A strategy-aware autonomous driving demonstration data evaluation and selection method and system

Abstract

The present application relates to the technical field of automatic driving training data processing, in particular to a strategy-aware automatic driving demonstration data evaluation and selection method and system. The method constructs mutually isolated real training set, candidate demonstration pool, independent calibration set and final evaluation set, trains the target driving strategy and obtains the scene-by-scene result on the calibration scene; extracts the sample-by-sample gradient of the candidate demonstration, calibration scene and geometric subset within the planning parameter subspace of the target strategy, forms the strategy-aware calibration direction through fixed random projection and local training geometry with damping, and obtains the value of each candidate demonstration relative to the target strategy; selects high-value demonstrations according to the preset budget, and re-trains the complete strategy model using the original training target after merging with the real training data. The present application realizes data evaluation that changes with the target strategy on a demonstration-by-demonstration basis without explicitly constructing a counter-influence matrix, and maintains data isolation between calibration, selection and final evaluation.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving training data processing technology, specifically to a method and system for estimating and selecting autonomous driving demonstration data based on policy perception. Background Technology

[0002] End-to-end autonomous driving systems typically utilize onboard sensors such as cameras to obtain environmental observations and directly predict the vehicle's future trajectory or vehicle control variables through neural networks. As the scale of real-world collected data and simulated synthetic data increases, the training data exhibits significant heterogeneity in terms of road structure, traffic participant interactions, supervision quality, and generation sources. Directly incorporating all candidate demonstrations into the training set allows the target driving strategy to simultaneously absorb beneficial, redundant, or conflicting supervisory information.

[0003] Existing data selection methods often prioritize data based on the rarity of samples, scene difficulty, coverage, or visual similarity. While these attributes describe the appearance or distribution of candidate demonstrations, they fail to characterize the changes in training parameters caused by these demonstrations, nor can they determine whether these parameter changes align with the planning requirements of a specific driving strategy. Therefore, the same demonstration may have different effects in different trajectory generation structures, and a uniform ranking based on inherent sample attributes cannot reflect this strategy dependency.

[0004] Another type of approach uses metrics such as training loss, uncertainty, or gradient norm to measure the importance of samples. These metrics are related to the current model training state, but typically represent optimization difficulty or gradient magnitude, without establishing a directional link between candidate demonstrations and downstream driving results. Directly differentiating the final driving evaluation metric is limited by the discreteness and non-differentiability of the evaluation process, as well as the high computational cost.

[0005] Per-sample retraining allows direct observation of policy changes after deleting or adding candidate demonstrations, but performing complete training and evaluation for each candidate demonstration separately is not feasible for large-scale image-based driving models and large-scale candidate pools. Classical local influence calculation requires handling second-order training geometry in a high-dimensional parameter space and may generate pairwise influence matrices of the size of the product of the number of candidate demonstrations and the number of calibration scenes, resulting in significant storage and solution costs.

[0006] Furthermore, mixing the calibration scenario used for generating data sorting with the final evaluation scenario can easily lead to the candidate demonstration selection process absorbing evaluation distribution information. Freezing some parameters to reduce evaluation overhead may incorrectly carry over this constraint to the final policy training, reducing the compatibility between the evaluation method and the native driving policy. Therefore, a technical solution is needed that can, under conditions of strict data set isolation, transfer scenario-level driving results to a differentiable supervised planning gradient and calculate the demonstration value of each candidate demonstration relative to the target driving policy in a scalable manner. Summary of the Invention

[0007] The technical problem to be solved by this invention is: under the conditions that the image-driven end-to-end autonomous driving strategy has high-dimensional parameters, the downstream driving results are not directly differentiable, and the number of candidate demonstrations is large, how to obtain the demonstration training value of each candidate demonstration corresponding to the target driving strategy without retraining each candidate demonstration separately or using the final evaluation set to participate in the selection, and to construct traceable budgeted training data based on the training value to support the subsequent training of the complete driving strategy model.

[0008] Technical solution To address the aforementioned technical problems, this invention provides a method for estimating and selecting demonstration data for autonomous driving based on policy awareness, comprising the following steps: Construct a real training set, a candidate demonstration pool consisting of synthetic driving demonstrations generated by driving pseudo-experts, and an independent calibration set pre-reserved and fixed from the real training data. Use the real training set to train the target driving strategy and obtain basic parameters. The target driving strategy is used to perform driving trajectory prediction and scenario-level driving result calculation for each calibration scenario in the independent calibration set. The scenario-level driving results are then centered and normalized according to the standard deviation of the entire calibration set to obtain the result weights. In subsequent gradient calculations, the result weights are used as constants for stopping gradients. Within the planning parameter subspace of the target driving strategy, based on the same basic parameters and the same supervised planning loss, the sample-by-sample gradients of the candidate demonstration pool, the independent calibration set, and the geometric subset obtained from the real training set are extracted respectively. The sample-by-sample gradient is mapped to a low-dimensional projected gradient using the same fixed random projection. A damped local training geometric matrix is ​​constructed based on the projected gradient of the geometric subset. A policy-aware calibration direction is constructed based on the projected gradient of the independent calibration set and the result weights. The preconditioning direction determined by the local training geometric matrix and the policy-aware calibration direction is solved. The inner product of the projected gradient of each candidate demonstration and the preconditioning direction is calculated to obtain the demonstration value of each candidate demonstration relative to the target driving strategy. The demonstration values ​​are sorted in descending order, and several synthetic driving demonstrations with the largest values ​​are selected according to the budget of the number of candidate demonstrations. The selected synthetic driving demonstrations are merged with the real training set, and the complete strategy model is retrained according to the original training objective of the target driving strategy. The retrained full policy model is used to process the vehicle perception input and generate the vehicle's future trajectory. The constraints of the planning parameter subspace only apply to the demonstration value estimation and do not constitute a restriction on the retraining of the full policy. The final evaluation set is not included in the data valuation and selection process.

[0009] Furthermore, the candidate demonstration pool consists only of synthetic driving demonstrations generated by driving pseudo-experts in the simulation environment, and the independent calibration set consists only of real driving scenarios reserved in advance in the partitioning of real training data and is not used for basic policy training; the final evaluation set does not participate in basic policy training, result weight generation, sample-by-sample gradient extraction, demonstration value calculation and candidate demonstration selection, so as to prevent the final evaluation information from entering the candidate demonstration selection process.

[0010] Furthermore, the scenario-level driving results use the collision-free index NC, drivable area compliance index DAC, driving direction compliance index DDC, and traffic signal compliance index TLC as multiplicative constraints, and route progress EP, collision time TTC, lane keeping LK, heading comfort HC, and extended comfort EC as weighted quality terms. Calculate scenario-level driving results: in, For the first Scene-level driving results for each scenario , , , , , , , and The first The values ​​of NC, DAC, DDC, TLC, EP, TTC, LK, HC, and EC for each scenario; , , , and represents the weight of each quality item, which is a non-negative value and its sum is 1; The weights of the results are calculated as follows: in and These represent the mean and non-zero standard deviation of all calibration scenario-level driving results, respectively. It is a positive stability constant; The variance of the scenario-level driving results in the independent calibration set is greater than zero, and the sum of the weights of the results of all calibration scenarios is zero. This causes the gradient components common to all calibration scenarios to cancel each other out in the weighted aggregation, while the differentiated planning gradients related to the higher and lower scenario-level driving results are retained.

[0011] Furthermore, the supervised planning loss is the native loss used by the target driving strategy to supervise the generation of the vehicle trajectory. The candidate demonstration, calibration scenario, and geometric subset all use the same supervised planning loss, parameter arrangement, basic parameters, and fixed random projection to ensure the consistency of the demonstration value estimation.

[0012] The planning parameter subspace includes trajectory generation head parameters that directly generate the vehicle trajectory, and excludes parameters corresponding to target detection loss and bird's-eye view semantic loss that do not directly supervise the vehicle trajectory during the sample-by-sample gradient extraction process, so as to avoid auxiliary tasks unrelated to trajectory planning from changing the meaning of demonstration value; the planning parameter subspace is only used for demonstration value estimation, and the complete policy model retains all native trainable parameters and training objectives during retraining.

[0013] In one embodiment of the present invention, when the target driving strategy is a diffusion trajectory planning strategy, the supervised planning loss includes modal classification loss and trajectory regression loss obtained by iterative denoising of the noisy trajectory anchor points, and averages the diffusion time steps and multiple random implementations of the noise variables.

[0014] Furthermore, the dimension of the planning parameter subspace is P, the projection dimension p is less than P, and the elements of the fixed random projection... Independently obeying a set of rules with a mean of zero and a variance of 1. The Gaussian distribution is used; the candidate demonstration, calibration scene and geometric subset share the projection seed, and the projection coefficients are regenerated sequentially according to the preset parameter block to avoid materializing the complete projection matrix in memory, thereby achieving scalable gradient projection on a large-scale candidate pool.

[0015] Furthermore, the local training geometric matrix is ​​composed of the sum of the normalized second moment of the projected gradient matrix of the geometric subset and the positive damping coefficient multiplied by the identity matrix, as follows: Among them, matrix It is the damped projected gradient second moment. Let be the projection gradient matrix, and let the projection dimension be . , The number of samples in the geometric subset. It is a positive damping coefficient. The identity matrix has the same dimensions as the projection dimension; The positive damping coefficient makes the local training geometry matrix positive definite and limits the numerical amplification of low-support projection directions; the policy-aware calibration direction is obtained by transposing the projection gradient matrix of the independent calibration set and then combining it with the resulting weight vector. Multiplying them together yields: ; in, To calibrate the projection gradient matrix; Solve linear equations The precondition direction is obtained, where Combine the downstream results of the target driving strategy in the calibration scenario with differentiable supervised planning gradients; The policy-aware direction is obtained after local training and geometric preconditioning. The demonstrative value of all candidate demonstrations is obtained simultaneously by performing a first-order matrix-vector multiplication of the candidate projection gradient matrix with the preconditioning direction. .in Indicates the first The local training value of each candidate demonstration relative to the current target driving strategy and the current independent calibration target. Indicates the first The original sample-by-sample gradient of the candidate demonstration 3D projection gradient, This indicates transpose.

[0016] Furthermore, the policy-aware calibration direction is obtained by multiplying the transposed projection gradient matrix of the independent calibration set with the result weight vector; after solving the linear equation formed by the local training geometric matrix and the policy-aware calibration direction to obtain the precondition direction, the demonstration value of all candidate demonstrations is obtained simultaneously by a first matrix-vector multiplication of the candidate projection gradient matrix and the precondition direction, without the need to store the pairwise influence matrix of the product of the number of candidate demonstrations and the number of calibration scenarios.

[0017] Preferably, the budget for the number of candidate demonstrations is B. The B driving demonstrations with the largest values ​​are selected from the candidate demonstration pool in descending order of demonstration value, where B is the rounded-up value corresponding to a preset proportion of the total number N in the candidate demonstration pool.

[0018] In some embodiments, when the same candidate demonstration pool is used for multiple driving strategies, the corresponding demonstration value is calculated using the scenario-level driving results, supervised planning loss, planning parameter subspace, fixed random projection, and local training geometric matrix of each driving strategy, and a candidate demonstration allowable list is generated for each driving strategy, so that the same candidate demonstration pool can form a combination of training data that changes with the target driving strategy. The multiple driving strategies include at least two of the following: diffusion trajectory planning strategy, direct regression trajectory planning strategy, and flow matching trajectory planning strategy; the direct regression trajectory planning strategy uses the supervised regression loss of the trajectory regression head as the supervised planning loss; the flow matching trajectory planning strategy uses the combined supervised loss of the flow generation head and the trajectory reconstruction head as the supervised planning loss.

[0019] This invention also provides a strategy-aware autonomous driving demonstration data estimation and selection system to implement the above-mentioned method. It includes a data isolation module, a basic policy training module, a calibration result acquisition module, a planning gradient extraction module, a projection geometry construction module, a demonstration value calculation module, a Top-B budget selection module, a full model retraining module, and a trajectory generation module.

[0020] Among them, the data isolation module establishes non-repeating data identifiers and persistent lists for the real training set, candidate demonstration pool, independent calibration set, final evaluation set and geometric subset, and provides them to subsequent modules; The basic strategy training module is data-connected to the data isolation module and is used to train the target driving strategy using the real training set to obtain basic parameters. Basic checkpoints are provided to the calibration result acquisition module and the planning gradient extraction module. The calibration result acquisition module is connected to the basic strategy training module and is used to perform scenario-level driving result calculation on each calibration scenario in the independent calibration set using the target driving strategy. The scenario-level driving results are then centered and normalized according to the standard deviation of the entire calibration set to generate the result weights of the stopping gradient. The planning gradient extraction module is used to extract the sample-by-sample gradients of the candidate demonstration pool, the independent calibration set, and the geometric subset based on the same basic parameters, the same supervised planning loss, and the same planning parameter subspace; the sample-by-sample gradients are streamed to the projection geometry construction module. The projection geometry construction module is data-connected to the planning gradient extraction module. It is used to map the sample-by-sample gradient to a low-dimensional projection gradient using the same fixed random projection, to obtain the candidate projection gradient, calibration projection gradient matrix and geometric subset projection gradient matrix. The geometric subset projection gradient matrix is ​​used to construct a damped local training geometry matrix. The calibration projection gradient matrix is ​​multiplied by the result weight vector to construct the policy-aware calibration direction. The demonstration value calculation module is data-connected to the projection geometry construction module and is used to solve the linear equation determined by the local training geometry matrix and the policy-aware calibration direction to obtain the precondition direction and calculate the demonstration value of each candidate demonstration. The Top-B budget selection module is connected to the demonstration value calculation module. It is used to sort the demonstration values ​​in descending order and select the composite driving demonstrations with the largest values ​​according to the number of candidate demonstrations to generate a candidate demonstration allowable list. The full model retraining module is connected to the Top-B budget selection module and the data isolation module. It is used to merge the synthetic driving demonstration with the real training set according to the candidate demonstration allowable list, unfreeze the upstream parameters in the valuation stage, and retrain the complete strategy model according to the original training objective. The trajectory generation module receives vehicle perception input, uses the retrained complete policy model to generate the vehicle's future trajectory, and provides it to the vehicle controller.

[0021] The system supports diffusion-based trajectory planning, direct regression-based trajectory planning, and flow-matching trajectory planning through a policy adaptation interface, enabling different policies to utilize their own result weights, supervised planning losses, parameter subspaces, and local training geometric computation demonstration values.

[0022] Beneficial effects Compared with the prior art, the present invention has the following advantages: (1) By using the scenario-level driving results of the target driving strategy itself to weight the calibration gradient, the demonstration value depends on both the training effect of the candidate demonstration and the planning behavior of the target strategy, thus avoiding the data value being regarded as an inherent attribute of the sample that is unrelated to the strategy.

[0023] (2) By stopping the gradient result weight, the non-differentiable scenario-level driving result is transformed into the calibration direction of the differentiable supervised planning loss, without requiring differentiation of the driving evaluation process itself.

[0024] (3) By using fixed random projection, damped local training geometry and solving linear equations, the high-dimensional sample-by-sample influence calculation is compressed into matrix-vector operations of candidate projection gradient and single precondition direction, without the need to construct all paired influence matrices.

[0025] (4) By identifying and isolating the real training set, candidate demonstration pool, independent calibration set and final evaluation set and persisting the list, the final evaluation information is prevented from entering the demonstration value and candidate selection process.

[0026] (5) By distinguishing the planning parameter subspace used for value estimation from the complete policy parameters used for final retraining, the native training capability of each target driving policy is preserved while controlling the valuation overhead.

[0027] (6) By using the strategy adaptation interface, different trajectory generation structures are valued separately, so that the same candidate demonstration pool can form a combination of training data that changes with the target driving strategy. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method for estimating and selecting autonomous driving demonstration data based on the strategy perception of the present invention. Figure 2 This is a schematic diagram illustrating the isolation relationship between the real training set, candidate demonstration pool, independent calibration set, geometric subset, and final evaluation set of this invention. Figure 3 This is a schematic diagram illustrating the calculation relationship between the projection gradient, policy-aware calibration direction, local training geometric matrix, and the demonstration value of each candidate demonstration in this invention. Figure 4 This is a schematic diagram of the unified valuation interface for diffusion-based, regression-based, and flow-matching driving strategies of the present invention. Figure 5 This is a schematic diagram of the overall structure of the autonomous driving demonstration data estimation and selection system of the present invention; Figure 6 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make corresponding adjustments to the data scale, driving strategy structure, random projection dimension, and training parameters without departing from the core technical concept of the present invention.

[0030] Terminology and Symbol Conventions: (1) Demonstration: A training unit consisting of vehicle observation history and corresponding reference vehicle trajectory; the observation history may include multi-view camera images, driving commands, vehicle speed and vehicle acceleration.

[0031] (2) Target driving strategy: an end-to-end neural network that accepts vehicle observations and outputs future vehicle trajectories; in different embodiments, a diffusion-type, regression-type, or flow-matching trajectory generation structure may be used.

[0032] (3) Scenario-level driving results: downstream driving evaluation values ​​of the target driving strategy in a calibrated scenario; the evaluation values ​​can be obtained by a combination of one or more indicators among safety, road compliance, driving progress, collision time, lane keeping and comfort.

[0033] (4) Planning parameter subspace: The set of parameters that directly participate in the generation of the vehicle trajectory in the target driving strategy; this set is used for demonstration value estimation and does not restrict the retraining of the final complete strategy.

[0034] (5) Strategy perception: Demonstration value is determined by the parameters of the target driving strategy, the supervised planning loss, the scenario-level driving results, and the local training geometry.

[0035] (6) Driving pseudo-expert: A programmed driving subject that generates a reference vehicle trajectory in a simulation environment based on preset driving rules, planners or controllable generation strategies; its output serves as a supervisory signal for the synthetic driving demonstration and does not represent an expert with the identity of a real driver.

[0036] (7) Training geometry: a local matrix used to describe the sensitivity of the training target to different parameter directions near the basic parameters; the present invention approximates the training geometry with the damped second moment of the gradient of the geometric subset projection and uses its inverse to the calibration direction.

[0037] (8) Hessian matrix: A matrix composed of the second-order partial derivatives of the training target with respect to the model parameters, used to represent the precise second-order curvature near the basic parameters; the present invention does not directly calculate the precise Hessian matrix.

[0038] Example 1: Overall Method Reference Figure 1 This embodiment provides an offline data processing flow consisting of data isolation, basic policy training, calibration result acquisition, planning gradient extraction, projective geometry construction, demonstration value calculation, budget selection, and complete policy retraining. This flow generates a candidate demonstration allowable list for subsequent end-to-end driving policy training without changing the input / output interface of the vehicle during online inference.

[0039] A method for estimating and selecting demonstration data for policy-aware autonomous driving includes the following steps: Step S1: Construct the basic strategy for dataset merging and training.

[0040] The real training set is defined as The candidate demonstration pool is Independent calibration set is , This represents the total number of candidate demonstrations in the candidate demonstration pool. This represents the total number of independent calibration scenarios, of which The first candidate demonstration pool Candidate demonstrations, For the independent calibration set Each calibration scenario consists of a candidate demonstration pool comprised solely of synthetic driving demonstrations generated by pseudo-experts in a simulated environment, while the independent calibration set comprises only real driving scenarios pre-reserved from the real training data. A pseudo-expert is a programmed driving entity that generates a reference vehicle trajectory based on preset driving rules, a planner, or a controllable generation strategy. , Indicates vehicle observation, This indicates that the trajectory is referenced from the vehicle's own trajectory. Training target driving strategy Obtain basic parameters .

[0041] Reference Figure 2 Independent calibration set Fixed before calculating the demonstration value, the final evaluation set It does not enter the training, calibration, valuation, or selection process.

[0042] The problem this invention aims to solve is budget. The target subset selection problem is defined by the budget, which refers to the maximum number of synthetic driving demonstrations allowed to be selected from the candidate demonstration pool into the training set, as follows: in This indicates the use of the real training set and candidate subsets. The parameters obtained during training, This represents the downstream driving utility of the trained policy. Since direct solution requires retraining on a large number of candidate subsets, this invention constructs a local estimate near the basic parameters.

[0043] Step S2: Obtain scenario-level driving results on the independent calibration set.

[0044] Using basic strategies Each scenario's execution trajectory prediction and scenario-level driving evaluation are obtained. Calculate the mean of the calibration results. and standard deviation And introduce a positive stability constant. Obtain the result weights .in, To avoid numerical instability in the denominator, this embodiment takes... .

[0045] In a scenario-level driving outcome implementation, the collision-free index (NC), drivable area compliance index (DAC), driving direction compliance index (DDC), and traffic signal compliance index (TLC) are used as multiplicative constraints, while route progress (EP), collision time (TTC), lane keeping (LK), heading comfort (HC), and extended comfort (EC) are used as weighted quality terms. Each index is first mapped to the same numerical range, and the weights of the quality terms are non-negative and summed to one. The scenario-level driving outcome can be formed by the following formula: in, , , , , , , , and The first The values ​​of NC, DAC, DDC, TLC, EP, TTC, LK, HC, and EC for each scenario; , , , and represents the weight of each quality item, which is a non-negative value and its sum is 1; Before standardization is implemented, the processor checks... Is the value greater than zero? If it equals zero, then rebuild or expand the independent calibration set until the scenario-level driving results have a non-zero standard deviation. (All) In subsequent differentiation, the gradient is fixed as a constant by stopping the gradient operation and does not participate in backpropagation.

[0046] Since the sum of the weights after standardization is zero, the gradient components common to all calibration scenarios cancel each other out in the weighted aggregation, while the differentiated planning gradients associated with the driving results of higher and lower scenario levels are preserved.

[0047] Step S3: Extract the sample-by-sample gradient under the unified supervised planning loss.

[0048] The target driving strategy parameters are divided into upstream parameters. and planning parameters ,in This refers to network parameters that do not directly output the vehicle trajectory, such as perceptual coding and feature fusion. This refers to the planning parameters that directly participate in the generation of the vehicle's trajectory. This represents the total number of scalar parameters in the planning parameter subspace. The upstream parameter values ​​and planning parameter values ​​in the basic checkpoints obtained from step S1 training are respectively denoted as... and During the valuation period, fixed and in The expected planning loss for each demonstration is defined using the native vehicle trajectory supervision loss of the target policy: in, This represents a driving demonstration. Indicates the first in the candidate demonstration pool One driving demonstration, This represents the native supervised planning loss of the target driving strategy. This represents a random variable in the original training objective. The upstream parameter values ​​are those obtained from the basic checkpoints trained in step S1, and are kept fixed during the demonstration value estimation. This represents the planning parameter values ​​in the basic checkpoints obtained from training in step S1. This indicates that the upstream parameters are fixed at... At that time, driving demonstration Regarding planning parameters Expected monitoring of planning losses, Indicates information about planning parameters gradient calculation, That is, taking the expectation of the random variable. Indicates the first Candidate demonstration items regarding planning parameters The sample-by-sample gradient; when the target strategy is a deterministic regression strategy, Empty. Candidate demonstrations, calibration scenarios, and geometric subsets use the same supervised planning loss, the same basic parameters, the same parameter arrangement, and the same operating mode. To avoid auxiliary tasks unrelated to trajectory planning altering the meaning of demonstration value, per-sample gradients can exclude the parameters and losses corresponding to object detection and bird's-eye view semantic segmentation.

[0049] To illustrate the local effect of the aforementioned valuation, an empirical target is defined for the real training set within the planning parameter subspace. and candidate demonstrations Apply infinitesimal weights : in for In basic planning parameters The second-order training geometry at the location, To improve candidate demonstrations based on the real training set Apply infinitesimal weights Then, the optimal solution of planning parameters obtained by minimizing the training objective is re-applied. This is then applied to the calibration scenario. Define monitoring planning loss and gradient The local impact of the candidate demonstration on this calibration loss is as follows: in As a candidate demonstration For calibration scenarios The local impact function value of monitoring planning loss. This indicates the positive contribution of the candidate demonstration to the calibration loss; when When the value is greater than zero, it indicates that an additional candidate demonstration is added in the local first-order sense. The weight will decrease in the calibration scenario. The loss of oversight and planning.

[0050] This invention does not directly materialize all of them. Instead, a scalable approximation is obtained through the projection, aggregation, and linear solution process in steps S4-S7.

[0051] Step S4: Form the low-dimensional projection gradient.

[0052] like Figure 3 As shown, let the dimension of the planning parameter subspace be... Choose less than projection dimension Pre-generated fixed random projection Its elements Independently obeying orders with a mean of zero and a variance of 1 The Gaussian distribution is applied to the candidate, calibration, and geometric subset gradients. : in, Indicates the first The original sample-by-sample gradient of each candidate demonstration. Indicates the first The original sample-by-sample gradient for each calibration scenario, and They represent the corresponding 3D projection gradient.

[0053] In an optional implementation, the projection matrix is ​​not stored as a whole; the processor regenerates the projection coefficients sequentially according to a fixed random seed and a fixed parameter block length, completing the gradient projection block by block. The candidate demonstration pool, independent calibration set, and geometric subset must reuse the same random seed, parameter order, and projection block.

[0054] Step S5: Construct damped local training geometry.

[0055] Select a geometric subset from the real training set When selecting, first exclude independent calibration sets, then perform stratified random sampling according to road structure, so that... It covers the main scene types in the real training set.

[0056] Let its projection gradient matrix be . The sample size is The projection dimension is Constructing the local training geometry matrix: matrix It is the damped projected gradient second moment, used to approximate the local training geometry near the fundamental parameters. The inverse action used to approximate the training geometry, rather than the inverse of the exact Hessian matrix. Positive damping coefficient. make It is a positive definite matrix, thus ensuring that the subsequent linear equations have a unique solution and restricting the geometric subset to support numerical amplification in weaker directions.

[0057] Step S6: Aggregate strategy to perceive calibration direction.

[0058] Let the calibration projection gradient matrix be... The resulting weight vector is , build: in Combine the downstream results of the target driving strategy in the calibration scenario with differentiable supervised programming gradients; The policy-aware direction is obtained after local training and geometric preconditioning.

[0059] This step only requires... A linear system is solved once.

[0060] Step S7: Calculate the demonstration value of each candidate demonstration and perform budget selection.

[0061] For each candidate demonstration ,calculate: This represents the local training value of a candidate demonstration relative to the current target driving strategy and the current independent calibration target. The value of all candidate demonstrations can be expressed by the candidate projection gradient matrix. and Obtained through a single matrix-vector multiplication, requiring no storage. The candidate-calibrated pairwise influence matrix.

[0062] processor according to Sort by largest to smallest, and select the largest value. A synthetic driving demonstration is performed. The selection results, along with candidate identifiers, source identifiers, and target strategy identifiers, are written into the permitted list.

[0063] Step S8: Perform full policy retraining.

[0064] Merge the candidate demonstrations corresponding to the allowed list with the complete real training set, restore all native trainable parameters, auxiliary tasks and training configurations of the target policy, use the basic parameters obtained in step S1 as initialization, unfreeze the upstream parameters in the evaluation stage, and make all native trainable parameters actually updated in the fixed training process; execute one hundred training cycles according to the native optimizer, learning rate scheduling and batch size corresponding to the target policy.

[0065] After retraining is completed, the target driving strategy receives images from the vehicle camera, driving commands, and the vehicle's motion status to generate the vehicle's future trajectory. The vehicle's future trajectory is used by the vehicle controller to generate steering, braking, and drive control commands. The final evaluation set is only used after the above training process is completed.

[0066] Example 2: Diffusion-based trajectory planning strategy Reference Figure 4 In one diffusion-based strategy embodiment, vehicle input includes current images from the left front, front, and right front cameras. These three images are stitched together and scaled to a preset size, then input to an image encoder along with driving commands, vehicle speed, and vehicle acceleration. The strategy output consists of multiple vehicle poses sampled at fixed time intervals over the next four seconds.

[0067] Diffusion-type track head maintenance Cleaning trajectory anchor points During training, the sampling diffusion time step is performed. and Gaussian noise The noise-adding anchor points are obtained as follows: in, This refers to the noisy anchor point of the k-th trajectory anchor point at the diffusion time step t; This represents the cumulative noise scheduling coefficient corresponding to time step t during the diffusion process; The noise sampled at the k-th anchor point is Gaussian noise. The transformation to normalize the anchor points to the standard normal distribution space, This is the inverse transformation of the above transformation; This indicates that the value will be truncated to [ ]. Clipping operation on the interval [1,1].

[0068] The trajectory head predicts the trajectories and modal scores for each modality based on the noisy anchor points and scene representation. The supervised modality is determined based on the average positional distance between the clean anchor points and the reference trajectory. The supervised planning loss of the diffusion strategy is composed of the modality classification loss and the trajectory regression loss of the supervised mode. : in, This represents the total number of random samples taken for the same driving demonstration; Indicates the random sampling sequence number; and They represent the first The diffusion time step and Gaussian noise obtained from the random sampling; Indicates the first The modal score vectors output by the trajectory head under each random sampling; This represents the supervisory mode determined based on the average positional distance between the cleaning trajectory anchor point and the reference trajectory; Represents the supervisory mode The corresponding one-hot label vector; This indicates that the trajectory head is for the supervised mode under the (r)th random sampling. The output is the predicted vehicle trajectory; This indicates that the reference is the vehicle's trajectory; Indicates to The expected supervised planning loss is obtained by averaging the diffusion-based supervised planning loss from each random implementation. , , respectively, are the weight coefficients of the modality classification loss and the trajectory regression loss, and are non-negative real numbers; Focal loss function is used for supervision of modality classification.

[0069] Candidate demonstrations, calibration scenarios, and geometric subsets each use the same number of time steps and noise samples. The planned parameter subspace can take a trajectory generation head, which includes anchor point encoding, time step conditional encoding, iterative decoding, and a trajectory prediction layer. An exemplary setup could employ twenty trajectory anchor points, a two-level decoder, and a projection dimension. Number of geometric subset samples Calibration set scene number Positive stability constant Positive damping coefficient and the budget for the number of candidate demonstrations A fixed random projection seed is reused among candidate, calibration, and geometric subsets. The above settings are for illustrative purposes only and do not constitute a limitation on the scope of protection.

[0070] The complete policy retraining uses the AdamW optimizer with an initial learning rate of It employs a three-cycle preheating cosine scheduling with a minimum learning rate. The weight decays to The batch size for each graphics processor is 64, and the training runs for 100 cycles across four graphics processors.

[0071] Example 3: Direct Regression Trajectory Planning Strategy In the direct regression strategy implementation, the upstream image encoder and the fusion network generate a scene representation. Trajectory regression head Directly output a vehicle trajectory within a reference time range. Demonstration value estimation uses trajectory regression head parameters as the planning parameter subspace and reference trajectory. and predicted trajectory The weighted L1 loss is used as the supervised planning loss. : in, is the weighting coefficient of the trajectory regression loss, and is a non-negative real number.

[0072] Since the trajectory regression loss does not include policy-specific random sampling variables, the planning gradient of a single demonstration can be directly calculated. The remaining steps for projection, local training geometry, policy-aware calibration direction, demonstration value of each candidate demonstration, and budget selection are the same as in Example 1.

[0073] In the exemplary setting of this embodiment, the projection dimension Number of geometric subset samples Calibration set scene number Positive stability constant Positive damping coefficient and the budget for the number of candidate demonstrations Fixed random projection seeds are reused among candidate, calibration, and geometric subsets.

[0074] Full policy retraining uses the Adam optimizer with an initial learning rate and a constant learning rate. Without using weight decay, the batch size is 64 per GPU, and the training is performed for 100 epochs on four GPUs.

[0075] Example 4: Flow Matching Trajectory Planning Strategy In the stream-matching strategy implementation, the stream generation head generates multiple trajectory proposals based on the scene representation and time variables, and the trajectory reconstruction head aggregates these trajectory proposals into the final vehicle trajectory. The planning parameter subspace includes both the stream generation head and the trajectory reconstruction head to avoid overlooking the final trajectory reconstruction role by attributing it solely to proposal generation.

[0076] The reference trajectory is represented as an incremental trajectory. Sampling prior noise and satisfaction Time variables, construct and target speed The weighted sum of flow generation loss and trajectory reconstruction loss constitutes the supervised planning loss. : in, , , respectively, are the weighting coefficients of the flow generation head loss and the trajectory reconstruction head loss, and are non-negative real numbers; For trajectory reconstruction head to reference trajectory The reconstruction prediction results; Candidate demonstrations, calibration scenarios, and geometric subsets use consistent noise and time variable sampling rules. For different target policies, even if the independent calibration scenario lists and candidate demonstration pools are the same, scenario-level driving results, result weights, sample-by-sample gradients, random projections, local training geometry, and demonstration values ​​are calculated separately, without directly comparing the absolute values ​​of demonstration values ​​between different policies.

[0077] In the exemplary setting of this embodiment, the projection dimension Number of geometric subset samples Calibration set scene number Positive stability constant Positive damping coefficient and the budget for the number of candidate demonstrations Fixed random projection seeds are reused among candidate, calibration, and geometric subsets.

[0078] The complete policy retraining uses the AdamW optimizer with an initial learning rate of It employs a three-cycle preheating cosine scheduling with a minimum learning rate. Without using weight decay, the batch size is 64 per GPU, and the training is performed for 100 epochs on four GPUs.

[0079] Experimental Verification: On the NavTest benchmark of NAVSIM v2 (NAVSIM v2 is an open-source simulation benchmark platform designed for end-to-end autonomous driving policy evaluation, and NavTest is a standardized test benchmark set for NAVSIM v2), using Example 2 as the target driving policy, real navtrain demonstrations and the SimScale synthetic candidate pool were used, and only the top 25% of synthetic demonstrations ranked by the method of this invention were retained to perform full policy retraining. The results show that the extended predictive-driver-model score (EPDMS) of the retrained policy is 88.3, ​​which is 3.5 points higher than the 84.8 obtained by training with only real data, and 2.8 points higher than the 85.5 obtained by training with the full synthetic candidate pool.

[0080] The Extended Predictive Driver Model Score (EPDMS) test results for Examples 2-4 are shown in Table 1: Table 1. Test results (EPDMS scores) for Examples 2-4 The above results demonstrate that, under the aforementioned implementation conditions, the policy-aware demonstration estimation and budget selection of the present invention can achieve higher comprehensive planning performance with less synthetic training data.

[0081] Example 5: Systems and Electronic Devices Reference Figure 5 The strategy-aware autonomous driving demonstration data valuation and selection system 100 includes a data isolation module, a basic policy training module 110, a calibration result acquisition module 120, a planning gradient extraction module 130, a projection geometry construction module 140, a demonstration value calculation module 150, a Top-B budget selection module 160, a full model retraining module 170, and a trajectory generation module 180.

[0082] The data input for the strategy-aware autonomous driving demonstration data estimation and selection system 100 includes real training sets. (101) Candidate Demonstration Pool (102) and independent calibration set (103). Among them, the real training set (101) Used to train the basic driving strategy and obtain a geometric subset from it for constructing the local training geometry; candidate demonstration pool (102) Synthetic driving demonstrations generated by driving pseudo-experts are used for calculating the demonstration value of each candidate demonstration and selecting the Top-B budget; independent calibration set (103) It consists of real driving scenarios reserved in advance in the division of real training data, and is used to generate result weights and calibration gradients. It is not used as basic strategy training data, candidate demonstrations or final evaluation data.

[0083] Each module can be implemented by one or more processors executing software instructions, or it can be implemented by the central processing unit and the graphics processing unit working together.

[0084] Among them, the data isolation module ( Figure 5 (Not shown in the image) To establish unique data identifiers and persistent lists for the real training set, candidate demonstration pool, independent calibration set, final evaluation set, and geometric subset, and provide them to subsequent modules.

[0085] The basic strategy training module 110 is data-connected to the data isolation module and is used to train the target driving strategy using the real training set to obtain basic parameters. Basic checkpoints are provided to the calibration result acquisition module 120 and the planning gradient extraction module 130.

[0086] The calibration result acquisition module 120 is connected to the basic strategy training module 110 for data connection. It is used to perform scenario-level driving result calculation on each calibration scenario in the independent calibration set using the target driving strategy, and to perform full calibration set centering and standard deviation normalization on the scenario-level driving results to generate the result weights of the stopping gradient.

[0087] The planning gradient extraction module 130 is used to extract the sample-by-sample gradients of the candidate demonstration pool, the independent calibration set, and the geometric subset based on the same basic parameters, the same supervised planning loss, and the same planning parameter subspace; the sample-by-sample gradients are provided to the projection geometry construction module 140 in a streaming manner. The projection geometry construction module 140 is data-connected to the planning gradient extraction module 130. It is used to map the sample-by-sample gradient to a low-dimensional projection gradient using the same fixed random projection, to obtain the candidate projection gradient, calibration projection gradient matrix and geometric subset projection gradient matrix. The geometric subset projection gradient matrix is ​​used to construct a damped local training geometry matrix. The calibration projection gradient matrix is ​​multiplied by the result weight vector to construct the policy-aware calibration direction. The demonstration value calculation module 150 is data-connected to the projection geometry construction module 140 and is used to solve the linear equation determined by the local training geometry matrix and the policy-aware calibration direction to obtain the precondition direction and calculate the demonstration value of each candidate demonstration. Top-B budget selection module 160 is data-connected to the demonstration value calculation module 150, and is used to sort the demonstration values ​​in descending order, and select several synthetic driving demonstrations with the largest values ​​according to the candidate demonstration quantity budget to generate a candidate demonstration allowable list. The full model retraining module 170 is data-connected to the Top-B budget selection module 160 and the data isolation module, and is used to merge the synthetic driving demonstration with the real training set according to the candidate demonstration allowable list, unfreeze the upstream parameters in the valuation stage, and retrain the complete strategy model according to the original training objective. The trajectory generation module 180 receives vehicle perception input, generates the vehicle's future trajectory using the retrained complete policy model, and provides it to the vehicle controller.

[0088] Reference Figure 6 The electronic device includes a processor, memory, a data interface, an accelerated computing unit, and a policy training interface. The memory stores the data list, basic parameters of the target policy, projection seeds, parameter arrangements, projection gradients, local training geometric matrices, demonstration values, and a list of allowed values. The data interface receives training and calibration data; the accelerated computing unit performs sample-by-sample gradient calculations, random projections, second-order moment accumulation, and linear equation solving; the policy training interface calls the native training program of the target driving policy to perform basic training and full retraining.

[0089] In a distributed implementation, different computing nodes process different partitions of the candidate demonstration pool and write the projected gradient rows into a storage structure with the candidate identifier as the key. Independent nodes accumulate the second moment of the geometric subset and the calibration direction, and after completing the linear solution, broadcast the preconditioning direction to each candidate partition. Each candidate partition only needs to perform the inner product of the projected gradient and the preconditioning direction to obtain the demonstration value, thereby avoiding the transmission of the original high-dimensional sample-by-sample gradient between nodes.

[0090] Implementation Boundaries and Fixed Conditions: (1) The candidate demonstration pool consists of synthetic driving demonstrations generated by driving pseudo-experts in the simulation environment; the independent calibration set consists of real scenes reserved in advance during training.

[0091] (2) The result weights are only centered and normalized using the standard deviation of the full calibration set; the variance of the scenario-level driving results of the independent calibration set must be greater than zero.

[0092] (3) The planning parameter subspace consists of trajectory generation head parameters that directly generate the vehicle trajectory; the candidate demonstration, calibration scenario and geometric subset use the same set of parameters, supervised planning loss, basic parameters and parameter arrangement.

[0093] (4) Random projection uses element variance as The fixed Gaussian projection is used, and the candidate demonstrations, calibration scenes, and geometric subsets share the same projection seed; the budget is only measured by the number of candidate demonstrations and the B-line synthetic driving demonstration with the highest value is selected.

[0094] (5) The final evaluation data will not be used in training, calibration, valuation or selection; the independent calibration results will only be used to construct the valuation direction and will not be used as training samples, candidate demonstrations or final evaluation data.

[0095] The above embodiments have provided explicit support for the data isolation, result weights, planning parameter subspace, fixed random projection, damped local training geometry, policy-aware calibration direction, demonstration value of each candidate demonstration, budget selection, complete policy retraining, system modules, electronic devices, and storage media in the claims.

[0096] 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. A method for estimating and selecting demonstration data for strategy-aware autonomous driving, characterized in that, Includes the following steps: Construct a real training set, a candidate demonstration pool consisting of synthetic driving demonstrations generated by driving pseudo-experts, and an independent calibration set pre-reserved and fixed from the real training data. Use the real training set to train the target driving strategy and obtain basic parameters. The target driving strategy is used to perform driving trajectory prediction and scenario-level driving result calculation for each calibration scenario in the independent calibration set. The scenario-level driving results are then centered and normalized according to the standard deviation of the entire calibration set to obtain the result weights. In subsequent gradient calculations, the result weights are used as constants for stopping gradients. Within the planning parameter subspace of the target driving strategy, based on the same basic parameters and the same supervised planning loss, the sample-by-sample gradients of the candidate demonstration pool, the independent calibration set, and the geometric subset obtained from the real training set are extracted respectively. The sample-by-sample gradient is mapped to a low-dimensional projected gradient using the same fixed random projection. A damped local training geometric matrix is ​​constructed based on the projected gradient of the geometric subset. A policy-aware calibration direction is constructed based on the projected gradient of the independent calibration set and the result weights. The preconditioning direction determined by the local training geometric matrix and the policy-aware calibration direction is solved. The inner product of the projected gradient of each candidate demonstration and the preconditioning direction is calculated to obtain the demonstration value of each candidate demonstration relative to the target driving strategy. The demonstration values ​​are sorted in descending order, and several synthetic driving demonstrations with the largest values ​​are selected according to the budget of the number of candidate demonstrations. The selected synthetic driving demonstrations are merged with the real training set, and the complete strategy model is retrained according to the original training objective of the target driving strategy. The retrained complete policy model is used to process the vehicle perception input and generate the vehicle's future trajectory. The constraints of the planning parameter subspace only affect the demonstration value estimation and do not constitute a constraint on the retraining of the complete policy. The final evaluation set is not included in the data valuation and selection process.

2. The method according to claim 1, characterized in that, The candidate demonstration pool consists only of synthetic driving demonstrations generated by driving pseudo-experts in the simulation environment. The independent calibration set consists only of real driving scenarios reserved in advance in the partitioning of real training data and is not used for basic policy training. The final evaluation set does not participate in basic policy training, result weight generation, sample-by-sample gradient extraction, demonstration value calculation and candidate demonstration selection, so as to prevent the final evaluation information from entering the candidate demonstration selection process.

3. The method according to claim 1, characterized in that, The scenario-level driving results use collision-free index (NC), drivable area compliance index (DAC), driving direction compliance index (DDC), and traffic signal compliance index (TLC) as multiplicative constraints, and route progress (EP), collision time (TTC), lane keeping (LK), heading comfort (HC), and extended comfort (EC) as weighted quality terms. Calculate scenario-level driving results: in, For the first Scene-level driving results for each scenario , , , , , , , and The first The values ​​of NC, DAC, DDC, TLC, EP, TTC, LK, HC, and EC for each scenario; , , , and represents the weight of each quality item, which is a non-negative value and its sum is 1; The weights of the results are calculated as follows: in and These represent the mean and non-zero standard deviation of all calibration scenario-level driving results, respectively. It is a positive stability constant; The variance of the scenario-level driving results in the independent calibration set is greater than zero, and the sum of the weights of the results of all calibration scenarios is zero. This causes the gradient components common to all calibration scenarios to cancel each other out in the weighted aggregation, while the differentiated planning gradients related to the higher and lower scenario-level driving results are retained.

4. The method according to claim 1, characterized in that, The supervised planning loss is the native loss used by the target driving strategy to supervise the generation of the vehicle trajectory. The candidate demonstration, calibration scenario and geometric subset all use the same supervised planning loss, parameter arrangement, basic parameters and fixed random projection to ensure the consistency of demonstration value estimation. The planning parameter subspace includes trajectory generation head parameters that directly generate the vehicle trajectory, and excludes parameters corresponding to target detection loss and bird's-eye view semantic loss that do not directly supervise the vehicle trajectory during the sample-by-sample gradient extraction process, so as to avoid auxiliary tasks unrelated to trajectory planning from changing the meaning of demonstration value; the planning parameter subspace is only used for demonstration value estimation, and the complete policy model retains all native trainable parameters and training objectives during retraining.

5. The method according to claim 4, characterized in that, When the target driving strategy is a diffusion-based trajectory planning strategy, the supervised planning loss includes modality classification loss and trajectory regression loss obtained by iteratively denoising the noisy trajectory anchor points, and averages the diffusion time steps and multiple random implementations of the noise variables.

6. The method according to claim 1, characterized in that, The dimension of the planning parameter subspace is P, the projection dimension p is less than P, and the elements of the fixed random projection... Independently obeying orders with a mean of zero and a variance of 1 The Gaussian distribution; the candidate demonstration pool, the independent calibration set and the geometric subset share the projection seed, and the projection coefficients are regenerated sequentially according to the preset parameter block to avoid materializing the complete projection matrix in the memory.

7. The method according to claim 1, characterized in that, The local training geometry matrix is ​​composed of the normalized second moment of the projected gradient matrix of the geometric subset and the sum of the positive damping coefficient multiplied by the identity matrix, as follows: Among them, matrix It is the damped projected gradient second moment. Let be the projection gradient matrix, and let the projection dimension be . , The number of samples in the geometric subset. It is a positive damping coefficient. The identity matrix has the same dimensions as the projection dimension; The positive damping coefficient makes the local training geometry matrix positive definite and limits the numerical amplification of low-support projection directions; the policy-aware calibration direction is obtained by transposing the projection gradient matrix of the independent calibration set and then combining it with the resulting weight vector. Multiplying them together yields: ； in, To calibrate the projection gradient matrix; Solve linear equations The precondition direction is obtained, where Combine the downstream results of the target driving strategy in the calibration scenario with differentiable supervised programming gradients; The policy-aware direction is obtained after local training and geometric preconditioning. The demonstrative value of all candidate demonstrations is obtained simultaneously by performing a first-order matrix-vector multiplication of the candidate projection gradient matrix with the preconditioning direction. ,in Indicates the first The local training value of each candidate demonstration relative to the current target driving strategy and the current independent calibration target. Indicates the first The original sample-by-sample gradient of the candidate demonstration 3D projection gradient, This indicates transpose.

8. The method according to claim 1, characterized in that, When using the same candidate demonstration pool for multiple driving strategies, the demonstration value of each driving strategy is calculated using its own scenario-level driving results, supervised planning loss, planning parameter subspace, fixed random projection, and local training geometric matrix, and a candidate demonstration allowable list is generated for each driving strategy. The multiple driving strategies include at least two of the following: diffusion trajectory planning strategy, direct regression trajectory planning strategy, and flow matching trajectory planning strategy.

9. A strategy-aware autonomous driving demonstration data estimation and selection system, characterized in that, It includes a data isolation module, a basic strategy training module, a calibration result acquisition module, a planning gradient extraction module, a projection geometry construction module, a demonstration value calculation module, a Top-B budget selection module, a full model retraining module, and a trajectory generation module; Among them, the data isolation module establishes non-repeating data identifiers and persistent lists for the real training set, candidate demonstration pool, independent calibration set, final evaluation set and geometric subset, and provides them to subsequent modules; The basic strategy training module is connected to the data isolation module and is used to train the target driving strategy using the real training set to obtain basic parameters; and to provide basic checkpoints to the calibration result acquisition module and the planning gradient extraction module. The calibration result acquisition module is connected to the basic strategy training module and is used to perform scenario-level driving result calculation on each calibration scenario in the independent calibration set using the target driving strategy. The scenario-level driving results are then centered and normalized according to the standard deviation of the entire calibration set to generate the result weights of the stopping gradient. The planning gradient extraction module is used to extract the sample-by-sample gradients of the candidate demonstration pool, the independent calibration set, and the geometric subset based on the same basic parameters, the same supervised planning loss, and the same planning parameter subspace; the sample-by-sample gradients are streamed to the projection geometry construction module. The projection geometry construction module is data-connected to the planning gradient extraction module. It is used to map the sample-by-sample gradient to a low-dimensional projection gradient using the same fixed random projection, to obtain the candidate projection gradient, calibration projection gradient matrix and geometric subset projection gradient matrix. The geometric subset projection gradient matrix is ​​used to construct a damped local training geometry matrix. The calibration projection gradient matrix is ​​multiplied by the result weight vector to construct the policy-aware calibration direction. The demonstration value calculation module is data-connected to the projection geometry construction module and is used to solve the linear equation determined by the local training geometry matrix and the policy-aware calibration direction to obtain the precondition direction and calculate the demonstration value of each candidate demonstration. The Top-B budget selection module is connected to the demonstration value calculation module. It is used to sort the demonstration values ​​in descending order and select the synthetic driving demonstrations with the largest values ​​according to the budget of the number of candidate demonstrations to generate a list of allowed candidate demonstrations. The full model retraining module is data-connected to the Top-B budget selection module and the data isolation module. It is used to merge the synthetic driving demonstration with the real training set according to the candidate demonstration allowable list, unfreeze the upstream parameters in the valuation stage, and retrain the complete strategy model according to the original training objective. The trajectory generation module receives vehicle perception input, uses the retrained complete policy model to generate the vehicle's future trajectory, and provides it to the vehicle controller.

Images