Interactive simulation method and system oriented to automatic driving test and centered on tested vehicle

Abstract

The invention relates to an interactive simulation method and system with a tested vehicle as the center for an automatic driving test, the automatic driving test is realized by jointly constructing and training a tested vehicle planning agent model and a background vehicle interaction model, and the method comprises the following steps: constructing a standardized input feature based on an automatic driving data set; a measured vehicle planning agent model with the scene data as input and the measured vehicle optimal planning track with the minimum weighting cost as output is constructed, and a weight reconstruction model is continuous probability distribution of learnable parameters; constructing a background vehicle interaction model which takes the scene data and the optimal planning track of the tested vehicle as input and takes the background vehicle interaction simulation track as output and takes the tested vehicle as the center; based on the automatic driving data set, carrying out joint training on the tested vehicle planning agent model and the background vehicle interaction model; and realizing an automatic driving closed-loop simulation test based on the trained tested vehicle planning agent model and the background vehicle interaction model. According to the method, the authenticity and interactivity of simulation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving simulation testing, in particular to an interactive simulation method and system for automatic driving testing with a vehicle under test as the center. BACKGROUND

[0002] The safety verification of autonomous vehicles (AVs) largely depends on large-scale simulation testing, and the high credibility of the test depends on the authenticity of the background traffic flow.

[0003] The core bottleneck of current simulation technology lies in the modeling of interactive behavior. Although existing simulation testing technology can learn the behavior of background vehicles from large-scale real driving data, it is mostly based on "person-person" interaction data, and the special subject identity and driving strategy of AVs are not explicitly modeled. In this case, the background vehicles in the simulation often respond to the AVs as human-driven vehicles (HVs), and their behavior decisions usually only depend on the current physical state (such as position, speed, etc.) of the vehicle under test (VUT), and it is difficult to reflect the predictive judgment and causal response of human drivers to the future intentions of AVs in real traffic environments. The lack of this interaction mode significantly weakens the ability of the simulation system to reproduce key safety scenarios, limiting the authenticity and interactivity of the test results.

[0004] In addition, in the data-driven modeling process, the training data comes from various sources and has different formats, and there are differences between different data sets in terms of feature space, semantic definition, and sampling resolution. Such multi-source data can cause inconsistencies in model input features, which in turn affects the generalization performance of the simulation model in different test environments.

[0005] Therefore, there is an urgent need in the current automatic driving simulation testing field for a new interactive simulation method and system to solve or partially solve the aforementioned problems. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art and provide an interactive simulation method and system for automatic driving testing with a vehicle under test as the center, which aims to consider the causal response of the background traffic flow based on the future planning intentions of the vehicle under test, and improve the robustness and generalization ability of the model in the multi-data source scenario through a standardized input interface, to realize more realistic and interactive automatic driving simulation testing.

[0007] The purpose of the present application can be achieved by the following technical solutions: 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: 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 standardized input features of the autonomous driving dataset, the test vehicle planning agent model and the background vehicle interaction model are jointly trained. Autonomous driving closed-loop simulation testing is achieved based on the trained vehicle planning agent model and background vehicle interaction model.

[0008] As a preferred technical solution, the vehicle-under-test planning proxy model is modeled as follows:

[0009]

[0010] 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.

[0011] 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.

[0012] As a preferred technical solution, the processing process of the background vehicle interaction model comprises the following steps: Encode the state of the measured vehicle and the state of the background vehicle using a long short-term memory network to obtain initial agent features; Encode the map information using a multi-layer perception machine to obtain initial map features; Encode the input optimal planning trajectory of the measured vehicle using a long short-term memory network to obtain initial planning trajectory features; Process the initial agent features through a self-attention mechanism to model the interaction relationship between multiple traffic participants to obtain agent interaction features; Through a cross-attention mechanism, the agent interaction features are taken as query vectors, and the initial map features are taken as key vectors and value vectors for feature fusion to model the interaction relationship between the agent and the map elements, and obtain agent-map interaction features representing the influence of map constraints on the agent behavior; Through a cross-attention mechanism, the agent interaction features are taken as query vectors, and the initial planning trajectory features are taken as key vectors and value vectors for feature fusion to model the response relationship of the background vehicle to the planning intention of the measured vehicle, and obtain agent-intention interaction features; Concatenate the agent interaction features, the agent-map interaction features, and the agent-intention interaction features, and obtain the multi-modal distribution of the background vehicle trajectory through decoding, i.e., the background vehicle interactive simulation trajectory.

[0013] As a preferred technical solution, in the background vehicle interaction model, the agent interaction features, the agent-map interaction features, and the agent-intention interaction features are obtained by the following formula:

[0014]

[0015]

[0016] wherein, represents the agent interaction features of the i-th agent obtained by fusing the initial features of multiple agents through a self-attention mechanism, represents the agent-map interaction features of the i-th agent obtained by modeling the interaction relationship between the agent and the map elements through a cross-attention mechanism, represents the agent-intention interaction features of the i-th agent obtained by modeling the response relationship of the background vehicle to the planning intention of the measured vehicle through a cross-attention mechanism, , ​​​respectively represent self-attention mechanism, cross-attention mechanism, is the number of agents, is the initial agent feature of the th agent, , respectively represent initial map features, initial planning trajectory features, representing attention calculation with left feature as query vector and right feature set as key vector and value vector.

[0017] As a preferred technical solution, the multi-modal distribution of the background vehicle trajectory is:

[0018]

[0019] wherein, is the multi-modal distribution of the th background vehicle trajectory, , respectively represent the th candidate trajectory corresponding to the th background vehicle trajectory, probability, is the total number of candidate trajectories.

[0020] As a preferred technical solution, the model parameter update of the measured vehicle planning agent model and the background vehicle interaction model is realized by calculating the combined loss function and performing back propagation, wherein the combined loss function includes background vehicle simulation trajectory loss and measured vehicle planning agent model loss.

[0021] As a preferred technical solution, the process of automatic driving test based on the trained measured vehicle planning agent model and background vehicle interaction model includes the following steps: In the current simulation step, the optimal planning trajectory of the measured vehicle is generated by using the trained measured vehicle planning agent model; The optimal planning trajectory and the obtained scene data are input into the trained background vehicle interaction model to generate a responsive simulation trajectory of the background vehicle, drive the background vehicle to move in the simulation environment, and realize automatic driving closed-loop simulation test.

[0022] As a preferred technical solution, in the process of constructing standardized input features based on the automatic driving dataset, it further includes: Data parsing and conversion to a preset structured dictionary data format are performed on the multi-source automatic driving dataset; Based on the converted dictionary data, a tensor is formed by feature extraction to obtain scene data; A local coordinate system is established with the current position of the measured vehicle as the origin and the direction as the coordinate axis, and the historical and future states of all traffic participants in the scene and all map elements are converted to the dynamic coordinate system centered on the measured vehicle; The scene data is uniformly cropped to a fixed time length and resampled at a fixed frequency to unify the time dimension; The state of each traffic participant at each time is uniformly represented as a standardized feature vector, which includes kinematic information; The static elements in the map are represented as a vectorized polyline composed of a series of coordinate points, and a standardized type identifier is assigned to each type of map element.

[0023] Another aspect of the present application provides an interactive simulation system centered on the measured vehicle for automatic driving test, which is used to implement the aforementioned interactive simulation method, and the system comprises: A shadow measured vehicle planning module is deployed with a measured vehicle planning agent model taking scene data as input and the optimal planning trajectory of the measured vehicle with the minimum weighted cost as output; An interactive module centered on the measured vehicle is deployed with a background vehicle interaction model centered on the measured vehicle taking scene data and the optimal planning trajectory of the measured vehicle as input and the simulation trajectory of the background vehicle as output.

[0024] Compared with the prior art, the present application has at least one of the following beneficial effects: (1) Significantly improve the authenticity and interactivity of simulation: the present application makes the background vehicle able to make predictive response based on the intention of the measured vehicle by taking the future planning trajectory of the measured vehicle as the conditional input of the background vehicle model, instead of only making passive reaction to its current state, thereby realizing the cause-and-effect interaction more in line with the real human driving logic, and effectively improving the restoration ability and risk discovery ability of simulation test for complex interaction scenarios.

[0025] (2) Explicit modeling of multi-level traffic interaction relationship: the present application 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 planning intention of the measured vehicle, thereby realizing explicit modeling of multi-level interaction relationship such as agent-agent, agent-map and agent-planning intention of the measured vehicle, and improving the expression ability of the model for interaction dependency relationship in complex traffic environment.

[0026] (3) Support diversified vehicle-under-test behavior modeling: By modeling the cost function weight as a learnable probability distribution, the vehicle-under-test planning agent model can generate diversified 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.

[0027] (4) Good generalization and adaptability: By standardizing the input features and unifying the data interface, the invention realizes compatible training under multiple data sources, improving the model's migration and stability in different scenarios and simulation platforms. As a complementary design of the method, this feature 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. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 Flowchart of the vehicle-under-test centered interactive simulation method for autonomous driving testing in the embodiment; Figure 2 Schematic diagram of interactive simulation in the embodiment; Figure 3 Visualization result of VUT driving style weight distribution of each data set in the embodiment; Figure 4 Schematic diagram of the results of interactive analysis in the embodiment; Figure 5 Schematic diagram of the results of the background vehicle responding differently to different VUT driving styles in the embodiment; Figure 6 Case analysis schematic diagram of high-risk mismatch scenarios in the embodiment. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should be within the scope of protection of the present application.

[0030] Embodiment 1 To solve the problems in the prior art, the embodiment provides an interactive simulation method for automatic driving test and centered on a vehicle under test, which aims to overcome the problem that in the prior art of automatic driving simulation, the background traffic flow lacks the ability to respond to the future intention of the vehicle under test in a causal manner. To this end, the embodiment provides a Vehicle-Under-Test-Centered Interactive Simulation (VCIS) method for automatic driving test, which jointly constructs and trains a vehicle under test planning agent model and a background vehicle interaction model. The vehicle under test planning agent model is used to simulate the planning logic of the vehicle under test and generate diversified future driving trajectories of the vehicle under test as an explicit expression of the future intention of the vehicle under test. The background vehicle interaction model takes the future trajectory as a key causal input, thereby driving the background vehicles to generate a response behavior that is closer to the driving logic in the real world.

[0031] Referring to Figure 1 and Figure 2 The method comprises the following steps: Step S1: scene data standardization and feature unification.

[0032] This step is used for structured analysis and feature standardization processing of the automatic driving data set, so as to ensure the consistency and compatibility of the model input. In order to ensure that the core model of the method has strong generalization ability and robustness, the step designs a standardized data input and feature unification mechanism, so that data sets of different sources have consistent representation forms, realizes the standardization of heterogeneous data sets, improves the stability and generalization ability under multi-source conditions, and provides data support and implementation basis for the implementation of the interactive simulation method.

[0033] Specifically, the step comprises the following sub-steps: S11: uniform data structure. First, the original driving data of different sources is parsed and converted into the ScenarioNet structured dictionary format which is commonly used in the industry and is in units of scenes. Then, the data in the dictionary format is further processed, the required information is extracted, and the tensor (Tensor) format suitable for the input of the deep learning model is converted.

[0034] S12: feature unification and coordinate transformation. On the basis of uniform data structure, the space-time, dynamic and static data features in the scene are comprehensively regularized, specifically including: Coordinate unification: a local coordinate system with the current position of the VUT as the origin and the direction as the x-axis is established, and all historical and future states of all traffic participants in the scene and all map elements are converted into the dynamic coordinate system centered on the VUT.

[0035] Temporal and spatial unification: all scene data is 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 the consistency of the time sequence dimension of the model input.

[0036] Agent feature unification: the state of each traffic participant at each time is uniformly represented as a standardized feature vector, which contains its position, speed, heading angle, size, and other core kinematic information.

[0037] 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 each type of map element is assigned a standardized type identifier.

[0038] Through the above processing, datasets of different sources can be input into the same simulation framework without the need to redesign the model structure, thereby supporting model training and verification across data sources.

[0039] Step S2: Construct a planning agent model of the vehicle under test.

[0040] The planning agent model of the vehicle under test refers to a functional model that can simulate the planning logic of the vehicle under test and output future trajectories instead of the real vehicle under test. Since the future behavior of the real vehicle under test cannot be obtained as input for the background vehicle model during simulation (reasoning), the planning agent model of the vehicle under test is used to simulate the planning behavior of the vehicle under test, and the future planning trajectory generated by the planning agent model is used as a realistic and reasonable reference for the future intention of the vehicle under test.

[0041] Specifically, step S2 includes the following sub-steps: S21: Construct a VUT planning agent model based on a differentiable optimization algorithm, which finds an optimal trajectory to minimize a weighted cost function :

[0042] wherein is a preset cost function term for quantifying different evaluation dimensions of the trajectory; is a weight vector that determines the driving style, is the scene state, is the map information.

[0043] S22: To achieve diversified driving styles, the weight vector is modeled as a continuous probability distribution (such as Gaussian distribution, Beta distribution, normal distribution, etc.) of learnable parameters: wherein

[0044] wherein, and are characteristic parameters of a probability distribution, is a random variable conforming to the preset distribution.

[0045] S23: learning the characteristic parameters of the continuous probability distribution through the training process and so that during simulation, the diversified VUT future planning trajectory with a specific style can be generated by sampling the distribution.

[0046] Step S3: constructing a background vehicle interaction model centered on the measured vehicle.

[0047] The background vehicle interaction model takes the future planning trajectory of the measured vehicle as the core condition to generate a responsive simulation trajectory of the background vehicle.

[0048] Specifically, step S3 includes the following sub-steps: The background vehicle interaction model aims to solve the following problem: given standardized scene data, including map information , the state of all agents, and the future planning trajectory of the VUT , generate the future simulation trajectory of all background vehicles : .

[0049] S31: constructing a Transformer-based encoder-decoder architecture as the background vehicle interaction model, which receives standardized data and the VUT future planning trajectory as input; S32: encoding the scene information, using a Long Short-Term Memory (LSTM) network to encode the historical trajectories of all agents, and using a Multilayer Perceptron (MLP) to encode the vectorized map elements, to obtain agent features and map features, respectively; S33: modeling the scene interaction relationship, first processing the agent state feature set through a self-attention mechanism to capture the interaction between agents:

[0050] Then, through a cross-attention mechanism, the agent features and map features are fused to incorporate static environmental information:

[0051] S34: encode the VUT future planning trajectory as planning features, and inject the information of planning features into the features of each background vehicle by using cross-attention mechanism again, to obtain the final features fused with the intention of VUT:

[0052] S35: decode the final features to generate a multi-modal distribution of each background vehicle future simulation trajectory, which consists of a candidate trajectory and its corresponding probability:

[0053] Step S4: jointly train the VUT planning agent model and the background vehicle interaction model.

[0054] The training is performed in an end-to-end manner, specifically, step S4 includes the following sub-steps: S41: design a combined loss function for collaborative optimization of the two models; wherein, is the background vehicle simulation trajectory loss, used to optimize the background vehicle interaction model; is the VUT planning agent model loss, used to optimize the weight distribution parameters and of the VUT planning agent model to achieve weight update of the cost function.

[0055] S42: calculate the gradient of the combined loss function with respect to all model parameters by using the backpropagation algorithm, and perform synchronous update using the optimizer.

[0056] Step S5: deploy the trained model in the simulation environment and perform interactive simulation test centered on the vehicle under test.

[0057] S51: at each simulation step, first generate the future planning trajectory of VUT by the VUT planning agent model; S52: then input the planning trajectory and the current scene information into the background vehicle interaction model; S53: finally generate the responsive simulation trajectory of all background vehicles by the interaction model, and drive them to move in the simulation environment.

[0058] The method will be described below with a specific example.

[0059] Firstly, step S1 is performed to standardize and unify the features of the autonomous driving scene data. In this embodiment, three mainstream large-scale autonomous driving datasets are selected for processing, namely the Waymo dataset, the Argoverse 2 dataset, and the nuPlan dataset. After standardization, the data enters the same training process in a unified structure without additional adaptation. This setting is used to verify the robustness and scene adaptability of the method in diverse traffic environments and driving styles.

[0060] Secondly, steps S2 and S3 are performed to build 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 use deep neural network architectures. The recurrent neural network used to encode the trajectory sequence is specifically an LSTM, and the neural network used to encode the map and perform feature fusion is a Transformer network containing an MLP and an attention mechanism. After training, the error between the planning trajectory generated by the VUT planning agent model and the real VUT trajectory in the dataset is small enough to ensure the authenticity of the simulated planning behavior, thereby providing a high-credibility conditional input for the background vehicle interaction model.

[0061] Thirdly, step S4 is performed to jointly train the VUT planning agent model and the background vehicle interaction model end-to-end. The Adam optimizer is used in the training process, and the weight coefficients of each part of the loss function are optimized through experiments to obtain the best performance.

[0062] Finally, step S5 is performed to deploy the model and perform interactive simulation testing and evaluation.

[0063] To verify the beneficial effects of the method, three groups of comparative models are designed for evaluation in this embodiment: 1) State Condition Baseline Model (SCB): This model uses only the current state of the VUT as input when generating the simulation trajectory of the background vehicle, without considering its future planning.

[0064] 2) Embodiment Method (Fixed Weight, VCIS-Fixed): This model uses the complete framework of the embodiment, but the cost function weight in the VUT planning agent model is a set of fixed deterministic parameters.

[0065] 3) Embodiment Method (Probabilistic Weight, VCIS-Dist): This model uses the complete framework of the embodiment, and the cost function weight in the VUT planning agent model is sampled by the learned probability distribution (i.e., modeling the continuous probability distribution in S22 as a multivariate Gaussian distribution).

[0066] To verify the robustness and generalization performance of the method under different traffic scenarios and driving styles, this embodiment tests the performance of the single joint training model on multiple typical scenario data. As shown in Table 1, the model maintains high simulation accuracy in different types of test data, indicating that the proposed method can adapt to diverse traffic environments.

[0067] Table 1 Simulation and planning error of single model on different data sets

[0068] Further, to analyze the ability of the model to characterize different driving style features, the weight distribution of the cost function learned by the model is visualized, as shown in Figure 3 The differences in different weight distributions in the figure show that the model can capture and distinguish various driving style features, thereby generating planning behaviors of the test vehicle with differentiated characteristics in simulation. This result shows that the method has good scene adaptability and generalization ability while ensuring interaction authenticity.

[0069] To verify the authenticity and interaction improvement effect brought by introducing the VUT future behavior (planning trajectory) in step S3, this embodiment performs comparative analysis. In terms of simulation authenticity, as shown in Table 2, the simulation error (average displacement error ADE is 0.6592 meters) of the method is significantly lower than that of the baseline model (ADE is 0.6877 meters). In terms of interaction, as shown in Figure 4 The box plot in the figure shows the 2-Wasserstein distance between the background vehicle trajectory distribution generated by the method (VCIS-Dist) and the baseline model (SCB) based only on the state, and the average distance measure of the method is higher, indicating that the method can generate more interactive background vehicle behavior. The above results jointly prove that the technical solution of this embodiment can improve the interaction and authenticity of simulation at the same time.

[0070] Table 2 Comparison of simulation and planning error of different models

[0071] To further verify that the background vehicle interaction model constructed in step S3 can respond differently to different driving styles, this embodiment performs a special test, and the results are shown in Figure 5 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.

[0072] 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.

[0073] Example 2 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: (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.

[0074] (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.

[0075] The application introduces the future intention of the measured vehicle into the background vehicle response model, significantly improves the reality and interactivity of the simulation, and the unified data input process further guarantees the high generalization ability of the model under the training of multi-source data, and finally provides a more adaptive simulation tool for the safety verification of the automatic driving system.

[0076] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for automated driving test-oriented interactive simulation with a testee vehicle as a center, characterized by, The method comprises the following steps: An autonomous driving dataset is obtained, and standardized input features are constructed based on the autonomous driving dataset; A tested vehicle planning agent model is constructed, taking scene data as input and the optimal planning trajectory of the tested vehicle with the minimum weighted cost as output, wherein the weight modeling is a continuous probability distribution of learnable parameters; A background vehicle interaction model is constructed, taking the scene data and the optimal planning trajectory of the tested vehicle as input and the interaction simulation trajectory of the background vehicle as output; The tested vehicle planning agent model and the background vehicle interaction model are jointly trained based on the standardized input features of the autonomous driving dataset; An autonomous driving closed-loop simulation test is realized based on the trained tested vehicle planning agent model and the background vehicle interaction model.

2. The method of claim 1, wherein, The tested vehicle planning agent model is modeled as: ， ， wherein, is a current time, is a reference, a simulation step, is a measured vehicle optimal planning trajectory in a simulation step, is a weighted cost function, is a planning trajectory, is a weight vector that determines a driving style and , is a number of terms of a cost function, is a preset first cost function term, is a scenario state, is map information, and is a feature parameter of a learnable probability distribution, is a random variable conforming to a preset distribution, , represents a Gaussian distribution, represents a tensor of all 1s, represents element-wise multiplication of vectors.

3. The method of claim 1, wherein, The scene data includes map information, the state of the tested vehicle, and the state of the background vehicle.

4. The method of claim 3, wherein, The processing process of the background vehicle interaction model comprises the following steps: The state of the tested vehicle and the state of the background vehicle are encoded using a long short-term memory network to obtain initial agent features; The map information is encoded using a multi-layer perception machine to obtain initial map features; The input optimal planning trajectory of the tested vehicle is encoded using a long short-term memory network to obtain initial planning trajectory features; The initial agent features are processed through a self-attention mechanism to model the interaction relationship between multiple traffic participants to obtain agent interaction features; The agent interaction features are taken as query vectors, and the initial map features are taken as key vectors and value vectors for feature fusion through a cross-attention mechanism to model the interaction relationship between the agent and the map elements, and agent-map interaction features representing the influence of map constraints on agent behavior are obtained; The agent interaction features are taken as query vectors, and the initial planning trajectory features are taken as key vectors and value vectors for feature fusion through a cross-attention mechanism to model the response relationship of the background vehicle to the planning intention of the tested vehicle, and agent-intention interaction features are obtained; The agent interaction features, the agent-map interaction features, and the agent-intention interaction features are spliced, and a multi-modal distribution of the background vehicle trajectory, i.e., the background vehicle interactive simulation trajectory, is obtained through decoding.

5. The method of claim 4, wherein, In the background vehicle interaction model, the agent interaction features, the agent-map interaction features, and the agent-intention interaction features are obtained by 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.

6. The method of claim 4, wherein, The multi-modal distribution of the background vehicle trajectory is: ， ， 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.

7. The method of claim 1, wherein, Model parameter updating of the tested vehicle planning agent model and the background vehicle interaction model is realized by calculating a combined loss function and performing back propagation, wherein the combined loss function comprises a background vehicle simulation trajectory loss and a tested vehicle planning agent model loss.

8. The method of claim 1, wherein, The process of realizing autonomous driving testing based on the trained tested vehicle planning agent model and the background vehicle interaction model comprises the following steps: In the current simulation step, the optimal planning trajectory of the tested vehicle is generated using the trained tested vehicle planning agent model; The optimal planning trajectory and the obtained scene data are input into the trained background vehicle interaction model to generate a responsive simulation trajectory of the background vehicle, drive the background vehicle to move in the simulation environment, and realize automatic driving closed-loop simulation testing.

9. The method of claim 1, wherein, In the process of constructing standardized input features based on the automatic driving dataset, the process further includes: Data analysis and conversion of the multi-source automatic driving dataset into a preset structured dictionary data format; Based on the converted dictionary data, a tensor is formed through feature extraction to obtain scene data; A local coordinate system is established with the current position of the measured vehicle as the origin and the direction as the coordinate axis, and the historical and future states of all traffic participants in the scene and all map elements are converted to a dynamic coordinate system with the measured vehicle as the center; The scene data is uniformly cropped to a fixed time length and resampled at a fixed frequency to unify the time dimension; The state of each traffic participant at each time is uniformly represented as a standardized feature vector, which includes kinematic information; Static elements in the map are represented as a vectorized polyline composed of a series of coordinate points, and each type of map element is assigned a standardized type identifier.

10. An interactive simulation system for autonomous driving test oriented to a testee vehicle, characterized by, The system for implementing the interactive simulation method according to any one of claims 1-9 comprises: A shadow measured vehicle planning module, which is deployed with a measured vehicle planning agent model taking scene data as input and the optimal planning trajectory of the measured vehicle with the minimum weighted cost as output; An interaction module centered on the measured vehicle, which is deployed with a background vehicle interaction model centered on the measured vehicle taking scene data and the optimal planning trajectory of the measured vehicle as input and the simulation trajectory of the background vehicle as output.

Images