Closed site automatic driving vehicle test method with adaptive adjustment of interaction difficulty

Abstract

The present application relates to a kind of closed field automatic driving vehicle test methods of interactive difficulty self-adaptive adjustment, method includes: the interactive ability evaluation index of measured vehicle and simulated traffic participant is calculated, the interactive difficulty parameter corresponding to current time is obtained based on interactive ability evaluation index;The cost function is updated based on the interactive difficulty parameter of current time, the reference trajectory of simulated traffic participant is obtained by using game algorithm based on updated cost function, the interactive ability evaluation index is calculated after executing reference trajectory, if the interactive ability evaluation index meets preset boundary criterion, then reference trajectory is used as test trajectory;Otherwise, the interactive difficulty parameter is updated based on the interactive ability evaluation index.Compared with prior art, the present application realizes the automatic adjustment of test difficulty, so as to realize the effective test of the interactive ability boundary of automatic driving vehicle.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle testing technology, and in particular to a closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty. Background Technology

[0002] Closed-site testing is a crucial verification method for autonomous vehicles before mass production. It can construct complex dynamic traffic scenarios in a controlled and safe environment to systematically verify the perception, decision-making, and control capabilities of autonomous driving systems in interactive scenarios. Currently, the mainstream closed-site testing method is predefined scenario testing, such as Chinese patent application CN116679580A, which provides a simulation verification method for intelligent connected vehicles in closed scenarios. By acquiring and mapping data in a closed scenario to a virtual model, it simulates the operating state of intelligent connected vehicles and generates risk warnings and control commands. These methods typically conduct experiments by pre-setting parameters such as the movement trajectory and speed of simulated traffic participants. While they offer advantages such as good test standardization and strong repeatability, they also have the following problems: 1) The behavior of simulated traffic participants is fixed, meaning that the simulated traffic participants used in the test (such as fake cars and pedestrians) cannot accurately depict the dynamic game characteristics of human drivers in real traffic environments. The tested vehicles often only need to rely on pre-stored strategies to pass the test smoothly, making it difficult to expose the decision boundaries and deep algorithmic defects of the autonomous driving system under extreme conditions in a timely manner; 2) There is a lack of efficient online adjustment mechanisms. Test parameters (such as entry timing, interaction distance, and degree of aggressiveness of behavior) are usually statically set before the test, which can easily lead to two extreme situations: On the one hand, the overall difficulty of the test scenario is too low, and the tested vehicle always operates within a large safety margin, which cannot fully examine the boundary performance of the system and wastes test resources; on the other hand, some scenario designs are too aggressive, exceeding the physical safety tolerance of the system, and posing safety risks such as causing real vehicle collisions.

[0003] In summary, existing testing technologies have significant limitations in terms of adaptive adjustment of interaction difficulty, coverage of high-risk interaction scenarios, dynamic adjustment of game strategies, and semantic interpretability of the testing process, making it difficult to support efficient and reliable verification of the boundaries of the interaction capabilities of autonomous driving systems. Therefore, providing a method that can solve the above problems is a technical issue that needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a closed-field autonomous vehicle testing method with adaptive adjustment of interaction difficulty. By constructing a game interaction model between the vehicle under test and simulated traffic participants, as well as a closed loop for interaction capability evaluation and difficulty adjustment, the testing difficulty is automatically adjusted, thereby effectively testing the interaction capability boundary of autonomous vehicles.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This invention provides a closed-course testing method for autonomous vehicles with adaptive adjustment of interaction difficulty. At any given time, the method includes:

[0007] S1. Obtain the pose information of the tested vehicle and simulated traffic participants at the current moment, calculate the interaction capability evaluation index based on the pose information, and obtain the interaction difficulty parameter corresponding to the current moment based on the interaction capability evaluation index.

[0008] S2. Update the cost function based on the interaction difficulty parameter at the current moment, and solve the updated cost function using a game theory algorithm to obtain the reference trajectory of the simulated traffic participant; the cost function includes the cost function of the tested vehicle and the cost function of the simulated traffic participant, and the cost function of the simulated traffic participant is a weighted sum of risk interaction cost, smoothness cost and trajectory rationality cost, and the game weight of each cost is calculated based on the interaction difficulty parameter.

[0009] S3. Obtain the pose information of the simulated traffic participants after executing the reference trajectory, calculate the interaction capability evaluation index based on the pose information, and if the interaction capability evaluation index meets the preset boundary criteria, then the reference trajectory is used as the test trajectory; otherwise, return to execute S1 based on the current pose information.

[0010] As a preferred technical solution, the tested vehicle and simulated traffic participants include:

[0011] The simulated traffic participants are treated as leaders, and the vehicles under test are treated as followers.

[0012] The control quantities for both leaders and followers include longitudinal acceleration and lateral acceleration, expressed as: , Indicates longitudinal acceleration. Indicates lateral acceleration;

[0013] The leader state and follower state are both represented as follows: And the discrete state transition function is:

[0014] ;

[0015] Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. This represents the discrete state transition function for simulating traffic participants or the vehicle being tested. Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. Indicates the first The control quantities of simulated traffic participants or the tested vehicles at any given time. Indicates the longitudinal distance along the reference path. Indicates longitudinal velocity. This indicates the lateral deviation along the tangent of the reference path. Indicates lateral velocity. Indicates transpose; Let represent the state transition matrix, and:

[0016] ;

[0017] Indicates a time interval; Let represent the control input matrix, and:

[0018] .

[0019] As a preferred technical solution, the cost function of the vehicle under test is:

[0020] ;

[0021] ;

[0022] ;

[0023] in, This represents the follower cost function value; Indicates followers from the first Time to the The amount of control at any given moment Indicates the length of the test round; Indicates the number of followers The state at any given moment; Indicates that the follower is on the reference trajectory at the 1st The state at any given moment; Indicates transpose; and Each represents the positive semidefinite weight matrix of the followers; Indicates the number of followers The amount of control at any given moment; Indicates non-negative weighting coefficients; Indicates the first Physical safety margin at any given moment; Represents the discrete state transition function of the follower; Indicates the first The actual distance between the vehicle being tested and the nearest simulated traffic participant at any given time; and They represent the first The longitudinal speed of the vehicle being measured and the simulated traffic participants at any given time; Indicates response time; Indicates the minimum safe distance when stationary; This represents the minimum longitudinal acceleration of the follower.

[0024] As a preferred technical solution, the cost function for simulating traffic participants is:

[0025] ;

[0026] ;

[0027] ;

[0028] in, Indicates that the leader is from the first Time to the The amount of control at any given moment Indicates the length of the test round; This represents the cost function for simulating traffic participants; , and These represent the game weights for risk interaction costs, smoothness costs, and trajectory rationality costs, respectively. Indicates the interaction difficulty parameter; Indicates the cost of risk interaction, and:

[0029] ;

[0030] Indicates the time of intrusion. Indicates followers. Indicates the leader, Indicates the potential collision point. Indicates the time it takes for the leader to arrive at the potential collision point. Indicates the time it takes for the followers to reach the potential collision point; Let $\frac{ ...

[0031] ;

[0032] Indicates the leader's longitudinal acceleration at time t, Indicates the leader's Lateral acceleration at any given moment; Let $\math$ represent the cost of trajectory rationality, and:

[0033] ;

[0034] Indicates the leader's The state at any given moment, This indicates that the leader is on the reference trajectory. The state at any given moment; This represents a positive semidefinite matrix that the leader uses to weight longitudinal, lateral offset, and velocity error. This represents the follower's optimal response. Indicates the leader's status. Indicates the follower state. Indicates the amount of control by followers; Indicates the leader in The amount of control at any given moment; This indicates the minimum time after the intrusion.

[0035] As a preferred technical solution, the method for obtaining the interaction difficulty parameter is as follows:

[0036] The system acquires the current interaction capability evaluation index, the trajectory of the tested vehicle and simulated traffic participants, the minimum physical safety margin, the test results, and the corresponding scene semantic information, and performs feature extraction and encoding to obtain a structured description. The encoding is based on a prompt template, which includes: scene type, current interaction difficulty parameter value, interaction capability evaluation index of this round of test, minimum safety margin, whether a collision has occurred, whether emergency braking has been triggered, and the interaction difficulty parameter value of the previous round and its corresponding test results.

[0037] Based on the structured description, a large language model is used to output text-based difficulty adjustment suggestions, which include the adjustment value of the interaction difficulty parameter at the next time step.

[0038] As a preferred technical solution, the method for updating the cost function is as follows:

[0039] Based on the The interaction difficulty parameter at each moment is used to update the game weights and the minimum post-intrusion time using a monotonic mapping relationship; the monotonic mapping relationship is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] in, , as well as The game weights represent the costs of risk interaction, smoothness, and trajectory rationality. Indicates the minimum time after intrusion; and These represent the minimum and maximum values ​​of the risk interaction cost game weights obtained after calibration; Indicates the interaction difficulty parameter; and These represent the maximum and minimum values ​​of the smoothness cost game weights obtained after calibration; and These represent the maximum and minimum values ​​of the trajectory rationality cost game weights obtained after calibration; and These represent the maximum and minimum values ​​of the minimum post-intrusion time obtained through calibration, respectively.

[0046] As a preferred technical solution, the method for calculating the interaction capability evaluation index is as follows:

[0047] ;

[0048] in, , and Let represent the non-negative weight coefficients, and . ; The security index is calculated based on the minimum post-intrusion time and physical security margin. The efficiency index is calculated based on the actual travel time and reference travel time of the tested vehicle in the current interaction segment. The stability index is calculated based on the longitudinal and lateral accelerations of the vehicle under test.

[0049] As a preferred technical solution, the method for calculating the aforementioned safety index is as follows:

[0050] ;

[0051] in, Indicates adjustable weights; followers and leaders Reaching the potential collision point Minimum post-intrusion time; This indicates a positive number that prevents the denominator from being zero; This represents the minimum value of the physical security margin.

[0052] As a preferred technical solution, the method for calculating the efficiency index is as follows:

[0053] ;

[0054] in, This indicates the actual travel time of the tested vehicle in the current interaction segment; This indicates the reference transit time of the tested vehicle in the current interaction segment.

[0055] As a preferred technical solution, the method for calculating the stability index is as follows:

[0056] ;

[0057] in, Indicates the total duration of the test rounds; and They represent the first Time and the The longitudinal acceleration of the vehicle being measured at any given time; Indicates the sampling period; and They represent the first Time and the The lateral acceleration of the vehicle being measured at all times.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1) To address the issues of unbalanced testing difficulty, wasted resources, or security risks caused by static settings of existing testing parameters, this invention introduces an interactive difficulty parameter. D By establishing a monotonic mapping relationship between the test scenario and the game weights and the PET safety lower bound, and combining this with a closed-loop adjustment mechanism based on the actual response of the vehicle under test, the intensity of the background vehicle interaction during the test can be automatically adjusted to adapt to the test, eliminating the need for frequent manual script modifications or scenario reconstruction. On the one hand, the test difficulty can be dynamically increased to the capability boundary of the vehicle under test, avoiding the waste of test resources due to excessive safety margins and fully examining the system's boundary performance. On the other hand, relying on vehicle dynamics constraints, minimum safe distance constraints, and PET safety lower bound control, it ensures that the scenario adjustment does not exceed the physical safety tolerance, avoids the risk of real vehicle collisions, and achieves dual assurance of test effectiveness and safety.

[0060] 2) Addressing the issues in existing tests where simulated traffic participants fail to capture the dynamic game-playing characteristics of human drivers in real traffic environments, and where tested vehicles can rely on pre-stored strategies to pass tests, this invention constructs a game-playing model. Simulated traffic participants act as leaders, and tested vehicles as followers. It introduces multi-dimensional game weights such as risk, smoothness, and trajectory rationality to simulate the dynamic interactive behavior of real traffic participants, breaking through the limitations of fixed behavior in simulated traffic participants. Simultaneously, it introduces interaction capability evaluation indicators. JQuantifying the safety margin, traffic efficiency, and control smoothness of the vehicle under test can replace traditional extensive testing. It can accurately capture the decision boundaries of the vehicle under test in dynamic game scenarios, avoid the vehicle under test relying on pre-stored strategies to avoid test risks, and improve the effectiveness of testing.

[0061] 3) Furthermore, this invention utilizes a large language model to achieve strategy-level parameter tuning. By inputting multi-dimensional information such as scene semantics, interaction capability evaluation results, and current interaction difficulty parameters, the large language model intelligently outputs difficulty adjustment suggestions and reasons for parameter tuning. Compared to fixed rules or simple threshold strategies, it can make autonomous decisions by comprehensively considering multiple rounds of test history and scene characteristics. This not only improves the adaptability and intelligence of scene adjustment but also solves the problems of unexplainable and difficult-to-trace traditional automatic parameter tuning processes by explaining the parameter tuning logic through natural language, making it easier for testers to optimize test strategies. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method of the present invention;

[0063] Figure 2 The interaction difficulty parameter is used when the test scenario of this invention is an intersection. D Mapping relationship with game parameters;

[0064] Figure 3 This is a schematic diagram of the initial reference trajectory and the trajectory after difficulty adjustment when the test scenario of this invention is an intersection. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0066] To address the problems existing in the background technology, a closed-site autonomous driving testing technology is proposed. This technology can perceive the interaction capability boundaries of the vehicle under test in real time based on its actual response, introduce a large language model to dynamically adjust the interaction difficulty, and generate simulated traffic participant trajectories that combine physical safety and behavioral adversarial behavior. This system improves the sufficiency and reliability of autonomous driving system performance boundary testing. In this embodiment, the method includes an offline modeling stage and an online testing stage, wherein the online testing stage process is as follows: Figure 1 As shown.

[0067] The offline modeling stage provided in this embodiment includes the following steps:

[0068] T1. Scene selection and parameter initialization.

[0069] According to the testing requirements, a closed test scenario is selected from the scenario library, such as an unprotected left turn scenario. The scenario library pre-stores road geometry information, lane lines, traffic light status and corresponding reference trajectories obtained from simulation or real vehicle tests. The initial reference trajectory of the test vehicle and simulated traffic participants in this scenario is selected. The reference trajectory includes a series of waypoints, corresponding expected speeds and timestamps.

[0070] T2. Construct a Stackelberg game model of the tested vehicle and simulated traffic participants.

[0071] By simulating traffic participants as leaders and the tested vehicles as followers, a Stackelberg game model is established in the Frenet coordinate system in a structured road scenario.

[0072] Let the game interaction model be: ,in, Let represent the set of game participants. , Indicates the leader, Indicates a follower; Let the set of cost functions be . , This represents the cost function for simulating traffic participants. This represents the cost function of the vehicle being tested. The set of control inputs, including leader control variables and follower control variables, is represented as: , Indicates longitudinal acceleration. Indicates lateral acceleration; The set of state spaces, including leader states and follower states, is described by the same discrete-time state-space equation, which is: And there are:

[0073] ;

[0074] Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. This represents the discrete state transition function for simulating traffic participants or the vehicle being tested. Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. Indicates state, Indicates the longitudinal distance along the reference path. Indicates longitudinal velocity. This indicates the lateral deviation along the tangent of the reference path. Indicates lateral velocity. Indicates transpose; Let represent the state transition matrix, and:

[0075] ;

[0076] Indicates the sampling time interval for state updates; Let represent the control input matrix, and:

[0077] .

[0078] Furthermore, in this game model, the interaction behavior of the tested vehicle and the simulated traffic participants in each test round can be described as a two-level optimization problem. Based on this, the cost function of the tested vehicle and the cost function of the simulated traffic participants can be constructed. Specifically, in this embodiment, the cost function includes the cost function of the tested vehicle and the cost function of the simulated traffic participants. The cost function of the simulated traffic participants is a weighted sum of risk interaction cost, smoothness cost, and trajectory rationality cost. The game weight of each cost is calculated based on the interaction difficulty parameter.

[0079] 1) Cost function of the vehicle under test.

[0080] The tested vehicle acts as a follower F, and its actual control quantities (i.e., longitudinal acceleration and lateral acceleration) are provided by its own autonomous driving system. In this embodiment, control commands are not directly issued to the tested vehicle. To predict the tested vehicle's response to the behavior of simulated traffic participants, the tested vehicle can be modeled as minimizing its own cost function under vehicle dynamics and safety constraints. Rational followers.

[0081] Suppose the vehicle under test is at discrete time... The state vector is The control input is The scenario library provides the reference trajectory state of the tested vehicle as follows: In the planning time domain The optimization problem for the tested vehicle can be expressed as:

[0082] ;

[0083] ;

[0084] ;

[0085] in, This represents the follower cost function value; Indicates followers from the first Time to the The amount of control at any given moment Indicates the length of the test round; Indicates the number of followers The state at any given moment; Indicates that the follower is on the reference trajectory at the 1st The state at any given moment; Indicates transpose; and Each represents the positive semidefinite weight matrix of the followers; Indicates the number of followers The amount of control at any given moment; Indicates non-negative weighting coefficients; Indicates the first Physical safety margin at any given moment; Represents the discrete state transition function of the follower; Indicates the first The actual distance between the vehicle being tested and the nearest simulated traffic participant at any given time; and They represent the first The longitudinal speed of the vehicle being measured and the simulated traffic participants at any given time; Indicates response time; Indicates the minimum safe distance when stationary; This represents the minimum longitudinal acceleration of the follower.

[0086] 2) Simulate the cost function of traffic participants.

[0087] Simulated traffic participants as leaders Its goal is to actively approach the vehicle under test while meeting physical safety constraints. This approach aims to make the post-intrusion time (PET) of both parties as close as possible to the preset safety lower bound, thereby generating a challenging strong interaction boundary scenario. Simultaneously, to ensure the behavior of simulated traffic participants is reasonable and physically feasible, this embodiment comprehensively considers three aspects—risk, smoothness, and trajectory rationality—when constructing the interaction utility function, which can be expressed as:

[0088] Risk interaction cost: , Indicates the time of intrusion. Indicates followers. Indicates the leader, Indicates the potential collision point. Indicates the time it takes for the leader to arrive at the potential collision point. This indicates the time it takes for the follower to reach the potential collision point; in addition, safety constraints must be imposed as follows: Thus, in minimizing the leader Cost function At that time, simulated traffic participants meet the requirements Under certain conditions, the tendency is to make PET as close as possible to This refers to strong interaction with the vehicle under test F to perform boundary testing.

[0089] Smoothing cost: , Indicates the leader's longitudinal acceleration at time t, Indicates the leader's Lateral acceleration at any given moment.

[0090] Cost of trajectory rationality: , Indicates the leader's The state at any given moment, This indicates that the leader is on the reference trajectory. The state at any given moment; This represents a positive semidefinite matrix that weights the leader's longitudinal, lateral offset, and velocity errors.

[0091] Since the cost function for simulated traffic participants is a weighted sum of risk interaction costs, smoothness costs, and trajectory rationality costs, the cost function for simulated traffic participants can be constructed as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] in, Indicates that the leader is from the first Time to the The amount of control at any given moment Indicates the length of the test round; This represents the cost function for simulating traffic participants; , and The game weights for risk interaction cost, smoothness cost, and trajectory rationality cost are respectively represented, where the weights are... Used to balance adversarialness and reference trajectory plausibility: when the interaction difficulty parameter When increasing, it can be appropriately reduced. This allows simulated traffic participants to deviate from the reference trajectory while meeting basic scenario constraints, thereby enhancing the intensity of interaction; Indicates the interaction difficulty parameter. Values ​​are taken from a preset range ,and The larger the value, the higher the weight of risk interaction costs, and the constraints on smoothness and reference trajectory compliance can be appropriately relaxed to form more intense game behavior; Indicates the cost of risk interaction; Indicates the cost of smoothness; Indicates the cost of trajectory rationality; This represents the follower's optimal response. Indicates the leader's status. Indicates the follower state. Indicates the amount of control by followers; Indicates the leader in The amount of control at any given moment; This indicates the minimum time after the intrusion.

[0096] T3. Define interaction difficulty parameters And its mapping relationship with game weights.

[0097] In this embodiment, a large language model is set as the strategy generation module for interaction difficulty parameters. Based on the interaction ability evaluation indicators observed during the testing process, it automatically provides difficulty adjustment suggestions for the next period, thereby achieving intelligent adaptive adjustment of the game weight parameters. The large language model used can be an open-source or commercial general-purpose language model, and can provide inference capabilities through local deployment or remote inference services, without being limited to a specific model architecture.

[0098] The method for obtaining the interaction difficulty parameter using the Big Prophet model is as follows:

[0099] T31. At the end of each test round, obtain the current interaction capability evaluation index, the trajectory of the tested vehicle and simulated traffic participants, the minimum physical safety margin, the test results, and the corresponding scene semantic information, and perform feature extraction and encoding to obtain a structured description. .

[0100] Specifically, the coding is based on a prompt template, which includes: scene type, current interaction difficulty parameter value, evaluation index of interaction capability in this round of testing, minimum safety margin, whether a collision occurred, whether emergency braking was triggered, and the interaction difficulty parameter value of the previous round and its corresponding test results. For example: "Scene type: XX; Current difficulty..." The safety indicators for this round of testing are: The efficiency index is The stability index is The minimum safety margin is meters; Collision occurred: Yes / No; Emergency braking was triggered: Yes / No; Previous round difficulty: = The corresponding test result is "XX"; in this way, the large language model can simultaneously acquire scene semantic information and quantitative evaluation indicators.

[0101] The evaluation indicators for interaction capabilities are as follows:

[0102] ;

[0103] , and Let represent the non-negative weight coefficients, and . ; The security index, calculated based on the minimum post-intrusion time and physical security margin, can be expressed as:

[0104] ;

[0105] Indicates adjustable weights; followers and leaders Reaching the potential collision point Minimum post-intrusion time; This indicates a positive number that prevents the denominator from being zero; This represents the minimum value of the physical safety margin; This efficiency metric, used to characterize the passage efficiency of a tested vehicle in a given interaction scenario, is calculated based on the actual passage time and reference passage time of the tested vehicle in the current interaction segment, and can be expressed as:

[0106] ;

[0107] This indicates the actual travel time of the tested vehicle in the current interaction segment; Indicates the reference transit time of the tested vehicle in the current interaction segment; This represents a stability index used to characterize the smoothness of the vehicle's control. It is calculated based on the vehicle's longitudinal and lateral accelerations and can be expressed as:

[0108] ;

[0109] Indicates the total duration of the test rounds; and They represent the first Time and the The longitudinal acceleration of the vehicle being measured at any given time; Indicates the sampling period; and They represent the first Time and the The lateral acceleration of the vehicle being measured at all times.

[0110] In addition, to characterize boundary scenarios, interactive capability evaluation metrics will be used. Normalized to the [0,1] interval, assuming the results are based on historical tests or simulation statistics. , They are respectively The typical minimum and maximum values ​​are used to define the normalization index:

[0111] ;

[0112] in, This represents a normalized interaction capability evaluation index; and two thresholds are given during the offline modeling phase. and (0< < <1), then we have: when < At that time, it was considered that the current test difficulty was too low and the interaction was far from the capability boundary; when > At that time, it was believed that the current test was too aggressive or close to the physical limits, belonging to a high-risk range that exceeded the boundaries of capability; when ≤ ≤ At that time, the current scene is considered to be in the boundary area of ​​the interactive capabilities of the vehicle under test, and can be recorded and tested as a boundary scene.

[0113] Test results include whether emergency braking was triggered, whether a collision occurred, or whether human intervention was required.

[0114] Scene semantic information includes the number of lanes, whether it is a lane merging or intersection scene, speed limit level, weather and road conditions, etc.

[0115] T32. Based on structured description, use a large language model to output text-based difficulty adjustment suggestions, including the adjustment value of the interaction difficulty parameter for the next time step.

[0116] Detailed difficulty adjustment suggestions could be: "The current scene has a relatively large safety margin in its interaction; the difficulty can be appropriately increased." "Adjust to 0.6", or "This round of testing has triggered emergency braking and the safety margin is close to 0, so the difficulty should be reduced." Adjusted to 0.35.

[0117] Obtain new interaction difficulty parameters and truncate and smooth them so that they always remain within the preset range [0,1] and the difficulty change between two adjacent rounds does not exceed the given step size, thereby avoiding sudden changes in test conditions due to excessive adjustment.

[0118] Furthermore, after obtaining the new interaction difficulty parameters, a monotonic mapping relationship is used. The game weight parameters are updated in real time, and then the cost function is updated based on the updated game weight parameters. The detailed monotonic mapping relationship is as follows:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] in, , as well as The game weights represent the costs of risk interaction, smoothness, and trajectory rationality. Indicates the minimum post-intrusion correction time; and These represent the minimum and maximum values ​​of the risk interaction cost game weights obtained after calibration; Indicates the interaction difficulty parameter; and These represent the maximum and minimum values ​​of the smoothness cost game weights obtained after calibration; and These represent the maximum and minimum values ​​of the trajectory rationality cost game weights obtained after calibration; and Let represent the maximum and minimum values ​​of the minimum post-intrusion time obtained through calibration, respectively; and for all maximum and minimum values ​​obtained through calibration, we have: , , , .

[0125] Therefore, for any < For the game weights constructed above, we can obtain: , Therefore, a larger difficulty parameter D This inevitably corresponds to higher risk weights, more relaxed smoothness and reference trajectory constraints, and smaller safety time margins, thus manifesting as more intense game-theoretic behavior and increasing the interaction difficulty parameter. Combination of game parameters A monotonic one-to-one mapping relationship is formed between them, such as Figure 2 As shown.

[0126] Through the above monotonic mapping relationship, it is ensured that as... Increase the risk game weight ; follow Increase and moderately decrease the smoothness game weights Weights in the game of trajectory rationality or correspondingly reduce the lower safety limit of PET. This enhances the competitive aspect of simulated traffic participants.

[0127] By using the updated interaction difficulty parameters and corresponding game weights in the next game algorithm, this invention realizes an adaptive adjustment mechanism for interaction difficulty based on the actual response of the tested vehicle. This allows the testing process to automatically approach the interaction capability boundary of the tested vehicle while ensuring physical safety, and automatically slows down the increase in difficulty or appropriately reduces the difficulty when approaching the boundary, thereby avoiding the risk of real-vehicle collision caused by excessive aggression.

[0128] For the online testing phase, Figure 3 Taking the intersection shown as an example, we perform optimization game theory, where (11) represents the tested vehicle, (12) represents the simulated traffic participant, (21) represents the initial reference trajectory of the simulated traffic participant, and (23) represents the driving trajectory of the tested vehicle; taking the first... Taking a specific time as an example, perform the following steps:

[0129] S1, Obtain the first The pose information of the tested vehicle and simulated traffic participants at any time is used to calculate the interaction capability evaluation index based on the pose information, and the interaction difficulty parameter corresponding to the current time is obtained based on the interaction capability evaluation index.

[0130] In detail, the interaction capability evaluation indicators, the trajectories of the tested vehicle and simulated traffic participants, the minimum physical safety margin, the test results, and the corresponding scene semantic information are input into the large language model. Based on the interaction difficulty parameter adjustment suggestions output by the large language model, the current interaction difficulty parameters are determined. Furthermore, the initial values ​​should be used in the first test round. .

[0131] S2. Update the cost function based on the interaction difficulty parameter at the current moment, and use a game theory algorithm to solve the updated cost function to obtain the reference trajectory of the simulated traffic participants.

[0132] In detail, within each planning time domain, given the initial state and current interaction difficulty parameters The Stackelberg game described above can be solved using the following steps:

[0133] S21. Given the leader control sequence, solve the follower cost function to obtain the optimal response control sequence of the tested vehicle. .

[0134] S22, Considering the optimal response of the follower Based on this, minimize the leader's cost function. The optimal control sequence for the leader is obtained by solving the problem. .

[0135] by Figure 3 To explain, the trajectory of the simulated traffic participant after optimization is represented by the curve (22), and the final collision point between the simulated traffic participant and the tested vehicle is the point shown in (CP).

[0136] In practical implementation, the bi-level optimization problem can be solved by iterative solution, equivalent single-level transformation or other numerical optimization methods. This invention does not limit the specific solution algorithm. For example, the bi-level optimization problem can be solved by a single-level transformation method based on KKT conditions, and the optimal reference trajectory and control input of the simulated traffic participants in the current planning time domain can be obtained by using the Ipopt solver.

[0137] S3. Obtain the pose information of the simulated traffic participants after executing the reference trajectory, calculate the interaction capability evaluation index based on the pose information. If the interaction capability evaluation index meets the preset boundary criteria, the reference trajectory is used as the test trajectory; otherwise, return to execute S1 based on the current pose information.

[0138] In detail, if the final calculated interaction capability index meets the following requirements: The key information within the current planning time domain will be recorded as boundary scenarios, including at least: the actual trajectories of the tested vehicle and simulated traffic participants; the corresponding control input sequences and interaction difficulty parameters. Interaction capability evaluation indicators J Its components; scene semantic information and road structure information, etc.; the above information is added to the scene library as new boundary scene samples for repeated calls in subsequent regression testing and capability assessment, and can also be used as training data or retrieval samples for subsequent large language model parameter tuning modules.

[0139] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty, characterized in that, For the At any given time, the method includes: S1. Obtain the pose information of the tested vehicle and simulated traffic participants at the current moment, calculate the interaction capability evaluation index based on the pose information, and obtain the interaction difficulty parameter corresponding to the current moment based on the interaction capability evaluation index. S2. Update the cost function based on the interaction difficulty parameter at the current moment, and solve the updated cost function using a game theory algorithm to obtain the reference trajectory of the simulated traffic participant; the cost function includes the cost function of the tested vehicle and the cost function of the simulated traffic participant, and the cost function of the simulated traffic participant is a weighted sum of risk interaction cost, smoothness cost and trajectory rationality cost, and the game weight of each cost is calculated based on the interaction difficulty parameter. S3. Obtain the pose information of the simulated traffic participants after executing the reference trajectory, calculate the interaction capability evaluation index based on the pose information, and if the interaction capability evaluation index meets the preset boundary criteria, then the reference trajectory is used as the test trajectory; otherwise, return to execute S1 based on the current pose information. The method for obtaining the interaction difficulty parameter is as follows: The system acquires the current interaction capability evaluation index, the trajectory of the tested vehicle and simulated traffic participants, the minimum physical safety margin, the test results, and the corresponding scene semantic information, and performs feature extraction and encoding to obtain a structured description. The encoding is based on a prompt template, which includes: scene type, current interaction difficulty parameter value, interaction capability evaluation index of this round of test, minimum safety margin, whether a collision has occurred, whether emergency braking has been triggered, and the interaction difficulty parameter value of the previous round and its corresponding test results. Based on the structured description, a large language model is used to output text-based difficulty adjustment suggestions, which include the adjustment value of the interaction difficulty parameter at the next time step.

2. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 1, characterized in that, The tested vehicles and simulated traffic participants include: The simulated traffic participants are treated as leaders, and the vehicles under test are treated as followers. The control quantities for both leaders and followers include longitudinal acceleration and lateral acceleration, expressed as: , Indicates longitudinal acceleration. Indicates lateral acceleration; The leader state and follower state are both represented as follows: And the discrete state transition function is: ； Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. This represents the discrete state transition function for simulating traffic participants or the vehicle being tested. Indicates the first Simulate the state of traffic participants or the tested vehicle at any given time. Indicates the first The control quantities of simulated traffic participants or the tested vehicles at any given time. Indicates the longitudinal distance along the reference path. Indicates longitudinal velocity. This indicates the lateral deviation along the tangent of the reference path. Indicates lateral velocity. Indicates transpose; Let represent the state transition matrix, and: ； Indicates a time interval; Let represent the control input matrix, and: 。 3. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 1, characterized in that, The cost function for the tested vehicle is: ； ； ； in, This represents the follower cost function value; Indicates followers from the first Time to the The amount of control at any given moment Indicates the length of the test round; Indicates the number of followers The state at any given moment; Indicates that the follower is on the reference trajectory at the 1st The state at any given moment; Indicates transpose; and Each represents the positive semidefinite weight matrix of the followers; Indicates the number of followers The amount of control at any given moment; Indicates non-negative weighting coefficients; Indicates the first Physical safety margin at any given moment; Represents the discrete state transition function of the follower; Indicates the first The actual distance between the vehicle being tested and the nearest simulated traffic participant at any given time; and They represent the first The longitudinal speed of the vehicle being measured and the simulated traffic participants at any given time; Indicates response time; Indicates the minimum safe distance when stationary; This represents the minimum longitudinal acceleration of the follower.

4. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 1, characterized in that, The cost function for the simulated traffic participants is: ； ； ； in, Indicates that the leader is from the first Time to the The amount of control at any given moment Indicates the length of the test round; This represents the cost function for simulating traffic participants; , and These represent the game weights of risk interaction cost, smoothness cost, and trajectory rationality cost, respectively. Indicates the interaction difficulty parameter; Indicates the cost of risk interaction, and: ； Indicates the time of intrusion. Indicates followers. Indicates the leader, Indicates the potential collision point. Indicates the time it takes for the leader to arrive at the potential collision point. Indicates the time it takes for the follower to reach the potential collision point; Let $\frac{ ... ； Indicates the leader's longitudinal acceleration at time t, Indicates the leader's Lateral acceleration at any given moment; Let $\math$ represent the cost of trajectory rationality, and: ； Indicates the leader's The state at any given moment, This indicates that the leader is on the reference trajectory. The state at any given moment; This represents a positive semidefinite matrix that the leader uses to weight longitudinal, lateral offset, and velocity error. This represents the follower's optimal response. Indicates the leader's status. Indicates the follower state. Indicates the amount of control by followers; Indicates the leader in The amount of control at any given moment; This indicates the minimum time after the intrusion.

5. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 1, characterized in that, The method for updating the cost function is as follows: Based on the The interaction difficulty parameter at each moment is used to update the game weights and the minimum post-intrusion time using a monotonic mapping relationship; the monotonic mapping relationship is as follows: ； ； ； ； ； in, , as well as The game weights represent the costs of risk interaction, smoothness, and trajectory rationality. Indicates the minimum time after intrusion; and These represent the minimum and maximum values ​​of the risk interaction cost game weights obtained after calibration; Indicates the interaction difficulty parameter; and These represent the maximum and minimum values ​​of the smoothness cost game weights obtained after calibration; and These represent the maximum and minimum values ​​of the trajectory rationality cost game weights obtained after calibration; and These represent the maximum and minimum values ​​of the minimum post-intrusion time obtained through calibration, respectively.

6. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 1, characterized in that, The method for calculating the aforementioned interaction capability evaluation index is as follows: ； in, , and Let represent the non-negative weight coefficients, and . ; The security index is calculated based on the minimum post-intrusion time and physical security margin. The efficiency index is calculated based on the actual travel time and reference travel time of the tested vehicle in the current interaction segment. The stability index is calculated based on the longitudinal and lateral accelerations of the vehicle under test.

7. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 6, characterized in that, The method for calculating the aforementioned safety indicators is as follows: ； in, Indicates adjustable weights; followers and leaders Reaching the potential collision point Minimum post-intrusion time; This indicates a positive number that prevents the denominator from being zero; This represents the minimum value of the physical security margin.

8. The closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 6, characterized in that, The method for calculating the efficiency index is as follows: ； in, This indicates the actual travel time of the tested vehicle in the current interaction segment; This indicates the reference transit time of the tested vehicle in the current interaction segment.

9. A closed-course autonomous vehicle testing method with adaptive adjustment of interaction difficulty according to claim 6, characterized in that, The method for calculating the stability index is as follows: ； in, Indicates the total duration of the test rounds; and They represent the first Time and the The longitudinal acceleration of the vehicle being measured at any given time; Indicates the sampling period; and They represent the first Time and the The lateral acceleration of the vehicle being measured at all times.

Images