A virtual-real scene fusion mobile robot three-dimensional training environment reconstruction method

Abstract

The application discloses a kind of mobile robot three-dimensional training environment reconstruction methods of virtual-real scene fusion, including using mobile robot with multiple sensors in real scene to collect real scene data, and the real scene data is preprocessed, and form preprocessed data;Modeling is carried out based on the preprocessed data generated in the above steps, to generate real three-dimensional scene model and to construct high-fidelity robot multi-level physical model in simulation software, and its physical parameters are set and corrected, and robot model is constructed.The application collects real scene data by multiple sensors and generates high-precision three-dimensional scene model, and at the same time, high-fidelity robot physical model is constructed, to realize the accurate restoration of virtual-real fusion environment.

Description

Technical Field

[0001] This invention belongs to the field of mobile robot 3D training environment technology, specifically referring to a method for reconstructing a mobile robot 3D training environment by fusing virtual and real scenes. Background Technology

[0002] In the field of 3D training environments for mobile robots, current technologies heavily rely on standardized simulation scenarios and preset physical parameters. The industry commonly employs high-precision simulation engines to build 3D virtual spaces, accurately simulating real-world physical rules (such as gravity, friction, and collision feedback) and environmental features (such as light intensity, terrain texture, and obstacle distribution) to provide robots with a training platform for core capabilities such as perception, planning, and control. This standardized simulation model, with its controllable parameters and reproducible scenarios, demonstrates high efficiency in robot training within structured environments, becoming a crucial support for driving the iteration of mobile robot technology.

[0003] However, when robots face complex 3D scenes in the real world, the limitations of existing training models become increasingly apparent. In scenarios like urban ruins, where building debris is randomly stacked and structural stability is unknown, the regularized obstacle models preset in the simulation environment are difficult to match the irregular shapes of real ruins. In dense forests, the dynamically changing vegetation, randomly falling branches and leaves, and instantaneous changes in light differ significantly from the fixed plant parameters and lighting settings in the simulation environment. Furthermore, in special scenarios such as inside nuclear power plants, there is not only a dense distribution of pipes and equipment, but also dynamic fluctuations in radiation fields, temperature, and humidity. These complex variables are difficult for standardized simulation environments to accurately reproduce.

[0004] These differences directly lead to the "simulation-reality gap": the decision-making logic and action patterns that robots develop through standardized training in simulation environments often fail in real, complex scenarios due to sudden changes in environmental parameters, irregular structural forms, and the randomness of dynamic disturbances. For example, the flat ground preset in a simulation environment contrasts sharply with the uneven terrain of real ruins, causing a significant decrease in the robot's motion control accuracy. Fixed sensor parameters cannot cope with sudden changes in light in a forest, thus affecting the accuracy of environmental recognition. This lack of adaptability makes existing training methods based on standardized simulations insufficient to meet the practical application needs of mobile robots in complex 3D scenes. The training effect falls significantly short of the expected goals, becoming a key bottleneck restricting the expansion of mobile robots into complex real-world environments and preventing the training effect from achieving the desired results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing mobile robot training methods, such as the gap between the simulation environment and the real world, and the difficulty in adapting robots to the complexity of the real three-dimensional environment, which leads to the failure of training results to meet expectations. This invention provides a novel method for reconstructing the three-dimensional training environment of mobile robots by integrating virtual and real scenes.

[0006] The objective of this invention is achieved through the following technical solution: a method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes, comprising the following steps:

[0007] S1. Use a mobile robot equipped with multiple sensors to collect real-world scene data in a real-world setting, and preprocess the real-world scene data to form preprocessed data.

[0008] S2. Based on the preprocessed data generated in step S1, model a real 3D scene model.

[0009] S3. Construct a multi-level physical model of a high-fidelity robot in simulation software, and set and correct its physical parameters to build a robot model.

[0010] S4. Based on the generated real 3D scene model, import the virtual object model and the constructed robot model. Use the coordinate alignment function to accurately align the virtual object and robot model with the real 3D scene, and set the physical parameters of the virtual object model and robot model.

[0011] S5. Parametrically model the parameters of each environmental element to form a virtual-real fusion 3D training environment;

[0012] S6. Train and iteratively update the robot model in the constructed virtual-real fusion 3D training environment.

[0013] Furthermore, the real scene data mentioned in step S1 includes at least lidar point cloud data, RGB-D camera image data, IMU inertial measurement data, and ground material data; wherein, preprocessing the lidar point cloud data involves removing noise points and outliers, preprocessing the RGB-D image data involves using a pinhole camera model to correct radial distortion, and preprocessing the ground material data involves standardizing it.

[0014] The "removal of noise points and outliers" process uses a statistical filtering algorithm to remove noise points.

[0015] The "generating a realistic 3D scene model" mentioned in step S2 includes at least: a point cloud model, which is formed by stitching together LiDAR point cloud data at different times using a point cloud registration algorithm, with a registration accuracy of centimeters and a root mean square error of less than 5cm; and a textured 3D point cloud model, which is formed by mapping RGB-D image data to textures, and requires labeling of different ground material areas.

[0016] The "construction of a multi-level physical model of a high-fidelity robot" mentioned in step S3 includes at least setting physical parameters such as joint torque, motor inertia, and body posture in high-precision simulation software, and converting the CAD model of the real robot into URDF to generate a USD model.

[0017] The phrase "precisely align the virtual object, robot model, and real 3D scene" in step S4 means that the positional deviation between the virtual object, robot model, and real 3D scene does not exceed 10cm.

[0018] The step S6, "training and iteratively updating the robot model in the constructed virtual-real fusion 3D training environment", specifically includes testing the robot model in the constructed training environment on tasks such as walking, climbing, and obstacle avoidance, recording data such as joint forces, motion trajectory, and posture stability, calculating average speed and posture error, and optimizing environmental parameters and iteratively updating them based on the test results.

[0019] The statistical filtering algorithm removes noise points under the following condition: when the neighborhood set of a point p is 1.5 times the standard deviation, it is determined to be a noise point and removed.

[0020] The process of "converting the CAD model of a real robot into URDF and then generating a USD model" uses the Euler-Lagrange equations to construct a robot motion model. This robot motion model is specifically as follows: ,in For joint angle, velocity, and acceleration, The inertia matrix, The Coriolis force matrix, For gravity, Joint torque (range) ), This is the disturbance torque.

[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0022] (1) This invention collects real scene data through multiple sensors and generates a high-precision three-dimensional scene model. At the same time, it constructs a high-fidelity robot physical model to achieve accurate restoration of the virtual-real fusion environment. The registration accuracy of the real three-dimensional scene model reaches the centimeter level and the root mean square error is less than 5cm. The physical parameters of the robot model conform to the real characteristics, making the training environment closer to the real world, providing a reliable scene basis for robot training, and reducing the deviation of training effect caused by environmental differences.

[0023] (2) The virtual-real fusion three-dimensional training environment of this invention can import virtual objects and combine them with the environmental elements such as ground material of real scenes for parameterized modeling, which can simulate a variety of complex scenes. When the robot performs tasks such as walking, climbing, and obstacle avoidance in it, it can record multi-dimensional data such as joint force and motion trajectory. By calculating the average speed and posture error to optimize parameters, the robot's adaptability to different scenes can be comprehensively improved, and the training effect is more significant.

[0024] (3) This invention performs targeted preprocessing on the collected real-scene data. For example, the lidar point cloud data is filtered using a statistical filtering algorithm (with a standard deviation multiple of 1.5) to remove noise points, the RGB-D image data is corrected for radial distortion, and the ground material data is standardized, effectively improving the data quality. Based on the high-quality data, a textured 3D point cloud model is generated and the ground material area is labeled, providing accurate data support for subsequent modeling and training, and improving the overall modeling accuracy.

[0025] (4) This invention utilizes simulation software to construct a training environment, allowing the robot model to be trained and iteratively updated in a virtual environment, eliminating the need for frequent operation of the physical robot in real-world scenarios. This not only reduces the wear and tear on the physical robot and lowers costs associated with equipment maintenance and site usage, but also avoids the safety risks that may arise from training in complex or dangerous real-world scenarios, providing a more economical and safer solution for robot training.

[0026] (5) In the robot model construction, this invention constructs a motion model using the Euler-Lagrange equations, precisely sets and corrects physical parameters such as joint torque, and iteratively optimizes the model by combining data such as posture stability recorded during training. This helps to deeply analyze the robot's motion laws, optimize motion control strategies, and enable the robot to have more stable posture, more reasonable joint forces, and better motion performance in practical applications, thereby improving its work efficiency and reliability. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0029] Example

[0030] like Figure 1 As shown in this embodiment, the method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes includes the following steps:

[0031] S1. Use a mobile robot equipped with multiple sensors to collect real-world scene data and preprocess the real-world scene data to form preprocessed data.

[0032] In this step, the real-world scene data includes at least LiDAR point cloud data, RGB-D camera image data, IMU inertial measurement data, and ground material data. LiDAR point cloud data is a collection of numerous three-dimensional coordinate points generated after a LiDAR (Light Detection and Ranging) sensor scans the target scene; it is a digital representation of three-dimensional spatial information. These points are calculated using information such as the time difference between laser pulse emission and reflection, and angle, collectively constituting the contours, positions, and morphological features of objects in the scene. Each point in the point cloud contains at least three key pieces of information: three-dimensional coordinates (X, Y, Z), representing the point's position in space, typically based on the sensor's own coordinate system or the world coordinate system; reflection intensity, the energy intensity of the laser pulse reflected back to the sensor, related to the object's surface material, roughness, and angle of incidence; and additional information, such as the scanning time, laser channel number, and color recorded by some LiDAR systems. The point cloud data in this embodiment features high density and high precision, three-dimensional spatial integrity, and strong environmental adaptability.

[0033] The RGB-D camera image data is a type of image data that integrates color (RGB) and depth information, simultaneously providing the color features and 3D spatial location information of objects in a scene. It is a crucial data format for achieving 3D scene perception. This RGB-D image data consists of two key components: an RGB color image and a depth map. The RGB color image is identical to the output of a regular color camera, containing pixel information from the red (R), green (G), and blue (B) channels, used to describe the color, texture, and other appearance features of objects in the scene. The depth map is presented as a grayscale image or a numerical matrix, where each pixel value represents the actual distance from that point to the camera (usually in millimeters or meters), directly reflecting the object's depth position in 3D space. These two data components are synchronized in time and aligned spatially (i.e., pixel-level alignment) to ensure that each pixel in the RGB image corresponds to the depth information at the same location in the depth image, thus forming fused "color + depth" data.

[0034] The inertial measurement unit (IMU) data is physical quantity data collected by the IMU sensor to describe the motion state of an object (acceleration, angular velocity, attitude, etc.), and is the core data source for motion tracking, navigation and positioning, and attitude estimation.

[0035] An IMU typically consists of two core sensors, whose measurement data includes accelerometer data and gyroscope data. Accelerometer data measures the linear acceleration of an object in three-dimensional space (X, Y, Z axes) (usually in m / s²), including both the object's own motion acceleration and a component of gravitational acceleration. By integrating the acceleration data, the object's velocity and displacement can be estimated. Gyroscope data measures the angular velocity of an object around its three-dimensional coordinate axes (roll, pitch, yaw) (usually in ° / s or rad / s), reflecting the speed and direction of the object's rotation. By integrating the angular velocity data, the object's rotation angle can be obtained, and thus, attitude changes can be deduced.

[0036] The ground material data is a collection of information describing the material properties of ground surfaces in real-world scenarios. It primarily characterizes the physical properties, appearance, and interactive behaviors of the ground, serving as crucial foundational data for constructing high-fidelity virtual-real integrated environments. This ground material data typically includes key information such as physical properties, appearance features, and environmental interaction characteristics. Physical properties include the coefficient of friction (affecting the robot's grip and slip probability), hardness / elasticity (determining the force feedback when the robot contacts the ground, such as collisions and vibrations), roughness (related to noise and wear during robot movement), and load-bearing capacity (whether it can support the robot's weight and prevent collapse in simulated scenarios). Appearance features include color, texture (e.g., the gray, rough texture of cement, the smooth, reflective surface of ceramic tiles), reflectivity and transparency (affecting the realism of lighting rendering in virtual scenes), and surface structure (e.g., the presence of cracks and unevenness, affecting robot path planning). Environmental interaction characteristics include water absorption and slip resistance (e.g., the impact of rainy weather changes on robot movement), electrical conductivity and thermal conductivity (functional simulation in special scenarios, such as the working environment of industrial robots).

[0037] The ground material data plays a particularly crucial role in this embodiment. It supports the physics engine simulation of the virtual scene (such as the feeling of a robot getting stuck when walking on "sand" or the slipping effect on "ice"), enhances the realism of robot training (by reproducing the differences in ground materials in the real world, the robot learns to adapt to different terrains in the virtual environment), and assists in the optimization of environmental parameters (such as adjusting the physical parameters of the training environment based on the robot's movement data on different ground materials to improve the training effect). It is a "bridge" connecting the real world and the virtual training environment, and its accuracy directly affects the effectiveness and reliability of robot training in the virtual-real fusion scenario.

[0038] In this embodiment, when preprocessing the aforementioned real-world scene data, the preprocessing of LiDAR point cloud data involves removing noise points and outliers. Preprocessing of RGB-D image data involves using a pinhole camera model to correct radial distortion. Preprocessing of ground material data involves standardization. Specifically, removing noise points and outliers from LiDAR point cloud data requires using a statistical filtering algorithm to remove noise points, with the mean distance between neighboring points being: Standard deviation: That is, when the neighborhood of a point p is set When the value is 1.5 times the standard deviation, it is identified as a noise point and removed.

[0039] The method of using a pinhole camera model to perform radial distortion correction on RGB-D image data specifically involves: based on the pinhole camera model, the radial distortion correction (x, y) represents the pixel coordinates of the distorted point. For the corrected coordinates, the distortion correction formula is: , .in , Principal point coordinates The radial distortion coefficients are used to set the distortion coefficients to obtain the corrected image data.

[0040] The standardization of ground material data refers to converting ground material information from different sources, in different formats, and at different levels into a unified format that conforms to preset rules or common standards. This eliminates interference caused by data differences and ensures the consistency and usability of the data in subsequent modeling, analysis, and simulation. Different acquisition methods (such as sensor measurement, manual input, and database retrieval) may generate data in different formats (such as CSV tables, JSON files, and custom text). Standardization requires converting this data into a unified format (such as a table format with unified fields) and clearly defining the data structure (such as including mandatory fields like "material category," "friction coefficient," and "RGB color value") to facilitate data storage, retrieval, and batch processing. The physical parameters of ground materials (such as friction coefficient, hardness, and roughness) may use different units (e.g., friction coefficient may be expressed as "dimensionless" or "N / m²," while roughness may be expressed as "micrometer" or "millimeters"). Standardization requires converting these parameters into unified units (e.g., unifying friction coefficient as dimensionless and roughness as "micrometer") to avoid errors in calculations or comparisons. The values ​​of the same physical parameter can vary greatly depending on the material type or measuring tool (e.g., the coefficient of friction may range from 0.1 to 2.0, and the reflectivity may range from 0 to 100%). Standardization maps values ​​to fixed intervals (e.g., [0,1]) or sets standard ranges according to material categories (e.g., the standard range for the coefficient of friction of "concrete" is 0.6 to 0.8), facilitating unified data analysis by subsequent models (e.g., threshold judgment of parameters by the physics engine). The naming of "ground materials" may be ambiguous (e.g., "cement ground" and "concrete ground" are actually the same material, and "plastic running track" may be abbreviated as "plastic ground"). Standardization requires the establishment of a unified classification label system (e.g., classified according to material composition as "concrete," "asphalt," "wood," "plastic," etc.), and clearly defines the characteristic parameter range corresponding to each category to avoid modeling errors caused by classification confusion.

[0041] S2. Based on the preprocessed data generated in step S1, a model is created to produce a realistic 3D scene model. In this step, the generated realistic 3D scene model includes at least a point cloud model and a textured 3D point cloud model. For the point cloud model, a point cloud registration algorithm is used to match the point clouds, stitching together LiDAR point cloud data from different times. The registration accuracy is at the centimeter level with a root mean square error of less than 5 cm.

[0042] The point cloud matching mentioned refers to the initial construction of a 3D scene based on preprocessed data using high-precision simulation reconstruction tools. In this point cloud matching process, a point cloud registration algorithm is needed to iteratively minimize the point cloud distance error to achieve registration. The objective function is: , where min R,tThis indicates an optimization of parameters R and t, seeking the combination of R (rotation matrix) and t (translation amount, which can be understood as the intercept in linear fitting) that minimizes the subsequent expression. This refers to summing the data points i from 1 to n, where n is the total number of data points. It calculates the square of the residual (error) of the i-th data point.

[0043] The textured 3D point cloud model is created by mapping RGB-D image data to textures, and requires annotation of different ground material areas. Specifically, it combines RGB-D image data to assign color information to the point cloud model, generating a textured 3D point cloud model. This involves processing the RGB image using image segmentation algorithms (such as semantic segmentation networks) to obtain the category information of objects in the scene (such as pedestrians, vehicles, buildings, different ground areas, etc.) and mapping it to the 3D point cloud model.

[0044] S3. Construct a multi-level physical model of a high-fidelity robot in simulation software, and set and correct its physical parameters to build the robot model. In this step S3, "constructing a multi-level physical model of a high-fidelity robot" includes at least setting physical parameters such as joint torque, motor inertia, and body posture in high-precision simulation software, and converting the CAD model of the real robot into URDF to generate a USD model.

[0045] The physical parameters involved specifically include: joint torque, which is set according to the actual performance of the robot's joint motors, and its range is... Motor inertia, set to Body posture is simulated and parameters are set via IMU, including the measurement range and accuracy of pitch, roll, and yaw angles. The robot model is then finely adjusted to ensure that parameters such as the range of motion and moment of inertia of each joint are consistent with those of the real robot. The CAD model of the real robot is converted to URDF and then to USD model. During the conversion process, the Euler-Lagrange equations are used to construct the robot's motion model. ,in, For joint angle, velocity, and acceleration, The inertia matrix, The Coriolis force matrix, For gravity, Joint torque (range) ), The disturbance torque is calculated. Scale and parameter matching are performed in the simulation environment, with the error controlled within 3%.

[0046] S4. Based on the generated realistic 3D scene model, import the virtual object model and the constructed robot model. Use the coordinate alignment function to accurately align the virtual object and robot models with the realistic 3D scene, and set the physical parameters of the virtual object and robot models. This step is to ensure that the virtual environment can realistically reflect its force and motion characteristics under various working conditions. For example, in a climbing scenario, calculate the force on the robot's legs based on the slope and accurately simulate it in the simulation.

[0047] This step specifically includes scene import and physical property simulation settings. Scene import refers to importing virtual object models onto the reconstructed realistic 3D scene. These virtual object models include, but are not limited to, virtual obstacles, dynamic pedestrians, and task objectives. Simultaneously, the constructed robot model is imported into the simulation training environment. Coordinate alignment refers to precisely aligning the virtual objects and robot models with the real scene using the coordinate alignment function of high-precision simulation software, with a positional deviation not exceeding 10cm. Setting the physical parameters of the virtual object and robot models in this step involves setting their physical properties. For example, the virtual pedestrian's mass is set to 60kg, and its walking speed ranges from 1 to 1.5m / s; the friction coefficient of the virtual obstacle is set to 0.5.

[0048] S5. Parametrically model the parameters of each environmental element to form a virtual-real fusion 3D training environment. This step's parametric modeling of environmental elements should at least include modeling parameters such as environmental parameter configuration, obstacle settings, sensor settings, and dynamic disturbance settings. Specifically, environmental parameter configuration should include parametric modeling of elements such as ground material, friction coefficient, slope variation, and environmental obstacles. In practical applications, ground material parameters are assigned to different areas based on collected data; for example, the friction coefficient for grass is set to 0.6–0.8, for gravel to 0.3–0.5, and for step surfaces to 0.7–0.9. Slope variation settings are adjustable parameters, ranging from 0° to 30°.

[0049] Obstacle setting modeling refers to setting up parametric models with different shapes, sizes, and locations. This model includes rapid switching between various scenes, from flat ground, grass, gravel, and steps to complex and irregular obstacles, with a switching response time set to 200ms.

[0050] Sensor setup modeling involves simultaneously modeling real-world factors such as joint sensor noise, delay, and dynamic disturbances. The joint sensor noise is set to Gaussian noise with a mean of 0 and a standard deviation of 0.01, and the sensor delay is set to 5–10 ms.

[0051] Dynamic disturbance modeling refers to simulating disturbance forces caused by external wind and uneven ground, with a range of ±5N. By comparing the motion data of real robots in the test scenario, the virtual environment parameters are optimized to ensure that the strategy can robustly converge under high noise and high uncertainty conditions during training. Robust convergence, in this context, means "stable convergence unaffected by disturbances," ensuring that the algorithm / system can reliably achieve its objectives (fitting models, controlling robots, etc.) in real-world complex scenarios. It is a key indicator for evaluating the practicality of the algorithm.

[0052] S6. Train and iteratively update the robot model in the constructed virtual-real fusion 3D training environment.

[0053] This step includes two parts: environmental verification testing and iterative updates. Environmental verification testing involves testing the robot model in a pre-constructed virtual-real 3D training environment, performing tasks such as walking, climbing, and obstacle avoidance. Data on the robot's joint forces, motion trajectory, and posture stability are recorded. The formulas for calculating the robot's average speed and posture error in the training scenario are: Where T is the test time and Φ is the attitude angle.

[0054] The iterative update refers to optimizing and updating the training environment based on the test time T and the attitude angle Φ. For example, if the robot slips severely when walking on gravel, the friction coefficient parameter of that area needs to be adjusted; if joint sensor noise causes unstable robot attitude control, the noise parameter needs to be recalibrated.

[0055] Through the steps described in this embodiment, the present invention can construct a high-fidelity, high-accuracy virtual-real fusion training environment. It can accurately simulate various elements in real-world scenarios, allowing the robot to perform diverse task training within it, with its performance truly reflecting the actual situation. Furthermore, continuous optimization ensures that the training strategy converges robustly under high noise and high uncertainty conditions, ultimately improving the robot's ability and adaptability to complete tasks such as walking, climbing, and obstacle avoidance in real-world scenarios.

[0056] As described above, the present invention can be well implemented.

Claims

1. A method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes, characterized in that, Includes the following steps: S1. Use a mobile robot equipped with multiple sensors to collect real-world scene data in a real-world setting, and preprocess the real-world scene data to form preprocessed data. S2. Based on the preprocessed data generated in step S1, model a real 3D scene model. S3. Construct a multi-level physical model of a high-fidelity robot in simulation software, and set and correct its physical parameters to build a robot model. S4. Based on the generated real 3D scene model, import the virtual object model and the constructed robot model. Use the coordinate alignment function to accurately align the virtual object and robot model with the real 3D scene, and set the physical parameters of the virtual object model and robot model. S5. Parametrically model the parameters of each environmental element to form a virtual-real fusion 3D training environment; S6. Train and iteratively update the robot model in the constructed virtual-real fusion 3D training environment; The real-world scene data mentioned in step S1 includes at least LiDAR point cloud data, RGB-D camera image data, IMU inertial measurement data, and ground material data. Preprocessing the LiDAR point cloud data involves removing noise points and outliers; preprocessing the RGB-D image data involves using a pinhole camera model to correct radial distortion; and preprocessing the ground material data involves standardization. The "removal of noise points and outliers" uses a statistical filtering algorithm to remove noise points. The condition for using the statistical filtering algorithm to remove noise points is: if the neighborhood set of a point p is 1.5 times the standard deviation, it is considered a noise point and is removed.

2. The method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes according to claim 1, characterized in that, The "generating a realistic 3D scene model" mentioned in step S2 includes at least: a point cloud model, which is formed by matching point clouds using a point cloud registration algorithm and stitching together LiDAR point cloud data at different times. The registration accuracy is at the centimeter level and the root mean square error is less than 5cm. And textured 3D point cloud models, which are generated by mapping RGB-D image data to textures, and require annotation of different ground material areas.

3. The method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes according to claim 1, characterized in that, The "construction of a multi-level physical model of a high-fidelity robot" mentioned in step S3 includes at least setting physical parameters such as joint torque, motor inertia, and body posture in high-precision simulation software, and converting the CAD model of the real robot into URDF to generate a USD model.

4. The method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes according to claim 1, characterized in that, The phrase "precisely aligning the virtual object, robot model, and real 3D scene" in step S4 means that the positional deviation between the virtual object, robot model, and real 3D scene does not exceed 10cm.

5. The method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes according to claim 1, characterized in that, The step S6, "training and iteratively updating the robot model in the constructed virtual-real fusion 3D training environment", specifically includes testing the robot model in the constructed training environment on tasks such as walking, climbing, and obstacle avoidance, recording data such as joint force, motion trajectory, and posture stability, calculating average speed and posture error, and optimizing environmental parameters and iteratively updating them based on the test results.

6. The method for reconstructing a 3D training environment for a mobile robot by fusing virtual and real scenes according to claim 3, characterized in that, The process of "converting the CAD model of a real robot into URDF and then generating a USD model" uses the Euler-Lagrange equations to construct a robot motion model. This robot motion model is specifically as follows: ,in For joint angle, velocity, and acceleration, The inertia matrix, The Coriolis force matrix, For gravity, Joint torque (range ±50N) m), This is the disturbance torque.

Images