A full-process self-deduction simulation test system and construction method for autonomous driving vehicles

Abstract

The present invention relates to a full-process self-deductive simulation test system and construction method for an autonomous driving vehicle, the system comprising: an element layer for constructing dynamic elements and static elements in a simulation test environment; a relationship layer for realizing the interaction between dynamic elements to obtain strong interaction relationships and weak interaction relationships; a component layer for constructing corresponding self-deductive test components according to strong interaction relationships and weak interaction relationships; a system layer for combining multiple self-deductive test components; and a scenario layer for generating test scenarios of different categories and quantities according to the set test purpose and conditions. The method comprises: setting dynamic and static elements; setting relationships between elements; combining different components from different elements and relationships between elements to construct a self-deductive test system; and screening unknown unsafe test scenarios from the self-deductive system according to the test purpose. Compared with the prior art, the present invention can effectively improve the completeness of generated test scenarios and improve the accuracy and efficiency of simulation tests for autonomous driving vehicles.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving vehicle performance testing, and in particular to a full-process self-deductive simulation testing system and a construction method for an autonomous driving vehicle. Background Art

[0002] Self-driving cars are the inevitable direction of future development, and a scientific and complete testing and evaluation system is crucial to improving the R&D efficiency of self-driving cars and promoting industrial innovation and development. Self-driving cars are currently in the early stages of development, and their systems and performance are not yet perfect, so there are potential dangers. A comprehensive and complete self-driving test process can help discover and eliminate potential dangers.

[0003] To complete the testing of the autonomous driving system, all the algorithms, software, hardware, subsystems, and entire vehicles contained in it must be tested layer by layer to form a full chain of testing. In terms of the selection of test methods, simulation testing, field testing, and road testing together constitute the "three pillars" of autonomous driving testing. Among them, field testing and road testing are only for the entire vehicle level, and the coverage of scene conditions is limited, especially for long-tail scenes, which are difficult to test through real vehicles. The autonomous driving simulation test can well make up for the shortcomings of the actual vehicle test. In addition to the scene coverage, it can also form a full chain of testing for test objects at different levels such as autonomous driving algorithms, software, hardware, subsystems, and entire vehicles.

[0004] In the existing invention technology, the simulation test for autonomous driving vehicles mainly adopts a script-based test environment, in which the test background car is predefined. Since it does not have a self-deduction effect, it is less complex and the test scene is less complete. It lacks challenges for the autonomous driving system and it is difficult to ensure the accuracy of the test. In addition, for an autonomous driving system with predictive capabilities, this predefined background car is easily learned and predicted by the autonomous driving system, resulting in the inability to achieve efficient testing results. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a full-process self-deductive simulation test system and construction method for an autonomous driving vehicle, which can effectively improve the accuracy and efficiency of the simulation test of the autonomous driving vehicle.

[0006] The object of the present invention can be achieved by the following technical scheme: a full-process self-deductive simulation test system for an autonomous driving vehicle, comprising an element layer, a relationship layer, a component layer, a system layer and a scenario layer connected in sequence, wherein the element layer is used to construct various basic dynamic elements and static elements in a simulation test environment;

[0007] The relationship layer is used to realize the interaction between dynamic elements, and obtain strong interaction relationships and weak interaction relationships;

[0008] The component layer is used to construct corresponding self-deductive test components according to strong interaction relationships and weak interaction relationships;

[0009] The system layer is used to combine multiple self-deductive test components;

[0010] The scenario layer is used to generate different categories and quantities of test scenarios according to the set test objectives and conditions.

[0011] Furthermore, the dynamic factor is specifically a driver model, and is divided into strong factors and weak factors according to the intelligence level and driving ability;

[0012] The static elements specifically include lanes, traffic environment, and weather.

[0013] Furthermore, the strong interaction relationship specifically refers to that when multiple elements interact, there is a driving purpose or driving benefit with a similarity exceeding a set corresponding threshold, and a close interaction relationship can be formed;

[0014] The weak interaction relationship specifically refers to that multiple elements have no purpose or a relatively single purpose when interacting, and the coordination relationship between them is weak.

[0015] Furthermore, the self-deductive test components include: components composed of strong elements with strong interactive relationships, components composed of strong elements with weak interactive relationships, components composed of weak elements with strong interactive relationships, and components composed of weak elements with weak interactive relationships.

[0016] Furthermore, the test scenarios include: collision scenarios, impending collision scenarios, rule violation scenarios, and unreasonable scenarios.

[0017] A method for constructing a full-process self-deduction simulation system for an autonomous driving vehicle comprises the following steps:

[0018] S1. Design a MAOB (Motivation-Ability-Opportunity-Behavior) logical framework to construct dynamic and static elements, where the dynamic element is the driver model.

[0019] S2. Divide dynamic factors into strong factors and weak factors according to the intelligence level and driving ability;

[0020] S3. For strong elements and weak elements, strong interaction relationship processing and weak interaction relationship processing are performed respectively to determine the strong and weak relationships corresponding to the strong and weak elements respectively;

[0021] S4, based on the strong and weak relationships corresponding to the strong and weak elements, a plurality of self-deductive test components are obtained by cross-linking, and combined into a self-deductive test system;

[0022] S5. Based on the self-deductive test system, different categories and numbers of test scenarios are deduced by setting the test objectives and conditions.

[0023] Furthermore, the step S1 specifically includes the following steps:

[0024] S11, Motivation, design the interactive motivation of the driver model, including cooperative, neutral, competitive, and provocative driving purposes;

[0025] S12, Ability, on the one hand, trains the driver model based on the "Level-k" training logic and multiple machine learning algorithms, so that it is endowed with designed interactive motivations for training strong elements;

[0026] On the other hand, the rule-based setting method is used to train weak elements;

[0027] S13, Opportunity, is used to combine elements-components-into a self-deductive test system, form a simulation test environment, enable multiple types of driver models to interact, and form a trigger mechanism for showing various interactive behaviors;

[0028] S14, Behavior, the triggered interactive behaviors include driving behaviors of going straight, accelerating and decelerating, turning, and sudden acceleration and deceleration.

[0029] Furthermore, the information of the self-deductive test component in step S4 includes: element characteristics, element biological analogy, element generation method, and characteristics of the element constituting the self-deductive component;

[0030] The self-deductive test components are divided into: components composed of strong elements + strong relationships; components composed of strong elements + weak relationships; components composed of weak elements + strong relationships; components composed of weak elements + weak relationships.

[0031] Furthermore, the information contained in the component consisting of the strong element + strong relationship is specifically:

[0032] (a) Element characteristics: Individual elements have high intelligence, elements have high heterogeneity, groups have strong interactions, activities have a wide range of time and space, elements have strong execution, and multiple elements have high coordination;

[0033] (b) Biological analogies of elements: wolf pack, hyena pack;

[0034] (c) Factor generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL (Multi-Agent Reinforcement Learning), AIRL (Adversarial Inverse Reinforcement Learning), and IRL (Inverse Reinforcement Learning);

[0035] (d) Characteristics of self-deductive test components: testability and diversity;

[0036] The information contained in the components consisting of strong elements + weak relationships is specifically:

[0037] (a) Element characteristics: Individual elements have high intelligence, high heterogeneity among elements, weak interaction among groups, wide spatial and temporal range of activities, weak element execution, and low coordination among multiple elements;

[0038] (b) Element biological analogy: lion pride, leopard group;

[0039] (c) Element generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL, AIRL, and IRL;

[0040] (d) Characteristics of self-deductive test components: testability and diversity.

[0041] Furthermore, the information contained in the component consisting of the weak elements + strong relationships is specifically:

[0042] (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, strong interaction among groups, narrow spatial and temporal scope of activities, strong element execution, and high degree of coordination among multiple elements;

[0043] (b) Element biological analogies: ant colonies, bee colonies;

[0044] (c) Factor generation method: driver model generated based on DL (Deep Learning) and IL (Imitation Learning);

[0045] (d) Characteristics of self-deductive test components: adjustability and replicability;

[0046] The information contained in the components consisting of weak elements + weak relationships is specifically:

[0047] (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, weak interaction between groups, narrow spatial and temporal scope of activities, weak element execution, and low coordination among multiple elements.

[0048] (b) Biological analogy: flock of sheep, school of fish

[0049] (c) Factor generation method: driver model generated based on utility function or rules

[0050] (d) Characteristics of self-deductive test components: adjustability and replicability.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. The present invention designs the element layer, relationship layer, component layer, system layer and scenario layer connected in sequence, uses the element layer to construct various basic dynamic elements and static elements in the simulation test environment; uses the relationship layer to realize the interaction between dynamic elements to obtain strong interaction relationships and weak interaction relationships; uses the component layer to construct corresponding self-deductive test components according to strong interaction relationships and weak interaction relationships; uses the system layer to combine multiple self-deductive test components; and uses the scenario layer to generate different categories and quantities of test scenarios according to the set test purposes and conditions. Therefore, a full-process test solution from elements, relationships, components to systems and finally used for testing scenarios is proposed, which can effectively improve the completeness of generated test scenarios and the accuracy of test results.

[0053] 2. The present invention defines the framework of the self-deductive simulation test system, then sets the dynamic and static elements, sets the relationship between the elements, and then combines different self-deductive test components from different elements and the relationship between the elements, then builds a self-deductive test system based on multiple self-deductive test components, and finally selects the test scene from the self-deductive system according to the test purpose. When setting the dynamic and static elements, the MOA theory is combined from the perspective of interdisciplinary studies, and the MAOB method framework is proposed based on the theory and derived according to the traffic environment and driving behavior, which can ensure the accuracy and comprehensiveness of dynamic and static elements; when constructing the self-deductive test system, the scheme of cross-combining strong and weak relationships and strong and weak elements to form the test system is proposed by analogy with the relationship between animal groups, so that the test system can reliably meet the four performance indicators of authenticity, testability, diversity, and adjustability. The present invention generates from elements, relationships, components, systems to key scenes, and uses biological relationships for analogy, combined with the relationship from motivation to behavior, to form the effect of self-deduction of elements in the test scene.

[0054] 3. The test scenarios designed by the present invention include: collision scenarios, impending collision scenarios, rule violation scenarios, and unreasonable scenarios. When it comes to autonomous driving testing, the unknown unsafe test cases produced can effectively improve the accuracy of the test results, improve the efficiency and challenge of the test, and can test the potential dangers of autonomous driving vehicles, which is conducive to the subsequent improvement of the safety and intelligence of autonomous driving vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the system framework of the present invention;

[0056] Figure 2 It is a schematic diagram of the method flow of the present invention;

[0057] Figure 3 Generate a road topology of a test system in an embodiment;

[0058] Figure 4 It is a schematic diagram of the element training generation process based on the MAOB method framework;

[0059] Figure 5 A schematic diagram of the information contained in the self-deductive test component consisting of elements and their relationships;

[0060] Figure 6a~6d Schematic diagram of four test scenarios in the present invention;

[0061] Figure 7 This is an edge test case generated based on the self-deductive test system in the embodiment. DETAILED DESCRIPTION

[0062] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Example

[0064] like Figure 1 As shown, a full-process self-deductive simulation test system for an autonomous driving vehicle includes an element layer, a relationship layer, a component layer, a system layer, and a scenario layer connected in sequence.

[0065] Element layer: used to construct various basic dynamic and static elements in the simulation test environment. The dynamic element is specifically the driver model. This solution also defines the dynamic driver model elements as strong elements and weak elements according to the intelligence level and driving ability. Strong elements refer to models with high intelligence level and high driving ability, and weak elements refer to models with low intelligence level and low driving ability. Static elements mainly refer to other elements such as lanes, traffic environment, and weather.

[0066] Relationship layer: The interactive relationship between dynamic elements, specifically referring to strong interactive relationships and weak interactive relationships.

[0067] Component layer: A subsystem with testing purposes composed of elements and relationships between elements.

[0068] System layer: A system with testing purposes based on a simulation environment. Specifically, it means a diverse and evolving traffic environment system composed of interactively intelligent elements (driver models) and their attributes in a continuous simulation space and time, which is used to test the intelligence of autonomous vehicles.

[0069] Scenario layer: Based on the self-deductive test system, set the purpose and conditions to be tested, and deduce different categories and numbers of test scenarios.

[0070] The construction method of the above system is as follows Figure 2 As shown, the following steps are included:

[0071] S1. Design the MAOB logical framework to construct dynamic elements and static elements, where the dynamic element is the driver model;

[0072] S2. Divide dynamic factors into strong factors and weak factors according to the intelligence level and driving ability;

[0073] S3. For strong elements and weak elements, strong interaction relationship processing and weak interaction relationship processing are performed respectively to determine the strong and weak relationships corresponding to the strong and weak elements respectively;

[0074] S4. Based on the strong and weak relationships corresponding to the strong and weak elements, multiple self-deductive test components are obtained and combined into a self-deductive test system. The performance objectives that the self-deductive test system needs to meet include the following four points: authenticity, testability, diversity, and adjustability;

[0075] S5. Based on the self-deductive test system, by setting the test objectives and conditions, different categories and numbers of test scenarios are deduced. The test scenarios must include the following two objectives: (1) to test the intelligence of the autonomous driving vehicle; (2) to include unknown unsafe edge cases. The conditions are set to filter out the following four categories of unknown unsafe scenarios for testing: collision scenario, impending collision scenario, rule violation scenario, and unreasonable scenario.

[0076] Among them, in step S1, the MAOB logic framework is based on the motivation-ability-opportunity logic relationship to train the elements so that they are given features. On this theoretical basis, this solution adds a layer of behavior layer logic to enable the driver model to have the designed and trained behaviors. Specifically:

[0077] Motivation: Design the interaction motivation of the driver model, such as cooperative, neutral, competitive, provocative and other driving purposes;

[0078] Ability: Train the driver model based on "Level-k" training logic and multiple machine learning algorithms to give it the interactive motivation in S21. The various machine learning algorithms include but are not limited to MARL, GAIL (Generative Adversarial Imitation Learning) and other methods for training strong elements, and rule-based methods for training weak elements;

[0079] Opportunity: combine elements-components-into a self-deductive test system, form a simulation test environment, enable multiple types of driver models to interact, and form a trigger mechanism to show a variety of interactive behaviors;

[0080] Behavior: The interactive behaviors triggered include but are not limited to driving behaviors such as going straight, accelerating and decelerating, turning, and sudden acceleration and deceleration.

[0081] In step S3, a strong interaction relationship refers to that multiple elements have the same or similar driving purpose when interacting, and can form a close interaction relationship such as cooperation; a weak interaction relationship refers to that multiple elements have no purpose or a relatively single purpose when interacting, and there is no cooperation between them.

[0082] In step S4, the elements in the component layer contain the following information: (a) element characteristics, (b) element biological analogy, (c) element generation method, and (d) characteristics of the elements constituting self-deductive components.

[0083] And from the four categories (strong and weak elements, strong and weak relationships) in the two dimensions of elements and relationships between elements, four components are cross-combined, namely: (1) components composed of strong individuals and strong relationship elements, (2) components composed of strong individuals and weak relationship elements, (3) components composed of weak individuals and strong relationship elements, and (4) components composed of weak individuals and weak relationship elements.

[0084] Combine the 4 elements in the component with the 4 components to form the following elements and components:

[0085] 1. Components consisting of strong elements + strong relationships:

[0086] (a) Element characteristics: Individual elements have high intelligence, elements have high heterogeneity, groups have strong interactions, activities have a wide range of time and space, elements have strong execution, and multiple elements have high coordination;

[0087] (b) Biological analogies of elements: wolf packs, hyena packs, etc.;

[0088] (c) Element generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL, AIRL, and IRL;

[0089] (d) Characteristics of self-deductive test components: testability and diversity;

[0090] 2. Components composed of strong elements + weak relationships:

[0091] (a) Element characteristics: Individual elements have high intelligence, high heterogeneity among elements, weak interaction among groups, wide spatial and temporal range of activities, weak element execution, and low coordination among multiple elements;

[0092] (b) Biological analogy of elements: lion pride, leopard group, etc.;

[0093] (c) Element generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL, AIRL, and IRL;

[0094] (d) Characteristics of self-deductive test components: testability and diversity;

[0095] 3. Components composed of weak elements + strong relationships:

[0096] (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, strong interaction among groups, narrow spatial and temporal scope of activities, strong element execution, and high degree of coordination among multiple elements;

[0097] (b) Element biological analogies: ant colonies, bee colonies;

[0098] (c) Element generation method: driver model generated based on DL and IL;

[0099] (d) Characteristics of self-deductive test components: adjustability and replicability;

[0100] 4. Components composed of weak elements + weak relationships:

[0101] (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, weak interaction among groups, narrow spatial and temporal scope of activities, weak element execution, and low coordination among multiple elements;

[0102] (b) Biological analogies of elements: flock of sheep, school of fish;

[0103] (c) Factor generation method: driver model generated based on utility function or rules;

[0104] (d) Characteristics of self-deductive test components: adjustability and replicability.

[0105] This embodiment applies the above technical solution, and the main contents are:

[0106] 1. Design five dimensions: element layer, relationship layer, component layer, system layer, and use case layer.

[0107] 2. Element layer: The tester first determines the dynamic and static elements of the self-deductive test system from the element layer. In this embodiment, the dynamic element selects the four-wheel vehicle driver model, and sets the strong element with a higher intelligence level and the weak element with a lower intelligence level; the static element is set as a highway with a section of four-lane straight road (Initialization Section), a section of ramp merging (On-Ramp), a section of curve (Curve Road), and the last section of four-lane straight road (Straight Road, with a total length of about 2.3 kilometers and a lane width of 3.6m, which meets national standards, such as Figure 3 shown.

[0108] Among them, the strong factors are trained based on "level-k" logic and deep reinforcement learning (DRL) algorithm, and the weak factors are based on rule settings, usually referred to as (Intelligent Driver Model, IDM).

[0109] 3. Relationship layer: the interactive relationship between dynamic elements. Specifically refers to strong interactive relationships and weak interactive relationships. In this embodiment, it is set to a strong interactive relationship, which means that the vehicle driver model used as the background vehicle in the test environment has a high level of intelligence and can cooperate with each other to achieve a certain driving purpose during interaction. Strong elements are generated by DRL training, and strong interactive relationships are trained by "Level-k" logic.

[0110] 4. Component layer: self-deductive test component 1 composed of strong elements and strong interaction relationships, and self-deductive test component 2 composed of weak elements and weak interaction systems. The number of elements in the two test components is 10 vehicle driver models respectively.

[0111] 5. System layer: The self-deductive test system generated based on the combination of self-deductive test components 1 and 2.

[0112] 6. Use case layer (i.e., scenario layer): Based on the self-deductive test system, set the purpose to be tested and the test case screening conditions, and deduce different categories and numbers of test cases. The embodiment sets the test purpose as the intelligence of the self-driving car. The screening conditions for the test cases are: collision scenario, imminent collision scenario, rule violation scenario, and unreasonable scenario. As shown in Table 1, the collision scenario refers to the scenario where the background vehicle (NPC) collides with the system under test. It has the highest importance level and represents the scenario that the test system cannot reasonably respond to, resulting in system failure. The Near-Crash scenario is a scenario that is close to a collision but does not actually occur. It is at the edge of the dangerous parameter space and can expand the scope of the key scenario library and serve as an exploration benchmark for collision scenario generation. The violation scenario is a scenario where the behavior of the automated driving system under test (SUT) violates traffic rules. These behaviors may cause the vehicle trajectory to be unable to be reasonably predicted by other traffic participants during vehicle interaction, resulting in unsafe incidents, such as continuous lane change scenarios and lane deviation scenarios. The unreasonable behavior scenario refers to the vehicle behavior that does not violate traffic rules, but its behavior is beyond the scope of human knowledge or is difficult to predict, such as vehicle loss of control scenarios and lane change scenarios to merge into ramps.

[0113] Table 1

[0114]

[0115] Based on the above steps, a self-deductive test system was generated and four types of test scenarios were derived. According to the generation method designed in this scheme, the self-deductive test system was constructed using the neutral, competitive, and cooperative driver models generated by training to generate scenarios.

[0116] Figure 4 The figure shows a schematic diagram of the element training generation process based on the MAOB method framework. In the MAOB relationship, neutrality, competition, and cooperation express the interaction motivation (Motivation) of the driver model. The DRL algorithm combined with the "Level-k" logic trains the above interaction capabilities (Ability) of these driver models. Combining these models and relationships to generate a self-deductive test system is to provide the model with interaction opportunities (Opportunity). The driving behavior (Behavior) that occurs when the driver model interacts in the self-deductive test system generates scenarios with different risks.

[0117] Figure 5 The corresponding information is a schematic diagram of the self-deductive test component composed of the elements and their relationships in the component layer. Figure 6a~6d They correspond to four types of edge test scenario examples.

[0118] In order to express the scene intuitively, the vehicle trajectory data is visualized, and the vehicle trajectory in the traffic environment is represented by a sequence of points of different shapes and colors. The four types of scene settings are shown in Table 1. Figure 7 As shown in the figure, the four vehicles A, B, C, and D in lanes ① and ② (dynamic elements) continue to go straight. These elements are weak elements generated based on IDM and rules. The relationship between them is a weak interaction relationship. They are all going straight, and there is no strong interaction relationship such as lane change. The four vehicles E, F, G in lanes ③ and ④ and ramp ⑤, as well as the test vehicle SUT, are strong elements, and they all changed lanes. The relationship between them is a strong interaction relationship. It can be seen that because the E car first changed from lane ④ to lane ③, the automatic driving system under test (SUT) changed from lane ③ to lane ④. However, at this time, the G car just changed from lane ⑤ to lane ④, so the SUT collided with the G car, showing that the self-deductive interactive system not only has causal interpretability, but also has good testability.

[0119] In summary, this solution includes a self-deductive system consisting of five dimensions and a theoretical system based on interdisciplinary studies. A self-deductive system consists of the following six steps: 1. Define the framework of the self-deductive simulation test system; 2. Set dynamic and static elements; 3. Set the relationship between elements; 4. Combine different components from different elements and the relationship between elements; 5. Build a self-deductive test system based on the above steps; 6. Select unknown unsafe test scenarios from the self-deductive system according to the test purpose. A theoretical system is to train elements based on the Motivation-Ability-Opportunity logical relationship so that they are given features. When this solution is used for autonomous driving testing, it can effectively improve the completeness of the generated test scenarios, and the unknown unsafe test cases produced can improve the accuracy of the test results.

Claims

1. A method for constructing a full-process self-deductive simulation test system for an autonomous driving vehicle, characterized in that: The full-process self-deductive simulation test system for autonomous driving vehicles includes an element layer, a relationship layer, a component layer, a system layer, and a scenario layer connected in sequence, wherein the element layer is used to construct various basic dynamic elements and static elements in the simulation test environment; The relationship layer is used to realize the interaction between dynamic elements, and obtain strong interaction relationships and weak interaction relationships; The component layer is used to construct corresponding self-deductive test components according to strong interaction relationships and weak interaction relationships; The system layer is used to combine multiple self-deductive test components; The scenario layer is used to generate different categories and quantities of test scenarios according to the set test objectives and conditions; The dynamic factor is specifically a driver model, and is divided into strong factors and weak factors according to intelligence level and driving ability; The static elements specifically include lanes, traffic environment, and weather; The strong interaction relationship specifically refers to that when multiple elements interact, there is a driving purpose or driving benefit with a similarity exceeding a set corresponding threshold, and a close interaction relationship can be formed; The weak interaction relationship specifically refers to that multiple elements have no purpose or a single purpose when interacting, and the coordination relationship between them is weak; The self-deductive test components include: components composed of strong elements with strong interaction relationships, components composed of strong elements with weak interaction relationships, components composed of weak elements with strong interaction relationships, and components composed of weak elements with weak interaction relationships; The method for constructing the full-process self-deductive simulation test system of the autonomous driving vehicle includes the following steps: S1. Design the M motivation - A ability - O opportunity - B behavior logic framework to construct dynamic elements and static elements, where the dynamic element is the driver model; S2. Divide dynamic factors into strong factors and weak factors according to the intelligence level and driving ability; S3. For strong elements and weak elements, strong interaction relationship processing and weak interaction relationship processing are performed respectively to determine the strong and weak relationships corresponding to the strong and weak elements respectively; S4, based on the strong and weak relationships corresponding to the strong and weak elements, a plurality of self-deductive test components are obtained by cross-linking, and combined into a self-deductive test system; S5. Based on the self-deductive test system, different categories and numbers of test scenarios are deduced by setting the test objectives and conditions; The step S1 specifically includes the following steps: S11, motivation, design the interactive motivation of the driver model, including cooperative, neutral, competitive, and provocative driving purposes; S12, Capability, on the one hand, trains the driver model based on "Level-k" training logic and multiple machine learning algorithms, so that it is endowed with designed interactive motivations for training strong elements; On the other hand, the rule-based setting method is used to train weak elements; S13, opportunity, for combining elements-components-into a self-deductive test system, forming a simulation test environment, allowing multiple types of driver models to interact, and forming a trigger mechanism for expressing various interactive behaviors; S14, behavior, the triggered interactive behaviors include driving behaviors of going straight, accelerating and decelerating, turning, and rapid acceleration and deceleration; The information of the self-deductive test component in step S4 includes: element characteristics, element biological analogy, element generation method, and characteristics of the element constituting the self-deductive component; The self-deductive test components are divided into: components composed of strong elements + strong relationships; components composed of strong elements + weak relationships; components composed of weak elements + strong relationships; components composed of weak elements + weak relationships; The information contained in the components consisting of strong elements + strong relationships is specifically: (a) Element characteristics: Individual elements have high intelligence, elements have high heterogeneity, groups have strong interactions, activities have a wide range of time and space, elements have strong execution, and multiple elements have high coordination; (b) Biological analogies of elements: wolf pack, hyena pack; (c) Element generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL, AIRL, and IRL; (d) Characteristics of self-deductive test components: testability and diversity; The information contained in the components consisting of strong elements + weak relationships is specifically: (a) Element characteristics: Individual elements have high intelligence, high heterogeneity among elements, weak interaction among groups, wide spatial and temporal range of activities, weak element execution, and low coordination among multiple elements; (b) Biological analogy of elements: lion pride, leopard group; (c) Element generation method: A driver model with high intelligence and adversarial characteristics generated based on MARL, AIRL, and IRL; (d) Characteristics of self-deductive test components: testability and diversity; The information contained in the components consisting of weak elements + strong relationships is specifically: (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, strong interaction among groups, narrow spatial and temporal scope of activities, strong element execution, and high degree of coordination among multiple elements; (b) Biological analogies: ant colonies, bee colonies; (c) Factor generation method: driver model generated based on DL and IL; (d) Characteristics of components that constitute self-deductive testing: adjustability and replicability; The information contained in the components consisting of weak elements + weak relationships is specifically: (a) Element characteristics: low intelligence of individual elements, high homogeneity among elements, weak interaction among groups, narrow spatial and temporal scope of activities, weak element execution, and low coordination among multiple elements; (b) Biological analogies of elements: flock of sheep, school of fish; (c) Factor generation method: driver model generated based on utility function or rules; (d) Characteristics of self-deductive test components: adjustability and replicability.

2. The method for constructing a full-process self-deductive simulation test system for an autonomous driving vehicle according to claim 1, characterized in that: The test scenarios include: collision scenarios, impending collision scenarios, rule violation scenarios, and unreasonable scenarios.

Images