An autonomous traffic scene generation and reproduction platform based on multi-software combination

Abstract

The application discloses an autonomous traffic scene generation and reproduction platform based on multi-software jointing, and belongs to the technical field of autonomous traffic system simulation testing. The platform comprises an integrated simulation framework, and has a scene generation layer, a scene reproduction layer and a simulation testing layer which are sequentially and cooperatively operated. The scene generation layer generates a structured traffic scene library by using a hybrid method comprising interactive group extraction, importance sampling optimization search and proximal policy optimization reinforcement learning. The scene reproduction layer integrates traffic scene visualization, high-precision vehicle dynamics and hybrid traffic flow simulation software through a unified interface specification. A joint simulation control center is realized based on a control logic modeling software, is used for coordinating data flow and simulation steps of each layer, and is integrated with a vehicle control unit model, an electronic stability program model and a roadside information transmission simulation module. The application solves the problems of disconnection between scene generation and reproduction, difficulty in multi-software cooperation and insufficient support for system-level interaction verification in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of autonomous transportation system simulation and testing technology, and in particular relates to an autonomous transportation scene generation and reproduction platform based on multi-software collaboration. Background Technology

[0002] With the development of intelligent connected vehicle and vehicle-road cooperative technologies, the verification of Autonomous Transportation Systems (ATS) increasingly relies on high-fidelity simulation testing. Currently, virtual simulation of transportation systems mainly relies on various specialized software to simulate specific functions, and on this basis, multi-software joint architectures have been developed. In terms of scene construction and visualization, software such as PreScan, CARLA, and VTD can build highly realistic virtual traffic environments and provide simulation models of various sensors such as cameras and LiDAR, focusing on testing autonomous driving perception and decision-making algorithms. In terms of high-precision simulation of vehicle dynamics, software such as Carsim and Adams have built-in complex tire and chassis models, which can accurately analyze the handling, ride comfort, and stability of vehicles under various operating conditions. In terms of macroscopic and mesoscopic traffic flow simulation, software such as SUMO and Vissim can effectively simulate the operating state of mixed traffic flow at the road network level based on custom car-following and lane-changing models. In addition, general-purpose modeling tools such as Simulink are good at the design and simulation of vehicle control logic and state estimation algorithms. To integrate the above advantages, the industry has seen the emergence of highly integrated simulation platforms such as IPG CarMaker and PanoSIM, or the coupling of the above software through custom interfaces to form a joint simulation solution, in order to take into account the multi-dimensional needs of scenarios, dynamics, control and traffic flow in a single test.

[0003] However, existing simulation technologies still have significant shortcomings. First, the limited coverage of single software disciplines leads to a break in the simulation chain. Software dynamics models focused on specific scenarios are simplified, while high-precision dynamics software lacks support for complex traffic environments and sensors, making it difficult to support closed-loop verification of the entire "perception-decision-control-execution" process. Second, when multi-software co-simulation is involved, data interaction is limited and simulation stability faces challenges. Different software differ in data formats, spatiotemporal coordinate systems, simulation step sizes, and communication protocols, requiring extensive customization during integration, resulting in poor compatibility and scalability. Furthermore, the closed-loop coupling of multiple software programs can easily introduce algebraic loops, leading to low simulation efficiency, numerical oscillations, and even difficulty in convergence. Finally, existing tools do not adequately support the verification needs unique to autonomous transportation systems. Traditional simulation tools lack convenient modeling and testing support for vehicle-to-vehicle and vehicle-to-infrastructure information interaction mechanisms, as well as the strong coupling and interaction between intelligent connected vehicles and the background mixed traffic flow, making it difficult to efficiently reproduce and evaluate system-level collaborative operation scenarios and the effectiveness of safety management algorithms. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an autonomous traffic scene generation and reproduction platform based on multi-software collaboration, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides an autonomous traffic scene generation and reproduction platform based on multi-software collaboration, comprising: An integrated simulation framework, comprising a scene generation layer, a scene reproduction layer, and a simulation testing layer that operate sequentially and collaboratively; The scene generation layer is used to generate a traffic scene library for system-level collaborative testing based on real traffic data, using a hybrid method that includes interaction group extraction, importance sampling optimization search, and near-end policy optimization reinforcement learning. The traffic scene library contains a structured description of static road networks, dynamic traffic participants, environmental conditions, and information interaction elements. The scene reproduction layer includes traffic scene visualization software, high-precision vehicle dynamics simulation software, and hybrid traffic flow simulation software integrated through a unified interface specification; The joint simulation control center, implemented based on control logic modeling software, is used to coordinate the data flow and simulation steps between the scene generation layer, scene reproduction layer and simulation test layer, and integrates a vehicle control unit model, an electronic stability program model and a roadside information transmission simulation module. The platform is configured to implement an autonomous transportation system testing process, from scene generation and multidisciplinary high-fidelity reproduction to closed-loop verification of safety control algorithms.

[0006] Preferably, the hybrid method used in the scene generation layer specifically includes: Based on the location of the central vehicle, lane topology, and preset spatial range, extract interaction groups containing multiple related traffic participants; Construct an importance function based on the level of danger and exposure frequency of a scene, and use this importance function to guide the optimization search to select key scene seeds from the initial scene set; Using the aforementioned key scenario seed as the initial environment, a background traffic participant agent is trained using a near-end policy optimization algorithm. By maximizing the reward function that incorporates driving style parameters, dynamic dangerous interaction scenarios are generated.

[0007] Preferably, when generating dynamic dangerous interaction scenarios, a multi-vehicle joint trajectory prediction model based on graph neural networks is adopted. The model obtains global interaction features through an interaction encoder, and outputs the cooperative actions of multiple vehicles in the interaction group through a joint strategy head. The joint trajectory conforming to physical constraints is obtained through a trajectory decoder.

[0008] Preferably, the unified interface specification includes: A first communication protocol and module for clock synchronization and pose data exchange with the traffic scene visualization software; A second communication protocol and module for initializing traffic flow and synchronizing state with the hybrid traffic flow simulation software; A third interface module for issuing control commands and providing vehicle status feedback to the high-precision vehicle dynamics simulation software.

[0009] Preferably, the first communication protocol and module includes a tire contact point information processing submodule, which is used to send tire coordinates to the traffic scene visualization software and receive the corresponding road surface elevation and friction coefficient data returned by the software, and provide them to the high-precision vehicle dynamics simulation software.

[0010] Preferably, the vehicle control unit model is used to calculate the driving torque based on the vehicle status and driving commands; the electronic stability program model uses a fuzzy PID controller to calculate the yaw moment compensation value to maintain vehicle stability.

[0011] Preferably, the roadside information transmission simulation module is used to simulate and generate incomplete traffic environment perception information and send it to the downstream vehicle decision model based on preset roadside equipment attributes and environmental conditions.

[0012] Preferably, the hybrid traffic flow simulation software is equipped with a parameterized driver behavior model that can simulate various driving styles, including aggressive, conservative, and moderate driving styles.

[0013] Preferably, the simulation test layer supports the access of external safety management algorithm models in the form of functional model units, and can partially or completely take over the driver decision-making functions in the joint simulation control center.

[0014] Preferably, the platform further includes a data management and analysis module, used to record simulation process data and perform calculation analysis and visualization output on the test results based on preset evaluation indicators.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs an integrated simulation framework encompassing a "scene generation layer, scene reproduction layer, and simulation testing layer." The platform is configured to achieve a testing process from scene generation and multidisciplinary high-fidelity reproduction to closed-loop verification of safety control algorithms. This technical solution can complete the entire process—from intelligent construction of test scenarios and multidisciplinary high-fidelity reproduction to the injection of control algorithms and performance evaluation—within a unified platform. This overcomes the shortcomings of existing technologies where scene generation, dynamics simulation, traffic flow simulation, and control logic verification tools are fragmented, making it difficult to support the entire "perception-decision-control-execution" chain and system-level collaborative interactive verification.

[0016] The scene generation layer of this invention employs a hybrid approach encompassing interaction group extraction, importance sampling optimization search, and Proximal Policy Optimization (PPO) reinforcement learning. This approach first characterizes multi-vehicle coupling relationships through interaction group extraction, then efficiently focuses on high-risk areas for seed scene search using importance sampling, and finally uses PPO reinforcement learning for dynamic adversarial evolution. This method integrates the advantages of rule-guided, optimization search, and data-driven learning, enabling the automated generation of test scenarios that are highly dangerous, possess reasonable behavioral logic, and exhibit diversity. Particularly for long-tail risk scenarios, it significantly improves the relevance, realism, and efficiency of scene generation.

[0017] The scene reproduction layer described in this invention integrates traffic scene visualization software, high-precision vehicle dynamics simulation software, and hybrid traffic flow simulation software through a unified interface specification. A joint simulation control center is responsible for coordinating data flow and simulation steps between layers. This technical solution effectively solves the heterogeneity issues of different commercial software in terms of data format, spatiotemporal coordinate system, simulation step size, and communication interface by defining a unified interaction protocol and coordination mechanism. This ensures high-fidelity, synchronous reproduction of complex scenes in a multi-software environment and suppresses simulation instability, inefficiency, or result distortion caused by poor data interaction or asynchronous timing.

[0018] This invention integrates a Vehicle Control Unit (VCU) model, an Electronic Stability Program (ESP) model, and a roadside information transmission simulation module within a co-simulation control center. This technical solution deeply integrates the core electronic control system model of new energy vehicles and the vehicle-road cooperative information interaction mechanism into the simulation closed loop. This enables the platform to not only test decision-making and planning algorithms but also to perform high-fidelity verification of vehicle actuator-level control response, stability control logic, and collaborative decision-making capabilities under conditions of receiving incomplete roadside information, thereby comprehensively supporting the development and testing of functions unique to autonomous transportation systems. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a diagram showing the overall structure of the simulation test platform according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the interaction group division process according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the interaction group division according to an embodiment of the present invention; Figure 4 This is a clustering diagram of driving styles according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the Trigger module and its corresponding S-Function in Simulink according to an embodiment of the present invention, wherein (a) is the overall structure of the trigger module and (b) is the core component of the trigger module; Figure 6 This is a schematic diagram of a vehicle information transmission module in Simulink according to an embodiment of the present invention, wherein (a) is the first part of the vehicle position and attitude information encapsulation component, (b) is the second part of the vehicle position and attitude information encapsulation component, and (c) is the vehicle position and attitude information transmission component. Figure 7 This is a schematic diagram of the tire grounding point information transmission module in Simulink according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the tire grounding point information receiving module in Simulink according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the Carsim data input module according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the Carsim data output and conversion module structure according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the SUMO data output and conversion module structure according to an embodiment of the present invention; Figure 12 This is a schematic diagram illustrating the operation steps of the autonomous transportation system platform according to an embodiment of the present invention; Figure 13 This is a schematic diagram of a dangerous scene generated based on reinforcement learning, according to an embodiment of the present invention. Figure 14 This is the Simulink interface of the autonomous transportation system simulation test platform according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 This embodiment provides an autonomous traffic scene generation and reproduction platform, aiming to solve the technical problems of inefficient scene generation, consistent reproduction, and repeatable evaluation when autonomous traffic systems conduct multi-scenario, cross-software refined simulation tests under multi-subject, multi-scale, and strongly coupled conditions involving "people-vehicle-road-loop". In existing technologies, capabilities such as vehicle dynamics simulation, control system simulation, mixed traffic flow simulation, roadside equipment and V2X communication simulation, and environmental and sensor simulation are often scattered across different commercial software or open-source simulators. The definitions of model accuracy, time stepping mechanism, coordinate system and map semantics, object and event description, and communication and signal control interfaces are inconsistent among various tools, making it difficult to complete the construction and management of multi-scale scenes from road networks to roads / intersections within a single platform. It is also difficult to jointly model and close-loop verify traffic control elements such as traffic light timing, traffic control measures, event triggering logic, and vehicle-to-vehicle / vehicle-to-infrastructure information interaction, resulting in difficulties in scene reuse, incomparable results, and the inability to stably reproduce key risk scenarios.

[0023] To this end, the platform proposed in this embodiment takes "scene generation—scene reproduction—simulation testing" as its main thread, and constructs a unified scene data model, interface specifications, and cross-software collaboration mechanism: At the scene generation layer, the platform addresses the needs of system-level collaborative testing by providing structured representations of roads and infrastructure, dynamic traffic participants, environment, and digital information (including signal control and communication parameters), enabling high-precision scene generation and version management for multiple types of traffic participants, multiple traffic control measures, and multi-scale road network structures; At the scene reproduction layer, the platform integrates multi-disciplinary simulation software or components such as vehicle dynamics simulation, control logic / algorithm simulation, mixed traffic flow loading, and roadside equipment and communication simulation. The platform utilizes a unified data interface and spatial coordinate transformation method to achieve synchronous transmission and consistent execution of map semantics, object states, event triggering, signals, and communication messages across different simulators. At the data interaction and synchronization layer, the platform introduces an algebraic loop processing mechanism and a time synchronization and step size coordination strategy to suppress numerical oscillations and instability caused by multi-software closed-loop coupling. This enables high consistency reproduction and repeatable evaluation of the same test scenario across different toolchains, ultimately obtaining high-precision, scalable autonomous traffic system simulation results. This meets the simulation testing needs of safety control algorithms in various multi-scale scenarios, multi-person / multi-vehicle mixed traffic, and vehicle-to-vehicle / road information interaction conditions.

[0024] The hardware and software configuration table of the scene generation and reproduction platform proposed in this invention is shown in Table 1: Table 1 The scenario generation and reproduction platform in this platform is structured as a simulation environment built upon multiple commercial simulation software programs. It integrates and constructs model resources such as driver decision-making and control models, VCU models, ESP models, and information transmission and interaction models. This enables joint simulations based on commercial simulation software including VTD, CarSim, and SUMO, as well as Simulink, to support multi-scenario joint simulation testing of safety management algorithms based on autonomous transportation systems. The overall structure and functions of each component of the simulation testing platform are as follows: Figure 1 As shown.

[0025] The real traffic scene library provides the scene generation module with a set of real scenes as training data for scene generation; the scene generation module provides the scene reproduction layer with a generated traffic scene library, including static scene information, dynamic element information, and interaction information; the scene reproduction module provides scene information for the hybrid traffic flow model, vehicle dynamics model, and driver model; the driver control model provides driving control strategies for the vehicle dynamics model and VCU model; the vehicle dynamics model provides vehicle pose information at various times for the road scene model; the VCU model provides four-wheel braking information for the vehicle dynamics model, and the ESP model provides four-wheel braking torque correction information for the VCU model; the hybrid traffic flow simulation model provides traffic flow information for the roadside information transmission model and road scene model; the roadside information transmission model provides simulated road scene information and traffic flow information for the driver control model.

[0026] This embodiment provides an autonomous traffic scene generation and reproduction platform based on multi-software collaboration, including: An integrated simulation framework, comprising a scene generation layer, a scene reproduction layer, and a simulation testing layer that operate sequentially and collaboratively; The scene generation layer is used to generate a traffic scene library for system-level collaborative testing based on real traffic data, using a hybrid method that includes interaction group extraction, importance sampling optimization search, and near-end policy optimization reinforcement learning. The traffic scene library contains a structured description of static road networks, dynamic traffic participants, environmental conditions, and information interaction elements. Furthermore, the hybrid method employed in the scene generation layer specifically includes: Based on the location of the central vehicle, lane topology, and preset spatial range, extract interaction groups containing multiple related traffic participants; Construct an importance function based on the level of danger and exposure frequency of a scene, and use this importance function to guide the optimization search to select key scene seeds from the initial scene set; Using the aforementioned key scenario seed as the initial environment, a background traffic participant agent is trained using a near-end policy optimization algorithm. By maximizing the reward function that incorporates driving style parameters, dynamic dangerous interaction scenarios are generated.

[0027] Furthermore, when generating dynamic dangerous interaction scenarios, a multi-vehicle joint trajectory prediction model based on graph neural networks is adopted. The model obtains global interaction features through an interaction encoder, and outputs the collaborative actions of multiple vehicles in the interaction group through a joint strategy head. The joint trajectory conforming to physical constraints is obtained through a trajectory decoder.

[0028] Specifically, the scene generation module, based on Python, imports into Simulink via FUM (Full Context) and its main function is to build a high-value scene library for algorithm testing based on real-world scene datasets. It also imports the scene reproduction module and the mixed traffic flow simulation module as prerequisites for program execution. Considering differences in autonomy levels, autonomous driving grades, and penetration rates, it generates a set of test scenes. It inputs the testing requirements for the risk identification or safety management algorithms to be tested, including traffic participants, static scene elements, and autonomy levels, and then filters and customizes the high-precision scene library required for the test. There are three main methods for constructing traffic scenarios.

[0029] (1) Directly select typical traffic scenarios. Typical traffic scenarios have a high frequency of occurrence, high probability of accidents / risks within the scenario, and high coverage of various traffic elements, such as urban road intersections and highway merging and diverging areas. Typical scenarios can reflect the algorithm performance in most cases; (2) Select and customize the required test scenarios. Based on typical traffic scenarios or default traffic scenarios, and according to the algorithm testing requirements, determine the quantity, location, and motion characteristics of dynamic and static elements within the scenario. The system is customized to reflect the performance of the algorithm under test in a specific scenario and its impact on the traffic system; (3) Efficient and automatic generation of test scenarios. Based on the need for accelerated testing, the system adopts methods such as rule constraints, element adversarial approaches, and knowledge-driven approaches to quickly generate generalized scenarios and risk / collision scenarios required for algorithm testing in batches, so as to comprehensively evaluate the algorithm performance based on the overall performance in multiple scenarios. Based on this, the system integrates a method for generating dangerous traffic scenarios for algorithm testing and verification needs.

[0030] a) Extraction of interactive groups: In autonomous driving hazardous scenario generation and adversarial testing, the proper division of interaction groups is fundamental to the entire system. Most researchers, when generating hazardous scenarios, adopt a two-vehicle interaction assumption, considering only the interaction between the central vehicle and the nearest target vehicle. Examples include following vehicles and lane-changing interference from adjacent lanes. However, in real-world road environments, vehicle behavior is often influenced by multi-vehicle coupling effects. For instance, in scenarios such as parallel driving in multiple lanes on highways, waiting at urban intersections, or merging, a vehicle's behavioral adjustment (such as deceleration or lane change) can lead to behavioral responses and spatial reconfiguration of multiple surrounding vehicles. Considering only two-vehicle interactions cannot capture this dynamic propagation effect, resulting in decreased robustness and realism in hazardous scenario generation.

[0031] Therefore, the interaction group extraction method proposed in this module aims to identify the set of multiple vehicles with the most behavioral relevance in the local space, providing constrained inputs for subsequent key scene search based on importance sampling and PPO reinforcement learning generation.

[0032] The interactive group extraction is structured around three layers of logic: geometric constraints, lane topology, and distance relationships. The extraction steps are as follows: 1. Using the center vehicle (HV) as a reference, determine its position, orientation, and lane; 2. Extract the center lane and its left and right adjacent lanes based on map features; 3. Filter surrounding vehicles (SVs) that meet the interaction conditions within the longitudinal (front-to-back) and lateral (sideways) spatial ranges. 4. Select the closest vehicles (e.g., the first 5 vehicles) based on distance priority to form an interaction group. The flowchart for dividing the interaction group is shown below. Figure 2 As shown.

[0033] This module uses a multi-layered filtering logic based on geometric constraints, map topology, and distance priority to automatically extract multi-vehicle interaction groups of the central vehicle at each time step, providing structured input for subsequent hazardous scenario generation. When a significant change in the behavior or position of any vehicle leads to a reconstruction of the interaction relationships, the original interaction group is disrupted, and the system automatically enters a new interaction phase, thus achieving the modeling of the "dynamic interaction evolution" in real traffic systems. An interaction group partitioning diagram is shown below. Figure 3 As shown.

[0034] b) Optimize search based on key scenarios using importance sampling: In generating hazardous scenarios for autonomous driving, a key challenge lies in efficiently identifying high-risk, low-probability, but high-impact test scenarios within a vast scene space. Traditional Monte Carlo methods rely on random sampling, whose samples are mostly concentrated on common safe scenarios, making it difficult to cover rare hazardous scenarios, thus resulting in low testing efficiency.

[0035] Importance sampling significantly improves the sampling efficiency for rare events by reconstructing the sampling distribution to favor "critical areas" (high-risk regions). Compared to traditional methods, importance sampling focuses on high-risk scenarios, reducing redundant samples, lowering variance, improving estimation accuracy, and accelerating the convergence of indicators such as accident rate and risk exposure. In the field of autonomous driving, this method is often used to accelerate the simulation of rare events and the generation of critical test scenarios.

[0036] 1. Importance sampling principle: Assume the probability distribution of the test scenario is as follows The performance metric function for the scenario is: Its mathematical expectation is: (1) In coarse Monte Carlo, the sample The estimator is: (2) However, when the probability of a critical event occurring is extremely low, the estimate converges slowly.

[0037] Therefore, an auxiliary distribution is introduced. This is called the importance function: (3) in, For importance weights, the estimator of the importance sampling method is: (4) when The closer to the ideal distribution (i.e., the concentrated area of ​​key events), the smaller the sampling variance and the faster the convergence.

[0038] 2. Construction of the scene importance function: In ISO 34505, risk scenario assessment is defined as a combination of scenario hazard level, exposure level, and complexity. Scenario importance is composed of two factors: scenario hazard level and exposure frequency. The scenario importance function is defined as follows: (5) in, These are scenario-based decision variables. It is the running design domain parameter, It refers to the degree of danger, that is, the probability of the target event occurring in the scene. It is the exposure frequency, which represents the probability of a scenario occurring in a real traffic environment.

[0039] The minimum normalized enhanced collision time (mnpETTC) is selected as the hazard level indicator: (6) in, The vertical relative distance, The velocity is relative. After normalization, we obtain: (7) (8) The minimum value is selected as the indicator of the level of danger in the scenario. The smaller the mnpETTC, the higher the collision risk.

[0040] The exposure frequency metric is estimated by the distance between the scene and the "set of public high-frequency scenes": (9) in, Indicates high-exposure areas. Normalization factor for each dimension To guide the search direction and improve the efficiency of key scene extraction, an auxiliary objective function is defined: (10) in, This is a weighting factor used to balance hazard and exposure indicators. It is achieved by minimizing... This allows for the search to identify the most critical test scenarios that are both high-risk and moderately expose the most vulnerable. The search process employs an optimization algorithm, iteratively extracting scenarios from the original scenario library that meet the requirements. High-importance samples are used to generate a refined set of key test scenarios.

[0041] 3. Hazardous scenario generation based on PPO: After key scenario extraction through interactive groups and screening based on importance sampling, a set of high-value key scenarios was obtained. In order to generate more adversarial and diverse dangerous scenarios for evaluating the robustness of the autonomous driving system, this module uses reinforcement learning (PPO) to train the background vehicle agent. Through the learned policy, the central vehicle is induced into a dangerous state within a physically acceptable range, thereby generating dynamic dangerous scenarios.

[0042] 1. PPO Algorithm Principle Proximal Policy Optimization (PPO) is a proximal policy optimization algorithm, belonging to the policy gradient method in reinforcement learning, used to solve reinforcement learning problems in continuous or discrete action spaces. Its core idea is to limit the magnitude of policy updates during updates, avoiding performance degradation due to excessive policy changes, thus achieving a good balance between stability and sample efficiency.

[0043] The key to PPO is using clipping to limit the policy ratio in each update, thereby ensuring training stability. A commonly used objective function is: (11) in, Indicates the probability ratio of the strategies. For the dominant function, To update the limit parameters, It is a clipping function that will Limitations between.

[0044] PPO is easily extended to multi-agent scenarios and is friendly to continuous action spaces. Through the above constraints, PPO can effectively learn high-risk behavioral patterns induced in complex traffic scenarios while maintaining the stability of policy updates.

[0045] 2. Reinforcement Learning Model Framework In scene generation, the PPO agent acts as a virtual participant that interferes with or triggers risks. The overall reinforcement learning model is as follows: 1) State Space The state vector contains the real-time motion state of all participants in the interaction group, the local road topology, the traffic light state, the relative distance matrix, and the predicted future trajectory.

[0046] 2) Action Space The motion vector is a continuously controlled variable, including longitudinal acceleration. Target velocity deviation Horizontal offset Lane change intention parameters The agent selects actions within permissible dynamic constraints to influence traffic evolution.

[0047] 3) Reward function design The reward function comprehensively considers three aspects: risk, rationality, and diversity. Common structure: (12) in, It is a risk score calculated based on TTC and minimum relative distance. Ensure that the evolution of risk scenarios is reasonable and continuous. Suppressing irrational actions, and Penalties for violating traffic rules.

[0048] 3. Driving style modeling In terms of driving style, based on speed, acceleration, and lane change frequency, drivers are categorized into aggressive, conservative, and moderate styles. A driving style clustering diagram is shown below. Figure 4 As shown.

[0049] To generate dangerous scenarios with diverse behavioral characteristics, this study models driving styles from both the strategy and reward levels.

[0050] 1) Rewards Reward function To further refine this, the feature vector is defined as: (13) in, It is based on the immediate danger level of Time-To-Collision. It is the risk level related to the minimum clearance distance. It is the relative velocity close to an instantaneous moment. For task completion rate, For comfort indicators, Used to encourage a gradual evolution of danger. It is a physical violation penalty. It is a penalty for violating traffic regulations.

[0051] Driving styles are divided into three types The weights consist of the baseline weights and the mapping matrix.

[0052] (14) 2) Strategy Level To generate more realistic and diverse scenarios and evaluate the system's robustness under different driving styles, driving style parameters are used as policy inputs or context variables: (15) in, This is the expanded system status. For time steps The original environmental state, Style encoding is represented by one-hot encoding embedding, which enables the same strategy to generate diverse behaviors in different style contexts.

[0053] 4. Multi-vehicle joint trajectory prediction In real-world scenarios, vehicles constitute a dynamic interactive system. Independent decision-making by a single vehicle makes it difficult to capture coordinated behaviors such as following, yielding, and lane changing, easily leading to trajectory conflicts or inconsistent behavior. Therefore, an end-to-end framework of "interaction encoding – joint decision-making – trajectory decoding" is adopted to ensure the consistency and rationality of multi-vehicle trajectories by modeling vehicle-to-vehicle and vehicle-to-infrastructure interactions. 1) Interactive encoder To depict the dynamic interaction relationships between multiple vehicles, each vehicle is considered as a node in a graph, and the node's characteristic is its current state. (Position, velocity, heading, acceleration, etc.), while edge features represent relative motion relationships (such as relative distance, relative velocity). A graph neural network (GNN) structure is used to encode the hidden vectors of the nodes: (16) in, Indicates vehicle Based on the interactive features, the encoder can employ a self-attention module with a graph attention mechanism to highlight key interactive pairs. As a map structural feature, lane topology relationships are represented by a directed graph, incorporating environmental constraints such as road speed limits and lane line types.

[0054] Subsequently, the hidden representations of all vehicles are aggregated to obtain global interaction features: (17) This global representation reflects the overall risk situation in the current scenario, providing contextual information for joint policy learning.

[0055] 2) Joint Strategy Head The goal of hazardous scenario generation is to learn strategies from a set of vehicles (especially background vehicles) to expose the main vehicle to high-risk but reasonable interaction scenarios. This is based on global interaction features. The policy network outputs the joint action distribution of multiple vehicles: (18) in, The driving style vector can be learned through an additional style encoder to adjust the mean and variance of the action distribution, so that the strategy can generate differentiated scenarios under different driving personalities.

[0056] For strongly coupled actions (such as merging and rear-end collision scenarios), a central controller structure is used to directly generate the joint action distribution, thereby better modeling the collaborative behavior under complex interactions: (19) 3) Track Decoder The combined action sequence is processed by a vehicle dynamics model or a neural network dynamics decoding module to obtain predictions of the future trajectories of multiple vehicles: (20) The trajectory decoder can incorporate kinematic constraints (such as nonholonomic constraints or acceleration boundaries) to ensure that the generated trajectory is physically feasible.

[0057] At each time step or throughout the entire prediction period, risk indicators are calculated based on the predicted trajectory, including: (twenty one) in, The minimum time interval (Time-to-Collision). Let be the collision probability. This indicates the safety margin between the main vehicle and the background vehicle.

[0058] The generator learns through PPO to maximize the "danger" reward while satisfying physical rationality constraints, so that the generated interactive scenarios can both induce potential risks and maintain a realistic driving style.

[0059] The scenario generation module, based on the PPO hazardous scenario generation method, can generate highly challenging dynamic scenarios by utilizing conditional driving styles and multi-vehicle joint trajectory modeling while preserving physical rationality, thereby meeting the simulation testing needs of autonomous transportation systems.

[0060] The scene reproduction layer includes traffic scene visualization software, high-precision vehicle dynamics simulation software, and hybrid traffic flow simulation software integrated through a unified interface specification; Furthermore, the unified interface specification includes: A first communication protocol and module for clock synchronization and pose data exchange with the traffic scene visualization software; A second communication protocol and module for initializing traffic flow and synchronizing state with the hybrid traffic flow simulation software; A third interface module for issuing control commands and providing vehicle status feedback to the high-precision vehicle dynamics simulation software.

[0061] Furthermore, the first communication protocol and module includes a tire contact point information processing submodule, which is used to send tire coordinates to the traffic scene visualization software and receive the corresponding road surface elevation and friction coefficient data returned by the software, and provide them to the high-precision vehicle dynamics simulation software.

[0062] Furthermore, the hybrid traffic flow simulation software is equipped with a parameterized driver behavior model, which can simulate various driving styles, including aggressive, conservative, and moderate driving styles.

[0063] Specifically, the core task of the Scene Reproduction Module (VTD) is to visualize the traffic scene given by the Scene Generation Module, including road alignment, scene environment, weather conditions, traffic facilities and equipment, and to visualize vehicle pose information. The model can be configured with diverse simulated road scenes according to the requirements of different types of tests.

[0064] A complete road scene consists of three files: XML, XODR, and OSGB files. The XML file is the configuration file for the simulation scene. It primarily defines the paths to road map files and 3D model files, the parameters of the test vehicle, its initial position and control method, and the placement of static obstacles such as stakes. This file should be stored in the path: / Data / Projects / ProjectsName / Scenarios / XODR files describe the geometry of roads in a simulation scenario.

[0065] The OSGB file is used to define the 3D visualization model of the scene. These two files are related and should be stored together in the following path: Data / Distros / Current / Databases / ScenarioName /

[0066] The joint simulation control center, implemented based on control logic modeling software, is used to coordinate the data flow and simulation steps between the scene generation layer, scene reproduction layer and simulation test layer, and integrates a vehicle control unit model, an electronic stability program model and a roadside information transmission simulation module. The platform is configured to implement an autonomous transportation system testing process, from scene generation and multidisciplinary high-fidelity reproduction to closed-loop verification of safety control algorithms.

[0067] Furthermore, the vehicle control unit model is used to calculate the driving torque based on the vehicle status and driving commands; the electronic stability program model uses a fuzzy PID controller to calculate the yaw moment compensation value to maintain vehicle stability.

[0068] Furthermore, the roadside information transmission simulation module is used to simulate and generate, and send incomplete traffic environment perception information to the downstream vehicle decision model, based on preset roadside equipment attributes and environmental conditions.

[0069] Furthermore, the simulation test layer supports the access of external safety management algorithm models in the form of functional model units, and can partially or completely take over the driver decision-making functions in the joint simulation control center.

[0070] Specifically, the core task of the Vehicle Dynamics module (CarSim) is to simulate and analyze the dynamic response of vehicles under different operating conditions in autonomous traffic scenarios, including acceleration, braking, steering, and suspension behavior. When Simulink and CarSim are used for co-simulation, CarSim, as the dynamics model, requires inputs regarding vehicle handling (steering wheel, accelerator, brakes, etc.), vehicle state (four-wheel z-axis information), and road condition-related information (friction coefficient). Through mathematical model calculations, it can output vehicle state information (vehicle coordinates, pose, acceleration, engine speed, etc.).

[0071] Mixed traffic flow refers to traffic flow composed of different types of vehicles, including but not limited to cars, motorcycles, bicycles, pedestrians, and public transportation. These vehicles operate simultaneously on the same routes or in nearby areas, forming a complex dynamic system. The core task of the Mixed Traffic Flow Module (SUMO) is to construct and simulate mixed traffic flows in traffic scenarios, including simulated travel paths, driving behaviors, and lane-changing behaviors of various traffic participants according to traffic rules, in order to create an operating environment for controlled vehicles within the traffic scenario. When Simulink and SUMO perform joint simulation, pre-set mixed traffic flow settings and traffic scenario information need to be input. Through mathematical model calculation and simulation, the position and attitude information (position coordinates, heading angle, acceleration, etc.) of each vehicle in the background traffic flow at each time point is output.

[0072] In the simulation of autonomous transportation system operation, the Simulink-VCU module mainly undertakes the function of torque calculation.

[0073] Torque calculation: Based on the current vehicle speed and accelerator pedal opening, the corresponding value is looked up in the two-dimensional interpolation table to calculate the driving torque in the forward state; on this basis, combined with the battery torque limit and the motor torque limit, the output torque is constrained, and finally the target torque required by the motor is generated.

[0074] The core function of the Simulink-ESP module is to ensure the stability of controlled vehicles in traffic scenarios. Its structure mainly includes a vehicle dynamics subsystem, a fuzzy PID control system, and a torque distribution subsystem. (1) Vehicle dynamics subsystem: The two-degree-of-freedom model in the vehicle reference model is selected as the ideal model to reflect the linear relationship between the driver's steering input and the vehicle's lateral acceleration, center of gravity sideslip angle and yaw rate.

[0075] (2) Fuzzy PID control system: Based on the analysis of the characteristics of the ESP system, the difference between the actual yaw rate and the ideal value, and the difference between the actual center of gravity sideslip angle and the ideal value are selected as input variables. The fuzzy controller outputs the yaw torque adjustment amount M, which is further processed by the PID controller to output the final yaw torque compensation value ΔM.

[0076] (3) Torque distribution subsystem: The braking side is determined based on the difference between the actual and ideal values ​​of the yaw rate. When the difference is greater than zero, the right wheel is braked; when the difference is less than zero, the left wheel is braked, thereby achieving effective compensation for the vehicle's yaw moment.

[0077] In addition, the ESP module involves several key vehicle parameters imported from the outside, including vehicle mass m, distance a from the center of gravity to the front axle, distance b from the center of gravity to the rear axle, front axle lateral stiffness k1, rear axle lateral stiffness k2, distance between the front and rear axles L, front wheel track Bf, rear wheel track Br, and road adhesion coefficient μ.

[0078] The Simulink driver model consists of two independent modules: lateral control and longitudinal control. These modules do not interfere with each other.

[0079] (1) Horizontal control module The standard lateral control module consists of three main parts: the FMU component, the Stanley lateral control subsystem, and the vehicle orientation change function. The functions of each component are described below: FMU Component: This component is associated with an external FMU file. Its main function is to calculate the positional and heading errors between the vehicle and the expected trajectory based on the vehicle's current position, speed, and yaw angle. In handling stability tests, the expected trajectory is dynamically generated based on the road centerline and stake positions; in dynamic tests, the expected trajectory is embedded in the FMU as a static trajectory. To change the expected trajectory, use a decompression tool such as 7-Zip to open the corresponding FMU file and replace the data.txt file in the resources folder. Furthermore, double-clicking the FMU component in Simulink allows you to view its control parameters. Taking the slalom test as an example, the parameters include simulation step size, delta, and k; the latter two parameters are used to adjust the amplitude of the vehicle's slalom maneuver.

[0080] Stanley's lateral control subsystem calculates the corresponding steering wheel angle based on the pose error output by the FMU and the vehicle's current speed and attitude. This subsystem requires external parameters, primarily for heading control calculations. Internally, the subsystem contains several adjustable gain parameters, as shown in the figure below, which are used to adjust the vehicle's steering sensitivity.

[0081] Lane change control component: Based on the vehicle's current lane and speed, it controls whether the vehicle changes lanes according to the set logic.

[0082] Vehicle heading conversion: Since the heading angle in CarSim is defined as 0 degrees due east and counterclockwise as the positive direction, while the lateral control algorithm uses 0 degrees due north and clockwise as the positive direction, it is necessary to convert the vehicle yaw angle output by CarSim to ensure coordinate system consistency.

[0083] (2) Longitudinal control module The main function of the longitudinal control module is to control the accelerator pedal opening to make the current vehicle speed as close as possible to the desired vehicle speed. The core of this module is the LonFeedforwardPI longitudinal control subsystem.

[0084] The LonFeedforwardPI longitudinal control subsystem: This subsystem, based on the current vehicle speed, acceleration, and expected vehicle speed, comprehensively considers driving resistance and feedback control mechanisms to output the accelerator pedal opening, achieving the desired acceleration effect. This subsystem requires external parameters, including vehicle mass, aerodynamic / rolling drag coefficients, and controller parameters.

[0085] Meanwhile, considering the need for testing the safety management algorithm of the autonomous transportation system, the decision-making part of the driver module can be replaced by an external safety management algorithm in the form of an FMU, or linked with the safety management algorithm, allowing the safety management algorithm to take over when the vehicle enters a risk / collision scenario, so as to achieve conflict resolution and safe operation of the transportation system.

[0086] The Trigger module, as the startup module for co-simulation, primarily packages simulation-related information for each simulation step and sends it to the VTD via port 48191 to initiate the VTD simulation process. It allows setting information such as simulation step size, pulse width, and sending port. The Trigger module and its corresponding S-Function in Simulink, such as... Figure 5 As shown.

[0087] The vehicle state information sending module is responsible for combining and packaging the vehicle pose information calculated by the vehicle dynamics module, including vehicle coordinates (rear axle center point coordinates), vehicle velocity (x, y, z), vehicle acceleration (x, y, z), pitch angle, yaw angle, and roll angle, and sending it to the VTD via port 48191 to update the vehicle pose information in the VTD. It is also responsible for transmitting the background traffic participant information simulated in the mixed traffic flow module. The vehicle information sending module in Simulink, such as... Figure 6 As shown.

[0088] The core of the vehicle information receiving module is to receive the vehicle pose information transmitted by the vehicle dynamics module and send it to the driver model and the vehicle dynamics model respectively as input for the next simulation step.

[0089] The core of the tire contact point information transmission module is to collect the x and y coordinates of the current vehicle's four wheels, package them, and send them via port 48271 to the VTD's odrgateway tool. This tool then performs coordinate retrieval to obtain the z-coordinate and friction coefficient of the corresponding point on the map. The tire contact point information transmission module in Simulink, such as... Figure 7 As shown.

[0090] The core of the tire contact point information receiving module is to read the VTD48270 port information and, by extracting the string, obtain the z-coordinate and friction coefficient of the corresponding point on the map corresponding to the four-wheel xy coordinates sent by the tire contact point information sending module in the previous simulation step. These are then parsed and sent to the vehicle dynamics module. The tire contact point information receiving module in Simulink, as shown... Figure 8 As shown.

[0091] (1) Carsim data input module This module is used to package all the input parameters required for the vehicle dynamics model (CarSim) and provide a unified simulation interface. The input parameters include: acceleration torque, brake pedal opening, steering wheel angle, four-wheel contact point height, and four-wheel contact point friction coefficients in the x / y axis directions.

[0092] This module packages all the input parameters required for the vehicle dynamics model (CarSim), providing a unified simulation interface. Input parameters include: acceleration torque, brake pedal opening, steering wheel angle, four-wheel contact point height, and four-wheel contact point friction coefficients in the x / y axes. The Carsim data input module, such as... Figure 9 As shown.

[0093] (2) Carsim data output and conversion module This module is used to generate data labels and convert formats for various parameters output from the vehicle dynamics model, so that subsequent modules can call and analyze them. In the handling stability test, the output data includes: vehicle coordinates (X, Y, Z axes), vehicle speed components (X, Y, Z axes), acceleration (longitudinal and lateral), pitch angle, roll angle, yaw angle, yaw rate, center of gravity sideslip angle, and four-wheel coordinates (X, Y axes). In addition, some parameters are converted, including: Vehicle speed synthesis: Calculate the resultant speed based on the speed components of the X and Y axes, which is then transmitted to the driver model, VCU, and ESP modules; Front and rear axle center point calculation: The position of the front and rear axle center points is calculated based on the four-wheel coordinates and transmitted to VTD and boundary conditions (only the rear axle center point is calculated in the dynamic test). VTD Data Packaging: Packs the key parameters that need to be transmitted to the VTD, including the acceleration components in each direction of the global coordinate system calculated based on the vehicle speed components (X, Y, Z).

[0094] The internal details of the module, including output data and data processing methods, can be freely adjusted according to the experiment type and the required data type. This platform constructs a corresponding Carsim data output and conversion module structure, such as... Figure 10 As shown.

[0095] SUMO data output and conversion module structure, such as Figure 11 As shown. The SUMO data input and output module is used to establish communication between SUMO and Simulink, and it is mainly divided into two parts: an information sending module and an information receiving module; 1) Information Sending Module: The information sending module is responsible for writing the mixed traffic flow information to be simulated, including the number of participants, the initial position, attitude, speed / acceleration, trajectory path, driving style, etc., and packaging it into a SUMO-compatible data format. This data is then sent to SUMO's Traci module via UDP to complete the initialization of the mixed traffic flow. Simultaneously, it sends triggers to SUMO to initiate each step of the simulation, ensuring spatiotemporal synchronization between SUMO and other simulation software.

[0096] 2) Information receiving module: The information receiving module is responsible for receiving the position and attitude information of each traffic participant in the mixed traffic flow obtained from each step of the SUMO simulation, including position coordinates, velocity, acceleration, heading angle, etc., and synchronizing it to VTD.

[0097] The roadside information transmission module simulates the traffic scene information and mixed traffic flow information that each roadside device may receive and can send to vehicles, based on the location, type, and level of the roadside devices placed in the traffic scene module, combined with other scene settings, including weather, visibility, and traffic environment. This information serves as the basis for vehicle decision-making and control. This module can simulate incomplete road information and information about uncontrolled vehicles and other traffic participants that roadside devices may receive, offering higher realism compared to the usual full input of scene information. The core function of the boundary conditions module is to set start and end conditions for simulation experiments, or to define the range of experimental data recording through logical condition judgments.

[0098] In this embodiment, the platform also includes a data management and analysis module, which is used to record simulation process data and perform calculation analysis and visualization output on the test results based on preset evaluation indicators.

[0099] The main function of the data output and analysis module is to output process data during the experiment and calculate various index values ​​required for simulation testing, based on boundary conditions. The output data is saved in .m and .excel formats, and the index values ​​are presented in chart form. During index calculation, the data to be calculated is first selected according to the boundary conditions, and then the index values ​​are calculated and visualized through data extraction and calculation.

[0100] This platform can build an autonomous traffic scenario library according to given requirements and reproduce it with high accuracy, realizing simulation testing of autonomous traffic systems based on multiple disciplines, including but not limited to risk identification algorithm testing, such as vehicle risk identification and illegal behavior judgment algorithms; and safety control algorithm testing, such as vehicle platooning operation control algorithms and multi-agent conflict resolution algorithms.

[0101] The operational steps for simulating an autonomous traffic system based on an autonomous traffic scene generation and reproduction platform are as follows. The operation flowchart is shown below. Figure 12 As shown, first, start Simulink, generate the corresponding test scenario library based on the runtime requirements or algorithm testing requirements, and import VTD and SUMO for initialization; then start VTD, confirm the settings, select and load the scenario required for testing; then start SUMO, confirm that SUMO successfully listens to the relevant ports in Simulink and loads the initial traffic flow; then start CarSim, select the vehicle dynamics model parameter settings, initial pose, and input / output corresponding to the simulation test, and send the model to Simulink; finally, complete the VTD runtime configuration and click the run button in Simulink to start the simulation test.

[0102] The technical effects of this embodiment are mainly reflected in the following aspects: 1. More accurate and realistic scene generation. This invention generates dynamic hazardous scenarios with high risk and high challenge through interactive group modeling, importance sampling, and reinforcement learning methods. This supports the systematic testing and robustness evaluation of autonomous traffic system safety management algorithms under complex traffic conditions. Compared with existing traffic scene generation technologies, this invention is more targeted, generates risk / collision scenarios more frequently, and has autonomous level elements, enabling it to serve the simulation operation of autonomous traffic systems.

[0103] 2. More comprehensive simulation scale coverage. This invention can realize simulation mapping of autonomous transportation systems at multiple scales, including micro, meso, and macro scales. It can support the modeling and simulation of complex traffic scenarios based on mixed traffic flows, and can generate a traffic scenario library for refined vehicle dynamics and vehicle decision-making and control simulations. It offers higher precision, a wider range of applicable scenarios, and greater flexibility. 3. Significant advantages of multi-software simulation modeling. This invention integrates multiple software programs, including VTD, Simulink, Carsim, SUMO, and Python, covering multidisciplinary modeling and joint simulation of traffic scene automatic generation, traffic scene and sensor simulation, vehicle dynamics, control systems, mixed traffic flow, and roadside equipment information transmission. This addresses the shortcomings of current autonomous traffic scene generation and reproduction methods, which suffer from incomplete coverage and weak realism. Furthermore, a unified data interface and mapping mechanism are proposed to resolve differences in data formats, spatial coordinate systems, and communication methods between different software programs. Through a strategy based on delay decoupling and optimized scheduling, the simulation instability caused by algebraic loops is effectively eliminated, improving simulation efficiency and convergence.

[0104] 4. End-to-end verification capability for autonomous transportation systems. This invention not only supports traffic scene generation and multi-software scene reproduction, but also deeply couples with vehicle control systems, driver decision-making, and mixed traffic flow, thereby achieving high-fidelity simulation and verification of the entire chain of "perception-decision-control-execution" for autonomous transportation systems, which is more complete than CarMaker and more refined than PanoSIM.

[0105] Example 2 This embodiment takes an actual operation case of autonomous transportation system simulation test as an example, and specifically includes the following through the technical solution of Embodiment 1: (1) Based on simulation testing or operational requirements, run the traffic scenario generation module, construct a test scenario library tailored to the needs, and import VTD and SUMO. The dataset used is the Waymo dataset, generating hazardous scenarios based on reinforcement learning, such as... Figure 13 As shown.

[0106] (2) According to the simulation test requirements, import the scenario required for simulation test into the scenario model (VTD) and construct the test traffic scenario based on VTD.

[0107] (3) According to the test requirements, integrate the scenario model, vehicle dynamics model, mixed traffic flow model, driver model, VCU, and ESP models; write corresponding Python programs to monitor Simulink data and send SUMO simulation data; improve the VCU model to calculate the driving torque output when moving forward by querying a two-dimensional interpolation table using the current vehicle speed and accelerator pedal opening. The Simulink interface of the autonomous transportation system simulation test platform is shown below. Figure 14 As shown.

[0108] (4) Based on the coordinates of the center points of the front and rear axles and the vehicle size parameters (length and width), calculate the vehicle outline position in real time, and make the following judgments based on the actual dynamic positions of each traffic participant in the scene and the unknown parameters of fixed objects / obstacles in the scene: a. When the vehicle outline collides with any obstacle / traffic participant, set the current frame's output_collision=1 (this will not directly terminate the simulation); the core idea of ​​this collision detection algorithm is to calculate the vertical distance (dist_long) from other traffic participants and fixed objects to the vehicle's longitudinal axis and the vertical distance (dist_lat) to the lateral axis. A collision is determined to have occurred when dist_lat < half the vehicle's width and dist_long < half the vehicle's length.

[0109] b. When the controlled vehicle is within the test area, set the current frame's output_inrange=1; c. The simulation process is terminated once the vehicle leaves the test area.

[0110] (5) Based on the operation simulation test, the operation of the autonomous transportation system based on VTD and SUMO is visualized.

[0111] (6) After the simulation test is completed, the test output data is extracted, processed and analyzed, and finally the performance indicators required by the test are obtained and the effect is evaluated. Finally, an evaluation report is generated for algorithm optimization and iteration.

[0112] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An autonomous traffic scene generation and reproduction system based on multi-software collaboration, characterized in that, include: An integrated simulation framework, comprising a scene generation layer, a scene reproduction layer, and a simulation testing layer that operate sequentially and collaboratively; The scene generation layer is used to generate a traffic scene library for system-level collaborative testing based on real traffic data, using a hybrid method that includes interaction group extraction, importance sampling optimization search, and near-end strategy optimization reinforcement learning. The traffic scene library contains a structured description of static road networks, dynamic traffic participants, environmental conditions, and information interaction elements. The hybrid approach used in the scene generation layer specifically includes: extracting interaction groups containing multiple associated traffic participants based on the location of the central vehicle, lane topology, and preset spatial range; constructing an importance function based on the scene's hazard level and exposure frequency, and using the importance function to guide optimization search to select key scene seeds from the initial scene set; using the key scene seeds as the initial environment, training the background traffic participant agent using a proximal policy optimization algorithm, and generating dynamic dangerous interaction scenes by maximizing the reward function that incorporates driving style parameters; The scene reproduction layer includes traffic scene visualization software, high-precision vehicle dynamics simulation software, and hybrid traffic flow simulation software integrated through a unified interface specification; The joint simulation control center, implemented based on control logic modeling software, is used to coordinate the data flow and simulation steps between the scene generation layer, scene reproduction layer and simulation test layer, and integrates a vehicle control unit model, an electronic stability program model and a roadside information transmission simulation module. The system is configured to implement an autonomous transportation system testing process, from scene generation and multidisciplinary high-fidelity reproduction to closed-loop verification of safety control algorithms.

2. The system according to claim 1, characterized in that, When generating dynamic dangerous interaction scenarios, a multi-vehicle joint trajectory prediction model based on graph neural networks is adopted. The multi-vehicle joint trajectory prediction model obtains global interaction features through an interaction encoder, and outputs the cooperative actions of multiple vehicles in the interaction group through a joint strategy head. The joint trajectory conforming to physical constraints is obtained through a trajectory decoder.

3. The system according to claim 1, characterized in that, The unified interface specification includes: A first communication protocol and module for clock synchronization and pose data exchange with the traffic scene visualization software; A second communication protocol and module for initializing traffic flow and synchronizing state with the hybrid traffic flow simulation software; A third interface module for issuing control commands and providing vehicle status feedback to the high-precision vehicle dynamics simulation software.

4. The system according to claim 3, characterized in that, The first communication protocol and module includes a tire contact point information processing submodule, which is used to send tire coordinates to the traffic scene visualization software and receive the corresponding road surface elevation and friction coefficient data returned by the software, and provide them to the high-precision vehicle dynamics simulation software.

5. The system according to claim 1, characterized in that, The vehicle control unit model is used to calculate the driving torque based on the vehicle status and driving commands; the electronic stability program model uses a fuzzy PID controller to calculate the yaw moment compensation value to maintain vehicle stability.

6. The system according to claim 1, characterized in that, The roadside information transmission simulation module is used to simulate and generate incomplete traffic environment perception information and send it to the downstream vehicle decision model based on preset roadside equipment attributes and environmental conditions.

7. The system according to claim 1, characterized in that, The hybrid traffic flow simulation software is equipped with a parameterized driver behavior model, which can simulate various driving styles, including aggressive, conservative, and moderate driving styles.

8. The system according to claim 1, characterized in that, The simulation test layer supports the access of external safety management algorithm models in the form of functional model units, and can partially or completely take over the driver decision-making functions in the joint simulation control center.

9. The system according to claim 1, characterized in that, The system also includes a data management and analysis module, which records simulation process data and performs calculations, analysis, and visualization output on the test results based on preset evaluation indicators.

Images