A closed-loop verification self-distillation training method and system for a visual language automatic driving model
By using a closed-loop validator to perform quality assessment and hierarchical processing on the trajectory planning task of the visual language autonomous driving model, generating and filtering self-teacher trajectories, and constructing supervised fine-tuning training data, the problems of unstable model decoding results and insufficient utilization efficiency of training samples are solved, thereby improving the stability and quality of trajectory planning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN AUTOMOTIVE RES INST BEIJING INST OF TECH (SHENZHEN RES INST OF NAT ENG LAB FOR ELECTRIC VEHICLES)
- Filing Date
- 2026-06-29
- Publication Date
- 2026-07-31
AI Technical Summary
Visual language autonomous driving models suffer from unstable single-decoding results in trajectory planning tasks, difficulty in effectively selecting high-quality self-generated trajectories, insufficient utilization of training samples, and a lack of systematic closed-loop validator integration methods.
The baseline trajectory results are evaluated for quality using a closed-loop validator. The sample data is stratified, candidate sampling budgets are allocated, multiple candidate trajectory results are generated, and the self-teacher trajectory is selected by scoring using the closed-loop validator. Supervised fine-tuning training data is then constructed for continued training.
It improves the effectiveness of training data and the stability of trajectory planning output, reduces the risk of low-quality samples entering the training set, enhances the probability of generating high-quality trajectory patterns, and improves the generation stability and pass rate of the model.
Smart Images
Figure CN122493414A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, and in particular to a closed-loop verification self-distillation training method and system for visual language autonomous driving models. Background Technology
[0002] With the application of visual language models in autonomous driving scenarios, these models can combine onboard visual information, driving status information, and task information to generate trajectory planning results. Compared to traditional planning methods that rely solely on single-episode supervised learning, visual language autonomous driving models typically have the ability to generate multiple candidate trajectory results in the same scenario.
[0003] However, in trajectory planning tasks, the model's single-run decoding results do not always consistently achieve high quality. For the same driving scenario, the model may generate multiple candidate trajectories under different sampling conditions, some of which are significantly better than the baseline decoding results. If training is still conducted using only the original labeled data or single-run inference results, it is difficult to fully utilize the model's inherent candidate trajectory generation capabilities, and it is also difficult to adaptively increase the exploration budget for low-quality samples.
[0004] Existing self-distillation or rejection sampling training methods typically rely on general reward models or task-level validation signals. However, in autonomous driving trajectory planning scenarios, a systematic solution for combining closed-loop validators to evaluate the quality of candidate trajectories and constructing self-teacher training data remains lacking for visual language autonomous driving models. Therefore, a training method is needed that integrates baseline evaluation, sample stratification, candidate sampling, closed-loop validation, self-teacher selection, and continued training to improve the effectiveness of training data and the stability of trajectory planning output. Summary of the Invention
[0005] This invention addresses the problems of unstable single-decoding results, difficulty in effectively selecting high-quality self-generated trajectories, and insufficient utilization efficiency of training samples in existing visual language autonomous driving models for trajectory planning tasks. It proposes a closed-loop verification self-distillation training method and system for visual language autonomous driving models.
[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0007] Step S1: Obtain sample data for trajectory planning training, and use the visual language autonomous driving model to be trained to perform baseline inference on the sample data to obtain the corresponding baseline trajectory result.
[0008] Step S2: Use a closed-loop validator to evaluate the quality of the baseline trajectory results, and divide the sample data into multiple difficulty levels based on the evaluation results;
[0009] Step S3: Allocate different candidate sampling budgets for sample data of different difficulty levels, and for samples whose baseline trajectory results do not meet the preset quality conditions, control the visual language autonomous driving model to generate multiple candidate trajectory results according to the corresponding candidate sampling budget.
[0010] Step S4: Use the closed-loop validator to score the candidate trajectory results, and determine the self-teacher trajectory from the multiple candidate trajectory results based on the score improvement of the candidate trajectory results relative to the baseline trajectory results and / or preset screening conditions;
[0011] Step S5: When there is no self-teacher trajectory that meets the conditions among the multiple candidate trajectory results, backtracking is performed on the corresponding sample to construct backtracked teacher samples;
[0012] Step S6: Construct supervised fine-tuning training data based on the self-teacher trajectory and the backtracking teacher samples, and use the supervised fine-tuning training data to continue training the visual language autonomous driving model in order to improve the trajectory planning capability of the visual language autonomous driving model.
[0013] In step S2, the sample data is divided into at least two quality levels based on the quality assessment results of the baseline trajectory results, and different candidate sampling budgets are configured for different quality levels. Multiple difficulty levels can be determined based on the trajectory quality scores output by the closed-loop validator. For samples whose quality scores meet the preset quality conditions, candidate sampling can be reduced or eliminated; for samples whose quality scores are below the preset quality conditions but still have some valid planning results, a medium candidate sampling budget can be allocated; for samples with low quality scores or invalid baseline trajectory results, a high candidate sampling budget can be allocated.
[0014] In step S4, the closed-loop validator outputs the trajectory quality score corresponding to the candidate trajectory result, and uses the score improvement of the candidate trajectory result relative to the baseline trajectory result, the absolute score value of the candidate trajectory result, or a combination of the two as the selection criteria for the self-teacher trajectory.
[0015] In step S4, the selection of the self-teacher trajectory can be based on at least one of the following: the improvement in score of the candidate trajectory result relative to the baseline trajectory result, the absolute score value of the candidate trajectory result, and a preset selection threshold. When multiple candidate trajectory results meet the selection criteria, the candidate trajectory result with the highest score can be selected as the self-teacher trajectory.
[0016] In step S5, the rollback process includes constructing rollback teacher samples using the real trajectories corresponding to the samples, and / or removing samples that have not obtained valid self-teacher trajector trajectories from the supervised fine-tuning training data.
[0017] In step S6, the supervised fine-tuning training data includes sample inputs, self-teacher trajectories selected from the multi-candidate trajectory results, and backtracked teacher samples retained in the backtracking process, while maintaining the pairing relationship between the sample inputs and the corresponding trajectory supervision targets. Samples for which no effective self-teacher trajectories are obtained and which are unsuitable for constructing backtracked teacher samples may not be included in the supervised fine-tuning training data.
[0018] A system for closed-loop verification self-distillation training of a vision-language autonomous driving model is provided, for implementing the aforementioned closed-loop verification self-distillation training of the vision-language autonomous driving model; the system includes:
[0019] The evaluation module is used to evaluate the quality of baseline trajectory results and candidate trajectory results, and output the corresponding trajectory quality score;
[0020] The sampling module is used to allocate candidate sampling budgets according to the difficulty level of the samples and control the visual language autonomous driving model to generate multiple candidate trajectory results.
[0021] The verification and filtering module is used to determine the self-taught trajectory from multiple candidate trajectory results based on trajectory quality scores;
[0022] The data construction module is used to construct supervised fine-tuning training data based on self-teacher trajectories and backtracked teacher samples;
[0023] The training module is used to continue training the visual language autonomous driving model using the supervised fine-tuning training data.
[0024] The present invention has the following advantages:
[0025] (1) The present invention uses a closed-loop validator to evaluate the quality of baseline trajectories and candidate trajectories, so that the selection of self-teacher samples has a clear task feedback basis and reduces the risk of low-quality self-generated samples entering the training set.
[0026] (2) The present invention stratifies the samples according to the quality of the baseline trajectory and allocates different candidate sampling budgets to samples of different levels, which can increase the probability of obtaining high-quality candidate trajectories for difficult samples while controlling the sampling cost.
[0027] (3) The present invention uses self-teacher trajectories and backtracked teacher samples together to construct supervised fine-tuning training data, so that the model can strengthen its own generated high-quality trajectory patterns and improve the stability and pass rate of trajectory planning generation.
[0028] (4) The present invention combines baseline inference, candidate sampling, closed-loop verification, data construction and continued training into a closed-loop training process, so that the training data generation process is consistent with the feedback of the trajectory planning task. Attached Figure Description
[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0030] Figure 1 This is a flowchart of the overall process of a closed-loop verification self-distillation training method for a visual language autonomous driving model according to the present invention.
[0031] Figure 2 This is a schematic diagram of the sample stratification, candidate sampling, and self-teacher screening process in this invention. Detailed Implementation
[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Based on the embodiments of the present invention, equivalent substitutions or modifications made by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0033] like Figure 1 As shown, this embodiment provides a closed-loop verification self-distillation training method for visual language autonomous driving models, including the following steps:
[0034] S1, Baseline Inference and Quality Assessment. Sample data for trajectory planning training is acquired and input into the visual language autonomous driving model to be trained, generating baseline trajectory results for the corresponding samples. Subsequently, a closed-loop validator is used to assess the quality of the baseline trajectory results, obtaining the corresponding trajectory quality score.
[0035] S2, Sample Stratification and Sampling Budget Allocation. Based on the trajectory quality score, the samples are divided into multiple difficulty levels. Different difficulty levels correspond to different candidate sampling budgets. Samples with higher quality may have their candidate sampling budget reduced or waived, while samples with lower quality are allocated higher candidate sampling budgets.
[0036] In one implementation, the multiple difficulty levels may include a high-quality level, an intermediate-quality level, and a low-quality level. The high-quality level corresponds to samples whose baseline trajectory results meet preset quality conditions; the intermediate-quality level corresponds to samples whose baseline trajectory results have partially valid planning results but do not meet preset quality conditions; and the low-quality level corresponds to samples whose baseline trajectory results are invalid or have low quality scores. The candidate sampling budgets corresponding to different levels can be determined by preset rules or dynamically adjusted based on the quality score output by the closed-loop validator.
[0037] S3, Multiple Candidate Trajectory Generation. For samples where the baseline trajectory results do not meet the preset quality conditions, the control vision-language autonomous driving model generates multiple candidate trajectory results according to the allocated candidate sampling budget to increase the probability of obtaining high-quality trajectories.
[0038] S4, Closed-loop verification scoring and self-teacher selection. The closed-loop verifier scores multiple candidate trajectory results separately, and the self-teacher trajectory is selected from the multiple candidate trajectory results based on the improvement of the candidate trajectory results relative to the baseline trajectory results and / or whether the preset selection conditions are met; when multiple candidate trajectory results meet the selection conditions, the candidate trajectory result with the highest score can be selected as the self-teacher trajectory.
[0039] S5, Failure Sample Backtracking. When no self-teacher trajectory that meets the conditions is found among multiple candidate trajectory results, backtracking is performed on the corresponding sample. The backtracking process may include constructing backtracked teacher samples using real trajectories, and / or removing samples that have not obtained valid self-teacher trajectories from the training data.
[0040] S6. Construct a training set and continue training. Based on the self-teacher trajectory and the backtracked teacher samples retained in the backtracking process, construct supervised fine-tuning training data, and use the supervised fine-tuning training data to continue training the visual language autonomous driving model, so as to improve the model's generation stability and output quality in trajectory planning tasks.
[0041] In one implementation, the supervised fine-tuning training data may include sample inputs, teacher source identifiers, and trajectory supervision targets. The teacher source identifiers are used to distinguish between self-taught trajectories and backtracked teacher samples; the trajectory supervision targets are paired with the sample inputs so that the visual language autonomous driving model can learn high-quality trajectory outputs for the corresponding scenario during continued training.
[0042] Furthermore, in one implementation, supervised fine-tuning of the training objective can be used to optimize the visual language autonomous driving model. Let the training sample set be D={(Input_k, Target_k)}, and the model parameters be θ, then the supervised fine-tuning training objective... It can be represented as: ;
[0043] Among them, Input k Target represents the input information of the k-th training sample. k P represents the corresponding trajectory supervision target. θ This represents the conditional probability distribution of the visual language autonomous driving model for the supervised target under parameter θ. By minimizing this supervised fine-tuning training objective, the model can learn high-quality trajectory output patterns corresponding to the self-teacher trajectory and the backtracked teacher samples.
[0044] like Figure 2 As shown, in this embodiment, the closed-loop validator is used to perform a unified quality assessment on the baseline trajectory results and candidate trajectory results. The closed-loop validator can output a score that characterizes the quality of trajectory planning, and determine whether the candidate trajectory results are superior to the baseline trajectory results based on the score results.
[0045] In one implementation, the closed-loop validator can output continuous scores corresponding to the candidate trajectory results. Let the no-fault collision index be NC, the drivable area compliance index be DAC, the vehicle's forward travel index be EP, the collision time index be TTC, the historical comfort index be HC, and the safety gating index be α, then the continuous score... It can be represented as: ;
[0046] In a preferred embodiment, α can be 0.5. The continuous score is used to reflect the overall quality of the candidate trajectory results in terms of both safety and performance.
[0047] In one implementation, the selection criteria for the self-taught trajectory include the score improvement of the candidate trajectory result relative to the baseline trajectory result, the absolute score value of the candidate trajectory result, or a combination of both. Let the score of the baseline trajectory result be Score_base, and the score of the j-th candidate trajectory result be Score_j, then the score improvement of the candidate trajectory result relative to the baseline trajectory result can be expressed as: ;
[0048] Furthermore, a candidate trajectory can be determined as a valid candidate trajectory when it meets the following conditions: ;
[0049] Where τ represents the quality threshold of the candidate trajectory result, and δ represents the minimum improvement threshold of the candidate trajectory result relative to the baseline trajectory result. Candidate trajectories whose scores do not meet the preset conditions are not included in the supervised fine-tuning training data.
[0050] In another implementation, when multiple candidate trajectory results all meet the filtering criteria, the candidate trajectory result with the highest score can be selected as the self-teacher trajectory. Let the candidate trajectory set be {Trajectory_1, Trajectory_2,..., Trajectory_n}, then the self-teacher trajectory can be represented as: ;
[0051] In another implementation, samples can be categorized based on the final processing result into samples not participating in enhancement, self-teacher samples, and backtracking samples. Samples not participating in enhancement correspond to those whose baseline trajectory results have reached a preset quality condition; self-teacher samples correspond to those whose high-quality trajectories were selected from the candidate trajectory results; and backtracking samples correspond to those for which no valid self-teacher trajectories were selected and therefore backtracking was performed. Backtracking samples can be further divided into samples that form backtracking teacher samples and samples that were removed. Figure 2 The skip, self-teacher sample, and rollback sample shown correspond to the different processing results mentioned above.
[0052] Corresponding to the above method, the present invention also provides a system for implementing closed-loop verification self-distillation training, the system comprising:
[0053] The evaluation module is used to evaluate the quality of baseline trajectory results and candidate trajectory results, and output the corresponding trajectory quality score;
[0054] The sampling module is used to allocate candidate sampling budgets according to the difficulty level of the samples and control the visual language autonomous driving model to generate multiple candidate trajectory results.
[0055] The verification and filtering module is used to determine the self-taught trajectory from multiple candidate trajectory results based on trajectory quality scores;
[0056] The data construction module is used to construct supervised fine-tuning training data based on self-teacher trajectories and backtracked teacher samples;
[0057] The training module is used to continue training the visual language autonomous driving model using the supervised fine-tuning training data.
[0058] The above modules can be deployed in the same computing device as software, or they can be implemented in a distributed manner by combining software and hardware. This invention does not limit this.
Claims
1. A closed-loop verification self-distillation training method for visual language-based autonomous driving models, characterized in that, Includes the following steps: Step S1: Obtain sample data for trajectory planning training, and use the visual language autonomous driving model to be trained to perform baseline inference on the sample data to obtain the corresponding baseline trajectory result. Step S2: Use a closed-loop validator to evaluate the quality of the baseline trajectory results, and divide the sample data into multiple difficulty levels based on the evaluation results; Step S3: Allocate different candidate sampling budgets for sample data of different difficulty levels, and for samples whose baseline trajectory results do not meet the preset quality conditions, control the visual language autonomous driving model to generate multiple candidate trajectory results according to the corresponding candidate sampling budget. Step S4: Use the closed-loop validator to score the candidate trajectory results, and determine the self-teacher trajectory from the multiple candidate trajectory results based on the score improvement of the candidate trajectory results relative to the baseline trajectory results and / or preset screening conditions; Step S5: When there is no self-teacher trajectory that meets the conditions among the multiple candidate trajectory results, backtracking is performed on the corresponding sample to construct backtracked teacher samples; Step S6: Construct supervised fine-tuning training data based on the self-teacher trajectory and the backtracking teacher samples, and use the supervised fine-tuning training data to continue training the visual language autonomous driving model in order to improve the trajectory planning capability of the visual language autonomous driving model.
2. The closed-loop verification self-distillation training method for a visual language-based autonomous driving model according to claim 1, characterized in that, In step S2, the sample data is divided into at least two quality levels based on the quality assessment results of the baseline trajectory results, and different candidate sampling budgets are configured for different quality levels.
3. The closed-loop verification self-distillation training method for a visual language-based autonomous driving model according to claim 1, characterized in that, In step S4, the closed-loop validator outputs the trajectory quality score corresponding to the candidate trajectory result, and uses the score improvement of the candidate trajectory result relative to the baseline trajectory result, the absolute score value of the candidate trajectory result, or a combination of the two as the selection criteria for the self-teacher trajectory.
4. The closed-loop verification self-distillation training method for a visual language-based autonomous driving model according to claim 1, characterized in that, In step S4, when multiple candidate trajectory results meet the preset screening conditions, the candidate trajectory result with the highest trajectory quality score is selected as the self-teacher trajectory.
5. The closed-loop verification self-distillation training method for a visual language-based autonomous driving model according to claim 1, characterized in that, In step S5, the rollback process includes constructing rollback teacher samples using the real trajectories corresponding to the samples, and / or removing samples that have not obtained valid self-teacher trajector trajectories from the supervised fine-tuning training data.
6. The closed-loop verification self-distillation training method for a visual language-based autonomous driving model according to claim 1, characterized in that, In step S6, the supervised fine-tuning training data includes sample input, self-teacher trajectories selected from multiple candidate trajectory results, and backtracked teacher samples retained in the backtracking process, while maintaining the pairing relationship between the sample input and the corresponding trajectory supervision target.
7. A system for closed-loop verification self-distillation training of a visual language-based autonomous driving model, characterized in that, A system for implementing closed-loop verification self-distillation training of a vision-language-oriented autonomous driving model as described in any one of claims 1 to 6; the system comprises: The evaluation module is used to evaluate the quality of baseline trajectory results and candidate trajectory results, and output the corresponding trajectory quality score; The sampling module is used to allocate candidate sampling budgets according to the difficulty level of the samples and control the visual language autonomous driving model to generate multiple candidate trajectory results. The verification and filtering module is used to determine the self-taught trajectory from multiple candidate trajectory results based on trajectory quality scores; The data construction module is used to construct supervised fine-tuning training data based on self-teacher trajectories and backtracked teacher samples; The training module is used to continue training the visual language autonomous driving model using the supervised fine-tuning training data.