End-to-end automatic driving model training method and device, computer equipment and medium

Abstract

The embodiment of the invention provides an end-to-end automatic driving model training method and device, computer equipment and a medium, and relates to the technical field of automatic driving, and the method comprises the steps: obtaining the driving behavior data of a driver and the first self-vehicle planning data of an end-to-end model, and when the difference value of the two kinds of data exceeds a first preset threshold value, determining the driving behavior of the driver; triggering a data recording mechanism, obtaining preliminary screening data of a period of time before and after the current moment, obtaining second self-vehicle planning data based on the preliminary screening data and an end-to-end model, and carrying out comparative learning on the second self-vehicle planning data and the driving behavior data to obtain a similarity score; screening corresponding data fragments when the similarity score is lower than a threshold value in the primarily screened data, obtaining a virtual driving scene according to the data fragments, and inputting the virtual driving scene into the VLA model to obtain driving related data; and training an end-to-end model by taking the driving-related data as training data. According to the scheme, the quality of training data and the efficiency and precision of model training are improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an end-to-end autonomous driving model training method, device, computer equipment and medium. Background Art

[0002] The field of autonomous driving focuses on using modular / non-modular deep learning models to achieve end-to-end intelligent driving tasks from perception to prediction to decision-making. Deep learning models are often data-driven, that is, they rely on a large amount of data to train the model, so that the end-to-end algorithm based on the deep learning model can have high accuracy and high robustness in different scenarios. In other words, the quality and quantity of data determine the upper limit of the deep learning model in the field of autonomous driving, and then determine the overall performance of the intelligent driving function. Large-scale, high-quality data often means a huge annotation cost, especially for end-to-end algorithms. Since the end-to-end model, whether it is a modular architecture design or a purely implicit architecture design, means that the overall architecture is strongly dependent on the deep learning model. At the same time, since this architecture needs to directly generate the planned path of the vehicle and even the vehicle control signal of the chassis from the original sensor data, the task is relatively complex, and the amount of data required is naturally more difficult than the perception task of generating environmental perception information from the original sensor data. The data demand is naturally greater. Only end-to-end algorithms that have been fully verified in as many and as rare scenarios as possible can shorten the process of deploying end-to-end algorithms on the vehicle, resulting in low model training efficiency and low accuracy. Summary of the invention

[0003] In view of this, an embodiment of the present invention provides an end-to-end autonomous driving model training method to solve the technical problems of low efficiency and low precision in end-to-end autonomous driving model training in the prior art. The method includes:

[0004] Acquire the driving behavior data of the driver during the driving of the ego vehicle and the first ego vehicle planning data output by the end-to-end autonomous driving model according to the sensor data;

[0005] Determine whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, trigger a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, wherein the preliminary screening data includes the sensor data, the driving behavior data, and the self-vehicle planning data;

[0006] obtaining second self-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and comparing and learning the second self-vehicle planning data with the driving behavior data in the preliminary screening data to obtain a similarity score;

[0007] Screening out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtaining a virtual driving scene according to the data segments;

[0008] Inputting the virtual driving scenario into the VLA model to obtain driving related data;

[0009] The virtual driving scene and the driving-related data are used as a first training data set, and the end-to-end autonomous driving model is trained using the first training data set.

[0010] The embodiment of the present invention also provides an end-to-end autonomous driving model training device to solve the technical problems of low efficiency and low precision of end-to-end autonomous driving model training in the prior art. The device includes:

[0011] A first acquisition module is used to acquire the driving behavior data of the driver during the driving of the self-vehicle and the first self-vehicle planning data output by the end-to-end automatic driving model according to the sensor data;

[0012] A second acquisition module is used to determine whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, trigger a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, wherein the preliminary screening data includes the sensor data, the driving behavior data, and the self-vehicle planning data;

[0013] A comparative learning module, configured to obtain second ego-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and perform comparative learning on the second ego-vehicle planning data and the driving behavior data in the preliminary screening data to obtain a similarity score;

[0014] A virtual driving scene acquisition module, used to screen out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtain a virtual driving scene according to the data segments;

[0015] A driving-related data acquisition module, used for inputting the virtual driving scene into the VLA model to obtain driving-related data;

[0016] A training module is used to use the virtual driving scene and the driving-related data as a first training data set, and to train the end-to-end autonomous driving model using the first training data set.

[0017] An embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned end-to-end autonomous driving model training methods to solve the technical problems of low efficiency and low accuracy of end-to-end autonomous driving model training in the prior art.

[0018] An embodiment of the present invention also provides a computer-readable storage medium, which stores a computer program for executing any of the above-mentioned end-to-end autonomous driving model training methods to solve the technical problems of low efficiency and low accuracy in end-to-end autonomous driving model training in the prior art.

[0019] Compared with the prior art, the beneficial effects that can be achieved by at least one of the above technical solutions adopted in the embodiments of this specification include at least the following: the present application strategically selects data from the data of the real vehicle to ensure the quality of the real vehicle data collection, and at the same time, the high-quality real vehicle data will be scored using the similarity of comparative learning, and the data is filtered according to the similarity score of comparative learning, which further improves the data quality. The virtual driving scene is constructed using the data filtered by the similarity score and the self-vehicle driving data in the virtual driving scene is generated in combination with the VLA large model, and the virtual driving scene and the relevant data output by the VLA large model are used as training data sets to train the end-to-end autonomous driving model, and the trained model is used to continue to plan the self-vehicle, thereby forming a data closed-loop process of the end-to-end algorithm. Therefore, the present application combines shadow mode, comparative learning, scene generation, large model and simulation tools, etc., to design a set of data closed-loop processes, which can effectively reduce the cost of data collection, improve data quality and the efficiency and accuracy of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 is a flow chart of an end-to-end autonomous driving model training method provided by an embodiment of the present invention;

[0022] Figure 2 is another flow chart of the end-to-end autonomous driving model training method provided by an embodiment of the present invention

[0023] Figure 3 is a structural block diagram of a computer device provided by an embodiment of the present invention;

[0024] Figure 4It is a structural block diagram of an end-to-end autonomous driving model training device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0026] The following describes the implementation methods of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work belong to the scope of protection of the present application.

[0027] In an embodiment of the present invention, an end-to-end autonomous driving model training method is provided, such as Figure 1 As shown, the method includes:

[0028] Step S101, obtaining driving behavior data of the driver during the driving of the self-vehicle and first self-vehicle planning data output by the end-to-end autonomous driving model according to sensor data;

[0029] Step S102, determining whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, triggering a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, the preliminary screening data including the sensor data, the driving behavior data and the self-vehicle planning data;

[0030] Step S103: obtaining second self-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and comparing and learning the second self-vehicle planning data with the driving behavior data in the preliminary screening data to obtain a similarity score;

[0031] Step S104, screening out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtaining a virtual driving scene according to the data segments;

[0032] Step S105: inputting the virtual driving scene into the VLA model to obtain driving related data;

[0033] Step S106: Use the virtual driving scene and the driving-related data as a first training data set, and use the first training data set to train the end-to-end autonomous driving model.

[0034] In this embodiment, when the driver's driving behavior data is different from the self-vehicle planning data output by the end-to-end model, the data recording mechanism is triggered, so that the data can be strategically selected from the real vehicle data to ensure the quality of the real vehicle data collection. At the same time, the high-quality real vehicle data will be scored using the similarity of comparative learning, and the data is filtered according to the similarity score of comparative learning, which further improves the data quality. The virtual driving scene is constructed using the data filtered by the similarity score and the self-vehicle driving data in the virtual driving scene is generated in combination with the VLA large model. The virtual driving scene and the relevant data output by the VLA large model are used as training data sets to train the end-to-end automatic driving model, and the trained model is used to continue to plan the self-vehicle, thereby forming a data closed-loop process of the end-to-end algorithm. Therefore, this application combines shadow mode, comparative learning, scene generation, large model and simulation tools, etc., to design a set of data closed-loop processes, which can effectively reduce the cost of data collection and improve the data quality and the efficiency and accuracy of model training.

[0035] In the specific implementation, first, during manual driving, the original sensor data (generally covering eye movement signals, camera data, lidar data, combined inertial navigation data and body chassis data, etc.) will be input into the end-to-end algorithm for real-time calculation. Once the difference between the self-vehicle planning result output by the end-to-end algorithm and the human driving behavior exceeds the set threshold, the data recording mechanism will be triggered to record the original sensor data of the ambiguous fragment within a period of time before and after the ambiguous time point, the human driving behavior and the calculation results of the end-to-end algorithm, and the recorded data will be used for subsequent data screening.

[0036] In one embodiment, the determining whether the difference between the driving behavior data and the vehicle planning data exceeds a first preset threshold value specifically includes the following steps:

[0037] An open-loop test evaluation index is used to determine whether a difference between the driving behavior data and the vehicle planning data exceeds a first preset threshold.

[0038] In practice, open-loop testing can be used to systematically evaluate the difference between the self-driving planning results output by the end-to-end algorithm and human driving behavior. Generally, the following steps can be followed:

[0039] (1) Define evaluation indicators. Define a set of reasonable evaluation indicators that can quantify the difference between the ego-vehicle planning results and human driving behavior. Common evaluation indicators include but are not limited to:

[0040] Trajectory deviation: Compare the distance difference between the trajectory generated by the algorithm and the actual trajectory of the human driver;

[0041] Speed ​​control: Compare the vehicle speed under algorithmic control with the speed selection of the human driver;

[0042] Acceleration / deceleration: Evaluate the consistency between the vehicle acceleration and deceleration and the human driver's behavior;

[0043] Decision latency: measuring the difference between the algorithm’s response time in making decisions and the reaction time of a human driver;

[0044] Collision risk: assessing the proximity of the algorithm-controlled vehicle to obstacles or other vehicles, and comparing it to the behavior of a human driver;

[0045] Comfort: measures the riding experience through the rate of change of acceleration;

[0046] (2) Setting a threshold. For each evaluation indicator, a reasonable threshold needs to be set. This threshold should reflect the maximum acceptable difference. If the difference exceeds this value, it is considered that the performance of the algorithm is significantly different from that of a human driver. The setting of the threshold can be determined based on the safety requirements of the application scenario, industry standards or empirical data, and is not particularly limited in this embodiment;

[0047] (3) Open-loop test design. In an open-loop test, once the initial conditions are given, the algorithm will run independently and will not be affected by subsequent inputs. Designing such a test scenario can ensure that the algorithm performs consistently under the same conditions;

[0048] (4) Collect test data. In the same test scenario, record the self-driving car planning results output by the algorithm and the actual behavior data of the human driver. Such data should include but are not limited to data points related to all the evaluation indicators mentioned above;

[0049] (5) Comparison and analysis. Compare the results of the algorithm output with the behavior of human drivers, and use statistical methods (such as mean comparison, variance analysis, etc.) to determine whether the difference between the two sets of data is significant and whether it exceeds the preset threshold. If the difference exceeds the preset threshold, it means that the behavior of the algorithm is significantly different from that of human drivers.

[0050] When it is detected through the above method that there is a large difference (ambiguity) between the driver's driving behavior data and the self-vehicle planning data output by the end-to-end autonomous driving model, the data recording mechanism can be triggered to record the original sensor data of the ambiguous fragment within a period of time before and after the time point when the ambiguity occurs, the human driving behavior and the calculation results of the end-to-end algorithm for use as screening data.

[0051] In one embodiment, obtaining second self-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and comparing and learning the second self-vehicle planning data with the driving behavior data in the preliminary screening data include:

[0052] Generate positive sample data and negative sample data based on the preliminary screening data, wherein the positive sample data includes data of changing the direction of the vehicle according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that does not interact with the vehicle, and the negative sample data includes data of changing the traffic light according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that interacts with the vehicle;

[0053] Inputting the positive sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the positive sample data;

[0054] Inputting the negative sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the negative sample data;

[0055] The second self-vehicle planning data based on the positive sample data and the second self-vehicle planning data based on the negative sample data are respectively compared with the driving behavior data in the preliminary screening data for learning.

[0056] In this embodiment, for the recorded ambiguous fragment data (preliminary screening data), positive samples and negative samples are generated respectively through different strategies, wherein the positive sample refers to slightly changing the state of the direction of the self-vehicle, etc., for the original scene, adding or deleting an intelligent agent or obstacle that will not interact with the self-vehicle, etc.; the negative sample refers to adding or deleting an intelligent agent or obstacle that will interact with the self-vehicle, etc., for the original scene, changing the indicator signal of the traffic light, etc. Then, the end-to-end algorithm is scored using the similarity definition of contrastive learning, that is, the end-to-end algorithm is operated on the positive and negative sample data respectively, and the output self-vehicle trajectory planning result is compared with the recorded driving behavior data of the manual driver. The scoring standard is that the second self-vehicle planning data on the positive sample should be as similar as possible to the driving behavior data, and the second self-vehicle planning data on the negative sample should be as different as possible from the driving behavior data, and the data with poor similarity score (data below the second preset threshold) is output. Poor similarity means that the end-to-end algorithm does not follow the expected driving behavior of human drivers on positive samples or does not follow the expected driving behavior of human drivers on negative samples. Such data is valuable for training end-to-end autonomous driving models. Therefore, we use the similarity of contrastive learning to score high-quality real-car data, and further filter data based on the similarity score of contrastive learning to improve data quality.

[0057] In one embodiment, obtaining a virtual driving scene according to the data segment includes:

[0058] Inputting the data fragment into a diffusion model to expand the scene and obtain an expanded scene;

[0059] The extended scene is rendered by a simulator to obtain the virtual driving scene.

[0060] During the specific implementation, the problem of reduced data volume introduced by multiple data screenings is solved. In order to solve this problem, a diffusion model is used to expand the scene by analogy with the high-quality data after multiple screenings. Then, the simulator is used to simulate the expanded scene into a virtual driving scene through the preset intelligent agent and map scene library and the intelligent agent's behavior strategy. In this scene, the Vision-Language-Action (VLA) large model is used to generate the scene planning trajectory and the description words of the current scene and action logic.

[0061] The diffusion model is a probability-based method that can be used to generate new data. It generates new similar data samples by adding noise to existing data (data fragments) and then gradually removing the noise to restore the data. In this process, the diffusion model will generate more simulated driving scenarios based on the existing high-quality data. These scenarios retain the key features of the original data while increasing diversity. These expanded scenarios are input into a simulator, which contains a preset map library and the behavior strategies of the agents. These agents can be simulated other vehicles, pedestrians, or other traffic participants. The simulator will simulate the real traffic environment based on the behavior strategies of the agents. At this stage, each agent will act according to the predetermined rules to form a dynamic virtual driving environment.

[0062] Since the previously collected driving behavior data of manual driving in the expanded scene is no longer applicable to the diffusion scene, the diffused scene data (video data, etc.) is input into the VLA large model to obtain the driving-related data of the ego vehicle in the simulation scene.

[0063] In one embodiment, the step of inputting the virtual driving scene data into a VLA model to obtain driving related data includes:

[0064] Inputting the virtual driving scene data into the VLA model, obtaining the simulation scene generated by the VLA model, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words;

[0065] The simulation scene, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words are used as the driving related data.

[0066] In the specific implementation, the Vision-Language-Action (VLA) large model is used in this virtual driving scene to plan the scene, generate trajectories, and generate descriptive words for the current scene and action logic. The simulator is used to allow the self-vehicle to interact with other intelligent agents with behavioral strategies in the virtual driving scene to execute the vehicle control instructions generated by the VLA large model, thereby generating the self-vehicle driving data in the virtual driving scene.

[0067] Specifically, the VLA model refers to the Vision-Language-Action (VLA) model, which is an artificial intelligence model designed to process multimodal information and can process three types of data: vision, language, and action. Specifically, in a simulation environment (virtual driving scene), the VLA model is used to generate planning trajectories and descriptive words for the current scene and action logic. Trajectory planning includes: the VLA model analyzes the current driving environment (visual input), understands the target (language input), and then generates a series of actions (outputs), which determine how the vehicle moves in the environment; Descriptive word generation: the VLA model also generates text descriptions of the current scene and action logic, which can be used to verify whether the vehicle's behavior is as expected. Then, execution and data generation are performed. The self-driving car will execute the control instructions generated by the VLA model in the simulation environment. In this process, the self-driving car interacts with other intelligent agents in the environment and performs corresponding actions. The entire interaction process will be recorded to generate self-driving car driving data in the virtual driving scene. These data can be used to further train and optimize the end-to-end model, so that the self-driving car can make more accurate decisions when facing different driving scenarios.

[0068] Furthermore, the method further comprises:

[0069] Manually reviewing the driving-related data;

[0070] When the manual review fails, the driver is manually driven in the virtual driving scene by using a virtual interactive device to obtain the driver's eye movement signal data;

[0071] The virtual driving scene, the self-vehicle driving data and the eye movement signal data are used as a second training data set, and the end-to-end autonomous driving model is trained using the second training data set.

[0072] When implementing it, refer to Figure 2 ,All the self-driving data are manually reviewed. If the evaluation indicators such as driving norms, safety and comfort are judged to be passed manually, the virtual ,driving scene, the self-driving data and the description words of the current scene and ,action logic generated by the VLA large model are used as the first ,training data set ( Figure 2 The qualified data in the self-driving car model 1) is input into the end-to-end autonomous driving model for training; if the evaluation indicators such as manual judgment of driving specifications, safety and comfort are not up to standard, an experienced driver is allowed to control the car in a virtual driving scene through a virtual interactive device (for example, a VR device), and his eye movement signals are captured during the process of controlling the car, so as to collect his focus at different times in the virtual driving scene, and then the virtual driving scene, the self-driving car driving data and the eye movement signals of the experienced driver are used as the second training data set ( Figure 2 The qualified data in 2) is input into the end-to-end autonomous driving model for training.

[0073] In this embodiment, the efficiency of annotation on the virtual driving scene is improved through diffusion model + simulator rendering + VLA large model automatic annotation, and the annotation quality of the virtual driving scene is ensured by manual review and selection. For data that fails the manual review, the corner case is retained by the human driver driving in the simulation scene. At the same time, multimodal data (including the acceleration and deceleration of the vehicle, steering angle and other chassis information, that is, the human driver's eye movement information) is used to train the end-to-end autonomous driving model. This method improves the efficiency of screening training data, while also improving data quality and further improving the training accuracy of the model.

[0074] In one embodiment, the manual review indicators include driving compliance, safety and comfort.

[0075] This application aims to solve the problem that there are few data closed-loop strategies for end-to-end models that generate self-vehicle planning and even control instructions from raw sensor data in the prior art. It triggers data collection through shadow mode, uses comparative learning scoring, and screens multiple times to improve the quality of end-to-end algorithm data; then expands rare scene data through diffusion models. The above strategies can ensure both the quality and quantity of data sets.

[0076] At the same time, through the combination of VLA large model and manual review and labeling, the VLA large model is used to improve the labeling efficiency while manual review is used to avoid the data bias and hallucination problems that may exist in the VLA large model, thereby improving data quality and further improving model training accuracy.

[0077] In this embodiment, a computer device is provided, such as Figure 3 As shown, it includes a memory 301, a processor 302, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, any of the above-mentioned end-to-end autonomous driving model training methods is implemented.

[0078] Specifically, the computer device may be a computer terminal, a server or a similar computing device.

[0079] In this embodiment, a computer-readable storage medium is provided, which stores a computer program for executing any of the above-mentioned end-to-end autonomous driving model training methods.

[0080] Specifically, computer-readable storage media include permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable storage media does not include transitory media such as modulated data signals and carrier waves.

[0081] Based on the same inventive concept, an end-to-end autonomous driving model training device is also provided in the embodiments of the present invention, as described in the following embodiments. Since the principle of solving the problem by the end-to-end autonomous driving model training device is similar to that of the end-to-end autonomous driving model training method, the implementation of the end-to-end autonomous driving model training device can refer to the implementation of the end-to-end autonomous driving model training method, and the repeated parts will not be repeated. As used below, the term "unit" or "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0082] Figure 4 is a structural block diagram of an end-to-end autonomous driving model training device according to an embodiment of the present invention. Figure 4 As shown, it includes: a first acquisition module 401, a second acquisition module 402, a comparative learning module 403, a virtual driving scene acquisition module 404, a driving related data acquisition module 405 and a training module 406. The structure is described below.

[0083] A first acquisition module 401 is used to acquire the driving behavior data of the driver during the driving of the ego vehicle and the first ego vehicle planning data output by the end-to-end autonomous driving model according to the sensor data;

[0084] A second acquisition module 402 is used to determine whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, trigger a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, wherein the preliminary screening data includes the sensor data, the driving behavior data, and the self-vehicle planning data;

[0085] A comparative learning module 403 is used to obtain second ego-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and compare and learn the second ego-vehicle planning data with the driving behavior data in the preliminary screening data to obtain a similarity score;

[0086] A virtual driving scene acquisition module 404 is used to screen out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtain a virtual driving scene according to the data segments;

[0087] A driving related data acquisition module 405 is used to input the virtual driving scene into the VLA model to obtain driving related data;

[0088] The training module 406 is used to use the virtual driving scene and the driving-related data as a first training data set, and use the first training data set to train the end-to-end autonomous driving model.

[0089] In one embodiment, the comparative learning module 403 is further used to:

[0090] Generate positive sample data and negative sample data based on the preliminary screening data, wherein the positive sample data includes data of changing the direction of the vehicle according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that does not interact with the vehicle, and the negative sample data includes data of changing the traffic light according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that interacts with the vehicle;

[0091] Inputting the positive sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the positive sample data;

[0092] Inputting the negative sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the negative sample data;

[0093] The second self-vehicle planning data based on the positive sample data and the second self-vehicle planning data based on the negative sample data are respectively compared with the driving behavior data in the preliminary screening data for learning.

[0094] In one embodiment, the virtual driving scene acquisition module 404 is further used to:

[0095] Inputting the data fragment into a diffusion model to expand the scene and obtain an expanded scene;

[0096] The extended scene is rendered by a simulator to obtain the virtual driving scene.

[0097] In one embodiment, the driving related data acquisition module 405 is further used to:

[0098] Inputting the virtual driving scene data into the VLA model, obtaining the simulation scene generated by the VLA model, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words;

[0099] The simulation scene, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words are used as the driving related data.

[0100] In one embodiment, the apparatus further comprises:

[0101] A manual review module, used for manually reviewing the driving-related data;

[0102] An eye movement signal acquisition module, used to perform manual driving of the driver in the virtual driving scene using a virtual interactive device to obtain eye movement signal data of the driver when the manual review fails;

[0103] The second training module is used to use the virtual driving scene, the self-vehicle driving data and the eye movement signal data as a second training data set, and use the second training data set to train the end-to-end autonomous driving model.

[0104] In one embodiment, the manual review module is further used to:

[0105] The manual review indicators include driving compliance, safety and comfort.

[0106] In one embodiment, the second acquisition module 402 is further configured to:

[0107] An open-loop test evaluation index is used to determine whether a difference between the driving behavior data and the vehicle planning data exceeds a first preset threshold.

[0108] The embodiments of the present invention achieve the following technical effects: the present application strategically selects data from the data of the real vehicle to ensure the quality of the real vehicle data collection, and at the same time, the high-quality real vehicle data will be scored using the similarity of comparative learning, and the data is filtered according to the similarity score of comparative learning, thereby further improving the data quality. The virtual driving scene is constructed using the data filtered by the similarity score, and the self-vehicle driving data in the virtual driving scene is generated in combination with the VLA large model, and the virtual driving scene and the relevant data output by the VLA large model are used as training data sets to train the end-to-end autonomous driving model, and the trained model is used to continue to plan the self-vehicle, thereby forming a data closed-loop process of the end-to-end algorithm. Therefore, the present application combines shadow mode, comparative learning, scene generation, large models and simulation tools, etc., to design a set of data closed-loop processes, which can effectively reduce data collection costs, improve data quality and the efficiency and accuracy of model training.

[0109] Obviously, those skilled in the art should understand that the modules or steps of the above-mentioned embodiments of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An end-to-end autonomous driving model training method, characterized in that: The method comprises: Acquire the driving behavior data of the driver during the driving of the ego vehicle and the first ego vehicle planning data output by the end-to-end autonomous driving model according to the sensor data; Determine whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, trigger a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, wherein the preliminary screening data includes the sensor data, the driving behavior data, and the self-vehicle planning data; obtaining second self-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and comparing and learning the second self-vehicle planning data with the driving behavior data in the preliminary screening data to obtain a similarity score; Screening out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtaining a virtual driving scene according to the data segments; Inputting the virtual driving scenario into the VLA model to obtain driving related data; The virtual driving scene and the driving-related data are used as a first training data set, and the end-to-end autonomous driving model is trained using the first training data set.

2. The end-to-end autonomous driving model training method according to claim 1, characterized in that: The obtaining of second self-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and comparing and learning the second self-vehicle planning data with the driving behavior data in the preliminary screening data, includes: Generate positive sample data and negative sample data based on the preliminary screening data, wherein the positive sample data includes data of changing the direction of the vehicle according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that does not interact with the vehicle, and the negative sample data includes data of changing the traffic light according to the original driving scene and data of adding or deleting an intelligent agent or obstacle that interacts with the vehicle; Inputting the positive sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the positive sample data; Inputting the negative sample data into the end-to-end autonomous driving model to obtain second self-vehicle planning data based on the negative sample data; The second self-vehicle planning data based on the positive sample data and the second self-vehicle planning data based on the negative sample data are respectively compared with the driving behavior data in the preliminary screening data for learning.

3. The end-to-end autonomous driving model training method according to claim 1, characterized in that: The step of obtaining a virtual driving scene according to the data segment comprises: Inputting the data fragment into a diffusion model to expand the scene and obtain an expanded scene; The extended scene is rendered by a simulator to obtain the virtual driving scene.

4. The end-to-end autonomous driving model training method according to claim 1, characterized in that: The step of inputting the virtual driving scene data into the VLA model to obtain driving related data includes: Inputting the virtual driving scene data into the VLA model, obtaining the simulation scene generated by the VLA model, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words; The simulation scene, the self-vehicle driving data of the self-vehicle in the simulation scene, and the action logic description words are used as the driving related data.

5. The end-to-end autonomous driving model training method according to claim 4, characterized in that: The method further comprises: Manually reviewing the driving-related data; When the manual review fails, the driver is manually driven in the virtual driving scene by using a virtual interactive device to obtain the driver's eye movement signal data; The virtual driving scene, the self-vehicle driving data and the eye movement signal data are used as a second training data set, and the end-to-end autonomous driving model is trained using the second training data set.

6. The end-to-end autonomous driving model training method according to claim 5, characterized in that: The manual review indicators include driving compliance, safety and comfort.

7. The end-to-end autonomous driving model training method according to claim 1, characterized in that: The determining whether the difference between the driving behavior data and the vehicle planning data exceeds a first preset threshold value includes: An open-loop test evaluation index is used to determine whether a difference between the driving behavior data and the vehicle planning data exceeds a first preset threshold.

8. An end-to-end autonomous driving model training device, characterized in that: include: A first acquisition module is used to acquire the driving behavior data of the driver during the driving of the self-vehicle and the first self-vehicle planning data output by the end-to-end automatic driving model according to the sensor data; A second acquisition module is used to determine whether the difference between the driving behavior data and the first self-vehicle planning data exceeds a first preset threshold value, and when the difference exceeds the first preset threshold value, trigger a data recording mechanism to obtain preliminary screening data within a preset time period adjacent to the current moment, wherein the preliminary screening data includes the sensor data, the driving behavior data, and the self-vehicle planning data; A comparative learning module, configured to obtain second ego-vehicle planning data based on the preliminary screening data and the end-to-end autonomous driving model, and perform comparative learning on the second ego-vehicle planning data and the driving behavior data in the preliminary screening data to obtain a similarity score; A virtual driving scene acquisition module, used to screen out data segments corresponding to when the similarity score is lower than a second preset threshold value from the primary screening data, and obtain a virtual driving scene according to the data segments; A driving-related data acquisition module, used for inputting the virtual driving scene into the VLA model to obtain driving-related data; A training module is used to use the virtual driving scene and the driving-related data as a first training data set, and to train the end-to-end autonomous driving model using the first training data set.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the end-to-end autonomous driving model training method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the end-to-end autonomous driving model training method according to any one of claims 1 to 7.

Images