Four-legged robot autonomous navigation method and system based on reinforcement learning

Abstract

The application provides a quadruped robot complex terrain autonomous navigation method and system based on reinforcement learning, and relates to the technical field of robot control. The method comprises the following steps: acquiring multi-modal perception data of the robot when walking in a complex terrain environment, generating a terrain semantic graph and a terrain elevation graph after processing, and extracting a geometric feature graph from the terrain elevation graph and a semantic feature graph from the terrain semantic graph; projecting the semantic feature graph and the geometric feature graph to the same overhead grid space with the robot as the center, and constructing multiple grid layers with different resolutions. Through time series modeling and end-to-end reinforcement learning control, the application overcomes the problem of perception and decision fragmentation in traditional methods, and significantly improves the autonomous navigation ability and dynamic adaptability of the quadruped robot in the complex terrain of the wild.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a method and system for autonomous navigation of quadruped robots in complex terrain based on reinforcement learning. Background Technology

[0002] Traditional quadruped robot navigation methods typically rely on precise environmental modeling and model-based motion planning. These methods first collect environmental data using sensors such as LiDAR and depth cameras, and construct 2D grid maps, 2D half-level maps, or 3D point cloud maps to describe the terrain geometry. Based on this, path planning and motion control are performed using predefined gait patterns or model-based predictive control and whole-body control. With the rapid development of deep learning and reinforcement learning technologies, data-driven navigation methods are gradually replacing traditional control strategies based on precise modeling, providing a new technological path for quadruped robots to adapt to complex terrains. Existing data-driven methods still face severe challenges in perceptual data processing and fusion, especially the effective integration of multi-source heterogeneous information, which has become a core technical bottleneck restricting the improvement of navigation decision quality. Quadruped robots in field navigation are usually equipped with a variety of heterogeneous sensors. These sensors continuously generate massive and heterogeneous data streams at different frequencies, coordinate systems, and data dimensions. How to design an efficient data processing framework to accurately align high-dimensional visual semantic information with precise geometric ranging information in spatiotemporal and deep feature fusion, and generate a unified understanding of the environment, is a key technical challenge for improving the quality of robot navigation decisions. Summary of the Invention

[0003] This invention provides a method and system for autonomous navigation of quadruped robots in complex terrain based on reinforcement learning. By fusion of spatiotemporally aligned multimodal perception data, temporal modeling and end-to-end reinforcement learning control, it overcomes the problem of separation between perception and decision-making in traditional methods, and significantly improves the autonomous navigation capability and dynamic adaptability of quadruped robots in complex terrain in the wild.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning, the method comprising: Step 1: Acquire multimodal perception data of the robot walking in complex terrain environment, generate terrain semantic map and terrain elevation map after processing, extract geometric feature map from terrain elevation map and semantic feature map from terrain semantic map; Step 2: Project the semantic feature map and geometric feature map onto the same top-down grid space centered on the robot to construct multiple grid layers with different resolutions. For each grid position, associate and combine the semantic features and geometric features at the same resolution to form a fused feature unit at that resolution. Step 3: Calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution in the neighborhood of each grid location. Based on this complexity, dynamically adjust the contribution weight of fused feature units at different resolutions, and generate a comprehensive environmental situation map after weighted integration. Step 4: Obtain comprehensive environmental situation map at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, analyze the dynamic evolution trend of the terrain and predict the terrain status of the area ahead to form an enhanced spatiotemporal situation map. Step 5: Obtain the robot's current pose and velocity commands. Combine the pose and velocity commands with the enhanced spatiotemporal situation map to form a state space. The state space is used to obtain the joint target angles through a reinforcement learning policy network and converted into driving torques by the underlying controller to achieve autonomous navigation.

[0005] Furthermore, multimodal perception data of the robot walking in complex terrain environments is acquired, processed to generate terrain semantic maps and terrain elevation maps, and geometric feature maps are extracted from the terrain elevation maps and semantic feature maps are extracted from the terrain semantic maps, including: Acquire multimodal perception data of the robot when it walks in complex terrain environment, wherein the multimodal perception data includes at least environmental images, laser ranging data, inertial measurement data and joint state data; The environmental image is analyzed at the pixel level to obtain a terrain semantic map and extract semantic feature maps from it; the laser ranging data is spatially interpolated and reconstructed to generate a terrain elevation map centered on the robot and extract geometric feature maps reflecting the degree of terrain undulation from it.

[0006] Furthermore, the semantic feature map and geometric feature map are projected onto the same top-down grid space centered on the robot, constructing multiple grid layers of different resolutions. For each grid location, the semantic features and geometric features at the same resolution are associated and combined to form a fused feature unit at that resolution, including: Using terrain semantic map and terrain elevation map as processing objects, coordinate transformation is used to project them onto the same top-view grid coordinate system centered on the robot's current position, resulting in a spatially aligned top-view semantic grid map and top-view elevation grid map. Based on the top-view semantic raster map and the top-view elevation raster map, multi-scale downsampling processing is performed to construct multiple raster layers including the original resolution, 1 / 2 resolution and 1 / 4 resolution, so that each resolution layer contains a corresponding semantic feature sub-map and a geometric feature sub-map. For each raster cell in each resolution layer, the semantic feature value and geometric feature value corresponding to the raster cell are associated and combined to generate a fused feature cell at the resolution.

[0007] Furthermore, the local terrain complexity is calculated based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution within the neighborhood of each grid location. The contribution weights of fused feature units at different resolutions are dynamically adjusted based on this complexity, and a comprehensive environmental situation map is generated after weighted integration. This also includes: Based on the obtained fused environment representation, for each grid location, a complexity index representing the local terrain complexity is calculated according to the rate of change of geometric features in the neighborhood of the grid location and the degree of semantic category mixing. Based on the complexity index of each grid position, the contribution weight of the fused feature unit at the original resolution, 1 / 2 resolution and 1 / 4 resolution at that position is dynamically determined; Based on the dynamically determined contribution weights, the fusion feature units at different resolutions at each grid location are weighted and summed to obtain the fusion feature value that integrates semantic category and geometric undulation information. The fused feature values ​​obtained after weighted integration of all grid locations are aggregated to construct a comprehensive environmental situation map that includes terrain category and geometric undulation.

[0008] Furthermore, by acquiring comprehensive environmental situation maps at multiple moments in a continuous time series, and comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, the dynamic evolution trend of the terrain is analyzed, and the terrain condition of the area ahead is predicted accordingly, forming an enhanced spatiotemporal situation map, including: Based on the comprehensive environmental situation map, situation maps of the current moment and multiple previous moments are continuously collected along the time dimension to construct a comprehensive environmental situation map sequence containing time-series information; For the situation map of two adjacent time points in the time series, the same spatial grid position is used as the comparison unit. The semantic category information and geometric elevation information at that position are extracted respectively, and the change in semantic category and the change in geometric elevation are calculated. Based on the semantic and geometric elevation changes of each grid position between adjacent time points, the dynamic evolution characteristics of the local terrain are analyzed, and areas where terrain type is changing and elevation is changing are identified, thus obtaining the analysis results of grid positions. Based on the analysis results of all grid positions, the dynamic evolution trend of the entire terrain scene in the time dimension is obtained.

[0009] Furthermore, by acquiring comprehensive environmental situation maps at multiple moments in a continuous time series, and comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, the dynamic evolution trend of the terrain is analyzed, and the terrain condition of the area ahead is predicted accordingly, forming an enhanced spatiotemporal situation map. This also includes: Based on the dynamic evolution trend of the terrain, for the grid area in front of the robot's current position that has not yet been entered, the semantic category and geometric elevation of the grid area in the future are predicted according to the evolution law of semantic change and geometric elevation change between adjacent time points, so as to obtain the terrain state prediction result of the area in front. The terrain condition prediction results of the area ahead are spatiotemporally aligned with the current comprehensive environmental situation map, and the predicted information and historical measured information are matched in the same overhead grid space to achieve position matching. Based on the predicted information after location matching and the historical measured information, confidence weights are assigned to the same grid location. Through weighted fusion, the predicted terrain status of the area in front is calculated with the comprehensive environmental situation map of the current and historical time, resulting in an enhanced spatiotemporal situation map that includes the measured terrain information of the historical time and the predicted terrain information of the future time.

[0010] Furthermore, the robot's current pose and velocity commands are acquired, and these commands, together with the enhanced spatiotemporal situational map, constitute a state space. This state space is then processed by a reinforcement learning policy network to acquire joint target angles, which are converted into driving torques by the underlying controller to achieve autonomous navigation. This includes: The robot's current pose and velocity commands are obtained, and the enhanced spatiotemporal situation map is combined with the robot's current pose information and velocity commands to form a joint state space that describes the robot's current environmental state and its own motion state. The joint state space is input into the navigation policy network, which is pre-trained using a reinforcement learning framework and performs forward reasoning based on the received joint state space to output the target motion angles of each joint of the robot. The target motion angle of each joint is input to the underlying PID controller, and the PID controller converts it into the driving torque corresponding to the motor of each joint; By driving the robot's joints with driving torque, the robot can perform corresponding movements, enabling it to autonomously select landing points and adjust its motion posture based on the terrain understanding and prediction information contained in the enhanced spatiotemporal situation map, thus completing navigation tasks in complex terrain.

[0011] Secondly, a reinforcement learning-based autonomous navigation system for quadruped robots in complex terrain includes: The acquisition module is used to acquire multimodal perception data of the robot when it walks in complex terrain environment. After processing, it generates terrain semantic map and terrain elevation map, and extracts geometric feature map from terrain elevation map and semantic feature map from terrain semantic map. The fusion module is used to project semantic feature maps and geometric feature maps onto the same top-down grid space centered on the robot, construct multiple grid layers with different resolutions, and associate and combine semantic features and geometric features at the same resolution for each grid position to form a fusion feature unit at that resolution. The calculation module is used to calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic features in the neighborhood of each grid location. Based on this complexity, the contribution weights of fused feature units at different resolutions are dynamically adjusted, and a comprehensive environmental situation map is generated after weighted integration. The analysis module is used to acquire comprehensive environmental situation maps at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, it analyzes the dynamic evolution trend of the terrain and predicts the terrain status of the area ahead, thereby forming an enhanced spatiotemporal situation map. The control module is used to acquire the robot's current pose and velocity commands. The pose and velocity commands, together with the enhanced spatiotemporal situation map, constitute the state space. The state space is used to acquire the joint target angles through a reinforcement learning policy network and is converted into driving torques by the underlying controller to achieve autonomous navigation.

[0012] The above-described solution of the present invention has at least the following beneficial effects: This invention generates semantic and geometric features from multimodal perception data and performs multi-resolution fusion within the same top-down grid space, overcoming the problems of information fragmentation and inability to represent collaboratively in traditional methods, thereby achieving a unified understanding of terrain material and physical structure. By dynamically adjusting the fusion weights based on local terrain complexity, it overcomes the problem that fixed weights cannot adapt to varying terrain, thus generating a comprehensive environmental situation map that accurately represents terrain differences. By analyzing semantic and elevation changes between adjacent moments in time series, it predicts the terrain state ahead and forms an enhanced spatiotemporal situation map, overcoming the problem that single-frame perception cannot cope with occlusion and terrain changes, thereby providing the robot with advanced environmental perception capabilities. Finally, the enhanced spatiotemporal situation map and the robot's state are jointly input into a reinforcement learning policy network to achieve end-to-end autonomous navigation from environmental understanding to motion control, significantly improving the quadruped robot's ability to traverse complex terrains and its navigation safety. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a reinforcement learning-based autonomous navigation method for quadruped robots in complex terrain, provided by an embodiment of the present invention.

[0014] Figure 2This is a schematic diagram of a quadruped robot autonomous navigation system for complex terrain based on reinforcement learning, provided by an embodiment of the present invention. Detailed Implementation

[0015] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0016] like Figure 1 As shown, an embodiment of the present invention proposes an autonomous navigation method for quadruped robots in complex terrain based on reinforcement learning. The method includes the following steps: Step 1: Acquire multimodal perception data of the robot walking in complex terrain environment, generate terrain semantic map and terrain elevation map after processing, extract geometric feature map from terrain elevation map and semantic feature map from terrain semantic map; Step 2: Project the semantic feature map and geometric feature map onto the same top-down grid space centered on the robot to construct multiple grid layers with different resolutions. For each grid position, associate and combine the semantic features and geometric features at the same resolution to form a fused feature unit at that resolution. Step 3: Calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution in the neighborhood of each grid location. Based on this complexity, dynamically adjust the contribution weight of fused feature units at different resolutions, and generate a comprehensive environmental situation map after weighted integration. Step 4: Obtain comprehensive environmental situation map at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, analyze the dynamic evolution trend of the terrain and predict the terrain status of the area ahead to form an enhanced spatiotemporal situation map. Step 5: Obtain the robot's current pose and velocity commands, which together with the enhanced spatiotemporal situation map constitute the state space. Input the state space into the reinforcement learning policy network, output the joint target angles, and convert them into driving torques through the underlying controller to achieve autonomous navigation.

[0017] In this embodiment of the invention, semantic and geometric features are generated by acquiring multimodal perception data, overcoming the incompleteness of information from a single sensor and achieving comprehensive perception of terrain material and physical structure. By projecting semantic and geometric features onto the same top-view grid space and constructing a multi-resolution fusion unit, the problem of semantic and geometric information being separated is overcome, achieving precise spatial alignment of the two types of information. By dynamically adjusting the fusion weights based on local terrain complexity, the inability of fixed weights to adapt to varying terrain is overcome, generating a comprehensive environmental situation map that accurately represents terrain differences. By analyzing semantic and elevation changes at adjacent moments and predicting the terrain state ahead, the limitations of single-frame perception in dealing with occlusion and terrain changes are overcome, providing advanced environmental perception capabilities. By inputting the enhanced spatiotemporal situation map and robot state into the reinforcement learning policy network, the problem of perception and control being separated is overcome, achieving end-to-end autonomous navigation from environmental understanding to motion control, significantly improving the quadruped robot's ability to traverse complex terrain and its navigation safety.

[0018] In a preferred embodiment of the present invention, step 1 above includes: To acquire multimodal perception data of the robot while it walks in complex terrain environments, the multimodal perception data includes at least environmental images, laser ranging data, inertial measurement data, and joint state data. Specifically, this includes: activating various sensing devices mounted on the quadruped robot, ensuring all sensing devices operate synchronously and their parameters are correctly calibrated; the environmental images are acquired in real-time by a high-definition camera mounted on the robot's head, with a acquisition frequency of 30 frames per second, covering a 180-degree field of view in front of the robot and an area 0 to 10 meters away from the robot; noise points in the images are filtered in real-time during the acquisition process; laser ranging data is acquired by lidar mounted on both sides of the robot's body, with the lidar acquiring data at a frequency of 1... The robot emits laser pulses at a frequency of 0 Hz to the surrounding environment. After receiving the reflected signals, it calculates the laser propagation time and converts it into distance data. The acquisition range covers a 360-degree area around the robot and a distance of 0 to 5 meters from the robot. Simultaneously, it records the emission angle and orientation information of the laser pulses. Inertial measurement data is collected by an inertial measurement unit (IMU) mounted inside the robot's torso. This unit collects the robot's acceleration, angular velocity, and attitude angle data in real time at a frequency of 100 times per second. During the acquisition process, a Kalman filter algorithm is used to eliminate vibration interference. The process consists of two iterative steps: prediction and update. The specific calculation formula for the prediction step is: [Formula omitted for brevity]. ,in, This is the predicted value of the system state at the current moment. Here is the state transition matrix. This is the optimal estimate of the system state at the previous moment. k Given the current time, and combining it with the state transition matrix A, the predicted state value at the current time is calculated; the error covariance matrix of the predicted state at the current time is calculated using the following formula: ,in, k Indicates the current moment. Let be the error covariance matrix of the predicted state at the current time. A Here is the state transition matrix. Let $\mathbf{ ... Let Q be the transpose of the state transition matrix A, and let Q be the process noise covariance matrix. The specific calculation formula for the update step is as follows: Calculate the Kalman gain at the current time step using the following formula: ,in, The Kalman gain at the current moment. Let be the error covariance matrix of the predicted state at the current time. k Indicates the current moment. Observation matrix The transpose of the matrix, For the observation matrix, The observation noise covariance matrix is ​​used; the Kalman gain is used to balance the confidence of the predicted and observed values. A larger gain indicates a higher weight for the observed values, and vice versa; the optimal estimate of the system state at the current time is calculated using the following formula: X =X - +K ( -H X - ),in, k Indicates the current moment. X This is the optimal estimate of the system state at the current moment. X - This is the predicted value of the system state at the current moment. K The Kalman gain at the current moment. The system's observations at the current moment, The observation matrix; via Kalman gain K For predicted values X - Make corrections to obtain a result that more closely approximates the true state of the system.X The optimal estimate is obtained; the error covariance matrix of the current optimal estimate state is updated using the following formula: ,in, Let be the error covariance matrix of the optimal estimated state at the current moment. It is the identity matrix. The Kalman gain at the current moment. For the observation matrix, The error covariance matrix is ​​the predicted state at the current moment. The entire process involves two steps: continuous iterative prediction and updating, gradually approximating the true state of the system to ensure that the data accurately reflects the robot's motion state and posture changes. Joint state data is collected by angle and torque sensors mounted on the joints of the robot's four limbs, recording the current rotation angle, rotation speed, and torque of each joint in real time. The collection frequency is 50 times per second, which is synchronously transmitted to the data processing module to achieve temporal and spatial synchronization with other sensing data. Finally, all collected environmental images, laser ranging data, inertial measurement data, and joint state data are summarized to form a complete multimodal sensing dataset.

[0019] The process involves pixel-level terrain category analysis of environmental images to obtain a terrain semantic map and extract semantic feature maps from it. Spatial interpolation reconstruction is performed on laser ranging data to generate a terrain elevation map centered on the robot, from which geometric feature maps reflecting terrain undulations are extracted. Specifically, this includes: preprocessing the acquired environmental images, which involves image grayscale conversion, Gaussian blur denoising, and image enhancement. Image grayscale conversion converts color environmental images to grayscale; Gaussian blur denoising smooths the grayscale image using a Gaussian filter, filtering out random noise and isolated noise points while preserving terrain contour information; and image enhancement improves image contrast using histogram equalization to highlight the boundary differences between different terrain categories. After preprocessing, pixel-level terrain category analysis is performed using a pre-defined terrain category template matching combined with pixel feature analysis. The specific process is as follows: First, a standard template library of pre-defined terrain categories is constructed, containing labels for all pre-defined terrain categories such as grass, rocks, soil, steps, puddles, and obstacles. For each terrain category, samples were collected under different lighting and angles to ensure the template's universality and comprehensiveness. Simultaneously, core pixel features of each standard template were extracted, including pixel grayscale value range, pixel grayscale variation patterns, and local pixel texture features, establishing a correspondence between the template and pixel features. During parsing, a single pixel in the preprocessed grayscale image was first selected as the parsing object. The grayscale value of this pixel, the grayscale variation trend of surrounding pixels, and local texture features were extracted. Then, the extracted pixel features were compared one by one with the standard pixel features of all terrain categories in the template library, calculating feature similarity. After comparison, the terrain category with the highest feature similarity was selected as the initial classification result for this pixel. A secondary verification was then performed, combining the template matching results. A comprehensive analysis of pixel features within a certain range around this pixel was conducted to determine if the classification results of the surrounding pixels were consistent with the initial classification result of this pixel. If consistent, the classification result of this pixel was confirmed; otherwise, the features were re-compared and the classification adjusted to ensure accuracy.During the parsing process, the category matching credibility of each pixel is calculated. Credibility is determined comprehensively by indicators such as feature similarity and consistency with surrounding pixels. Pixel classification results with a credibility greater than 90% are retained, while blurry or unreliable pixels are removed. Finally, a terrain semantic map is generated, where the value of each pixel corresponds to its terrain category, achieving a visual representation of environmental terrain categories. Semantic feature maps are extracted from the terrain semantic map. The extraction process uses a convolutional neural network to extract features from the terrain semantic map. Through multi-layer convolutional operations, global and local features of terrain categories are gradually extracted, including the distribution patterns of terrain categories, the boundary contours of different terrain categories, and the area proportion of each terrain category. The features are normalized to ensure that the feature values ​​are on the same order of magnitude, ultimately generating a semantic feature map. This semantic feature map accurately represents the semantic information of different terrain categories. The collected laser ranging data undergoes preprocessing, which includes two steps: data denoising and data completion. Data denoising removes abnormal data points caused by obstacles or abnormal signal reflections by setting distance and signal strength thresholds. The distance threshold is set to 0.05 meters to 5 meters, and the signal strength threshold is set to 20 dBm to 80 dBm, retaining valid ranging data. Data completion addresses missing data areas encountered during laser radar acquisition by using a linear interpolation algorithm. The calculation formula is as follows: ,in, Synchronized data values ​​under a unified time base This refers to the sensor data value from the previous time point. The sensor data value at the next time point. To ensure the continuity and integrity of the ranging data, a unified time point is established to supplement the data. After preprocessing, spatial interpolation reconstruction is performed to generate a terrain elevation map. A three-dimensional coordinate system is established with the robot's current position as the origin, the horizontal X and Y axes as the horizontal axis, and the vertical Z axis as the vertical axis. The laser ranging data is converted into three-dimensional spatial coordinate points in this coordinate system. The X and Y coordinates of each coordinate point correspond to the position on the horizontal plane, and the Z coordinate corresponds to the terrain elevation of that position. A weighted average interpolation method based on the elevation correlation of adjacent coordinate points is used to interpolate the three-dimensional spatial coordinate points. The specific process is as follows: First, determine the effective three-dimensional spatial coordinate points within a certain range around each grid point to be interpolated, i.e., the actual elevation points collected by the laser radar. Calculate the straight-line distance between the grid point to be interpolated and each surrounding effective coordinate point. The closer the distance, the higher the weight of the effective coordinate point; the farther the distance, the lower the weight. The weight is determined by the inverse relationship of distance to ensure that the surrounding effective points have a greater impact on the interpolation result. Multiply the elevation value of each surrounding effective coordinate point by its corresponding weight. Then, sum all the product results to obtain the elevation value of the grid point to be interpolated. This method is used to sequentially calculate the elevation values ​​of all grid points within the entire detection range. The grid size is set to 0.1m × 0.1m to ensure the terrain elevation map accurately reflects the subtle undulations of the terrain. Finally, a terrain elevation map centered on the robot is generated. The value of each grid in the terrain elevation map corresponds to the terrain elevation at that location. Geometric feature maps reflecting the degree of terrain undulation are extracted from the terrain elevation map. The extraction process includes calculating three core indicators: terrain slope, terrain roughness, and terrain undulation. Terrain slope is obtained by calculating the ratio of the elevation difference between adjacent grids to the horizontal distance, reflecting the degree of terrain inclination. Terrain roughness is obtained by calculating the elevation standard deviation of neighboring grids around a given grid, reflecting the smoothness of the terrain surface. Terrain undulation is obtained by calculating the difference between the highest and lowest grid elevations in a certain area, reflecting the overall undulation of the terrain in that area. The calculated terrain slope, terrain roughness, and terrain undulation are then fused and normalized to generate a geometric feature map.

[0020] In this embodiment of the invention, a complete multimodal perception data acquisition and preprocessing process ensures the accuracy, completeness, and synchronization of various perception data such as environmental images and laser ranging data, overcoming the shortcomings of incomplete and easily interfered data acquisition by a single sensor. By generating a terrain semantic map through pixel-level terrain category parsing and extracting semantic feature maps, accurate identification and semantic representation of terrain material categories are achieved. By generating a terrain elevation map through spatial interpolation reconstruction of laser ranging data and extracting geometric feature maps, accurate characterization of the terrain's physical structure and undulation is achieved. The synchronous generation of semantic feature maps and geometric feature maps ensures that the robot can comprehensively and accurately understand complex terrain environments during navigation, providing data support for the safety and reliability of autonomous navigation.

[0021] In a preferred embodiment of the present invention, step 2 above includes: Using terrain semantic maps and terrain elevation maps as processing objects, coordinate transformation is used to project them onto the same top-view raster coordinate system centered on the robot's current position, resulting in spatially aligned top-view semantic raster maps and top-view elevation raster maps. Specifically, this includes: defining the original coordinate systems of the terrain semantic map and terrain elevation map. The terrain semantic map is generated from the environmental image through pixel-level analysis, and its original coordinate system is the image pixel coordinate system, with the top-left corner of the image as the origin, the horizontal axis to the right as the x-axis, and the vertical axis downwards as the y-axis. The coordinates of each pixel correspond to its position in the image, and the numerical value corresponds to the terrain semantic category; The terrain elevation map is generated from laser ranging data through spatial interpolation. Its original coordinate system is the robot's local 3D coordinate system, with the robot's current position as the origin, the horizontal axis as the x and y axes, and the vertical axis as the y-axis. Each coordinate point corresponds to its position in real space, and the value corresponds to the terrain elevation. A common top-view grid coordinate system is then established centered on the robot's current position. This coordinate system is a 2D plane coordinate system, with the robot's current center of gravity as the origin, the horizontal axis pointing forward as the positive x-axis, and the horizontal axis pointing to the left as the positive y-axis. The grid size is consistent with the grid size of the subsequent terrain elevation map, both set to 0.1 meters. The grid area is 0.1 meters, covering a range of 0 to 5 meters around the robot, ensuring complete coverage of the terrain range the robot needs to perceive during its movement. Next, coordinate transformation and projection are performed on the terrain semantic map. First, the real spatial range corresponding to the terrain semantic map is obtained. Combining the installation height, shooting angle, and focal length parameters of the robot's high-definition camera, the coordinates of each pixel in the image pixel coordinate system are converted into spatial coordinates in the robot's local 3D coordinate system. During the conversion, errors caused by image distortion must be eliminated to ensure the accuracy of the real spatial position corresponding to each pixel. The converted 3D spatial coordinates are projected onto a top-view grid coordinate system centered on the robot. Vertical top-view projection is used, ignoring the vertical axis elevation information in the 3D spatial coordinate system and retaining only the horizontal and vertical axis position information. The spatial position of each pixel is matched to the corresponding grid cell in the top-view grid coordinate system. Simultaneously, the terrain semantic category value of that pixel is assigned to the corresponding grid cell. For multiple pixels projected to the same grid cell, the semantic category with the highest frequency is selected as the grid cell. For semantic values, if they appear with the same frequency, the value with the higher semantic category priority is selected, ultimately generating a top-view semantic raster map. Simultaneously, the terrain elevation map undergoes coordinate transformation and projection. The original coordinate system of the terrain elevation map is the robot's local 3D coordinate system, so no 3D coordinate transformation is needed; it is directly projected onto the same top-view raster coordinate system. The projection process also uses vertical top-view projection, matching the horizontal and vertical axis positions of each 3D spatial coordinate point in the terrain elevation map to the corresponding raster cell in the top-view raster coordinate system, and assigning the elevation value of that coordinate point to the corresponding raster cell. For multiple 3D coordinate points projected to the same raster cell, the average elevation value of that raster cell is calculated to ensure the accuracy and rationality of the elevation data. After projection, the top-view semantic raster map and the top-view elevation raster map are spatially aligned. The spatial positions of corresponding raster cells in the two raster maps are checked to ensure complete alignment. If deviations exist, they are corrected by fine-tuning the coordinate projection parameters to ensure that the same raster cell corresponds to the same position in real space, ultimately resulting in a spatially aligned top-view semantic raster map and a top-view elevation raster map.

[0022] Based on the top-view semantic raster map and the top-view elevation raster map, multi-scale downsampling processing is performed to construct multiple raster layers including the original resolution, 1 / 2 resolution, and 1 / 4 resolution. This ensures that each resolution layer contains a corresponding semantic feature sub-map and a geometric feature sub-map. Specifically, the parameters of the original resolution raster layer are defined. The original resolution is the initial resolution of the top-view semantic raster map and the top-view elevation raster map, with a raster size of 0.1m × 0.1m. The original resolution raster layer directly uses the spatially aligned top-view semantic raster map and the top-view elevation raster map, where the top-view semantic raster map serves as the semantic feature sub-map at the original resolution, and the top-view elevation raster map serves as the geometric feature sub-map at the original resolution. No downsampling processing is required; the original raster data and feature information are directly preserved. Then, 1 / 2 resolution downsampling processing is performed to construct a 1 / 2 resolution raster layer. The downsampling processing is applied to both the top-view semantic raster map and the top-view elevation raster map. The processing method employed is the neighborhood raster merging method. Specifically, for a top-view semantic raster map, a merging unit is formed by combining 2×2 adjacent original raster cells. The four original raster cells within each merged unit are merged to form a new raster cell. The semantic feature value of the new raster cell is selected from the semantic category values ​​with the highest frequency among the four original raster cells within the merged unit. If the frequencies are the same, the value with the higher semantic category priority is selected. For a top-view elevation raster map, a similar method is used, where a merging unit is formed by combining 2×2 adjacent original raster cells. The elevation values ​​of the four original raster cells within the merged unit are calculated using an arithmetic mean. Specifically, the elevation values ​​of the four original raster cells within the merged unit are added together to obtain a total elevation value. This sum is then divided by 4, and the result is the elevation value of the new raster cell, i.e., the geometric feature value. This method reduces the size of the original resolution raster map to half its original size, resulting in a raster size of 0.2 meters × 0.A 2-meter depth is used to create a 1 / 2 resolution semantic feature sub-map and a 1 / 2 resolution geometric feature sub-map, which together constitute a 1 / 2 resolution raster layer, ensuring that the semantic and geometric features of this layer correspond to the original resolution layer. Next, a 1 / 4 resolution downsampling process is performed to construct a 1 / 4 resolution raster layer. Downsampling is also performed on both the top-view semantic raster map and the top-view elevation raster map, using the neighboring raster merging method. Specifically, 4×4 adjacent original raster cells are used as a merging unit, and the 16 original raster cells within this unit are merged to form a new raster cell. For the top-view semantic raster map, the semantic feature value of the new raster cell is selected from the semantic category value with the highest frequency among the 16 original raster cells in the merging unit. If the frequencies are the same, the value with the higher semantic category priority is selected. For the top-view elevation raster map, the geometric feature value of the new raster cell is selected from the arithmetic mean of the elevation values ​​of the 16 original raster cells in the merging unit. The specific calculation process is as follows: The elevation values ​​of all 16 original raster cells within the merged unit are summed to obtain a total elevation value. This sum is then divided by 16 to calculate the geometric feature value of the new raster cell. This method reduces the size of the original resolution raster image to one-quarter of its original size, resulting in a raster size of 0.4 meters × 0.4 meters. This forms a one-quarter resolution semantic feature sub-image and a geometric feature sub-image, which together constitute a one-quarter resolution raster layer. After downsampling, consistency checks are performed on the three resolution raster layers to ensure that the semantic and geometric feature sub-images in each resolution layer are precisely aligned in space. The number and position of raster cells in the semantic and geometric feature sub-images within the same resolution layer are completely consistent. Raster cells between different resolution layers can achieve spatial correspondence and matching. Finally, multiple raster layers are constructed, including the original resolution, half-resolution, and one-quarter resolution layers. Each resolution layer contains one semantic feature sub-image and one geometric feature sub-image.

[0023] For each raster cell in each resolution layer, the semantic and geometric feature values ​​corresponding to the raster cell are associated and combined to generate a fused feature cell at each resolution. Specifically, this involves: defining the composition of each resolution layer, where each layer contains a semantic feature sub-map and a geometric feature sub-map, with the number and position of raster cells perfectly aligned. Each raster cell in each resolution layer corresponds to a unique semantic feature value in the semantic feature sub-map and a unique geometric feature value in the geometric feature sub-map. The semantic feature value represents the category information of the terrain corresponding to the raster cell, and the geometric feature value represents the elevation information of the terrain corresponding to the raster cell. Feature association is then performed on the original resolution raster layer. The process involves selecting each raster cell in the original resolution layer one by one. First, the semantic feature value corresponding to the raster cell is extracted from the semantic feature sub-image of that layer. The semantic feature value is the corresponding value of the preset terrain category. Then, the geometric feature value corresponding to the raster cell is extracted from the geometric feature sub-image of that layer. The geometric feature value is the elevation value of the raster cell. After extraction, the semantic and geometric feature values ​​are standardized to a uniform value range of 0 to 1, eliminating the influence of differences in the magnitude of different feature values. The standardization process uses linear normalization. Based on the semantic and geometric feature value ranges of all raster cells, the normalization of the feature value of each raster cell is calculated separately. As a result, after standardization, the normalized semantic feature value and normalized geometric feature value of the raster unit are associated and combined using feature concatenation. The two feature values ​​are concatenated sequentially to form a two-dimensional feature vector. This two-dimensional feature vector is the fused feature unit of the raster unit at the original resolution. The fused feature unit simultaneously contains the terrain category information and terrain elevation information corresponding to the raster unit, achieving the initial fusion of semantic and geometric features. Following the above process, all raster units in the original resolution layer are processed one by one to generate all fused feature units at the original resolution. All fused feature units are arranged according to the spatial position of their corresponding raster units to form a fused feature layer at the original resolution. Subsequently, according to... The same process is used to process the half-resolution raster layer and the quarter-resolution raster layer separately. For the half-resolution layer, each raster cell is selected one by one, and the semantic feature value and geometric feature value in the corresponding semantic feature sub-map and geometric feature sub-map of that raster cell are extracted. The two feature values ​​are standardized to eliminate differences in numerical magnitude. Then, the two normalized feature values ​​are associated and combined by feature concatenation to generate a fused feature cell at half-resolution. After processing all raster cells in sequence, a fused feature layer at half-resolution is formed. For the quarter-resolution layer, the same method is used to extract the semantic feature value and geometric feature value of each raster cell, perform standardization, and then associate and combine them.The process generates fused feature units at one-quarter resolution, ultimately forming a fused feature layer at one-quarter resolution. During feature association and combination, it is crucial to ensure that the fused feature units within each resolution layer maintain spatial consistency with the semantic and geometric feature sub-maps of that layer. Fusion feature units across different resolution layers must be correlated via spatial coordinates. Simultaneously, the terminology for semantic and geometric feature values ​​must be consistent, and their numerical values ​​accurate. This process ultimately completes the generation of fused feature units across all resolution layers.

[0024] In this embodiment of the invention, the terrain semantic map and terrain elevation map are projected onto the same top-view grid coordinate system through coordinate transformation, achieving precise spatial alignment of semantic features and geometric features, overcoming the problem of feature incoordination caused by spatial misalignment; multiple resolution grid layers are constructed through multi-scale downsampling, preserving subtle terrain features at the original resolution while capturing global terrain features through low-resolution layers, achieving multi-scale representation of terrain features; by associating and combining the semantic and geometric features of grid units, fused feature units are generated, organically combining terrain category information and elevation information, overcoming the deficiency of incomplete representation by a single feature, effectively improving the quadruped robot's perception and understanding of complex terrain environments, and providing a guarantee for the stability and reliability of autonomous navigation.

[0025] In a preferred embodiment of the present invention, step 3 above includes: Based on the constructed fusion feature units at the original resolution, 1 / 2 resolution, and 1 / 4 resolution, the fusion feature units at all grid positions in each resolution layer are aggregated to form a fusion feature map corresponding to the resolution. Specifically, this involves clarifying the distribution pattern of fusion feature units in each resolution layer. The original resolution, 1 / 2 resolution, and 1 / 4 resolution layers each contain a different number of fusion feature units, and each fusion feature unit corresponds to a unique grid position in its respective resolution layer. The fusion feature units exist in the form of two-dimensional feature vectors, containing both semantic and geometric feature information corresponding to the grid position. All fusion feature units have been standardized, with values ​​uniformly between 0 and 1, ensuring comparability between fusion feature units at different resolutions. First, fusion feature units are aggregated for the original resolution layer. The fusion feature units corresponding to all grid units in the original resolution layer are extracted one by one. According to the spatial position of the grid units in the top-view grid coordinate system, from left to right and from top to bottom, all fusion feature units are arranged sequentially, strictly corresponding to each other during the arrangement process. The original raster position of each fused feature unit is determined to ensure that the arranged fused feature units are completely consistent with the raster positions of the original resolution raster layer, without any misalignment, omission, or duplication. After arrangement, a two-dimensional feature matrix is ​​formed, which is the fused feature map corresponding to the original resolution. Each element in the matrix corresponds to a fused feature unit, and the number of rows and columns of the matrix are exactly the same as the number of raster rows and columns of the original resolution raster layer, which can completely represent the fused feature information of all raster positions at the original resolution. Subsequently, following the same process, the fused feature units of the half-resolution layer and the quarter-resolution layer are collected respectively. For the half-resolution layer, the fused feature units corresponding to all raster units in the layer are extracted, and arranged sequentially according to their spatial positions in the top-view raster coordinate system to form a two-dimensional feature matrix. This matrix is ​​the fused feature map corresponding to the half-resolution layer, and the number of rows and columns of the matrix are consistent with the number of raster rows and columns of the half-resolution raster layer. The raster positions can be matched with the raster positions of the original resolution fused feature map through spatial coordinates.For the quarter-resolution layer, all fusion feature units are extracted and arranged sequentially according to their spatial location to form a two-dimensional feature matrix, which serves as the fusion feature map corresponding to the quarter-resolution layer. The number of rows and columns in this matrix is ​​consistent with the number of raster rows and columns in the quarter-resolution raster layer, ensuring that the fusion feature maps of the three resolutions are accurately correlated in spatial location. After aggregation, the fusion feature maps of the three resolutions are verified to check whether the fusion feature units in each fusion feature map are complete and whether the arrangement order is correct, ensuring that there is no misalignment between the fusion feature units and their corresponding raster positions, and that the spatial correspondence between the fusion feature maps of different resolutions is accurate. At the same time, it is confirmed that the semantic and geometric feature terminology of all fusion feature units is consistent and the values ​​are correct, ultimately forming the fusion feature maps corresponding to the original resolution, half-resolution, and quarter-resolution layers.

[0026] The fused feature maps are concatenated along the feature channel direction to construct a concatenated feature tensor containing multi-scale information. Specifically, this involves: defining the structural features of the fused feature maps at three resolutions; the original resolution, half-resolution, and quarter-resolution fused feature maps are all two-dimensional feature matrices, with each element being a two-dimensional fused feature vector. That is, each fused feature unit contains two feature channels, corresponding to the semantic feature channel and the geometric feature channel, respectively. The three fused feature maps have different spatial dimensions, but their spatial positions can be precisely matched, and the type and numerical range of the feature channels are consistent, providing a prerequisite for feature channel concatenation. Before performing the cascading operation, the spatial dimensions of the three fused feature maps at different resolutions are unified. Using the spatial dimensions of the original resolution fused feature map as the standard, upsampling is performed on the half-resolution and quarter-resolution fused feature maps to ensure that the number of raster rows and columns of the three fused feature maps are completely consistent and that their spatial positions are precisely aligned. The upsampling process uses neighborhood interpolation. For the half-resolution fused feature map, the fused feature unit of each raster cell is copied and expanded into 2×2 raster cells, ensuring that the expanded raster size is consistent with the original resolution. For the quarter-resolution fused feature map, each raster cell is... The fusion feature unit of the grid cell is replicated and expanded into 4×4 grid cells to achieve spatial size uniformity with the original resolution fusion feature map. During upsampling, the values ​​of the fusion feature units remain unchanged to ensure no loss or distortion of feature information. After spatial size uniformity is achieved, a cascading operation is performed along the feature channel direction. The cascading operation targets fusion feature units at corresponding grid positions in the three fusion feature maps. Specifically, the fusion feature units corresponding to the grid cells at the same spatial position in the three fusion feature maps are selected, and the feature channels of these three fusion feature units are stitched together. The two feature channels of the original resolution fusion feature unit are then combined. Two feature channels of a half-resolution fusion feature unit and two feature channels of a quarter-resolution fusion feature unit are sequentially concatenated to form a six-dimensional feature vector. This six-dimensional feature vector contains semantic and geometric feature information of the same spatial location at three different resolutions. Following the above method, all grid positions are processed one by one, and the three fusion feature units corresponding to each grid position are concatenated along the feature channels to form a six-dimensional feature vector. Then, the six-dimensional feature vectors of all grid positions are arranged sequentially according to their spatial positions to form a three-dimensional feature structure. This three-dimensional feature structure is the cascaded feature tensor containing multi-scale information.

[0027] For each grid position in the cascaded feature tensor, an optimization algorithm based on the quasi-Newton method is introduced. The optimal weight parameters for fusing features of different resolutions at that position are determined through iterative calculation. Specifically, for each grid position in the cascaded feature tensor, the weight parameters need to be calculated iteratively independently to ensure that the weight parameters are adapted to the terrain features at that position. The quasi-Newton method is mainly used to solve for the optimal values ​​of the weight parameters. The specific steps are as follows: Initial value setting formula: , ;in, The initial weight parameter vector is set to 1 / 3 for each initial weight parameter to ensure that the initial contribution weights of the three resolutions are the same and that the sum is 1. The inverse of the initial approximate Hessian matrix. It is a 3×3 identity matrix; This is a general term for weighting parameters. m This indicates the current iteration number; the vector contains three elements, corresponding to the weight parameters for the original resolution, half-resolution, and quarter-resolution, respectively. ;in, w This is a general term for weighting parameters. Corresponding to the original resolution weights, Corresponding to half-resolution weights, Corresponding to quarter-resolution weights; Search direction calculation formula: ,in, This is the search direction vector for the current iteration. m Indicates the current iteration number. This is the inverse of the approximate Hessian matrix at the current iteration number. This is the error gradient vector at the current iteration number. The term "approximate inverse matrix of Hessian matrix" is used to refer to the general term for the matrix that approximates the inverse of the Hessian matrix. This refers to the general term for error gradients; the formula for determining the step size is: ,in, This represents the step size at the current iteration number. m Indicates the current iteration number. This indicates the search for minimizing the fusion error function within the parentheses. value; To determine the step size variable for updating the weight parameters, This is a general term for fusion error. This is the search direction vector for the current iteration. w The term "weight parameters" is a general term; the formula for updating weight parameters is: ;in, This is the weight parameter vector for the next iteration. This is the weight parameter vector for the current iteration number. w This is a general term for weighting parameters.m Indicates the current iteration number; This represents the step size at the current iteration number. Let be the search direction vector for the current iteration; the error gradient update formula is: ,in, w This is a general term for weighting parameters. This is a general term for error gradients. This is the error gradient vector for the next iteration. Represents the fusion error function E(w) exist Find the gradient at the given point. m Indicates the current iteration number. This is a general term for fusion error. This is the weight parameter vector for the next iteration; the update formula is approximately the inverse of the Hessian matrix: H +1 = H + (1 + (y H y ) / (s y )) (s s ) / (s y ) -(H y s + s y H ) / (s y );in,H +1 This is the approximate inverse of the Hessian matrix for the next iteration. H This is the inverse of the approximate Hessian matrix at the current iteration number. m Indicates the current iteration number; y This is the vector of error gradient changes in the current iteration. ;in, This is a general term for error gradients. m Indicates the current iteration number. s This is the vector of changes in the weight parameters during the current iteration. , T Represents the transpose of a vector; Iteration termination determination formula: or ,in, This represents the maximum absolute value of the difference between the corresponding elements in the weight parameter vector of the next iteration and the current iteration. This is the weight parameter vector for the next iteration. This is the weight parameter vector for the current iteration. m Indicates the current iteration number. The threshold for the change in the weighting parameter is set to 0.001. The iteration threshold is set to 50. The judgment process is as follows: if the maximum update amount of the weight parameter is less than 0.001, it indicates that the weight parameter has stabilized and no further iteration is needed; if the current iteration number is less than 0.001, it indicates that the weight parameter has stabilized and no further iteration is needed. m Reaching 50 iterations indicates that the maximum iteration limit has been reached, and iteration stops. At this point, the current weight parameter vector... This is the optimal weight parameter vector for that grid location. The entire iterative process proceeds step by step according to the above formula, starting from the initial weight parameters, sequentially calculating the search direction, step size, updated weight parameters, error gradient, and approximate Hessian matrix inverse, until the iteration termination condition is met. Finally, the optimal weight parameters for each grid location are obtained, ensuring that the weight parameters are adapted to the terrain features of the corresponding grid location. Weight parameters are initialized for each grid location. The three original resolutions, half-resolution, and quarter-resolution correspond to three weight parameters, with the initial weight parameters all set to one-third, meaning the initial contribution weights of the features at the three resolutions are the same. The initial weight parameters must satisfy the constraint that the sum is one to ensure the rationality of the weight allocation. Simultaneously, iteration termination conditions are set, including an iteration number threshold and a weight parameter change threshold. The iteration number threshold is set to 50, and the weight parameter change threshold is set to 0.001. When the iteration number reaches... mWhen the weight parameter change is less than 0.001 after 50 iterations or between two consecutive iterations, the iteration stops, and the current weight parameter is determined to be the optimal weight parameter. After the iteration calculation begins, the current iteration is calculated. m The feature fusion error corresponding to the weight parameters is, where, m This indicates the current iteration number. The feature fusion error is obtained by calculating the deviation between the fused feature and the ideal feature. The ideal feature is set based on the actual terrain conditions at the grid location, combined with the semantic category and elevation information of that location, to determine the ideal fused feature value. Then, the current iteration... m The fused feature values ​​corresponding to the weighted parameters are compared with the ideal feature values, and the absolute error between the two is calculated. This absolute error is the value for the current iteration. m The fusion error corresponding to the weight parameters is calculated. The smaller the fusion error, the closer the weight parameters are to their optimal values. Based on the fusion error, the weight parameters are updated using a quasi-Newton method. The specific process is as follows: based on the current iteration... m The fusion error is used to calculate the update direction of the weight parameters, and the update direction is determined by the current iteration. m The product of the approximate Hessian matrix inverse and the error gradient is used to determine the Hessian matrix. Direct calculation of the Hessian matrix is ​​unnecessary; the current iteration's matrix is ​​constructed using the error from the previous iteration and the changes in the weight parameters. m The approximate Hessian matrix inverse is then combined with the current iteration. m The error gradient is used to obtain the update step size and direction of the weight parameters. The weight parameters for the next iteration are then calculated. The updated weight parameters need to be re-normalized to ensure that the sum of the three weight parameters is still one, satisfying the weight allocation constraints. Following this process, error calculation and weight parameter updates are repeated. After each iteration, the iteration termination condition is checked; if not, the iteration ends. m Increment by 1 and continue to the next iteration. If the condition is met, stop the iteration and determine the current weight parameter as the optimal weight parameter for that grid position. Each grid position completes the iterative calculation independently according to this process to ensure that the optimal weight parameter for each grid position can adapt to its own terrain features, avoiding the problem that a uniform weight cannot adapt to different terrain areas. Finally, the optimal weight parameters for all grid positions are determined.

[0028] Based on the optimal weight parameters, the fusion feature units at different resolutions at the same grid position are weighted and summed to generate fusion feature values ​​of multi-scale semantic and geometric information. This process iterates through all grid positions to obtain a fusion environment representation. Specifically, this includes: identifying the optimal weight parameters and three resolution fusion feature units corresponding to each grid position. Each grid position corresponds to three optimal weight parameters, corresponding to fusion feature units at the original resolution, half resolution, and quarter resolution, respectively, and the sum of the three weight parameters is one. The magnitude of each weight parameter corresponds to the contribution of the corresponding resolution feature to the fusion at that grid position; the larger the weight parameter, the higher the contribution of the corresponding resolution feature. First, a grid position in the cascaded feature tensor is selected, and the three resolution fusion feature units corresponding to that grid position are extracted. Each fusion feature unit is a two-dimensional feature vector containing semantic feature values ​​and geometric feature values. Simultaneously, the three optimal weight parameters corresponding to that grid position are extracted, each corresponding to a fusion feature unit at one of the three resolutions. Then, a weighted summation calculation is performed, which is divided into two parts: a weighted summation of semantic feature values ​​and a weighted summation of geometric feature values. The formula for the weighted summation of semantic feature values ​​is: ,in, This represents the fused semantic feature value of the raster position. The optimal weight parameters corresponding to the fused feature units at the original resolution are obtained through iterative calculation using the quasi-Newton method. The optimal weight parameters corresponding to the half-resolution fused feature units are also obtained through iterative calculation using the quasi-Newton method. This represents the optimal weight parameters corresponding to the quarter-resolution fused feature unit, and + + = 1, which satisfies the constraints of weight allocation; This represents the semantic feature values ​​in the original resolution fused feature unit. This represents the semantic feature value in the half-resolution fused feature unit. The semantic feature values ​​in the quarter-resolution fused feature unit are represented; the weighted summation formula for geometric feature values ​​is: .in, This represents the fused geometric feature value of the grid location. The optimal weight parameters corresponding to the fused feature units at the original resolution are obtained through iterative calculation using the quasi-Newton method. The optimal weight parameters corresponding to the half-resolution fused feature units are also obtained through iterative calculation using the quasi-Newton method. This represents the optimal weight parameters corresponding to the quarter-resolution fused feature unit, and + + =1, which satisfies the weight allocation constraint; This represents the geometric feature values ​​in the fused feature units of the original resolution. This represents the geometric feature value in the half-resolution fused feature unit. This represents the geometric feature values ​​in a quarter-resolution fused feature unit, which will be fused with semantic feature values. and fusion geometric eigenvalues G The features are concatenated to form a two-dimensional fusion feature vector, which represents the fusion feature value of the multi-scale semantic and geometric information of the grid location. Following the above process, all grid locations in the cascaded feature tensor are processed one by one. Each grid location is weighted and summed using the optimal weight parameters to generate a corresponding fusion feature value, ensuring that the fusion feature value of each grid location can integrate the advantages of features from three resolutions. After the fusion feature values ​​of all grid locations have been processed, they are arranged sequentially according to their spatial position in the top-view grid coordinate system to form a two-dimensional feature matrix. This two-dimensional feature matrix represents the fusion environment.

[0029] Based on the obtained fused environment representation, for each grid location, a complexity index representing the local terrain complexity is calculated according to the rate of change of geometric features and the degree of semantic category mixing within the neighborhood of the grid location. Specifically, this includes: clarifying the core structure of the fused environment representation, which is a two-dimensional feature matrix, with each grid location corresponding to a two-dimensional fused feature vector containing fused semantic feature values. R and fusion geometric eigenvalues G The fused semantic feature value R integrates semantic feature information from the original resolution, half-resolution, and quarter-resolution, corresponding to the steps described above. t1 , t2 , t3 Weighted summation result, incorporating geometric eigenvalues G This integrates geometric feature information from three resolutions, corresponding to the steps described above. h1 、 h2 、 h3 The weighted summation results show that all feature values ​​have been standardized. Before calculating the complexity index for each grid location, the neighborhood of that grid location is determined. The neighborhood is uniformly set to a 3×3 grid area, that is, centered on the current grid location and including the eight adjacent grid cells in the up, down, left, right, and diagonal directions. This ensures that the neighborhood can fully reflect the terrain feature changes around the current grid location. After determining the neighborhood, the fused geometric feature values ​​of all grid cells within the neighborhood are extracted one by one. G and fusion semantic feature values RThese two metrics are used to calculate the rate of change of geometric features and the degree of semantic category mixing, respectively. Together, they constitute the local terrain complexity index. First, the rate of change of geometric features within the neighborhood is calculated: the fused geometric feature values ​​of all raster cells within the neighborhood are extracted. G Calculate these G The average value is used as the average level of the neighborhood geometric features; then the average value of each adjacent grid cell is calculated. G Value and current grid position G The absolute differences of the values ​​are used to obtain eight absolute differences. The average of these eight absolute differences is calculated as the average change of the geometric feature. The average change of the geometric feature is divided by the average geometric feature of the neighborhood to obtain the geometric feature change rate. The result is rounded to four decimal places. Then, the degree of semantic category mixing within the neighborhood is calculated: first, the fused semantic feature value R of all raster cells within the neighborhood is extracted, according to... R The value corresponds to the terrain semantic category, and the number of terrain semantic categories appearing in the neighborhood is counted; then, the number of semantic categories appearing is divided by the total number of preset terrain categories to obtain the semantic category mixing degree. The local terrain complexity index is calculated using the arithmetic mean method: the geometric feature change rate and semantic category mixing degree of the current grid position are added together to obtain the sum of the two, and then the sum is divided by 2 to obtain the complexity index of the grid position. The complexity index of each grid position is calculated independently according to the above process to ensure that the complexity index of each position can accurately reflect the actual complexity of its local terrain.

[0030] Based on the complexity index of each grid location, the contribution weights of the fused feature units at the original resolution, 1 / 2 resolution, and 1 / 4 resolution at that location are dynamically determined. Specifically, this includes: different contribution weights for the fused feature units at the three resolutions have different complexity indices; clearly defined constraints for weight allocation; and the three contribution weights corresponding to each grid location are obtained from the above steps. These correspond to the fused feature units at the original resolution, half resolution, and quarter resolution, respectively. The sum of the three weights must be 1. The weight parameter constraints in the weighted summation formula above remain consistent. The specific process for determining the dynamic weights is as follows: based on the numerical range of the complexity index, the weights are allocated in three gradients. First gradient: When the complexity index is ≥0.7, it is determined that the local terrain is highly complex. At this time, the focus is on the subtle features of the original resolution, and the weight allocation is as follows: Original resolution contribution weight =0.6, half-resolution contribution weight =0.3, quarter-resolution contribution weight =0.1; Second gradient: When 0.3 < complexity index < 0.7, it is judged as moderately complex local terrain. At this time, both subtle features and global features are taken into account, and the weight allocation is as follows: Original resolution contribution weight =0.4, half-resolution contribution weight =0.4, quarter-resolution contribution weight =0.2; Third gradient: When the complexity index is ≤0.3, the local terrain is considered simple. In this case, the focus is on low-resolution global features, and the weights are allocated as follows: Original resolution contribution weight =0.2, half-resolution contribution weight =0.3, quarter-resolution contribution weight =0.5; Each grid position determines its gradient based on its own complexity index value, thereby obtaining the corresponding gradient. Three contribution weights are determined. After the weights are determined, it is necessary to check again whether the sum of the three weights is 1. After the contribution weights of all grid positions are dynamically determined, a dynamic weight set exclusive to each grid position is formed.

[0031] Based on dynamically determined contribution weights, the fusion feature units at different resolutions at each grid location are weighted and summed to obtain a fusion feature value that integrates semantic category and geometric undulation information. Specifically, this includes: identifying the core data for each grid location, with each grid location corresponding to three dynamically determined contribution weights. Each grid cell corresponds to a fusion feature unit at the original resolution, half the resolution, and a quarter the resolution, respectively, with the sum of the three weights being 1. Each grid cell position corresponds to a fusion feature unit at three resolutions. Each fusion feature unit is a two-dimensional feature vector containing both semantic and geometric feature values. The original resolution fusion feature unit contains semantic feature values. t1 and geometric eigenvalues h1 Half-resolution fused feature units contain semantic feature values t2 and geometric eigenvalues h2 A quarter-resolution fused feature unit contains semantic feature values. t3 and geometric eigenvalues h3 All eigenvalues ​​are standardized values, maintaining complete consistency with the eigenvalue terminology in the weighted summation formula above. The weighted summation calculation is still divided into two parts: semantic eigenvalue weighted summation and geometric eigenvalue weighted summation. Following the weighted summation formula above, the semantic eigenvalue weighted summation is calculated using the following formula: ;in, This represents the fused semantic feature value of the grid position after dynamic weight adjustment. The contribution weights are dynamically determined for the current grid position, and their sum is 1. These are the semantic feature values ​​from the three resolution fused feature units, consistent with the previous text; the geometric feature values ​​are weighted and summed, calculated using the following formula: Where G represents the fused geometric feature value of the grid position after dynamic weight adjustment. The weights are exactly the same as those in the weighted summation formula of the semantic feature values ​​mentioned above, and are the contribution weights that are dynamically determined for the current grid position. These are the geometric feature values ​​from the three resolution fused feature units, consistent with the previous text. After weighted summation, the fused semantic feature values ​​are... The fusion geometric feature value G is concatenated to form a two-dimensional fusion feature vector, which is the fusion feature value of the grid position. Compared with the fusion feature value in the previous fusion environment representation, this fusion feature value is calculated based on dynamic weights and can adapt to the complexity of local terrain. Each grid position completes the weighted summation calculation according to the above process to generate the corresponding fusion feature value.

[0032] The fused feature values ​​obtained after weighted integration of all grid locations are aggregated to construct a comprehensive environmental situation map containing terrain category and geometric relief. Specifically, this involves: extracting the fused feature values ​​obtained after dynamic weighted summation of all grid locations; arranging all fused feature values ​​sequentially according to their spatial position in the top-view grid coordinate system, following a left-to-right and top-to-bottom order, strictly corresponding to the original grid position of each fused feature value during the arrangement process to ensure that the arranged fused feature values ​​are completely consistent with the grid positions in the top-view grid coordinate system, without misalignment, omissions, or duplications, maintaining consistency with the aggregation logic of the fused feature map mentioned earlier; after arrangement, a two-dimensional feature matrix is ​​formed, which is the core structure of the comprehensive environmental situation map. The number of rows and columns of the matrix is ​​completely consistent with the number of rows and columns of the original resolution fused feature map and the fused environmental representation. Each grid cell in the comprehensive environmental situation map corresponds to a fused feature value, among which the fused semantic feature value... Corresponding to specific terrain categories, through The range of values ​​can directly determine the terrain material type of the grid location; the fused geometric feature value G corresponds to the terrain elevation of the corresponding grid location, and the magnitude of the G value can directly determine the terrain undulation of the grid location. The combination of the two achieves an integrated representation of terrain category and geometric undulation information. To improve the practicality and readability of the comprehensive environmental situation map, it needs to be standardized to ensure that the fused semantic feature values ​​of all grid units are consistent. Both the fused geometric feature value G and the fused feature value G are kept between 0 and 1. If there are values ​​outside this range due to calculation errors, normalization adjustment is required to correct the values ​​to between 0 and 1, ensuring the consistency and comparability of the feature values. At the same time, the effectiveness of the integrated environmental situation map is verified, checking whether the fused feature value of each grid position is reasonable, whether it is suitable for its local terrain complexity, whether the fused feature value corresponds correctly with the grid position, and whether the spatial position is consistent with the original top-view grid coordinate system, ensuring no misalignment, no error, and no redundant data. After verification, the integrated environmental situation map is visualized and optimized, and the fused semantic feature value is used. The corresponding terrain categories are marked with different colors, and the degree of terrain undulation corresponding to the fused geometric feature value G is distinguished by the color depth. The darker the color, the more severe the terrain undulation, and the lighter the color, the flatter the terrain. The final comprehensive environmental situation map is constructed.

[0033] In this embodiment of the invention, by constructing a multi-resolution fusion feature map and using the quasi-Newton method to iteratively optimize the optimal weight of each grid position, the defect of traditional fixed weights being unable to adapt to different terrain features is overcome, and accurate fusion of multi-scale features is achieved. Based on the fusion environment representation, the local terrain complexity index is calculated, and the contribution weight of each resolution feature is dynamically adjusted, overcoming the limitation that a single resolution cannot take into account both details and the global situation, and generating a comprehensive environmental situation map that can simultaneously express terrain category and geometric undulation.

[0034] In a preferred embodiment of the present invention, step 4 above includes: Based on the comprehensive environmental situation map, situation maps of the current moment and multiple previous moments are continuously collected along the time dimension to construct a comprehensive environmental situation map sequence containing time-series information. Specifically, this includes: using the robot's current operating moment as the baseline moment, setting the length of the time-series sequence, including the current moment and a fixed number of previous historical moments in the collection range to ensure that the time-series sequence can fully reflect the continuous change of the terrain over a period of time; extracting the comprehensive environmental situation map corresponding to each moment in chronological order, ensuring that each comprehensive environmental situation map maintains the same top-down grid coordinate system, the same grid size, and the same spatial coverage, guaranteeing that the situation maps at different moments are completely consistent in spatial structure; arranging the extracted comprehensive environmental situation maps in chronological order from early to late to form an ordered time-series data set, which is the comprehensive environmental situation map sequence containing time-series information. Each element in the sequence corresponds to a comprehensive environmental situation map at a fixed moment, and the time interval between adjacent elements is consistent with the collection frequency of the robot's perception data to ensure that the time-series information is continuous and evenly distributed.

[0035] For situation maps at two adjacent moments in a time series, using the same spatial grid position as the comparison unit, semantic category information and geometric elevation information at that position are extracted respectively, and the changes in semantic category and geometric elevation are calculated. Specifically, this includes: accurately selecting a set of comprehensive environmental situation maps at two adjacent moments from the time series comprehensive environmental situation map sequence, clearly distinguishing between the reference situation map and the comparison situation map, with the comprehensive environmental situation map at the previous moment as the reference benchmark and the comprehensive environmental situation map at the next moment as the comparison object, ensuring that the situation maps at the two moments are continuous in time and consistent in space, both using the same set of top-down grid coordinate system, the same grid size, and the same spatial coverage, initiating the grid position traversal process, and traversing all corresponding spatial grid positions in the two situation maps one by one in the order from left to right and from top to bottom, ensuring that each grid position can be compared one by one without omission, duplication, or misalignment. For each grid location being compared, semantic category information and geometric elevation information are extracted: From the reference situation map of the previous moment, the fused semantic feature value and fused geometric feature value corresponding to the grid location are extracted. Both feature values ​​are calculated using the weighted summation formula mentioned above, and both feature values ​​have been standardized. Then, from the comparison situation map of the next moment, the fused semantic feature value and fused geometric feature value corresponding to the same grid location are extracted. These two feature values ​​are also calculated using the two weighted summation formulas mentioned above, and the calculation logic and standardization standards are completely consistent with the corresponding feature values ​​in the reference situation map, ensuring the comparability of feature values ​​between the two moments. After the feature value extraction is completed, the change in semantic category and the change in geometric elevation are calculated: For semantic category information, the fused semantic feature value of the grid location at the next moment is compared with the fused semantic feature value of the same grid location at the previous moment, and the absolute difference between the two feature values ​​is calculated. This absolute difference is the semantic category change of the current grid location. The calculation of changes in geometric elevation is strictly performed, retaining four decimal places to ensure accuracy. For geometric elevation information, the same calculation logic as for semantic category changes is used. The fused geometric feature value of the grid position at the next moment is compared with the fused geometric feature value of the same grid position at the previous moment, and the absolute difference between the two feature values ​​is calculated. This absolute difference is the geometric elevation change of the current grid position. Following this process, the semantic category change and geometric elevation change calculations for all grid positions are completed sequentially. After each grid position is calculated, the two changes for that grid position are associated with the corresponding grid row and column numbers. After all grid positions have been calculated, the semantic category changes of all grid positions are arranged sequentially according to their spatial location, forming a semantic change distribution covering the entire detection range. Simultaneously, the geometric elevation changes of all grid positions are arranged sequentially according to their spatial location, forming a geometric elevation change distribution covering the entire detection range. Both change distributions are consistent with the grid structure of the integrated environmental situation map.

[0036] Based on the semantic and geometric elevation changes of each raster location between adjacent time points, the dynamic evolution characteristics of the local terrain are analyzed, identifying areas where terrain type is changing and areas where elevation is changing. The analysis results for each raster location are obtained, specifically including: setting judgment thresholds for semantic and geometric elevation changes. These thresholds are set based on the standardized range of 0 to 1 for the fused semantic and geometric feature values, as well as the calculation characteristics of semantic category changes and geometric elevation changes; both are absolute differences between 0 and 1. Specific values ​​were determined through multiple field tests. The semantic change judgment threshold was set to 0.1, and the geometric elevation change judgment threshold was set to 0.08. After setting these thresholds, the calculation results for each raster location are compared with the corresponding thresholds. A comparison is made to ensure consistent judgment logic and standards. When the semantic category change of a certain grid location is greater than 0.1, the grid location is determined to have a terrain type conversion trend and is marked as a terrain type conversion grid. When the geometric elevation change of a certain grid location is greater than 0.08, the grid location is determined to have a terrain undulation change trend and is marked as an elevation change grid. For grids that simultaneously meet the conditions of semantic category change greater than 0.1 and geometric elevation change greater than 0.08, they are marked as comprehensive change grids. During the marking process, a unique label is assigned to each change type. The marking results of all grid locations are arranged from left to right and from top to bottom according to their spatial position in the top-view grid coordinate system to form a grid-level analysis result of the local terrain dynamic evolution characteristics.

[0037] Based on the analysis results of all grid positions, the dynamic evolution trend of the entire terrain scene in the time dimension is obtained. Specifically, this includes: using the robot's current position as a reference, delineating the range of grid areas that have not yet been entered. The specific delineation method is as follows: spatially aggregate the grid-level analysis results, statistically analyze the clustering degree of terrain type conversion grids and elevation change grids in adjacent areas, divide the spatially continuously distributed change grids into independent change regions, and for each change region, combine the change amount records of earlier moments in the time series to determine the persistence and direction of the change trend of the region, determine the expansion direction, expansion speed, and change intensity of the change region, and determine the grid areas with no obvious changes as terrain stable regions. By combining the evolution direction and change intensity of all change regions with the distribution range of stable regions, the overall dynamic evolution trend of the entire terrain scene in the time dimension is formed, clarifying which regions in the terrain scene are continuously stable, which regions change gradually, and which regions change rapidly.

[0038] Based on the dynamic evolution trend of the terrain, for the grid area that the robot has not yet entered in front of its current position, the semantic category and geometric elevation of the grid area in the future are predicted according to the evolution pattern of semantic and geometric elevation changes between adjacent time points, resulting in a terrain state prediction result for the area in front. Specifically, this includes: obtaining the robot's precise grid position in the top-down grid coordinate system at the current time through the robot's own localization module, and determining the grid row and column number corresponding to this position; setting the forward detection distance, and combining the grid size of the comprehensive environmental situation map with the previous information to delineate the grid area that has not yet been entered in front. The delineated area is a rectangular grid area extending forward a fixed number of rows and left and right a fixed number of columns from the current grid position of the robot, while clarifying the range of grid row and column numbers in the delineated area. After delineation, the row and column numbers of all grids in the area are recorded to form a list of grids in the area that has not yet been entered in front, and the semantic changes of all grids in the area are extracted in the historical time sequence. After extracting the semantic and geometric elevation changes, an independent change record file is created for each raster. This file clearly records the semantic and geometric elevation changes of the raster at each adjacent time point, along with the corresponding time. Based on the determined overall dynamic evolution trend of the terrain scene, the change pattern of each raster position is fitted. The specific operations are as follows: For each raster's change record file, the semantic and geometric elevation change data for that raster at all adjacent time points are compiled. Trend fitting is performed on both types of change data. During the fitting process, time is used as the horizontal axis and change amount as the vertical axis, employing a linear fitting method to fit the change curves of the raster's semantic and geometric elevation changes over time. The magnitude of the raster's change over time is determined through the fitted curves, specifically by calculating the slope of the fitted curve and the goodness of fit, ensuring that the goodness of fit is not less than 0.8. If the goodness of fit is less than 0...8. Then, the fitting method is readjusted, and a second fitting is used to supplement and correct until the fitting degree meets the requirements, ensuring that the fitted change pattern can accurately reflect the historical change trend of the raster. Based on the fitted change pattern, the fused semantic feature value and fused geometric feature value at a fixed future time are calculated. The specific calculation process is as follows: Set the future prediction time, which is a fixed duration after the current time. This duration is consistent with the time interval of adjacent times to ensure the continuity and rationality of the prediction. For the calculation of the fused semantic feature value, first extract the fused semantic feature value of the raster at the current time and calculate it using the weighted summation formula mentioned above. Combined with the semantic change change pattern obtained from the fitting, calculate the cumulative value of semantic change from the current time to the future prediction time. Add the fused semantic feature value at the current time to the cumulative value of semantic change to obtain the predicted fused semantic feature value of the raster at the future prediction time. For the calculation of the fused geometric feature value, the same calculation logic as the fused semantic feature value is used: extract the fused geometric feature value of the raster at the current time and calculate it using the weighted summation formula mentioned above. The weighted summation formula above calculates, combined with the fitted geometric elevation change pattern, the cumulative value of geometric elevation change from the current time to the future prediction time is calculated. The fused geometric feature value at the current time is added to the cumulative value of geometric elevation change to obtain the predicted fused geometric feature value of the grid at the future prediction time. For each grid whose prediction value is calculated, the predicted fused semantic feature value, predicted fused geometric feature value, and corresponding grid row and column number are associated and recorded. After the prediction values ​​of all grids in the unexplored area are calculated, the prediction results of all grids are arranged according to their spatial position in the top-view grid coordinate system, strictly following the order from left to right and from top to bottom. During the arrangement, the row and column numbers of each grid are strictly matched to ensure that the arranged prediction results are completely consistent with the grid structure and spatial position of the comprehensive environmental situation map, including the number of grid rows, columns, grid size, and top-view grid coordinate system. After arrangement, a complete two-dimensional feature matrix is ​​formed, which is the predicted terrain state result of the area ahead.

[0039] The terrain condition prediction results for the area ahead are spatiotemporally aligned with the current integrated environmental situation map. Position matching is achieved by combining the predicted information with historical measured information within the same top-view grid space. Specifically, this involves: maintaining the same top-view grid coordinate system for both the terrain condition prediction results and the current integrated environmental situation map, ensuring complete consistency in grid row and column numbers, spatial dimensions, and physical locations; mapping each grid cell in the prediction results to its corresponding spatial location on the current situation map according to its row and column number, ensuring a one-to-one correspondence between the predicted and measured information within the same spatial grid; directly performing grid position matching for overlapping areas between the predicted and measured regions; and extending the predicted area outwards at the same grid scale for areas exceeding the current measured range, maintaining continuous and unified spatial coordinates. After completing the position matching of all grid cells, precise spatial alignment between the predicted and measured information is achieved.

[0040] Based on the predicted information after location matching and historical measured information, confidence weights are assigned to the same grid location. Weighted fusion is then used to calculate the predicted terrain state of the foreground area with the current and historical integrated environmental situation maps, resulting in an enhanced spatiotemporal situation map containing both historical measured terrain information and future predicted terrain information. Specifically, for each grid location, corresponding confidence weights are assigned to both the measured and predicted information. The confidence weight for the measured information is primarily determined by the sustained stability of the historical measured data for that grid location. The historical measured data originates from the fusion feature values ​​of all historical and current times in the previously constructed temporal integrated environmental situation map sequence, all calculated using the aforementioned formula. First, the fluctuations of the fused semantic and geometric feature values ​​of the grid location across all historical and current times are statistically analyzed. The standard deviation of the feature values ​​is then calculated. A smaller standard deviation indicates more stable historical measured data for the grid location, resulting in a higher confidence weight for the measured information. Conversely, a larger standard deviation indicates greater fluctuations in historical measured data, resulting in a lower confidence weight for the measured information. The confidence weight for the measured information is set between 0.5 and 0.9. Specifically, when the standard deviation of the feature value is ≤0.05, the historical measured data is considered continuously stable, and a confidence weight of 0.9 is assigned. When 0.5 < standard deviation of the feature value ≤0.1, the historical measured data is considered basically stable, and a confidence weight of 0.9 is assigned. The weight is 0.7; when the standard deviation of the eigenvalue is >0.1, it is determined that the historical measured data fluctuates greatly, and the confidence weight of the measured information is assigned 0.5. For the confidence weight of the predicted information, the core is to assign a value based on the clarity of the grid position change pattern and the consistency of the evolution trend. The change pattern comes from the fitting curve of the semantic change and geometric elevation change obtained in the previous fitting. The consistency of the evolution trend comes from the degree of matching between the change trend of the grid position and the dynamic evolution trend of the overall terrain scene. The fitting degree of the fitting curve is judged. When the fitting degree is ≥0.9, the change pattern is judged to be clear. When the fitting degree is 0.8≤fit degree<0.9, the change pattern is judged to be relatively clear. When the fitting degree is <0.8, the change pattern is judged to be ambiguous. Then the grid position is judged to be clear. The consistency between the changing trend of grid positions and the overall terrain evolution trend is considered. If they are completely consistent, the consistency is considered high; if they are basically consistent, the consistency is considered medium; and if they are inconsistent, the consistency is considered low. The confidence weight of the prediction information is set from 0.1 to 0.5, and is assigned a comprehensive value based on the clarity of the change pattern and the consistency of the evolution trend: when the change pattern is clear and the consistency of the evolution trend is high, the confidence weight of the prediction information is 0.5; when the change pattern is clear but the consistency of the evolution trend is medium, or when the change pattern is relatively clear and the consistency of the evolution trend is high, the confidence weight of the prediction information is 0.3; and when the change pattern is ambiguous or the consistency of the evolution trend is low, the confidence weight of the prediction information is 0.1. Simultaneously ensure that the sum of the confidence weights of the measured information and the predicted information at the same grid location is 1. If there is a slight deviation due to decimal calculation, normalization adjustment is required to ensure that the weight allocation meets the constraints. After the weights are assigned, for the same grid location, the semantic feature value and geometric feature value are calculated separately. The calculation process strictly follows the weighted summation logic and is consistent with the calculation logic of the fused feature value mentioned above. The specific calculation process is as follows: For the weighted fusion of semantic feature values, first extract the measured fused semantic feature value of the grid location, which is calculated by the above formula. The measured fused semantic feature value is obtained from the prediction result of the terrain state of the area ahead. Multiply the measured fused semantic feature value by the corresponding measured information confidence weight to obtain the measured semantic weight value; multiply the predicted fused semantic feature value by the corresponding predicted information confidence weight to obtain the predicted semantic weight value; add the measured semantic weight value and the predicted semantic weight value to obtain the fused semantic feature value of the grid location. For the weighted fusion of geometric feature values, the calculation method is completely consistent with the weighted fusion of semantic feature values. The calculation logic involves first extracting the measured fused geometric feature value of the grid position using the formula above. Then, the measured fused geometric feature value is multiplied by the corresponding measured information confidence weight to obtain the measured geometric weighted value. Next, the predicted fused geometric feature value is multiplied by the corresponding predicted information confidence weight to obtain the predicted geometric weighted value. Finally, the measured geometric weighted value and the predicted geometric weighted value are added to obtain the fused geometric feature value for the grid position. After the weighted calculation of a single grid position is completed, the fused semantic feature value and the fused geometric feature value of that grid position are compared with the... The corresponding raster row and column numbers are associated and recorded. Then, the raster traversal process is started. Following the order from left to right and from top to bottom, all raster positions are traversed one by one. The above weight assignment and weighted calculation process is repeated to ensure that the semantic feature value and geometric feature value of each raster position can be weighted and fused without omission or duplicate calculation. After the weighted calculation of all raster positions is completed, the semantic feature value and geometric feature value of all raster positions are fused and strictly follow the overall environmental situation according to their spatial position in the top-view raster coordinate system. Figure 1 The arrangement follows a consistent rule, proceeding from left to right and from top to bottom. During the arrangement process, the row and column numbers of each grid are strictly matched to ensure that the fused feature values ​​after arrangement correspond completely with the spatial position of the grid. Furthermore, the number of grid rows, columns, and grid size are all consistent with the comprehensive environmental situation map. After the arrangement is completed, a complete two-dimensional feature matrix is ​​formed, which is the enhanced spatiotemporal situation map.

[0041] In this embodiment of the invention, a time-series sequence is constructed by collecting multi-moment comprehensive environmental situation maps along the time dimension, overcoming the deficiency of single-frame static perception in being unable to capture dynamic changes; terrain type conversion areas and elevation change areas are identified based on the amount of change, achieving accurate positioning and classification of change areas; the terrain state of areas not yet entered ahead is predicted by combining the overall dynamic evolution trend, overcoming the limitation that current perception cannot predict the terrain ahead, and providing advanced environmental cognition capabilities; by spatiotemporally aligning and weighting the prediction results with measured information, an enhanced spatiotemporal situation map is generated, realizing an integrated representation of historical information, current state, and future prediction, significantly improving the quadruped robot's environmental perception and navigation prediction capabilities in dynamic and complex terrain.

[0042] In a preferred embodiment of the present invention, step 5 above includes: The robot's current pose and velocity commands are acquired, and the augmented spatiotemporal situation map is combined with the robot's current pose and velocity commands to form a joint state space describing the robot's current environmental state and its own motion state. Specifically, this includes: real-time acquisition of the robot's pose information using its onboard positioning module and inertial measurement unit. This pose information includes the robot's spatial coordinates in the top-view grid coordinate system, its pitch angle, roll angle, and yaw angle. After acquisition, this information is standardized, mapping all values ​​to a range of zero to one to ensure consistency with the value range of the feature values ​​in the augmented spatiotemporal situation map. Simultaneously, from the robot's upper layer... The motion control module acquires the current velocity command, which includes the linear velocity command for the robot's forward direction and the angular velocity command for the turning direction. The velocity command is also standardized to ensure that the numerical range is consistent with the pose information and environmental feature information. After information acquisition and standardization, the enhanced spatiotemporal situation map is expanded into a one-dimensional feature sequence in the form of a two-dimensional feature matrix. The standardized pose information and the standardized velocity command are then concatenated to the end of the one-dimensional feature sequence to form a continuous and dimensionally unified feature combination. This feature combination fully includes the spatiotemporal features of the environment, the robot's own pose state, and motion command information. After being arranged in a fixed dimensional order, it constitutes the joint state space used for navigation decision-making.

[0043] The joint state space is input into the navigation policy network, which is pre-trained using a reinforcement learning framework. Based on the received joint state space, it performs forward inference and outputs the target motion angles for each joint of the robot. Specifically, this involves: dimensionality normalization of the constructed joint state space. First, the preset dimension parameters of the navigation policy network's input layer are read to clarify the core parameters such as feature dimension and number of channels. Then, the one-dimensional feature sequence of the joint state space is adjusted according to the network input requirements to ensure its size perfectly matches the dimension of the navigation policy network's input layer. If the dimension of the joint state space matches the dimension of the input layer... Dimension mismatch is addressed using zero-padding or feature truncation. Zero-padding sets the padding value to 0, while feature truncation prioritizes preserving spatiotemporal features to ensure no core features are lost. After dimensionality normalization, the normalized joint state space is input to the input layer of the navigation policy network via a data interface. During input, the data transmission rate is synchronized with the network inference rate. The navigation policy network consists of multiple convolutional layers, batch normalization layers, activation layers, and fully connected layers stacked sequentially. The convolutional layers each contain three convolutional layers with a kernel size of 3×3 and a stride of 1. Padding is performed using… samePadding ensures that the feature map size is not reduced after convolution; batch normalization layer is used to normalize the features after convolution; activation layer uses ReLU activation function to introduce non-linear features; the fully connected structure contains 2 fully connected layers, the output dimension of the first fully connected layer is set to 128, and the output dimension of the second fully connected layer is consistent with the number of robot joints. Before being deployed in practical applications, the network underwent comprehensive training in diverse terrain scenarios using a reinforcement learning framework. The training scenarios covered various complex terrain types, including flat terrain, undulating terrain, semantically ambiguous terrain, and dynamically changing terrain. During training, terrain adaptability, gait stability, and travel efficiency were the core optimization objectives. A reward and punishment mechanism was implemented: positive rewards were given when the robot smoothly traversed complex terrain areas, while negative penalties were imposed when the robot experienced gait instability, trajectory deviation, or inability to pass. Through multiple rounds of iterative training, parameters such as convolutional kernel parameters and fully connected layer weights were continuously updated, enabling the network to output reasonable joint movements based on environmental conditions and motion commands. After training, the network was tested to ensure that its inference accuracy was no less than 95%, meeting practical navigation requirements. After receiving the joint state space, the network sequentially executes the forward computation process of each layer: the first convolutional layer performs convolution operations on the input joint state space to extract terrain features from the enhanced spatiotemporal situation map and filter redundant information; through batch processing... The normalization layer normalizes the convolutional features, and then performs a non-linear transformation through a ReLU activation layer. This convolution, batch normalization, and activation process is repeated to complete feature extraction across three convolutional layers, yielding a high-level terrain feature vector. The extracted terrain feature vector is then input into a fully connected layer. The first fully connected layer fuses the terrain features, robot pose information, and velocity command information, converting them into a high-dimensional feature vector. The second fully connected layer maps this high-dimensional feature vector to an output vector corresponding to the number of robot joints. In the final stage of forward inference, the network output layer generates a corresponding number of consecutive values ​​according to the robot joint numbers. Each value corresponds to a target motion angle for one robot joint. All target motion angles are strictly controlled within the physical motion limits of the robot joints. The motion angle range for each joint is pre-set based on the robot's mechanical structure parameters to ensure that the target motion angle does not exceed the joint's maximum range of motion. After output, the values ​​are generated according to the robot joint numbers from 1 to... n The target motion angles are organized in sequence to form an ordered sequence of target joint angles. Each element in the sequence corresponds one-to-one with a joint number, and a joint number identifier is added to provide clear and identifiable motion commands for the underlying control unit.

[0044] The target motion angles of each joint are input to the underlying PID controller, which then converts them into the driving torque corresponding to each joint motor. Specifically, this involves: synchronously sending the target joint angle sequence output from the navigation strategy network to the robot's underlying embedded control unit via the communication interface between the navigation strategy network and the underlying embedded control unit. During transmission, a CRC check is used. After receiving the target joint angle sequence, the underlying embedded control unit first performs a CRC check on the received data. Once the data is confirmed to be correct, it begins parsing the joint number and corresponding target angle in the sequence, and then processes the data according to the joint number from 1 to... nThe target motion angle for each joint is sent to its corresponding independent PID controller in sequence. Each joint is equipped with an independent PID controller. After receiving the target motion angle, the independent PID controller of each joint immediately starts the workflow, reading the current actual joint angle of the corresponding joint motor in real time through the motor encoder. The sampling frequency of the motor encoder is set to 20Hz. After sampling, the actual joint angle is first converted to a unit consistent with the target motion angle, and then the difference between the target motion angle and the current actual joint angle is calculated to obtain the joint angle deviation value. The deviation value is calculated by subtracting the current actual joint angle from the target motion angle. Angle: If the deviation value is positive, it means the actual joint angle is less than the target angle, and the driving torque needs to be increased to move the joint towards the target angle; if the deviation value is negative, it means the actual joint angle is greater than the target angle, and the driving torque needs to be decreased to adjust the joint to the target angle. After obtaining the joint angle deviation value, the PID controller sequentially performs proportional, integral, and derivative operations based on this deviation value. The proportional, integral, and derivative operation coefficients are all preset according to the mechanical characteristics of the robot joint, specifically set to a proportional operation coefficient of 5.0, an integral operation coefficient of 0.1, and a derivative operation coefficient of 0.5. This parameter setting can balance the controller's response speed and control precision. After completing the proportional, integral, and derivative operations, the results of these three operations are weighted and summed according to preset weights. The weight of the proportional operation result is 0.6, the weight of the integral operation result is 0.2, and the weight of the derivative operation result is 0.2. Through this weight allocation, a preliminary control output is obtained. The preliminary control output needs to be further converted into a driving torque that the motor can recognize. Based on parameters such as the reduction ratio and transmission efficiency of the robot's joint transmission structure, as well as the rated output range of the motor, the preliminary control output needs to be mapped and converted. The unit of the converted driving torque value is uniformly Newton-meter. During the conversion process, two constraint processing needs to be performed simultaneously: one is to perform amplitude limiting processing on the driving torque. First, the torque value is strictly controlled within the maximum allowable output threshold range of the motor, and the maximum allowable output torque of the motor is set according to the rated parameters of the motor. Second, the torque change rate is constrained, and the maximum torque change rate is set to 0.5 N·m / ms. Since each joint is equipped with an independent PID controller, the PID controller of each joint will independently complete the above deviation calculation, PID calculation and torque conversion process, without being affected by the control process of other joints. After the driving torque command of all joints is calculated, the underlying embedded control unit performs synchronous calibration on all torque commands, generates driving torque commands corresponding to each joint, and waits to send them to the drive motors of each joint.

[0045] By driving torque to propel each joint of the robot to perform corresponding movements, the robot can autonomously select landing points and adjust its motion posture based on terrain understanding and prediction information contained in the enhanced spatiotemporal situation map, thus completing navigation tasks in complex terrains. Specifically, the underlying embedded control unit synchronously sends the corresponding driving torque commands for each joint to the joint drive motors. After receiving the torque commands, the motors output the corresponding torque according to the command magnitude, driving the links and joint structures to perform rotational movements. During joint movements, the robot dynamically adjusts its body posture and landing point position in real time based on historical measured terrain information and future predicted terrain information contained in the enhanced spatiotemporal situation map. In areas with significant elevation changes, the robot automatically lowers its center of gravity and reduces its stride. In areas with complex terrain semantics, it automatically avoids unstable terrain. In areas where terrain changes are predicted, it adjusts its gait and posture in advance. Throughout the journey, the robot continuously executes state input, network inference, torque output, and joint movement processes based on the updated results of the joint state space. It constantly corrects its trajectory and joint posture, autonomously completing foothold selection, gait switching, obstacle avoidance, and posture stabilization control. The robot gradually traverses complex terrain areas while continuously moving, advancing along the preset direction until it completes the autonomous navigation task in the overall complex terrain environment.

[0046] In this embodiment of the invention, by constructing an enhanced spatiotemporal situation map that integrates historical measured information and future prediction information, the robot possesses the ability to perceive dynamic terrain in advance, effectively improving navigation safety and predictability. Unifying environmental information, robot pose, and motion commands into a joint state space provides a complete and standardized input basis for navigation decisions. The navigation strategy network trained based on reinforcement learning can adapt to complex terrain changes and output stable and reasonable joint motion commands. Precise torque output is achieved through a PID controller, ensuring smooth and reliable joint motion. The overall solution enables the robot to autonomously select safe landing points, adjust its motion posture in real time, and avoid potential danger zones in advance in complex dynamic terrain, significantly improving the stability, adaptability, and passage efficiency of autonomous navigation.

[0047] like Figure 2 As shown, embodiments of the present invention also provide an autonomous navigation system for quadruped robots in complex terrain based on reinforcement learning, comprising: The acquisition module is used to acquire multimodal perception data of the robot when it walks in complex terrain environment. After processing, it generates terrain semantic map and terrain elevation map, and extracts geometric feature map from terrain elevation map and semantic feature map from terrain semantic map. The fusion module is used to project semantic feature maps and geometric feature maps onto the same top-down grid space centered on the robot, construct multiple grid layers with different resolutions, and associate and combine semantic features and geometric features at the same resolution for each grid position to form a fusion feature unit at that resolution. The calculation module is used to calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic features in the neighborhood of each grid location. Based on this complexity, the contribution weights of fused feature units at different resolutions are dynamically adjusted, and a comprehensive environmental situation map is generated after weighted integration. The analysis module is used to acquire comprehensive environmental situation maps at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, it analyzes the dynamic evolution trend of the terrain and predicts the terrain status of the area ahead, thereby forming an enhanced spatiotemporal situation map. The control module is used to acquire the robot's current pose and velocity commands. The pose and velocity commands, together with the enhanced spatiotemporal situation map, constitute the state space. The state space is used to acquire the joint target angles through a reinforcement learning policy network and is converted into driving torques by the underlying controller to achieve autonomous navigation.

[0048] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0049] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0050] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0051] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning, characterized in that, The method includes: Step 1: Acquire multimodal perception data of the robot walking in complex terrain environment, generate terrain semantic map and terrain elevation map after processing, extract geometric feature map from terrain elevation map and semantic feature map from terrain semantic map; Step 2: Project the semantic feature map and geometric feature map onto the same top-down grid space centered on the robot to construct multiple grid layers with different resolutions. For each grid position, associate and combine the semantic features and geometric features at the same resolution to form a fused feature unit at that resolution. Step 3: Calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution in the neighborhood of each grid location. Based on this complexity, dynamically adjust the contribution weight of fused feature units at different resolutions, and generate a comprehensive environmental situation map after weighted integration. Step 4: Obtain comprehensive environmental situation map at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, analyze the dynamic evolution trend of the terrain and predict the terrain status of the area ahead to form an enhanced spatiotemporal situation map. Step 5: Obtain the robot's current pose and velocity commands. Combine the pose and velocity commands with the enhanced spatiotemporal situation map to form a state space. The state space is used to obtain the joint target angles through a reinforcement learning policy network and converted into driving torques by the underlying controller to achieve autonomous navigation.

2. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 1, characterized in that, Multimodal perception data of the robot navigating in complex terrain environments is acquired, processed, and used to generate terrain semantic maps and terrain elevation maps. Geometric feature maps are extracted from the terrain elevation maps, and semantic feature maps are extracted from the terrain semantic maps, including: Acquire multimodal perception data of the robot when it walks in complex terrain environment, wherein the multimodal perception data includes at least environmental images, laser ranging data, inertial measurement data and joint state data; The environmental image is analyzed at the pixel level to obtain a terrain semantic map and extract semantic feature maps from it; the laser ranging data is spatially interpolated and reconstructed to generate a terrain elevation map centered on the robot and extract geometric feature maps reflecting the degree of terrain undulation from it.

3. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 2, characterized in that, The semantic and geometric feature maps are projected onto the same top-down grid space centered on the robot, constructing multiple grid layers of different resolutions. For each grid location, the semantic and geometric features at the same resolution are associated and combined to form a fused feature unit at that resolution, including: Using terrain semantic map and terrain elevation map as processing objects, coordinate transformation is used to project them onto the same top-view grid coordinate system centered on the robot's current position, resulting in a spatially aligned top-view semantic grid map and top-view elevation grid map. Based on the top-view semantic raster map and the top-view elevation raster map, multi-scale downsampling processing is performed to construct multiple raster layers including the original resolution, 1 / 2 resolution and 1 / 4 resolution, so that each resolution layer contains a corresponding semantic feature sub-map and a geometric feature sub-map. For each raster cell in each resolution layer, the semantic feature value and geometric feature value corresponding to the raster cell are associated and combined to generate a fused feature cell at the resolution.

4. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 3, characterized in that, The local terrain complexity is calculated based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution within the neighborhood of each grid location. The contribution weights of fused feature units at different resolutions are dynamically adjusted according to this complexity, and a weighted integration is performed to generate a comprehensive environmental situation map, including: Based on the fusion feature units constructed at the original resolution, 1 / 2 resolution, and 1 / 4 resolution, the fusion feature units at all grid positions in each resolution layer are aggregated to form a fusion feature map corresponding to the resolution. The fused feature maps are concatenated along the feature channel direction to construct a concatenated feature tensor containing multi-scale information. For each grid position in the cascaded feature tensor, an optimization algorithm based on the quasi-Newton method is introduced to determine the optimal weight parameters for the fusion of features of different resolutions at that position through iterative calculation. Based on the optimal weight parameters, the fusion feature units at different resolutions at the same grid position are weighted and summed to generate fusion feature values ​​of multi-scale semantic and geometric information. By traversing all grid positions, the fusion environment representation is obtained.

5. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 4, characterized in that, The local terrain complexity is calculated based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution within the neighborhood of each grid location. The contribution weights of fused feature units at different resolutions are dynamically adjusted according to this complexity. After weighted integration, a comprehensive environmental situation map is generated, which also includes: Based on the obtained fused environment representation, for each grid location, a complexity index representing the local terrain complexity is calculated according to the rate of change of geometric features and the degree of semantic category mixing within the neighborhood of the grid location. Based on the complexity index of each grid position, the contribution weight of the fused feature unit at the original resolution, 1 / 2 resolution and 1 / 4 resolution at that position is dynamically determined; Based on the dynamically determined contribution weights, the fusion feature units at different resolutions at each grid location are weighted and summed to obtain the fusion feature value that integrates semantic category and geometric undulation information. The fused feature values ​​obtained after weighted integration of all grid locations are aggregated to construct a comprehensive environmental situation map that includes terrain category and geometric undulation.

6. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 5, characterized in that, A comprehensive environmental situation map is acquired at multiple points in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent times, the dynamic evolution trend of the terrain is analyzed, and the terrain condition of the area ahead is predicted accordingly, forming an enhanced spatiotemporal situation map, including: Based on the comprehensive environmental situation map, situation maps of the current moment and multiple previous moments are continuously collected along the time dimension to construct a comprehensive environmental situation map sequence containing time-series information; For the situation map of two adjacent time points in the time series, the same spatial grid position is used as the comparison unit. The semantic category information and geometric elevation information at that position are extracted respectively, and the change in semantic category and the change in geometric elevation are calculated. Based on the semantic and geometric elevation changes of each grid position between adjacent time points, the dynamic evolution characteristics of the local terrain are analyzed, and areas where terrain type is changing and elevation is changing are identified, thus obtaining the analysis results of grid positions. Based on the analysis results of all grid positions, the dynamic evolution trend of the entire terrain scene in the time dimension is obtained.

7. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 6, characterized in that, This involves acquiring comprehensive environmental situation maps across multiple time points in a continuous time series, analyzing the dynamic evolution trend of the terrain by comparing semantic and geometric elevation changes at the same spatial location at adjacent times, and predicting the terrain condition of the area ahead to form an enhanced spatiotemporal situation map. It also includes: Based on the dynamic evolution trend of the terrain, for the grid area in front of the robot's current position that has not yet been entered, the semantic category and geometric elevation of the grid area in the future are predicted according to the evolution law of semantic change and geometric elevation change between adjacent time points, so as to obtain the terrain state prediction result of the area in front. The terrain condition prediction results of the area ahead are spatiotemporally aligned with the current comprehensive environmental situation map, and the predicted information and historical measured information are matched in the same overhead grid space to achieve position matching. Based on the predicted information after location matching and the historical measured information, confidence weights are assigned to the same grid location. Through weighted fusion, the predicted terrain status of the area in front is calculated with the comprehensive environmental situation map of the current and historical time, resulting in an enhanced spatiotemporal situation map that includes the measured terrain information of the historical time and the predicted terrain information of the future time.

8. The method for autonomous navigation of a quadruped robot in complex terrain based on reinforcement learning according to claim 7, characterized in that, The robot acquires its current pose and velocity commands, and combines these commands with an enhanced spatiotemporal situational map to form a state space. This state space is then processed by a reinforcement learning policy network to obtain joint target angles, which are converted into driving torques by the underlying controller to achieve autonomous navigation. This process includes: The robot's current pose and velocity commands are obtained, and the enhanced spatiotemporal situation map is combined with the robot's current pose information and velocity commands to form a joint state space that describes the robot's current environmental state and its own motion state. The joint state space is input into the navigation policy network, which is pre-trained using a reinforcement learning framework and performs forward reasoning based on the received joint state space to output the target motion angles of each joint of the robot. The target motion angle of each joint is input to the underlying PID controller, and the PID controller converts it into the driving torque corresponding to the motor of each joint; By driving the robot's joints with driving torque, the robot can perform corresponding movements, enabling it to autonomously select landing points and adjust its motion posture based on the terrain understanding and prediction information contained in the enhanced spatiotemporal situation map, thus completing navigation tasks in complex terrain.

9. A quadruped robot autonomous navigation system for complex terrain based on reinforcement learning, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, include: The acquisition module is used to acquire multimodal perception data of the robot when it walks in complex terrain environment. After processing, it generates terrain semantic map and terrain elevation map, and extracts geometric feature map from terrain elevation map and semantic feature map from terrain semantic map. The fusion module is used to project semantic feature maps and geometric feature maps onto the same top-down grid space centered on the robot, construct multiple grid layers with different resolutions, and associate and combine semantic features and geometric features at the same resolution for each grid position to form a fusion feature unit at that resolution. The calculation module is used to calculate the local terrain complexity based on the rate of change of geometric features and the degree of heterogeneity of semantic feature distribution in the neighborhood of each grid location. Based on this complexity, the contribution weights of fused feature units at different resolutions are dynamically adjusted, and a comprehensive environmental situation map is generated after weighted integration. The analysis module is used to acquire comprehensive environmental situation maps at multiple moments in a continuous time series. By comparing semantic and geometric elevation changes at the same spatial location at adjacent moments, it analyzes the dynamic evolution trend of the terrain and predicts the terrain status of the area ahead, thereby forming an enhanced spatiotemporal situation map. The control module is used to acquire the robot's current pose and velocity commands. The pose and velocity commands, together with the enhanced spatiotemporal situation map, constitute the state space. The state space is used to acquire the joint target angles through a reinforcement learning policy network and is converted into driving torques by the underlying controller to achieve autonomous navigation.

Images