Test-oriented interactive simulation method and system for autonomous driving test

Abstract

The application relates to a kind of interactive simulation methods and systems for automatic driving test with the center of the vehicle under test, which realizes automatic driving test by jointly constructing and training vehicle under test planning agent model and background vehicle interaction model, the method comprises the following steps: constructing standardized input features based on automatic driving dataset;Construct the vehicle under test planning agent model with the minimum weighted cost optimal planning trajectory of the vehicle under test as the output with scene data as the input, wherein the weight modeling is a continuous probability distribution of learnable parameters;Construct the background vehicle interaction model with the center of the vehicle under test with the background vehicle interaction simulation trajectory as the output with scene data and the optimal planning trajectory of the vehicle under test as the input;Based on automatic driving dataset, jointly train the vehicle under test planning agent model and the background vehicle interaction model;Based on the trained vehicle under test planning agent model and the background vehicle interaction model, realize automatic driving closed-loop simulation test.The method improves the authenticity and interactivity of simulation.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving simulation testing technology, and in particular to an interactive simulation method and system centered on the vehicle under test for autonomous driving testing. Background Technology

[0002] Safety verification of autonomous vehicles (AVs) relies heavily on large-scale simulation testing, and the high reliability of these tests depends on the realism of the background traffic flow.

[0003] The core bottleneck of current simulation technology lies in modeling interactive behaviors. While existing simulation testing technologies can learn the behavior of background vehicles from large-scale real-world driving data, they are mostly based on human-to-human interaction data, lacking explicit modeling of the specific agency identity and driving strategies of autonomous vehicles (AVs). In this situation, the background vehicle in the simulation often treats the autonomous vehicle as a human-driven vehicle (HV) and responds accordingly. Furthermore, its behavioral decisions typically rely solely on the current physical state (such as position and speed) of the vehicle under test (VUT), failing to reflect the predictive judgments and causal responses of human drivers to the future intentions of AVs in real-world traffic environments. This lack of interaction significantly weakens the simulation system's ability to reproduce critical safety scenarios, limiting the realism and interactivity of the test results.

[0004] Furthermore, in data-driven modeling, training data comes from diverse sources and has varying formats, with differences between different datasets in terms of feature space, semantic definition, and sampling resolution. This multi-source data can easily lead to inconsistencies in the model's input features, thereby affecting the generalization performance of the simulation model under different testing environments.

[0005] Therefore, the current field of autonomous driving simulation testing urgently needs a new interactive simulation method and system to solve or partially solve the aforementioned problems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an interactive simulation method and system centered on the vehicle under test for autonomous driving testing. It aims to consider the background traffic flow and make causal responses based on the future planning intentions of the vehicle under test, and improve the robustness and generalization ability of the model in multi-data source scenarios through standardized input interfaces, so as to achieve more realistic and interactive autonomous driving simulation testing.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] One aspect of the present invention provides a vehicle-centric interactive simulation method for autonomous driving testing, which achieves autonomous driving testing by jointly constructing and training a planning agent model for the vehicle under test and an interaction model for the background vehicle. The method includes the following steps:

[0009] Obtain an autonomous driving dataset, and construct standardized input features based on the autonomous driving dataset;

[0010] Construct a vehicle planning proxy model that takes scene data as input and outputs the optimal planned trajectory of the vehicle under test with the minimum weighted cost. The weights are modeled as a continuous probability distribution of learnable parameters.

[0011] Construct a background vehicle interaction model centered on the test vehicle, with scene data and the optimal planned trajectory of the test vehicle as input and the background vehicle interaction simulation trajectory as output;

[0012] Based on the standardized input features of the autonomous driving dataset, the test vehicle planning agent model and the background vehicle interaction model are jointly trained.

[0013] Autonomous driving closed-loop simulation testing is achieved based on the trained vehicle planning agent model and background vehicle interaction model.

[0014] As a preferred technical solution, the vehicle-under-test planning proxy model is modeled as follows:

[0015]

[0016]

[0017] in, For the current moment Based on the baseline, simulation steps The optimal planned trajectory for the tested vehicle within the test area. For the weighted cost function, To plan the trajectory, To determine the weight vector of driving style and , Let be the number of terms in the cost function. For the preset first A cost function term, For scene state, For map information, and These are the feature parameters of a learnable probability distribution. For random variables that conform to a preset distribution, , Indicates a Gaussian distribution. A tensor representing all 1s. This represents element-wise multiplication of vectors.

[0018] As a preferred technical solution, the scene data includes map information, the status of the vehicle under test, and the status of the background vehicles.

[0019] As a preferred technical solution, the processing of the background vehicle interaction model includes the following steps:

[0020] The state of the tested vehicle and the state of the background vehicle are encoded using a long short-term memory network to obtain the initial agent features;

[0021] The map information is encoded using a multilayer perceptron to obtain initial map features;

[0022] The optimal planned trajectory of the test vehicle is encoded using a long short-term memory network to obtain the initial planned trajectory features;

[0023] The initial agent features are processed by a self-attention mechanism to model the interaction relationships between multiple traffic participants and obtain agent interaction features.

[0024] By using the cross-attention mechanism, the interaction features of the agent are used as the query vector, and the initial map features are used as the key vector and value vector for feature fusion. The interaction relationship between the agent and map elements is modeled to obtain agent-map interaction features that characterize the influence of map constraints on agent behavior.

[0025] By using the cross-attention mechanism, the interaction features of the intelligent agent are used as the query vector, and the initial planning trajectory features are used as the key vector and value vector for feature fusion. The response relationship between the background vehicle and the planning intention of the test vehicle is modeled to obtain the intelligent agent-intention interaction features.

[0026] By splicing together the agent interaction features, agent-map interaction features, and agent-intent interaction features, the multimodal distribution of the background vehicle trajectory is obtained through decoding, i.e., the background vehicle interactive simulation trajectory.

[0027] As a preferred technical solution, in the background vehicle interaction model, the agent interaction features, agent-map interaction features, and agent-intent interaction features are obtained using the following formula:

[0028]

[0029]

[0030]

[0031] in, This represents the first value obtained after fusing initial features from multiple agents through a self-attention mechanism. Agent interaction characteristics of an agent; This represents the first step obtained by modeling the interaction between the agent and map features through the cross-attention mechanism. Agent-map interaction characteristics of an agent; This represents the first step obtained by modeling the relationship between the background vehicle and the test vehicle's planning intention response through the cross-attention mechanism. Agent-intent interaction characteristics of an agent; , These represent self-attention mechanism and cross-attention mechanism, respectively. For the number of agents, For the first Initial agent characteristics of an agent , These represent the initial map features and the initial planned trajectory features, respectively. This indicates that attention is calculated using the left-hand features as the query vector and the right-hand feature set as the key and value vectors.

[0032] As a preferred technical solution, the multimodal distribution of the background vehicle trajectory is as follows:

[0033]

[0034]

[0035] in, Indicates the first Multimodal distribution of background vehicle trajectories , They represent the first The first background vehicle trajectory corresponds to the first Each candidate trajectory and its probability This represents the total number of candidate trajectories.

[0036] As a preferred technical solution, the model parameters of the test vehicle planning agent model and the background vehicle interaction model are updated by calculating the combined loss function and performing backpropagation. The combined loss function includes the background vehicle simulation trajectory loss and the test vehicle planning agent model loss.

[0037] As a preferred technical solution, the process of conducting autonomous driving tests based on the trained planning agent model of the test vehicle and the interaction model of the background vehicle includes the following steps:

[0038] In the current simulation step, the optimal planned trajectory of the test vehicle is generated using the trained test vehicle planning agent model;

[0039] The optimal planned trajectory and the acquired scene data are input into the trained background vehicle interaction model to generate the background vehicle's responsive simulation trajectory, driving the background vehicle to move in the simulation environment, thus realizing closed-loop simulation testing of autonomous driving.

[0040] As a preferred technical solution, the process of constructing standardized input features based on the autonomous driving dataset further includes:

[0041] Data parsing is performed on multi-source autonomous driving datasets, and the data is converted into a preset structured dictionary data format;

[0042] Based on the transformed dictionary data, a tensor is formed through feature extraction to obtain scene data;

[0043] Establish a local coordinate system with the current position of the vehicle under test as the origin and the orientation as the coordinate axis, and transform the historical and future states of all traffic participants in the scene as well as all map elements into a dynamic coordinate system centered on the vehicle under test;

[0044] Scene data is uniformly cropped to a fixed time length and resampled at a fixed frequency to achieve uniformity in the time series dimension;

[0045] The state of each traffic participant at each moment is uniformly represented as a standardized feature vector, which includes kinematic information;

[0046] Static elements in the map are represented as vectorized polylines consisting of a series of coordinate points, and a standardized type identifier is assigned to each type of map element.

[0047] Another aspect of the present invention provides an interactive simulation system centered on the vehicle under test for autonomous driving testing, used to implement the aforementioned interactive simulation method, the system comprising:

[0048] The shadow vehicle planning module is equipped with a vehicle planning proxy model that takes scene data as input and outputs the optimal planned trajectory of the vehicle with the minimum weighted cost.

[0049] The interactive module centered on the vehicle under test is equipped with a background vehicle interactive model that takes scene data and the optimal planned trajectory of the vehicle under test as input and the simulated trajectory of the background vehicle as output.

[0050] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0051] (1) Significantly improves the realism and interactivity of the simulation: This invention uses the future planned trajectory of the test vehicle as the conditional input of the background vehicle model, enabling the background vehicle to make predictive responses based on the test vehicle's intentions, rather than just passively reacting to its current state. This achieves causal interaction that is more in line with real human driving logic, effectively improving the simulation test's ability to reproduce complex interactive scenarios and detect risks.

[0052] (2) Explicit modeling of multi-level traffic interaction relationships: This invention introduces self-attention mechanism and cross-attention mechanism in the background vehicle interaction model. The self-attention mechanism is used to model the mutual influence relationship between multiple traffic participants, and the cross-attention mechanism is used to model the interaction relationship between traffic participants and map elements, and between traffic participants and the planning intention of the test vehicle, respectively. This enables explicit modeling of multi-level interaction relationships such as agent-agent, agent-map, and agent-test vehicle intention, and improves the model's ability to express interaction dependencies in complex traffic environments.

[0053] (3) Support for diverse vehicle behavior modeling: By modeling the cost function weights as a learnable probability distribution, the vehicle planning agent model can generate diverse trajectories with different driving styles, providing more targeted conditional inputs for the interaction model, which helps to test the safety and robustness of the autonomous driving system under different driving strategies.

[0054] (4) Good generalization and adaptability: This invention achieves compatible training under multiple data source conditions through standardized input features and a unified data interface, improving the model's transferability and stability in different scenarios and simulation platforms. This feature, as a supplementary design to this method, effectively reduces the impact of structural and semantic differences between different data sources on model performance, providing a key guarantee for actual deployment and expansion. Attached Figure Description

[0055] Figure 1 This is a flowchart of an interactive simulation method centered on the vehicle under test for autonomous driving testing, as described in the embodiment.

[0056] Figure 2 This is a schematic diagram of the interactive simulation in the embodiment;

[0057] Figure 3 This is a visualization of the VUT driving style weight distribution for each dataset in the example;

[0058] Figure 4 This is a schematic diagram illustrating the results of the interaction analysis in the example;

[0059] Figure 5 This is a schematic diagram illustrating the results of the background vehicle responding differently to different VUT driving styles in the embodiment.

[0060] Figure 6 This is a case study illustration of a high-risk mismatch scenario in the examples. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0062] Example 1

[0063] To address the aforementioned problems in existing technologies, this embodiment provides a vehicle-under-test-centered interactive simulation method for autonomous driving testing. This method aims to overcome the lack of causal response capability of background traffic flow to the future intentions of the vehicle under test in existing autonomous driving simulation technologies. To this end, this embodiment provides a vehicle-under-test-centered interactive simulation (VCIS) method for autonomous driving testing. It jointly constructs and trains a planning agent model for the vehicle under test and an interaction model for the background vehicle. The planning agent model for the vehicle under test simulates the planning logic of the vehicle under test, generating diverse future driving trajectories as an explicit expression of its future intentions. The interaction model for the background vehicle uses this future trajectory as a key, causal conditional input, thereby driving the background vehicle to produce response behaviors that more closely resemble real-world driving logic.

[0064] See Figure 1 and Figure 2 This method includes the following steps:

[0065] Step S1: Standardize and unify scene data.

[0066] This step involves structured parsing and feature standardization of the autonomous driving dataset to ensure consistency and compatibility of model inputs. To ensure the core model of this method possesses strong generalization ability and robustness, this step designs a standardized data input and feature unification mechanism, enabling datasets from different sources to have a consistent representation, achieving standardization of heterogeneous datasets, improving stability and generalization ability under multi-source conditions, and providing data support and a foundation for the implementation of interactive simulation methods.

[0067] Specifically, this step includes the following sub-steps:

[0068] S11: Unified Data Structure. First, raw driving data from different sources is parsed and converted into the industry-standard ScenarioNet structured dictionary format, which is based on scenarios. Then, this dictionary-formatted data is further processed to extract the required information and convert it into a tensor format suitable for deep learning model input.

[0069] S12: Feature Unification and Coordinate Transformation. Based on a unified data structure, a comprehensive standardization of spatiotemporal, dynamic, and static data features within the scene is performed, specifically including:

[0070] Coordinate unification: Establish a local coordinate system with the current position of the VUT as the origin and the orientation as the x-axis, and transform the historical and future states of all traffic participants in the scene, as well as all map elements, into a dynamic coordinate system centered on the VUT.

[0071] Spatiotemporal uniformity: All scene data are uniformly cropped to a fixed time length (e.g., 5s, 7s, 10s) and resampled at a fixed frequency (e.g., 5Hz, 10Hz, 20Hz) to ensure that the temporal dimension of the model input is consistent.

[0072] Unified agent features: The state of each traffic participant at each moment is uniformly represented as a standardized feature vector, which contains core kinematic information such as its position, speed, heading angle, and size.

[0073] Map feature unification: Static elements such as lane lines in high-precision maps are uniformly represented as vectorized polylines composed of a series of coordinate points, and a standardized type identifier is assigned to each type of map element.

[0074] Through the above processing, datasets from different sources can be input into the same simulation framework without redesigning the model structure, thereby supporting model training and validation across data sources.

[0075] Step S2: Construct the planning agent model for the vehicle under test.

[0076] The vehicle-under-test planning proxy model refers to a functional model that can simulate the planning logic of the vehicle under test and output the future driving trajectory instead of the real vehicle under test. Since the future behavior of the real vehicle under test cannot be obtained as the input of the background vehicle model during the simulation (inference) stage, the vehicle-under-test planning proxy model is used to simulate the planning behavior of the vehicle under test. The future planning trajectory it generates serves as a real and reasonable reference for the background traffic flow to the future intention of the vehicle under test.

[0077] Specifically, step S2 includes the following sub-steps:

[0078] S21: Construct a VUT planning surrogate model based on a differentiable optimization algorithm, which finds an optimal trajectory. To minimize a weighted cost function :

[0079]

[0080] in, This is a pre-defined cost function term used to quantify different evaluation dimensions of the trajectory; It is a weighted vector that determines driving style. For scene state, This is map information.

[0081] S22: To achieve diverse driving styles, the weight vector will be... Model it as a continuous probability distribution with learnable parameters (such as Gaussian distribution, Beta distribution, normal distribution, etc.):

[0082] in

[0083] in, and These are the characteristic parameters of the probability distribution. These are random variables that conform to a preset distribution.

[0084] S23: By learning the feature parameters of a continuous probability distribution during training. and This allows for the generation of diverse VUT future planning trajectories with specific styles by sampling the distribution during simulation.

[0085] Step S3: Construct a background vehicle interaction model centered on the vehicle under test.

[0086] The background vehicle interaction model uses the future planned trajectory of the test vehicle as the core condition to generate the responsive simulation trajectory of the background vehicle.

[0087] Specifically, step S3 includes the following sub-steps:

[0088] The background vehicle interaction model aims to solve the following problem: given standardized scene data, including map information ,all The state of each agent And VUT's future planning trajectory Generate all The future simulation trajectory of the background car : .

[0089] S31: Construct a Transformer-based encoder-decoder architecture as the background vehicle interaction model, which receives standardized data and the VUT's future planned trajectory as input;

[0090] S32: Encode scene information by using a Long Short-Term Memory (LSTM) network to encode the historical trajectories of all agents and a Multilayer Perceptron (MLP) to encode vectorized map elements, thereby obtaining agent features and map features respectively;

[0091] S33: Modeling scene interaction relationships, firstly, the self-attention mechanism is used to process the agent state feature set to capture the interaction between agents:

[0092]

[0093] Then, the agent features and map features are fused through a cross-attention mechanism to incorporate static environmental information:

[0094]

[0095] S34: Inject the future intent of the VUT as the center, encode the future planned trajectory of the VUT as planning features, and again use the cross-attention mechanism to inject the information of the planning features into the features of each background vehicle, to obtain the final features that integrate the VUT's intent:

[0096]

[0097] S35: Decode the final features to generate a multimodal distribution of the future simulation trajectory of each background vehicle, which is generated by... Composition of candidate trajectories and their corresponding probabilities:

[0098]

[0099] Step S4: Jointly train the planning agent model of the vehicle under test and the interaction model of the background vehicle.

[0100] Training is performed in an end-to-end manner. Specifically, step S4 includes the following sub-steps:

[0101] S41: Design a combination loss function This is used for collaborative optimization of two models; among them, The simulation trajectory loss of the background vehicle is used to optimize the interaction model of the background vehicle; The loss parameter for the VUT planning surrogate model is used to optimize the weight distribution parameters of the VUT planning surrogate model. and This is to achieve cost function weight updates.

[0102] S42: Calculate the gradient of the combined loss function with respect to all model parameters using the backpropagation algorithm, and then update it synchronously using the optimizer.

[0103] Step S5: Deploy the trained model in the simulation environment and perform an interactive simulation test centered on the vehicle under test.

[0104] S51: In each simulation step, the future planning trajectory of the VUT is first generated by the VUT planning agent model;

[0105] S52: Then input the planned trajectory and current scene information into the background vehicle interaction model;

[0106] S53: Finally, the interactive model generates responsive simulation trajectories for all background vehicles and drives them to move in the simulation environment.

[0107] The following example illustrates this method.

[0108] First, step S1 is performed to standardize and unify the features of the autonomous driving scenario data. In this embodiment, three mainstream large-scale autonomous driving datasets in the industry are selected for processing: the Waymo dataset, the Argoverse 2 dataset, and the nuPlan dataset. After standardization, the data enters the same training process with a unified structure, requiring no additional adaptation. This setting is used to verify the robustness and scenario adaptability of the method under diverse traffic environments and driving styles.

[0109] Next, steps S2 and S3 are executed to construct and train the VUT planning agent model and the background vehicle interaction model. In this embodiment, both the VUT planning agent model and the background vehicle interaction model adopt a deep neural network architecture. Specifically, the recurrent neural network used to encode the trajectory sequence is an LSTM, and the neural network used to encode the map and perform feature fusion is a Transformer network containing MLP and attention mechanisms. After training, the planned trajectory generated by the VUT planning agent model has a sufficiently small error compared to the real VUT trajectory in the dataset, ensuring the realism of its simulated planning behavior, thereby providing a highly reliable conditional input for the background vehicle interaction model.

[0110] Next, step S4 is executed to perform end-to-end joint training of the VUT planning agent model and the background vehicle interaction model. The Adam optimizer is used during training, and the weight coefficients of each part of the loss function are fine-tuned experimentally to obtain the best performance.

[0111] Finally, proceed to step S5 to deploy the model and perform interactive simulation testing and evaluation.

[0112] To verify the beneficial effects of this method, three sets of comparative models were designed for evaluation in this embodiment:

[0113] 1) State Condition Baseline Model (SCB): This model uses only the current state of the VUT as input when generating the background vehicle simulation trajectory, without considering its future planning.

[0114] 2) Method of this embodiment (fixed weight, VCIS-Fixed): This model uses the complete framework of this embodiment, but the cost function weights in its VUT planning agent model are a fixed set of deterministic parameters.

[0115] 3) Method of this embodiment (probability weight, VCIS-Dist): The model uses the complete framework of this embodiment, and the cost function weights in its VUT planning agent model are sampled from the learned probability distribution (that is, the continuous probability distribution in S22 is modeled as a multivariate Gaussian distribution).

[0116] To verify the robustness and generalization performance of the proposed method under different traffic scenarios and driving styles, this embodiment tested the performance of a single jointly trained model on various typical scenario data. As shown in Table 1, the model maintained high simulation accuracy in different types of test data, indicating that the proposed method can stably adapt to diverse traffic environments.

[0117] Table 1. Simulation and planning errors of a single model on different datasets.

[0118]

[0119] Furthermore, to analyze the model's ability to characterize different driving styles, the weight distribution of the cost function learned by the model was visualized, and the results are as follows: Figure 3 As shown in the figure, the differences in weight distributions indicate that the model can capture and distinguish various driving style features, thereby generating differentiated vehicle planning behaviors in simulation. This result demonstrates that the proposed method possesses good scene adaptability and generalization ability while ensuring the realism of the interaction.

[0120] To verify the improved realism and interactivity brought about by introducing VUT future behavior (planned trajectory) in step S3, this embodiment conducted a comparative analysis. Regarding simulation realism, as shown in Table 2, the simulation error of this method (average displacement error ADE of 0.6592 meters) is significantly lower than that of the baseline model (ADE of 0.6877 meters). Regarding interactivity, as... Figure 4As shown in the figure, the box plot illustrates the 2-Wasserstein distance between the background vehicle trajectory distribution generated by the method of this invention (VCIS-Dist) and the state-based baseline model (SCB). The method of this invention has a higher average distance metric, indicating that it can generate more interactive background vehicle behavior. These results collectively demonstrate that the technical solution of this embodiment can simultaneously improve the interactivity and realism of the simulation.

[0121] Table 2 Comparison of simulation and planning errors for different models

[0122]

[0123] To further verify that the background vehicle interaction model constructed in step S3 can respond differently to different driving styles, this embodiment conducted a specific test, and the results are as follows: Figure 5 As shown. Figure 5 The figure shows the 2-Wasserstein distance between the background vehicle trajectory response distribution and the baseline model response distribution when the VUT plans using different driving styles learned from three datasets. The figure clearly shows that the response distributions induced by the three different input styles are significantly separated. This result demonstrates that the background vehicle interaction model of our method can sensitively identify the driving style of the VUT and produce statistically distinguishable and matching interaction behaviors.

[0124] To verify that this method can serve as an effective security verification tool for discovering and quantifying security risks caused by cognitive mismatch, this embodiment designed and analyzed a high-risk intersection scenario during step S5, and the results are as follows. Figure 6 As shown, (a) represents a scenario where the expected behavior matches the actual behavior, while (b) and (c) represent scenarios where the expected behavior does not match the actual behavior. In this scenario, the background vehicle learns an expectation from historical data: the VUT will adopt a conservative strategy when going straight through the intersection, slowing down to yield to vehicles turning left. Figure 6 (a) demonstrates that when the VUT's actual behavior matches this expectation (both in Waymo style), the two parties interact safely without collision. However, Figure 6 (b) and Figure 6 (c) This demonstrates that when the actual behavior of the VUT deviated from expectations, adopting the more aggressive Argoverse 2 style and nuPlan style respectively, the VUT failed to slow down and yield, resulting in collisions with the background vehicle turning left at 3.6s and 4.6s respectively. This case study strongly proves that the simulation method of this embodiment can reproduce safety-critical events caused by the mismatch between expected and actual driving behavior, providing an important means for the safety assessment of autonomous driving systems.

[0125] Example 2

[0126] Building upon Example 1, this example provides an interactive simulation system centered on the vehicle under test for autonomous driving testing, used to implement the interactive simulation method of Example 1. (See also...) Figure 2 The system includes:

[0127] (1) The shadow vehicle planning module is equipped with a vehicle planning proxy model that takes scene data as input and outputs the optimal planning trajectory of the vehicle with the minimum weighted cost.

[0128] (2) The interactive module centered on the vehicle under test is equipped with a background vehicle interactive model centered on the vehicle under test, which takes scene data and the optimal planned trajectory of the vehicle under test as input and the simulated trajectory of the background vehicle as output.

[0129] This invention significantly improves the realism and interactivity of the simulation by incorporating the future intentions of the vehicle under test into the background vehicle response model; its unified data input process further ensures the high generalization ability of the model under multi-source data training, and ultimately provides a more adaptable simulation tool for the safety verification of autonomous driving systems.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A vehicle-centric interactive simulation method for autonomous driving testing, characterized in that, Autonomous driving testing is achieved by jointly constructing and training a planning agent model for the vehicle under test and an interaction model for the background vehicle. The method includes the following steps: Obtain an autonomous driving dataset, and construct standardized input features based on the autonomous driving dataset; Construct a vehicle planning proxy model that takes scene data as input and outputs the optimal planned trajectory of the vehicle under test with the minimum weighted cost. The weights are modeled as a continuous probability distribution of learnable parameters. Construct a background vehicle interaction model centered on the test vehicle, with scene data and the optimal planned trajectory of the test vehicle as input and the background vehicle interaction simulation trajectory as output; Based on the autonomous driving dataset, standardized input features are constructed, and the planning agent model of the test vehicle and the interaction model of the background vehicle are jointly trained. Autonomous driving closed-loop simulation testing is achieved based on the trained test vehicle planning agent model and background vehicle interaction model. The proposed vehicle planning agent model is modeled as follows: ， ， in, For the current moment Based on the baseline, simulation steps The optimal planned trajectory for the tested vehicle within the test area. For the weighted cost function, To plan the trajectory, To determine the weight vector of driving style and , Let be the number of terms in the cost function. For the preset first A cost function term, For scene state, For map information, and These are the feature parameters of a learnable probability distribution. For random variables that conform to a preset distribution, , Indicates a Gaussian distribution. A tensor representing all 1s. This represents element-wise multiplication of vectors. The processing steps of the background vehicle interaction model include the following: The state of the tested vehicle and the state of the background vehicle are encoded using a long short-term memory network to obtain the initial agent features; The map information is encoded using a multilayer perceptron to obtain initial map features; The optimal planned trajectory of the test vehicle is encoded using a long short-term memory network to obtain the initial planned trajectory features; The initial agent features are processed by a self-attention mechanism to model the interaction relationships between multiple traffic participants and obtain agent interaction features. By using the cross-attention mechanism, the interaction features of the agent are used as the query vector, and the initial map features are used as the key vector and value vector for feature fusion. The interaction relationship between the agent and map elements is modeled to obtain agent-map interaction features that characterize the influence of map constraints on agent behavior. By using the cross-attention mechanism, the interaction features of the intelligent agent are used as the query vector, and the initial planning trajectory features are used as the key vector and value vector for feature fusion. The response relationship between the background vehicle and the planning intention of the test vehicle is modeled to obtain the intelligent agent-intention interaction features. By splicing together the agent interaction features, agent-map interaction features, and agent-intent interaction features, the multimodal distribution of the background vehicle trajectory is obtained through decoding, i.e., the background vehicle interactive simulation trajectory.

2. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, The scene data includes map information, the status of the vehicle being tested, and the status of the background vehicles.

3. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, In the background vehicle interaction model, the agent interaction features, agent-map interaction features, and agent-intent interaction features are obtained using the following formula: ， ， ， in, This represents the first [aspect] obtained by fusing initial features from multiple agents through a self-attention mechanism. Agent interaction characteristics of an agent; This represents the first step obtained by modeling the interaction between the agent and map features through the cross-attention mechanism. Agent-map interaction characteristics of an agent; This represents the first step obtained by modeling the relationship between the background vehicle and the test vehicle's planning intention response through the cross-attention mechanism. Agent-intent interaction characteristics of an agent; , These represent self-attention mechanism and cross-attention mechanism, respectively. For the number of agents, For the first Initial agent characteristics of an agent , These represent the initial map features and the initial planned trajectory features, respectively. This indicates that attention is calculated using the left-hand features as the query vector and the right-hand feature set as the key and value vectors.

4. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, The multimodal distribution of the background vehicle trajectory is as follows: ， ， in, Indicates the first Multimodal distribution of background vehicle trajectories , They represent the first The first background vehicle trajectory corresponds to the first Each candidate trajectory and its probability This represents the total number of candidate trajectories.

5. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, The model parameters of the test vehicle planning agent model and the background vehicle interaction model are updated by calculating the combined loss function and performing backpropagation. The combined loss function includes the background vehicle simulation trajectory loss and the test vehicle planning agent model loss.

6. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, The process of conducting autonomous driving tests based on the trained vehicle-under-test planning agent model and the background vehicle interaction model includes the following steps: In the current simulation step, the optimal planned trajectory of the test vehicle is generated using the trained test vehicle planning agent model; The optimal planned trajectory and the acquired scene data are input into the trained background vehicle interaction model to generate the background vehicle's responsive simulation trajectory, driving the background vehicle to move in the simulation environment, thus realizing closed-loop simulation testing of autonomous driving.

7. The interactive simulation method centered on the test vehicle for autonomous driving testing according to claim 1, characterized in that, The process of constructing standardized input features based on the autonomous driving dataset also includes: Data parsing is performed on multi-source autonomous driving datasets, and the data is converted into a preset structured dictionary data format; Based on the transformed dictionary data, a tensor is formed through feature extraction to obtain scene data; Establish a local coordinate system with the current position of the vehicle under test as the origin and the orientation as the coordinate axis, and transform the historical and future states of all traffic participants in the scene as well as all map elements into a dynamic coordinate system centered on the vehicle under test; Scene data is uniformly cropped to a fixed time length and resampled at a fixed frequency to achieve uniformity in the temporal dimension; The state of each traffic participant at each moment is uniformly represented as a standardized feature vector, which includes kinematic information; Static elements in the map are represented as vectorized polylines consisting of a series of coordinate points, and a standardized type identifier is assigned to each type of map element.

8. An interactive simulation system centered on the test vehicle for autonomous driving testing, characterized in that, The system for implementing the interactive simulation method as described in any one of claims 1-7 includes: The shadow vehicle planning module is equipped with a vehicle planning proxy model that takes scene data as input and outputs the optimal planned trajectory of the vehicle with the minimum weighted cost. The interactive module centered on the vehicle under test is equipped with a background vehicle interactive model that takes scene data and the optimal planned trajectory of the vehicle under test as input and the simulated trajectory of the background vehicle as output.

Images