A method and system for generating human-machine collaborative conflict test cases

Abstract

The application provides a human-machine collaborative conflict test case generation method and system, and the generation method comprises the following steps: S1: collecting vehicle motion state, driving operation behavior and driving personnel physiological state, performing multi-modal data fusion modeling, constructing a personalized driving style model, and calculating driving style parameters; S2: real-time detection of human-machine collaborative conflict behavior, and quantification of human-machine collaborative conflict level; S3: inputting the human-machine collaborative conflict level and the driving style parameters as condition variables into a generative adversarial imitation learning model to generate human-machine collaborative conflict test cases; S4: verifying the dynamics rationality of the human-machine collaborative conflict test cases through a physical engine; and S5: distributing and iterating the human-machine collaborative conflict test case generation model by using a federated learning mechanism. The human-machine collaborative conflict test case generated by the method can truly simulate the human-machine responsibility game and behavior contradiction existing in the actual driving process.

Description

Technical Field

[0001] This invention relates to the field of automotive automation control technology, and more specifically, to a method and system for generating human-machine collaborative conflict test scenarios. Background Technology

[0002] With the development of autonomous driving technology, testing and verification have become crucial for ensuring its safety and reliability. Currently, autonomous driving testing mainly relies on two methods: real-vehicle testing and virtual simulation testing. While real-vehicle testing can reflect real-world road performance, it is costly, time-consuming, and subject to environmental and regulatory limitations, making it difficult to cover a vast range of scenarios. As autonomous driving systems evolve to higher levels, the requirements for testing scale and complexity are increasing dramatically, highlighting the limitations of real-vehicle testing.

[0003] CN114492157B provides a method for generating autonomous driving test scenarios based on personalized driver models. Based on genetic and evolutionary principles and combined with reinforcement learning, it simulates various stages of a human driver's driving experience, progressively training driver models at different developmental stages, including novice driver models, experienced driver models, and skilled driver models. Based on these developmental models, and according to the different personalized driving characteristics of human drivers, it further trains personalized driver models, including aggressive driver models, conservative driver models, provocative driver models, and cooperative driver models. According to testing requirements, the driver models at different developmental stages and the different personalized driver models are combined in corresponding proportions to generate appropriate target test scenarios.

[0004] CN114862156B provides a method for generating customized test scenarios for emotion-driven personalized driver models, including: selecting specific emotions that affect driving behavior based on existing emotion models; pre-applying specific emotional influences to test drivers, and then collecting corresponding driving data of test drivers on a driving simulator; performing imitation learning and reinforcement learning Q-Learning on the collected driving data to obtain generalized driving data of different styles; combining driving data of different styles according to a specific ratio, determining hyperparameters through a fireworks optimization algorithm, thereby generating efficient test scenarios for autonomous vehicles; determining verification indicators for efficient test scenarios; selecting the decision-making method of the system under test, and verifying it through the verification indicators.

[0005] However, the aforementioned methods primarily focus on testing scenarios involving driver proficiency and emotional state, failing to effectively model the dynamic interactive behaviors between human drivers and autonomous driving systems in typical adversarial situations such as control struggles and command conflicts. Therefore, the generated test cases cannot realistically simulate the power struggles and behavioral conflicts between humans and machines during actual driving, limiting their effectiveness in evaluating human-machine collaborative safety and coordination capabilities. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for generating human-machine collaborative conflict test scenarios. By using driving style and human-machine collaborative conflict behavior as conditional variables, a generative adversarial imitation learning model dynamically generates human-machine collaborative conflict test cases based on personalized driving style modeling. This can realistically simulate the human-machine power and responsibility game and behavioral contradictions that exist in actual driving.

[0007] To achieve the above objectives, this invention provides a method for generating human-machine collaborative conflict test cases. The method includes the following steps: S1: Collecting vehicle motion state, driving operation behavior, and driver physiological state; performing multimodal data fusion modeling to construct a personalized driving style model and calculate driving style parameters; S2: Real-time detection of human-machine collaborative conflict behavior and quantification of the human-machine collaborative conflict level; S3: Using the human-machine collaborative conflict level and driving style parameters as conditional variables, inputting them into a generative adversarial learning model, which dynamically generates human-machine collaborative conflict test cases; S4: Verifying the dynamic rationality of the human-machine collaborative conflict test cases through a physics engine and eliminating invalid test cases; S5: Using a federated learning mechanism to distribute and iterate the human-machine collaborative conflict test case generation model.

[0008] The technical effects achieved by adopting this solution are as follows: By integrating vehicle motion state, driving operation behavior, and driver physiological state, and performing multimodal data fusion modeling, a deep quantification and personalized representation of driver style is achieved, providing a precise data foundation for generating highly realistic test cases; by real-time detection of human-machine collaborative conflict behavior and quantification of human-machine collaborative conflict levels, human-machine conflict scenarios are introduced into the test cases; then, using driver style and human-machine collaborative conflict behavior as conditional variables, a generative adversarial imitation learning model dynamically generates human-machine collaborative conflict test cases based on personalized driving style modeling, which can realistically simulate the human-machine power and responsibility game and behavioral contradictions that exist in actual driving processes. Combined with a physics engine and federated learning mechanism, it provides strong support for the safety verification of autonomous driving systems.

[0009] Preferably, the driving style parameters are calculated using the following weighted fusion formula:

[0010] K=α1×K_vehicle+α2×K_operations+α3×K_physiological

[0011] Where α1, α2, and α3 are weighting coefficients, satisfying α1+α2+α3=1, K is the driving style parameter, K_vehicle is the vehicle motion state score, K_operation is the driving operation behavior score, and K_physiological is the driver's physiological state score.

[0012] The technical effects achieved by adopting this solution are as follows: By fusing data from three dimensions—vehicle motion state, driving operation behavior, and driver physiological state—and using a weighted summation method, the generated test cases not only reflect the driver's long-term behavioral characteristics but also capture instantaneous decision-making changes caused by physiological factors such as fatigue and tension, greatly improving the realism and personalization of human-machine conflict scenario simulations. Secondly, the adjustability of the weight coefficients α1, α2, and α3 allows for flexible customization of generation strategies according to different testing objectives, thereby enhancing the ability to generate targeted test cases for specific high-risk scenarios while ensuring comprehensive test case coverage.

[0013] Preferably, the driver's physiological state includes at least one of heart rate variability, blink frequency, and skin conductance; the vehicle's motion state includes at least one of following distance, longitudinal acceleration, lateral acceleration, and lane departure; and the driving operation behavior includes at least one of steering wheel angular velocity, steering wheel torque, accelerator pedal opening, and emergency braking frequency.

[0014] The technical effects achieved by adopting this solution are as follows: By precisely defining the specific composition of multimodal data, physiological states include physiological indicators such as heart rate variability, blink frequency, and skin conductance; vehicle motion states include dynamic parameters such as following distance, longitudinal / lateral acceleration, and lane departure; and driving operation behaviors include control characteristics such as steering wheel angular velocity, torque, accelerator pedal opening, and emergency braking frequency. The synergistic use of these data can capture the driver's implicit physiological states (such as tension and fatigue) and operational intentions, making the calculation of driving style parameters both physiologically realistic and behaviorally accurate, thereby generating conflict scenarios that better fit the complex decision-making mechanisms of humans. Moreover, these data ensure the collectability, interpretability, and reproducibility of the model input data, significantly improving the transparency and reliability of the test case generation process, and ultimately providing high-value test cases for human-machine cooperative driving systems that combine physiological realism and engineering practicality.

[0015] Preferably, driving styles are categorized based on driving style parameters:

[0016] Conservative type: K < 0.3;

[0017] Robust type: 0.3≤K≤0.7;

[0018] Aggressive type: K>0.7.

[0019] The technical effects achieved by adopting this solution are as follows: By clearly classifying driving style parameters (conservative: K<0.3; robust: 0.3≤K≤0.7; aggressive: K>0.7), the standardization and quantifiable grading of driver behavior characteristics are realized, thereby significantly improving the relevance and coverage of test case generation. On the one hand, this classification method can accurately match the typical behavior patterns of drivers with different styles, ensuring that the generated conflict test cases can realistically reproduce the cautious hesitation of conservative drivers, the balanced decision-making of robust drivers, and the risk-taking tendency of aggressive drivers, greatly enhancing the diversity and realism of human-computer interaction conflict scenarios.

[0020] Preferably, the human-machine collaborative conflict behaviors include at least one of the following: the angle difference between the driver's steering wheel torque and the autonomous driving system control command; the difference between the driver's pedal operation and the autonomous driving system control command; the difference between the driver's gaze on the road and the target road of the autonomous driving system control command; changes in the driver's heart rate variability; and changes in the driver's skin conductance response.

[0021] The technical effects achieved by adopting this solution are as follows: By clearly defining human-machine collaborative conflict behavior as encompassing multiple dimensions—including operational (steering wheel torque angle difference, pedal operation differences), cognitive (road target gaze deviation), and physiological (heart rate variability changes, skin conductance response changes)—it achieves multimodal deep fusion perception and quantification of conflict behavior. By introducing driver gaze deviation and real-time physiological stress responses (such as reduced heart rate variability indicating tension and enhanced skin conductance response indicating alertness), it can capture implicit discrepancies between the driver and the autonomous driving system in intent understanding and emotional response earlier and more accurately, thereby achieving advanced prediction and graded warning of potential conflicts. This multimodal fusion-based definition of conflict behavior provides extremely rich and realistic human-like feature inputs for generating adversarial learning models, enabling them to generate high-fidelity, full-dimensional human-machine conflict test cases that not only involve operational opposition but also cognitive dissonance and physiological stress. This significantly improves the test coverage and verification depth of autonomous driving systems in the face of the complexity of human-machine co-driving.

[0022] Preferably, the conflict level in human-machine collaboration is calculated using the following formula: Conflict_Level = β1 × Δθ norm +β2×Pedal_Conflict norm +β3×Eye_deviation norm +β4×(1 / HRV norm )+β5×GSR_meannorm ;

[0023] Where Conflict_Level is the human-machine collaborative conflict level, and β1, β2, β3, β4, and β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm For the normalized steering wheel operation angle difference, Pedal_Conflict norm Eye_deviation is a normalized metric for pedal operation variation. norm For normalized gaze path bias, HRV norm For normalized heart rate variability, GSR_mean norm This represents the normalized average value of the skin conductance response.

[0024] The technical effects achieved by adopting this solution are as follows: By introducing a calculation formula for human-machine collaborative conflict levels, a refined and multi-dimensional objective assessment of conflict intensity is achieved. First, the calculation formula comprehensively integrates external behavioral conflicts (steering wheel angle difference, pedal operation difference), cognitive attention distraction (road gaze deviation), and internal physiological stress responses (reduced heart rate variability, enhanced skin conductance response). This makes the determination of conflict levels no longer limited to a single operational disagreement, but deeply integrates the driver's physiological state, greatly improving the accuracy and robustness of the assessment. Second, through normalization processing and flexible configuration of weight coefficients (β1 to β5), the dimensional differences of multi-source data are eliminated, and the assessment focus is dynamically adjusted according to the type of conflict (such as control struggle, attention conflict), thereby accurately distinguishing different conflict intensities from mild intervention to severe confrontation. Finally, this quantified conflict level provides stable and reliable conditional input for generating adversarial imitation learning models, ensuring that the generated test cases can accurately cover human-machine conflict scenarios with different levels of tension.

[0025] Preferably, the level of human-machine collaboration conflict is classified according to the level of human-machine collaboration conflict:

[0026] Low conflict: Conflict_Level < 0.4;

[0027] Conflict Level: 0.4 ≤ Conflict_Level < 0.7

[0028] High conflict level: Conflict_Level ≥ 0.7.

[0029] The technical effects achieved by adopting this solution are as follows: By quantifying the human-machine collaborative conflict level into three distinct levels—low, medium, and high—standardized and refined hierarchical management of conflict intensity is realized. This grading method transforms continuous comprehensive conflict metrics into discrete semantic labels, enabling precise matching and differentiated processing of test case generation and autonomous driving system response strategies (e.g., logging is sufficient for low-level conflicts, warnings are issued for medium-level conflicts, and emergency takeover is triggered for high-level conflicts), greatly enhancing the relevance and effectiveness of testing. Furthermore, this grading mechanism provides clear constraints for generating adversarial learning models, enabling them to generate conflict scenarios covering the entire spectrum from minor daily disagreements to extremely intense confrontations, thereby systematically verifying the reliability and safety of human-machine interaction under different levels of human-machine conflict in autonomous driving systems.

[0030] Preferably, the adversarial training of the generative adversarial imitation learning model includes: the generator receiving random noise and the condition variable, and outputting synthetic scene data; the discriminator receiving real scene data or the synthetic scene data and the condition variable, and outputting scalar probability values; and adversarial training is performed by fixing the generator, updating the discriminator, and fixing the discriminator, and updating the discriminator, until Nash equilibrium is reached.

[0031] The technical effects achieved by adopting this solution are as follows: This adversarial training mechanism forces the generator to learn the underlying distribution of real human-machine conflict scenarios under strict constraints of specific conflict levels and driving styles, thereby generating high-quality synthetic data that conforms to physical laws and is rich in detail. This effectively solves the problems of high cost of acquiring real conflict data and scarcity of long-tail scenarios. The discriminator ensures a high degree of consistency between the generated scenarios and the target conditions by jointly judging the matching of data and conditions (such as whether the generated aggressive driving style scenario corresponds to a high conflict level). Through a dynamic optimization process of alternating training until Nash equilibrium, the generator and discriminator continuously evolve in the game, and the final generated data approaches or even surpasses the real data distribution in terms of realism and diversity. This provides an extensive, reasonably distributed test scenario library for autonomous driving systems that includes a large number of extreme cases.

[0032] Preferably, the federated learning mechanism further includes: the central server initializing the human-machine collaboration conflict test case generation model and distributing it to each node; each node training and customizing the human-machine collaboration conflict test case generation model distributed to each node based on local data; each node encrypting and uploading the model parameters of the human-machine collaboration conflict test case generation model to the central server for secure aggregation; and updating the human-machine collaboration conflict test case generation model and distributing it to each node for the next iteration.

[0033] The technical effects achieved by adopting this solution are as follows: This mechanism utilizes the data value of all nodes to improve the generalization ability and robustness of the model, while strictly protecting data privacy. At the same time, it can accurately focus on core testing objectives such as personalized driving styles and human-machine conflicts, and ultimately efficiently generate test cases that can expose system weaknesses, providing a powerful impetus for the safety verification of autonomous driving systems.

[0034] This invention also provides a human-machine collaborative conflict test scenario generation system for implementing any of the above technical solutions. The system includes: a central server and multiple distributed nodes communicatively connected to the central server; wherein the distributed nodes are deployed in vehicle terminals or simulators; the central server runs a human-machine collaborative conflict test case generation model; the distributed nodes run local human-machine collaborative conflict test case generation models; both the human-machine collaborative conflict test case generation model and the local human-machine collaborative conflict test case generation model include the following functional modules: a driving style modeling model for calculating driving style parameters; a human-machine conflict detection model for real-time detection and calculation of the level of human-machine collaborative conflict; a generative adversarial learning model for generating human-machine collaborative conflict test cases; and a verification model for verifying the dynamic rationality of the human-machine collaborative conflict test cases; the distributed nodes also include: a data acquisition module for acquiring vehicle motion state, driving operation behavior, and driver physiological state.

[0035] The technical effects achieved by adopting this solution are as follows: The system deploys distributed nodes in vehicle terminals or simulators to directly and locally collect multimodal data. The data acquisition module is directly embedded in the terminal, achieving millisecond-level real-time data acquisition. The driving style modeling model and the human-machine conflict detection model calculate driving style parameters and conflict levels in real time, significantly improving system response speed. The adversarial learning model generates personalized test cases using driving style parameters and conflict levels. The verification module uses an integrated physics engine to generate test cases for dynamic rationality verification, automatically eliminating invalid test cases and ensuring the quality of generated test cases. Therefore, this system, through an efficient, self-learning, and rationally guaranteed automated method, can generate realistic, diverse, and high-risk human-machine conflict test scenarios on a large scale. Attached Figure Description

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0037] Figure 1 A flowchart illustrating the steps of the method for generating human-machine collaborative conflict test cases provided in this application embodiment. Detailed Implementation

[0038] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This invention provides a method for generating human-machine collaborative conflict test cases, comprising the following steps: S1: Collecting vehicle motion state, driving operation behavior, and driver physiological state; performing multimodal data fusion modeling to construct a personalized driving style model and calculating driving style parameters; S2: Real-time detection of human-machine collaborative conflict behavior and quantification of the human-machine collaborative conflict level; S3: Using the human-machine collaborative conflict level and driving style parameters as conditional variables, inputting them into a generative adversarial learning model, which dynamically generates human-machine collaborative conflict test cases; S4: Verifying the dynamic rationality of the human-machine collaborative conflict test cases through a physics engine and automatically eliminating invalid test cases; S5: Using a federated learning mechanism to distribute and iterate the human-machine collaborative conflict test case generation model. See details below. Figure 1 As shown.

[0040] This solution integrates vehicle motion state, driving operation behavior, and driver physiological state, and performs multimodal data fusion modeling to achieve in-depth quantification and personalized representation of driver style, providing a precise data foundation for generating highly realistic test cases. By real-time detection of human-machine collaborative conflict behavior and quantification of human-machine collaborative conflict levels, human-machine conflict scenarios are introduced into the test cases. Driver style and human-machine collaborative conflict behavior are then used as conditional variables, and a generative adversarial learning model dynamically generates human-machine collaborative conflict test cases based on personalized driving style modeling. This realistically reflects the human-machine power and responsibility game and behavioral contradictions that exist in actual driving. Combined with a physics engine and federated learning mechanism, it solves the problems of missing human-machine interaction modeling, insufficient scenario realism, and data silos in traditional testing methods. Its core breakthrough lies in incorporating driver psychological motivation and physiological indicators into conflict decision-making and generating high-fidelity test scenarios through interpretable mathematical mapping, providing a more reliable verification framework for autonomous driving systems.

[0041] In some embodiments of this application, the driving style parameters are calculated using the following weighted fusion formula:

[0042] K=α1×K_vehicle+α2×K_operations+α3×K_physiological

[0043] Where α1, α2, and α3 are weighting coefficients, satisfying α1+α2+α3=1, K is the driving style parameter, K_vehicle is the vehicle motion state score, K_operation is the driving operation behavior score, and K_physiological is the driver's physiological state score.

[0044] Specifically, the comprehensive evaluation index of driving style, the driving style parameter K, is obtained by weighted summation of the vehicle motion state score K_vehicle, the driving operation behavior score K_operation, and the driver's physiological state score K_physiological. The core idea behind vehicle motion state scoring is to quantify the intensity and stability of vehicle motion, reflecting the direct impact of driving behavior on vehicle dynamics. Aggressive driving behavior is typically characterized by greater longitudinal and lateral acceleration; a higher score indicates more intense and unstable vehicle motion. The core idea behind driving operation behavior scoring is that frequent steering wheel adjustments, large-angle turns, and frequent hard braking are typical signs of aggressive driving or inattentiveness; a higher score indicates more aggressive and unstable driving operations. The core idea behind driver physiological state scoring is that decreased heart rate variability (HRV), increased skin conductance (GSR), or changes in eye movement patterns (such as gaze or blink frequency) objectively reveal increased cognitive load and fatigue. This allows monitoring these indicators to more accurately assess a driver's suitability for driving. Higher values ​​generally indicate a poorer physiological state, suggesting the driver is more stressed, fatigued, or distracted. In other words, for the same driving behavior, some drivers may have relatively stable physiological indicators (skilled, confident), while others may exhibit extremely high skin conductance (stressed, risk-taking). Therefore, physiological data helps distinguish between different intrinsic motivations and states under externally similar behaviors.

[0045] This solution integrates data from three dimensions—vehicle motion state, driving operation behavior, and driver physiological state—using a weighted summation method. This allows the generated test cases to not only reflect the driver's long-term behavioral characteristics but also capture instantaneous decision-making changes caused by physiological factors such as fatigue and tension, greatly enhancing the realism and personalization of human-machine conflict scenario simulations. Secondly, the adjustability of the weight coefficients α1, α2, and α3 allows for flexible customization of generation strategies according to different testing objectives, thereby ensuring comprehensive test case coverage while enhancing the ability to generate targeted test cases for specific high-risk scenarios.

[0046] In some embodiments of this application, the driver's physiological state includes at least one of heart rate variability, blink frequency, and skin conductance; the vehicle's motion state includes at least one of following distance, longitudinal acceleration, lateral acceleration, and lane departure; and the driving operation behavior includes at least one of steering wheel angular velocity, steering wheel torque, accelerator pedal opening, and emergency braking frequency.

[0047] Specifically, heart rate variability (HRV) can be obtained through a steering wheel capacitive sensor, a smartwatch / wearable device, or a sensor built into the seat; blink frequency can be obtained through non-contact recognition using an in-cabin camera combined with computer vision algorithms; skin conductance response (GSR) can be measured through a dedicated wristband or contact electrodes on the steering wheel; longitudinal / lateral acceleration can be calculated from data such as the vehicle's inertial measurement unit or wheel speed sensors; lane departure can be obtained through the lane recognition system of the vehicle's camera or a high-precision positioning module (such as RTK-GPS); following distance can be calculated by emitting a laser beam from a lidar and measuring the reflection time or by identifying the vehicle ahead from a point cloud; steering wheel angular velocity, steering wheel torque, accelerator pedal opening, and emergency braking frequency can be obtained by reading from the vehicle's CAN bus.

[0048] Based on the above data, the formula for calculating the vehicle motion state score K_vehicle is: K_vehicle = a 1× long_std+a2×lat_std+a3×lane_deviation_std+a4×K_d;

[0049] Where long_std is the normalized standard deviation of longitudinal acceleration, reflecting the severity of acceleration and deceleration; lat_std is the normalized standard deviation of lateral acceleration, reflecting the severity of steering; lane_deviation_std is the normalized standard deviation of the distance between the vehicle's center of gravity and the lane centerline, reflecting the stability of lane keeping ability; K_d is the vehicle distance score; a1, a2, a3, and a4 are weighting coefficients, and satisfy a1+a2+a3+a4=1.

[0050] Furthermore, when K_d=0 (when THW≥T_safe, the following distance is very safe); K_d=1 (when THW≤T_danger, the following distance is extremely dangerous); K_d changes linearly between T_danger and T_safe. THW is the headway, the time required for two vehicles to reach each other at their current speeds; T_safe is the safe time threshold, and T_danger is the dangerous time threshold; T_safe can be set to 3 seconds or more (e.g., 3.5 seconds), and T_danger can be set to 1 second or less (e.g., 0.8 seconds).

[0051] The formula for calculating the driving operation behavior score K_operation is as follows:

[0052] K_operations=b1×steering_std+b2×brake_event_rate+b3×throttle_std;

[0053] Where steering_std is the normalized standard deviation of steering wheel angular velocity, reflecting the smoothness or abruptness of steering operation; brake_event_rate is the normalized frequency of emergency braking events, reflecting the aggressiveness of driving and safety awareness; throttle_std is the normalized standard deviation of accelerator pedal opening, reflecting the smoothness of throttle operation. The higher the value, the more frequent the "deep and shallow" throttle operation; b1, b2, and b3 are weighting coefficients, and satisfy b1+b2+b3=1.

[0054] The formula for calculating the driver's physiological state score K_physiological is as follows:

[0055] K_physiological=c1×(1 / HRV_std)+c2×GSR_mean+c3×blink_rate;

[0056] HRV_std represents the standard deviation of normalized heart rate variability. A decrease in HRV is usually associated with stress and fatigue, indicating a poor state of mind. Therefore, the higher the 1 / HRV_std value, the worse the driver's state of mind. GSR_mean represents the mean of normalized skin conductance response. An increase in GSR usually indicates emotional arousal, tension, or stress. blink_rate_norm represents blink rate. An abnormally high or low blink rate may be related to fatigue or high concentration. c1, c2, and c3 are weighting coefficients, and they satisfy c1+c2+c3=1.

[0057] This solution precisely defines the specific composition of multimodal data. Driver physiological states include indicators such as heart rate variability, blink rate, and skin conductance; vehicle motion states include dynamic parameters such as following distance, longitudinal / lateral acceleration, and lane departure; and driving behavior includes control characteristics such as steering wheel angular velocity, torque, accelerator pedal opening, and emergency braking frequency. The collaborative use of these data can capture the driver's latent physiological states (such as tension and fatigue) and operational intentions, ensuring that the calculation of driving style parameters possesses both physiological realism and behavioral accuracy. This generates conflict scenarios that better reflect complex human decision-making mechanisms. Furthermore, this data ensures the collectability, interpretability, and reproducibility of the model input data, significantly improving the transparency and reliability of the test case generation process. Ultimately, it provides high-value test cases for human-machine cooperative driving systems that combine physiological realism and engineering practicality.

[0058] In some embodiments of this application, driving styles are classified according to the driving style parameter (K):

[0059] Conservative type: K < 0.3;

[0060] Robust type: 0.3≤K≤0.7;

[0061] Aggressive type: K>0.7.

[0062] When the driving style is conservative, the vehicle speed is stable, acceleration and deceleration are gentle, the following distance is long, and the frequency of lane changes and overtaking is extremely low; when the driving style is steady, the driving behavior is close to average, with no obvious risk tendency; when the driving style is aggressive, the vehicle speed is relatively high, with frequent rapid acceleration and braking, close following distance, and active lane changes and overtaking.

[0063] This solution achieves standardized and quantifiable classification of driver behavior characteristics through explicit classification of driving style parameters (conservative: K<0.3; robust: 0.3≤K≤0.7; aggressive: K>0.7), thereby significantly improving the relevance and coverage of test case generation. On the one hand, this classification method can accurately match the typical behavioral patterns of drivers with different styles, ensuring that the generated conflict test cases can realistically reproduce the cautious hesitation of conservative drivers, the balanced decision-making of robust drivers, and the risk-taking tendency of aggressive drivers, greatly enhancing the diversity and realism of human-computer interaction conflict scenarios.

[0064] In some embodiments of this application, human-machine cooperative conflict behaviors include at least one of the following: the angle difference between the driver's steering wheel torque and the control command of the autonomous driving system, the difference between the driver's pedal operation and the control command of the autonomous driving system, the difference between the road the driver is looking at and the target road of the autonomous driving system control command, the change in the driver's heart rate variability, and the change in the driver's skin conductance response.

[0065] Among them, the angle difference Δθ between the driver's steering wheel torque and the autonomous driving system control command is the difference between the steering wheel torque angle applied by the driver (θ_human) and the target angle (θ_ads) of the autonomous driving system (ADS) control command (Δα=θ_human-θ_ads). This indicator directly represents the intensity of the immediate struggle and intentional divergence between the human and machine in terms of lateral control of the vehicle. The difference in accelerator / brake pedal operation captures the direct intentional confrontation between the driver and ADS in terms of longitudinal control of the vehicle. Specifically, it is manifested as the driver pressing the brake pedal deeply while the ADS system is still executing the acceleration command, or the human requests acceleration while the ADS system is in the process of acceleration. Execution braking; gaze deviation is the angle of deviation between the driver's gaze focus and the target road planned by the ADS system, captured by an eye tracker. This indicator profoundly reveals the driver's decreased trust in or misunderstanding of the ADS decision; a decrease in HRV is a strong signal, indicating that the driver is in a state of high stress. This physiological state reduces their patience and tolerance, amplifies dissatisfaction with the system's behavior, and thus directly increases the level of human-machine conflict; the peak frequency and amplitude of GSR are perfect indicators for quantifying instantaneous emotional reactions. For example, a sudden SCR spike can directly correspond to a specific human-machine conflict event (such as emergency system takeover or false alarm).

[0066] This solution achieves multi-modal deep fusion perception and quantification of conflict behavior by clearly defining human-machine collaborative conflict behavior as encompassing the operational layer (steering wheel torque angle difference, pedal operation differences), the cognitive layer (road target gaze deviation), and the physiological layer (heart rate variability changes, skin conductance response changes). By introducing driver gaze target deviation and real-time physiological stress responses (such as decreased heart rate variability indicating tension, and increased skin conductance response indicating alertness), it can capture implicit discrepancies between the driver and the autonomous driving system in intent understanding and emotional response earlier and more accurately. This enables advanced prediction and graded warning of potential conflicts. This multi-modal fusion-based definition of conflict behavior provides extremely rich and realistic human-like feature inputs for generative adversarial learning models, enabling them to generate high-fidelity, full-dimensional human-machine conflict test cases that not only involve operational opposition but also cognitive dissonance and physiological stress. This significantly improves the test coverage and verification depth of autonomous driving systems in the face of the complexity of human-machine co-driving.

[0067] In some embodiments of this application, the human-machine collaboration conflict level is calculated using the following formula: Conflict_Level = β1 × Δθ norm +β2×Pedal_Conflict norm +β3×Eye_deviation norm +β4×(1 / HRV norm )+β5×GSR_mean norm;

[0068] Where Conflict_Level is the human-machine collaborative conflict level, and β1, β2, β3, β4, and β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm For the normalized steering wheel operation angle difference, Pedal_Conflict norm Eye_deviation is a normalized metric for pedal operation variation. norm For normalized gaze path bias, HRV norm For normalized heart rate variability, GSR_mean norm This represents the normalized average value of the skin conductance response.

[0069] This solution achieves a refined and multi-dimensional objective assessment of conflict intensity by introducing a calculation formula for human-machine collaborative conflict levels. Firstly, the formula comprehensively integrates external behavioral conflicts (steering wheel angle difference, pedal operation differences), cognitive attention distraction (road gaze deviation), and internal physiological stress responses (reduced heart rate variability, enhanced skin conductance). This ensures that conflict level determination is no longer limited to a single operational disagreement but deeply incorporates the driver's physiological state, significantly improving the accuracy and robustness of the assessment. Secondly, through normalization and flexible configuration of weighting coefficients (β1-β5), it eliminates the dimensional differences between multi-source data and allows for dynamic adjustment of the assessment focus based on conflict type (e.g., control struggle, attention conflict), thus accurately distinguishing different conflict intensities from minor intervention to severe confrontation. Finally, this quantified conflict level provides stable and reliable input conditions for generating adversarial imitation learning models, ensuring that the generated test cases accurately cover human-machine conflict scenarios with varying levels of tension.

[0070] In some embodiments of this application, the conflict level is classified according to the level of human-machine collaboration conflict:

[0071] Low conflict: Conflict_Level < 0.4;

[0072] Conflict Level: 0.4 ≤ Conflict_Level < 0.7

[0073] High conflict level: Conflict_Level ≥ 0.7.

[0074] This technical solution quantifies human-machine collaborative conflict levels into three distinct grades: low (<0.4), medium ([0.4, 0.7)), and high (≥0.7), achieving standardized and refined hierarchical management of conflict intensity. This grading method transforms continuous comprehensive conflict metrics into discrete semantic labels, enabling precise matching and differentiated processing of test case generation and autonomous driving system response strategies (e.g., logging only for low-level conflicts, issuing warnings for medium-level conflicts, and triggering emergency takeover for high-level conflicts), greatly enhancing the relevance and effectiveness of testing. Furthermore, this grading mechanism provides clear constraints for generating adversarial learning models, allowing them to generate conflict scenarios covering the entire spectrum from minor daily disagreements to extremely intense confrontations, thereby systematically verifying the reliability and safety of human-machine interaction under different levels of human-machine conflict in autonomous driving systems.

[0075] Low conflict: Human and machine intentions are basically consistent, operation is smooth, and it is a normal co-driving state; Medium conflict: There is an identifiable operational disagreement, the driver may be slightly distracted or dissatisfied, and there is a potential risk; High conflict: There is a serious disagreement between human and machine intentions, danger is about to occur or has already occurred, and the driver may be in a tense, angry or dangerous operating state.

[0076] In some embodiments of this application, adversarial training of the generative adversarial imitation learning model includes: a generator receiving random noise and the condition variable, and outputting synthetic scene data; a discriminator receiving real scene data or the synthetic scene data and the condition variable, and outputting a scalar probability value representing the confidence level of the data authenticity; and performing adversarial training by fixing the generator, updating the discriminator, and fixing the discriminator, and updating the discriminator, with training continuing until Nash equilibrium is reached.

[0077] Specifically, the core of the Generative Adversarial Imitation Learning (GAIL / CGAIL) model is to learn the distribution characteristics of real data through adversarial training between the generator and the discriminator. Specifically, the generator (G) receives random noise z and optional condition variables c (driving style parameter K value, human-machine collaboration conflict level value), and outputs synthetic scene data G(z|c). The discriminator (D) receives real scene data (s,a) or generated scene data G(z|c), and the corresponding condition variable c, and outputs a scalar probability value representing the confidence level of the data's authenticity. The adversarial training process employs an alternating optimization strategy. First, the generator is fixed, and the discriminator is updated. A batch of real-world state-action scenario data (s,a) and its corresponding condition variable c are sampled from real data. Simultaneously, a batch of generated scenarios G(z|c) is sampled from the generator. Gradient ascent is used to maximize V(D,G) and update the parameters of D, improving its ability to distinguish between real and generated data. Second, the discriminator is fixed, and the generator is updated: a batch of random noise z and the condition variable c are sampled. The parameters of G are updated using policy gradient methods (such as REINFORCE) or direct backpropagation (when G is a differentiable network), maximizing the confidence of generated data passing through D, logD(G(z|c)). Training continues until Nash equilibrium is reached or performance requirements are met. Here, Nash equilibrium refers to a state where the game between the generator and discriminator reaches a stable state, and the parameters of both the generator and discriminator tend to stabilize, with the training loss no longer changing significantly. When the system reaches Nash equilibrium, the discriminator's average output probability for generated data is close to 0.5, meaning that the discriminator considers the generated data to be real with a 50% probability, indicating that generated data and real data are indistinguishable. This means the scalar probability value reaches 50%.

[0078] Furthermore, CGAIL's conditional generation mechanism is its core advantage in generating test cases for autonomous driving. It allows the test generation process to be closely integrated with personalized driving styles and the dynamic quantification of human-machine conflict, achieving precise and controllable scenario generation. In CGAIL, the K-value serves as a conditional variable input to the generator and discriminator, directly influencing the "personality" and behavioral patterns of the generated scenarios. By adjusting the K-value, CGAIL can seamlessly cover a spectrum of driving styles from conservative to aggressive, generating matching test scenarios and achieving a comprehensive assessment of the risks that drivers of different styles may face. When the driving style is conservative, if the conditional variable c contains a low K-value, the generator tends to generate low-risk scenarios, such as a large following distance (THW>2.5s), gentle acceleration changes (longitudinal acceleration <1.5m / s²), early and gradual lane changes (steering angular velocity <10° / s), and conservative responses to surrounding vehicle behavior. When the driving style is aggressive, if the variable c contains a high K value, the generator will generate high-risk scenarios, such as extremely short following distance (THW<1.0s), intense acceleration and braking behavior (longitudinal acceleration>3.0m / s²), frequent sudden lane changes (steering angular velocity>30° / s), and provocative interactive behaviors (such as close-range cutting in and vying for right-of-way).

[0079] The dynamic quantitative metric Conflict_Level reflects the degree of difference between the autonomous driving system and the human driver's control intentions in real time. In CGAIL, the target Conflict_Level can be used as a condition variable to directly guide the generator to create test scenarios with a specific conflict intensity. After receiving the target Conflict_Level through the condition variable, the generator adjusts the scenario parameters to achieve that conflict level. For example, to achieve a high conflict level, the generator may shorten the cutting distance of the vehicle in front, increase the speed of the background vehicle, or create visual obstructions to delay the system's perception, thereby inducing a decision-making divergence between the driver and the system. When generating a low-conflict scenario with Conflict_Level < 0.4, the generator will generate a human-machine conflict... Figure 1 Contextualized scenarios include: smooth acceleration / deceleration consistent with driver expectations, lane-changing maneuvers performed by the system matching driver expectations, and system responses to the environment matching driver judgment. When generating high-conflict scenarios (Conflict_Level > 0.7), the generator creates scenarios with severe discrepancies between human and machine intentions to expose system weaknesses, such as emergency takeover requests where the system fails to respond to danger in time, forcing the driver to intervene urgently (e.g., grabbing the steering wheel, slamming on the brakes); erroneous intervention where overly conservative system decisions cause driver discomfort (e.g., unnecessary emergency braking), leading to conflict; and control struggles where the autonomous driving system and driver operate the vehicle simultaneously and in opposite directions (e.g., the driver brakes while the system accelerates).

[0080] This adversarial training mechanism forces the generator to learn the underlying distribution of real human-machine conflict scenarios under strict constraints of specific conflict levels and driving styles, thereby generating high-quality synthetic data that conforms to physical laws and is rich in detail. This effectively solves the problems of high cost of acquiring real conflict data and scarcity of long-tail scenarios. The discriminator ensures a high degree of consistency between the generated scenarios and the target conditions by jointly judging the matching of data and conditions (such as whether the generated aggressive driving style scenario corresponds to a high conflict level). Through a dynamic optimization process of alternating training until Nash equilibrium, the generator and discriminator continuously evolve in the game, and the final generated data approaches or even surpasses the real data distribution in terms of realism and diversity. This provides autonomous driving systems with a test scenario library that is widely covered, reasonably distributed, and contains a large number of extreme cases.

[0081] In some embodiments of this application, the federated learning mechanism further includes: the central server initializing the human-machine collaboration conflict test case generation model and distributing it to each node; each node training and customizing the human-machine collaboration conflict test case generation model distributed to each node based on local data; each node encrypting and uploading the model parameters of the human-machine collaboration conflict test case generation model to the central server for secure aggregation; and updating the human-machine collaboration conflict test case generation model and distributing it to each node for the next iteration.

[0082] Federated learning is essentially a distributed machine learning technique. Its goal is to achieve a distributed training model where "the data remains still while the model moves," while ensuring data privacy, security, and legal compliance. Specifically, a central server initializes a human-machine collaborative conflict test case generation model, which includes a driving style modeling model, a human-machine conflict detection model, and a generative adversarial learning model. This model is then distributed to all vehicles or simulation test nodes participating in the federated learning process. Each vehicle or node uses locally collected driving data to train and fine-tune these global models locally. The local training data includes multi-dimensional information such as the node's personalized driving style K-value, the real-time calculation of the human-machine conflict level, and the vehicle's motion state. To prevent the leakage of sensitive information, after local training, each vehicle or node only shares the model parameters... Data updates (such as gradients or weights) are encrypted and uploaded to a central aggregation server, while the original driving data remains locally. After collecting encrypted parameters from multiple nodes, the central server executes a secure aggregation algorithm (such as FedAvg) to generate an updated global model. Simultaneously, weights are assigned based on the data quality, quantity, or importance of each node, preventing excessive bias in data from certain nodes from negatively impacting the global model's performance. The updated global model is then redistributed to each node, which uses it for the next round of local training. This process repeats continuously, allowing the global model to evolve by absorbing the wisdom of all nodes, without requiring centralized storage of any original data. This mechanism leverages the data value of all nodes to enhance the model's generalization ability and robustness while strictly protecting data privacy. It also precisely focuses on core testing objectives such as personalized driving styles and human-machine conflicts, ultimately efficiently generating test cases that expose system weaknesses, providing a powerful impetus for the safety verification of autonomous driving systems.

[0083] This invention also provides a human-machine collaborative conflict test scenario generation system for implementing any of the above technical solutions. The system includes: a central server and multiple distributed nodes communicatively connected to the central server; wherein the distributed nodes are deployed in vehicle terminals or simulators; the central server runs a human-machine collaborative conflict test case generation model; the distributed nodes run local human-machine collaborative conflict test case generation models; both the human-machine collaborative conflict test case generation model and the local human-machine collaborative conflict test case generation model include the following functional modules: a driving style modeling model for calculating driving style parameters; a human-machine conflict detection model for real-time detection and calculation of the level of human-machine collaborative conflict; a generative adversarial learning model for generating human-machine collaborative conflict test cases; and a verification model for verifying the dynamic rationality of the human-machine collaborative conflict test cases; the distributed nodes also include: a data acquisition module for acquiring vehicle motion state, driving operation behavior, and driver physiological state.

[0084] This system directly collects multimodal data locally by deploying distributed nodes in vehicle terminals or simulators. The data acquisition module is directly embedded in the terminal, achieving millisecond-level real-time data acquisition. The driving style modeling model and human-machine conflict detection model calculate driving style parameters and conflict levels in real time, significantly improving system response speed. An adversarial learning model generates personalized test cases using driving style parameters and conflict levels. The verification module uses an integrated physics engine to perform dynamic rationality checks on the generated test cases, automatically eliminating invalid cases to ensure the quality of the generated test cases. Therefore, this system, through an efficient, self-learning, and rationally guaranteed automated method, can generate realistic, diverse, and high-risk human-machine conflict test scenarios on a large scale.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating test cases for human-machine collaborative conflicts, characterized in that, The generation method includes the following steps: S1: Collect vehicle motion status, driving operation behavior and driver physiological state, perform multimodal data fusion modeling, construct personalized driving style model, and calculate driving style parameters; S2: Real-time detection of human-machine collaborative conflict behaviors and quantification of the level of human-machine collaborative conflict; S3: Based on the human-machine collaborative conflict level and the driving style parameters as condition variables, the generator of the adversarial imitation learning model accepts random noise and the condition variables, and outputs synthetic scene data; the discriminator of the adversarial imitation learning model receives real scene data or the synthetic scene data and the condition variables, and outputs a scalar probability value representing the confidence level of the data authenticity; by fixing the generator, the discriminator is updated; then the discriminator is fixed and the generator is updated, adversarial training is performed, and the training continues until Nash equilibrium is reached, thereby dynamically generating the human-machine collaborative conflict test cases; S4: Verify the dynamic rationality of the human-machine collaboration conflict test cases through a physics engine and eliminate invalid test cases; S5: The federated learning mechanism is used to distribute and iterate the human-machine collaborative conflict test case generation model.

2. The generation method according to claim 1, characterized in that, The driving style parameters are calculated using the following weighted fusion formula: K =α1×K_vehicle+α2×K_operations+α3×K_physiological Where α1, α2, and α3 are weighting coefficients, satisfying α1+α2+α3=1, K is the driving style parameter, K_vehicle is the vehicle motion state score, K_opersation is the driving operation behavior score, and K_physiological is the driver's physiological state score.

3. The generation method according to claim 2, wherein the driver's physiological state includes at least one of heart rate variability, blink frequency, and skin conductance response; The vehicle motion state includes at least one of following distance, longitudinal acceleration, lateral acceleration, and lane departure; The driving operation behavior includes at least one of the following: steering wheel angular velocity, steering wheel torque, accelerator pedal opening, and emergency braking frequency.

4. The generation method according to claim 2, characterized in that, Driving styles are classified according to the aforementioned driving style parameters: Conservative type: K < 0.3; Robust type: 0.3≤K≤0.7; Aggressive type: K>0.

7.

5. The generation method according to claim 1, characterized in that, The human-machine collaborative conflict behaviors include at least one of the following: the angle difference between the driver's steering wheel torque and the control command of the autonomous driving system; the difference between the driver's pedal operation and the control command of the autonomous driving system; the difference between the road the driver is looking at and the target road of the autonomous driving system control command; changes in the driver's heart rate variability; and changes in the driver's skin conductance response.

6. The generation method according to claim 5, characterized in that, The level of conflict in human-machine collaboration is calculated using the following formula: Conflict_Level=β1×Δθ norm +β2×Pedal_Conflict norm +β3×Eye_deviation norm +β4×(1 / HRV norm )+β5×GSR_mean norm ; Where Conflict_Level is the human-machine collaborative conflict level, and β1, β2, β3, β4, and β5 are weight coefficients, satisfying β1+β2+β3+β4+β5=1, Δθ norm For the normalized steering wheel operation angle difference, Pedal_Conflict norm Eye_deviation is a normalized metric for pedal operation variation. norm For normalized gaze path bias, HRV norm For normalized heart rate variability, GSR_mean norm This represents the normalized average value of the skin conductance response.

7. The generation method according to claim 6, characterized in that, The human-machine collaboration conflict level is classified according to the aforementioned human-machine collaboration conflict level: Low conflict: Conflict_Level < 0.4; Conflict Level: 0.4 ≤ Conflict_Level < 0.7 High conflict level: Conflict_Level ≥ 0.

7.

8. The generation method according to claim 1, characterized in that, The federated learning mechanism also includes: The central server initializes the human-machine collaboration conflict test case generation model and distributes it to each node; Each node trains and personalizes the model for generating human-machine collaborative conflict test cases distributed to each node based on local data. Each node encryptedly uploads the model parameters of the human-machine collaboration conflict test case generation model to the central server for secure aggregation; Update the test case generation model for human-machine collaboration conflicts and distribute it to each node for the next iteration.

9. A system for generating human-machine collaborative conflict test scenarios, used to implement the generation method as described in any one of claims 1 to 8, characterized in that, The system includes: A central server and multiple distributed nodes that are communicatively connected to the central server; The distributed nodes are deployed in the vehicle terminal or simulator; The central server runs a human-machine collaboration conflict test case generation model; The distributed nodes run a local human-machine collaboration conflict test case generation model; Both the human-machine collaboration conflict test case generation model and the local human-machine collaboration conflict test case generation model include the following functional modules: Driving style modeling model, used to calculate driving style parameters; A human-machine conflict detection model is used to detect and calculate the level of human-machine collaborative conflict in real time. Generate an adversarial imitation learning model to generate test cases for human-machine collaborative conflict; The validation model is used to verify the dynamic rationality of human-machine collaborative conflict test cases; The distributed node also includes a data acquisition module, used to acquire vehicle motion status, driving operation behavior, and driver physiological status.

Images