Intelligent control method for single-hand and double-hand grabbing task of humanoid robot

Abstract

The invention provides an intelligent control method for a single-hand and double-hand grabbing task of a humanoid robot, and the method comprises the steps: obtaining interlayer connection strength distribution and gap width data of nested objects, and constructing an initial connection strength distribution diagram and a gap width record chart; according to the connection strength distribution diagram and the gap width record chart, the boundary of a continuous reachable area and a discrete isolated area is recognized, and the deviation degree of the gap transformation amplitude and a critical threshold value is evaluated; based on the deviation degree, the priority order of the grabbing hierarchy is adjusted, a target hierarchy is determined, and a finger insertion path is planned; the finger insertion path is combined with real-time contact force distribution to evaluate the contact degree, and the path offset distance is dynamically adjusted to generate a to-be-grabbed path; and a final grabbing level and a grabbing path are determined based on the weak position coordinates, and the humanoid robot is guided to execute the grabbing action.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent control method for single-handed grasping tasks of a humanoid robot. Background Technology

[0002] Humanoid robot grasping technology is a crucial research area in robotics, possessing irreplaceable value in industrial production, domestic services, and operations in complex environments. Especially when handling nested objects, the accuracy and stability of the grasp directly impact the success rate and safety of the task. Research in this field has not only driven advancements in robotics but also offered possibilities for solving complex operational problems in real life. However, current methods for grasping nested objects often struggle to balance grasping stability with protecting the object's structure. Many solutions tend to simply select a single layer for grasping, neglecting the differences in connection strength between different layers and the potential damage that may be caused to the object during grasping. This oversight leads to the robot potentially damaging the object's structure due to improper grasping strategies, especially when dealing with complex nested objects. A deeper technical challenge lies in the crucial factor of the directional distribution of connection strength between layers. In nested structures, the connection strength between each layer and adjacent layers is often not uniform but exhibits directional differences; some areas are tightly connected, while others are more fragile. If the robot's fingers accidentally touch these fragile areas while passing through the outer layers, it can lead to damage to the entire structure. The non-uniform distribution of this connection strength significantly increases the complexity of grasping path planning, as the robot needs to ensure grasping depth while avoiding critical points that could cause damage. Take carbon fiber composite laminates, commonly used in aerospace manufacturing, as an example. Each layer of carbon fiber fabric is laid alternately at specific angles such as 0°, 90°, and ±45° and bonded to an epoxy resin matrix via hot pressing. The tensile strength along the fiber direction can reach over 1500 MPa, while the interlaminar shear strength perpendicular to the layup direction is often less than 80 MPa. This nearly twenty-fold strength difference makes it extremely easy for the robot's fingers to cause delamination if force is applied along the interlaminar shear direction. For instance, when grasping a nested object composed of multiple thin shells, the robot may need to insert its fingers into deeper layers to ensure a firm grasp. However, the connection strength in some parts of the outer layers is low. If the finger deviates slightly during insertion, it may cause the outer layer to crack, or even affect the integrity of the inner structure. This contradiction between grasping depth and structural protection makes it difficult for the robot to find a safe operating path in actual tasks. Therefore, how to accurately identify the directional distribution of interlayer connection strength when grasping nested objects, and effectively avoid vulnerable areas while dynamically adjusting the grasping hierarchy decision, has become a key problem that this study urgently needs to solve. Summary of the Invention

[0003] This invention provides an intelligent control method for humanoid robots to perform single-handed grasping tasks, mainly including: Obtain the interlayer connection strength distribution and gap width data of nested objects, and construct an initial connection strength distribution map and gap width record table; Based on the connection strength distribution map and the gap width record table, identify the boundaries between continuously accessible regions and discrete isolated regions, and evaluate the deviation of the gap transition amplitude from the critical threshold. Based on the degree of deviation, the priority of the grabbing level is adjusted, the target level is determined, and the finger insertion path is planned. The degree of contact is assessed by combining the finger insertion path with the real-time contact force distribution, and the path offset distance is dynamically adjusted to generate the proposed grasping path. The deviation between the real-time opening size and the initial opening size is evaluated based on the proposed grabbing path. After updating the hierarchical priority sorting based on the deviation between the real-time opening size and the initial opening size, the coordinates of the weak position to be grabbed are identified. The final grasping level and grasping path are determined based on the coordinates of the weak points, guiding the humanoid robot to perform the grasping action.

[0004] Furthermore, the step of obtaining the interlayer connection strength distribution and gap width data of nested objects, and constructing an initial connection strength distribution map and gap width record table, includes: A full-range scan of the nested object is performed to collect the optical reflection intensity values ​​of the nested layer edges at various angles, and the force distribution of the interlayer connection strength is determined based on the reflection intensity values. The stress distribution of the interlayer connection strength is divided into two or more sector regions, and the average value and standard deviation of the stress value in each region are calculated to form a contact resistance value distribution matrix. The contact resistance value distribution matrix is ​​used to identify the outline of the opening into which a finger can be inserted, the gap depth value is measured along the boundary, and the depth value sequence is recorded to form a gap width spectrum; A three-dimensional point cloud model is constructed based on the gap width spectrum and the coordinates of the spatially accessible restricted area. Key locations are extracted to generate the connection strength distribution map, and the gap depth value sequence is organized to form the gap width record table.

[0005] Furthermore, based on the connection strength distribution map and the gap width record table, the boundaries between continuously accessible regions and discrete isolated regions are identified, and the deviation of the gap transition amplitude from the critical threshold is evaluated, including: The location with the highest intensity value is selected from the connection intensity distribution map as the seed point. The continuous reachable region is expanded using the region growing algorithm. Points with intensity differences exceeding a preset threshold are marked as discrete isolated points. The spatial coordinate sequence of the region boundary line is determined. Extract the gap width value of sampling points on both sides of the region boundary line along the region boundary line, and calculate the difference between the average values ​​on both sides of the region boundary line as the transformation amplitude value; The distribution of the transformation amplitude values ​​in each layer is statistically analyzed to determine the critical threshold point for the gap transformation; The difference between the critical threshold point and the actual transition amplitude value is used as the degree of deviation between the gap transition amplitude and the critical threshold.

[0006] Furthermore, the distribution of the transition amplitude values ​​in each layer is statistically analyzed to determine the critical threshold point for gap transition, including: statistically analyzing the frequency of the transition amplitude values ​​in each layer, constructing a distribution histogram, and determining the gap width corresponding to the segmentation point with the largest inter-class variance in the distribution histogram as the critical threshold point for gap transition.

[0007] Furthermore, based on the degree of deviation, the priority ranking of the grasping layers is adjusted to determine the target layer and plan the finger insertion path, including: The hierarchical weight values ​​of each nested layer are calculated based on the deviation of the gap transition amplitude from the critical threshold. The hierarchical weight values ​​are then arranged using a sorting algorithm to form an adjusted priority sequence. Extract the first priority level from the priority sequence as the target level, obtain the force distribution data corresponding to the target level, and mark the points with force values ​​lower than a preset threshold as weak points; A danger zone map is generated based on the spatial coordinates of the weak points. Based on the danger zone map, a path planning algorithm is used to search for the shortest path from the current location to the target grasping point, thus forming the finger insertion path.

[0008] Furthermore, generating a danger zone map based on the spatial coordinates of the weak points includes: extracting the three-dimensional spatial coordinates of the weak points to form a spatial coordinate set, using a convex hull algorithm to calculate the minimum outer envelope of all points in the spatial coordinate set, and expanding outward by a preset safety distance based on the outer envelope to generate a danger zone map that needs to be avoided.

[0009] Furthermore, based on the danger zone map, a path planning algorithm is used to search for the shortest path from the current location to the target grab point to form the finger insertion path, including: calculating the detour trajectory segment in the normal direction at the intersection point of the path and the danger zone, connecting the original path and the detour trajectory segment to obtain the finger insertion path.

[0010] Furthermore, by assessing the degree of contact through the finger insertion path combined with real-time contact force distribution, and dynamically adjusting the path offset distance to generate a proposed grasping path, including: Real-time contact force distribution data is collected along the finger insertion path. The contact force distribution data is matched with the spatial coordinates of the weak location points. The proportion of sensing units falling into the weak area is counted to obtain the contact degree ratio. The path offset is calculated based on the difference between the contact ratio and the preset safety threshold. The coordinates of the control points on the finger insertion path are corrected based on the path offset, and the corrected control point sequence is connected by an interpolation algorithm to generate the proposed grasping path.

[0011] Furthermore, based on the proposed grabbing path, the deviation between the real-time opening size and the initial opening size is evaluated. After updating the hierarchical priority ranking based on the deviation between the real-time opening size and the initial opening size, the coordinates of the weak points to be grabbed are identified, including: For the control points on the proposed grasping path, calculate the difference between the real-time opening size and the initial opening size to obtain the deviation ratio; Update the gap width record table according to the deviation ratio; Based on the updated gap width record table, the weights of each level are recalculated to generate a dynamically updated level priority sort. Based on the change in the hierarchical priority sorting, the weak location identification is re-executed to obtain an updated set of weak location coordinates.

[0012] Furthermore, based on the coordinates of the weak points, the final grasping level and grasping path are determined, guiding the humanoid robot to perform the grasping action, including: Select the layer with the highest priority from the updated layer priority sort as the final crawling layer; Based on the coordinates of the weak points corresponding to the final crawling level, a crawling path is generated that bypasses the weak points. The humanoid robot's fingertips are guided to the target position via the grasping path. The closing force is adjusted according to the gap width of the final grasping level, and the closing action is performed to complete the grasping.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an intelligent control method for humanoid robots performing single-handed grasping tasks, aiming to solve the challenges of grasping path planning caused by uneven distribution of connection strength and differences in gap width in complex nested structures, as well as the impact of real-time contact force on grasping stability. This invention constructs an initial connection strength distribution map and a gap width recording table to identify the boundaries between continuously accessible and discrete isolated regions, extracts the gap transition amplitude and determines critical thresholds, and then assesses the degree of deviation to dynamically adjust the grasping layer priority and optimize the finger insertion path. Simultaneously, this invention combines real-time contact force distribution to assess the degree of contact, dynamically updates the path offset distance and opening size deviation, repeatedly identifies weak points and plans bypass paths, and finally determines the optimal grasping layer and path, ensuring the accuracy and stability of the humanoid robot's grasping actions. This invention significantly improves the adaptability and success rate of grasping complex nested objects, providing efficient and reliable technical support for intelligent robot operation. Attached Figure Description

[0014] Figure 1 This is a flowchart of an intelligent control method for single-handed grasping tasks of a humanoid robot according to the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0016] like Figure 1 This embodiment of an intelligent control method for a humanoid robot to perform single-handed grasping tasks may specifically include: Step S101: Obtain the interlayer connection strength distribution and gap width data of nested objects, and construct an initial connection strength distribution map and gap width record table.

[0017] A humanoid robot's vision sensor performs a 360-degree surround scan of a multi-layered, thin-shell nested object, collecting optical reflection intensity values ​​at the edges of the nested layers from various angles. Based on the correspondence between reflection intensity values ​​and material stress, the force distribution of interlayer connection strength in each direction is obtained. This force distribution is divided into eight sector regions at 45-degree intervals. The average and standard deviation of the force values ​​within each sector region are calculated to obtain a contact resistance value distribution matrix. The numerical gradient changes in the contact resistance value distribution matrix identify the boundary contour of the opening where a finger can be inserted. A laser rangefinder measures the gap depth value every 5 millimeters along the boundary contour, recording the depth value sequence to form a gap width spectrum. If the depth change gradient of adjacent measurement points exceeds a preset threshold, the coordinates of that location are marked as a spatially restricted area. A three-dimensional point cloud model is constructed based on the coordinates of the spatially restricted area and the gap width spectrum data. Key locations where the interlayer connection strength exceeds the threshold are extracted from the point cloud model as visual calibration points. An interpolation algorithm connects these visual calibration points to generate an intensity distribution map. Simultaneously, the gap depth value sequence is organized according to the nesting hierarchy to form a gap width record table.

[0018] Specifically, in one implementation, a humanoid robot uses a high-resolution industrial camera in conjunction with a ring-shaped LED light source to perform a 360-degree surround scan of a multi-layered, thin-shell nested object. The light source illuminates the object's surface at a fixed incident angle. When stress concentration occurs within the material, its crystal structure undergoes minute deformation, resulting in a change in surface reflectivity.

[0019] Specifically, for every 10 MPa increase in stress, the reflection intensity changes by approximately 0.8%. Using a pre-calibrated stress-reflection intensity lookup table, the collected reflection intensity values ​​are converted into corresponding interlayer bonding strength values.

[0020] For example, the robot divides a 360-degree circle into eight sector regions at 45-degree intervals, with each region covering approximately 400 measurement points. Nested objects, such as fairings, consist of three layers of carbon fiber prepreg thermoformed into thin shells, each approximately 0.5 mm thick, for a total thickness of approximately 1.5 mm. The nesting method involves the inner, middle, and outer layers concentrically nested together, bonded together using thermosetting epoxy resin adhesive. In the carbon fiber composite nested structure, the reflection intensity value along the fiber layup direction is typically in the range of 85-92 units, corresponding to a connection strength of 1200-1500 MPa; while the reflection intensity perpendicular to the fiber direction is only 12-18 units, corresponding to a connection strength of 60-80 MPa. By calculating the average and standard deviation of all measurement points within each sector region, an 8×2 contact resistance value distribution matrix is ​​formed, where regions with a standard deviation exceeding 15 units are marked as areas of uneven intensity distribution.

[0021] It should be noted that the laser rangefinder uses a 905 nm wavelength pulsed laser to perform a helical scan along the opening boundary contour. The sensor emits laser pulses and receives reflected signals, calculating the precise gap depth value through time-of-flight. When the gradient value obtained by dividing the depth difference between adjacent measurement points by the 5 mm spacing exceeds 0.3, it indicates a sharp depth change at that location, typically corresponding to areas with weak interlayer adhesion or delamination defects. The three-dimensional coordinates of these locations are recorded as areas with limited spatial accessibility.

[0022] Preferably, the 3D point cloud model is constructed using an octree data structure to organize the spatial point cloud data, with each node storing the average intensity value and depth information within that spatial region. When extracting visual calibration points, locations with a connection intensity exceeding 800 MPa are selected as key support points. These discrete calibration points are connected using a bicubic spline interpolation algorithm to generate a continuous intensity distribution map. The interpolation weights are dynamically adjusted based on the Euclidean distance between adjacent calibration points, with closer points receiving larger weights, thus ensuring a smooth transition in the intensity distribution.

[0023] Step S102: Identify the boundaries between continuously accessible regions and discrete isolated regions based on the connection strength distribution map and the gap width record table, and evaluate the deviation of the gap transition amplitude from the critical threshold.

[0024] Based on the intensity values ​​in the initial connectivity strength distribution map, the location with the highest intensity value is selected as the seed point. A region growing algorithm is used to expand outwards from the seed point. When the intensity difference between adjacent points is less than a preset first threshold, it is marked as a continuously reachable region. When the intensity difference exceeds the first threshold and the number of surrounding connected points is less than a preset second threshold, it is marked as a discrete isolated point. The spatial coordinate sequence of the region boundary line is determined by the boundary line between the continuously reachable regions and the discrete isolated points. The gap width recording table structure is expanded, and the gap values ​​of a preset number of sampling points on both sides of the boundary are remeasured and recorded along the spatial coordinate sequence of the region boundary line. The average gap width of the sampling points on the continuously reachable region side and the average gap width of the sampling points on the discrete isolated region side are calculated respectively. The difference between the two average values ​​is the transition amplitude value at the gap abrupt change point. Nested layers are distinguished using a depth threshold layering method. The gap depth sequence is organized into inner, middle, and outer layers to obtain the distribution of the transition amplitude value in each nested layer. A distribution histogram is constructed by statistically analyzing the frequency of the transition amplitude value in each layer. The Otsu method is used to automatically determine the dividing point with the largest inter-class variance in the histogram. The gap width corresponding to this dividing point is used as the critical threshold point for gap transition. The percentage value of the deviation is obtained by dividing the difference between the critical threshold point and the actual change amplitude value by the critical threshold point.

[0025] Specifically, in one implementation, the seed point is selected based on a global maximum search in the initial connectivity strength distribution map. The robot scans the entire strength distribution map, identifying all candidate points with strength values ​​exceeding 1200 MPa, and selects the location with the highest eight-neighbor average strength as the initial seed point. The region growing algorithm starts from this seed point and expands outward layer by layer, checking adjacent pixels. When the strength difference between a neighboring point and a point within the current region is less than 50 MPa, the neighboring point is marked as part of a continuously reachable region and added to the expansion queue. The expansion process continues until all neighboring points satisfying the strength difference condition have been processed.

[0026] Specifically, when identifying discrete isolated points, the algorithm examines each point not marked as a continuously reachable region and counts the number of points within its eight neighborhoods that belong to continuously reachable regions. If this number is less than a preset threshold of three, and the strength difference between the point and the nearest continuously reachable region point exceeds 100 MPa, it is marked as a discrete isolated point. This dual-determination mechanism avoids misclassification that may be caused by a single threshold determination. The boundary between continuously reachable regions and discrete isolated points is extracted using an edge detection operator, forming a spatial coordinate sequence containing hundreds of coordinate points. In practical applications of carbon fiber composites, continuously reachable regions typically correspond to areas where the fiber layup is tightly connected, and the interlayer resin matrix in these regions is fully cured, forming a stable bonding interface. Discrete isolated points, on the other hand, often appear at layup edges or locations with manufacturing defects, such as bubbles, delamination, or resin-deficient areas. When robotic fingers apply force to these discrete isolated point areas, they can easily induce localized stress concentration, leading to interlayer delamination or fiber breakage.

[0027] For example, when sampling gap values ​​along the region boundary line, the robot starts from a starting point on the boundary line and selects a sampling position every 10 mm. At each sampling position, a normal distance of 50 mm is extended to both the continuously accessible region side and the discrete isolated region side. The gap width is measured every 10 mm along this normal path, obtaining a gap value sequence of 5 sampling points on each side. By calculating the arithmetic mean of the gap widths at each sampling point, the gap transition characteristics at that boundary position are obtained. The transition amplitude reflects the degree of abrupt change in the interlayer gap when transitioning from a continuously accessible region to a discrete isolated region. For example, in the laminate structure of aerospace manufacturing, the gap width in the normal region is typically 0.1-0.2 mm, while in the defective region it may reach 0.8-1.2 mm, with a transition amplitude value exceeding 0.6 mm.

[0028] Preferably, the application of Otsu's method involves adaptive segmentation of the histogram of transition amplitude values. The algorithm first counts the transition amplitude values ​​at all boundary positions in each nested layer, constructing a frequency histogram with a group interval of 0.05 mm. Otsu's method calculates the inter-class variance between the two categories at each possible segmentation threshold. Specifically, a threshold t is set to divide the histogram into two categories: transition amplitude values ​​less than t are classified as normal transitions, and values ​​greater than or equal to t are classified as abnormal transitions. The algorithm calculates the means μ1 and μ2 of the two categories, as well as their respective probability weights ω1 and ω2, with the inter-class variance σ² = ω1 × ω2 × (μ1 - μ2)². After iterating through all possible t values, the t that maximizes the inter-class variance is selected as the final critical threshold point. The advantage of this method is that it eliminates the need for manual threshold setting and can adaptively determine the optimal segmentation point based on the actual data distribution.

[0029] In one possible implementation, the deviation is calculated as a relative deviation. If the actual change in a boundary location is 0.75 mm, and the critical threshold determined by Otsu's method is 0.5 mm, then the deviation is calculated as (0.75 - 0.5) / 0.5 × 100% = 50%. Locations with a deviation exceeding 30% are considered to have significant structural anomalies, and the robot should avoid these high-risk areas during path planning.

[0030] Understandably, by establishing this recognition mechanism based on region growing and Otsu's method, the robot can automatically adapt to the feature distribution of different types of nested structures without needing to set separate judgment thresholds for each material. In actual grasping tasks, the deviation level provides the robot with a quantitative risk assessment indicator, enabling it to minimize damage to nested structures while ensuring grasping stability.

[0031] Step S103: Adjust the priority sorting of the grabbing layers based on the degree of deviation, determine the target layer and plan the finger insertion path.

[0032] Based on the percentage value of the deviation, a negative exponential mapping relationship is used to calculate the hierarchical weight value of each nested layer. When the deviation is P, the weight value is calculated as exp(-P / 100). The greater the deviation, the lower the weight. The bubble sort algorithm is used to arrange the hierarchical weight values ​​from high to low, forming an adjusted priority sequence. The first-ranked hierarchical number is extracted from the priority sequence as the target hierarchical number. The force distribution data in each direction corresponding to the target hierarchical number is obtained. Each sampling point in the force distribution data is traversed. When the force value of the sampling point is lower than a preset threshold, the sampling point is marked as a weak point. The three-dimensional spatial coordinates of the weak points are extracted to form a spatial coordinate set. The convex hull algorithm is used to calculate the minimum outer envelope of all points in the spatial coordinate set. Based on the outer envelope, a preset safety distance is extended outward to generate a map of the danger zone to be avoided. According to the danger zone map, A The path planning algorithm searches for the shortest path from the current position of the finger to the target level grasping point. If the path intersects with a dangerous area, the detour trajectory segment in the normal direction is calculated at the intersection point. The original path and the detour trajectory segment are connected to obtain the optimized finger insertion path.

[0033] Specifically, in one implementation, the negative exponential mapping exp(-P / 100) is chosen based on the non-linear relationship between the degree of deviation and the risk of grasping. When the degree of deviation P is 0, exp(0)=1 indicates that the level has no risk and has the highest weight value; when P increases to 100, exp(-1)≈0.368, and the weight value drops to one-third of the original; when P reaches 200, exp(-2)≈0.135, and the weight value further decays. This exponential decay characteristic ensures that levels with slight deviations still maintain high priority, while the weight of levels with severe deviations drops sharply. In the nested structure of carbon fiber composite materials, levels with a deviation of less than 20% usually indicate good interlayer bonding, and the weight value remains above 0.82; a deviation of more than 50% means there is a significant risk of layering, and the weight value drops below 0.61. The bubble sort algorithm arranges the weight values ​​from high to low by comparing and exchanging adjacent elements, forming a new priority sequence. The original default order of inner layer priority is adjusted to outer layer relative safety priority.

[0034] Specifically, the identification of weak points is based on the statistical characteristics of force distribution data. After the robot acquires the force distribution data at the target level, this data is stored in the form of a two-dimensional matrix, where each element represents the connection strength value at that location. Each sampling point in the matrix is ​​traversed, and when the force value at a point is lower than 40% of the material's yield strength, it is marked as a weak point. In practical applications of aerospace manufacturing, the yield strength of carbon fiber laminates is approximately 800 MPa; therefore, the threshold is set at 320 MPa. Areas below this value are prone to localized damage when subjected to contact force from the robot's fingers.

[0035] It's important to note that applying the convex hull algorithm in 3D space involves calculating the minimum outer envelope of all vulnerable points. The algorithm selects four non-coplanar points from the spatial coordinate set to construct an initial tetrahedron. Then, it checks the remaining points one by one. If a point is outside the current convex hull, the convex hull structure is updated, deleting all faces visible from that point and adding a new face with that point as its vertex. By iteratively processing all vulnerable points, a convex polyhedron containing all dangerous points is finally obtained. Based on this, a safety distance of 15 mm is extended outward along the normal vector of the convex hull surface, forming a larger envelope as a map of the danger zone. This safety distance takes into account the diameter of the robot's finger and positioning errors, ensuring that the finger will not accidentally touch vulnerable areas during movement.

[0036] For example, A The path planning algorithm uses the Manhattan distance as the heuristic function. Let the Euclidean distance from the current node to the target node be d, and the heuristic function h(n) = d × 1.2, where h(n) is the heuristic function value, d is the Manhattan distance, and the coefficient 1.2 is used to compensate for path curvature in three-dimensional space. The algorithm maintains two lists: an open list stores nodes to be explored, and a closed list records explored nodes. Starting from the current position of the finger, the algorithm calculates the cost function f(n) = g(n) + h(n) for adjacent nodes, where g(n) is the actual path length from the starting point to the current node. Each time, the node with the smallest f value is selected for expansion until the target point is found or the open list is empty.

[0037] Preferably, the intersection detection between the path and the danger zone employs a ray casting algorithm. The path is discretized into a series of line segments, and each segment is checked for intersection with the convex hull of the danger zone. If an intersection is detected, the algorithm calculates the normal vector of the surface at the intersection point and generates a detour control point by offsetting 25 mm along the normal vector direction. A Bézier curve connects the original path points, the detour control points, and subsequent path points to form a smooth detour trajectory segment. The control point weights of the Bézier curve are dynamically adjusted according to the curvature requirements to ensure the continuity and smoothness of finger movement.

[0038] In one possible implementation, the optimized finger insertion path is represented by a sequence of path points, each containing three-dimensional coordinates and the desired finger pose angle. The spacing between path points is set to 5 mm to ensure the precision of robot control. During actual execution, the robot uses fifth-order polynomial interpolation to generate a continuous trajectory, allowing the finger to smoothly transition from its current position to the target grasping point while avoiding all marked danger zones. Through this dynamic priority adjustment based on deviation and intelligent path planning, the robot can find a safe and stable grasping scheme in complex nested structures, effectively reducing the risk of structural damage caused by improper grasping.

[0039] Step S104: The degree of contact is evaluated by combining the finger insertion path with the real-time contact force distribution, and the path offset distance is dynamically adjusted to generate a proposed grasping path.

[0040] By optimizing the finger insertion path, real-time contact force distribution data is collected from a pressure sensor array at the fingertip. The force values ​​of each sensing unit in the contact force distribution data are matched with the spatial coordinates of weak points. The ratio of the number of sensing units falling within the weak area to the total number of sensing units is calculated to obtain the contact degree ratio. Based on the difference between the contact degree ratio and a preset safe contact threshold, a proportional-integral controller is used to calculate the path offset. When the contact degree ratio exceeds the safe contact threshold, the controller output value is multiplied by an adjustment coefficient to obtain the offset distance, resulting in a dynamic adjustment value. The spatial coordinates of each control point on the original insertion path are corrected according to the dynamic adjustment value. The corrected control point sequence is then connected using a cubic spline interpolation algorithm to obtain the proposed grasping path.

[0041] Specifically, in one implementation, the pressure sensor array at the fingertip employs an 8×8 matrix layout with 64 piezoresistive sensing units, each covering a sensing area of ​​2 square millimeters. The sensors acquire contact force distribution data in real time at a sampling frequency of 100Hz, and each sensing unit outputs a force signal within the range of 0-10N. The spatial matching process maps the local coordinate system of the sensor array to the global coordinate system using a coordinate transformation matrix. When the mapped coordinates of a sensing unit fall within a 5-millimeter radius of a weak point, that sensing unit is marked as being in a weak area. The number of all marked sensing units is counted and divided by 64 to obtain the percentage of contact intensity.

[0042] Specifically, the proportional-integral controller's proportional coefficient Kp is set to 2.5, the integral coefficient Ki is set to 0.8, and the safe contact threshold is set to 15%. When the contact ratio is 20%, the difference from the threshold is 5%, and the controller's proportional output is 5% × 2.5 = 12.5 mm. The integral output accumulates the error values ​​of the first 5 sampling periods and multiplies them by Ki; the two are then added together to obtain the total offset. The adjustment coefficient dynamically changes according to the current finger insertion depth. When the insertion depth is less than 30 mm, the coefficient is 1.0; when the depth is between 30 and 60 mm, the coefficient linearly increases to 1.5, ensuring greater avoidance space during deep insertion.

[0043] It should be noted that when processing the corrected control points, the cubic spline interpolation algorithm first calculates the first and second derivatives between adjacent control points to ensure the continuity and smoothness of the curve at connection points. The original path contains 20 control points, each offset in the normal vector direction according to the dynamic adjustment value of its corresponding position. The interpolation process uses natural boundary conditions, i.e., the second derivative of the first and last points is zero, and the polynomial coefficients of each segment are obtained by solving a system of tridiagonal linear equations.

[0044] For example, in the actual grasping of carbon fiber composite materials, when the robot finger approaches an area with delamination defects, the corresponding sensing unit detects an abnormally low contact force value, typically only about 30% of that in normal areas. The control system immediately calculates the required offset distance of 15 mm and completes path correction within 0.2 seconds, allowing the finger to bypass the danger zone and continue insertion, ultimately forming a smooth S-shaped pseudo-grasping path.

[0045] Step S105: Evaluate the deviation between the real-time opening size and the initial opening size based on the proposed grabbing path, update the hierarchical priority sorting based on the deviation between the real-time opening size and the initial opening size, and identify the coordinates of the weak position to be grabbed.

[0046] For each control point on the target capture path, the real-time opening size value at that point is collected. The difference between the real-time opening size value and the initial opening size is calculated, and the deviation ratio is obtained by dividing the difference by the initial value. The corresponding entry in the gap width record table is searched based on the deviation ratio to confirm the deviation. If the deviation ratio exceeds a preset threshold of 0 to 1, the gap width of each nested layer is rescanned, the gap width record table is updated, and the weights of each level are recalculated using the same negative exponential mapping method as the initial weight calculation (i.e., weight equals e raised to the power of the negative deviation ratio, where e is a natural constant), resulting in a dynamically updated layer priority ranking. Based on the changes in the layer priority ranking, the weak location identification process (i.e., the process of identifying coordinates based on weight ranking) is re-executed to obtain an updated set of weak location spatial coordinates.

[0047] Specifically, in one implementation, the proposed grasping path includes 25 control points, each spaced 8 millimeters apart. As the robot executes the path, it measures the opening size at each control point in real time using a laser displacement sensor. The sensor emits a laser beam perpendicular to the surface of the nested object, measures the time difference between emission and reception of the reflected signal, and converts this time difference into a distance value, which is the current opening size. The initial opening size, typically 15-20 millimeters, is recorded as a baseline value before path planning. The deviation ratio is calculated using a relative deviation formula: the difference between the real-time value and the initial value divided by the initial value and then multiplied by 100%.

[0048] Specifically, when the deviation exceeds a preset threshold of 25%, the system determines that the nested structure has undergone significant deformation. This deformation may originate from the contact force applied by the robot's fingers causing local separation between layers, or from the elastic recovery of the material itself. The system immediately initiates a gap width rescanning procedure, using the same laser ranging method as the initial scan, measuring the gap depth every 5 millimeters along the boundary contour of each nested layer, and updating the values ​​in the gap width record table.

[0049] It should be noted that the recalculation of hierarchical weights maintains the same negative exponential mapping relationship as the initial calculation, exp(-P / 100), where P is the degree of deviation after the update. If the deviation of a certain level increases from 30% to 45%, its weight value decreases from exp(-0.3)≈0.74 to exp(-0.45)≈0.64, and its position in the new priority ranking will decrease accordingly. This dynamic adjustment mechanism allows the system to respond to structural changes in real time, avoiding the selection of levels that have already undergone deformation as crawling targets.

[0050] For example, during the grasping process of a nested carbon fiber composite structure, when the finger approached the third layer, the opening size was detected to have decreased from the initial 18 mm to 13 mm, a deviation of 28%. A rescan revealed that the gap between the second and third layers had widened from 0.3 mm to 0.8 mm, indicating interlayer separation in that area. Through recalculation, the third layer, originally ranked second, dropped to fourth, and the system selected the fifth layer, which had risen in ranking, as the new target layer.

[0051] Preferably, the re-identification of weak points follows the same process as the initial identification, but is only performed on levels where the priority order has changed. Force distribution data for the new target level is acquired, regions with force values ​​below a threshold are identified as new weak points, and the set of spatial coordinates for these weak points is updated.

[0052] Step S106: Determine the final grasping level and grasping path based on the coordinates of the weak position, and guide the humanoid robot to perform the grasping action.

[0053] Based on dynamically updated hierarchical priority ranking, the first-ranked hierarchical level is selected as the final grasping level. According to the spatial coordinates of the weak point corresponding to the final grasping level, a grasping path is generated by connecting the current position to the target grasping point while avoiding the radius of the weak point coordinates. The humanoid robot's fingertips are guided along this grasping path to the target position. The fingertips adjust their closing force according to the gap width of the target hierarchical level, executing a closing action to complete the grasp.

[0054] Specifically, in one implementation, the final grabbing level is determined based on a dynamically updated priority ranking, directly selecting the first-ranked level as the target. Path generation employs an elliptical arc bypass method, starting from the current position, with the target grabbing point as the endpoint, and the weak point coordinates as the obstacle center. First, the ellipse parameters are calculated: with the starting point A and ending point B as foci, the major axis is 1.5 times the distance between A and B, and the minor axis is set based on a 10mm bypass radius. Then, a smooth curve path within a 10mm bypass radius is generated, selecting the arc segment that does not cross the obstacle as the final path.

[0055] Specifically, as the fingertip moves along a sequence of path points, each path point contains its position coordinates and desired speed. The movement speed is dynamically adjusted based on the distance to the weak point; the closer to the weak point, the slower the speed, typically varying within the range of 5-20 mm / s. The closing force is adjusted based on the gap width of the target layer. A larger closing force is applied when the gap width is less than 0.5 mm, and a smaller closing force is used when the gap width is greater than 1 mm. The torque output of the servo motor is adjusted in real time through feedback from a force sensor to achieve adaptive gripping control.

[0056] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for intelligent control of humanoid robots performing single-handed grasping tasks, characterized in that, include: Obtain the interlayer connection strength distribution and gap width data of nested objects, and construct an initial connection strength distribution map and gap width record table; Based on the connection strength distribution map and the gap width record table, identify the boundaries between continuously accessible regions and discrete isolated regions, and evaluate the deviation of the gap transition amplitude from the critical threshold. Based on the degree of deviation, the priority of the grabbing level is adjusted, the target level is determined, and the finger insertion path is planned. The degree of contact is assessed by combining the finger insertion path with the real-time contact force distribution, and the path offset distance is dynamically adjusted to generate the proposed grasping path. The deviation between the real-time opening size and the initial opening size is evaluated based on the proposed grabbing path. After updating the hierarchical priority sorting based on the deviation between the real-time opening size and the initial opening size, the coordinates of the weak position to be grabbed are identified. The final grasping level and grasping path are determined based on the coordinates of the weak points, guiding the humanoid robot to perform the grasping action.

2. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 1, characterized in that, The process of obtaining the interlayer connection strength distribution and gap width data of nested objects, and constructing an initial connection strength distribution map and gap width record table, includes: A full-range scan of the nested object is performed to collect the optical reflection intensity values ​​of the nested layer edges at various angles, and the force distribution of the interlayer connection strength is determined based on the reflection intensity values. The stress distribution of the interlayer connection strength is divided into two or more sector regions, and the average value and standard deviation of the stress value in each region are calculated to form a contact resistance value distribution matrix. The contact resistance value distribution matrix is ​​used to identify the outline of the opening into which a finger can be inserted, the gap depth value is measured along the boundary, and the depth value sequence is recorded to form a gap width spectrum; A three-dimensional point cloud model is constructed based on the gap width spectrum and the coordinates of the spatially accessible restricted area. Key locations are extracted to generate the connection strength distribution map, and the gap depth value sequence is organized to form the gap width record table.

3. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 1, characterized in that, The step of identifying the boundary between continuously accessible regions and discrete isolated regions based on the connection strength distribution map and the gap width record table, and evaluating the deviation of the gap transition amplitude from the critical threshold, includes: The location with the highest intensity value is selected from the connection intensity distribution map as the seed point. The continuous reachable region is expanded using the region growing algorithm. Points with intensity differences exceeding a preset threshold are marked as discrete isolated points. The spatial coordinate sequence of the region boundary line is determined. Extract the gap width value of sampling points on both sides of the region boundary line along the region boundary line, and calculate the difference between the average values ​​on both sides of the region boundary line as the transformation amplitude value; The distribution of the transformation amplitude values ​​in each layer is statistically analyzed to determine the critical threshold point for the gap transformation; The difference between the critical threshold point and the actual transition amplitude value is used as the degree of deviation between the gap transition amplitude and the critical threshold.

4. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 3, characterized in that, The process of statistically analyzing the distribution of the transition amplitude values ​​across each layer and determining the critical threshold point for gap transition includes: statistically analyzing the frequency of occurrence of the transition amplitude values ​​in each layer, constructing a distribution histogram, and determining the gap width corresponding to the segmentation point with the largest inter-class variance in the distribution histogram as the critical threshold point for gap transition.

5. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 1, characterized in that, The process of adjusting the priority ranking of the grasping layers based on the degree of deviation, determining the target layer, and planning the finger insertion path includes: The hierarchical weight values ​​of each nested layer are calculated based on the deviation of the gap transition amplitude from the critical threshold. The hierarchical weight values ​​are then arranged using a sorting algorithm to form an adjusted priority sequence. Extract the first priority level from the priority sequence as the target level, obtain the force distribution data corresponding to the target level, and mark the points with force values ​​lower than a preset threshold as weak points; A danger zone map is generated based on the spatial coordinates of the weak points. Based on the danger zone map, a path planning algorithm is used to search for the shortest path from the current location to the target grasping point, thus forming the finger insertion path.

6. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 5, characterized in that, The step of generating a danger zone map based on the spatial coordinates of the weak points includes: extracting the three-dimensional spatial coordinates of the weak points to form a spatial coordinate set, using a convex hull algorithm to calculate the minimum outer envelope of all points in the spatial coordinate set, and expanding outward by a preset safety distance based on the outer envelope to generate a danger zone map that needs to be avoided.

7. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 5, characterized in that, The step of searching for the shortest path from the current location to the target grab point using a path planning algorithm based on the danger zone map to form the finger insertion path includes: calculating the detour trajectory segment in the normal direction at the intersection point of the path and the danger zone, connecting the original path and the detour trajectory segment to obtain the finger insertion path.

8. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 5, characterized in that, The process of assessing the degree of contact by combining the finger insertion path with real-time contact force distribution, and dynamically adjusting the path offset distance to generate a proposed grasping path includes: Real-time contact force distribution data is collected along the finger insertion path. The contact force distribution data is matched with the spatial coordinates of the weak location points. The proportion of sensing units falling into the weak area is counted to obtain the contact degree ratio. The path offset is calculated based on the difference between the contact ratio and the preset safety threshold. The coordinates of the control points on the finger insertion path are corrected based on the path offset, and the corrected control point sequence is connected by an interpolation algorithm to generate the proposed grasping path.

9. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 1, characterized in that, The process of evaluating the deviation between the real-time opening size and the initial opening size based on the proposed crawling path, updating the hierarchical priority ranking based on the deviation between the real-time opening size and the initial opening size, and identifying the coordinates of the weak points to be crawled includes: For the control points on the proposed grasping path, calculate the difference between the real-time opening size and the initial opening size to obtain the deviation ratio; Update the gap width record table according to the deviation ratio; Based on the updated gap width record table, the weights of each level are recalculated to generate a dynamically updated level priority sort. Based on the change in the hierarchical priority sorting, the weak location identification is re-executed to obtain an updated set of weak location coordinates.

10. The intelligent control method for single-handed grasping tasks of a humanoid robot as described in claim 1, characterized in that, The process of determining the final grasping level and grasping path based on the coordinates of weak points, and guiding the humanoid robot to perform the grasping action, includes: Select the layer with the highest priority from the updated layer priority sort as the final crawling layer; Based on the coordinates of the weak points corresponding to the final crawling level, a crawling path is generated that bypasses the weak points. The humanoid robot's fingertips are guided to the target position via the grasping path. The closing force is adjusted according to the gap width of the final grasping level, and the closing action is performed to complete the grasping.

Images