A traffic management road test robot simulation method based on a spatial intelligent world model

Abstract

The application relates to the technical field of robot simulation testing, and discloses a traffic management road test robot simulation method based on a spatial intelligent world model, which comprises the following steps: firstly, acquiring multi-modal environment data collected by a traffic management road test robot in a target road area, and constructing a three-dimensional spatial intelligent world model comprising a geometry layer, a semantic layer, a physical layer, a rule layer, a dynamic layer and an environment layer; in the model, the physical layer, the environment layer and the dynamic layer are parameterized configured, and a simulation test scene comprising road conditions, environment conditions, dynamic traffic participants and behavior trajectories of the dynamic traffic participants is generated under the constraint of the rule layer; a digital twin of the robot is controlled to execute a simulation test task and obtain an execution result, and a test evaluation result is generated to guide actual operation of the traffic management road test robot. Through the three-dimensional spatial intelligent world model and parameterized scene generation, efficient simulation testing and safe operation guidance of the traffic management road test robot in a complex traffic environment are realized.

Description

Technical Field

[0001] This application relates to the field of robot simulation and testing technology, and in particular to a simulation method for traffic control road testing robots based on a spatial intelligent world model. Background Technology

[0002] Traffic management road testing robots are widely used for tasks such as road traffic facility inspection, intersection traffic condition monitoring, traffic violation assistance, and road anomaly detection. The algorithm verification and performance evaluation of existing traffic management road testing robots mainly rely on real-world road testing. However, real-world road testing suffers from high testing costs, significant risks, insufficient scenario coverage, and difficulty in reproducing extreme weather and sudden traffic events. For example, robots need to maintain stable perception, localization, planning, and control capabilities under complex conditions such as potholes, water accumulation, slopes, slippery roads, nighttime conditions, rain, snow, fog, mixed pedestrian and vehicle traffic, non-motorized vehicle weaving, and pedestrian obstruction. However, these scenarios are difficult to construct in batches on real roads and are difficult to test repeatedly, resulting in low robot system verification efficiency and high safety risks.

[0003] In existing technologies, several simulation systems are used for testing autonomous vehicles or general-purpose robots. For example, some solutions automatically search for and generate accident scenarios in road accident areas using high-precision map processing and swarm intelligence algorithms. However, this method is mainly geared towards accident prevention testing for autonomous vehicles, and since the scenarios are automatically generated by algorithms, their correspondence with real roads is weak, making it difficult to recreate the complex topological structure of real intersections. Another example is a solution that adjusts the size of the simulation environment by obtaining the offset between the robot's actual and estimated positions. However, this method is only suitable for adjusting the size of simple geometric environments and lacks the interactivity and randomness similar to real traffic flow.

[0004] Therefore, there is an urgent need for a simulation testing method that can improve the realism of road traffic scenarios. Summary of the Invention

[0005] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a simulation method for traffic management road testing robots based on a spatial intelligent world model, in order to solve the technical problems in the prior art such as insufficient reproduction of simulation scenarios with real roads, weak simulation capability of road surface micro-conditions, lack of real interaction of dynamic traffic participant behavior, difficulty in using simulation results to guide real road testing, and lack of a closed-loop construction mechanism from real data collection to simulation testing.

[0006] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted in this application include: This application provides a simulation method for traffic control roadside testing robots based on a spatial intelligent world model, the specific steps of which include: S1, acquire multimodal environmental data collected by the traffic management road testing robot in the target road area; Based on the multimodal environment data, a three-dimensional spatial intelligent world model of the target road area is constructed; the three-dimensional spatial intelligent world model includes a geometric layer, a semantic layer, a physical layer, a rule layer, a dynamic layer, and an environment layer, which are used to simulate the three-dimensional scene of the target road area and its traffic rules, physical attributes, and the behavior of dynamic traffic participants; S2, In the three-dimensional intelligent world model, based on the geometric layer and semantic layer, multiple simulation test scenarios are generated by parameterizing the physical layer, environmental layer and dynamic layer, and under the constraints of the rule layer, including road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories. S3, the digital twin of the traffic management road test robot is controlled to perform simulation test tasks in the multiple simulation test scenarios respectively, and generate corresponding test evaluation results based on the simulation test tasks; S4. Based on the test evaluation results, generate the safety operation parameter configuration of the traffic management road test robot under multiple simulation test scenarios in the target road area.

[0007] Optionally, in some embodiments of this application, constructing a three-dimensional spatial intelligent world model of the target road area based on the multimodal environment data includes: The target road area is divided into spatial grids to obtain multiple grid cells; The multimodal environment data is mapped to the multiple grid cells according to their spatial location, so that each grid cell is associated with data corresponding to that spatial location in the geometric layer, semantic layer, physical layer, rule layer, dynamic layer and environment layer, thereby constructing the three-dimensional intelligent world model.

[0008] Optionally, in some embodiments of this application, the geometric layer includes three-dimensional structural data for describing roads, curbs, buildings, and traffic facilities; the semantic layer includes category and attribute data for describing lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs, and guardrails; the physical layer includes data for describing road surface friction coefficients, slopes, collision bodies, and rigid body properties; the rule layer includes data for describing lane travel directions, traffic light cycles, and traffic rules; the dynamic layer includes data for describing the position, speed, trajectory, and behavioral state of traffic participants; and the environmental layer includes data for describing weather conditions, lighting conditions, visibility, and road surface slippage.

[0009] Optionally, in some embodiments of this application, step S2 includes: In the three-dimensional intelligent world model, road conditions are generated based on configurable road condition parameters; environmental conditions are generated based on configurable weather and illumination parameters; and dynamic traffic participants and their behavioral trajectories are generated under the constraints of the rule layer based on configurable traffic flow and behavior model parameters. By combining different road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories, multi-condition combinations are generated, and corresponding simulation test scenarios are generated for each condition combination to execute simulation test tasks.

[0010] Optionally, in some embodiments of this application, step S2 further includes: Obtain the user's natural language scene description, perform semantic parsing on the natural language scene description through a large language model, and extract scene parameters, which include at least one of road condition parameters, weather parameters, lighting parameters, traffic flow parameters, and behavior model parameters; Based on the scenario parameters, the parameterized configuration of the road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories is executed to generate the corresponding simulation test scenario model.

[0011] Optionally, in some embodiments of this application, generating road conditions based on configurable road condition parameters includes: Based on at least one of the road condition parameters, the geometry and / or physical properties of the road surface are modified within the drivable area of ​​the traffic control and road testing robot in the three-dimensional intelligent world model to generate the road condition. The road condition parameters include pothole depth, pothole radius, slope angle, water accumulation depth, crack width, road surface friction coefficient, speed bump height, or manhole cover height difference.

[0012] Optionally, in some embodiments of this application, the environmental conditions in the three-dimensional intelligent world model include weather conditions, lighting conditions, and sensor degradation conditions. The sensor degradation conditions include simulating dirt on a camera lens, raindrop obstruction, or image blurring, as well as simulating point cloud attenuation or noise increase of a lidar in rainy or foggy weather.

[0013] Optionally, in some embodiments of this application, the three-dimensional spatial intelligent world model further includes a preset linkage rule base, which includes a mapping relationship between environmental layer parameter changes and physical layer parameters and dynamic layer parameters. The linkage rule base is used to automatically maintain the physical consistency between the corresponding parameters in the physical layer and the dynamic layer and the parameters in the environmental layer when the weather parameters or light parameters in the environmental layer change.

[0014] Optionally, in some embodiments of this application, in the three-dimensional intelligent world model, the behavioral trajectory includes abnormal behaviors that violate the constraints of the rule layer, and the abnormal behaviors include at least one of running a red light, driving against traffic, suddenly crossing the road, changing lanes illegally, cutting in at intersections, or non-motorized vehicles occupying motor vehicle lanes.

[0015] Optionally, in some embodiments of this application, the safe operation parameter configuration includes at least one of the following operation parameters: recommended driving speed and speed limit instructions under different road conditions and environmental conditions; recommended turning radius on road sections with different curvatures; and detour paths generated based on the location of obstacles in the target road area.

[0016] (III) Beneficial Effects This application discloses a traffic management road testing robot simulation method based on a spatial intelligent world model. The method involves a traffic management road testing robot collecting multimodal environmental data in a target road area and constructing a three-dimensional spatial intelligent world model containing geometric, semantic, physical, rule, dynamic, and environmental layers based on the multimodal environmental data. This method enables the simulation scene to be highly consistent with the real road environment and improves the simulation fidelity. By generating road conditions based on preset road condition parameters in the three-dimensional intelligent world model, adjusting weather and lighting parameters in the environment layer to generate environmental conditions, and generating dynamic traffic participants and their behavioral trajectories under the constraints of traffic rule layer in the three-dimensional intelligent world model based on preset traffic flow and behavior model parameters, a simulation test scenario is obtained. This allows for the parameterized and systematic generation of complex test scenarios, compensating for the deficiency that real road tests cannot reproduce extreme conditions. By controlling the digital twin of the traffic management road test robot to perform simulation test tasks in the simulated test scenario and obtaining the execution results, the robot's ability to perform specialized tasks such as road surface defect detection, traffic violation recognition, and traffic light recognition in complex traffic environments can be comprehensively evaluated. By generating test evaluation results based on the execution results, and optimizing the real road test strategy of the traffic management road test robot based on the test evaluation results to guide actual road tests, a closed-loop feedback mechanism for actual test optimization can be formed, reducing the cost and risk of actual road tests and improving test efficiency and safety. Attached Figure Description

[0017] Figure 1 This is a flowchart of a traffic control roadside robot simulation method based on a spatial intelligent world model, as described in this application. Figure 2 This is a flowchart illustrating the data processing and hierarchical construction of a three-dimensional intelligent world model in one embodiment of this application; Figure 3 This is a flowchart illustrating the generation of a multidimensional simulation test scenario in one embodiment of this application. Detailed Implementation

[0018] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0019] In existing technologies, simulation testing methods for traffic management roadside testing robots can be mainly categorized into the following three types: The first category is the pure real-world road testing method, which places the robot in a real road environment for algorithm verification and performance evaluation. This type of method has problems such as high testing costs, high risks, and insufficient scene coverage. Extreme or complex working conditions such as potholes, water accumulation, rain, snow, fog, nighttime, and mixed traffic of pedestrians and vehicles are difficult to construct in batches and repeat for testing, resulting in low efficiency of robot system verification and high safety risks.

[0020] The second category is parametric scene simulation methods. These methods utilize simulation platforms such as CARLA to generate road, weather, and traffic participant scenarios by manually setting or randomly sampling parameters. These methods typically rely on manually constructed scenarios or simple parameter combinations, resulting in insufficient fidelity in the geometric, semantic, and physical properties of the scenario models compared to real intersections. They also lack the ability to automatically construct high-fidelity world models based on real-world data. The behavior of dynamic traffic participants is mostly based on preset trajectories, lacking a systematic simulation of illegal behaviors such as running red lights and driving against traffic. Furthermore, the simulation results are difficult to use to guide the optimization of real robot road testing strategies, lacking a closed-loop feedback mechanism.

[0021] The third category is general digital twin simulation methods; these methods build digital twin models for industrial equipment or autonomous vehicles to conduct simulation testing or optimize control strategies. However, these methods lack a testing system designed for specific tasks of traffic management road testing robots (such as road defect detection, traffic violation evidence collection, and traffic light status recognition). The digital twin models are typically built based on simple operational data (such as temperature and voltage), lacking structured modeling of multi-level information such as road geometry, road surface physical properties, traffic rules, and ambient lighting. The generation of dynamic traffic participant behavior is not deeply coupled with the traffic rule layer, making it impossible to effectively test the robot's ability to recognize complex traffic events. Furthermore, they lack a closed-loop mechanism to automatically convert simulation evaluation results into actual test strategies.

[0022] To address this, this application provides a simulation method for a traffic management road test robot based on a spatial intelligent world model. The method first utilizes a traffic management road test robot to collect multimodal environmental data in a target road area, and then constructs a three-dimensional spatial intelligent world model of the target road area based on the collected data. On this basis, the method generates a simulation test scenario model containing different road conditions, environmental conditions, and dynamic traffic participant behavior trajectories by parameterizing the physical, environmental, and dynamic layers of the model and under the constraints of the rule layer. Then, a digital twin of the traffic management road test robot executes preset simulation test tasks in this simulation test scenario model, obtains the execution results, and generates test evaluation results. Finally, based on the test evaluation results, a safe operation parameter configuration for the traffic management road test robot in the target road area is generated to guide the actual operation of the traffic management road test robot in the target road area. This method replaces the traditional testing method that relies on real roads with simulation, reducing testing costs and safety risks, and improving test coverage and efficiency.

[0023] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0024] Example 1: The traffic management road test robot simulation method based on the spatial intelligent world model provided in this application embodiment can be deployed on cloud servers, edge computing nodes or local workstations to conduct comprehensive functional testing and performance evaluation of traffic management road test robots in a simulation environment, so as to guide the safe operation and task execution of traffic management road test robots in actual deployment. Figure 1 The flowchart below shows a simulation method for a traffic control road testing robot based on a spatial intelligent world model, according to this application. Figure 1 As shown, the traffic control road testing robot simulation method based on the spatial intelligent world model includes: Step S1: Obtain multimodal environmental data collected by the traffic management road testing robot in the target road area; construct a three-dimensional spatial intelligent world model of the target road area based on the multimodal environmental data.

[0025] Figure 2 This is a flowchart illustrating the data processing and hierarchical construction of the three-dimensional intelligent world model in this embodiment. In a preferred embodiment, the construction process of the three-dimensional intelligent world model includes the following steps: S1.1, Acquisition and spatiotemporal alignment fusion of multimodal environmental data; Traffic control roadside testing robots operate in target road areas, collecting multimodal environmental data through their onboard multi-sensor system. This multimodal environmental data includes at least two of the following: images captured by cameras, point clouds collected by lidar, inertial data collected by IMUs, and positioning data collected by RTK. During the data acquisition process, the RTK positioning data achieves centimeter-level accuracy (fixed solution state), providing a high-precision spatial position reference.

[0026] The multimodal environmental data is subjected to spatiotemporal alignment and fusion processing. Based on the timestamp of the RTK positioning data, the time synchronization of image, point cloud and inertial data is completed; the spatial transformation relationship between the image coordinate system, lidar coordinate system, IMU coordinate system and global coordinate system is unified.

[0027] In a preferred embodiment, the spatial transformation relationship is obtained through a self-calibration method in a non-calibrated space. Specifically, the geometric constraint features of road structural elements (including the coplanarity and parallelism of lane lines, curbs, and stop lines) are extracted from the multimodal environment data; based on the geometric constraint features, a set of geometric consistency equations across the screen is constructed, the intrinsic and extrinsic parameters of the camera are automatically solved, the two-dimensional image pixels are mapped to a unified three-dimensional spatial coordinate system, and joint calibration is performed with the lidar coordinate system, IMU coordinate system, and RTK global coordinate system to achieve spatial coordinate system unification of the multimodal environment data.

[0028] S1.2, Generation of geometric layers; Based on spatiotemporally aligned point cloud data, a road geometric mesh for the target road region is generated using a 3D reconstruction algorithm, resulting in the geometric layer of a 3D intelligent world model. The road geometric mesh includes the 3D structure of the road surface, curbs, buildings, and traffic facilities.

[0029] In a preferred embodiment, during the reconstruction of road geometry using multimodal environmental data, geometric constraint features (including at least one of lane lines, curbs, and stop lines) of road structural elements in the multimodal environmental data are simultaneously extracted. Based on these geometric constraint features, a cross-frame geometric consistency equation set is constructed, and the camera's intrinsic and extrinsic parameters are automatically calculated. Furthermore, based on the calibrated and calculated camera intrinsic and extrinsic parameters, the road structural elements in the two-dimensional image are mapped to three-dimensional space, simultaneously generating lane line and stop line annotation data in the semantic layer, as well as initial estimates of road surface slope and smoothness for the corresponding area in the physical layer. The confidence level of the calibration parameters serves as the local confidence weight of the three-dimensional intelligent world model in that area, used as a priority reference for subsequent model updates.

[0030] For areas where self-calibration is not used or is not applicable, based on spatiotemporally aligned image data, a semantic segmentation network is used to perform semantic recognition and annotation on lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs, and guardrails in the target road area, thereby obtaining the semantic layer of the three-dimensional intelligent world model.

[0031] S1.3, Generation of semantic layer, physical layer, rule layer, dynamic layer and environment layer; Based on spatiotemporally aligned image data (and point cloud data as an aid), semantic recognition and annotation are performed on lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs and guardrails in the target road area through semantic segmentation networks and / or object detection algorithms to obtain the semantic layer of the three-dimensional intelligent world model.

[0032] Based on spatiotemporally aligned point cloud and image data, the physical layer of the 3D intelligent world model is obtained by extracting road surface friction coefficient, slope, collision body, and rigid body properties through road surface texture analysis, point cloud intensity analysis, and physical parameter inversion algorithms. The rule layer of the 3D intelligent world model is obtained by analyzing lane markings, traffic light timing, and traffic sign information. The dynamic layer of the 3D intelligent world model is obtained by extracting the position, speed, trajectory, and behavioral state of traffic participants through target detection and multi-target tracking algorithms. The environmental layer of the 3D intelligent world model is obtained by extracting weather conditions, lighting conditions, visibility, and road surface slippage through environmental perception modules (including light sensors, rain sensors, etc.) or weather and lighting information parsed from images.

[0033] The three-dimensional intelligent world model comprises six levels, corresponding to six information dimensions: geometric information, semantic information, physical attributes, traffic rules, dynamic objects, and environmental conditions. It is used to simulate the three-dimensional scene of the target road area and its traffic rules, physical attributes, and the behavior of dynamic traffic participants.

[0034] Specifically, the geometry layer is used to store three-dimensional structural data of roads, curbs, buildings, and traffic facilities; the semantic layer is used to store category and attribute data of lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs, and guardrails; the physical layer is used to store data on road surface friction coefficient, slope, collision body, and rigid body properties; the rule layer is used to store data on lane travel direction, traffic light cycle, and traffic rules; the dynamic layer is used to store data on the position, speed, trajectory, and behavioral state of traffic participants; and the environment layer is used to store data on weather conditions, lighting conditions, visibility, and road surface slippage.

[0035] The geometric and semantic layers remain fixed after the model is built. The geometric layer reconstructs the three-dimensional geometric structure of roads, curbs, buildings, and traffic facilities based on point clouds and images from multimodal environmental data. The semantic layer labels the categories and attributes of lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs, and guardrails based on image semantic segmentation results. The geometric and semantic layers are static base layers of the model and will not be changed after construction, so as to ensure the consistency between the simulation scene and the real road environment in terms of spatial structure and traffic signs.

[0036] The physical, environmental, and dynamic layers are editable. The physical layer stores road surface friction coefficient, slope, collision body, and rigid body properties. By adjusting road condition parameters (such as pothole depth, water depth, and friction coefficient), the geometry and physical properties of the road surface can be altered to simulate different road conditions. The environmental layer stores weather conditions, lighting conditions, visibility, and road surface slippage. By adjusting weather and lighting parameters, different environmental conditions such as rain, snow, fog, nighttime, and backlighting can be simulated. The dynamic layer stores the position, speed, trajectory, and behavioral state of traffic participants. By adjusting traffic flow parameters and behavioral model parameters, the movement trajectories of pedestrians, non-motorized vehicles, and motorized vehicles can be generated.

[0037] Under the constraints of the rule layer, the behavioral trajectories of dynamic traffic participants must comply with lane direction, traffic light cycles, and traffic rules. Simultaneously, by setting the violation probability in the behavioral model parameters, violations such as running red lights, driving against traffic, and jaywalking can be generated to test the traffic management road test robot's ability to recognize traffic events.

[0038] By combining different parameter values ​​for the physical, environmental, and dynamic layers, multiple different simulation test scenario models can be generated in batches. Each model corresponds to a specific combination of road conditions, environmental conditions, and dynamic traffic participant behaviors. These scenario models are used by the digital twin of the traffic management road test robot to perform simulation test tasks, in order to evaluate the robot's perception, decision-making, and operational capabilities under different conditions.

[0039] In a preferred embodiment, the three-dimensional intelligent world model further includes a preset linkage rule base. The linkage rule base stores preset mapping relationships between environment layer parameters, physical layer parameters, and dynamic layer parameters, and is pre-built based on at least one of the following methods: Based on physical laws: According to the physical relationship between road surface materials, meteorological conditions and tire adhesion coefficient, a mapping relationship between precipitation intensity and friction coefficient, and a mapping relationship between temperature and road surface icing probability are constructed. Based on engineering experience: Based on actual road test data and domain expert knowledge, we constructed a mapping relationship between fog concentration and visibility distance, and a mapping relationship between light intensity and camera exposure parameters; User-defined: Users can set or adjust the mapping relationship between environment layer parameters, physical layer parameters, and dynamic layer parameters according to specific testing needs, in order to adapt to the simulation requirements of different scenarios.

[0040] The above mapping relationship is pre-stored in the linkage rule base. When the weather parameters or light parameters in the environment layer change, the linkage rule base automatically calls the corresponding mapping relationship to trigger the linkage adjustment of the corresponding parameters in the physical layer and the dynamic layer.

[0041] For example, the linkage rule base may include at least one of the following linkage rules: When the precipitation parameter in the environmental layer switches from "no precipitation" to "moderate rain", the linkage rule base automatically reduces the friction coefficient of the corresponding road surface area in the physical layer and simultaneously adjusts the speed distribution range of traffic participants in the dynamic layer. When the lighting parameters in the environment layer switch from "daytime" to "nighttime", the linked rule base automatically adjusts the road surface reflectivity parameters in the physical layer and synchronously updates the camera exposure parameters in the sensor model. When the temperature parameter in the environmental layer is lower than the preset threshold, the linkage rule base automatically reduces the friction coefficient of the corresponding road surface area in the physical layer and activates the tire slippage model.

[0042] By leveraging a linked rule base, when weather or lighting parameters in the environmental layer change, the system automatically triggers adjustments to the corresponding parameters in the physical and dynamic layers, eliminating the need for users to manually configure each parameter individually. This not only improves the efficiency of simulation scene generation but also avoids issues where parameter configurations fail to conform to physical laws due to human oversight.

[0043] In one implementation, when multiple mapping relationships in the linkage rule base generate different adjustment instructions for the same physical layer parameter (e.g., precipitation rules require a reduction in the friction coefficient, while user-defined rules require maintaining the friction coefficient unchanged), the system determines the effective mapping relationship based on a preset priority strategy. The priority strategy includes at least one of the following: physical law mappings have higher priority than user-defined mappings, and mappings directly triggered by the environment layer have higher priority than those indirectly triggered. This priority strategy avoids parameter configuration uncertainties caused by rule conflicts, ensuring the controllability and physical rationality of the linkage adjustment.

[0044] It should be noted that the linkage rule base is an optional implementation method in this application. Even without setting up a linkage rule base, users can manually configure parameters for the physical layer, environment layer, and dynamic layer separately to generate simulation test scenario models. The role of the linkage rule base is to automatically optimize the efficiency and accuracy of parameter configuration.

[0045] In a preferred embodiment, the three-dimensional intelligent world model divides the drivable road area into spatial grids, discretizing the drivable road area into multiple grid units. Each grid unit is simultaneously associated with all data corresponding to that spatial location in the geometric layer, semantic layer, physical layer, rule layer, dynamic layer, and environment layer.

[0046] Specifically, for any given grid cell, the associated data includes: the road elevation and surface normal vector at that location in the geometry layer; the road type at that location in the semantic layer; the friction coefficient, slope, and rolling resistance at that location in the physical layer; the access rights and speed limits at that location in the rules layer; the presence and movement status of traffic participants at that location in the dynamic layer; and the current light intensity and road surface slipperiness at that location in the environment layer.

[0047] With the above-described raster configuration, when traversing the drivable road area, the digital twin can read all six levels of data at once based on the current raster cell, without needing to query each level separately and then perform spatial matching. This maintains the flexibility of independent storage and editing for each level while achieving efficient full-dimensional data acquisition during spatial queries, thus improving simulation efficiency.

[0048] In a preferred embodiment, the geometric layers of the three-dimensional intelligent world model are implemented using 3D Gaussian Splatting technology, reconstructing a high-fidelity three-dimensional scene of the target road area from multi-view images. Based on this, collider and rigid body properties are bound to the surfaces of objects in the 3D Gaussian Splatting model to support physical interaction and collision detection of the digital twin in the simulation environment.

[0049] Step S2: In the three-dimensional intelligent world model, based on the geometric layer and semantic layer, multiple simulation test scenarios are generated by parameterizing the physical layer, environment layer and dynamic layer, and under the constraints of the rule layer, including road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories.

[0050] Specifically, Figure 3 This is a flowchart illustrating the generation of the multidimensional simulation test scenario in this embodiment. The generation of the multidimensional simulation test scenario includes the following processes: S2.1, Generation of road conditions; Based on configurable road condition parameters, road conditions are generated in the physical layer of the three-dimensional intelligent world model. These parameters include at least one of the following: pothole depth, pothole radius, slope angle, water depth, crack width, road surface friction coefficient, speed bump height, or manhole cover height difference. For example, setting a pothole depth of 5cm and a radius of 30cm at a specified location will change the road surface geometry of the corresponding area in the physical layer, and the road surface friction coefficient in that area can be automatically or manually adjusted according to the water depth. By adjusting these parameters, various road conditions such as potholes, slopes, water accumulation, cracks, slippery surfaces, speed bumps, or abnormal manhole covers can be generated in the three-dimensional intelligent world model.

[0051] S2.2, Generation of environmental conditions; Environmental conditions are generated in the environment layer of the three-dimensional intelligent world model based on configurable weather and lighting parameters. These environmental conditions include weather conditions, lighting conditions, and sensor degradation conditions.

[0052] The environmental layer also includes visibility and road slipperiness. Visibility is related to fog concentration, and road slipperiness is related to precipitation parameters. Both are automatically updated as weather parameters change.

[0053] Weather conditions include at least one of rainfall, snowfall, and fog concentration to simulate different weather conditions such as precipitation, snowfall, and heavy fog; lighting conditions include at least one of lighting angle, lighting intensity, or nighttime light intensity to simulate different lighting environments such as daytime, dusk, nighttime, and backlighting.

[0054] Sensor degradation conditions are an extension of environmental operating conditions, used to simulate the decline in sensor perception performance caused by environmental factors. Specifically, when precipitation or high humidity is present in the environment, the camera sensor's image may experience partial occlusion or blurring due to raindrops adhering to the lens; when there is heavy fog or haze, the point cloud data of the LiDAR will be attenuated, manifested as a decrease in the number of effective point clouds and an increase in noise; when in backlight or strong light environments, the camera's image quality will suffer from local overexposure or decreased contrast due to overexposure. By simulating these sensor degradation phenomena, the perception robustness of traffic management road testing robots under harsh environmental conditions can be evaluated in simulation tests.

[0055] For example, if the weather parameters are set to moderate rain (rainfall of 10 mm / h) and the illumination parameters are set to dusk (illumination angle of 15°), the corresponding area in the environmental layer will exhibit the illumination and precipitation effects of a rainy dusk, resulting in reduced visibility and increased road slipperiness.

[0056] S2.3, Generation of dynamic traffic participants and their behavioral trajectories; Based on configurable traffic flow parameters and behavior model parameters, dynamic traffic participants and their behavioral trajectories are generated in the dynamic layer of the three-dimensional intelligent world model, wherein the behavioral trajectories of the dynamic traffic participants are constrained by the rule layer.

[0057] Traffic flow parameters include at least one of the following: participant type, number, initial location, target location, speed range, or acceleration range. Behavioral model parameters include the participant's behavior type (at least one of normal passage, running a red light, going against traffic, crossing the road, changing lanes, turning, stopping, yielding, or sudden acceleration / deceleration) and its corresponding behavior trigger probability.

[0058] For example, traffic flow parameters are set as follows: 10 pedestrians, 5 non-motorized vehicles, and 8 motorized vehicles are generated on the east side of the intersection; behavioral model parameters are set as follows: non-motorized vehicles have a 0.3 probability of going against traffic, pedestrians have a 0.1 probability of running a red light, and motorized vehicles have a 0.9 probability of passing normally. Then, corresponding traffic participants are generated in the dynamic layer, and their respective behavioral trajectories are formed under the constraints of the rule layer. Most participants obey traffic lights and lane rules and pass normally, while a small number of participants exhibit behaviors such as going against traffic or running red lights.

[0059] In a preferred embodiment, the generation of the dynamic traffic participants and their behavioral trajectories employs a three-dimensional trajectory continuity constraint method, specifically including: The three-dimensional spatial location and appearance features of the dynamic traffic participants in consecutive time frames are obtained; based on the three-dimensional spatial location and appearance features, the same traffic participants in adjacent time frames are associated and matched to identify the same target across frames. When the 3D spatial position of the traffic participant is interrupted or abruptly changed in consecutive time frames (e.g., due to obstruction by trees, buildings, or other vehicles causing loss of observation, or due to target association loss caused by switching between camera perspectives), the interrupted trajectory is restored by reconstructing the breakpoints based on the position, velocity, and direction of movement before and after the interruption, generating a continuous 3D trajectory sequence. This continuous 3D trajectory sequence is output as the behavioral trajectory of the dynamic traffic participant.

[0060] By using the three-dimensional trajectory continuity constraint method described above, the trajectory continuity and identity consistency of dynamic traffic participants can be guaranteed when they are occluded or observed from different perspectives. This makes the behavioral trajectories of traffic participants in the simulation scenario closer to real physical laws and improves the credibility of the simulation test.

[0061] In a preferred embodiment, to further improve the continuity and identity consistency of traffic participant trajectories under occlusion or cross-view conditions, at least one of the following methods can be used to optimize the trajectory: (1) Trajectory completion based on generative models; When a traffic participant's trajectory is detected to be interrupted due to occlusion, a pre-trained generative model generates multiple candidate trajectories for that participant during the interruption period, based on the historical trajectory sequence before the interruption and the current scene context information (including road topology and the motion state of surrounding traffic participants). When the participant reappears, the multiple candidate trajectories are matched with the actual observed trajectory, and the candidate trajectory with the highest matching degree is selected as the completion result to restore the continuous trajectory during the interruption period.

[0062] (2) Multi-objective cooperative trajectory optimization based on spatiotemporal graph convolution; Multiple dynamic traffic participants in the current scenario are constructed as a spatiotemporal graph, where nodes represent participants and edges represent spatial distances and motion correlations between participants. When the trajectory of any participant is interrupted, collaborative optimization is performed based on the participant's historical trajectory and the trajectory information of its neighboring participants to generate a continuous trajectory that satisfies spatial consistency constraints.

[0063] (3) Robust trajectory generation based on multi-sensor fusion; The system integrates multimodal perception data from cameras and lidar mounted on traffic control roadside robots to acquire the three-dimensional spatial position and appearance features of traffic participants. When a participant cannot be observed by the camera due to obstruction, the three-dimensional spatial position of the participant is obtained based on the point cloud data of the lidar to maintain trajectory continuity; when the point cloud data of the lidar is attenuated due to rain or fog, supplementary observation is performed based on the image data of the camera.

[0064] The above-mentioned preferred methods can be used individually or in combination according to the actual needs of the scenario, so as to improve the realism and continuity of the dynamic traffic participant behavior trajectory in the simulation scenario.

[0065] S2.4, Generation of multi-condition combination and simulation test scenario model; By combining different road conditions, environmental conditions, and dynamic traffic participant trajectories, multi-condition combinations are generated. For each condition combination, parameters are configured in the physical, environmental, and dynamic layers of the 3D intelligent world model to generate a corresponding simulation test scenario model. Each simulation test scenario model corresponds to a specific condition combination and is used to execute one or a set of preset simulation test tasks.

[0066] For example, the combination of multiple operating conditions includes, but is not limited to: A combination of minor potholes and daytime with low traffic volume; The combination of deep potholes, rainy weather, and non-motorized vehicle traffic; The combination of slopes, nighttime, and pedestrian obstruction; The combination of glare from standing water, backlighting, and vehicles turning right.

[0067] By employing the aforementioned multi-condition configuration method, multiple different simulation test scenario models can be generated in batches based on the same three-dimensional intelligent world model. This covers various traffic conditions that may occur in the target road area, avoiding the tedious process of manually building scenarios one by one and improving the scenario coverage and efficiency of simulation testing. Simultaneously, under the constraints of the rule layer, the behavioral trajectories of dynamic traffic participants conform to the basic framework of traffic rules, ensuring the realism of the simulation scenarios and supporting the simulation of violations by adjusting behavioral model parameters to test the event recognition capabilities of traffic management road test robots in complex traffic environments.

[0068] In a preferred embodiment, the simulation test scenario model is generated using an automated parameter combination method. Specifically, the system arranges and combines preset road condition parameters, environmental condition parameters, and dynamic traffic participant behavior parameters to generate multiple simulation test scenario models in batches. This automated parameter combination method is applicable to at least one of the following scenarios: when a systematic performance verification (such as regression testing) of the traffic management road testing robot is required under a large number of parameter combinations; when boundary exploration of a preset parameter space is required; or when reproducing and verifying known failure scenarios is required. The advantage of this method is that it can quickly generate a large number of scenarios covering different parameter values, making it suitable for batch testing tasks with high requirements for testing efficiency and scenario coverage.

[0069] In another preferred embodiment, the simulation test scenario model is generated using a large language model-driven approach. Specifically, the system acquires a natural language scene description input by the user (e.g., "I want to test the robot's performance when it encounters an electric vehicle running a red light at a rainy, flooded intersection"), performs semantic parsing on the natural language scene description using a large language model, extracts scene parameters (including road condition parameters, environmental condition parameters, and dynamic traffic participant behavior parameters), and generates a corresponding simulation test scenario model in the three-dimensional intelligent world model based on the extracted scene parameters.

[0070] The large language model-driven approach is applicable to at least one of the following scenarios: user requirements are unstructured natural language descriptions; users lack professional knowledge of scenario parameter configuration; unknown scenarios need to be generated beyond the preset parameter combination space; or creative scenario generation is required. The advantage of this approach is that it lowers the barrier to entry, enabling non-professional users to create complex test scenarios using natural language and generate edge scenario combinations that are difficult for humans to conceive of based on the reasoning capabilities of the large language model.

[0071] For example, the system automatically selects the generation method based on the format of the user input: when the user input is structured parameters (such as parameter lists or configuration files), the automated parameter combination method is used; when the user input is an unstructured natural language description, the large language model-driven method is used. Both generation methods share the same three-dimensional intelligent world model and subsequent simulation testing process. The generated simulation test scenario models are used by the digital twin of the traffic management road test robot to perform simulation testing tasks.

[0072] In practical applications, when users describe scenarios using natural language, it is often difficult to provide complete scenario parameters at once (e.g., only describing weather and road conditions, but not mentioning traffic flow, probability of violations, etc.). If scenarios are generated directly based solely on LLM parsing results, the number of generated scenarios is limited and cannot cover multiple parameter variations under the same scenario type; if only automated parameter combinations are relied upon, users need to manually set a large number of parameter value ranges, resulting in a high barrier to entry. Therefore, in a preferred embodiment, the LLM-driven approach and the automated parameter combination approach are used in conjunction, specifically including: First, an initial simulation test scenario model is generated using the LLM-driven approach. Specifically, the user inputs a natural language scenario description (e.g., "I want to test the robot's performance at an intersection with rain, water accumulation, and many non-motorized vehicles"). The LLM performs semantic parsing on the natural language description, extracts scenario parameters (including road condition type, environmental condition type, and dynamic traffic participant behavior type), and generates one or more initial simulation test scenario models in the three-dimensional intelligent world model based on the extracted scenario parameters.

[0073] Then, based on the scene parameters in the initial simulation test scenario model, the key parameters in the initial scenario are expanded in batches using the automated parameter combination method. Specifically, the system identifies expandable parameters in the initial simulation test scenario model (including at least one of the following: pothole depth, water depth, and friction coefficient in road condition parameters; rainfall, fog concentration, and illumination angle in environmental condition parameters; and traffic flow and violation probability in dynamic traffic participant behavior parameters), sets an expansion range for each expandable parameter, and arranges and combines the expandable parameters within the expansion range to generate multiple expanded simulation test scenario models.

[0074] For example, when the user inputs a natural language description of "intersection with water accumulation on a rainy day", the LLM driving method first generates an initial scene with a rainfall of 10mm / h and a water depth of 2cm. The automated parameter combination method further expands the rainfall parameter to four values: 5mm / h, 10mm / h, 20mm / h, and 50mm / h, and expands the water depth to four values: 0cm, 1cm, 2cm, and 5cm, combining to generate 16 extended scenes, covering the complete working condition range from "light rain with no water accumulation" to "heavy rain with severe water accumulation".

[0075] Through the aforementioned collaborative approach, the semantic understanding capabilities of LLM are utilized to quickly generate initial scenarios that conform to user intent. Furthermore, automated parameter expansion achieves systematic coverage of the key parameter space, avoiding the tedious manual configuration of numerous parameters and improving the scenario coverage and testing efficiency of the simulation test. The expanded simulation test scenario model is used by the digital twin of the traffic management road test robot to perform simulation test tasks.

[0076] Step S3: Control the digital twin of the traffic management road test robot to execute simulation test tasks in the multiple simulation test scenarios, and generate corresponding test evaluation results based on the simulation test tasks.

[0077] Specifically, the digital twin of the traffic management road testing robot includes at least one of the following: a robot geometric model, a kinematic model, a dynamic model, a sensor model, a perception algorithm model, a localization algorithm model, a path planning model, and a control model. This model is used to simulate the motion characteristics and perception decision-making capabilities of a real robot in a simulation environment. The simulation test tasks are used to simulate the actual work performed by the traffic management road testing robot in a real road environment. These simulation test tasks include at least one of the following: road surface defect detection, traffic violation recognition, traffic light status recognition, or traffic sign and marking integrity detection. For example, in a simulation test scenario, the digital twin can simultaneously perform two tasks: road surface pothole detection and red-light violation vehicle recognition, to evaluate the robot's comprehensive processing capabilities under complex working conditions.

[0078] In a preferred embodiment, when the digital twin executes multiple simulation test tasks simultaneously, the system dynamically adjusts the execution priority and resource allocation weight of each task based on the current road conditions, environmental conditions, and the robot's own state. The dynamic task priority adjustment mechanism includes at least one of the following methods: In rainy weather, the impact of potholes and water accumulation on vehicle traffic safety increases significantly. The system automatically raises the priority of the road defect detection task and allocates more sensor data processing resources to it to ensure the detection accuracy and response speed of potholes and water accumulation. During peak traffic hours or when there is a large volume of pedestrians and vehicles at intersections, the system automatically increases the priority of traffic violation identification tasks, increasing the computing power invested in identifying violations such as running red lights, driving against traffic, and illegal lane changes.

[0079] When the robot's battery level drops below a preset threshold, the system automatically prioritizes the task of returning to the charging point to the highest level, ensuring that the robot can safely return before its battery runs out.

[0080] Through the above-mentioned dynamic task priority adjustment mechanism, the task scheduling rationality and overall execution efficiency of the traffic management road test robot under the conditions of multiple tasks in parallel and resource competition can be verified in simulation tests.

[0081] Furthermore, test evaluation results are generated based on the execution results of the digital twin in various simulation test scenarios. These results are used to quantitatively evaluate the comprehensive performance of the traffic management road test robot under different road conditions, environmental conditions, and combinations of traffic participants.

[0082] The test evaluation results include at least one of the following: road surface anomaly detection accuracy, traffic target recognition accuracy, positioning error, obstacle avoidance success rate, collision risk value, task completion time, path length, energy consumption, traffic incident missed detection rate, and scene coverage.

[0083] Among them, the road surface anomaly detection accuracy rate is used to evaluate the robot's accuracy in identifying road defects such as potholes, cracks, and water accumulation; the traffic target recognition accuracy rate is used to evaluate the robot's accuracy in detecting and classifying pedestrians, non-motorized vehicles, and motorized vehicles; the positioning error is used to evaluate the robot's pose estimation accuracy in the simulation scenario; the obstacle avoidance success rate is used to evaluate the robot's ability to bypass obstacles; the collision risk value is used to evaluate the robot's driving safety in dynamic traffic flow; the task completion time, path length, and energy consumption are used to evaluate the robot's task execution efficiency; the traffic event missed detection rate is used to evaluate the completeness of the robot's identification of illegal behaviors such as running red lights and driving against traffic; and the scene coverage rate is used to evaluate the degree of coverage of the current simulation test scene combination for various working conditions in the target road area.

[0084] In a preferred embodiment, the test evaluation results may also include task completion rate and resource utilization efficiency indicators in multi-task parallel scenarios, in order to evaluate the robot's comprehensive task scheduling performance under complex working conditions.

[0085] Step S4: Based on the test evaluation results, generate the safety operation parameter configuration of the traffic management road test robot under multiple simulation test scenarios in the target road area.

[0086] Specifically, step S4 extracts operational parameters that enable the robot to complete tasks safely and efficiently based on the execution results and test evaluation results of the digital twin in multiple simulation test scenarios, forming a safe operational parameter configuration for the current target road area. This safe operational parameter configuration is used to guide the actual operation of the traffic management road test robot in the target road area.

[0087] In one implementation, the safe operating parameter configuration includes at least one of the following: recommended driving speed under different road and environmental conditions, speed limit instructions, recommended turning radius on road sections with different curvatures, and detour paths generated based on obstacle locations.

[0088] (1) Generation of recommended driving speeds and speed limit instructions; For each simulated test scenario in the target road area, the robot's safe driving speed for traversing that scenario is extracted from the test evaluation results. For scenarios involving road conditions such as potholes, water accumulation, slopes, and slippery surfaces, the maximum safe speed for the robot to traverse the corresponding road segment is extracted. For scenarios involving environmental conditions such as rain, snow, fog, and nighttime, the recommended speed range for the corresponding environmental conditions is extracted. By combining the test results from multiple simulation scenarios, recommended driving speeds and speed limit commands are generated for the target road area under different combinations of road and environmental conditions.

[0089] For example, under dry, flat road conditions with sufficient sunlight, the recommended driving speed can be set to 5 m / s; under conditions of water accumulation, potholes, and rain, the recommended driving speed can be set to 2 m / s, with an additional speed limit instruction. For road sections with significant curvature, the system generates a speed limit value for the corresponding curvature based on the speed record of the digital twin safely navigating the curve.

[0090] (2) Recommended generation of turning radius; During simulation testing, the driving trajectory and turning radius of the digital twin were recorded when turning on road sections with different curvatures. For each curved road section that the digital twin safely passed, the corresponding turning radius value was extracted. By combining the test results of multiple simulation scenarios, recommended turning radii for the target road area under different curvatures were generated.

[0091] (3) Generation of detour routes; The location of fixed obstacles in the target road area is identified through a semantic layer. Fixed obstacles include stone blocks, guardrails, lampposts, protruding or missing manhole covers, construction barriers, etc. In simulation testing, the digital twin is controlled to perform path planning to avoid obstacles, and the driving trajectory when safely bypassing them is recorded. Successfully bypassed paths are extracted from the test evaluation results to generate a bypass path from the starting point to the target point. This bypass path is used to guide the traffic management road test robot to avoid fixed obstacles in actual operation.

[0092] For example, when there are stone blocks in the target road area with a spacing smaller than the robot's width, the system controls the digital twin to plan a path to bypass the stone blocks in the simulation. After verifying that it passes safely, the bypass path is included in the safe operation parameter configuration for path planning in actual operation.

[0093] In one implementation, if all simulation tests under a certain working condition combination fail to obtain safe operating parameters (e.g., all driving speeds in the target road area under maximum rainfall cause the robot to slip or overturn), the system outputs a prompt message indicating that no safe operating parameters are configured for that working condition combination, allowing the user to confirm or adjust the robot's hardware parameters.

[0094] In one implementation, the safety operation parameter configuration is generated and stored in the form of a structured file. This structured file can be parsed by the task scheduling system of the traffic management roadside robot and automatically loaded when the robot arrives at the target road area, thereby realizing the automatic configuration and updating of operation parameters.

[0095] Because the geometry (including intersection curvature, lane width, obstacle location, etc.) and physical characteristics (including road surface friction coefficient, slope, etc.) of each target road area are different, the safety operation parameter configurations for different target road areas are also different, and each configuration needs to be generated separately. Therefore, the safety operation parameter configuration is customized for the target road area and is only applicable to that specific target road area, and cannot be directly reused for other road areas.

[0096] S5. In a preferred embodiment, this embodiment further includes a self-calibration step for the digital twin.

[0097] When the traffic control road test robot performs actual road tests according to the aforementioned safe operating parameters, it records real motion data (including actual speed, actual acceleration, actual turning radius, actual wheel speed, etc.). Simultaneously, in a simulation environment, a digital twin is controlled to perform the same task with the same operating parameters, generating simulated motion data. The real motion data is compared with the simulated motion data to calculate kinematic parameter deviations (such as wheel diameter error, odometer drift, steering response delay, etc.).

[0098] When the deviation of the kinematic parameters exceeds a preset threshold, the kinematic model parameters (such as wheel diameter, steering gain, inertia parameters, etc.) in the digital twin are corrected according to the deviation, so that the kinematic characteristics of the digital twin are consistent with those of the real robot.

[0099] The above self-calibration steps can eliminate the digital twin mismatch problem caused by wear and tear or parameter drift of the real robot hardware, ensuring that the simulation test results accurately map to the real robot.

[0100] In a preferred embodiment, a simulation confidence scoring step is also included.

[0101] A confidence score for the simulation test task is generated based on at least one of the following: the last update time of the 3D intelligent world model, the distribution difference between the real multimodal environment data and the simulated multimodal environment data, and the correlation between the simulation test scenario model and the 3D intelligent world model. The confidence score characterizes the reliability of the current simulation test results for real-world road testing.

[0102] When the confidence score is lower than a preset threshold, the system outputs a model calibration suggestion, prompting the user to re-collect multimodal environment data or perform a model update operation.

[0103] This embodiment uses a traffic control road testing robot to collect multimodal environmental data in a target road area and automatically constructs a three-dimensional intelligent world model of the area. This model includes six levels: geometric, semantic, physical, rule, dynamic, and environmental. It can comprehensively represent the static structure (roads, buildings, markings), dynamic objects (traffic participants), physical properties (friction coefficient, slope), and traffic rules (traffic light cycle, traffic direction) of the target road area. Compared with traditional manual modeling methods, this embodiment does not require manual measurement or modeling and can generate a simulation scene that is highly consistent with the real road, providing an accurate basic environment for simulation testing. By parameterizing the physical, environmental, and dynamic layers in a 3D intelligent world model and generating dynamic traffic participant behavior trajectories under the constraints of the rule layer, multiple simulation test scenarios with different combinations of road conditions, environmental conditions, and traffic participants can be flexibly generated. By adjusting road condition parameters (such as pothole depth, water depth, and friction coefficient), environmental condition parameters (such as rainfall and illumination angle), and behavioral model parameters (such as traffic flow and violation probability), extreme conditions that are difficult to construct or pose safety hazards in real roads can be generated in batches in the virtual world, including potholes and water accumulation, rain, snow, fog, mixed pedestrian and vehicle traffic, and sudden traffic events. This solves the problem of insufficient coverage of extreme scenarios in traditional road testing and significantly improves the diversity and controllability of test scenarios. In this embodiment, a digital twin of the traffic management road test robot is controlled to perform preset simulation test tasks (including road defect detection, traffic violation recognition, traffic light status recognition, etc.) in a simulated test scenario, and quantitative test evaluation results (including detection accuracy, recognition recall, positioning error, obstacle avoidance success rate, collision risk value, etc.) are generated. By using a digital twin to replace the real robot for testing, the robot's operating data and performance evaluation under various working conditions can be obtained in a virtual environment. This allows for a comprehensive assessment of the robot's perception, decision-making, planning, and control capabilities in complex traffic environments, avoiding direct reliance on high-cost and high-risk real road tests for parameter tuning, and reducing testing costs and safety risks. Based on test evaluation results from multiple simulation scenarios, a safe operating parameter configuration is generated for the target road area. This includes recommended driving speeds, speed limits, recommended turning radii for road sections with different curvatures, and detour paths generated based on obstacle locations under different road and environmental conditions. Since the geometry and physical characteristics of each target road area are different, the safe operating parameter configuration is customized for that specific area and applies only to that particular region. This configuration guides the safe operation and task execution of traffic management road testing robots in actual deployment, enabling the robot to automatically match optimal operating parameters based on current road conditions and environmental conditions, readily handling various situations that may arise on real roads, thus improving the robot's operational safety and task execution efficiency. In summary, the embodiments of this application realize a complete technical path from real data acquisition, high-fidelity world model construction, parameterized scene generation, dedicated task simulation to customized safe operation parameter configuration, effectively overcoming technical problems in the prior art such as insufficient simulation scene fidelity, difficulty in reproducing extreme working conditions, and disconnect between simulation results and actual road tests.

[0104] Example 2: This embodiment provides a traffic management road testing robot simulation system based on a spatial intelligent world model, used to execute the method described in Embodiment 1 above. This system takes the engineering application of traffic management road testing robots in actual road scenarios as its background, constructs a complete simulation test closed loop to generate safe operation parameter configurations for specific target road areas, guiding the road testing robot to make optimal safe operation decisions when facing various complex scenarios.

[0105] Traffic management road testing robots are deployed on urban roads or at intersections to perform tasks such as inspecting road traffic facilities, assisting in the collection of evidence for traffic violations, and detecting road anomalies. In real road environments, these robots need to cope with complex conditions such as potholes and puddles, rain, snow, fog, mixed pedestrian and vehicle traffic, and sudden traffic incidents. These conditions are difficult to construct in large quantities on real roads. Directly debugging through real road testing is not only costly and inefficient, but also poses safety risks.

[0106] This system can be deployed on cloud servers, edge computing nodes, or local workstations to solve the above problems through simulation.

[0107] The actual operating trajectory of traffic management road testing robots in target road areas is relatively fixed, and their working range is usually limited to specific drivable areas (such as fixed inspection routes, designated intersection duty areas, etc.). Based on this engineering characteristic, the target road area mentioned in this embodiment refers to the drivable area covered by the traffic management road testing robot in actual work, and the system's simulation modeling and parameter configuration are all customized around this specific area.

[0108] This system specifically includes the following modules: Multimodal data acquisition module: Installed on the traffic management roadside testing robot, it is used to collect multimodal environmental data in the target road area; 3D Spatial Intelligent World Model Construction Module: Used to construct a 3D spatial intelligent world model of the target road area based on the multimodal environment data; Parametric scene generation module: used to generate a simulation test scene model containing road conditions, environmental conditions and dynamic traffic participants by adjusting the parameters of the physical layer, environment layer and dynamic layer in the three-dimensional intelligent world model. Simulation test execution module: Used to control the digital twin of the traffic management road test robot to execute preset simulation test tasks in the simulation test scenario model, obtain execution results, and generate test evaluation results; Safe operation parameter configuration generation module: used to generate the safe operation parameter configuration of the traffic management road test robot in the target road area based on the test evaluation results; Closed-loop iterative update module: used to update the three-dimensional intelligent world model based on newly collected multi-source environmental data after the actual road test is performed.

[0109] In practical engineering applications, this system operates according to the following process: First, the traffic control road testing robot runs along a fixed inspection route in the target road area, collecting images, point clouds, inertial data, and RTK positioning data of the area using its onboard cameras, LiDAR, IMU, RTK, and other sensors. Based on this data, the system constructs a three-dimensional intelligent world model of the target road area. This model comprises six layers: a static base layer consisting of a geometric layer and a semantic layer, which remain fixed after construction to ensure consistency between the simulation scene and the real road environment in terms of spatial structure and traffic signs; and a dynamic overlay layer consisting of a physical layer, a rule layer, a dynamic layer, and an environment layer, supporting subsequent parametric editing and scene generation.

[0110] Based on this, the system generates various simulation test scenarios within a three-dimensional intelligent world model. Specifically: Different road surface conditions can be simulated by adjusting road condition parameters (such as pothole depth, water depth, friction coefficient, etc.) in the physical layer. Different environmental conditions are simulated by adjusting weather and illumination parameters (such as rainfall, fog concentration, and illumination angle) in the environmental layer. By adjusting the traffic flow parameters and behavior model parameters in the dynamic layer, the movement trajectories of pedestrians, non-motorized vehicles, and motorized vehicles are generated. Under the constraints of the rule layer, the traffic compliance of the behavior trajectories is ensured. At the same time, violations such as running red lights and driving against traffic can be generated by setting the probability of violations.

[0111] By combining different parameter values ​​from the aforementioned layers, the system generates a large number of different simulation test scenario models in batches, each model corresponding to a specific combination of working conditions. The digital twin of the traffic management road test robot executes simulation test tasks in these scenario models, obtains execution results, and generates quantitative test evaluation results.

[0112] Based on the evaluation results, the system generates a safe operating parameter configuration for the target road area. This configuration includes: recommended driving speeds, speed limit instructions, recommended turning radii for different road curvature sections, and detour paths generated based on obstacle locations, etc., to guide traffic management road testing robots in making decisions based on current road conditions and environmental conditions during actual deployment.

[0113] The system ultimately outputs the configuration of safe operating parameters for the target road area. In actual operation, the traffic management road test robot can automatically match the corresponding safe operating parameters (such as speed, turning radius, whether to detour, etc.) based on the currently perceived road conditions and environmental conditions, and make the optimal safe operating decisions in real time. It can intelligently and efficiently deal with various complex conditions that may occur on real roads, avoiding direct reliance on high-cost and high-risk real road tests for parameter debugging.

[0114] In addition, the system supports continuous iterative optimization. After the traffic management road test robot performs real road test tasks according to the configuration of safe operation parameters, it collects new multimodal environmental data. The system uses an incremental iterative algorithm to automatically compare the data differences, only updates the local model for the spatial areas that have changed, and re-executes the simulation test task based on the updated model to continuously optimize the configuration of safe operation parameters.

[0115] This embodiment provides a traffic control roadside robot simulation system based on a spatial intelligent world model. By collecting multimodal environmental data in the target road area through traffic management road testing robots, a three-dimensional spatial intelligent world model containing geometric, semantic, physical, rule, dynamic and environmental layers is automatically constructed. Extreme working conditions that are difficult to construct in real roads are generated in batches in the virtual world to achieve high-fidelity simulation testing. To address the relatively fixed operating trajectories of traffic management road testing robots in actual engineering projects, this system generates specific safety operation parameter configurations for the drivable areas of the target road region. These configurations include recommended driving speeds, speed limits, recommended turning radii, and detour routes under different operating conditions. This configuration applies only to this specific area, enabling the robot to automatically match optimal operating parameters based on current road conditions and environmental conditions. This allows it to cope with various complex situations that may arise on real roads, avoiding the drawbacks of directly relying on high-cost, high-risk real-world road testing for parameter tuning. The system described in this embodiment, together with the method described in Embodiment 1, constitutes a complete integrated hardware and software solution. Embodiment 1 describes the complete steps of simulation testing from a methodological perspective; this embodiment describes the deployment and operation of each module from a system architecture and engineering application perspective, as well as the application value of the system in actual road test robot scenarios. The two support each other and are implemented collaboratively, together constituting the technical solution of this application.

[0116] Example 3: Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores computer programs, and the processor executes the instructions stored in the memory so that the computer device can perform a traffic control road testing robot simulation method based on a spatial intelligent world model as described in the above embodiments.

[0117] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A simulation method for traffic control roadside testing robots based on a spatial intelligent world model, characterized in that, include: S1, acquire multimodal environmental data collected by the traffic management road testing robot in the target road area; Based on the multimodal environmental data, a three-dimensional intelligent world model of the target road area is constructed. This model comprises a geometric layer, a semantic layer, a physical layer, a rule layer, a dynamic layer, and an environmental layer, used to simulate the three-dimensional scene of the target road area, its traffic rules, physical attributes, and the dynamic behavior of traffic participants. Specifically, the geometric layer includes three-dimensional structural data describing roads, curbs, buildings, and traffic facilities; the semantic layer includes category and attribute data describing lane lines, stop lines, pedestrian crossings, traffic lights, traffic signs, and guardrails; the physical layer includes data describing road surface friction coefficient, slope, collision body, and rigid body properties; the rule layer includes data describing lane direction, traffic light cycle, and traffic rules; the dynamic layer includes data describing the position, speed, trajectory, and behavioral state of traffic participants; and the environmental layer includes data describing weather conditions, lighting conditions, visibility, and road surface slippage. S2, In the three-dimensional intelligent world model, based on the geometric layer and semantic layer, multiple simulation test scenarios are generated by parameterizing the physical layer, environmental layer and dynamic layer, and under the constraints of the rule layer, including road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories. S3, the digital twin of the traffic management road test robot is controlled to perform simulation test tasks in the multiple simulation test scenarios respectively, and generate corresponding test evaluation results based on the simulation test tasks; S4. Based on the test evaluation results, generate the safety operation parameter configuration of the traffic management road test robot under multiple simulation test scenarios in the target road area.

2. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 1, characterized in that, The step of constructing a three-dimensional intelligent world model of the target road area based on the multimodal environment data includes: The target road area is divided into spatial grids to obtain multiple grid cells; The multimodal environment data is mapped to the multiple grid cells according to their spatial location, so that each grid cell is associated with data corresponding to that spatial location in the geometric layer, semantic layer, physical layer, rule layer, dynamic layer and environment layer, thereby constructing the three-dimensional intelligent world model.

3. The traffic control roadside robot simulation method based on a spatial intelligent world model according to claim 1, characterized in that, Step S2 includes: In the aforementioned three-dimensional intelligent world model, road conditions are generated based on configurable road condition parameters; Environmental conditions are generated based on configurable weather and lighting parameters; Based on configurable traffic flow parameters and behavior model parameters, dynamic traffic participants and their behavioral trajectories are generated under the constraints of the rule layer. By combining different road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories, multi-condition combinations are generated, and corresponding simulation test scenarios are generated for each condition combination to execute simulation test tasks.

4. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 3, characterized in that, Step S2 also includes: Obtain the user's natural language scene description, perform semantic parsing on the natural language scene description through a large language model, and extract scene parameters, which include at least one of road condition parameters, weather parameters, lighting parameters, traffic flow parameters, and behavior model parameters; Based on the scenario parameters, the parameterized configuration of the road conditions, environmental conditions, dynamic traffic participants and their behavioral trajectories is executed to generate the corresponding simulation test scenario model.

5. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 4, characterized in that, The process of generating road conditions based on configurable road condition parameters includes: Based on at least one of the road condition parameters, the geometry and / or physical properties of the road surface are modified within the drivable area of ​​the traffic control and road testing robot in the three-dimensional intelligent world model to generate the road condition. The road condition parameters include pothole depth, pothole radius, slope angle, water accumulation depth, crack width, road surface friction coefficient, speed bump height, or manhole cover height difference.

6. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 5, characterized in that, In the three-dimensional intelligent world model, the environmental conditions include weather conditions, lighting conditions, and sensor degradation conditions; The sensor degradation conditions include simulating dirt on a camera lens, raindrop obstruction, or image blurring, as well as simulating point cloud attenuation or noise increase of a lidar in rainy or foggy weather.

7. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 6, characterized in that, The three-dimensional intelligent world model also includes a preset linkage rule library, which includes a mapping relationship between environmental layer parameter changes and physical layer parameters and dynamic layer parameters. The linkage rule base is used to automatically maintain the physical consistency between the corresponding parameters in the physical layer and the dynamic layer and the parameters in the environmental layer when the weather parameters or light parameters in the environmental layer change.

8. The traffic control roadside robot simulation method based on a spatial intelligent world model according to claim 7, characterized in that, In the three-dimensional intelligent world model, the behavioral trajectory includes abnormal behaviors that violate the constraints of the rule layer. The abnormal behaviors include at least one of the following: running a red light, driving against traffic, suddenly crossing the road, changing lanes illegally, cutting in at intersections, or non-motorized vehicles occupying motor vehicle lanes.

9. The traffic control road testing robot simulation method based on a spatial intelligent world model according to claim 1, characterized in that, The safe operating parameter configuration includes at least one of the following operating parameters: Recommended driving speeds and speed limits under different road and environmental conditions; Recommended turning radius for road sections with different curvatures; A detour path is generated based on the location of obstacles in the target road area.

Images