An end-to-end automatic driving training method based on closed-loop capability boundary

Abstract

The present application relates to the technical field of automatic driving model training and training data optimization, in particular to an end-to-end automatic driving training method based on closed-loop capability boundary. It comprises the following steps: S1: scene feature extraction and unified scene space construction; S2: multi-stage closed-loop evaluation and capability boundary positioning; S3: boundary ball construction and training scene mapping; S4: dynamic sampling probability generation based on boundary ball; S5: model training and capability boundary closed-loop updating. The present application realizes dynamic revaluation of training data value, so that the training data value can dynamically change with the model training stage, improves the data utilization efficiency under the limited training budget, and balances boundary reinforcement training and training stability.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving model training and training data optimization technology, and in particular to an end-to-end autonomous driving training method based on closed-loop capability boundaries. Background Technology

[0002] With the development of intelligent connected and autonomous driving technologies, autonomous driving systems are gradually evolving from traditional modular architectures to end-to-end architectures. End-to-end autonomous driving methods directly establish mappings between sensor observations, vehicle states, navigation information, and future trajectories or driving actions through neural networks. This reduces information transmission losses between modules such as perception, prediction, planning, and control, and has become an important technological direction for autonomous driving decision-making and planning. Since end-to-end models primarily rely on driving data to learn road environment, traffic interactions, and driving behavior patterns, the scale, quality, and organization of the training data directly affect the model's driving performance and generalization ability.

[0003] As the capabilities for collecting, storing, and processing real-vehicle data continue to improve, the scale of end-to-end autonomous driving training data continues to expand. However, a large number of samples in real-world road data are concentrated in high-frequency scenarios such as going straight, following other vehicles, and regular lane changes. There is strong similarity and repetition among different samples, while the proportion of complex interactions and rare traffic conditions is relatively low. When training using random or uniform sampling methods, many similar scenarios are repeatedly selected, preventing new data from being fully converted into improvements in model performance. Therefore, the challenge facing end-to-end autonomous driving training is gradually shifting from simply expanding the data scale to improving the effective utilization of training data.

[0004] Existing end-to-end autonomous driving data training methods mainly fall into two categories: The first is conventional training methods based on the original data distribution, which construct training batches through random sampling, uniform sampling, or fixed-ratio sampling. This type of method is simple to implement and can maintain the overall data distribution, but it struggles to distinguish the actual impact of different scenarios on model training. The second is data filtering methods based on scenario attributes or model performance. These methods use indicators such as traffic density, interaction intensity, time-to-collision ratio, prediction error, collisions, boundary violations, and task failures to filter out difficult scenarios, long-tail scenarios, or failed scenarios, and increase their training proportion. This type of method can increase the frequency of complex scenarios during the training process, improving the relevance of the training.

[0005] However, existing methods typically determine sample value based on the scene's inherent attributes or the model's performance at a certain stage, employing relatively fixed data selection rules and sampling weights. As training progresses, the model's performance in different scenes continuously changes, and the role of the same scene in model training also shifts. Some scenes may have been stably mastered by the model, offering minimal gains from continued training; some scenes are in a state where the model can attempt to handle them but hasn't yet achieved stable success, providing relatively effective training signals; and some scenes may remain unhandled for a long time, and directly increasing their training proportion may not yield stable results. Therefore, the training value of a scene is not only related to its inherent difficulty but also closely related to the model's processing level at the current training stage.

[0006] In recent years, closed-loop evaluation has been increasingly used to analyze the actual driving performance of end-to-end autonomous driving models in specific scenarios. Compared with open-loop evaluation based on pre-recorded trajectory comparison, closed-loop evaluation can record collisions, boundary violations, stalls, and task completion statuses that occur after the model executes the planning results, more directly reflecting the model's ability to handle different scenarios. Existing research shows that as the capabilities of autonomous driving systems change, the distribution of corresponding challenge scenarios also changes, requiring continuous updates to the scenario selection range based on new evaluation results. However, current research mostly uses closed-loop evaluation for performance verification, defect discovery, or key scenario generation, and lacks a dynamic training method that unifies closed-loop evaluation results, training scenario localization, and data sampling adjustments.

[0007] Furthermore, autonomous driving scenarios typically encompass multi-dimensional information such as road structure, traffic participants, vehicle motion states, and interaction relationships. There is often a lack of direct correspondence between key scenarios identified in closed-loop evaluations and large-scale real-world training data. How to transform model performance in evaluation scenarios into data selection criteria usable for training, how to identify samples from the training data pool that closely resemble the characteristics of the current key scenarios, and how to adjust the distribution of training data in a timely manner as model performance changes remain challenges that need to be addressed in end-to-end autonomous driving training. Summary of the Invention

[0008] The purpose of this invention is to provide an end-to-end autonomous driving training method based on closed-loop capability boundaries, addressing issues such as low training data utilization efficiency, mismatch between difficult scenario selection and the current model state, and difficulty in dynamically adjusting training data sampling rules as model capabilities change in existing end-to-end autonomous driving training. This method, based on the closed-loop driving performance of the end-to-end autonomous driving model in different scenarios, identifies boundary scenarios that the model has not yet fully mastered but have learning value through joint analysis of model capabilities and scenario difficulty. Based on this, the sampling distribution of real training data is adjusted, ensuring that the training focus is continuously updated as the model's capabilities improve.

[0009] In its specific implementation, this invention first extracts multi-dimensional features such as road structure, traffic participants, vehicle motion state, and scene interaction relationships to uniformly represent the evaluation and training scenarios. It then constructs a low-dimensional scene space that can measure scene similarity through dimensionality reduction. Subsequently, it conducts closed-loop evaluations on end-to-end autonomous driving models at different training stages. Based on the model's pass and fail results in each evaluation scenario, it establishes a success-failure matrix and jointly estimates the model's capabilities and scene difficulty to determine the boundary scenarios that match the current model's capabilities. Furthermore, it clusters the boundary scenarios and constructs a boundary spherical neighborhood, mapping the capability boundaries identified in the evaluation stage to the real training data pool and selecting training samples located inside and near the boundary sphere.

[0010] During the training phase, this invention dynamically adjusts the sampling weights of training samples based on the distance between the training samples and the center of the boundary sphere, as well as the local coverage area of ​​the boundary sphere. This ensures that scenes closer to the current capability boundary receive a higher sampling probability, while replaying training with a lower probability is reserved for scenes outside the boundary sphere, thus maintaining the original data distribution and the driving capabilities already acquired by the model. After completing the first phase of training, a closed-loop evaluation is conducted again, and the capability state is re-estimated, boundary scenes are located, and sampling weights are adjusted based on the updated model performance, thereby forming an iterative training process of "closed-loop evaluation—capability calibration—boundary mapping—dynamic training".

[0011] Through the above design, this invention can directly transform closed-loop evaluation results into training data selection criteria, changing the value judgment of training samples from a fixed scenario difficulty or single failure result to a dynamic matching between scenario difficulty and the model's current capabilities. This method is applicable to end-to-end autonomous driving models with different structures and training stages. Without changing the original model training objectives and data sources, it increases the training proportion of boundary scenarios, reduces repetitive training of already mastered scenarios, and avoids long-term concentration of training resources on scenarios that are difficult to learn effectively at the current stage. This improves the data utilization efficiency, closed-loop driving performance, and generalization ability of end-to-end autonomous driving models in complex scenarios.

[0012] The objective of this invention is achieved through the following technical solution: An end-to-end autonomous driving training method based on closed-loop capability boundaries includes the following steps: S1: Scene feature extraction and unified scene space construction; A training scene set and a capability assessment scene set are obtained. A unified scene feature vector is extracted from each scene in the training scene set and the capability assessment scene set. Standardization parameters are calculated based on the training scene set. Standardization processing is performed on the training scene set and the capability assessment scene set. Then, the standardized feature vector of the training scene set is used to fit a dimension reduction projection relationship. Principal component analysis and whitening processing are used to map the training scene set and the capability assessment scene set to the same low-dimensional scene space to obtain the low-dimensional vector representation of each scene. S2: Multi-stage closed-loop evaluation and capability boundary positioning; The model is obtained from the stage models saved in multiple training stages of the end-to-end autonomous driving model. The closed-loop evaluation of each stage model is performed on the capability assessment scenario set. The model-scenario success or failure matrix is ​​constructed based on the comparison results of the closed-loop comprehensive score of each scenario and the preset pass threshold. The capability-difficulty probability model is established using the success or failure matrix. The capability parameters of each stage model and the difficulty parameters of each assessment scenario are jointly estimated. The predicted pass probability is calculated based on the current stage model's capability parameters and the difficulty parameters of each assessment scenario. Assessment scenarios whose predicted pass probability meets the preset probability interval conditions are identified as capability boundary scenarios. S3: Boundary sphere construction and training scene mapping; The capability boundary scene is mapped to the low-dimensional scene space, and clustering is performed on the low-dimensional vector of the capability boundary scene to determine the boundary-related training scene. S4: Dynamic sampling probability generation based on boundary sphere; Distance weights are generated based on the distance between each boundary-related training scenario and the corresponding boundary sphere center. After normalization, the distance weights are fused with the basic replay distribution of the entire training pool to generate the dynamic sampling distribution of the current training stage. S5: Model training and closed-loop update of capability boundaries; Training scenarios are sampled from the training data pool according to the dynamic sampling distribution, training batches are constructed and the end-to-end autonomous driving model is updated. After the model update reaches the preset stage, the closed-loop evaluation is re-executed and the capability boundary is updated in a closed loop. The closed-loop iterative training continues until the training termination condition is met, and then the trained end-to-end autonomous driving model is output.

[0013] Specifically, step S1 includes: Step S11: Obtain the training scenario set and the capability assessment scenario set; Obtain a set of training scenarios for training parameters of end-to-end autonomous driving models, as well as a set of capability evaluation scenarios for closed-loop evaluation of models at different training stages.

[0014] The training scenario set is represented as follows: (1) The set of capability assessment scenarios is represented as follows: (2) in, Represents the training scenario set. Indicates the first One training scenario, Indicates the total number of training scenarios; This represents a set of ability assessment scenarios. Indicates the first A capability assessment scenario, This indicates the total number of capability assessment scenarios.

[0015] Each scenario includes at least road environment information, vehicle status information, and traffic participant status information, and may also include one or more of the following: historical trajectory, navigation information, traffic control information, and scenario evaluation information. The training scenario set and the capability assessment scenario set are independent of each other. The capability assessment scenario set is used only for closed-loop evaluation and model capability boundary localization, and is not directly used for parameter updates of end-to-end autonomous driving models.

[0016] Step S12: Construct a unified scene feature vector; Following a unified feature definition and arrangement order, features are extracted from each scene in both the training scene set and the capability assessment scene set, and the original scene feature vectors are constructed as follows: (3) in, This represents any training scenario or capability assessment scenario. Representing a scene The original scene feature vector; Representing a scene The 3D scene features, ; The dimension representing the features of the original scene.

[0017] The scene features consist of the following four categories, totaling 21 dimensions: (1) Four dimensions of road structure characteristics: average curvature, maximum curvature, route complexity and route length; (2) 11 dimensions of traffic participant characteristics: number of vehicles, number of pedestrians, total number of participants, distance of the nearest participant to the vehicle, number of forward participants, distance of the nearest future participant, number of the maximum number of forward participants in the future, participant speed variance, number of participants in front, number of participants on the left, and distance of the nearest vehicle; (3) Four-dimensional characteristics of the vehicle's motion state: current speed, vehicle acceleration modulus, historical speed variance and historical average heading angle change rate; (4) Scene interaction and risk characteristics 2D: minimum collision time of scene and maximum required deceleration of scene.

[0018] The training scenario and the capability assessment scenario use the exact same 21-dimensional feature definition and arrangement order so that the two scenarios obtain the same original feature vector with the same physical meaning.

[0019] Step S13: Scene feature standardization; Since different scene features have different physical dimensions and numerical ranges, in order to reduce the impact of feature scale differences on scene similarity measurement, the original scene features are standardized.

[0020] Calculate the first based on the training scenario set. Mean of scene features: (4) And calculate the first Standard deviation of scene features: (5) in, Indicates the first The mean of the scene features in the training scene set. Indicates the first The standard deviation of the scene features in the training scene set.

[0021] For any scenario , its first The dimensional standardized feature is represented as: (6) in, Representing a scene The Dimensional standardization features This represents a preset positive number, used to avoid division by zero when the standard deviation is zero.

[0022] This leads to the scene. Standardized scene feature vector: (7) Both the training scenario set and the capability assessment scenario set use the same set of mean parameters calculated from the training scenario set. and standard deviation parameter Standardization is performed to ensure that the two types of scenarios are at the same feature scale.

[0023] Step S14: Scene feature dimensionality reduction and unified mapping; A dimensionality reduction mapping relationship is established for the standardized training scene features, and the same dimensionality reduction mapping relationship is used to process the training scene and the capability assessment scene, mapping them to the same low-dimensional scene space.

[0024] The unified mapping process is represented as follows: (8) in, This represents the scene feature dimensionality reduction and scaling function; Representing a scene Vector representation in a unified low-dimensional scene space; Let represent the dimension of the low-dimensional scene space, and satisfy . .

[0025] Specifically, this invention constructs a unified scene space using standardization, principal component analysis, and whitening processing, and only uses the training scene to fit the standardized parameters and projection relationships; the capability assessment scene and the verification scene use the same dimensionality reduction projection matrix as the training scene for dimensionality reduction projection and perform whitening processing using the same whitening parameters, and do not participate in the fitting calculation of standardized parameters and projection relationships.

[0026] Calculate the covariance matrix based on the standardized feature vectors of the training scenario: (9) in, The covariance matrix representing the characteristics of the standardized training scenario.

[0027] For covariance matrix Perform eigenvalue decomposition: (10) in, This represents an orthogonal matrix composed of the eigenvectors of the covariance matrix; This represents a diagonal matrix composed of corresponding eigenvalues. The eigenvalues ​​are arranged in descending order as follows: Before selection The eigenvectors corresponding to the eigenvalues ​​form the dimension-reduced projection matrix: (11) and the corresponding eigenvalue matrix (12) in, Indicates the first Principal component direction, , This represents the corresponding eigenvalue. This indicates the constructor for a diagonal matrix.

[0028] When whitening is applied, the scene The low-dimensional vector representation is as follows: (13) in, express An identity matrix of order 1. This indicates the preset whitening stability parameter, used to avoid numerical instability when the eigenvalue is too small or zero; superscript This indicates that each diagonal element in the diagonal matrix is ​​raised to the power of -1 / 2. Whitening is used to adjust the scale of different principal component directions, reducing the impact of principal component variance differences on scene distance calculations.

[0029] The training scenario and the capability assessment scenario use the same standardized parameters, dimensionality reduction projection matrix, and scale adjustment parameters, and do not establish separate scene spaces.

[0030] Specifically, step S2 includes: Step S21: Obtain the model for multiple training phases; The system acquires multiple model checkpoints saved at different training stages of the end-to-end autonomous driving model. These checkpoints can come from the pre-training stage, fine-tuning stage, iterative training stage, or other training stages that can reflect changes in the model's capabilities.

[0031] Step S22: Perform closed-loop evaluation and construct a success or failure matrix; The models at each stage are subjected to closed-loop evaluation on the same boundary evaluation scenario set. Based on preset pass / fail conditions, the evaluation results of the models in each scenario are represented as pass or fail, and a model-scenario success / failure matrix is ​​constructed. Success / failure variables can be represented as: (14) in, Indicates the scene number. This indicates the training phase; the closed-loop composite score will be used. Compared with the preset threshold Compare, and when the threshold is reached, let Otherwise .

[0032] Step S23: Jointly estimate model capability and scenario difficulty; A capability-difficulty probability model is established using the success-failure matrix, calibrating the model's capability and scenario difficulty to a unified latent scale. In a preferred embodiment, the Rasch model of the Item Response Theory (IRT) model can be used to describe the probability of the model passing the scenario: (15) in, For the first Capability parameters of the stage model For the first Difficulty parameters for each scenario For the first The phase model is through the first The predicted probability of a scenario. This is a monotonic probability mapping function. Model capability parameters and scene difficulty parameters can be obtained through maximum likelihood estimation, maximum a posteriori estimation, or other parameter estimation methods.

[0033] In one embodiment of the present invention, the monotonic probability mapping function may be a Logistic function. .when When the model predicts the scenario, the probability is 0.5; when In this case, the probability of passing the prediction increases as the difference between ability and difficulty increases. For a given success-failure matrix, the model ability parameters and scenario difficulty parameters can be estimated by maximizing the Bernoulli likelihood or minimizing the negative log-likelihood with a regularization term.

[0034] Step S24: Locate the current model's capability boundary; The current capability boundary scenario is determined based on the difference between the current model's capability and the scenario difficulty, or based on the model's predicted probability of passing the scenario. The capability boundary scenario refers to a scenario that the current model has a certain probability of passing but has not yet mastered stably. The capability boundary can be determined in one of the following forms: (16) (17) in, and These represent the lower and upper bounds of the capability boundary probability band, respectively. When the number of boundary candidates is less than the preset minimum number, the probability interval is expanded simultaneously by a preset step size (lowering the lower bound and raising the upper bound) until the quantity constraint is met. This is a preset capability-difficulty difference threshold, used to limit how closely the scene difficulty matches the current model's capability. The smaller the value, the narrower the capability boundary.

[0035] Equations (16) and (17) represent two alternative methods for determining the capability boundary; when When using the Logistic function, the predicted probability range corresponding to equation (17) is: .

[0036] Specifically, step S3 includes: Step S31: Obtain a low-dimensional representation of the boundary scene; Map the capability boundary scene determined in step S2 back to the low-dimensional scene space constructed in step S1 to obtain a low-dimensional vector representation of the capability boundary scene.

[0037] Step S32: Determine the center of the boundary sphere; Let the first The set of capability boundary scenarios for each stage is The number of its scenes is The number of clusters is denoted as .right The low-dimensional vectors of each boundary scene are clustered using K-Means, with each cluster center serving as the center of a boundary sphere; according to Adaptive determination, and with This is the upper limit.

[0038] Step S33: Determine the radius of the boundary sphere; First, the basic radius is determined based on the nearest neighbor distance of the training scene near the center of the sphere. Then, the radius is adjusted according to the normalized scene difficulty of the corresponding boundary scene. Adjacent boundary spheres are merged and verified, and isolated clusters are processed as point boundaries and the nearest training scene is retained as support samples.

[0039] Step S34: Retrieve boundary-related training scenarios from the training data pool; Calculate the distance between the training scene and the center of each boundary sphere, and determine the training scenes that satisfy the boundary sphere neighborhood condition as boundary-related training scenes. The neighborhood condition can be expressed as: (18) in, For the first The vector representation of each training scene in the low-dimensional scene space. For the first The center of the boundary sphere, For the first The radius of the boundary sphere, This is the scene space distance metric function.

[0040] Step S35: Handle overlapping multiple boundary spheres; When a training scenario falls into multiple boundary balls simultaneously, record the set of boundary balls corresponding to that training scenario. The distance and radius of each boundary sphere are used to form an overlapping relationship between the training scene and multiple boundary spheres.

[0041] Specifically, step S4 includes: Step S41: Calculate the in-sphere distance weights for the training scenario; Sampling weights are generated based on the distance between the training scene and the center of the boundary sphere. The closer the training scene is to the center of the boundary sphere, the higher its weight; the closer it is to the edge of the sphere, the lower its weight. These sampling weights can be generated using a Gaussian decay function, a linear decay function, or other functions that monotonically decrease with distance.

[0042] The weighting of the distance inside the sphere is expressed as follows: (19) in, For the first The training scenario and the first The distance between the centers of the boundary spheres For the first The radius of the boundary sphere.

[0043] Step S42: Merge the weights of multiple boundary spheres; The distance weights of multiple boundary spheres are fused based on the overlapping relationships recorded in step S35. For training scenarios corresponding to only one boundary sphere, the distance weight corresponding to that boundary sphere is directly used; for training scenarios corresponding to multiple boundary spheres, the maximum value of the distance weights of each boundary sphere is taken as the comprehensive boundary weight. To avoid oversampling of overlapping regions due to repeated weight accumulation, where Indicates the first The training scenario and the first Each boundary sphere weight.

[0044] Step S43: Set the base replay probability for the entire training pool; Establish a basic replay distribution across all training scenarios. And ensure that each effective training scene has a non-zero base replay probability to maintain original scene coverage and reduce ability forgetting; the fusion of the base replay distribution and the boundary-related distribution is performed in step S44.

[0045] Step S44: Generate a normalized dynamic sampling distribution; The boundary weight distribution is fused with the basic replay distribution to generate the first... The dynamic sampling distribution of the stage. The dynamic sampling distribution can be expressed as: (20) in, For the first Phase 1 The sampling probability of each training scene. To incorporate boundary weights, To ensure uniform playback distribution across the entire training pool, The base playback ratio; in one specific embodiment, We set it to 0.50, meaning the probability quality is 50% for the boundary correlation distribution and 50% for the base replay distribution of the entire training pool.

[0046] Specifically, step S5 includes: Step S51: Construct training batches based on dynamic sampling probabilities; Based on the dynamic sampling probability obtained in step S4, training scenarios are sampled from the training data pool to construct training batches for model parameter updates.

[0047] Step S52: Update the end-to-end autonomous driving model; The end-to-end autonomous driving model is updated using the training batches obtained from sampling. This invention does not limit the specific structure of the end-to-end model and can be applied to diffusion trajectory planning models, regression planning models, Transformer planning models, and other end-to-end planning or control models.

[0048] Step S53: Re-evaluate the closed-loop performance of the updated model; After the model training reaches the preset stage, the updated model is put back into the boundary evaluation scenario set for closed-loop evaluation to obtain new scenario pass or failure results.

[0049] Step S54: Update the capacity boundary and sampling distribution; Based on the updated success-failure matrix, the model capability parameters and scene difficulty parameters are re-estimated, and then the capability boundary scene, boundary sphere center, boundary sphere radius, training scene comprehensive distance weight, and dynamic sampling probability are updated sequentially. Thus, the capability boundary migrates as the model training progresses, and the distribution of training data is dynamically adjusted according to the changes in model capability, forming a closed-loop iterative training process.

[0050] Step S55: Training terminates and model output; When the preset number of training rounds is reached, the preset training budget is reached, the model performance converges, or the model capability increment is lower than the preset threshold, training stops and the trained end-to-end autonomous driving model is output.

[0051] Beneficial effects Compared with the prior art, the present invention has the following advantages: (1) Dynamically determine the value of training data. This invention determines the capability boundary based on the relationship between the current capability of the model and the difficulty of the scene, so that the value of training data can change dynamically with the model training stage, avoiding static selection of training samples based solely on the absolute difficulty of the scene or the result of a single failure.

[0052] (2) Improve data utilization efficiency under limited training budget. This invention increases the sampling probability of training scenarios near the current capability boundary, reduces repeated sampling of simple scenarios that have been stably mastered by the model and scenarios that are significantly beyond the current capability, so that training resources are used more for scenarios that can generate effective learning signals. Experiments show that, under the same scenario set and evaluation rules, the closed-loop comprehensive evaluation index (EPDMS) of the planning models DiffusionDrive and LTF increases from 84.5 to 87.3 and from 79.1 to 82.4, respectively.

[0053] (3) Achieve effective mapping of closed-loop evaluation results to the training data pool. This invention maps the boundary evaluation scenario and the training scenario to a unified scenario space, retrieves training scenarios that are close to the current capability boundary from the training data pool through the boundary ball, and transforms the capability boundary information obtained from the closed-loop evaluation into training sampling weights.

[0054] (4) Balancing boundary reinforcement training with training stability. This invention combines high-weight sampling of the boundary ball with low-probability replay of the entire training pool, so that the scene outside the boundary ball still maintains a non-zero sampling probability, thereby reducing the ability forgetting and training fluctuations caused by excessive concentration of training distribution.

[0055] (5) Achieve closed-loop dynamic update of capability boundaries. This invention re-evaluates the closed-loop after model training and updates the model capability, boundary scene, boundary ball and sampling probability, so that the distribution of training data can dynamically shift as the model capability improves. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the overall training process of the present invention. Figure 2 This is a schematic diagram illustrating the construction of scene features and the unified space in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-stage closed-loop evaluation and success / failure matrix in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the model capability and scene difficulty boundary localization in an embodiment of the present invention; Figure 5 This is a schematic diagram of dynamic sampling weights and basic playback in an embodiment of the present invention; Figure 6 This is a schematic diagram of capability boundary migration according to an embodiment of the present invention. Detailed Implementation

[0057] The technical solution provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0058] Example An end-to-end autonomous driving training method based on closed-loop capability boundaries, the overall process of which is as follows: Figure 1 As shown. During implementation, the processor sequentially completes scene feature extraction, unified scene space construction, multi-stage closed-loop evaluation, joint estimation of model capability and scene difficulty, boundary sphere construction, dynamic sampling, and model update, and forms a closed-loop training process through repeated evaluation and updates.

[0059] S1: Scene Feature Extraction and Unified Scene Space Construction Step S11: Obtain the training scenario set and the capability assessment scenario set; A training scenario set, a capability assessment scenario set, and an independent verification scenario set are established separately. In this embodiment, the training scenario set includes... One scenario is used for parameter updates of the end-to-end autonomous driving model; the capability assessment scenario set includes... The three scene sets are used for closed-loop evaluation, capability estimation, and capability boundary localization of the model at each training stage, and do not directly participate in model parameter updates; the independent validation scene set contains 5912 scenes, which are used only for final validation after training. The three scene sets are deduplicated according to scene identifiers and remain independent of each other.

[0060] Step S12: Construct a unified scene feature vector; Using the same feature definition, calculation method, and arrangement order for the three scene sets, any scene is represented as a 21-dimensional original feature vector. The 21-dimensional features consist of four categories: 4-dimensional road structure features, including average curvature, maximum curvature, route complexity, and route length; 11-dimensional traffic participant features, including the number of vehicles, pedestrians, total number of participants, distance to the nearest participant, number of forward participants, future nearest participant distance, future maximum number of forward participants, participant velocity variance, number of participants ahead, number of participants to the left, and distance to the nearest vehicle; 4-dimensional vehicle motion state features, including current speed, vehicle acceleration modulus, historical velocity variance, and historical average heading angle change rate; and 2-dimensional scene interaction and risk features, including the scene minimum collision time and the scene maximum required deceleration. For missing scene elements, corresponding feature values ​​are generated according to a unified rule.

[0061] Step S13: Scene feature standardization; Arrange the original features of the training scene into a matrix by row. The mean of each feature dimension is calculated using only the training matrix. and standard deviation And standardized according to equations (4) to (7). The capability assessment scenario and the independent verification scenario use only the same set of training scenarios. , The transformation is performed without recalculating the statistics. For features with a standard deviation of zero or close to zero, a pre-defined positive number is added to the denominator to ensure the stability of the standardization process.

[0062] Step S14: Scene feature dimensionality reduction and unified mapping; Only the standardized training scene is used to fit the dimension-reduced projection relationship, and the principal component orientation and scale adjustment parameters are determined according to equations (8) to (13). In this embodiment, the 21-dimensional original features are reduced to dimensional, projection matrix is The low-dimensional matrix of the training scenario is The capability assessment scenario and the independent verification scenario use the same projection matrix, the same set of eigenvalues, and the same whitening stabilization parameters to perform transformations, resulting in the following: and The low-dimensional matrix does not participate in the fitting of the projection relationship. The process is as follows: Figure 2 As shown, in the scene, blue vehicles (EGOs) represent autonomous vehicles, and the remaining gray vehicles represent other traffic participants. By extracting scene features such as road structure, traffic participant relationships, autonomous vehicle trajectories, and trajectory interaction patterns, corresponding scene feature vectors are constructed. These feature vectors are then subjected to dimensionality reduction processing, mapping different scenes to a unified scene space. In this scene space, scenes represented by pentagrams can represent a class of scenes similar to their scene feature vectors; triangles represent similar scenes that can be represented by this scene (representative scenes), and crosses represent other scenes that cannot be represented by this scene.

[0063] S2: Multi-stage closed-loop evaluation and capability boundary positioning: Step S21: Obtain models from multiple training phases; In this embodiment, a preheating model is first trained for 20 rounds. Then, 16 consecutive training phases are performed, with each phase consisting of 5 rounds of training and saving one phase model. Therefore, the phase number is... A total of A model for closed-loop evaluation The total number of training rounds is Each stage of the model uses the same input definition, output definition, and closed-loop evaluation conditions; this invention does not limit the specific network structure of the end-to-end model, nor does it require the use of a specific training framework or optimizer.

[0064] Step S22: Perform closed-loop evaluation and construct a success or failure matrix; Each stage model is sequentially placed into the same capability assessment scenario set to perform closed-loop evaluation. For the first stage... The first stage model and the second Each assessment scenario is evaluated based on a closed-loop comprehensive score. With threshold The comparison results are used to generate success and failure variables according to equation (14). In this embodiment, the EPDMS percentage result output by the closed-loop evaluation is divided by 100 to obtain the closed-loop comprehensive score; the same index calculation and normalization rules are used for each stage. All stage models use the same initial scenario state, simulation time domain, and index calculation rules, thus forming a success or failure matrix with rows corresponding to 12,146 evaluation scenarios and columns corresponding to 17 stage models. This process is as follows: Figure 3 As shown, where, For the first Capability parameters of the stage model For the first The difficulty parameters for each scene.

[0065] (twenty one) In the formula, Indicates the number of capability assessment scenarios. The success or failure matrix in this embodiment represents the number of stages in the model. .

[0066] Step S23: Jointly estimate model capability and scenario difficulty; A capability-difficulty probability model is established using the success-failure matrix and equation (15). To enable the capability parameters at different training stages and the difficulty parameters for different scenarios to be solved together, maximum likelihood estimation or maximum a posteriori estimation with regularization terms can be used. For the observed results Its Bernoulli log-likelihood term is expressed as: (twenty two) When using a quadratic regularization term, the objective function can be minimized as shown in the following equation: (twenty three) In the formula, Let be the vector of stage model capability parameters to be estimated. Let be the vector of scene difficulty parameters to be estimated. Let be the given observation success / failure matrix. Therefore, the optimization independent variable of equation (23) is: and , Given the observed data; the estimation result is defined as follows: and satisfy .

[0067] This embodiment takes , The regularity coefficient in equation (23) Under constraints Seeking to obtain Minimum capability and difficulty parameters. During the solution process, both types of parameters can be updated alternately or simultaneously. After each update, zero-mean centering is performed on the scene difficulty, and the model capability is correspondingly shifted to maintain... constant.

[0068] Step S24: Locate the current model capability boundary; Based on the current stage capability parameters and difficulty parameters of each assessment scenario Calculate the probability of prediction passing This embodiment will Scenes identified as capability boundary candidates are considered. When the number of candidates is less than 300, the probability lower bound is decreased and the probability upper bound is increased in steps of 0.05 until the number of candidates reaches 300 or the probability interval reaches the allowable range. When the number of candidates is too large, scenes that are close to the capability center and have complete scene information are prioritized. The localization results are as follows: Figure 4 As shown.

[0069] S3: Boundary Sphere Construction and Training Scene Mapping Step S31: Obtain a low-dimensional representation of the boundary scene; Extract the capability boundary scenarios determined in step S24 from the capability assessment scenario set, and transform them into unified low-dimensional vectors using the standardized parameters and dimensionality reduction mapping relationships obtained in steps S13 and S14. This step does not refit the normalizer or dimensionality reduction model to ensure that the distance between the boundary scenarios and the training scenarios has a consistent meaning.

[0070] Step S32: Determine the center of the boundary sphere; Clustering is performed on the 12-dimensional vectors of the capability boundary scenarios, with each cluster center serving as the center of a boundary sphere. This embodiment is based on the number of capability boundary scenarios. Adaptive determination of cluster number .

[0071] In the formula, For the first The set of capability boundary scenarios for each stage The number of boundary scenes contained in this set. Cluster number; constant This is the minimum number of clusters used in this embodiment, a constant. A preset scaling parameter for adjusting the number of clusters based on the size of the boundary scene set.

[0072] If the number of boundary scenes is small, the boundary scene or its local center can be used as the sphere center directly; after determining the sphere center, sphere centers that are too close to each other representing the same local boundary region are merged.

[0073] Step S33: Determine the radius of the boundary sphere; For each boundary sphere center, calculate its distance to the 12-dimensional vector of the training scene. In this embodiment, the median distance from the sphere center to the five nearest neighbors of the training scene is multiplied by 1.5 to obtain the base radius. ,in, For the first The center of the boundary sphere, Indicates distance in a unified 12-dimensional scene space The set formed by the 5 most recent training scene vectors, This indicates taking the median of the distance values ​​within the set. For the first The base radius of each boundary sphere is then determined. The radius is adjusted based on the normalized average difficulty of the corresponding boundary scene, with a difficulty gain coefficient of 0.35. Each boundary sphere should cover at least 300 training scenes; if this is insufficient, the radius is gradually increased. Isolated clusters containing no more than two boundary scenes are treated as point boundaries, and at least the 10 training scenes closest to the sphere's center are retained as support samples.

[0074] Step S34: Retrieve boundary-related training scenarios from the training data pool; Calculate the low-dimensional vector for each training scene sequentially. With each boundary center The distance between them. Training scenarios that satisfy the neighborhood conditions shown in Equation (18) are identified as boundary-related training scenarios, and their corresponding sphere centers, distances, and sphere radii are recorded. This process maps the capability boundaries in the capability assessment scenarios back to the training data pool that can be used for model updates.

[0075] Step S35: Handle overlapping multiple boundary spheres; When multiple boundary balls fall into the same training scenario, record the set of boundary balls corresponding to that training scenario. The distance between it and the center of each boundary sphere, as well as the radius of each boundary sphere, form a multi-boundary sphere overlapping relationship.

[0076] S4: Dynamic sampling probability generation based on boundary spheres: Step S41: Calculate the in-sphere distance weights for the training scenario; For the boundary-related training scene obtained in step S34, calculate its intrasphere distance weight according to equation (19). The closer the training scene is to the center of the boundary sphere, the higher the weight; the closer it is to the edge of the boundary sphere, the lower the weight. In addition to the Gaussian decay function, other weight functions that monotonically decrease with distance and are non-negative within the sphere can also be used. The dynamic weight relationship is as follows: Figure 5 As shown.

[0077] Step S42: Merge the weights of multiple boundary spheres; For training scenarios involving only one boundary sphere, the corresponding intra-sphere distance weights are used directly; based on the boundary sphere set recorded in step S35... The maximum distance weight is taken as the comprehensive boundary weight, that is... .

[0078] Step S43: Set the base replay probability for the entire training pool; Establish a basic replay distribution across all training scenarios. This embodiment uses a uniform base playback distribution, that is... This ensures that each effective training scenario has a non-zero base replay probability, in order to maintain training data coverage and reduce model forgetting.

[0079] Step S44: Generate a normalized dynamic sampling distribution; The comprehensive boundary weights are normalized to a boundary-related distribution, and then mixed with the basic replay distribution according to equation (20) to obtain the dynamic sampling distribution for the current training stage. In this embodiment, the base playback ratio of equation (20) is taken as... That is, the probability quality is such that the boundary correlation distribution and the full training pool base replay distribution each account for 50%, and the sum of the dynamic sampling probabilities is 1.

[0080] S5: Model Training and Closed-Loop Update of Capability Boundaries Step S51: Construct training batches based on dynamic sampling probabilities; Based on the dynamic sampling probabilities obtained in step S44, scenarios are extracted from the training data pool, and training batches are constructed according to the batch organization method required by the end-to-end autonomous driving model. In this embodiment, the sample budget for each round is equal to the total number of training scenarios, i.e. The training batch size is 64 samples; when using weighted sampling with replacement, each round is based on the current stage. Reconstruct the training sequence. Alternatively, stratified sampling can be used while maintaining the overall probability ratio.

[0081] Step S52: Update the end-to-end autonomous driving model; The end-to-end autonomous driving model parameters are updated using the constructed training batches. The model continues to use its original inputs, outputs, loss functions, and optimization methods. This invention provides the model with a more focused capability boundary learning signal by changing the sampling probability of the training scenario, without relying on a specific network structure or planning head form.

[0082] Step S53: Re-evaluate the closed-loop performance of the updated model; After the model completes a preset training phase, the updated phase model is saved, and closed-loop evaluation is re-executed on a capability assessment scenario set consistent with the aforementioned evaluation conditions. The success or failure matrix is ​​updated based on the new scenario scores, enabling the evaluation results of the previous and subsequent phases to be compared on the same scale.

[0083] Step S54: Update the capacity boundary and sampling distribution; Re-estimate model capability based on updated success and failure matrix and scene difficulty Then, the capability boundary scene, boundary sphere center, boundary sphere radius, training scene comprehensive boundary weights, and dynamic sampling probabilities are updated sequentially. The updated sampling distribution is used in the next training phase, allowing the capability boundary to migrate as the model's capabilities change, and forming... Figure 6 The closed-loop update process is shown. Figure 6In the diagram, solid dots represent boundary-related training scenarios sampled with high probability in the current stage, hollow dots represent other training scenarios retained according to the base probability, and dashed circles represent the current capability boundary region. From stage 1 to stage 3, the boundary region migrates as the model's capability improves, and high-probability sampled scenarios synchronously move to the vicinity of the new boundary.

[0084] Step S55: Training terminates and model output; This embodiment reaches a total training limit of 100 rounds after completing 16 five-round training phases. If the training budget is exhausted, the closed-loop performance reaches the target, or the model capability increment is lower than the preset threshold before this, the training can be stopped early. After termination, the trained end-to-end autonomous driving model is output, and the final closed-loop verification is performed on an independent verification scenario set containing 5912 scenarios to confirm the model's performance in difficult scenarios and the overall scenario distribution.

[0085] It should be noted that this invention does not change the network structure, loss function, or inference process of the basic model; it only adjusts the selection of training samples according to the dynamic sampling probability. For real-world scenarios, uniform features are extracted and standardized and dimensionality reduced according to step S1. This result is only used for capability boundary localization and sampling weight calculation; model training still uses the original scene data.

[0086] Experimental verification results: To verify the effectiveness of the method of this invention, this embodiment uses a single-stage closed-loop evaluation benchmark containing 12,146 real-world scenarios, with the diffusion-based end-to-end planning model DiffusionDrive and the regression-based end-to-end planning model LTF as the base models. Evaluation metrics include no-fault collision (NC), drivable area compliance (DAC), driving direction compliance (DDC), traffic light compliance (TLC), vehicle progress (EP), collision time margin (TTC), lane keeping (LK), historical comfort (HC), extended comfort (EC), and the closed-loop comprehensive evaluation metric (EPDMS). All methods use the same scenario set and metric calculation rules.

[0087] Tables 1 and 2 list the two types of basic models and the closed-loop evaluation results after 100 training rounds using the method of this invention. The performance improvements of the DiffusionDrive (diffusion-based) and LTF (regressive end-to-end) planning models after using the method of this invention are compared. Table 1 compares the performance of No-Fault Collision (NC), Drivable Area Compliance (DAC), Direction of Travel Compliance (DDC), Traffic Light Compliance (TLC), and Vehicle Progress (EP). Table 2 compares the performance of Collision Time Margin (TTC), Lane Keeping (LK), Historical Comfort (HC), Extended Comfort (EC), and Closed-Loop Comprehensive Evaluation Metric (EPDMS).

[0088] Table 1. Closed-loop evaluation results of the method of the present invention on two types of end-to-end planning models (Part 1) Table 2. Closed-loop evaluation results of the method of the present invention on two types of end-to-end planning models (Part 2) As shown in Table 2, based on the DiffusionDrive basic model, EPDMS improved from 84.5 to 87.3, an absolute improvement of 2.8; based on the LTF basic model, EPDMS improved from 79.1 to 82.4, an absolute improvement of 3.3. The comprehensive evaluation metrics of both models were improved, indicating that the present invention's method of dynamically adjusting the sampling focus of the training scenario based on the current capability boundary does not rely on a single planning model structure.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. After reading this specification, those skilled in the art can make various modifications or variations to the technical solutions without departing from the concept of the present invention, and all such modifications or variations fall within the protection scope of the present invention.

Claims

1. An end-to-end autonomous driving training method based on closed-loop capability boundaries, characterized in that, include: S1: Scene feature extraction and unified scene space construction; A training scene set and a capability assessment scene set are obtained. A unified scene feature vector is extracted from each scene in the training scene set and the capability assessment scene set. Standardization parameters are calculated based on the training scene set. Standardization processing is performed on the training scene set and the capability assessment scene set. Then, the standardized feature vector of the training scene set is used to fit a dimension reduction projection relationship. Principal component analysis and whitening processing are used to map the training scene set and the capability assessment scene set to the same low-dimensional scene space to obtain the low-dimensional vector representation of each scene. S2: Multi-stage closed-loop evaluation and capability boundary positioning; The process involves acquiring the stage models saved during multiple training phases of the end-to-end autonomous driving model, performing closed-loop evaluation on the capability assessment scenario set for each stage model, jointly estimating the capability parameters of each stage model and the difficulty parameters of each assessment scenario, and determining the capability boundary scenarios. S3: Boundary sphere construction and training scene mapping; The capability boundary scene is mapped to the low-dimensional scene space, and clustering is performed on the low-dimensional vector of the capability boundary scene to determine the boundary-related training scene. S4: Dynamic sampling probability generation based on boundary sphere; Distance weights are generated based on the distance between each boundary-related training scenario and the corresponding boundary sphere center. After normalization, the distance weights are fused with the basic replay distribution of the entire training pool to generate the dynamic sampling distribution of the current training stage. S5: Model training and closed-loop update of capability boundaries; Training scenarios are sampled from the training data pool according to the dynamic sampling distribution, training batches are constructed and the end-to-end autonomous driving model is updated. After the model update reaches the preset stage, the closed-loop evaluation is re-executed and the capability boundary is updated in a closed loop. The closed-loop iterative training continues until the training termination condition is met, and then the trained end-to-end autonomous driving model is output.

2. The end-to-end autonomous driving training method according to claim 1, characterized in that, The unified scene feature vector is a 21-dimensional feature vector, including: 4-dimensional road structure features, including average curvature, maximum curvature, route complexity, and route length; 11-dimensional traffic participant features, including the number of vehicles, pedestrians, total number of participants, distance to the nearest participant, number of forward participants, distance to the nearest future participant, maximum number of forward future participants, participant speed variance, number of participants ahead, number of participants to the left, and distance to the nearest vehicle; 4-dimensional vehicle motion state features, including current speed, vehicle acceleration modulus, historical speed variance, and historical average heading angle change rate; and 2-dimensional scene interaction and risk features, including the scene minimum collision time and the scene maximum required deceleration.

3. The end-to-end autonomous driving training method according to claim 1, characterized in that, In step S1, the training scenario set and the capability evaluation scenario set are processed as follows: Step S13: Scene feature standardization; Since different scene features have different physical dimensions and numerical ranges, in order to reduce the impact of feature scale differences on scene similarity measurement, the original scene features are standardized. Calculate the first based on the training scenario set. The mean and standard deviation of the scene features are used to obtain the scene. Standardized scene feature vector: （7） in, Representing a scene The Dimensional standardization features ; The dimension representing the features of the original scene; Both the training scenario set and the capability assessment scenario set use the same set of mean parameters calculated from the training scenario set. and standard deviation parameter Standardize; Step S14: Scene feature dimensionality reduction and unified mapping; A dimensionality reduction mapping relationship is established for the standardized training scene features, and the same dimensionality reduction mapping relationship is used to process the training scene and the capability assessment scene, mapping them to the same low-dimensional scene space.

4. The end-to-end autonomous driving training method according to claim 3, characterized in that, A unified scene space is constructed using standardization, principal component analysis, and whitening. Only the training scene is used to fit the standardized parameters and projection relationships. The capability assessment scene and the validation scene use the same dimension reduction projection matrix as the training scene for dimension reduction projection and perform whitening processing using the same whitening parameters. They do not participate in the fitting calculation of standardized parameters and projection relationships. The covariance matrix is ​​calculated based on the standardized feature vectors of the training scenario. The covariance matrix is ​​then decomposed into eigenvalues, and the eigenvalues ​​are arranged in descending order. Before selection The eigenvectors corresponding to each eigenvalue constitute the dimension-reduced projection matrix and the corresponding eigenvalue matrix; Whitening is applied to the projected vectors to adjust the scale of each principal component direction and reduce the impact of principal component variance differences on scene distance calculation. The training scenario and the capability assessment scenario use the same standardized parameters, dimensionality reduction projection matrix, and scale adjustment parameters, and do not establish separate scene spaces.

5. The end-to-end autonomous driving training method according to claim 1, characterized in that, Step S2 is as follows: Step S21: Obtain the model for multiple training phases; Obtain multiple model checkpoints saved by the end-to-end autonomous driving model at different training stages; the model checkpoints come from the pre-training stage, the fine-tuning stage, and the iterative training stage; Step S22: Perform closed-loop evaluation and construct a success or failure matrix; The models at each stage are subjected to closed-loop evaluation on the same boundary evaluation scenario set. Based on the preset pass conditions, the evaluation results of the models in each scenario are represented as pass or fail, and a model-scenario success or failure matrix is ​​constructed. Success or failure variables can be represented as: （14） in, Indicates the scene number. This indicates the training phase; the closed-loop composite score will be used. Compared with the preset threshold Compare, and when the threshold is reached, let Otherwise ; Step S23: Jointly estimate model capability and scenario difficulty; A capability-difficulty probability model is established using the success-failure matrix, calibrating the model capability and scenario difficulty to a unified potential scale; Step S24: Locate the current model's capability boundary; The current capability boundary scenario is determined based on the difference between the current model capability and the scenario difficulty, or based on the model's predicted probability of passing the scenario.

6. The end-to-end autonomous driving training method according to claim 5, characterized in that, The Rasch model is used to describe the probability of the model passing the scene: （15） in, For the first Capability parameters of the stage model For the first Difficulty parameters for each scenario For the first The phase model is through the first The predicted probability of a scenario. It is a monotonic probability mapping function; The monotonic probability mapping function adopts the Logistic function. ;when When the model predicts the scenario, the probability is 0.5; when At that time, the probability of passing the prediction increases as the difference between ability and difficulty increases; Given a success or failure matrix, the model capability parameters and scenario difficulty parameters are estimated by maximizing the Bernoulli likelihood or minimizing the negative log-likelihood with regularization. The capability boundary is determined in one of the following forms: （16） （17） in, and These represent the lower and upper bounds of the capability boundary probability band, respectively; when the number of boundary candidates is less than the preset minimum number, the probability interval is expanded simultaneously by a preset step size until the quantity constraint is met; where, This is a preset capability-difficulty difference threshold, used to limit the degree of closeness between the scene difficulty and the current model capability.

7. The end-to-end autonomous driving training method according to claim 1, characterized in that, Step S3 is as follows: Step S31: Obtain a low-dimensional representation of the boundary scene; Map the capability boundary scene back to a low-dimensional scene space to obtain a low-dimensional vector representation of the capability boundary scene; Step S32: Determine the center of the boundary sphere; Let the first The set of capability boundary scenarios for each stage is The number of its scenes is The number of clusters is denoted as ;right The low-dimensional vectors of each boundary scene are clustered using K-Means, with each cluster center serving as the center of a boundary sphere; according to Adaptive determination, and with The upper limit; Step S33: Determine the radius of the boundary sphere; First, the basic radius is determined based on the nearest neighbor distance of the training scene near the center of the sphere. Then, the radius is adjusted according to the normalized scene difficulty of the corresponding boundary scene. Adjacent boundary spheres are merged and verified, and isolated clusters are processed as point boundaries and the nearest training scene is retained as support samples. Step S34: Retrieve boundary-related training scenarios from the training data pool; Calculate the distance between the training scene and the center of each boundary sphere, and determine the training scenes that satisfy the boundary sphere neighborhood condition as boundary-related training scenes; the neighborhood condition is expressed as: （18） in, For the first The vector representation of each training scene in the low-dimensional scene space. For the first The center of the boundary sphere, For the first The radius of the boundary sphere, This is a scene space distance metric function; Step S35: Handle overlapping multiple boundary spheres; When a training scenario falls into multiple boundary balls simultaneously, record the set of boundary balls corresponding to that training scenario. The distance and radius of each boundary sphere are used to form an overlapping relationship between the training scene and multiple boundary spheres.

8. The end-to-end autonomous driving training method according to claim 1, characterized in that, Step S4 includes: Step S41: Calculate the in-sphere distance weights for the training scenario; Sampling weights are generated based on the distance between the training scene and the center of the boundary sphere; the sampling weights are generated using a function that monotonically decreases with distance. Step S42: Merge the weights of multiple boundary spheres; For training scenarios that correspond to only one boundary ball, the distance weight corresponding to that boundary ball is directly used; for training scenarios that correspond to multiple boundary balls, the maximum value of the distance weights of each boundary ball is taken as the comprehensive boundary weight. Step S43: Set the base replay probability for the entire training pool; Establish a basic replay distribution across all training scenarios. Each effective training scenario has a non-zero base replay probability; Step S44: Generate a normalized dynamic sampling distribution; The boundary weight distribution is fused with the basic replay distribution to generate the first... The dynamic sampling distribution of the stage; the dynamic sampling distribution is expressed as: （20） in, For the first Phase 1 Sampling probability of each training scene To incorporate boundary weights, To ensure uniform playback distribution across the entire training pool, Based on the replay ratio.

9. The end-to-end autonomous driving training method according to claim 8, characterized in that, The weight of the distance inside the sphere is expressed as follows: （19） in, For the first The training scenario and the first The distance between the centers of the boundary spheres.

10. The end-to-end autonomous driving training method according to claim 1, characterized in that, Step S5 is as follows: Step S51: Construct training batches based on dynamic sampling probabilities; Based on the dynamic sampling probability obtained in step S4, training scenarios are sampled from the training data pool to construct training batches for model parameter updates. Step S52: Update the end-to-end autonomous driving model; The parameters of the end-to-end autonomous driving model are updated using the training batches obtained from sampling. Step S53: Re-evaluate the closed-loop performance of the updated model; After the model training reaches the preset stage, the updated model is put back into the boundary evaluation scenario set for closed-loop evaluation to obtain new scenario pass or failure results. Step S54: Update the capacity boundary and sampling distribution; Based on the updated success and failure matrix, the model capability parameters and scene difficulty parameters are re-estimated, and then the capability boundary scene, boundary sphere center, boundary sphere radius, training scene comprehensive distance weight and dynamic sampling probability are updated in sequence. Step S55: Training terminates and model output; When the preset number of training rounds is reached, the preset training budget is reached, the model performance converges, or the model capability increment is lower than the preset threshold, training stops and the trained end-to-end autonomous driving model is output.

Images