Text-driven automatic driving scene automatic generation method

Abstract

The invention belongs to the technical field of automatic driving scene automatic generation methods, and particularly relates to a text-driven automatic driving scene automatic generation method, which comprises the following steps of: receiving an automatic driving scene natural language description input by a user, processing the natural language description, extracting scene key elements, and generating a text-driven automatic driving scene; the scene key elements comprise road types, traffic participants, environmental conditions and detailed information of each element, matching and calling a scene template based on a hierarchical scene library, synthesizing the extracted scene key elements and the scene template to generate a complete scene description, converting the complete scene description into a domain-specific language script, and displaying the domain-specific language script on the basis of the scene template. And integrating the DSL script to a simulation platform, running the generated test scene in the simulation platform, monitoring the behavior of the automatic driving system in the test scene in real time, and generating an evaluation report based on a predefined evaluation index.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving scene automatic generation method, and particularly relates to a text-driven automatic driving scene automatic generation method. BACKGROUND

[0002] Traditional scene generation relies on manual design by engineers, which not only consumes a large amount of manpower and time cost, but also is difficult to cover a large number of edge scenes in complex traffic environments, resulting in insufficient diversity and comprehensiveness of test scenes, and difficulty in fully verifying the performance of the automatic driving system in extreme or rare scenes.

[0003] Weak natural language processing capability and inaccurate scene element extraction: the existing automatic scene generation method has limited processing capability for user input natural language description, and is difficult to accurately analyze semantic information, resulting in incomplete or biased extraction of key scene elements, which affects the authenticity and pertinence of the generated scene.

[0004] Lack of hierarchical and dynamic updating mechanism in scene library: the existing scene library is mostly stored in a flat manner, without forming a hierarchical structure covering road topology, traffic participants, climate conditions and other dimensions, and lacking a dynamic behavior model trained based on real traffic data, which is difficult to simulate the complex interactive behavior of traffic participants; at the same time, the scene library update relies on manual intervention, and cannot be automatically optimized according to the test results, with poor adaptability.

[0005] Complex integration with simulation platform, insufficient executability: the scene description generated by the existing method is difficult to directly convert into a script executable by the simulation platform, and needs to be manually written or adjusted in a field-specific language script, resulting in a cumbersome process from scene design to execution, reducing the test efficiency; and lacking a unified adaptation layer and interface, it is difficult to be compatible with different simulation platforms.

[0006] Inadequate evaluation and feedback mechanism: during the scene test process, there is a lack of real-time monitoring and scientific evaluation system for the behavior of the automatic driving system, which cannot timely discover potential problems of the system; at the same time, the linkage updating mechanism of test results, scene library and evaluation index is missing, which makes it difficult for the test framework to self-optimize, affecting the continuous improvement of test effect, and therefore a text-driven automatic driving scene automatic generation method is designed to solve the above problems. SUMMARY

[0007] The purpose of the present application is to provide a text-driven automatic driving scene automatic generation method to solve the above technical problems.

[0008] Therefore, the present application provides a text-driven automatic driving scene automatic generation method, comprising the following steps:

[0009] Step 1: receiving the user input automatic driving scene natural language description;

[0010] Step two: processing the natural language description to extract scene key elements, including road type, traffic participants, environmental conditions, and detailed information of each element;

[0011] Step three: matching and calling scene templates based on the hierarchical scene library, synthesizing the extracted scene key elements with the scene templates to generate a complete scene description;

[0012] Step four: converting the complete scene description into a domain-specific language script;

[0013] Step five: integrating the DSL script into the simulation platform and running the generated test scene in the simulation platform;

[0014] Step six: real-time monitoring of the behavior of the autonomous driving system in the test scene and generating an evaluation report based on predefined evaluation indicators;

[0015] Step seven: storing the test scene, DSL script, and evaluation report, and updating the hierarchical scene library and evaluation indicators according to the evaluation report.

[0016] Preferably, the hierarchical scene library maintains and updates scene elements covering road topology, traffic participants, and climate conditions, and includes dynamic behavior models trained based on real traffic data, which are used to simulate the acceleration, deceleration, lane changing, and pedestrian random crossing behaviors of traffic participants;

[0017] The hierarchical scene library generates scene templates through template matching technology based on scene element similarity calculation, with the calculation formula being:

[0018]

[0019] where e i is a scene element, t j is a time and space constraint.

[0020] Preferably, processing the natural language description includes the following steps:

[0021] Tokenizing and part-of-speech tagging the natural language description;

[0022] Determining scene key elements and element relationships through dependency syntax analysis and semantic role labeling;

[0023] Refining scene key element information using large language models and multi-stage context few-shot learning;

[0024] The few-shot learning generates scene element extraction rules by providing scene generation examples containing natural language input and corresponding scene element output.

[0025] Preferably, the large language model is GPT-4, and the detailed information of the scene key elements includes specific sections of highways, brands and models of traffic participants, and rainfall intensity.

[0026] Preferably, the process of converting the complete scene description into a DSL script includes converting the complete scene description into an intermediate representation through lexical analysis and syntax analysis, converting the intermediate representation into DSL instructions based on predefined DSL syntax rules, and generating a structured DSL script.

[0027] The generation algorithm is:

[0028] DSLscript = generate(parse(scene description));

[0029] Wherein the parse function is used to parse the scene description, and the generate function is used to generate the DSL instructions.

[0030] Preferably, the process of integrating the DSL script into the simulation platform includes developing an adaptation layer to adapt to the instruction set and interface specification of the simulation platform, and the adaptation layer includes an error handling mechanism.

[0031] A standardized interface is provided for scene loading, script execution and result acquisition, and the simulation platform includes Carla.

[0032] Preferably, the process of real-time monitoring and generating evaluation reports includes collecting sensor data of the autonomous driving system, including vehicle position, speed and acceleration.

[0033] Based on safety, robustness and efficiency indicators, the safety indicators include the number of collisions and the number of emergency brakes, and the evaluation algorithm is:

[0034]

[0035] Wherein, is the weight, v i is the actual value, v max is the maximum value.

[0036] Preferably, the process of storage and update includes using a relational database to store structured data and a non-relational database to store unstructured data.

[0037] Based on the evaluation report, common problems of the system are identified, and the scene templates in the scene library and the weights of the evaluation indicators are updated.

[0038] Preferably, the example input of the multi-stage context few-shot learning is a scene of generating an automatic driving unprotected left turn, and the example output includes road scene, lane, signal light, traffic flow initial position and self-driving end point information.

[0039] The beneficial effects of the present application are:

[0040] Based on the scene element extraction technology of natural language processing and large language model, without manual design of scene details, the text description input by the user is quickly converted into structured scene elements, greatly reducing the labor cost and time cost of scene generation. The hierarchical scene library covers road topology, traffic participants, climate conditions and other multi-dimensional elements, and integrates dynamic behavior models. Combined with template matching and synthesis technology, it can generate a variety of scenes covering urban roads, highways and special weather, solving the problem of limited scene coverage of traditional methods.

[0041] Through dependency syntax analysis, semantic role labeling and multi-stage context few-shot learning, the key elements in natural language are accurately extracted and refined into specific parameters, ensuring that the generated scene is highly consistent with the real traffic environment. During the scene synthesis process, time, space and event consistency verification is performed to avoid logical contradictions and ensure the feasibility of the scene, providing real and effective input for the automatic driving system test.

[0042] Based on the domain-specific language script generation technology, the complete scene description is automatically converted into an executable script compatible with simulation platforms such as Carla, without the need for manual code writing or adjustment. This solves the problem of poor compatibility between scenes and platforms, improves the continuity of the test process, and ensures stable execution of the script on different platforms through the integration of the simulation platform module through the adaptation layer and standardized interface, improving the generality and scalability of the method.

[0043] The real-time monitoring and evaluation module quantitatively evaluates the behavior of the automatic driving system through multi-dimensional indicators, generates a report containing potential problems and improvement suggestions, and realizes instant feedback of the test process, facilitating quick positioning of system defects. The data storage and update module dynamically optimizes the scene library and evaluation indicators based on the test results, enabling the test framework to have self-improvement capabilities. As the number of tests increases, the scene coverage becomes more comprehensive and the evaluation becomes more accurate, continuously improving the test effect of the automatic driving system. Combined with the few-shot learning technology, only a small number of examples are needed to generate complex scenes, especially suitable for the generation of edge scenes, making up for the defects of traditional methods in covering edge scenes, further ensuring the safety and reliability of the automatic driving system. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a system flowchart;

[0045] Figure 2To prompt the tool chain system diagram;

[0046] Figure 3 To data storage and update module flow chart;

[0047] Figure 4 To the hierarchical structure diagram of the scene. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0049] It should be noted that all the terms for indicating directionality and positionality in the present application, such as “upper”, “lower”, “left”, “right”, “front”, “back”, “vertical”, “horizontal”, “inner”, “outer”, “top”, “low”, “lateral”, “longitudinal”, “center”, etc., are only used to explain the relative positional relationship, connection condition, etc. between components in a certain state (as shown in the drawings), and are only for the convenience of describing the present application, and thus cannot be understood as a limitation on the present application. In addition, the description of “first”, “second”, etc. in the present application is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features.

[0050] In the description of the present application, unless explicitly specified and limited, the terms “mounting”, “connection”, “connecting” should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integrally connected; can be mechanical connection; can be directly connected, or indirectly connected through an intermediate medium; can be the communication inside two elements. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0051] In the description of the present application, the description of the terms “one embodiment”, “some embodiments”, “illustrative embodiment”, “example”, “specific example”, or “some examples” means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0052] A hierarchical scenario library module that maintains and updates a hierarchical scenario library, covering road topology, traffic participants, weather conditions, and other elements.

[0053] It matches and generates detailed scenario templates based on the output of the scenario element extraction module, ensuring that the generated test scenarios are comprehensive and diverse. Other components can call materials from the scenario library according to the corresponding interface for scenario generation and construction.

[0054] A natural language processing and scenario element extraction module that is responsible for processing natural language descriptions provided by users, including word segmentation, part-of-speech tagging, and semantic analysis.

[0055] Through large language models (LLM) and multi-stage context few-shot learning, this module extracts and refines key elements in scenarios such as road types, traffic participants, and environmental conditions, ensuring the accuracy and comprehensiveness of each scenario element. The input is a natural language description (such as "a car changes lanes on a highway in the rain"), and the output is detailed scenario element information (such as road type: highway, weather: rain, traffic participant: car, detailed information: specific section of the highway, brand and model of the car, rainfall intensity in the rain).

[0056] A scenario synthesis module that is responsible for synthesizing templates in the scenario library with extracted scenario elements to generate complete scenario descriptions.

[0057] By organically combining various elements, a coherent and detailed test scenario is generated, ensuring wide test coverage and strong practicality. The input is the detailed scenario element information provided by the hierarchical scenario library and the natural language processing module, and the output is the complete scenario content (such as a car changing lanes on a certain highway section in the rain).

[0058] A DSL script generation module that automatically generates domain-specific language (DSL) scripts to convert scenario descriptions into executable test scripts.

[0059] By parsing the complete scenario description and using predefined DSL syntax rules, test scripts suitable for simulation platforms are generated. The input is the complete scenario content (such as a car changing lanes on a certain highway section in the rain), and the output is the DSL script (such as a script file for the Carla simulation platform). The schematic diagram of the integration process is shown in Figure 1 .

[0060] A simulation platform integration module that is responsible for integrating the generated DSL script with the simulation platform (such as Carla) to ensure that the script can be successfully executed in the simulation environment. This module also provides an interface for loading and running test scenarios in the simulation platform.

[0061] The input is a DSL script (such as a script file for the Carla simulation platform), and the output is a test scenario running in the simulation platform.

[0062] The real-time monitoring and evaluation module monitors the behavior of the autonomous driving system in real-time while the simulation platform runs the test scenario, and evaluates it based on predefined evaluation indicators.

[0063] It provides immediate feedback, helping to identify potential problems and shortcomings of the system, ensuring the reliability and safety of the autonomous driving system in various scenarios. The input is a test scenario running in the simulation platform, and the output is an evaluation report (including the performance of the autonomous driving system, potential problems, and improvement suggestions).

[0064] The data storage and update module is responsible for storing all generated test scenarios and evaluation reports, and updating the scenario library and evaluation indicators based on test results to continuously improve the performance and adaptability of the test framework. It ensures that the framework has the ability to self-improve, thereby continuously optimizing the testing process. The input is the evaluation report and the generated test scenario, and the output is the updated scenario library and evaluation indicators, and the improved performance of the test framework.

[0065] The schematic diagram of the whole process is shown in Figure 1 ;

[0066] The hierarchical scenario library module maintains and updates a scenario library that covers road topology, traffic participants, and weather conditions. This module integrates multiple data sources to create a detailed scenario library, covering various road types from urban roads to highways, as well as pedestrians, vehicles, and other traffic participants. Road topology information is obtained through geographic information systems (GIS) and high-definition maps, ensuring detailed data such as intersections, curves, and lane divisions. These data are imported into the system through standardized interfaces, forming a hierarchical road network structure that allows the system to flexibly generate diverse test scenarios.

[0067] To simulate complex traffic environments in the real world, the scenario library also includes dynamic behavior models. These models are trained based on real traffic data using machine learning algorithms, accurately simulating the dynamic behavior and interaction rules of various traffic participants.

[0068] For example, the acceleration, deceleration, lane changing, and other behaviors of vehicles, as well as the random crossing behavior of pedestrians, are precisely modeled. In addition, the weather conditions module introduces meteorological data to simulate various weather conditions such as sunny, rainy, and snowy days, ensuring that test scenarios can cover different environmental conditions.

[0069] The generation of scene templates is achieved through template matching techniques. This technique matches existing templates in the scene library based on the output of the scene element extraction module and generates a base scene. Specifically, the system calculates the similarity of the extracted elements to the templates in the scene library and matches based on the highest similarity template. The scene matching algorithm selects the most appropriate template for matching based on similarity calculation of scene elements, and its formula is as follows:

[0070]

[0071] wherein, is the weight, f i is the feature similarity.

[0072] The input is detailed scene element information (such as specific sections of the highway, car brands and models, and rainfall intensity in rainy weather), and the output is the matched scene template (such as a specific highway section containing dynamic information of the car and weather conditions). The scene library module also provides an interface so that other components can call the library materials for scene generation and construction. This modular design not only improves the flexibility of the system, but also ensures the comprehensiveness and diversity of the test scenarios.

[0073] The hierarchical structure of the scene is shown in Figure 2 We divide it into seven levels.

[0074] The natural language processing and scene element extraction module is responsible for processing the natural language description provided by the user, including word segmentation, part-of-speech tagging, and semantic parsing. First, this module uses natural language processing (NLP) techniques to perform word segmentation and part-of-speech tagging on the input text, identifying important components such as scene-related nouns and verbs. By using modern NLP tools such as spaCy or NLTK, efficient and accurate text processing is achieved. The purpose of word segmentation and part-of-speech tagging is to convert natural language into structured data, making subsequent semantic parsing more accurate.

[0075] After completing the basic text processing, the module will perform dependency syntax analysis and semantic role labeling to understand the structure of the sentence and the relationship between elements. Dependency syntax analysis reveals the dependency relationships between words in a sentence by constructing a dependency tree; semantic role labeling determines the role of each word in an event, such as action, subject, object, etc. This step uses dependency parsers and semantic role labelers such as pre-trained models like BERT to ensure the accuracy of semantic understanding.

[0076] To further improve the accuracy of scene element extraction, this module uses large language models (LLM) and multi-stage context few-shot learning techniques.

[0077] By utilizing large language models such as GPT-4, the system can learn and reason from a small number of examples, extracting specific scene elements.

[0078] This method takes advantage of the powerful language understanding and generation capabilities of large language models. By providing a few examples, the model can generate new scene element extraction rules, ensuring the accuracy and comprehensiveness of the extraction results.

[0079] In specific technical implementation, the module first preprocesses the natural language description (such as "on a rainy day, a car changes lanes on the highway") to obtain word segmentation and part-of-speech tagging results. Then, through dependency syntax analysis and semantic role labeling, it determines the key elements in the scene and their relationships. Finally, it uses a large language model for context-based few-shot learning to refine and perfect the scene element information. The extracted scene element information includes road type, weather, traffic participants, and other detailed information.

[0080] Specific examples of few-shot learning:

[0081] Example input: Generate a scene of automatic driving without protection for left turn.

[0082] Example output: Road scene: intersection; lane: left lane; traffic light: green light; initial position of traffic flow: opposite lane; self-driving endpoint: left turn lane.

[0083] The input is a natural language description, and the output is detailed scene element information. These include road types (such as highways), weather conditions (such as rainy days), traffic participants (such as cars), and more detailed information (such as specific sections of the highway, car brands and models, and rainfall intensity on rainy days). Through this processing flow, the module can efficiently and accurately convert natural language descriptions into structured scene element information, providing a foundation for subsequent scene synthesis.

[0084] The scene synthesis module is responsible for synthesizing templates in the scene library with extracted scene elements to generate complete scene descriptions. By organically combining various elements, it generates a coherent and detailed test scene, ensuring wide test coverage and strong practicality. The module first receives the hierarchical scene library and detailed scene element information provided by the natural language processing module, and then integrates different scene elements into a coherent whole through template synthesis technology.

[0085] The core of the template synthesis technique is to select the most suitable template for synthesis based on the matching degree of scene elements and scene library templates. The system uses logical rules and heuristic algorithms to organically combine scene elements such as road types, traffic participants, weather conditions, etc., to generate a complete scene. For example, if the extracted scene elements include rain, highway, and car lane changing, the system will select the template with the highest matching degree for these elements to generate a specific scene description.

[0086] To ensure the logical consistency and practical feasibility of the synthesized scene, the scene synthesis module also performs a series of verifications. First, the system checks the temporal consistency of the scene, such as whether the time sequence of events is reasonable; second, it checks the spatial consistency, such as whether the positions and movement trajectories of traffic participants on the road are reasonable; finally, it checks the consistency of events, such as whether traffic rules are followed and whether the events in the scene conform to actual traffic logic. These checks are implemented through a series of rules and algorithms to ensure that the generated scene is not only logically coherent but also applicable in actual testing.

[0087] The specific scene synthesis algorithm includes logical rules and heuristic methods, which can be formulated as follows:

[0088]

[0089] where e i is the scene element, t j is the time and space constraint.

[0090] The input is the hierarchical scene library and the detailed scene element information provided by the natural language processing module, and the output is a complete scene description (such as a car changing lanes on a highway section in the rain).

[0091] The synthesized scene not only covers all the extracted elements, but also ensures its feasibility and logical consistency through verification. This method ensures the comprehensiveness and diversity of the test scene, providing rich test conditions and environments for the testing of autonomous driving systems. Through such a process, the scene synthesis module plays a key role in the testing framework, generating high-quality test scenes and improving test coverage and accuracy.

[0092] The DSL script generation module automatically generates domain-specific language (DSL) scripts to convert scene descriptions into executable test scripts. This module parses the complete scene description and uses predefined DSL syntax rules to generate test scripts suitable for simulation platforms. First, the module parses the elements in the scene description according to the syntax rules of the DSL, converting them into corresponding DSL instructions.

[0093] The process of syntax parsing is to build a syntax tree, analyze the elements and their relationships in the scene description, and ensure that each element can be correctly identified and converted. This module uses syntax analysis techniques in compiler theory to convert natural language scene descriptions into intermediate representations through lexical analysis and syntax analysis. Then, through predefined DSL syntax rules, the intermediate representation is converted into DSL instructions. These instructions cover all operations required by the simulation platform, such as vehicle acceleration, deceleration, lane change, etc.

[0094] After parsing is complete, the DSL script generation module generates a structured DSL script based on the parsing results. This script contains all necessary instructions and configurations according to the requirements of the simulation platform, ensuring that the script can be executed correctly on the simulation platform. To ensure the correctness and execution efficiency of the generated script, the module also performs a series of optimizations, such as instruction merging and redundant instruction deletion.

[0095] The specific DSL script generation algorithm is as follows:

[0096] DSLscript = generate(parse(scene description));

[0097] Where the parse function is used to parse the scene description, and the generate function is used to generate DSL instructions. The input is a complete scene description (such as a car changing lanes on a highway section in the rain), and the output is a DSL script (such as a script file for the Carla simulation platform).

[0098] The generated DSL script not only covers all scene elements, but also ensures that each instruction meets the execution requirements of the simulation platform. Through this method, the DSL script generation module can quickly and accurately convert scene descriptions into executable test scripts, providing a reliable execution foundation for the testing of autonomous driving systems.

[0099] DSL script instance:

[0100] Mainly includes scene initialization (initialize actions: set weather, change simulator settings, etc.);

[0101] Storyline (actions: start trigger, end trigger, behavior group: actor, type, behavior (event, start trigger, action, etc.));

[0102] Stop trigger.

[0103] The simulation platform integration module is responsible for integrating the generated DSL scripts with simulation platforms such as Carla, ensuring that the scripts can be successfully executed in the simulation environment. This module also provides interfaces for loading and running test scenarios in the simulation platform. First, the module develops an appropriate adaptation layer based on the requirements of different simulation platforms such as Carla, ensuring that DSL scripts can be recognized and correctly executed by the platform.

[0104] The development of the adaptation layer involves a deep understanding of the simulation platform, including the platform's instruction set, execution process, and interface specifications. By analyzing the execution process of the simulation platform, the module development team formulates a design scheme for the adaptation layer, ensuring that each instruction of the DSL script can be correctly parsed and executed. In addition, the adaptation layer also includes error handling mechanisms to ensure that problems during script execution can be detected and handled in a timely manner, avoiding the impact on test results.

[0105] To facilitate the calling of other modules, the simulation platform integration module provides standardized interfaces. These interfaces cover functions such as scene loading, script execution, and result acquisition, enabling the entire test framework to efficiently interact with the simulation platform. Specifically, the interface design adopts object-oriented programming principles, providing simple and easy-to-use interface methods by encapsulating the underlying implementation. For example, the scene loading interface can accept DSL script files, automatically load and initialize test scenarios; the script execution interface is responsible for starting script execution and monitoring the execution status in real time.

[0106] The input is the DSL script (such as the script file of the Carla simulation platform), and the output is the test scenario running in the simulation platform. Through this modular design, the simulation platform integration module not only ensures the smooth execution of scripts, but also provides efficient interfaces for easy calling and integration of other modules.

[0107] The real-time monitoring and evaluation module monitors the behavior of the autonomous driving system in real time when running test scenarios in the simulation platform, and evaluates it according to predefined evaluation indicators. This module provides immediate feedback to help identify potential problems and shortcomings of the system, ensuring the reliability and safety of the autonomous driving system in various scenarios. Through sensor data and log information, the behavior of the autonomous driving system is monitored in real time, capturing key events and abnormal situations.

[0108] Real-time monitoring involves data collection and processing from multiple sensors of the autonomous driving system. By integrating sensor data interfaces, the module can obtain real-time status of the system, including vehicle location, speed, acceleration, and other dynamic information. Using data processing algorithms, these real-time data are analyzed to identify the behavior patterns and potential anomalies of the system. For example, by analyzing the trajectory data of the vehicle, it can be determined whether it is driving along the expected route; by monitoring speed and acceleration data, it can be detected whether the acceleration and deceleration behavior of the system is reasonable.

[0109] Based on predefined evaluation indicators, the module evaluates the performance of the autonomous driving system. Evaluation indicators cover safety, robustness, and efficiency, among other aspects. For example, safety indicators include the number of collisions, the number of emergency brakes, etc.; robustness indicators include the performance of the system under different weather conditions; efficiency indicators include travel time, fuel consumption, etc. Through comprehensive evaluation of these indicators, an evaluation report of the system is generated.

[0110] The specific evaluation algorithm is as follows:

[0111]

[0112] where, is the weight, v i is the actual value, v max is the maximum value.

[0113] The input is the test scenario running in the simulation platform, and the output is the evaluation report (including the performance of the autonomous driving system, potential problems, and improvement suggestions).

[0114] Through real-time monitoring and evaluation, the module can provide immediate feedback to help the development team discover and solve problems in the system in a timely manner, improving the reliability and safety of the autonomous driving system. This method ensures the comprehensiveness and effectiveness of the test, providing an important basis for the optimization and improvement of the autonomous driving system.

[0115] The data storage and update module is responsible for storing all generated test scenarios and evaluation reports, and updating the scenario library and evaluation indicators based on test results to continuously improve the performance and adaptability of the test framework. This module ensures that the framework has the ability to improve itself, thereby continuously optimizing the testing process. First, use a database system to store the generated test scenarios, DSL scripts, and evaluation reports, ensuring data integrity and traceability.

[0116] The database system adopts a combination of relational databases and non-relational databases to ensure efficient storage and fast access of data. Relational databases are used to store structured data such as scene descriptions, evaluation reports, etc., while non-relational databases are used to store unstructured data such as sensor logs, video recordings, etc. By designing a reasonable database schema, the module can efficiently manage and retrieve large-scale test data.

[0117] Then, according to the test results, dynamically update the scene library and evaluation indicators, optimize the scene template and evaluation method. The module analyzes the data in the evaluation report to identify common problems and bottlenecks in the system, adjusts the templates in the scene library, and adds new test scenarios to cover more test conditions and environments. In addition, the evaluation indicators will also be adjusted according to the actual test results to ensure the scientificity and effectiveness of the evaluation system. For example, if a certain indicator performs poorly in multiple scenarios, the module will redefine and adjust the weight of that indicator.

[0118] The specific update algorithm is as follows:

[0119] new_template = update(existing_template, test_results);

[0120] Where the update function is used to update the scene template according to the test results. The input is the evaluation report and the generated test scenario, and the output is the updated scene library and evaluation indicators, and the improved performance of the test framework.

[0121] In this way, the data storage and update module not only ensures the integrity and traceability of the data, but also realizes the self-optimization and continuous improvement of the test framework. This dynamic updating mechanism enables the test framework to continuously adapt to new testing requirements and environments, improving the comprehensiveness and effectiveness of the test, and providing a solid foundation for the development and validation of autonomous driving systems.

[0122] The embodiments of the present application are described above in combination with the drawings, and the embodiments and features in the present application can be combined with each other without conflict, and the present application is not limited to the above specific embodiments, which are only illustrative and not limiting. Those skilled in the art can make many forms without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, which are all within the protection of the present application.

Claims

1. A text-driven automated driving scenario automatic generation method, characterized in that: The method comprises the following steps: Step 1: receiving a user inputted natural language description of an autonomous driving scenario; Step 2: processing the natural language description to extract scenario key elements, including road type, traffic participants, environmental conditions, and detailed information of each element; Step 3: matching and calling scenario templates based on a hierarchical scenario library, synthesizing the extracted scenario key elements with the scenario templates to generate a complete scenario description; Step 4: converting the complete scenario description into a domain-specific language (DSL) script; Step 5: integrating the DSL script into a simulation platform and running the generated test scenario in the simulation platform; Step 6: monitoring the behavior of the autonomous driving system in the test scenario in real time and generating an evaluation report based on predefined evaluation indicators; Step 7: storing the test scenario, DSL script, and evaluation report, and updating the hierarchical scenario library and evaluation indicators based on the evaluation report.

2. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The hierarchical scenario library maintains and updates scenario elements covering road topology, traffic participants, and weather conditions, and includes dynamic behavior models trained based on real traffic data, which are used to simulate the acceleration, deceleration, lane changing, and pedestrian random crossing behaviors of traffic participants; The hierarchical scenario library generates scenario templates through template matching technology based on scenario element similarity calculation, with the calculation formula being: where e i is a scene element, t j is time and spatial constraints.

3. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: Processing the natural language description includes the following steps: Tokenizing and part-of-speech tagging the natural language description; Determining scenario key elements and their relationships through dependency syntax analysis and semantic role labeling; Refining scenario key element information using a large language model and multi-stage context few-shot learning; The few-shot learning provides scenario generation examples to enable the model to generate scenario element extraction rules, with the examples including natural language input and corresponding scenario element output.

4. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The large language model is GPT-4, and the detailed information of the scenario key elements includes specific sections of a highway, the brand and model of traffic participants, and rainfall intensity.

5. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The process of converting the complete scenario description into a DSL script includes converting the complete scenario description into an intermediate representation form through lexical analysis and syntax analysis, converting the intermediate representation form into DSL instructions based on predefined DSL syntax rules, and generating a structured DSL script; The generation algorithm is: DSLscript = generate(parse(scene description)); Where the parse function is used to parse the scenario description, and the generate function is used to generate DSL instructions.

6. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The process of integrating the DSL script into the simulation platform includes developing an adaptation layer to adapt to the instruction set and interface specifications of the simulation platform, with the adaptation layer including an error handling mechanism; Providing a standardized interface for scenario loading, script execution, and result acquisition, with the simulation platform including Carla.

7. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The process of real-time monitoring and generating an evaluation report includes collecting sensor data of the autonomous driving system, including vehicle position, speed, and acceleration; Based on the safety, robustness and efficiency indicators, the safety indicators include the number of collisions and the number of emergency brakes, and the evaluation algorithm is: wherein, is the weight, v i is the actual value, v max is the maximum value.

8. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The storage and updating process includes: using a relational database to store structured data and a non-relational database to store unstructured data; Based on the evaluation report, the common problems of the system are identified, and the scene templates and the weights of the evaluation indicators in the scene library are updated.

9. The text-driven automated driving scene automatic generation method according to claim 1, characterized in that: The example input of the multi-stage context few-shot learning is to generate an automatic driving unprotected left turn scene, and the example output includes road scene, lane, signal light, traffic flow initial position and self-driving endpoint information.

Images