective labels in the visited states. ALVINN [14], which showed that neural networks could DART [71] injects structured noise into expert demonstrations map raw camera images directly to steering commands. during collection, HG-DAgger [72] improves scalability by PilotNet [15] scaled this idea with deep CNNs trained on learning an intervention rule that requests expert control large datasets and demonstrated that end-to-end policies only in states predicted to be risky, and MEGA-DAgger [73] could generalize across diverse road conditions. A key extends DAgger to settings with multiple imperfect experts. limitation of these reactive approaches, that is the inability to Beyond interactive correction, Urban Driver [37] presents reason about goals, was addressed by Conditional Imitation an offline approach that builds a differentiable data-driven Learning (CIL) [16]. CIL conditions the policy on high-level simulator from perception outputs and HD maps. navigation commands through a branched architecture and Despite its simplicity and scalability, imitation learning enables goal-directed behavior. ChauffeurNet [34] moved faces several persistent challenges with direct implications away from raw sensor input to mid-level bird’s-eye view for simulation quality. Distributional mismatch can cause representations and introduced trajectory perturbation as a BC-trained agents behave unrealistically when the simulation form of data augmentation to reduce distributional shift. diverges from recorded scenarios, limiting their utility as “Learning by Cheating” [70] proposed a two-stage, teacher- reactive traffic participants [74]. This may arise from covariate student training paradigm in which a privileged agent with shift, dataset bias, and causal confusion. Interactive correction access to ground-truth environment information supervises methods such as DAgger [35] mitigate this but require a purely vision-based sensorimotor agent. Their framework repeated expert queries in simulation, which is expensive substantially outperforming prior methods on the CARLA [5] and difficult to scale for high-dimensional observations and and NoCrash benchmarks. TransFuser [36] then brought multi- long-horizon tasks. Data coverage limitations compound this modal perception into this framework by fusing camera and problem: expert data quality and consistency are critical LiDAR features through transformer attention at multiple but is typically unavailable for rare edge cases, which are resolutions. precisely the scenarios most relevant for AV safety testing. A fundamental limitation of standard behavior cloning Together, these challenges mean that IL-based agents may is covariate shift. The policy is trained on expert-induced match logged driving distributions while failing to respond state distributions, whereas during deployment the agent may plausibly to interventions, which is the core requirement for
6
VOLUME ,
TABLE 2: Summary of single-agent behavior modeling methods for mixed autonomy traffic simulation. Method
Training Signal
Loop
Strengths
Limitations
Representative Studies
Imitation Learning Expert / Behavior demonstrations Cloning (supervised)
OL, CL
Simple and scalable; direct data-to-policy mapping; strong baselines with large datasets
Covariate shift and compounding errors; causal confusion; limited robustness to out-of-distribution states
ALVINN [14], PilotNet [15], CIL [16], ChauffeurNet [34], DAgger [35], TransFuser [36], Urban Driver [37] CoverNet [38], MultiPath [39], TNT [40], DenseTNT [41], LaPred [42], PRIME [43], MID [12] Kendall et al. [18], Roach [44], BC-SAC [45], Think2Drive [46], HAIMDRL [47], Safe RL [17] MaxEnt IRL [48], GAIL [49], Kuefler et al. [50], PSGAIL [51], AIRL [52], RAIL [53] NMP [54], UniAD [55], VAD [56], SparseDrive [57], InterFuser [58], ThinkTwice [59], NAVSIM [60] World Models [61], Dreamer [62], DreamerV3 [63], MILE [64], AdaWM [65], Imagine-2Drive [66] Agent-Driver [10], DiLu [11], LanguageMPC [67], GPTDriver [68], DriveGPT4 [69]
Trajectory Prediction
Logged trajectories (supervised)
OL
Multi-modal forecasting; mature standardized benchmarks; captures distributional structure of future motion
Marginal prediction; non-reactive; displacement metrics miss interactive quality
Reinforcement Learning
Reward signal (trial-and-error)
CL
Discovers novel strategies via exploration; no expert demonstrations needed; inherently reactive
Sample inefficiency; reward engineering difficulty; safety during exploration; simto-real gap
IRL / Adversarial Imitation
Expert demos (adversarial matching)
CL
Recovers interpretable reward functions; robust to compounding errors; captures driving preferences
Training instability; mode collapse; sensitivity to discriminator design; scaling difficulty
End-to-End Learned Planners
Expert demos + perception labels
OL, CL
Joint perception–planning; end-to-end gradient flow; unified architecture
Ego-status bias; open-loop metric gap; limited closed-loop generalization
Model-Based Policy Learning
Learned world model + reward
CL
Sample-efficient via imagination; decouples dynamics learning from policy optimization
Model bias; compounding errors in imagined rollouts; sensitive to model accuracy
Foundation Model Approaches
Pretrained LLM/VLM + prompts
OL, CL
Commonsense reasoning; interpretable chain-of-thought; instruction-following capabilities
Latency bottleneck; grounding gap; hallucination risk; impractical for multiagent at scale
Loop: OL = open-loop, CL = closed-loop. Methods listed under both can operate in either mode depending on deployment context.
closed-loop simulation. Future directions include uncertainty quantification for detecting out-of-distribution operation, retrieval-based methods that ground predictions in similar recorded situations, and hybrid approaches that use IL as initialization for reinforcement or adversarial training. The next section discusses trajectory prediction, a related family that shares the supervised training signal of behavior cloning but differs in deployment role and the core problem it addresses. B. Trajectory Prediction
Trajectory prediction forecasts future positions of agents given observed motion history and environmental context. Although both trajectory prediction and imitation learning learn from recorded expert data and can produce future waypoints as output, they differ in role and training objective. An imitation learning policy is trained as an actor that produces actions for an agent to execute. A trajectory prediction model is trained as a passive observer that forecasts where other agents will be; because it does not act, it does not compound its own errors across time steps in the same way. When a trajectory prediction model is deployed in closed-loop by re-querying it at each time step, it is operationally equivalent to an imitation learning policy, but without the training-time mechanisms, such as DAgger or adversarial correction, that handle the resulting distributional mismatch. Single-agent trajectory prediction models learn a conditional distribution over future trajectories: P (Y | X, C), where X = {x1 , x2 , . . . , xTh } represents the observed trajectory over history horizon Th , C VOLUME ,
encodes contextual information, and Y = {y1 , y2 , . . . , yTf } denotes the predicted future trajectory over forecast horizon Tf . These models are predominantly evaluated in open-loop settings, where predicted trajectories are compared against ground-truth recordings using displacement metrics. Foundational approaches established key paradigms for encoding agent history, scene context, and multi-modal output. CoverNet [38] treats prediction as classification over a fixed set of trajectory anchors derived from expert data. An important aspect is multimodality: at any decision point, an agent may turn left, go straight, or turn right, each representing a valid future. MultiPath [39] extended CoverNet with learnable anchor trajectories corresponding to distinct behavioral modes to enable efficient multi-modal predictions. Goal-conditioned approaches improve accuracy for long-horizon predictions by decomposing the task into two sequential stages: the model first predicts a plausible goal location, such as a target position or waypoint, and then generates a trajectory conditioned on that endpoint. This two-stage structure anchors the trajectory to a specific destination, which constrains the prediction space and reduces the uncertainty that accumulates in free rollouts over long time horizons. TNT [40] and DenseTNT [41] predict target goal locations or probability distributions from lane centerlines and generate trajectories conditioned on selected endpoints. Similarly, LaPred [42] leveraged lane-aware representations to structure predictions in highway and urban environments. Recently, generative models offer alternative approaches to capturing trajectory multimodality. PRIME [43] improved ro-
7
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
bustness under imperfect tracking by constraining predictions to dynamically feasible trajectories. MID [12] formulated the trajectory prediction task as a reverse process of motion indeterminacy diffusion that progressively refines noisy samples into determinate trajectories. These generative approaches capture richer distributional structure than deterministic or anchor-based methods, but typically at higher computational cost. Despite significant progress, single-agent trajectory prediction faces fundamental limitations for simulation applications. The independence assumption among agents could potentially generate inconsistent or colliding trajectories. Rare events and safety-critical scenarios remain challenging due to underrepresentation in training data. As a result, models also show limited generalization across geographic regions and driving cultures. These limitations motivate joint multiagent prediction methods that explicitly model inter-agent dependencies, and closed-loop agent models that adapt to evolving simulation states. C. Reinforcement Learning
Reinforcement learning (RL) provides a framework for learning driving policies through trial-and-error interaction with the environment, guided by reward signals rather than relying on labeled data. In RL, an agent observes state st at time step t, executes action at according to policy πθ (at |st ) parameterized by θ, receives scalar reward rt , and transitions to new state st+1 . The objective is to maximize expected cumulative discounted reward: " T # X t J(θ) = Eπθ γ rt (2) t=0
and showed that RL reward signals significantly improve robustness in challenging scenarios. More recently, Li et al. [46] proposed Think2Drive, a model-based RL approach using latent world models, which illustrates the potential of imagination-based policy learning for driving. Safety-aware formulations have also emerged as viable paradigms. ShalevShwartz et al. [17] introduced hard safety constraints into RL driving policies, and Wen et al. [76] developed Parallel Constrained Policy Optimization to enforce safety constraints during training. Recently, human-in-the-loop approaches offer complementary efficiency gains. HAIM-DRL [47] treats humans as AI mentors providing sparse interventions during training to guide the exploration toward safer and more efficient policies. Sample efficiency in RL is a primary concern. RL training requires millions of environment interactions with heavy simulation reliance and associated sim-to-real gap issues. Reward engineering is also non-trivial; simple rewards may lead to unintended behaviors while complex multi-objective rewards are hard to tune and may produce brittle policies. Generalization to unseen environments is also a challenge similar to other learning-based methods. For mixed autonomy simulation specifically, current RL methods focus primarily on ego-vehicle control rather than generating diverse traffic agent behaviors. Therefore, future directions could focus on sample-efficient algorithms through world models and offline RL, improved sim-to-real transfer, safe exploration guarantees, and generative diverse and human-like behaviors. D. Inverse Reinforcement Learning and Adversarial Imitation Learning
Inverse reinforcement learning (IRL) infers reward functions from expert demonstrations by assuming that observed behavior is approximately optimal with respect to unknown rewards. Unlike behavior cloning, which directly maps states to actions, IRL recovers the underlying objectives driving expert behavior, and therefore, enables generalization to novel states by optimizing the recovered rewards. Also, unlike standard RL, which requires manually specifying reward functions, IRL extracts these objectives from data, which is a significant advantage for driving, where human preferences involve complex trade-offs between speed, comfort, safety, and social norms. Mathematically, given expert demonstrations DE = {(si , ai )}N i=1 , IRL seeks to find reward function rϕ parameterized by ϕ such that the expert policy outperforms alternatives: h i ϕ∗ = arg max EπE [rϕ ] − max Eπ [rϕ ] (3)
where γ ∈ [0, 1] is the discount factor and the expectation is taken over trajectories generated by policy πθ . For automated driving, reward functions encode objectives such as progress toward goal, lane keeping, collision avoidance, and passenger comfort. Unlike imitation learning, which requires expert demonstrations, RL discovers policies through exploration. This is a strength enabling discovery of novel strategies but also a drawback as it requires extensive interaction with the environment to learn a robust policy. RL methods are inherently closed-loop because the agent continuously receives environmental feedback and adapts its actions. Foundational work demonstrated RL viability for increasingly complex driving tasks. Kendall et al. [18] achieved the first successful deep RL application to real-world driving by training DDPG agents for lane following using monocular π ϕ camera input. Toromanoff et al. [75] introduced implicit affordances combined with Rainbow-IQN-Apex, achieving where πE denotes the expert policy and the expectation is the first RL agent capable of end-to-end urban driving over trajectories. However, this formulation is inherently illincluding traffic light detection. Zhang et al. [44] introduced posed, which means that many reward functions can explain Roach, an RL expert trained using PPO that serves as the same behavior. Maximum Entropy IRL [48] resolves a superior “coach” for imitation learning. Waymo’s BC- this ambiguity by selecting reward functions that make the SAC [45] combined behavior cloning initialization with expert demonstrations most likely under maximum entropy Soft Actor-Critic on over 100,000 miles of real-world data trajectory distributions. 8
VOLUME ,
Adversarial imitation learning methods reframe the imitation problem as a two-player game: a policy generates behavior while a discriminator tries to distinguish it from expert demonstrations. Generative Adversarial Imitation Learning GAIL [49] bypasses explicit reward recovery by directly matching the learner’s state-action occupancy measure to that of the expert. Instead, it directly trains the policy to produce the same distribution of (state, action) pairs that the expert visits. The discriminator is what drives this. it learns to tell the policy’s behavior apart from the expert’s, and the policy is trained to fool it, which pushes the two distributions closer together. The GAIL objective is:
lenges worth noting. Training instability remains significant, particularly in multi-agent settings where discriminators must distinguish expert behavior while each agent’s behavior is highly dependent on surrounding vehicles, which makes the discriminator’s feedback noisy and unreliable. Multi-agent extensions that explicitly address discriminator instability arising from irrelevant agent interactions are discussed in Section C. Moreover, recovered rewards may not transfer across different road geometries or traffic densities, which limits generalization. For multi-agent simulation, scaling remains difficult. Training independent GAIL agents leads to non-stationary dynamics as each agent’s learning changes the environment for others, and centralized approaches can min max Eπθ [log Dψ (s, a)] πθ Dψ (4) become computationally expensive and disregard the decentralized nature of mixed autonomy traffic. Future directions + EπE [log(1 − Dψ (s, a))] − λH(πθ ) may include more stable multi-agent adversarial training, where Dψ is a discriminator parameterized by ψ distin- combining IRL with offline RL to reduce environment guishing expert from learner state-action pairs, and H(πθ ) interaction requirements, and learning hierarchical reward denotes the entropy of policy πθ , weighted by λ > 0. structures that separate strategic intent from tactical control. Adversarial Inverse Reinforcement Learning (AIRL) [52] Also, integrating safety constraints into the reward recovery extends this framework by structuring the discriminator to process seems a promising direction. recover a disentangled reward function that transfers across environments with different dynamics. For mixed autonomy E. End-to-End Learned Planners simulation, these methods are particularly valuable as they End-to-end learned planners represent a major paradigm learn interpretable reward functions capturing human driving within single-agent AV behavior modeling, directly mapping preferences and enable generation of realistic and diverse sensor observations to planned trajectories through unified traffic agent behaviors without requiring hand-crafted reward differentiable architectures. By jointly learning perception, engineering. prediction, and planning, these systems produce driving IRL and adversarial IL methods are inherently closedpolicies that can serve as the automated ego vehicle in mixed loop during training, as the learner must interact with autonomy simulation, as alternative to rule-based autopilots the environment to generate trajectories for discriminator in tools like CARLA or SUMO. While end-to-end planners comparison. Once a reward function is recovered, it can be are often trained using imitation learning objectives, they used to train new policies via standard RL in closed-loop are treated as a distinct section in this survey because they settings, or the GAIL-trained policy itself can serve as a are defined by their architectural scope and can be trained closed-loop simulation agent. However, recovered reward with IL, RL, or hybrid objectives. However, IL is defined by functions can also be applied in open-loop contexts, for its learning signal, that is mimicking expert behavior, and instance, to score or rank pre-generated trajectory candidates applies regardless of architecture. End-to-end planners learn without environmental feedback. a mapping: Kuefler et al. [50] first applied GAIL to highway driving πθ : (I, M, E) → τ (5) using NGSIM data and demonstrated that adversarial imitation produces more realistic lane-keeping and car-following where I represents sensor inputs, M denotes map information, behavior than behavior cloning, particularly over longer E captures ego vehicle state, and τ = {y1 , y2 , . . . , yTp } is the rollout horizons where BC suffers from compounding errors. planned trajectory over planning horizon Tp . The policy πθ PS-GAIL [51] extended this to multi-agent settings via shared is parameterized by neural network weights θ. These models policy networks and curriculum learning. Bhattacharyya et can be evaluated in either open-loop mode or closed-loop al. [53] further augmented GAIL with semantic rewards mode. (RAIL) to disentangle latent driving style factors for controlNMP [54] established the foundational idea that planning lable behavior generation. Huang et al. [77] modeled driving quality depends on rich intermediate scene representations, intentions as discrete latent variables to enable learning encoding LiDAR and HD maps into learned cost volumes personalized rewards that generalize to unseen conditions. that score candidate trajectories. TransFuser [36], [80] and Wang et al. [78] and Sackmann et al. [79] both applied AIRL InterFuser [58] demonstrated that multi-modal transformer to highway scenarios, the former augmenting with semantic fusion across camera and LiDAR streams produces inrewards for stability, the latter recovering interpretable reward terpretable, safety-aware AV behavior. ST-P3 [81] introfunctions capturing diverse driving styles. duced spatial-temporal feature learning that tightly couples For simulation applications, especially in the context of scene understanding with trajectory generation. A pivotal mixed autonomy traffic, IRL and adversarial IL face chal- contribution, UniAD [55], introduced a planning-oriented VOLUME ,
9
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
philosophy by hierarchically organizing tracking, mapping, motion forecasting, and occupancy prediction to facilitate planning. It established the paradigm that intermediate task supervision improves planning quality. VAD [56] and SparseDrive [57] demonstrated that efficient vectorized and sparse representations can sustain high behavioral quality, a property that matters for simulation deployments where computational budget is shared across many concurrently running agents. ThinkTwice [59] addressed the refinement problem by conditioning trajectory generation on imagined future scenes and produced more consistent and anticipatory AV behavior. PP&P [82] contributed the complementary insight that inspectable intermediate representations enable verification of what the agent perceives before its behavioral decisions are executed. This property is particularly important for simulation-based AV testing workflows. More recently, diffusion-based architectures have been applied directly to the planning task by leveraging iterative denoising to model multi-modal driving behavior distributions more expressively than regression-based decoders. Diffusion Planner [83] jointly models ego planning and neighboring vehicle prediction within a single diffusion transformer, using classifier guidance to adapt driving style (safety, comfort, speed) without retraining. DiffusionDrive [84] addresses the computational bottleneck of iterative denoising through truncated diffusion with anchored Gaussian priors, reducing the reverse process to two steps and enabling real-time inference. End-to-end planners have shown impressive benchmark performance; however, recent studies have exposed fundamental evaluation concerns. Zhai et al. [85] and Zhiqi et al. [86] demonstrated that simple MLP-based models using only ego vehicle state can achieve comparable L2 displacement error to sophisticated perception-based methods on nuScenes. This suggests that standard open-loop metrics may not adequately measure planning capability. This “ego status bias” arises because datasets are dominated by straight driving scenarios where maintaining current velocity is nearly optimal. NAVSIM [60] proposed a middle ground through non-reactive simulation with safety-aware metrics (progress, time-to-collision), finding that simpler methods can match complex architectures on challenging scenarios. In other words, the gap between open-loop metrics and closed-loop performance means models with low displacement error may still produce unsafe behavior when deployed reactively. For simulation applications, end-to-end planners provide architectural insights for building traffic agents but require closed-loop training mechanisms to serve as truly reactive simulation participants.
environment for every policy update. A common formulation introduces latent state zt summarizing observation history, using encoder eθ mapping observations to latent states and latent transition model pθ modeling dynamics in latent space: zt = eθ (o≤t ),
zt+1 ∼ pθ (zt+1 | zt , at )
(6)
where o≤t = {o1 , o2 , . . . , ot } denotes the observation history up to time t, and at is the action taken at time t. The ego policy πϕ (at | zt ) parameterized by ϕ is optimized by rolling out imagined trajectories in latent space and maximizing expected return: "H # X k J(ϕ) = Ezt ,{at ,zt+1 } γ r(zt+k , at+k ) (7) k=0
where H is the imagination horizon, r(zt , at ) is a reward function operating on latent states, and the expectation is over latent rollouts generated by the learned dynamics model and policy. This separates learning predictive models of the world from learning policies through imagination within those models. Compared to model-free RL, which requires direct environment interaction for every gradient step, policy improvement here happens through simulated rollouts inside learned models, which substantially improves sample efficiency but makes performance sensitive to model accuracy and bias. Model-based approaches are fundamentally closed-loop methods. The learned world model provides the feedback loop during imagination-based training, and the resulting policy is deployed as a reactive agent that conditions actions on current observations. However, the world model itself can also be used in open-loop fashion, for instance, to generate multiple candidate future trajectories under different action sequences for offline scoring or trajectory ranking without environmental feedback. The key distinction from the generative world models surveyed in Section A is functional rather than architectural: here the world model serves as a training substrate for a specific ego policy, whereas Section A covers world models deployed as reusable simulation infrastructure serving multiple downstream applications including policy evaluation, synthetic data generation, and counterfactual testing. Some model families, such as latent dynamics models and autoregressive token predictors, could appear in both contexts. Ha and Schmidhuber [61] introduced the foundational paradigm of training recurrent generative world models and optimizing compact controllers using extracted latent features. The Dreamer line of work established scalable recipes for jointly learning latent dynamics and improving behavior through differentiable imagination. Dreamer [62] learns a recurrent state-space model (RSSM) and optimizes F. Model-Based Policy Learning actors using value gradients backpropagated through imagined Model-based policy learning trains driving policies by leverag- trajectories. DreamerV2 [87] strengthened this approach ing learned dynamics models, also known as “world models”, with discrete latent representations to improve model expresto predict how the environment evolves under candidate siveness. DreamerV3 [63] demonstrated robustness across actions, rather than requiring direct interaction with the diverse domains with a single hyperparameter configuration 10
VOLUME ,
and established world-model-based RL as a general-purpose and most current systems operate at reduced frequency or approach. For automated driving specifically, MILE [64] with asynchronous decision-making. adapts imagination-based learning to urban driving by jointly This class of models has become very popular in recent learning compact latent world models and ego policies from years. Agent-Driver [10] proposes an LLM-centered driving offline demonstrations. AdaWM [65] targets the distribution agent equipped with tool libraries (for perception and premismatch between pretrained dynamics models and planning diction queries), memory modules storing past experiences, policies and proposed adaptive finetuning strategies that and explicit chain-of-thought reasoning steps to yield interimprove policy performance without retraining the world pretable decisions. DiLu [11] develops a knowledge-driven model from scratch. Imagine-2-Drive [66] pairs high-fidelity framework by combining LLMs with explicit Reasoning world models with diffusion-based policy actors and leverage and Reflection modules. The Reasoning module generates the multi-modal generation capability of diffusion models to decisions from scene descriptions, and the Reflection module represent diverse action modes within the imagined rollouts. evaluates outcomes and accumulates experience memory. World models and model-based approaches are becoming This enables closed-loop improvement through self-correction. increasingly popular, but they also face challenges. Model LanguageMPC [67] uses LLMs as high-level decision-makers bias and compounding errors, where policies can learn to that output situation-dependent parameters, such as target exploit inaccuracies in the learned dynamics can produce speed and desired lane, mapped into actionable commands actions that appear optimal in imagination but fail in the through Model Predictive Control; effectively using LLMs real environment. Learning dynamics that remain valid under to select or adapt control objectives while preserving the interventions is particularly difficult simply because the model optimization structure and constraint satisfaction guarantees has not observed the consequences of actions outside the of MPC. data distribution. Multi-modality in traffic evolution requires Beyond high-level reasoning, recent work explores whether uncertainty-aware rollouts and robust planning that does not LLMs can directly generate trajectories or control signals. rely only to a single imagined future. Future directions include GPT-Driver [68] reformulates motion planning as a language stronger uncertainty calibration to detect when imagination modeling problem, representing planner inputs and outputs as diverges from reality, hybrid training that mixes offline world language tokens and leveraging GPT-3.5 to generate driving model priors with online correction through limited real trajectories through coordinate descriptions. DriveGPT4 [69] interaction, and explicit safety constraints that remain stable extends this paradigm to interpretable end-to-end driving. It under distribution shift between imagined and real dynamics. processes multi-frame video inputs through a multimodal LLM to predict low-level vehicle control signals while providing natural language explanations for vehicle actions. G. Foundation Model Approaches Reliability is a primary concern in foundation models as Foundation model approaches leverage large pretrained LLM outputs can be sensitive to prompt phrasing. Moreover, models, such as large language models (LLMs) or visionmapping language-level reasoning into geometrically and language models (VLMs), as high-level decision-making dynamically valid actions remains difficult as language or reasoning modules for ego vehicles. The key idea is representations do not natively encode spatial relationships that pretraining on internet-scale data provides these models with the precision required for collision avoidance. Halwith commonsense reasoning, semantic understanding, and lucination, where models generate plausible-sounding but instruction-following capabilities that are difficult to acquire factually incorrect reasoning about scene elements, poses from driving data alone. safety risks that are difficult to detect without independent A generic formulation treats the foundation model as a verification. Latency constraints limit closed-loop deployment. conditional decision function over structured scene represenCurrent evaluation is most limited to small scenario sets tations: or with open-loop metrics. For mixed autonomy simulation, yt = fθ P(o≤t , m, g), Mt (8) foundation models offer intriguing possibilities for generating where o≤t denotes the observation history, m represents diverse, human-like reasoning patterns for traffic agents, but the map context, g encodes navigation goals or mission the computational cost of running LLM inference for every specifications, P(·) is a prompt builder that structures these agent at every timestep makes direct application to multiinputs into a format suitable for the foundation model, Mt agent simulation impractical with current architectures. Future denotes memory or retrieved knowledge at time t, and directions include uncertainty-aware decision interfaces that fθ is the pretrained foundation model parameterized by θ. express confidence alongside decisions, formal safety wrapThe output yt can take several forms: discrete maneuver pers that verify LLM outputs against physical constraints decisions (e.g., “change lane left,” “yield to pedestrian”), before execution, more efficient inference, and standardized constraints or cost weights for downstream optimizers, or closed-loop evaluation protocols measuring. draft trajectory waypoints. Foundation model approaches can operate in both open-loop and closed-loop settings. However, the computational cost of LLM inference poses a significant bottleneck for real-time closed-loop deployment, VOLUME ,
11
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
H. Discussions on Single-Agent Methods
A. Joint Trajectory Forecasting
The fundamental limitation shared across many single-agent methods is the independence assumption. When a singleagent policy trained via imitation learning, reinforcement learning, or any other approach is deployed in a simulator where surrounding agents follow fixed replay or simple rulebased models, the resulting traffic dynamics may diverge substantially from what would occur in the real world. The ego policy is effectively optimized against an environment whose transition dynamics ptrain do not match the true interactive dynamics preal , and this mismatch can remain hidden during development because open-loop evaluation metrics, which compare predicted trajectories against recorded ground truth, are insensitive to interactive effects. A model that achieves low displacement error may still produce unsafe or unrealistic behavior when other agents begin responding to its actions rather than following scripted trajectories. This method-agnostic challenge point toward several promising research directions. Tighter integration between ego-policy learning and surrounding traffic modeling represents the most natural bridge to the multi-agent methods discussed in Section V. Hybrid architectures that combine the semantic reasoning capabilities of foundation models with the dynamics grounding of model-based approaches and the safety guarantees of constrained optimization could address complementary weaknesses. Evaluation methodology also requires attention: closed-loop benchmarks that measure not just trajectory accuracy but counterfactual robustness would provide far more meaningful assessments of simulation readiness than current displacement-based metrics. Finally, uncertainty quantification across all method families would enable conservative fallback behaviors when models operate outside their competence, a prerequisite for any deployment in safety-critical simulation infrastructure.
Joint trajectory forecasting models predict the conditional distribution of future motion for multiple agents simultaneously and explicitly capture inter-agent dependencies that single-agent prediction methods (Section B) ignore. Mathematically, given past states for N agents and map context m, the goal is to learn:
V. MULTI-AGENT METHODS
Multi-agent methods explicitly model multiple traffic participants and their interactions, capturing the “social contract” of driving: how agents anticipate, respond to, and coordinate with each other. Unlike the single-agent methods of Section IV, which treat surrounding traffic as part of the environment, multi-agent approaches model the bidirectional influence between agents. We organize multi-agent methods along two architectural axes. The first distinguishes joint trajectory forecasting (Section A), which predicts coordinated futures for multiple agents. The second divides interactive simulation into centralized approaches (Section B), where a single joint model produces globally consistent scene evolution, and decentralized approaches (Section C), where independent per-agent policies yield emergent coordination. We further discuss the emerging role of foundation models for multiagent coordination (Section D). Table 3 provides a summary of methods and studies reviewed in this section. 12
P (Y 1:N | X 1:N , m)
(9)
where X i = {xi1 , xi2 , . . . , xiTh } represents the observed trajectory history for agent i over horizon Th , Y i = {y1i , y2i , . . . , yTi f } denotes the predicted future trajectory for agent i over forecast horizon Tf , and Y 1:N = {Y 1 , Y 2 , . . . , Y N } denotes the joint future for all agents. The critical distinction from single-agent trajectory prediction is how inter-agent dependencies are modeled within the joint P (Y 1:N | X 1:N , m). Factorizing this as QN distribution i 1:N , m) (predicting each agent independently i=1 P (Y | X given shared context) captures contextual influence but not the constraint that future trajectories must be jointly consistent. True joint prediction instead models correlations across agents, producing consistent and collision-free future trajectories, where, for instance, one agent yielding is paired with another proceeding. This joint consistency is essential for simulation. Joint trajectory forecasting models are predominantly evaluated in open-loop settings, comparing predicted trajectories against ground-truth recordings using displacement metrics such as minADE and minFDE. However, they can also be deployed in closed-loop simulation by re-querying the model at each time step with updated observations. However, this receding-horizon approach introduces challenges around temporal consistency because the model is not trained to maintain coherent behavior across successive re-invocations. This dual-use nature makes joint forecasting a bridge between pure prediction and interactive simulation. Foundational work in this domain established key paradigms for encoding agent interactions. Social LSTM [88] introduced social pooling mechanisms capturing spatial dependencies among agents using LSTM encoders with gridbased pooling of neighboring hidden states. Social GAN [89] incorporated adversarial training with global pooling, using a variety loss that promotes diverse yet plausible multiagent futures by encouraging the generator to cover multiple behavioral modes. DESIRE [90] encoded diverse intents through stochastic latent variables with scene-level refinement that conditions trajectory generation on both learned intent and environmental context. Trajectron++ [91] modeled heterogeneous agent types through dynamic interaction graphs with edge types reflecting semantic relationships (vehicle– vehicle, vehicle–pedestrian). SoPhie [128] combined social attention over agent interactions with physical attention to scene context, jointly attending to who matters and where constraints lie. PRECOG [92] used normalizing flows to produce diverse, goal-conditionable joint futures with exact VOLUME ,
TABLE 3: Summary of multi-agent interaction methods for mixed autonomy traffic simulation. Method
Core Idea
Strengths
Limitations
Representative Studies
Joint Trajectory Forecasting (predominantly open-loop) Joint Trajectory Prediction
Predicts joint future trajectories for multiple agents; uses graph networks, attention, or transformers to model inter-agent dependencies
Captures inter-agent dependencies; multi-modal outputs; strong benchmark performance; joint consistency
Does not model reactive feedback; joint Social LSTM [88], Social consistency not guaranteed under marginal GAN [89], DESIRE [90], Tralosses; gap to interactive simulation jectron++ [91], PRECOG [92], VectorNet [93], LaneGCN [94], Scene Transformer [95], HiVT [96], MTR++ [97], GameFormer [98]
Centralized Neural Simulation (closed-loop) Diffusion-Based Simulators
Models multi-agent High diversity; multi-modal generation; evolution via iterative controllable via guidance; collision-aware denoising; samples sampling from learned joint trajectory distributions with guidance for controllability Token-Based / Discretizes motion into Scalable transformer architectures; handles Autoregressive tokens; casts simulation variable agent counts; real-time inference as conditional language possible modeling with next-token prediction Regression-Based Single learned model Global scene consistency; captures multiReactive produces globally agent dependencies; computationally consistent scene efficient; responds to ego deviations updates via continuous regression; rolls forward step-by-step
Slow sampling (iterative denoising); hard MotionDiffuser [13], CTG [99], constraint satisfaction difficult; closed-loop CTG++ [100] stability over long horizons
Discretization error; compounding token mistakes; serialization order asymmetry; constraint satisfaction requires constrained decoding
MotionLM [101], Trajeglish [102], BehaviorGPT [103], SMART [104]
Long-horizon drift; limited multi-modality (point estimates); counterfactual validity concerns from passive logs
SimNet [105], TrafficSim [106]
Emergent coordination; robustness through competition; automatic difficulty scaling
Non-stationarity; credit assignment difficulty; reward specification; may converge to non-human equilibria
Safe MARL [17], Flow [107], Nocturne [108], Data-Reg. Self-Play [109], Robust SelfPlay [110]
Reproduces human interaction styles; scalable via parameter sharing; captures emergent behaviors
Training instability with global discriminators; diversity collapse; irrelevant interaction misguidance
Principled interaction modeling; captures strategic behavior; interpretable equilibrium concepts
Utility specification difficulty; equilibrium selection ambiguity; computational cost; bounded rationality needed
PS-GAIL [51], RAIL/BurnInfoGAIL [111], Symphony [112], DecompGAIL [113] Kita [114], Sadigh et al. [115], iLQGames [116], GameFormer [98], Levelk [117], [118], BeTop [119]
Balances realism and controllability; behavior steering via conditioning; stable closed-loop rollouts
Controllability can leave data support; reward misspecification; calibration of latent controls
TrafficBots [120], TrafficBots V1.5 [121], CtRL-Sim [122], TrajGen [123]
Grounding difficulty; temporal inconsistency; scalability bottleneck; verification of joint outputs
AgentsCoDriver [124], KoMA [125], CCMA [126], Wang et al. [127]
Decentralized Simulation (closed-loop) Multi-Agent RL / Each agent learns a Self-Play decentralized policy; CTDE paradigm for training; self-play creates automatic curriculum Multi-Agent Decentralized policies Imitation / GAIL trained to match expert interaction patterns via adversarial discrimination Game-Theoretic Models interaction as Models strategic reasoning; Nash or Stackelberg equilibria capture anticipation and mutual influence Hybrid DataCombines imitation Driven Agents learning base with RL fine-tuning or return conditioning; controllable behavior knobs
Agentic Foundation-Model Coordination (decentralized, centralized, or hierarchical) Agentic FoundationModel Coordination
VOLUME ,
LLM/VLM agents Semantic reasoning; emergent social with communication, norms; interpretable decisions; flexible memory, and reflection coordination modules coordinate multi-vehicle behavior; architectures could be decentralized, centralized, and hierarchical
13
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
likelihood computation, which enables principled probabilistic evaluation. As the field scaled to larger agent counts and more complex road topologies, vectorized map and scene-graph representations became dominant. VectorNet [93] encoded both map elements and agent trajectories as polylines processed through hierarchical graph networks. LaneGCN [94] constructed explicit lane graphs to preserve topological connectivity through multi-scale graph convolutions that propagate information along and across lanes. TNT [40] decomposed prediction into target endpoint prediction from lane centerlines and trajectory completion conditioned on selected goals, structuring multi-modality around discrete spatial anchors. In recent years, transformer-based architectures unified agent interaction modeling and scene reasoning through attention mechanisms. Scene Transformer [95] introduced a unified architecture with masking strategies to enable flexible querying, which means the same model can predict marginal or joint futures by adjusting which agents are masked during decoding. AgentFormer [129] combined transformers with latent variables using stochastic attention to model both social interactions and multi-modal intent simultaneously. HiVT [96] introduced a two-stage architecture separating local context extraction from global interaction modeling by using translation- and rotation-invariant representations that improve generalization across scenes. Wayformer [130] applied efficient attention mechanisms (including latent query attention) for scalable processing of large agent sets. QCNet [131] and QCNeXt [132] emphasized structured scene querying with query-centric designs to further improve computational efficiency. MTR-A [133] introduced the Motion Transformer framework by combining learned intention points with iterative motion refinement, which resulted in significant performance improvement. MTR++ [97] extended this foundation with symmetric scene modeling and strengthened intent querying. MotionDiffuser [13] represents a growing diffusionbased direction in the latest studies. It applies denoising diffusion to joint trajectory generation for controllable, highdiversity sampling with the ability to incorporate constraints during the reverse process. GameFormer [98] incorporates game-theoretic reasoning into the transformer framework by modeling strategic interactions where agents’ predicted futures reflect anticipation of others’ responses rather than independent extrapolation. To summarize, joint trajectory forecasting has become increasingly popular, following the recent Waymo Open Dataset Challenges [26], [134]. Despite this interest, core limitation of joint trajectory forecasting for simulation is still the mismatch between distributional realism and causal reactivity. These models can achieve low displacement error by matching data distributions under passive observation, yet the predicted futures do not reflect how agents would react if the ego vehicle deviated from its logged behavior. They model correlations observed in data but not the causal mechanisms
14
generating those correlations. This is important, especially when simulation is used to evaluate what happens under counterfactual ego actions. Nevertheless, joint forecasting is still valuable in several simulation-adjacent roles. Multimodal joint predictions can seed downstream planners or closed-loop simulators with plausible initial futures, which reduces the search space for trajectory optimization. They can also support scenario mining, where identifying interactions with high uncertainty or predicted conflict flags situations that warrant closer testing. Finally, aggregated occupancy or flow forecasts derived from joint predictions offer lightweight, riskaware scene summaries that planners can consume without reasoning over individual agent trajectories explicitly. Subsequent studies could explore training objectives that explicitly penalize joint inconsistency (e.g., collision-aware losses), uncertainty-aware scoring that aligns predicted probabilities with empirical frequencies of behavioral modes, tighter bridges to closed-loop evaluation through recedinghorizon deployment with temporal consistency regularization, and metrics that measure closed loop and intervention sensitivity rather than only log-matching accuracy. B. Centralized Modeling and Simulation
Centralized simulation models scene evolution through a single learned model that jointly determines how all agents move. Rather than assigning independent policies to each agent. A centralized simulator takes the complete scene state as input and produces the next scene state, enforcing inter-agent consistency through shared computation. The key distinction from decentralized simulation (Section C) is that coupling between agents is enforced within the model architecture, which means a single forward pass produces globally consistent next states rather than emerging from the interaction of separate per-agent policies. In the following subsections, the major classes of centralized modeling and simulation methods are reviewed.
1) Diffusion-Based Simulators
Diffusion-based methods generate realistic multi-agent traffic by starting from random noise and gradually refining it into plausible joint trajectories. The approach defines a denoising process over a joint trajectory tensor Y ∈ RN ×Tf ×d stacking future states for all N agents over a short forecast horizon Tf in d-dimensional state space. Let Y(0) denote the clean joint trajectory and Y(n) denote the noised version at diffusion step n. The forward process progressively corrupts trajectories by adding Gaussian noise, and the model learns to reverse this corruption. The training objective minimizes: 2 (n) min En,ϵ ϵ − ϵθ (Y , n, m, c) (10) θ
where ϵ ∼ N (0, I) is the noise added to trajectories, ϵθ is a learned noise prediction network, n indexes diffusion steps (distinct from simulation time t), m represents map context, and c denotes conditioning signals including initial scene VOLUME ,
state and ego actions. At inference, the model samples from Gaussian noise and iteratively denoises: 1 1 − αn Y(n−1) = √ ϵθ (Y(n) , n, m, c) +σn z Y(n) − √ αn 1 − ᾱn (11) where αn and ᾱn are noise schedule parameters at diffusion step n, σn is the noise level, and z ∼ N (0, I). The model can be steered toward desired behaviors in two ways: by training it with specific conditions baked into c, or by nudging the denoising process at inference time using gradient-based guidance toward goals like collision avoidance. For closedloop simulation, the model is re-run at every timestep with the latest scene state as input, so traffic agents continuously react to what the ego vehicle does. In this class of models, CTG [99] demonstrates diffusion generating socially consistent multi-agent trajectories with controllability through conditioning signals. CTG++ [100] extends with language conditioning to enable semantically specified scene evolution valuable for scenario-based testing. MotionDiffuser [13], while primarily being a forecasting model (discussed in Section A), demonstrated architectural principles (joint diffusion over all agents with constraint-based guidance) that directly inform the design of diffusion-based simulators. Building on these foundations, diffusion-based approaches have been extended with increasingly expressive controllability interfaces for scenario-level generation, including differentiable cost function guidance and languagebased conditioning. These controllable synthesis methods are discussed in Section 1 as they serve primarily a scenario generation role. The primary challenge for diffusion-based simulators is computational cost. Iterative denoising requires multiple neural network evaluations per simulation step, and this cost multiplies when the sampler must be re-invoked at each rollout timestep for closed-loop operation. Closedloop stability presents a second challenge, as even locally realistic samples can drift over long horizons when successive re-conditionings accumulate small inconsistencies. A third concern is that guidance-based controllability provides soft rather than hard constraint satisfaction. For instance, biasing the sampling distribution toward collision avoidance does not guarantee it, which is problematic for safety-critical simulation. Promising directions include accelerated sampling through diffusion distillation (reducing the number of denoising steps), consistency models that generate highquality samples in fewer steps, uncertainty-aware guidance that becomes conservative under distributional ambiguity, and hybrid architectures combining diffusion sampling for multimodal diversity with deterministic components enforcing hard physical constraints.
2) Token-Based and Autoregressive Simulators
Token-based simulators discretize continuous agent motion into sequences of discrete tokens, which is basically recasting VOLUME ,
multi-agent simulation as conditional sequence modeling analogous to language generation. This formulation leverages the scalability and expressiveness of transformer architectures developed for natural language processing by applying them to the structured sequential prediction problem of traffic evolution. Let V denote a learned vocabulary of motion tokens obtained through vector quantization of trajectory segments, and let zti ∈ V represent the discretized motion token for agent i at time t. Given map context m encoding road geometry and topology, an autoregressive simulator learns the conditional distribution over the next-step tokens for all N agents, factorized over a serialization of tokens within the timestep: N Y (j) (<j) 1:N 1:N 1:N pθ zt+1 | z≤t , m = pθ zt+1 | z≤t , m, zt+1 (12) j=1 1:N z≤t
where denotes all motion tokens up to time t for (<j) all agents, and zt+1 denotes tokens already generated at timestep t + 1 before serial position j . This intra-timestep factorization is critical. Later tokens in the serialization are conditioned on earlier tokens within the same timestep, which implicitly models within-step agent interactions through the autoregressive ordering. Training uses standard cross-entropy next-token prediction over the full sequence. For closed-loop simulation, tokens are sampled sequentially at each timestep, decoded back to continuous states through the inverse of the vector quantization mapping, and the process iterates with the decoded states forming the conditioning for the next step. MotionLM [101] introduced the paradigm of discretizing trajectories into motion tokens and training autoregressive transformers for joint multi-agent futures. Trajeglish [102] refined this approach by modeling traffic as next-token prediction with fine spatial resolution and explicit accounting for intra-timestep agent interactions through careful serialization ordering. BehaviorGPT [103] proposed next-patch prediction (NP3), where multi-step motion patches are predicted as single units rather than individual tokens. This diminishes the “copying shortcut” problem where models learn to simply repeat the previous token rather than generating meaningful motion updates. SMART [104] extended the tokenization framework with explicit road tokens encoding map structure within the same vocabulary by using decoder-only transformers. It demonstrates scalability across datasets, cross-dataset generalization, and real-time inference capability. The central tension in token-based simulation is between discretization fidelity and sequence tractability. Finer-grained tokenization (larger vocabularies and higher spatial resolution) reduces quantization error but increases sequence length, which increases the computational cost and exacerbates exposure to compounding errors over long rollouts. The serialization order of agents within each timestep introduces an arbitrary asymmetry, which means the first agent’s token is generated without conditioning on other agents’ current actions, while later agents benefit from this information. Compounding error is particularly critical since a single 15
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
incorrect token can shift the agent to an implausible state and functions, and transition dynamics: potentially cascading into increasingly unrealistic rollouts. st+1 ∼ p(st+1 | st , a1t , . . . , aN (13) t ) Further research could focus on hybrid approaches combining discrete token generation with continuous refinement where st ∈ S is the global state, ait ∈ Ai is the action of layers, learned serialization orderings that minimize informa- agent i, and p represents the environment transition dynamics tion asymmetry, training paradigms incorporating scheduled that depend on all agents’ joint actions. It is worth noting that sampling or rollout-aware losses to improve long-horizon while the world’s transition depends on all joint actions, each robustness, and constrained beam search or rejection sampling agent’s policy can only condition on its own observations. for enforcing hard physical constraints. Each agent i selects actions based on local observations: ait ∼ πθi (· | oit ), where oit is agent i’s (partial) observation of st . The key architectural distinction from centralized simulation 3) Regression-Based Reactive Simulators (Section B) is that no single model jointly determines all Regression-based reactive simulators directly predict continu- agents’ next states. When deployed reactively, decentralized methods operate in ous state updates for all agents through learned regression functions, without the intermediate discretization of token- a closed-loop fashion, as each agent conditions its actions on based methods or the iterative sampling of diffusion models. current observations. However, they can equally be trained via These methods emphasize computational efficiency and open-loop objectives such as behavior cloning or maximum deterministic prediction, which makes them well suited for likelihood, and evaluated through fixed-horizon rollouts without external intervention. The subsubsections below real-time closed-loop deployment. SimNet [105] established the paradigm of learning reactive organize decentralized approaches by how agent policies traffic simulation by training on large-scale perception outputs are obtained, that is through reward optimization (MARL), to produce agents that respond to ego deviations during demonstration matching (imitation and adversarial learning), closed-loop rollout. The architecture encodes scene context strategic reasoning (game theory), or combination of these through rasterized representations and predicts per-agent approaches (hybrid methods). trajectory continuations. This work demonstrated that datadriven reactive simulation is feasible at scale. TrafficSim [106] advanced this approach by explicitly modeling agent-to-agent 1) Multi-Agent Reinforcement Learning and Self-Play dependencies through graph neural network architectures Multi-agent reinforcement learning (MARL) learns decenand shows that interaction-aware scene representations sub- tralized driving policies by optimizing expected return in stantially improve the realism of joint rollouts compared to Markov games. Each agent i has its own reward function ri and learns policy πθi to maximize expected cumulative independent per-agent prediction. The challenges in regression-based reactive methods is the discounted return: "∞ # gap between short-horizon accuracy and long-horizon stabilX t 1 N ity, where compounding prediction errors accumulate over Ji (θi ) = Eπθ1 ,...,πθN γ ri (st , at , . . . , at ) (14) t=0 extended rollouts. Because regression-based methods typically produce point estimates rather than distributions over next where γ ∈ [0, 1] is the discount factor and the expectation states, they capture limited multi-modality. Causal validity is over joint trajectories generated by all agent policies is another concern: learning from passive observational logs simultaneously. The fundamental difficulty is that each agent’s does not uniquely identify how agents would respond to coun- optimization landscape depends on all other agents’ evolving terfactual ego actions, and the model may learn correlations policies, which creates a non-stationary learning problem that break under intervention. Future directions include multi- where the environment effectively changes as co-learners modal regression heads that output mixture distributions over update their behavior. next states, intervention-aware training that augments logged A dominant paradigm for managing this non-stationarity is data with synthetic ego deviations, uncertainty estimation centralized training with decentralized execution (CTDE). that flags when predictions are unreliable, and curriculum- During training, each agent has access to a centralized based unrolling strategies that progressively increase rollout critic Qϕ (st , a1 , . . . , aN ) parameterized by ϕi that observes t t i horizons during training to improve long-horizon stability. the global state and all agents’ actions, providing more C. Decentralized Multi-Agent Methods
stable value estimates than would be possible from local observations alone. At execution time, each agent acts using only its local observation:
Decentralized multi-agent simulation assigns each traffic participant its own policy, letting scene-level behavior emerge ait ∼ πθi (· | oit ) (15) from the coupled execution of independent decision-makers. Formally, this setting is modeled as a Markov game (also This asymmetry (global information for learning, local called stochastic game) with N agents, global state space information for acting) enables agents to internalize the S , per-agent action spaces {Ai }N effects of other agents’ behavior during training while i=1 , per-agent observation 16
VOLUME ,
remaining deployable in partially observed settings. Selfplay, a special formulation of MARL where agents train against copies of themselves or a population of past policy checkpoints, provides an automatic curriculum: as policies improve, the training environment becomes correspondingly more challenging and produce increasingly robust interaction strategies. Foundational work established MARL’s viability for traffic applications. Shalev-Shwartz et al. [17] proposed a safetyoriented multi-agent formulation that decomposes driving into learned high-level “desires” (e.g., target lane, desired speed) and constraint-satisfying trajectory planners that ensure feasibility and aims to address the challenge that unconstrained RL policies may produce physically implausible or dangerous actions. Flow [107] provided a computational framework for studying mixed autonomy traffic through MARL and demonstrates that even small penetration rates of RL-controlled vehicles can produce emergent improvements in traffic throughput. Nocturne [108] introduced a purposebuilt multi-agent driving environment designed for MARL research by featuring 2D partially observed scenarios running at over 2,000 steps per second. This efficiency enables the large-scale experiments (millions of episodes) that MARL algorithms require. Building on the need for high-throughput multi-agent environments, recent simulators, such as GPUDrive [2] and PufferDrive [135] achieve substantially higher efficiency and flexibility. Recent work has focused on producing human-compatible rather than merely reward-optimal behavior. The observation that self-play alone can converge to superhuman or non-human strategic equilibria motivates data-regularized approaches. Human-Regularized PPO [109] addresses this by augmenting the self-play objective with demonstration regularization, which penalizes policies that deviate from recorded human driving distributions while still benefiting from self-play’s robustness to adversarial interactions. This anchoring to human data is critical for simulation applications where the goal is reproducing realistic mixed traffic rather than optimal traffic. Complementing this direction, Robust Autonomy from Self-Play [110] demonstrated that policies emerging from large-scale self-play exhibit robustness to diverse interaction partners, including out-of-distribution agents not encountered during training. However, compared to [109], the reward design and shaping is quite extensive. Bronstein et al. [136] showed that structuring the action space through learned latent representations simplifies multi-agent policy learning and improves transfer across scenarios. One of the main challenges for MARL in driving simulation is non-stationarity and credit assignment. Each agent’s learning changes the environment for all others. This makes convergence difficult and potentially causes oscillatory training dynamics. Reward specification is also demanding. Small changes in reward weights can produce qualitatively different emergent behaviors, and multi-objective rewards balancing safety, progress, comfort, and rule compliance are difficult to
VOLUME ,
tune. Future directions include stronger CTDE architectures with uncertainty-aware critics that distinguish epistemic from aleatoric uncertainty in other agents’ behavior, hybrid training that anchors MARL policies to real driving logs through behavioral priors, population-based training that maintains a diverse pool of driving styles preventing convergence to a single behavioral mode, and explicit safety constraints that remain stable under multi-agent distributional shift.
2) Multi-Agent Imitation and Adversarial Learning
Multi-agent imitation learning extends the imitation and adversarial methods discussed for single agents (Section D) to decentralized multi-agent settings. The key difference from single-agent adversarial imitation is that the discriminator must evaluate joint behavior and see if the collection of agents’ actions matches the distributional properties of real multi-agent traffic. Given expert demonstrations DE = {(sj , a1:N )}M j j=1 containing M samples of global states sj and joint actions a1:N = {a1j , . . . , aN j j } for all N agents, the goal is to learn per-agent policies {πθi }N i=1 whose joint rollouts are indistinguishable from expert behavior. In the multi-agent GAIL formulation, the objective is: min max Eπθ log Dψ (s, a1:N ) {πθi } Dψ
N X + EπE log(1 − Dψ (s, a1:N )) − λ H(πθi ) i=1
(16) where Dψ is a discriminator evaluating joint state-action tuples, πE represents the expert’s joint behavior, and H(πθi ) is the entropy of agent i’s policy weighted by λ > 0. Training typically follows the CTDE paradigm. The discriminator operates on global state-action tuples during training, while execution remains decentralized with each agent acting from local observations ait ∼ πθi (· | oit ). PS-GAIL [51] first scaled adversarial imitation to multiagent driving by introducing parameter sharing across agents. All agents use the same policy network conditioned on agentspecific context. Parameter sharing is combined with curriculum learning that progressively increases rollout horizons during training to improve long-horizon interaction stability. The extended version [111] systematized this framework by adding Reward Augmented Imitation Learning (RAIL) for injecting domain knowledge (e.g., lane-keeping and speed maintenance rewards alongside the adversarial signal) and Burn-InfoGAIL for disentangling latent driving style factors. Symphony [112] addressed a key failure mode of multiagent adversarial learning, which is diversity collapse where all agents converge to similar behavior. This is achieved by combining imitation learning with discriminator-guided parallel beam search during rollout and hierarchical goal mechanisms that maintain distinct agent intents over long horizons. DecompGAIL [113] identified another failure mode specific to multi-agent settings: “irrelevant interaction mis17
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
guidance”, where a global discriminator penalizes an agent for ego vehicle acts as leader, choosing its policy π ego while behavior that appears unrealistic only because of other agents’ anticipating that other agents will best-respond: errors. The solution decomposes discrimination into egoπ ego,∗ = arg max J ego (π ego , BR−ego (π ego )) (19) map components (evaluating map compliance independently) π ego and ego-neighbor components (evaluating interaction realism ego separately) to prevent irrelevant signals from destabilizing where BR−ego (π ) denotes the collection of best-response policies for all non-ego agents given the leader’s policy. individual agent learning. Counterfactual validity is still the most fundamental Computing best responses requires solving each follower’s concern about multi-agent imitation. Since training data optimization problem given the leader’s committed strategy, consists of passive logs, the discriminator can only evaluate making the Stackelberg formulation a bilevel optimization. Game-theoretic approaches can operate in both openwhether joint behavior matches the observed distribution of loop and closed-loop modes. Open-loop game solutions interactions, not whether agents respond correctly to novel compute complete strategy profiles (sequences of actions) interventions. Training stability is more difficult than in singlebefore execution, suitable for short-horizon interactions where agent settings because the adversarial dynamics must simulreplanning is unnecessary. Closed-loop (feedback) strategies taneously shape multiple policies whose interactions create condition actions on the current state at each timestep and additional non-stationarity beyond the standard generatorproduce reactive policies that adapt to the evolving interaction. discriminator oscillation. The diversity-versus-realism tension Kita [114] modeled highway merging as a game where is also concerning. Mode collapse in multi-agent settings can gap acceptance decisions depend on strategic reasoning about lead to loss of behavioral variety and interaction variety, which the mainline driver’s likely response, demonstrating that itself can generate traffic where all agents exhibit similar human merge behavior is better explained by game theory responses to the same situations. Moving forward, it will be beneficial to explore factored discriminator architectures than by fixed gap thresholds. Sadigh et al. [115] formulated that isolate sources of unrealism, hybrid training combining automated planning as actively influencing human actions behavior cloning pre-training with adversarial fine-tuning for through Stackelberg games and showed that an ego vehicle stability, uncertainty-aware discriminators that reduce gradient can leverage its commitment power to create more favorable magnitude in states where expert behavior is ambiguous, and interactions (e.g., nudging forward to claim right-of-way). evaluation protocols that explicitly test intervention response iLQGames [116] introduced efficient iterative linear-quadratic by measuring how agent behavior changes when the ego solvers for general-sum dynamic games, which resulted in producing feedback (closed-loop) Nash equilibrium strategies vehicle deviates from logged trajectories. in real time. This is important for planning at the speeds required for driving. GameFormer [98] incorporated gametheoretic structure into transformer-based architectures for 3) Game-Theoretic and Strategic Interaction Models joint interactive prediction and planning to capture how Game-theoretic models formalize driving interaction as agents’ predicted futures influence each other. strategic reasoning among decision-makers with coupled An important insight is that human drivers do not reason objectives and provide a principled framework for underwith unlimited strategic depth. Bounded rationality models standing phenomena such as yielding, merging negotiation, using level-k reasoning [117], [118] assume that level-0 and implicit coordination at unsignalized intersections. While agents follow simple heuristic policies (e.g., maintain current MARL also operates in the Markov game framework, the key speed), level-1 agents best-respond to level-0 behavior, leveldistinction is that MARL learns policies through trial-and2 agents best-respond to level-1, and so on. This framework error interaction, whereas game-theoretic approaches compute provides a controlled spectrum from reactive heuristics to or approximate equilibrium solutions analytically without deep strategic reasoning. This (arguably) better matches the requiring extensive simulation experience. heterogeneity of real driving populations where some drivers Mathematically, each agent i seeks to optimize its utility plan ahead while others react instinctively. BeTop [119] Ji that depends on all agents’ strategies: introduces behavioral topology as a structured representation " T # X for multi-agent interactions and reasons about relative spatialJi (π1 , . . . , πN ) = Eπ1 ,...,πN γ t ri (st , a1t , . . . , aN t ) temporal relationships between agents (who yields to whom) t=0 (17) rather than absolute positions. By modeling topological where T is the interaction horizon. A Nash equilibrium is relations explicitly, BeTop captures the strategic structure ∗ a joint policy (π1∗ , . . . , πN ) where no agent can improve its of driving interactions in a form that is both interpretable and compatible with downstream planning. utility by unilateral deviation: Utility specification is a demanding task in game theoretic ∗ ∗ Ji (π1∗ , . . . , πi∗ , . . . , πN ) ≥ Ji (π1∗ , . . . , πi , . . . , πN ), ∀πi , ∀i approaches. Small changes in reward weights can qualitatively (18) flip equilibrium behavior (e.g., switching from yielding to In Stackelberg formulations, which is natural for modeling aggressive merging). Equilibrium selection poses a further ego-vehicle planning among reactive human drivers, the problem. Many driving interactions admit multiple Nash 18
VOLUME ,
equilibria (e.g., both “I yield, you go” and “you yield, I go” may be equilibria), and selecting among them requires additional assumptions about conventions, precedence, or focal points. Computational cost is significant for dynamic games with many agents and long horizons, as the state space grows combinatorially. Most critically, the assumption of rationality (even bounded rationality) may not capture the full range of human driving behavior, such as distraction, confusion, and rule violations. Future directions could include differentiable game solvers with learned utility functions calibrated from driving data, data-driven bounded rationality models that learn the distribution of reasoning depths from observed interactions, and evaluation protocols testing whether game-theoretic agents produce realistic equilibrium selection patterns under distribution shift.
not correspond to any observed human driving pattern. Calibration of latent controls to interpretable and consistent behaviors remains difficult. It is often unclear what specific latent values correspond to a specific observable driving characteristic. Future frameworks should focus on evaluation frameworks that measure the controllability-versus-realism trade-off explicitly, and try to disentangle latent spaces where individual dimensions correspond to interpretable behavioral attributes. D. Agentic and Foundation-Model Coordination
Foundation model approaches, introduced for single-agent ego planning in Section G, are increasingly explored for multi-agent coordination as reasoning layers that mediate interactions across multiple traffic participants. The foundation model in this context is aimed at maintaining temporal and inter-agent consistency across multiple simultaneous reasoning processes, avoiding producing over-coordinated 4) Hybrid Data-Driven Agents Hybrid approaches combine data-driven imitation with struc- behaviors that would be unrealistic without explicit vehicletured mechanisms for controllability and physical feasibility. to-vehicle communication, and scaling its reasoning across The motivation is that pure imitation produces realistic but variable numbers of participants with heterogeneous intents. The general formulation treats the foundation model as uncontrollable agents, pure RL produces controllable but producing high-level decisions for agent i conditioned on potentially unrealistic agents, and hybrid methods seek the multi-agent scene context: best of both by starting from data-driven foundations and adding targeted modifications. yti = fθi P o1:N , m, g 1:N , h1:N , Mit (20) t t TrafficBots [120] demonstrated the effectiveness of goaland-personality conditioning for scalable multi-agent simu- where o1:N represents observations for all N agents, m t lation. It trains closed-loop policies on large-scale driving denotes map context, g 1:N represents goals for all agents, datasets where each agent is conditioned on its extracted h1:N denotes intent or high-level plan representations, P(·) t destination and a latent personality inferred via CVAE. The is a prompt builder structuring multi-agent context, Mit approach scales efficiently to dense urban scenes because represents memory or retrieved knowledge for agent i, and yti the shared policy handles variable agent counts. TrafficBots is the high-level decision output subsequently grounded into V1.5 [121] improved upon this foundation with stronger a control action ait = C i (yti , sit ) through a low-level controller transformer-based scene encoders and training techniques C i . An important design tension concerns the information (including better data augmentation and loss scheduling) structure. Providing each agent’s model with observations of for improved closed-loop performance on the Waymo Sim all other agents o1:N enables globally informed reasoning t Agents benchmark [26]. CtRL-Sim [122] introduced return- but is unrealistic. Systems sharing full scene information risk conditioned offline RL to multi-agent simulation setting. It producing coordination that implicitly assumes omniscience trains on logged data annotated with per-timestep rewards or communication. Therefore, restricting each agent to local for safety, progress, and comfort, then conditioning the observations oit is more faithful to real driving although it policy on desired return levels at inference. This enables limits the model’s reasoning about distant or occluded agents. test designers to generate specific behavioral profiles (e.g., In this class of models, AgentsCoDriver [124] formulates a tailgating agent, a hesitant merger) by specifying return multi-vehicle collaborative driving as a Decentralized Partially targets without retraining. TrajGen [123] takes a two-stage Observable Markov Decision Process (Dec-POMDP) and approach, first generating diverse candidate trajectories using equips each LLM-based agent with five modules: observation, a prediction model and then applying RL-based refinement to reasoning engine, cognitive memory, reinforcement reflection, modify trajectories under kinematic constraints while avoiding and communication. The communication module allows collisions. agents to determine when to communicate (based on potential The central tension in hybrid approaches is between trajectory conflicts) and what to convey (via an LLMcontrollability and realism. As behavioral conditioning moves instantiated message generator). This provides a principled agents further from the modes well-represented in training framework for selective information sharing rather than data, the policy operates increasingly out-of-distribution, and broadcasting the global state. The reinforcement reflection predicted behaviors may become implausible. Aggressive module evaluates outcomes and updates experience memory return targets or extreme personality samples can produce to enable a form of lifelong learning across episodes. “sim-native” behaviors that are internally consistent but do KoMA [125] proposes a knowledge-driven framework where VOLUME ,
19
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
multiple LLM-powered agents analyze surrounding vehicles’ case. Scalability is a fundamental bottleneck. Running LLM behaviors to infer intentions. A multi-step planning module inference for every agent at every timestep is computationally structures reasoning through a goal–plan–action hierarchy. prohibitive for large-scale traffic simulation; even hierarchical KoMA demonstrates that structured multi-agent reasoning designs face challenges when the number of interacting agents with shared knowledge substantially outperforms single-agent grows. Verification and safety assurance require checking not LLM approaches in highway merging scenarios. only individual outputs but their joint consistency. Evaluation CCMA [126] introduces a hierarchical architecture by requires closed-loop, intervention-based metrics assessing integrating reinforcement learning for individual-level op- how multi-agent systems respond when individual agents are timization with fine-tuned LLMs for regional and global perturbed. Future directions include lightweight reasoning coordination. Rather than using LLMs for direct control, architectures, such as distilled models, cached reasoning CCMA leverages them to dynamically tune reward functions patterns; formal verification layers checking joint physical to balance traffic flow, cooperation, comfort, and safety. This consistency before execution; integration of LLM reasoning hybrid RL-LLM design addresses the latency problem by with game-theoretic solvers for principled equilibrium comconfining LLM inference to less frequent strategic decisions. putation; and hybrid architectures allocating expensive LLM Wang et al. [127] investigated whether LLM agents playing reasoning selectively to complex interactions while learned Markov games in multi-agent driving scenarios can develop neural policies manage routine driving. emergent social norms. Their experiments showed that social norms do emerge, with agents adopting conservative E. Discussion on Multi-Agent Methods policies in collision-prone situations and suggested that LLMs’ pretraining on human-generated text encodes implicit social Multi-agent methods explicitly model interaction but introreasoning relevant to driving coordination. A comprehensive duce challenges absent in the single-agent setting (Section H). survey by Hu et al. [137] systematically categorizes LLM- While single-agent methods face the fundamental limitation based multi-agent driving systems by interaction mode (multi- of treating other agents as non-responsive environment comvehicle, vehicle-infrastructure, agent-human) and architectural ponents, multi-agent methods must satisfy three simultaneous requirements that create inherent tensions: (1) interaction pattern (centralized, decentralized, hierarchical). These systems can operate in both open-loop and closed- realism, where agents respond plausibly to each other and loop modes. In open-loop mode, the foundation model to ego interventions; (2) scene-level consistency, where joint generates a complete interaction plan for all agents from rollouts remain collision-free and map-compliant, unless an initial scene description, useful for scenario generation scenarios specifically demand otherwise; and (3) scalability and offline analysis. In closed-loop mode, agents are queried and controllability, where simulators remain computationally at each timestep with updated observations; however, the tractable and behaviorally steerable across varying traffic computational cost of LLM inference has led current imple- densities, road geometries, and testing objectives. The most fundamental challenge cutting across all multimentations to reduce the query frequency (1–2 Hz). agent approaches is arguably counterfactual validity under Several trends are emerging in this domain. First, systems intervention. Most models train from passive observational are shifting from free-form language outputs toward struclogs that do not uniquely identify how agents would react tured, testable interfaces with well-defined output schemas. when the ego vehicle deviates from its recorded behavior. In Second, multi-agent coordination increasingly employs expractice, this means a simulator may produce highly realistic plicit communication abstractions (message protocols, shared traffic when the ego follows its logged trajectory but generate memory pools) and ensuring that the level of coordination implausible reactions when the ego deviates, such as agents is realistic given assumed communication capabilities. Third, failing to brake or yielding when they should not. This is a hybrid architectures coupling LLM-level reasoning with RLlimitation that is particularly problematic for scenario-based trained or safety-certified low-level controllers are gaining testing and verification of automated vehicles. Addressing this more attention, as they separate the semantic understanding requires either interventional training data (rare and expensive where LLMs excel from the real-time reactivity where neural to collect), data augmentation strategies, or explicit causal controllers are more practical. Fourth, the question of whether models of agent response. Additionally, the compounding LLMs can serve as models of human driving cognition (rather error in multi-agent settings is amplified by interaction than just optimal planners) is gaining attention, with social feedback. When one agent’s prediction error changes the norm emergence studies suggesting that LLMs’ implicit input distribution for all others, cascading effects emerge knowledge of human behavior may be directly applicable to simulating human-like traffic participants. Despite these that are absent in single-agent rollouts. Current mitigation progresses and attentions, these methods face open challenges. strategies include unrolled training losses, scheduled sampling, Grounding remains the primary concern. Translating language- and explicit reset mechanisms, but maintaining stable multilevel multi-agent reasoning into geometrically precise, dy- agent rollouts over horizons exceeding 10 seconds remains namically feasible, and mutually consistent trajectories for open, particularly in dense urban traffic. The realism-controllability trade-off is an interesting area multiple agents is significantly harder than the single-agent to explore. Test engineers require controllable agents to 20
VOLUME ,
create specific stress-testing scenarios, but controllability conditioned transition model: mechanisms can push agent behavior beyond the support of pθ (ot+1 | o≤t , a≤t , c) (21) logged data, also known as “sim-native” behaviors. Sim-native behaviors are internally consistent but do not necessarily corre- where ot is an observation (multi-camera images, LiDAR, spond to observed human driving patterns. Parameter sharing, BEV features, or learned latent), o≤t = {o1 , o2 , . . . , ot } while essential for scalability, can reduce behavioral diversity denotes observation history up to time t, at is an action unless heterogeneity is explicitly maintained. Quantifying or planned motion, a≤t = {a1 , a2 , . . . , at } denotes action where the boundary lies between controlled-but-realistic and history, c represents structured conditioning such as map controlled-but-implausible remains an open problem. Related context, agent states, or high-level commands, and θ repto this, evaluation and benchmarking constitute perhaps the resents model parameters. Modern driving world models most critical bottleneck limiting the field’s ability to compare often factorize generation through tokenizers and generative and validate multi-agent methods. Open-loop prediction backbones (autoregressive next-token prediction or diffusion metrics such as minADE and minFDE do not guarantee in latent space), which helps scale to long horizons and closed-loop simulation quality. Closed-loop benchmarks such high-dimensional outputs. World models function as general-purpose environment as WOSAC [26] and nuPlan [25] represent important progress simulators that can serve multiple downstream applications. but remain sensitive to evaluation infrastructure assumptions, For behavior modeling specifically, generative world models including how ego behavior is specified, how collisions enable three capabilities that traditional simulators lack: are detected, and what constitutes realistic interaction. Key imagination-based training, where behavioral policies can evaluation dimensions that remain insufficiently addressed be optimized through rollouts inside the learned model include intervention sensitivity, long-horizon distributional without requiring access to the real environment or a handstability, calibration of multi-modality (whether behavioral crafted simulator; closed-loop evaluation, where learned diversity and mode probabilities match real-world frequenbehavior models are tested against environment responses that cies), controllability fidelity (whether steering mechanisms reflect data-driven dynamics rather than scripted rules; and produce intended effects without degrading realism), and data augmentation, where the world model generates novel geographic and cultural transfer (whether models trained in scenarios that expand the training distribution for behavior one city generalize to different road geometries, traffic rules, models beyond what is available in recorded logs. and driving norms). Developing comprehensive evaluation protocols addressing these dimensions is an important open research problem. 1) Video-Based and Latent World Models
Video-based world models treat driving simulation as conditional video generation, producing future visual observaVI. ENVIRONMENT-LEVEL SIMULATION METHODS Sections IV–V addressed how individual agents and groups tions conditioned on actions and structured controls. Let of agents make decisions and generate trajectories. This ot ∈ RC×H×W denote a multi-view observation at time section shifts focus to the environment side of the simulation t. The observation is encoded into a latent representation loop. The key architectural distinction from agent-level zt ∈ Rd or discrete tokens. In autoregressive form with methods is that the modeling target is the environment state discrete tokenization, let zt = {zt,1 , zt,2 , . . . , zt,Kt } represent itself—observations, occupancy fields, or complete scene Kt tokens at time t from vocabulary V . The autoregressive configurations—rather than individual agent actions or poli- world model learns: Kt T Y cies. We organize this section into generative world models Y pθ (z1:T ) = pθ (zt,k | z<t,· , zt,<k , u) (22) (Section A), which learn environment dynamics from data, t=1 k=1 and traffic scenario generation (Section B), which constructs initial conditions and test configurations for simulation. It where z<t,· denotes all tokens before time t, zt,<k denotes is worth noting that many of the methodological paradigms, tokens at time t before position k , and u aggregates conego especially in traffic scenario generation, are shared with trol variables (ego actions a , map context m, semantic methods presented in multi-agent and single-agent sections; constraints c). In diffusion form, models denoise latent reprehowever, the goal here is not to model how agents behave but sentations conditioned on past observations and controls. The to construct and generate the environments in which they are key differentiator from occupancy-based models discussed trained and tested. Table 4 provides a summary of paradigms in the next section is targeting view-level appearance and temporal coherence directly, which is useful for training and discussed in this section. evaluating end-to-end autonomy stacks that operate on raw sensor inputs. A. Generative World Models GAIA-1 [138] demonstrated that large-scale generative Driving world models learn data-driven simulators of how models can produce controllable scenario generation with the environment evolves under ego vehicle actions and emergent understanding of scene dynamics, geometry, and optional control variables. A common formulation is an action- agent interactions. GAIA-2 [19] extends this foundation with VOLUME ,
21
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
TABLE 4: Summary of neural simulation environment and scene generation methods for mixed autonomy traffic simulation. Method
Output Space Core Idea
Strengths
Limitations
Representative Studies
GAIA-1 [138], GAIA2 [19], Drive-WM [139], DriveDreamer [140], DriveDreamer-2 [141], Vista [142], Copilot4D [143], MUVO [144], DrivingWorld [145], GenAD [146] OccWorld [147], OccSora [148], DOME [149], T3 Former [150], I 2 World [151], DriveOccWorld [152], OccLLaMA [153]
Generative World Models Video-Based & Latent World Models
Multi-view images; latent tokens; video frames
Learns action-conditioned video or latent generation via autoregressive token prediction or latent diffusion; produces future visual observations for end-toend evaluation
Photorealistic rollouts; end-to-end stack evaluation; controllable via actions, text, and structured inputs; emergent scene dynamics
Long-horizon drift; high computational cost; causal consistency under intervention not guaranteed; image quality metrics miss driving-critical correctness
OccupancyBased World Models
3D/4D semantic occupancy grids; BEV fields
Predicts future semantic occupancy conditioned on ego actions via autoregressive or diffusion generation in voxel or triplane space
Geometric and semantic structure; directly supports planning cost evaluation; captures free-space and drivablearea evolution
No agent identity or intent; compounding topology errors; memory-intensive voxel grids; expensive 4D labels
Generates traffic configurations ranging from static actor placements to full dynamic scenarios; controllable methods expose control interfaces (language, temporal logic, guidance functions, retrieval tags) for steerable generation Discovers failure-inducing scenarios via optimizationbased search (RL adversaries, importance sampling, fuzzing) or safety-targeted generative synthesis (guided diffusion, latent-space optimization with learned traffic priors)
Data-driven scene diversity; user-specified controllability; intuitive language interfaces; compositional control; supports scenario augmentation and targeted testing
Realism–control trade-off; stronger constraints push off data manifold; unstandardized control interfaces; scene initialization methods lack dynamic validation; limited closed-loop validation
SceneGen [154], TrafficGen [155], DriveSceneGen [156], CTG [99], CTG++ [100], Scenario Diffusion [157], LCTGen [158], ChatSim [159], RealGen [160]
Efficient failure discovery; targeted stress testing; identifies safety-critical edge cases; generative methods maintain realism via learned priors; complements naturalistic evaluation
Validity–criticality trade-off; adversarial drift to implausible scenarios; coverage gaps; failure finding ̸= risk estimation; guided generation provides soft not hard constraint satisfaction
AST [161], [162], Accelerated Eval. [163], STRIVE [164], KING [165], AdvDiffuser [166], CAT [167], AVFUZZER [168], CCDiff [169], CaDRE [170], BridgeGen [171], MADS [172]
Traffic Scenario Generation Scene Initialization & Scenario Synthesis
Static scene layouts; multi-agent trajectories; map + traffic configurations
Safety-Critical Generation & Adversarial Falsification
Scenario parameters; adversarial agent trajectories; safety-critical multi-agent configurations
Output Space indicates the primary representation generated by each method family. Generative world models produce environment-level observations, while scenario generation methods produce simulation inputs (initial conditions, agent configurations, and trajectory specifications).
latent diffusion and achieved multi-camera consistency, finegrained control over agent configurations and environmental factors, and geographic diversity across multiple countries. GAIA-2 and its successor GAIA-3 enable the generation of safety-critical scenarios that are rare in naturalistic data but essential for behavior model evaluation. Drive-WM [139] demonstrated that world models can directly serve planning by rolling out futures under different maneuvers and scoring with image-based rewards. This established a pathway from world model predictions to behavioral decisions. DriveDreamer [140] proposed diffusion-based world models trained on real-world data with structured constraints (3D bounding boxes, HD maps, ego actions) to improve controllability. Its extension, DriveDreamer-2 [141] added LLM interfaces converting natural language user intent into trajectories for customized driving video generation. Vista [142] emphasized high fidelity and versatile controllability for generalizable driving world models by demonstrating long-horizon actionconditioned video prediction across diverse driving domains. Additional architectural contributions address scalability and multimodal fusion. Copilot4D [143] learns unsupervised world models by tokenizing LiDAR observations and predicting future tokens via discrete diffusion. MUVO [144] bridges video-based and geometry-aware representations by fusing camera and LiDAR inputs to predict future observations 22
in multiple formats (images, point clouds, 3D occupancy grids). It shows that multimodal sensor fusion improves both camera and LiDAR forecasting quality. DrivingWorld [145] introduces spatial-temporal fusion mechanisms for longduration rollouts within GPT-style autoregressive frameworks, and GenAD [146] proposes a generalized predictive model that rolls latent states forward to sample ego-conditioned futures, which enables end-to-end planning through world model imagination. While these video-level world models produce visually compelling rollouts, their utility for behavior modeling depends on whether the generated dynamics are causally consistent under intervention. Evaluation remains a further challenge and interesting area to investigate. Standard image quality metrics (FID, FVD) do not measure driving-critical correctness, and standardized closed-loop benchmarks testing causal consistency under interventions are lacking. In addition, current models face long-horizon drift, where compounding errors accumulate over extended rollouts. Computational cost leads high-resolution multi-view generation become expensive for real-time closed-loop deployment. Hybrid supervision combining perceptual and dynamics losses, structured latent spaces separating geometry from appearance, and uncertaintyaware generation that flags unreliable rollouts could be considered for future research. A further open question is VOLUME ,
how to incorporate physical priors, such as vehicle dynamics, collision geometry, and road constraints, into video generation objectives to ensure rollouts are within physically valid space.
avoid collisions, and respect road boundaries because there are assessments that are more natural in occupancy space than in pixel space. However, although occupancy captures geometry current approaches do not explicitly capture agent identity or intent. In other words, an occupied voxel does 2) Occupancy-Based World Models not carry information about which agent occupies it, what Occupancy-based world models represent environment state that agent’s goal is, or how it would respond to interventions. using semantic occupancy fields (typically 3D voxel grids This makes occupancy world models complementary to, but or BEV representations with height) and learn to predict not a replacement for, the agent-level behavior models of future evolution conditioned on ego action. Let Ot ∈ Sections IV–V. Additional challenges include compound{1, . . . , C}H×W ×Z denote semantic occupancy at time t, ing error in closed-loop rollouts, representation trade-offs where H , W , and Z are spatial dimensions and C is (voxel grids are memory-intensive while tokenizers introduce the number of semantic classes (road, vehicle, pedestrian, compression loss), and expensive 4D occupancy labels. background). An occupancy world model learns: Future research could focus on semi-supervised training pθ Ot+1:t+Tf | O1:t , at:t+Tf −1 (23) schemes that reduce label requirements, hybrid pipelines connecting occupancy rollouts to neural rendering for sensorwhere O1:t represents occupancy history and at:t+Tf −1 de- realistic evaluation, and explicit agent-identity tracking within notes future ego actions over forecast horizon Tf . Compared occupancy representations. with video world models, occupancy models trade pixellevel realism for geometric and semantic structure closer to planning cost functions, such as free space, drivable area, B. Traffic Scene and Scenario Generation While world models focus on learning environment dynamics, dynamic occupancy. OccWorld [147] learns scene tokenizers for 3D occupancy scene and scenario generation focus on constructing the using GPT-like spatiotemporal transformers to autoregres- traffic configurations, such as initial states and trajectory sively generate future scene tokens jointly with ego motion, specifications. These methods address two complementary which established the foundational autoregressive paradigm needs. First, populating simulation with diverse, realistic for occupancy prediction. The other paradigm, diffusion-based traffic scenes that let users steer generation toward specific approaches, improve generation fidelity and controllability. configurations. Second, evaluating AV safety by generating OccSora [148] treats 4D occupancy generation as a core and discovering rare, failure-inducing conditions. We organize simulation primitive using diffusion-style generation for this subsection accordingly, focusing on these two purposes. long sequences with semantic structure, and DOME [149] uses continuous occupancy latents via an occupancy VAE with spatiotemporal diffusion transformers and trajectory- 1) Scene Initialization and Scenario Synthesis based resampling for strengthened controllability. Efficiency- This section covers methods that generate traffic configufocused architectures address the computational cost of dense rations, ranging from static scene layouts to full dynamic 3D prediction. T3 Former [150] compresses 3D occupancy scenarios with explicit control interfaces. These methods into triplanes and predicts triplane deltas autoregressively cover a spectrum from unconditional scene initialization, for real-time rollouts. I 2 -World [151] decouples tokenization which generates plausible starting conditions by sampling into intra-frame and inter-frame components, which reduces from learned data distributions, to controllable scenario redundancy. Drive-OccWorld [152] intends to Connecting synthesis, which exposes control variables that steer generated occupancy world models to planning and foundation-model outcomes toward user-specified constraints while maintaining interfaces. It adapts vision-centric 4D occupancy forecasting consistency with real-world traffic statistics. for end-to-end planning by evaluating candidate trajectories against occupancy-based cost functions. OccLLaMA [153] a: Scene initialization and unconditional scenario generation. introduces unified occupancy-language-action modeling for Scene initialization methods generate the starting conditions dynamics modeling, language conditioning, and multi-task for simulation without producing subsequent dynamic behavoutputs. It jointly tokenizes 3D occupancy, natural language ior. SceneGen [154] introduces autoregressive models that instructions, and ego actions into a shared discrete token sequentially insert actors (with class, position, bounding box, space, which enables a single model to simultaneously and velocity) given ego state and HD map, which leads to predict future scene states and plan actions. This unified generating realistic static traffic snapshots that can serve as token representation allows language commands to directly initial conditions for downstream simulation or trajectory condition occupancy rollouts that bridges scene understanding prediction models. Extending beyond static initialization, and decision-making within a single generative framework. TrafficGen [155] generates both initial actor layouts and For behavior modeling, occupancy world models offer dynamic multi-agent trajectories via learning from fragmented a geometric interface that is particularly well suited for real-world logs and supporting scenario augmentation through evaluating whether simulated agents maintain safe distances, sampling from the learned distribution. DriveSceneGen [156] VOLUME ,
23
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
further advances this direction by jointly synthesizing both static map elements and dynamic traffic participants from scratch, which reduces reliance on preselected map regions and enabling generation of entirely novel road geometries paired with plausible traffic. These methods generate diverse scenarios by sampling from learned data distributions but do not offer explicit user control over the generated outcomes, which is a limitation addressed by the controllable methods discussed next. b: Controllable scenario synthesis.
language-commanded modification of the visual appearance of driving scenes (for instance, adding, removing, or repositioning rendered vehicles). While ChatSim’s controllability interface is relevant to scenario specification, its primary contribution lies in sensor-level scene editing rather than traffic behavior synthesis. RealGen [160] introduces retrievalaugmented generation retrieving template scenarios, including rare tagged cases for compositional behavior synthesis. A key concern in scene and scenario generation methods is the progression from unconditional to controllable generation because stronger constraints push samples off the data manifold. Moreover, evaluation metrics and pipelines that measure both constraint satisfaction and realism are lacking. For scene initialization methods, ensuring that generated starting conditions lead to realistic spawn of vehicles and plausible dynamic evolution when paired with downstream behavior models is still an open problem. Future research directions include compositional control building scenarios from reusable primitives, hybrid retrieval-plus-generation systems targeting rare events while producing novel variations, tighter coupling between scenario synthesis and downstream evaluation objectives, and unified frameworks that jointly generate scene layouts, initial conditions, and controllable dynamic trajectories within a single generative pipeline. Also, verification against traffic flow theory foundations seems essential for ensuring the realism of generated scenes and scenarios.
Controllable scenario synthesis learns generative distributions over scenarios by exposing control variables that steer generated outcomes. Assume a scenario consisting of HD map m and multi-agent trajectories τ = {τ 1 , . . . , τ N }, controllable synthesis learns the conditional distribution, pθ (τ | m, c), where c encodes user intent or constraints, which can take multiple forms, such as semantic tokens, behavior tags, goal endpoints, interaction types, rule specifications, or natural language descriptions. Control is applied either by explicit conditioning during training or through guided sampling that biases generation toward satisfying constraints via auxiliary score functions. In this class, CTG [99] proposes conditional diffusion with sampling-time guidance using differentiable constraints including temporal-logic specifications, demonstrating that diffusion models can generate socially consistent multi-agent trajectories while offering controllability through differentiable cost function guidance. In practice, users may specify constraints such as collision avoidance or goal reaching, and 2) Safety-Critical Scenario Generation and Adversarial the reverse diffusion process is steered toward satisfying them Falsification without retraining. CTG++ [100] extends this framework with While the methods of Section 1 generate diverse and language-based conditioning to translate natural language controllable traffic scenarios for general-purpose simulation, a scenario descriptions into guidance signals that steer multi- particularly important sub-problem is the efficient discovery agent trajectory generation. Scenario Diffusion [157] frames of safety-critical scenarios, such as near-misses, or edge controllability through conditioning tokens and map context cases. This section surveys methods that address this subusing latent diffusion. These foundational controllable gener- problem through two complementary strategies: optimizationation mechanisms support several safety-critical generation based adversarial search, which uses reinforcement learning, methods discussed in Section 2, where the same architectural importance sampling, or fuzzing to search the scenario space principles are specialized toward adversarial and failure- for failure-inducing conditions; and safety-targeted generative inducing objectives. synthesis, which specializes the generative models introduced in Section 1 by directing their sampling toward safety-critical c: Language-conditioned and retrieval-augmented regions of the scenario distribution. The central challenge approaches. shared across both strategies is balancing criticality (finding Language-conditioned approaches provide intuitive interfaces scenarios that expose failures) with validity (ensuring those for scenario specification, bridging natural language intent scenarios remain physically plausible and represent situations and generated traffic behavior. LCTGen [158] uses natural that could actually occur). language as supervision and combines LLMs with generaIt is important to distinguish the purpose of these methods tive decoders for map selection, initial traffic distributions, from the adversarial imitation methods of Section D and and agent dynamics,. InteractTraj [173] translates language multi-agent adversarial learning of Section C: adversarial descriptions into structured codes conditioning trajectory imitation learns realistic agent policies by matching expert generators. ChatSim [159] operates at a different level of distributions, whereas the methods here use optimization to abstraction: rather than generating traffic behavior, it intro- test those policies by searching for failure-inducing conditions. duces collaborative LLM agents for editable photo-realistic The generated scenarios serve as evaluation infrastructure 3D scene rendering via neural radiance fields, which enables rather than as behavior models themselves. 24
VOLUME ,
a: Optimization-based adversarial search. targeted. For modeling hazardous interactions in mixed A first family of methods treats safety-critical scenario autonomy traffic, specifically, MADS [172] develops a multidiscovery as explicit optimization. Adaptive Stress Testing agent driving simulation framework designed to characterize (AST) [161], [162] casts failure search as MDPs where adver- safety-critical encounters between automated and manual saries select disturbance actions that drive the system toward vehicles, which allows systematic evaluation of AV safety failure while keeping trajectories probable under nominal under realistic adversarial conditions involving human-driven disturbance models. Accelerated evaluation methods [163] vehicles. The validity-criticality trade-off remains the central chalmodify surrounding vehicle behavior and use importance sampling for statistically meaningful risk estimates with lenge for safety-critical scenario generation methods. Adfar fewer simulation miles than naïve Monte Carlo. Multi- versarial search and safety-targeted generation can drift into agent adversarial RL extends this by training adversarial unrealistic behaviors unless constrained by strong priors. On agents that exploit AV policy weaknesses [174], [175], the other hand, overly conservative validity constraints may while hierarchical RL architectures manage reward sparsity prevent discovery of genuine edge cases. Coverage accountand promote diverse scenario generation [176]. Closed-loop ing is a further concern. Finding many failures does not Adversarial Training (CAT) [167] iterates between safety- necessarily imply meaningful operational domain coverage, critical scenario generation and driving policy optimization and the relationship between generated scenario distributions using probabilistic factorization, and through this, turn the log- and real-world occurrence frequencies requires statistical replay into adversarial scenario. Fuzzing frameworks provide frameworks that connect failure finding to principled risk a complementary approach, mutating scenario parameters estimation [180]. For safety-targeted generative methods, the guided by surrogate scores. AV-FUZZER [168] uses genetic challenge of ensuring that guided sampling produces causally algorithms with learned fitness functions, DriveFuzz [177] valid agent reactions, and not just statistically plausible combines mutation-based fuzzing with neural guidance, and trajectories that happen to be dangerous, is shared with TM-fuzzer [178] targets traffic management scenarios with the counterfactual validity concerns discussed in Section C. Worthwhile directions for investigation include tighter intelearned surrogate models. gration between optimization-based search and generative priors (using learned models to constrain the feasible search b: Safety-targeted generative synthesis. space while optimization targets criticality), likelihood-aware A second family specializes the generative scenario synthesis adversarial methods that yield unbiased risk estimates rather frameworks of Section 1 by directing their sampling toward than simply finding failures, multi-objective frameworks that safety-critical regions. These methods leverage learned traffic simultaneously optimize for criticality, diversity, and validity, priors to ensure that adversarial scenarios remain realistic and standardized evaluation protocols that measure both the rather than drifting into implausible configurations. realism and the coverage of generated safety-critical scenario STRIVE [164] uses VAE-based traffic models for gradient- sets. based optimization in latent space. The learned prior constrains the search to realistic behaviors even as the objective pushes toward dangerous configurations. KING [165] uses C. Discussions on Environment-level Simulations proxy dynamics and kinematics gradients for efficient adver- Neural simulation environments promise advances in realism sarial perturbation of recorded scenarios. Hao et al. [179] and controllability by learning environment dynamics through combine GAIL-based driver representations with PPO-based generative world models and scenario inputs through dataoptimization to balance adversariality and naturalness. Ad- driven synthesis. However, deploying these components as vDiffuser [166] builds on the guided diffusion foundations reliable infrastructure for behavior modeling raises challenges established by CTG (Section 1) and replaced general-purpose distinct from those faced by agent-level methods. controllability objectives with auxiliary reward functions A fundamental challenge is counterfactual validity under that steer the diffusion sampling process specifically toward intervention. Neural simulators must remain correct when safety-critical scenarios. CCDiff [169] similarly extends ego policy changes, agents are edited, or scene conditions the controllable diffusion paradigm and targeted long-tail are modified. This is inherently difficult because most and safety-critical settings by injecting causal structure training data is observational and does not uniquely identify and decomposition into the generation process for closed- counterfactual reactions. World models can produce visually loop evaluation. CaDRE [170] emphasizes controllability convincing rollouts that respond incorrectly when the ego plus diversity for safety-critical generation, specifically ad- vehicle deviates from logged behavior. Similarly, in scenario dressing mode collapse around single failure patterns to generation, adversarial search can exploit simulator weakensure a broad coverage of the safety-critical scenario space. nesses, producing unrealistic high-criticality failures unless BridgeGen [171] bridges data-driven and knowledge-driven validity constraints are strong and calibrated. approaches by combining statistical patterns from driving Long-horizon stability and compounding error presents data with explicit domain knowledge about safety-critical a second challenge. Closed-loop testing requires rollouts over situations to generate scenarios that are both realistic and many seconds with repeated feedback. Generative models VOLUME ,
25
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
suffer from drift where small early errors move simulated perceive, attend to, and evaluate the driving environment distributions out-of-support, which produces increasingly including risk assessment and trust toward other agents implausible states, particularly acute for high-dimensional (Section B), and how physical priors can be encoded representations and in multi-agent settings where interaction into neural architectures to enforce dynamic plausibility feedback creates coupling between agents’ errors. The gap (Section C). We conclude with a discussion of integration between environment fidelity and behavioral realism is challenges and promising directions for bridging cognitive specific to the environment-level perspective. A world model and data-driven paradigms (Section D). Table 5 provides a may generate visually or geometrically plausible scenes summary of the reviewed methods and studies in this section. whose behavioral content (how agents move and interact) does not match the quality of purpose-built agent models. A. Bounded Rationality and Decision-Making Models Conversely, sophisticated agent-level behavior models may be undermined by environment models that provide unrealistic Standard reinforcement learning (Section C) and gamefeedback. Aligning environment-level and agent-level fidelity theoretic models (Section C) assume agents that maximize is essential but rarely addressed explicitly. A further gap expected utility. However, behavioral research demonstrate is risk estimation. Adversarial failure finding differs from that human decision-making under uncertainty may be estimating real-world risk, which necessitates statistical governed by bounded rationality [9], which means drivers frameworks connecting generated scenario distributions to use heuristics, overweight rare catastrophic outcomes, exhibit loss aversion, and make choices that are satisfactory rather real-world occurrence frequencies. Promising future directions include: (1) hybrid simula- than optimal. Incorporating these biases into behavior models tors combining explicit structure (maps, kinematics, traffic produces agents whose decisions are not merely statistically rules) with learned residual generation; (2) uncertainty-aware plausible but psychologically supported. Cumulative Prospect Theory (CPT) [181] provides a simulation exposing confidence estimates and supporting formal framework for bounded rationality. CPT replaces the conservative rollout when out-of-distribution; (3) structured linear utility and probability weighting of expected utility controllability interfaces aligned with engineering workflows; theory with an S-shaped value function, and a probability (4) causal and interventional training objectives going beyond weighting function that overweights small probabilities and matching observational distributions to explicitly model underweights large ones. Applied to driving, this means counterfactual responses; and (5) standardized evaluation that drivers disproportionately weigh the small probability protocols that assess both environmental realism and behavof a catastrophic collision relative to the certain time cost ioral fidelity of generated scenarios. of driving slowly. Sun et al. [183] show that the CPTbased formulation yields an interpretable explanation of VII. COGNITIVE AND PHYSICS INFORMED AI METHODS behaviors such as cautious versus risk-seeking responses The data-driven methods surveyed in Sections IV to VI during interaction, and it can reproduce different driving optimize for trajectory accuracy against recorded behavior. tendencies by tuning CPT parameters. Risk allostasis theHowever, they often cannot explain why human drivers ory [184] and its computational instantiations [185], [186] behave as they do. A model that matches logged trajectories provide a complementary perspective. They suggest that rather with low displacement error may still fail to reproduce the than optimizing any utility function, drivers continuously cognitive mechanisms that generate those trajectories. This adjust their behavior to maintain a subjective target risk distinction matters for mixed autonomy simulation in at level. Models incorporating this homeostatic mechanism yield least three ways. First, purely data-driven agents cannot higher fidelity to naturalistic driving data than efficiencypredict human error modes, such as inattention, misjudg- maximizing algorithms. Prospect-balancing theory [182] ment, and overload failures that dominate real-world crash further formalizes this insight by conceptualizing speed choice causation. Second, without cognitive grounding, behavioral as a trade-off between the prospect of arriving sooner and the diversity in simulation relies on arbitrary latent variable prospect of avoiding a collision, with subjective probability sampling (Section C) rather than on principled variation distortions favoring cautious choices. of psychologically meaningful parameters such as reaction For the dynamic and timely decisions characteristic of time, risk tolerance, or attentional capacity. Third, as mixed driving, such as gap acceptance at intersections, merge autonomy traffic introduces novel interaction patterns between initiation, and pedestrian yielding, drift-diffusion models humans and AVs (patterns absent from historical driving logs) (DDMs) from cognitive psychology offer a process-level cognitively grounded models offer a basis for extrapolation account of how decisions unfold over time [187]. In DDMs, and generalization that distribution-matching methods lack. a decision variable accumulates noisy evidence toward This section reviews methods that incorporate validated one of two boundaries (e.g., “accept gap” versus “wait”). theories from cognitive psychology, human factors, and traffic Markkula et al. [188] demonstrated that variable-drift DDMs physics into learning-based behavior models. We organize accurately capture not just mean response times but the full the discussion around three complementary perspectives: distribution of response times in gap acceptance scenarios. how drivers decide under uncertainty (Section A), how they This distributional accuracy is critical for safety validation, 26
VOLUME ,
TABLE 5: Summary of cognitive and physics-informed AI methods for mixed autonomy traffic simulation. Method
Theory Basis
Core Idea
Strengths
Limitations
Representative Studies
Interpretable style variation via CPT parameters; reproduces conservative/aggressive profiles without separate rewards; psychologically validated
Parameter calibration from driving data is difficult; limited to single-decision settings; does not model temporal dynamics of choices
Tversky & Kahneman [181], Schmidt et al. [182], Sun et al. [183]
High fidelity to naturalistic carfollowing and speed choice; captures risk compensation effects; simple parameterization
Target risk level is latent and hard to estimate; limited to longitudinal control; does not explain strategic interaction
Fuller [184], Mohammadian et al. [185], Kashifi [186]
Captures full RT distribution, including dangerous late decisions; perceptual inputs (looming, TTC) as natural drift rates; differentiable neural implementations Principled heterogeneity in interactive behavior; avoids full-rationality assumption; calibratable from observed interactions
Binary decisions only; limited multi-alternative extensions; requires careful specification of evidence signals; calibration from naturalistic data is challenging
Ratcliff [187], Markkula et al. [188], Zgonnikov et al. [189], McDonald et al. [190], Fengler et al. [191]
Requires specifying level-0 heuristic; distribution over levels is scenario-dependent; computational cost grows with k
Stahl & Wilson [117], Chong et al. [118], Li et al. [192]
Interpretable parameters with psychological meaning; predict error modes (inattention, overload); principled generalization to novel scenarios; validated against human data Prevents causal confusion; provides explainability; captures perceptual biases; large-scale gaze datasets available
Hand-crafted production rules; operate in simplified environments; difficult to scale to high-fidelity scenarios; limited continuous control
Anderson [193], Salvucci [194], Cao et al. [195], Ebadi et al. [196], Laird [197], Zhou et al. [198]
Gaze ̸= attention (covert shifts missed); dataset bias toward normal driving; limited integration with closed-loop control
Palazzi et al. [199], Xia et al. [200], Fang et al. [201], Kultrera et al. [202]
Robust to high-speed/lowlatency scenarios; interpretable intermediate affordances; validated dual-task degradation predictions (R2 > 0.90)
SNNs lack mature training frameworks; affordance definition requires domain expertise; resource models calibrated from lab, not naturalistic settings
Ma et al. [203], Chen et al. [204], Wickens [205], Horrey & Wickens [206], Engström et al. [207]
Implicit safety guarantees; reduced data requirements; captures subjective risk biases (overestimation of head-on, underestimation of lateral); interpretable field visualization
Risk field calibration is driverspecific; limited to reactive control; does not model strategic anticipation; field superposition assumptions may be simplistic
Wang et al. [208], Ji et al. [209], Kolekar et al. [210]
Essential for mixed autonomy transition modeling; ML-based prediction from physiological and behavioral features; captures asymmetric trust repair after failures
Trust is latent and difficult to measure at scale; models calibrated from simulator studies, not naturalistic driving; limited integration with trajectory-level behavior models
Lee & See [211], Hoff & Bashir [212], Ayoub et al. [213], Kaufman et al. [214]
Limited to longitudinal models (car-following); physics priors may be too rigid for complex urban scenarios; fusion strategy selection is ad hoc
Mo et al. [215], Mo et al. [216], Xu et al. [217], Geng et al. [218], Wang et al. [219]
Bounded Rationality and Decision-Making Models ProspectCumulative Theoretic Prospect Decision Models Theory (CPT); Prospect Balancing Risk Allostasis & Homeostatic Models
Risk Allostasis Theory; TaskCapability Interface
Drift-Diffusion Models (DDMs)
Sequential Sampling; Evidence Accumulation
Level-k Reasoning
Cognitive Hierarchy; Behavioral Game Theory
Replaces expected utility with S-shaped value function and nonlinear probability weighting; captures loss aversion and overweighting of rare catastrophic outcomes in driving decisions Drivers adjust behavior to maintain a subjective target risk level rather than optimizing utility; explains systematic deviations from time-optimal behavior Decision variable accumulates noisy perceptual evidence toward boundaries; models the temporal dynamics and variability of gap acceptance, braking, and yielding decisions Agents have heterogeneous strategic depth: level-0 follows heuristics, level-k best-responds to level-(k−1); models realistic variation in anticipatory reasoning
Cognitive Architectures, Attention, Risk Perception, and Trust Models Cognitive Architectures (ACT-R, SOAR)
Attention & Gaze Prediction
Brain-Inspired & Affordance Models
Risk Fields & Potential-Based Models
Trust Dynamics in Human–AV Interaction
ACT-R; SOAR; Unified computational theories Production of perception, memory, and Systems motor execution with constrained processing resources; predict realistic latencies and error modes from cognitive bottlenecks Selective Models where human drivers Attention; look using learned gaze maps Visual Saliency; combining bottom-up saliency Foveal– and top-down task cues; weights Peripheral features by predicted attention for Processing downstream models Ecological Spiking neural networks mimic Perception; visual cortex; affordance-based Multiple representations extract drivingResource relevant variables; resource theory Theory; predicts dual-task interference Cognitive and distraction effects Control Hypothesis Driving Safety Surrounding objects and Field; Artificial road elements emit repulsive Potential Fields potentials; trajectories minimize cumulative risk exposure; learned risk fields capture subjective risk perception differing from objective TTC Trust Models how human drivers adjust Calibration; behavior based on evolving Three-Layer trust toward AVs; captures Trust Model dispositional, situational, and learned trust influencing headway, gap acceptance, and interaction willingness
Physics-Informed Learning PhysicsIDM; OVM; Informed Deep FVDM; CarLearning (PIDL) Following Theory
Encodes classical traffic flow Outperforms pure physics models as structural priors or and pure neural models; regularization in neural networks; strong in sparse-data regimes; combines physics rigor with data- enforces kinematic plausibility; driven flexibility for car-following distributional estimates via PIDLand trajectory prediction GAN
Theory Basis indicates the primary cognitive, psychological, or physical theory grounding each method family. Methods are organized by whether they address decision-making under uncertainty, cognitive processing and evaluation including risk perception and trust, or physical and kinematic constraints.
VOLUME ,
27
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
where the probability of rare late responses determines crash risk. Subsequent work extended DDMs to looming-based evidence accumulation for braking decisions [189], [190]. They showed that perceptual variables (optical expansion rate, time-to-collision) serve as natural drift-rate inputs. Neural implementations of evidence accumulation [191] enable endto-end learning of accumulation dynamics from behavioral data, which bridges the gap between cognitive process models and scalable deep learning architectures. The level-k reasoning framework introduced in Section C for game-theoretic planning has deep roots in cognitive science [117], [118]. Its relevance here is that it provides a principled cognitive model of strategic depth heterogeneity: some drivers plan several steps ahead (high-k reasoners anticipating others’ responses), while others react to immediate stimuli (level-0 heuristic followers). Calibrating the distribution of reasoning levels from naturalistic data could produce traffic populations with realistic heterogeneity in interactive scenarios. For simulation, the key advantage of bounded rationality models is that behavioral diversity emerges from variation in interpretable cognitive parameters, such as risk sensitivity, evidence accumulation rate, decision boundary, reasoning depth, rather than from obscured latent variables. This enables test engineers to construct specific driver profiles (e.g., “distracted driver with slow evidence accumulation and high risk tolerance”) grounded in psychological foundations, which could support targeted scenario design for safety-critical evaluation.
response times. SOAR-ACPPO [198] combined a rule-based cognitive control [197] with deep reinforcement learning for lane change decisions. It uses SOAR to structure the decision space and RL to optimize within it. This hybrid paradigm addresses the problem of hand-crafting production rules for complex continuous control while preserving the interpretability and cognitive fidelity that pure RL lacks. For simulation, cognitive architectures offer three advantages that neural networks alone cannot provide. First, their parameters have direct psychological meaning (e.g., visual sampling frequency, memory decay rate, production rule latency). This property allows principled transfer across scenarios and populations. Second, they predict error modes, the specific failures arising from cognitive overload, inattention, or memory interference. This is important because these are the human behaviors that AVs must handle safely. Third, they support principled generalization to novel scenarios that lack training data, because behavior is generated from cognitive mechanisms rather than pattern-matched to observed distributions. b: Attention and perception models.
Human drivers do not process visual scenes uniformly: foveal vision captures detail in a narrow central region while peripheral vision detects motion and salience, and selective attention determines which scene elements receive processing resources [220]. This stands in contrast to the uniform feature extraction assumed by most neural driving models, which process entire bird’s-eye-view representations or multicamera inputs with equal weight. Large-scale driving attention B. Cognitive Architectures and Attention-Guided Models datasets, such as DR(eye)VE [199] (74 sequences with eyeWhile bounded rationality models address the decision- tracking and physiological signals), BDD-A [200] (over 1,200 making process itself (how drivers evaluate options, accu- videos with gaze annotations), and DADA-2000 [201] (2,000 mulate evidence, and select actions) this subsection covers accident videos with driver attention labels), could help the perceptual and cognitive processing from which those with training attention prediction networks that model where decisions arise, such as the processing bottlenecks and human drivers actually look. capacity limitations that cause errors. It also covers two Xia et al. [200] fused bottom-up visual saliency (sceneclosely related aspects of cognitive evaluation: how drivers driven conspicuity) with top-down task-driven cues (goalassess risk through subjective threat perception, and how they directed attention allocation) by using predicted gaze maps calibrate trust when interacting with automated vehicles. to weight feature importance in downstream driving models. This attention-guided weighting prevents causal confua: Cognitive architectures. sion [74] where end-to-end models learn spurious correlations Cognitive architectures provide unified computational theories (e.g., associating stopped vehicles with red traffic lights in of human information processing. ACT-R (Adaptive Control the background rather than with the vehicle itself). Kultrera of Thought-Rational) [193] has become one of the most et al. [202] demonstrated that attention-based architectures extensively validated architectures for driving. Salvucci’s provide explainability for end-to-end steering and help verify foundational work [194] established a comprehensive driver whether a model’s failures correspond to plausible human model based on ACT-R by integrating control (steering and perceptual limitations. speed regulation), monitoring (situation assessment through Brain-inspired architectures push this further. DeepDrivvisual sampling), and decision-making (tactical choices ing [204] operationalized Gibson’s ecological theory of such as lane changes). Later, QN-ACTR [195] integrated affordances and trained CNNs to estimate driving-relevant queueing network theory to model multitask performance affordances (distance to lane markings, angles to surrounding and predict how secondary tasks degrade driving, and QN- vehicles) as intermediate representations rather than mapping ACTR-SA [196] extended situation awareness modeling, directly from images to control. The aim was to produce which resulted in better performance in predicting brake more interpretable and transferable behavior models. Wickens’ 28
VOLUME ,
Multiple Resource Theory [205] provides a complementary framework for predicting dual-task interference. When secondary tasks (phone use, conversation, navigation) compete for the same perceptual or cognitive resources as driving, performance degrades predictably. Horrey and Wickens [206] could predict driver attention allocation under dual-task conditions with high accuracy, and Engström et al.’s Cognitive Control Hypothesis [207] refined this by showing that cognitive load selectively impairs executive-dependent subtasks, such as hazard detection and tactical decisions, while preserving automatized behaviors like lane-keeping and speed maintenance. Collectively, these brain-inspired and resourcetheoretic approaches demonstrate that encoding the structure of human cognition and attention into computational models yields not only more faithful representations of driver behavior but also principled predictions of the context-dependent failure modes that autonomous vehicles must be validated against. c: Risk fields and potential-based models.
The Driving Safety Field (DSF) theory [208] conceptualizes driving behavior as navigation through subjective risk fields generated by surrounding objects, road boundaries, and traffic infrastructure. Each obstacle and road element emits a repulsive potential that decays with distance and depends on relative kinematics, and the driver’s trajectory minimizes cumulative exposure to the resulting composite field. Deep learning extensions learn to map sensor inputs to continuous potential surfaces, with ego-vehicles performing gradient descent on learned risk. Building on this, recent approaches employ graph representations paired with deep reinforcement learning to navigate these fields. By modeling traffic participants and road topology as interconnected graphs, these DRL agents can efficiently process complex, multi-agent potential fields to learn safe navigation policies [209]. This formulation provides implicit safety guarantees while drastically reducing data requirements compared to unconstrained end-to-end learning. Kolekar et al. [210] introduced the Driver’s Risk Field (DRF), a learned two-dimensional field representing subjective collision probability that varies across drivers and scenarios. The DRF captures the empirical finding that drivers’ perceived risk differs systematically from objective risk metrics such as TTC. Drivers overestimate risk from vehicles approaching on a collision course while underestimating risk from vehicles in adjacent lanes.
closely after observing smooth driving but maintain excessive headway after witnessing an abrupt stop, and why trust recovery after automation failures is slow and asymmetric. In a similar work, Hoff and Bashir [212] distinguish dispositional trust (personality-based baseline), situational trust (contextdependent adjustment), and learned trust (experience-based calibration). Machine learning approaches increasingly enable datadriven trust modeling. Ayoub et al. [213] used XGBoost with SHAP explanations to predict trust from driving context and behavioral indicators, which provided accuracy and interpretability regarding which factors most influence trust dynamics. Kaufman et al. [214] achieved high accuracy in classifying trust levels using random forests trained on physiological and behavioral features. For simulation, trust-aware agent models are useful for representing the transition period of mixed autonomy deployment. As AV penetration rates increase, human drivers gradually calibrate their interaction strategies. Integrating trust dynamics into the hybrid data-driven agents discussed in Section C represents a promising direction for generating realistic mixed traffic populations. C. Physics-Informed Learning
A complementary approach to cognitive grounding is encoding physical laws and traffic flow theory directly into neural architectures. While cognitive models and bounded rationality methods constrain how agents process information and make decisions, physics-informed methods constrain what behaviors are physically and dynamically plausible. This addresses the well-known problem that purely data-driven models can generate trajectories violating basic kinematics, producing negative headways, or ignoring road geometry. a: Physics-informed deep learning for traffic behavior.
The physics-informed deep learning (PIDL) paradigm, introduced by Mo et al. [215], encodes established car-following models, including Intelligent Driver Model (IDM), Optimal Velocity Model (OVM), and Full Velocity Difference Model (FVDM), directly into neural network architectures as structural priors or regularization terms. Rather than learning carfollowing behavior from scratch and with no prior knowledge from data, PIDL networks are constrained to produce outputs consistent with the functional forms of these classical models while retaining the flexibility to capture residual complexity that physics alone misses. They demonstrate that physicsd: Trust and adaptation in human–AV interaction. informed networks substantially outperform both pure physics For mixed autonomy traffic, an important behavioral di- models and pure neural networks, particularly in sparse-data mension is trust: how human drivers calibrate their ex- regimes where training samples are limited. Xu et al. [217] pectations and adjust their behavior when interacting with extended this paradigm with improved IDM formulations automated vehicles whose driving patterns may differ from and complementary fusion strategies. Geng et al. proposed a familiar human norms. Lee and See [211] establish three physics-informed Transformer architecture (PIT-IDM) [218] trust dimensions: performance (competence-based), process by combining the long-range temporal modeling of self(understanding-based), and purpose (intent-based). Their attention with IDM-based structural constraints for highway framework explains why the same driver may follow an AV trajectory prediction and achieved improvements over both VOLUME ,
29
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
physics-uninformed transformers and standalone IDM. Mo et al. [216] developed a physics-informed generative adversarial network for uncertainty quantification in car-following. They produced point predictions and calibrated distributional estimates of acceleration conditioned on traffic regime. Most recently, the Knowledge-Informed Deep Learning (KIDL) paradigm [219] distills high-level driving knowledge from large language models into lightweight neural car-following models, which combines LLM-derived priors about traffic regulations and driving norms with data-driven calibration for improved cross-dataset generalization.
capacity) within a single training framework would produce agents that are simultaneously physically plausible and cognitively realistic. A further open challenge is calibration at scale. Cognitive model parameters (risk sensitivity, attention capacity, trust propensity) are typically estimated from small-sample laboratory experiments. Developing methods to calibrate these parameters from large-scale naturalistic driving datasets would unlock the full potential of cognitively grounded simulation. The recent availability of datasets combining vehicle trajectories with driver physiological signals (eye tracking, electrodermal activity) opens new possibilities for this calibration. D. Integration Challenges and Future Directions To summarize, cognitive and physics-informed methods The methods reviewed above offer complementary benefits. provide value that purely data-driven approaches cannot Bounded rationality models explain decision variability replicate. They predict the variability and failure modes that through psychologically meaningful parameters, cognitive matter most for safety, their parameters carry psychological architectures predict processing limitations and error modes, or physical meaning that allows principled transfer across attention models capture perceptual biases, risk fields rep- populations and scenarios, and they capture the bounded resent subjective threat evaluation, trust models represent rationality that fundamentally distinguishes human driving adaptation in human–AV interaction, and physics-informed from the optimization-based behavior of automated systems. methods enforce dynamic plausibility. However, each operates Therefore, the integration challenge represents one of the at a different level of abstraction and computational cost, and most important and valuable research directions in behavior integrating them into cohesive, scalable simulation agents modeling for automated vehicle simulation. remains an open challenge. A key challenge is the tradeoff between cognitive validity and data-driven scalability. Cognitive architectures such as ACT-R produce interpretable, VIII. CHRONOLOGICAL LANDSCAPE OF AI METHODS theory-grounded behavior but require hand-crafted production FOR DRIVING BEHAVIOR MODELING rules and operate in simplified environments. Conversely, data- Figure 3 presents a timeline view of the AI methods reviewed driven models scale to complex scenarios but lack cognitive in this survey, organized by methodological family and foundations, which results in agents that match trajectory mapped to the period in which key contributions emerged. The distributions without capturing the generative mechanisms timeline reveals three broad eras of development. The founbehind them. dational era (pre-2016) was marked by pioneering but mainly Several promising integration strategies could be consid- isolated efforts, most notably early neural network–based ered. First, instead of building full cognitive architectures, key imitation learning and recurrent architectures for trajectory cognitive constraints (processing delays, attention bottlenecks, prediction. The transformers era (roughly 2021–2022) brought bounded lookahead) can be encoded as structural priors or a convergence of advances: transformer-based architectures regularization terms in neural network training, which is anal- permeated nearly every methodological family simultaneogous to how PIDL encodes physics. Evidence accumulation ously, from prediction and imitation learning to multi-agent dynamics, for instance, can be implemented as differentiable reinforcement learning and reactive simulation, while latent neural layers [191] within end-to-end architectures. This world models and traffic scenario generation methods also combines DDM-like decision processes with learned feature matured during this period. The most recent generative extraction. Second, LLM-enhanced cognitive modeling seems AI era (2023–2024) is characterized by an explosion in to be a promising direction [221]. Wu et al. [222] explored both breadth and diversity of approaches: diffusion-based using large language models to assist in developing and simulation, occupancy and video world models, LLM/VLMextending cognitive architectures, leveraging LLMs’ broad based planning, game-theoretic prediction, and controllable knowledge to generate production rules and parameter scenario generation all emerged in a compressed timeframe. hypotheses that would otherwise require extensive manual Three patterns in this evolution are particularly noteworthy. effort. Third, cognitive architectures can generate synthetic First, there is a clear top-down temporal cascade across behavioral data, such as rare error modes and edge cases abstraction levels: agent-level methods such as imitation underrepresented in naturalistic datasets, to augment training and prediction matured earliest, followed by environmentsets for neural models. It has been demonstrated by recent level approaches like simulation and world models, and most work using cognitive-based driver models to produce critical recently by foundation model methods that attempt to unify scenarios for AV testing [223]. Fourth, combining physics- perception, reasoning, and planning. Second, the timeline informed constraints (kinematic feasibility, road compliance) illustrates an accelerating methodological convergence. While with cognitive constraints (bounded reaction time, attentional early work in each family developed largely in isolation, 30
VOLUME ,
Neural Driving
Imitation Prediction
RNN-LSTM
RL/IL
Transformer Fusion
End –to-End Planning
GNNs
Transforme rs
GameTheoretic
Modelbased RL
Multi-Agent
Conditional IL
Deep IL
GAIL
ReturnConditioned
Reactive Sim
Simulation
Autoregressive
Diffusion Latent World Models
World Models Foundation
Video World Models
VLM Planning
LLM Agents
Scenario Gen
Adversarial
Physics-Cognitive
Cognitive Foundation
1989 1992
Foundational
DDMs & Risk Fields
~~
2016
2017
2018
2019 - 2020
Traffic Gen
Controllable
PIDL
Cog-Physics Hybrids
2021 - 2022
2023 – 2024
Transformers era
Occupancy World Models
Gen. AI
2025-2026
FIGURE 3: Timeline of model development in driving simulation and virtual evaluation
recent methods increasingly draw on shared architectural a: Displacement metrics. building blocks (e.g., transformers, diffusion processes) and The most widely adopted metrics measure Euclidean distance blur traditional boundaries between prediction, planning, between predictions and ground-truth. Average Displacement and simulation. This convergence suggests that the field Error (ADE) computes the mean L2 distance across all is moving toward more holistic frameworks. However, the predicted timesteps: cognitive and physics-informed methods seem to follow Tf 1 X a distinct developmental trajectory compared to the data∥ŷt − yt ∥2 (24) ADE = Tf t=1 driven families. Rather than advancing through architectural innovations (from RNNs to transformers to diffusion models), Final Displacement Error (FDE) measures accuracy at the this family evolved from established theoretical foundations prediction horizon endpoint: in psychology, human factors, and traffic flow theory toward computational integration with deep learning. This dual FDE = ∥ŷTf − yTf ∥2 (25) evolution highlights a maturing field that must reconcile the For multi-modal predictions generating K trajectory hyrapid scalability of data-driven architectures with the slower potheses {Ŷ(k) }K k=1 , the minimum variants select the bestbut more principled development of theory-grounded models matching mode: to achieve truly realistic mixed autonomy simulation. Tf 1 X (k) minADEK = min ∥ŷt − yt ∥2 , (26) k∈{1,...,K} Tf t=1 IX. EVALUATION AND BENCHMARKING
For learning-based behavior models in mixed autonomy traffic, the core question is whether simulated rollouts are (1) behaviorally realistic for both automated vehicles and human drivers, (2) interactive under feedback (agents react plausibly to each other), and (3) safe and rule-compliant under controlled stress. This section presents evaluation protocols, metrics, and benchmarking practices organized by evaluation paradigm and application context.
A. Open-Loop Evaluation Metrics
Open-loop evaluation compares predicted trajectories or actions against ground-truth recordings without environmental feedback. Given predicted trajectory Ŷ = {ŷ1 , . . . , ŷT } and ground-truth Y = {y1 , . . . , yT } over prediction horizon T , standard metrics quantify displacement accuracy, multi-modal coverage, and probabilistic calibration. VOLUME ,
minFDEK =
min k∈{1,...,K}
(k)
∥ŷTf − yTf ∥2
(27)
Common choices include K ∈ {1, 5, 6, 10} depending on benchmark conventions [23], [24]. b: Miss rate and probabilistic metrics.
Miss Rate (MR) measures the fraction of predictions where no hypothesis falls within distance threshold δ of ground-truth: (k) δ MRK = ⊮ min ∥ŷT − yT ∥2 > δ (28) k
Typical thresholds include δ = 2.0m for the Waymo Open Motion Dataset [23]. The Brier minimum FDE used in Argoverse 2 [24] jointly evaluates displacement and probability calibration: (k)
brier-minFDE = min ∥ŷT − yT ∥2 + (1 − pk )2 k
(29) 31
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
where pk is the predicted probability for mode k . For methods c: Comfort and rule compliance. outputting trajectory distributions, Negative Log-Likelihood Longitudinal Jerk and Lateral Acceleration measure ride (NLL) measures how well the predicted distribution covers comfort through motion smoothness. Metrics are typically ground-truth. Soft mAP from the Waymo Motion Predic- reported as maximum values or fraction of time exceeding tion Challenge adapts object detection Average Precision comfort thresholds (e.g., jerk > 4 m/s3 , lateral acceleration to trajectory forecasting, using soft assignment based on > 3 m/s2 ). Traffic rule adherence is evaluated through Red distance thresholds across multiple horizons and incorporating Light Violation Rate, Speed Limit Compliance, and Lane precision-recall curves over semantic trajectory modes. Violation Rate, aggregated into composite Traffic Rule Score. c: Map and kinematic compliance.
C. Multi-Agent and Interaction Metrics
Beyond positional accuracy, practical evaluation requires Multi-agent simulation requires metrics capturing not only checking physical feasibility. Off-road Rate measures the individual agent quality but also scene-level consistency and fraction of predicted positions falling outside drivable areas. interaction realism [26]. Overlap Rate (OR) quantifies the frequency of predicted agent trajectories whose bounding boxes overlap with other agents a: Scene-level consistency. or environment, signaling physically implausible predictions. Scene Collision Rate measures collisions between any pair Kinematic Feasibility checks whether predicted trajectories of simulated agents (excluding ego), indicating whether joint satisfy velocity, acceleration, and curvature bounds consistent rollouts produce physically consistent traffic. For trajectory with vehicle dynamics. prediction, Joint minADE/FDE evaluates coordinated accuracy across multiple agents simultaneously rather than B. Closed-Loop Evaluation Metrics averaging independent scores. Closed-loop evaluation deploys models as interactive agents within simulation by measuring realized behavior over ex- b: Interaction realism. tended rollouts. Unlike open-loop metrics comparing against The Waymo Open Sim Agents Challenge (WOSAC) [26] fixed recordings, closed-loop metrics assess how agents introduced metrics specifically targeting behavioral realism respond to evolving scenarios and interventions [25], [26]. for simulation agents. The Realism Meta-Metric combines multiple distributional comparisons between simulated and a: Safety metrics. real traffic, computed via negative log-likelihood of logged Collision Rate (CR) measures the fraction of episodes futures under densities estimated from model rollouts. The involving collisions: NLL objective WOSAC aims to approximate is: Nep |D| 1 X 1 X ∗ 1[collision in episode i] (30) CR = NLL = − log qworld (o≥1,i | o<1,i ) (32) Nep i=1 |D| i=1 where Nep is number of episodes. To avoid scoring the full high-dimensional future o≥1,i , Post-Encroachment Time (PET) measures the time gap WOSAC parameterizes scenarios using a set of component between consecutive occupancy of the same conflict region measurements and computes a time-series likelihood for each by different agents, and is commonly used to quantify component metric m as an average (in log-space) over time, intersection conflicts: masked by validity vt : exit enter exit P PETij (R) = tenter (R) − t (R) if t (R) ≥ t (R) j i j i 1{vt } NLLt tP m = exp − (33) (31) t 1{vt } where R denotes a conflict region, texit (R) is the time agent i i exits R, and tenter (R) is the time agent j enters R (with i After obtaining component metrics for each measurement, j WOSAC aggregates them into a single composite metric being the first agent to traverse R and j the second). MK: N M M b: Progress and task completion. X 1 XX K K M = w m , wj = 1 (34) j i,j Route Completion Rate measures the fraction of intended N M i=1 j=1 j=1 route successfully traversed. Goal Achievement Rate indicates whether agents reach designated destinations within time where N is the number of scenarios, M = 9 is the number of limits. The nuPlan benchmark [25], [224] introduces a component metrics, and K = 32 is the number of stochastic Closed-Loop Score (CLS) as a weighted aggregate combining rollouts. The metric aggregates three categories: Kinematic progress (fraction of route completed), safety-related viola- Metrics comparing distributions of linear speed, linear acceltions (collision, off-road, wrong direction), comfort metrics eration magnitude, angular speed, and angular acceleration (acceleration, jerk), speed-limit adherence, and minimum magnitude; Interactive Metrics comparing TTC distributions, TTC, with multiplier penalties that zero the score for any distance to nearest object, and collision likelihood; and collision, off-road departure, or failure to make progress. Map-Based Metrics comparing distance to road edge, road 32
VOLUME ,
departure frequency, and off-road rate. Collisions and road departures are typically double-weighted to emphasize safety.
diverse urban driving with full sensor suite annotations. For closed-loop planning evaluation, nuPlan [25], [224] represents a paradigm shift by offering 1,282 hours of data with scenario taxonomies, reactive simulation capabilities, and planningc: Interaction-specific evaluation. For targeted interaction assessment, metrics include Gap specific metrics across four cities with distinct driving cultures. Acceptance Rate at merging scenarios, Yielding Compliance These datasets collectively enable the training and rigorous at unprotected turns, Time Headway (THW) distributions for evaluation of the behavior modeling methods reviewed in this car-following behavior, and Lane Change Success Rate. The survey, with dataset selection depending on the target task INTERACTION dataset [225] and highD dataset [226] em- (prediction vs. planning vs. simulation), required scenario phasize such scenario-specific metrics, with INTERACTION diversity, and evaluation paradigm (open-loop vs. closedcovering diverse international driving contexts (roundabouts, loop). intersections, merging) and highD providing naturalistic highway trajectories from German highways for calibration B. Simulation Tools and validation. Table 7 summarizes widely used simulation tools for developing and evaluating learning-based driving methods. Tool selection depends on the target application, required fidelity X. SIMULATION TOOLS AND DATASETS This section reviews the simulation platforms, datasets, and level, dataset compatibility, and computational constraints. benchmarks used to train and evaluate learning-based traffic High-throughput state-level simulators (GPUDrive, Waymax, Nocturne) are most suitable for methods requiring billions of behavior models in mixed autonomy. environment steps, typical in reinforcement learning, self-play, and large-scale ablations. Sensor-level simulators (CARLA, A. Datasets for Behavior Modeling AWSIM, Isaac Sim) are preferred when contributions depend Large-scale naturalistic driving datasets have been essential on perception realism, sensor modeling, or full autonomy in advancing AI-based behavior modeling for automated stack integration. Traffic-level simulators (SUMO, CityFlow) vehicles. These datasets provide the trajectory data, sensor support network-wide analysis and mixed autonomy penerecordings, and contextual information necessary for training tration studies. Neural sensor simulators (UniSim, VISTA) and evaluating prediction, planning, and simulation models. generate realistic sensor outputs from real driving logs, which Table 6 summarizes the major publicly available datasets orallows closed-loop evaluation under counterfactual scenarios. ganized by their primary characteristics: geographic coverage, recording duration, agent diversity, and key features. Early trajectory datasets such as NGSIM [227] established foundational benchmarks for car-following and lane-changing analysis. However, they were limited in scale and scenario diversity. The drone-based recording methodology pioneered by the highD dataset [226] and its successors (inD [228], rounD [229], exiD [230]) enabled high-precision trajectory extraction. This set of datasets provide naturalistic behavior data for highways, intersections, roundabouts, and merging scenarios across German road networks. The INTERACTION dataset [225] extended this approach internationally and covers diverse driving cultures and highly interactive scenarios, including adversarial and cooperative maneuvers with semantic map annotations. Finally, pNEUMA [231] provides a large-scale drone-based trajectory dataset recorded over the congested central business district of Athens using a coordinated swarm of 10 drones over five days. It contains high-frequency trajectories across multiple traffic modes (cars, buses, taxis, powered two-wheelers, and heavy vehicles). Modern large-scale datasets have shifted toward multimodal sensor fusion and standardized benchmarking. The Waymo Open Motion Dataset (WOMD) [23] provides over 570 hours of driving data across six U.S. cities with high-fidelity trajectory annotations, HD maps, and curated interactive splits emphasizing multi-agent scenarios. Argoverse 2 [24] offers 250,000 motion forecasting scenarios with 3D lane boundaries and ground height information, while nuScenes [232] provides VOLUME ,
XI. DISCUSSION
The methods reviewed in this survey cover a wide range. While each section has identified challenges specific to individual method families, a broader view of the field reveals several cross-cutting themes that can guide future efforts in the domain of AI-driven mixed autonomy traffic simulation. This section synthesizes these themes into nine actionable challenges and future directions that go beyond individual methods. A. The Causality Gap: From Correlation to Counterfactual Validity
Perhaps the most pervasive challenge across all methods surveyed is the reliance on observational training data to build models that must function under interventional conditions. Imitation learning, trajectory prediction, multiagent imitation, and even generative world models all learn from passively recorded driving logs, while the primary use case for simulation demands counterfactual reasoning: what would surrounding agents do if the ego vehicle behaved differently from what was recorded? This gap between observational training and interventional deployment is not merely a technical limitation of any single method family; it is a fundamental constraint that limits causal validity under policy interventions. Current approaches address this indirectly through data augmentation, adversarial perturbation, 33
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
TABLE 6: Major publicly available datasets for trajectory prediction, behavior modeling, and planning in automated driving. Duration indicates total recording time; scenarios indicates number of extracted segments for benchmarking where applicable. Dataset
Location(s)
Year
Duration / Scenarios
Freq. (Hz)
Environment
Agent Types
Key Features
Foundational; car-following analysis High precision (<10 cm); 5,600 lane changes Naturalistic intersection behavior Complex interaction patterns
Drone-Based Trajectory Datasets NGSIM [227]
CA, GA (USA)
2005–06
90 min
10
Highway, arterial
Vehicles
highD [226]
Germany
2018
25
Highway (6 sites)
Cars, trucks
inD [228]
Germany
2020
25
rounD [229]
Germany
2020
pNEUMA [231]
Athens (Greece)
2018
16.5 h / 110k vehicles 10 h / 11.5k road users 6 h / 13k road users 12.5 h / 0.5M trajectories
Urban intersections (4 sites) Roundabouts (3 sites) Urban CBD (congested)
exiD [230]
Germany
2022
16 h / 69k vehicles 25
Vehicles, cyclists, pedestrians Vehicles, cyclists, pedestrians Cars, taxis, buses, PTWs, medium/heavy vehicles Vehicles
2019
16.5 h / 40k+ trajectories
INTERACTION [225] USA, China, Germany
25 25
10
Highway ramps (7 sites) Mixed (roundabouts, Vehicles intersections, highways)
Swarm of 10 drones over 5 days; dense multimodal trajectories over large urban network Merging and lane change scenarios International; adversarial & cooperative scenarios; semantic maps
Vehicle-Mounted Sensor Datasets (Motion Forecasting) nuScenes [232]
Boston, Singapore
2020
5.5 h / 1,000 2 (keyframes) Urban scenes 320 h / 324k 10 Urban scenarios — / 250k scenarios 10 Urban
Argoverse 1 [233]
Miami, Pittsburgh
2019
Argoverse 2 [24]
6 U.S. cities
2023
Waymo Open Motion [23]
6 U.S. cities
2021
570+ h / 103k segments
10
Urban, suburban
Lyft Level 5 [234]
Palo Alto (USA)
2021
1,118 h / 170k scenes
10
Urban
1,282 h / 75 scenario types
20
Urban (4 cities)
— / 5,700 scenarios
varies
Highway, urban
Vehicles, Full sensor suite; 23 classes; pedestrians, cyclists 3D annotations Vehicles First large-scale; HD maps with centerlines 30 classes 3D lane boundaries; ground height; 6s prediction Vehicles, Interactive split; realism pedestrians, cyclists meta-metric; WOSAC benchmark Vehicles, Semantic maps; large scale pedestrians, cyclists
Planning and Closed-Loop Evaluation Datasets nuPlan [25], [224]
CommonRoad [235]
Boston, Pittsburgh, 2021–24 Las Vegas, Singapore Synthetic + 2017 recorded
Closed-loop benchmark; Vehicles, pedestrians, cyclists reactive agents; planning metrics Vehicles Motion planning benchmark; formal specifications
or self-play, but none provide principled guarantees of produce unsafe or unrealistic behavior when deployed as counterfactual validity. Bridging this gap will likely require reactive simulation agents. Closed-loop benchmarks such as several measures, such as large-scale collection of interven- WOSAC and nuPlan represent important progress, but they tional driving data (e.g., through controlled AV deployments introduce their own complications: sensitivity to evaluation that deliberately probe human responses), development of infrastructure assumptions, dependence on how ego behavior causal inference frameworks adapted to multi-agent sequential is specified, lack of standardized intervention protocols, and decision-making, and hybrid architectures that combine data- in the case of WOSAC, reliance on available fine-grained driven pattern matching with structured causal models that datasets. More fundamentally, the field lacks consensus on encode domain knowledge about driver response mechanisms. what “realistic simulation” means operationally. Is the goal The intersection of causal machine learning and traffic to match distributional statistics of real traffic (a statistical simulation remains underdeveloped, but it is a high-impact fidelity criterion), to produce plausible reactions to arbitrary research direction. ego interventions (a counterfactual validity criterion), or to reproduce the logged data or generative cognitive mechanisms behind human driving? These are distinct objectives that may B. The Evaluation Crisis A recurring finding throughout this survey is that evaluation require different models, training paradigms, and evaluation methodology lags behind modeling methodology. Open- metrics. Developing a unified evaluation framework that loop displacement metrics (minADE, minFDE) remain the addresses behavioral realism, interaction quality, counterdominant evaluation paradigm despite mounting evidence factual robustness, and long-horizon stability simultaneously, that they correlate poorly with closed-loop simulation quality. while remaining computationally tractable and reproducible Models that achieve state-of-the-art prediction accuracy can 34
VOLUME ,
TABLE 7: Simulation tools commonly used with learning-based driving methods. Simulation level indicates output granularity: sensor (camera/LiDAR images), state (bounding boxes, trajectories), or traffic (aggregate flow). Background agents refers to how non-ego vehicles are modeled during simulation. Tool
Primary Application
Sim Level
Supported Datasets
Background Agents
Throughput
Very high (GPU-native, 1M+ steps/s) High (JAX/GPU/TPU batching) Moderate (multi-process)
State-Level / Data-Driven Simulators GPUDrive
Multi-agent RL training
State
WOMD
Learned policies, rule-based
Waymax
State
WOMD
Log replay, IDM, learned
State
nuPlan
Nocturne
Motion planning, prediction evaluation Planning benchmark and evaluation Multi-agent RL research
State
WOMD-derived scenarios
highway-env
RL algorithm prototyping
State
Synthetic (procedural)
IDM, log replay, learned (SMART) Log replay, learned (BC baseline) IDM, MOBIL
MetaDrive
RL training, generalization
State
SMARTS
Multi-agent RL, social driving State
Procedural, WOMD, nuScenes, nuPlan, Lyft NGSIM, Argoverse, WOMD, SUMO
Social agents, rule-based, log replay
Moderate (multi-instance)
OpenStreetMap, custom SUMO networks Custom networks
Car-following (Krauss, IDM) SUMO models Rule-based
High (large-scale networks) Moderate (RL overhead) High (city-scale)
Autopilot (rule-based), traffic manager Rule-based Scripted behaviors Scripted, AI traffic
Moderate (rendering-limited)
Log-based with viewpoint synthesis Neural reconstruction, actor manipulation
Dataset-dependent
nuPlan Devkit
High (C++ core, >2k steps/s)
High (vectorized environments) Rule-based (IDM), log replay High (1k+ FPS)
Traffic-Level Simulators SUMO Flow CityFlow
Traffic flow, mixed autonomy Traffic RL for traffic control Traffic Traffic signal optimization Traffic
Sensor-Level / 3D Simulators CARLA
End-to-end driving, perception Sensor
Custom, OpenDRIVE maps
AWSIM Isaac Sim BeamNG.tech
Autoware stack testing Perception, robotics Vehicle dynamics, crash testing
Sensor Sensor Sensor
Lanelet2 maps Custom environments Custom scenarios
Moderate Moderate (GPU rendering) Moderate
Neural / Data-Driven Sensor Simulators VISTA 2.0
Data-driven policy learning
Sensor
Custom logs (MIT AVT)
UniSim
Closed-loop counterfactual testing
Sensor
PandaSet, custom logs
State State
OpenDRIVE/OpenSCENARIO Scripted trajectories Recorded + synthetic Replay, interactive
Moderate (GPU inference)
Scenario Engines and Benchmarks esmini CommonRoad
OpenSCENARIO execution Motion planning benchmark
across research communities, is one of the most important infrastructure challenges facing the field. C. The Scalability, Fidelity, and Controllability Trilemma
Lightweight Offline evaluation
observed human driving. To the best of our knowledge, no current method satisfactorily resolves all three requirements simultaneously. Future architectures may need to adopt hierarchical designs that allocate computational resources adaptively. They can use expensive high-fidelity reasoning selectively for complex interactions and use lightweight policies for routine driving. They also need calibrated uncertainty estimates that indicate when controlled behaviors depart from training data.
Across method families, a persistent three-way tension emerges between scalability, behavioral fidelity, and controllability. High-fidelity multi-agent simulators like diffusionbased reactive simulators produce realistic joint rollouts but face computational costs that limit their applicability to large-scale testing campaigns. Scalable approaches such as parameter-shared policies or autoregressive token-based D. The Mixed Autonomy Representation Problem simulators achieve throughput suitable for reinforcement While “mixed autonomy” is the motivating context for this learning and large-scale ablations, but may sacrifice behav- entire survey, surprisingly few methods explicitly model ioral diversity or interaction quality. Controllable methods that the heterogeneous behavioral dynamics that arise when allow test engineers to specify desired behavioral profiles AVs and human drivers share road infrastructure. Most (e.g., aggressive merging, hesitant yielding) risk pushing current approaches either model all agents as human-like agents beyond the support of training data into “sim-native” (ignoring the distinct behavioral signatures of AVs) or focus behaviors that are internally consistent but not grounded in exclusively on ego-vehicle planning (ignoring how human VOLUME ,
35
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
drivers adapt to AV presence). The critical missing ingredient nominal driving conditions, with safety-critical events (nearis the co-adaptation between human drivers and AVs: as misses, erratic behaviors, AV-human conflicts) dramatically AV penetration rates increase, human drivers encounter underrepresented relative to their importance for simulationAVs more frequently and may adjust their gap acceptance, based AV testing. Addressing these limitations requires following distances, and interaction strategies. Trust dynamics, concerted data collection efforts across diverse geographies as reviewed in Section VII, offer one formalization of and interaction contexts, as well as synthetic data generation this adaptation, but integration into scalable multi-agent techniques that can augment rare-event coverage without simulation remains an open research direction. Constructing sacrificing distributional validity. Moreover, longitudinal simulation environments that capture the evolving dynamics collection of AV datasets is essential for studying and of the transition period, rather than assuming static behav- modeling behavioral adaptation in mixed autonomy traffic. ioral distributions, is essential for realistic testing of AV deployment strategies. This requires not only new modeling G. Toward Unified Architectures approaches but also new datasets that capture longitudinal The chronological analysis presented in Section VIII reveals human-AV interaction patterns across varying AV penetration accelerating methodological convergence: transformers, difrates, which currently do not exist at scale. fusion processes, and autoregressive generation now serve as shared building blocks across prediction, simulation, world E. Bridging the Research and Deployment Gap modeling, and scenario generation. This convergence suggests Currently, a notable disconnect exists between the methods that the traditional boundaries between single-agent and multideveloped in the research community and what simulation agent methods, between agent-level and environment-level platforms actually deploy. Mainstream simulation tools simulation, and between behavior modeling and scenario (CARLA, SUMO, VISSIM) still predominantly rely on generation are becoming increasingly artificial. The logical rule-based car-following and lane-changing models or log endpoint of this trajectory is unified simulation architectures replay for background traffic, despite the rich landscape of that jointly generate environment evolution, multi-agent learned behavior models reviewed in this survey. This gap behavior, and scenario conditions within a single learned persists for practical reasons: learned models require careful framework. This can combine the strengths of world models integration with map representations, traffic rule enforcement, (environment dynamics), reactive simulators (interaction fiand scenario specification workflows that research prototypes delity), and scenario generators (controllability) into cohesive rarely address. They also lack the interpretability and pre- systems. Early steps in this direction are visible in recent dictability that safety engineers require for systematic test works that combine diffusion-based scene generation with design. Addressing these issues demands not only continued reactive rollout capabilities. However, the risk of monolithic improvement in model accuracy and efficiency but also architectures is reduced interpretability and reduced ease attention to engineering concerns, such as standardized APIs of debugging. A more promising path may be modular for behavior model integration, modular architectures that but differentiable pipelines where specialized components allow mixing learned and rule-based agents within the same (perception, prediction, interaction reasoning, and physical simulation, and validation frameworks that quantify when a dynamics) are connected but remain individually interpretable learned model can be trusted as a substitute for a calibrated and replaceable. rule-based model. The emergence of data-driven simulation platforms (Waymax, GPUDrive, Nocturne) [2], [108], [236] H. Simulation-to-Real Transfer Gap designed specifically for learned behavior models represents a Despite the increasing realism of AI-driven traffic simulators, promising trend, but these remain largely confined to research a persistent challenge remains the gap between simulation settings. performance and real-world behavior. Many behavior models are trained and evaluated in simulated environments whose dynamics, interaction patterns, and long-tail events only partially The datasets underpinning the methods reviewed in this reflect real traffic conditions. As a result, policies that perform survey exhibit notable geographic and cultural concentration. well in simulation may fail to generalize when exposed to The dominant benchmarks, such as Waymo Open Motion the variability, uncertainty, and rare edge cases encountered Dataset, Argoverse, LevelX Dataset, nuScenes, and nuPlan, in real deployments. This issue is particularly pronounced in are collected primarily in a handful of U.S. and European mixed autonomy traffic, where subtle differences in human cities, and Singapore. Driving behavior, however, is deeply driving behavior, local traffic culture, and infrastructure influenced by local traffic culture, road infrastructure, regula- can significantly affect interaction dynamics. Bridging this tory environments, and social norms. Models trained on U.S. simulation-to-real gap requires improved environment fidelity, highway data may produce implausible behavior in the dense, better calibration against real-world data, and evaluation mixed-mode traffic of South and Southeast Asian cities, or in frameworks that explicitly measure transferability between the roundabout-heavy networks of European towns. Beyond simulated and real-world settings. Hybrid approaches combingeographic bias, existing datasets overwhelmingly capture ing data-driven learning with physics-based constraints and F. Data Limitations and Geographic Bias
36
VOLUME ,
real-world validation pipelines represent promising directions for improving the reliability of simulation-trained models. I. The Role of Cognitive Grounding
Finally, the cognitive and physics-informed methods reviewed in this survey highlight a fundamental research direction. The dominant research trend is toward larger data-driven models that match behavioral distributions with increasing fidelity. However, for mixed autonomy simulation specifically, understanding why drivers behave as they do, not just reproducing what they do, carries distinct practical value. Cognitively grounded models offer three capabilities that purely data-driven approaches currently lack: (1) principled generation of human error modes critical for safety testing, (2) interpretable behavioral diversity parameterized by psychologically meaningful variables rather than obscured latent information, (3) and the ability to generalize to novel interaction patterns (such as those arising from AV introduction) that lie outside training data distributions. The integration challenge is also worth noting. Cognitive architectures operate at different abstraction levels and computational costs than modern deep learning. Nevertheless, hybrid approaches that encode cognitive constraints as inductive biases within neural architectures, rather than building full cognitive simulations, represent a pragmatic and promising approach that could combine the scalability of data-driven methods with the explanatory power of cognitive science. XII. CONCLUSION
This survey provides a comprehensive and structured review of artificial intelligence methods for modeling mixed automated and human traffic in simulation environments, organized along three complementary axes: agent-level behavior models (single-agent and multi-agent), environment-level simulation methods (world models and scenario generation), and cognitive and physics-informed approaches. Across these methodological families, several cross-cutting challenges were identified: the gap between open-loop trajectory accuracy and closed-loop interactive realism remains fundamental, as matching recorded driving logs does not guarantee plausible behavior under counterfactual interventions; the tension between data-driven fidelity and user-specified controllability limits the practical utility of generative approaches for systematic testing; long-tail safety-critical scenarios remain underrepresented in both training data and evaluation benchmarks; and standardized validation frameworks that can establish trust in learned simulators for safety assurance are still lacking. Looking ahead, the field is converging toward hybrid architectures that combine the scalability of data-driven learning with the interpretability of cognitive models and the guarantees of physics-based constraints, while foundation models and diffusion-based generation are opening new possibilities for unified perception–prediction–planning frameworks. We argue that bridging the transportation engineering and machine learning communities will be essential VOLUME ,
for realizing the full potential of AI-driven simulation in enabling the safe deployment of automated vehicles in mixed autonomy traffic. This could be achieved by jointly addressing behavioral realism, counterfactual validity, domain generalization, and integration with safety assurance pipelines. REFERENCES [1] Haneen Farah, Irene Postigo, Bahman Madadi, Matin Moarefianpour, Gerben van Beukel, Bart van Arem, and Johan Olstam. Modeling automated driving in microscopic traffic simulations for traffic performance evaluations: Aspects to consider and state of the practice. IEEE Open Journal of Intelligent Transportation Systems, 3:536–546, 2022. [2] Saman Kazemkhani, Aarav Pandya, Daphne Cornelisse, Brennan Shacklett, and Eugene Vinitsky. Gpudrive: Data-driven, multi-agent driving simulation at 1 million fps. In International Conference on Learning Representations (ICLR), 2025. Poster. [3] Pablo Álvarez López, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flötteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 2575–2582. IEEE, 2018. [4] PTV Group. Ptv vissim – multimodal traffic simulation software. Available at: https://www.ptvgroup.com/en-us/products/ptv-vissim. [5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1–16. PMLR, 2017. [6] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 62(2):1805, 2000. [7] Arne Kesting, Martin Treiber, and Dirk Helbing. General lanechanging model mobil for car-following models. Transportation Research Record, 1999(1):86–94, 2007. [8] Wilko Schwarting, Javier Alonso-Mora, and Daniela Rus. Social behavior for autonomous vehicles. Proceedings of the National Academy of Sciences, 116(50):24972–24978, 2019. [9] Herbert A. Simon. A behavioral model of rational choice. The Quarterly Journal of Economics, 69(1):99–118, 1955. [10] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. [11] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. Dilu: A knowledgedriven approach to autonomous driving with large language models. In International Conference on Learning Representations (ICLR), 2024. [12] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17113– 17122, 2022. [13] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, and Dragomir Anguelov. Motiondiffuser: Controllable multiagent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644–9653, 2023. [14] Dean A Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems, volume 1, 1989. [15] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. [16] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693–4700. IEEE, 2018. [17] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. Safe, multi-agent reinforcement learning for autonomous driving. arXiv preprint arXiv:1610.03295, 2016.
37
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
[18] Alex Kendall, Jeffrey Hawke, David Janz, Przemyslaw Mazur, Daniele Reda, John-Mark Allen, Vinh-Dieu Lam, Alex Bewley, and Amar Shah. Learning to drive in a day. In 2019 International Conference on Robotics and Automation (ICRA), pages 8248–8254. IEEE, 2019. [19] Lloyd Russell, Anthony Hu, Lorenzo Bertoni, George Fedoseev, Jamie Shotton, Elahe Arani, and Gianluca Corrado. Gaia-2: A controllable multi-view generative world model for autonomous driving. arXiv preprint arXiv:2503.20523, 2025. Technical report. [20] Yueyuan Li, Wei Yuan, Songan Zhang, Weihao Yan, Qiyuan Shen, Chunxiang Wang, and Ming Yang. Choose your simulator wisely: A review on open-source simulators for autonomous driving. IEEE Transactions on Intelligent Vehicles, 9(5):4861–4876, 2024. [21] Xuan Di and Rongye Shi. A survey on autonomous vehicle control in the era of mixed-autonomy: From physics-based to AI-guided driving policy learning. Transportation Research Part C: Emerging Technologies, 125:103008, 2021. [22] Wilko Schwarting, Javier Alonso-Mora, and Daniela Rus. Planning and decision-making for autonomous vehicles. Annual Review of Control, Robotics, and Autonomous Systems, 1(1):187–210, 2018. [23] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurélien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large scale interactive motion forecasting for autonomous driving: The Waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9710–9719, 2021. [24] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493, 2023. [25] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint, 2021. arXiv:2106.11810. [26] Nico Montali, John Lambert, Paul Mougin, Alex Kuefler, Nick Rhinehart, Michelle Li, Cole Gulino, Tristan Emrich, Zoey Yang, Shimon Whiteson, Brandyn White, and Dragomir Anguelov. The waymo open sim agents challenge. Advances in Neural Information Processing Systems (NeurIPS), Track on Datasets and Benchmarks, 2023. arXiv:2305.12032. [27] Sampo Kuutti, Richard Bowden, Yaochu Jin, Phil Barber, and Saber Fallah. A survey of deep learning applications to autonomous vehicle control. IEEE Transactions on Intelligent Transportation Systems, 22(2):712–733, 2021. [28] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3):362–386, 2020. [29] B. Ravi Kiran, Ibrahim Sobh, Victor Talpaert, Patrick Mannion, Ahmad A. Al Sallab, Senthil Yogamani, and Patrick Pérez. Deep reinforcement learning for autonomous driving: A survey. IEEE Transactions on Intelligent Transportation Systems, 23(6):4909–4926, 2022. [30] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10164–10183, 2024. [31] Qianwen Chao, Huikun Bi, Weizi Li, Tianlu Mao, Zhaoqi Wang, Ming C. Lin, and Zhigang Deng. A survey on visual traffic simulation: Models, evaluations, and applications in autonomous driving. Computer Graphics Forum, 39(1):287–308, 2020. [32] Di Chen, Meixin Zhu, Hai Yang, Xuesong Wang, and Ye Wang. Datadriven traffic simulation: A comprehensive review. IEEE Transactions on Intelligent Vehicles, 9(4):4730–4748, 2024. [33] Wenhao Ding, Chejian Xu, Mansur Arief, Haohong Lin, Bo Li, and Ding Zhao. A survey on safety-critical driving scenario generation— a methodological perspective. IEEE Transactions on Intelligent Transportation Systems, 24(7):6971–6988, 2023. [34] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. ChauffeurNet: Learning to drive by imitating the best and synthesizing the worst. In Robotics: Science and Systems, 2019. [35] Stéphane Ross, Geoffrey J. Gordon, and J. Andrew Bagnell. A reduction of imitation learning and structured prediction to no-regret online
38
learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS), volume 15 of Proceedings of Machine Learning Research, pages 627–635, 2011. [36] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7077– 7087, 2021. [37] Oliver Scheel, Luca Bergamini, Maciej Wolczyk, Błażej Osiński, and Peter Ondruska. Urban driver: Learning to drive from real-world demonstrations using policy gradients. In Proceedings of the 5th Conference on Robot Learning (CoRL), volume 164 of Proceedings of Machine Learning Research, pages 718–728. PMLR, 2022. [38] Tung Phan-Minh, Elena Corina Grigore, Freddy A Boulton, Oscar Beijbom, and Eric M Wolff. Covernet: Multimodal behavior prediction using trajectory sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14074– 14083, 2020. [39] Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypotheses for behavior prediction. In Conference on Robot Learning (CoRL), pages 86–99, 2020. [40] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. Tnt: Target-driven trajectory prediction. In Conference on Robot Learning (CoRL), pages 895–904, 2021. [41] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15283– 15292, 2021. [42] ByeoungDo Kim, Seong Hyeon Park, Seokhwan Lee, Elbek Khoshimjonov, Dongsuk Kum, Junsoo Kim, Jeong Soo Kim, and Jun Won Choi. Lapred: Lane-aware prediction of multi-modal future trajectories of dynamic agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14636–14645, 2021. [43] Haoran Song, Di Luan, Wenchao Ding, Michael Y Wang, and Qifeng Chen. Learning to predict vehicle trajectories with model-based planning. In Conference on Robot Learning, pages 1035–1045. PMLR, 2022. [44] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15222–15232, 2021. [45] Eli Bronstein, Supratik Srinivasan, Sushant Paul, Aman Sinha, Matthew O’Kelly, Payam Nikdel, and Shimon Whiteson. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5711– 5718. IEEE, 2022. [46] Qifeng Li, Xiaosong Jia, Li Chen, Junchi Yan, Ping Luo Liu, and Hongyang Li. Think2drive: Efficient reinforcement learning by thinking in latent world model for quasi-realistic autonomous driving (in carla-v2). In European Conference on Computer Vision, pages 41–57. Springer, 2024. [47] Zilin Huang, Zihao Sheng, Chengyuan Ma, and Sikai Chen. Human as AI mentor: Enhanced human-in-the-loop reinforcement learning for safe and efficient autonomous driving. Communications in Transportation Research, 4:100127, 2024. [48] Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse reinforcement learning. In AAAI Conference on Artificial Intelligence, volume 8, pages 1433–1438, 2008. [49] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. [50] Alex Kuefler, Jeremy Morton, Timothy Wheeler, and Mykel Kochenderfer. Imitating driver behavior with generative adversarial networks. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 204–211. IEEE, 2017. [51] Raunak P Bhattacharyya, Derek J Phillips, Blake Wulfe, Jeremy Morton, Alex Kuefler, and Mykel J Kochenderfer. Multi-agent imitation learning for driving simulation. In 2018 IEEE/RSJ International
VOLUME ,
Conference on Intelligent Robots and Systems (IROS), pages 1534– 1539. IEEE, 2018. [52] Justin Fu, Katie Luo, and Sergey Levine. Learning robust rewards with adverserial inverse reinforcement learning. In International Conference on Learning Representations, 2018. [53] Raunak P Bhattacharyya, Derek J Phillips, Changliu Liu, Jayesh K Gupta, Katherine Driggs-Campbell, and Mykel J Kochenderfer. Simulating emergent properties of human driving behavior using multiagent reward augmented imitation learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 789–795. IEEE, 2019. [54] Wenyuan Zeng, Wenjie Luo, Simon Suo, Abbas Sadat, Bin Yang, Sergio Casas, and Raquel Urtasun. End-to-end interpretable neural motion planner. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8660–8669, 2019. [55] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planningoriented autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17853– 17862, 2023. [56] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized scene representation for efficient autonomous driving. In IEEE/CVF International Conference on Computer Vision, pages 8340–8350, 2023. [57] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Haoran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 8795–8801. IEEE, 2025. [58] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726–737, 2023. [59] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21983–21994, 2023. [60] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. NAVSIM: Datadriven non-reactive autonomous vehicle simulation and benchmarking. In Advances in Neural Information Processing Systems (NeurIPS), 2024. [61] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In Advances in Neural Information Processing Systems 31, pages 2451–2463. Curran Associates, Inc., 2018. [62] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. In International Conference on Learning Representations (ICLR), 2020. [63] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy P. Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. [64] Anthony Hu, Gianluca Corber, Matthias Muller, Dragomir Anguelov, Sergio Casas, Raquel Urtasun, and Alex Kendall. Model-based imitation learning for urban driving. In Advances in Neural Information Processing Systems, volume 35, pages 20703–20716, 2022. [65] Hang Wang, Xin Ye, Feng Tao, Chenbin Pan, Abhirup Mallik, Burhaneddin Yaman, Liu Ren, and Junshan Zhang. Adawm: Adaptive world model based planning for autonomous driving. In International Conference on Learning Representations (ICLR), 2025. [66] Anant Garg and K Madhava Krishna. Imagine-2-drive: Leveraging high-fidelity world models via multi-modal diffusion policies. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4188–4195. IEEE, 2025. [67] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. Languagempc: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. [68] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. GPT-Driver: Learning to drive with GPT. In Advances in Neural Information Processing Systems (NeurIPS), 2023. [69] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, KwanYee K. Wong, Zhenguo Li, and Hengshuang Zhao. DriveGPT4:
VOLUME ,
Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. [70] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krähenbühl. Learning by cheating. In Conference on Robot Learning, pages 66–75, 2020. [71] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Proceedings of the 1st Conference on Robot Learning (CoRL), volume 78 of Proceedings of Machine Learning Research, pages 143–156, 2017. [72] Michael Kelly, Chelsea R. Sidrane, Katherine R. Driggs-Campbell, and Mykel J. Kochenderfer. HG-DAgger: Interactive imitation learning with human experts. In 2019 International Conference on Robotics and Automation (ICRA), 2019. [73] Xiatao Sun, Shuo Yang, Mingyan Zhou, Kunpeng Liu, and Rahul Mangharam. Mega-dagger: Imitation learning with multiple imperfect experts. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 1–8. IEEE, 2024. [74] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In IEEE/CVF International Conference on Computer Vision, pages 9329–9338, 2019. [75] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7153–7162, 2020. [76] Long Wen, Junqiang Duan, Shengbo Li, Shaobing Xu, and Huei Peng. Safe reinforcement learning for autonomous vehicles through parallel constrained policy optimization. In IEEE International Conference on Intelligent Transportation Systems (ITSC), pages 1–7, 2020. [77] Zhiyu Huang, Jingda Wu, and Chen Lv. Driving behavior modeling using naturalistic human driving data with inverse reinforcement learning. IEEE Transactions on Intelligent Transportation Systems, 23(8):10239–10251, 2021. [78] Pin Wang, Dapeng Liu, Jiayu Chen, Hanhan Li, and Ching-Yao Chan. Decision making for autonomous driving via augmented adversarial inverse reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1036–1042. IEEE, 2021. [79] Moritz Sackmann, Henrik Bey, Ulrich Hofmann, and Jörn Thielecke. Modeling driver behavior using adversarial inverse reinforcement learning. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1683–1690. IEEE, 2022. [80] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformerbased sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. [81] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. ST-P3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533–549, 2022. [82] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In European Conference on Computer Vision (ECCV), pages 414–430. Springer, 2020. [83] Yinan Zheng, Ruiming Liang, Kexin Zheng, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Weihao Gu, Rui Ai, Shengbo Eben Li, Xianyuan Zhan, and Jingjing Liu. Diffusion-based planning for autonomous driving with flexible guidance. In The Thirteenth International Conference on Learning Representations, 2025. [84] Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, and Xinggang Wang. Diffusiondrive: Truncated diffusion model for endto-end autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12037–12047, 2025. [85] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. [86] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864–14873, 2024.
39
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
[87] Danijar Hafner, Timothy P. Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations (ICLR), 2021. [88] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. [89] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2255–2264, June 2018. [90] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 336–345, 2017. [91] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data. In European conference on computer vision, pages 683–700. Springer, 2020. [92] Nicholas Rhinehart, Rowan McAllister, Kris Kitani, and Sergey Levine. Precog: Prediction conditioned on goals in visual multi-agent settings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2821–2830, 2019. [93] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525–11533, 2020. [94] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In European Conference on Computer Vision (ECCV), pages 541–556, 2020. [95] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David Weiss, Ben Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting multiple agent trajectories. In International Conference on Learning Representations (ICLR), 2022. [96] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8813–8823, 2022. [97] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. MTR++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. [98] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Gametheoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3903–3913, 2023. arXiv:2303.05760. [99] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 3560–3566. IEEE, 2023. [100] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. In Proceedings of The 7th Conference on Robot Learning, volume 229 of Proceedings of Machine Learning Research, pages 144–177. PMLR, 2023. [101] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S. Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. [102] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. In International Conference on Learning Representations (ICLR), 2024. [103] Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue.
40
Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2024. [104] Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: Scalable multi-agent real-time simulation via next-token prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2024. [105] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Błażej Osiński, Hugo Grimmett, and Peter Ondruska. Simnet: Learning reactive self-driving simulations from real-world observations. In IEEE International Conference on Robotics and Automation (ICRA), pages 5119–5125, 2021. [106] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. Trafficsim: Learning to simulate realistic multi-agent behaviors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400–10409, 2021. [107] Cathy Wu, Abdul Rahman Kreidieh, Kanaad Parvate, Eugene Vinitsky, and Alexandre M Bayen. Flow: A modular learning framework for mixed autonomy traffic. IEEE Transactions on Robotics, 38(2):1270– 1286, 2021. [108] Eugene Vinitsky, Nathan Lichtlé, Xiaomeng Yang, Brandon Amos, and Jakob Foerster. Nocturne: A scalable driving benchmark for multiagent reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2022. [109] Daphne Cornelisse and Eugene Vinitsky. Human-compatible driving partners through data-regularized self-play reinforcement learning. arXiv preprint arXiv:2403.19648, 2024. [110] Marco Cusumano-Towner, David Hafner, Alex Hertzberg, Brody Huval, Aleksei Petrenko, Eugene Vinitsky, Erik Wijmans, Taylor W. Killian, Stuart Bowers, Ozan Sener, Philipp Krähenbühl, and Vladlen Koltun. Robust autonomy emerges from self-play. CoRR, abs/2502.03349, 2025. [111] Raunak Bhattacharyya, Blake Wulfe, Derek J. Phillips, Alex Kuefler, Jeremy Morton, Ransalu Senanayake, and Mykel J. Kochenderfer. Modeling human driving behavior through generative adversarial imitation learning. IEEE Transactions on Intelligent Transportation Systems, 24(3):2874–2887, 2023. [112] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In IEEE International Conference on Robotics and Automation (ICRA), 2022. [113] Ke Guo, Haochen Liu, Xiaojun Wu, and Chen Lv. DecompGAIL: Learning realistic traffic behaviors with decomposed multi-agent generative adversarial imitation learning. arXiv preprint arXiv:2510.06913, 2025. [114] Hideyuki Kita. A merging-giveway interaction model of cars in a merging section: A game theoretic analysis. Transportation Research Part A: Policy and Practice, 33(3–4):305–312, 1999. [115] Dorsa Sadigh, Anca D Dragan, S Shankar Sastry, and Sanjit A Seshia. Planning for autonomous cars that leverage effects on human actions. In Robotics: Science and Systems, volume 2, page 5, 2016. [116] David Fridovich-Keil, E. Ratner, L. Peters, A. D. Dragan, and C. J. Tomlin. Efficient iterative linear-quadratic approximations for nonlinear multi-player general-sum differential games. In IEEE International Conference on Robotics and Automation (ICRA), pages 1475–1481, 2020. [117] Dale O. Stahl and Paul W. Wilson. On players’ models of other players: Theory and experimental evidence. Games and Economic Behavior, 10(1):218–254, 1995. [118] Juin-Kuan Chong, Colin F Camerer, and Teck-Hua Ho. Cognitive hierarchy: A limited thinking theory in games. In Experimental Business Research: Marketing, Accounting and Cognitive Perspectives Volume III, pages 203–228. Springer, 2005. [119] Haochen Liu, Li Chen, Yu Qiao, Chen Lv, and Hongyang Li. Reasoning multi-agent behavioral topology for interactive autonomous driving. Advances in Neural Information Processing Systems, 37:92605–92637, 2024. [120] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In IEEE International Conference on Robotics and Automation, pages 1522–1529. IEEE, 2023.
VOLUME ,
[121] Zhejun Zhang, Christos Sakaridis, and Luc Van Gool. Trafficbots v1.5: Traffic simulation via conditional vaes and transformers with relative pose encoding. arXiv preprint arXiv:2406.10898, 2024. [122] Luke Rowe, Roger Girgis, Anthony Gosselin, Bruno Carrez, Florian Golemo, Felix Heide, Liam Paull, and Christopher Pal. Ctrl-sim: Reactive and controllable driving agents with offline reinforcement learning. Proceedings of Machine Learning Research, 270:3600–3621, 2024. [123] Qianwen Zhang, Yang Gao, Yixuan Zhang, Yun Guo, Dongpu Ding, Yunpeng Wang, Peng Sun, and Ding Zhao. TrajGen: Generating realistic and diverse trajectories with reactive and feasible agent behaviors for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(12):24474–24487, 2022. [124] Senkang Hu, Zhengru Fang, Yiqin Deng, Yun-Peng Mao, Yue Fang, and Sam Kwong. AgentsCoDriver: Large language model empowered collaborative driving with lifelong learning. arXiv preprint arXiv:2404.06345, 2024. [125] Kemou Jiang, Xuan Cai, Zhiyong Cui, Aoyong Li, Yilong Ren, Haiyang Yu, Hao Yang, Daocheng Fu, Licheng Wen, and Pinlong Cai. Koma: Knowledge-driven multi-agent framework for autonomous driving with large language models. IEEE Transactions on Intelligent Vehicles, 2024. [126] Miao Zhang, Zhenlong Fang, Tianyi Wang, Shuai Lu, Xueqian Wang, and Tianyu Shi. Ccma: A framework for cascading cooperative multiagent in autonomous driving merging using large language models. Expert Systems with Applications, 282:127717, 2025. [127] Boxuan Wang, Haonan Duan, Yanhao Feng, Xu Chen, Yongjie Fu, Zhaobin Mo, and Xuan Di. Can LLMs understand social norms in autonomous driving games? In 2024 IEEE International Automated Vehicle Validation Conference (IAVVC), pages 1–4. IEEE, 2024. [128] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1349–1358, 2019. [129] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M. Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9813–9823, October 2021. [130] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In IEEE International Conference on Robotics and Automation (ICRA), pages 2980–2987, 2023. [131] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863–17873, 2023. [132] Zikang Zhou, Zihao Wen, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Qcnext: A next-generation framework for joint multi-agent trajectory prediction. arXiv preprint arXiv:2306.10508, 2023. [133] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr-a: 1st place solution for 2022 waymo open dataset challenge–motion prediction. arXiv preprint arXiv:2209.10033, 2022. [134] Waymo. A public dataset of driving behaviors and traffic interactions from autonomous vehicles. Available at: https://waymo.com/open, 2021. [135] Daphne Cornelisse*, Spencer Cheng*, Pragnay Mandavilli, Julian Hunt, Kevin Joseph, Waël Doulazmi, Valentin Charraut, Aditya Gupta, Joseph Suarez, and Eugene Vinitsky. PufferDrive: A fast and friendly driving simulator for training and evaluating RL agents, 2025. [136] Eli Bronstein, Mark Palatucci, Dominik Notz, Brandyn White, Alex Kuefler, Yiren Lu, Supratik Paul, Payam Nikdel, Paul Mougin, Hongge Chen, Justin Fu, Austin Abrams, Punit Shah, Evan Racah, Benjamin Frenkel, Shimon Whiteson, and Dragomir Anguelov. Hierarchical model-based imitation learning for planning in autonomous driving. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8652–8659. IEEE, 2022. [137] Yaozu Wu, Dongyuan Li, Yankai Chen, Renhe Jiang, Henry Peng Zou, Wei-Chieh Huang, Yangning Li, Liancheng Fang, Zhen Wang, and Philip S. Yu. Multi-agent autonomous driving systems with large language models: A survey of recent advances. arXiv preprint arXiv:2502.16804, 2025.
VOLUME ,
[138] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. Technical report. [139] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14749– 14759, 2024. [140] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving. In European Conference on Computer Vision, pages 174–192, 2024. [141] Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. Drivedreamer-2: Llmenhanced world models for diverse driving video generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 10412–10420, 2025. [142] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. Advances in Neural Information Processing Systems, 37:91560–91596, 2024. [143] Lunjun Zhang, Yuwen Xiong, Ze Yang, Sergio Casas, Rui Hu, and Raquel Urtasun. Copilot4d: Learning unsupervised world models for autonomous driving via discrete diffusion. arXiv preprint arXiv:2311.01017, 2023. [144] Daniel Bogdoll, Yitian Yang, Tim Joseph, Melih Yazgan, and J Marius Zollner. Muvo: A multimodal generative world model for autonomous driving with geometric representations. In 2025 IEEE Intelligent Vehicles Symposium (IV), pages 2243–2250. IEEE, 2025. [145] Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. Drivingworld: Constructing world model for autonomous driving via video gpt. arXiv preprint arXiv:2412.19505, 2024. [146] Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu, Jia Zeng, Ping Luo, et al. GenAD: Generalized predictive model for autonomous driving. arXiv preprint arXiv:2403.09630, 2024. [147] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3D occupancy world model for autonomous driving. In European Conference on Computer Vision (ECCV), 2024. [148] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4D occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. [149] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. DOME: Taming diffusion model into high-fidelity controllable occupancy world model. arXiv preprint arXiv:2410.10429, 2024. [150] Haoran Xu, Peixi Peng, Guang Tan, Yiqian Chang, Yisen Zhao, and Yonghong Tian. Temporal triplane transformers as occupancy world models. arXiv preprint arXiv:2503.07338, 2025. [151] Zhimin Liao, Ping Wei, Ruijie Zhang, Shuaijia Chen, Haoxuan Wang, and Ziyang Ren. I2-world: Intra-inter tokenization for efficient dynamic 4d scene forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 25810–25819, 2025. [152] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 9327–9335, 2025. [153] Julong Wei, Shanshuai Yuan, Pengfei Li, Qingda Hu, Zhongxue Gan, and Wenchao Ding. OccLLaMA: An occupancy-language-action generative world model for autonomous driving. arXiv preprint arXiv:2409.03272, 2024. [154] Shuhan Tan, Kelvin Wong, Shenlong Wang, Sivabalan Manivasagam, Mengye Ren, and Raquel Urtasun. Scenegen: Learning to generate realistic traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 892–901, 2021.
41
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
[155] Lan Feng, Quanyi Li, Zhenghai Peng, Shuhan Tan, and Bolei Zhou. TrafficGen: Learning to generate diverse and realistic traffic scenarios. In IEEE International Conference on Robotics and Automation (ICRA), pages 3567–3575, 2023. [156] Shuo Sun, Zekai Gu, Tianchen Sun, Jiawei Sun, Chengran Yuan, Yuhang Han, Dongen Li, and Marcelo H Ang. Drivescenegen: Generating diverse and realistic driving scenarios from scratch. IEEE Robotics and Automation Letters, 9(8):7007–7014, 2024. [157] Ethan Pronovost, Meghana Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nicholas Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. [158] Shuhan Tan, Boris Ivanovic, Xinshuo Weng, Marco Pavone, and Philipp Krähenbühl. Language conditioned traffic generation. In Conference on Robot Learning (CoRL), 2023. PMLR, volume 229. [159] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via collaborative llm-agents. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15077–15087, 2024. [160] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. In European Conference on Computer Vision (ECCV), 2024. [161] Mark Koren, Saud Alsaif, Ritchie Lee, and Mykel J. Kochenderfer. Adaptive stress testing for autonomous vehicles. In IEEE Intelligent Vehicles Symposium (IV). IEEE, 2018. [162] Ritchie Lee, Ole J. Mengshoel, Anshu Saksena, Ryan W. Gardner, Daniel Genin, Joshua Silbermann, Michael Owen, and Mykel J. Kochenderfer. Adaptive stress testing: Finding likely failure events with reinforcement learning. Journal of Artificial Intelligence Research (JAIR), 69:1165–1201, 2020. [163] Ding Zhao, Henry Lam, Huei Peng, Shan Bao, David J LeBlanc, Kazutoshi Nobukawa, and Christopher S Pan. Accelerated evaluation of automated vehicles safety in lane-change scenarios based on importance sampling techniques. IEEE transactions on intelligent transportation systems, 18(3):595–607, 2016. [164] Davis Rempe, Jonah Philion, Leonidas J Guibas, Sanja Fidler, and Or Litany. Generating useful accident-prone driving scenarios via a learned traffic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17305– 17315, 2022. [165] Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, and Andreas Geiger. King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In European Conference on Computer Vision (ECCV). Springer, 2022. [166] Yuting Xie, Xianda Guo, Cong Wang, Kunhua Liu, and Long Chen. Advdiffuser: Generating adversarial safety-critical driving scenarios via guided diffusion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9983–9989. IEEE, 2024. [167] Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. Cat: Closed-loop adversarial training for safe end-to-end driving. In Conference on Robot Learning (CoRL), pages 2357–2371. PMLR, 2023. [168] Guanpeng Li, Yiran Li, Saurabh Jha, Timothy Tsai, Michael Sullivan, Siva Kumar Sastry Hari, Zbigniew Kalbarczyk, and Ravishankar Iyer. Av-fuzzer: Finding safety violations in autonomous driving systems. In IEEE International Symposium on Software Reliability Engineering (ISSRE). IEEE, 2020. [169] Haohong Lin, Xin Huang, Tung Phan, David Hayden, Huan Zhang, Ding Zhao, Siddhartha Srinivasa, Eric Wolff, and Hongge Chen. Causal composition diffusion model for closed-loop traffic generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27542–27552, 2025. [170] Peide Huang, Wenhao Ding, Benjamin Stoler, Jonathan Francis, Bingqing Chen, and Ding Zhao. Cadre: Controllable and diverse generation of safety-critical driving scenarios using real-world trajectories. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 5474–5481. IEEE, 2025. [171] Kunkun Hao, Lu Liu, Wen Cui, Jianxing Zhang, Songyang Yan, Yuxi Pan, and Zijiang Yang. Bridging data-driven and knowledge-driven approaches for safety-critical scenario generation in automated vehicle validation. arXiv preprint arXiv:2311.10937, 2023.
42
[172] Yang Liu, Jian Sun, and Shuo Chen. A multi-agent driving-simulation approach for characterizing hazardous vehicle interactions between AVs and manual vehicles. Accident Analysis & Prevention, 2024. [173] Junkai Xia, Chenxin Xu, Qingyao Xu, Yanfeng Wang, and Siheng Chen. Language-driven interactive traffic trajectory generation. Advances in Neural Information Processing Systems, 37:77831–77859, 2024. [174] Aizaz Sharif and Dusica Marijan. Adversarial deep reinforcement learning for improving the robustness of multi-agent autonomous driving policies. In 2022 29th Asia-Pacific Software Engineering Conference (APSEC), pages 61–70. IEEE, 2022. [175] Sampo Kuutti, Saber Fallah, and Richard Bowden. Training adversarial agents to exploit weaknesses in deep control policies. In IEEE International Conference on Robotics and Automation (ICRA), pages 108–114. IEEE, 2020. [176] Bing Zhu, Rui Tang, Jian Zhao, Peixing Zhang, Wenxu Li, Xinran Cao, and Siyuan Li. Critical scenarios adversarial generation method for intelligent vehicles testing based on hierarchical reinforcement architecture. Accident Analysis & Prevention, 215:108013, 2025. [177] Seulbae Kim, Major Liu, Junghwan Rhee, Yuseok Jeon, Yonghwi Kwon, and Chung Hwan Kim. Drivefuzz: Discovering autonomous driving bugs through driving quality-guided fuzzing. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2022. [178] Shenghao Lin, Fansong Chen, Laile Xi, Gaosheng Wang, Rongrong Xi, Yuyan Sun, and Hongsong Zhu. Tm-fuzzer: fuzzing autonomous driving systems through traffic management. Automated Software Engineering, 2024. [179] Kun Hao, Wei Cui, Yuxiang Luo, Lei Xie, Yuxuan Bai, Jian Yang, Songyang Yan, Yue Pan, and Zhiwei Yang. Adversarial safety-critical scenario generation using naturalistic human driving priors. IEEE Transactions on Intelligent Vehicles, 2024. [180] Saeed Rahmani, Sabine Rieder, Erwin de Gelder, Marcel Sonntag, Jorge Lorente Mallada, Sytze Kalisvaart, Vahid Hashemi, and Simeon C Calvert. A systematic review of edge case detection in automated driving: Methods, challenges and future directions. arXiv preprint arXiv:2410.08491, 2024. [181] Amos Tversky and Daniel Kahneman. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4):297–323, 1992. [182] Martin Schmidt-Daffy. Prospect balancing theory: Bounded rationality of drivers’ speed choice. Accident Analysis & Prevention, 63:49–64, 2014. [183] Liting Sun, Wei Zhan, Yeping Hu, and Masayoshi Tomizuka. Interpretable modelling of driving behaviors in interactive driving scenarios based on cumulative prospect theory. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 4329–4335. IEEE, 2019. [184] Ray Fuller. Driver control theory: From task difficulty homeostasis to risk allostasis. In Handbook of traffic psychology, pages 13–26. Elsevier, 2011. [185] Saeed Mohammadian, Zuduo Zheng, Mazharul Haque, and Ashish Bhaskar. Net-rat: Non-equilibrium traffic model based on risk allostasis theory. Transportation research part A: policy and practice, 174:103731, 2023. [186] Mohammad Tamim Kashifi. Modelling intra-driver behavioral adaptation through risk sensitivity and regime transitions: A taskdifficulty car-following model. arXiv preprint arXiv:2510.09886, 2025. [187] Roger Ratcliff. A theory of memory retrieval. Psychological Review, 85(2):59–108, 1978. [188] Gustav Markkula, Richard Romano, Ruth Madigan, Charles W Fox, Oscar T Giles, and Natasha Merat. Models of human decision-making as tools for estimating and optimizing impacts of vehicle automation. Transportation Research Record, 2672(37):153–163, 2018. [189] Arkady Zgonnikov, David Abbink, and Gustav Markkula. Should I stay or should I go? cognitive modeling of left-turn gap acceptance decisions in human drivers. Human Factors, 66(5):1399–1413, 2022. [190] Anthony D McDonald, Abhijit Sarkar, Jeffrey S Hickman, Hananeh Alambeigi, Gustav Markkula, and Tobias Vogelpohl. Modeling driver behavior during automated vehicle platooning failures. January 2021. [191] Alexander Fengler, Lakshmi N Govindarajan, Tony Chen, and Michael J Frank. Likelihood approximation networks (LANs) for
VOLUME ,
fast inference of simulation models in cognitive neuroscience. eLife, 10:e65074, 2021. [192] Nan Li, Dave W Oyler, Mengxuan Zhang, Yildiray Yildiz, Ilya Kolmanovsky, and Anouck R Girard. Game theoretic modeling of driver and vehicle interactions for verification and validation of autonomous vehicle control systems. IEEE Transactions on control systems technology, 26(5):1782–1797, 2017. [193] John R Anderson, Daniel Bothell, Michael D Byrne, Scott Douglass, Christian Lebiere, and Yulin Qin. An integrated theory of the mind. Psychological review, 111(4):1036, 2004. [194] Dario D Salvucci. Modeling driver behavior in a cognitive architecture. Human Factors, 48(2):362–380, 2006. [195] Shi Cao and Yili Liu. Queueing network-adaptive control of thought rational (QN-ACTR): An integrated cognitive architecture for modelling complex cognitive and multi-task performance. International Journal of Human Factors Modelling and Simulation, 4(3-4):199–224, 2014. [196] Yalda Ebadi, Ganesh Pai, Donald Fisher, and Fred Feng. Modelling level 1 situation awareness in driving: A cognitive architecture approach. Transportation Research Part C: Emerging Technologies, 166:104752, 2024. [197] John E Laird. The Soar cognitive architecture. MIT Press, 2012. [198] Rongliang Zhou, Haotian Cao, Jiakun Huang, Xiaolin Song, Jing Huang, and Zhi Huang. Hybrid lane change strategy of autonomous vehicles based on soar cognitive architecture and deep reinforcement learning. Neurocomputing, 611:128669, 2025. [199] Andrea Palazzi, Davide Abati, Simone Calderara, Francesco Solera, and Rita Cucchiara. Predicting the driver’s focus of attention: The DR(eye)VE project. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1720–1733, 2018. [200] Y. Xia, D. Zhang, J. Kim, K. Nakayama, K. Zipser, and D. Whitney. Predicting driver attention in critical situations. In Asian Conference on Computer Vision, pages 658–674. Springer, 2018. [201] Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, and Hongkai Yu. DADA-2000: Can driving accident be predicted by driver attention? analyzed by a benchmark. IEEE Transactions on Intelligent Transportation Systems, 23(6):4973–4986, 2021. [202] Luca Kultrera, Lorenzo Seidenari, Federico Becattini, Stefano Pini, and Alberto Del Bimbo. Looking at the right stuff: Guided semanticgaze for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11349–11358, 2020. [203] D. Ma, L. Shen, and Z. Deng. Traffic prediction based on braininspired spiking neural network. IEEE Transactions on Intelligent Transportation Systems, 23(10):17757–17769, 2021. [204] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. DeepDriving: Learning affordance for direct perception in autonomous driving. In IEEE International Conference on Computer Vision, pages 2722–2730, 2015. [205] Christopher D Wickens. Multiple resources and mental workload. Human Factors, 50(3):449–455, 2008. [206] William J Horrey, Christopher D Wickens, and Kristen P Consalus. Modeling drivers’ visual attention allocation while interacting with in-vehicle technologies. Journal of Experimental Psychology: Applied, 12(2):67–78, 2006. [207] Johan Engström, Gustav Markkula, Trent Victor, and Natasha Merat. Effects of cognitive load on driving performance: The cognitive control hypothesis. Human Factors, 59(5):734–764, 2017. [208] J. Wang, J. Ding, S. Zhang, and K. Li. Driving safety field theory modeling and its application in traffic safety. Transportation Research Part C: Emerging Technologies, 50:92–109, 2015. [209] S. Ji, A. Khajepour, and W. Melek. Navigating with graph representations for potential fields in deep reinforcement learning. In IEEE International Conference on Robotics and Automation (ICRA), pages 7596–7602. IEEE, 2020. [210] Sarvesh Kolekar, Joost de Winter, and David Abbink. Human-like driving behaviour emerges from a risk-based driver model. Nature Communications, 11:4850, 2020. [211] John D Lee and Katrina A See. Trust in automation: Designing for appropriate reliance. Human Factors, 46(1):50–80, 2004. [212] Kevin Anthony Hoff and Masooda Bashir. Trust in automation: Integrating empirical evidence on factors that influence trust. Human Factors, 57(3):407–434, 2015. [213] Jackie Ayoub, X Jessie Yang, and Feng Zhou. Modeling dispositional and initial learned trust in automated vehicles with predictability and
VOLUME ,
explainability. Transportation Research Part F: Traffic Psychology and Behaviour, 77:102–116, 2021. [214] Robert A Kaufman, Emi Lee, Manas Satish Bedmutha, David Kirsh, and Nadir Weibel. Predicting trust in autonomous vehicles: Modeling young adult psychosocial traits, risk-benefit attitudes, and driving factors with machine learning. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–24, 2025. [215] Zhaobin Mo, Rongye Shi, and Xuan Di. A physics-informed deep learning paradigm for car-following models. Transportation Research Part C: Emerging Technologies, 130:103305, 2021. [216] Zhaobin Mo and Xuan Di. Uncertainty quantification of car-following behaviors: physics-informed generative adversarial networks. In the 28th ACM SIGKDD in conjunction with the 11th International Workshop on Urban Computing (UrbComp2022), 2022. [217] Le Xu, Jianxiao Ma, and Yuchen Wang. A car-following model considering the effect of following vehicles under the framework of physics-informed deep learning. Journal of Advanced Transportation, 2022(1):3398862, 2022. [218] Maosi Geng, Junyi Li, Yingji Xia, and Xiqun Michael Chen. A physics-informed transformer model for vehicle trajectory prediction on highways. Transportation research part C: emerging technologies, 154:104272, 2023. [219] Chengming Wang, Dongyao Jia, Wei Wang, Dong Ngoduy, Bei Peng, and Jianping Wang. A knowledge-informed deep learning paradigm for generaliz-able and stability-optimized car-following models. Communications in Transportation Research, 5:100211, 2025. [220] Marisa Carrasco. Visual attention: The past 25 years. Vision Research, 51(13):1484–1525, 2011. [221] Yuan Gao, Mattia Piccinini, Yuchen Zhang, Dingrui Wang, Korbinian Moller, Roberto Brusnicki, Baha Zarrouki, Alessio Gambi, Jan Frederik Totz, Kai Storms, et al. Foundation models in autonomous driving: A survey on scenario generation and scenario analysis. IEEE Open Journal of Intelligent Transportation Systems, 2026. [222] Qiuxuan Wu, Peiyao Wu, and Robert West. Comparing LLMs for prompt-enhanced ACT-R and Soar model development: A case study in cognitive simulation. In AAAI Symposium Series, pages 285–292. AAAI, 2024. [223] Abdul Jawad and Jim Whitehead. Accident scenario generation using driver behavior model. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 4210–4215. IEEE, 2024. [224] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. Towards learning-based planning: The nuplan benchmark for realworld autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 629–636. IEEE, 2024. [225] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kümmerle, Hendrik Königshof, Christoph Stiller, Arnaud de La Fortelle, and Masayoshi Tomizuka. INTERACTION dataset: An INTERnational, adversarial and cooperative moTION dataset in interactive driving scenarios with semantic maps. arXiv preprint, 2019. arXiv:1910.03088. [226] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highd dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of trajectory prediction and driving behavior models. 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 1620–1627, 2018. [227] J. Colyar and J. Halkias. NGSIM – next generation simulation program: Data analysis methodology. In FHWA Report, 2007. [228] Julian Bock, Robert Krajewski, Tobias Moers, Steffen Runde, Lennart Vater, and Lutz Eckstein. The ind dataset: A drone dataset of naturalistic road user trajectories at german intersections. In 2020 IEEE Intelligent Vehicles Symposium (IV), pages 1929–1934. IEEE, 2020. [229] Robert Krajewski, Tobias Moers, Julian Bock, Lennart Vater, and Lutz Eckstein. The round dataset: A drone dataset of road user trajectories at roundabouts in germany. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1–6. IEEE, 2020. [230] Tobias Moers, Lennart Vater, Robert Krajewski, Julian Bock, Adrian Zlocki, and Lutz Eckstein. The exid dataset: A real-world trajectory dataset of highly interactive highway scenarios in germany. In 2022
43
Rahmani et al.: Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic
IEEE Intelligent Vehicles Symposium (IV), pages 958–964. IEEE, 2022. [231] Emmanouil Barmpounakis and Nikolas Geroliminis. On the new era of urban traffic monitoring with massive drone data: The pNEUMA largescale field experiment. Transportation Research Part C: Emerging Technologies, 111:50–71, 2020. [232] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618– 11628, 2020. [233] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Shun Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, and Deva Ramanan. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. [234] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference on Robot Learning, pages 409–418. PMLR, 2021. [235] Matthias Althoff, Markus Koschi, and Stefanie Manzinger. Commonroad: Composable benchmarks for motion planning on roads. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 719–726. IEEE, 2017. [236] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In Advances in Neural Information Processing Systems, volume 36, pages 7730–7742, 2023.
Saeed Rahmani (Graduate Student Member, IEEE) is a PhD Candidate at Delft University of Technology and a visiting scholar at Oxford University. He received the B.Sc. and M.Sc. degrees in civil engineering, transportation and traffic systems engineering. Since 2022, He has been pursuing a Ph.D. degree, focusing on motion planning and decision-making for automated vehicles focusing on mixed automated and human traffic.
Dr. Eugene Vinitsky is an Assistant Professor at the Tandon School of Engineering, New York University, New York, United States. He received his Ph.D. in Control from the University of California, Berkeley. His research applies multi-agent reinforcement learning to robotics and transportation, with a focus on synthesizing complex, human-like behavior from unsupervised interaction between groups of learning agents and designing scalable data-driven simulators.
Prof.Dr.ir. Bart van Arem was appointed as full professor Transport Modelling at the department of Transport and Planning in 2009 and serves as Pro Vice Rector for Doctoral Affairs of TU Delft since 2021 He was head of the department Transport & Planning from 2010 till 2017 and served as director of the TU Delft Transport Institute from 2012-2021. His research focuses on analysing and modelling the implications of intelligent transportation systems, such as automated, electric and shared vehicles.
Dr.ir. Simeon C. Calvert is an Associate Professor of traffic and network management at TU Delft. He is the director of the Automated Driving & Simulation (ADaS) Lab in the Department of Transport & Planning and co-leads the Delft AI Lab on urban mobility behaviour: CiTy- AI. From 2010 to 2016, he worked as a Research Scientist with TNO. His research interests include ITS, impacts of vehicle automation, traffic management, traffic flow theory, and network analysis.
Shiva Rasouli is a Ph.D. candidate and Research Assistant in the Department of Industrial and Systems Engineering at the University of Michigan, Dearborn, Michigan, USA. Her current research interests include affective computing, humancomputer interaction, driver behavior, and user state estimation to facilitate human and automation teaming.
Daphne Cornelisse is a Ph.D. student at the Tandon School of Engineering, New York University, New York, United States. Her research applies reinforcement learning and imitation learning to multi-agent systems. She is particularly focused on developing effective, human-compatible simulation agents for autonomous driving, aiming to scale selfplay for reliable traffic modeling and evaluation.
44
VOLUME ,