A humanoid robot control method, device and equipment based on vector entanglement

Abstract

The application provides a humanoid robot control method, device and equipment based on vector entanglement, core action sequences are extracted by performing frequency statistical analysis on a control task; a basic action vector is constructed, a spiral entanglement action vector chain is generated in a partial overlap entanglement mode; an action semantic atlas is constructed by using dynamic semantic mapping and a deviation prediction path is generated; non-integer period actions are identified by a deviation feature vector, and an overlap node is extracted from the entanglement vector chain to generate a complete execution sequence by borrowing; a deviation propagation network is constructed to analyze a deviation diffusion path, identify a deviation amplification node and a convergence node, and extract available components to generate an action enhancement factor; the enhancement factor is optimized in cooperation with the entanglement vector chain to generate an enhanced control sequence, and the transformation from passive elimination to active utilization of the deviation is realized, and the continuity and energy efficiency of the action of the humanoid robot are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of humanoid robot control, in particular to a humanoid robot control method and device based on vector entanglement and equipment. BACKGROUND

[0002] As a highly bionic intelligent system, humanoid robots have broad application prospects in industrial manufacturing, service assistance, medical rehabilitation and other fields. However, the multi-joint coordinated control, complex motion planning and environmental adaptability of humanoid robots have always been the key technical bottlenecks restricting their practicality. Especially when performing precise operation tasks, small control deviations will be amplified through the kinematic chain, resulting in task failure or system instability.

[0003] Most existing humanoid robot control methods use traditional deviation elimination strategies, regarding any deviation from the predetermined trajectory as an error that needs to be corrected. This method ignores the beneficial effects that some deviations may bring, such as gravity deviation that can assist in descending motion, inertia deviation that can accelerate turning, etc. In addition, existing methods lack effective compensation mechanisms when dealing with non-complete periodic motions, making it difficult to achieve continuity and smoothness of motion sequences. Therefore, a method is needed to solve at least one of the above problems. SUMMARY

[0004] The present application provides a humanoid robot control method and device based on vector entanglement and equipment, aiming to realize the coupling and correlation between actions by constructing an entangled action vector chain, analyze the diffusion and convergence characteristics of deviations using a deviation propagation network, convert traditional deviations that need to be eliminated into available control resources, and complete non-integral periodic actions through borrowing mechanism, ultimately achieving intelligent and adaptive control of humanoid robots.

[0005] The first aspect of the present application provides a humanoid robot control method based on vector entanglement, comprising the following steps:

[0006] Receiving a humanoid robot control task, performing frequency statistical analysis on the control task to extract a core action sequence, constructing a basic action vector according to the core action sequence, and generating an entangled action vector chain through partial overlap entanglement using the basic action vector;

[0007] Performing dynamic semantic mapping using the entangled action vector chain to construct an initial action semantic atlas, evolving and updating the initial action semantic atlas according to real-time execution feedback, and generating a deviation prediction path through the updated action semantic atlas;

[0008] Performing humanoid robot actions according to the deviation prediction path to generate actual execution deviation data, comparing and analyzing the actual execution deviation data and the deviation prediction path to generate a deviation feature vector;

[0009] identify non-integral period actions using the deviation feature vector, extract entangled action vectors that overlap with the non-integral period actions from the entangled action vector chain, and generate a complete execution sequence by borrowing and fusing the entangled action vectors;

[0010] Construct a deviation propagation network combining the complete execution sequence and the deviation feature vector, analyze the diffusion path of deviation in the action chain through the deviation propagation network to obtain deviation amplification nodes and deviation convergence nodes, perform deviation regulation analysis on the deviation amplification nodes and the deviation convergence nodes to obtain available components, and generate an action enhancement factor according to the available components;

[0011] Co-optimize the action enhancement factor and the entangled action vector chain to generate an enhanced control sequence, and use the enhanced control sequence to complete the driving of the humanoid robot.

[0012] The second aspect of the present application proposes a humanoid robot control device based on vector entanglement, comprising:

[0013] A vector construction module is configured to receive a humanoid robot control task, perform frequency statistical analysis on the control task to extract a core action sequence, construct a basic action vector according to the core action sequence, and generate an entangled action vector chain by partial overlap entanglement using the basic action vector.

[0014] A semantic mapping module is configured to perform dynamic semantic mapping using the entangled action vector chain to construct an initial action semantic graph, evolve and update the initial action semantic graph according to real-time execution feedback, and generate a deviation prediction path through the updated action semantic graph.

[0015] A deviation analysis module is configured to perform humanoid robot actions according to the deviation prediction path to generate actual execution deviation data, compare and analyze the actual execution deviation data and the deviation prediction path to generate a deviation feature vector.

[0016] An action completion module is configured to identify non-integral period actions using the deviation feature vector, extract entangled action vectors that overlap with the non-integral period actions from the entangled action vector chain, and generate a complete execution sequence by borrowing and fusing the entangled action vectors.

[0017] A deviation enhancement module is configured to construct a deviation propagation network combining the complete execution sequence and the deviation feature vector, analyze the diffusion path of deviation in the action chain through the deviation propagation network to obtain deviation amplification nodes and deviation convergence nodes, perform deviation regulation analysis on the deviation amplification nodes and the deviation convergence nodes to obtain available components, and generate an action enhancement factor according to the available components.

[0018] A control optimization module is configured to optimize the action enhancement factor and the entangled action vector chain to generate an enhanced control sequence, and to drive the humanoid robot by using the enhanced control sequence.

[0019] The third aspect of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the humanoid robot control method based on vector entanglement disclosed in the first aspect when executing the program.

[0020] The beneficial effects of the present application are reflected in the following points: first, the present application realizes the structured expression and dynamic coupling of the action sequence by constructing a spiral entangled action vector chain. The entanglement quantifies the correlation strength of adjacent actions, and the spiral topology preserves the time sequence and spatial proximity, so that the humanoid robot can predict and adjust the subsequent action according to the current action state, improving the fluency and accuracy of multi-joint coordinated control, especially in complex action conversion, the continuity of motion can be maintained. Second, the present application realizes an active utilization mechanism of bias, identifies bias amplification nodes and convergence nodes through a bias propagation network, and converts traditional factors that need to be eliminated such as gravity bias and inertia bias into beneficial resources for auxiliary action execution, not only reducing energy consumption, but also making the humanoid robot action more natural and smooth, fully utilizing environmental forces and inertial forces in specific actions such as descending and turning. Finally, the problem of action interruption and non-integer period execution is effectively solved by the borrowing fusion mechanism, which realizes the intelligent completion of interrupted actions by dynamically extracting surplus force, speed and direction components from adjacent entangled nodes, ensuring the continuous execution of tasks, so that the humanoid robot can maintain the integrity of the action when facing sudden disturbances or abnormal situations, and enhance the task adaptability in complex environments.

[0021] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0022] The drawings herein show specific examples of the technical solutions described in the present application, and constitute part of the specification together with the specific embodiments, for explaining the technical solutions, principles and effects of the present application.

[0023] Unless specifically stated or defined otherwise, the same reference signs in different drawings represent the same or similar technical features, and different reference signs may also be used to represent the same or similar technical features.

[0024] Figure 1 is a flowchart of a humanoid robot control method based on vector entanglement.

[0025] Figure 2 is a structural block diagram of a humanoid robot control device based on vector entanglement according to the present application.

[0026] Figure 3 is a structural schematic diagram of a computer device according to the present application. DETAILED DESCRIPTION

[0027] In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular sequences of steps, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, circuits, and

[0028] It will be understood that the term "includes," "including," "has," "having," "comprises," "comprising" or "contains," "containing" when used in this specification and in the following claims, specifies the presence of stated features, integers, steps, operations, elements, or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or groups thereof.

[0029] It will also be understood that the term "and / or," when used in this specification and in the following claims, can be interpreted to mean one or more of the listed items can be present, alone or in combination, and that the combinations are included as possible combinations.

[0030] As used in this specification and in the claims, the term "if" can be interpreted as meaning "when," or "once," or "in response to determining," or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [a described condition or event] is detected" can be interpreted to mean "once it is determined," or "in response to determining," or "once [the described condition or event] is detected," or "in response to detecting [the described condition or event]," depending on the context.

[0031] In addition, the terms "first," "second," "third," etc. are used herein only to describe different instances of elements, and are not intended to imply or suggest relative importance of the elements.

[0032] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0033] The technical solutions of the embodiments of this application will be described below.

[0034] like Figure 1 As shown, this embodiment of the invention provides a humanoid robot control method based on vector entanglement, including the following steps S110-S160:

[0035] Step S110: Accept the humanoid robot control task, perform frequency statistical analysis on the control task to extract the core action sequence, construct basic action vectors based on the core action sequence, and generate an entangled action vector chain using the basic action vectors through partial overlap and entanglement.

[0036] The system receives task instructions from the host control system via a humanoid robot control interface, constructing a multi-layered task parsing framework. Control tasks are defined using a structured description language, including key elements such as task type, target object, execution requirements, and constraints. Task types cover typical application scenarios such as grasping operations, assembly operations, material handling, and precision positioning. The task parsing process first performs semantic understanding, converting natural language or symbolic task descriptions into machine-executable instruction sequences. For example, upon receiving the task "Move workpiece A from position P1 to position P2 and place it precisely," the system identifies the material handling task type and extracts parameters such as the starting position P1 (x1, y1, z1), the target position P2 (x2, y2, z2), and workpiece attributes (mass m = 2.5 kg, dimensions 200 × 150 × 80 mm). The task verification mechanism checks the completeness and feasibility of the instructions, including workspace accessibility verification, load capacity verification, and path collision-free detection. Task priorities are set based on urgency and importance, using a four-level classification: Urgent and Important (Priority 1), Urgent and Moderate (Priority 2), Routine and Important (Priority 3), and Routine and Moderate (Priority 4). Time constraints extract the start time, end time, maximum allowed execution time, and other time-series requirements for each task. An abnormal task identification mechanism filters out unreasonable or dangerous instructions, such as load requirements exceeding the humanoid robot's capabilities or action sequences that violate safety regulations.

[0037] Frequency statistical analysis is performed on the control tasks to extract core action sequences. Based on the received control tasks, a hierarchical decomposition strategy is adopted to refine complex tasks into action sets of different granularities. The multi-granularity decomposition follows the progressive principle of "whole-part-detail". Coarse-grained decomposition divides the task into major action stages, each of which represents a relatively independent functional unit, such as approaching, grasping, transferring, placing, and returning stages for the carrying task. Medium-grained decomposition further refines each coarse-grained action into specific movement sequences, and fine-grained decomposition reaches the joint movement level. The determination of decomposition granularity adopts the information entropy criterion H = -∑p(i)log(p(i)), and the decomposition stops when the H value changes less than 5%. For the three granularities of action sets generated, statistical analysis methods are used to identify high-frequency repeated action patterns. The coarse-grained frequency calculation formula is f_coarse(k) = n_k / N_total, where n_k is the number of occurrences of action k, and N_total is the total number of tasks. It is found that the "approaching-grasping-transferring-placing" sequence appears with a frequency of 85% in carrying tasks. Medium-grained analysis identifies action transition rules by calculating conditional probability P(a_j|a_i). When P(a_j|a_i) > 0.7, it indicates that there is a strong association between actions. Fine-grained statistics reveal the common characteristics of underlying control. Cluster analysis of joint angle sequences finds that 70% of the grasping actions use similar trajectory patterns. Cross-granularity association analysis establishes the mapping relationship of three layers of actions, and by considering the statistical results of each granularity, multi-level core action sequences are extracted.

[0038] According to the core action sequence, a basic action vector is constructed. Based on the extracted core action sequence, a multi-dimensional basic action vector is constructed to fully represent the action characteristics of the humanoid robot. The basic action vector is defined as a four-dimensional composite structure V_base=(F, V, D, M), where F is the force vector component, V is the velocity vector component, D is the direction vector component, and M is the memory component. The force vector component F=[f_x, f_y, f_z, τ_x, τ_y, τ_z] includes three-axis force and three-axis torque, which fully describes the mechanical characteristics of the action, such as the "precise grasp" action, which is represented by a vertical downward clamping force and a moderate wrist torque. The velocity vector component V=[v_linear, a_linear, v_angular, a_angular] covers the linear velocity, linear acceleration, angular velocity and angular acceleration of the end effector, reflecting the dynamic characteristics of the action. The direction vector component D is represented by a quaternion D=[q_w, q_x, q_y, q_z], which avoids the gimbal lock problem of Euler angles and ensures the continuity and uniqueness of the pose representation. The memory component M=[t_duration, success_rate, energy_cost, stability_index] records the average execution time, historical success rate, energy consumption index and stability score, so that the system can learn and optimize from historical execution. Vector normalization ensures the comparability of different dimensions, and a timestamp marks the position of each vector in the core sequence. Through multi-dimensional feature extraction and vector representation, a basic action vector is successfully constructed, which can fully reflect the essential characteristics of the action.

[0039] In some embodiments, the basic action vector is used to generate an entangled action vector chain by partially overlapping entanglement, including: calculating the vector angle and modulus difference between time-series adjacent vectors in the basic action vector to generate a similarity matrix; determining the vector overlap area according to the similarity matrix; performing nonlinear weighted fusion on the vector components in the overlap area to generate entanglement coefficients; and using the entanglement coefficients to construct an entangled action vector chain by spiral connection.

[0040] With the constructed basic motion vectors, the similarity measure between time-adjacent vectors is calculated to generate a similarity matrix. Time-adjacent is defined as two consecutive actions in the core motion sequence, and their correlation needs to be evaluated. The vector angle uses cosine similarity calculation: cos(θ)=(V_i·V_j) / (||V_i||×||V_j||), where V_i and V_j are adjacent basic motion vectors, θ is the angle between the two vectors, · represents vector dot product, and ||·|| represents vector length. Since the basic motion vector contains multiple components, the angle of each component is calculated and then weighted. The length difference reflects the change in action strength: Δ_mag=|||V_i||-||V_j||| / max(||V_i||,||V_j||), Δ_mag is the normalized length difference. For example, in the transition from "quick approach" to "precise positioning", the speed component decreases significantly, indicating a change in action mode. The similarity comprehensive index S_ij integrates direction consistency and strength matching. The similarity matrix S is a symmetric matrix, and the element S_ij represents the similarity between action i and action j, with a value range of [0, 1]. The structural characteristics of the matrix reflect the inherent correlation pattern of the action sequence. The calculation of the similarity matrix not only considers the vector relationship at a single time, but also captures the dynamic change trend in a short time through a sliding window. The sparsity of the matrix reflects the modularization degree of the action sequence, and high sparsity indicates that the actions are relatively independent.

[0041] Based on the generated similarity matrix, the overlap region between adjacent vectors is determined. The determination of the overlap region is based on the similarity threshold and the action duration. When the similarity exceeds the preset threshold, it is considered that there is an overlap. The time range of the overlap is dynamically determined according to the action duration and the similarity size, and the higher the similarity, the larger the overlap ratio. Spatial overlap is determined by the intersection of the workspaces of the two actions. The overlap degree of the vector components is different, and the overlap characteristics of the force component, speed component, and direction component are different, which need to be processed separately. For example, two consecutive operation actions may be highly consistent in the direction of force, but there is a significant difference in the speed curve. The features of the overlap region include key parameters such as overlap center and overlap gradient. Boundary condition processing ensures smooth transition of the overlap region and avoids abrupt changes. Abnormal overlap detection identifies unreasonable overlap patterns, such as completely opposite directions but high similarity. During the identification of the overlap region, special attention is paid to the key moments of action transition, which often contain rich coordination information. The data density in the overlap region is usually higher than that in the non-overlap region, which requires more detailed sampling and processing.

[0042] In the determined overlapping region, the vector components are nonlinearly weighted and fused to generate the entanglement coefficient representing the correlation strength of the action. The calculation of the entanglement coefficient uses a comprehensive evaluation model: C_ij=α×S_ij×O_ij×exp(-β×Δt_ij), where C_ij is the entanglement coefficient between vectors i and j, with a value range of [0, 1]; S_ij is the similarity index calculated as described above, which is obtained based on the vector angle and the difference in module length; O_ij is the coverage of the overlapping region, reflecting the degree of overlap of the two actions in space and time, with a value of [0, 1]; Δt_ij is the time interval |t_i-t_j|, representing the time difference between adjacent actions; α is the weighting coefficient, usually taking a value of 0.9; β is the time decay parameter, taking a value of 0.1, controlling the influence of time distance on entanglement strength. The nonlinear fusion uses exponential decay characteristics to avoid the limitations of simple linear combination. The force component fusion considers the continuity constraint of force to ensure that there is no sudden change in force during the transition process. The velocity component fusion uses spline interpolation to ensure the smoothness of the motion. The direction component fusion uses quaternion spherical interpolation to maintain the continuity of the attitude change. The memory component fusion uses a probability updating method to integrate historical execution information. The coefficient normalization ensures the comparability of different action pairs. The key of nonlinear fusion is to maintain the physical feasibility of the action, and the fusion result must satisfy the kinematics and dynamics constraints of humanoid robots.

[0043] For example, the entanglement coefficient is used to construct an entangled action vector chain in a spiral connection mode, including: determining the radial distribution of the spiral according to the force vector component in the basic action vector; determining the tangential velocity of the spiral using the velocity vector component in the basic action vector; generating a spiral trajectory equation by combining the radial distribution and the tangential velocity; calculating the spatial position of each vector according to the spiral trajectory equation and the entanglement coefficient; and forming a spiral entangled action vector chain by using the spatial position and the entanglement coefficient for weighted connection.

[0044] According to the force vector component in the base motion vector, the radial distribution of the helix structure is determined. The size of the force vector component directly affects the helix radius, and a mapping relationship r(t) = r_0 + k_F x ||F(t)|| is established, where r(t) is the helix radius at time t, r_0 is the base radius, k_F is the force-radius mapping coefficient, and ||F(t)|| is the force vector length. The radial distribution reflects the change in the mechanical strength of the motion, and large force motions correspond to larger helix radii, while fine operations correspond to smaller radii. For example, the helix radius of a heavy lifting motion is significantly larger than that of a precision assembly motion. The radial change rate dr / dt reflects the dynamic adjustment process of the force, and rapid changes indicate the switching of the motion mode. The continuity of the radial distribution is ensured by a smoothing function, avoiding structural instability caused by sudden changes in the radius. The limitation of abnormal radial values prevents extreme cases, setting boundaries for the maximum and minimum radii. Statistical analysis shows that the radial distribution of typical operation tasks presents a bimodal feature, corresponding to the two main modes of force control and position control. The physical significance of the radial distribution is to intuitively reflect the working strength of the humanoid robot, and the operator can understand the current load state by observing the expansion and contraction of the helix. Radial mutation points often correspond to key switching moments of the task, which need special attention.

[0045] Using the velocity vector component in the base motion vector, the tangential motion characteristics of the helix are determined. The velocity vector component controls the winding speed of the helix, and a relationship v_θ(t) = ω_0 + k_V x ||V(t)|| is established, where v_θ(t) is the tangential angular velocity, ω_0 is the base angular velocity, k_V is the velocity mapping coefficient, and ||V(t)|| is the velocity vector length. The tangential velocity determines the density of the helix, with high-speed motion producing sparse helices and low-speed motion producing dense helices. For example, the helix spacing is larger during the fast movement stage, and the helix spacing is tightened during the precise positioning stage. The change in angular velocity reflects the urgency and precision requirements of the motion. The acceleration component affects the rate of change of the tangential velocity, which is used to predict the velocity trend. The upper and lower limits of the tangential velocity prevent excessive winding or loosening. The direction information of the velocity vector is used to determine the rotation direction of the helix, with positive velocity producing right-handed rotation and negative velocity producing left-handed rotation. A dynamic adjustment mechanism corrects the tangential velocity based on real-time feedback. The change pattern of the tangential velocity is closely related to the task rhythm, with uniform tangential velocity indicating a stable execution process, and frequent changes indicating task difficulty or environmental interference. The change in the density of the helix provides a natural segmentation marker for the motion sequence. Through the conversion of the velocity vector to the tangential motion, the tangential velocity of the helix is determined.

[0046] The complete helical trajectory equation is generated in combination with the determined radial distribution and tangential velocity. The helical trajectory is represented in a parametric form: x(t) = r(t)cos(0(t)), y(t) = r(t)sin(0(t)), z(t) = h(t), where x, y, z are Cartesian coordinates, 0(t) is the cumulative phase angle, and h(t) is the axial height. The phase angle is obtained by integrating the tangential velocity: 0(t) = J v_0(t)d t. The axial height h(t) is linearly related to the execution progress and is modulated by the memory component. The continuity of the trajectory is guaranteed by the differentiability of the parametric equation. The curvature and torsion calculations are used to evaluate the geometric properties of the trajectory. For example, the helical trajectory generated for a complete grasp-pick-place sequence exhibits the feature of expanding first and then contracting. The trajectory optimization reduces unnecessary bending by the principle of energy minimization. The boundary conditions set the start and end points of the trajectory to ensure consistency with the task requirements. The discretization of the parametric equation supports numerical calculation and visualization. The generation process of the helical trajectory fully considers the motion constraints of the humanoid robot, ensuring that the generated trajectory is reachable in the joint space. The geometric features of the trajectory are related to the complexity of the task, with regular helices corresponding to simple tasks and irregular spatial curves for complex tasks.

[0047] According to the generated helical trajectory equation and the entanglement coefficient, the spatial positions of each basic motion vector in three-dimensional space are calculated. The calculation of the entanglement coefficient considers the similarity, overlap degree, and time decay characteristics: C_ij = a x S_ij x O_ij x exp(-b x At_ij), where S_ij is the similarity index calculated above, O_ij is the coverage degree of the overlapping area, At_ij is the time interval, a = 0.9 and b = 0.1 are empirical parameters. Spatial positioning is achieved by substituting discrete time points into the trajectory equation, and the spatial coordinates P_i = (x_i, y_i, z_i) of the ith vector at time t_i are obtained. The entanglement coefficient affects the relative positions of the vectors, and a high entanglement coefficient leads to close spatial positions, forming a tight entanglement. The position adjustment algorithm ensures that the distance between adjacent vectors meets the entanglement strength: d_ij = d_base x (1 - C_ij), where d_ij is the spatial distance between vectors i and j, and d_base is the baseline distance. For example, when C_ij = 0.8, the actual distance d_ij = 0.2 x d_base, indicating that the strongly entangled action pair is close to overlapping on the helix; when C_ij = 0.2, d_ij = 0.8 x d_base, the weakly entangled action pair maintains a significant interval. The uniformity of the spatial distribution is achieved through iterative optimization to avoid local over-concentration. The distribution pattern of the spatial positions reflects the internal structure of the action sequence, with dense areas corresponding to highly coordinated action groups and sparse areas representing relatively independent actions. The conversion relationship between the base coordinate system of the humanoid robot and the world coordinate system is considered during the position calculation process.

[0048] The complete helical entangled action vector chain is formed by establishing the weighted connection relationship between vectors through spatial position and entanglement coefficient. The connection is built based on neighborhood relationship and entanglement strength. The vectors at adjacent positions are connected by the entanglement coefficient C_ij, and the connection weight is directly equal to the entanglement coefficient value. The connection topology adopts a directed graph structure, the nodes are basic action vectors at spatial positions, and the edges are weighted connections. The direction of the edge represents the time sequence relationship of action execution. Strong entanglement connection (C_ij>0.7) forms the main link, which embodies the close coupling between actions. Medium entanglement connection (0.3≤C_ij≤0.7) forms an auxiliary link, which provides coordination support between actions. Weak entanglement connection (C_ij<0.3) serves as a supplementary link, which maintains the integrity of the action sequence. The helical connection is characterized by maintaining the continuity of the time sequence and the proximity of the space. The actions at adjacent time points are physically close on the helix, which facilitates information transmission and state sharing. The time-varying characteristics of the connection weight allow the entanglement strength to be dynamically adjusted according to the execution state, realizing adaptive link management. For example, in a precision assembly task, the high entanglement coefficient 0.85 of the "alignment" and "insertion" actions forms a strong connection, ensuring the close coordination of the two actions. Through weighted connection based on the entanglement coefficient and the organization of the helical topology, a complete helical entangled action vector chain is finally constructed.

[0049] In step S120, the initial action semantic graph is constructed by dynamic semantic mapping using the entangled action vector chain. The initial action semantic graph is evolved and updated according to real-time execution feedback, and the deviation prediction path is generated based on the updated action semantic graph.

[0050] Specifically, the initial action semantic graph is constructed by dynamic semantic mapping using the entangled action vector chain. The helical connection and entanglement coefficient in the entangled action vector chain reflect the time sequence relationship and coupling strength of the actions, which need to be converted into a semantic network that can be understood and analyzed. Semantic mapping first identifies the key nodes in the vector chain, i.e. the positions with high entanglement coefficients. These positions usually correspond to the transition points of action patterns. The semantic labeling of the nodes is based on the dominant component of the basic action vector. When the force component is dominant, it is labeled as "force control type"; when the velocity component is dominant, it is labeled as "motion type"; and when the direction component is dominant, it is labeled as "positioning type". The construction of edges directly uses the entanglement coefficient. When the entanglement coefficient of two nodes exceeds a certain threshold, a connection edge is established, and the edge weight is equal to the entanglement coefficient value. The graph adopts a directed graph structure G=(V,E), where V is the set of action nodes, E is the set of transition edges, and the direction represents the execution order of the actions. For example, in a precision assembly task, the high entanglement coefficient between the "alignment" node and the "insertion" node forms a strong association edge, reflecting the close coordination relationship between the two actions. The initial graph also contains the hierarchical information of the actions, with coarse-grained actions as parent nodes and fine-grained actions as child nodes, forming a multi-level semantic structure.

[0051] According to the feedback data in the actual execution process of the humanoid robot, the initial action semantic graph is dynamically updated to reflect the real execution characteristics. Real-time execution feedback includes multi-dimensional data such as action completion time, execution accuracy, force control error, and energy consumption. These data are used to adjust the node attributes and edge weights in the graph. The node attribute update adopts the exponential weighted moving average method, and the new execution data is weighted and fused with the historical data to ensure that the graph can adapt to the changes in execution characteristics. The adjustment of the edge weight reflects the actual effect of action transfer. When the execution effect of a certain action sequence is better than expected, the weight of the corresponding edge increases; otherwise, it decreases. The discovery mechanism of new nodes identifies new action patterns that appear during execution. When a certain unrecorded action combination repeatedly appears and the effect is stable, it is added to the graph as a new node. The topology optimization of the graph includes removing long-unused nodes and low-weight edges to maintain the simplicity and efficiency of the structure. The evolution rate adopts an adaptive strategy, with fast learning at the beginning of system operation and gradually reducing the update frequency as experience accumulates. Version control records the evolution history of the graph, supporting performance comparison and abnormal tracing. The update process maintains the connectivity and integrity of the graph, avoiding isolated nodes or broken paths. Through incremental update and structure optimization driven by real-time feedback, the evolution and update of the action semantic graph are realized.

[0052] In some embodiments, the bias prediction path generated by the updated action semantic graph comprises: extracting historical bias nodes from the updated action semantic graph to construct a bias evolution subgraph; calculating the bias transfer probability between nodes using the bias evolution subgraph to generate a bias state matrix; generating a time series bias sequence based on the bias state matrix; and mapping the time series bias sequence to an execution time axis to form a bias prediction path.

[0053] From the updated action semantic graph, nodes with significant bias are identified, and a subgraph dedicated to bias analysis is constructed. The identification of bias nodes is based on error statistics in execution feedback. When the position error, force error, or time deviation of a certain action exceeds the allowable range, the node is marked as a bias node. The types of bias are divided into systematic bias and random bias. Systematic bias shows persistent one-way error, and random bias shows fluctuating error. The bias subgraph not only includes the bias nodes themselves, but also includes their predecessor and successor nodes, which are used to analyze the causes and propagation effects of bias. The construction of the subgraph retains the connection relationships and weight information between these nodes in the original graph, while adding bias-related attribute labels. Bias intensity is normalized to unify the dimension, facilitating the comparison and analysis of different types of bias. Time markers record the typical time and duration of bias occurrence, which helps to predict the pattern of bias occurrence.

[0054] The transfer probability of bias between nodes is calculated by statistical analysis of historical data in the subgraph of bias evolution. The bias transfer probability P_ij represents the probability that node j will also be biased given that node i is biased. It is calculated by the conditional probability formula P_ij = N_ij / N_i, where N_ij is the number of times node j is also biased after node i is biased, and N_i is the total number of times node i is biased. The bias state is divided into multiple levels, from no bias to severe bias. Each node is in one of these states at a certain time. State transition not only considers the direct influence between adjacent nodes, but also considers the cumulative effect and decay characteristics of bias. The matrix construction process ensures row normalization, meeting the mathematical properties of probability matrices. Sparse processing retains the main transition path and ignores the transition relationship with extremely small probability. Time-varying characteristics are reflected through time-based statistics, and the transition probabilities of different task stages may differ. The stability of the system is verified by the characteristic analysis of the matrix, and the maximum eigenvalue less than 1 indicates that the bias will not be amplified indefinitely.

[0055] Based on the bias state matrix, a multi-step Markov prediction is performed to generate the time series bias sequence in the future time period. The prediction starts from the current bias state and calculates the state probability distribution at each future time step through the power operation of the state transition matrix. The prediction step is determined according to the task characteristics, usually choosing the number of steps that can cover a complete action sequence. The initial state vector is set according to the latest bias detection results, which can be a single state or a probability distribution. As the prediction step increases, uncertainty accumulates, probability distribution tends to spread, and the confidence of long-term prediction decreases accordingly. The most likely path is determined by the dynamic programming algorithm, which gives the state evolution trajectory with the highest probability. Branch prediction considers multiple possible evolution scenarios, providing comprehensive information for risk assessment. The prediction results include the probability values of each state at each time, forming a complete time series bias sequence. The identification of key turning points helps to determine the timing of intervention, and when the prediction shows that the bias will increase sharply, it needs to be adjusted in advance.

[0056] The discrete time deviation sequence is mapped to a continuous execution time axis to form a deviation prediction path that can be used for real-time control. The time mapping requires the typical execution duration of each action in the semantic atlas, which has been statistically completed during the atlas update process. The execution time of each action has some fluctuations, and the statistical average value is taken as the nominal duration. The distribution of deviation intensity on the time axis is continuous by interpolation method, and cubic spline interpolation is commonly used to ensure the smoothness of the curve. The labeling of key time points includes the deviation start time, peak time, and recovery time, which are crucial for control decisions. The confidence interval of the path reflects the uncertainty of the prediction, which is represented by the upper and lower boundary curves. The sliding update of the time window supports online prediction, and each control cycle is based on the latest data to refresh the prediction path. Abnormal events are marked as special in the path to alert the system to potential risks. The compressed storage of path data only retains key turning points and feature points, reducing storage and computing overhead.

[0057] In step S130, actual execution deviation data is generated by executing humanoid robot actions according to the deviation prediction path, and a deviation feature vector is generated by comparing and analyzing the actual execution deviation data and the deviation prediction path.

[0058] Specifically, actual execution deviation data is generated by executing humanoid robot actions according to the deviation prediction path. The deviation prediction path provides expected deviation values at each time point and key time nodes, including deviation start time, peak time, and recovery time, which need to be monitored during the execution process. The humanoid robot executes the action sequence defined in the entangled action vector chain while adjusting the control parameters according to the prediction path. When the prediction path shows that a position deviation peak will occur at a certain time, the control system adjusts the position loop gain in advance to prepare for compensation. High-frequency sampling is used for data collection during execution, and the sampling frequency of key parameters such as position, force, and velocity is dynamically adjusted according to the type of action. For example, in the precision assembly stage, the position sampling frequency is increased to 1 kHz to capture small deviations; in the fast moving stage, the focus is on monitoring the changes in speed and acceleration. Real-time calculation of deviation data uses the reference trajectory comparison method, actual deviation Δ(t) = X_actual(t) - X_reference(t), where X represents the state vector, including position, attitude, force, and other multi-dimensional information. Data caching mechanism ensures complete time series record, including normal execution data and abnormal mutation data. The guiding role of the prediction path is reflected in the proactive adjustment of the control strategy, rather than passive response.

[0059] The actual execution deviation data and the deviation prediction path are compared and analyzed to generate a deviation feature vector. The comparison and analysis are carried out in two dimensions of time and amplitude. First, time alignment is performed to ensure that the prediction data and the actual data are compared on the same time reference. The key time points marked in the prediction path are matched with the actual deviation evolution process, and the time error At = t_actual - t predicted is calculated. The comparison of the deviation amplitude uses a combination of point-to-point difference calculation and statistical analysis. Not only the instantaneous error is concerned, but also the distribution characteristics of the error are analyzed. The prediction accuracy evaluation indicators include mean absolute error MAE, root mean square error RMSE, and correlation coefficient, etc. The comparison of the deviation evolution mode identifies the trend differences between the prediction and the actual, such as the prediction of gradual growth but the actual of stepwise mutation. Feature extraction is carried out from multiple angles: dynamic features include deviation growth rate, peak value characteristics, and duration; statistical features include mean, variance, skewness, and kurtosis; frequency domain features are obtained by FFT analysis to obtain the main frequency component; and correlation features reflect the coupling relationship between different types of deviations. Abnormal deviations are identified that are not covered by the prediction path. These information has important value for model improvement. The deviation feature vector V_dev = [f_1, f_2,..., f_n] is constructed by integrating all the extracted features, where f_i represents the i-th feature component. Feature normalization ensures the comparability of different dimensional features.

[0060] In step S140, the non-integer period action is identified using the deviation feature vector, and the entangled action vector that overlaps with the non-integer period action is extracted from the entangled action vector chain. The entangled action vector is fused by borrowing place to generate a complete execution sequence.

[0061] Specifically, the non-integer period action is identified using the deviation feature vector. The deviation feature vector contains multi-dimensional information such as dynamic features, statistical features, and frequency domain features. These features collectively reveal the integrity state of the action execution. The determination of the non-integer period action is based on multiple indicators: when the duration feature in the deviation feature vector is significantly shorter than the expected value, the deviation growth rate shows abnormal truncation, and the frequency domain feature shows the absence of the fundamental frequency component, it indicates that the action has not completed a complete period. For example, in a precision insertion task, the standard execution time of the "insertion" action is 2.5 seconds, but the deviation feature vector shows that the actual execution time is only 1.8 seconds, the phase analysis shows that the termination phase φ_end = 1.45π < 2π, and the force feature shows a sharp rise, indicating that the action is interrupted due to abnormal resistance. The cause analysis of the interruption is identified through the mutation mode of the deviation feature, which distinguishes between active protection interruption and passive interference interruption. The time stamp marker records the precise time of the interruption and the proportion of the execution completed.

[0062] The entangled action vector chain is extracted from the entangled action vector chain. The spiral structure of the entangled action vector chain preserves the spatial relationship and entanglement strength between nodes, providing a topological basis for finding overlapping nodes. The determination of the overlapping relationship considers multiple dimensions: the entanglement coefficient reflects the coupling strength between nodes, and a high entanglement coefficient indicates a close functional relationship; the spatial distance is measured by the arc length on the spiral trajectory, and nodes with similar functions are more likely to share resources; the functional similarity is evaluated by comparing the components of the basic action vector, and similar vector components are suitable for complementarity. After determining the position of the non-integer period action in the vector chain, the search algorithm searches for candidate overlapping nodes in its neighborhood. The search range not only includes the directly adjacent predecessor and successor nodes, but also extends to the spatially adjacent nodes in the spiral structure. The timing compatibility ensures that the overlapping nodes do not conflict in time.

[0063] In some embodiments, the borrowing and fusion of the entangled action vector generates a complete execution sequence, including: locating adjacent entangled nodes in the entangled action vector chain according to the non-integer period action; extracting borrowable vector components from the adjacent entangled nodes to generate a compensation vector set; time-aligning the compensation vector set and the non-integer period action to generate a fusion weight matrix; and performing vector fusion operations through the fusion weight matrix to generate a complete execution sequence.

[0064] According to the non-integer period action, the adjacent entangled nodes are located in the entangled action vector chain. The data structure of the entangled vector chain records the index, spatial coordinates and connection relationship of each node, and quickly locates the non-integer period action node through index lookup. The definition of adjacent nodes is not limited to the sequence relationship before and after, but also includes the geometric proximity in the spiral space. The neighborhood search algorithm takes the non-integer period node as the center and expands the search range in order of increasing distance. The first-order neighbor includes the direct predecessor and successor, which can be directly accessed through the linked list structure; the second-order neighbor is indirectly obtained through the connection relationship of the first-order neighbor; the spatial neighbor is identified by calculating the Euclidean distance on the spiral trajectory. The entanglement strength is used as the weight of the adjacency relationship, and the nodes with strong entanglement are preferentially included in the adjacent set. The node state check ensures the availability of the adjacent nodes. The size of the adjacent node set is dynamically determined according to the degree of missing of the non-integer period action.

[0065] For example, the extraction of borrowable vector components from the adjacent entangled nodes to generate a compensation vector set includes: analyzing the vector richness of the adjacent entangled nodes to identify borrowable components with a richness exceeding a threshold; determining the borrowing priority for the missing type of the non-integer period action, with force components being preferred to velocity components, and velocity components being preferred to direction components; constructing a borrowing weight based on the borrowing priority and the entangled action vector chain, with closer nodes having a larger borrowing weight; and dynamically extracting vector components from the borrowable components to form a compensation vector set according to the borrowing weight.

[0066] The located adjacent entangled nodes are analyzed to evaluate the use state and surplus degree of each component in the base action vector V_base=(F, V, D, M). The surplus degree is calculated based on the ratio of the current value of the component to the upper limit of the capacity. When the use rate of a certain component is lower than the set threshold, it is considered that there is surplus that can be borrowed. For example, the adjacent "precise positioning" node has just completed execution, and the force component F still has a residual value of 30% available for borrowing, and the speed component V has a kinetic energy surplus of 50% due to inertia; while the "force control adjustment" node to be executed, all its force components are in reserve state, and the surplus degree is 90%. The dynamic characteristics of the surplus degree consider the decay law of the component. The force component decays faster, and the direction component is relatively stable. The time-varying surplus degree is updated in real time to ensure that the borrowing decision is based on the latest state. The safety margin is reserved to prevent excessive borrowing from affecting the node's own function. After comprehensive evaluation of the resource state of adjacent nodes, the borrowable components with surplus degree exceeding the threshold are identified.

[0067] According to the specific missing situation of non-integer period action, the borrowing priority strategy of vector component is formulated. The missing type analysis is based on the state at the time of interruption. If the force control is insufficient, the force component demand is the highest, and if the trajectory deviates, the direction component is more important. The basic principle of borrowing priority is that the force component is prior to the speed component, and the speed component is prior to the direction component, which reflects the importance of the decreasing execution stability. Dynamic adjustment of priority according to task characteristics, precise assembly task increases the weight of force component, fast moving task increases the weight of speed component. The memory component is usually not involved in borrowing as auxiliary information to maintain the independence of each action. Priority quantization uses normalized weight to ensure the comparability of different components. Feedback adjustment mechanism corrects the priority according to the borrowing effect. Based on task-oriented analysis and dynamic adjustment, the borrowing priority is determined.

[0068] Based on the borrowing priority and the topological relationship between nodes, the quantitative borrowing weight is obtained. The aforementioned borrowing priority "force component is prior to velocity component, velocity component is prior to direction component" needs to be converted into a numerical priority weight P_priority, which is assigned in a decreasing equal difference manner. The force component obtains the highest weight value, the velocity component is next, and the direction component is the lowest, ensuring that the weight difference can reflect the priority relationship. Key parameters required for weight construction are extracted from the entangled action vector chain: for adjacent nodes, the entanglement coefficient C_ij uses the value determined when the helical connection is constructed; for non-adjacent nodes, the Euclidean distance or helical arc length distance d_ij between nodes is recalculated based on the spatial coordinates P_i and P_j recorded in the vector chain. Borrowing weight considers three factors: priority weight reflects the importance of components, distance decay ensures priority borrowing from adjacent nodes, and entanglement strength reflects the functional correlation between nodes. Distance decay adopts an exponential form exp(-d_ij / σ), which makes the contribution of remote nodes decrease rapidly; entanglement strength amplifies the role of high-coupling nodes through a power function C_ij^γ. The comprehensive weight calculation formula is W_ij=P_priority×exp(-d_ij / σ)×C_ij^γ, and then normalized in the standard interval to form a comparable quantitative index. The weight matrix is organized according to the two-dimensional structure of node-component, with rows corresponding to adjacent nodes and columns corresponding to vector component types.

[0069] According to the borrowing weight, the required vector component is extracted from the available component to form the compensation set. The extraction process uses a greedy algorithm, which starts with the highest weight node-component combination. The determination of the extraction amount is based on the degree of non-integer period action loss and the node's surplus capacity, following the principle of minimal sufficiency. The extraction of force components considers direction consistency, the extraction of velocity components maintains motion continuity, and the extraction of direction components avoids abrupt changes. The extraction order follows the weight ordering queue, and the remaining demand and node surplus state are updated after each extraction. Conflict detection is implemented during the extraction process. When multiple non-integer period actions request the same node resource at the same time, an arbitration mechanism is used to coordinate allocation. The multi-source extraction strategy allows different types of components to be obtained from different nodes, achieving complementary advantages. The constraints of the extraction process ensure that the normal function of the providing node is not affected. The compensation vector set is organized in a structured form, recording the source, quantity, and timeliness of each component. After systematic extraction based on weight, the extracted vector components are arranged in time sequence to generate the compensation vector V_compensate(t), where t represents the time point of the compensation action. This vector contains force, velocity, direction, and other component information borrowed from multiple adjacent nodes.

[0070] The extracted compensation vector set is precisely aligned with the non-integer period action in the time dimension to construct a fusion weight matrix. The core of time alignment is to determine the action law of the compensation component on the time axis. The completed part of the non-integer period action remains unchanged, and only the missing part after interruption is compensated. The alignment process divides the compensation period into three stages: transition period, fusion period and stable period. The design of the fusion weight matrix W(t) considers the time-varying characteristics, and the weight function adopts an S-shaped curve w(t)=1 / (1+exp(-k(t-t_c))), where t_c is the switching center time, and k controls the transition speed. The synchronization of the time reference is realized through the unified clock stamp, which ensures that the original action and the compensation component are aligned in the same time reference system. The parameters of the weight function are adjusted adaptively according to the type of action, and slow transition is adopted for precise action, and steep switching is allowed for fast action. The rows of the matrix correspond to different vector components, and the columns correspond to time sampling points. The boundary condition ensures the continuity of the weight at the interruption time.

[0071] The original non-integer period vector and the compensation vector set are subjected to weighted fusion operation through the fusion weight matrix. The fusion operation adopts a matrix form V_complete(t)=W(t)×[V_original(t);V_compensate(t)], where V_original(t) is the completed part of the original non-integer period vector, V_compensate(t) is the compensation vector extracted from the adjacent node, [V_original(t);V_compensate(t)] represents the vertical splicing of the two vectors to form an extended input vector, and W(t) is the corresponding time-varying fusion weight matrix. The fusion process maintains the continuity of the action, and the part before the interruption point completely retains the original vector, and the smooth transition is realized through weighted combination after the interruption point. Independent fusion at the component level ensures that the optimization of each dimension does not interfere with each other. The fusion calculation adopts an incremental update method, which only processes the fusion at the current time at each time step, reducing the computational complexity. The fused vector sequence covers the complete action period, restoring the interrupted function. According to the fusion operation and optimization processing of the system, the final complete execution sequence is generated.

[0072] In step S150, the deviation propagation network is constructed by combining the complete execution sequence and the deviation feature vector, the deviation amplification node and the deviation convergence node are obtained by analyzing the diffusion path of the deviation in the action chain through the deviation propagation network, the available components are obtained by analyzing the deviation regulation of the deviation amplification node and the deviation convergence node, and the action enhancement factor is generated according to the available components.

[0073] Specifically, the deviation propagation network is constructed by combining the complete execution sequence and the deviation feature vector. The complete execution sequence provides the action flow after the borrow fusion repair, in which each action node is arranged in the execution order, and these actions are directly converted into a node set of the network, and the total number of nodes is equal to the number of actions in the sequence. The deviation feature vector V_dev records the quantitative characteristics of various types of deviations, from which key components such as the deviation growth rate r_dev and the maximum deviation D_max are extracted for determining node attributes, where r_dev is obtained by calculating the first derivative of the deviation time series, and D_max is the peak value in the deviation sequence. The edges of the network are established according to the time sequence relationship of the execution sequence, and directed edges are established between adjacent actions, and the edge weight w_ij is determined by calculating the correlation coefficient of the deviations at adjacent times in the deviation feature vector, w_ij = p(V_dev(i), V_dev(j)), where p is the correlation coefficient. For indirect effects across actions, when the deviation feature vector shows that the deviations of certain two non-adjacent actions are highly correlated (p>0.7), a jump connection edge is added. For example, the three actions of "grabbing, carrying, and placing" in the complete execution sequence are converted into three network nodes, and the deviation feature vector shows that the correlation coefficient of the force deviation during grabbing and the placement accuracy is 0.82, so a direct connection with a weight of 0.82 is established between the grabbing and placing nodes.

[0074] In some embodiments, the obtaining of the deviation amplification node and the deviation convergence node by analyzing the diffusion path of the deviation in the action chain through the deviation propagation network comprises: inputting the deviation feature vector as an initial excitation to the deviation propagation network; calculating the cumulative intensity and the propagation speed of the deviation feature vector at each node through a network propagation algorithm; identifying the node with continuously increasing cumulative intensity and accelerating propagation speed as a deviation amplification node; and identifying the node with gradually decaying cumulative intensity and slowing down propagation speed as a deviation convergence node.

[0075] The bias feature vector is injected as an initial excitation signal into the corresponding node of the bias propagation network. Each component in the bias feature vector is mapped to the source node in the network that produces the bias as the starting state of the propagation analysis. The excitation strength is determined according to the severity of the bias, and the peak bias, growth rate, and other indicators in the bias feature vector are directly converted into excitation amplitude. The injection method considers the spatial and temporal distribution of the bias, and instantaneous bias uses pulse excitation, and sustained bias uses step excitation. In the case of multiple source excitation, different bias sources may produce coherent superposition or destructive interference effects. The timing characteristics of the excitation preserve the dynamic evolution information of the original bias. The initial state setting of the network ensures that all non-excited nodes are in the baseline state. The boundary conditions of excitation propagation are set according to task constraints to prevent the analysis results from exceeding a reasonable range. The calibration process of the initial excitation verifies the integrity and accuracy of the input signal. After the mapping and parameter setting of the system, the conversion of the bias feature vector to network excitation is completed.

[0076] The network propagation algorithm is used to analyze the dynamic evolution process of the bias excitation at each node. Based on the signal flow graph theory, the initial excitation injected as described above is used as the input signal of the network. At the initial time, the bias value of the excitation node D_in(i,0) is equal to the injected excitation strength, and the other nodes D_in(j,0)=0. The propagation process is carried out in discrete time steps, and the bias output of node i at the k+1 step is calculated as follows: D_out(i,k+1)=f_i(Σw_ji×D_out(j,k)), where D_out(j,k) is the output of node j at the k step (also the input of the downstream node), f_i is the transfer function of node i, and w_ji is the connection weight from node j to i. In this way, the initial excitation propagates through the network connections to form the dynamic evolution of the bias. The calculation of the cumulative strength uses the integral form: S_i(t)=S_i(0)+∫D_i(τ)dτ, where S_i(0) is the initial strength of the excitation node, equal to the corresponding component value of the bias feature vector, and S_i(0)=0 for non-excited nodes. The propagation speed v_i is determined by calculating the peak bias arrival time, v_i=d_i / Δt_i, where d_i is the topological distance between nodes, and Δt_i is the time delay of the bias peak propagating from the upstream node to node i. The cumulative strength and propagation speed data of each node are finally obtained.

[0077] Based on the characteristics of cumulative intensity and propagation velocity, the deviation amplification nodes in the network are identified. The deviation amplification nodes are characterized by output deviation greater than input deviation, and this amplification effect increases with time. The identification criteria include: the first derivative of cumulative intensity dS_i / dt>0 is positive for a long time, indicating that the deviation is growing; the acceleration of propagation velocity d²v_i / dt²>0, indicating that the influence range is rapidly expanding; the transfer function gain of the node G_i=S_i(t_out) / S_i(t_in)>1, indicating that the node has inherent amplification characteristics. Typical deviation amplification nodes appear in gait switching, center of gravity shift, and dual-arm coordination. For example, during the dynamic walking of a humanoid robot, a small angle deviation of the supporting ankle joint will be transmitted upward through the kinematic chain and amplified several times at the torso position, eventually leading to a loss of balance; when the dual arms cooperate to carry heavy objects, the force control deviation of the single arm will disrupt the coordination of the dual arms, leading to the risk of load tilting and falling. The distribution of amplification nodes presents an aggregation feature, usually concentrated in key conversion points of the kinematic chain, such as the multi-degree-of-freedom coupling positions of the hip joint and shoulder joint. The time-varying amplification characteristics indicate that certain nodes only exhibit amplification effects under specific conditions, such as the significant increase in deviation amplification effect of the waist node during rapid turning. The significance of identifying deviation amplification nodes lies in locating the weak links of the system, as these nodes are the source of deviation out of control and need to be monitored and optimized.

[0078] Using a similar analysis method, the convergence nodes in the deviation propagation network are identified. Deviation convergence nodes have the ability to absorb and attenuate deviations, characterized by output deviation less than input deviation. Convergence characteristics include: cumulative intensity decreases over time, showing an exponential decay trend, where λ is the decay coefficient; the propagation velocity gradually decreases until it stops; the node transfer function has low-pass filtering characteristics, with a gain G_i<1. The convergence mechanism is derived from adaptive adjustment of motion, redundant degrees of freedom, and compliant control. The force control loop absorbs position deviation through compliance, and visual servoing eliminates cumulative error through real-time correction. Convergence nodes often appear in feedback control loops, fault tolerance mechanism activation points, and human-machine collaboration interfaces. The convergence rate reflects the efficiency of the node in eliminating deviations, and fast convergence nodes can reduce deviations to an acceptable level within a few control periods. Stability analysis of convergence ensures that there is no oscillation caused by excessive correction. The capacity limit of the convergence node determines the maximum deviation that can be handled. The value of identifying deviation convergence nodes lies in discovering the self-healing ability of the system, as these nodes can block the deviation propagation chain and are key resources for improving system robustness.

[0079] For the identified amplification nodes and convergence nodes, the control mechanism is analyzed to obtain the available components. The goal of the control analysis is to understand the bias handling mechanism of the nodes and to mine factors that can be used to improve system performance. For amplification nodes, the amplification reasons are analyzed: kinematic amplification caused by structural constraints, such as the inverted pendulum structure of a humanoid robot that amplifies the bottom bias at the top; excessive response caused by improper control gain setting; coupling resonance of environmental disturbance and system inherent frequency, etc. For convergence nodes, the convergence mechanism is studied: the active compensation strategy of the ankle joint eliminates the bias by adjusting the support torque in real time; the passive compliance characteristics of the flexible spine naturally absorb the impact; the fusion of vision and inertial sensors provides accurate state estimation, etc. These convergence mechanisms are valuable resources of the system and can be applied to other nodes. The exploration of the control space identifies adjustable parameters, such as control gain K_p, damping coefficient ζ, filter parameter τ, etc. The adjustment of these parameters can change the bias handling characteristics of the nodes. Sensitivity analysis reveals the degree of influence of each parameter on the characteristics of the nodes, ∂G / ∂K_p represents the sensitivity of the gain to the amplification rate, and high sensitivity parameters have greater optimization potential. The constraint conditions define the feasible region of the control, ensuring that the adjustment does not damage the stability of the system. The optimal control strategy is determined through multi-objective optimization, balancing bias suppression and system performance. The available components are all elements that can be used to improve the system identified through the above analysis, including adjustable parameter range, constraint condition, optimized control strategy, etc. The identified adjustable parameters, constraint conditions and control strategies are quantitatively collected, and finally a multi-dimensional dataset of the control characteristics of each node is obtained as the available components.

[0080] In some embodiments, the generating the action enhancement factor according to the available components comprises: performing principal component analysis on the available components to extract key bias features; constructing a bias enhancement mapping relationship using the key bias features and the core action sequence; generating a standard enhancement parameter according to the bias enhancement mapping relationship; and generating an action enhancement factor acting on action execution by vector space transformation of the standard enhancement parameter.

[0081] The key deviation features are extracted by principal component analysis (PCA) on the available components. The available component dataset contains multi-dimensional parameters extracted from the deviation amplification nodes and convergence nodes. First, the standardization of each dimension parameter is performed: X_norm = (X - μ) / σ, where X is the original data, μ is the mean, and σ is the standard deviation, to eliminate the influence of dimension difference. The covariance matrix of the standardized data is calculated, and the principal component direction is obtained by eigenvalue decomposition. According to the size of the eigenvalue, the first k principal components with cumulative contribution rate exceeding the set threshold are selected. The principal component loading matrix W shows the contribution weight of the original parameters to the principal components. The first principal component usually reflects the overall control ability, the second principal component corresponds to the fast response characteristic, and the third principal component is related to the stability index. The key deviation features are obtained by projection transformation: Y = X_norm × W, which realizes the effective compression from the multi-dimensional parameter space to the key feature space. The feature selection process not only reduces the computational complexity, but also highlights the key factors that have the most significant impact on the system performance.

[0082] The key deviation features and the core action sequence are used to establish a deviation-enhanced nonlinear mapping relationship. The core action sequence provides a standard action template A_core = [a1, a2,..., an], where ai represents the i-th action, and each action contains basic attributes such as position, velocity, and force. The key deviation features Y represent the quantitative description of the weak links of the system, and a mapping function from the deviation features to the enhancement strategy needs to be established. According to the deviation type classification, a special mapping relationship is constructed: the position deviation adopts trajectory optimization mapping E_pos = k1 · y1 + k2 · y1² / (1 + α|y1|), where E_pos is the position enhancement, y1 is the position deviation feature, k1 and k2 are mapping coefficients, and α is a saturation parameter. The impedance adjustment mapping and dynamic compensation mapping are used for force deviation and velocity deviation, respectively, forming a complete mapping function family. The mapping coefficients are obtained by training historical data, and the least squares method is used for fitting to ensure the accuracy of the mapping. The mapping relationship also considers the coupling effect between actions, and the enhancement of adjacent actions is connected by a smoothing function to avoid the sudden change of the enhancement strategy. The key transition points such as posture switching and contact establishment are set with enhancement weight factors w_critical to improve the priority of enhancement. The established mapping relationship forms a structured enhancement lookup table, realizing the fast conversion from the deviation features to the enhancement strategy.

[0083] According to the bias-enhanced mapping relationship, the mapping output is converted into directly applicable standard enhancement parameters. The standard enhancement parameters include a position compensation vector Δp = [Δpx, Δpy, Δpz], a force control gain correction coefficient α_f, a velocity curve optimization parameter β_v, and an acceleration limit parameter a_max, etc., where Δpx, Δpy, Δpz are the position compensation components of the three coordinate axes, respectively. The parameter standardization process considers the physical constraints of the actuators, and the position compensation is limited within the workspace: ||Δp||≤p_max, where p_max is the maximum allowed compensation amplitude, and the force gain correction maintains system stability. The enhancement intensity grading design provides a multi-grade selection mechanism, which automatically selects the appropriate enhancement level according to the severity of the bias, ensuring that the enhancement effect matches the bias level. The parameter timing arrangement is described by a time function: P_std(t) = P_base + P_enhance · h(t - t_start), where P_std(t) is the standard parameter at time t, P_base is the base parameter, P_enhance is the enhancement amplitude, h(t) is the enhancement activation function, and t_start is the enhancement start time. The sigmoid form is used to achieve smooth start and end. Compatibility verification ensures that the enhancement parameters will not cause control conflicts or system oscillations through simulation checks. The verification process includes stability analysis and performance evaluation. The standardized parameter format ensures compatibility of different types of actuators, and the modular design of the parameters supports flexible combination and adjustment.

[0084] The standard enhancement parameters are mapped to the humanoid robot execution space through a vector space transformation to generate action enhancement factors that can directly drive the actuators. Based on the kinematic and dynamic models of the humanoid robot, the transformation matrix T is composed of the Jacobian matrix J and the inertia matrix M: T = J T · M, where J is the Jacobian matrix, M is the inertia matrix, and J T is the transpose of the Jacobian matrix. The Jacobian matrix J describes the velocity mapping relationship from the joint space to the Cartesian space, and its transpose J T provides the conversion from the Cartesian force to the joint torque. The inertia matrix M reflects the mass distribution and inertia characteristics of each link of the humanoid robot, ensuring that the enhancement factors meet the dynamic constraints. The enhancement factor calculation formula is: F_enhance = T·P_std = J T ·M·P_std, where F_enhance is the enhancement factor vector, P_std is the standard enhancement parameter vector, and the output F_enhance contains the torque enhancement components of each joint. Kinematic singular points are handled during the transformation, and when the Jacobian matrix determinant approaches zero, a damped least squares method is used to avoid numerical instability. Time-varying enhancement factors generate continuous trajectories through interpolation, and cubic spline interpolation is used to ensure acceleration continuity, ensuring the smoothness of the enhancement process. The generation of action enhancement factors realizes the active optimization and adaptive adjustment of system performance. Taking a precision assembly task as an example, when it is predicted that the "insertion" action may fail due to position deviation, the enhancement factor adjusts the torque distribution and velocity curve of the relevant joints in advance, improving the success rate of task execution. The enhancement factor not only compensates for the inherent deviation amplification characteristics of the system, but also utilizes the regulation mechanism of the deviation convergence node to extend the local convergence ability to the global, improving the robustness of the overall system. This active enhancement mechanism realizes the transition from passive fault tolerance to active optimization, enabling the humanoid robot to prevent potential performance degradation in complex industrial environments and maintain stable and efficient task execution capabilities.

[0085] In step S160, the action enhancement factor and the entangled action vector chain are optimized together to generate an enhancement control sequence, and the humanoid robot is driven using the enhancement control sequence.

[0086] Specifically, first, specific control adjustment instructions are extracted from the action enhancement factor, including torque compensation amplitude, trajectory offset, attitude adjustment angle, response time window, and other key parameters. These parameters come from the deep analysis results of the deviation amplification node and the convergence node. At the same time, the spiral topological structure information is read from the entangled action vector chain to obtain the spatial coordinates of each action node, the connection relationship between adjacent nodes, the entanglement strength value between nodes, and the time sequence execution order. These structured data provide a navigation basis for the accurate positioning of the enhancement factor. The cooperative optimization process matches the enhancement factor with the vector chain node one by one through the action identification code. Each enhancement factor carries a target node identifier that is compared with the node number stored in the vector chain. After successful matching, the enhancement parameters are injected into the corresponding basic action vector. For nodes in high entanglement areas, the enhancement process needs to consider the chain reaction of adjacent nodes. By reading the entanglement coefficient value, the influence range is determined. When the entanglement coefficient is high, the enhancement of a single node will simultaneously adjust the execution parameters of its associated nodes to ensure the coordination of the action combination is not disrupted. The optimization strategy adopts a hierarchical processing method, which decomposes the complex vector chain structure into task layer target adjustment, action layer parameter correction, and joint layer driving instruction. Each level receives enhancement information of the corresponding granularity to form a complete control chain from macro planning to micro execution. For example, when handling precision assembly tasks, the force control precision improvement instruction in the enhancement factor is assigned to the "insertion" action node, and the high-entanglement associated node "alignment" also receives coordination enhancement instructions to ensure the synchronous optimization of the two actions. Finally, an enhanced control sequence that integrates the deviation utilization mechanism is generated.

[0087] The generated enhanced control sequence is used to drive the humanoid robot to perform tasks, achieving active utilization of deviations and performance optimization. Through the real-time controller, the sequence instructions are converted into motor driving signals. The deviation utilization mechanism converts the deviations that need to be eliminated into beneficial adjustment resources, such as using gravity deviation to assist the descent action and using inertia deviation to accelerate the turning process. The real-time monitoring module tracks the execution state, maintains the current strategy when the actual deviation is consistent with the predicted path, and activates the compensation mechanism when there is a deviation. The enhancement effect is reflected in the reduction of action completion time, the improvement of precision, and the reduction of energy consumption. Especially when handling non-integer period actions, the enhancement of the borrowed integration mechanism enables the originally interrupted action to be completed smoothly. The execution verification confirms the effective suppression of the identified deviation amplification node and the full utilization of the convergence node, and the running stability is maintained throughout the process. The accurate execution of the enhanced control sequence and the clever transformation of the deviation ultimately successfully drive the humanoid robot to complete the optimization control of deviation utilization.

[0088] In order to perform the above-mentioned method embodiment corresponding to the vector entanglement-based humanoid robot control method, to realize the corresponding functions and technical effects. Referring to Figure 2 , Figure 2A structural block diagram of a humanoid robot control device 200 based on vector entanglement provided by an embodiment of the present application is shown. For ease of illustration, only the parts related to the present embodiment are shown. The humanoid robot control device 200 based on vector entanglement provided by the embodiment of the present application includes:

[0089] A vector construction module 201 is configured to receive a humanoid robot control task, perform frequency statistical analysis on the control task to extract a core action sequence, construct a basic action vector according to the core action sequence, and generate an entangled action vector chain by using the basic action vector through a partial overlap entanglement manner.

[0090] A semantic mapping module 202 is configured to perform dynamic semantic mapping by using the entangled action vector chain to construct an initial action semantic graph, perform evolution update on the initial action semantic graph according to real-time execution feedback, and generate a deviation prediction path through the updated action semantic graph.

[0091] A deviation analysis module 203 is configured to perform humanoid robot action according to the deviation prediction path to generate actual execution deviation data, compare and analyze the actual execution deviation data and the deviation prediction path to generate a deviation feature vector.

[0092] An action completion module 204 is configured to identify a non-integer period action by using the deviation feature vector, extract an entangled action vector that overlaps with the non-integer period action from the entangled action vector chain, and perform borrow fusion on the entangled action vector to generate a complete execution sequence.

[0093] A deviation enhancement module 205 is configured to construct a deviation propagation network by combining the complete execution sequence and the deviation feature vector, analyze a diffusion path of deviation in an action chain through the deviation propagation network to obtain a deviation amplification node and a deviation convergence node, perform deviation regulation analysis on the deviation amplification node and the deviation convergence node to obtain an available component, and generate an action enhancement factor according to the available component.

[0094] A control optimization module 206 is configured to perform collaborative optimization on the action enhancement factor and the entangled action vector chain to generate an enhanced control sequence, and complete driving of the humanoid robot by using the enhanced control sequence.

[0095] The humanoid robot control device 200 based on vector entanglement described above can implement the humanoid robot control method based on vector entanglement of the method embodiment described above. The optional items in the method embodiment described above are also applicable to the present embodiment, which will not be described in detail herein. The remaining contents of the present embodiment can be referred to the contents of the method embodiment described above, which will not be described in detail herein.

[0096] As Figure 3As shown, the third embodiment of the present application further provides a computer device, comprising a memory 301, a processor 302, and a computer program stored in the memory 301 and capable of running on the processor 302, characterized in that the processor 302 implements the steps of the vector entanglement-based humanoid robot control method according to the first embodiment of the present application when executing the program.

[0097] The above embodiments are intended to exemplarily reproduce and deduce the technical solutions of the present application, and to completely describe the technical solutions, objects and effects of the present application, so as to make the public more thoroughly and comprehensively understand the disclosed content of the present application, and not to limit the protection scope of the present application.

[0098] The above embodiments are not exhaustive enumeration based on the present application, and there can be a plurality of other embodiments not listed. Any replacement and improvement made without violating the concept of the present application is within the protection scope of the present application.

Claims

1. A humanoid robot control method based on vector entanglement, characterized in that, include: Receive a humanoid robot control task, perform frequency statistical analysis on the control task to extract core action sequences, construct basic action vectors based on the core action sequences, and generate an entangled action vector chain using the basic action vectors through partial overlap and entanglement. An initial action semantic graph is constructed by using the entangled action vector chain for dynamic semantic mapping. The initial action semantic graph is then updated based on real-time execution feedback. A deviation prediction path is generated using the updated action semantic graph. The humanoid robot performs actions according to the deviation prediction path to generate actual execution deviation data. The actual execution deviation data and the deviation prediction path are compared and analyzed to generate a deviation feature vector. The deviation feature vector is used to identify non-integer periodic actions. Entangled action vectors that overlap with the non-integer periodic actions are extracted from the entangled action vector chain. The entangled action vectors are then fused by borrowing to generate a complete execution sequence. A deviation propagation network is constructed by combining the complete execution sequence and the deviation feature vector. The propagation network is used to analyze the diffusion path of the deviation in the action chain to obtain deviation amplification nodes and deviation convergence nodes. Deviation regulation analysis is performed on the deviation amplification nodes and deviation convergence nodes to obtain usable components. Based on the usable components, an action enhancement factor is generated. The motion enhancement factor and the entangled motion vector chain are collaboratively optimized to generate an enhanced control sequence, which is then used to drive the humanoid robot.

2. The method as described in claim 1, characterized in that, Using the aforementioned basic action vectors, an entangled action vector chain is generated through partial overlap entanglement, including: Calculate the vector angle and magnitude difference between temporally adjacent vectors in the basic action vector to generate a similarity matrix; The vector overlap region is determined based on the similarity matrix; Within the overlapping region, the vector components are nonlinearly weighted and fused to generate entanglement coefficients; The entanglement coefficients are used to construct an entangled action vector chain through a spiral connection method.

3. The method as described in claim 1, characterized in that, The updated action semantic graph generates biased prediction paths, including: Historical deviation nodes are extracted from the updated action semantic graph to construct a deviation evolution subgraph; The deviation evolution subgraph is used to calculate the deviation transition probabilities between nodes and generate a deviation state matrix; Based on the aforementioned deviation state matrix, a multi-step prediction is performed to generate a time-series deviation sequence; The time-series deviation sequence is mapped to the execution time axis to form a deviation prediction path.

4. The method as described in claim 1, characterized in that, Borrowing and fusing the entangled action vectors to generate a complete execution sequence includes: Based on the non-integer periodic action, locate adjacent entangled nodes in the entangled action vector chain; Extract the borrowable vector components from the adjacent entangled nodes to generate a compensation vector set; The compensation vector set and the non-integer cycle action are time-aligned to generate a fusion weight matrix; The complete execution sequence is generated by performing vector fusion operations using the fusion weight matrix.

5. The method as described in claim 1, characterized in that, The deviation propagation network is used to analyze the propagation path of the deviation in the action chain to obtain the deviation amplification node and the deviation convergence node, including: The deviation feature vector is used as the initial excitation input to the deviation propagation network; The cumulative strength and propagation speed of the deviation feature vector at each node are calculated using a network propagation algorithm. Nodes whose cumulative intensity continuously increases and whose propagation speed accelerates are identified as deviation amplification nodes; Nodes whose cumulative intensity gradually decreases and whose propagation speed slows down are identified as deviation convergence nodes.

6. The method as described in claim 1, characterized in that, The generation of motion enhancement factors based on the available components includes: Principal component analysis was performed on the available components to extract key deviation features; A deviation enhancement mapping relationship is constructed using the key deviation features and the core action sequence; Standard enhancement parameters are generated based on the aforementioned deviation enhancement mapping relationship; The standard enhancement parameters are transformed into action enhancement factors that act on the execution of the action through vector space transformation.

7. The method as described in claim 2, characterized in that, The entangled action vector chain is constructed using the entanglement coefficients in a spiral connection manner, including: The radial distribution of the spiral is determined based on the force vector component in the basic motion vector, and the tangential velocity of the spiral is determined using the velocity vector component in the basic motion vector. The helical trajectory equation is generated by combining the radial distribution and the tangential velocity; The spatial position of each vector is calculated based on the spiral trajectory equation and the entanglement coefficient; A spiral entangled action vector chain is formed by weighting the connection of the spatial locations using entanglement coefficients.

8. The method as described in claim 4, characterized in that, Extracting borrowable vector components from the adjacent entangled nodes to generate a compensation vector set includes: Analyze the vector redundancy of the adjacent entangled nodes and identify the borrowable components whose redundancy exceeds a threshold; For the missing type of the non-integer cycle action, the borrowing priority is determined, with force component taking priority over velocity component, and velocity component taking priority over direction component. The borrowing priority and the entangled action vector chain are combined to construct the borrowing weight, and the closer the nodes are, the greater the borrowing weight. Based on the borrowing weight, vector components are dynamically extracted from the borrowable components to form a compensation vector set.

9. A humanoid robot control device based on vector entanglement, characterized in that, include: The vector construction module is used to receive humanoid robot control tasks, perform frequency statistical analysis on the control tasks to extract core action sequences, construct basic action vectors based on the core action sequences, and generate entangled action vector chains using the basic action vectors through partial overlap and entanglement. The semantic mapping module is used to construct an initial action semantic graph by performing dynamic semantic mapping using the entangled action vector chain, to evolve and update the initial action semantic graph based on real-time execution feedback, and to generate a deviation prediction path through the updated action semantic graph. The deviation analysis module is used to execute humanoid robot actions according to the deviation prediction path to generate actual execution deviation data, and compare and analyze the actual execution deviation data with the deviation prediction path to generate a deviation feature vector. The action completion module is used to identify non-integer periodic actions using the deviation feature vector, extract entangled action vectors that overlap with the non-integer periodic actions from the entangled action vector chain, and perform borrowing fusion on the entangled action vectors to generate a complete execution sequence. The deviation enhancement module is used to construct a deviation propagation network by combining the complete execution sequence and the deviation feature vector, analyze the diffusion path of the deviation in the action chain through the deviation propagation network to obtain deviation amplification nodes and deviation convergence nodes, perform deviation regulation analysis on the deviation amplification nodes and deviation convergence nodes to obtain usable components, and generate action enhancement factors based on the usable components. The control optimization module is used to collaboratively optimize the motion enhancement factor and the entangled motion vector chain to generate an enhanced control sequence, and to use the enhanced control sequence to drive the humanoid robot.

10. A computer device, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the method as described in any one of claims 1 to 8.

Images