An automatic driving simulation test method and system based on a large model agent

Abstract

The application relates to the technical field of automatic driving simulation data processing, and discloses an automatic driving simulation test method and system based on a large model Agent, wherein the method comprises the following steps: obtaining an automatic driving test requirement text; performing risk scenario combination; performing test case arrangement preprocessing; performing automatic driving closed-loop simulation testing; and performing abnormal attribution and iterative control. Compared with the prior art, the automatic driving simulation test relies on manual disassembly of test requirements, especially under complex risk conditions such as sudden crossing of traffic participants, and cannot realize continuous linkage among test cases, test execution and abnormal attribution in a simulation scene. Since the test requirement is structurally analyzed by the large model Agent, and the test task description set is sequentially used for risk scenario combination, test case arrangement and closed-loop simulation execution, continuous closed-loop processing of automatic driving simulation testing from requirement analysis to test iterative control is realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving simulation data processing technology, and in particular to an autonomous driving simulation testing method and system based on a large model agent. Background Technology

[0002] Currently, the development and verification of autonomous driving systems heavily rely on simulation testing, which allows for efficient and safe evaluation of system performance in a virtual environment. However, existing autonomous driving simulation testing methods have many shortcomings. For example, the construction of test scenarios typically depends on engineers manually defining them or selecting them from a limited scenario library. This approach is not only inefficient but also struggles to cover the massive and complex long-tail risk scenarios of the real world, especially those complex risk scenarios intertwined with various traffic participants, environmental conditions, and dynamic events. Existing technologies cannot fully meet the needs for systematic, high-coverage, and high-efficiency verification of autonomous driving systems.

[0003] Furthermore, the process of transforming unstructured natural language testing requirements (such as regulatory requirements, functional definitions, and accident report descriptions) into executable simulation test cases currently relies heavily on manual interpretation and orchestration. This results in low automation, susceptibility to subjective bias, and difficulty in ensuring a complete and accurate mapping of test cases to the original requirements. Simultaneously, post-test result analysis often remains at the pass / fail level, lacking in-depth attribution of abnormal behavior and intelligent iterative optimization capabilities based on feedback. Therefore, there is an urgent need for a method and system capable of automating and intelligently constructing scenarios, orchestrating test cases, and iteratively optimizing them even when facing complex natural language testing requirements, in order to improve the coverage, efficiency, and intelligence level of autonomous driving simulation testing. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to propose an autonomous driving simulation testing method based on a large-scale agent model. This method aims to solve the technical problem in existing autonomous driving simulation testing that relies on manual breakdown of test requirements, especially under complex risk conditions such as sudden crossings by traffic participants, where continuous linkage between test cases, test execution, and anomaly attribution in the simulation scenario cannot be achieved.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an autonomous driving simulation test method based on a large model agent.

[0006] The aforementioned autonomous driving simulation testing method based on a large model agent includes: Step S10: Obtain the autonomous driving test requirement text, perform test task parsing based on the autonomous driving test requirement text, and output the test task description set; Step S20: Based on the test task description set, perform risk scenario combination and output the simulation scenario set to be tested; Step S30: Perform test case orchestration preprocessing based on the set of simulation scenarios to be tested, and output a sorted test case set; Step S40: Execute autonomous driving closed-loop simulation test based on the sorted test case set, and output the test process dataset; Step S50: Perform anomaly attribution and iterative control based on the test process dataset, and output the test iteration control set.

[0007] Preferably, step S10, which involves obtaining the autonomous driving test requirement text, performing test task parsing based on the autonomous driving test requirement text, and outputting a test task description set, specifically includes: Step S101: Input the autonomous driving test requirement text into a preset large model Agent. The large model Agent extracts keywords from the autonomous driving test requirement text according to the preset functional verification vocabulary, risk event vocabulary, road type vocabulary, environmental condition vocabulary and test object vocabulary to obtain a set of requirement keywords. Step S102: Map the set of requirement keywords to the functional items to be verified, road type, environmental conditions, traffic participant type, target risk event, expected vehicle behavior, test termination conditions and result judgment indicators to obtain the initial test task table; Step S103: Classify the test tasks according to the initial test task table and generate a test task description set.

[0008] Preferably, step S20, which involves performing risk scenario combinations based on the test task description set and outputting a set of simulation scenarios to be tested, specifically includes: Step S201: Based on the road type in the test task description set, call the corresponding basic road scene from the preset road scene library; Step S202: Based on the environmental conditions and target risk events in the test task description set, write the weather parameters, illumination parameters, road surface adhesion parameters and visibility parameters into the basic road scene to obtain the environment-enhanced road scene; Step S203: Based on the traffic participant types, expected vehicle behaviors, and target risk events in the test task description set, bind the traffic participant behavior trajectories to the environment-enhanced road scene to obtain a candidate simulation scene set; Step S204: Select candidate simulation scenarios based on the candidate simulation scenario set to generate a set of simulation scenarios to be tested.

[0009] Preferably, step S204, the step of filtering candidate simulation scenarios based on the candidate simulation scenario set to generate a set of simulation scenarios to be tested, includes: Calculate the scenario risk coverage value for each candidate simulation scenario in the candidate simulation scenario set; When the scenario risk coverage value is greater than or equal to the preset scenario retention threshold, the corresponding candidate simulation scenario will be written into the simulation scenario set to be tested. When the scenario risk coverage value is less than the scenario retention threshold, scenario perturbation processing is performed on the corresponding candidate simulation scenario, and when the perturbed scenario risk coverage value is greater than or equal to the scenario retention threshold, the perturbed candidate simulation scenario is written into the simulation scenario set to be tested. The scene disturbance processing includes at least one of the following: reducing visibility parameters, reducing road surface adhesion parameters, increasing the speed of traffic participants, shortening the initial distance between traffic participants and target vehicles, and advancing the triggering time of risk events. The scenario risk coverage value is calculated according to the following formula:

[0010] in, Indicates the first The scenario risk coverage value of each candidate simulation scenario. Indicates the first The matching values ​​of the candidate simulation scenarios and the road types in the test task description set. Indicates the first The severity values ​​of environmental parameters in each candidate simulation scenario. Indicates the first Traffic participant density values ​​in candidate simulation scenarios Indicates the first The matching values ​​between traffic participant behaviors and target risk events in candidate simulation scenarios. Indicates the historical test data and the first Normalized values ​​of failure frequencies for candidate simulation scenarios that are the same or similar to those of other scenarios. , , , , All are preset weighting coefficients and satisfy the following: .

[0011] Preferably, step S30, which involves performing test case orchestration preprocessing based on the set of simulation scenarios to be tested and outputting a sorted test case set, specifically includes: Step S301: Construct a scene coverage matrix based on the set of simulation scenarios to be tested. The scene coverage matrix includes road type, weather type, illumination status, road surface adhesion level, traffic participant type, target risk event, and functional items to be verified. Step S302: Generate an initial test case set based on the scene coverage matrix and the set of simulation scenes to be tested, and associate each initial test case with a simulation scene to be tested; Step S303: Obtain the scenario risk coverage value, coverage matrix cell status, risk event triggering complexity, and historical failure frequency corresponding to the initial test cases, and sort the initial test cases based on the scenario risk coverage value, coverage matrix cell status, risk event triggering complexity, and historical failure frequency to generate a sorted test case set; When there are matrix cells in the scenario coverage matrix that are not associated with initial test cases, the large model agent selects the simulation scenario with the highest matching degree from the set of simulation scenarios to be tested, and generates supplementary test cases by adjusting the initial speed of the target vehicle, the initial position of the traffic participant, the trigger time of the traffic participant, and environmental parameters. The supplementary test cases are then incorporated into the sorted test case set.

[0012] Preferably, step S40, which involves performing autonomous driving closed-loop simulation testing based on the sorted test case set and outputting the test process dataset, specifically includes: Step S401: According to the execution order of the sorted test case set, read the associated simulation scenario number, initial state of the target vehicle, initial state of the traffic participants, risk event trigger time and test termination condition of the current test case in sequence, and load the corresponding simulation scenario to be tested; Step S402: Connect the target autonomous driving test object to the loaded simulation scene to be tested, so that the target autonomous driving test object receives simulation sensor data and outputs perception results, planned trajectory and control commands; Step S403: Trigger changes in traffic participant behavior and environmental parameters according to the risk event triggering time in the current test case, and collect target vehicle status, traffic participant status, environmental status, perception results, planned trajectory and control commands according to a fixed sampling period to generate a test process dataset.

[0013] Preferably, step S403 further includes: calculating the instantaneous safe distance deviation between the target vehicle and the m-th traffic participant at the t-th sampling time, wherein the instantaneous safe distance deviation is calculated according to the following formula:

[0014] in, Indicates the first The target vehicle at the sampling time and the first Instantaneous safe distance deviation between traffic participants Indicates the first The target vehicle at the sampling time and the first The actual distance between traffic participants Indicates the first The target vehicle speed at each sampling time. Indicates the first The target vehicle acceleration at each sampling time point Indicates the first Road surface adhesion parameters at each sampling time This indicates the minimum safe distance under the current speed, acceleration, and road adhesion parameters; When the instantaneous safe distance deviation is less than zero, the target vehicle status and traffic participant status at the corresponding sampling time are written into the safe distance anomaly table; When the lateral offset between the target vehicle trajectory and the road boundary line is greater than the lane departure threshold, the corresponding sampling time is written into the lane departure anomaly table. When the curvature change of the target vehicle's planned trajectory between two adjacent sampling times is greater than the planning mutation threshold, the corresponding sampling time is written into the planning mutation anomaly table. When the acceleration, deceleration or steering angular velocity in the control command exceeds the corresponding control threshold, the corresponding sampling time will be written into the control over-limit exception table. The safety distance anomaly table, lane departure anomaly table, planning mutation anomaly table, and control over-limit anomaly table are incorporated into the test process dataset.

[0015] This invention also provides an autonomous driving simulation testing system based on a large-scale agent model, comprising: The test task parsing module is used to obtain autonomous driving test requirement text, perform test task parsing based on the autonomous driving test requirement text, and output a test task description set. The simulation scenario combination module is used to perform risk scenario combination based on the test task description set and output the simulation scenario set to be tested. The test case orchestration module is used to perform test case orchestration preprocessing based on the set of simulation scenarios to be tested, and output a sorted set of test cases. The closed-loop simulation test module is used to perform autonomous driving closed-loop simulation tests based on the sorted test case set and output the test process dataset. The anomaly attribution iteration module is used to perform anomaly attribution and iteration control based on the test process dataset, and output a test iteration control set.

[0016] The present invention also provides an autonomous driving simulation test device based on a large model agent. The autonomous driving simulation test device based on a large model agent includes: a memory, a processor, and an autonomous driving simulation test program based on a large model agent stored in the memory and executable on the processor. When the autonomous driving simulation test program based on a large model agent is executed by the processor, it implements an autonomous driving simulation test method based on a large model agent.

[0017] The present invention also provides a computer program product, the computer program product including an autonomous driving simulation test program based on a large model agent, the autonomous driving simulation test program based on a large model agent implementing an autonomous driving simulation test method based on a large model agent when executed by a processor.

[0018] The beneficial effects of this invention are as follows: 1. This invention introduces a large-scale model agent as the core intelligent agent, achieving automated and intelligent generation and combination of natural language testing requirements, structured test task descriptions, and high-coverage risk simulation scenarios. This method leverages the powerful semantic understanding and knowledge extraction capabilities of the large-scale model to accurately parse unstructured requirements. Combined with a pre-set multi-dimensional vocabulary (function, risk, road, environment, object), it maps requirements into quantifiable and executable test elements. This fundamentally changes the inefficient traditional model that relies on human experience to construct test scenarios, significantly improving the automation level of test scenario construction and the accuracy of requirement coverage.

[0019] 2. This invention, through the design of a dynamic filtering and perturbation mechanism based on scenario risk coverage values, and an intelligent test case orchestration and supplementation strategy based on the scenario coverage matrix, can automatically generate and optimize test case sets, ensuring that test cases maximize coverage of the target risk space under limited resources. Simultaneously, in closed-loop simulation testing, multiple safety and performance indicators (such as instantaneous safe distance deviation, lane departure, planning abrupt changes, and control exceedances) are calculated in real time, and a structured anomaly data table is generated, providing a refined data foundation for subsequent in-depth attribution. Combined with anomaly attribution and iterative control based on test process data, a complete intelligent testing closed loop of "requirement analysis – scenario generation – test case orchestration – test execution – analysis iteration" is formed, achieving self-optimization and continuous evolution of the testing process, and significantly improving the comprehensiveness, depth, and efficiency of autonomous driving system simulation verification. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the first embodiment of an autonomous driving simulation testing method based on a large model agent according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the autonomous driving simulation test method based on a large model agent according to the present invention. The first embodiment of the autonomous driving simulation test method based on a large model agent according to the present invention is presented.

[0024] In the first embodiment, the autonomous driving simulation testing method based on a large model agent includes: Step S10: Obtain the autonomous driving test requirement text, perform test task parsing based on the autonomous driving test requirement text, and output the test task description set; It should be noted that the "autonomous driving test requirement text" in this step refers to unstructured text requirements in natural language, describing the functions, performance indicators, test conditions, or risk events to be verified in the autonomous driving system. For example, "Test the vehicle's response capability when faced with a sudden braking by a vehicle in front at a slippery urban intersection in rainy weather." "Test task parsing" refers to the process of extracting and mapping structured information from the above text using a pre-set large-scale model (Agent) combined with a functional verification vocabulary, a risk event vocabulary, a road type vocabulary, an environmental condition vocabulary, and a test object vocabulary. The output "test task description set" is a structured dataset containing key test elements parsed from the requirement text, such as the functions to be verified, road types, environmental conditions, traffic participant types, target risk events, expected vehicle behavior, test termination conditions, and result judgment indicators. This provides precise and actionable input for subsequent steps.

[0025] Understandably, this step, as the starting point and input interface of the entire testing process, plays a crucial role in transforming vague and unstructured user or regulatory requirements into standardized test task descriptions that are understandable and executable by machines. Through the semantic understanding capabilities of the large-scale model agent, the standardization and element-based representation of the requirement language are achieved, ensuring that all subsequent steps are based on a unified and clear task objective. This forms the foundation for the automation and intelligence of the entire method.

[0026] It should be understood that, compared to the traditional approach that relies entirely on test engineers manually reading requirements documents and writing test case outlines or scenario descriptions based on experience, this step automates requirements parsing through a large-scale model agent. This not only significantly improves the efficiency of requirements conversion and avoids test deviations caused by inconsistencies or omissions in human interpretation, but also enables the handling of more complex and colloquial requirements descriptions, expanding the input forms of test requirements. This allows testing activities to respond more directly and quickly to diverse sources of requirements such as product definitions, security assessments, or incident reviews.

[0027] For example, when the input requirement text is "Verify the lane-changing avoidance function of an autonomous truck encountering fog and a disabled vehicle blocking the lane on a mountain highway at night," the large model agent in this step will extract keywords: the function to be verified is "lane-changing avoidance," the road type is "mountain highway," the environmental conditions include "night" and "fog" (corresponding to low light and low visibility parameters), the types of traffic participants include "autonomous vehicle (truck)" and "disabled vehicle (stationary obstacle)," the target risk event is "stationary obstacle blocking the lane," the expected vehicle behavior may include "obstacle recognition," "initiating lane change," and "safely completing lane change," the test termination condition may be "successfully passing the lane change" or "collision occurring," and the result evaluation indicators may include "time to collision (TTC)" and "lane-changing trajectory smoothness," etc. These elements together constitute a structured test task description, driving the subsequent scenario construction.

[0028] Step S20: Based on the test task description set, perform risk scenario combination and output the simulation scenario set to be tested; It should be noted that the "risk scenario combination" in this step refers to the process of intelligently selecting and combining basic elements from the preset road scenario library, environmental parameter library, and behavioral model library to construct a specific simulation scenario based on the structured test task description set output in the previous step. The "simulation scenario set to be tested" is the output of this step, which contains one or more specific 3D virtual scene files that can be loaded into the simulation engine and their parameter configurations. Each scene accurately reflects the road topology (such as intersections, high-speed curves), environmental conditions (such as rain and snow, nighttime lighting), and dynamic traffic participants and their preset behavioral trajectories (such as vehicles cutting in, pedestrians crossing), and in particular, embeds the triggering logic of the "target risk event".

[0029] Understandably, the technical advantage of this step lies in its automated and intelligent mapping from abstract test elements to concrete simulation instances. It dynamically assembles basic scenario modules and injects risk elements based on task requirements, thereby generating a large number of targeted and diverse test scenarios. This is not merely the "construction" of scenarios, but also the "combination" and "enhancement" based on risk objectives, ensuring that the generated scenarios can effectively bear and trigger the risk events to be verified.

[0030] It should be understood that, compared to the traditional approach of manually building scenarios in simulation software or selecting from a limited library of fixed scenarios, this step automatically calls and combines the basic resource library in a programmatic and parameterized manner. This enables the rapid generation of a large number of scenario variants that conform to specific risk patterns, greatly expanding the coverage and generation efficiency of test scenarios. In particular, combined with the scenario risk coverage value calculation and perturbation mechanism in step S204, the generated candidate scenarios can be quantitatively evaluated and optimized, filtering out redundant or low-risk scenarios. Furthermore, by perturbing key parameters (such as visibility, vehicle distance, and speed), more extreme or more easily exposed system defects can be generated, thereby achieving higher risk coverage with fewer test resources.

[0031] For example, regarding the task description of "encountering a disabled vehicle in nighttime fog on a mountain highway" generated in the previous step, this step first calls a model of a curved, sloping road section that matches the characteristics of a "mountain highway" from the road scene library as the base road scene. Then, based on the conditions of "nighttime" and "fog," the illumination parameters are set to dark mode, the visibility parameters are significantly reduced, and the road surface humidity parameter may be increased to simulate slippery conditions, resulting in an environmentally enhanced road scene. Next, based on the type of traffic participant, a stationary vehicle model (disabled vehicle) is placed at a specific location in front of the lane, and a "stationary" behavior trajectory is bound to it. Finally, the risk coverage value of this scene is calculated. If the value is high (e.g., due to the high risk of the combination of low visibility and stationary obstacles), it is directly included in the test set; if the value is low, it may be perturbed by "shortening the initial distance between the vehicle and the disabled vehicle" or "further reducing visibility" to generate a variant scene with a higher risk before being included in the test set.

[0032] Step S30: Perform test case orchestration preprocessing based on the set of simulation scenarios to be tested, and output a sorted test case set; It should be noted that the "test case orchestration preprocessing" in this step refers to the intelligent process of further planning the test execution logic and sequence based on the generated set of simulation scenarios to be tested. The "sorted test case set" is the output of this step. It not only encapsulates each simulation scenario to be tested into an independently executable test case (associated with scenario files, self-test initial states, test termination conditions, etc.), but more importantly, it prioritizes these test cases. The sorting is based on factors including the scenario risk coverage value, the unique test dimensions (unit states) covered by the test case in the scenario coverage matrix, the logical complexity of risk event triggering, and the failure frequency of the scenario or similar scenarios in historical tests.

[0033] Understandably, the technical benefit of this step lies in its optimized scheduling of test resources. By intelligently prioritizing test cases, tests with higher risks, more unique coverage dimensions, or a history of higher failure rates can be executed first. This allows testing activities to discover potential defects faster and earlier, improving testing efficiency. Simultaneously, by constructing a "scenario coverage matrix" and checking its completeness, this step can automatically identify blind spots in test coverage and drive the large model agent to generate "supplementary test cases" to fill these blind spots, thereby ensuring that the test case set provides the most comprehensive possible coverage across multiple dimensions (roads, weather, risk type, etc.).

[0034] It should be understood that, compared to traditional techniques where test cases are executed in alphabetical or random order of scenario names, lacking optimized scheduling methods, this step introduces an intelligent orchestration strategy based on multi-objective optimization (risk, coverage, historical data). This makes test execution no longer a blind traversal, but a guided exploration, capable of hitting system weaknesses with a higher probability and faster speed. Simultaneously, automated coverage matrix analysis and test case supplementation mechanisms solve the problem of systematically ensuring combined coverage when manually designing test cases, improving the comprehensiveness and rigor of the tests.

[0035] For example, suppose the set of simulation scenarios to be tested contains 100 scenarios, covering various combinations such as urban roads / highways, sunny / rainy days, and vehicle intrusion / pedestrian crossing. This step first constructs a multi-dimensional scenario coverage matrix. Then, it finds that the combination cell "highway + heavy rain + pedestrian crossing" in the matrix is ​​empty (because it is rare in reality but needs to be verified theoretically). At this point, the large model agent will select the closest "highway + rainy day" scenario from the scenario library and generate a supplementary test case by adjusting parameters (such as adjusting the rainfall intensity to the maximum to simulate heavy rain, adding a pedestrian model on the roadside and setting its crossing behavior). Finally, during the sorting, a scenario that has historically frequently caused test failures, "urban intersection + rainy day + right-turning vehicle conflict," and the newly generated "highway + heavy rain + pedestrian crossing" scenario, which is high-risk and has high coverage uniqueness, may be given a high execution priority and placed at the top of the test queue.

[0036] Step S40: Execute autonomous driving closed-loop simulation test based on the sorted test case set, and output the test process dataset; It should be noted that the "autonomous driving closed-loop simulation test" in this step refers to the process of treating the autonomous driving system (software algorithm or model) under test as the controlled object, connecting it to the simulation environment, forming a complete "perception-planning-control-simulation environment" closed loop, and driving this closed loop to run according to the test case settings. The "test process dataset" is the core output of this step. It is a collection of full-link, time-series data collected at a fixed sampling period (e.g., 10 milliseconds) during the simulation operation. This data includes not only the state (position, speed, acceleration, heading angle) of the target vehicle and all traffic participants, and the environmental state, but more importantly, it includes the perception results inside the autonomous driving system (e.g., a list of detected obstacles), the future trajectory output by the planning module, and the final execution commands (accelerator, brake, steering angle) issued by the control module.

[0037] Understandably, the technical advantage of this step lies in its ability to capture holographic data on the dynamic responses of autonomous driving systems under pre-defined risk scenarios. It not only records the system's final pass / fail outcome but also the complete decision-making chain and external environment interaction process leading to that outcome. This fine-grained data acquisition, including internal intermediate states, provides an indispensable data foundation for the next step of in-depth "anomaly attribution." Simultaneously, the calculation and recording of real-time safety indicators can quickly pinpoint dangerous moments and anomaly types, improving the efficiency of problem localization.

[0038] It should be understood that, compared to traditional simulation testing, which often only records the final result (such as whether a collision occurred or whether the vehicle deviated from the road) or a few macroscopic indicators, this step, through high-frequency, end-to-end data acquisition, achieves a transformation of the testing process from a "black box" to a "gray box," and even partially a "white box," observation. This allows analysts or subsequent automated analysis modules to trace the root cause of anomalies; for example, whether a collision was caused by a missed detection by a sensor, an unreasonable planned trajectory, or a control response delay. This deep data acquisition capability is a prerequisite for effective iterative optimization and a key link in achieving intelligent testing closed-loop in this invention.

[0039] For example, when executing a test case of "the vehicle in front suddenly decelerates," the simulation environment triggers the emergency braking behavior of the vehicle in front at a specific moment. The closed-loop simulation in this step begins running and continuously collects data. At a certain sampling moment, the test data set may record: the environmental status shows the brake lights of the vehicle in front are illuminated; the perception output of the autonomous driving system shows the vehicle in front is correctly identified and the distance is rapidly decreasing; the trajectory output by the planning module is a smooth deceleration curve; and the control command issues a large deceleration request. Simultaneously, real-time calculations reveal that the instantaneous safe distance deviation Lmt is less than zero for the first time, therefore this moment and the related vehicle status are recorded in the "safe distance anomaly table." These data collectively constitute a complete snapshot of this "dangerous approach" event, awaiting subsequent analysis.

[0040] Step S50: Perform anomaly attribution and iterative control based on the test process dataset, and output the test iteration control set.

[0041] It should be noted that the "anomaly attribution and iterative control" in this step refers to performing root cause analysis on anomalies exposed during testing (such as safe distance warnings, collisions, lane departures, etc.) based on the rich test process dataset collected in the previous step, and automatically adjusting the test strategy or parameters according to the analysis results to drive the next round of more effective testing. The "test iteration control set" is the output of this step, which may contain various forms of control instructions, such as: indicating the need to add more relevant scenarios for a specific type of perception miss (such as small objects at night); suggesting the direction of adjusting the scenario perturbation parameters (such as further shortening the vehicle distance to test the limits of the planning module); marking certain scenarios that have passed the test as having low risk coverage values ​​and thus being downgraded in priority; and even feeding back to the requirements analysis module in step S10, suggesting the addition of certain risk dimensions that were not previously considered.

[0042] Understandably, the technical benefit of this step lies in endowing the entire testing process with the ability to learn and continuously optimize. By feeding test results (especially anomalous data) back to the front-end stages such as scenario generation and test case orchestration, a complete intelligent closed loop of "test-analysis-optimization" is formed. This makes testing activities no longer static and one-off, but dynamic and evolving, automatically focusing on the system's weaknesses as testing progresses, continuously improving the relevance and effectiveness of the tests.

[0043] It should be understood that, compared to traditional testing processes where test execution and result analysis are often disconnected, and the analysis conclusions cannot directly and automatically guide the updating and optimization of test cases, this step achieves an automated feedback loop from test data to test strategy. It solves the key problem of "how to use the results of one test to guide the next, better test." Through this iterative control, the system can learn from failures like an experienced testing expert, continuously adjusting its testing focus, thereby enabling faster and deeper discovery of defects in autonomous driving systems with limited testing resources, significantly accelerating the system's verification and maturation process.

[0044] For example, after analyzing the "test process dataset" from multiple rounds of testing, the attribution module in this step might discover that over 60% of "safe distance anomalies" are fundamentally related to a significant decrease in the perception module's recognition distance for distant stationary obstacles under "heavy rain + backlight" conditions. Based on this attribution result, the "test iteration control set" might generate the following instructions: 1) Send a request to the simulation scene combination module (S20) to significantly increase the combined weight of the "heavy rain" and "backlight" environmental conditions when generating future scenes, and generate more scene variations under these conditions. 2) Send an instruction to the test case orchestration module (S30) to increase the priority of all test cases containing the "heavy rain + backlight" conditions. 3) Record this attribution conclusion for subsequent targeted optimization by the autonomous driving algorithm team. In this way, the next round of testing will automatically focus on the discovered system weaknesses.

[0045] Example 2: Furthermore, the present invention provides an autonomous driving simulation testing system based on a large-model agent, employing an autonomous driving simulation testing method based on a large-model agent as described in the above embodiments, which can solve a technical problem in autonomous driving simulation testing based on a large-model agent. The beneficial effects of the autonomous driving simulation testing system based on a large-model agent provided by the present invention are the same as those of the autonomous driving simulation testing method based on a large-model agent provided in the above embodiments, and other technical features of the autonomous driving simulation testing system based on a large-model agent are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0046] Example 3: This invention provides an autonomous driving simulation test device based on a large-model agent. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which are then executed to enable the at least one processor to perform the autonomous driving simulation test method based on a large-model agent described in Example 1. The autonomous driving simulation test device based on a large-model agent in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This autonomous driving simulation test device based on a large-model agent is merely an example and should not limit the functionality or scope of the invention. An autonomous driving simulation test device based on a large-model agent may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can execute various appropriate actions and processes according to a program stored in read-only memory or a program loaded from a storage device into random access memory. The random access memory also stores various programs and data required for the operation of the autonomous driving simulation test device based on a large-model agent. The processing unit, read-only memory, and random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including, for example, magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the autonomous driving simulation test device based on a large-model agent to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows an autonomous driving simulation test device based on a large-model agent with various systems, it should be understood that it is not required to implement or possess all the systems shown. It can be implemented alternatively or with more or fewer systems.

[0047] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for autonomous driving simulation testing based on a large-model agent. The computer program product provided by this invention can solve a technical problem in autonomous driving simulation testing based on a large-model agent. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the autonomous driving simulation testing method based on a large-model agent provided in the above embodiments, and will not be repeated here.

[0048] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0049] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0050] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention or its equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for simulating and testing autonomous driving based on a large-scale agent model, characterized in that, The methods include: Step S10: Obtain the autonomous driving test requirement text, perform test task parsing based on the autonomous driving test requirement text, and output a test task description set; wherein, the step of obtaining the autonomous driving test requirement text, performing test task parsing based on the autonomous driving test requirement text, and outputting a test task description set specifically includes: The autonomous driving test requirement text is input into a preset large model Agent. The large model Agent extracts keywords from the autonomous driving test requirement text according to a preset functional verification vocabulary, risk event vocabulary, road type vocabulary, environmental condition vocabulary, and test object vocabulary to obtain a set of requirement keywords. The set of required keywords is mapped to functional items to be verified, road types, environmental conditions, traffic participant types, target risk events, expected vehicle behavior, test termination conditions, and result judgment indicators to obtain the initial test task table; The test tasks are classified according to the initial test task table, and a test task description set is generated. Step S20: Execute risk scenario combinations based on the test task description set to output a set of simulation scenarios to be tested; wherein, the step of executing risk scenario combinations based on the test task description set to output a set of simulation scenarios to be tested specifically includes: Based on the road type in the test task description set, the corresponding basic road scene is called from the preset road scene library; Based on the environmental conditions and target risk events in the test task description set, weather parameters, illumination parameters, road surface adhesion parameters, and visibility parameters are written into the basic road scene to obtain an environment-enhanced road scene. Based on the traffic participant types, expected vehicle behaviors, and target risk events in the test task description set, the traffic participant behavior trajectories are bound to the environment-enhanced road scenario to obtain a candidate simulation scenario set; Based on the candidate simulation scenario set, candidate simulation scenarios are selected to generate a set of simulation scenarios to be tested; Step S30: Perform test case orchestration preprocessing based on the set of simulation scenarios to be tested, and output a sorted test case set; Step S40: Execute autonomous driving closed-loop simulation test based on the sorted test case set, and output the test process dataset; Step S50: Perform anomaly attribution and iterative control based on the test process dataset, and output the test iteration control set.

2. The autonomous driving simulation testing method based on a large-scale agent model as described in claim 1, characterized in that, The steps of selecting candidate simulation scenarios from the candidate simulation scenario set and generating a set of simulation scenarios to be tested include: Calculate the scenario risk coverage value for each candidate simulation scenario in the candidate simulation scenario set; When the scenario risk coverage value is greater than or equal to the preset scenario retention threshold, the corresponding candidate simulation scenario will be written into the simulation scenario set to be tested. When the scenario risk coverage value is less than the scenario retention threshold, scenario perturbation processing is performed on the corresponding candidate simulation scenario, and when the perturbed scenario risk coverage value is greater than or equal to the scenario retention threshold, the perturbed candidate simulation scenario is written into the simulation scenario set to be tested. The scene disturbance processing includes at least one of the following: reducing visibility parameters, reducing road surface adhesion parameters, increasing the speed of traffic participants, shortening the initial distance between traffic participants and target vehicles, and advancing the triggering time of risk events. The scenario risk coverage value is calculated according to the following formula: in, Indicates the first The scenario risk coverage value of each candidate simulation scenario. Indicates the first The matching values ​​of the candidate simulation scenarios and the road types in the test task description set. Indicates the first The severity values ​​of environmental parameters in each candidate simulation scenario. Indicates the first Traffic participant density values ​​in candidate simulation scenarios Indicates the first The matching values ​​between traffic participant behaviors and target risk events in candidate simulation scenarios. Indicates the historical test data and the first Normalized values ​​of failure frequencies for candidate simulation scenarios that are the same or similar to those of other scenarios. , , , , All are preset weighting coefficients and satisfy the following: .

3. The autonomous driving simulation testing method based on a large-model agent as described in claim 1, characterized in that, Step S30, which involves performing test case orchestration preprocessing based on the set of simulation scenarios to be tested and outputting a sorted test case set, specifically includes: Step S301: Construct a scene coverage matrix based on the set of simulation scenarios to be tested. The scene coverage matrix includes road type, weather type, illumination status, road surface adhesion level, traffic participant type, target risk event, and functional items to be verified. Step S302: Generate an initial test case set based on the scene coverage matrix and the set of simulation scenes to be tested, and associate each initial test case with a simulation scene to be tested; Step S303: Obtain the scenario risk coverage value, coverage matrix cell status, risk event triggering complexity, and historical failure frequency corresponding to the initial test cases, and sort the initial test cases based on the scenario risk coverage value, coverage matrix cell status, risk event triggering complexity, and historical failure frequency to generate a sorted test case set; When there are matrix cells in the scenario coverage matrix that are not associated with initial test cases, the large model agent selects the simulation scenario with the highest matching degree from the set of simulation scenarios to be tested, and generates supplementary test cases by adjusting the initial speed of the target vehicle, the initial position of the traffic participant, the trigger time of the traffic participant, and environmental parameters. The supplementary test cases are then incorporated into the sorted test case set.

4. The autonomous driving simulation testing method based on a large-scale agent model as described in claim 1, characterized in that, Step S40, which involves performing autonomous driving closed-loop simulation testing based on the ordered test case set and outputting the test process dataset, specifically includes: Step S401: According to the execution order of the sorted test case set, read the associated simulation scenario number, initial state of the target vehicle, initial state of the traffic participants, risk event trigger time and test termination condition of the current test case in sequence, and load the corresponding simulation scenario to be tested; Step S402: Connect the target autonomous driving test object to the loaded simulation scene to be tested, so that the target autonomous driving test object receives simulation sensor data and outputs perception results, planned trajectory and control commands; Step S403: Trigger changes in traffic participant behavior and environmental parameters according to the risk event triggering time in the current test case, and collect target vehicle status, traffic participant status, environmental status, perception results, planned trajectory and control commands according to a fixed sampling period to generate a test process dataset.

5. The autonomous driving simulation testing method based on a large-model agent as described in claim 4, characterized in that, Step S403 also includes: in the first Calculate the target vehicle and the first sampling time. The instantaneous safe distance deviation between traffic participants is calculated according to the following formula: in, Indicates the first The target vehicle at the sampling time and the first Instantaneous safe distance deviation between traffic participants Indicates the first The target vehicle at the sampling time and the first The actual distance between traffic participants Indicates the first The target vehicle speed at each sampling time. Indicates the first The target vehicle acceleration at each sampling time point Indicates the first Road surface adhesion parameters at each sampling time This indicates the minimum safe distance under the current speed, acceleration, and road adhesion parameters; When the instantaneous safe distance deviation is less than zero, the target vehicle status and traffic participant status at the corresponding sampling time are written into the safe distance anomaly table; When the lateral offset between the target vehicle trajectory and the road boundary line is greater than the lane departure threshold, the corresponding sampling time is written into the lane departure anomaly table. When the curvature change of the target vehicle's planned trajectory between two adjacent sampling times is greater than the planning mutation threshold, the corresponding sampling time is written into the planning mutation anomaly table. When the acceleration, deceleration or steering angular velocity in the control command exceeds the corresponding control threshold, the corresponding sampling time will be written into the control over-limit exception table. The safety distance anomaly table, lane departure anomaly table, planning mutation anomaly table, and control over-limit anomaly table are incorporated into the test process dataset.

6. An autonomous driving simulation testing system based on a large-model agent, applied to the autonomous driving simulation testing method based on a large-model agent as described in any one of claims 1 to 5, characterized in that, The autonomous driving simulation and testing system based on large-model agents includes: The test task parsing module is used to obtain autonomous driving test requirement text, perform test task parsing based on the autonomous driving test requirement text, and output a test task description set; wherein, the steps of obtaining autonomous driving test requirement text, performing test task parsing based on the autonomous driving test requirement text, and outputting a test task description set specifically include: The autonomous driving test requirement text is input into a preset large model Agent. The large model Agent extracts keywords from the autonomous driving test requirement text according to a preset functional verification vocabulary, risk event vocabulary, road type vocabulary, environmental condition vocabulary, and test object vocabulary to obtain a set of requirement keywords. The set of required keywords is mapped to functional items to be verified, road types, environmental conditions, traffic participant types, target risk events, expected vehicle behavior, test termination conditions, and result judgment indicators to obtain the initial test task table; The test tasks are classified according to the initial test task table, and a test task description set is generated. The simulation scenario combination module is used to perform risk scenario combination based on the test task description set and output a set of simulation scenarios to be tested; wherein, the step of performing risk scenario combination based on the test task description set and outputting the set of simulation scenarios to be tested specifically includes: Based on the road type in the test task description set, the corresponding basic road scene is called from the preset road scene library; Based on the environmental conditions and target risk events in the test task description set, weather parameters, illumination parameters, road surface adhesion parameters, and visibility parameters are written into the basic road scene to obtain an environment-enhanced road scene. Based on the traffic participant types, expected vehicle behaviors, and target risk events in the test task description set, the traffic participant behavior trajectories are bound to the environment-enhanced road scenario to obtain a candidate simulation scenario set; Based on the candidate simulation scenario set, candidate simulation scenarios are selected to generate a set of simulation scenarios to be tested; The test case orchestration module is used to perform test case orchestration preprocessing based on the set of simulation scenarios to be tested, and output a sorted set of test cases. The closed-loop simulation test module is used to perform autonomous driving closed-loop simulation tests based on the sorted test case set and output the test process dataset. The anomaly attribution iteration module is used to perform anomaly attribution and iteration control based on the test process dataset, and output a test iteration control set.

7. An autonomous driving simulation testing device based on a large-scale agent model, characterized in that, The autonomous driving simulation test device based on a large model agent includes: a memory, a processor, and an autonomous driving simulation test program based on a large model agent stored in the memory and executable on the processor. When the autonomous driving simulation test program based on a large model agent is executed by the processor, it implements an autonomous driving simulation test method based on a large model agent as described in any one of claims 1 to 5.

8. A computer program product, characterized in that, The computer program product includes an autonomous driving simulation test program based on a large model agent. When the autonomous driving simulation test program based on the large model agent is executed by the processor, it implements an autonomous driving simulation test method based on a large model agent as described in any one of claims 1 to 5.

Images