A method, system and storage medium for controlling multi-layer nested service robots

By constructing a multi-layered nested self-generating system based on operation closure, the problems of separation of training and execution, catastrophic forgetting, and easy failure of safety constraints in the service robot control system are solved. The system achieves endogenous steady state and topological growth of autonomous learning, thereby improving the system's autonomous adaptability and safety.

CN123103648APending Publication Date: 2026-11-17罗浩
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611326946.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing service robot control systems rely on external loss-driven mechanisms, separate training and execution, suffer from catastrophic forgetting, have fixed capability ceilings, and are prone to failure due to safety constraints, lacking endogenous steady-state and topology growth capabilities.

Method used

A multi-layered nested self-generated system based on operation closure is constructed. It adopts a four-layer nested architecture and is driven by both discrete normalized Ricci flow and Hamiltonian flow in two time scales. The steady state is endogenously determined by the fixed point of operation identity. The security constraints are encoded as topological infeasible regions and the identity continuity is protected by topological invariants.

Benefits of technology

It achieves endogenous steady-state autonomous learning, zero-forgetting incremental learning, enhanced safety response, and capability topology growth, solving the problems of separation of training and execution, catastrophic forgetting, and easy failure of safety constraints in existing technologies, thereby improving the system's autonomous adaptability and security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103648A_ABST
    Figure CN123103648A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-layered nested service robot control method, system, and storage medium, belonging to the interdisciplinary field of service robot and complex system control. The invention employs a nested architecture of four or more layers, mapping multimodal perception signals to joint configuration manifolds to form a unified configuration-contact topology, and intrinsically defining the perception boundary through closure verification; mapping basic actions to homotopy-like closed trajectories to construct an action topology set, using a Lipschitz constant less than 1 as the convergence criterion; mapping task instructions to topological deformation perturbations, converging to a steady-state fixed point via Hamiltonian flow and decoding the sub-target sequence; and implementing topological-level safety control and topological growth at the metacognitive layer, maintaining identity continuity through topological invariants. This invention achieves intrinsic steady-state operation, zero-forgetting lifelong learning, intrinsic safety control, and interpretable reasoning in service robots, adapting to open scenarios such as home care and elderly care services.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of service robot and complex system control, specifically relating to a multi-layer nested service robot control method, system, and storage medium based on operational closure constraints and topological dynamics. Background Technology

[0002] The current mainstream control paradigm in the service robot field is a "perception-planning-action" architecture centered on deep learning. The "brain" layer uses a large language model or visual language model for task decomposition, while the "cerebellum" layer uses reinforcement learning, diffusion strategies, or model predictive control to generate motion trajectories. The two layers are connected through a predefined interface. This architecture has achieved significant engineering progress, but it has fundamental flaws.

[0003] First, parameter updates rely on an external loss function. Existing reinforcement learning schemes guide policy search through an external reward function, the diffusion policy learns actions by fitting the demonstration trajectory distribution, and model predictive control optimizes the trajectory by minimizing the external cost function. After system deployment, there is no intrinsic learning motivation, training and execution are separated, and it cannot continuously adapt to open environments.

[0004] Secondly, there is a lack of unified steady-state constraints between the layers. Semantic planning at the cerebral level and motor control at the cerebellum level are two independent systems, connected by a manual interface. When environmental disturbances pass through the interlayer interface, the disturbance signal lacks a unified convergence anchor point at the splicing interface, easily amplifying and failing layer by layer. Existing technologies lack a global identity steady-state constraint that runs through all layers.

[0005] Third, incremental learning inevitably leads to catastrophic forgetting. All tasks share the same set of parameter weights, and training for new tasks will overwrite the parameter configurations of older tasks. If a robot learns to "grab a cup" and then tries to learn to "grab a pen," the old grasping strategy is overwritten by the parameters, resulting in a significant drop in task success rate. Existing solutions cannot achieve zero forgetting after incremental learning.

[0006] Fourth, the capability ceiling is determined by the initial architecture. The existing system's action and task libraries are fixed during the training phase and cannot be expanded after deployment. The system itself lacks the motivation and mechanism to break through its own topology, and any capability improvement requires manual redesign of the architecture and retraining.

[0007] Existing Riemannian manifold motion planning techniques model the joint configuration space as a Riemannian manifold, using geodesics as the optimal trajectory. Geometric fabrics techniques embed safety constraints into the motion equations to improve stability. However, both have fundamental limitations: First, the Riemannian manifold is merely a geometric constraint tool for the configuration space, with its evolutionary momentum entirely derived from the optimization of external cost functions. The system lacks endogenous steady state and endogenous boundaries, and its essence remains an externally organized structure. Second, geometric fabrics are merely external constraint enhancements, without changing the core logic of external driving forces, and they do not possess autonomous growth or lifelong learning capabilities. Third, all existing manifold-related technologies are optimizations of the parameter space, and capability expansion inevitably involves rewriting the original parameters, failing to fundamentally eliminate catastrophic forgetting.

[0008] Therefore, existing technologies have not built service robot control systems from the underlying constraints of operational closure, nor have they formed a complete engineering chain of "endogenous steady state - topological growth - identity continuity". Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing service robot systems, such as reliance on external loss-driven mechanisms, separation of training and execution, catastrophic forgetting, fixed capability ceilings, and easy failure of external safety constraints. The invention provides a multi-layered nested service robot control method, system, and storage medium based on closed-loop operation.

[0010] The core idea of ​​this invention is to construct a service robot as a multi-layered, nested, self-generating system, consisting of at least a perception-action layer, a motion pattern layer, a task planning layer, and a metacognitive layer, based solely on operational closure constraints as the underlying principle. System evolution is driven by both discrete-timescale normalized Ricci flow and Hamiltonian flow; steady state is intrinsically determined by the fixed point of operational identity; capability enhancement is achieved through topological growth; safety constraints are encoded as topologically infeasible regions; and identity continuity is protected by topological invariants. Operational closure is an intrinsic property of the self-generating system, characterized by composite closure and continuous productivity: composite closure means that all legal state transformations do not exceed the intrinsically defined operational boundaries, and environmental inputs only trigger state adjustments as structural disturbances, without directly rewriting the system's internal operational rules; continuous productivity means that the system has the ability to continuously generate new operations that cannot be reduced to existing composite operation equivalence classes. The rate of new operation generation can decrease in steady state, but this ability is not structurally eliminated; the inability to be reduced to existing composite operation equivalence classes means that there is at least one reachable state between the new operation and any existing composite operation, such that the transformation effects at that state are different. The preferred embodiment of the service robot described in this invention is that each layer is constructed as one or more self-generated systems based on operation closure. For details of the technical solutions related to constructing self-generated systems, please refer to the invention patent application "Construction Method, System and Storage Medium of General Self-generated System Based on Operation Closure" filed by the applicant on the same day. This invention is a specific application of the general self-generated system paradigm in the service robot scenario. The exclusive innovations include: unified alignment of joint configuration manifolds for multimodal perception, engineering determination of action homotopy, human-machine collaborative topology switching, and robot-specific diffusion-MPC self-generated transformation.

[0011] First aspect: A control method for multi-layered nested service robots

[0012] The control method provided in this paper is applicable to service robots that employ at least a four-layer nested control architecture. The four layers, arranged from closest to furthest along the physical interaction distance, are the perception-action layer, the motion pattern layer, the task planning layer, and the metacognition layer. The method includes the following core steps:

[0013] S1. The computation unit maps the multimodal sensing signals to the joint configuration manifold, forming a unified configuration-contact topology using the joint configuration manifold as a unified reference topological space. The computation unit performs closure checks on the set of atomic operations, intrinsically defining the topological feasible domain boundary of sensing-action on the joint configuration manifold. The operation closure refers to the system's intrinsic properties, including composite closure and continuous productivity. Composite closure means that all legal state transitions do not exceed the intrinsically defined operation boundary, and environmental inputs only serve as structural disturbances to trigger state adjustments without directly rewriting the system's internal operation rules. Continuous productivity means that the system has the ability to continuously generate new operations that cannot be reduced to the equivalence class of existing legal operations. The rate of new operation generation can be reduced in steady state, but this ability is not structurally eliminated. The inability to be reduced to the equivalence class of existing legal operations means that there is at least one reachable state between the new operation and any existing legal operation, such that the transformation effects of the two operations at that state are different.

[0014] Continuous productivity verification is achieved by observing changes in the operation library within an observation window: an observation window with a length of 1000 evolution cycles is set, and the number of newly appearing non-mergeable operations in the set of legal operations within the window is counted. When the number of new operations is greater than zero, continuous productivity is considered satisfied; when the number of new operations is zero for three consecutive observation windows, the system is considered to have degenerated into a closed set, triggering structural adjustments.

[0015] Furthermore, after the multimodal sensing signals are mapped to the joint configuration manifold, topological invariant cross-validation is performed: the Betty number sequences extracted from the visual signals and the Betty number sequences extracted from the force signals are compared with the Betty number sequences of the fused unified configuration-contact topology. When the deviation between the Betty number sequence of any single modality and the fusion result is greater than 10%, that modality is resampled and reprojected until the deviation is less than 10%. The 10% deviation threshold is calibrated based on the typical noise level of the multimodal sensor, corresponding to a topological feature mismatch rate of less than 0.1%.

[0016] Furthermore, the perception-action layer adopts a dual-loop architecture: the fast loop performs servo control at a frequency of no less than 100Hz, responsible for force control, collision response, and joint limit; the slow loop performs topological steady-state evolution at a frequency of 10Hz to 100Hz; under normal operating conditions, the fast loop operates autonomously within the topological feasible region defined by the slow loop; when continuous disturbance of the fast loop causes the topological invariant of the slow loop to fluctuate beyond a preset threshold, the slow loop initiates metric adjustment to reconstruct the feasible region; under emergency operating conditions, the fast loop directly triggers hardware-level safety braking and simultaneously sends boundary disturbance signals to the slow loop, which then re-converges to a steady state based on the new boundary.

[0017] S2. The computational unit maps basic movements to homotopic closed trajectories on the joint configuration manifold, constructing a set of movement topologies for the motion mode layer; the Lipschitz constant of the motion mode layer closed-loop composite mapping in the information field space is less than 1 as the motion convergence criterion. The closed-loop composite mapping refers to the intra-layer closed-loop mapping composed of the mapping from the information field state to the energy mechanism, the mapping from the energy mechanism to the material structure, and the mapping from the material structure to the information field state in sequence.

[0018] Furthermore, when the basic motion is mapped to a homotopy-class closed trajectory, persistent homology barcodes are calculated for the configuration space loops corresponding to the two motion trajectories. When the bottleneck distance between the two barcodes is less than 5% of the configuration space feature scale, they are determined to be of the same motion homotopy class. The configuration space feature scale is the root mean square value of the total range of joint angles. The 5% bottleneck distance threshold is based on the homotopy class classification confusion rate, corresponding to a classification error rate of less than 0.1%.

[0019] Furthermore, candidate trajectories for the motion topology are generated by the diffusion model, and the model predicts and controls the physical constraints of joint limits and force limits on the candidate trajectories. The trajectory scoring function is set to a score equal to 1 minus the Lipschitz constant of the motion pattern layer closed-loop composite mapping, with maximizing the steady-state convergence of the motion closed loop replacing minimizing the external tracking error as the training objective. The diffusion model adopts a two-stage training approach: "demonstration initialization + reinforcement learning fine-tuning". The first stage uses teleoperation demonstration trajectories for supervised learning to train basic denoising capabilities; the second stage uses... The score serves as the reward signal, and a policy gradient algorithm is used to micro-survey the denoising network; the trajectory is calculated in real time at each step during the denoising sampling process. If the value exceeds the threshold, the step is resampled to ensure that all generated trajectories satisfy the steady-state convergence trend. During topology growth, the metric parameters and topological invariants of the original closed submanifold remain unchanged, and the newly added topology does not modify the original topological structure.

[0020] S3. The computing unit maps the task instructions into the topological deformation of the task planning layer and uses it as a situational perturbation. Through Hamiltonian flow evolution, the situation of the task planning layer is converged to a unique steady-state topological fixed point, i.e., the operation identity fixed point. The executable sub-target sequence is decoded from the steady-state situation, and each sub-target corresponds to a standard action mode of the motion mode layer.

[0021] Furthermore, task instructions are mapped to topological deformation of the task planning layer. Specifically, the speech signal is converted into text semantic features through speech recognition, and the task vocabulary is mapped to the feature deformation pattern of the task planning layer topology through a semantic topology encoder. Semantic logical combination and topological algebra operations satisfy a homomorphic relationship, that is, the semantic combination of multiple actions corresponds to the connectivity and operation of the homotopy class of the corresponding actions. The semantic topology encoder adopts a 3-layer fully connected network structure. Seed training uses 1000 pairs of "speech instruction-task topology" paired samples. The loss function is a weighted sum of "topological deformation convergence error + semantic classification cross-entropy". The optimizer is Adam, with an initial learning rate of 0.001, a batch size of 32, and 30 training rounds for convergence. After deployment, incremental fine-tuning is performed based on whether the instruction execution successfully converges to the corresponding fixed point as a self-supervised signal. Successful instruction execution is a positive sample, and failure is a negative sample. Incremental fine-tuning is performed every 100 task executions with a fine-tuning learning rate of 0.0001. The parameters of the first two layers are frozen, and only the third layer is updated to ensure the stability of the basic semantic topology.

[0022] As a preferred embodiment, the semantic topology encoder can also be established using a language emergence mechanism based on multimodal perception synchronous association: physical field perception data and language perception data are synchronously encoded within the same time window, and the association mapping between language symbols and task topology is established endogenously through Hebbian learning with operational identity constraints, without the need for initial seed training.

[0023] Furthermore, the specific method for decoding the sub-target sequence from the steady-state situation is as follows: starting from the target fixed point converged to the situation in the task planning layer, backtracking back to the initial situation along the lowest energy path of the topological constraints, and taking the key topological nodes on the path that correspond to the action mode switching or contact state change as sub-targets, with each sub-target corresponding to a standard action mode in the motion mode layer.

[0024] Furthermore, if the task planning layer situation still cannot converge to the operation identity fixed point within the preset maximum number of iterations, the system outputs a "cannot be executed" response, without forcibly generating planning results, thus avoiding task planning illusions.

[0025] Furthermore, the identity potential term of the Hamiltonian in the task planning layer is constructed using a Gaussian mixture model, with each task fixed point corresponding to a Gaussian potential well, the center of which represents the steady-state distribution of that task. The potential well depth is positively correlated with the steady-state convergence; the potential energy function is of the form of... ,in The second-order Wasserstein distance. For the potential well depth, The potential well width is defined as follows: after each task converges, the parameters of the corresponding Gaussian potential well are updated with the steady-state distribution as the center, thereby realizing the endogenous update of the potential energy field.

[0026] S4. The operation unit encodes the security rules into permanent infeasible regions in the metacognitive layer topology at the metacognitive layer. These infeasible regions are propagated to each layer through downward causality. The Betti number sequence and Euler characteristic of each layer are used as topological invariants to maintain the continuity of the system identity.

[0027] Information transmission between layers is achieved solely through the coupling deformation of the M-topology of each layer. The information field state rules and energy mechanism driving force distribution processes of each layer remain within their respective layers and do not directly penetrate the interlayer boundaries. The M-topology, namely the material structure topology, is the topological configuration carrier of each layer and the sole medium for intralayer evolution and interlayer interaction.

[0028] The permanently infeasible region is an immutable subset of the topologically feasible region boundary. The perception-action layer endogenously delineates the complete topologically feasible region boundary in S1, and the metacognitive layer encodes the security-related portion of it as a permanently infeasible region in S4. The permanently infeasible region is a component of the feasible region boundary's type identity and cannot be changed under local perturbations; the remaining portion of the feasible region boundary is a component of the pattern identity and can adaptively adjust during steady-state evolution.

[0029] The steps for encoding and mapping safety rules to permanently infeasible regions are as follows: First, quantify the physical constraints in the safety rules into constraint inequalities in the state space: upper bound constraints of force limits corresponding to joint torque components. The lower limit of spatial distance corresponds to the repulsion domain constraint at the end position. The second step is to map each constraint inequality to a boundary function on the joint configuration manifold. ,in The critical surface constraining the inequality, i.e. The isosurface in the configuration space; the third step is to take the union of the critical surfaces of all security constraints to form the boundary of the permanently infeasible region in the metacognitive layer topological space; the fourth step is that the permanently infeasible region is identified by topological invariants: the Betti number sequence and Euler characteristic of the region remain constant in normal evolution and can only be changed by updating the security rules of the metacognitive layer.

[0030] Furthermore, the specific link of the downward causal transmission is as follows: the safety topology hard boundary of the metacognitive layer is first transformed into the flux boundary condition of the upper energy topology, and then mapped to the total flux constraint of the lower energy topology through inter-layer topological coupling. Finally, the lower energy flow shapes the evolution of the lower M topology and limits the lower feasible domain. The entire process is transmitted across layers only through topological configuration deformation, and the safety rule information and driving force allocation process remain within this layer.

[0031] The engineering implementation of inter-layer coupling is as follows: the upper-layer M topology is passed down in the form of "boundary constraint vectors," with the vector dimension being the boundary degrees of freedom of the lower-layer topology space; the lower layer updates the boundary curvature and feasible region of its own manifold through the boundary constraint vectors, realizing the constraint of the upper-layer rules on the lower layer. Inter-layer synchronization adopts event-triggered synchronization, and coupling signals are only passed down when the upper-layer topological invariants change. Under normal circumstances, the lower layer operates autonomously, ensuring real-time performance.

[0032] Furthermore, the topology-level safety constraints adopt a two-level safety architecture: topology-level safety is the feasible region constraint at the software layer, ensuring that all trajectories generated by the planning layer are within the safety boundary. The boundary points of the feasible region of the manifold correspond one-to-one with the physical thresholds of joint angles, torques, and velocities, eliminating over-limit commands from the planning source; the hardware safety loop is an independent physical-level redundancy, directly reading sensor data to trigger emergency stops, independent of the software topology safety, and hardware safety can still provide a backup when topology-level safety fails. The permanently infeasible region is the endogenous immutable boundary of the corresponding hierarchical topology space, and the motion state can only evolve within the feasible region manifold, and cannot break through the manifold boundary constraints by adjusting parameters.

[0033] Furthermore, it also includes a topology growth step to expand the topology set of the corresponding level: when used for the motion pattern layer, topology growth expands the action topology set by adding action closed submanifolds; when used for the task planning layer, topology growth expands the task concept set by adding task topology equivalence classes; when used for the metacognition layer, topology growth expands the strategy library by adding strategy topology patterns.

[0034] The complexity criterion may also include causal gap detection: construct an action-task causal graph, and when it is detected that the existing action closed submanifold cannot cover the causal chain corresponding to the topological equivalence class of the new task, it is determined to be a causal gap, triggering topology growth.

[0035] Furthermore, the complexity criterion for topology growth is: the local complexity exceeds N times the global mean for k consecutive evolution steps, where k ranges from 5 to 20 and N ranges from 2 to 5; or the current task success rate is continuously lower than a preset threshold; the energy criterion is that the system energy reserve is higher than the critical threshold. In the pure electric drive scenario, the system energy reserve is the weighted value of the power reserve and the computing power reserve. In the electromechanical coupling scenario, it also includes the load reserve of the actuator.

[0036] Further, the topology growth is specifically performed according to the following steps: Step a: The initial metric tensor of the newly added closed submanifold is set to the metric value corresponding to the global average curvature, and the initial closed-loop composite mapping Lipschitz constant is set to a preset value greater than 1; Step b: In the boundary region between the new topology and the adjacent topology, local metric smoothing iteration is performed, and in each step, the metric tensor of the boundary region nodes is updated to its weighted average with the metric tensors of the neighboring nodes; Step c: The maximum magnitude of the curvature gradient in the boundary region is calculated. When the maximum magnitude is less than a preset curvature threshold, the bonding is deemed successful. The preset curvature threshold is based on 5% of the global average curvature. Dynamic calibration; Step d: Perform operational closure verification on the new topology, including composite closure verification, compression factor verification, and boundary independence verification; Composite closure verification refers to the proportion of the pairwise composite results of the new topology and the existing topology belonging to the legal set being no less than 95%; Compression factor verification refers to the Lipschitz constant of the closed-loop composite mapping of the new topology converging to less than 1; Boundary independence verification refers to the bottleneck distance of the persistent cohomology barcode between the new topology and the existing topology being greater than 5% of the feature scale of the configuration space; Step e: After the verification passes, the new topology is solidified into a closed submanifold of the corresponding level, the Euler characteristic of this level undergoes an integer quantization transition, and the identity fingerprint of this level is updated; Step f: If any of the steps b to d fails the verification, the newly added topology is automatically deleted, and the system reverts to the state before growth; If the overall organizational entropy fluctuation amplitude after growth exceeds twice the baseline fluctuation amplitude before growth, growth is judged to have failed, and automatic reversion occurs. After reversion, the identity fingerprint of the corresponding level is restored to the state before growth; The organizational entropy refers to the information entropy of the situation distribution, which is used to characterize the steady-state degree of the system.

[0037] Furthermore, a dual-verification closed loop is performed after topology growth: statistical performance verification includes three levels: logical consistency verification, digital twin deduction verification, and practical operation performance verification. Logical consistency verification means that the new topology has no operational conflicts with existing rules and the composite closure degree is ≥95%. Digital twin deduction verification means that the deviation of the simulation trajectory in 50 steps is less than 15%. Practical operation performance verification means that the average score of 20 actual tests is greater than 0.7. Geometric topology verification includes three levels: volume compactness verification, dual reconstruction accuracy verification, and joint Chern number topology verification. Volume compactness verification means that the fluctuation is less than 5% for 5 consecutive periods. Dual reconstruction accuracy verification means that the root mean square error is less than 5%. Joint Chern number topology verification means that in the joint topology structure formed by the new topology and the physical properties of the ontology, the joint Chern number transitions from 0 to a non-zero stable integer and remains stable for 10 consecutive observation periods. Only when both verifications pass can the new topology be confirmed as valid growth. Topologies that fail verification are automatically rolled back to the state before growth.

[0038] Furthermore, during the topology growth process, the metric parameters and topological invariants of the original closed submanifold remain unchanged, and the newly added topology does not modify the original topological structure, thus achieving zero-forgetting incremental learning from a geometric mechanism. Compared with existing incremental learning schemes, techniques such as EWC regularization and experience replay alleviate forgetting through parameter constraints or data replay, but are essentially still trade-offs in parameter space, and forgetting cannot be eradicated. This scheme achieves absolute zero forgetting by adding the ability to expand independent closed submanifolds, without modifying the original topological structure at all.

[0039] Furthermore, each layer constitutes an independent operational closed loop, i.e., a self-generating control subsystem; each layer contains three functional primitives: information field state C, energy mechanism E, and M topology; information field state C encodes all rules and constraints within the layer, energy mechanism E encodes the driving force distribution and dissipation process of the layer, and M topology carries the configuration organization and connection relationship of the layer; within the layer, closed-loop composite mappings are formed by C→E mapping, E→M mapping, and M→C mapping in sequence, with the goal of approximating the identity mapping of the information field state space; when the Lipschitz constant of the closed-loop composite mapping in the information field state space is less than 1, the layer satisfies the operational closure steady-state criterion, and the evolution within the layer converges to a unique operational identity fixed point.

[0040] Furthermore, each layer adopts a dual-timescale evolution mechanism: the slow timescale updates the metric tensor of each layer with discrete normalized Ricci flow, and the long timescale drives curvature homogenization and steady-state geometry optimization; the fast timescale drives the state order pair evolution with Hamiltonian flow to achieve rapid response of the state within the layer; the evolution process is decomposed into self-indicative dissipative components and other-indicative conservative components. The self-indicative components drive the state within the layer to regress to the operational identity fixed point, and the other-indicative components provide exploratory fluctuations.

[0041] Furthermore, each layer adopts an endogenous time quantum as the smallest discrete time unit for the evolution of that layer. The endogenous time quantum is the discrete time interval of a single-step topological evolution of that layer. The endogenous time quantum of the perception-action layer is 1ms, and the endogenous time quantum of the metacognition layer is 1s. The endogenous time period between layers satisfies the integer frequency division relationship. The evolution of different levels proceeds on the discrete time axis defined by the endogenous time quantum of that layer, and the alignment between layers is achieved through time synchronization topological coupling.

[0042] The typical values ​​for endogenous time quantum are as follows: The perception-action layer needs to perform topological steady-state evolution at a frequency of no less than 100Hz (corresponding to a time quantum of no more than 10ms). Considering the real-time requirements of fast-loop servo control, 1ms is taken as the minimum time unit for slow-loop topological evolution. The metacognitive layer is responsible for long-cycle strategy evaluation and safety management, with a typical evaluation cycle on the order of seconds. 1s is taken as the minimum evolution time unit for the metacognitive layer. The typical value for the motion pattern layer is 10ms, and the typical value for the task planning layer is 100ms, satisfying the integer frequency division constraint between layers.

[0043] Furthermore, the interlayer coupling channel satisfies the small gain stability condition: the product of the Lipschitz constant of the adjacent layer closed-loop composite mapping and the interlayer coupling gain is less than 1, which is used to ensure the stability of global input and output after multi-layer nesting; the interlayer coupling does not transmit the original perceptual data and semantic rules, but only transmits the topological boundary deformation parameters.

[0044] Furthermore, it also includes an upward emergence step: after the lower-level steady-state topological patterns undergo operational closure verification, they emerge upward as independent elements of the upper-level state space; specifically: multiple lower-level steady-state situations are probabilistically expanded into empirical probability measures of the lower-level situation space; local high-density regions are extracted as candidate patterns; three operational closure verifications are performed on each candidate pattern: the composite closure degree is not less than 95%, the self-explained gradient flow compression factor is less than 1, and the ratio of internal to external coupling strength is not less than 2; the candidate patterns that pass the verification collapse into independent elements of the upper-level state space.

[0045] The ratio of internal to external coupling strength refers to the ratio of the average coupling strength between nodes within a candidate pattern to the average coupling strength between a candidate pattern node and external nodes. Internal coupling strength is the average of the semantic association weights between nodes in that pattern, and external coupling strength is the average of the semantic association weights between a pattern node and external nodes in the network. A ratio of 2 or higher indicates that the candidate pattern possesses sufficient cohesion and can emerge upwards as an independent element.

[0046] Furthermore, it also includes a human-machine collaborative topology switching step: when the visual sensor detects that a human body has entered the preset interaction space, the system automatically switches to the human-machine collaborative topology, endogenously shrinks the force limit, speed limit, and acceleration limit to the preset safety threshold, and reduces the contact stiffness; when the human body leaves the interaction space, it automatically restores the normal topology; the shrinkage ratio can be adjusted according to the scene safety level.

[0047] Furthermore, it also includes a body-boundary dual non-intrusive identity authentication step: collecting the system boundary output sequence, constructing the Laplace matrix on the boundary state space, calculating the eigenvalue spectrum, counting the algebraic multiplicity of zero eigenvalues ​​and near-zero eigenvalues ​​with moduli less than a preset threshold, mapping the multiplicity pattern to the Betti number sequence of each layer, generating a topological identity fingerprint, and comparing it with the baseline topological identity fingerprint; when the topological charge undergoes an unexpected integer quantum jump, the identity verification is determined to have failed.

[0048] Furthermore, it also includes five-dimensional coherence monitoring: the system continuously monitors the coherence of each layer's material structure, energy mechanism, information field state, time synchronization, and spatial alignment in five dimensions; the coherence of each dimension is the normalized deviation between the current state of that dimension and the baseline steady state, with a value from 0 to 1, where 0 represents complete coherence and 1 represents complete mismatch; faults are located through a five-dimensional coherence cross-validation matrix: if a single dimension first shows an anomaly and the associated dimensions lag behind by at least two corresponding layer-level endogenous time quanta, the single dimension is determined to be the root cause of the fault; if multiple dimensions show anomalies simultaneously and the timing difference is less than one corresponding layer-level endogenous time quantum, it is determined to be a common-source fault in the inter-layer coupling channel.

[0049] Furthermore, each layer is coupled sequentially along the data flow direction and deployed on heterogeneous hardware: the perception-action layer is deployed on an onboard embedded processor or FPGA, the motion mode layer and task planning layer are deployed on an onboard GPU, and the metacognition layer is deployed on an edge computing node or the cloud; the real-time closed loop relies only on the onboard hardware, and the cloud is only responsible for the synchronization of slow timescale topology growth strategies and the updating of long-cycle meta-rules.

[0050] Second aspect: A multi-layered nested service robot control system

[0051] The control system provided in this aspect is deployed on a service robot onboard heterogeneous computing architecture with a multimodal sensor array, computing unit, storage unit and actuator. It includes at least a four-layer nested control architecture coupled sequentially along the data flow direction. The four layers are, in order from near to far along the physical interaction distance, the perception-action layer, the motion mode layer, the task planning layer and the metacognition layer. The perception-action layer is coupled to the motion mode layer through a real-time bus, the motion mode layer is coupled to the task planning layer through a high-speed data interface, and the task planning layer is coupled to the edge metacognition layer through Ethernet.

[0052] The multimodal signal input terminal of the perception-action layer is connected to the sensor array, and the topology output terminal is coupled to the configuration space input terminal of the motion mode layer. This is used to map the multimodal perception signals to the joint configuration manifold, form a unified configuration-contact topology with the joint configuration manifold as a unified reference topology space, perform closure verification on the set of atomic operations, and endogenously define the perception-action boundary.

[0053] The motion mode layer's action output is coupled to the actuator drive, which maps the basic motion to a homotopy-like closed trajectory on the joint configuration manifold, constructing a motion topology set. The motion mode layer's closed-loop composite mapping Lipschitz constant being less than 1 is used as the motion convergence criterion.

[0054] The command input terminal of the task planning layer is connected to the voice interaction module, and the planning output terminal is coupled to the target input terminal of the motion mode layer. It is used to map the task command into the topological deformation mode and use it as a situational perturbation. Through Hamiltonian flow evolution, the situation is converged to the operational identity fixed point, and the sub-target sequence is decoded from the fixed point.

[0055] The security control terminal of the metacognitive layer is coupled downward to the constraint interface of each layer. It is used to encode the security rules into permanent infeasible regions in the metacognitive layer topology and propagate them downward through causality. When the complexity criterion and the energy criterion are satisfied at the same time, topological growth is performed to maintain the continuity of the system identity with the Betti number sequence and Euler characteristic of each layer.

[0056] Information is transmitted between layers only through the coupling deformation of the M-topology of each layer. The information field state rules and energy mechanism driving force allocation process of each layer are retained within their respective layers. Each layer contains three functional primitives: information field state, energy mechanism, and M-topology, forming a closed-loop composite mapping.

[0057] Furthermore, the system also includes at least one of the following: a topology growth module, a human-machine collaborative topology state switching module, a body boundary duality authentication module, and a fault rollback module; the topology growth module is used to perform topology growth by adding a closed submanifold when the complexity criterion and energy criterion are satisfied simultaneously; the human-machine collaborative topology state switching module is used to automatically contract the force limit and velocity boundary when the human body enters the preset interaction space; the body boundary duality authentication module is used to non-invasively extract the topology identity fingerprint through the boundary output spectrum features; and the fault rollback module is used to automatically restore the topology to the state before growth when instability occurs after growth.

[0058] Third aspect: A computer-readable storage medium

[0059] The computer-readable storage medium provided in this aspect stores a computer program thereon, which, when executed by a processor, implements all the steps of the multi-layer nested service robot control method based on operation closure described in the first aspect.

[0060] This invention can be widely applied to service robot scenarios that require autonomous maintenance of steady state, continuous learning, and intrinsic safety control, such as home services, elderly care, medical assistance, and catering services. Compared with existing technologies, it has the following technological advancements: (1) For the first time, a multi-layered nested self-generated control architecture for service robots is constructed based on operational closure constraints. Each layer is an independent operational closed loop, and information is transmitted between layers through topological coupling deformation, which completely solves the semantic-action gap problem caused by the splicing of each layer in existing technologies. (2) Motion learning uses the Lipschitz constant of the closed-loop composite mapping being less than 1 as the convergence criterion, and the learning target is intrinsic steady state rather than fitting the demonstration distribution. After learning new actions, the topological structure of old actions is not affected, thus eliminating catastrophic forgetting from a mechanistic perspective. In lifelong learning scenarios, the success rate of old actions decreases by less than 3%, compared to a decrease of 20% to 50% in existing reinforcement learning schemes, and a decrease of 10% to 20% in incremental learning schemes such as EWC. (3) Capability improvement is achieved through topological growth rather than parameter coverage. Topology growth expands the topology set of the corresponding level by adding closed submanifolds without modifying the existing topology structure, thus ensuring zero forgetting from the mechanism. After growth, the topology invariants undergo integer quantization transitions, and the identity is continuously traceable. The system can continuously learn new actions and tasks autonomously after deployment, breaking through the capability limit of the initial architecture. (4) The security constraints are encoded as permanent infeasible regions of each layer of topology, which belong to the topology-level endogenous boundary rather than software condition judgment, and cannot be broken from the mechanism. In the human-computer interaction scenario, the force control overshoot is reduced by more than 60%, and the security response delay is reduced by 40%. (5) The reasoning process corresponds to the steady-state evolution of each layer of situation. Each step has quantitative indicators such as topology invariants and organizational entropy that can be verified. The topology of the entire reasoning path is traceable, and the action decision can be explained. When the task situation cannot converge, the output is "cannot be executed", eliminating illusions from the topology level. (6) The inter-layer coupling satisfies the small gain stability condition. The global input and output stability after multi-layer nesting has clear mathematical guarantees, avoiding the problem of disturbance amplification layer by layer in the existing splicing architecture. Attached Figure Description

[0061] Figure 1 is a flowchart of the steps of a multi-layer nested service robot control method according to an embodiment of the present invention. Figure 2 is a schematic diagram of a four-layer nested self-generated system architecture according to an embodiment of the present invention. Figure 3 is a schematic diagram of intra-layer CEM closed-loop composite mapping according to an embodiment of the present invention. Figure 4 is a schematic diagram of inter-layer topological coupling transmission according to an embodiment of the present invention. Figure 5 is a schematic diagram of dual-timescale evolution and endogenous time quantum according to an embodiment of the present invention. Figure 6 is a flowchart of topology growth triggering and dual verification according to an embodiment of the present invention.

[0062] Figure labels: S1 - Perception-Action Layer; S2 - Motion Pattern Layer; S3 - Task Planning Layer; S4 - Metacognition Layer; 21 - Information Field State C; 22 - Energy Mechanism E; 23 - M Topology; 24 - C→E Mapping; 25 - E→M Mapping; 26 - M→C Mapping; 31 - Fast Loop; 32 - Slow Loop; 41 - Diffusion Model; 42 - Model Predictive Control; 51 - Semantic Topology Encoder; 52 - Hamiltonian Flow Evolution; 53 - Backward Decoding; 61 - Safe Topology Hard Boundary; 62 - Downward Causal Propagation; 63 - Topology Growth; 64 - Dual Verification. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0064] In this embodiment, a six-axis collaborative robotic arm is used as the hardware platform, and the hardware parameters are benchmarked against the industrial-grade UR5e collaborative robotic arm: the joint angle range is... It has a rated load of 5kg and a repeatability of ±0.03mm. The sensor configuration includes an Intel Realsense D435i RGB-D camera, a six-dimensional force sensor (range ±100N), a 16×16 tactile array, and a 17-bit absolute joint encoder. The computing platform is an onboard NVIDIA Jetson AGX Orin + FPGA real-time control board, conforming to the mainstream industrial configuration of current service robots. All numerical parameters are calibrated based on this hardware platform, ensuring industrial-grade reproducibility.

[0065] Example 1: Full-process control of the cup-feeding task

[0066] This embodiment completes the full-link control of the typical nanny task of "feeding water with a cup", covering all core steps from S1 to S4, and verifies the end-to-end operational effectiveness of the multi-layer nested architecture.

[0067] The initial configuration and initial action topology set contain five seed action homotopy classes: grab, drop, translate, rotate, and push. These seed actions are generated by extracting homotopy classes from 300 sets / classes of teleoperation data. The initial Betty number sequence for the perception-action layer is as follows: ,in This corresponds to a connected component in the configuration space. Independent closed loops corresponding to the six joint degrees of freedom. Eulerian characteristic of the initial motion topology of the motion pattern layer. The value is -5, according to the formula. Calculation, where Let be the genus number of the action topology space. This represents the number of topological boundary endpoints; in this embodiment... , Substituting into The values ​​are based on the topological genus and boundary number statistics corresponding to the initial 5 homotopy classes of actions. The initial task concept set of the task planning layer contains three task equivalence classes: pick, place, and organize. The initial safety rules of the metacognitive layer contain three permanently infeasible regions: upper limit of human contact force, lower limit of human spatial distance, and dangerous action exclusion set. The configuration space feature scale is taken as the root mean square value of the total range of joint angles, and the total range of six-axis joints is... The characteristic scale was calculated to be 0.8 rad, and the value was taken from the root mean square statistic of the six-dimensional joint angle range.

[0068] S1 Execution Process: Multimodal topology alignment and operation closure verification inputs include 1280×720 resolution scene point clouds acquired by an RGB-D camera, 16×16 pressure distribution acquired by a fingertip tactile array, real-time data from a six-dimensional force sensor, and joint encoder position data. First, multimodal topology alignment is performed: using the joint configuration manifold as a unified reference space, visual signals are processed by a point cloud registration algorithm (using ICP registration from the Open3D open-source library) to identify the geometric features and contact boundaries of the object surface, transforming them into object contact constraints in the configuration space as global boundary corrections for the unified configuration-contact topology; force and tactile signals are transformed by a contact force distribution estimation algorithm into local curvature corrections in the configuration space as local deformations for the unified configuration-contact topology. Topological feature alignment of signals at different sampling rates is achieved based on the scale invariance of persistent cohomology. Persistent cohomology calculation is implemented using the GUDHI open-source library, with the scale parameter set to... , .

[0069] After alignment, perform topological invariant cross-validation: the Betti number sequence extracted from the visual signal is as follows: The Betty number sequence extracted from the tactile signals is The fused topological Betti number sequence is The deviation calculation formula is: Deviation The calculated deviation between the visual and fusion results was 0%, while the deviation between the tactile and fusion results was [missing value]. If the threshold is exceeded (10%), a tactile resampling is performed. After resampling, the tactile Betty number sequence becomes... All deviations are less than 10%, and the alignment is effective.

[0070] Then, an operation closure check is performed: the number of basic operations is 12, which is less than 20, so a full enumeration check is performed. The total number of second-order operation combinations is... There are a total of 3rd-order operation combinations. All operations passed the closure check. For example, a second-order composite operation of "state read + parameter load" results in a state read operation with parameters, which is a valid operation. Similarly, a third-order composite operation of "boundary scan + address jump + state read" results in a directed boundary state read operation, also a valid operation, thus verifying closure. The operation set also satisfies operation distinguishability: for any two unequal operations, there exists at least one initial state that makes the final state different, and the operation boundaries can be intrinsically determined.

[0071] Continuous productivity verification: An observation window of 1000 evolution cycles is set, and the number of newly added non-mergeable operations is counted within the window. In the first 1000 cycles of this embodiment, the system endogenously derives the "displacement grab" sub-operation based on the "grab" action. This sub-operation has at least one state transition effect that is different from any existing operation, and is judged as a new non-mergeable operation, thus satisfying continuous productivity.

[0072] S2 Execution Process: Action Topology Construction and Convergence Criterion. The "Cup-Holding" action trajectory candidate generation uses a one-dimensional diffusion Transformer model (modified from the Diffuser open-source framework). Using the previous round's action trajectory as input, 100 candidate trajectories are generated after 20 denoising iterations. Model Predictive Control (MPC) performs joint limit and force constraint checks on the candidate trajectories based on a rigid body dynamics model, eliminating candidate trajectories that violate torque constraints and retaining 20 feasible trajectories.

[0073] Action homotopy classification: Calculate persistent homology barcodes for the two optimal candidate trajectories of the "holding a cup" action. The barcode of trajectory A is the persistent feature interval of order 0. First-order feature persistence interval The barcode of trajectory B has a persistent 0-order feature interval. First-order feature persistence interval The bottleneck distance is calculated as the minimum-maximum distance between all feature point pairs of the two barcodes, yielding a bottleneck distance of 0.03. The configuration space feature scale is 0.8 rad, and the 5% threshold is... If 0.03 < 0.04, the two trajectories are determined to belong to the same action homotopy class, and the action is solidified into a closed submanifold of the motion mode layer.

[0074] Lipschitz constant calculus of closed-loop composite mapping in motion mode layer: closed-loop composite mapping The action is applied to the motion pattern layer information field space, which is a 6-dimensional joint moment distribution space. Two arbitrary situational state vectors are taken from the information field space. , The formula for calculating the Lipschitz constant is as follows: The maximum singular value of the composite mapping Jacobian matrix is ​​approximated by calculating the maximum singular value. After 20 rounds of endogenous iteration, the Lipschitz constant of the "grab" action in the initial action library converges from the initial value of 0.85 to 0.47, and is stable at less than 1, indicating that the action has reached a steady state of operational identity.

[0075] The trajectory scoring function is For the initially non-convergent "cup-holding" action, the initial... The rating is After 20 iterations Converging to 0.47, the score improved to The training objective is to maximize the score, which is equivalent to maximizing the steady-state convergence of the motion closed loop.

[0076] S3 Execution Process: Task Planning and Sub-goal Decoding. The input task instruction "fed water with a cup" is converted into 128-dimensional text semantic features through speech recognition (using Whisper Lightweight). These features are then mapped to the topological feature deformation pattern of the task planning layer using a semantic topology encoder. The semantic topology encoder is a 3-layer fully connected network with an input dimension of 128 and an output dimension of 30, corresponding to the deformation vector in the task topology space. This deformation pattern corresponds to a combination of four semantic actions: "grasp," "translate," "tilt," and "reset." The semantic-topological homomorphic relationship maps this to the connected sum of the homotopy classes of the four actions, generating the initial situation at the task planning layer.

[0077] The situational awareness at the task planning layer evolves along the Hamiltonian flow. The situational awareness space is derived from the task concept activation distribution space, and the situational awareness sequence pairs include the distribution. With conjugate generation potential Hamiltonian ,in For kinetic energy term, This is the identity potential energy term. The identity potential energy is constructed using a Gaussian mixture model, with each task fixed point corresponding to a Gaussian potential well, the center of which is the steady-state distribution of that task. The potential well depth is positively correlated with the steady-state convergence; the potential energy function is of the form of... ,in The second-order Wasserstein distance. For the potential well depth, The potential well width is defined as follows: after each task converges, the parameters of the corresponding Gaussian potential well are updated with the steady-state distribution as the center, thereby realizing the endogenous update of the potential energy field.

[0078] The discrete update of the Hamiltonian flow uses Sinhalese iteration, and the formula is as follows: in and The step size coefficient is 0.01, and its value is determined based on the balance between task space dimension and convergence stability. Initial situation. To ensure uniform distribution, It has zero potential. Each evolutionary step performs a discrete Hamiltonian flow update, and the organization entropy is determined during the evolution process. The value gradually decreases from the initial value of 0.98, and after 8 iterations, it drops to 0.15 and remains stable. The rate of change is less than the threshold of 0.5% for 3 consecutive steps, indicating that the situation has converged to the target fixed point, and the planning is completed.

[0079] Sub-target decoding: Starting from the target fixed point, backtracking to the initial state along the lowest energy path of the topological constraints. The four key topological nodes on the path correspond to action mode switching (grab → translate), contact state change (translate → tilt), action mode switching (tilt → reset), and contact state change (reset → initial), forming four sub-targets: Sub-target 1 - Grab the cup, Sub-target 2 - Translate to the mouth of the service object, Sub-target 3 - Tilt to feed water, Sub-target 4 - Reset. Each sub-target corresponds to a standard action mode in the motion mode layer.

[0080] S4 Execution Process: The safety control and steady-state maintenance metacognitive layer encodes three types of safety rules into permanently infeasible regions: upper limit of human contact force (50N), lower limit of human spatial distance (10cm), and dangerous action exclusion set. These permanently infeasible regions are endogenous immutable boundaries of the corresponding hierarchical topological space; motion states can only evolve within the feasible manifold and cannot be overcome by adjusting parameters. When the visual sensor detects that the head of the service object enters the preset interaction space (distance < 50cm from the end point), the system automatically switches to a human-machine collaborative topological state, reducing the force limit by 30%, the speed limit by 50%, and the acceleration limit by 50%, thereby lowering the contact stiffness.

[0081] The safety control downward causal transmission chain: The safety topology hard boundary of the metacognitive layer is first transformed into the flux boundary condition of the upper energy topology, and then mapped to the total flux constraint of the lower energy topology through inter-layer topological coupling. Finally, the lower energy flow shapes the evolution of the lower M topology, limiting the lower feasible domain. In this embodiment, the upper limit of human contact force is reduced from 50N in normal working conditions to 35N in the human-machine collaborative topology, the upper limit of joint velocity is reduced from 120° / s to 60° / s, and the upper limit of joint acceleration is reduced from 240° / s² to 120° / s².

[0082] Throughout the entire service task execution, the topological invariants at each layer remain constant. The Betty number sequence of the perception-action layer is maintained. The motion pattern layer Euler indicator number is maintained The task topology equivalence classes in the task planning layer remain stable. After the task is completed, the system returns to a safe standby steady state.

[0083] Example 2: Full Process of Topology Growth

[0084] This embodiment verifies the entire topology growth process of the newly added "pinch" action, and verifies the zero-forgetting characteristic and identity continuity.

[0085] When the system continuously executes the "medicine feeding" task, the existing action topology set cannot stably cover the "pill picking" action. The local information entropy exceeds twice the global average for eight consecutive evolution steps, satisfying the complexity criterion; the system has a power reserve of 60%, a computing power reserve of 55%, and a mechanical load of 40%, with the overall energy reserve exceeding the 30% critical threshold, satisfying the energy criterion and triggering topology growth.

[0086] Step a: Initialize the initial metric tensor of the newly added "pinch" action closed submanifold to the metric value corresponding to the global average curvature, which is calibrated to -2.0. The initial closed-loop composite mapping Lipschitz constant is set to 1.2 (greater than 1), indicating that the new action is initially non-convergent. During topology growth, the metric parameters and topological invariants of the original closed submanifold remain unchanged, and the newly added topology does not modify the original topological structure.

[0087] Step b: Boundary smoothing. In the topological boundary region between the "pinch" action and the adjacent "grab" action, perform 5 steps of local metric smoothing iterations. The formula for each iteration is as follows: ,in The self-weight is set to 0.3, and the neighborhood nodes are set to 6 neighbors. After executing 5 steps, the curvature distribution of the boundary region is continuous, with no curvature singularities.

[0088] Step c: The maximum magnitude of the curvature gradient in the interface region is calculated to be 0.08. The global average curvature is -2.0, and the 5% threshold is... 0.08 < 0.1, therefore the adhesion is deemed acceptable.

[0089] Step d: Closure Verification. Composite Closure Verification: 10 pairs of composites between the new action and the existing 5 actions were performed. 9.8 of these composite results belonged to the set of valid actions, a proportion of 98% > 95%, passing. Compression Factor Verification: After 15 rounds of endogenous iterations, the Lipschitz constant of the closed-loop composite mapping converged from the initial 1.2 to 0.52 < 1, passing. Boundary Independence Verification: The bottleneck distance between the new action and the persistent homology barcode of the homotopy class of the "grab" action was 0.05, greater than the 5% threshold of 0.04 for the configuration space feature scale, ensuring it is an independent new homotopy class, passing. All three verifications passed.

[0090] Step e: The identity update action is solidified into a closed submanifold of the motion pattern layer. The Eulerian characteristic of the motion pattern layer is derived from... Leap to An integer quantization transition occurs with a transition value of -1, corresponding to the addition of one genus. The L2 layer identity fingerprint is updated, recording the correspondence between the fingerprint before and after growth.

[0091] Step f: Dual verification statistical performance verification: Logical self-consistency verification passed; the new topology has no operational conflicts with existing rules; composite closure ≥ 95%; the deviation of the 50-step digital twin derivation trajectory is 12%, less than the 15% threshold; the average score of 20 actual tests is 0.78, greater than the 0.7 threshold, passing. Geometric topology verification: Volume compactness fluctuates by 2.1% over 5 consecutive periods, less than the 5% threshold; the root mean square error of dual reconstruction is 2.5%, less than the 5% threshold; in the joint Chern number topology verification, in the joint topology structure formed by the new topology and the ontological physical properties, the joint Chern number jumps from 0 to 1 and remains stable for 10 consecutive observation periods, indicating topological self-consistency, passing. All dual verifications passed, and the newly added topology is confirmed as effective growth.

[0092] After growth validation with zero forgetting, each of the original five seed actions was repeated 100 times, with an average success rate of 96.2%; the baseline success rate before growth was 97.1%, a decrease of only [missing data]. The forgetting rate is much smaller than the 20%~50% of existing reinforcement learning schemes, and also better than the 12% forgetting rate of the EWC incremental learning scheme, verifying the zero forgetting characteristic.

[0093] Example 3: Parameter Endpoint Verification

[0094] This embodiment verifies the effectiveness of the endpoint values ​​of the topology growth complexity criterion, covering... , , , All endpoints.

[0095] , Scenario: Growth is triggered when the local entropy exceeds twice the global mean for 5 consecutive steps. The growth success rate is 92%, and the steady-state fluctuation after growth is 3.2%, both of which meet the design requirements.

[0096] , Scenario: Growth is triggered when the local entropy exceeds 5 times the global mean for 20 consecutive steps, with a growth success rate of 98%, a steady-state fluctuation of 1.1% after growth, and a false trigger rate of less than 0.1%.

[0097] The verification results show that , The invention objective can be achieved across the entire range, and the parameter range is supportive.

[0098] Example 4: Body-World Dual Non-Intrusive Identity Authentication

[0099] This embodiment verifies the ability to complete identity verification without reading internal parameters. The system boundary output sequence is acquired, including joint current trajectories and end-effector displacement trajectories sampled at 1000Hz. A 4th-order Laplace matrix is ​​constructed on the boundary state space. ,in For degree matrix, Given the adjacency matrix, the eigenvalue spectrum is calculated as follows: , , , Set the near-zero threshold to 0.01, and count the algebraic multiplicity of zero eigenvalues ​​and near-zero eigenvalues ​​to obtain the pattern. The multiplicity pattern is mapped to a sequence of Betty numbers at each level. The identity fingerprint is matched with the baseline topology fingerprint, completing non-invasive identity authentication. The motion pattern layer parameters are artificially perturbed, causing the Betty number to change... The verification method failed due to a mismatch in the multiple patterns during re-authentication.

[0100] Example 5: Numerical Calculus of Multimodal Topology Alignment

[0101] The six-axis robotic arm's configuration manifold is a six-dimensional torus. Visual signals are registered using point clouds to extract the surface topology of the object; the Betty number sequence is... The tactile signal is used to extract the topology of the contact area through contact force distribution, and the Betty number sequence is... The unified configuration after fusion – the contact topological Betti number sequence is: According to the deviation formula, the visual deviation is: The normalized bias rate was 33.3%; the tactile bias was also 33.3%, both exceeding the 10% threshold and triggering resampling. The resampled visual Betty number sequence is... The tactile Betty number sequence is All deviations are 0%, and the alignment is effective.

[0102] Example 6: Action Homology Criterion

[0103] Calculate persistent coherence barcodes for the two candidate trajectories of the "grabbing" action: the 0th-order feature persistence interval of trajectory A. First-order feature persistence interval The persistent 0th-order feature of trajectory B First-order feature persistence interval Bottleneck distance is calculated as follows: The configuration space feature scale is 0.8 rad, the 5% threshold is 0.04, and 0.03 < 0.04, so the two trajectories are determined to belong to the same homotopy class. The bottleneck distance between the "rotation" trajectory and the "grab" trajectory is calculated to be 0.52, which is greater than 0.04, so they are determined to be different homotopy classes.

[0104] Example 7: Numerical Calculation Example of Lipschitz Constant for Closed-Loop Composite Mapping

[0105] The motion mode layer information field space is a 6-dimensional joint moment distribution space. Two arbitrary situational state vectors are taken. , Original distance Closed-loop composite mapping After the action, the result is , Distance after mapping The Lipschitz ratio for a single group is [value missing]. After sampling across the entire space and taking the maximum value, the Lipschitz constant for the "grabbing" action after 20 iterations is 0.47, which is consistently less than 1. After topological growth, the Lipschitz constant for the newly added "pinch" action oscillates around 0.9 in the first 5 iterations, gradually decreasing from the 6th iteration onwards, and converging to 0.52 after the 15th iteration, which is less than 1, indicating that the new action has achieved steady-state solidification.

[0106] Example 8: Hamiltonian Flow Situation Evolution Example

[0107] The task planning layer's situation space is 3D, and the task concept activation distribution space is included. The situation order pairs contain distributions. With conjugate generation potential Hamiltonian ,in This is the same potential energy term. Initial state. (Uniform distribution) (Zero potential). Step size Step 1 Iteration: Therefore ; Therefore Accordingly, steps 2-8 are iterated as follows:

[0108] Step 2 iteration: , Step 3 Iteration: , Step 4 Iteration: , Step 5 Iteration: , Step 6 Iteration: , Step 7 Iteration: , Step 8 Iteration: , After 8 iterations, the distribution... Organizational entropy If the rate of change is less than 0.5% for three consecutive steps, it is determined that the convergence has reached the target fixed point, and the planning is completed.

[0109] Example 9: Human-Machine Collaborative Topology Switching

[0110] When the RGB-D camera detects a human entering a preset 50cm interaction space using a human detection algorithm, the system automatically switches to a human-computer collaborative topology. The force limit is reduced from 50N to 35N (a 30% reduction), the speed limit from 120° / s to 60° / s (a 50% reduction), and the acceleration limit from 240° / s² to 120° / s² (a 50% reduction). Contact stiffness is reduced by 30% by adjusting the stiffness parameters of the force control loop. When the human leaves the interaction space, the system automatically returns to the normal topology. The switching process is propagated step-by-step through a downward causal link, with no parameter jumps and a smooth transition, verifying the inherent safety capabilities of the human-computer interaction.

[0111] Example 10: Safety Performance Comparison Test

[0112] To verify the quantitative conclusions regarding the beneficial effects, a control test was set up: On the same hardware platform, human-machine interaction force control tests were conducted using both a traditional add-on safety rule scheme and the proposed topology-level safety scheme. The interference condition was a sudden application of a 20N step force signal at the end effector. Each test was repeated 50 times, and the average value was taken. Test results: The add-on safety scheme had an average force control overshoot of 45% and an average safety response delay of 15ms; the proposed scheme had an average force control overshoot of 17% and an average safety response delay of 8ms. The calculated reduction in overshoot was... The response latency was reduced by a certain percentage. This supports the conclusion that "force control overshoot is reduced by more than 60% and safety response delay is reduced by 40%" in the beneficial effects.

[0113] Example 11: Distributed Deployment and Latency Testing

[0114] The perception-action layer is deployed on an onboard FPGA, the motion mode layer and task planning layer are deployed on an onboard GPU, and the metacognitive layer is deployed on edge computing nodes. Real-time closed-loop operation relies solely on onboard hardware; the cloud is only responsible for synchronizing slow-timescale topology growth strategies. To ensure security, the metacognitive layer is deployed onboard with the task planning layer in normal deployments, and only deployed on directly connected edge nodes in scenarios with limited computing resources. Furthermore, critical security control paths do not traverse off-board links. Actual measured latency of the FPGA three-stage pipeline: Stage 1 multimodal topology construction and persistent coherence computation latency < 1ms; Stage 2 discrete Ricci flow single-step iteration and W-entropy verification latency < 5ms; Stage 3 steady-state determination latency < 1ms; total latency < 10ms, meeting the 100Hz slow loop requirement. Fast loop servo latency < 1ms, meeting the 1000Hz real-time control requirement.

[0115] Example 12: Semantic Topology Encoder Special Test

[0116] This embodiment verifies the training effect and self-supervised optimization capability of the semantic topology encoder. Seed training configuration: 1000 pairs of "voice command - task topology" samples were used. The loss function was a weighted sum of "topology deformation convergence error (weight 0.7) + semantic classification cross-entropy (weight 0.3)". The optimizer was Adam, with an initial learning rate of 0.001, a batch size of 32, and convergence after 30 training rounds. Test results: The topology mapping accuracy for 100 test commands was 94%, and the command understanding success rate was 92%. Self-supervised optimization: After deployment, incremental fine-tuning was performed every 100 tasks with a fine-tuning learning rate of 0.0001. The parameters of the first two network layers were frozen, and only the third output layer was updated. After 30 days of continuous operation, the personalized command understanding accuracy for the target family improved to 98%, while the basic semantic topology stability remained unchanged.

[0117] Example 13: Two-stage training and testing of the diffusion model

[0118] This embodiment verifies the training effect of the self-biochemical diffusion model. Phase 1: Demonstration Initialization: Supervised learning is performed using 500 teleoperation demonstration trajectories to train the diffusion model's basic denoising capability. The denoising steps are 20, and the training epochs are 50. The initial trajectory generation success rate is 85%. Phase 2: Reinforcement Learning Fine-tuning: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The rating serves as a reward signal, and the reward function is: The denoising network is fine-tuned using a policy gradient algorithm; the trajectory corresponding to each step is calculated in real time during the denoising sampling process. If the value is greater than the threshold of 0.9, the step is resampled to ensure that the generated trajectories all satisfy the steady-state convergence trend. After 30 rounds of fine-tuning, the steady-state convergence rate of the trajectory increased from 62% to 91%, verifying the effectiveness of endogenous steady-state training.

[0119] Example 14: Safety Testing in Extreme Scenarios

[0120] This embodiment verifies the effectiveness of topology-level security under large disturbances. Test scenario: A software parameter tampering attack was launched, attempting to change the force limit threshold from 50N to 100N. Since the security constraint is a manifold topology boundary, the motion state can only evolve within the feasible region. The parameter tampering could not break through the manifold boundary, and the actual output force at the end point was still limited to within 50N. Simultaneously, the hardware security loop triggered a secondary fallback. Test results: All 10 parameter tampering attacks failed to break the security constraint. The maximum force control output was 49.7N, and the security mechanism was 100% effective, verifying the unbypassable nature of topology-level security.

[0121] Example 15: Five-dimensional coherence fault location test

[0122] A simulated torque anomaly fault at the third joint was used to monitor changes in five-dimensional coherence: the material structure dimension showed anomalies first, with coherence rising to 0.8; after a lag of two L1 endogenous time quanta, the energy mechanism dimension showed anomalies, with coherence rising to 0.6. This conformed to the judgment rule of "single-dimensional anomaly first, followed by related dimensions lagging," identifying the material structure dimension as the root cause of the fault, locating it to the third joint, and triggering self-repair. The location accuracy exceeded 95%, validating the effectiveness of the fault monitoring.

[0123] The above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit it. Those skilled in the art should understand that any modifications, equivalent substitutions, or improvements made based on the technical solutions of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-layered nested service robot control method, applied to a service robot equipped with a multimodal sensor array, a computing unit, a storage unit, and an actuator, wherein the service robot adopts a multi-layered nested control architecture, comprising at least a perception-action layer, a motion pattern layer, a task planning layer, and a metacognitive layer arranged sequentially from near to far along the physical interaction distance, characterized in that... Includes the following steps: S1. Map the multimodal sensing signals to the joint configuration manifold, and form a unified configuration-contact topology with the joint configuration manifold as the unified reference topological space; S1. Perform a composite closure check on the set of atomic operations to intrinsically delineate the topological feasible region boundary of perception-action on the joint configuration manifold; S2. Map basic actions to homotopic closed trajectories on the joint configuration manifold to construct the action topology set of the motion mode layer; use the Lipschitz constant of the intra-layer state transition composite mapping in the state space being less than 1 as the action convergence criterion; the intra-layer state transition composite mapping is a closed-loop mapping composed of multiple order state transition mappings acting on the intra-layer state space in sequence, and its Lipschitz constant is used to characterize the compression convergence characteristics of the mapping; S3. Map task instructions to the topological deformation of the task planning layer and use it as a situational perturbation, and use Hamiltonian flow evolution to make the situation of the task planning layer converge to a unique steady-state fixed point, and decode the executable sub-objective sequence from the steady-state situation, each sub-objective corresponding to a standard action mode of the motion mode layer; S4. Encode safety rules into permanent infeasible regions in the metacognitive layer topology, and these infeasible regions are propagated to each layer through downward causality. The system identity continuity is maintained by using the Betti number sequence and Euler characteristic of each layer as topological invariants; the information transmission between layers is achieved only through the coupling deformation of the topological configuration of each layer, and the information rules and driving force allocation process of each layer remain within their respective layers and do not directly penetrate the inter-layer boundaries.

2. The multi-layer nested service robot control method according to claim 1, characterized in that, Each layer of the service robot has its own set of atomic operations and set of composite operations. The set of composite operations is generated by finite composition of each atomic operation set. The set of composite operations and their composition at each layer satisfy operation closure constraints. Operation closure is characterized by composite closure and continuous productivity: composite closure means that all legal state transitions do not exceed the intrinsically defined operation boundary, and environmental inputs only serve as structural disturbances to trigger state adjustments, without directly rewriting the internal operation rules of the system; continuous productivity means that the system has the ability to continuously generate new operations that cannot be reduced to existing composite operation equivalence classes. The rate of new operation generation can be reduced in steady state, but this ability is not structurally eliminated; the inability to be reduced to existing composite operation equivalence classes means that there is at least one reachable state between the new operation and any existing composite operation, such that the transformation effects of the two operations at that state are different.

3. The multi-layer nested service robot control method according to claim 2, characterized in that, The operational closure is achieved through the closed-loop coupling of three functional primitives: information field state C, energy mechanism E, and material structure M. Information field state C encodes all rules and constraints within the layer; energy mechanism E encodes the distribution and dissipation of driving forces within the layer; and material structure M carries the topological configuration of the layer. Within the layer, closed-loop composite mappings are formed sequentially by C→E, E→M, and M→C mappings, aiming to approximate the identity mapping in the information field space. When the Lipschitz constant of the closed-loop composite mapping in the information field space is less than 1, the layer satisfies the operational closure steady-state criterion, and the evolution within the layer converges to a unique operational identity fixed point. Interlayer coupling is achieved only through the topological deformation of material structure M; information field state C and energy mechanism E remain within the layer and do not directly penetrate the interlayer boundary. The topological configuration of material structure M is the topological configuration described in claim 1, abbreviated as M topology.

4. The multi-layer nested service robot control method according to claim 1, characterized in that, Each layer employs a multi-timescale evolution mechanism, including slow-timescale and fast-timescale mechanisms: the slow-timescale mechanism updates the metric tensor of each layer with discrete normalized Ricci flow, driving curvature homogenization and steady-state geometry optimization over long timescales; the fast-timescale mechanism drives the evolution of state order pairs with Hamiltonian flow, enabling rapid response of the state within the layer; the evolution process is decomposed into self-explaining dissipative components and other-explaining conservative components, with the self-explaining components driving the state within the layer to regress to the operational identity fixed point, and the other-explaining components providing exploratory fluctuations.

5. The multi-layer nested service robot control method according to claim 1, characterized in that, The specific link of the downward causal transmission in step S4 is as follows: the safety topology hard boundary of the metacognitive layer is first transformed into the flux boundary condition of the upper energy topology, and then mapped to the total flux constraint of the lower energy topology through inter-layer topological coupling. Finally, the lower energy flow shapes the evolution of the lower M topology and limits the lower feasible domain. The entire process is transmitted across layers only through topological configuration deformation, and the safety rule information and driving force allocation process are retained within this layer.

6. The multi-layer nested service robot control method according to claim 1, characterized in that, The interlayer coupling channel satisfies the small gain stability condition: the product of the Lipschitz constant of the adjacent layer closed-loop composite mapping and the interlayer coupling gain is less than 1, which is used to ensure the stability of global input and output after multi-layer nesting; the interlayer coupling does not transmit the original perceptual data and semantic rules, but only transmits the topological boundary deformation parameters.

7. The multi-layer nested service robot control method according to claim 1, characterized in that, It also includes an upward emergence step: after the lower-level steady-state topological patterns are verified by operation closure, they emerge upward as independent elements of the upper-level state space; specifically: multiple lower-level steady-state situations are probabilistically expanded into empirical probability measures of the lower-level situation space; local high-density regions are extracted as candidate patterns; three operation closure checks are performed on each candidate pattern: the composite closure degree is not less than 95%, the self-explained gradient flow compression factor is less than 1, and the ratio of internal to external coupling strength is not less than 2; the candidate patterns that pass the checks collapse into independent elements of the upper-level state space.

8. The multi-layer nested service robot control method according to claim 1, characterized in that, In step S1, after the multimodal sensing signals are mapped to the joint configuration manifold, topological invariant cross-validation is performed: the Betty number sequence extracted from the visual signal and the Betty number sequence extracted from the force signal are compared with the Betty number sequence of the fused unified configuration-contact topology. When the deviation between the Betty number sequence of any single mode and the fusion result is greater than 10%, the mode is resampled and reprojected until the deviation is less than 10%.

9. The multi-layer nested service robot control method according to claim 1, characterized in that, In step S1, the perception-action layer adopts a multi-loop architecture, including a fast loop and a slow loop architecture: the fast loop performs servo control at a frequency of not less than 100Hz, and is responsible for force control, collision response and joint limit; the slow loop performs topological steady-state evolution at a frequency of 10Hz to 100Hz; under normal operating conditions, the fast loop operates autonomously within the topological feasible region defined by the slow loop; when the continuous disturbance of the fast loop causes the topological invariant of the slow loop to fluctuate beyond a preset threshold, the slow loop initiates metric adjustment to reconstruct the feasible region; under emergency operating conditions, the fast loop directly triggers hardware-level safety braking, and at the same time sends a boundary disturbance signal to the slow loop, and the slow loop re-converges to a steady state based on the new boundary.

10. The multi-layer nested service robot control method according to claim 1, characterized in that, In step S2, the basic motion is mapped to a homotopy-class closed trajectory, which includes: calculating persistent homology barcodes for the configuration space loops corresponding to the two motion trajectories respectively; when the bottleneck distance between the two barcodes is less than 5% of the configuration space feature scale, they are determined to be the same motion homotopy class; wherein, the configuration space feature scale is the root mean square value of the total range of joint angles.

11. The multi-layer nested service robot control method according to claim 1, characterized in that, In step S2, the candidate trajectories of the motion topology are generated by the diffusion model, and the model predicts and controls the joint limit and force limit physical constraint verification of the candidate trajectories. The trajectory scoring function is set to score equal to 1 minus the Lipschitz constant of the motion mode layer closed-loop composite mapping, so as to maximize the steady-state convergence of the motion closed loop instead of minimizing the external tracking error as the training objective. During the topology growth process, the metric parameters and topological invariants of the original closed submanifold remain unchanged, and the newly added topology does not modify the original topological structure.

12. The multi-layer nested service robot control method according to claim 1, characterized in that, In step S3, the task instructions are mapped to the topological deformation of the task planning layer. Specifically, the speech signal is converted into text semantic features through speech recognition, and the task vocabulary is mapped to the feature deformation pattern of the task planning layer topology through the semantic topology encoder. The semantic logic combination and topological algebra operation satisfy the homomorphic relationship, that is, the semantic combination of multiple actions corresponds to the connectivity and operation of the homotopy class of the corresponding actions. The semantic topology encoder is initially trained with instruction-task topology pairing data as seeds. After deployment, the parameter optimization is performed using whether the instruction execution successfully converges to the corresponding fixed point as a self-supervised signal.

13. The multi-layer nested service robot control method according to claim 1, characterized in that, The specific method for decoding the sub-target sequence from the steady-state situation in step S3 is as follows: starting from the target fixed point converged to the situation in the task planning layer, backtracking back to the initial situation along the lowest energy path of the topological constraints, and taking the key topological nodes on the path that correspond to the action mode switching or contact state change as sub-targets, with each sub-target corresponding to a standard action mode in the motion mode layer.

14. The multi-layer nested service robot control method according to claim 1, characterized in that, If the task planning layer situation fails to converge to the operation identity fixed point within the preset maximum number of iterations, the system outputs a "cannot be executed" response, without forcibly generating planning results, thus avoiding task planning illusions.

15. The multi-layer nested service robot control method according to claim 1, characterized in that, The security rule is encoded as a permanently infeasible region in the metacognitive layer topology. This permanently infeasible region is an endogenous immutable boundary of the corresponding hierarchical topological space. The motion state can only evolve in the feasible domain manifold and cannot break through the manifold boundary constraint by adjusting parameters.

16. The multi-layer nested service robot control method according to claim 1, characterized in that, It also includes a topology growth step to expand the topology set of the corresponding level: when used in the motion pattern layer, topology growth expands the action topology set by adding action closed submanifolds; when used in the task planning layer, topology growth expands the task concept set by adding task topology equivalence classes; when used in the metacognition layer, topology growth expands the strategy library by adding strategy topology patterns.

17. The multi-layer nested service robot control method according to claim 16, characterized in that, The complexity criterion for topology growth is: the local complexity exceeds N times the global mean for k consecutive evolution steps, where k is 5 to 20 and N is 2 to 5; or the success rate of the current task is continuously lower than a preset threshold. The energy criterion is that the system energy reserve is higher than the critical threshold. In the pure electric drive scenario, the system energy reserve is the weighted value of the power reserve and the computing power reserve. In the electromechanical coupling scenario, it also includes the load reserve of the actuator.

18. The multi-layer nested service robot control method according to claim 16, characterized in that, Topology growth is performed as follows: Step a: The initial metric tensor of the newly added closed submanifold is set to the metric value corresponding to the global average curvature, and the initial closed-loop composite mapping Lipschitz constant is set to a preset value greater than 1; Step b: In the boundary region between the new topology and the adjacent topology, local metric smoothing iteration is performed, and in each step, the metric tensor of the boundary region node is updated to its weighted average with the metric tensor of the neighboring node; Step c: The maximum magnitude of the curvature gradient in the boundary region is calculated. When the maximum magnitude is less than the preset curvature threshold, the bonding is deemed successful. The preset curvature threshold is dynamically calibrated based on 5% of the global average curvature. Step d: Perform operational closure verification on the new topology, including composite closure verification, compression factor verification, and boundary independence verification. Composite closure verification means that the proportion of pairwise composite results of the new topology and the existing topology belonging to the valid set is not less than 95%. Compression factor verification means that the Lipschitz constant of the closed-loop composite mapping of the new topology converges to less than 1. Boundary independence verification means that the bottleneck distance of the persistent cohomology barcode of the new topology and the existing topology is greater than 5% of the feature scale of the configuration space. Step e: After the verification is passed, the new topology is solidified into a closed submanifold of the corresponding level, the Euler characteristic of the level undergoes an integer quantization transition, and the identity fingerprint of the level is updated. Step f: If any of the steps b to d fails the verification, the newly added topology is automatically deleted and the system reverts to the state before growth. If the fluctuation amplitude of the overall organizational entropy after growth exceeds twice the baseline fluctuation amplitude before growth, the growth is judged to have failed, and the system automatically reverts. After reverting, the identity fingerprint of the corresponding level is restored to the state before growth. The organizational entropy refers to the information entropy of the situation distribution, which is used to characterize the steady state of the system.

19. The multi-layer nested service robot control method according to claim 18, characterized in that, After topology growth, a dual-verification closed loop is performed: statistical performance verification includes three levels: logical consistency verification, digital twin deduction verification, and practical operation performance verification. Logical consistency verification means that the new topology has no operational conflicts with existing rules and the composite closure degree is ≥95%. Digital twin deduction verification means that the deviation of the simulation trajectory in 50 steps is less than 15%. Practical operation performance verification means that the average score of 20 actual tests is greater than 0.

7. Geometric topology verification includes three levels: volume compactness verification, dual reconstruction accuracy verification, and joint Chern number topology verification. Volume compactness verification means that the fluctuation is less than 5% for 5 consecutive periods. Dual reconstruction accuracy verification means that the root mean square error is less than 5%. Joint Chern number topology verification means that in the joint topology structure formed by the new topology and the physical properties of the ontology, the joint Chern number transitions from 0 to a non-zero stable integer and remains stable for 10 consecutive observation periods. Only when both verifications pass is the newly added topology confirmed as valid growth. Topologies that fail verification are automatically rolled back to the state before growth.

20. The multi-layer nested service robot control method according to claim 16, characterized in that, During topology growth, the metric parameters and topological invariants of the original closed submanifold remain unchanged, and the newly added topology does not modify the original topological structure, thus achieving zero-forgetting incremental learning from a geometric mechanism.

21. The multi-layer nested service robot control method according to claim 1, characterized in that, It also includes a human-machine collaborative topology switching step: when the visual sensor detects that a human body has entered the preset interaction space, the system automatically switches to the human-machine collaborative topology, internally shrinks the force limit, speed limit, and acceleration limit to the preset safety threshold, and reduces the contact stiffness; when the human body leaves the interaction space, it automatically restores the normal topology; the shrinkage ratio can be adjusted according to the scene safety level.

22. The multi-layer nested service robot control method according to claim 1, characterized in that, It also includes a body-boundary dual non-intrusive identity authentication step: collecting the system boundary output sequence, constructing the Laplace matrix on the boundary state space, calculating the eigenvalue spectrum, counting the algebraic multiplicity of zero eigenvalues ​​and near-zero eigenvalues ​​with moduli less than a preset threshold, mapping the multiplicity pattern to the Betti number sequence of each layer, generating a topological identity fingerprint, and comparing it with the baseline topological identity fingerprint; when the topological charge undergoes an unexpected integer quantum jump, the identity verification is determined to have failed.

23. The multi-layer nested service robot control method according to claim 1, characterized in that, It also includes five-dimensional coherence monitoring: the system continuously monitors the coherence of each layer's material structure, energy mechanism, information field state, time synchronization, and spatial alignment in five dimensions; the coherence of each dimension is the normalized deviation between the current state of that dimension and the baseline steady state, with a value from 0 to 1, where 0 represents complete coherence and 1 represents complete mismatch; the fault is located by the five-dimensional coherence cross-validation matrix: if a single dimension first shows an anomaly and the associated dimension lags behind by at least 2 corresponding layer endogenous time quanta, the single dimension is determined to be the root cause of the fault; if multiple dimensions show anomalies simultaneously and the timing difference is less than 1 corresponding layer endogenous time quantum, it is determined to be a common source fault of the inter-layer coupling channel.

24. The multi-layer nested service robot control method according to claim 1, characterized in that, Each layer uses an endogenous time quantum as the smallest discrete time unit for its evolution. The endogenous time quantum is the discrete time interval of a single-step topological evolution of that layer. The endogenous time quantum of the perception-action layer is no longer than 1 ms, and the endogenous time quantum of the metacognition layer is no longer than 1 s. The endogenous time period between layers satisfies the integer frequency division relationship. The evolution of different layers proceeds on the discrete time axis defined by the endogenous time quantum of that layer, and the layers are aligned through time synchronization topological coupling.

25. A multi-layered nested service robot control system, deployed on a service robot onboard heterogeneous computing architecture comprising a multimodal sensor array, a computing unit, a storage unit, and an actuator, characterized in that, The system comprises a perception-action layer, a motion mode layer, a task planning layer, and a metacognitive layer, sequentially coupled along the data flow direction. The perception-action layer and motion mode layer are deployed on the airborne end and coupled via a real-time bus. The motion mode layer and task planning layer are deployed on the airborne end and coupled via a high-speed data interface. The metacognitive layer is deployed on the airborne end or on an edge computing node directly connected to the airborne end. The task planning layer is coupled to the metacognitive layer via a high-speed communication link. The multimodal signal input of the perception-action layer is connected to a sensor array, and its topology output is coupled to the configuration space input of the motion mode layer. This is used to map multimodal perception signals to a joint configuration manifold, forming a unified configuration-contact topology using the joint configuration manifold as a unified reference topology space. Closure checks are performed on the set of atomic operations, intrinsically defining the perception-action boundary. The action output of the motion mode layer is coupled to the actuator drive end, used to map basic actions to homotopy-like closed trajectories on the joint configuration manifold, constructing an action topology set. The motion mode layer's closed-loop composite mapping has a Lipschitz constant less than 1. As a criterion for action convergence; the command input end of the task planning layer is connected to the voice interaction module, and the planning output end is coupled to the target input end of the motion mode layer. It is used to map the task command into the topological deformation mode and use it as a situational perturbation. Through Hamiltonian flow evolution, the situation is converged to the operation identity fixed point, and the sub-target sequence is decoded from the fixed point. The security control terminal of the metacognitive layer is coupled downward to the constraint interface of each layer. It is used to encode security rules into permanent infeasible regions in the metacognitive layer topology and propagate them downward causally. When the complexity criterion and energy criterion are satisfied at the same time, topological growth is performed. The system identity continuity is maintained by the Betti number sequence and Euler characteristic of each layer. Information is transmitted between layers only through the coupling deformation of the M-topology of each layer. The information field state rules and energy mechanism driving force allocation process of each layer are retained within their respective layers. Each layer contains three functional primitives: information field state, energy mechanism, and M-topology, which form a closed-loop composite mapping.

26. The multi-layer nested service robot control system according to claim 25, characterized in that, It also includes at least one of the following: a topology growth module, a human-machine collaborative topology state switching module, a body boundary duality authentication module, and a fault rollback module; the topology growth module is used to perform topology growth by adding a closed submanifold when the complexity criterion and the energy criterion are satisfied at the same time; the human-machine collaborative topology state switching module is used to automatically shrink the force limit and velocity boundary when the human body enters the preset interaction space; the body boundary duality authentication module is used to non-invasively extract the topology identity fingerprint through the boundary output spectrum features; and the fault rollback module is used to automatically restore the topology to the state before growth when the topology becomes unstable after growth.

27. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-layer nested service robot control method based on operation closure as described in any one of claims 1 to 24.