Automatic driving test scene condition controllable generation method based on table diffusion model

Abstract

The present application relates to the technical field of automatic driving simulation test scene generation, and relates to an automatic driving test scene condition controllable generation method based on a table diffusion model. Each vehicle cut-in event segment is extracted from a natural driving data set to determine cut-in data. A vehicle kinematics cut-in trajectory model is established, and each cut-in data is parameter fitted to represent each seven-dimensional continuous dynamic parameter vector. Each seven-dimensional continuous dynamic parameter vector is matched with a discrete logic condition label respectively to construct a cut-in logic scene training data set. Based on the cut-in logic scene training data set, a condition controllable generation model is trained. A target condition label is determined, a target scene parameter vector is generated through the condition controllable generation model, and an automatic driving test scene is determined through the vehicle kinematics cut-in trajectory model. A cut-in trajectory model based on vehicle kinematics is proposed, and the generation result can be directly restored to a specific cut-in trajectory that can be used for simulation.

Description

Technical Field

[0001] This specification relates to the field of autonomous driving simulation test scenario generation technology, and in particular to a method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model. Background Technology

[0002] Safety verification of autonomous driving technology is a crucial step in its commercialization. To comprehensively evaluate the performance of autonomous driving systems under various driving conditions, it is necessary to construct test datasets containing a rich variety of scenarios in a simulation environment. Among these, the cut-in scenario is one of the most common and safety-significant interaction scenarios, involving surrounding vehicles changing lanes from adjacent lanes to enter front of the autonomous vehicle. Research shows that 44% of the driver of the autonomous vehicle has already taken braking action when the cutting vehicle is still in its original lane and has not yet crossed the line. This means that the trajectory of the target vehicle in the initial stage of the cut-in affects the driver's behavior. Therefore, how to efficiently and with high quality generate large-scale test scenarios covering different road types, risk levels, and other conditions is an urgent problem to be solved in the field of autonomous driving simulation testing.

[0003] However, in actual engineering, the number of entry scenario samples directly obtained from natural driving data is limited, and they are mainly conventional low-risk scenarios with a very low proportion of high-risk boundary scenarios. Directly using these limited raw samples is difficult to meet the needs of comprehensive testing.

[0004] Therefore, this specification provides a method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model. Summary of the Invention

[0005] This specification provides a method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model, in order to partially solve the aforementioned problems existing in the prior art.

[0006] The following technical solution is adopted in this specification: This manual provides a method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model, including: S1. Extract vehicle entry event segments from the natural driving dataset and determine the entry data corresponding to the vehicle entry event segments; S2. Establish a vehicle kinematics entry trajectory model and perform parameter fitting on each entry data, representing each entry data as a seven-dimensional continuous dynamic parameter vector; S3. Match each seven-dimensional continuous dynamic parameter vector with discrete logic condition labels to construct a training dataset for the logic scenario. S4. Based on the aforementioned entry logic scenario, a training dataset is used to generate a model under controllable training conditions; S5. Determine the target condition label, generate the target scene parameter vector through the trained conditionally controllable generation model, and input the target scene parameter vector into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario.

[0007] Based on the aforementioned technical means, this scheme proposes a seven-parameter entry trajectory model based on vehicle kinematics. It uses a complete sine cycle to describe lateral acceleration and piecewise uniform acceleration to describe longitudinal motion, compressing the high-dimensional entry trajectory point sequence into a seven-dimensional continuous parameter vector. Each of the seven parameters has a clear kinematic meaning, reducing the learning dimensionality of the generative model while preserving the physical characteristics of the entry behavior. The generated results can be directly reconstructed into a specific entry trajectory suitable for simulation using the kinematic model. A dynamic-static joint entry logic scene representation is constructed, uniformly encoding the seven-dimensional continuous dynamic parameters and discrete logical condition labels such as road type into structured scene samples. This representation naturally transforms the condition generation problem into a problem of learning the condition distribution within a joint distribution, without needing to decompose the static environment into a high-dimensional geometric description, while preserving static semantic information directly related to the test task. A unified, condition-controllable generation framework is implemented, allowing a single model to generate scenes of the corresponding category based on given discrete condition labels, eliminating the need to train independent models for each category of conditions. The condition label design has good scalability, allowing for the inclusion of more discrete test conditions through label combinations.

[0008] Furthermore, the vehicle kinematics entry trajectory model includes a lateral motion model and a longitudinal motion model; The expressions for the seven-dimensional continuous dynamic parameter vector, the lateral motion model, and the longitudinal motion model are as follows:

[0009]

[0010]

[0011]

[0012]

[0013] in, It is a seven-dimensional continuous dynamic parameter vector. To input the starting speed of the main vehicle, The initial relative speed between the cutting vehicle and the main vehicle. The initial longitudinal relative distance, For the duration of the cut-in, and These are the longitudinal accelerations of the vehicle during the first and second halves of its cut-in journey, respectively. This represents the lateral acceleration amplitude of the cutting vehicle. , , These are the expressions for the lateral acceleration, lateral velocity, and lateral position of the lateral motion model, respectively. The calculation expression is a longitudinal motion model. The horizontal position at the start of the cut. This refers to the lateral acceleration during the cutting process. , The initial longitudinal velocity of the cutting vehicle.

[0014] Furthermore, the process of training the conditionally controllable generative model includes a forward diffusion process and a backward diffusion process; the forward diffusion process gradually injects Gaussian noise into the seven-dimensional continuous dynamic parameter vector, and the discrete logic condition label is used as the control signal input; the backward diffusion process gradually denoises based on the Gaussian noise prediction value and generates the scene parameter vector under the discrete logic condition label; an adaptive layer normalization condition modulation mechanism, a multi-level attention mechanism for feature relationship modeling, and a Sigmoid noise scheduling strategy are introduced into the conditionally controllable generative model.

[0015] Based on the aforementioned technical methods, three targeted improvements were made to the table diffusion model framework: layer-by-layer deep injection of conditional information was achieved through an adaptive layer normalization conditional modulation mechanism; the kinematic coupling relationship between the seven parameters was captured through two-level feature relationship modeling of the input layer and intermediate layer; and information preservation and noise injection were balanced during the diffusion process through Sigmoid noise scheduling. These three improvements complement each other, significantly enhancing the consistency between the generated scene distribution and the real data distribution.

[0016] Furthermore, the conditionally controllable generative model includes a forward diffusion module, a conditional encoding module, an input layer feature relationship modeling module, a conditional denoising network, and an adaptive layer normalization conditional modulation module. The conditionally controllable generative model is trained using the following method, wherein: From the training dataset of the aforementioned entry logic scenario, a seven-dimensional continuous dynamic parameter vector and discrete logic condition labels are determined as training samples; The training samples are input into the conditionally controllable generative model to be trained. The seven-dimensional continuous dynamic parameter vector in the training samples is passed through the forward diffusion module. Gaussian noise is injected in the forward diffusion module according to the diffusion time step and noise scheduling parameters to obtain noise parameter samples. The discrete logic condition labels and diffusion time steps in the training samples are passed through the condition encoding module to obtain joint condition representations. The noise parameter samples are sequentially passed through the input layer feature relationship modeling module and the conditional denoising network. In the conditional denoising network, the predicted noise is determined by the modulation of the joint conditional representation. Based on the difference between the predicted noise and the Gaussian noise, the loss is determined, and the conditionally controllable generative model to be trained is trained with the goal of minimizing the loss.

[0017] Furthermore, the modulation of the conditional denoising network after the joint conditional representation specifically includes: For the conditional denoising network, the first The layer hidden representation generates translation and scaling terms through a conditional modulation network, and adaptively modulates the layer-normalized hidden representation. The computational expression for adaptive modulation of the layer-normalized hidden representation is as follows:

[0018] in, For layer normalization operation, This is element-wise multiplication; For translation terms, For scaling items, For joint condition representation, For the first Layered hidden representation.

[0019] Furthermore, the multi-level attention mechanism feature relationship modeling includes input layer feature relationship modeling; The noise parameter samples are processed by the input layer feature relationship modeling module, specifically including: The input layer feature relationship modeling module maps each dimension of the continuous dynamic parameter vector in the noise parameter sample to a feature token and superimposes positional encoding. Each feature token is combined into a feature sequence, and the updated representation is calculated through multi-head self-attention. The representation is then projected back into the original input space to form a residual connection.

[0020] Furthermore, the multi-level attention mechanism feature relationship modeling also includes intermediate layer feature interaction attention modeling; In the conditional denoising network, the hidden representation of the widest hidden layer of the conditional denoising network is projected onto the token space and rearranged. After normalizing and modulating the token using joint conditional representation, multi-head self-attention computation is performed. The result of the multi-head self-attention computation is then projected back into the hidden space to form the residual output.

[0021] Furthermore, the noise scheduling parameters in the Sigmoid noise scheduling strategy are defined as follows:

[0022] in, For the Sigmoid function, Let t be the diffusion time step number, and t be the current diffusion time step. and These are the preset range values ​​for the noise dispatch coefficient. The noise scheduling coefficient is the diffusion time step t.

[0023] Furthermore, the target scene parameter vector is input into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario, specifically including: The target scene parameter vector is input into the vehicle kinematics entry trajectory model to determine the target entry trajectory; According to the preset physical constraints, the target entry trajectory is screened, and an autonomous driving test scenario is formed based on the target entry trajectory that meets the physical constraints.

[0024] Furthermore, the physical constraints include initial speed constraints of the main vehicle, relative speed constraints, initial distance constraints, cut-in time constraints, longitudinal acceleration constraints, lateral acceleration constraints, lateral displacement constraints, and collision relationship constraints.

[0025] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This scheme proposes a seven-parameter entry trajectory model based on vehicle kinematics. It uses a complete sine cycle to describe lateral acceleration and piecewise uniform acceleration to describe longitudinal motion, compressing the high-dimensional entry trajectory point sequence into a seven-dimensional continuous parameter vector. Each of the seven parameters has a clear kinematic meaning, reducing the learning dimensionality of the generative model while preserving the physical characteristics of the entry behavior. The generated results can be directly reconstructed into a specific entry trajectory suitable for simulation using the kinematic model. A joint dynamic and static entry logic scene representation is constructed, uniformly encoding the seven-dimensional continuous dynamic parameters and discrete logical condition labels such as road type into structured scene samples. This representation naturally transforms the condition generation problem into learning the condition distribution within a joint distribution, without needing to decompose the static environment into a high-dimensional geometric description, while preserving static semantic information directly related to the test task. A unified, condition-controllable generation framework is implemented, allowing a single model to generate scenes of the corresponding category based on given discrete condition labels, eliminating the need to train independent models for each category of conditions. The condition label design has good scalability, allowing for the inclusion of more discrete test conditions through label combinations. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1A flowchart illustrating the method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model, as provided in the embodiments of this specification. Figure 2 This is a schematic diagram illustrating one possible entry point for this specification. Figure 3 This specification provides a schematic diagram of a generated entry trajectory based on a vehicle kinematics entry trajectory model. Figure 4 A comparison diagram of the distribution of a seven-dimensional continuous dynamic parameter vector generated by a conditionally controllable model provided in this specification and the distribution of the original data; Figure 5 This specification provides a corresponding Figure 1 A schematic diagram of the structure of an electronic device. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0028] In embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0029] Since the number of entry scenario samples directly obtained from natural driving data is limited and mainly consists of conventional low-risk scenarios with a very low proportion of high-risk boundary scenarios, it is difficult to meet the needs of comprehensive testing by directly using these limited raw samples.

[0030] The existing scene generation methods have the following main drawbacks: Firstly, methods based on statistical parametric models, such as Gaussian mixture models combined with Markov chain Monte Carlo sampling or Hamiltonian Monte Carlo sampling, generate data by sampling after fitting the parameter distribution. However, Gaussian mixture models require pre-setting the number of mixture components and have limited ability to fit complex structures of multi-parameter joint distributions; Markov chain Monte Carlo sampling has a slow convergence speed, and its sampling efficiency decreases significantly as the parameter dimension increases; while Hamiltonian Monte Carlo sampling accelerates convergence by introducing gradient information, it still struggles to accurately characterize the joint distribution structure of multiple parameters when dealing with strong coupling relationships between parameters, resulting in a large deviation between the generated scene distribution and the real data.

[0031] Secondly, methods based on deep generative models, such as Conditional Generative Adversarial Networks (CGNs) and Variational Table Autoencoders (VAEs), suffer from limitations. Conditional Generative Adversarial Networks learn the data distribution through adversarial training, but suffer from training instability and pattern collapse, resulting in limited generation quality at small sampling scales, with little improvement as the sampling scale increases. Variational Autoencoders learn the latent distribution through a variational inference framework, but their assumed prior distribution may not match the actual data distribution, limiting generation quality. Both methods exhibit insufficient adaptability when handling tabular data with a mixture of continuous parameters and discrete conditions.

[0032] Third, from the perspective of scene representation, when directly using the original trajectory point sequence for modeling, the data dimensionality is high, and different scenes are difficult to align due to differences in duration and sampling frequency, which is not conducive to the learning and training of the generative model. Although existing studies have proposed parameterized trajectory models, the physical meaning of some model parameters is not clear enough, or static environmental semantics such as road type are not included in the unified representation.

[0033] Fourth, from the perspective of conditional generation, most existing methods are unconditional, meaning they sample from the overall data distribution without specifying generation conditions; or they require training independent generation models for each type of condition, lacking a unified, conditionally controllable generation framework. This approach not only increases the cost of model training and maintenance but also fails to utilize the common information of scenarios under different conditions.

[0034] Therefore, this specification provides a method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model.

[0035] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0036] Figure 1 The flowchart of the method for generating controllable test scenario conditions based on a table diffusion model provided in the embodiments of this specification includes the following steps: S1: Extract vehicle entry event segments from the natural driving dataset and determine the entry data corresponding to the vehicle entry event segments.

[0037] This specification describes the process of controllably generating autonomous driving test scenario conditions based on a table-based diffusion model. In the embodiments described herein, this process can be executed by a server. However, this specification does not limit the type of device or platform used to perform this process; for example, a personal computer, mobile terminal, or other such device or platform can also be used. For ease of description, the following description uses a server as the executing entity.

[0038] In one or more embodiments of this specification, the server can extract vehicle cut-in event segments from the natural driving dataset, determine the cut-in data corresponding to the vehicle cut-in event segments, and the cut-in data includes the motion state information of the target vehicle (main vehicle) and surrounding vehicles (cut-in vehicles) during the cut-in process, providing a data foundation for subsequent cut-in trajectory modeling and scene parameterization.

[0039] S2: Establish a vehicle kinematics entry trajectory model and perform parameter fitting on each entry data, representing each entry data as a seven-dimensional continuous dynamic parameter vector.

[0040] In one or more embodiments of this specification, the server can establish a vehicle kinematics entry trajectory model and use nonlinear least squares method to fit parameters to each entry data, representing each entry data as a seven-dimensional continuous dynamic parameter vector.

[0041] Figure 2 This is a schematic diagram illustrating one possible entry point for this specification. For example... Figure 2 As shown, the light orange and orange cars represent the main cars before and after the cut-in, respectively; the silver and blue cars represent the cutting-in cars; and the blue car represents the cutting-in car after the cut-in.

[0042] The vehicle kinematics entry trajectory model includes a lateral motion model and a longitudinal motion model.

[0043] The expressions for the seven-dimensional continuous dynamic parameter vector, the lateral motion model, and the longitudinal motion model are as follows:

[0044]

[0045]

[0046]

[0047]

[0048] in, It is a seven-dimensional continuous dynamic parameter vector. To input the starting speed of the main vehicle, The initial relative speed between the cutting vehicle and the main vehicle. The initial longitudinal relative distance, For the duration of the cut-in, and These are the longitudinal accelerations of the vehicle during the first and second halves of its cut-in journey, respectively. This represents the lateral acceleration amplitude of the cutting vehicle. , , These are the expressions for the lateral acceleration, lateral velocity, and lateral position of the lateral motion model, respectively. The calculation expression is a longitudinal motion model. The horizontal position at the start of the cut. This refers to the lateral acceleration during the cutting process. , Let this be the initial longitudinal velocity of the vehicle entering the collision. In the vehicle's kinematic collision trajectory model, lateral motion is described by a complete sine cycle for lateral acceleration, and longitudinal motion is described by the midpoint of the collision duration. The motion is divided into two segments of uniform acceleration. The lateral motion model ensures that the lateral velocity of the cutting vehicle is zero at both the beginning and end, and the jerk is bounded and continuous.

[0049] S3: Match each seven-dimensional continuous dynamic parameter vector with discrete logic condition labels to construct a training dataset for the logic scenario.

[0050] S4: Based on the training dataset of the aforementioned entry logic scenario, a model is generated under controllable training conditions.

[0051] In one or more embodiments of this specification, the server can match each seven-dimensional continuous dynamic parameter vector with discrete logic condition labels to construct a training dataset for the logic scenario.

[0052] Each vehicle entry event segment is represented as a joint form of dynamic parameters and discrete logic conditions. ,in It is a seven-dimensional continuous dynamic parameter vector. These are discrete logic condition labels used to represent the static environmental semantic information of road types. All samples are organized into a continuous feature matrix. and the corresponding conditional label vector This forms a training dataset for the logical scenarios of entry points. .

[0053] Discrete logic condition labels can incorporate combinations of road type, risk level, or other discrete test conditions by expanding the label set, enabling scene generation under multiple conditions within the same generative model without the need to train a separate generator for each type of condition.

[0054] Then, based on the training dataset of the entry logic scenario, the model is generated under controllable training conditions.

[0055] The training process of the conditionally controllable generative model includes forward diffusion and backward diffusion. The forward diffusion process progressively injects Gaussian noise into the seven-dimensional continuous dynamic parameter vector, with discrete logic condition labels serving as the control signal input. The backward diffusion process progressively denoises the predicted Gaussian noise values ​​and generates scene parameter vectors under discrete logic condition labels. The conditionally controllable generative model incorporates an adaptive layer normalization conditional modulation mechanism, a multi-level attention mechanism for feature relationship modeling, and a Sigmoid noise scheduling strategy.

[0056] The conditionally controllable generative model includes a forward diffusion module, a conditional encoding module, an input layer feature relationship modeling module, a conditional denoising network, and an adaptive layer normalization conditional modulation module. It is constructed within the framework of a table-based denoising diffusion probability model. The server can train the conditionally controllable generative model using the following methods: From the input logic scenario training dataset, a seven-dimensional continuous dynamic parameter vector and discrete logic condition labels are selected as training samples. These training samples are then input into the conditionally controllable generative model to be trained. The seven-dimensional continuous dynamic parameter vector from the training samples is passed through a forward diffusion module, where Gaussian noise is injected based on the diffusion time step and noise scheduling parameters to obtain noise parameter samples. The discrete logic condition labels and diffusion time steps from the training samples are passed through a conditional encoding module to obtain a joint conditional representation. Subsequently, the noise parameter samples pass sequentially through an input layer feature relationship modeling module and a conditional denoising network. In the conditional denoising network, denoising features are extracted under the modulation of the joint conditional representation to determine the predicted noise. Finally, based on the difference between the predicted noise and the Gaussian noise, the loss is determined, and the conditionally controllable generative model to be trained is trained with the goal of minimizing the loss.

[0057] It is worth noting that the loss calculation can be Gaussian mean square error loss. :

[0058] In the formula, This is Gaussian noise sampled during the forward diffusion process. The conditional denoising network predicts the noise level, i.e., it predicts the noise. This represents the noise parameter sample at the t-th diffusion time step.

[0059] Specifically, the modulation after joint conditional representation in the conditional denoising network involves embedding the diffusion time step... Embedding with Discrete Logic Condition Labels Adding them together yields the joint conditional representation. back( ), for the conditional denoising network Layer hidden representation, through conditional modulation network Generate translation terms and scaling items And adaptively modulate the hidden representation after layer normalization.

[0060] The computational expression for adaptive modulation of the layer-normalized hidden representation is as follows:

[0061] in, For layer normalization operation, For element-wise multiplication, For translation terms, For scaling items, For joint condition representation, For the first Layer Hidden Representation. Unlike the original model that injects conditions only once at the input, AdaLN (Adaptive Layer Normalization) allows conditional information to remain effective across all layers of the network, transforming the input-level additive signal into a layer-by-layer modulated signal. The conditional modulation network employs a zero-initialization strategy, ensuring that the network approaches an identity mapping in the early stages of training, thus preventing excessively strong conditional perturbations from causing optimization instability.

[0062] In this specification, the feature relationship modeling of the multi-level attention mechanism includes input layer feature relationship modeling. Noise parameter samples pass through the input layer feature relationship modeling module. Thus, through this module, each dimension of the continuous dynamic parameter vector in the noise parameter sample (a seven-dimensional continuous dynamic parameter vector) is mapped to a feature token and then positional encoding is superimposed, resulting in... , , Then, combine the feature tokens into a feature sequence. The updated representation is calculated using multi-head self-attention. ( ), and then represent The model projects back into the original input space and forms residual connections. This structure allows the model to learn before entering the conditional denoising network. , , , , , and The dependencies between them should be understood, rather than treating them as unrelated numerical dimensions.

[0063] Furthermore, the multi-level attention mechanism feature relationship modeling also includes intermediate layer feature interaction attention modeling. In the intermediate hidden layers of the conditional denoising network, the hidden representation of the widest hidden layer of the conditional denoising network is used... Projecting onto the token space and rearranging it, we get ( , For learnable positional encoding, (This is the projection vector). Then, after normalizing and modulating the token using the joint conditional representation, multi-head self-attention computation is performed to obtain... ( ,in, This indicates the use of joint conditions to represent the pair. The token representation obtained after normalization modulation, and then the result of multi-head self-attention calculation. The model projects back into the hidden space and forms the residual output. By modeling the feature relationships between the input layer and the intermediate layer, the model learns parameter correlations at the input stage and continues to update feature interaction information in the higher-level representation space, thereby capturing the kinematic coupling relationships between the seven dynamic parameters.

[0064] Of course, in this specification, the noise scheduling parameters in the Sigmoid noise scheduling strategy can be defined as follows:

[0065] in, For the Sigmoid function, Let t be the diffusion time step number, and t be the current diffusion time step. and These are the preset range values ​​for the noise dispatch coefficient. Let be the noise scheduling coefficient at diffusion time step t. Compared with linear or cosine scheduling, sigmoid scheduling changes more slowly in the early and late stages of diffusion, and transitions smoothly in the middle stage. This can reduce the information destruction caused by noise injection that is too fast or too slow, and helps to balance global distribution fitting and local detail recovery.

[0066] S5: Determine the target condition label, generate the target scene parameter vector through the trained conditionally controllable generation model, and input the target scene parameter vector into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario.

[0067] In one or more embodiments of this specification, the server may determine the target condition label, generate the target scene parameter vector through the trained conditionally controllable generative model, and input the target scene parameter vector into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario.

[0068] In the process of generating the target scene parameter vector through the trained conditionally controllable generative model, the backdiffusion sampling module of the conditionally controllable generative model initializes the current sample with Gaussian random noise. The target conditional label and the current diffusion time step are input into the conditional encoding module to form a joint conditional representation. The conditional denoising network predicts the noise at the current time step under the modulation of the joint conditional representation. The backdiffusion sampling module updates the current sample with the sample from the previous time step based on the predicted noise, until the generated seven-dimensional continuous dynamic parameter vector is obtained. Finally, the generated seven-dimensional continuous dynamic parameter vector is input into the vehicle kinematics entry trajectory model to recover the corresponding entry trajectory and determine the autonomous driving test scenario.

[0069] Furthermore, in one or more embodiments of this specification, the server inputs the target scene parameter vector into the vehicle kinematics entry trajectory model to determine the target entry trajectory. Then, the target entry trajectory can be screened according to preset physical constraints, and an autonomous driving test scenario can be formed based on the target entry trajectory that satisfies the physical constraints.

[0070] The physical constraints include initial speed constraints of the main vehicle, relative speed constraints, initial distance constraints, cut-in duration constraints, longitudinal acceleration constraints, lateral acceleration constraints, lateral displacement constraints, and collision relationship constraints. For example, physical constraints could include: the initial speed of the main vehicle must be within a reasonable driving speed range; the relative speed must ensure that the speed of the cutting-in vehicle is positive; the initial distance must be positive; the cut-in duration must be within a reasonable range; the longitudinal acceleration must be within the range that vehicle dynamics can reach; the lateral acceleration amplitude must meet the lateral stability requirements; the total lateral displacement must match the lane width constraints; and no collision must occur between the main vehicle and the cutting-in vehicle during the entire cut-in process.

[0071] based on Figure 1 The table-based diffusion model for autonomous driving test scenario conditions controllable generation method is shown. This scheme proposes a seven-parameter entry trajectory model based on vehicle kinematics, as follows: Figure 2As shown, a complete sine cycle is used to describe lateral acceleration, and piecewise uniform acceleration is used to describe longitudinal motion, compressing the high-dimensional entry trajectory point sequence into a seven-dimensional continuous parameter vector. Each of the seven parameters has a clear kinematic meaning, reducing the learning dimensionality of the generative model while preserving the physical characteristics of the entry behavior. The generated results can be directly reconstructed into a specific entry trajectory for simulation using the kinematic model. A joint dynamic and static entry logic scene representation is constructed, uniformly encoding the seven-dimensional continuous dynamic parameters and discrete logical condition labels such as road type into structured scene samples. This representation naturally transforms the condition generation problem into a problem of learning the condition distribution within a joint distribution, without needing to decompose the static environment into a high-dimensional geometric description, while preserving static semantic information directly related to the test task. A unified, conditionally controllable generation framework is implemented, allowing a single model to generate scenes of the corresponding category based on given discrete condition labels, without needing to train an independent model for each category of conditions. The condition label design has good scalability, allowing more discrete test conditions to be incorporated through label combinations.

[0072] Based on the table diffusion model framework, three targeted improvements were made: a layer-by-layer deep injection of conditional information was achieved through an adaptive layer normalization conditional modulation mechanism; the kinematic coupling relationship between the seven parameters was captured by modeling the feature relationship between the input layer and the intermediate layer; and information preservation and noise injection were balanced during the diffusion process through Sigmoid noise scheduling. These three improvements complement each other, significantly improving the consistency between the generated scene distribution and the real data distribution.

[0073] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A method for generating controllable test scenario conditions for autonomous driving based on a table diffusion model is provided.

[0074] like Figure 3 As shown, Figure 3 From left to right, the images show the entry trajectories of all generated samples from 20,000 samplings, highway condition labeled samples, and urban road condition labeled samples. The horizontal axis represents the longitudinal position of the entering vehicle, and the vertical axis represents the lateral position. Each colored curve corresponds to an entry trajectory obtained by reconstructing a seven-dimensional continuous dynamic parameter vector from a vehicle kinematics entry trajectory model. The curve colors are mapped according to the initial speed of the main vehicle, with blue to red indicating an increase in initial speed; the black horizontal dashed line represents the boundary of adjacent lanes. The curves expand in the positive and negative directions of the vertical axis to reflect two lateral entry directions, respectively. The distribution range of the curves in the longitudinal and lateral directions reflects the differences in entry distance, duration, speed, and lateral motion intensity of the generated scene. The proportion of warm-colored trajectories is relatively high in the highway sub-map, while the proportion of cool-colored trajectories is relatively high in the urban road sub-map, indicating that the target road type conditions differentiate the main vehicle speed and the corresponding trajectory distribution.

[0075] like Figure 4 As shown, the seven subplots, arranged from left to right and top to bottom, sequentially present the one-dimensional probability distributions of the initial speed, initial relative speed, initial longitudinal relative distance, cut-in duration, longitudinal acceleration in the first half of the trajectory, longitudinal acceleration in the second half of the trajectory, and lateral acceleration amplitude. The horizontal axis of each subplot represents the value and unit of the corresponding parameter, while the vertical axis represents the probability density. The blue histogram bars represent the original distribution obtained from fitting natural driving data, and the orange histogram bars represent the distribution of valid samples generated by the conditionally controllable generative model and retained after parameter range screening. The degree of overlap between the two colors within the same bin reflects the consistency of the marginal distribution of that parameter.

[0076] This instruction manual also provides Figure 5 The diagram shows a schematic structural representation of the electronic device. Figure 5 As shown, at the hardware level, this electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above. Figure 1 A method for generating controllable test scenario conditions for autonomous driving based on a table diffusion model is provided.

[0077] Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0078] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0079] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0080] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0081] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0087] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0088] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic or disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0089] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0090] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0092] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0093] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model, characterized in that, include: S1. Extract vehicle entry event segments from the natural driving dataset and determine the entry data corresponding to the vehicle entry event segments; S2. Establish a vehicle kinematics entry trajectory model and perform parameter fitting on each entry data, representing each entry data as a seven-dimensional continuous dynamic parameter vector; S3. Match each seven-dimensional continuous dynamic parameter vector with discrete logic condition labels to construct a training dataset for the logic scenario. S4. Based on the aforementioned entry logic scenario, a training dataset is used to generate a model under controllable training conditions; S5. Determine the target condition label, generate the target scene parameter vector through the trained conditionally controllable generation model, and input the target scene parameter vector into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario.

2. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 1, characterized in that, The vehicle kinematics entry trajectory model includes a lateral motion model and a longitudinal motion model; The expressions for the seven-dimensional continuous dynamic parameter vector, the lateral motion model, and the longitudinal motion model are as follows: in, It is a seven-dimensional continuous dynamic parameter vector. To input the starting speed of the main vehicle, The initial relative speed between the cutting vehicle and the main vehicle. The initial longitudinal relative distance, For the duration of the cut-in, and These are the longitudinal accelerations of the vehicle during the first and second halves of its cut-in journey, respectively. This represents the lateral acceleration amplitude of the cutting vehicle. , , These are the expressions for the lateral acceleration, lateral velocity, and lateral position of the lateral motion model, respectively. The calculation expression is a longitudinal motion model. The horizontal position at the start of the cut. This refers to the lateral acceleration during the cutting process. , The initial longitudinal velocity of the cutting vehicle.

3. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 1, characterized in that, The process of training a controllable generative model includes a forward diffusion process and a backward diffusion process; the forward diffusion process gradually injects Gaussian noise into the seven-dimensional continuous dynamic parameter vector, and the discrete logic condition label is used as the control signal input; The reverse diffusion process gradually denoises the predicted Gaussian noise value and generates scene parameter vectors under discrete logic condition labels; an adaptive layer normalization condition modulation mechanism, a multi-level attention mechanism for feature relationship modeling, and a Sigmoid noise scheduling strategy are introduced into the conditionally controllable generation model.

4. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 3, characterized in that, The conditionally controllable generative model includes a forward diffusion module, a conditional coding module, an input layer feature relationship modeling module, a conditional denoising network, and an adaptive layer normalization conditional modulation module. The conditionally controllable generative model is trained using the following method, wherein: From the training dataset of the aforementioned entry logic scenario, a seven-dimensional continuous dynamic parameter vector and discrete logic condition labels are determined as training samples; The training samples are input into the conditionally controllable generative model to be trained. The seven-dimensional continuous dynamic parameter vector in the training samples is passed through the forward diffusion module. Gaussian noise is injected in the forward diffusion module according to the diffusion time step and noise scheduling parameters to obtain noise parameter samples. The discrete logic condition labels and diffusion time steps in the training samples are passed through the condition encoding module to obtain joint condition representations. The noise parameter samples are sequentially passed through the input layer feature relationship modeling module and the conditional denoising network. In the conditional denoising network, the predicted noise is determined by the modulation of the joint conditional representation. Based on the difference between the predicted noise and the Gaussian noise, the loss is determined, and the conditionally controllable generative model to be trained is trained with the goal of minimizing the loss.

5. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 4, characterized in that, The modulation processed by the joint conditional representation in the conditional denoising network specifically includes: For the conditional denoising network, the first The layer hidden representation generates translation and scaling terms through a conditional modulation network, and adaptively modulates the layer-normalized hidden representation. The computational expression for adaptive modulation of the layer-normalized hidden representation is as follows: in, For layer normalization operation, This is element-wise multiplication; For translation terms, For scaling items, For joint condition representation, For the first Layered hidden representation.

6. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 4, characterized in that, The multi-level attention mechanism feature relationship modeling includes input layer feature relationship modeling; The noise parameter samples are processed by the input layer feature relationship modeling module, specifically including: The input layer feature relationship modeling module maps each dimension of the continuous dynamic parameter vector in the noise parameter sample to a feature token and superimposes positional encoding. Each feature token is combined into a feature sequence, and the updated representation is calculated through multi-head self-attention. The representation is then projected back into the original input space to form a residual connection.

7. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 6, characterized in that, The multi-level attention mechanism feature relationship modeling also includes intermediate layer feature interaction attention modeling; In the conditional denoising network, the hidden representation of the widest hidden layer of the conditional denoising network is projected onto the token space and rearranged. After normalizing and modulating the token using joint conditional representation, multi-head self-attention computation is performed. The result of the multi-head self-attention computation is then projected back into the hidden space to form the residual output.

8. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 4, characterized in that, The noise scheduling parameters in the Sigmoid noise scheduling strategy are defined as follows: in, For the Sigmoid function, Let t be the diffusion time step number, and t be the current diffusion time step. and These are the preset range values ​​for the noise dispatch coefficient. The noise scheduling coefficient is the diffusion time step t.

9. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 1, characterized in that, The target scene parameter vector is input into the vehicle kinematics entry trajectory model to determine the autonomous driving test scenario, specifically including: The target scene parameter vector is input into the vehicle kinematics entry trajectory model to determine the target entry trajectory; According to the preset physical constraints, the target entry trajectory is screened, and an autonomous driving test scenario is formed based on the target entry trajectory that meets the physical constraints.

10. The method for controllable generation of autonomous driving test scenario conditions based on a table diffusion model as described in claim 9, characterized in that, The physical constraints include the initial speed constraint of the main vehicle, the relative speed constraint, the initial distance constraint, the cutting-in time constraint, the longitudinal acceleration constraint, the lateral acceleration constraint, the lateral displacement constraint, and the collision relationship constraint.

Images