A Survey on Testbench‐Based Vehicle‐in‐the‐Loop Simulation Testing for Autonomous Vehicles: Architecture, Principle, and Equipment

Autonomous vehicles (AVs) must be thoroughly tested to ensure safety and reliability before marketing. Simulation‐based testing has gained widespread recognition as the essential approach for AV testing by providing sufficient testing scenarios in the virtual environment. Vehicle‐in‐the‐loop (VIL) simulation has the ability to perform comprehensive tests and validations for the AVs’ overall behaviors while keeping significant testing accuracy and efficiency through the combination of the virtual scenarios and the physical AV. This article provides an overview of representative studies on testbench‐based VIL simulation testing for AVs, mainly focusing on utilizing testbenches to simulate realistic road conditions, and using physical signal stimulation methods and related equipment to generate sensors’ physical signals. This article first summarizes current AV testing studies, identifying existing issues and flaws of the state‐of‐the‐art methods and tools. Afterward, the testbench‐based VIL is addressed around architecture, principles, advantages, and characteristics. Then, the road condition simulation and the sensor physical signal generation in VIL are discussed in depth from structure, principle, and corresponding advanced equipment. Finally, research gaps between cutting‐edge technologies and AV testing applications in industrialization are identified to facilitate future research in this direction.


Introduction
The autonomous vehicles (AVs) is a deep integration of the autonomous driving (AD) system and a physical automobile. [1]igure 1 demonstrates the architecture of an AV, as a complicated cyber-physical system, [2] its subsystems (including localization, perception, decision-making, planning, control, etc.) couple and cooperate with each other considerably. [3,4]The extremely complex inner mechanism makes the AV's behaviors highly random and uninterpretable. [5]][9][10] Simulation-based testing has been broadly recognized as indispensable, [11] since it has the ability to provide sufficient realistic scenarios in a virtual world for the AV-under-test (AVUT) to carry out safety validation, [12] simultaneously maintaining a considerable testing efficiency. [13]Although pure virtual simulation testing has the ability to flexibly generate high-fidelity scenarios according to specific requirements, [14] the related testing results are still not accurate enough, because the virtual environment is completely constructed according to various mathematical models. [15]ehicle-in-the-loop (VIL) simulation testing has the ability to refine testing accuracy and efficiency simultaneously by combining the virtual scenarios and the whole physical AV, and it has been arousing more interest in academia and industry. [16]In VIL simulation testing, the AVUT is considered the testing object connected to the virtual scenarios to carry out vehicle-level testing. [13]The AVUT's comprehensive performance can be evaluated by continuously observing the AVUT's overall driving behaviors and the state information of each subsystem. [1]IL testing is generally performed in an enclosed open test field or on an indoor testbench. [16]Testbench-based VIL is more controllable, cost-efficient, and safer, which is applied more extensively. [17]As the testbench-based VIL is performed in a laboratory environment, and the whole AVUT is integrated into the virtual scenarios, it is necessary to provide drivable roads and actual sensor data for the physical AVUT, virtual scenario construction, road condition simulation, and sensor physical signal generations consequently become three essential technologies in testbench-based VIL testing.As shown in Figure 2, the virtual scenario provides requisite semantic information for the AVUT, that is, the virtual road conditions and the virtual perception data.The road condition (mostly involving the adhesion coefficient, the slope, and the curvature) mainly provides driving conditions for the AVUT, [18] which is used for simulating dynamics driving behaviors (longitudinal/lateral motions, steering motions, etc.) and evaluating the AVUT's automobile performance (such as power, resistance, [19] fuel consumption, [20,21] emission, [22,23] turning radius, slip rate, [24] etc.);, the perception data of various sensors (e.g., camera, radar, LiDAR, GNSS, etc.) is the primary input of the AD system, [25] which is used for evaluating the AVUT's intelligence performance.Since the whole AVUT is embedded into the testing loop as the testing object, it is necessary to convert the two types of semantic information into the close-to-reality road conditions and the realistic sensor

Testbench-Based VIL Simulation Testing
perception data through relevant equipment, that is, testbenches and sensor physical signal generators, to ensure the accuracy of the testing results.
All kinds of utilized testing equipment in testbench-based VIL have their unique mechanical designs and complex operating principles, causing the reality of simulated road conditions and the precision of the generated sensor physical signal to be distinct, also making the accuracy of testing results dissimilar.In addition, the utilized equipment generally has significantly high manufacturing costs.Consequently, selecting applicable testing equipment in testbench-based VIL is salutary to improve the accuracy and efficiency of AV testing.
There have been very few literature reviews in the study of VIL simulation testing for AVs at present.With the emerging studies on testbench-based VIL testing, a comprehensive survey that assists researchers in choosing appropriate sensor physical signal generators and testbenches to perform VIL testing, and further helps identify future research directions, is extensively demanded.To address the need, this article reviews representative studies of testbench-based VIL simulation for AV testing.The reviewed literature is categorized based on the essential technologies in testbench-based VIL.Rather than merely enumerating relevant studies, this article concentrates on demonstrating the characteristics of these categories through the selected representative studies, thereby revealing connections and relationships among related studies.
The major contributions of this article are as follows: 1) discussing the architecture, principle, advantages, and characteristics of testbench-based VIL, field-based VIL, and sensor physical signal generation; 2) focusing on investigating the mechanical designs, operating principles, road condition simulation characteristics, and potential improvements of different kinds of testbench (chassis dynamometer and powertrain dynamometer); 3) analyzing the process and relevant state-of-the-art equipment of camera and radar physical signal generation; 4) and facilitating researchers in this area to obtain a more focused view of related implementation approaches in testbench-based VIL.
This article is organized as follows.Section 2 briefly summarizes the current state of AV testing and identifies existing issues and challenges of the cutting-edge testing methods and tools, laying the fundament for the subsequent studies on testbenchbased VIL and related contents.In Section 3, the architecture, principle, advantages, and characteristics of testbench-based VIL are discussed.Section 4 analyzes two kinds of road condition simulation (field-based VIL and testbench-based VIL) and mainly investigates the mechanical design, operating principle, road condition simulation characteristics, and potential improvements of various testbenches.Section 5 analyzes the principle of sensor physical signal generation methods and concludes relevant equipment of camera and radar physical signal generators.Section 6 concludes this article and proposes improvement directions for future studies.

Overview of Autonomous Vehicle Testing
The foundation of testbench-based VIL simulation testing still lies in the performance validation of road vehicles.Based on the automobile functional safety validation and verification procedure, this section comprehensively discusses the related characteristics of testing methods and simulation tools and indicates the importance of simulation tools for AV testing.Finally the challenges and improvements of existing prevailing simulation tools are identified and the fundament for the subsequent studies on testbench-based VIL and related contents is laid.

Safety Validation of AVs
As the functional structures of AVs are complicated, [12] continuous testing is essential for identifying failures and hazards during the development lifecycle of an AV product. [26,27]According to ISO 26 262 (2018), [28] Figure 3 illustrates the procedure of vehicle functional safety management during the development lifecycle.Based on the requirement specifications and the safety concepts for the AVUT, the development item is divided into five stages according to the functional complexity from high to low: vehicle, system, subsystem, functions and modules, hardware and software. [29]Contrarily, the implementation and validation process of functions are performed from the lower stage to the higher stage.Several lower-stage items integrate a higherstage item and finally construct an AV. [30]As a result, sufficient tests, verifications, and validations for all stages are necessary.Specifically, unit, integration, system, and acceptance testing are gradually performed for different stages. [31]After completing the safety validation, the AV product will enter the stage of production and operation.

Testing Methods of Validation
Although the five stages are initially divided for the functional safety validation of road vehicles, they have been extended to test safety, functionalities, and other performances of AVs and advanced driver assistance systems (ADAS) due to the comprehensive and efficient verification and validation procedure. [32]ome testing methods are necessary to carry out test cases during verification and validation.There are generally three methods in AV testing: on-road testing, [33] closed-facility testing, [34] and simulation testing; [35] each has its own advantages and disadvantages.Corresponding testing methods are adopted according to the different needs of testing.
For the validation of AV's safety and functionalities, Feng et al. indicate that safety, economic and time costs, and efficiency are the keys to the AV testing problem. [36,37]In terms of simulation testing, Yan et al. highlight that the fidelity of the constructed virtual environment (especially the driving environment) determines the authenticity of testing results. [38]Regarding generating simulation scenarios, Ding et al. point out five important indicators: fidelity, efficiency, diversity, transferability, and controllability. [39]ased on the multiple indicators discussed above, combined with the related requirements of AV development and testing, Table 1 lists the characteristics of these three methods from six aspects: safety, cost, authenticity, controllability, diversity, and efficiency.Please note that the requirements of different testing items are dissimilar, and the corresponding evaluation criteria for the six indicators are also various.Therefore, Table 1 merely qualitatively compares the three methods from the six indicators and does not list the quantitative evaluation criteria for each indicator.
Among the six indicators, this safety mainly refers to the likelihood of injury to the operators and surrounding traffic during the test.Cost is the total expenditure of money, time, and labor in testing (including preparation, execution, and postmaintenance).Authenticity describes the testing scenario's fidelity and also means the testing results' accuracy.Controllability is the ability to flexibly configure scenario elements for generating custom-designed scenarios or reproducing specific scenarios.
Diversity is the abundance of testing scenario's variety and quantity.Efficiency refers to the entire test time spent to achieve required testing accuracy, it is primarily determined by the controllability and diversity of the testing scenarios.
As shown in Table 1, using either testing method alone is unable to satisfy the requirements of AV development and testing fully.In practice, many institutions adopt a continuousfeedback-loop testing framework by combining these three testing methods to verify and validate AVs iteratively, [40] as shown in Figure 4. [41,42] First, AVs are manually driven on public roads, collecting realistic environmental data with vehicle sensors (e.g., GPS, cameras, radars, and LiDARs) and simultaneously recording naturalistic driving data (e.g., position, speed, steering wheel angle, etc.).Next, based on the acquired data, virtual scenarios are constructed to train and update AVs' algorithms, models, and software in simulation testing.Then, to test some functionalities of AVs, closed-facility testing provides specific and realistic scenarios.Finally, on-road testing is utilized to validate the AVs' comprehensive performance (mainly including safety and intelligence).Then, based on the testing results, the testing cycle is conducted iteratively to improve the AVs' weaknesses and further optimize the strengths until the testing results satisfy a specific standard in acceptance testing.
As shown in Figure 3, simulation testing stands as the most pivotal part among the three methods, appropriate utilization of it can effectively shorten the development and testing cycle. [43]imulation testing has been broadly recognized as an indispensable method for improving the safety and efficiency of AV testing. [36]

State-of-the-Art Simulation Tools for AV Testing
To thoroughly validate AVUT's performance, there are effective testing tools for different stages to guarantee accuracy and efficiency during testing.More specifically, throughout the entire development lifecycle of AVs, software-in-the-loop (SIL) and hardware-in-the-loop (HIL) are generally employed to verify different testing objects at various stages. [44,45]IL in this article also includes model-in-the-loop, [46] both are purely virtual simulation tools.SIL focuses on testing AVUT's algorithms, models, as well as software stacks, it is generally applied in the preliminary phases of simulation testing, [47] its purpose is to evaluate the correctness and reliability of the functionalities executed by the testing objects in the virtual world. [48]nce the performed functionalities in SIL simulation are satisfactory, the software is embedded into relevant hardware for the execution of HIL. [45]Within the simulation platform, HIL provides an emulated plant to be controlled, which represents the physical subsystems such as sensors, actuators, and mechanical components with mathematical models. [49]During HIL testing, the hardware controller (hardware under test) is connected to the simulation platform and interacts with the emulated plant in real-time, and the plant provides input to the controller.Likewise, the controller executes its software and outputs relevant control signals to the plant.The behaviors of the plant change and its updated states are fed back to the controller. [50]IL normally works at the intermediate phases of AV simulation testing to enhance the testing quality. [31]The purpose of HIL is to systematically test the hardware controller's performance and stability and also evaluate the hardware controller's dependability by disturbance and fault injection. [51]n the one hand, since the embedded systems are connected to the testing loop, the accuracy and reliability of HIL have been greatly improved compared with SIL. [52,53]On the other hand, HIL has been verified to have similar accuracy as the test drive (including on-road testing and closed-facility testing), while being safer, more efficient, and less costly.For instance, in Shao's experimental studies on engine fuel consumption and emission, [54] the testing error between HIL and actual AV is about 1%.
Nevertheless, since an AV is a complicated cyber-physical system integrated by several subsystems, SIL and HIL can only test partial functionalities and subsystems of the AVUT, failing to test the comprehensive performance of the entire vehicle.Therefore, further enhancements on HIL are still necessary to facilitate testing the overall performance of AVs.

Testbench-Based Vehicle-in-the-Loop for AV Testing
Considering the automobile's dynamics characteristics, the hardware controller in HIL is gradually extended from engines, sensors, electric control units, etc., to the whole AVUT, called VIL simulation. [52,55]gure 4. Continuous feedback loop testing framework. [41,42].1.Architecture of Testbench-Based VIL Simulation Testing Figure 5 illustrates the structure and procedure of the VIL simulation testing.There are three parts in VIL: the virtual world, the real world, and the master control system. [16]he virtual world mainly includes the virtual environment, virtual sensor models, and the virtual AVUT.The virtual environment is constructed based on the required data during on-road testing and closed-field testing. [56]The virtual environment provides various testing scenarios as the input of VIL testing, related technologies involve 3D modeling of roads and buildings, object shading and image rendering, and dynamic traffic generation. [57,58]The virtual sensor models simulate the perception principles of relevant sensors and perceive the semantic information of the surrounding environment in the virtual world.The virtual AVUT is an avatar of the physical AVUT in the virtual world, the driving behaviors/states of the physical AVUT (e.g., velocity, acceleration, steering angle, and others) are entirely mapped into the virtual AVUT.
The virtual world is generally constructed through various simulation software and frameworks.[66] Each category of simulation software has unique features.The game engine-based frameworks are adept at constructing a highfidelity virtual environment, as a highly realistic image rendering is supported by Unreal Engine or Unity.The precision in modeling visual and LiDAR sensors is also notable, attributed to advanced ray tracing capabilities.Commercial software such as PreScan and CarSim enjoys widespread employment among automotive original equipment manufacturers and suppliers globally.Their strength lies in the integration of multifarious vehicle dynamics and kinematics models, coupled with numerous empirical parameters.PreScan also provides an extensive selection of sensor models, for example, radar, LiDAR, monocular and binocular cameras, etc.The commercial software for ADAS testing is instrumental in enabling users to quickly utilize and debug tested systems, meanwhile producing simulation results that are close to reality.Traffic flow simulation software is concentrated on providing sophisticated vehicle behavior models, including lane changing, following, collaborative group behavior, and conflict scenarios.Additionally, they are capable of simulating large-scale traffic situations, involving tens of thousands of vehicles.This category of software predominantly serves as a third-party joint simulation platform, providing AD algorithms with randomly complex dynamic traffic environments.Table 2 lists the characteristics of the three categories of simulation software.The real world chiefly contains the physical AVUT, the road condition simulation subsystem, and the sensor-in-the-loop subsystem.The physical AVUT operates within a specific space (generally driving on a testbench), the road condition simulation subsystem performs realistic road conditions, actual dynamics characteristics, and driving behaviors through the testbench for the physical AVUT.The sensor-in-the-loop subsystem generates sensor physical signals (e.g., camera, radar, GNSS, etc.) based on the received virtual perception data for the physical AVUT.
The master control system is mainly responsible for data management (collection, analysis, transmission, etc.), virtual traffic scenario generation, AVUT performance evaluation, and simulation state monitoring.

Principle of Testbench-Based VIL Simulation Testing
During the VIL simulation testing, the AVUT initially remains stationary on the testbench, the virtual AVUT is also accordingly stationary in the virtual world.When the tests begin, all the virtual sensor models will operate and acquire perception data within the virtual environment and then transmit the virtual perception data to the sensor-in-the-loop subsystem through the master control system; simultaneously, according to the road conditions where the virtual AVUT is located in the virtual environment, the master control system will also transmit the virtual road condition information to the road condition simulation subsystem.
In the real world, once the sensor-in-the-loop subsystem receives the virtual perception data from the master control system, it will generate corresponding physical perception signals (such as millimeter waves (MMW), images, etc.) by various sensor physical signal generators to stimulate related sensors, thus achieving the perception of the physical AVUT in the virtual environment.At the same time, the road condition simulation subsystem will perform and simulate realistic road conditions based on the received virtual road conditions.
Subsequently, the AVUT makes a series of decisions and controls on the basis of the generated physical perception data from the sensor-in-the-loop subsystem and performs AD under the simulated road conditions.Meanwhile, the road condition simulation subsystem is continuously collecting the driving states of the AVUT and mapping the states to the virtual AVUT.Then the driving behaviors of the virtual AVUT are generated from the mapped states and injected into the virtual environment, synchronously driving the updates of the parameters for all entities within the virtual environment.
Afterward, the virtual sensor models perceive the surrounding environment within the updated virtual environment, and a new iteration of in-the-loop testing will be performed.

Advantage of VIL Simulation Testing
Figure 6 demonstrates the testing quality of four testing tools utilized during AV development. [52]It is evident that, except for the environment and traffic, the objects involved in the testing process are identical across VIL and test drive, which maximizes the authenticity of VIL testing results.Meanwhile, utilizing the virtual environment and traffic also empowers VIL with the advantage of virtual simulation testing, that is, enhanced safety, efficiency, and cost-effectiveness. [67]ompared with SIL and HIL, the physical AVUT is connected to the testing loop, facilitating evaluation of the AVUT's comprehensive performance through observing its overall driving behaviors.The connection also reduces some deviations caused by vehicle model errors.The simulated realistic road conditions ensure the authenticity of the automobile dynamics characteristics. [68]V testing principally aims to ensure safety and reliability and validate its application performance while driving on the roads.VIL provides the physical AVUT with the closest driving conditions to actual driving by combining the virtual environment and the whole AVUT.In industrial applications, especially automobile manufacturers with the most stringent requirements for simulation testing, SIL, HIL, and VIL are applied in different development stages, VIL is generally performed after HIL to validate the AVUT's overall performance.From SIL to HIL, then to VIL, the three tools collaboratively contribute to forming a complete testing toolchain. [69]

Accuracy Analysis of Testbench-Based VIL Simulation Testing
Based on the architecture and principle of VIL, two external factors will directly affect the AVUT's testing results.1) The physical perception data generated by various equipment of the sensor-inthe-loop subsystem, which serves as the input for the AVUT.
2) The road conditions provided by the testbench, which serve as the driving conditions for the AVUT.Both kinds of data are derived and converted from the semantic information in the virtual environment.
Consequently, two primary factors will influence the accuracy of VIL testing results.1) In the virtual world, the fidelity of the virtual scenarios and the authenticity of the virtual sensor models will fundamentally (also indirectly) affect the results.2) In the real world, the performance of various testing equipment in the sensor-in-the-loop subsystem and the road condition simulation subsystem will directly impact the AVUT's driving behaviors and testing results, such as the range accuracy of simulated radar targets, the simulation error of road adhesion coefficient, the control error of road curvature simulation, etc.

Road Condition Simulation in VIL
The essence of VIL is to integrate the physical AVUT into the testing loop, thus requiring drivable test tracks for the vehicle.As mentioned in Introduction, according to the manner of providing road conditions, VIL is generally divided into two categories: [16] the field-based VIL and the testbench-based VIL.During the field-based VIL testing process, the AVUT is operating in an enclosed and empty ground or an enclosed area, the closed field will provide actual road conditions for the AVUT; during the testbench-based VIL testing, the AVUT is driven on an indoor testbench, the testbench will simulate road conditions and automobile's dynamic behaviors through some complex mechanical structures.
There are several differences in the principles of the two categories of VIL while providing road conditions, causing their application situations and testing focuses to be accordingly various.In addition, various testbenches have their unique mechanical designs and operating principles, causing the reality of simulated road conditions to be distinct.
This section discusses field-based VIL and testbench-based VIL respectively, then analyzes and compares the characteristics of the two manners of VIL, and finally summarizes their application scopes.Slightly different from the structure and principle in Figure 4, there is no road condition simulation subsystem in the fieldbased VIL.As the physical AVUT operates on actual roads, it is unnecessary to obtain the virtual road conditions and transmit them to the real world.Instead, the virtual road models are typically constructed based on the actual road conditions. [70]The absence of the road condition simulation subsystem also makes the information of AVUT's localization and driving states only be measured by its GNSS sensors and transmitted to the virtual world with suitable networks. [71]Furthermore, driving in the real world also makes it not feasible to generate sensor physical signals in field-based VIL.Instead, the input of the physical AVUT is replaced by the virtual sensor models' perception data.such as autonomous emergency braking (AEB), lane-keeping assist systems (LKAS), forward collision warning (FCW), and so on.[75][76][77] Experimental results of major studies showed that the difference between field-based VIL and real road testing is small and acceptable, and has almost no impact on the testing results.Park et al. even pointed out that the testing error between the field-based VIL and the real road test is about 3-4%. [52]In this kind of field-based VIL, the AVUT is not constrained by test tracks and can maneuver freely within the empty ground, facilitating the evaluation of multiple testing items with few traffic participants or high speed.Nonetheless, the absence of road constraints affects the accuracy of testing results, particularly in large-scale ITS environments.
To improve the testing accuracy, several studies conducted VIL testing in an enclosed field, meanwhile reproducing the virtual world according to the construction of the closed field, especially the road networks.For example, Tettamanti et al., [78] Griggs et al., [79] and Feng et al. all replicated the road layouts of closed campuses in microtraffic flow simulation software (e.g., SUMO, VISSIM) [80][81][82] and combined with simulated large-scale traffic flows to perform VIL for testing some functions (e.g., a traffic jam assist function, a speed advisory system, the decisionmaking subsystem, and the communication ability of AVs).Solmaz et al. constructed a comprehensive field-based VIL testing platform for testing a traffic management scheme of cooperative vehicle infrastructure systems. [83,84]Experimental results demonstrated that the reproduced road networks standardized the behaviors and actions of all traffic participants, enhancing the authenticity of the traffic conditions in the testing scenarios.As a result, the reproduced road networks and simulated traffic flows in field-based VIL can effectively and accurately validate some functions of AVUT and ITS, especially in large-scale virtual traffic environments.
The apparent advantage of field-based VIL is its ability to provide actual road conditions for AVUT.Nevertheless, due to the structural problems of field-based VIL, there are still some prominent flaws that could be improved.1) Most existing studies concentrated more on simulation-reality synchronization while ignoring the reality of the virtual environments, such as 3D modeling and virtual sensor modeling.The ground truth of the virtual world was directly transmitted to the AVUT, that is, the perception process with the virtual sensors was ignored.Therefore, the testing results were more ideal and were inconsistent with the actual situations.2) The actual environment cannot be fully controlled, it is challenging to provide various realistic conditions flexibly, for example, rainy days, cloudy days, slippery roads, and it is impossible to reproduce the realworld environment for the AVUT.3) It is unsuitable for longterm and long-distance testing due to the limitation of the field's size.
In summary, field-based VIL can only cover limited test cases.It is more appropriate for validating certain ADAS functions and ITS applications in a real-world environment with some specific conditions.

Testbench-Based VIL
In the realm of traditional vehicle testing, testbenches are necessary for functional testing of the entire vehicle, especially in the end-of-line testing stage.For instance, the chassis dynamometer was commonly utilized for testing the braking ability and the antilock brake system of vehicles.Inspired by the testing and diagnosing methods for traditional vehicles, the testbench-based VIL was first presented by Verhoeff et al. of TNO, [85] called "vehicle-hardware-in-the-loop (VEHIL)".In the VEHIL testing method, a full-scale AVUT equipped with sensors was mounted on a chassis dynamometer.The chassis dynamometer (also called "roller dynamometer" or "roller testbench") was used to provide a realistic road load for the AVUT according to the road information of the virtual environment, that is, the AVUT on the chassis dynamometer behaved almost equal to actual tests on the road.Meanwhile, the AVUT was tested by interacting with the virtual environment.A vehicle platoon experiment was conducted with the VEHIL method, and testing results showed that it was feasible and necessary to implement the VEHIL method for AV testing.The VEHIL method provided innovative ideas for ADAS testing, which can ensure safety, reliability, and dynamics performance during testing process.
To simulate realistic road conditions for the AVUT, testbenchbased VIL testing has been continuously employed and improved for the development of AVs over the last two decades.These utilized testbenches were generally subdivided into two types according to their mechanical structure design: [86] chassis dynamometer and powertrain dynamometer.The main difference between them is whether the testbench is designed with rollers for providing the road load, which also makes the realizable dynamic behaviors dissimilar.Figure 8 illustrates the difference between them.

Chassis Dynamometer
A chassis dynamometer has been used since the concept of VEHIL was put forward, and its mechanical structure design has been improved to meet the development of AVs.Gietelink et al. of TNO utilized the VEHIL with a chassis dynamometer to perform a series of functional testing and fault diagnosis for several ADAS applications, [54,[87][88][89] for example, adaptive cruise control (ACC), cooperative ACC, FCW, precrash system (PCS), and a driver information and warning system.The testing work demonstrated the versatility and effectiveness of the VEHIL in validating complex ADAS functionalities under controllable conditions and contributing to the enhancement of vehicle safety features.Verburg et al. of TNO further made an expansion to the VEHIL for implementing lateral maneuvers of the AVUT on a chassis dynamometer. [90]This study marked a significant step in simulating more dynamic and realistic driving scenarios on a roller dynamometer, thus broadening the range of testable maneuvers and enhancing the realism of vehicle dynamics.It also specified a research direction for the subsequent VIL testing.Albers and Du proposed a validation framework with a chassis dynamometer to configure and specify the complex validation environment. [91]The configurations of all the utilized testbenches are generally similar, for instance, the drum configuration is the four-wheel-independent drive, adjustable wheelbase for various sizes of passenger vehicles, and can support maximum velocity, acceleration, and deceleration for regular passenger vehicles.Most importantly, these testbenches have a wonderful performance on response time (≤10 ms) to respond to the AVUT's driving actions in real time.The responsiveness is crucial for ensuring accurate vehicle dynamics characteristics, thus providing reliable and valid test results.Table 3 summarizes the main specifications of the utilized testbench of VEHIL. [87]alko et al. utilized a chassis dynamometer with two independent axles from Horiba Corporation to perform VIL testing for ADAS prototyping and validation. [92,93]The Horiba testbench can further simulate a friction force equivalent to a 5% gradient on the road surface by adjusting torques on the rollers.This capability is particularly important for testing vehicles' performance on various road conditions, which is beneficial to further enhance the comprehensiveness of AV/ADAS validation.
All of the chassis dynamometers mentioned above belong to the traditional roller testbench, which does not support the physical steering of vehicles and can only implement longitudinal motions.Although Gietelink et al. have demonstrated the feasibility of VHEIL for lateral maneuvers on the traditional chassis dynamometer, [88] only the rear wheels of the vehicle were placed on the rollers.Its front wheels were lifted and did not come into contact with rollers.Therefore, the lateral motions of the vehicle were actually simulated according to the single-track bicycle model and the magic formula tire model with the measured steering angle and the speeds of rotation of the rear wheels. [94,95]his testing manner is feasible to control the orientation of the front wheels on the traditional chassis dynamometer.Still, the testing results need to be more accurate, especially for AV functional testing.
To deal with this problem, the DÜRR group developed a new chassis dynamometer with rotating roller sets, [96,97] called X-Road-Curve, [98] which is capable of supporting the steering operations of the AVUT.Table 4 lists the technical indicators of X-Road-Curve.When the AVUT is driving on the  X-Road-Curve, it is also able to automatically keep the AVUT in the center of the testbench without any adjustment on the steering wheel, even if the AVUT is steering.The most significant advantage of the X-Road-Curve is the capability for testing some AV functions requiring lateral maneuvers, such as LKAS and lane-changing operations.The designed steering functionality is essential for the comprehensive assessment of ADAS/AV, as it can replicate real-world driving conditions with common lateral movements.On the basis of the X-Road-Curve, Solmaz et al. proposed a novel hybrid-testing paradigm that utilized VIL to conduct calibration, functional testing, and failure diagnosis for ACC and LKA functions. [99,100]Experimental results showed that X-Road-Curve had the ability to effectively support the steering movements of the AVUT.Furthermore, during the process of vehicle steering, the front axle's rollers could promptly respond to the rotation of the wheels while maintaining the AVUT's center of gravity (COG) at a stable position.The characteristic of the X-Road-Curve can avoid a side drift of the AVUT and ensure safety.In summary, the X-Road-Curve is effective for the validation of AVs' various functionalities and performance, although the testing accuracy needs to be improved especially in simulating complex driving behaviors.
In addition, as the testbench-based VIL is usually conducted in an indoor laboratory, both traditional and steerable testbenches need to simulate longitudinal inertia and resistance forces (e.g., wind resistance, rolling resistance) on different road gradients for the AVUT by applying and adjusting torques on the rollers, [101] that is, the resultant force of these forces is merely implemented through emulating the tire forces.However, it needs to be more accurate to simulate the resistance forces by depending on the rollers alone, because the attitude of AVUT will affect its comprehensive performance, especially in some complex road conditions, such as ramp bridges and slope roads.
To provide more realistic road conditions on the testbench, Huayan Traffic Technology Corporation in China developed a nine-degree-of-freedom (9-DOF) steerable chassis dynamometer, [102] which is capable of simulating pitch, roll, and yaw angle for the AVUT, Table 5 lists the technical specifications of this chassis dynamometer.Zhao et al. from Chang'an University utilized the 9-DOF steerable testbench to perform VIL testing in a virtual environment with various road conditions (e.g., straight roads, slope roads, curve roads) for ACC, AEB, LKAS, PCS, and lane-changing functions. [103,104]Experimental results indicated that the 9-DOF steerable testbench was able to provide a more realistic driving posture for the AVUT by providing more road conditions, although the accuracy and latency still needed to be enhanced.Compared to the X-Road-Curve, the steering performance of the 9-DOF testbench is better due to its ability to support a broader range of steering angles.Moreover, there is no significant shaking and oscillation during the steering process, with the error margin of the steering angle controlled within 2 deg. [103]n summary, the 9-DOF steerable testbench represents an advancement in AV testing technology.It offers a more authentic replication of real-world driving conditions, enabling a more effective and thorough assessment of various AV functions.Despite its current limitations in terms of accuracy and response time, its superior steering and attitude simulation performance and stability during testing provide an effective foundation for future applications in AV development and validation.

Powertrain Dynamometer
In terms of the powertrain dynamometer, it has the advantage of steerability, several manners can be utilized to provide highly dynamic steering motions. [105,106]The powertrain dynamometer is not equipped with any roller, the AVUT is lifted on the testbench, and its tires do not come into contact with any surface.Therefore, it is necessary to construct a precise wheel model (including a tire model) to ensure the reality of the physical road conditions.In the design of the powertrain dynamometer, four load engines are utilized to load torques on the wheel hubs of the AVUT individually for simulating the driving resistance and inertia according to the wheel models.
Schyr et al. of AVL Corporation in Austria designed a powertrain dynamometer and proposed a novel VIL testing framework called "DrivingCube" for efficient validation of AV applications. [105,107,108]Experiment results showed that the DrivingCube powertrain dynamometer was able to simulate road loads accurately, and the vehicle's performances of steering, accelerating, and breaking on it were comparable to those on actual roads.Furthermore, the testing results in ACC, AEB, and LKA scenarios were also close to on-road testing.Li et al. also proposed a cosimulation VIL framework with a highly dynamic powertrain testbench (called R2R, Road to Rig) from KS Engineers Corporation,. [86,109,110]The R2R testbench had the ability to implement authentic nonlinear vehicle dynamics and high tire slip values, allowing for flexibly simulating highly dynamic driving behaviors, thereby covering an extensive range of testing scenarios.Institute of Vehicle System Technology in Karlsruhe Institute of Technology (KIT) designed a powertrain testbench to perform VIL test for ADAS and AD functions. [111]able 6 lists the technical specifications of the KIT's powertrain dynamometer.Based on the steerable powertrain testbench, Kurz et al. further utilized Lagrange's II.equation to model pitch dynamics of the AVUT driving on the testbench. [112]xperimental results showed that the powertrain testbench can accurately simulate the vehicles' pitch motion and ensure the precision during validating ADAS and AD function.Nevertheless, the accurate testing results fully depend on the precision of the vehicle model.In contrast, the chassis dynamometer is more authentic, it is widely used in industry due to its mature technology.However, it requires a more complicated mechanical structure to implement steering motions, and corresponding reality also needs to be improved.In addition, although the AVUT is operating on the testbench and simulating longitudinal/ lateral motions, its center of mass is slightly oscillating at a stationary position.The dynamic behaviors of the AVUT are thus reduced, causing the longitudinal/lateral acceleration, as well as the yaw rate, to not be accurately measured by the inertial sensors.It is necessary to obtain these perception data from the virtual environment.
Taking an overview of road condition simulation in testbenchbased VIL, future studies are suggested to focus on the following.1) Refining the mechanical structure of the chassis dynamometer to simulate more dynamic behaviors (e.g., longitudinal/lateral acceleration of the AVUT's COG, yaw rate, etc.) of the AVUT, that is, increasing the DOFs of the testbench is important.
2) Simulating the longitudinal and lateral acceleration of the virtual AVUT according to its motions and transmitting the information to the actual inertial sensors for generating AVUT's realistic dynamic behaviors is a step forward.3) Existing testbenches all suffer from latency issues.Reducing the overall response times, especially in the simulation of steering angle, pitch angle, and roll angle, is another improving direction for the future.The communication latency and mechanical latency are also supposed to be improved.

Summary of Road Condition Simulation
Although field-based VIL can provide the most realistic road conditions, its application scope is limited.As a result, it is better suited for discovering issues with the tested functionality.Regarding AVUT's performance testing, it still has shortcomings in sensor and automobile performance testing, because it lacks relevant testing equipment (road condition simulator and sensor physical signal generator) to obtain actual information.
In contrast, testbench-based VIL is more appropriate for AV testing and diagnosis because it is able to perform repeatable, long-term, and long-distance testing tasks, which is beneficial for improving testing efficiency.Although the authenticity of the road conditions simulated by testbenches may be slightly lower, the simulation errors between field-based testing and testbench-based testing have no significant impacts on the testing results.Furthermore, testbench-based VIL is also able to enable the combination with sensor-in-the-loop testing to generate precise physical sensor data for the AVUT, [113] which can provide more realistic test cases and further enhance testing accuracy.Finally, as testbench-based VIL is generally carried out in a laboratory environment, it is safer to comprehensively validate the functionalities and overall performance of the AVUT, which contributes to enhancing testing quality. [69]

Sensor Physical Signal Generation in Testbench-Based VIL
The perception data of the surrounding environment is necessary for the AVUT to trigger various functions, [25] for example, ACC, CACC, AEB, FCW, etc.As the whole AVUT is the hardware controller in the VIL testing platform, its perception subsystem must consequently be connected to the closed testing loop by HIL simulation; the sensor physical signal generation as the interface between the AVUT and the virtual world provides real-time sensor signals for the AVUT.
The virtual world generates the original perception data in VIL, and this process does not involve the operation of vehicle sensors.To integrate the physical sensors into the testing loop, a sensor physical signal generator is necessary to stimulate the vehicle sensors to generate corresponding physical sensor data according to the virtual perception data.
Sensor physical signal generators generally have complex operating principles and significantly high manufacturing costs; assembling various sensor signal generators reasonably based on related testing requirements is salutary to increase the accuracy of testing results. [114]

Principle of Sensor Signal Generation Based on Virtual Environment
In the VEHIL testing method, [85] a real-world traffic environment was established in a large room (200 m Â 40 m), and some robot vehicles equipped with vehicle bodies were utilized to simulate the motions of other vehicles relative to the AVUT.The shape, reflection properties, and radar cross sections of the vehicle body are similar to that of a standard passenger vehicle.During VEHIL testing, the robot vehicles moved at high speed in front of the AVUT, and the AVUT was also equipped with a camera and a radar to perceive the surrounding environment.The sensor data were transmitted to the decision-making subsystem to trigger ADAS functions.The sensor data in VEHIL was exact.Nevertheless, establishing such a large-scale laboratory traffic environment was costly and dangerous for both AVUT and the persons involved due to the high-speed robot vehicles. [92]ith the development of SIL and HIL, purely virtual simulation environment was widely used to validate sensor models and various physical sensors.As shown in Figure 9, there are three different channels in simulation testing to generate sensor data from the virtual environment, [83] the relationship among these three channels is progressively dependent.

Full Simulation
The full simulation is generally used in SIL testing.The ground truth of target parameters (e.g., position, radial velocity, azimuth, and so on) that needs to be measured in the virtual environment is directly transmitted to the AVUT.No sensor models or sensor equipment are involved, the generated sensor data is thus highly ideal.This channel is suitable for the early stage of functional development and implementation, it provides fundamental data for the other two channels.

Sensor Modeling
Various sensor models are established in the virtual environment according to the operating principle of the corresponding sensor equipment.The input of the sensor models is the ground truth generated in the full simulation channel, the output of the models is the measure of the target parameters, and there are usually errors or noises in the measurements to improve the authenticity of the sensor data.The sensor data generated by the virtual sensor models are finally transmitted to the AVUT.
Sensor modeling is mainly utilized for HIL testing, it has been quite mature and applied by many AD simulation software, such as PreScan, CARLA, AirSim, etc.These pieces of software all support the virtual simulation of radar, LiDAR, ultrasonic radar, camera, and other sensors.

Physical Signal Stimulation
Physical signal stimulation is generally applied for VIL testing. [115]Based on the sensor models in the virtual environment, the virtual sensor data is transmitted into an external sensor signal injection device (also called sensor physical signal generator), which produces corresponding physical signals, for example, MMW, images, laser, etc.Then, the produced physical signals will stimulate correlative vehicle sensors to output physical sensor data.
In the physical signal stimulation approach, the generated physical sensor data actually describes the states of the virtual detected targets.However, for the AVUT, the data is considered to describe the states of the targets in the real world.As shown in Figure 10, whether the perception information comes from the real environment or the virtual environment, it is eventually received by the vehicle sensors, and these sensors will output corresponding physical sensor data to the AVUT (mainly to the decision-making and motion planning subsystems).The AVUT thus "thinks" it is driving in a real environment, that is, the AVUT is "fooled" by various sensor signal generators in the testbench-based VIL testing platform.The physical signal stimulation approach in VIL keeps the sensor equipment in the testing loop, so it is essentially a sensor HIL simulation method, which is also proper for sensor testing and diagnosis.

Camera Physical Signal Generation
It is relatively easy to implement signal stimulation for the camera.For example, some studies utilized a Mobileye smart camera as the only perception input to the ADAS functions. [83,84]A screen (or a huge curtain) is used to show the scenery captured by the virtual camera in real time.The camera is fixed behind the windshield and oriented toward the curtain.According to the distance to the curtain, calibration is necessary for the camera to fit its field of view (FOV).During VIL testing, the camera is stimulated by the sceneries projected onto the curtain, and corresponding ADAS functions are further triggered through the camera's signals.Rossi et al. utilized a Mobileye 560 smart camera installed on the windshield to validate the ADAS system. [116]VL utilized projection onto canvas in the "DrivingCube" platform to test the ACC and LKA systems. [117]Experimental results illustrated that the camera stimulation method enables near realworld validation in a safe environment.While this stimulation manner is convenient to implement, its accuracy and reliability leave much to be desired.On one hand, the lighting conditions of the testing environment will affect the camera's imaging quality.On the other hand, if real-world objects unrelated to the testing process enter the camera's FOV, the perception results will affect the AVUT's decision-making.This manner is suitable for situations where the camera is fixed to the AVUT and cannot be removed.
In another manner, a high-definition display screen and the camera are both fixed in an obscura, the distance between them can be adjusted by a rail on where the camera is mounted, as shown in Figure 11.The advantages of this stimulation manner include ensuring controllability and repeatability, the ability to control the lighting conditions inside the obscura, and isolating external interference.The disadvantage is that it requires the removal of the AVUT's camera and placement inside the obscura.Nevertheless, this stimulation manner in an obscura is still recognized and widely employed in many camera-inthe-loop testing studies.Reway et al. proposed a camera-in-the-loop testing method with a camera obscura to validate object detection algorithms of ADAS. [119]Experimental results showed that the proposed method can quantify the performance of the tested detection algorithms, especially suitable in some severe weather conditions and some unwanted effects on the lenses. [120]

Radar Physical Signal Generation
A target simulator with an anechoic box is necessary for the automotive radar sensor stimulation.The radar is usually fixed in the anechoic box to avoid the generation of false targets in the laboratory environment, as shown in Figure 12.The virtual sensor data is sent to the target simulator (i.e., radar physical signal   generator), and the target simulator also simultaneously receives the MMW from the radar on the AVUT.According to the virtual radar data, the target simulator calculates corresponding echo signals to simulate the parameters (e.g., relative distance, radial velocity, azimuth, elevation, and radar cross section, etc.) of the virtual targets, and the echo is used to stimulate the radar for triggering ADAS functions.[123] Since its core technology is not the focus of this article, this article will not introduce these studies in detail.
Automotive RTS has been increasingly attractive in testbenchbased VIL testing due to the high precision and mature technology.In addition, some commercial test equipment manufacturers have also been bringing corresponding products to the market, for example, National Instrument (NI) Corporation in America, [124] dSPACE Corporation in Germany, [125] and so on.Gadringer and Gruber first proposed a solution to incorporate an RTS into the "DrivingCube" and validated the performance of the RTS. [126]Zhu et al. proposed an MMW radar in-the-loop testing platform for intelligent vehicles. [121]The utilized NI RTS was able to generate accurate and authentic radar echo signals.
The average error of the echo simulation in distance and velocity are 0.28 m and 0.14 m s À1 , much better than the actual radar detection's 0.98 m and 0.21 m s À1 .One limitation of the utilized RTS is that it can only generate one virtual target.Zhao et al. also integrated a single-target radar simulator and a camera target simulator in their 9-DOF steerable testbench-based VIL platform to test and diagnose the radar and camera sensor in an ACC driving scenario, [102][103][104]125] as shown in Figure 13. In thir recent studies, a double-target radar simulator based on NI equipment is utilized for testing in more complex scenarios.There are two sets of target simulation groups, each generates one virtual radar target.Each target simulation group contains a turntable controlled by a servo motor, an MMW radio head (model number is mmRH-3608) fixed on the turntable, a vector signal transceiver (model number is PXle-5840), and a variable delay generator (model number is NI-5692).dSPACE has produced an RTS that can generate detection data of five targets simultaneously.[127] It is capable of simulating lateral movement of the targets by controlling the antenna motor.Another feature of this RST is that it can adjust the height and grazing angle of the radar inside the anechoic box according to the virtual scenario.[128] Additionally, the RTS can operate in synchronization with other simulators during the testing process, such as a camera simulator.For testing in a more complex traffic environment, Diewald et al. from KIT Institute of Vehicle System Technology implemented a multichannel, multiecho RTS in the fully integrated VIL platform, [106,129] the simulator had the ability to simulate 4-6 virtual targets.Furthermore, the maximum angle error in angle simulation was only 0.2 deg.In AVL's latest "DrivingCube" VIL system, the utilized RTS of Rohde & Schwarz (R&S) can simulate up to 8 independently moving targets from 2 to 300 m. [117,130,131] This advancement highlights the growing sophistication of RTS technology, allowing for more complex and various testing scenarios.The ability to simulate multiple targets and their movements adds depth to the testing scenarios, enabling more comprehensive validation of AV systems under diverse conditions.
The evolution of RTS technology and its integration into VIL platforms has significantly enhanced the capability to accurately simulate and test various radar-based AV/ADAS functionalities. [132]The precision, versatility, and synchronicity of RTS with other sensor signal generators are pivotal for ensuring the reliability and safety of AV.
As far as LiDAR stimulation, it has not been used in VIL testing because traditional LiDAR-related technology is advanced and highly difficult.It was generally replaced by virtual point cloud or virtual LiDAR modeling.

GNSS Signal Generation
The AVUT always remains fixed on the testbench during VIL simulation testing, its positioning information is thus almost unchanged.High-precision positioning information is the fundamental of AV, [133] aiding AD system in making accurate navigation decisions.To thoroughly and comprehensively validate the performance of AVUT driving on the testbench, it is essential to provide GNSS signal.
Based on the high-definition map constructed in the virtual world, the master control system transmits the virtual AVUT's positioning information to the GNSS simulator.The GNSS simulator generates radio frequency signals that are similar to those emitted by real GNSS satellites, such as GPS, Galileo, and BeiDou, including orbital and timing information. [134]The generated GNSS signal is able to stimulate the onboard positioning devices to achieve the synchronization of positioning information and time between the virtual and real world.This synchronization is primary for ensuring the accuracy and reliability of the simulation testing, as it bridges the gap between virtual and physical environments.Furthermore, the GNSS simulator can also set various parameters according to the virtual world, such as weather conditions, attenuations, and interference.
The technology of GNSS simulators has become quite advanced, with many well-developed products available for AV HIL testing, offered by companies such as Spirent and R&S. [135,136]Nevertheless, there have been almost few studies or companies using GNSS simulators in testbench-based VIL.Rossi et al. integrated GPS simulator with the VEHIL platform to provide real-time positioning information to the AVUT. [116]VL integrated R&S GNSS simulator into the "DrivingCube" VIL system.The generated GNSS signals are consistent with that in the simulation environment, facilitating the validation of ADAS and AD functionalities. [137,138]he integration of GNSS simulators in testbench-based VIL enhances the fidelity and comprehensiveness of AV testing, contributing to more effective and reliable testing of AV systems.However, the relatively low uptake in the industry suggests a potential area for growth and further exploration, particularly in enhancing the realism and accuracy of AV testing processes.

Summary of Sensor Physical Signal Generation
Sensor physical signal generation is fundamental for testbenchbased VIL, considerable advancements have been made in the method of sensor stimulation; nonetheless, there are still some shortcomings that necessitate further enhancement.1) The camera physical signal generation requires progress to support stereo cameras, fisheye cameras, and multicamera fusion testing.2) Automotive RTSs are suggested to concentrate on generating bidirectional (longitudinal and lateral) movements and multitarget (more than 10 targets, [139] best to 64 targets) signals.Such enhancements would substantially improve the authenticity of radar-based driving system testing.3) Improvements in accuracy and latency reduction are essential for GNSS simulators.4) Physical signal generation of LiDAR is extensively demanded.Therefore, future studies can attempt to generate physical signals for solid-state LiDAR.5) Sensor physical signal generators typically have high costs, it is preferable to reduce the cost of related equipment if possible in the future.It would make testbench-based VIL testing more accessible and feasible for a wider range of applications.
To summarize, while sensor physical signal generation technologies for testbench-based VIL have seen considerable progress, enhancements in supporting advanced sensor combinations, improving multi-target simulation capabilities, introducing LiDAR signal generation, and cost reduction are crucial steps forward.Addressing these problems will further refine the efficacy and applicability of sensor physical signal generation in the validation of AV/ADAS.

Conclusion and Future Work
This article summarizes current studies concerning testbenchbased VIL simulation testing for AVs.AV testing-related procedures, methods, and tools are first concluded to analyze existing issues and improvements.Then, the architecture, principle, advantages, and characteristics of testbench-based VIL are discussed, highlighting the importance of road condition simulation and sensor physical signal generation in testbench-based VIL testing.Most importantly, the structures, mechanical designs, operating principles, and characteristics of various testbenches and sensor signal generators are investigated in detail.
The main advantages of testbench-based VIL simulation testing include 1) preserving highly realistic vehicle dynamics characteristics, 2) ensuring close-to-reality perception ability through sensor signal generation, 3) operating in highly safe and reproducible conditions, particularly suitable for verifying the performance boundary of AV/ADAS in some critical scenarios, and 4) considerable accuracy and efficiency.It is obvious that testbench-based VIL is feasible, providing an ingenious approach to efficiently speed up the verification and validation process of AV/ADAS.
Given these advantages, testbench-based VIL simulation testing bridges the gap between simulation and proving ground.It has gained widespread recognition in the industry, particularly among several AD testing companies (e.g., AVL, KS Engineers, and DÜRR) in Europe that are promoting this method.Meanwhile, these companies collaborate with some universities (e.g., the Technical University Berlin, Karlsruhe Institute of Technology, Technical University of Darmstadt, Graz University of Technology, and Clemson University), and the academic community is thus gradually adopting testbenchbased VIL.In summary, testbench-based VIL is able to facilitate the validation process, implementing a full-size VIL testing platform equipped with a testbench, and various physical sensor signal generators are meaningful approaches for AV/ADAS validation.
Nevertheless, to the best of our knowledge, there has not been a full-scale VIL platform that can be directly applied to test AVs in industrialization.One reason is that the supportable DOFs of existing testbenches are not adequate enough, another reason is that LiDAR-in-the-loop testing (mainly LiDAR signal generator) is significantly difficult and costly to implement in VIL.In addition, existing automotive RTS can generate radar data for merely a few targets, which cannot completely describe the actual traffic conditions.Hence, future studies can be devoted to filling these technology gaps and establishing a full-scale VIL testing platform to test AVs comprehensively, accurately, and efficiently.

Figure 3 .
Figure 3. Procedure of AV functional safety management during development lifecycle.

Figure 5 .
Figure 5. Architecture and principle of testbench-based VIL for AV testing.

Figure 6 .
Figure 6.Testing quality comparison of four testing tools utilized during AV development.

Figure 7 .
Figure 7. Structure and principle of field-based VIL testing.

Figure 8 .
Figure 8. Difference in mechanical structure design between a) chassis dynamometer and b) powertrain dynamometer.

Figure 9 .
Figure 9. Three different channels of data generation in simulation testing.The dashed line is virtual data or signal and the solid line is physical data or signal.

Figure 10 .
Figure 10.Comparison of the sensor physical signal generation process between on-road testing and VIL testing.

Figure 11 .
Figure 11.Process of camera stimulation in VIL.

Figure 12 .
Figure 12.Process of RTS in VIL.

Figure 13 .
Figure 13.Illustrations of the VIL platform with the 9-DOF chassis dynamometer, a RTS, and a camera target simulator a) Overall view of the VIL platform; b) 3-DOF chassis dynamometer; c) Radar target simulator; d) Camera target simulator.

Table 2 .
Comparison of three categories of simulation software.
4.1.2.Studies on Field-Based VIL Testing Based on the characteristics of field-based VIL, numerous studies have utilized field-based VIL for functional testing of ADAS and applications of intelligent transportation systems (ITS) in urban traffic environments.Some studies performed field-based VIL in an enclosed and empty ground.For instance, Park et al. and Miquet et al. utilized VIL to test ADAS functions,

Table 3 .
Specifications of the chassis dynamometer of VEHIL.

Table 4 .
Technical indicators of X-Road-Curve.

Table 6 .
Technical specifications of the KIT Powertrain Testbench.