A local terrain passability decision method and system for a humanoid robot

Abstract

The present application relates to the technical field of robot control, in particular to a kind of local terrain passability decision method and system of humanoid robot, comprising: the two-dimensional elevation grid graph of robot is constructed;According to the elevation value difference between each grid and adjacent grid in foot bottom coverage window, the distribution density of effective grid in foot bottom coverage window is combined to obtain local height continuity index;The grid coordinates and elevation value corresponding to each grid in foot bottom coverage window are carried out feature decomposition to obtain fitting feature vector, and the deviation between fitting feature vector and the direction of gravity of robot is obtained to obtain local surface inclination angle;Fall foot risk assessment value is obtained, and the position distribution of the center grid of foot bottom coverage window in two-dimensional elevation grid graph and fall foot risk assessment value are used to determine the passability decision result of robot.The present application realizes the safe, stable, adaptive fall foot closed-loop control of humanoid robot from perception to execution in full scene and full terrain.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a method and system for determining the local terrain accessibility of a humanoid robot. Background Technology

[0002] When humanoid robots conduct autonomous navigation and gait planning in unknown environments, they need to accurately assess the treadability of obstacles along their path to generate movement decisions such as stepping, crossing, or detouring, ensuring the stability of the robot's movement and avoiding safety risks such as slipping and falling. The spatial tilt angle of the landing area is a core geometric parameter that determines the robot's foot friction efficiency and stable support, and is directly related to the robot's safe landing boundary.

[0003] Existing methods rely on 3D point cloud data collected by depth sensing modules. By calculating statistical indicators such as the standard deviation of elevation and elevation dispersion within a local area, they quantify the degree of surface undulation and thus determine the feasibility of landing. However, this method only relies on the dispersion of elevation values ​​for safety assessment and cannot effectively distinguish between smooth slopes with continuous elevation changes and flat, level surfaces. For gently sloping terrain with minimal elevation undulations, the standard deviation of elevation often falls within a preset safety threshold range, leading the system to mistakenly classify it as a stable, footholdable surface. When the overall tilt angle of such terrain exceeds the frictional stability limit of the robot's feet, the robot is prone to lateral slippage after stepping on it, causing instability or even a fall. This severely restricts the operational safety and environmental adaptability of humanoid robots in complex terrains.

[0004] Therefore, existing decision-making methods based solely on elevation dispersion in a single dimension cannot effectively avoid the risks of instability caused by foot edge suspension, local stress concentration, and lateral slippage on smooth slopes, and are therefore insufficient to meet the application requirements of humanoid robots for autonomous walking in complex and unknown environments. Summary of the Invention

[0005] To address the problem that existing decision-making methods based solely on a single dimension of elevation dispersion have poor rationality, the present invention aims to provide a method and system for determining the local terrain drivability of a humanoid robot. The specific technical solution adopted is as follows: In a first aspect, the present invention provides a method for determining the local terrain drivability of a humanoid robot, comprising: Based on the three-dimensional point cloud data collected in real time during the humanoid robot's movement, a two-dimensional elevation grid map of the robot is constructed; wherein, the two-dimensional elevation grid map includes the elevation value and the number of projected point clouds in each grid. Based on the size parameters of the robot's foot, a foot coverage window is constructed on a two-dimensional elevation grid map. Based on the elevation value difference between each grid and its adjacent grids in the foot coverage window, combined with the distribution density of the effective grids in the foot coverage window, a local height coherence index is obtained. The grid coordinates and elevation values ​​corresponding to each grid in the foot cover window are decomposed to obtain the fitting feature vector of the local terrain of the foot cover window. Based on the deviation between the fitting feature vector and the opposite direction of the robot's gravity, the local surface tilt angle is obtained. By integrating the local height coherence index and the local surface tilt angle, the landing risk assessment value of the foot cover window is obtained. Based on the position distribution of the central grid of the foot cover window in the two-dimensional elevation grid map and the landing risk assessment value, the robot's accessibility decision result is determined.

[0006] Preferably, the step of obtaining a local height coherence index based on the elevation difference between each grid cell and its adjacent grid cells within the plantar coverage window, combined with the distribution density of effective grid cells within the plantar coverage window, specifically includes: The grids with a number of projected point clouds greater than zero within the foot cover window are considered valid grids. Based on the elevation difference between each effective grid cell and its adjacent effective grid cells in the plantar coverage window, the elevation transition weight between effective grid cells is analyzed, and an adjacency matrix is ​​constructed. Based on the sum of all elements in the same row of the adjacency matrix, a diagonal matrix is ​​constructed to obtain the connection strength matrix. The connectivity eigenvalues ​​are obtained by performing eigenvalue decomposition on the difference matrix between the connectivity strength matrix and the adjacency matrix. The connectivity eigenvalues ​​are the second smallest eigenvalues ​​in the eigenvalue decomposition results, arranged in ascending algebraic order. The local height coherence index is obtained based on the connectivity feature value and the proportion of effective grids in the plantar coverage window.

[0007] Preferably, the step of analyzing the elevation transition weights between effective grids and constructing an adjacency matrix based on the elevation differences between each effective grid and its adjacent effective grids within the plantar coverage window specifically includes: Obtain the nonlinear coefficient of the elevation value difference between each effective grid and its adjacent effective grid, and perform negative correlation processing on the nonlinear coefficient to obtain the elevation transition weight between each effective grid and its adjacent effective grid; Using each valid raster as a row and column, an adjacency matrix is ​​constructed based on the elevation transition weights between each valid raster and its adjacent valid raster.

[0008] Preferably, the step of obtaining the local height coherence index based on the connectivity feature value and the proportion of effective grids in the plantar coverage window specifically includes: The local height coherence index is determined by multiplying the connectivity feature value and the percentage of effective grids within the plantar coverage window.

[0009] Preferably, the step of performing feature decomposition on the grid coordinates and elevation values ​​corresponding to each grid cell within the plantar coverage window to obtain a fitted feature vector of the local terrain within the plantar coverage window specifically includes: Based on the grid coordinates and elevation values ​​of each grid, construct the three-dimensional coordinate values ​​of each grid. The 3D coordinate values ​​of all valid grids within the plantar coverage window are decomposed into features. The feature vector corresponding to the smallest feature value in the feature decomposition result is used as the fitting feature vector of the local terrain of the plantar coverage window.

[0010] Preferably, the step risk assessment value of the foot cover window obtained by integrating the local height coherence index and the local surface tilt angle specifically includes: The local surface tilt angle and local height consistency index are normalized respectively; Based on the normalized local surface tilt angle and the normalized local height coherence index, the landing risk assessment value of the plantar coverage window is determined.

[0011] Preferably, determining the robot's drivability decision based on the positional distribution of the center grid of the foot cover window in the two-dimensional elevation grid map and the foot landing risk assessment value specifically includes: The grid cell containing the centroid of the foot cover window is taken as the foot landing grid cell of the foot cover window; Landing grids with a risk assessment value less than a preset risk threshold are designated as safety grids. Connectivity analysis is performed on all safety grids, and the connected region with the most grids in the connected region analysis results is designated as a continuous landing region. When the number of grids in a continuous foothold area is greater than or equal to a preset area threshold, the continuous foothold area can be stepped on; when the number of grids in a continuous foothold area is less than the preset area threshold, the continuous foothold area cannot be stepped on.

[0012] Preferably, the construction of the two-dimensional elevation grid map of the robot specifically includes: Perform a vertical projection operation on all 3D point cloud data in the robot's gravity direction, count the maximum height of all 3D point cloud data in the same grid as the elevation value of the grid, count the total number of all 3D point cloud data in the same grid as the number of projected point clouds of the grid, and construct a two-dimensional elevation grid map based on the coordinate position, elevation value and number of projected point clouds of each grid.

[0013] Preferably, constructing a foot cover window on a two-dimensional elevation grid map based on the size parameters of the robot's foot specifically includes: The length of the sliding detection window is obtained by rounding down the ratio between the designed length of the robot's foot and the side length of the grid; the width of the sliding detection window is obtained by rounding down the ratio between the designed width of the foot and the side length of the grid; and the sliding detection window is used to traverse the two-dimensional elevation grid map grid by grid step to obtain the foot coverage window.

[0014] In a second aspect, the present invention provides a local terrain accessibility decision system for a humanoid robot, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program, when executed by the processor, implements the steps of a local terrain accessibility decision method for a humanoid robot.

[0015] The embodiments of the present invention have at least the following beneficial effects: This invention first transforms disordered three-dimensional point cloud data into regular two-dimensional elevation raster data, while simultaneously preserving terrain height information and point cloud distribution density information, namely elevation values ​​and the number of projected point clouds, thus achieving dimensionality reduction and structuring processing of terrain data.

[0016] Then, based on the robot's foot size parameters, a corresponding foot coverage window is constructed on a two-dimensional elevation grid map. Combining the elevation differences between adjacent grids within the window with the effective grid distribution density, a local height continuity index is calculated, which quantifies the degree of height continuity of the local terrain. This index can stably reflect the height continuity of the local terrain and provide a quantitative basis for assessing the structural stability of the landing area.

[0017] Furthermore, feature decomposition is performed on the grid coordinates and elevation values ​​of each grid within the foot coverage window to extract a fitted feature vector that represents the overall orientation of the local terrain. By calculating the deviation of this vector from the direction of gravity, the local surface tilt angle is obtained, which accurately represents the slope level of the local terrain and provides an independent slope dimension index for the assessment of the risk of foot slippage.

[0018] Finally, by integrating the local height coherence index and the local surface tilt angle, the landing risk assessment value of the foot coverage window is calculated. Then, by combining the position distribution of the center grid of each window in the two-dimensional elevation grid map, the drivability decision result of the robot is determined, which effectively prevents the risk of instability caused by the foot edge being suspended or local stress concentration. This achieves safe, stable, and adaptive landing closed-loop control of the humanoid robot from perception to execution in all scenarios and all terrains. Attached Figure Description

[0019] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the steps of a local terrain accessibility decision method for a humanoid robot provided by the present invention; Figure 2 This is a flowchart illustrating the steps involved in obtaining the local height coherence index provided by the present invention. Detailed Implementation

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] The following description, in conjunction with the accompanying drawings, details the specific scheme of the local terrain accessibility decision-making method and system for a humanoid robot provided by the present invention.

[0023] Please see Figure 1 The diagram illustrates a flowchart of a method for determining the local terrain accessibility of a humanoid robot according to an embodiment of the present invention. The method includes the following steps: Step S100: Based on the three-dimensional point cloud data collected in real time during the humanoid robot's movement, construct a two-dimensional elevation grid map of the robot; wherein, the two-dimensional elevation grid map includes the elevation value and the number of projected point clouds in each grid.

[0024] During autonomous movement of a humanoid robot, the raw 3D point cloud data collected by the front-end sensing device is disordered and scattered, making it difficult to directly match with the foot bearing capacity for refined local footing safety assessments. Furthermore, blank areas without data can easily be confused with flat terrain. This step, based on the real-time 3D point cloud data collected during the robot's movement, constructs a 2D elevation grid map of the robot through spatial projection and grid quantization. This grid map synchronously records the elevation value and the number of projected point clouds for each grid cell, completely preserving terrain height information and data distribution density information. The structured 2D grid data provides a unified computational basis for subsequent local terrain analysis based on foot dimensions. The number of projected point clouds also clearly distinguishes effective topographic data from blank areas, avoiding confusion in data state determination.

[0025] Based on this, firstly, during the humanoid robot's movement, depth images are acquired using the robot's front-facing depth camera. Then, an intrinsic inverse perspective projection is used to generate a set of three-dimensional point cloud data composed of the three-dimensional point cloud data during the movement. It should be understood that the method for acquiring the three-dimensional point cloud data of the humanoid robot is a well-known technology in the field of robot control, and will only be briefly introduced here.

[0026] Preferably, in one embodiment of the present invention, while acquiring 3D point cloud data, the gravity direction vector output by the IMU is read, and the opposite direction of the gravity direction vector is defined as the Z-axis. A gravity-aligned 3D coordinate system is constructed. Specifically, an absolutely horizontal plane that is perpendicular and orthogonal to the Z-axis is constructed and defined as the XY plane. By combining the Z-axis and the XY plane, a gravity-aligned 3D coordinate system completely unaffected by fuselage tilt is generated, and all 3D point cloud data is transformed to this 3D coordinate system. It should be noted that coordinate transformation is a well-known technique and will not be described in detail here.

[0027] Considering that humanoid robots may pitch or tilt when moving outdoors or indoors due to uneven ground or their own gait, directly using the point cloud height values ​​in the camera's original coordinate system would misinterpret the height changes caused by the robot's tilt as terrain undulations. For example, when the robot tilts to the left, the point cloud height on the left side of the ground would be raised.

[0028] The aforementioned coordinate system transformation operation eliminates perspective distortion interference caused by changes in the robot's dynamic posture, ensuring that the height coordinates of all point clouds are strictly measured based on the horizontal plane (gravitational equipotential surface) of the physical world, making the subsequent quantification of terrain undulation comparable across frames and scenes.

[0029] Preferably, in one embodiment of the present invention, the specific process of constructing a two-dimensional elevation grid map of the robot includes: Perform a vertical projection operation on all 3D point cloud data in the robot's gravity direction, count the maximum height of all 3D point cloud data in the same grid as the elevation value of the grid, count the total number of all 3D point cloud data in the same grid as the number of projected point clouds of the grid, and construct a two-dimensional elevation grid map based on the coordinate position, elevation value and number of projected point clouds of each grid.

[0030] Specifically, the terrain in front of the robot may include cardboard boxes, slopes, and perforated grids, resulting in a continuous but disordered 3D point cloud in space. This high-dimensional, disordered data structure cannot directly participate in subsequent matrix convolution or graph network topology operations. Therefore, the complex 3D surface contact problem is reduced to a 2D elevation sampling problem, significantly reducing the computational complexity of subsequent sliding window feature solving.

[0031] First, based on a preset grid side length resolution (e.g., 0.01 meters), a regular orthogonal grid network is divided on the XY plane of the gravity-aligned coordinate system, and each grid cell is assigned a global row and column index. The row and column indices of each grid cell are then used as global two-dimensional grid coordinates. For example, the grid coordinates of a single grid cell can be represented as... ,in This indicates the row number of the raster in the XY plane. This indicates the column number of the grid in the XY plane.

[0032] Then, all the 3D point cloud data in the gravity-aligned coordinate system are projected vertically downward along the Z-axis (gravity direction) onto the 2D grid network in the XY plane. After completing the vertical projection operation of all the 3D point cloud data, each grid in the 2D grid is traversed, and the maximum height of all the 3D point cloud data in the same grid is counted as the elevation value of the grid. The total number of all the 3D point cloud data in the same grid is counted as the number of projected point clouds of the grid.

[0033] The maximum height of all 3D point cloud data within the same grid is determined by extracting the maximum height value of all 3D point cloud data in the original 3D space. Specifically, during the statistical process, if the projected point cloud count of a grid is 0, it indicates that no 3D point cloud data falls into the specific location corresponding to that grid. This means the grid location is a visual blind spot or the sensor has not detected any entity. To prevent numerical confusion between this blind spot and the real physical ground with a height of 0, such as a flat floor, in this embodiment, the elevation value of the grid corresponding to a projected point cloud count of 0 is set to a preset maximum negative number, such as -999. This indicates that the value is far below the physical height that any real terrain might have, and such grids are recorded as blind spot grids. Grids with a projected point cloud count greater than 0 are recorded as valid grids. The weight between this blind spot grid and any subsequent valid grids will approach zero due to exponential decay, thus mathematically isolating invalid data while retaining the valid zero elevation of a real flat ground for calculation.

[0034] It should be understood that in a two-dimensional elevation grid map, each grid corresponds to a grid coordinate, an elevation value, and a number of projected point clouds. The elevation value represents the most prominent physical surface at the grid's location, i.e., the top of an obstacle's shell or the ground surface, indicating the grid's absolute elevation. The number of projected point clouds represents the sensor's detection coverage density at the local location of that grid.

[0035] Step S200: Based on the size parameters of the robot's foot, construct a foot coverage window on a two-dimensional elevation grid map. Based on the elevation value difference between each grid and its adjacent grids in the foot coverage window, and combined with the distribution density of the effective grids in the foot coverage window, obtain a local height coherence index.

[0036] Obstacle surfaces may contain both scattered areas with missing data and actual structural fractures and holes. Relying solely on height dispersion statistics is insufficient to distinguish between these two types of areas, and random fluctuations in the effective data size can lead to instability in the criteria for determining continuity features. This step constructs a foot coverage window of corresponding size on a two-dimensional elevation grid map based on the robot's foot size parameters. Height transition features are characterized by combining the elevation differences between each grid cell and its adjacent cells within the window, while also incorporating the distribution density of the effective grid cells within the window for correction. Finally, a local height continuity index is calculated. This index objectively reflects the degree of height continuity of the local terrain, and the density correction mechanism eliminates feature shifts caused by fluctuations in the number of effective grid cells, stably distinguishing between scattered missing data and actual structural fractures.

[0037] Preferably, in one embodiment of the present invention, the specific process of constructing a foot cover window on a two-dimensional elevation grid map based on the size parameters of the robot's foot includes: The length of the sliding detection window is obtained by rounding down the ratio between the designed length of the robot's foot and the side length of the grid; the width of the sliding detection window is obtained by rounding down the ratio between the designed width of the foot and the side length of the grid; and the sliding detection window is used to traverse the two-dimensional elevation grid map grid by grid step to obtain the foot coverage window.

[0038] Specifically, the physical dimensions of the robot's foot determine the force boundary of the robot's foot. Areas far from the landing point in the global terrain map do not have a substantial impact on the current support stability. Based on the robot's foot size parameters, a foot coverage window is constructed on a two-dimensional elevation grid map. The foot coverage window reflects the actual physical projection range of the corresponding robot foot landing, avoiding the inclusion of irrelevant distant terrain in the evaluation and ensuring the foot-specificity of the evaluation.

[0039] It should be understood that the robot's foot size parameters include the designed length and width of the foot, which can be directly read by the robot system. The length of the sliding detection window is obtained by rounding up the ratio of the designed foot length to the side length of the grid, and the width of the sliding detection window is obtained by rounding up the ratio of the designed foot width to the side length of the grid. Thus, the size of the sliding detection window can be determined using the robot's foot size parameters.

[0040] Using a sliding detection window, with a step size of one grid cell, the system traverses a 2D elevation grid map. Each sliding window corresponds to a foot cover window. It's important to understand that for local areas where the full size of the sliding detection window cannot be constructed, foot assessment is not performed at such boundary locations. This sliding window detection method reduces the complexity of global surface analysis to several independent, small-scale local planes and topological analyses, making real-time operation possible on embedded low-computing-power platforms. The size of the sliding detection window avoids decision-making errors caused by an evaluation range that is too large (introducing irrelevant distant terrain) or too small (missing edge support points). The dense sliding with a step size of one grid cell ensures that every grid cell on the map is evaluated once as the geometric center of a potential foothold, thus accurately detecting local terrain areas traversable by the robot's feet.

[0041] It should be noted that the calculation process for the local height coherence index and the subsequent local surface tilt angle is exactly the same within the corresponding foot cover window for each slide. This embodiment uses the feature analysis process of any one foot cover window as an example for illustration.

[0042] If the number of projected point clouds within the grid cells of the foot cover window is too sparse, it indicates that there is a significant terrain gap in the local area corresponding to that foot cover window. Forcibly continuing to calculate local height coherence indices and subsequent local surface tilt angles would lead the algorithm into meaningless loops or degradation. Therefore, a density proportion interception and abnormal distribution exclusion mechanism is introduced.

[0043] Preferably, before analyzing the local height coherence index and subsequent local surface tilt angle of the plantar coverage window, this embodiment of the invention further includes a security verification process for the point cloud coverage density of the plantar coverage window, specifically including: The ratio between the number of effective graticules within the foot cover window and the total number of all graticules within the foot cover window is taken as the point cloud cover density of the foot cover window; When the point cloud coverage density is less than the preset safety coverage threshold, the foot coverage window is directly marked as an absolutely prohibited footing label; when the point cloud coverage density is greater than or equal to the preset safety coverage threshold, the feature analysis process of local height coherence index and local surface tilt angle in the subsequent steps is executed.

[0044] It should be noted that the method for obtaining the effective grid has been described in detail in step S100, and will not be repeated here. The value of the safety coverage threshold is set by the safety experience value to prevent stepping on the hollow ground. For example, it is set to 70%, which means that at least 70% of the actual terrain in the local area corresponding to the foot coverage window supports the robot's footing.

[0045] When the point cloud coverage density of the foot cover window is less than the preset safe coverage threshold, it indicates that if a foot is placed within the local area of ​​the foot cover window, a large area of ​​data-free deep pits appears within the current foot cover area, representing a real structural gap, such as encountering a hollow sewer grid. To prevent the robot from misjudging this hole as safe terrain due to continued calculation, the final risk assessment result is directly output, marking the foot cover window as an absolutely prohibited landing area. Subsequent feature analysis steps are not performed, and the window is directly slid to perform a new round of feature analysis for the next foot cover window.

[0046] When the point cloud coverage density of the foot cover window is greater than or equal to the preset safe coverage threshold, it indicates that the overall surface structure of the terrain within the current foot cover area is basically intact, and the small amount of missing data is most likely due to scattered visual blind spots caused by material reflection. In addition, further investigation is conducted to check for anomalies in the geometric distribution of the effective grids within the foot cover window on the two-dimensional plane.

[0047] Specifically, the system checks whether there are fewer than three valid graticules within the plantar coverage window, and whether all valid graticules are aligned on the same geometric line. That is, if the total number of valid graticules within the plantar coverage window is less than three, or if all valid graticules within the plantar coverage window are collinear, subsequent feature analysis cannot be performed, and the 3D principal component analysis algorithm cannot be implemented in the subsequent local surface tilt angle acquisition method. Therefore, in this case, the final risk assessment result is directly output, the plantar coverage window is marked as an absolutely prohibited landing area, subsequent feature analysis steps are not performed, and the window is directly slid to the next plantar coverage window for a new round of feature analysis.

[0048] When the point cloud coverage density of the foot cover window is greater than or equal to the preset safe coverage threshold, and the total number of effective grids in the foot cover window is greater than or equal to three, and there are no cases where all effective grids in the foot cover window are collinear, the foot cover window passes the safety check. The foot cover window is then used as a safe and computable valid detection object for local height coherence index, local surface tilt angle, and subsequent analysis steps.

[0049] Preferably, such as Figure 2 As shown, in one embodiment of the present invention, the specific process of obtaining the local height coherence index based on the elevation difference between each grid and its adjacent grids within the plantar coverage window, combined with the distribution density of the effective grids within the plantar coverage window, includes: Step S201: Based on the elevation value difference between each effective grid and its adjacent effective grids in the plantar coverage window, analyze the elevation transition weight between effective grids and construct an adjacency matrix.

[0050] When a robot's foot treads on local terrain, the continuous transmission of the foot's support reaction force depends on the coplanarity of adjacent contact grids. The smaller the elevation difference between each grid and its adjacent grids, the more likely the two adjacent grids belong to the same continuous solid surface, allowing for a smooth transition. When the elevation difference exceeds the ankle height tolerance threshold, a real fault or cliff exists on the surface. However, considering that the scattered blind spots caused by material reflection from the depth camera can result in missing local grid data, this embodiment uses the effective grids within the foot's coverage window as nodes and the elevation difference between adjacent effective grids as a metric to construct an undirected graph network characterizing the continuity of the terrain surface structure. In this undirected graph network, the weight of the connecting edges monotonically decreases as the elevation difference increases, ensuring that the graph network reflects the true connection state of the physical entity surface, rather than spurious transition signals affected by sensor blind spots.

[0051] Specifically, the first step is to obtain the nonlinear coefficient of the elevation difference between each effective grid and its adjacent effective grid, and to perform negative correlation processing on the nonlinear coefficient to obtain the elevation transition weight between each effective grid and its adjacent effective grid.

[0052] As a specific example, the method for obtaining the elevation transition weight between any two graticles is exactly the same. This embodiment takes any valid graticle and any adjacent valid graticle as an example for illustration. For any valid graticle, each valid graticle within its eight-neighborhood is considered as its adjacent valid graticle. The method for obtaining the elevation transition weight can be expressed by the formula: in, This represents the elevation transition weight between an effective raster a and its adjacent effective raster b within the plantar coverage window, where the adjacent effective raster b is within the eight-neighborhood of effective raster a. This represents the elevation value of effective grid cell a within the plantar coverage window. This represents the elevation value of the adjacent valid grid cell b. This indicates the preset ankle height tolerance threshold. This represents an exponential function with the natural constant e as its base.

[0053] The nonlinear coefficient representing the elevation difference between effective grid a and its adjacent effective grid b is used to amplify the influence of elevation undulation faults between adjacent grids. An exponential decay model is used to represent the change in elevation value as the difference in elevation values ​​decreases. When the elevation transition weight approaches 0, it approaches 1, indicating that two adjacent valid grid cells are on the same horizontal plane and there is no structural fault in between. When the elevation value difference... When the height exceeds the ankle height tolerance threshold, the elevation transition weight decreases sharply to 0, indicating that there is a vertical cliff or steep slope between two adjacent effective grids. When the foot crosses this point, it will cause an impact collision or be completely suspended in the air, which is a geometrical abrupt change that cannot be smoothly transitioned.

[0054] It should be noted that the preset ankle height tolerance threshold represents the maximum height of terrain undulation that the robot's ankle joint is allowed to overcome. This value is set by the robot at the factory. For example, in this embodiment, the robot's ankle height tolerance threshold is 0.02 meters.

[0055] It should be understood that the method for obtaining valid grids has already been introduced in step S100, namely, grids with a number of projected point clouds greater than 0 within the foot cover window are recorded as valid grids, which will only be briefly described here. A valid grid represents the existence of a definite and detectable physical reflective surface element at that grid location, indicating the presence of physical matter at that grid location. It should also be noted that the blind zone grids involved in step S100 do not participate in node numbering and edge construction; they are only represented as missing nodes in the adjacency matrix, with the corresponding row and column elements all being zero. This design ensures that blind zones are physically isolated at the topological level and will not transmit false height transition information to the continuous network due to eigenvalue decomposition.

[0056] The second step involves constructing an adjacency matrix, using each valid raster as a row and column, based on the elevation transition weights between each valid raster and its adjacent valid raster.

[0057] Specifically, each valid grid cell is used as a node, and the elevation transition weights between adjacent valid grid cells are used as edge weights to construct connecting edges between adjacent valid grid cells, forming a graph network that represents the local surface topology. This graph structure ensures that valid grid cells with real-world spatial adjacency relationships have definite connections, providing a topological input describing the overall surface connectivity for subsequent calculations of the local height coherence index. The edge weights of the graph structure characterize the degree of coplanar continuity between adjacent valid grid cells.

[0058] Based on this, using all valid rasters as the row and column indices of the matrix, the elevation transition weights between each valid raster and its adjacent valid rasters are filled into the corresponding positions of the matrix row and column indices to construct an adjacency matrix A. In adjacency matrix A, the matrix elements corresponding to each valid raster and its non-adjacent valid rasters are directly set to 0. This adjacency matrix numerically records the mechanical coupling relationships between all valid rasters on the local terrain surface, providing standardized mathematical input for subsequent matrix eigenvalue decomposition.

[0059] As a concrete example, suppose there are K valid grid cells within the plantar coverage window. Number all valid grid cells from 1 to K. Traverse any two valid grid cells numbered i and j. If their positions on the 2D grid plane satisfy the condition of direct eight-neighborhood, calculate the elevation transition weight between the two adjacent valid grid cells and use this elevation transition weight as the element in the i-th row and j-th column of the adjacency matrix. If the two are not adjacent, assign the corresponding element a value of 0. After traversal, a K×K adjacency matrix A is generated.

[0060] The adjacency matrix A as a whole represents the vertical coplanar transition distribution pattern between all adjacent grates on the local terrain surface, and the non-zero elements describe the mechanical coupling strength between all effective grates on the local terrain surface that have a direct spatial adjacency relationship.

[0061] Step S202: Based on the sum of all elements in the same row of the adjacency matrix, construct a diagonal matrix to obtain the connection strength matrix.

[0062] Specifically, for any valid raster in the adjacency matrix, the sum of all elements in the row of the adjacency matrix containing that valid raster is calculated as the global transition value of that valid raster. The arrangement of the connection strength matrix is ​​exactly the same as that of the adjacency matrix, with each valid raster having an element value only in the diagonal position. That is, the global transition value of each valid raster is filled into the diagonal position of the row and column containing that valid raster.

[0063] For example, for an effective grid numbered i, the global transition value of that effective grid is used as the element in the i-th row and i-th column of the connectivity strength matrix. The global transition value characterizes the sum of the coupling strengths between an effective grid and all its neighboring effective grids. This value quantifies the total strength of the mechanical coupling between the effective grid and the surrounding solid surfaces within its local spatial neighborhood. For each effective grid, the more coplanar neighboring effective grids it has and the higher its elevation transition weight, the larger the global transition value, indicating that the effective grid is more deeply embedded in the continuous surface and is less prone to isolated deformation under stress. The connectivity strength matrix D characterizes the support embedding depth of each effective grid relative to its neighboring effective grids, and the distribution of its diagonal elements reflects the degree of force convergence of each micro-element within the foot cover area in the topological network.

[0064] Step S203: Perform eigenvalue decomposition on the difference matrix between the connection strength matrix and the adjacency matrix to obtain connectivity eigenvalues. The connectivity eigenvalues ​​are the second smallest eigenvalues ​​in the eigenvalue decomposition results arranged in ascending algebraic order.

[0065] Specifically, the connection strength matrix D and the adjacency matrix A are subtracted element by element, that is, L=DA, where L represents the K×K nonnormalized Laplace matrix.

[0066] The Laplacian matrix L is the topological potential matrix of a graph network. The diagonal elements of the Laplacian matrix L are equal to the diagonal elements of the connection strength matrix D. For example... , Let represent the value of the element in the i-th row and i-th column of the Laplace matrix L. The element value in the i-th row and i-th column of the connection strength matrix D represents the potential energy stored in the effective grid at index i anchored to the overall network; off-diagonal elements , Let represent the value of the element in the i-th row and j-th column of the Laplace matrix L. The element value in the i-th row and j-th column of the adjacency matrix A represents the negative coupling potential energy between the effective grid at index i and the effective grid at index j. The larger its absolute value, the greater the external energy input required to break the connection. The eigenvalue decomposition of the Laplace matrix L decomposes the overall connectivity state of the graph network into several independent vibration modes, where the smallest eigenvalue is always zero, corresponding to the rigid body mode of synchronous translation of all effective grids.

[0067] The second smallest eigenvalue represents the minimum sum of the weights of all connecting edges that must be cut to divide the current graph network into two disconnected subgraphs. It quantifies the topological integrity margin of the local terrain surface against physical fractures. The higher the value, the less obvious the weak cutting path inside the surface, and the better the structural integrity. When the value approaches zero, it indicates that there is a natural cutting path formed by a high-drop fault inside the surface, and the structure is divided into several independent fragments.

[0068] It should be understood that in this embodiment, the result of eigenvalue decomposition of the Laplacian matrix L is to arrange all eigenvalues ​​in ascending order and take the second smallest eigenvalue as the connected eigenvalue.

[0069] Step S204: Based on the connectivity feature value and the proportion of effective grids in the plantar coverage window, obtain the local height coherence index.

[0070] Specifically, a local height coherence index is determined based on the product of the connectivity feature value and the proportion of effective grids within the foot cover window. In this embodiment, the ratio between the number of all effective grids within the foot cover window and the total number of all grids is calculated as the proportion of effective grids within the foot cover window, which is also the point cloud coverage density recorded in step S200. This reflects the proportion of the actual effective contact area of ​​the robot's foot, and the product of the connectivity feature value and the proportion of effective grids within the foot cover window is then used as the local height coherence index.

[0071] By calculating the product between connectivity eigenvalues ​​and the proportion of nodes, the topological connectivity strength is quantified within the total area of ​​the sole. If only a small number of nodes form a highly connected subset within a window, but a large area is a blind spot or a hole, this product will significantly decrease, accurately reflecting the proportion of the actual area of ​​the sole that can receive support.

[0072] The local height coherence index characterizes the overall integrity of the load-bearing topology network formed by all effective grids through spatial adjacency within the robot's foot cover window. This index quantifies the ability of the local terrain surface to resist structural fractures caused by vertical faults, cracks, and depressions, that is, the strength of the local terrain surface in maintaining the continuity of mechanical transmission in the horizontal direction.

[0073] Step S300: Perform feature decomposition on the grid coordinates and elevation values ​​corresponding to each grid in the foot cover window to obtain the fitting feature vector of the local terrain of the foot cover window. Based on the deviation between the fitting feature vector and the opposite direction of the robot's gravity, obtain the local surface tilt angle.

[0074] Some terrain surfaces are flat but have an overall inclination. When the inclination angle exceeds the foot's friction bearing limit, it can cause a risk of sideslip. Relying solely on the consistent height feature cannot identify such hidden dangers on smooth slopes, and the calculation of surface orientation is prone to directional ambiguity, affecting the accuracy of inclination angle assessment. This step performs feature decomposition on the grid coordinates and elevation values ​​corresponding to each grid in the foot coverage window, and fits a fitted feature vector representing the overall orientation of the local terrain. By calculating the deviation between this fitted feature vector and the opposite direction of the robot's gravity, the local surface inclination angle is obtained. This indicator independently quantifies the macroscopic degree of terrain inclination, making up for the insensitivity of pure height features to smooth slopes. It can effectively identify the sidelip risk of terrain, improve the dimensions of terrain safety assessment, and provide an independent feature dimension for subsequent assessment of the sidelip risk when the robot's foot steps on it.

[0075] Preferably, in one embodiment of the present invention, the specific process of performing feature decomposition on the grid coordinates and elevation values ​​corresponding to each grid in the foot cover window to obtain the fitting feature vector of the local terrain of the foot cover window includes: Based on the grid coordinates and elevation values ​​of each grid, construct the three-dimensional coordinate values ​​of each grid. The 3D coordinate values ​​of all valid grids within the plantar coverage window are decomposed into features. The feature vector corresponding to the smallest feature value in the feature decomposition result is used as the fitting feature vector of the local terrain of the plantar coverage window.

[0076] Considering that the grid coordinates and elevation values ​​of the grid within the foot cover window are the row and column positions of the grid and the stored values, and do not have the meaning of length in physical space, the grid coordinates in the two-dimensional grid network are converted into the spatial position coordinates of the robot in the physical space, so that the subsequent geometric calculations have actual physical scale.

[0077] Specifically, for each grid cell within the foot coverage window, the product of the grid cell's horizontal coordinate index and the grid cell's side length, as well as the product of the grid cell's vertical coordinate index and the grid cell's side length, are calculated to obtain the X-axis and Y-axis coordinates of the three-dimensional coordinates. The height value of each grid cell is then used as the Z-axis coordinate to obtain the three-dimensional coordinates of each grid cell.

[0078] This process restores the dimensionless coordinate index to the actual X-axis and Y-axis position coordinates on the physical plane, and recombines the XY plane coordinates with the absolute elevation values ​​to generate the three-dimensional coordinate points of the grid in the gravity-aligned three-dimensional coordinate system with physical units.

[0079] The original local point cloud is scattered within the physical area covered by the foot window, and its absolute coordinate values ​​vary significantly depending on the window's position in the overall map (e.g., larger coordinate values ​​when the window is far away and smaller coordinate values ​​when it is close). Directly mixing the absolute coordinates of different locations into the covariance calculation would cause the variance information to be dominated by the absolute coordinate values, obscuring the contribution of the terrain's undulations. Therefore, this embodiment of the invention includes a decentralization normalization process for the three-dimensional coordinate values ​​before performing feature decomposition.

[0080] As a concrete example, the mathematical average of the three-dimensional coordinate values ​​of all valid grids within the plantar coverage window in the X, Y, and Z dimensions is calculated to obtain the spatial centroid coordinates of all valid grids. Using these spatial centroid coordinates as a reference, the three-dimensional coordinate values ​​of each valid grid within the plantar coverage window are subtracted from these spatial centroid coordinates to achieve zero-mean centering, resulting in the normalized three-dimensional coordinate values ​​of each valid grid within the plantar coverage window.

[0081] It should be understood that the centering operation translates the reference source points of all valid rasters to the spatial centroid corresponding to the foot cover window, ensuring that the subsequent calculation of the covariance matrix is ​​entirely based on the degree of dispersion of each point relative to the centroid, thus eliminating the translation effect caused by the absolute position of the window. This guarantees that the extracted terrain orientation features only reflect the geometry of the local terrain and are independent of the window's position in the global map.

[0082] Furthermore, feature decomposition is performed on the normalized 3D coordinate values ​​of all effective raster within the plantar coverage window. For example, a 3×3 covariance matrix is ​​constructed, where each element represents the covariance of the normalized 3D coordinate values ​​of all effective raster in two different coordinate dimensions. The covariance matrix is ​​then decomposed using the singular value decomposition algorithm. The feature vector corresponding to the minimum eigenvalue among all the decomposed feature vectors is used as the fitting feature vector of the local terrain of the plantar coverage window.

[0083] Geometrically, the eigenvector of the minimum eigenvalue represents the minimum direction of change in the 3D coordinate values ​​after normalization of all valid grids. Physically, the eigenvector corresponding to the minimum eigenvalue points to the macroscopic surface normal direction of the local terrain where the foot cover window is located. The fitted eigenvector represents the direction of the total normal constraint reaction force exerted by the terrain surface on the foot rubber when the robot's foot makes contact.

[0084] Preferably, in one embodiment of the present invention, the specific process of obtaining the local surface tilt angle based on the deviation between the fitted feature vector and the opposite direction of the robot's gravity includes: calculating the absolute value of the cosine of the angle between the fitted feature vector and the unit vector in the opposite direction of the robot's gravity, and using the angle obtained by inverse solving of the absolute value as the local surface tilt angle.

[0085] It should be understood that the opposite direction of gravity refers to the direction opposite to the direction of gravity, which indicates the direction perpendicular to the Z-axis in the gravity-aligned coordinate system. For example, the cosine of the angle between the fitted feature vector and the unit vector in the positive direction of the Z-axis in the gravity-aligned coordinate system can be obtained by calculating the dot product between the fitted feature vector and the unit vector in the opposite direction of the robot's gravity.

[0086] Furthermore, since the eigenvectors obtained from singular value decomposition only indicate the direction of the line, without distinguishing whether it points to the sky or the earth's center, directly calculating the dot product may yield negative values ​​or obtuse angles greater than 90 degrees. Taking the absolute value forces the mapping of this angle to an acute angle range of 0 to 90 degrees, which is physically equivalent to the dihedral angle (i.e., slope angle) between the macroscopic terrain plane and the absolute horizontal plane. When the local surface tilt angle approaches 0 degrees, the terrain plane is strictly perpendicular to the direction of gravity, and the robot's foot only needs to provide a vertical support force equal to gravity, with the tangential component being zero. As the local surface tilt angle increases, the projection component of the gravity vector onto the tangential direction of the plane increases accordingly, and this tangential component is entirely balanced by the static friction between the foot rubber and the terrain surface. Once the local surface tilt angle exceeds the friction angle determined by the maximum static friction coefficient, irreversible macroscopic lateral sliding will occur on the foot.

[0087] The local surface tilt angle reflects the acute angular deviation between the fitted feature vector and the vertically upward reference Z-axis (i.e., the opposite direction of gravity) in the gravity-aligned coordinate system. A local surface tilt angle of 0° represents an absolutely horizontal terrain plane, while a local surface tilt angle of 90° represents an absolutely vertical terrain plane. This calculation method ensures that regardless of whether the fitted normal vector points towards the ground or the sky, the tilt angle obtained from the analysis uniquely corresponds to the true dihedral angle between the macroscopic terrain plane and the gravity equipotential surface, making the assessment of sideslip risk more accurate.

[0088] Step S400: The landing risk assessment value of the foot cover window is obtained by integrating the local height coherence index and the local surface tilt angle. Based on the position distribution of the center grid of the foot cover window in the two-dimensional elevation grid map and the landing risk assessment value, the drivability decision result of the robot is determined.

[0089] Landing safety is affected by multiple factors, including height fracture and lateral slippage. Single-dimensional features cannot fully cover the risks, and single-point risk assessment does not consider the load-bearing requirements of continuous support areas, making it difficult to directly support the robot's gait decisions. This step integrates local height continuity indicators and local surface tilt angles to comprehensively calculate the landing risk assessment value of the foot coverage window. Then, by combining the positional distribution of the central grid of each foot coverage window on a two-dimensional elevation grid map, the risk level and spatial arrangement patterns of different areas are considered to ultimately determine the robot's drivability decision. This method combines single-window risk quantification with global spatial position distribution, ensuring the comprehensiveness of single-point safety assessment while also leveraging grid positions to understand the overall safety pattern of the terrain. The output drivability decision is more aligned with the robot's navigation and gait control needs.

[0090] Preferably, in one embodiment of the present invention, the specific process of obtaining the foot landing risk assessment value of the plantar coverage window by integrating the local height coherence index and the local surface tilt angle includes: The local surface tilt angle and local height consistency index are normalized respectively; Based on the normalized local surface tilt angle and the normalized local height coherence index, the landing risk assessment value of the plantar coverage window is determined.

[0091] It should be noted that the local surface tilt angle and local height consistency index are normalized using the minimax normalization method, which is a well-known technique and will not be elaborated upon here. The global maximum and minimum values ​​required in the normalization process are obtained by statistically analyzing the sets of local surface tilt angles and local height consistency indices output after traversing all foot coverage windows within the complete two-dimensional elevation raster map constructed in step S100. After the local surface tilt angles and local height consistency indices of all windows have been calculated and temporarily stored, the global extreme values ​​need to be extracted uniformly and then substituted back into each window to complete the normalization and risk value fusion. This processing mode ensures that the normalization results reflect the relative risk ranking within the current complete perception scene, rather than local absolute values.

[0092] It should be further explained that when the global maximum and global minimum values ​​of a certain dimension of the local surface tilt angle and local height coherence index are equal, it indicates that there is no spatial distribution difference of the feature in the corresponding dimension in the current scene. At this time, the normalization result of all windows of the feature in the corresponding dimension is uniformly set to 0.5, which means that the feature in this dimension does not provide any information on the relative superiority or inferiority in the current scene, and maintains numerical neutrality in subsequent risk fusion.

[0093] The normalized local surface tilt angle represents the standardized position of the macroscopic slope of the local terrain relative to the maximum and minimum slopes within the scene. The numerical value of the normalized local surface tilt angle represents the normalized tangential driving force required to maintain plantar static balance.

[0094] The normalized local height coherence index characterizes the structural integrity (fracture resistance) of a local terrain relative to the standardized positions of the best and worst coherence within the scene. The numerical value of the normalized local height coherence index represents the relative ability to resist structural collapse or stress fracture of the foot.

[0095] Furthermore, the method for obtaining the landing risk assessment value of the plantar coverage window can be expressed by the formula: in, This indicates the risk assessment value for foot landing within the plantar coverage window. This represents the normalized local surface tilt angle of the foot cover window. The normalized local height coherence index representing the plantar coverage window. This represents a preset minimum positive number to prevent the denominator from being 0. In this embodiment, the preset minimum positive number can be in the range of 0.001 to 0.01, preferably 0.005. This value is an empirical value obtained through a large number of data experiments.

[0096] The landing risk assessment value characterizes the normalized lateral slip driving force that a unit of structural integrity must withstand. The smaller the value, the higher the anti-slip capability of the landing area of ​​the plantar coverage window at the cost of lower structural risk; the larger the value, the higher the landing area of ​​the plantar coverage window is in a high-risk location.

[0097] Preferably, in one embodiment of the present invention, the specific process of determining the accessibility decision result of the plantar coverage window based on the foot risk assessment value and the grid area distribution of the plantar coverage window includes: The first step is to use the grid where the centroid of the foot cover window is located as the foot landing grid of the foot cover window. The foot landing grid reflects the landing point of the robot's foot on the horizontal plane, realizing the physical mapping from local area evaluation to point command. The grid coordinates of the foot landing grid are the expected landing position that the robot planning layer will finally execute.

[0098] It should be understood that, since the sliding step size is set to one grid, each grid on the map will sequentially become the geometric center of a window at a certain moment. Therefore, by combining the foot risk assessment values ​​of the foot coverage windows where all foot grids are located with the corresponding foot grids, a risk assessment map is constructed.

[0099] More specifically, using the grid coordinates of the landing grid corresponding to each foot coverage window as an index, the landing risk assessment value of the landing grid is filled into the corresponding position of a blank matrix with the same number of rows and columns as the two-dimensional elevation grid map in step S100. For edge grids not covered by the centroid of any foot coverage window (i.e., grid positions where the foot template cannot be completely accommodated at the window sliding range boundary), the risk value at that position is directly set to the preset maximum penalty value (this value is much larger than all landing risk assessment values, for example, it can be 999), and marked as an impassable area. This operation generates a two-dimensional risk matrix that is strictly aligned in spatial dimensions with the two-dimensional elevation grid map in step S100, defined as a local terrain navigability risk map, whose spatial resolution is completely consistent with the input elevation map, and the value at any grid position represents the comprehensive navigability risk when that grid is the landing center.

[0100] The second step is to designate the landing grids whose landing risk assessment values ​​are less than the preset risk thresholds as safety grids.

[0101] Specifically, when the risk assessment value of the landing grid is less than the preset risk threshold, it means that the structural continuity and slope characteristics of the terrain where the landing grid is located meet the basic conditions of static equilibrium, and structural collapse or macroscopic slippage is not expected to occur if the foot steps on the position.

[0102] Furthermore, landing grids with a risk assessment value greater than or equal to a preset risk threshold are designated as hazardous grids.

[0103] Specifically, when the landing risk assessment value corresponding to the landing grid is greater than or equal to the preset risk threshold, it indicates that the terrain where the landing grid is located does not meet the static safety margin and will be excluded in subsequent decision-making.

[0104] It should be noted that the risk threshold represents the maximum level of risk that the robot is allowed to land on. In this embodiment, it is preferably 0.4, which is an empirical value obtained through a large number of experiments.

[0105] The third step is to perform connectivity analysis on all safety grids and select the connectivity region with the most grids in the connectivity analysis results as the continuous landing region.

[0106] It should be noted that connected component analysis algorithms are well-known techniques and will only be briefly introduced here. For example, safe grids can be marked as 1 and dangerous grids as 0. The binarized labels corresponding to all landing grids are used to construct a binary map. The eight-neighbor connected component clustering method is then used to perform connected component analysis on the binary map, that is, to merge adjacent safe grids into a single connected component. In a local scene, the location of the largest continuous low-risk grid provides the most ample foot containment space and the most sufficient landing tolerance redundancy, and is the optimal support candidate that the robot can obtain within its current perception range. Therefore, the connected component with the largest number of grids in the connected component analysis result is taken as the continuous landing region.

[0107] Specifically, when there is only one connected component in the connected component analysis results, that connected component is directly taken as the continuous landing region.

[0108] When the robot's foot rubber comes into contact with the terrain surface, a continuous distribution of support reaction forces needs to be obtained on a solid surface across a continuous space. If the safety grids are spatially discrete and isolated, and are interrupted by dangerous grids, a stable support envelope cannot be formed when the foot actually steps on the ground due to some areas being suspended or sudden changes in force boundaries. Connectivity analysis, by detecting the spatial adjacency between grids, aggregates all directly adjacent safety grids into several spatially isolated continuous blocks, thereby filtering out candidate locations that can physically accommodate complete foot contact.

[0109] Fourth, when the number of grids in the continuous foothold area is greater than or equal to the preset area threshold, the continuous foothold area can be stepped on; when the number of grids in the continuous foothold area is less than the preset area threshold, the continuous foothold area cannot be stepped on.

[0110] When the number of grid cells in a continuous landing area is greater than or equal to a preset area threshold, it indicates that the width and depth of the continuous landing area have sufficient physical containment space in the horizontal direction, enough to completely accommodate the robot's foot and provide it with a complete support reaction force distribution boundary. At this time, it is determined that the continuous landing area meets the dual requirements of static bearing capacity and dynamic landing point deviation redundancy, and the local terrain traversability decision result corresponding to the continuous landing area is output as "walkable".

[0111] Furthermore, the geometric center coordinates of the continuous foot-landing area can be used as the desired foot-landing point, packaged into a stepping execution command, and sent to the lower-level drive controller of the humanoid robot to control the robot to complete the stepping action. The final output foot position command is the grid coordinate of the foot-landing grid, and the robot's drive controller performs foot trajectory interpolation and joint torque calculation accordingly.

[0112] When the number of grids in a continuous foothold area is less than a preset area threshold, it indicates that the continuous foothold area spatially manifests as a fragmented, isolated platform, a sharp protrusion with an extremely small area, or a narrow strip of land that is not wide enough to accommodate the sole of the foot. If the robot's foot steps on such terrain, its sole area will not be completely covered by the support surface, and the edge of the sole will be suspended outside the boundary of the support surface, resulting in insufficient contact area and uneven pressure distribution, which can easily trigger a flipping torque or local slippage instability.

[0113] At this point, the continuous footholding area is determined to not meet the geometric boundary conditions for stable foot contact. The local terrain drivability decision for this area is output as "untouchable," and the generation of any foothold coordinates in this area is prohibited. Furthermore, the entire local terrain drivability risk map can be passed to the upper-level path planner: based on the continuous risk gradient values ​​of each grid in the risk map, the planner re-searches for safe paths while avoiding all high-risk areas. This mechanism ensures that the rejection of a single window only contributes one high-risk grid marker to the global risk map. The final drivability decision is jointly determined by the area determination of the entire connected domain and the planner's global path search, thus forming a complete decision-making and control closed loop from perception and assessment to safe passage.

[0114] It should be noted that the area threshold represents the minimum physical area required to completely accommodate the robot's foot without edge overhang. As a concrete example, this area threshold can be obtained using the following formula: in, Indicates the area threshold. and These represent the robot's foot design length and foot design width, respectively. Indicates the side length of the grid. This indicates the preset area allowance coefficient, which represents the safety magnification factor to prevent slippage from stepping on the edge. For example, it is set to 1.5, and this value is an empirical value obtained through a large number of tests.

[0115] This represents the minimum projected area of ​​the foot, defined by the robot's foot size, and signifies the theoretical minimum support area required for the foot to fully conform to a flat surface. An area reserve factor is used to amplify the allowable safe zone area, providing geometrical accommodation for horizontal deviations in the landing point caused by inertia and joint control delays during dynamic walking, preventing the heel or toe from dangling beyond the edge of the support surface. This is achieved through calculation... Converting continuous physical area into discrete grid counts enables area qualification judgment at the grid quantity level.

[0116] In summary, the embodiments of the present invention achieve independent quantification of terrain structure integrity and sideslip risk in two dimensions by constructing multi-dimensional physical features (local height coherence index and local surface tilt angle). Through global normalization, data-driven fusion generates a comprehensive risk assessment value, and finally outputs executable gait commands. This enables the robot to effectively distinguish between reflective blind spots and real holes in all terrain scenarios, accurately identify hidden dangers on smooth slopes, and adaptively eliminate differences in feature dimensions, forming a stable and reliable drivability decision-making closed loop from perception to execution.

[0117] This invention also provides a local terrain accessibility decision system for a humanoid robot, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is executed by the processor, it implements the steps of a local terrain accessibility decision method for a humanoid robot.

[0118] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for determining the local terrain drivability of a humanoid robot, characterized in that, The method includes the following steps: Based on the three-dimensional point cloud data collected in real time during the humanoid robot's movement, a two-dimensional elevation grid map of the robot is constructed; wherein, the two-dimensional elevation grid map includes the elevation value and the number of projected point clouds in each grid. Based on the size parameters of the robot's foot, a foot coverage window is constructed on a two-dimensional elevation grid map. Based on the elevation value difference between each grid and its adjacent grids in the foot coverage window, combined with the distribution density of the effective grids in the foot coverage window, a local height coherence index is obtained. The grid coordinates and elevation values ​​corresponding to each grid in the foot cover window are decomposed to obtain the fitting feature vector of the local terrain of the foot cover window. Based on the deviation between the fitting feature vector and the opposite direction of the robot's gravity, the local surface tilt angle is obtained. By integrating the local height coherence index and the local surface tilt angle, the landing risk assessment value of the foot cover window is obtained. Based on the position distribution of the central grid of the foot cover window in the two-dimensional elevation grid map and the landing risk assessment value, the robot's accessibility decision result is determined.

2. The method for determining the local terrain drivability of a humanoid robot according to claim 1, characterized in that, The local height coherence index is obtained by combining the elevation difference between each grid and its adjacent grids in the plantar coverage window with the distribution density of the effective grids in the plantar coverage window. Specifically, this includes: The grids with a number of projected point clouds greater than zero within the foot cover window are considered valid grids. Based on the elevation difference between each effective grid cell and its adjacent effective grid cells in the plantar coverage window, the elevation transition weight between effective grid cells is analyzed, and an adjacency matrix is ​​constructed. Based on the sum of all elements in the same row of the adjacency matrix, a diagonal matrix is ​​constructed to obtain the connection strength matrix. The connectivity eigenvalues ​​are obtained by performing eigenvalue decomposition on the difference matrix between the connectivity strength matrix and the adjacency matrix. The connectivity eigenvalues ​​are the second smallest eigenvalues ​​in the eigenvalue decomposition results, arranged in ascending algebraic order. The local height coherence index is obtained based on the connectivity feature value and the proportion of effective grids in the plantar coverage window.

3. The method for determining the local terrain drivability of a humanoid robot according to claim 2, characterized in that, The step involves analyzing the elevation transition weights between effective graticles based on the elevation differences between each effective graticle and its adjacent effective graticles within the plantar coverage window, and constructing an adjacency matrix. Specifically, this includes: Obtain the nonlinear coefficient of the elevation value difference between each effective grid and its adjacent effective grid, and perform negative correlation processing on the nonlinear coefficient to obtain the elevation transition weight between each effective grid and its adjacent effective grid; Using each valid raster as a row and column, an adjacency matrix is ​​constructed based on the elevation transition weights between each valid raster and its adjacent valid raster.

4. The method for determining the local terrain drivability of a humanoid robot according to claim 2, characterized in that, The local height coherence index is obtained based on the connectivity feature value and the proportion of effective grids in the plantar coverage window, specifically including: The local height coherence index is determined by multiplying the connectivity feature value and the percentage of effective grids within the plantar coverage window.

5. The method for determining the local terrain drivability of a humanoid robot according to claim 4, characterized in that, The step of performing feature decomposition on the grid coordinates and elevation values ​​corresponding to each grid cell in the foot cover window to obtain the fitted feature vector of the local terrain of the foot cover window specifically includes: Based on the grid coordinates and elevation values ​​of each grid, construct the three-dimensional coordinate values ​​of each grid. The 3D coordinate values ​​of all valid grids within the plantar coverage window are decomposed into features. The feature vector corresponding to the smallest feature value in the feature decomposition result is used as the fitting feature vector of the local terrain of the plantar coverage window.

6. The method for determining the local terrain drivability of a humanoid robot according to claim 1, characterized in that, The method of integrating the local height coherence index and the local surface tilt angle to obtain the foot coverage window landing risk assessment value specifically includes: The local surface tilt angle and local height consistency index were normalized respectively; Based on the normalized local surface tilt angle and the normalized local height coherence index, the landing risk assessment value of the plantar coverage window is determined.

7. The method for determining the local terrain drivability of a humanoid robot according to claim 1, characterized in that, The determination of the robot's drivability decision based on the positional distribution of the central grid of the foot cover window in the two-dimensional elevation grid map and the foot landing risk assessment value specifically includes: The grid cell containing the centroid of the foot cover window is taken as the foot landing grid cell of the foot cover window; Landing grids with a risk assessment value less than a preset risk threshold are designated as safety grids. Connectivity analysis is performed on all safety grids, and the connected region with the most grids in the connected region analysis results is designated as a continuous landing region. When the number of grids in a continuous foothold area is greater than or equal to a preset area threshold, the continuous foothold area can be stepped on; when the number of grids in a continuous foothold area is less than the preset area threshold, the continuous foothold area cannot be stepped on.

8. The method for determining the local terrain drivability of a humanoid robot according to claim 1, characterized in that, The two-dimensional elevation grid map for constructing the robot specifically includes: Perform a vertical projection operation on all 3D point cloud data in the robot's gravity direction, count the maximum height of all 3D point cloud data in the same grid as the elevation value of the grid, count the total number of all 3D point cloud data in the same grid as the number of projected point clouds of the grid, and construct a two-dimensional elevation grid map based on the coordinate position, elevation value and number of projected point clouds of each grid.

9. The method for determining the local terrain drivability of a humanoid robot according to claim 1, characterized in that, The step of constructing a foot cover window on a two-dimensional elevation grid map based on the size parameters of the robot's foot specifically includes: The length of the sliding detection window is obtained by rounding down the ratio between the designed length of the robot's foot and the side length of the grid; the width of the sliding detection window is obtained by rounding down the ratio between the designed width of the foot and the side length of the grid; and the sliding detection window is used to traverse the two-dimensional elevation grid map grid by grid step to obtain the foot coverage window.

10. A local terrain drivability decision-making system for a humanoid robot, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of a local terrain accessibility decision method for a humanoid robot as described in any one of claims 1-9.

Images