Workshop on Computational Models of Human Road User Behavior for Evaluation of Automated Vehicles

Abstract

Retrospective evaluations of automated vehicle (AV) safety performance typically compare AV crash rates to those of the general driving population. However, this may include crashes involving risky, impaired, distracted, or otherwise non-compliant driver behavior. An alternative approach is to benchmark AV performance against a behaviorally defined subset of competent and compliant human drivers.

This work introduces an initial behavioral reference model derived from U.S. national crash data (NCD) that classifies prudent, attentive, and sober (PAS) drivers by removing vehicle crashes with evidence of imprudent, inattentive, or non-sober driving. PAS drivers are identified using exclusion criteria derived from variables available in police-reported crash data, including indicators of impairment, distraction, speeding, and traffic violations. Example findings are given for light passenger vehicles in police-reported crashes.

Beyond supporting retrospective AV safety comparisons, the PAS framework provides an empirical characterization of behavior consistent with emerging concepts of "careful and competent" human driving. As such, it can be used to complement behavioral definitions used in regulatory and industry approaches to AV evaluation. The resulting PAS population will serve as a data-driven benchmark for estimating the safety impacts of specific behavioral constraints and for quantifying the crash reductions associated with excluding specific forms of unsafe driving behavior.

The presentation will discuss how PAS relates to computational definitions of desirable driving, including the UNECE Competent and Careful Human Driver framework and the NIEON driver behavior model, and examine which behavioral constructs from such models can be operationalized using NCD. PAS-based filtering will help inform the development and evaluation of behavioral reference models used in AV safety assessment, contributing to more interpretable and policy-relevant AV safety benchmarks.

Abstract

This study examines how competent and careful human drivers perceive risk and initiate braking in safety-critical motorway cut-in scenarios. It relies on evidence from controlled test-track experiments in which professional drivers reacted to a robotic surrogate vehicle executing sudden, unindicated cut-ins under varying speeds, gaps, and lateral velocities. The data are first used to evaluate the two human benchmark models in UNECE Regulation No. 157. Although no crash occurred in the experiments, regulatory models classified about 68% of the manoeuvres as unpreventable, revealing a mismatch between model predictions and observed behaviour. This mismatch mainly reflects an inadequate representation of risk perception and delayed reaction onsets.

A new driver model is therefore derived from the data. The analysis identifies the dynamic cues governing risk perception, with cut-in yaw rate emerging as the dominant predictor. Reaction onset is represented as a discrete-time hazard process and selected through double Monte Carlo cross-validation, separating model discovery from out-of-sample assessment. Braking behaviour is modelled using a Gipps-based formulation, which proved more accurate than the regulatory models and the IDM in reproducing the observed emergency braking dynamics.

The resulting model provides a more credible human-driver benchmark than regulatory models for automated-driving safety assessment in safety-critical cut-in situations.

Abstract

Evaluation of automated vehicles depends on models of human road user behavior. Data-driven and generative models dominate practice and are assessed with likelihood-based distribution-matching scores. We argue such scores, dominated by routine driving, are ill-suited to credible AV evaluation: a model can score well while failing on the rare, safety-relevant events that contribute little to the mean.

We illustrate failure modes these metrics miss using SMART, a state-of-the-art trajectory generation transformer model, on the Waymo Open Dataset. In sparsely represented high-speed scenes, the model overshoots laterally and leaves the road, and when tested in emergency conflicts it produces implausible responses that do not match human brake reaction. We thus propose extending evaluation metrics to better reflect the full range of relevant driving behavior.

We then present the latest developments in driveBOT, a mechanistic cognitive driver model built on the cogniBOT architecture, representing perception, cognition, and motor control without fitting to benchmark data. It reproduces routine traffic and replicates naturalistic-driving effects, including surprise-modulated emergency braking, and now adds inattentive and impaired driving through distraction and alcohol-impairment variants. We discuss its use as a behavioral reference model, arguing that hybrid mechanistic and AI-based approaches are needed for credible AV behavior modeling.

Abstract

Human driving behavior is heterogeneous, and reproducing this heterogeneity in simulation is essential for automated vehicle (AV) validation. We present Style-Conditioned Self-Play (SCSP), a closed-loop multi-agent model that makes driving style an explicit, controllable, per-agent input grounded in human trajectories. In the proposed method, a variational autoencoder learns a style latent from human kinematic data, and a multi-agent policy is then trained by combining PPO self-play with style-conditioned behavior cloning.

This ties each style value to real driving behavior, so that changing the input changes the agent's behavior in a continuous and predictable way. At test time, different agents can be given different driving styles using a single trained policy. We assess SCSP along three properties relevant to AV validation. Behavioral realism is confirmed by a WOSAC score above 0.7. Style controllability is demonstrated by showing that a single style input can move agents behavior smoothly from a cautious, low-risk style to an aggressive, higher-risk one, without loss in realism.

Relevance for AV validation is shown by evaluating an ego planner in heterogeneous traffic generated by SCSP. Exposing the planner to heterogeneous traffic doubles its collision rate from 2.2% to 4.6% and drops its goal completion rate from 96.5% to 87.2%, relative to homogeneous traffic. These results show SCSP provides a scalable, human-grounded solution for generating heterogeneous traffic.

Abstract

Driving is fundamentally an interaction between humans and intelligent systems. Yet, computational representations of drivers remain fragmented, often focusing on isolated aspects such as gaze behavior, or takeover actions. We argue that the next frontier of human-centered automation is the development of Driver World Models (DWMs): unified autoregressive models that characterize drivers’ attention, and actions as a function of dynamic traffic context.

In this talk, we first introduce the DWM conceptual framework which represents the driver as a temporally evolving system whose internal states, external behaviors, interactions with automation, and surrounding context jointly shape driving behavior. We will then present a prototype of the aforementioned framework based on autoregressive generative models trained on multimodal behavioral and psycho-physiological driver data collected from controlled semi-automated driving experiments.

Using DWM, we further distinguish between General DWMs, which characterize population-level driver behavior and can generate realistic virtual road users for simulation and automated vehicle evaluation, and Personalized DWMs, which capture the unique characteristics of individual drivers offering a pathway toward humanized automated vehicles.

We conclude by discussing the research directions for establishing DWM as a new paradigm for computational human behavior modeling and human-centered intelligent transportation systems.

Abstract

Driving requires humans to continuously perceive, predict, and act under uncertainty while pursuing multiple goals, from collision avoidance to efficient and socially coordinated interaction with other road users. However, existing computational models that explain human driving behavior are typically fragmented, focusing on specific scenarios or isolated aspects of behavior. Active inference has recently emerged as a principled computational and generalizable framework for modeling such adaptive behavior through the minimization of expected free energy.

Here, we present a series of active inference models that together provide a unified account of human driving behavior. We mainly focus on how active inference can offer a closed-loop explanation of collision avoidance across diverse traffic scenarios, reproducing empirical findings on response timing, maneuver selection, and execution. We then briefly discuss extensions of the same framework to interactive driving behavior (and the role of traffic rules and communication therein) and the modeling of affective state.

These extensions demonstrate how active inference naturally generalizes to increasingly complex aspects of driving, including interactions between road users and the influence of emotional state on decision-making, without requiring a fundamentally different modeling approach. Together, these results highlight active inference's promise for understanding and modeling human driving behavior.

Abstract

This presentation introduces TNO’s research on competent driving as a reference for assessing Automated Driving Systems (ADS). While regulations increasingly refer to the behaviour of a “competent and careful human driver,” the concept remains difficult to define and measure. TNO approaches competent driving from a Safety-II perspective, focusing not only on avoiding failures or collisions, but also on preserving the everyday interactions that make traffic safe, predictable and socially acceptable.

The work translates competent driving into practical specifications for ADS assessment. Driving manuals, traffic rules and expert knowledge are formalized into mathematical descriptions of behaviour, covering temporal, spatial, ordinal and dynamic relationships relevant to operational and tactical driving. These specifications are calibrated using empirical data from professional CCV-D1 drivers, who provide a candidate reference for competent highway driving.

The resulting assertion ranges can be used to evaluate whether automated vehicles behave within defensible bounds of competent driving, rather than simply checking compliance with fixed rules. Deviations can then be assessed in terms of frequency, severity, duration and recovery. The talk aims to stimulate expert discussion on the assumptions and limitations of this approach, and on how competent driving can be made explicit, measurable and useful for the safe introduction of automated vehicles.

Abstract

Consumer-testing campaigns such as Euro NCAP offer established procedures for ADAS but none for ADS, and existing protocols do not capture how automated behaviour compares to real human drivers. This contribution presents a scenario-based virtual-testing and grading framework, developed in the Horizon Europe project i4Driving (Deliverable D6.6, led by RDW), extending Euro NCAP/UN R157-01-style evaluation to ADS and ADAS consumer testing.

The framework follows a Scenario-Environment-Analysis structure (D6.1), aligned with the SUNRISE Safety Assurance Framework, and grades collision-avoidance, regulatory safety-distance/deceleration criteria, and comfort, weighted by scenario difficulty. The i4Driving probabilistic human driver model library, calibrated and validated against real-driver data, serves as the benchmark for interpreting automated behaviour.

A key outcome is that human-driver models provide an essential reference point for interpreting ADS/ADAS behaviour: a human baseline indicates whether a system struggles in scenarios that are also challenging for humans, or whether deficiencies are system-specific, and helps identify cases where a system is technically safe but uncomfortable, relevant to user acceptance.

Overall, this work supports virtual testing as a scalable, flexible tool for consumer-testing and shows that human-baseline models can strengthen future ADS/ADAS evaluation protocols by providing a fair, realistic, behaviour-based benchmark.

Abstract

An important aspect in evaluating an automated driving system (ADS) is how it behaves in safety critical scenarios. Computational driver behaviour models are created for ADS validation, evaluation, and benchmarking. However, it remains unclear how such models should be scored when used as reference models.

We develop and demonstrate a task-oriented scoring framework for driver behaviour models in collision avoidance scenarios. The framework decomposes collision avoidance behaviour into safety-relevant aspects of the response and compares model behaviour with human reference behaviour. We apply it to six human collision avoidance scenarios and evaluate four driver model variants: a no-response baseline, a reaction-time braking model, and deterministic and probabilistic optimal-control models.

For each model, we calculate safety-relevant behavioural metrics and assess whether the model reproduces the effects of scenario conditions compared to human data. Preliminary results show that the framework provides both an overall model ranking and a diagnostic breakdown of where models align with or differ from human data. This helps identify whether a model reproduces specific aspects of collision avoidance behaviour, rather than judging it as good or bad.

We also highlight open questions, including a more robust response- time definition, pass/fail scoring criteria for practical automated vehicle benchmarking, and future application of the framework to automated vehicle policies.

Abstract

Safety and comfort of automated driving systems (ADS) heavily rely on understanding characteristics of manual human driving behavior. Acceleration patterns play a central role in this, and being able to statistically model these patterns would enable a behavioral reference which can be used to support comfort and motion plausibility in ADS design. This paper presents a large-scale analysis of how positive and negative longitudinal acceleration distributions vary with longitudinal vehicle speed, using more than 64 million data points collected across several countries, primarily in Europe and the United States. The data set forms the basis for probabilistic modeling of how human acceleration behavior changes across speeds, with boundaries captured by gamma-based statistical models. The results disclose a progressive narrowing of acceleration variance as speed increases, accompanied by a simultaneous decrease in maximum acceleration. Despite the contraction observed for extreme values, the 99th percentile of the acceleration distribution remains relatively constant within ± 2.5 m/s2 across speeds. This study provides a continuous estimation of the joint speed-acceleration distribution and derives a corresponding speed-acceleration envelope. The findings can inform the design and safety assessment of ADS and advanced driver assistance systems (ADAS), contributing toward human-like longitudinal control strategies.

Abstract

Human behavior models are essential as behavior references and for simulating human agents in virtual safety assessment of automated vehicles (AVs), yet current models face a trade-off between interpretability and flexibility. Generalpurpose large language models (LLMs) offer a promising alternative: a single model potentially deployable without parameter fitting across diverse scenarios. However, what LLMs can and cannot capture about human driving behavior remains poorly understood. We address this gap by embedding two general-purpose LLMs (OpenAI o3 and Google Gemini 2.5 Pro) as standalone, closed-loop driver agents in a simplified onedimensional merging scenario and comparing their behavior against human data using quantitative and qualitative analyses. Both models reproduce human-like intermittent operational control and tactical dependencies on spatial cues. However, neither consistently captures the human response to dynamic velocity cues, and safety performance diverges sharply between models. A systematic prompt ablation study reveals that prompt components act as model-specific inductive biases that do not transfer across LLMs. These findings suggest that generalpurpose LLMs could potentially serve as standalone, readyto- use human behavior models in AV evaluation pipelines, but future research is needed to better understand their failure modes and ensure their validity as models of human driving behavior.

Abstract

Computational models of careful and competent human drivers are essential for scenario-based evaluation of automated driving systems (ADS). However, most existing safety reference models primarily focus on longitudinal braking, neglecting the role of evasive steering in human collision avoidance. This paper proposes a hybrid Fuzzy-Safety Model (FSM-H) that integrates longitudinal mitigation and lateral avoidance within a unified behavioral framework. The braking component is governed by Proactive Fuzzy Safety (PFS) metrics, representing the erosion of longitudinal safety margins, while the steering component is driven by Criticality Fuzzy Safety for lane-change (CFS-LC), capturing lateral conflict severity and maneuver feasibility. A finite-state architecture models the sequential escalation from nominal driving to braking and, when necessary, to evasive steering, incorporating percep- tion–reaction time and lane-check delays to reflect human de- cision processes. The model is evaluated in reconstructed high- criticality cut-in scenarios and compared with braking-only and steering-only reference strategies. Results show that the hybrid approach expands the preventability envelope while maintain- ing behavioral plausibility and computational tractability. The proposed framework provides a transparent and explainable human reference model suitable for simulation-based ADS safety benchmarking and regulatory assessment.