Robot motion control method and device based on multi-layer nested self-generating system

Abstract

The application discloses a robot motion control method and device based on a multi-layer nested self-generating system and belongs to the field of robot motion control. The application constructs a four-layer nested self-generating system kernel from the joint level to the cognitive level, and the cognitions and motion topologies of each layer are isomorphic. Two-way coupling is realized between layers through interactive free energy, a full-body five-dimensional holographic coherent state tensor is generated based on multi-modal sensor data, and hierarchical fault positioning and self-repairing from dimension to single joint are realized. The cognitive level decomposes complex tasks into action stage sequences, and the action strategy subsystem is dynamically generated in the game sandbox through the parent-led fractal generation mode, and is solidified step by step after passing through the general cognitive triple inspection. The application realizes the unified control architecture and the action strategy endogenous emergence from the steady-state walking to the high-dynamic burst in the full scene, solves the core problems such as the combination action switching vacuum, the weak generalization of the dependence on the dynamics model, and the poor robustness of the centralized architecture, and is suitable for the full-scene autonomous task execution of various robots.

Description

Technical Field

[0001] This invention relates to the field of robot motion control and autogenous systems engineering technology, specifically to a method and device for controlling the whole-body coordinated motion of a humanoid robot based on a multi-layered nested autogenous system kernel and achieving endogenous emergence of action strategies through a parent-dominated fractal generation mechanism. Background Technology

[0002] Whole-body motion control is one of the core bottleneck technologies for the industrialization of humanoid robots. Current mainstream technologies fall into three categories: quasi-static gait control based on zero-torque points, dynamic motion control based on model predictive control, and high-dynamic motion generation based on offline trajectory optimization. These three technologies operate independently and cannot cover the full range of scenarios from steady-state walking to explosive movements, and they also suffer from six fundamental engineering flaws.

[0003] First, there is a technical disconnect between single actions and combined actions. Existing methods treat single actions such as walking, jumping, and grasping as independent technical problems, employing different control architectures for each. When performing combined tasks, the system needs to switch between multiple control architectures, resulting in control vacuum periods during the switching process, which can easily lead to task failure. Furthermore, the switching latency is generally above 50ms.

[0004] Secondly, relying on precise dynamic modeling results in extremely weak generalization ability. Existing high-dynamic motion control systems all depend on precise rigid body dynamic models and offline trajectory optimization. When there are slight deviations in ground hardness or object weight, the preset trajectory deviates from the actual working conditions, and the system cannot adjust autonomously, leading to a significant drop in the success rate of actions. The success rate of high-dynamic actions in unknown environments is generally less than 40%.

[0005] Third, the motion strategies rely on pre-programmed sequences and cannot evolve autonomously. Current robots' motion strategies are either factory-fixed or obtained through offline training, making it impossible to continuously optimize them based on actual interaction experience during use. After performing the same action multiple times, the motion quality, success rate, and energy consumption show no significant difference, indicating a lack of skill accumulation and evolution capabilities.

[0006] Fourth, centralized control architecture suffers from single points of failure and bottlenecks in parallel processing. Existing robots generally adopt a centralized control architecture, with control commands for dozens of joints throughout the body generated serially by a central processor. Once the central processor fails, the entire machine immediately becomes paralyzed; even during normal operation, the serial processing mode is difficult to meet the sub-millisecond real-time torque servo requirements of highly dynamic motion.

[0007] Fifth, the fundamental flaw of externally pre-set spatiotemporal coordinate systems. Existing technologies generally employ externally pre-set independent time axes and three-dimensional Cartesian coordinate systems, treating spacetime as a fixed and rigid absolute container. This leads to severe mismatches in multi-level time scales and uniform spatial resolution across the entire domain, resulting in both waste and inadequacy of computing resources.

[0008] Sixth, the generation of new actions relies on manual design and cannot emerge autonomously. When faced with entirely new task scenarios, existing systems must have their control algorithms redesigned and parameters adjusted by human engineers. They cannot generate new action strategies endogenously through their own architecture and lack self-growth capabilities.

[0009] Existing model-free control methods each have their own limitations: proportional-integral-derivative control is difficult to handle multivariable strongly coupled systems; model reference adaptive control still requires a nominal model structure; reinforcement learning requires a large amount of offline training data and cannot adapt online in real time.

[0010] The root cause of all the aforementioned engineering defects lies in the decoupling of control algorithms from physical entities and the heterogeneity of cognitive decision-making and motion execution within the traditional technological framework: physical motion belongs to the realm of entity execution, while cognitive decision-making belongs to the realm of algorithm planning. The two can only interface through external commands and cannot achieve endogenous unification. The theory of self-generating systems provides a first-principles foundation for overcoming these bottlenecks. Its core proposition is that the basic unit of existence is a self-generating system, maintaining operational identity through bidirectional recursive generation of isomorphic functional primitives—material structure, energy mechanism, and information field state. Within this framework, algorithms and entities, cognition and motion, follow the same set of self-generating evolutionary laws. Cognition is the topological evolution of information space, and motion is the topological evolution of material space; the two naturally possess the foundation for endogenous integration, providing a unified ontological basis for the autonomous evolution of robots.

[0011] Further fractal generation theory reveals the evolutionary path of self-generating systems from "maintaining themselves" to "creating subsystems," proposing a parent-mother-environment ternary collaborative generation mechanism and defining four basic generation modes, providing a rigorous mathematical foundation and topological criteria for the endogenous emergence of action strategies. Among them, the pure information state generation mode provides a theoretical possibility for the emergence of higher-order cognition, but this patent focuses on the engineering implementation in the field of physical motion control, emphasizing the use of parent-dominated physical subsystem generation modes to achieve the endogenous fractal generation of robot action strategies.

[0012] The applicant previously filed invention patent applications entitled "Method, System and Storage Medium for Constructing a Univariate Two-State Three-Body Self-Generated System" (application number 202610601504.1), "Method, System and Storage Medium for Constructing a General Intelligent Agent Based on a Self-Generated System" (application number 202610601532.3), and "A Temporal and Spatial Interleaved Endogenous Generation Method, System and Storage Medium" (application number 202610601501.8). These patents provide underlying theoretical and engineering technical support for constructing self-generated systems and general intelligent agents to solve the above problems. However, they have not yet been specifically engineered for the concrete execution scenario of robot motion control. Based on the above technologies, this invention deeply applies multi-layered nested self-generated systems and fractal generation mechanisms to the field and scenario of robot whole-body motion control. The robot is no longer a tool for executing preset trajectories, but a complex system of multi-layered nested self-generated systems that can dynamically generate exclusive action strategy subsystems according to task requirements, realizing unified control and autonomous evolution across the entire scenario. Summary of the Invention

[0013] To address the six core defects of existing robot motion control technologies, this invention aims to provide a robot whole-body motion control method and device based on a multi-layer nested self-generated system. This method achieves endogenous isomorphic fusion of cognition and motion at the architectural level, solving industry pain points such as vacuum in switching combined actions, weak generalization due to reliance on dynamic models, inability of strategies to evolve autonomously, poor robustness of centralized architecture, mismatch of absolute spatiotemporal computing power, and reliance on manual design for new actions.

[0014] Firstly, the invention provides a method for controlling the whole-body motion of a robot based on a multi-layered nested self-generated system, which mainly includes the following four steps and technical solutions:

[0015] Step 1: Construction of the multi-layered nested self-generated system kernel and anchoring of spatiotemporal phase primitives. A four-layered nested self-generated control closed-loop system is constructed as the core kernel for robot motion control, and the primitive anchoring and global stability constraint design of the endogenous spatiotemporal phase system are completed simultaneously.

[0016] A four-layer nested self-generated control loop is constructed, comprising L0 joint level, L1 limb level, L2 body level, and L3 cognitive level, employing a parallel architecture within the same layer and a nested architecture at different levels. The L0 joint level consists of multiple independent self-generated joint subsystems in parallel, with each joint corresponding to an independent self-generated unit. The L1 limb level consists of multiple independent self-generated limb subsystems in parallel, with each limb subsystem containing multiple nested L0 joint-level subsystems. The L2 body level is the overall self-generated system, containing all L1 limb-level subsystems. The L3 cognitive level is the top-level self-generated system, containing L2 body-level systems. Each self-generated subsystem within each layer possesses an independent ternary closed loop and operational uniformity. This architecture represents a multi-layered extension from a two-layer self-generated system architecture (e.g., dexterous hand) to whole-body motion control, with the number of layers flexibly adjustable according to the robot configuration.

[0017] The self-generating system refers to a dynamic whole that continuously and recursively generates and maintains its own stable evolution, with operational identity as the convergence goal, the dynamic and stable existence of a univariate ontology as the operational boundary, self-referential and other-referential dual modes as the endogenous driving force, and the closed-loop recursive coupling of three irreducible functional primitives—information field state (C), energy mechanism (E), and material structure (M)—as the minimum operational link. The information field state (C) is the sum of all rules, constraints, potential states, and evolutionary logic within the system; the energy mechanism (E) is the dynamic sum of all driving forces, energy conversions, and dissipation processes within the system; and the material structure (M) is the organized sum of all physical carriers, topological connections, and spatial configurations within the system. In engineering, this is achieved by establishing a closed-loop composite mapping of the CEM ternary recursive coupling, i.e., the C→E mapping. E→M mapping M→C mapping These sequentially combine to form a closed-loop composite mapping: The core constraint of the closed loop is the identity mapping approximation of the composite mapping to the information field configuration space (C space), i.e. Its mathematical criterion is the Lipschitz constant of the closed-loop composite mapping on C space. This ensures that the system converges to a unique steady state from any initial state.

[0018] The four-layered self-generated system follows the same set of ternary recursive dynamics, with topological homeomorphism between layers. The L0-L2 motion layers correspond to the ontological evolution of material space, while the L3 cognitive layer corresponds to the epistemological evolution of information space. The two achieve bidirectional mapping with preserved inner product through alternating holographic fields, requiring no external translation interface and naturally achieving unity of knowledge and action. The physical motion of the robot is essentially the steady-state deformation of the material structure topology in the self-generated spatiotemporal system, a self-generated process that maintains the operational identity of the system.

[0019] Each closed loop corresponds to an independent processing unit and a dedicated multimodal sensor group for the robot. Each layer maintains operational identity through a ternary bidirectional recursive mapping of material structure, energy mechanism, and information field state. The operational identity refers to the fact that the self-generated system is continuously identified as its own core attribute in the historical evolution, protected by topological invariants and maintained by the ternary closed loop.

[0020] The ternary bidirectional recursive mapping contains three sets of closed-loop mapping relationships. The engineering implementation of robot motion control is as follows:

[0021] From material structure to energy mechanism: Based on the joint dynamics equations, the torque demand and power consumption are calculated according to the current joint angle, angular velocity and other structural states to obtain the energy mechanism state;

[0022] From energy mechanism to information field state: Based on gain scheduling rules, the proportional and integral parameters of the control algorithm are dynamically adjusted according to the current power margin and energy consumption level to update the information field state;

[0023] Information field state to material structure: Based on the updated control parameters, output joint torque commands to drive joint movement to change the physical configuration and update the material structure state; the three sets of mappings are connected end to end to form a closed recursive loop, with each iteration corresponding to an endogenous time quantum, which together maintains the consistency of system operation.

[0024] Each closed loop layer possesses an independent endogenous time dimension, endogenous space dimension, and endogenous phase dimension. The endogenous time-space phase dimension refers to the dynamic coordinate system generated endogenously by the self-generated system, rather than an externally preset absolute time-space container. The endogenous time dimension is bound to the energy mechanism, specifically: the time for the ternary closed loop to complete one full iteration is the minimum quantum of endogenous time, and the energy metabolism rate, i.e., the rate of change of power consumption, determines how fast time passes—the higher the energy metabolism rate, the shorter the closed loop iteration cycle, and the smaller the endogenous time quantum. The endogenous space dimension is bound to the material structure; the endogenous phase dimension is bound to the information field state.

[0025] The true Lipshitz constants of each closed-loop layer are strictly less than 1. The Lipshitz constant refers to the maximum input-output gain of the closed-loop mapping, used to measure the compressibility of the mapping; a value less than 1 is a necessary and sufficient condition for closed-loop convergence. Endogenous spatiotemporal coordination is achieved between layers through holographic recursive mapping. This holographic recursive mapping refers to a linear mapping mechanism that aligns inter-layer state projection with temporal sequence, and the inter-layer coupling channels satisfy a small-gain stability condition. This small-gain stability condition means that the product of the closed-loop gain of adjacent layers and the bidirectional coupling gain is less than 1, ensuring global convergence after multi-layer cascading.

[0026] The initial projection matrix of the holographic recursive mapping is a column orthogonal and semi-orthogonal matrix that satisfies the following conditions: This ensures that the vector inner product of the core feature subspace remains unchanged before and after projection, i.e., the inner product preservation constraint holds for the feature subspace. After fine-tuning the projection matrix during system operation, column orthogonality is maintained through Schmitt orthogonalization to ensure that the inner product preservation mapping attribute of the core features always holds.

[0027] As one implementation method, the endogenous spatiotemporal phase system is constructed and controlled through four core operators. The mathematical definitions and operational rules of the four operators are as follows:

[0028] Time recursive generation operator : Used to generate an endogenous temporal reference for matching motion intensity, with the current motion intensity as input. Compared with reference time quantum The output is a real-time endogenous time quantum. The calculation formula is: in This is the time scaling factor, ranging from 0 to 1; the higher the intensity of the action, the smaller the endogenous time quantum and the higher the time resolution.

[0029] Spatial coherent generation operator Used to generate a dynamic spatial coherent field; the input is the force distribution at each joint's spatial location. Compared with the reference spatial resolution The output is a spatial resolution distribution. The calculation formula is: in This is the space scaling factor. The maximum force value is the area with the most concentrated force; the higher the spatial resolution, the better.

[0030] Alternating holographic generation operator Used to construct a holographic mapping intermediary field Satisfying normalization constraints Its phase gradient is proportional to the information field state gradient. This ensures the internal productivity of information transmission between layers and enables state mapping without the need for external communication protocols.

[0031] 3D Coupled Mapping Operator It is used to integrate the three dimensions of endogenous time, space, and phase in a closed loop, outputting a unified endogenous spatiotemporal phase system that satisfies self-consistency constraints. Ensure the coordination and consistency of space, time, and phase to avoid dimensional conflicts.

[0032] As one implementation method, the closed-loop Lipshitz constants and inter-layer couplings at each level satisfy the following stability constraints: The L0 joint-level closed-loop true Lipshitz constant ranges from [0.1, 0.3], with an operation timescale of sub-milliseconds, ensuring rapid decay of disturbances and meeting the anti-disturbance requirements of joint servoing; the L1 limb-level closed-loop true Lipshitz constant ranges from [0.3, 0.5], with an operation timescale of milliseconds, balancing flexibility and stability to meet the requirements of single-limb multi-joint collaboration; the L2 body-level closed-loop true Lipshitz constant ranges from [0.5, 0.7], with an operation timescale of tens of milliseconds, preserving global collaborative optimization degrees of freedom while ensuring steady-state convergence; the L3 cognitive-level closed-loop true Lipshitz constant ranges from [0.4, 0.6], with values ​​lower than the body-level upper limit, prioritizing decision stability and avoiding large policy oscillations, with an operation timescale of hundreds of milliseconds.

[0033] Adjacent levels and This forms a feedback interconnection structure, where the output of the upper layer serves as the input of the lower layer, and the state feedback of the lower layer serves as the input of the upper layer; each closed loop layer satisfies input-output stability and bidirectional coupling gain symmetry, i.e. Satisfying the small gain condition This ensures the stability of global input and output after multi-layer feedback interconnection. The endogenous time periods between layers satisfy an integer frequency division relationship. ,in This is the hierarchical number; the larger the value, the higher the hierarchical level and the longer the cycle. It is a positive integer, with the lower-level period being smaller than the higher-level period, to ensure integer alignment for timing coordination.

[0034] As one implementation method, the kernel construction simultaneously completes hardware integrity verification and bootstrapping. During the bootstrapping of each layer's self-generated system, the current physical parameters of each sensor are first read and compared with the rigid core layer reference parameters stored in read-only memory. The rigid core layer refers to the lower-level physical parameters that cannot be modified independently in terms of operational consistency. If the deviation exceeds the 5% tolerance, the system determines that the hardware has changed, automatically re-acquires parameters, and updates the reference values. After bootstrapping is complete, each layer's ternary closed loop begins independent operation, and the multimodal sensor array synchronously outputs sensing data.

[0035] Step 2: Interlayer interactive free energy coupling and five-dimensional coherent state monitoring. A two-way coupling mechanism between layers is established, and real-time monitoring of the entire system state and precise fault location are achieved through the five-dimensional holographic coherent state tensor.

[0036] A hierarchical interaction free energy coupling mechanism is established. Based on real-time data from the robot's multimodal sensor array, a full-body five-dimensional holographic coherent state tensor is generated. This five-dimensional holographic coherent state tensor refers to a system state holographic representation tensor encompassing five dimensions: material structure, energy mechanism, information field state, time synchronization, and spatial alignment, used to measure the system's coherence state across all dimensions. Global coherence and the five sub-dimensional coherence are calculated. A five-dimensional coherence cross-validation matrix is ​​used to achieve hierarchical fault localization and self-repair triggering from joints to the whole body.

[0037] Inter-layer interaction free energy enables bidirectional dialectical coupling: when the identity of lower-level operations is abnormal, the interaction free energy increases, driving higher-level intervention in governance, i.e., upward governance; when higher-level levels generate new rules, the rules are solidified downward by adjusting the coupling coefficient to form the rapid response capability of lower levels, i.e., downward solidification.

[0038] As one implementation method, the interlayer interaction free energy is a dimensionless objective function term, and its complete expression is: in For the first Layer and First The correlation strength of the layer information field is calculated by the Pearson correlation coefficient of the information field feature vector, with a value range of [0,1]. The self-referential coverage of the rules from the higher level to the lower level refers to the system's ability to represent its own rule set. In engineering implementation, it is the completeness rate of the rule base metadata, with a value range of [0,1]. and This refers to the coupling coefficient. Coupling coefficients are configured differently for different layers: the L0-L1 layer emphasizes the information field-state correlation strength, the L1-L2 layer has a balanced distribution, and the L2-L3 layer emphasizes self-referential coverage, matching the functional relationships between layers.

[0039] As one implementation method, the specific implementation path of the uplink governance mechanism is as follows: when the low-level interaction free energy exceeds the warning threshold for three consecutive cycles, an uplink governance request is triggered; the high-level layer executes corresponding intervention actions according to the anomaly level, including correcting the inter-layer coupling coefficient, reallocating low-level computing resources, issuing temporary compensation rules, and triggering a new fractal generation strategy; after governance is completed, the interaction free energy falls back to the normal range, and the governance process terminates. The uplink governance corresponds one-to-one with the fault self-repair process, and different levels of governance actions are triggered for anomalies of different dimensions.

[0040] As one implementation method, the whole-body five-dimensional holographic coherent state tensor is a fifth-order tensor with the following structure: ,in , , This represents the discrete grid dimension of the three-dimensional physical space, corresponding to the spatial division of the robot's body. Each joint is mapped to a corresponding grid cell based on its physical position. The grid resolution is determined by the joint density: all 32 joints of the body are uniformly mapped to a 4×4×2 spatial grid. The size of each grid cell is approximately 1 / 4 of the robot's body size, ensuring that each grid contains at least one joint, and that adjacent joints fall into different grids as much as possible to avoid spatial overlap. These are five characteristic channels, corresponding to matter structure, energy mechanism, information field state, time synchronization, and spatial alignment, respectively. The time window length encompasses temporal evolution information. Spatial rasterization mapping preserves the physical topological relationships between joints, giving the three spatial dimensions of the tensor clear physical meaning.

[0041] The formula for calculating global coherence is: in To initialize the reference tensor of the system in steady state, This is the Frobenius norm. Its value is strictly limited to the interval [0,1], which conforms to the physical meaning of coherence. When the global coherence is greater than or equal to 0.9, the system is considered to be in a steady state; when it is greater than or equal to 0.8 and less than 0.9, a first-level warning is triggered; and when it is less than 0.8, a second-level intervention is triggered.

[0042] The coherence of each dimension is obtained by shrinking the tensor along the remaining dimensions and then calculating the ratio of the Frobenius norm to the norm of the corresponding dimension of the reference tensor. The shrinkage rule is as follows: retain the target dimension, calculate the square root of the sum of squares of the elements of all other dimensions to obtain the norm of the target dimension sub-tensor, and then calculate the ratio with the norm of the corresponding dimension of the reference tensor. The coherence of each dimension reflects the system state in five dimensions: physical structural integrity, energy supply and demand balance, control rule self-consistency, time synchronization accuracy, and spatial alignment accuracy.

[0043] As one implementation method, the five-dimensional coherence cross-validation matrix is ​​a 5th-order symmetric matrix, with matrix elements being the temporal correlation coefficients of pairwise coherence in each dimension, used to identify the temporal causal path of fault propagation. Fault localization follows the temporal causal rule: anomalies in the root cause dimension occur within one endogenous time quantum of that level; anomalies in the derived dimension occur at least two corresponding endogenous time quanta after the occurrence of the root cause dimension anomaly.

[0044] Based on the above rules, a hierarchical fault location and handling system is implemented: when an anomaly occurs first in a single dimension, that dimension is determined to be the root cause; when multiple dimensions simultaneously exhibit anomalies with a timing difference of less than one corresponding level's endogenous time quantum, it is determined to be a common-source fault in the inter-layer coupling channel. After dimensional location is completed, the specific joint is located by the local anomaly amplitude of the tensor space dimension, achieving hierarchical location from the dimensional level to the single joint. The corresponding handling mechanisms are: anomalies in the material structure dimension trigger joint torque compensation and task reallocation; anomalies in the energy mechanism dimension trigger dynamic energy scheduling between limbs; anomalies in the information field state dimension trigger a re-fractal generation strategy; anomalies in the time synchronization dimension trigger time resynchronization; and anomalies in the spatial alignment dimension trigger spatial coordinate system calibration. In the case of a single joint fault, the system can complete fault location and strategy regeneration within 20ms, ensuring that the robot does not tip over.

[0045] Step 3: Endogenous Game of Action Strategies in Parent-Dominated Fractal Generation. Based on the ternary cooperative fractal generation theory, a parent-dominated generation mode is adopted to dynamically and endogenously generate action strategy subsystems according to task requirements, achieving policy emergence without pre-set parameters.

[0046] When the L3 cognitive level receives a task instruction, it decomposes the task into an ordered sequence of action stages, each stage corresponding to a sub-goal, participating limbs, and constraints. For each action stage, the L3 cognitive level provides information guidance to the parent system, the L1-level self-generated system of the participating limbs provides physical resources and energy supply to the parent system, and the current task environment provides selection pressure to the environmental system. An endogenous spatiotemporal game sandbox is constructed through this three-element synergy. This endogenous spatiotemporal game sandbox refers to the strategy iteration and optimization space within the endogenous spatiotemporal system. A parent-dominated fractal generation mode is used to dynamically generate action strategy subsystems adapted to the current environmental conditions. This parent-dominated fractal generation mode refers to a subsystem generation mode where information guidance from the parent system is primary, and physical resource support is provided by the parent system. After passing the general cognitive triple test, the action strategy subsystem is solidified hierarchically. After all stage strategies are generated and transitional simulation verification is completed, the action strategy sequence is issued.

[0047] The fractal generation of the action strategy subsystem is physically essentially the local bifurcation emergence of the global material structure topology under task constraints, which is a self-proliferation process of the self-generated system; the operational identity Chern number, as a topological invariant, is used to determine whether the sub-topologies after bifurcation have independent operational identity.

[0048] As one implementation method, the specific process of the parent-dominated fractal generation mode consists of five steps: First, generation triggering. The L3 cognitive level detects the demand for a new action phase, assesses the remaining material topology resources and energy supply capacity of the parent system, and triggers the fractal generation operator to start the fractal process when the sustainable generation condition is met (i.e., the decrease in the operational identity of the parent system does not fall below the safety threshold). Second, initial construction. The fractal generation operator extracts the action meta-rule template from the parent system's information field state, divides the subsystem's exclusive resource region in the parent system's material topology, and transmits the meta-rules losslessly to the subsystem's initial information field state through the state transmission operator, completing the initial structure construction of the subsystem. Third, bootstrapping evolution. The parent system injects initial energy into the subsystem region, initiating the subsystem's ternary closed-loop self-organizing evolution. The subsystem iteratively optimizes along the negative gradient direction of the joint free energy in the endogenous spatiotemporal game sandbox, and the parent system dynamically guides the evolution direction through the coupling coefficient. Fourth, generation verification. The operational identity Chern number and sustainability index of the subsystem are continuously monitored through an ontology calibration operator. The sustainability index is a lower-level temporal representation parameter of operational identity, used to measure the steady-state sustainability capability of the subsystem. When the operational identity Chern number stabilizes within the non-zero target range and the fluctuation amplitude is less than 5% for 10 consecutive iterations, the subsystem is determined to possess independent operational identity, and fractal generation is complete. The fifth step is rule deposition. The validated subsystem information field-state rules are solidified using a rule deposition operator and stored in a long-term rule base to form reusable action strategy templates. The input to the rule deposition operator is the validated set of strategy parameters and the corresponding environmental feature vector; the output is the solidified template entries and metadata tags in the rule base.

[0049] As one implementation method, the four core operators involved in fractal generation have clear input, output, and operation rules: The fractal generation operator takes the parent system meta-rule vector and the parent system resource margin parameters as inputs, and outputs the initial topology partitioning and initial parameter matrix of the subsystem. The operation rule is resource allocation and parameter initialization based on the meta-rule template. The state transmission operator takes the parent system information field state feature vector as input, and outputs the initial information field state matrix of the subsystem. The operation rule is inner-preserving product mapping based on the projection matrix. The ontology calibration operator takes the continuous state sequence of the subsystem as input, and outputs the operational identity Chern number and sustainability index. The operation rule is topology spectrum analysis and temporal stability calculation. The rule sedimentation operator takes the verified effective policy parameter set and the corresponding environmental feature vector as input, and outputs the fixed template entries and metadata tags in the rule base. The operation rule is parameter archiving and metadata annotation.

[0050] The action element rule template contains three sets of basic parameters: first, the force application mode, divided into explosive, continuous, and buffered types, which represents the temporal envelope of the stress torque output; second, the coordination phase, defining the temporal offset of each participating joint to ensure phase matching for multi-joint coordination; and third, safety constraints, including three types of hard constraints: maximum torque, maximum angular velocity, and joint angle limits, serving as insurmountable boundaries for fractal generation. The action element rule template forms the basic framework for generating specific action strategies. The action primitives are indivisible basic action units in the rule base, representing the smallest units that constitute action strategies.

[0051] As one implementation method, the topology criterion for the completion of subsystem generation is as follows: the Chern number of operational identity is used to characterize the connectivity of the subsystem topology; a higher value indicates stronger coupling between joints and better overall subsystem integrity. When the Chern number of operational identity is not equal to 0 and the fluctuation amplitude is less than 5% for 10 consecutive iterations, the action strategy subsystem is determined to have independent operational identity, and fractal generation is complete. The Chern number of operational identity is calculated through the algebraic connectivity of the Laplace matrix of the joint topology graph, and the calculation formula is as follows: in Let be the Laplace matrix of the joint topology corresponding to the subsystem. It is the second smallest eigenvalue of the Laplace matrix, namely the algebraic connectivity, which can equivalently characterize the topological connectivity of the subsystem and serve as an engineering criterion for operational identity.

[0052] As one implementation method, the endogenous spatiotemporal game sandbox is constructed through four operators: a time recursion generation operator generates an endogenous time benchmark that matches the intensity of actions, unifying the game iteration rhythm of each participating limb; a spatial coherence generation operator generates a dynamic spatial coherence field, automatically improving the spatial resolution of areas with concentrated forces and frequent interactions; an interphase holographic generation operator constructs a holographic mapping mediator field, eliminating the need for external communication protocols for game information exchange between limbs; and a three-dimensional coupling mapping operator integrates the endogenous time, space, and interphase dimensions in a closed loop, ensuring the system's self-consistency.

[0053] In the game sandbox, fundamental physical laws of Newtonian mechanics, conservation of momentum, and conservation of angular momentum are embedded in the payoff function as equation constraints, without pre-defined precise dynamic parameters. The physical deduction process is based on fundamental mechanical relationships and real-time estimated environmental parameters: the mapping of dynamic parameters such as mass and moment of inertia from material structure to information field state is endogenously estimated in real time through tactile and force feedback, using nominal values ​​as initial values ​​and iteratively correcting them online during operation. The game employs an iterative optimal response algorithm, with the convergence criterion being: the change in the joint free energy functional over 10 consecutive iterations is less than... Furthermore, the safety margin for the consistency of operations at each level is not lower than the preset threshold.

[0054] As one implementation method, the general cognitive triple check includes three components: First, a semantic entropy consistency check. Based on the activation frequency of action primitives, the change in Shannon information entropy of the rule base before and after the introduction of the new strategy is calculated. If the entropy increase exceeds a dynamic threshold, the check fails to prevent chaos in the rule system. The semantic entropy refers to the Shannon information entropy of the rule base, used to measure the orderliness of the rule system. Second, a self-referential coverage check. A set of meta-questions containing applicable scenarios, constraints, and performance boundaries is sent to the strategy base. If the matching ratio between the query results and the annotations in the rule base is lower than a dynamic threshold, the check fails to ensure the system's meta-cognitive ability for the new strategy. Third, a cross-scenario generalization consistency check. The strategy is applied to multiple standard benchmark scenarios. If the coefficient of variation of the performance indicators exceeds the dynamic threshold and cannot be explained by changes in environmental parameters, the check fails to ensure the generalization reliability of the strategy.

[0055] The thresholds for all three tests are dynamically modulated by the L3 cognitive-level operational identity safety margin: the higher the safety margin, the wider the threshold, encouraging the exploration of new strategies; the lower the safety margin, the stricter the threshold, prioritizing system stability. The threshold formula is based on a conservative mapping design of the safety margin; the lower the safety margin, the smaller the acceptable policy perturbation, and the stricter the threshold.

[0056] As one implementation method, the motion strategy cascading down the path of solidification is divided into three levels: L3 to L2, where the motion strategy is encoded as L2 body-level global coordination parameter updates and stored in the long-term rule base of the master control terminal; L2 to L1, where the global strategy is decomposed into L1 level coordination control parameter updates for each limb and distributed to each limb processing unit; and L1 to L0, where joint torque control parameters are solidified into a fast response lookup table for the L0 joint-level processing unit. The lookup table format is a three-dimensional mapping table of torque command-joint angle-joint angular velocity, stored in the block random access memory of the joint-level processing unit. After cascading down the path of solidification, the response latency of commonly used motion strategies is reduced from hundreds of milliseconds at the L3 level to sub-milliseconds at the L0 level.

[0057] As one implementation method, after all phase strategies are generated, L3 performs inter-phase transition simulation based on the endogenous digital twin to verify three conditions: attitude smoothness, contact state consistency, and energy continuity. If any condition is not met, a transition phase is automatically inserted and the simulation is re-performed until all conditions are met before issuing the action strategy sequence.

[0058] Step 4: Multi-layered collaborative evolution and proactive stability maintenance. This involves achieving multi-layered collaboration, smooth phase transitions, and proactive stability control during the action execution process.

[0059] During the execution of the action, each layer of the self-generated system evolves independently along the negative gradient direction of the joint free energy functional; the L2 body level monitors the transition conditions of the stage in real time, and triggers the switching of strategies at each level synchronously when the conditions are met; the endogenous spatiotemporal parameters are dynamically adapted with the action stage, and each layer estimates the closed-loop stability margin in real time and actively adjusts the control parameters to maintain the system operating in the stable region.

[0060] The independent variables of the joint free energy functional are the ternary state vector of each layer, the control policy parameters, and the endogenous spatiotemporal parameters. Negative gradient evolution means that the control parameters and spatiotemporal parameters are iteratively updated along the direction of decreasing functional value. Gradient calculation is approximated using the central difference method: a positive and negative symmetric small perturbation with an amplitude of 0.1% of the nominal value of each control parameter to be optimized is applied, and the functional value under each perturbation is calculated. Approximate gradient calculation, balancing computational accuracy and real-time performance.

[0061] As one implementation method, the joint free energy functional is a dimensionless objective function with all terms normalized, and its complete expression is:

[0062] in For the level l operational identity safety margin, Let be the self-consistent deviation rate of the l-th layer ternary closed loop. For the energy consumption of the lth layer, This represents the maximum energy consumption of the baseline at layer l. The weights are initially satisfied with the normalization constraint. After the energy term weights are dynamically adjusted, the weights of the other two terms are scaled synchronously according to their original proportions, maintaining a weight sum of 1. The renormalization formula is:

[0063] The weight of the energy term is dynamically modulated by the remaining battery capacity, and the modulation function is: The SoC represents the battery's state of charge, with a value ranging from 0 to 1. When the battery is fully charged, its weight decreases to prioritize performance; when the battery is low, its weight increases to prioritize the survival of the device.

[0064] The first term of the functional is the stability term, which has the highest weight and priority, ensuring system stability first; the second term is the self-consistency term, which measures the self-consistency deviation of the ternary closed loop; the third term is the energy consumption term, which measures the energy consumption of the system; and the fourth term is the inter-layer interaction term, which measures the coupling cost between layers.

[0065] As one implementation method, each level employs an online perturbation injection method to estimate the normalized relative Lipshitz constant in real time. The specific process is as follows: A zero-mean Gaussian small perturbation is added to the control input in each control cycle; the change in system state output is measured, and the instantaneous gain is calculated; after smoothing by exponential moving average, the result is divided by the benchmark maximum gain to obtain the estimated normalized relative Lipshitz constant. The benchmark maximum gain is the factory-calibrated closed-loop limit gain of the system, corresponding to the steady-state output gain under the maximum allowable input. The normalized value is a relative gain ratio, used to characterize the degree to which the system approaches the stability boundary. When the estimated value exceeds the warning threshold of 0.9, the system actively adjusts the control parameters, including reducing the iteration step size, increasing joint damping, and reducing the intensity of actions, to pull the system back to the stable region; when the estimated value exceeds the intervention threshold of 0.95, a secondary intervention is triggered, freezing non-core rule updates and prioritizing the survival of the system itself.

[0066] As one implementation method, during action execution, the L2 body level continuously monitors the transition conditions of the current stage. When the conditions are met, a stage switching signal is broadcast to all participating limbs, and each limb synchronously switches to the new stage strategy in the next endogenous time cycle. The switching process has no control vacuum period and the delay is less than 1ms.

[0067] Endogenous spatiotemporal parameters dynamically adapt with each movement stage: In the high-dynamic burst stage, the L0-level endogenous time quantum is shortened to 50%~70% of the nominal value, and the spatial resolution of the force-generating joint area is improved to 150%~200% of the nominal value, ensuring control accuracy; In the steady-state maintenance stage, the endogenous time quantum is restored, the spatial resolution is homogenized, and energy consumption is reduced; In the underactuated motion stage, by using the angular momentum conservation constraint based on the endogenous spatiotemporal phase system, the recursive least squares method is used to identify the equivalent mass of each limb in real time and adjust the topological distribution of limb matter, thereby changing the rotational inertia tensor of the system and realizing fine-tuning of posture without the need for a pre-set precise rigid body dynamics model.

[0068] The intrinsic implementation path of angular momentum constraints in underactuated scenarios is as follows: the total angular momentum of the system is calculated by summing the joint velocities of each limb and the real-time estimated equivalent mass; the mapping between limb configuration and rotational inertia adopts a discrete form using the parallel axis theorem, and adjusting the topological distribution of the limbs changes the eigenvalues ​​of the rotational inertia tensor, thereby adjusting the rotational angular velocity. The equivalent mass parameter is identified online from the mapping of material structure to information field state, without the need for preset precise values, and can adapt to load changes and hardware wear.

[0069] Secondly, corresponding to the above method, the present invention also provides a robot motion control device based on a multi-layer nested self-generated system, comprising four functional modules. Each module is interconnected through an internal data bus and can acquire data collected by a multimodal sensor array through the internal bus, corresponding to the four core aspects of the above method. Each module is bound to a corresponding hardware processing unit for implementation: Multi-layer endogenous kernel construction and spatiotemporal anchoring module: used in step S1 of the above method, constructing a four-layer nested self-generated system kernel and completing spatiotemporal phase primitive anchoring; Inter-layer coupling and five-dimensional coherence monitoring module: used to execute step S2 of the above method, establishing an inter-layer interactive free energy coupling mechanism and realizing hierarchical fault location and self-repair; Fractal generation and strategy accumulation module: used to execute step S3 of the above method, completing task decomposition, parent-dominated fractal strategy generation, cognitive verification, and progressive solidification; Multi-layer collaboration and active stability maintenance module: used to execute step S4 of the above method, realizing multi-level collaborative evolution, smooth switching of action stages, and active stability maintenance.

[0070] Thirdly, the present invention provides corresponding computer-readable storage media for electronic devices and robot motion control, covering the entire subject matter of methods, devices, products, and storage media.

[0071] An electronic device includes a processor and a memory storing computer program instructions; when executed by the processor, the computer program instructions cause the processor to perform the steps of any of the above-described robot whole-body motion control methods based on a multi-layered nested self-generated system.

[0072] A computer-readable storage medium for robot motion control is characterized in that it stores computer program instructions thereon, which, when executed by a processor of a robot distributed processing architecture, implement the steps of any of the above-mentioned robot whole-body motion control methods based on a multi-layered nested self-generated system.

[0073] Compared with the prior art, the technological advancements and core beneficial effects of the present invention are as follows:

[0074] Unified control architecture for all scenarios: It adopts a four-layer nested self-generated system unified architecture, covering all scenario motion requirements from steady-state walking to high-dynamic bursts, with a combined task stage switching latency of less than 1ms, completely eliminating the control vacuum period;

[0075] Model-free environment adaptation capability: Based on the endogenous spatiotemporal game sandbox and online parameter identification, no need to pre-set a precise dynamic model, the adaptation time in unknown ground environments is shortened to less than 20ms, and the success rate of high dynamic actions on soft ground is improved by an order of magnitude.

[0076] Endogenous evolution of action strategies: The endogenous emergence of action strategies is achieved through a parent-dominated fractal generation mechanism. After being solidified step by step, the skills are accumulated. The more times the action is executed, the higher the stability and the lower the energy consumption, and the more autonomous evolution capability it has.

[0077] Distributed robust fault-tolerant architecture: The distributed architecture with parallel and nested layers can locate and redistribute tasks within 20ms when a single joint failure occurs, and the whole machine has no single point of failure, significantly improving the failure rate.

[0078] Endogenous spatiotemporal dynamic computing power allocation: Endogenous spatiotemporal parameters are dynamically adapted to the intensity of the action, increasing resolution in high-stress areas and reducing computing power consumption in steady-state areas, resulting in energy savings of more than 35% compared to the global fixed resolution solution;

[0079] New autonomous action generation capability: Based on the ternary collaborative fractal generation framework, it can autonomously generate action strategies for new task scenarios without the need for manual debugging of algorithms and parameters, which greatly reduces the cost of implementing new scenarios. Attached Figure Description

[0080] Figure 1 This is a diagram of a four-layer nested self-generated system kernel whole-body collaborative control architecture according to an embodiment of the present invention. It shows the nesting relationship, hardware carrier, time scale and Lipshitz constant range of the four layers from L0 to L3, and marks the organization forms of parallelism at the same layer and nesting at different levels, as well as the corresponding carriers of the ternary closed loop at each level.

[0081] Figure 2 This is a schematic diagram of the interlayer interactive free energy coupling and five-dimensional coherent monitoring mechanism according to an embodiment of the present invention, showing the uplink governance and downlink solidification path of interlayer bidirectional coupling, as well as the spatial grid construction of the five-dimensional holographic coherent state tensor and the hierarchical fault location logic.

[0082] Figure 3 This is a flowchart of task decomposition and fractal action strategy generation according to an embodiment of the present invention, showing the complete process from task instruction reception, stage division, ternary cooperative fractal generation, triple verification to strategy issuance, and marking the theoretical correspondence of M-topology local bifurcation.

[0083] Figure 4 This is a schematic diagram of the action phase transition simulation and smooth switching mechanism according to an embodiment of the present invention, illustrating the connection relationship of multi-stage actions, the transition simulation verification dimension, and the zero-vacuum synchronous switching principle.

[0084] Figure 5 This is a diagram of the joint free energy functional and multi-level cooperative maintenance structure according to an embodiment of the present invention, illustrating the four main components of the functional, the dynamic weight modulation rule, and the gradient descent mechanism of multi-level cooperative evolution.

[0085] Figure 6 This is a schematic diagram of the high-dynamic motion endogenous spatiotemporal dynamic adaptation mechanism of an embodiment of the present invention, comparing the endogenous temporal quantum differences, spatial resolution distribution differences and computing power allocation logic between the burst phase and the steady-state phase;

[0086] Figure 7 This is a hardware architecture diagram of a whole-body coordinated motion control device according to an embodiment of the present invention, showing the hierarchical dedicated configuration of the distributed processing unit level, the multimodal sensor array, and the interactive connection topology of the internal data bus. Detailed Implementation

[0087] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0088] Multi-layered nested self-generated system kernel construction and spatiotemporal anchoring

[0089] This section constructs a four-layer nested self-generated control closed-loop system as the core of robot motion control. Each layer is a complete self-generated system, possessing an independent ternary closed loop, operational uniformity, and endogenous spatiotemporal system. The layers achieve coordination through holographic recursive mapping.

[0090] Four-layer architecture and hardware mapping. The four-layer architecture adopts a parallel, nested organizational form within the same layer, forming a systematic extension with the two-layer self-generated system architecture of the dexterous hand:

[0091] L0 Joint Level: The system comprises 32 independently driven joints corresponding to 32 parallel sub-autogenous systems. Each joint operates an independent joint-level field-programmable gate array (FPGA) processing unit, responsible for single-joint torque servoing and state sensing. This level features a minimum closed-loop Lipshitz constant, ensuring rapid convergence under disturbances; the operation timescale is sub-millisecond, with a typical value of 0.8ms. A dedicated sensor suite includes joint encoders and motor current sensors, with a sampling frequency of 10kHz.

[0092] L1 Limb Level: Six limbs correspond to six parallel sub-systems, including the left leg, right leg, left arm, right arm, torso, and head. Each limb operates an independent limb-level microcontroller unit, internally nested with a corresponding number of L0 joint-level subsystems, responsible for single-limb multi-joint coordinated control and local posture maintenance. This level's closed-loop Lipshitz constant is moderate, balancing flexibility and stability; the operation timescale is in the millisecond range, with a typical value of 8ms. A dedicated sensor suite includes a six-dimensional force sensor for the limbs and a local inertial measurement unit, with a sampling frequency of 1kHz.

[0093] L2 Body Level: The robot operates as a single self-generated system within a body-level master control processing unit, which internally houses all six L1 limb-level subsystems. These subsystems are responsible for the coordinated control of multiple limbs throughout the body, overall posture estimation, and motion phase switching. The closed-loop Lipshitz constant at this level is close to but less than 1, preserving global optimization degrees of freedom. The operation timescale is in the tens of milliseconds range, with a typical value of 20ms. A dedicated sensor suite includes a full-body master inertial measurement unit, a visual depth camera, and a LiDAR sensor, with a sampling frequency of 30-100Hz.

[0094] L3 Cognitive Level: As a single top-level policy generation engine, it runs within a cognitive-level accelerated processing unit, nested within an L2 body-level system. It is responsible for the stage decomposition of complex tasks, fractal generation of action policies, and hierarchical accumulation. The closed-loop Lipshitz constant at this level is at the same level as L2 but with a lower upper limit, ensuring decision stability; the operation timescale is on the order of hundreds of milliseconds.

[0095] A multimodal sensor array is a collective term for dedicated sensor groups at each level. All sensors are synchronously sampled via a bus to provide sensing input for each closed loop.

[0096] Global stability constraint design. Based on the small gain theorem, a hierarchical stability system is designed to satisfy the functional positioning of each layer while ensuring global convergence after multi-layer coupling.

[0097] The single system is stable in both input and output: the true Lipshitz constant of each closed loop is strictly less than 1, corresponding to a finite L2 gain. The single system satisfies input-output stability, and the convergence of the single system can be independently proven by Banach's fixed-point theorem.

[0098] Feedback interconnection topology: Adjacent layers form a feedback interconnection structure, with the upper layer control output serving as the lower layer input and the lower layer state feedback serving as the upper layer input, thus forming a standard feedback interconnection system.

[0099] The low-gain condition must be satisfied: the bidirectional coupling channel gain between adjacent layers must be symmetrical, i.e. Satisfying the small gain condition According to the small gain theorem, if the closed-loop gain product of two input-output stable subsystems interconnected through feedback is less than 1, then the overall system is input-output stable.

[0100] Global convergence guarantee: The composite closed-loop gain after multi-layer cascading is the product of the gains of each layer and the coupling channel, which is strictly less than 1, and it converges globally to a unique fixed point.

[0101] This design retains the optimization freedom of the high-level system while strictly ensuring global stability through small gain constraints, thus completely resolving the logical contradiction between traditional layer-by-layer incremental design and the stability of compression mapping.

[0102] Spacetime Phase Element Anchoring Rules

[0103] The endogenous time dimension is bound to the energy mechanism: the time for the ternary closed loop to complete one full iteration is the minimum quantum of endogenous time. The rate of energy metabolism, i.e., the rate of change in power consumption, determines how quickly time passes—the higher the rate of energy metabolism, the shorter the closed-loop iteration cycle, and the smaller the endogenous time quantum. The endogenous time cycles between levels satisfy an integer frequency division relationship, with the cycle of lower layers being 1 / N of that of higher layers, ensuring integer alignment for temporal coordination.

[0104] Endogenous spatial dimensions are bound to material structure: the spatial scale of the smallest interactive unit within the system is the smallest quantum of endogenous space, and the frequency of interaction between limbs determines the local resolution. The interlayer spatial field satisfies a holographic projection relationship, with the lower-layer spatial field being a high-resolution projection of the higher-layer spatial field.

[0105] Endogenous interphase dimension and information field binding: A holographic mapping mediation field connecting spacetime is constructed in Hilbert space, providing an information interaction basis without external protocols for multi-party games. The interphase field satisfies normalization constraints, and the Pearson correlation coefficient between its probability gradient and the information field gradient is not less than 0.95.

[0106] The mathematical form of the holographic recursive mapping is a column orthogonal projection matrix that satisfies the inner product-preserving constraint of the eigenspace. The initial projection matrix is ​​obtained based on the kinematic calibration, and during system operation, it can be autonomously fine-tuned through the mapping from material structure to information field state to adapt to the kinematic parameter drift caused by mechanical wear.

[0107] Hardware integrity verification and boot process

[0108] When each layer of the self-generated system bootstraps, it first performs a hardware integrity check: reading the current physical parameters of each sensor and comparing them with the rigid core layer reference parameters stored in the read-only memory. If the deviation exceeds the 5% tolerance, the system determines that the hardware has been changed, automatically re-acquires parameters, and updates the reference values. After the bootstrap is completed, each layer of the ternary closed loop begins to operate independently, and the multimodal sensor array synchronously outputs sensing data.

[0109] When using a self-generated sensor system, the sensor simultaneously completes its own ternary closed-loop construction and self-calibration, outputting high-quality data that has undergone intrinsic calibration, further improving the system's sensing accuracy.

[0110] Interlayer interaction free energy coupling and five-dimensional coherent state monitoring

[0111] This section establishes a bidirectional coupling mechanism between layers, and uses a five-dimensional holographic coherent state tensor to achieve real-time monitoring of the entire system state and precise fault location.

[0112] Interlayer interaction free energy mechanism

[0113] Interlayer interaction can achieve bidirectional dialectical coupling, serving as both a driving signal for upward governance and a transmission channel for downward consolidation. Its complete expression is dimensionless, with all terms mapping to the interval [0,1]. Among them, the correlation strength of information field state The Pearson correlation coefficient is calculated using the eigenvectors of the two-layer information field, with a value ranging from [0,1]. A larger value indicates a stronger correlation between the two control rules. Self-referential coverage... is the completeness rate of the meta-representation of the higher-level rules to the lower-level rules, with a value range of [0,1]. The larger the value, the stronger the control ability of the higher-level rules over the lower-level rules.

[0114] Coupling coefficient differentiation configuration: L0-L1 layer Emphasis is placed on the strength of information field-state correlation to ensure accurate transmission of control commands; L1-L2 layer interlayer Balanced allocation, taking into account both coordination and physical autonomy; L2-L3 level It emphasizes self-referential coverage to ensure the consistency of the overall strategy.

[0115] The complete implementation mechanism of uplink governance is as follows: when the low-level interaction free energy exceeds the warning threshold for three consecutive cycles, an uplink governance request is triggered; the high-level layer executes corresponding intervention actions according to the anomaly level: mild anomalies adjust the coupling coefficient, moderate anomalies issue compensation rules, and severe anomalies trigger a new fractal generation strategy; after governance is completed, the interaction free energy falls back to the normal range, and the governance process terminates. Uplink governance corresponds one-to-one with the fault self-healing process, with different levels of governance actions triggered by anomalies of different dimensions.

[0116] Construction of 5D holographic coherent state tensor

[0117] The whole-body five-dimensional holographic coherent state tensor is a fifth-order tensor with the following structure: The three-dimensional space is constructed using a body space grid: the robot's entire physical space is divided into... A three-dimensional grid is used, where each joint is mapped to a corresponding grid cell based on its actual physical position. Each grid cell stores the feature data of the corresponding joint. The grid resolution is determined by the joint density: all 32 joints of the body are uniformly mapped to a 4×4×2 spatial grid, with each grid cell being approximately 1 / 4 the size of the fuselage. This ensures that each grid contains at least one joint, and adjacent joints are placed in different grids as much as possible to avoid spatial overlap. This structure allows the three spatial dimensions to correspond to the three coordinate axes of real physical space, possessing clear physical meaning and topological correspondence.

[0118] The feature dimension contains five channels, corresponding to matter structure, energy mechanism, information field state, time synchronization, and spatial alignment, respectively; the time dimension is a sliding observation window that covers temporal evolution information.

[0119] Global coherence is calculated using the ratio of tensor Frobenius norms, with an upper bound constraint added to ensure physical consistency. Reference Tensor Data was collected under the system's initial steady-state conditions and used as a reference benchmark for coherence. The rule for calculating the dimensional coherence is as follows: retain the target feature dimension, and shrink all other dimensions, that is, calculate the sum of squares of all elements in the other dimensions and take the square root to obtain the Frobenius norm of the target dimension subtensor, and then compare it with the norm of the corresponding dimension of the benchmark tensor.

[0120] Cross-validation fault location mechanism

[0121] A 5th-order symmetric cross-validation matrix is ​​constructed, with matrix elements being the temporal correlation coefficients of pairwise coherence, used to identify the temporal causal path of fault propagation. Fault localization follows the temporal causal rule: the anomaly in the root cause dimension occurs at least two corresponding level endogenous time quanta earlier than the derived dimension; simultaneous anomalies in multiple dimensions are identified as common-source faults.

[0122] The fault localization process is divided into two levels: the first level determines the fault dimension type through five-dimensional cross-validation; the second level locates the specific joint grid corresponding to the anomaly by using the local anomaly amplitude in the tensor space dimension, thereby pinpointing the specific joint. This hierarchical localization mechanism enables precise fault localization from the whole body to a single joint.

[0123] The tiered handling mechanism corresponds one-to-one with the fault dimensions, ensuring that the system automatically triggers the corresponding repair process when an anomaly occurs, without manual intervention. In single-joint fault scenarios, the system can complete the location and strategy regeneration within 20ms, maintaining the stability of the entire machine and preventing it from tipping over.

[0124] Parent-dominated fractal generation action strategy endogenous game

[0125] This part is the core innovation of the invention. Based on the ternary collaborative fractal generation theory, it realizes the endogenous emergence of action strategies without the need for manual pre-setting of trajectories and control algorithms.

[0126] Task decomposition and phase division

[0127] When the L3 cognitive level receives a complex task instruction, it first performs semantic-physical mapping decomposition:

[0128] Extract core semantic elements from the instructions, such as the target of the operation, the start point, the end point, and the operation type.

[0129] Retrieve relevant action strategy templates that have been accumulated in the long-term rule base;

[0130] Detect the contact state and energy continuity conflict between adjacent strategies, and automatically insert a transition phase;

[0131] For each action phase, clearly define the sub-goals, participating limbs, stability constraints, and energy constraints, and set quantifiable transition conditions.

[0132] Basic safety rules, as rigid operational uniformity content, are preset in the system's initial rule base by the designer, serving as inviolable boundaries for fractal generation. Based on this, the system gradually expands the rule base by repeatedly executing autonomous optimization transition strategies.

[0133] Ternary Collaborative Fractal Generation Mechanism

[0134] For each action stage, a three-element collaborative framework of "parent system-mother system-environment system" is constructed, and a parent-led generation modality is adopted to generate action strategy subsystems.

[0135] The parent system is at the L3 cognitive level, providing action meta-rules and goal guidance, and determining the functional positioning and rule direction of the subsystems;

[0136] The parent system is an L1-level self-generating system that participates in the limbs, providing material topological resources and energy supply, and providing physical carriers for the subsystems;

[0137] The environmental system, based on the current task scenario and physical environment, applies selection pressure through sensing data to shape the environmental adaptability of subsystems.

[0138] The complete generation process is divided into five stages: generation triggering, initial construction, bootstrapping evolution, generation verification, and rule accumulation, corresponding to the coordinated scheduling of four core operators.

[0139] The motion element rule template contains three sets of basic parameters: first, the force application mode, which is divided into three types: explosive, continuous, and buffered, representing the temporal envelope of the stress torque output; second, the coordination phase, which defines the temporal offset of each participating joint to ensure phase matching of multi-joint coordination; and third, the safety constraints, which include three types of hard constraints: maximum torque, maximum angular velocity, and joint angle limit, serving as insurmountable boundaries for fractal generation.

[0140] The Chern number, representing operational identity, is a discrete topological criterion for the operational identity of a subsystem. It characterizes the degree of connectivity of the subsystem's topology; a higher value indicates stronger coupling between joints and better overall subsystem integrity. In engineering, it is calculated using the algebraic connectivity of the joint topological Laplace matrix. The algebraic connectivity gradually increases and eventually stabilizes during the iteration process, reflecting the subsystem's generation process from loose to tight. When the Chern number stabilizes in the non-zero interval and fluctuates by less than 5%, it signifies that the subsystem possesses independently identifiable operational identity, and fractal generation is complete.

[0141] Endogenous Spatiotemporal Game Sandbox Operation Mechanism

[0142] The game sandbox constructs an endogenous spatiotemporal system through four operators, providing evolutionary space for fractal generation. The four operators are respectively responsible for the generation of time reference, dynamic adjustment of spatial resolution, construction of interphase intermediate fields, and three-dimensional closed-loop fusion, which together constitute a complete endogenous spatiotemporal phase operating environment.

[0143] In the game sandbox, the fundamental physical laws of Newtonian mechanics, conservation of momentum, and conservation of angular momentum are embedded in the payoff function as equality constraints. The physical derivation process is based on fundamental mechanical relationships: conservation of momentum and angular momentum are hard constraints that must be satisfied, and Newton's second law serves as the differential equation for state derivation. Dynamic parameters such as mass and moment of inertia are estimated in real time from sensing data, with nominal values ​​used as initial values ​​and iteratively corrected online during operation; therefore, there is no need to pre-set an accurate dynamic model.

[0144] The game payoff function is a local simplification of the joint free energy functional in a single-stage game scenario. The negative value of the payoff function corresponds to the local free energy in the current stage. Maximizing the payoff is equivalent to minimizing the local free energy, which is consistent with the global optimization objective.

[0145] The game employs an iterative optimal response algorithm, where each participant updates its strategy along the negative gradient of its own payoff function until a Nash equilibrium is reached. The convergence criterion satisfies both a free energy change threshold and a safety margin lower bound, ensuring that the generated strategy is both optimal and safe.

[0146] Triple verification and downward solidification

[0147] The fractal generation strategy must undergo triple security checks to ensure that it does not disrupt the overall operational consistency of the system.

[0148] Semantic entropy test: Discretize the rule base into a set of action primitives, calculate the Shannon entropy with the activation frequency of each primitive as the probability distribution, and judge it as failing if the entropy increases beyond the threshold;

[0149] Self-referential coverage test: Generate a set of meta-questions containing three categories: applicable scenarios, constraints, and performance boundaries. If the query matching degree is lower than the threshold, it is judged as failing.

[0150] Cross-scenario generalization test: The strategy is tested on a standard benchmark scenario set. If the performance coefficient of variation exceeds the threshold and there is no corresponding environmental parameter explanation, it is judged as failing.

[0151] The testing threshold is dynamically adjusted according to the system's safety margin, encouraging exploration when the system is stable and ensuring safety when the system is critical, thus achieving a dynamic balance between exploration and stability.

[0152] The validated strategies are solidified hierarchically, gradually evolving from slow, cognitive-level decision-making to fast, joint-level responses. The physical implementation of this solidification is as follows: L3→L2 rules are written to the main control memory rule base; L2→L1 rules are decomposed and distributed to the limb MCUs; and L1→L0 rules are solidified into FPGA lookup tables. The lookup table format is a three-dimensional mapping table of torque command-joint angle-joint angular velocity, stored in the block random access memory of the joint-level processing unit. After repeated execution of frequently used actions, response latency is significantly reduced, and energy consumption and success rate are continuously optimized.

[0153] After all phase strategies are generated, inter-phase transition simulations are performed to ensure the smoothness and continuity of action transitions. The transition simulation is based on an endogenous digital twin and verifies continuity from three dimensions: attitude, contact, and energy. If it fails, a transition phase is automatically inserted to avoid instability caused by switching shocks.

[0154] Multi-layered Co-evolution and Active Stability Maintenance

[0155] This section enables real-time collaboration and proactive stability control during the execution of actions.

[0156] Joint Free Energy Co-evolution

[0157] The independent variable space of the joint free energy functional contains three types of variables: the ternary state vector of each layer, the control policy parameters, and the endogenous spatiotemporal parameters. Gradient calculation is performed on the control parameters and the spatiotemporal parameters. In engineering, the gradient of the functional with respect to the control parameters is approximated by the small perturbation injection method, and the parameters are updated along the negative gradient direction to achieve gradient descent iteration.

[0158] Each layer of the self-generated system independently calculates its own operational consistency safety margin and evolves autonomously along the negative gradient direction of the joint free energy functional. Inter-layer coordination is achieved through interactive free energy. The joint free energy functional integrates four major optimization objectives: stability, self-consistency, energy consumption, and inter-layer coupling. All terms are normalized to be dimensionless, and the optimal trade-off between multiple objectives is achieved through dynamic adjustment of weights.

[0159] The weight of the energy term is dynamically modulated according to the remaining battery capacity, and the modulation function is: The SoC (System-on-Chips) represents the battery's state of charge, ranging from 0 to 1. When the battery level is above 80%, the weight is approximately 0.6 times the baseline value, prioritizing performance; when the battery level is below 20%, the weight is approximately 1.4 times the baseline value, prioritizing the battery's survival, achieving a dynamic balance between performance and battery life. After adjusting the energy item weights, the stability and self-consistency items are scaled proportionally to maintain a total weight of 1.

[0160] Smooth transition between stages and dynamic spatiotemporal adaptation

[0161] Phase switching is triggered by a unified L2 body-level broadcast, with each limb switching synchronously in the next control cycle, achieving zero-vacuum transition. The switching latency is less than 1ms, far superior to the tens of milliseconds of switching latency in traditional architectures.

[0162] Endogenous spatiotemporal parameters dynamically adapt to different action phases: During the burst phase, the L0-level endogenous time quantum is shortened to 50%~70% of the nominal value, and the spatial resolution of the force-generating joint area is improved to 150%~200% of the nominal value, ensuring control accuracy; during the steady-state phase, the resolution is reduced to save energy. In underactuated scenarios, attitude fine-tuning is achieved through angular momentum conservation constraints: The equivalent mass of each limb is identified in real time using the recursive least squares method, and the rotational inertia tensor is changed by adjusting the limb configuration, thereby adjusting the rotational angular velocity. This does not rely on rigid body dynamics models and naturally adapts to environments where the model is unknown.

[0163] Active stability maintenance mechanism

[0164] Each level employs an online perturbation injection method to estimate the normalized relative Lipshitz constant in real time. The perturbation amplitude is carefully designed to be small enough not to affect normal control performance, and large enough to produce measurable state changes. Experiments show that the impact of perturbation injection on control performance is less than 1%, which is completely acceptable.

[0165] The Lipshitz constant is calculated using a normalization method: the instantaneous gain is divided by the reference maximum gain to obtain the relative value. The reference maximum gain is the factory-calibrated closed-loop limiting gain, corresponding to the steady-state output gain of the system under the maximum permissible input. The normalized value is the relative Lipshitz constant of the closed-loop mapping, characterizing the degree to which the system approaches the stability boundary.

[0166] The two-tiered threshold system corresponds to two-tiered responses: the early warning threshold triggers parameter fine-tuning, while the intervention threshold triggers rule freezing. This mechanism enables the system to proactively detect instability trends and intervene in advance, preventing failures. Compared to traditional post-failure repair architectures, stability is significantly improved.

[0167] All embodiments of this invention are based on a 32-joint humanoid robot prototype platform, with the following hardware configuration:

[0168] L0 Joint Level: Xilinx Artix-7 Field Programmable Gate Array, 1 per joint, 32 in total, operating at 200MHz;

[0169] L1 limb level: STMicroelectronics STM32H750 microcontroller unit, 1 per limb, 6 in total, operating frequency 480MHz;

[0170] L2 body level: NVIDIA Jetson Orin NX main controller, operating frequency 1.5GHz;

[0171] L3 cognitive level: Xilinx Kria K26 field-programmable gate array, operating frequency 200MHz;

[0172] Multimodal sensor array: 17-bit high-precision joint encoder, six-dimensional force sensor, 9-axis inertial measurement unit, distributed tactile array, visual depth camera, LiDAR, with a sampling frequency of not less than 1kHz;

[0173] Communication bus: 100 Mbps Ethernet, synchronized using the IEEE 1588 Precision Time Protocol with a synchronization accuracy of no more than 1 microsecond.

[0174] Example 1: Deployment and Parameter Calculation of a Multi-Layer Nested Kernel for a 32-Joint Humanoid Robot

[0175] This embodiment fully demonstrates the industrial deployment process of a four-layer self-generated system kernel, including five stages: hardware deployment, parameter calibration, closed-loop verification, inter-layer mapping calibration, and benchmark tensor generation. All calculation processes provide complete input, output, and numerical results.

[0176] Step 1, Hardware Deployment and Clock Synchronization

[0177] Thirty-two joint-level processing units are installed on their respective joint drive boards, six limb-level processing units are installed in their respective limb control boxes, and the body-level main control unit and cognitive-level acceleration unit are installed in the torso main control cabin. All nodes are connected via a 100 Mbps Ethernet star topology and connected to the core switch.

[0178] After deployment, system-wide clock synchronization was performed: the precise time protocol master node was started at the body level, and each slave node synchronized and calibrated its local clock. Measurements were taken under the following conditions: 32 nodes across the entire system were synchronized, with a synchronization period of 1ms, at 30% load on a 100Mbps network. The maximum time synchronization error for the entire system was 0.8 microseconds, meeting the design requirements.

[0179] Step 2: Calibration of L0 ankle joint closed-loop parameters and calculation of Lipschitz constant

[0180] Taking the left ankle joint as an example, single-joint autogenous system calibration and closed-loop stability verification were performed.

[0181] Rigid core layer parameter acquisition: Ankle joint physical parameters, including motor torque constant, were acquired using factory-calibrated equipment. Reduction ratio Angle sensor resolution Maximum angular velocity The above parameters are written into the joint-level read-only memory as a reference for the rigid core layer.

[0182] Initial ternary closed-loop construction:

[0183] Material structural state: including joint angles angular velocity ,temperature Three physical quantities, sampling frequency 1kHz;

[0184] Energy mechanism state: motor winding current Bus voltage The data is collected in real time by the motor drive chip;

[0185] Information field state: torque-angle mapping coefficient matrix, initial values ​​are set based on classical proportional-integral-differential parameters, proportional coefficients Integral coefficient .

[0186] Reference maximum gain calibration: Under the maximum permissible torque input, the steady-state angular velocity output of the system is measured to obtain the reference maximum gain. This reference value is the steady-state angular velocity output gain of the system under the maximum permissible torque input, obtained from factory calibration, and serves as the normalization reference for the normalized relative Lipschitz constant.

[0187] Closed-loop true Lipschitz constant calculation: The online perturbation injection method is used for calculation, and the complete calculation process is as follows:

[0188] No. During the control cycle, a zero-mean Gaussian disturbance is injected into the torque control input. The disturbance amplitude is 0.1% of the rated torque, which will not affect normal operation;

[0189] Collect the change in joint angular velocity in the next cycle ;

[0190] Calculate instantaneous gain: ;

[0191] Normalization process: ;

[0192] Exponential moving average smoothing: smoothing coefficient Previous period estimate ;

[0193] Current estimate: Approximately 0.125.

[0194] The calculated result is the true Lipschitz constant. It is within the design range of [0.1, 0.3], satisfying the joint-level stability constraints.

[0195] Step 3: L1 left-leg stage closed-loop calibration and low-gain condition verification

[0196] The left leg comprises three L0 joint subsystems: ankle, knee, and hip, forming a limb-level self-generating closed loop.

[0197] Limb-level ternary state:

[0198] Material structure: angle vectors of 3 joints, six-dimensional force signal of the foot, and leg posture angle;

[0199] Energy mechanism state: energy distribution matrix of 3 joints, total power consumption;

[0200] Information field state: single-leg gait coordination rules, support phase moment distribution coefficient.

[0201] Calculation of the closed-loop true Lipschitz constant: Using the same perturbation injection method, a perturbation is injected into the leg torque distribution parameters, and the end-effector position change is measured. After smoothing, the limb-level closed-loop true Lipschitz constant is obtained. It falls within the design range of [0.3, 0.5].

[0202] Low-gain condition verification: Bidirectional coupling gain from L0 to L1 Verify the small gain condition: If the conditions are met, the global input and output of the L0-L1 feedback interconnection system are stable.

[0203] Step 4, Inter-layer holographic recursive mapping calibration

[0204] Calculate the spatial holographic projection matrix from L0 to L1 Based on the leg kinematics model, the joint spaces of the three joints are mapped to the Cartesian space of the limb's end effector, generating an initial column orthogonal projection matrix. The matrix size is 3×3, and the initial values ​​are as follows: Each column of the matrix is ​​orthogonal, satisfying The column orthogonality constraint ensures that the inner product of the core feature subspace is preserved. During system operation, this matrix can be autonomously fine-tuned through the mapping from material structure to information field state to adapt to the kinematic parameter drift caused by mechanical wear. After fine-tuning, column orthogonality is maintained through Schmidt orthogonalization.

[0205] Step 5: Generation of the baseline five-dimensional holographic coherent state tensor

[0206] Data Acquisition Conditions: The robot stands steadily on a level, hard surface; all sensors are calibrated; and the system is in a no-load steady state. Steady-State Determination: Steady state is determined when the global coherence fluctuation is less than 0.01 for 100 consecutive frames. Acquisition Process: 1250 frames of data are continuously acquired for 1 second; the average is then used to generate the baseline tensor. .

[0207] Spatial Dimension: The 32 joints of the whole body are mapped to a 4×4×2 three-dimensional spatial grid, corresponding to the spatial division of the body in front, back, left, right, up, and down directions;

[0208] Feature dimensions: 5 channels, corresponding to matter, energy, information, time, and space respectively;

[0209] Time dimension: Sliding window length This corresponds to an 8ms time scale.

[0210] Calculate the Frobenius norm of the reference tensor This serves as the baseline value for subsequent global coherence calculations. Under steady-state conditions, the initial global coherence value is 1.0, and the coherence values ​​for all five sub-dimensions are also 1.0.

[0211] Step 6, Verification of Adaptive Fine-tuning of Mechanical Wear

[0212] After the system simulated mechanical wear conditions for 100 hours, the joint kinematic parameters drifted by 3%. The holographic projection matrix autonomously fine-tuned and corrected the drift, and the end-effector positioning error decreased from 2.1 mm to 0.3 mm, verifying the adaptive fine-tuning effect.

[0213] Example 2: Control Flow of Backflip High Dynamic Movement

[0214] This embodiment fully demonstrates the entire process of a backflip action, from task input to fractal generation, verification and execution, and feedback optimization, including numerical calculations of key formulas and intermediate results.

[0215] Step 1: Task Input and Stage Decomposition

[0216] The L3 cognitive level receives the task instruction: "Perform a backflip in place." It then performs a semantic-physical mapping decomposition, resulting in four ordered action stages:

[0217] Preparatory squatting phase: The sub-goal is to lower the center of gravity to the lowest point, with the knee joint angle less than 40 degrees, and the limbs involved are both legs, arms, and torso;

[0218] Explosive take-off phase: The sub-goal is to obtain maximum vertical velocity and backward angular momentum, and the limbs involved are both legs and both arms;

[0219] Aerial attitude phase: underactuated state, completes 360-degree flip, attitude deviation is less than 5 degrees, and the whole body is involved;

[0220] Landing cushioning phase: Both feet land smoothly, the impact force is distributed throughout the body, the center of gravity falls within the supporting polygon, and the entire body participates in the impact.

[0221] Each stage is subject to corresponding stability and energy constraints, and quantization transition conditions are set for adjacent stages.

[0222] Step 2, Parent-dominated fractal generation calculus during the burst takeoff phase

[0223] Taking the core burst jump phase as an example, the iterative process of parent-dominated fractal generation is fully demonstrated.

[0224] Ternary initialization:

[0225] Parent system: L3 cognitive level, injected with jump meta-rule template, information injection intensity 0.9 critical value, belongs to parent-dominated modality;

[0226] Parent system: Dual-legged L1 level self-generating system, pre-allocated 80% of computing resources and peak energy supply;

[0227] Environmental system: Current ground environment, with initial ground stiffness estimated to be 1.0 times that of the baseline hard ground.

[0228] Iterative optimization process: An iterative optimal response algorithm is used, with the payoff function being the negative of the joint free energy. Maximizing the payoff is equivalent to minimizing the free energy. Convergence occurs after 21 iterations. The complete calculations for the first three iterations are shown below.

[0229] First iteration:

[0230] Initial strategy: knee joint torque distribution coefficients: left knee 0.5, right knee 0.5; ankle joint torque distribution coefficient: 0.5.

[0231] Physical deduction: Based on the conservation of momentum and Newtonian mechanics, the takeoff velocity and angular momentum are deduced by combining initial parameters;

[0232] Prediction result: Vertical velocity Backward angular momentum ;

[0233] Payoff function calculation: Weights set to vertical velocity 0.5, angular deviation 0.3, and safety margin 0.2; Payoff function value ;

[0234] Strategy update: Adjust the torque distribution along the gradient direction of the payoff function. Update the knee joint distribution coefficient to 0.52 for the left knee and 0.48 for the right knee, and update the ankle joint coefficient to 0.53.

[0235] Second iteration:

[0236] Updated strategy: Left knee 0.52, right knee 0.48, ankle joint 0.53;

[0237] Physical deduction: Vertical velocity angular momentum ;

[0238] Payment function value This represents an 8.1% increase compared to the first time.

[0239] Continue updating the strategy along the gradient.

[0240] 21st iteration:

[0241] Final strategy: Left knee 0.58, right knee 0.42, ankle 0.61, hip 0.45;

[0242] The change in the payoff function over 10 consecutive iterations is less than , satisfying the convergence criterion;

[0243] Final payment function value The safety margins at all levels are higher than the thresholds;

[0244] The game converges, yielding a Nash equilibrium strategy.

[0245] Generate verification: Calculate the operational identity Chern number of the action strategy subsystem: The subsystem involves three joints: ankle, knee, and hip. Construct a 3rd-order joint topological Laplacian matrix and calculate its second smallest eigenvalue, i.e., algebraic connectivity. ,Right now After 10 consecutive iterations of monitoring, the Chern number stabilized in the range of 1.78 to 1.85, with a fluctuation range of 3.8%, which is less than the 5% criterion, thus satisfying the stability criterion and completing the fractal generation.

[0246] The entire fractal generation process took 87ms, which meets the real-time requirements.

[0247] Step 3, General Cognitive Triple Test Calculation

[0248] A triple check is performed on the generated takeoff strategy. The calculation process is as follows:

[0249] Semantic entropy consistency test:

[0250] The existing rule base contains 128 action primitives, with information entropy. ;

[0251] After introducing the new strategy, two new action primitives are added: information entropy. ;

[0252] Entropy increase ;

[0253] Current L3 safety margin Dynamic threshold ;

[0254] 0.03 < 0.1825, the test is passed.

[0255] Self-reference coverage test:

[0256] Generate 100 meta-questions in 3 categories, including 30 applicable scenarios, 40 constraints, and 30 performance boundaries;

[0257] A total of 94 results were found that met the matching criteria.

[0258] Self-referential coverage ;

[0259] Dynamic threshold ;

[0260] 0.94 > 0.79, the test is passed.

[0261] Cross-scenario generalization consistency check:

[0262] Tests were conducted in three benchmark scenarios: hard ground, soft ground, and sloping ground.

[0263] The takeoff success rates were 98%, 92%, and 90%, respectively, with a coefficient of variation of 12%.

[0264] The deviation can be reasonably explained by the change in ground stiffness;

[0265] Dynamic threshold ;

[0266] 0.12 < 0.255, the test is passed.

[0267] All three tests passed, and the strategy was marked as effective.

[0268] Step 4: Active stabilization control calculation during the landing buffer phase

[0269] The online Lipschitz estimation and active adjustment process at the moment of landing are displayed.

[0270] At the moment of impact, the L0 ankle joint closed loop is subjected to a large disturbance, and the normalized relative Lipschitz constant is calculated in real time:

[0271] The instantaneous gain ratio corresponding to the change in state caused by the landing impact is 0.82;

[0272] Smoothed estimate ;

[0273] 0.92 > 0.9 warning threshold, triggering a Level 1 warning.

[0274] The system proactively performs stability adjustments:

[0275] The iteration step size was reduced by 20%;

[0276] The joint damping coefficient was increased by 15%;

[0277] The intensity of the action is limited to 80% of the original plan;

[0278] The adjusted normalized Lipshitz estimate for the next period fell back to 0.78, returning to the stable range.

[0279] The entire adjustment process is completed within a control cycle of 0.8ms, achieving proactive stabilization.

[0280] Step 5, Step-by-Step Curing Delay Verification

[0281] After the strategy is verified, it is implemented and solidified step by step. The actual measured response latency data for each level is as follows:

[0282] L3 level generation call latency: 120ms, requires triggering fractal generation and verification;

[0283] L2 rule base call latency: 15ms, reading global parameters from the main control rule base;

[0284] L1 level limb recall latency: 2ms, reading coordination parameters from limb unit;

[0285] L0 level lookup table call latency: 0.3ms, directly reads FPGA fixed parameters to output torque.

[0286] After being solidified step by step, the response delay of commonly used actions was reduced from 120ms to 0.3ms, realizing the transformation from cognitive decision-making to instinctive response.

[0287] Step 6, Action Execution Results and Performance Data

[0288] Result of a complete backflip:

[0289] The takeoff height was 0.62m, and the takeoff time was 0.71s;

[0290] The landing attitude deviation was 3.2 degrees, which is less than the design threshold of 5 degrees;

[0291] The peak impact force on the ankle joint is 3200N, which meets the design expectations.

[0292] The lowest global coherence value throughout the entire process is 0.92, which is always higher than the steady-state threshold of 0.9.

[0293] The initial success rate was 70%, and after three adaptive updates of ground parameters, the stable success rate reached 92%.

[0294] After 100 executions, the strategy was solidified to L1 and L0 levels, improving the stability of implementation to 98% and reducing energy consumption by 15%.

[0295] Example 3: Performance Comparison and Verification with Traditional Solutions

[0296] The comparative scheme adopts the industry-standard rigid body model predictive control framework, with a prediction time of 10 steps and a control cycle of 1ms. It runs on the same Jetson Orin NX hardware platform as this scheme, uses a nominal rigid body dynamics model, and has no online adaptive mechanism.

[0297] In this embodiment, the scheme of the present invention and the traditional model predictive control scheme were compared and tested on the same hardware platform. Each experiment was repeated 100 times and the average value was taken. The data differences were tested by independent samples t test, and the significance level was 0.01, which is statistically significant. The test results are detailed in Table 1.

[0298] Table 1: Performance Comparison Table

[0299] Combined task phase switching delay (ms) 0.8±0.1 65±8.2 Reduced by 98.8% Gait adaptation time to unknown terrain (ms) 18±2.3 520±45 Shortened by 96.5% Backflip stability success rate on soft ground (%) 92±2.5 35±4.8 An increase of 162.9% Success rate of first attempt at a backflip on soft ground (%) 70±5.0 12±3.2 An increase of 483.3% Walking task completion rate (%) after single joint failure 87±3.1 0 - Steady-state walking average power consumption (W) 156±12 240±18 Reduced by 35.0% 1000-hour trouble-free rate (%) 99.7±0.1 76.5±2.2 Increased by 30.3%

[0300] Note: Traditional model predictive control schemes cannot redistribute torque after a single joint failure, directly causing the entire machine to tip over, resulting in a 0% task completion rate.

[0301] Extreme operating condition tests show that the performance degradation of this solution is significantly less than that of the traditional solution under low temperature of -20 degrees Celsius, high temperature of 60 degrees Celsius, and 50Hz vibration environment, demonstrating obvious robustness advantages.

[0302] Example 4: Verification of Single Joint Fault Location and Self-Repair Process

[0303] This embodiment demonstrates the complete localization and self-repair process of a fault in the left ankle joint encoder, verifying the industrial application effectiveness of the fault localization function.

[0304] Fault settings

[0305] At 100ms, the encoder signal of the left ankle joint is manually disconnected to simulate a sensor malfunction.

[0306] Fault detection and localization process

[0307] At 100.8ms (after one L0 cycle), the coherence of the material structure dimension of the left ankle joint decreased from 1.0 to 0.62, while the other dimensions remained normal;

[0308] At 102.4ms (after 2 L0 cycles), the coherence of the energy mechanism dimension drops to 0.78, while the information field state dimension remains normal.

[0309] Five-dimensional cross-validation matrix analysis: The material structure dimension is abnormally earlier than the energy dimension by two L0 level endogenous time quanta, and the material structure dimension is determined to be the root cause;

[0310] Spatial dimension localization: The abnormal amplitude of the grid corresponding to the left ankle joint is the largest, and the faulty joint is located as the left ankle joint;

[0311] The total positioning time was 18.2ms. This time corresponds to approximately 23 L0 control cycles (0.8ms / cycle), which is much less than the body-level gait cycle of approximately 500ms, thus meeting the real-time requirements.

[0312] Self-repair execution

[0313] The self-repair mechanism is triggered immediately after the location is determined:

[0314] The faulty encoder signal was shielded, and the ankle joint angle was estimated using the motor current and the status of adjacent joints.

[0315] The leg moment coefficient is redistributed to reduce the load on the left ankle joint and increase the compensatory force of the knee and hip joints;

[0316] Adjust gait parameters and appropriately reduce walking speed to ensure walking stability.

[0317] Repair effect

[0318] After repair, the robot continued to perform walking tasks with a stable gait and no tipping. The walking speed decreased by 15%, and the task completion rate was 87%, which verified the system's robustness under single joint failure.

[0319] Supplementary Verification: Energy Dimension Fault Localization. A simulated voltage drop fault in the left leg motor revealed an initial anomaly in the energy mechanism dimension's coherence. Following this, anomalies occurred in the material structure dimension after two L1-level endogenous time quanta. The system determined the energy dimension as the root cause, locating it to the left leg motor and triggering interlimb energy scheduling compensation. The localization took 21ms.

[0320] The above embodiments are intended to illustrate the full-scene adaptability, engineering feasibility, and significant beneficial effects of the present invention. Those skilled in the art can make non-creative modifications and scene adaptations to the embodiments based on the core principles of the present invention, all of which fall within the protection scope of the present invention.

Claims

1. A method for controlling the whole-body motion of a robot based on a multi-layered nested self-generated system, characterized in that, The process includes the following steps: S1, constructing a multi-layered nested self-generated system kernel and completing spatiotemporal phase primitive anchoring: constructing L0 joint level, L1 limb level, L2 body level, and L3... The cognitive-level four-layer nested self-generated control closed loop adopts a parallel architecture within the same layer and a hierarchical nested architecture. Each closed loop corresponds to an independent processing unit and a dedicated multimodal sensor group for the robot. Each layer maintains operational identity through a ternary bidirectional recursive mapping of information field state, energy mechanism, and material structure. This operational identity means that the self-generated system is continuously identified as its own core attribute in the diachronic evolution, protected by topological invariants and maintained by the ternary closed loop. Each closed loop has an independent endogenous time dimension, endogenous space dimension, and endogenous phase dimension. The endogenous spatiotemporal phase dimension refers to the dynamic coordinate system generated endogenously by the self-generated system, rather than an externally preset absolute spatiotemporal container. The endogenous time dimension is bound to the energy mechanism, the endogenous space dimension is bound to the material structure, and the endogenous phase dimension is bound to the information field state. The Lipshitz constant of each closed loop is strictly less than 1. The Lipshitz constant refers to the maximum input-output gain of the closed loop mapping, used to measure the compressibility of the mapping. Its value is less than 1, which is a necessary and sufficient condition for the closed loop to converge. Interlayers achieve endogenous spatiotemporal coordination through holographic recursive mapping. The holographic recursive mapping refers to a linear inner product-preserving mapping mechanism that realizes interlayer state projection and temporal alignment. The interlayer coupling channel satisfies the small gain stability condition, which means that the product of the closed-loop gain of adjacent layers and the coupling gain is less than 1, to ensure the global convergence after multi-layer cascading. S2, an interlayer interactive free energy coupling mechanism is established. Based on the real-time data of the robot's multimodal sensor array, a full-body five-dimensional holographic coherent state tensor is generated. The five-dimensional holographic coherent state tensor refers to a system state holographic representation tensor covering five dimensions: material structure, energy mechanism, information field state, time synchronization, and spatial alignment, to measure the coherence state of the system in all dimensions. The global coherence is calculated in five sub-dimensions: material structure, energy mechanism, information field state, time synchronization, and spatial alignment. The five-dimensional coherence cross-validation matrix is ​​used to realize hierarchical fault localization and self-repair triggering from joints to the whole body. S3, when the cognitive level receives a task instruction, it decomposes the task into an ordered sequence of action stages, with each stage corresponding to a sub-goal, participating limbs, and constraints. For each action stage, with the cognitive level as the parent system, the limb-level self-generated system participating in the action as the mother system, and the current task environment as the environmental system, an endogenous spatiotemporal game sandbox is constructed through a parent-dominated fractal generation mode. The parent-dominated fractal generation mode refers to a subsystem generation mode guided primarily by information from the parent system and supported by physical resources from the mother system. The endogenous spatiotemporal game sandbox refers to the strategy iteration and optimization space under the endogenous spatiotemporal system. Action strategy subsystems adapted to the current environmental conditions are dynamically generated. After passing the general cognitive triple test, the action strategy subsystems are solidified hierarchically. After all stage strategies are generated and transition simulation verification is completed, the action strategy sequence is issued. In S4, during the action execution, each layer of self-generated system evolves independently along the negative gradient direction of the joint free energy functional. The body level monitors the stage transition conditions in real time, and when the conditions are met, the strategy switching of each level is triggered synchronously. The endogenous spatiotemporal parameters are dynamically adapted with the action stage, and each layer estimates the closed-loop stability margin in real time and actively adjusts the control parameters to maintain the system operating within the stable region.

2. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, The specific process of the parent-dominated fractal generation mode in step S3 is as follows: (1) Generation triggering: The cognitive level detects the demand for a new action stage, evaluates the material topology resources and energy mechanism margin of the parent system, and triggers the fractal generation operator when the sustainable generation conditions are met; (2) Initial construction: The fractal generation operator extracts the action meta-rule template from the information field state of the parent system, divides the physical resource region of the subsystem in the material topology of the parent system, and transmits the meta-rules to the initial information field state of the subsystem without loss through the state transmission operator; (3) Bootstrap evolution: The parent system injects initial energy into the subsystem region to start the subsystem. The system's ternary closed-loop self-organizing evolution is iteratively optimized along the negative gradient direction of the joint free energy; (4) Generation verification: The operation identity Chern number of the subsystem is continuously monitored by the ontology calibration operator. When the Chern number is stable in the target value range and the sustainability index continuously meets the threshold, the subsystem generation is determined to be complete; (5) Rule sedimentation: The valid subsystem information field rules are solidified by the rule sedimentation operator to form a reusable action strategy template. The input of the rule sedimentation operator is the valid strategy parameter set and the corresponding environmental feature vector, and the output is the solidified template entries and metadata tags in the rule base.

3. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 2, characterized in that, The topology criterion for a completed subsystem generation is as follows: the Chern number of operational identity is used to characterize the connectivity of the subsystem topology; a higher value indicates stronger coupling between joints and better overall subsystem integrity. When the Chern number of operational identity fluctuates by less than 5% for 10 consecutive iterations and remains stable within the non-zero target range, the action strategy subsystem is deemed to possess independent operational identity, and fractal generation is complete. The Chern number of operational identity is calculated using the Laplace matrix algebraic connectivity of the joint topology graph, and the calculation formula is as follows: in Let be the Laplace matrix of the joint topology corresponding to the subsystem. It is the second smallest eigenvalue of the Laplace matrix, i.e., the algebraic connectivity.

4. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, In step S1, the closed-loop Lipshitz constants and inter-layer couplings at each level satisfy the following stability constraints: the true Lipshitz constant for the joint-level closed loop ranges from [0.1, 0.3], with an operation timescale of sub-milliseconds; the true Lipshitz constant for the limb-level closed loop ranges from [0.3, 0.5], with an operation timescale of milliseconds; the true Lipshitz constant for the body-level closed loop ranges from [0.5, 0.7], with an operation timescale of tens of milliseconds; the true Lipshitz constant for the cognitive-level closed loop ranges from [0.4, 0.6], with values ​​lower than the upper limit for the body-level, prioritizing decision stability and avoiding significant policy oscillations, with an operation timescale of hundreds of milliseconds; adjacent levels... and The bidirectional coupling gain between them is symmetrical, that is Satisfying the small gain condition This ensures stable global input and output after multi-layer feedback interconnection; the endogenous time periods between layers satisfy integer frequency division relationships. ,in This is the hierarchical number; the larger the value, the higher the hierarchical level and the longer the cycle. It is a positive integer, and the lower-level period is smaller than the higher-level period.

5. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, In step S2, the five-dimensional coherence cross-validation matrix is ​​a 5th-order symmetric matrix, and the matrix elements are the temporal correlation coefficients of pairwise coherence, which are used to identify the temporal causal path of fault propagation. Fault localization follows these rules: if a single dimension exhibits an anomaly first, and the associated dimension lags behind by at least two corresponding level endogenous time quanta, the single dimension is determined to be the root cause of the fault; if multiple dimensions exhibit anomalies simultaneously and the timing difference is less than one corresponding level endogenous time quantum, it is determined to be a common source fault in the inter-layer coupling channel; after locating the dimension, the specific joint is located by the local anomaly amplitude of the tensor space dimension, achieving hierarchical fault localization from the dimension level to the single joint; the corresponding handling mechanisms are: anomalies in the material structure dimension trigger joint torque compensation and task reallocation; anomalies in the energy mechanism dimension trigger dynamic energy scheduling between limbs; anomalies in the information field state dimension trigger a re-fractal generation strategy; anomalies in the time synchronization dimension trigger time resynchronization; and anomalies in the spatial alignment dimension trigger spatial coordinate system calibration.

6. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, In step S3, the endogenous spatiotemporal game sandbox is constructed through four operators: a time recursion generation operator generates an endogenous time benchmark matching the intensity of actions, unifying the game iteration rhythm of each participating limb; a spatial coherence generation operator generates a dynamic spatial coherence field, automatically improving spatial resolution in areas with concentrated forces and frequent interactions; an interphase holographic generation operator constructs a holographic mapping intermediary field, eliminating the need for external communication protocols for game information exchange between limbs; and a three-dimensional coupling mapping operator integrates the endogenous time, space, and interphase dimensions in a closed loop, ensuring the system's self-consistency. In the game sandbox, the fundamental physical laws of Newtonian mechanics, conservation of momentum, and conservation of angular momentum are embedded in the payoff function as constraint terms, without including preset precise dynamic parameters, and the environmental parameters are estimated in real-time from the mapping of material structure to information field state.

7. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, In step S4, the joint free energy functional is a dimensionless objective function with all terms normalized, and its complete expression is: in For the first Layer operation identity safety margin For the first Self-consistent deviation rate of a three-element closed-loop layer. For the first Layer energy consumption, For the first Maximum energy consumption of the base layer; initial weights satisfy normalization constraints. After the energy term weights are dynamically adjusted, the other two terms are scaled synchronously according to their original proportions, maintaining a weight sum of 1; the expression for the interlayer interaction free energy is: in The interlayer information field state correlation strength, This refers to the self-referential coverage of rules from higher levels to lower levels. and is the coupling coefficient.

8. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, The general cognitive triple test in step S3 includes: (1) Semantic entropy consistency test: calculate the change in Shannon information entropy of the rule base before and after the introduction of the new strategy based on the activation frequency of the action primitive. If the entropy increase exceeds the dynamic threshold, it is judged as failing; the action primitive is the basic action unit that cannot be further divided in the rule base and is the smallest unit that constitutes the action strategy; (2) Self-referential coverage test: send a set of meta-questions containing three categories: applicable scenarios, constraints, and performance boundaries to the strategy base. If the matching degree ratio between the query result and the annotation of the rule base is lower than the dynamic threshold, it is judged as failing; (3) Cross-scenario generalization consistency test: apply the strategy to multiple sets of standard benchmark scenarios. If the coefficient of variation of the performance index exceeds the dynamic threshold and cannot be explained by the change in environmental parameters, it is judged as failing; the thresholds of the three tests are dynamically modulated by the cognitive-level operation identity safety margin: the higher the safety margin, the wider the threshold to encourage exploration; the lower the safety margin, the stricter the threshold to ensure stability.

9. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, The step-by-step path for solidifying the motion strategy in step S3 is as follows: from cognitive level to body level: the motion strategy is encoded as global coordination parameter updates at the body level and stored in the long-term rule base of the main control unit; from body level to limb level: the global strategy is decomposed into coordination control parameter updates at each limb level and distributed to each limb processing unit; from limb level to joint level: the joint torque control parameters are solidified into a fast response lookup table for the joint-level processing unit. The lookup table format is a three-dimensional mapping table of torque command-joint angle-joint angular velocity, and it is stored in the block random access memory of the joint-level processing unit. After being solidified step by step, the response latency of commonly used action strategies is reduced from hundreds of milliseconds at the cognitive level to sub-milliseconds at the joint level.

10. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, In step S4, each level uses an online perturbation injection method to estimate the normalized relative Lipshitz constant in real time: a zero-mean Gaussian perturbation is added to the control input in each control cycle, the change in system state output is measured, the instantaneous gain is calculated and smoothed by exponential moving average, and then divided by the benchmark maximum gain to obtain the estimated value of the normalized relative Lipshitz constant. When the estimated value exceeds the warning threshold of 0.9, the system actively reduces the iteration step size, increases joint damping, and reduces the intensity of the action; when the estimated value exceeds the intervention threshold of 0.95, it freezes the update of non-core rules and prioritizes the survival of the ontology.

11. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, Automatic adaptation of endogenous spatiotemporal parameters in high-dynamic motion scenarios: During the burst phase, the endogenous time quantum of joints is automatically shortened to 50%~70% of the nominal value, and the spatial resolution of the force concentration area is automatically improved to 150%~200% of the nominal value; During the steady-state maintenance phase, the endogenous time quantum is automatically restored, and the spatial resolution is homogenized to save energy; During the underactuated motion phase, the body level uses the angular momentum conservation constraint based on the endogenous spatiotemporal phase system, adopts the recursive least squares method to identify the equivalent mass of each limb in real time and adjust the topological distribution of limb material, changes the rotational inertia tensor of the system, and realizes fine-tuning of posture without the need for a preset precise rigid body dynamics model.

12. The robot whole-body motion control method based on a multi-layer nested self-generated system according to claim 1, characterized in that, The method operates on a distributed multi-level processing architecture, which includes a joint-level field-programmable gate array processing unit, a limb-level microcontroller unit processing unit, a body-level main control processing unit, and a cognitive-level acceleration processing unit, each corresponding to the execution of control logic at four levels. The multimodal sensor array contains dedicated sensor groups for each level, with a sampling frequency of not less than 1kHz, providing sensing input for each closed loop.

13. A robot motion control device based on a multi-layered nested self-generated system, characterized in that, It comprises four main functional modules, each interconnected via an internal system data bus, and all capable of acquiring data from a multimodal sensor array through the internal bus: Multi-layer endogenous kernel construction and spatiotemporal anchoring module: used to execute step S1 of claim 1, constructing a four-layer nested self-generated system kernel and completing spatiotemporal phase primitive anchoring; Inter-layer coupling and five-dimensional coherence monitoring module: used to execute step S2 of claim 1, establishing an inter-layer interactive free energy coupling mechanism and achieving hierarchical fault location and self-repair; Fractal generation and strategy accumulation module: used to execute step S3 of claim 1, completing task decomposition, parent-dominated fractal strategy generation, cognitive verification, and hierarchical solidification. Multi-level collaboration and active stability maintenance module: used to execute step S4 as described in claim 1, to realize multi-level collaborative evolution, smooth switching of action phases and active stability maintenance.

14. An electronic device, characterized in that, include: processor; And a memory storing computer program instructions; when the computer program instructions are executed by the processor, the processor performs the steps of the method as described in any one of claims 1 to 12.

15. A computer-readable storage medium for robot motion control, characterized in that, It stores computer program instructions that, when executed by a processor of a robot's distributed processing architecture, implement the steps of the method as described in any one of claims 1 to 12.

Images