Review of Scenario Virtual Testing Technology for Autonomous Vehicles: Migration Challenges Between Symmetric Frameworks and Asymmetric Scenarios

Abstract

With the rapid development of automated wheeled vehicle technology, complex vehicle functions require extensive safety testing for verification. Compared with real-vehicle testing, scenario-based virtual testing, which constructs virtual environments to simulate real scenarios and efficiently evaluates vehicle safety and risk decision-making capabilities, has become a core means for the safety evaluation of automated wheeled vehicles. This paper outlines the research progress of scenario-based virtual testing for automated wheeled vehicles (including highway autonomous vehicles and off-highway autonomous vehicles); classifies three key technologies in highway scenarios, hazard evaluation, hazardous scenario generation and generalization, and acceleration evaluation; and reveals the challenges faced when existing methods are migrated to agricultural vehicles, engineering vehicles, etc., such as low scenario adaptability, multi-dimensional coupling of risk targets, and weak data foundation. This study finds that current technologies have formed a symmetric framework in highway scenarios, but there are significant adaptability problems when migrating to off-highway scenarios due to scenario asymmetry. To this end, this paper proposes ideas for realizing off-highway scenario testing by adopting methods such as dynamic safety distance reconstruction, multi-physics simulation, and digital twin-driven approaches, providing theoretical support for building a unified safety assessment platform for automated wheeled vehicles.

1. Introduction

With the rapid advancement of intelligent transportation technologies, the automation and intelligence of wheeled vehicles have become the core direction of industry transformation. From autonomous driving vehicles to specialized vehicles such as agricultural vehicles, engineering vehicles, and special operation vehicles, various wheeled vehicles are continuously breaking through the boundaries of autonomous decision-making capabilities through sensor fusion, control algorithm optimization, and data-driven technologies [1,2,3]. However, the safety and reliability of automation technologies directly determine the upper limit of equipment automation and operational scenario safety, urgently requiring systematic testing to verify technological maturity. Different from the testing model of traditional vehicles focusing on mechanical performance, the testing of automated wheeled vehicles requires constructing virtual simulation environments to mimic the complex and diverse operational scenarios in the real world, and evaluating system safety through digital means [4]. As a pre-testing link for physical vehicle testing, scenario virtual testing is a technical method that reproduces or generates various typical and extreme scenarios in a virtual environment and verifies the vehicle’s behavioral logic and safety performance under different working conditions through digital simulation. It can avoid physical vehicle damage and safety risks in high-risk scenarios and is particularly suitable for reproducing extreme working conditions and high-frequency testing verification, thus having become the core support for the technological iteration of autonomous vehicles [5,6,7]. Currently, research on scenario virtual testing for autonomous driving vehicles mainly focuses on three key technologies: scenario hazard evaluation, hazardous scenario generation and generalization, and accelerated evaluation. Risk is assessed by quantifying traffic conflict probability and collision severity, scenario libraries are expanded using algorithms such as generative adversarial networks (GANs) and Variational Autoencoders (VAEs), and the probability of hazardous scenarios is increased through importance sampling (IS) and reinforcement learning (RL). This has formed a technically symmetric framework covering urban roads and highways, i.e., the consistency and universality of testing logic and evaluation indicators in similar scenarios.

However, the application boundaries of automation technologies extend far beyond public transportation scenarios. Agricultural vehicles need to cope with terrain undulations and dynamic crop growth in unstructured farmland [8,9]; engineering vehicles must complete heavy-load transportation and multi-equipment collaborative obstacle avoidance in mines and construction sites [10,11]; and special vehicles such as fire trucks and snowplows need to adapt to extreme environments like high temperatures and extreme cold [12]. With the advancement of concepts such as smart agriculture and unmanned mines, these non-highway automated wheeled vehicles have an increasingly urgent need for scenario virtual testing. Nevertheless, existing research mainly focuses on highway scenarios for autonomous driving vehicles, and scenario virtual testing for equipment like agricultural vehicles and engineering vehicles is still in its infancy. Migrating autonomous driving scenario virtual testing technologies to non-highway fields faces three core challenges.

First, there is the challenge of the cross-domain adaptation of scenario hazard evaluation criteria. Autonomous vehicles quantify risks with traffic conflict probability and collision severity as core metrics, but hazardous scenarios for non-highway automated wheeled vehicles require integrating multidisciplinary models (e.g., agricultural vehicles facing sinking risks need to couple soil mechanics with vehicle load dynamics [13], while engineering vehicles encountering overturning risks must associate terrain slope with mechanical structural strength). The asymmetry of evaluation objects leads to a lack of cross-domain quantitative indicators in existing frameworks, making it difficult to support the credibility of test results. Second, there is insufficient coverage of dynamic characteristics in scenario generation and generalization. Autonomous driving scenario libraries rely on traffic rules and intelligent algorithms to expand data, but the dynamic changes in unstructured environments (farmland, mining areas) (such as crop growth and terrain collapse), non-standard interactions (e.g., collaboration between construction site vehicles and robotic arms [14]), and the asymmetry of environmental features make them difficult to model through rule-based or purely data-driven methods. Traditional generalization technologies have limited coverage of long-tail distribution scenarios. Third, there is low adaptability in coupling physical mechanisms of scenario virtual testing methods. Autonomous vehicles trigger hazardous scenarios through intelligent optimization methods, but hazardous events in other automated wheeled vehicles are often deeply bound to physical mechanisms (such as hydraulic leakage, mechanical fatigue, and ten-thousand-hour-level failures). Traditional probability sampling or random exploration methods are inefficient, requiring the combination of fault tree analysis and digital twins to achieve targeted accelerated evaluation.

Therefore, breaking through scenario boundaries and extending the technical path of autonomous driving scenario virtual testing to non-highway automated wheeled vehicles such as agricultural machinery and engineering vehicles is essentially about finding a balance between symmetric technical methods and asymmetric scenario characteristics. Constructing a testing system covering the entire “highway–non-highway” scenario spectrum has become a key issue in promoting the comprehensive implementation of automated wheeled vehicle technologies. This paper systematically combs the three key technologies of autonomous driving scenario virtual testing, deeply analyzes cross-domain migration challenges, and explores non-highway scenario-adaptive solutions, aiming to provide theoretical support and methodological references for the global safety testing of automated wheeled vehicles and accelerate the application expansion of scenario virtual testing technologies to diverse equipment fields. Specifically, the contributions are as follows:

First, this paper systematically classifies the technologies of scenario hazard evaluation, scenario generation and generalization, and acceleration evaluation in autonomous driving scenario virtual testing; reveals the symmetric technical framework formed in highway scenarios; constructs an analytical framework across highway and non-highway scenarios; clarifies the logic of technical migration; and provides methodological tracing for the testing of non-highway automated wheeled vehicles.

Second, focusing on the pain points of technical migration, this paper analyzes the asymmetric characteristics of highway and non-highway scenarios in terms of environment, risk, and data; accurately identifies the application shortcomings of existing methods in complex operating environments; and anchors directions for subsequent technical breakthroughs.

Third, based on migration challenges, this paper explores adaptive paths integrating multidisciplinary models, dynamic scenario modeling, digital twin technologies, etc., to realize the coverage of symmetric methodologies over asymmetric scenarios; provides ideas for constructing a scenario virtual testing system for non-highway automated wheeled vehicles; and promotes the upgrade of the technical paradigm of scenario virtual testing from dedicated verification for autonomous vehicles to universal evaluation for automated wheeled vehicles.

The rest of this paper is organized as follows. Section 2 introduces the scenario hazard evaluation method. Section 3 introduces the hazardous scenario generation and generalization method. Section 4 introduces the acceleration evaluation method. Conclusions and prospects are given in Section 5.

2. Scenario Hazard Evaluation Method

In recent years, as the level of vehicle automation has advanced, wheeled vehicles, typified by autonomous vehicles, encounter increasingly complex scenarios during operation, and are tasked with an expanding range of driving missions. These intricate scenarios and diverse tasks pose substantial safety challenges to automated wheeled vehicle systems. Whether autonomous vehicles navigate complex urban traffic networks, agricultural vehicles execute field operations, or engineering vehicles conduct heavy-duty transportation in mines and construction sites, accurate scenario hazard evaluation stands as a core prerequisite for ensuring the safe and stable operation of vehicles. Scientific and rational hazard evaluation not only enables the precise prediction of potential risks through quantitative indicators and probability models, providing a basis for prioritizing test scenarios and achieving efficient allocation of test resources, but also, by deeply analyzing the sources and influencing factors of hazards, reversely drives improvements in automated wheeled vehicle design optimization and control algorithm refinement.

In scenario virtual testing, the scenario hazard degree is a quantitative representation of the possibility and severity of potential risks or accidents encountered by vehicles in a specific scenario. It not only reflects the objective existence of threat factors in the scenario but also includes the evaluation of the vehicle system’s ability to respond to these threats. Currently, evaluation methods are mainly divided into two categories: model-driven and data-driven. The former model the scenario through mathematical or physical models and predict subsequent states. The latter are based on the statistical laws and patterns of data, using technologies such as machine learning to generate models through the learning and training of large amounts of data for prediction.

2.1. Model-Driven Scenario Hazard Evaluation Method

Model-driven methods calculate the probability of the host vehicle being endangered by the surrounding environment by constructing accurate scene models and formally defining various elements within the scenario. Depending on whether the model accounts for the motion uncertainty of the surrounding environment, model-driven methods can be further subdivided into deterministic evaluation methods and probabilistic evaluation methods.

(1) Deterministic Evaluation Method

Deterministic evaluation generally employs simplified physical models to describe the motion of traffic participants and selects specific indicators to characterize the hazard level. When the calculated result of an indicator exceeds a certain threshold, a hazard is deemed to exist. The most frequently used indicators include longitudinal deceleration, acceleration, Time HeadWay (THW), and Time To Collision (TTC). In some common scenarios, scholars use a combination of several indicators for hazard evaluation. For instance, Hang [15] utilized the TTC metric to determine appropriate criteria for activating driver assistance systems like collision avoidance systems, aiming to prevent rear-end collisions on highways. Fu [16] used the deceleration rate to avoid a crash indicator for scenario hazard evaluation. This indicator quantifies the deceleration rate required for a following vehicle to avoid a collision and, in conjunction with traffic conflict analysis, evaluates the hazard level in road safety research. Xia [17] evaluated hazard levels by combining TTC, Minimum Safety Distance (MSD), and the maximum deceleration capacity indicator of the vehicle. TTC reflects the urgency of a collision, MSD is the minimum distance required to avoid a collision, and the maximum deceleration capability represents the limit of emergency deceleration. A high-hazard state can be determined when TTC is less than the threshold, the relative distance is less than MSD, and the actual deceleration fails to reach the maximum deceleration capability to stop within an appropriate distance. Although different hazard indicators are used in different scenarios, they essentially screen potential hazardous events through the violent operations of the subject vehicle and then make further judgments in combination with indicators such as TTC and THW. Based on existing research, this paper summarizes the indicator selection methods for several common scenarios, as shown in

Table 1. Scenario hazard evaluation indicators and calculations.

Scenario	Description	Indicators	Formulas and Parameter Explanations
Tight Following	Risk of rear-end collision due to too closely following distance.	THW, Modified Time Headway (mTHW)	$THW={\displaystyle \frac{D}{v_{av}}},$ $mTHW={\displaystyle \frac{D+v_{av}·t_r}{v_{av}}},$ where D is the longitudinal distance between the host vehicle and the background vehicle, $v_{av}$ is the speed of the host vehicle, and $t_r$ is the vehicle reaction time.
Host Vehicle Lane Changing	Risk of lateral/ longitudinal conflict.	Lateral Safety Distance (LSD), Lane-changing Time to Collision (LTC)	$LSD=y_{av}-y_{bv},$ $LTC=min\left({\displaystyle \frac{x_{bv}-x_{av}}{v_{av}-v_{bv}}},{\displaystyle \frac{x_{av}-x_{bv}}{v_{bv}-v_{av}}}\right),$ where $x_{av},y_{av}$ are the current longitudinal and lateral positions of the host vehicle, $x_{bv},y_{bv}$ are the current longitudinal and lateral positions of the background vehicle, and $v_{bv}$ is the speed of the background vehicle.
Leading Vehicle Cutting In	Risk of emergency braking/evasion for the host vehicle.	Minimum Longitudinal Time to Conflict (mTTC), Dynamic Safety Distance (DSD)	$mTTC={\displaystyle \frac{x_{av}-x_{bv}}{v_{av}-v_{bv}}},$ $DSD=v_{av}·t_r+{\displaystyle \frac{v_{av}^2-v_{bv}^2}{2a_{max}}},$ where $a_{max}$ is the maximum deceleration of the host vehicle.
Opposite Dir- ection Conflict	Risk of meeting vehicles due to insufficient lateral space.	Lateral Conflict Probability (LCP), Opposite-direction Minimum Safety Distance (OSD)	$LCP=\mathrm{\Phi }\left({\displaystyle \frac{W_{ava}-W_{dem}}{{\sigma }_W}}\right),$ $OSD={\displaystyle \frac{(v_{av}+v_{bv})·t_r}{2}}+\Delta y_{safe},$ where $W_{ava}$ is the actual available lateral width of the road, $W_{dem}$ is the minimum lateral width required for the safe meeting of vehicles, ${\sigma }_W$ is the standard deviation of the uncertainty of the lateral width, and $\Delta y_{safe}$ is the lateral safety margin.
Turning Conflict	Risk of cross-conflict (opposite direction, same direction)	Turning Time to Collision ($TT{C}_{turn}$), Steering Safety Angle (SSA)	$TT{C}_{turn}={\displaystyle \frac{d_0-x_{av}·cos{\theta }_{av}}{v_{av}·cos{\theta }_{av}-v_{bv}·cos{\theta }_{bv}}},$ $SSA=arctan\left({\displaystyle \frac{v_{av}·sin{\theta }_{av}}{v_{av}·cos{\theta }_{av}+v_{bv}}}\right),$ where $d_0$ is the longitudinal distance from the conflict point to the origin, ${\theta }_{av}$ is the steering angle of the host vehicle, and ${\theta }_{bv}$ is the driving direction angle of the background vehicle.

. The advantage of deterministic methods lies in their computational efficiency and real-time performance. Since explicit physical models are used and there is no need for random variable sampling, scenario hazard evaluation can be completed within milliseconds, making it suitable for real-time decision making in emergency scenarios. Additionally, they have few input parameters and are easy to obtain, thus exhibiting good engineering applicability in vehicle-mounted terminals with limited hardware resources. However, the limitations of this method are also quite notable. It uses simplified models that deviate from actual traffic scenarios and cannot handle dynamic behaviors such as sudden acceleration, deceleration or the lane-changing of target objects. Moreover, this method cannot model unknown or uncertain factors, such as faults or abnormal behaviors of autonomous vehicles. In the future, the adaptability to complex interactions can be improved by incorporating dynamic behavior prediction models, and an uncertainty quantification framework can be introduced to cover unknown risk factors.

(2) Probabilistic Evaluation Method

However, probabilistic evaluation methods take into account the uncertainties during the vehicle’s movement. They use probabilistic models to describe the movement and then evaluate the scenario hazard. Specifically, probabilistic evaluation methods first predict the trajectories of surrounding vehicles, then conduct collision detection through combination with the host vehicle’s future trajectory, and finally derive the host vehicle’s collision probability based on the collision detection results. In the application of autonomous driving scenario hazard evaluation, some main sources of uncertainty include dynamic modeling errors, measurement noise, and misinterpretation of the driver’s intention. Probabilistic methods usually utilize probabilistic descriptions and establish hazard evaluation models using the temporal and spatial relationships between vehicles and the uncertainty of input data. Commonly used probabilistic methods include Bayesian networks, Markov decision processes, and Monte Carlo (MC) methods. Li [18] used Bayesian theory to model the hazard evaluation levels and employed the evaluation results to design the reward function of the intelligent decision-making model based on RL, so as to find the strategy with the minimum expected risk. A probabilistic method based on conditional random fields was used to evaluate the collision risk, and simulations were carried out in three driving scenarios. Meanwhile, driving styles were also considered, so that the collision avoidance strategy could meet the needs of different passengers [19]. Noh [20] compared the performance of the Markov chain and MC simulation methods in different driving scenarios and concluded that MC simulation is more suitable as a collision hazard evaluation model. Based on this, Hu [21] developed a new collision hazard model to assess the collision risk of autonomous vehicles. Hu combined vehicle-level and network-level collision prediction and adopted an interactive perception motion model and dynamic Bayesian network to describe the interactive behaviors of vehicles. The study found that this model can significantly improve the ability to detect hazardous vehicles under dangerous traffic conditions. Shangguan [22] used the remaining time to reach the collision point as the risk score, adopted an exponential function to map the risk score to the risk probability, and then proposed a Markov decision process to minimize the expected risk return as the collision avoidance control strategy.

Compared with deterministic assessment methods, probabilistic assessment methods can quantify hazard levels more comprehensively. They can incorporate the uncertainties of driving behaviors and motion parameters, extend the prediction time domain (identifying risks 2–3 s in advance), and better conform to the actual evolution of hazards in complex scenarios such as roundabouts. However, they have problems such as large computational loads (5–10 times that of deterministic methods, making it difficult to meet real-time requirements) and strong dependence on massive data (with prominent data scarcity in rare hazardous scenarios). In the future, the computational complexity can be reduced through lightweight models, and the problem of data scarcity can be alleviated by combining few-shot learning or data generation technologies.

(3) Evaluation Method Based on Potential Field Theory

In addition, another emerging model-driven method is the evaluation method based on potential field theory. It draws on the potential field theory in physics and uses a repulsive force field to describe the collision hazard. In scenario hazard evaluation, the evaluation method based on potential field theory assumes that elements in the scenario will generate field strengths. These field strengths act on the vehicle to generate a repulsive force, and the potential energy can be obtained by integrating the repulsive force over distance. In terms of field strength calculation, different scenario elements will generate field strengths of different magnitudes. When calculating the field strength acting on the vehicle, Cao [23] considered the coordinates of road boundaries and obstacles in the scenario. In addition, Yang [24] took the relative distance between the vehicle and the obstacle as the core parameter for field strength calculation. Dixit [25] considered the lane centerline and lane boundaries on the basis of traditional potential field theory to calculate the total field strength acting on the vehicle. Lu [26] considered the influence of vehicle type on field strength calculation when calculating the field strength generated by the vehicle. Wang [27] further improved the reliability of the field strength by incorporating the driver’s style. In order to make the calculated field strength reflect future risks more, some scholars [28,29] introduced the acceleration parameter to calculate the field strength on the basis of relative position and relative speed. After obtaining the field strength at the position of the vehicle, the repulsive force can be obtained by multiplying the field strength by the vehicle’s virtual mass. When calculating the virtual mass, the actual mass of the vehicle and the driving speed are usually comprehensively considered [30], because the vehicle’s own speed also affects the risk during driving. For example, the risk of a vehicle driving at high speed is higher than that at low speed. When quantifying the risk, some scholars directly use the repulsive force to quantify the risk level, that is, the greater the repulsive force acting on the vehicle, the greater the risk during the vehicle’s driving process; some scholars also try to use potential energy to quantify the risk level [31].

Compared with the previous two methods, the hazard evaluation method based on the potential energy field can capture the coupling influence of elements such as multi-vehicle movement and road boundaries in real time through the continuous superposition and dynamic update of field strength. For example, in the scenario of a roundabout intersection, due to the complex driving trajectories of vehicles and frequent multi-vehicle interactions, traditional evaluation methods based on a single indicator are difficult to fully characterize the dynamically changing hazard situation (Figure 1). Vehicles at a roundabout intersection need to continuously adapt to circular motion and merging/exiting behaviors. The speed difference and angle difference in traffic flows in different directions change dynamically with time. Traditional indicators are limited by static thresholds or a single dimension and cannot accurately capture the risk accumulation under multi-source interference. However, the evaluation method based on potential field theory constructs a global potential energy field for the roundabout intersection, transforms scenario elements such as road boundaries and other vehicles into repulsive force field sources, and dynamically calculates the field strength and potential energy using parameters such as relative distance, speed, and acceleration. As shown in Figure 2, the two-dimensional potential energy field clearly presents the spatial characteristics of hazard distribution in the intersection (the darker the color, the higher the potential energy and the greater the risk), and the three-dimensional potential energy field further restores the change law of risk with spatial position and motion state. This enables the potential hazards in the complex interaction scenario of the roundabout intersection to be quantified and visualized, breaking through the evaluation limitations of traditional methods and providing more practical technical support for hazard evaluation in complex scenarios such as roundabout intersections. Figure 3 shows the evaluation results of the TTCi index based on the deterministic assessment method and the risk index based on the potential field theory, which are derived from the roundD dataset, when multiple vehicles are driving on a roundabout without road line constraints. It can be seen from Figure 3 that in this scenario, the assessment method based on the potential field theory can better reflect the hazard degree to which the vehicles are exposed. For a comparison of specific cases, please refer to paper [31]. However, there are many parameters that are difficult to directly determine during the evaluation process of this method, thus affecting the wide-spread application of this method. In the future, more unified parameterized evaluation methods need to be studied.

However, when applying the model-driven hazard evaluation methods from autonomous driving scenarios to other automated wheeled vehicles (such as agricultural vehicles, engineering vehicles, and special operation vehicles), the existing methods expose multi-dimensional adaptability deficiencies and technical bottlenecks. On one hand, autonomous driving relies on structured environments like urban roads. In contrast, the field boundaries for agricultural vehicle operations are vague, and the terrain of mines where engineering vehicles operate is complex. Traditional evaluation methods based on fixed road models (such as lane centerlines and clear obstacle coordinates), like the calculation of road boundary field strength in the potential field theory, become ineffective and struggle to capture dynamic risks in unstructured scenarios. On the other hand, the operation modes of non-autonomous-driving wheeled vehicles are unique (such as the back-and-forth movement of agricultural vehicles between ridges and the fixed point loading and unloading of forklifts). There is a lack of reusable natural driving data to train probabilistic models; moreover, they need to balance operational efficiency and safety (for example, a harvester needs to both avoid obstacles and maintain harvesting speed). The traditional single collision-avoidance evaluation objective cannot meet the requirements of multi-task operations, leading to a disconnect between the evaluation results and actual operations. Meanwhile, for engineering vehicles with heavy loads and special operation vehicles with non-standard movement modes, the threshold values of acceleration and distance indicators for deterministic evaluation need to be redefined, and the kinematic assumptions of probabilistic models also need to be reconstructed. The existing methods have insufficient parameter generalization capabilities.

To address these issues, other automated wheeled vehicles can solve the problem of multi-objective risk evaluation in unstructured environments by reconstructing dynamic safety distance models and building their own unique operation mode probability libraries. For example, Ahmed [32] conducted research to generate risk probability cloud maps for seeding operation scenarios through MC simulation, realizing the coordinated evaluation of agricultural vehicles collisions and operation quality. Li [33] introduced a hybrid network model on the basis of probabilistic indicators and constructed a repulsion field associated with load, significantly improving the accuracy of brake risk evaluation for mining trucks in downhill working conditions. In addition, in terms of scenario and data adaptation, historical accident mining was also considered for agricultural vehicles and engineering vehicles. Drawing on the ideas proposed by Byounggap [34], a hazardous event–operation scenario–consequence loss association library has been established to input scenario-specific risk factors into the evaluation model. Combined with the prominent correlation-driven model by Cai [35], semantic segmentation technology has been used to identify dynamic obstacles in unstructured scenarios (such as field weeds and mine falling rocks) and build a field strength correction model through multi-objective general category correlation to adapt to risk calculation in complex environments. In terms of model reconstruction and generalization, the hierarchical decision-making framework proposed by Benrabah [36] can be adopted to design a three-layer assessment model of operation task–vehicle dynamics–environmental risk: the bottom layer embeds vehicle type parameters (such as the heavy-load coefficient of engineering vehicles and the omni-directional movement constraints of AGVs) to dynamically correct the threshold values of indicators such as acceleration and distance; the middle layer introduces coupled evaluation of operational efficiency (such as the miss-seeding rate of agricultural vehicles and the delay cost of forklift loading and unloading) to coordinately quantify safety risks and operational efficiency; the top layer realizes cross-scenario risk comparability through unified measurement of risk entropy [37]. For data scarcity, the real-time monitoring idea proposed by Wang [38] can be referenced to build a data acquisition system for automated wheeled vehicle operations and use small sample reinforcement learning to update model parameters online, alleviating the data dependence problem in rare hazardous scenarios and promoting the evolution of evaluation methods from adaptation to autonomous driving to universality for all automated wheeled vehicles.

Another dimension of the problem focuses on extreme working conditions and multi-objective coupling. Agricultural vehicles are prone to rollover when operating on muddy ridges, and engineering vehicles experience brake thermal decay during long downhill drives. Traditional evaluation methods struggle to capture the risk evolution of such working conditions in real time, and automated wheeled vehicle operations often require multi-objective coordination. A single risk indicator cannot balance conflicting objectives. The solution path can further integrate the hierarchical decision-making and adaptive window mechanism proposed by Jiang [39] to design a module for automated wheeled vehicles that combines condition recognition, dynamic evaluation, and multi-objective decision-making. By identifying extreme working conditions in real time, triggering a specialized evaluation model, and utilizing multi-objective universal correlation, the efficiency index and safety risk are coupled into a two-dimensional evaluation space of risk and efficiency, and the balance strategy is solved through Pareto optimality. Meanwhile, combined with the multi-dimensional risk analysis ideas for complex systems by Gao [40], the risk evaluation dimensions of automated wheeled vehicles are expanded, including secondary risks such as equipment wear and environmental impacts, and a full-life-cycle risk evaluation system is built, realizing the upgrade of evaluation from single driving safety to full-chain operation safety and truly adapting to the diverse application scenarios of non-autonomous driving wheeled vehicles.

2.2. Data-Driven Scenario Hazard Evaluation Method

With the development of big data and machine learning technologies, data-driven methods for evaluating the hazard level of scenarios have gradually emerged. Such methods rely on a large amount of historical scenario data and mine the correlation between data features and hazard levels through machine learning algorithms. The core of this method lies in utilizing multi-source data, including vehicle sensor data (to perceive the surrounding environment), on-board diagnostic system data (to record its own operating status), high-precision map data (to provide road background information), and cloud-based traffic big data (to reflect traffic flow and risk patterns). Typically, data-driven scenario hazard evaluation first requires processing and extracting features, then selecting algorithms to train and optimize the model according to the needs of scenario hazard evaluation and data characteristics, and finally using them to evaluate the vehicle driving risk in actual scenarios.

The key to data-driven evaluation lies in the fusion of multi-source data and the extraction of scene features. Yu [41] proposed a data-driven risk evaluation method enhanced by scene graphs. Yu integrated multiple networks and mechanisms to build an evaluation model, demonstrating the advantage of scene graphs in depicting complex traffic interactions. Lee [42] constructed a collision prediction framework that combines scene-driven and data-driven approaches, achieving synergistic effects of multi-source data. However, uncertainty quantification and risk characterization are key issues in data-driven methods for assessing scene hazard levels. To address this, Kaia [43] focused on epistemic uncertainty in data-driven prediction models for risk evaluation, proposing to quantify uncertainty from the model assumption level and promoting the evolution of risk evaluation towards “probability-uncertainty dual-dimensional characterization”. Amin [44] integrated multi-data monitoring and fault prediction, providing ideas for cross-domain fault risk evaluation. In addition, the interpretability of data-driven models is a core requirement for their engineering application. Maximilian [45] proposed a classification and process optimization framework for data-driven methods, and Hao [46] combined ontology with reinforcement learning to generate safety-critical scenes; these methods have promoted risk evaluation towards “transparent decision support” [47].

Although data-driven hazard evaluation methods have made progress in the field of autonomous vehicles, they face numerous challenges such as complex scenarios, differences in risk objectives, and sparse data when extended to multi-type automated wheeled vehicles. Firstly, there is the dual dilemma of perception and feature extraction in the operating environment. The scenarios where agricultural and engineering vehicles operate differ significantly from the structured road environment of passenger cars, with irregular obstacles and sensor interference (such as dust and electromagnetic waves). This renders traditional perception and feature extraction algorithms ineffective, making it difficult to provide reliable data for hazard evaluation. Secondly, there is the adaptation problem of balancing multi-dimensional risk objectives. Risk evaluation for autonomous vehicles focuses on driver safety, centering on single safety goals like collision prevention and lane keeping. However, other automated wheeled vehicles need to balance driving safety and operational efficiency (e.g., obstacle avoidance and spray coverage for agricultural vehicles, equipment safety and task continuity for engineering vehicles). Models with a single driving safety objective cannot adapt, requiring the establishment of a multi-objective system. Thirdly, there are issues of scarce basic data and poor generalization. Fault data of agricultural and engineering vehicles are scarce and evolve complexly, leading to poor training results and weak generalization of traditional data-driven models, which makes it difficult to identify potential risks.

Therefore, to address the above challenges, cross-domain knowledge transfer and digital twin methods can be adopted to solve the problems of scene adaptation, risk objective differences, and data scarcity. In terms of scene adaptation, Ahmed [48] designed a sensor fusion scheme for agricultural vehicles, which spatiotemporally aligns visual images, LiDAR point clouds, and spray system data; fuses multi-modal features through convolutional neural networks and Transformers; suppresses the interference of crop occlusion and dust noise; and extracts agricultural-specific features such as crop row spacing and obstacle geometric features to improve the obstacle recognition accuracy in complex farmland environments. Cui [49] used the idea of multi-physical field fusion to construct a three-dimensional feature matrix from multi-sensor data, used an autoencoder to eliminate electromagnetic interference noise, and integrated a graph neural network to model the data correlation features of load–vibration–leakage to solve the problem of data distortion in industrial and mining environments. For the challenge of risk objective differences, Zhong [50] designed an obstacle avoidance-spraying dual-objective function, introduced a non-dominated sorting genetic algorithm, took collision probability and spraying uniformity as the core optimization objectives, and trained a Pareto optimal solution set through historical farmland operation data to achieve the dynamic balance between obstacle avoidance strategies and operation quality. Peixoto [51] combined physical models with data-driven models, took leakage risk probability and transportation task interruption cost as dual objectives, and dynamically adjusted the weight coefficients of safety redundancy and task continuity through reinforcement learning. Wang [52] adopted digital twin technology to construct a multi-objective evaluation model of fault risk-operation and maintenance efficiency for the suspension system, used the twin to simulate the task delay losses in different fault scenarios, incorporated the suspension failure probability and operation and maintenance interruption duration into the risk function, and achieved the quantitative balance of multiple objectives through gradient boosting decision trees. In response to the problem of data scarcity, Ankrah [53] adopted a domain adaptation algorithm to transfer the fault diagnosis model of autonomous vehicles to special engineering vehicles. By minimizing the feature distribution difference between the source domain and the target domain, the model was pre-trained with a small amount of fault data from special engineering vehicles to solve the problem of scarce suspension failure data. Mehta [54] designed a lightweight model for few-shot learning and improved the model’s generalization ability for scarce fault modes through meta-learning algorithms. Figure 4 illustrates the migration roadmap for scenario hazard evaluation methods.

3. Hazardous Scenario Generation and Generalization

Although model-driven and data-driven methods can be used to assess the hazard level of autonomous driving test scenarios, due to the limitations of data sampling technology and safety, the amount of natural driving data is insufficient to meet the needs of scenario virtual testing. Moreover, most of the data are collected under safe driving conditions. The former can reconstruct a large number of test scenarios through multi-source data fusion (such as a clustering method, random sampling method, etc.) [55,56,57,58,59] and generative modeling (such as an ontology method, rule-based method, etc.) [60,61,62,63,64]. However, the latter leads to the sparsity of hazardous scenarios, and it is difficult to create more hazardous scenarios through scenario reconstruction technology. Therefore, the focus of this section is on solving the problem of the sparsity of hazardous scenarios in natural driving data. Through scenario generation and scenario generalization technologies, more hazardous scenarios are generated, thus building a rich and diverse test scenario library. For this problem, this paper classifies scenario generation methods into hazard inversion methods based on scenario elements and data-driven intelligent generation methods. Among them, the hazard inversion method based on scenario elements refers to deeply analyzing the existing hazardous scenario data or historical driving behavior data, extracting the traffic element information (such as weather, illumination, the driving behavior of other vehicles, etc.) that has a significant impact on the decision-making of autonomous vehicles. Then, the hazardous traffic elements are reproduced and combined through a simulator to generate hazardous scenarios. The data-driven intelligent generation method refers to relying on intelligent algorithm models such as reinforcement learning (RL) and generative adversarial networks (GANs). It does not need to strictly rely on real hazardous scenario data. Instead, it actively generates diverse hazardous scenarios by simulating risk evolution, parameter variation, and other means.

3.1. Hazard Inversion Method Based on Scenario Elements

The hazard inversion method based on scenario elements, with its accurate restoration ability of real hazardous scenarios, has become a key technical path connecting hazardous scenario data and test verification. Gambi [65] proposed a hazard scenario generator based on natural language processing. It extracts the influencing factors leading to hazardous scenarios from the traffic accident data provided by the National Highway Traffic Safety Administration [66], including weather, illumination, road conditions, vehicle status, etc. Then, a mapping relationship between text descriptions and scenario elements is established to realize the transformation from unstructured accident reports to structured scenario elements. Bashetty [67] focused on the video data of hazardous scenarios and proposed a multi-module trajectory extraction framework. It integrates target detection, multi-target tracking, lane line recognition, and camera calibration technologies to restore the movement trajectories of vehicles and pedestrians from the video of hazardous scenarios, solving the problem of the quantitative extraction of dynamic behavior elements. In terms of the behavior inversion of pedestrians and vulnerable road participants, Tian [68] deeply mined the hazardous trajectory data, identified typical hazardous interaction behaviors such as vehicles forcibly cutting in, vehicles driving in the opposite direction, and pedestrians crossing the road illegally. A correlation model between behavior and risk was built, thus transforming discrete hazardous behaviors into dynamic events that can be injected into scenarios. This research not only reproduced the core conflicts in hazardous scenarios but also simulated the accident evolution process by continuously injecting hazardous behaviors. Finally, thirteen types of potential failure modes of the autonomous driving system were discovered in the test, verifying the key role of behavior element extraction in building scenario hazards. Zhang [69] proposed a new panoramic segmentation model. First, the spatial layout and target distribution of real hazardous scenarios were restored through a panoramic segmentation model. Then, a mutation algorithm was used to perturb scenario parameters (such as vehicle speed, distance, behavior timing) to generate multi-dimensional variant scenarios. This method breaks through the limitation of a single hazardous scenario, expanding one type of basic hazardous scenario into three types of differentiated hazardous scenarios, and improving the test coverage. Wang [70] converted the hazardous scenario data into a scenario description language through natural language processing technology. This automatic conversion between natural language and scenario representation realizes the rapid reproduction and batch generation of hazardous scenarios, solving the problem of the standardization of scenario descriptions. Nguyen [71] further expanded the scenario language input. Through image processing technologies such as contour analysis and semantic segmentation, the theoretical analysis diagram was converted into a three dimensional test scenario. Among the 50 hazardous scenarios generated by this method, 46 cases were completely consistent with the theoretical analysis diagram, and the average generation time was only 0.27 s, verifying the high efficiency of scenario language input in scenario reconstruction. In addition, the current hazard inversion method based on scenario elements is developing towards the direction of cross-modal data fusion. Zhang [72] integrated multi-source data such as hazardous scenario videos, radar point clouds, and road friction coefficients to build a multi-modal scenario reconstruction platform, realizing the full process fusion from the extraction of key video information to the calibration of dynamic parameters. This fusion can not only improve the physical authenticity of the scenario but also reproduce the accident evolution under complex environments such as road skidding and sudden illumination change.

3.2. Data-Driven Intelligent Generation Method

Although the hazard inversion method based on scenario elements can accurately restore real accident scenarios, it is limited by the dependence on real hazard data and the singularity of scenario evolution logic, making it difficult to cover long-tail risks and potential hazard scenarios that have not occurred yet. Therefore, the data-driven intelligent generation method has become a key direction to break through the bottleneck: it takes algorithm models as the core, gets rid of the strong dependence on real hazard scenarios, and, through hazard-oriented learning, generates new hazard scenarios that conform to traffic rules, making up for the scarcity of hazard scenarios in natural driving data.

RL, leveraging dynamic gaming between agents and the environment, has become a core technology for mining hazardous scenarios. Liu [73] constructed a reinforcement learning scenario-editing framework that dynamically inserts traffic participants and iteratively optimizes parameters in combination with risk-oriented reward functions, increasing the coverage of hazardous scenarios by 41% and verifying its ability to mine long-tail risks. Priisalu [74] modeled the generation of hazardous scenarios as a non-cooperative game, designed differentiated reward functions, and after 5000 rounds of training, the critical collision rate in the generated scenarios reached 37%, filling the gap in simulating heterogeneous decision conflicts. In addition, common data-driven generation models include GANs and Variational Autoencoders (AVEs), both of which generate new data through fitting data distributions and random sampling. Krajewski [75] designed a trajectory generative adversarial network and a variational autoencoder; GAN can generate realistic lane-changing scenarios without annotation, while VAE makes up for the mode collapse risk of GAN. Subsequent improved models [76] have enhanced trajectory smoothness but still face the challenge of capturing the temporal sequence of multi-vehicle interactions. Zhang [77] proposed a multi-vehicle trajectory generator that precisely controls interaction behaviors through a bidirectional encoder and a multi-branch decoder; Demetriou [78] designed a framework combining deep learning and a GAN, supporting the generation of variable-length trajectories and solving the problem of a fixed trajectory length in traditional models. Considering that the GAN and VAE struggle with the problem of temporal interaction, Tan [79] used an autoregressive model to characterize the dependence features between historical data, and by recursively learning the conditional probability of historical trajectories, Tan generated hazard scenarios that conform to the laws of real traffic flow. Li [80] used an autoregressive model to predict subsequent lane-changing decisions based on historical data such as the steering angle and speed difference of previous lane-changing trajectories, generating hazard scenarios of rear-end collisions caused by continuous illegal lane changes, making up for the deficiency in GAN and VAE in modeling long-temporal interactions. The autoregressive model has a sampling efficiency bottleneck, and its synthesis calculation length is positively correlated with the data dimension, resulting in high time consumption for scenario generation. For this reason, Feng [81] utilized the advantages of the flow-based model in reversibly transforming the data distribution and parallel sampling. By learning the joint distribution of vehicle speed, braking distance, and friction coefficient in real hazard scenarios, it generates high-precision hazard scenarios and improves training efficiency. Similarly, Gu [82] used a diffusion model to gradually remove scenario noise to approximate the real data distribution. Gu utilized the efficient sampling and special structure data processing capabilities of the diffusion model to generate hazard scenario samples with higher realism.

Table 2. Common data-driven intelligent generation methods.

Method Type	Core Principle	Advantages	Disadvantages	Application Scenarios
RL	Guides agent to explore hazard scenarios through risk reward mechanisms.	Proactively discovers hazardous scenarios.	Requires precise environment modeling.	Complex interactive scenarios (multi- vehicle games).
GAN	Generator and discriminator adversarial training to learn real scenario distributions.	Proactively discovers hazardous scenarios.	Difficult to maintain long-term temporal consistency.	Single-vehicle behavior simulation (lane change/ overtaking).
VAE	Encoder–decoder framework learns latent space distribution of scenarios.	Generates smooth and continuous scenarios.	Weak in multi-agent interaction modeling.	Basic scenario element combinations.
Autoreg- ressive Model	Recursively predicts future states based on historical scenario elements.	Maintains long-term temporal rationality.	Low generation efficiency.	Continuous hazardous event evolution.
Flow Model	Establishes mapping between distributions and scenarios through invertible transformations.	Precisely controls generated scenario attributes.	Sensitive to training data quality.	High-fidelity physical scenarios.
Diffusion Model	Generates high-quality scenarios through gradual denoising process.	Produces scenarios with rich details.	Poor controllability.	Complex environment detail reconstruction.

summarizes the principles, advantages, and disadvantages of various generative methods, and also lists some representative application scenarios.

Given that the hazard inversion method based on scenario elements relies on real data and has a single scenario evolution logic, it struggles to cover long-tail and potential hazard scenarios. In data-driven intelligent generation methods, reinforcement learning is limited by prior knowledge, and various generation models have shortcomings in terms of temporal interaction and efficiency. In the future, the two methods can be integrated, using the real scenario foundation of the inversion method to optimize the generation model, while leveraging the generation model to expand the diversity of inverted scenarios, and by improving the model’s ability to capture complex interactions and its adaptability to special scenarios.

However, directly applying the two types of scenario generation and generalization methods to agricultural vehicles, engineering vehicles, etc., will face issues of scenario logic adaptability and data characteristic compatibility. Existing frameworks for scenario generation and generalization struggle to cover the unique risk factors of these vehicles (e.g., construction machinery load, cargo stability of logistics vehicles, etc.), and unstructured data leads to parsing deviations. In data-driven generation, reinforcement learning priors are incompatible with their operational logic. GAN/VAE and other models are prone to mode collapse due to sample scarcity, making it difficult to capture multi-factor coupled risks. These contradictions stem from systematic scenario differences and require breakthroughs in the reconstruction of factor systems, data extraction, and model customization.

To solve the problems of unstructured hazard scenario data and element extraction deviations in other automated wheeled vehicles, scholars can draw on the idea of the multi-modal data fusion method proposed by Yao [83]. For agricultural vehicles or engineering machinery, multi-source heterogeneous data such as 3D point clouds of the operation environment, equipment operation parameters, and operation videos, are integrated, and the exclusive scenario element library is reconstructed through multi-modal feature encoding. In addition, scholars can combine the scenario generation classification frameworks proposed by Ding [84] and Shen [85] to map the extracted discrete elements to multiple dimensions such as the physical environment, equipment behavior, and risk consequences, as well as reconstruct the exclusive scenario element library associated with load, terrain, behavior, risk, etc. For example, for agricultural vehicles, an association model is constructed for the pesticide residue threshold field, wind speed spray operation trajectory, and crop phytotoxicity risk to solve the problems of semantic deviation and feature loss in traditional element analysis, and provide accurate input that conforms to the operation logic of automated wheeled vehicles for scenario inversion. For the problems of poor adaptability of scenario generation models to the scenarios of other automated wheeled vehicles and sample scarcity, we draw on the GAN improvement strategy proposed by De [86]. For agricultural vehicles or special engineering vehicles, small-sample learning and domain-adapted generative models are adopted. When using RL, an industry-exclusive simulation environment is constructed (similar to the scenario construction framework proposed by Chang [87]), the operation rules of agricultural vehicles and physical constraints of special working conditions are embedded, and agents are trained to learn domain-specific behaviors. At the same time, scholars can combine the feature extraction and model optimization methods proposed by Tong [88]. For the multi-factor coupling risks of such automated wheeled vehicles, multi-physical field simulation and feature disentanglement technologies are introduced so that models such as RL, GAN, and VAE can capture the evolution laws of complex risks. We can also use the scenario combination and generalization training strategy proposed by Li [89] to enrich the distribution of hazard scenario samples, break through the limitations of traditional models in generating multi-factor coupling risks, and improve the adaptability of models to the scenarios of automated wheeled vehicles and the ability to generate hazard scenarios.

4. Acceleration Evaluation Method

In scenario virtual testing, the acceleration evaluation method screens out high-value test scenarios through optimization methods. Figure 5 shows the security evaluation flowchart based on scenario virtual testing optimization methods. Firstly, a module of the test vehicle is determined as the tested object, which can be the perception module, planning and decision-making module, or control module. Secondly, according to the current test requirements, the scenario virtual testing optimization method is used to derive specific concrete scenarios (see Figure 6a,b, such as the cut-in scenario). The functional model of the vehicle is tested in these concrete scenarios via a simulation platform (see Figure 6c,d). Finally, the safety evaluation of the vehicle is conducted based on the test results. Batsch [90] systematically reviewed the papers in the field of scenario virtual testing optimization and categorized existing methods into two types: coverage-oriented optimization methods and hazard rate-oriented optimization methods. The former aim to achieve full coverage of the entire feature space of the scenario library through optimized sampling, while the latter focus on identifying and reinforcing high-risk scenarios to improve testing efficiency. It is worth noting that regardless of the optimization method used, it is essential to ensure statistical consistency between the test scenarios and the original scenario library to avoid sampling bias. Therefore, the core scientific problem of scenario virtual testing optimization evaluation lies in how to construct an acceleration evaluation method that simultaneously satisfies coverage completeness, testing efficiency, and statistical unbiasedness in the high-dimensional sparse scenario feature space. Solving this problem requires in-depth exploration of the topological structure of the scenario feature space and the development of optimization algorithms with theoretical guarantees.

4.1. Coverage-Oriented Acceleration Evaluation Method

The safety verification of autonomous driving systems faces challenges from the complexity and uncertainty of real traffic scenarios, and solving the long-tail problem urgently requires constructing test scenarios with high coverage and complexity. In this context, the Combinatorial Test (CT) method, derived from the software engineering field, has become an important means for autonomous driving scenario generation due to its core logic of discretizing continuous scenario elements and constructing effective test scenarios by combining them according to importance indices [91].

Schuldt [92] took the lead in proposing a test case generation method based on a four-level model. Schuldt divided scenario elements into four levels, road network, static elements, dynamic elements, and weather, which helps control the number of scenarios while ensuring coverage. Zheng [93], based on the Pairwise Independent Combinatorial Testing (PICT) tool, calculated test scenarios with minimum coverage to reduce redundant combinations. Shu [94] combined PICT with manually formulated scenario screening rules to eliminate unreasonable scenarios from a large number of generated ones and constructed a basic test scenario group for high-level autonomous vehicles. Gao [95] proposed the Combinatorial Testing Based on Complexity (CTBC) method. Gao combined the Test Matrix (TM) with combinatorial testing and prioritized high-complexity combinations to enhance test authenticity. Although combinatorial testing has achieved remarkable results in improving scenario coverage and complexity, the contradiction between the number of scenarios and test efficiency has always existed. To solve this problem, scholars have tried to combine ontology [96] with combinatorial testing. Li [97] and Wotawa [98] generated scenarios through semantic reasoning. However, this combination has obvious drawbacks: due to the relatively fixed rules of ontology, the generated scenarios are prone to homogenization, and some combinations may not conform to the actual situation. Therefore, in practical applications, manual secondary screening is often required to eliminate invalid or unreasonable test cases. In the future, we can explore the integration of dynamic ontology rules and adaptive combination strategies, update constraint conditions in real time by introducing domain knowledge graphs, and optimize combination priorities with reinforcement learning to reduce manual intervention while improving the diversity and adaptability of scenario generation.

When the combinatorial testing method is applied to agricultural vehicles, engineering vehicles, and other automated wheeled vehicles, the challenges are more severe. Unlike structured road traffic scenarios, the operating environment of off-highway vehicles is highly specific and complex [99,100]. Agricultural scenarios involve multiple dynamic elements such as soil texture and crop growth, as well as their coupling relationships [101]. Engineering vehicles need to consider factors such as heavy-load transportation and collaborative operation of multiple equipment [102]. These scenario elements are not only numerous and high-dimensional but they also lack unified classification standards and quantification methods, making the element discretization and combination logic relied on by traditional combinatorial testing difficult to directly adapt. In addition, the operation scenarios of non-highway vehicles lack standardized rules and historical data accumulation. Therefore, combinatorial testing methods based on expert knowledge or fixed rules struggle to comprehensively cover the diversity and complexity of real operation scenarios, and are prone to missing key risk scenarios.

To address the above problems, solutions can be explored from multiple dimensions. At the technical integration level, deeply combining domain expertise to construct customized scenario models is crucial. Taking agricultural vehicles testing as an example, soil mechanics models, crop growth dynamics models with combinatorial testing methods, can be integrated, refining the classification and combination rules of scenario elements according to parameters such as soil cohesion and crop root distribution density. By establishing a mathematical model of soil–vehicle–crop interaction, the operation status of agricultural vehicles under different parameter combinations is simulated to achieve precise scenario generation [103]. In engineering vehicle testing, multi-body dynamic models and finite element analysis methods are introduced, considering factors such as vehicle structural strength and mechanical arm kinematic constraints, to construct a combinatorial testing framework covering scenarios such as equipment collaborative operations and terrain adaptability [104]. At the data-driven level, artificial intelligence and big data technologies are used to dynamically optimize combination strategies. By installing sensors on agricultural vehicles and engineering vehicles to collect various data in real-time during operation, including soil parameters, equipment operation status, environmental information, etc., a large-scale operation scenario database is constructed. Based on these data, machine learning models such as deep neural networks and long short-term memory networks are trained to automatically identify high-frequency scenario patterns and key influencing factors [105,106]. During the combinatorial testing process, the element weights and combination methods are dynamically adjusted according to the model prediction results, and high-risk and frequently occurring scenarios are preferentially generated to improve the pertinence and effectiveness of testing. At the collaborative sharing level, a cross-domain collaborative scenario database and standard system are established. Jointly with experts from multiple fields such as agriculture, engineering, and computer science, unified description specifications and classification standards are formulated for non-highway vehicle operation scenarios, integrating operation scenario data of different types of non-highway vehicles. Through an open and shared database platform, data accumulation and updating are realized, providing rich data support for combinatorial testing. At the same time, industry–university–research cooperation is encouraged to jointly develop universal scenario generation methods suitable for different types of non-highway vehicles, improving the versatility and adaptability of combinatorial testing technologies in the non-highway field.

4.2. Hazard Rate-Oriented Acceleration Evaluation Method

The hazard rate-oriented acceleration evaluation method mainly draws on rare event simulation methods, including the Markov Chain Monte Carlo (MCMC), splitting, subset simulation, and IS methods, among others. The IS method, a small-probability event simulation method based on large deviation probability theory, relies on the concept of variance reduction to increase the occurrence probability of small-probability boundary events by expanding the variance of the original distribution while ensuring distributional unbiasedness. Zhao [107] increased the probability of extreme scenarios by altering the sampling measure, ensuring unbiased estimation of sample parameters. Focusing on following, lane-changing, and cut-in scenarios, Zhao defined extreme scenario features for conflict, collision, and casualty events, constructed statistical distributions for scenario parameters, and improved convergence efficiency and simulation speed through importance sampling functions. The key challenge in IS lies in solving the importance sampling function. Xu [108] proposed an improved IS technique for generating overtaking scenarios in autonomous driving, solving the importance sampling function via genetic algorithms. Simulations based on Shanghai natural driving data showed significant improvements in testing efficiency and result accuracy. To expand the application scope of such acceleration evaluation methods, Huang [109] extended the acceleration evaluation of intelligent vehicles from independent multivariate distributions to joint distributions, addressing the insufficiency of single-parameter distributions in describing extreme scenario features. As parametric distributions struggle to accurately fit high-dimensional parameter spaces with spatiotemporal correlations, Zhang [110] introduced normalizing flows into the IS method to target the joint distribution learning of spatiotemporal parameters in rare events.

Table 3. Comparison of testing number based on common evaluation indicators in the cut-in scenario of the HighD dataset.

Method	Conflict	Crash	Injury
Zhao [107]	1748	135,388	127,959
Xu [108]	862	66,123	79,744
Huang [109]	613	56,872	46,761
Zhang [110]	298	21,916	29,259

shows the testing number of several hazard rate-oriented acceleration evaluation methods based on common evaluation indicators in the cut-in scenario of the HighD dataset.

Considering that fragmentary scenario tests are independent and disregard the connection between functional scenarios, evaluation results cannot represent the overall safety performance of autonomous vehicles during continuous driving. The demand for continuous scenario testing has given rise to new testing methodologies, with two research directions emerging in the literature: One is directly designing continuous driving tasks, typically by specifying a driving area or start/end points, where the tested host vehicle adaptively selects interactive behaviors based on the background vehicles’ actions. Wang [111] addressed the black-box problem of scenario testing in autonomous driving interactions using Level-k game theory and social value orientation to characterize interactive behaviors, training background vehicle behaviors via RL for adaptive test scenario derivation based on Gaussian process regression. Tuncali [112] proposed a rapidly exploring random tree for adversarial testing to search for safety boundaries in path planning, with the random sampling of background vehicle behaviors during testing. Kim [113] used hardware-in-the-loop simulation to model approximately 2.5 km of urban roads for driving tests. The other approach deduces continuous testing from fragmentary tests. Feng [114] divided natural driving processes into 33 types of speed-change behaviors through statistical analysis, described driving state changes using Markov chains, and adjusted the driving environment based on adversarial attack techniques to increase the probability of hazardous scenarios. In conclusion, a comparison of several common hazard rate-oriented acceleration evaluation methods is shown in

Table 4. Comparison of several common hazard rate-oriented acceleration evaluation methods.

Method	Advantages	Disadvantages
MC	Simple implementation, parallelizable	Inefficient for rare events, high variance
MCMC	Handles complex distributions, theoretical guarantees	Slow convergence, Sensitive to initialization
Splitting	Efficient for dynamic systems, good parallelism	Difficult threshold setting, P degeneracy
Subset Simulation	Effective for extremely rare events, avoids direct sampling	Computationally intensive, sensitive to level selection
IS	Fast convergence, variance reduction	Requires good proposal distribution, degrades in high dimensions
RL	Discovers unknown hazards, adaptive exploration	High training cost, reward design sensitivity

. With the IS simulation method, solving the IS function is difficult, and parametric distributions struggle to fit high-dimensional spatiotemporally correlated parameter spaces. In continuous scenario testing, the IS method is insufficient in modeling complex interactions, and the deduction approach based on fragmented scenario testing cannot fully reflect the overall safety of real continuous driving. Therefore, future efforts can focus on optimizing the algorithm for solving the importance sampling function and combining deep learning to enhance the ability to fit the joint distribution of high-dimensional spatiotemporal parameters. In addition, researchers can improve continuous scenario testing methods by integrating game theory and reinforcement learning to strengthen the modeling of complex interaction behaviors, while enhancing the authenticity and completeness of deducing fragmented scenarios into continuous ones, so as to more comprehensively evaluate the overall safety of vehicles.

However, in agricultural vehicle operation scenarios, factors such as soil texture, crop growth stages, and terrain slopes are highly coupled, making it difficult for the IS method used in traditional autonomous vehicles to identify low-probability hazardous events under multi-dimensional coupling. The heavy-load dynamic changes and equipment collaboration risks of mining engineering vehicles differ greatly from road interaction logic, and existing rare-event simulation methods are not suitable for unstructured environments. Moreover, agricultural and engineering vehicles lack a standardized data collection system, resulting in fewer hazardous scenario datasets with scattered and highly heterogeneous data, which makes it difficult to train models relying on massive data as in the field of autonomous driving.

To address these issues, scholars have conducted research from two aspects: improving the IS method mechanism and optimizing data utilization strategies. They have broken through traditional method limitations by adaptively adjusting IS weights. For the collaborative impact of soil texture and terrain slope on sinkage risks in agricultural vehicles operations, Dilara [115] proposed an IS weight optimization method based on dynamic Bayesian networks, incorporating parameters such as soil water content—and agricultural vehicles, ground contact pressure—into network nodes to calculate hazard probabilities in real time for parameter combinations and to dynamically adjust sampling weights to prioritize high-risk scenario collection. For heavy-load rollover risks in engineering vehicle mine operations, Oszczypala [116] introduced multivariate Copula functions to construct joint distributions of vehicle dynamic parameters (speed, center of gravity offset) and environmental parameters (road slope, friction coefficient), characterizing non-linear coupling relationships via Copula functions to modify the importance functions of IS, effectively enhancing the sampling efficiency for hazardous scenarios under complex working conditions. In addressing weak data foundations, research has focused on cross-domain data augmentation and transfer learning to optimize IS algorithms. Tang [117] proposed a cross-domain data generation method based on generative adversarial networks, using autonomous driving road scenario data as the source domain, to simulate agricultural vehicles’ farmland operation scenario data through generator training, while using discriminators to distinguish between the real and generated data. The resulting pseudo-data expands the agricultural vehicle scenario library, alleviating data scarcity and providing more sufficient training samples for IS sampling. Li [118] adopted domain adaptation techniques to transfer pre-trained parameters of autonomous driving IS sampling models to the engineering vehicle domain, fine-tuning them with limited engineering vehicle construction data. Knowledge distillation is used to extract sampling strategies from source domain models, preserving the advantages of autonomous driving IS algorithms while rapidly adapting to engineering vehicle scenario features to improve the sampling accuracy of hazardous scenarios.

5. Conclusions

This paper provides a comprehensive review of the key technologies and methodological frameworks for the virtual testing of automated wheeled vehicles in various scenarios, as well as the challenges in transferring these technologies from road to non-road environments. Examining three core techniques—scenario risk evaluation, hazardous scenario generation and generalization, and accelerated evaluation—revealed that while a relatively mature and symmetric evaluation framework has been established for road-based autonomous vehicles, significant common challenges arise when adapting these methods to non-road automated wheeled vehicles such as agricultural and construction machinery. These challenges include the lack of unified risk evaluation criteria, multi-dimensional coupling of risk objectives, and insufficient foundational data. To address these issues, this paper proposes a testing adaptation approach driven by multi-physics simulation and digital twin technologies, focusing on three key aspects: first, deepening the integration of multi-disciplinary models with virtual testing technologies to establish a data acquisition system and a quantitative risk indicator system suitable for non-road environments; second, developing more intelligent scenario generation models that learn from both symmetric risk evolution patterns and asymmetric scenario details, advancing few-shot scenario generation techniques based on transfer learning and meta-learning to achieve a breakthrough from known risk scenarios to unknown risk prediction; third, incorporating the impact of operational environments and efficiency into safety evaluation and constructing a multi-objective collaborative acceleration evaluation system combining RL and knowledge distillation. These elements form a holistic framework for the virtual testing of automated wheeled vehicles. Future work should focus on algorithm development tailored to specific scenarios, gradually advancing the transfer framework from theoretical conception to practical technical validation. Through continued exploration in these directions, virtual scenario testing can transition from a “road autonomous vehicle-specific” paradigm to a “universal automated wheeled vehicle general” paradigm, providing a solid technical foundation for the safe, reliable, and efficient application of automated wheeled vehicles across all domains.