A long tail driving scene generation method based on neural reachability boundary guidance

Abstract

The present application belongs to the technical field of automatic driving car testing, and in particular relates to a long-tail driving scene generation method based on neural reachability boundary guidance. The steps include: S1, establishing a joint traffic dynamics and safety objective function; S2, training and conservatively calibrating a neural reachability safety value function to obtain a continuous safety margin; S3, training a natural driving diffusion model with road and dynamics constraints; S4, introducing a critical safety boundary guide in diffusion denoising, so that the scene converges to a critical state with a safety margin close to zero, and the naturalness, feasibility and diversity are considered; S5, closed-loop testing is carried out, scenes that are theoretically solvable but fail in the tested system are selected as defect evidence, and the minimum counterfactual safety modification is output. The present application generates physically feasible and challenging long-tail scenes through boundary attraction rather than one-way danger maximization, supporting the safety verification and industrialization landing of the automatic driving system.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle testing technology, specifically a method for generating long-tail driving scenarios based on neural reachability boundary guidance. Background Technology

[0002] Safety verification for autonomous vehicles requires identifying low-probability, high-risk long-tail traffic scenarios within a limited testing budget. Natural driving data contains a sparse number of safety-critical interactions, and direct random replay is insufficient to cover the system's capability boundaries. Therefore, parameter optimization, adversarial search, or generative models are typically used to construct hazardous scenarios.

[0003] Existing methods often use collision rate, minimum collision time, or planning loss as the sole objective, easily generating scenarios where background traffic vehicles actively collide, the lead vehicle cannot avoid collisions regardless of its behavior, or its trajectory violates dynamic constraints. While such scenarios are dangerous, the danger is already unavoidable at the start of the test, making it difficult to prove that the tested system has any improveable capability defects. Physical constraints can guarantee the continuity of speed, acceleration, and steering, but cannot determine whether feasible avoidance strategies still exist given control boundaries, road adhesion, and response delays. Data-driven diffusion models can maintain the naturalness and diversity of traffic behavior, but critical risk samples in natural distributions are still scarce, and relying solely on random diffusion sampling is inefficient. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a long-tail driving scenario generation method based on neural reachability boundary guidance. It employs a neural Hamilton-Jacobi reachability model to calculate the vehicle's safety margin under the current traffic conditions, defining a state with a safety margin close to zero as a critical state. Near this state, the vehicle can still avoid danger through reasonable braking, steering, and other control operations. As the safety margin continues to decrease, the vehicle's hazard avoidance space gradually shrinks, and the danger may eventually become unavoidable. During the natural driving diffusion sampling process, boundary attraction is used instead of unidirectional hazard maximization. After closed-loop testing, only scenarios that are theoretically solvable but where the tested system fails are retained, while the minimum counterfactual safety modification is output. The results of this invention can support the safety verification of autonomous vehicles in large-scale physically feasible long-tail scenarios, supporting the industrialization of autonomous vehicles.

[0005] The technical solution of this invention is described below in conjunction with the accompanying drawings:

[0006] This invention provides a method for generating long-tail driving scenarios based on neural reachability boundary guidance, comprising the following steps:

[0007] S1. Establish the joint traffic dynamics and safety objective function;

[0008] Based on typical driving scenarios, the relative positions, speeds, headings, control inputs, and bounded disturbances of the vehicle and key traffic participants are extracted to establish a joint dynamic system and a safety objective function.

[0009] S2. Train and conservatively calibrate the neural reachability safety value function;

[0010] The neural safety value function is trained using Hamilton-Jacobi-Isaacs partial differential equation residuals, terminal conditions, and state domain boundaries. Conservative calibration is performed using the error quantiles of a low-dimensional grid reachability benchmark to obtain continuous safety margins and state gradients.

[0011] S3. Train a natural driving diffusion model with road and dynamic constraints;

[0012] Learn the natural distribution of multi-agent trajectories and initial scene conditions from real driving logs, and eliminate unreasonable trajectories through discrete kinematics, road boundaries, and comfort constraints;

[0013] S4. Perform critical safety boundary guidance during diffusion denoising;

[0014] The generated trajectory is converted into a joint dynamic state sequence, and the minimum safety margin and gradient are calculated according to the conservative safety value function to make the scene converge to the specified critical risk zone while maintaining naturalness, physical feasibility and diversity.

[0015] S5. Conduct closed-loop testing, solveable failure screening, and counterfactual diagnosis;

[0016] In closed-loop simulation, generated scenarios are executed. Based on theoretical solvability and the classification of the test system results, only samples where reasonable avoidance strategies still exist but the test system fails are considered as core defect evidence. The minimum scenario modification that restores the system to safety is then searched.

[0017] Furthermore, the specific method of S1 is as follows:

[0018] S11. Constructing joint traffic state and dynamics;

[0019] The dynamics of the combined transportation system is expressed as equation (1):

[0020] (1)

[0021] In the formula, The joint state vector includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and necessary response delay between the vehicle and key traffic participants. Regarding continuous physical time The derivative; For the dynamics of vehicle-traffic interaction; This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; For continuous physical time;

[0022] S12. Construct the security objective function;

[0023] The terminal safety objective function is constructed based on the vehicle surround box spacing, as shown in equation (2):

[0024] (2)

[0025] In the formula, Joint state vector The terminal security objective function; For state The minimum bounding box distance between the corresponding vehicle and key traffic participants; To preset a safe distance threshold; When the value is less than or equal to zero, the state belongs to the target set that is either in collision range or has insufficient safety distance.

[0026] Furthermore, the specific method of S2 is as follows:

[0027] S21. Establish the Hamilton-Jacobi-Isaacs security value function;

[0028] The safety value function satisfies the Hamilton-Jacobi-Isaacs equation shown in equation (3):

[0029] (3)

[0030] In the formula, The safety value function corresponding to the joint state and remaining time; The remaining time from the current state to the end of the evaluation; for about The partial derivatives; min is the minimization operation for the vehicle to choose favorable control within the permissible control domain; max is the maximization operation for other participants and disturbances to choose unfavorable effects within their respective permissible domains; This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; for Regarding the joint state vector gradient; superscript Transpose of a vector; It is the dot product of vectors;

[0031] The neural safety value function adopts the terminal condition shown in equation (4):

[0032] (4)

[0033] In the formula, The neural safety value when the remaining time equals zero; These are all the trainable parameters of the neural safety value network; It is a joint state vector that includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and response delay between the vehicle and key traffic participants; Joint state vector The terminal security objective function;

[0034] S22. Train and conservatively calibrate neural value functions;

[0035] The training objective of the neural value function is expressed as equation (5):

[0036] (5)

[0037] In the formula, Overall training loss in neural accessibility; The partial differential equation residuals of the left-hand side of equation (3) relative to zero; The weight of the partial differential equation residuals on the left-hand side of equation (3) relative to zero; This is the terminal condition error corresponding to equation (4); The weight of the terminal condition error corresponding to equation (4); This represents the boundary constraint error when the state exceeds the training domain. The weights for the boundary constraint error when the state exceeds the training domain; The physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time; The weights of the physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time;

[0038] A conservative safety value is obtained using the benchmark error quantile, as shown in equation (6):

[0039] (6)

[0040] In the formula, For the joint state and remaining time The corresponding conservative safety value; For parameters A defined neural safety value function output; For the grid reachability benchmark and the absolute error of the neural value function quantiles; For the allowable calibration failure rate;

[0041] In low-dimensional main scenes where mesh solving is feasible, the neural value function is compared with the mesh Hamilton-Jacobi results, and the errors in the overall region and the critical boundary neighborhood are statistically analyzed separately; conservative safety values ​​are uniformly used for scene generation. The road adhesion coefficient, maximum braking capacity, maximum steering ratio, and control delay are used as input conditions for the value function to support different vehicle capabilities and road conditions.

[0042] Furthermore, the specific method of S3 is as follows:

[0043] S31. Establish the representation and diffusion process of natural driving trajectory;

[0044] No. The traffic participants in the first The state of each discrete trajectory at each time point is represented by equation (7):

[0045] (7)

[0046] In the formula, For the first The traffic participants in the first The state vector at each discrete trajectory moment; and These refer to the longitudinal and transverse positions in the plane, respectively. Scalar velocity; It is longitudinal acceleration; For heading angle; This refers to the yaw rate; For traffic participant serial numbers; For discrete trajectory time indices; superscript Transpose of a vector;

[0047] Multi-agent trajectory of the target Perform forward noise addition, as shown in equation (8):

[0048] (8)

[0049] In the formula, For the first Latent variables of noise trajectory at each diffusion level; For the first The cumulative trajectory fidelity coefficient for each diffusion level, with values ​​ranging from 0 to 1; For noise-free target multi-agent trajectory; To and Standard Gaussian noise vectors of the same dimension; The diffusion level number is used; the square root is used to ensure that the signal and noise terms meet the predetermined variance schedule.

[0050] The natural driving diffusion model uses the denoising loss shown in equation (9):

[0051] (9)

[0052] In the formula, This is the denoising loss for the natural driving diffusion model; To address training scenarios and diffusion levels and random noise Find the expected value; For parameters A defined noise prediction network; These are all the trainable parameters of the diffusion model; For road map condition tensors; For the historical multi-agent trajectory condition tensor; This is a vector of conditions for scenario type and target risk level; The square of the second norm of a vector;

[0053] S32. Constraints on the physical executability of the generated trajectory;

[0054] The trajectory physical residual is calculated by equation (10):

[0055] (10)

[0056] In the formula, The discrete dynamic residuals of the trajectory; ∑ represents the summation over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. , and The first A vector is obtained by stacking the positions, velocities, and accelerations of all traffic participants at each moment, where The component comes from equation (7). and ; For the first From the moment to the first The sampling interval at each time point; Update the residual weights to speed; It is the square of the second norm of the vector.

[0057] Furthermore, the specific method of S4 is as follows:

[0058] S41. Calculate the continuous critical safety margin of the generated trajectory;

[0059] The minimum safety margin of the entire trajectory is obtained by using the smoothing minimum operation, as shown in equation (11):

[0060] (11)

[0061] In the formula, To generate trajectory The smooth minimum conservative safety margin; The smoothing temperature coefficient is positive. is the natural logarithm function; exp is the natural exponential function; ∑ is the summation over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. Index for discrete sampling time; A conservative, safe value; For the trajectory In the The joint traffic state vector obtained by converting the index of discrete sampling time points; For the first The remaining time corresponding to each discrete sampling time index; For noise-free target multi-agent trajectory;

[0062] With target safety margin Construct the attraction function of the critical boundary, as shown in equation (12):

[0063] (12)

[0064] In the formula, For trajectory The critical boundary attracts the target; To predetermine the target safety margin; squaring penalizes scenarios where the safety margin deviates from the target value, and a leading negative sign maximizes the safety margin. Equivalent to approaching the target safety margin;

[0065] S42. Integrating risk, road, physical, and diversity guidance in diffusion sampling;

[0066] The synthesis guidance direction of the kth denoising level is expressed as Equation (13):

[0067] (13)

[0068] In the formula, For the first The comprehensive trajectory guidance direction of each noise reduction level; The basic denoising direction given for the natural diffusion model; , , and The first The boundary, path rules, dynamics, and novelty-guided weights of each denoising level; To track Find the gradient; Losses related to lane boundaries, right-of-way, and traffic rules; The reward is for novelty relative to the existing library of failure scenarios;

[0069] The scene novelty reward is calculated by equation (14):

[0070] (14)

[0071] In the formula, Candidate trajectories Novelty reward; This is for calculating the minimum value; This refers to the sequence number of the existing failed scenarios; For existing failure scenario libraries; For the first in the failure scenario library Trajectory; The trajectory embedding distance function is constructed based on the safety value curve, key relative state, response delay, and control action. For noise-free target multi-agent trajectory;

[0072] Furthermore, the specific method of S5 is as follows:

[0073] S51. In closed-loop simulation, the determination of theoretical solvability fails.

[0074] The high-value failure indicator of the candidate scenario is expressed as Equation (15):

[0075] (15)

[0076] In the formula, Candidate trajectories High-value solvable failure indicators; This is an indicator function that takes a value of 1 if the condition is true and 0 otherwise. The minimum conservative safety margin to be recalculated using the actual closed-loop trajectory of the system under test; The threshold for determining whether a solution is theoretically solvable; This is the failure indication function for the autonomous driving system under test. It is set to 1 when a collision, boundary violation, unacceptable emergency braking, or mission failure occurs, and 0 otherwise. For noise-free target multi-agent trajectory;

[0077] Theoretically solvable and system-successful samples serve as evidence of system pass and capability boundary samples; theoretically solvable but system-failed samples serve as evidence of core defects; theoretically unsolvable and system-failed samples are only used for safety boundary calibration and cannot be directly attributed to the tested system; theoretically unsolvable but system-successful samples are used to check whether the value function is overly conservative or whether the system control capability exceeds the modeling range. Here, system success means that the tested autonomous driving system does not experience a collision, unacceptable emergency braking, or driving task failure during the closed-loop simulation and completes the predetermined driving task; otherwise, it is judged as system failure.

[0078] S52, Minimal counterfactual safety modifications for search failure scenarios;

[0079] For theoretically solvable failure scenarios, search for the minimum safe counterfactual, as shown in equation (16):

[0080] (16)

[0081] In the formula, The minimum scenario parameter modification vector required to ensure successful recovery of the tested system while maintaining theoretical solvability; The operation corresponding to the independent variable is performed to find the minimum value of the objective function; This is the original scene parameter vector; Modify the vector for the scene parameters to be optimized; This is a diagonal weight matrix used to normalize scene parameters with different dimensions. Let be the vector 2 norm; st represents the constraint that follows it. The trajectory generated and executed in closed loop using the modified scene parameters;

[0082] Equation (17) is used to evaluate the efficiency of solvable failure detection:

[0083] (17)

[0084] In the formula, To improve the efficiency of solving failure detection; The number of theoretically solvable and reproducible high-value system failure scenarios discovered within a given testing budget; This represents the total number of candidate scenarios for closed-loop execution. The larger the value, the more efficient it is at identifying system capability gaps under the same simulation budget;

[0085] Equation (18) is used to evaluate the critical boundary hit rate:

[0086] (18)

[0087] In the formula, This refers to the critical boundary hit rate. To ensure the minimum safety margin falls within the target range The number of generated scenes; The total number of valid candidates that pass the naturalness, road rule, and dynamics checks; To preset the target safety margin; The target risk zone is half-width;

[0088] High-value failure scenarios are clustered according to safety value curves, key relative states, system response delays and control actions. For the same cluster, the sample with the smallest trajectory embedding distance from the cluster center, the smooth minimum conservative safety margin located within the target risk zone, and the ability to reproduce system failure through repeated closed-loop simulation are selected as representative boundary samples. Several counterfactual neighborhood samples obtained by adjusting scenario parameters are retained near the sample. The next round of generation prioritizes exploring uncovered mechanisms through the novelty reward of formula (14), forming an iterative test process of generation, closed-loop execution, failure diagnosis and active updating.

[0089] During implementation, scenario variables include initial longitudinal spacing, initial lateral spacing, speeds of both parties, cut-in start time, lateral speed of the cutting vehicle, maximum braking capacity of the vehicle, road adhesion coefficient, and total perception-control time delay. First, a low-dimensional grid reachability solver is used to obtain the baseline safety boundary, then the neural value function is trained and sampling is made more densely in the boundary neighborhood. The diffusion model learns the natural distribution of cut-in timing, speed changes, and lateral motion from real cut-in segments.

[0090] After closed-loop execution, the assumed vehicle response in the generation phase is replaced with the actual vehicle trajectory, the joint state is reconstructed, and a conservative safety value is calculated. If the generation phase determines that the solution is solvable but the actual closed-loop trajectory enters the danger zone, it is necessary to distinguish between the state deterioration caused by the response of the tested system and the neural value function estimation error. The state deterioration is considered a valid failure, and the neural value function estimation error is entered into the boundary calibration set. All core defect samples are executed at least a preset number of times, and only when the failure results and key risk indicators appear stably are they entered into the high-value scenario library.

[0091] The beneficial effects of this invention are as follows:

[0092] This invention provides a long-tail driving scenario generation method based on neural reachability boundary guidance. It uses a neural Hamilton-Jacobi reachability model to calculate the vehicle's safety margin under the current traffic conditions, defining a state with a safety margin close to zero as a critical state. Near this state, the vehicle can still avoid danger through reasonable braking, steering, and other control operations. As the safety margin continues to decrease, the vehicle's hazard avoidance space gradually shrinks, and the danger may eventually become unavoidable. The method employs boundary attraction rather than unidirectional hazard maximization during natural driving diffusion sampling, and after closed-loop testing, only scenarios that are theoretically solvable but where the tested system fails are retained, while outputting the minimum counterfactual safety modification. The results of this invention can support the safety verification of autonomous vehicles in large-scale physically feasible long-tail scenarios, supporting the industrialization of autonomous vehicles. Attached Figure Description

[0093] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0094] Figure 1 This is a flowchart of the present invention;

[0095] Figure 2 A schematic diagram defining the safety boundary and risk zone;

[0096] Figure 3 This is a schematic diagram of the theoretically solvable failure screening matrix. Detailed Implementation

[0097] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0098] Example 1

[0099] See Figure 1 This invention provides a method for generating long-tail driving scenarios based on neural reachability boundary guidance, comprising the following steps:

[0100] S1. Establish the joint traffic dynamics and safety objective function;

[0101] Based on typical driving scenarios, the relative positions, speeds, headings, control inputs, and bounded disturbances of the vehicle and key traffic participants are extracted to establish a joint dynamic system and a safety objective function. The specific method is as follows:

[0102] S11. Constructing joint traffic state and dynamics;

[0103] The dynamics of the combined transportation system is expressed as equation (1):

[0104] (1)

[0105] In the formula, The joint state vector includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and necessary response delay between the vehicle and key traffic participants. Regarding continuous physical time The derivative; For the dynamics of vehicle-traffic interaction; This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; For continuous physical time;

[0106] S12. Construct the security objective function;

[0107] The terminal safety objective function is constructed based on the vehicle surround box spacing, as shown in equation (2):

[0108] (2)

[0109] In the formula, Joint state vector The terminal security objective function; For state The minimum bounding box distance between the corresponding vehicle and key traffic participants; To preset a safe distance threshold; When the value is less than or equal to zero, the state belongs to the target set that is either in collision range or has insufficient safety distance.

[0110] S2. Train and conservatively calibrate the neural reachability safety value function;

[0111] The neural safety value function is trained using the Hamilton-Jacobi-Isaacs partial differential equation residuals, terminal conditions, and state domain boundaries. A conservative calibration is then performed using the error quantiles of a low-dimensional grid reachability benchmark to obtain continuous safety margins and state gradients. The specific method is as follows:

[0112] S21. Establish the Hamilton-Jacobi-Isaacs security value function;

[0113] The safety value function satisfies the Hamilton-Jacobi-Isaacs equation shown in equation (3):

[0114] (3)

[0115] In the formula, The safety value function corresponding to the joint state and remaining time; The remaining time from the current state to the end of the evaluation; for about The partial derivatives; min is the minimization operation for the vehicle to choose favorable control within the permissible control domain; max is the maximization operation for other participants and disturbances to choose unfavorable effects within their respective permissible domains; This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; for Regarding the joint state vector gradient; superscript Transpose of a vector; It is the dot product of vectors;

[0116] The neural safety value function adopts the terminal condition shown in equation (4):

[0117] (4)

[0118] In the formula, The neural safety value when the remaining time equals zero; These are all the trainable parameters of the neural safety value network; It is a joint state vector that includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and response delay between the vehicle and key traffic participants; When the value is less than or equal to zero, the state belongs to the set of targets that are either in collision range or have insufficient safety distance.

[0119] S22. Train and conservatively calibrate neural value functions;

[0120] The training objective of the neural value function is expressed as equation (5):

[0121] (5)

[0122] In the formula, Overall training loss in neural accessibility; The partial differential equation residuals of the left-hand side of equation (3) relative to zero; The weight of the partial differential equation residuals on the left-hand side of equation (3) relative to zero; This is the terminal condition error corresponding to equation (4); The weight of the terminal condition error corresponding to equation (4); This represents the boundary constraint error when the state exceeds the training domain. The weights for the boundary constraint error when the state exceeds the training domain; The physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time; The weights of the physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time;

[0123] A conservative safety value is obtained using the benchmark error quantile, as shown in equation (6):

[0124] (6)

[0125] In the formula, For the joint state and remaining time The corresponding conservative safety value; For parameters A defined neural safety value function output; For the grid reachability benchmark and the absolute error of the neural value function quantiles; The allowable calibration failure rate.

[0126] Based on the conservative safety value obtained from equation (6), when the conservative safety value is greater than the preset positive threshold, the system has a significant safety recovery margin; when the conservative safety value is greater than or equal to zero and not greater than the preset positive threshold, the state is located near the safety side of the zero level set, belonging to the critical solvable risk zone; when the conservative safety value is less than zero, under the given dynamics, control boundary, and disturbance assumptions, the danger is judged as unavoidable. The main test target is located near the safety side of the zero level set, rather than entering the unavoidable danger zone; on the low-dimensional main scene that can be solved by mesh, the neural value function is compared with the mesh Hamilton-Jacobi result, and the errors of the overall region and the critical boundary neighborhood are statistically analyzed respectively. The conservative safety value is used uniformly for scene generation. This reduces the risk of misjudging theoretically unsolvable states as solvable ones. Road adhesion coefficient, maximum braking capacity, maximum steering ratio, and control delay can be used as inputs to the value function to support different vehicle capabilities and road conditions.

[0127] S3. Train a natural driving diffusion model with road and dynamic constraints;

[0128] The method learns the natural distribution of multi-agent trajectories and initial scene conditions from real driving logs, and eliminates unreasonable trajectories through discrete kinematics, road boundaries, and comfort constraints. The specific method is as follows:

[0129] S31. Establish the representation and diffusion process of natural driving trajectory;

[0130] No. The traffic participants in the first The state of each discrete trajectory at each time point is represented by equation (7):

[0131] (7)

[0132] In the formula, For the first The traffic participants in the first The state vector at each discrete trajectory moment; and These refer to the longitudinal and transverse positions in the plane, respectively. Scalar velocity; It is longitudinal acceleration; For heading angle; This refers to the yaw rate; For traffic participant serial numbers; For discrete trajectory time indices; superscript Transpose of a vector;

[0133] The conditions for the diffusion model include road centerline, lane topology, historical trajectories, target scene type, and risk level. The map uses a polyline encoder to generate road markings, and multi-agent trajectories are represented using object labels. Training data comes from natural interactions in real driving logs; oversampling collisions are not forced during the training phase to ensure that the diffusion prior preserves the distribution of real traffic behavior.

[0134] Multi-agent trajectory of the target Perform forward noise addition, as shown in equation (8):

[0135] (8)

[0136] In the formula, For the first Latent variables of noise trajectory at each diffusion level; For the first The cumulative trajectory fidelity coefficient for each diffusion level, with values ​​ranging from 0 to 1; For noise-free target multi-agent trajectory; To and Standard Gaussian noise vectors of the same dimension; The diffusion level number is used; the square root is used to ensure that the signal and noise terms meet the predetermined variance schedule.

[0137] The natural driving diffusion model uses the denoising loss shown in equation (9):

[0138] (9)

[0139] In the formula, This is the denoising loss for the natural driving diffusion model; To address training scenarios and diffusion levels and random noise Find the expected value; For parameters A defined noise prediction network; These are all the trainable parameters of the diffusion model; For road map condition tensors; For the historical multi-agent trajectory condition tensor; This is a vector of conditions for scenario type and target risk level; The square of the second norm of a vector;

[0140] S32. Constraints on the physical executability of the generated trajectory;

[0141] The trajectory physical residual is calculated by equation (10):

[0142] (10)

[0143] In the formula, For trajectory discrete dynamic residuals; To sum over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. , and The first A vector is obtained by stacking the positions, velocities, and accelerations of all traffic participants at each moment, where The component comes from equation (7). and ; For the first From the moment to the first The sampling interval at each time point; Update the residual weights to speed; It is the square of the second norm of the vector.

[0144] S4. Perform critical safety boundary guidance during diffusion denoising;

[0145] The generated trajectory is converted into a joint dynamic state sequence. The minimum safety margin and gradient are calculated based on a conservative safety value function to make the scenario converge toward a specified critical risk zone, while maintaining naturalness, physical feasibility, and diversity. The specific method is as follows:

[0146] S41. Calculate the continuous critical safety margin of the generated trajectory;

[0147] The minimum safety margin of the entire trajectory is obtained by using the smoothing minimum operation, as shown in equation (11):

[0148] (11)

[0149] In the formula, To generate trajectory The smooth minimum conservative safety margin; The smoothing temperature coefficient is positive. exp is the natural logarithm function; exp is the natural exponential function. To sum over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. Index for discrete sampling time; A conservative, safe value; For the trajectory In the The joint traffic state vector obtained by converting the index of discrete sampling time points; For the first The remaining time corresponding to each discrete sampling time index; For noise-free target multi-agent trajectory;

[0150] With target safety margin Construct the attraction function of the critical boundary, as shown in equation (12):

[0151] (12)

[0152] In the formula, For trajectory The critical boundary attracts the target; To predetermine the target safety margin, a smaller positive value indicates a critical but still recoverable condition, while a value near zero indicates a near-critical condition. The squaring operation penalizes scenarios where the safety margin deviates from the target value, and a leading negative sign maximizes it. Equivalent to approaching the target safety margin;

[0153] Unlike directly minimizing the safety value, Equation (12) simultaneously suppresses both overly safe and overly dangerous scenarios, keeping the samples within the target safety margin. A defined critical risk zone is established to avoid generating a large number of theoretically inevitable collision scenarios. This is achieved by changing... It can form a continuous difficulty level from obviously safe, stressed, critically solvable to boundary calibration.

[0154] S42. Integrating risk, road, physical, and diversity guidance in diffusion sampling;

[0155] The synthesis guidance direction of the kth denoising level is expressed as Equation (13):

[0156] (13)

[0157] In the formula, For the first The comprehensive trajectory guidance direction of each noise reduction level; The basic denoising direction given for the natural diffusion model; , , and The first The boundary, path rules, dynamics, and novelty-guided weights of each denoising level; To track Find the gradient; Losses related to lane boundaries, right-of-way, and traffic rules; The reward is for novelty relative to the existing library of failure scenarios;

[0158] The scene novelty reward is calculated by equation (14):

[0159] (14)

[0160] In the formula, Candidate trajectories Novelty reward; This is for calculating the minimum value; This refers to the sequence number of the existing failed scenarios; For existing failure scenario libraries; For the first in the failure scenario library Trajectory; The trajectory embedding distance function is constructed based on the safety value curve, key relative state, response delay, and control action. For noise-free target multi-agent trajectory;

[0161] S5. Conduct closed-loop testing, solveable failure screening, and counterfactual diagnosis;

[0162] In closed-loop simulation, generated scenarios are executed. Based on theoretical solvability and the results of the tested system, only samples where, given vehicle dynamics, control input boundaries, road constraints, response delays, and disturbances, there still exists at least one set of executable driving strategies that can avoid collisions, boundary crossings, or other safety failures, but the tested system fails in actual operation, are considered as core defect evidence. The minimum scenario modification that restores the system to safety is then searched. The specific method is as follows:

[0163] S51. In closed-loop simulation, the determination of theoretical solvability fails.

[0164] The high-value failure indicator of the candidate scenario is expressed as Equation (15):

[0165] (15)

[0166] In the formula, Candidate trajectories High-value solvable failure indicators; This is an indicator function that takes a value of 1 if the condition is true and 0 otherwise. The minimum conservative safety margin to be recalculated using the actual closed-loop trajectory of the system under test; The threshold for determining whether a solution is theoretically solvable; This is the failure indication function for the autonomous driving system under test. It is set to 1 when a collision, boundary violation, unacceptable emergency braking, or mission failure occurs, and 0 otherwise. For noise-free target multi-agent trajectory;

[0167] Theoretically solvable and system-successful samples serve as evidence of system pass and capability boundary samples; theoretically solvable but system-failed samples serve as evidence of core defects; theoretically unsolvable and system-failed samples are only used for safety boundary calibration and cannot be directly attributed to the system under test; theoretically unsolvable but system-successful samples are used to check whether the value function is overly conservative or whether the system control capability exceeds the modeling range.

[0168] S52, Minimal counterfactual safety modifications for search failure scenarios;

[0169] For theoretically solvable failure scenarios, search for the minimum safe counterfactual, as shown in equation (16):

[0170] (16)

[0171] In the formula, The minimum scenario parameter modification vector required to ensure successful recovery of the tested system while maintaining theoretical solvability; The operation corresponding to the independent variable is performed to find the minimum value of the objective function; This is the original scene parameter vector; Modify the vector for the scene parameters to be optimized; This is a diagonal weight matrix used to normalize scene parameters with different dimensions. Let be the vector 2 norm; st represents the constraint that follows it. The trajectory generated and executed in closed loop using the modified scene parameters;

[0172] Equation (17) is used to evaluate the efficiency of solvable failure detection:

[0173] (17)

[0174] In the formula, To improve the efficiency of solving failure detection; The number of theoretically solvable and reproducible high-value system failure scenarios discovered within a given testing budget; This represents the total number of candidate scenarios for closed-loop execution. The larger the value, the more efficient it is at identifying system capability gaps under the same simulation budget;

[0175] Equation (18) is used to evaluate the critical boundary hit rate:

[0176] (18)

[0177] In the formula, This refers to the critical boundary hit rate. To ensure the minimum safety margin falls within the target range The number of generated scenes; The total number of valid candidates that pass the naturalness, road rule, and dynamics checks; To preset the target safety margin; The target risk zone is half-width;

[0178] High-value failure scenarios are clustered according to safety value curves, key relative states, system response delays and control actions. For the same cluster, the sample with the smallest trajectory embedding distance from the cluster center, the minimum smooth conservative safety margin located within the target risk zone, and the ability to reproduce system failure through repeated closed-loop simulation are selected as representative boundary samples. Several counterfactual neighborhood samples obtained by adjusting scenario parameters are retained near the sample. The next round of generation prioritizes exploring uncovered mechanisms through the novelty reward of formula (14), forming an iterative testing process of generation, closed-loop execution, failure diagnosis and active updating.

[0179] During implementation, scenario variables include initial longitudinal spacing, initial lateral spacing, speeds of both vehicles, cut-in start time, lateral speed of the cutting vehicle, maximum braking capacity of the vehicle, road adhesion coefficient, and total perception-control latency. First, a low-dimensional mesh reachability solver is used to obtain the baseline safety boundary. Then, a neural value function is trained and sampling is intensified in the boundary neighborhood. The diffusion model learns the natural distribution of cut-in timing, speed changes, and lateral motion from real cut-in segments.

[0180] For each target security level, according to the corresponding target security margin Candidate scenarios are generated in batches within a given range. A clearly defined safety level is used to verify naturalness and normal functionality; a positive small safety margin level serves as the primary capability boundary test; a near-zero level is used to calibrate the limiting response; and an unsolvable level is used only to check the safety value function and the simulator. This classification avoids mixing statistics of different risk levels and can verify whether the failure probability of the system under test monotonically increases as the safety margin decreases.

[0181] After closed-loop execution, the assumed vehicle response from the generation phase is replaced with the actual vehicle trajectory to reconstruct the joint state and calculate a conservative safety value. If the generation phase determines the system is solvable but the actual closed-loop trajectory enters the danger zone, it is necessary to distinguish between state degradation caused by the system response and neural value function estimation error; the former is considered a valid failure, while the latter is added to the boundary calibration set. All core defect samples are executed at least a preset number of times, and only when the failure results and key risk indicators appear stably are they added to the high-value scenario library.

[0182] Example 2

[0183] To verify the effectiveness of the critical safety boundary guidance method of this invention, uniform random sampling and critical boundary guidance sampling were performed for highway vehicle cut-in scenarios, generating 10,000 candidate scenarios for each method. In the experimental setup, the initial longitudinal spacing was set to 15–45 m, the vehicle speed to 20–30 m / s, the cutting-in vehicle speed to 18–28 m / s, the cut-in start time to 0.5–3.0 s, the cut-in lateral speed to 0.3–1.2 m / s, the road adhesion coefficient to 0.6–0.9, the maximum braking deceleration to 6–8 m / s², and the perception-control response delay to 0.2–0.6 s. The target safety margin was set to 0.20 m, and the risk zone half-width to 0.05 m, i.e., the target risk range was [0.15 m, 0.25 m].

[0184] Experimental results show that uniform random sampling yielded 8390 valid candidates, of which only 6 scenarios hit the target risk zone, resulting in a critical boundary hit rate of 0.07%. Critical boundary guided sampling yielded 10000 valid candidates, of which 4699 scenarios hit the target risk zone, achieving a hit rate of 46.99%, approximately 657 times that of uniform random sampling. To simulate a slight deviation in the execution capability of the tested system, the response delay was increased by 0.01 s in the closed-loop proxy model, and the actual braking capability was set to 97% of the theoretical braking capability. Ultimately, among the 4699 critical scenarios, 1624 scenarios were found to be theoretically solvable but where the tested controller failed, accounting for 34.56% of the hit scenarios.

[0185] In the representative scenario, the initial longitudinal spacing was 30.641 m, the vehicle speed was 26.754 m / s, the speed of the intervening vehicle was 18.689 m / s, the intervening start time was 2.139 s, the lateral speed was 1.003 m / s, the road adhesion coefficient was 0.635, and the response delay was 0.368 s. Its theoretical safety margin was 0.202 m, placing it within the target risk zone. However, the safety margin of the closed-loop proxy controller was -0.040 m, indicating that theoretically, the danger could still be avoided through sufficient braking. However, the tested controller entered an unsafe state due to the deviation between its response and braking capabilities. These results demonstrate that critical boundary guidance can significantly improve the efficiency of generating target risk scenarios within a limited testing budget and can effectively identify system capability defects that are highly sensitive to response delay and braking capability.

[0186] In summary, this invention employs a neural Hamilton-Jacobi reachability model to calculate the safety margin of a vehicle under current traffic conditions, defining a state with a safety margin approaching zero as a critical state. Near this state, the vehicle can still avoid danger through reasonable braking, steering, and other control operations. As the safety margin continues to decrease, the vehicle's hazard avoidance space gradually shrinks, and the danger may eventually become unavoidable. During natural driving diffusion sampling, boundary attraction is used instead of unidirectional hazard maximization. After closed-loop testing, only theoretically solvable scenarios where the tested system fails are retained, while outputting the minimum counterfactual safety modification. The results of this invention can support the safety verification of autonomous vehicles in large-scale physically feasible long-tail scenarios, supporting the industrialization of autonomous vehicles.

[0187] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating long-tail driving scenes based on neural reachability boundary guidance, characterized in that, Includes the following steps: S1. Establish the joint traffic dynamics and safety objective function; Based on typical driving scenarios, the relative positions, speeds, headings, control inputs, and bounded disturbances of the vehicle and key traffic participants are extracted to establish a joint dynamic system and a safety objective function. S2. Train and conservatively calibrate the neural reachability safety value function; The neural safety value function is trained using Hamilton-Jacobi-Isaacs partial differential equation residuals, terminal conditions, and state domain boundaries. Conservative calibration is performed using the error quantiles of a low-dimensional grid reachability benchmark to obtain continuous safety margins and state gradients. S3. Train a natural driving diffusion model with road and dynamic constraints; Learn the natural distribution of multi-agent trajectories and initial scene conditions from real driving logs, and eliminate unreasonable trajectories through discrete kinematics, road boundaries, and comfort constraints; S4. Perform critical safety boundary guidance during diffusion denoising; The generated trajectory is converted into a joint dynamic state sequence, and the minimum safety margin and gradient are calculated according to the conservative safety value function to make the scene converge to the specified critical risk zone while maintaining naturalness, physical feasibility and diversity. S5. Conduct closed-loop testing, solveable failure screening, and counterfactual diagnosis; In closed-loop simulation, generated scenarios are executed. Based on theoretical solvability and the classification of the test system results, only samples where reasonable avoidance strategies still exist but the test system fails are considered as core defect evidence. The minimum scenario modification that restores the system to safety is then searched.

2. The method for generating long-tail driving scenes based on neural reachability boundary guidance according to claim 1, characterized in that, The specific method of S1 is as follows: S11. Constructing joint traffic state and dynamics; The dynamics of the combined transportation system is expressed as equation (1): (1) In the formula, The joint state vector includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and necessary response delay between the vehicle and key traffic participants. Regarding continuous physical time The derivative; For the dynamics of vehicle-traffic interaction; This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; For continuous physical time; S12. Construct the security objective function; The terminal safety objective function is constructed based on the vehicle surround box spacing, as shown in equation (2): (2) In the formula, Joint state vector The terminal security objective function; For state The minimum bounding box distance between the corresponding vehicle and key traffic participants; To preset a safe distance threshold; When the value is less than or equal to zero, the state belongs to the target set that is either in collision range or has insufficient safety distance.

3. The method for generating long-tail driving scenes based on neural reachability boundary guidance according to claim 1, characterized in that, The specific method of S2 is as follows: S21. Establish the Hamilton-Jacobi-Isaacs security value function; The safety value function satisfies the Hamilton-Jacobi-Isaacs equation shown in equation (3): (3) In the formula, The safety value function corresponding to the joint state and remaining time; The remaining time from the current state to the end of the evaluation; for about The partial derivatives; min is the operation of minimizing the favorable control for the vehicle within the permissible control domain; max is the operation of maximizing the adverse effects for other participants and disturbances within their respective permissible domains. This is the vehicle control input vector; Input vectors for the behavior of other traffic participants; It is a bounded disturbance vector consisting of road adhesion changes, execution errors, and state estimation errors; for Regarding the joint state vector gradient; superscript Transpose of a vector; It is the dot product of vectors; The neural safety value function adopts the terminal condition shown in equation (4): (4) In the formula, The neural safety value when the remaining time equals zero; These are all the trainable parameters of the neural safety value network; It is a joint state vector that includes the relative longitudinal distance, relative lateral distance, speed, heading difference, and response delay between the vehicle and key traffic participants; Joint state vector The terminal security objective function; S22. Train and conservatively calibrate neural value functions; The training objective of the neural value function is expressed as equation (5): (5) In the formula, Overall training loss in neural accessibility; The partial differential equation residuals of the left-hand side of equation (3) relative to zero; The weight of the partial differential equation residuals on the left-hand side of equation (3) relative to zero; This is the terminal condition error corresponding to equation (4); The weight of the terminal condition error corresponding to equation (4); This represents the boundary constraint error when the state exceeds the training domain. The weights for the boundary constraint error when the state exceeds the training domain; The physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time; The weights of the physical monotonicity error of the safety value as a function of distance, braking capacity and remaining time; A conservative safety value is obtained using the benchmark error quantile, as shown in equation (6): (6) In the formula, For the joint state and remaining time The corresponding conservative safety value; For parameters A defined neural safety value function output; For the grid reachability benchmark and the absolute error of the neural value function quantiles; For the allowable calibration failure rate; In low-dimensional main scenes where mesh solving is feasible, the neural value function is compared with the mesh Hamilton-Jacobi results, and the errors in the overall region and the critical boundary neighborhood are statistically analyzed separately; conservative safety values ​​are uniformly used for scene generation. The road adhesion coefficient, maximum braking capacity, maximum steering ratio, and control delay are used as input conditions for the value function to support different vehicle capabilities and road conditions.

4. The method for generating long-tail driving scenes based on neural reachability boundary guidance according to claim 1, characterized in that, The specific method of S3 is as follows: S31. Establish the representation and diffusion process of natural driving trajectory; No. The traffic participants in the first The state of each discrete trajectory at each time point is represented by equation (7): (7) In the formula, For the first The traffic participants in the first The state vector at each discrete trajectory moment; and These refer to the longitudinal and transverse positions in the plane, respectively. Scalar velocity; It is longitudinal acceleration; For heading angle; This refers to the yaw rate; For traffic participant serial numbers; For discrete trajectory time indices; superscript Transpose of a vector; Multi-agent trajectory of the target Perform forward noise addition, as shown in equation (8): (8) In the formula, For the first Latent variables of noise trajectory at each diffusion level; For the first The cumulative trajectory fidelity coefficient for each diffusion level, with values ​​ranging from 0 to 1; For noise-free target multi-agent trajectory; To and Standard Gaussian noise vectors of the same dimension; The diffusion level number is used; the square root is used to ensure that the signal and noise terms meet the predetermined variance schedule. The natural driving diffusion model uses the denoising loss shown in equation (9): (9) In the formula, This is the denoising loss for the natural driving diffusion model; To address training scenarios and diffusion levels and random noise Find the expected value; For parameters A defined noise prediction network; These are all the trainable parameters of the diffusion model; For road map condition tensors; For the historical multi-agent trajectory condition tensor; This is a vector of conditions for scenario type and target risk level; The square of the second norm of a vector; S32. Constraints on the physical executability of the generated trajectory; The trajectory physical residual is calculated by equation (10): (10) In the formula, The discrete dynamic residuals of the trajectory; ∑ represents the summation over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. , and The first A vector is obtained by stacking the positions, velocities, and accelerations of all traffic participants at each moment, where The component comes from equation (7). and ; For the first From the moment to the first The sampling interval at each time point; Update the residual weights to speed; It is the square of the second norm of the vector.

5. The method for generating long-tail driving scenes based on neural reachability boundary guidance according to claim 1, characterized in that, The specific method of S4 is as follows: S41. Calculate the continuous critical safety margin of the generated trajectory; The minimum safety margin of the entire trajectory is obtained by using the smoothing minimum operation, as shown in equation (11): (11) In the formula, For trajectory The smooth minimum conservative safety margin; The smoothing temperature coefficient is positive. is the natural logarithm function; exp is the natural exponential function; ∑ is the summation over all adjacent discrete time points; This represents the total number of discrete sampling times for the trajectory. Index for discrete sampling time; A conservative, safe value; For the trajectory In the The joint traffic state vector obtained by converting the index of discrete sampling time points; For the first The remaining time corresponding to each discrete sampling time index; For noise-free target multi-agent trajectory; With target safety margin Construct the attraction function of the critical boundary, as shown in equation (12): (12) In the formula, For trajectory The critical boundary attracts the target; To predetermine the target safety margin; squaring penalizes scenarios where the safety margin deviates from the target value, and a leading negative sign maximizes the safety margin. Equivalent to approaching the target safety margin; S42. Integrating risk, road, physical, and diversity guidance in diffusion sampling; The synthesis guidance direction of the kth denoising level is expressed as Equation (13): (13) In the formula, For the first The comprehensive trajectory guidance direction of each noise reduction level; The basic denoising direction given for the natural diffusion model; , , and The first The boundary, path rules, dynamics, and novelty-guided weights of each denoising level; To track Find the gradient; Losses related to lane boundaries, right-of-way, and traffic rules; The reward is for novelty relative to the existing library of failure scenarios; The scene novelty reward is calculated by equation (14): (14) In the formula, Candidate trajectories Novelty reward; This is for calculating the minimum value; This refers to the sequence number of the existing failed scenarios; For existing failure scenario libraries; For the first in the failure scenario library Trajectory; The trajectory embedding distance function is constructed based on the safety value curve, key relative state, response delay, and control action. For noise-free target multi-agent trajectory.

6. The method for generating long-tail driving scenes based on neural reachability boundary guidance according to claim 1, characterized in that, The specific method of S5 is as follows: S51. In closed-loop simulation, the determination of theoretical solvability fails. The high-value failure indicator of the candidate scenario is expressed as Equation (15): (15) In the formula, Candidate trajectories High-value solvable failure indicators; This is an indicator function that takes a value of 1 if the condition is true and 0 otherwise. The minimum conservative safety margin to be recalculated using the actual closed-loop trajectory of the system under test; The threshold for determining whether a solution is theoretically solvable; This is the failure indication function for the autonomous driving system under test. It is set to 1 when a collision, boundary violation, unacceptable emergency braking, or mission failure occurs, and 0 otherwise. For noise-free target multi-agent trajectory; Theoretically solvable and system-successful samples serve as evidence of system pass and capability boundary samples; theoretically solvable but system-failed samples serve as evidence of core defects; theoretically unsolvable and system-failed samples are only used for safety boundary calibration and cannot be directly attributed to the tested system; theoretically unsolvable but system-successful samples are used to check whether the value function is overly conservative or whether the system control capability exceeds the modeling range; where system success means that the tested autonomous driving system does not experience a collision, unacceptable emergency braking, or driving task failure during the closed-loop simulation and completes the predetermined driving task, otherwise it is judged as system failure; S52, Minimal counterfactual safety modifications for search failure scenarios; For theoretically solvable failure scenarios, search for the minimum safe counterfactual, as shown in equation (16): (16) In the formula, The minimum scenario parameter modification vector required to ensure successful recovery of the tested system while maintaining theoretical solvability; The operation corresponding to the independent variable is performed to find the minimum value of the objective function; This is the original scene parameter vector; Modify the vector for the scene parameters to be optimized; This is a diagonal weight matrix used to normalize scene parameters with different dimensions. Let be the vector 2 norm; st represents the constraint that follows it. The trajectory generated and executed in closed loop using the modified scene parameters; Equation (17) is used to evaluate the efficiency of solvable failure detection: (17) In the formula, To improve the efficiency of solving failure detection; The number of theoretically solvable and reproducible high-value system failure scenarios discovered within a given testing budget; This represents the total number of candidate scenarios for closed-loop execution. The larger the value, the more efficient it is at identifying system capability gaps under the same simulation budget; Equation (18) is used to evaluate the critical boundary hit rate: (18) In the formula, This refers to the critical boundary hit rate. To ensure the minimum safety margin falls within the target range The number of generated scenes; The total number of valid candidates that pass the naturalness, road rule, and dynamics checks; To preset the target safety margin; The target risk zone is half-width; High-value failure scenarios are clustered according to safety value curves, key relative states, system response delays and control actions. For the same cluster, samples with small trajectory embedding distance from the cluster center, smooth minimum conservative safety margin located within the target risk zone, and stable reproduction of system failure through repeated closed-loop simulation are selected as representative boundary samples. Several counterfactual neighborhood samples obtained by slightly adjusting the scenario parameters are retained near these samples. The next round of generation prioritizes exploring uncovered mechanisms through the novelty reward of formula (14), forming an iterative test process of generation, closed-loop execution, failure diagnosis and active updating. During implementation, scenario variables include initial longitudinal spacing, initial lateral spacing, speeds of both parties, cut-in start time, lateral speed of the cutting vehicle, maximum braking capacity of the vehicle, road adhesion coefficient, and total perception-control time delay. First, a low-dimensional grid reachability solver is used to obtain the baseline safety boundary, then the neural value function is trained and sampling is made more densely in the boundary neighborhood. The diffusion model learns the natural distribution of cut-in timing, speed changes, and lateral motion from real cut-in segments. After closed-loop execution, the assumed vehicle response in the generation phase is replaced with the actual vehicle trajectory, the joint state is reconstructed, and a conservative safety value is calculated. If the generation phase determines that the solution is solvable but the actual closed-loop trajectory enters the danger zone, it is necessary to distinguish between the state deterioration caused by the response of the tested system and the neural value function estimation error. The state deterioration is considered a valid failure, and the neural value function estimation error is entered into the boundary calibration set. All core defect samples are executed at least a preset number of times, and only when the failure results and key risk indicators appear stably are they entered into the high-value scenario library.

Images