A body intelligent mobile robot based on semantic guidance and multi-strategy fusion knowledge variation system and algorithm

Abstract

This invention relates to a knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, comprising: a semantic encoder, which encodes the topological structure and node and / or edge semantic information of a process knowledge graph into a unified structural semantic joint embedding vector; a mutation operator, which performs multi-level constraint verification through symbolic reasoning and lightweight physical simulation to generate a candidate knowledge graph set; a fitness evaluator, which performs multi-objective performance pre-evaluation on the candidate knowledge graph set generated by mutation and calculates the comprehensive fitness; and a selection mechanism, which dynamically guides the direction of knowledge evolution to form a closed-loop knowledge evolution process. This application achieves dynamic automatic evolution and closed-loop optimization of the process knowledge graph through collaborative work, enabling dynamic automatic evolution of the process knowledge graph, improving exploration efficiency, ensuring safe verification, and forming a closed-loop knowledge evolution process, thereby enhancing the autonomous adaptability of multi-armed mobile robots in dynamic industrial scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a knowledge mutation system and algorithm for an embodied intelligent mobile robot based on semantic guidance and multi-strategy fusion. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, multi-arm mobile robots are widely used in complex, unstructured environments. Traditional methods rely on manual teaching or pre-programmed skill sets, resulting in static knowledge representation and experience-dependent skill generation. These methods cannot self-adjust or expand according to new tasks and environments, and their generalization ability is limited. In recent years, reinforcement learning and digital twin technologies have been used for robot skill learning, but most employ random trial and error or simple parameter optimization, leading to problems such as low exploration efficiency, large sample requirements, poor interpretability of generated skills, and difficulty in ensuring safety and physical feasibility. Furthermore, existing methods typically separate "knowledge" from "execution," preventing the knowledge base from automatically evolving under execution feedback. This results in slow updates to the robot's skill set, making it unable to adapt to rapidly changing task requirements.

[0003] In summary, the existing technology mainly has the following problems: 1. Static knowledge bases cannot evolve on their own: Traditional process skill bases rely on manual writing or offline optimization and cannot be automatically updated and expanded under the drive of execution feedback; 2. Random search is inefficient and uninterpretable: Existing reinforcement learning methods have a huge exploration space in complex multi-armed collaborative tasks, high sample requirements and difficult-to-interpret generation strategies; 3. Lack of safety guarantees in skill generation: Blind mutation may produce strategies that violate the laws of physics or safety regulations and cannot be directly applied to real robots; 4. Separation of knowledge and execution: Most systems separate planning and control, and disconnect knowledge generation from performance feedback, making it impossible to achieve closed-loop evolution; 5. Insufficient cross-task generalization ability: Manually designed skills rely on specific scenario parameters and are difficult to transfer to new tasks or new environments. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a knowledge mutation system and knowledge mutation algorithm for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, which can realize the dynamic automatic evolution of process knowledge graph, improve exploration efficiency, ensure safety verification, and form a closed-loop knowledge evolution process, thereby enhancing the autonomous adaptability of multi-arm mobile robots in dynamic industrial scenarios.

[0005] To address the aforementioned technical problems, this invention provides a knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, used for the automatic evolution of the process knowledge graph of a multi-armed mobile robot. The system includes: a semantic encoder configured to encode the topological structure and semantic information of nodes and / or edges of the process knowledge graph into a unified structural semantic joint embedding vector through graph neural network and symbolic and / or linguistic embedding, thereby providing a semantically aware graph representation for mutation; and a mutation operator configured to, based on the structural semantic joint embedding vector, perform addition, deletion, merging, splitting, and subgraph reorganization of nodes and / or edges in the structural dimension, and perform bounded random perturbation and... Yesian optimization guides the adjustment, and multi-level constraint verification is performed through symbolic reasoning and lightweight physical simulation to generate a set of candidate knowledge graphs. The fitness evaluator is configured to predict multi-objective performance based on graph neural network regression model and Gaussian process, and combine semantic consistency score and feasibility indicator to pre-evaluate the multi-objective performance of the candidate knowledge graph set generated by mutation, and calculate the comprehensive fitness. The selection mechanism is configured to dynamically guide the direction of knowledge evolution based on the comprehensive fitness, using elite retention, Bayesian optimization exploration, graph structure cross-reorganization and evolutionary state management strategies, to achieve a balance between the inheritance of high-quality knowledge and the efficient exploration of new knowledge, forming a closed-loop knowledge evolution process.

[0006] The embodied intelligent mobile robot provided in this application is based on a knowledge mutation system that combines semantic guidance and multi-strategy fusion. Through the collaborative work of a semantic encoder, mutation manipulator, fitness evaluator, and selection mechanism, it achieves dynamic automatic evolution and closed-loop optimization of the process knowledge graph. This enables dynamic automatic evolution of the process knowledge graph, improves exploration efficiency, ensures safety verification, and forms a closed-loop knowledge evolution process, thereby enhancing the autonomous adaptability of the multi-arm mobile robot in dynamic industrial scenarios.

[0007] In one embodiment of the present invention, the process knowledge graph is G_t=(V,E); where V is a set of nodes, E is a set of edges, each node v_i contains an attribute vector A_i, and each edge e_j contains a relation type R_j.

[0008] In one embodiment of the present invention, the semantic encoder employs a graph convolutional network or a graph attention network as a structural feature extractor, and takes the adjacency matrix A_d and the node feature matrix X as input. Through multi-layer graph convolution operations H^(l+1)=σ(D^{-1 / 2}A_dD^{-1 / 2}H^(l)W^(l)), it aggregates neighbor structural information layer by layer to obtain the structural embedding vector S_i∈ ^d_s; where D is the degree matrix, W^(l) is the learnable weight matrix, and σ is the activation function.

[0009] In one embodiment of the present invention, the semantic encoder performs domain word embedding or pre-trained language model encoding on the node v_i and the relation type R_j, mapping each symbolized text t_k to a first semantic vector T_k∈ ^d_t, and combined with ontology embedding, the relation type R_j is vectorized as O_j∈ ^d_o.

[0010] In one embodiment of the present invention, the semantic encoder introduces a multi-head attention mechanism in the fusion layer, calculates the attention weights α{i,j}=softmax(f(S_i,T_i,O_j)) between the structural embedding vector S_i and the second semantic vector T_i and the relation vector O_j, and generates a weighted sum F_i=∑jα_{i,j}·[S_i‖T_i‖O_j]; where ‖ represents vector concatenation, and the final structural semantic joint embedding vector E_i∈ is obtained after linear transformation and normalization. ^d_e.

[0011] In one embodiment of the present invention, the mutation operator includes two collaborative working units: a structural mutation module and a parameter mutation module. The structural mutation module adopts a dual-channel mechanism of graph structure editing operation and symbolic reasoning verification. The graph structure editing operation guides the node mutation operation, and the symbolic reasoning verification drives the edge mutation verification. Through the subgraph recombination operation of the VF2 subgraph isomorphism algorithm, the symbolic reasoning verification is called after each edge operation to verify the logical legality of the graph structure. The parameter mutation module adopts a hybrid strategy of bounded random perturbation and Bayesian guided adjustment. It performs continuous parameter perturbation on node attributes by Gaussian noise injection, uniform interval sampling, or discretized enumeration, and performs associated parameter mutation on edge attributes guided by historical performance data combined with current fitness prediction. After all parameter mutations, the symbolic reasoning verification is called again to verify whether the parameters meet the predefined constraints.

[0012] In one embodiment of the present invention, the fitness evaluator includes: a performance prediction submodule, a semantic consistency evaluation submodule, a feasibility verification submodule, and a comprehensive scoring and decision submodule; the performance prediction submodule adopts a hybrid architecture combining a graph neural network regression model and a Bayesian performance predictor, concatenating the candidate structural semantic joint embedding vector with the environment state vector as the model input, and outputting multi-objective performance prediction values ​​and prediction mean and variance; the semantic consistency evaluation submodule calculates the graph embedding cosine similarity and ontology relation distance of the graph before and after mutation, normalizes and weights the sum to obtain the semantic consistency score SC∈[0,1]; the feasibility verification submodule performs symbolic logic verification through a solver and calls a lightweight physics engine to perform physical simulation verification, doubly checking the physical realizability of the candidate graph; the comprehensive scoring and decision submodule normalizes, weights and fuses, and thresholds the outputs of the performance prediction submodule, the semantic consistency evaluation submodule, and the feasibility verification submodule.

[0013] In one embodiment of the present invention, the selection mechanism includes: an elite retention submodule, a Bayesian optimization exploration submodule, a crossover and recombination submodule, and an evolutionary state management submodule; the elite retention submodule employs a deterministic Top-K screening combined with a historical performance weighting strategy, selects an elite set E_set in descending order of fitness vector main scores, and calculates the historical performance index (HPI) of candidates in the elite set E_set. i =γ·Average success rate of all executions + (1 γ)·The inverse of the stability variance of each execution, marking candidates whose HPii is higher than a preset threshold as G_parent or G_i; where γ is an adjustable weight; the Bayesian optimization exploration submodule uses a Gaussian process combined with an expected improvement strategy to construct a model for non-elite candidates and calculate the expected improvement value EI(X)=∫_{y}^{∞}(y The algorithm y)·p(y|X)dy selects the top K2 candidates with the largest expected improvement value as the exploration set Expl_set; where y is the current best fitness value and p(y|X) is the probability density function predicted by the Gaussian process; the crossover and recombination submodule performs subgraph isomorphic matching based on the VF2 algorithm on the elite set and the exploration set, and generates offspring candidate G_child by performing structural single-point crossover, parametric arithmetic crossover or BLX-α crossover in the matching region. After feasibility verification, the legal offspring are added to the elite set; the evolutionary state management submodule establishes an evolutionary state table EST, records the key data of each round of evolution, dynamically adjusts the parameters in the evolutionary strategy configuration according to the convergence index and diversity index in the evolutionary state table EST, and feeds the updated evolutionary state table EST back to the Bayesian optimization model and mutation operator.

[0014] This invention also provides a knowledge mutation algorithm for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, comprising the following steps: First, the semantic encoder receives the process knowledge graph at the current moment and uses graph neural networks and symbolic and / or language embedding techniques to encode its topological structure and semantic information on nodes and / or edges into a unified semantic joint embedding vector. Next, the mutation operator uses semantic joint embedding vectors to perform controlled and interpretable mutations on the knowledge graph in both structural and parametric dimensions. In the structural dimension, it performs operations such as adding, deleting, merging, splitting, and reorganizing subgraphs of nodes and / or edges. In the parametric dimension, it applies bounded random perturbations to continuous parameters and guides the adjustment through Bayesian optimization. At the same time, the mutation operator calls the symbolic inference engine and lightweight physical simulation to perform multi-level constraint verification to ensure that the generated candidate knowledge graph is logically valid, physically feasible, and semantically consistent. Subsequently, the fitness evaluator performs a rapid, multi-dimensional pre-evaluation of the candidate knowledge graph set generated by the mutation operator. It uses a graph neural network regression model and a Gaussian process to predict the multi-objective performance of each candidate knowledge graph in the target execution environment. It also calculates a comprehensive fitness score by combining semantic consistency score and feasibility flag. Finally, the selection mechanism dynamically guides the direction of knowledge evolution based on the comprehensive fitness score output by the fitness evaluator. It ensures that high-fitness, high-quality knowledge is inherited to the next generation through an elite retention strategy; intelligently discovers and explores new regions with high potential but not yet fully validated through Bayesian optimization of the exploration strategy; generates offspring candidates that combine the advantages of the parent generation with the novelty of exploration through graph structure cross-recombination; and dynamically adjusts the evolution strategy through evolutionary state management to achieve rapid convergence and continuous maintenance of diversity in the knowledge evolution process.

[0015] Compared with the prior art, the above-mentioned embodied intelligent mobile robot of the present invention, based on a knowledge mutation system and algorithm of semantic guidance and multi-strategy fusion, has the following advantages: 1. Overcoming the limitations of manual teaching and random search, it enables automatic and semantically guided proactive evolution of process knowledge graphs, generating entirely new skill combinations and constraint configurations, significantly expanding the robot's capability boundaries.

[0016] 2. By pre-evaluating fitness and verifying multi-level constraints, low-potential and infeasible variations are eliminated before simulation or real-world trial and error, significantly reducing computational and physical experiment costs.

[0017] 3. Combine symbolic reasoning with physical simulation verification to ensure that the mutation results conform to process logic, safety specifications and physical laws, and avoid generating dangerous or invalid skills. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an architecture diagram of the embodied intelligent mobile robot's knowledge mutation system and algorithm based on semantic guidance and multi-strategy fusion, as presented in this invention. Figure 2 This is the internal architecture of the mutation operator of the present invention. Detailed Implementation

[0020] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0021] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] It should be noted that the following description covers various aspects of embodiments within the scope of the appended claims. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0023] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0025] Traditional multi-arm mobile robots rely on static human instruction or pre-programmed skill sets, which limits the self-evolution of knowledge and generalization capabilities. Existing reinforcement learning methods suffer from low exploration efficiency, large sample requirements, poor skill interpretability, and difficulties in ensuring safety and physical feasibility. Furthermore, the separation of knowledge and execution leads to slow skill set updates, making it difficult to adapt to rapidly changing task requirements.

[0026] To address this, this application proposes a knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, used for the automatic evolution of the process knowledge graph of multi-armed mobile robots. The system includes: a semantic encoder, a mutation operator, a fitness evaluator, and a selection mechanism. The semantic encoder is configured to encode the topological structure and semantic information of nodes and / or edges of the process knowledge graph into a unified structural semantic joint embedding vector through graph neural network fusion with symbolic and / or linguistic embeddings, thereby providing a semantically aware graph representation for mutation. The mutation operator is configured to add, delete, merge, split, and reorganize nodes and / or edges in the structural dimension, and in the parameter dimension, based on the structural semantic joint embedding vector. Bounded random perturbation and Bayesian optimization are applied to guide adjustments, and multi-level constraint verification is performed through symbolic reasoning and lightweight physical simulation to generate a set of candidate knowledge graphs. The fitness evaluator is configured to predict multi-objective performance based on graph neural network regression models and Gaussian processes, and combines semantic consistency scores and feasibility indicators to pre-evaluate the multi-objective performance of the candidate knowledge graph set generated by mutation, and calculate the comprehensive fitness. The selection mechanism is configured to dynamically guide the direction of knowledge evolution based on comprehensive fitness, using elite retention, Bayesian optimization exploration, graph structure cross-reorganization, and evolutionary state management strategies to achieve a balance between the inheritance of high-quality knowledge and the efficient exploration of new knowledge, forming a closed-loop knowledge evolution process.

[0027] Among them, the process knowledge graph refers to a graph-structured data model used to describe the knowledge required for multi-arm mobile robots to perform specific process tasks. It contains nodes (representing entities, concepts, or operations) and edges (representing relationships or attributes between them). Graph Neural Networks (GNNs) are deep learning models designed to directly process graph-structured data. They learn node and graph representations by aggregating neighbor node information, effectively capturing graph topology and node features. Symbolic and / or linguistic embeddings are techniques for mapping discrete symbolic information (such as text, category labels) or natural language text to a continuous low-dimensional vector space. These vectors capture the original symbolic or linguistic semantic information, facilitating processing by machine learning models. Structural-semantic joint embedding vectors are unified vector representations generated by a semantic encoder, integrating the process knowledge graph's topological information with node and / or edge semantic information, providing a comprehensive semantically aware graph representation for subsequent knowledge mutation operations. Symbolic reasoning refers to the process of automatically deriving new conclusions or verifying the consistency of existing knowledge through logical operations based on predefined logical rules and a knowledge base. In this system, it is used to verify the logical validity of the knowledge graph. Lightweight physics simulation: refers to the rapid simulation of robot actions or technological processes using simplified physical models or efficient computational methods to verify the physical feasibility and security of knowledge graphs. Graph neural network regression model: refers to a regression prediction model that utilizes a graph neural network architecture. It takes graph-structured data as input and learns graph feature representations to predict one or more continuous target values, such as multi-objective performance indicators. Gaussian process: a non-parametric probabilistic model used to model functions. It provides the prediction mean and prediction variance, thus quantifying prediction uncertainty, and is often used in Bayesian optimization and uncertainty estimation. Elite retention: a selection strategy in evolutionary algorithms that directly replicates high-fitness individuals (elites) from the current population to the next generation to ensure high-quality knowledge inheritance and algorithm convergence. Bayesian optimization exploration: a global optimization strategy that constructs a probabilistic surrogate model of the objective function (e.g., a Gaussian process) and uses a sampling function (e.g., desired improvement) to guide the selection of the next sampling point, efficiently exploring the search space and discovering the optimal solution. Graph-based crossover and recombination: A genetic operation based on graph-structured data, referring to the generation of new offspring knowledge graphs by exchanging or combining the subgraph structures of two parent knowledge graphs, thereby integrating the advantages of different parents and increasing population diversity. Evolutionary state management strategy: Refers to the real-time monitoring and recording of key indicators (such as convergence speed, diversity, and fitness distribution) during knowledge evolution, and the dynamic adjustment of evolutionary algorithm parameters or strategy configuration methods based on these indicators to optimize evolutionary efficiency and effectiveness.

[0028] The embodied intelligent mobile robot proposed in this application is a knowledge mutation system based on semantic guidance and multi-strategy fusion. It is an algorithm system that uses a process knowledge graph as a carrier and achieves automatic knowledge evolution through multi-strategy fusion under the combined effect of semantic guidance, structural perception, physical feasibility and safety constraints.

[0029] Its overall principle is as follows Figure 1 As shown, robot process knowledge is represented as a structured, computable knowledge graph. A semantic encoder fuses the graph's topological structure and node / edge semantic information into a unified vector representation, providing a semantic graph representation that is "computable, comparable, and interpretable" for mutation. Based on this, a mutation operator performs controlled, interpretable, and multi-modal mutation operations in both structural and parameter dimensions, including adding, deleting, merging, splitting, and reorganizing subgraphs, as well as parameter perturbation and intelligent guidance adjustments. At each step, multi-level constraint verification is performed through symbolic reasoning and lightweight physical simulation to ensure the mutation results are logically, physically, safely, and semantically valid. Subsequently, a fitness evaluator uses graph neural network regression and Gaussian process prediction to pre-evaluate the multi-objective performance of candidate mutations, and calculates the comprehensive fitness by combining semantic consistency and feasibility indicators, achieving low-cost and highly reliable candidate selection. Finally, the selection mechanism, through collaborative decision-making involving elite retention, Bayesian optimization exploration, graph structure cross-reorganization, and evolutionary state management, intelligently guides the exploration of high-potential new regions while maintaining verified high-quality knowledge, achieving rapid convergence and continuous diversity maintenance in the knowledge evolution process.

[0030] This system achieves automatic evolution of the process knowledge graph for multi-armed mobile robots through the synergistic operation of semantic encoders, mutation operators, fitness evaluators, and selection mechanisms. This system effectively addresses the issues of static knowledge bases and insufficient generalization ability in skill generation found in traditional methods, improving the efficiency and interpretability of knowledge exploration. Simultaneously, multi-level constraint verification ensures the physical feasibility and safety of generated skills and achieves closed-loop integration of knowledge and execution, enabling the robot's skill base to dynamically adapt to complex and ever-changing task requirements.

[0031] Furthermore, the semantic encoder: The input process knowledge graph G_t=(V,E) is subjected to structured parsing, where V is the node set and E is the edge set. Each node v_i contains an attribute vector A_i (such as object type, action type, physical parameters, constraint threshold), and each edge e_j contains a relation type R_j (such as "precondition", "post-effect", "constraint association"). Graph Convolutional Network (GCN) or Graph Attention Network (GAT) is used as the structure feature extractor. The adjacency matrix A_d and the node feature matrix X are used as inputs. Through multi-layer graph convolution operation H^(l+1)=σ(D^{-1 / 2}A_dD^{-1 / 2}H^(l)W^(l)), where D is the degree matrix, W^(l) is the learnable weight matrix, and σ is the activation function, the neighbor structure information is aggregated layer by layer to obtain the structure embedding vector S_i∈ ^d_s; Simultaneously, the node labels and edge relationships are encoded using domain word embedding or pre-trained language models (such as BERT, RoBERTa), mapping each symbolized text t_k to a first semantic vector T_k∈ ^d_t, and combined with ontology embedding (such as TransE, ComplEx) to vectorize the relation R_j to O_j∈ ^d_o; Subsequently, a multi-head attention mechanism is introduced in the fusion layer to calculate the attention weights α{i,j}=softmax(f(S_i,T_i,O_j)) between the structural embedding S_i and the second semantic vector T_i and the relation vector O_j, and to generate a weighted sum F_i=∑jα_{i,j}·[S_i‖T_i‖O_j], where ‖ represents vector concatenation. After linear transformation and normalization, the final structural-semantic joint embedding vector E_i∈ ^d_e; The implementation path is as follows: ① Receive the original graph structure and attribute / relation text of the knowledge graph; ② Perform GNN structure encoding and symbol / language embedding in parallel; ③ Perform attention weighting and feature concatenation in the fusion layer; ④ Output the joint embeddings E_i and E_j of each node v_i and edge e_j, which can be directly called by the knowledge mutation module, fitness evaluator and selection mechanism to realize the end-to-end conversion from symbolic process knowledge to computable and comparable vectorized representation, thereby maintaining structural consistency and semantic rationality in the knowledge mutation process, and supporting intelligent search and performance prediction based on vector space similarity.

[0032] Specifically, a process knowledge graph G_t=(V,E) is a knowledge base that represents process knowledge in a graph structure. It models entities in the process domain and their relationships as nodes and edges. G_t=(V,E) is the standard form of graph representation in graph theory, where G_t represents the process knowledge graph at time t, V represents the node set, and E represents the edge set. This explicit graph structure definition provides a unified and standardized data interface and processing framework for subsequent semantic encoding, mutation operations, and fitness evaluation. For example, it can be stored and represented in a computer using data structures such as adjacency matrices or adjacency lists. Here, V is the node set, which is the collection of all entities in the process knowledge graph. These entities can be specific process steps, robot actions, tools, workpieces, environmental parameters, or constraints, etc. The node set defines all independent information units in the knowledge graph and is the basic element for constructing the graph. Nodes can be represented by unique identifiers (e.g., UUIDs or integer IDs) and associated with their attribute data. E represents the edge set, a collection of relationships between all entities in the process knowledge graph. These relationships can be sequential, causal, compositional, or attribute-based, etc. The edge set connects nodes, representing semantic associations between them, thus constructing a complex knowledge network. Edges can be represented as ordered or unordered pairs (u, v) and associated with their relation types. Each node v_i contains an attribute vector A_i, which is a set of numerical or symbolic information describing the features of node v_i. For example, a node representing a "grabbing" action might have attribute vectors containing parameters such as gripping force, gripping speed, and gripping posture; a node representing a "workpiece" might have attribute vectors containing information such as size, material, and weight. Attribute vectors provide rich semantic details for nodes, enabling the semantic encoder to capture the deeper meaning of nodes and providing a basis for the mutation operator to adjust parameters. Attribute vectors can be floating-point arrays, one-hot encoded vectors, or word embedding vectors, etc. Each edge e_j contains a relation type R_j, which is a label or identifier describing the semantic relationship represented by edge e_j. For example, relation types can be "precedes", "requires", "has_part", or "is_a". Relation types clearly define the specific semantic connection between the two nodes connected by the edge, which is crucial for understanding the logical structure of the knowledge graph and performing accurate reasoning. Relation types can be string labels, integer codes, or represented by pre-trained relation embedding vectors.

[0033] The above technical solution clearly defines the mathematical structure and components of the process knowledge graph, providing a precise data foundation for the knowledge mutation system and thus solving the operational inconsistency problem caused by the lack of a clearly defined graph structure. Specifically, by clearly defining the process knowledge graph as G_t=(V,E), where V is the node set and E is the edge set, each node v_i contains an attribute vector A_i, and each edge e_j contains a relation type R_j, this application provides a unified and precise data foundation for the semantic encoder, mutation operator, fitness evaluator, and selection mechanism. The semantic encoder can accurately encode the graph's topological structure and semantic information such as node attributes and edge relation types into a unified structural semantic joint embedding vector based on the explicit node set and edge set, avoiding the encoding inaccuracy problem caused by structural ambiguity. The mutation operator, based on well-defined nodes and edges, can precisely perform operations such as adding, deleting, merging, splitting, and reorganizing subgraphs at the structural dimension. At the parameter dimension, it guides adjustments through bounded random perturbations and Bayesian optimization based on node attribute vectors and edge relationship types. Simultaneously, it utilizes explicit relationship types for symbolic reasoning verification to ensure the validity and logical legitimacy of mutations. The fitness evaluator can perform multi-objective performance pre-evaluation based on a well-structured and semantically clear candidate knowledge graph. Combining semantic consistency scoring (based on explicit node attributes and edge relationships) and feasibility indicators, it calculates a more accurate comprehensive fitness. The selection mechanism, based on a more accurate comprehensive fitness, dynamically guides the direction of knowledge evolution, achieving a balance between the inheritance of high-quality knowledge and the efficient exploration of new knowledge. This explicit definition fundamentally solves the processing inconsistencies, inefficiencies, and logical errors caused by undefined graph structures, significantly improving the robustness, interpretability, and evolutionary efficiency of the knowledge mutation system.

[0034] Furthermore, Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) are deep learning models specifically designed for processing graph-structured data, capable of effectively learning features from non-Euclidean data. They generate node representations by aggregating neighbor information. Specifically, GCNs capture local graph structure information by defining convolution operations on the graph, aggregating the features of each node with the features of its neighboring nodes. Graph Attention Networks introduce an attention mechanism, allowing the model to assign different weights based on the importance of neighboring nodes when aggregating neighbor information, thus capturing graph structure features more flexibly. Besides GCNs and GATs, those skilled in the art can also use other graph neural network models as structural feature extractors, such as GraphSAGE, which generates node embeddings by sampling and aggregating neighbor features, or Graph Isomorphism Networks (GINs), which demonstrate strong capabilities in distinguishing different graph structures. These networks all achieve efficient extraction and representation learning of graph structure information. The adjacency matrix A_d is a standard way to represent graph structures, where the elements A_d[i,j] of the matrix indicate whether there is a connection (or the weight of the connection) between node i and node j. It directly encodes the topological relationships of the process knowledge graph. The node feature matrix X contains the attribute vector A_i of each node v_i in the graph. These attributes can be the node type, functional description, parameter range, etc., providing rich semantic information for the graph neural network. These two inputs together form the basis for the graph neural network to process graph data, ensuring that the model can simultaneously consider the graph's connectivity and the intrinsic attributes of the nodes. The multi-layer graph convolution operation H^(l+1)=σ(D^{-1 / 2}A_dD^{-1 / 2}H^(l)W^(l)) is the core mechanism in the graph neural network used to update the node representation layer by layer. Here, H^(l) represents the node feature matrix of the l-th layer, and H^(l+1) represents the node feature matrix of the next layer obtained after one convolution operation. D^{-1 / 2}A_d is a symmetric normalization of the adjacency matrix A_d, designed to address the problem of excessively large differences in node degree in graph neural networks. It ensures that information propagation on the graph is not biased by node degree, thus stabilizing the training process and improving model performance. W^(l) is a learnable weight matrix that performs a linear transformation on the aggregated features in each convolutional layer, serving as a key parameter for the model to learn graph structure patterns. σ is the activation function, such as ReLU, Sigmoid, or Tanh, which introduces non-linear transformation capabilities, enabling the model to learn and represent more complex feature patterns and enhancing its expressive power. Through this multi-layer operation, nodes can progressively aggregate information from their multi-hop neighbors, thereby capturing a broader contextual structure. Layer-by-layer aggregation refers to the graph neural network stacking multiple graph convolutional layers, allowing each node to receive and integrate information from its neighboring nodes at each layer.The first layer aggregates information from direct neighbors, the second layer aggregates information from two-hop neighbors (i.e., neighbors of neighbors), and so on. This mechanism allows nodes to obtain their local and even global structural context information, thereby generating more representative structural embedding vectors. The structural embedding vector S_i∈. ^d_s is a low-dimensional dense vector of dimension d_s obtained after processing by the graph neural network. This vector compactly encodes the topological position of node v_i in the process knowledge graph, its neighbor relationships, and the contextual structure information obtained through multi-layer aggregation. As an abstract representation of the node's structural features, it provides structural input for subsequent semantic fusion and mutation operations. The degree matrix D is a diagonal matrix, where the diagonal elements D[i,i] represent the degree of node i (i.e., the number of edges connected to node i). In graph convolution operations, the degree matrix D is used to normalize the adjacency matrix to balance the influence of nodes with different degrees in information propagation, preventing nodes with high degrees from dominating the information flow, thereby ensuring fair and effective information propagation on the graph. The learnable weight matrix W^(l) is a parameter automatically adjusted by the graph neural network model during training through backpropagation and optimization algorithms. It determines how to perform linear transformations and feature mappings on the aggregated neighbor features in each convolution operation. By learning these weights, the model can adaptively capture and extract the graph structure patterns and features most useful for the task. The activation function σ introduces nonlinearity, enabling graph neural networks to learn and represent complex nonlinear relationships. Without the activation function, multi-layer graph convolution operations would be equivalent to a single-layer linear transformation, thus limiting the model's expressive power.

[0035] Through the above technical solutions, the semantic encoder can effectively solve the challenges of structural information extraction and aggregation in the process of encoding process knowledge graphs. Using graph convolutional networks or graph attention networks as structural feature extractors enables the model to automatically learn and capture the topological features of the graph, avoiding the tediousness and limitations of manual feature engineering in traditional methods. Using the adjacency matrix A_d and the node feature matrix X as input ensures a comprehensive consideration of the graph structure and node attributes. The multi-layer graph convolution operation H^(l+1)=σ(D^{-1 / 2}A_dD^{-1 / 2}H^(l)W^(l)), combined with the degree matrix D, the learnable weight matrix W^(l), and the activation function σ, achieves layer-by-layer aggregation of neighbor structural information. This allows each node to fully capture the contextual dependencies of its multi-hop neighbors, thus overcoming the problem that traditional methods cannot fully capture the contextual dependencies of neighbor nodes. By introducing the learnable weight matrix W^(l), the structural information extraction process becomes learnable, enabling adaptive optimization of feature representations based on data. The final structure embedding vector S_i∈ ^d_s not only compactly encodes the structural semantic information of nodes but also possesses stronger representational capabilities, significantly improving the quality of semantically aware graph representations. This high-quality structural embedding vector S_i provides more accurate and richer structural information for subsequent semantic fusion (such as fusion with language embeddings) and mutation operations, thereby improving the accuracy and efficiency of the entire knowledge mutation system and laying a solid foundation for the automatic evolution of process knowledge graphs for multi-arm mobile robots.

[0036] Furthermore, the semantic encoder performs domain word embedding or pre-trained language model encoding on nodes v_i and relation types R_j, aiming to transform these symbolic textual information into machine-understandable numerical vectors, thereby capturing their deep semantics. Domain word embedding refers to training word vector models on large-scale text corpora of specific domains (such as robotics or manufacturing processes), enabling the generated word vectors to reflect the domain-specific lexical relationships and semantics. For example, models such as Word2Vec, GloVe, or FastText can be used to learn their distributed representations by analyzing the co-occurrence patterns of words in domain-specific texts. Another approach is to use pre-trained language model encoding, such as large-scale language models like BERT, RoBERTa, or GPT. These models are pre-trained on massive amounts of general text and can capture rich linguistic knowledge and contextual semantics. By inputting the descriptive text of node v_i or relation type R_j into these pre-trained models, the output context embedding vector is extracted as its semantic representation. Subsequently, each symbolic text t_k is mapped to a first semantic vector T_k∈ ^d_t. This process is a direct result of the above encoding steps, that is, the text description t_k corresponding to node v_i or relation type R_j is encoded into a fixed-dimensional numerical vector T_k through domain word embedding or a pre-trained language model. For example, if domain word embedding is used, for a text t_k composed of multiple words, the embedding vectors of all its words can be averaged or weighted to obtain T_k. If a pre-trained language model is used, the model can directly output a vector representing the semantics of the entire text t_k as T_k. This first semantic vector T_k∈ The dimension d_t of ^d_t depends on the chosen embedding model. Furthermore, combining ontology embedding, the relation type R_j is vectorized as O_j∈ ^d_o. Ontology embedding aims to structurally represent concepts, attributes, and relations defined in a domain ontology as low-dimensional vectors. This not only captures the literal semantics of relation type R_j, but more importantly, it incorporates the logical position of the relation within the overall domain knowledge system and its association with other concepts. For example, knowledge graph embedding techniques (such as TransE, ComplEx, RotatE, etc.) can be used to train a domain ontology, mapping each relation type R_j in the ontology to a vector O_j. Another approach is to analyze the ontology's descriptive logic (such as OWL) or its graph structure, and then use graph neural networks (such as GCN, GAT) to learn the embeddings of relation types, enabling O_j to reflect the position of R_j in the ontology hierarchy and its inherent reasoning capabilities.

[0037] Through the above technical solution, the semantic encoder can transform symbolic text t_k into a first semantic vector T_k with rich semantic information by embedding domain terms or encoding the relation type R_j with pre-trained language models. Simultaneously, by combining ontology embedding to vectorize the relation type R_j into O_j, the accuracy and consistency of relation semantics are further enhanced. These semantic vectors T_k and O_j work synergistically with the aforementioned structural embedding vector S_i, enabling the semantic encoder to generate a more comprehensive and accurate joint structural-semantic embedding vector. This deeper semantic awareness effectively solves the problem of incomplete semantic representation caused by relying solely on structural embedding, thus providing a reliable foundation for the subsequent semantically aware and precise mutation of the mutation operator. This significantly improves the generation quality of the candidate knowledge graph set and enhances the accuracy and reliability of the fitness evaluator's multi-objective performance pre-evaluation.

[0038] Furthermore, multi-head attention is a neural network component that allows a model to learn multiple attention weights in parallel across different representation subspaces, thereby capturing dependencies between different positions in the input sequence from various perspectives. Each "head" independently computes attention, and then the outputs of these heads are concatenated and linearly transformed to obtain the final output. In this application, the purpose of introducing a multi-head attention mechanism is to capture the complex interactive information between structural embedding vectors, node semantic vectors, and relation vectors more comprehensively and flexibly, avoiding the biases that may be caused by a single attention mechanism, thus improving the flexibility and comprehensiveness of the fusion. For example, each attention head can employ a dot-product attention mechanism, computing attention scores through query, key, and value vectors, and then concatenating the outputs of multiple heads and performing a linear projection; alternatively, a feedforward neural network can be used to compute attention scores, combining query and key vectors through a hidden layer, and then obtaining attention weights through an activation function and a softmax layer. The calculation of attention weights α{i,j}=softmax(f(S_i,T_i,O_j)) is a core step in the multi-head attention mechanism. It quantifies the relative importance of different input elements (structural embedding vector S_i, second semantic vector T_i, and relation vector O_j) to the target output during the fusion process. The function f is typically a learnable transformation, such as a feedforward neural network, that maps the input vector to a scalar fraction, then normalizes it to a probability distribution using the softmax function, ensuring that the sum of all weights is 1. This mechanism dynamically determines the contribution of structural, node semantic, and relational information to the joint embedding, rather than using fixed weights. This allows the model to adaptively adjust the fusion strategy based on the current input features, more accurately reflecting the contribution of different elements to the joint embedding. When generating the weighted sum F_i=∑jα_{i,j}*[S_i‖T_i‖O_j], after calculating the attention weights α{i,j}, these weights are applied to the corresponding input feature vectors (here, the concatenated [S_i‖T_i‖O_j]), integrating information through weighted summation. This means that the information of feature vectors assigned high weights plays a more dominant role in the final fusion result. The weighted summation operation integrates structural, semantic, and relational information, and preserves the integrity of the original features based on the concatenated vector [S_i‖T_i‖O_j], avoiding information loss and providing rich input for subsequent processing. For example, the weighted summation of vectors can be performed directly according to the formula F_i=∑jα_{i,j}*[S_i‖T_i‖O_j]; or, before the weighted summation, a gating mechanism (such as the Sigmoid activation function) can be introduced to dynamically adjust the contribution of each concatenated vector according to the attention weights, achieving more refined information filtering and integration.Finally, the fused feature vector F_i is adjusted and normalized through linear transformation and normalization. Linear transformation typically refers to matrix multiplication of the vector using a learnable weight matrix to change its dimension or project it onto a new feature space. Normalization adjusts the numerical range of the vector to a specific interval (e.g., [0,1] or a mean of 0 and a variance of 1) to stabilize the training process and improve the model's generalization ability. This step ensures that the output structural semantic joint embedding vector E_i has numerical stability, dimensional consistency, and meets the input requirements of subsequent modules (such as the mutation operator), providing a high-quality structural semantic joint embedding foundation for the mutation system. For example, a fully connected layer (i.e., linear transformation) can be used to map the weighted sum F_i to the target dimension, and then LayerNormalization can be applied to normalize the result; alternatively, if F_i is considered as data with spatial structure, a 1D convolutional layer can be used for linear transformation, followed by Batch Normalization for normalization.

[0039] By introducing a multi-head attention mechanism into the fusion layer of the semantic encoder, this application can dynamically calculate the attention weights α{i,j} between the structural embedding vector S_i, the second semantic vector T_i, and the relation vector O_j, thereby achieving a more refined and comprehensive fusion of these heterogeneous information sources. This dynamic weight allocation mechanism enables the model to adaptively adjust the contributions of different information sources according to the actual situation of the input features, avoiding the limitation of representational ability caused by the inability of traditional simple fusion methods to capture complex dependencies. Specifically, by generating a weighted sum F_i=∑jα_{i,j}*[S_i‖T_i‖O_j], this application can effectively integrate structural, node semantic, and relational information, highlighting key information while preserving the integrity of the original features, providing high-quality input for subsequent linear transformation and normalization. The resulting joint structural and semantic embedding vector E_i can thus more accurately represent the topological structure and semantic content of the process knowledge graph, significantly improving the accuracy of controlled and interpretable mutations in the structural and parameter dimensions of the mutation operator, thereby improving the efficiency and quality of knowledge evolution.

[0040] Reference Figure 2As shown, the overall idea of ​​the mutation operator is to take the structure and parameters of the process knowledge graph as the mutation objects. Through controllable, interpretable, and semantically compliant graph structure operations and parameter perturbation mechanisms, while maintaining the logical consistency of the knowledge graph, it actively generates new knowledge graph variants with potential performance improvements or functional expansions. Its input is the knowledge graph G_t at the current time (including node set V, edge set E, node attribute A, edge attribute R), mutation strategy configuration C (such as mutation probability, structural perturbation strength, parameter perturbation range), and symbolic reasoning constraint set K (such as causal rules, safety boundaries, physical feasible regions). The output is a set of candidate knowledge graphs {G'1, G'2, ..., G'_k} after structural mutation and parameter mutation. Its goal is to achieve the active evolution of the knowledge graph, that is, to automatically explore new combinations of process actions, constraint configurations, or object attribute settings under the performance feedback of the execution layer, thereby breaking through the limitations of existing manual design or random search, and providing a diverse and high-quality knowledge candidate set for subsequent fitness evaluation and selection mechanisms.

[0041] The mutation operator is divided into two collaborative working units: a structural mutation module and a parameter mutation module. The structural variation module employs a dual-channel mechanism of graph structure editing operations and symbolic reasoning verification: First, based on graph traversal algorithms (such as DFS / BFS), mutable nodes v_i are identified. According to their type (such as "basic action", "process constraint", "object attribute"), the corresponding mutation sub-function is called. This includes adding new nodes (such as inserting a "detect workpiece position" node, whose attributes are generated by the domain thesaurus and historical successful trajectory statistics), deleting redundant nodes (such as removing "transition action" nodes that have no successor and no independent function), merging similar nodes (such as merging "light force gripping" and "standard gripping" into a "adjustable force gripping" node, taking the union of parameter ranges and introducing a continuous adjustment factor), and splitting composite nodes (such as splitting the "grip-rotate-place" node into three sub-nodes and automatically generating front / back edges).

[0042] Secondly, edge operations are performed on the connection relationships of the edges, including adding new edges (such as inserting a "detection" edge between "grab" and "place" and setting the "detection precision" attribute), deleting invalid edges (such as dangling edges caused by node deletion), and reconnecting edges (such as changing "grab → place" to "grab → detect → place" to introduce an intermediate check step). After each edge operation, the symbolic reasoning engine (such as the Z3 constraint solver) is called to verify the logical legality of the graph structure to ensure that the "precondition-posteffect" causal chain, acyclic, and contradictory constraints are satisfied.

[0043] Finally, subgraph reorganization is performed on frequently occurring subgraph patterns (such as "grab-rotate-place"). All matching instances are located using the VF2 subgraph isomorphism algorithm. Selected instances are then locally rearranged (e.g., swapping the order of "rotate" and "detect") or migrated across graphs (e.g., migrating the "assembly" subgraph to the "packaging" task graph, with semantic adaptation of node / edge types and attributes guided by the domain ontology mapping table).

[0044] The parameter mutation module employs a hybrid strategy of bounded random perturbation and Bayesian guided adjustment: Continuous parameter perturbation is applied to node attributes A_i (such as force threshold, velocity, angle, and time parameters). The perturbation amplitude is controlled by the "parameter perturbation intensity λ" in the mutation strategy configuration C. The perturbation methods include Gaussian noise injection (A_i' = A_i + ε, ε ~ N(0, λσ_i)), uniform interval sampling (A_i' ∈ [A_i - δ, A_i + δ]), and discretization enumeration (such as changing "grabbing force" from {light, medium, heavy} to {light, medium, heavy, adjustable}).

[0045] The edge attributes R_j (such as constraint threshold, priority, dependency weight) are associated with parameter mutation. For example, the "synchronization weight" or "time delay" of the "grab → rotate" edge is adjusted. The mutation value is guided by historical performance data and current fitness prediction. The possible fitness improvement after parameter mutation is predicted by the Gaussian process regression model in Bayesian optimization. High potential areas are selected for perturbation first.

[0046] After all parameters have been mutated, the symbolic reasoning engine is called again to verify whether the parameters meet the predefined safety / physical constraints (such as "the gripping force must be greater than the weight of the workpiece" and "the rotation angle cannot exceed the limits of the robotic arm joint"). Invalid mutations will be filtered or rolled back.

[0047] Finally, the results of structural and parameter mutations are integrated by the mutation candidate pool manager to generate a set of candidate knowledge graphs {G'1, G'2, ..., G'_k} that are structurally valid, parameter-feasible, and semantically consistent. These graphs are then used by the fitness evaluator for performance prediction and selection mechanisms for evolutionary decision-making.

[0048] Graph structure editing operations refer to a series of operations that modify the topological structure of a knowledge graph. This can include, but is not limited to, basic graph operations such as adding, deleting, merging, and splitting nodes, as well as adding, deleting, and modifying edge types. For example, this can be achieved through a predefined set of graph editing rules, which, based on domain knowledge or heuristics, guide effective structural modifications to generate new knowledge structures with potential value. Symbolic reasoning verification is a mechanism based on formal logic or domain ontology rules to check the legality and consistency of the knowledge graph's structure or parameters. It can be implemented using a rule-based reasoning engine containing a set of predefined logical rules (e.g., based on OWL or Datalog rules) to check whether entities and relationships in the graph conform to ontology constraints or domain common sense. Another approach is to use a first-order logic or descriptive logic inference engine to convert the knowledge graph into logical expressions, and then verify its logical consistency through theorem proofs or model checks to ensure the semantically reasonable nature of the mutated graph. The dual-channel mechanism refers to a working mode in which mutation generation and verification are performed in parallel or alternately within the structural mutation module. Graph structure editing operations serve as the primary channel for generating mutations, while symbolic reasoning verification acts as an auxiliary channel, performing real-time or near-real-time checks on the legality of the generated mutations. For example, symbolic reasoning verification can be triggered immediately after each graph structure editing operation (such as adding or deleting nodes / edges), forming a feedback loop to ensure that each mutation conforms to preset logical constraints, thereby avoiding the generation of invalid or illegal graph structures. The subgraph recombination operation of the VF2 subgraph isomorphism algorithm utilizes the VF2 algorithm to efficiently identify isomorphic subgraphs in the graph and performs structural recombination based on these subgraphs to generate new graph structures. The VF2 algorithm can efficiently search for all isomorphic subgraphs between two graphs, thereby identifying structural patterns that can be replaced, merged, or split. Recombination operations can involve replacing the identified subgraph with another predefined subgraph or a subgraph generated through mutation, or using the VF2 algorithm to identify subgraphs with specific patterns in the current graph, and then copying, moving, or interleaving these subgraphs with other subgraphs to generate candidate graphs with novel structures. Bounded random perturbation refers to making random, small-amplitude modifications to parameters within a preset numerical range. For continuous parameters, Gaussian noise injection can be used, which adds a random number following a Gaussian distribution to the current parameter value, ensuring the result is within preset upper and lower limits. For discrete parameters, uniform interval sampling can be used, which randomly selects a new value within the parameter's allowed range, or discretization enumeration can be performed to select a value from a predefined finite set, increasing the diversity of parameter variation. Bayesian guided adjustment uses Bayesian optimization methods to intelligently adjust the direction and magnitude of parameter variation based on historical performance data and current fitness predictions.Bayesian optimization efficiently searches for optimal parameters by constructing a probabilistic surrogate model of the objective function (such as a Gaussian process) and using a sampling function (such as the desired improvement in EI) to guide the location of the next parameter sampling. For example, by combining data such as the success rate of historical mutations and fitness improvement, a Bayesian network or Gaussian process regression model can be used to predict the performance gains that different parameter mutation directions may bring, and the mutation with the largest expected gain can be prioritized. Related parameter mutation is parameter mutation targeting edge attributes. Its adjustment not only considers itself but also incorporates historical performance data and current fitness predictions. For example, a regression model can be established by analyzing the correlation between different edge attribute values ​​and system performance in historical data. During mutation, based on the fitness prediction of the current graph, this model guides the direction and magnitude of edge attribute parameter adjustment.

[0049] Through the above technical solution, the mutation operator is divided into two collaborative working units: a structural mutation module and a parameter mutation module. This achieves division of labor and cooperation in mutation operations, effectively solving the problem of insufficient constraint verification during the mutation process. The structural mutation module adopts a dual-channel mechanism of graph structure editing operation and symbolic reasoning verification. Graph structure editing operation guides node mutation operation, allowing for the addition, deletion, and other editing of nodes in the knowledge graph. Symbolic reasoning verification drives edge mutation verification, performing logical legality verification of the graph structure after each edge operation by calling symbolic reasoning verification to ensure that the mutated graph is logically valid. Subgraph recombination operation is implemented through the VF2 subgraph isomorphism algorithm, improving the efficiency and accuracy of mutation. The parameter mutation module adopts a hybrid strategy of bounded random perturbation and Bayesian guided adjustment. It performs continuous parameter perturbation on node attributes through Gaussian noise injection, uniform interval sampling, or discretized enumeration, providing diverse parameter change methods. At the same time, it performs association parameter mutation on edge attributes guided by historical performance data and current fitness prediction. Based on historical data and prediction to guide the mutation direction, it improves the targeting and efficiency of mutation. After all parameters are mutated, symbolic reasoning verification is invoked again to confirm whether the parameters meet the predefined constraints, ensuring that the mutated parameters are physically feasible and safe. This design ensures that the mutation operation, when generating candidate knowledge graphs, can both efficiently explore new knowledge and strictly guarantee logical validity and physical feasibility, thereby improving the overall effect of knowledge evolution, avoiding the generation of invalid or dangerous knowledge graphs, and significantly improving the efficiency and security of knowledge evolution.

[0050] In this application, the fitness evaluator performs a rapid, multi-dimensional, and interpretable pre-evaluation of candidate knowledge graph variants generated by the mutation operator without relying on high-cost real-world trial and error. This pre-evaluation aims to predict the overall performance of these variants in the target execution environment (simulation or reality). The input to the evaluator is the candidate knowledge graph set {G′1, G′2, …, G′}. kThe database consists of a historical execution database D_hist (containing task execution trajectory, performance metrics, and environment status) and task objectives and constraint configurations C_task (such as objective success rate, time limit, safety margin, and energy budget). The output is each candidate graph G′. i The multidimensional fitness vector F i ={f1, f2, …, f m (e.g., success rate prediction, time cost prediction, energy consumption prediction, safety score, semantic consistency score, feasibility indicator), its functional goal is to eliminate low-potential or infeasible variants in the early stages of evolution, guide the selection mechanism to focus on high-value knowledge variants, thereby significantly reducing the computational and physical costs of subsequent simulation or real testing, while ensuring that the direction of knowledge evolution conforms to the mission objectives and process constraints.

[0051] The fitness evaluator of this application consists of four parts: a performance prediction submodule, a semantic consistency evaluation submodule, a feasibility verification submodule, and a comprehensive scoring and decision-making submodule, forming a multi-level, multi-index comprehensive evaluation pipeline.

[0052] The performance prediction submodule employs a hybrid architecture combining a graph neural network regression model (GNN-Regressor) and a Bayesian performance predictor: First, the candidate knowledge graph G′ is... i Structure-semantic joint embedding vector E i The input X is concatenated with the environment state vector S_env (provided by the environment observation module, including object pose, material, interference factors, target tolerance, etc.) to form the model input X. i =[E i ‖S_env]; Secondly, X i The input is a pre-trained GNN-Regressor (such as a regression network based on GraphSAGE or GAT, with the loss function being the mean squared error MSE). This model is trained on the historical execution database D_hist and can directly output G′. i The multi-objective performance prediction values ​​under the target environment include the expected task success rate P_succ, expected execution time T_exec, expected energy consumption E_cons, and expected safety score S_safe (0~1, higher for greater safety). Simultaneously, to improve the prediction robustness in small sample or new task scenarios, a Bayesian performance predictor is introduced: [The text then abruptly shifts to a different topic:] ...using X... i Using historical performance data as the training set as input, a Gaussian process (GP) regression model is constructed, and the predicted mean μ and variance σ² of each performance index are output to measure the prediction uncertainty. Confidence weights are also introduced into the comprehensive score.

[0053] The semantic consistency evaluation submodule assesses the rationality of knowledge variants in terms of process semantics by calculating the semantic distance and ontology relation preservation of the graphs before and after mutation: G′ is calculated using graph embedding cosine similarity. i The average similarity Sim_sem between the node / edge embedding vectors of the original graph G_t and the graph itself; G′ is calculated using an ontology relation distance metric (such as a score function based on TransE / ComplEx). i The relation R_j in the domain ontology and the standard relation R in the domain ontology. 0 The difference of _j is Dist_onto; after normalizing Sim_sem and Dist_onto, the weighted sum is obtained to obtain the semantic consistency score SC∈[0,1]. If SC is lower than the preset threshold (e.g. 0.8), it is judged as semantic drift and penalized in the comprehensive score.

[0054] The feasibility verification submodule checks the physical realizability of candidate graphs through both lightweight physical simulation and symbolic logic verification: in the symbolic logic verification layer, G′ i Submitted to the Z3 / SMT solver, it executes a predefined first-order logic rule base (such as " Action node a, Preconditions include nodes p, p→a, "gripping force ≥ workpiece weight", and "joint angle within mechanical limits". If a logical conflict or out-of-bounds error is detected, the feasibility flag Feas is set to False. In the lightweight physics simulation layer, a simplified physics engine (such as PyBullet's rigid body kinematics mode) is called to apply G′. i The corresponding task sequence is subjected to fast kinematic / dynamic simulation to detect whether physical infeasible states such as self-collision, singular configuration, force / torque exceeding limits, or motion inaccessibility occur. If detected, Feas is also set to False.

[0055] The comprehensive scoring and decision-making submodule normalizes, weights, fuses, and filters the outputs of each submodule: it performs Min-Max normalization on the performance prediction values ​​{P_succ, T_exec, E_cons, S_safe} to the [0,1] interval, and sets weights {w1, w2, w3, w4} according to the priority in the task objective C_task (e.g., w1=0.4, w2=0.2, w3=0.1, w4=0.3), and calculates the comprehensive performance score Perf_score=∑w i ·norm(P i The semantic consistency score (SC) and the feasibility indicator (Feas = 1 for True and 0 for False) are incorporated into the overall evaluation formula: F i=α·Perf_score + β·SC + γ·Feas, where α, β, and γ are adjustable coefficients (e.g., α=0.6, β=0.2, γ=0.2); for F i Candidate maps below a preset evolutionary threshold are directly discarded, only F... i The top-N (e.g., N=5) candidates are sent to the selection mechanism for subsequent evolutionary decisions, and detailed predicted values ​​and confidence levels of each indicator are output for manual review and knowledge feedback.

[0056] Through the aforementioned multi-level, multi-source data fusion evaluation mechanism, the fitness evaluator achieves low-cost, high-reliability, and highly interpretable knowledge variant pre-screening, ensuring that the evolutionary direction of the double helix engine is both efficient and robust, and significantly reducing ineffective trial and error on simulations or real robots.

[0057] Through the above technical solutions, the fitness evaluator of this application can perform multi-dimensional and refined evaluation of candidate knowledge graphs. The performance prediction submodule adopts a hybrid architecture combining a graph neural network regression model and a Bayesian performance predictor. It can fully utilize the joint embedding vector of structural semantics and the environment state vector, not only predicting multi-objective performance, but also providing the mean and variance of the prediction, thereby quantifying the uncertainty of the prediction. This enables the system to evaluate potential performance more intelligently, avoiding the inefficiency of traditional random trial and error and improving exploration efficiency. The semantic consistency evaluation submodule ensures the semantic rationality of the mutation process by calculating the cosine similarity of the graph embeddings and the distance of the ontology relationship before and after mutation. This effectively prevents the generation of semantically incoherent or contradictory knowledge to the domain ontology, ensuring the logical self-consistency of the knowledge graph. The feasibility verification submodule performs symbolic logic verification through a solver and physical simulation verification by calling a lightweight physics engine, performing a double check on the candidate graph. This mechanism can strictly filter out physically unrealizable or logically illegal knowledge, greatly enhancing the security and reliability of the generated knowledge, making it safer to apply to actual robot systems. The comprehensive scoring and decision-making submodule normalizes, weights, fuses, and thresholds the outputs of the aforementioned submodules, providing a comprehensive and operable comprehensive fitness score. This enables the selection mechanism to make decisions based on a unified and reliable evaluation result, thereby achieving the effective inheritance of high-quality knowledge and the efficient exploration of new knowledge, forming a closed-loop knowledge evolution process. This solves the problem of the separation between knowledge and execution, and improves the reliability and security of knowledge evolution.

[0058] In this application, the selection mechanism, based on the multi-dimensional performance prediction and constraint satisfaction information output by the fitness evaluator, selects the knowledge variants with the highest potential for achieving high success rate, high efficiency, and high security in the target environment through a strategy combining intelligent search and elite retention among multiple candidate knowledge graph variants. This guides the next round of evolution of the double helix engine. Its input is the candidate knowledge graph set {G′1, G′2, …, G′}. k} and its corresponding fitness vector {F1, F2, …, F k The system includes a comprehensive performance score, semantic consistency score, feasibility indicator, prediction confidence, etc., historical evolutionary data D_evo (including parent graphs selected in each iteration, child performance, and environmental change records), and evolutionary strategy configuration C_evo (such as elite retention rate, exploration rate, and Bayesian optimization objective function). The output is the selected parent knowledge graph G_parent and a set of selected Top-N candidate graphs {G1, G2, …, G_N}. Its functional goal is to ensure that verified high-quality knowledge is not lost while intelligently guiding the exploration of high-potential but not fully verified mutation regions, achieving rapid convergence and continuous diversity maintenance in the knowledge evolution process, thereby significantly improving the evolutionary efficiency and final skill quality of the double helix engine.

[0059] The selection mechanism consists of four parts: an elite retention submodule, a Bayesian optimization exploration submodule, a cross-recombination submodule, and an evolutionary state management submodule, forming a collaborative decision-making pipeline of "preserving the best + exploring new + combining + managing".

[0060] The elite retention submodule employs a deterministic Top-K selection strategy combined with historical performance weighting to ensure that highly fit and stable knowledge variants directly enter the next generation: First, the candidate set is sorted according to the fitness vector F i The main scores (such as overall performance scores) are sorted in descending order, and the top K1 candidates (e.g., K1=3) are selected as the elite set E_set; then, for each candidate G′ in E_set... i Query historical evolutionary data D_evo and calculate its historical performance index (HPI). i HPI i =γ·(average success rate of all executions)+(1 γ)·(reciprocal of the stability variance of each execution), where γ is an adjustable weight (e.g., 0.7), HPI i To mitigate misjudgments caused by short-term fluctuations, higher priority is given to retaining top performers with stable long-term performance; finally, HPI in E_set... i Candidates exceeding a preset threshold are directly labeled as G_parent or G_i, ensuring that high-quality knowledge is not lost due to noise in a single round of evaluation.

[0061] The Bayesian optimization exploration submodule employs a Gaussian process (GP) + expected improvement (EI) strategy to intelligently select high-potential exploration points from candidates outside the elite set: each non-elite candidate G′ j Structure-semantic joint embedding vector E j The input X is concatenated with the environment state vector S_env. j , with F j The main score is used as the objective function value y j Construct or update the GP model; for candidate regions that have not yet been fully evaluated, calculate the expected improvement EI(X) = ∫_{y}^{∞} (y y)·p(y|X) dy*, where y is the current best fitness value and p(y|X) is the probability density function predicted by GP; select the top K2 (e.g., K2=2) candidates with the largest EI values ​​as the exploration set Expl_set. Although these candidates are not currently optimal, they have significant potential for improvement in prediction and are used to maintain evolutionary diversity and break through local optima.

[0062] The cross-recombination submodule performs graph structure cross-combination and parameter mixing between the elite set and the exploration set to generate offspring candidates that combine the advantages of the parent generation with the novelty of exploration: For candidates in G_parent and Expl_set, subgraph isomorphic matching + V F2 algorithm* is used to locate the common substructures that can be crossed; within the matching region, single-point cross-combination is performed on the structural part (such as swapping the "detection" node parameters of G_parent with the "detection" node structure of Expl_set), and arithmetic cross-combination or BLX-α cross-combination is performed on the parameter part (such as taking the force threshold as the mean of the two parents ± α·difference range) to generate offspring candidate G_child*; G_child is quickly verified for feasibility by calling the symbolic inference engine (Z3) and lightweight physics simulation, and only offspring that pass the verification are added to the Top-N selection set to ensure that exploration does not violate the legitimacy of knowledge.

[0063] The evolutionary state management submodule is responsible for recording, updating, and dynamically adjusting the evolutionary state to ensure the adaptability and backtrackability of the selection mechanism: It establishes an evolutionary state table EST, recording the selected parent G_parent, the top candidate G_i, and its fitness F in each round. i HPI iInformation such as EI value and environment state S_env is collected; based on the convergence index in EST (such as the improvement of optimal fitness in N consecutive rounds < ε) and diversity index (such as the average cosine similarity of the graph embeddings of the Top-N candidates > θ), the parameters in C_evo are dynamically adjusted: if convergence is achieved, the exploration rate is reduced and the elite retention ratio is increased; if diversity is insufficient, the exploration rate is increased and more cross-graph migration candidates are introduced; at the end of each round, EST is updated and fed back to the Bayesian optimization model and mutation operator, so that the evolution rhythm and direction of the entire double helix engine have meta-learning characteristics, that is, "learning how to evolve".

[0064] Finally, the selection mechanism outputs G_parent (as the baseline graph for the next round of mutation) and {G1, …, G_N}* (as the set of tasks to be executed in the performance spiral), and updates the evolutionary state simultaneously, providing core decision support for the continuous, efficient, and adaptive knowledge evolution of the double helix engine.

[0065] Through the above technical solutions, the selection mechanism proposed in this application can effectively solve the problems of insufficient diversity, low exploration efficiency, poor offspring feasibility, and lack of dynamic adjustment during evolution. The elite preservation submodule, by combining a historical performance weighting strategy, not only ensures the inheritance of high-fitness, high-quality knowledge but also avoids the reduction in population diversity caused by solely relying on current fitness, thus maintaining population richness. The Bayesian optimization exploration submodule utilizes Gaussian processes and expectation improvement strategies to intelligently identify and explore regions with high potential, significantly improving exploration efficiency and avoiding the inefficiency of random search. The crossover and recombination submodule, through subgraph isomorphic matching based on the VF2 algorithm, ensures structural compatibility and performs crossover operations within the matching region, generating more feasible offspring candidates. Through rigorous feasibility checks, it effectively avoids the generation of invalid or infeasible offspring, improving evolutionary efficiency. The evolutionary state management submodule records and analyzes evolutionary data in real time, and dynamically adjusts evolutionary strategy parameters based on convergence and diversity indices. This enables the entire evolutionary process to adapt to the needs of different stages, ensuring rapid convergence and continuous diversity of knowledge evolution. Thus, in the automatic evolution of the process knowledge graph of multi-arm mobile robots, a dynamic balance is achieved between the effective inheritance of high-quality knowledge and the efficient exploration of new knowledge, further improving the closed-loop knowledge evolution process.

[0066] Based on the above system, this application proposes a knowledge mutation algorithm for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, including the following steps: First, the semantic encoder receives the process knowledge graph at the current moment and uses graph neural networks and symbolic and / or language embedding techniques to encode its topological structure and semantic information on nodes and / or edges into a unified semantic joint embedding vector. Next, the mutation operator uses semantic joint embedding vectors to perform controlled and interpretable mutations on the knowledge graph in both structural and parametric dimensions. In the structural dimension, it performs operations such as adding, deleting, merging, splitting, and reorganizing subgraphs of nodes and / or edges. In the parametric dimension, it applies bounded random perturbations to continuous parameters and guides the adjustment through Bayesian optimization. At the same time, the mutation operator calls the symbolic inference engine and lightweight physical simulation to perform multi-level constraint verification to ensure that the generated candidate knowledge graph is logically valid, physically feasible, and semantically consistent. Subsequently, the fitness evaluator performs a rapid, multi-dimensional pre-evaluation of the candidate knowledge graph set generated by the mutation operator. It uses a graph neural network regression model and a Gaussian process to predict the multi-objective performance of each candidate knowledge graph in the target execution environment. It also calculates a comprehensive fitness score by combining semantic consistency score and feasibility flag. Finally, the selection mechanism dynamically guides the direction of knowledge evolution based on the comprehensive fitness score output by the fitness evaluator. It ensures that high-fitness, high-quality knowledge is inherited to the next generation through an elite retention strategy; intelligently discovers and explores new regions with high potential but not yet fully validated through Bayesian optimization of the exploration strategy; generates offspring candidates that combine the advantages of the parent generation with the novelty of exploration through graph structure cross-recombination; and dynamically adjusts the evolution strategy through evolutionary state management to achieve rapid convergence and continuous maintenance of diversity in the knowledge evolution process.

[0067] The core functions of the above algorithm include: 1. Knowledge Representation and Semanticization: Transforming symbolic technological knowledge into computable and comparable graph embedding representations, providing a semantic awareness foundation for intelligent mutation; 2. Proactive knowledge innovation: Automatically generate new process skills, action combinations, and constraint configurations in the structural and parameter dimensions to achieve self-expansion and optimization of the knowledge base; 3. Multi-objective performance pre-evaluation: Predict the success rate, efficiency, energy consumption, and safety of candidate skills before simulation or real trial and error, so as to achieve efficient screening; 4. Safety and Feasibility Assurance: Through dual verification using symbolic logic and physical simulation, the generated skills are ensured to comply with process rules, safety specifications, and physical laws; 5. Intelligent guidance for evolutionary direction: A dynamic balance is achieved between the preservation of elites and exploration, realizing rapid convergence and continuous diversity in the knowledge evolution process, and avoiding local optima and knowledge degradation.

[0068] The algorithm described above aims to achieve the following objectives: 1. Breaking through the limitations of traditional manual teaching and static process libraries, enabling robots to have autonomous evolution capabilities in process knowledge, and automatically generating and optimizing skills based on task feedback and environmental changes; 2. Significantly improves the sample efficiency and exploration efficiency of skill generation, reducing reliance on high-cost real-world trial and error; 3. Ensure the interpretability, security, and physical feasibility of generated skills, and avoid ineffective or dangerous strategies resulting from random search; 4. Construct a knowledge evolution closed loop that is semantically driven, structurally perceptive, and constraint-guaranteed, so that the robot system has the ability to generalize and continuously adapt across tasks and scenarios. 5. To provide a core technology for intelligent robot evolution that can be engineered and continuously optimized for fields such as intelligent manufacturing, flexible assembly, and special operations.

[0069] The following example will provide a more detailed explanation of the above technical solution: For example, in a scenario where a multi-arm mobile robot performs complex electronic component assembly tasks, the robot needs to precisely mount multiple components of different shapes and sizes onto a circuit board. The initial process knowledge graph may only contain assembly steps and parameters for standard components and fixed environments. However, when encountering new component models, minor adjustments to the production line layout, or unexpected obstacles, the robot's original static knowledge base cannot adapt, leading to assembly failure or inefficiency. Traditional solutions often require manual reprogramming or extensive random trial and error, which is inefficient and poses safety hazards.

[0070] This system first receives the current-time process knowledge graph through a semantic encoder. For example, this graph might contain nodes such as "pick up resistor," "place resistor at position A," and "weld resistor," as well as edges such as "pick up pre-placement" and "place using tool B." Each node v_i contains attribute vectors A_i, such as component size and gripping torque, and each edge e_j contains relation types R_j, such as "order" and "use." The semantic encoder uses a graph convolutional network as a structural feature extractor, taking the adjacency matrix and node feature matrix as input. Through multi-layer graph convolution operations, it aggregates neighbor structural information layer by layer to obtain a structural embedding vector. Simultaneously, it performs domain word embedding on nodes v_i and relation types R_j, mapping each symbolic text to a first semantic vector, and combines this with ontology embedding to vectorize relation type R_j. In the fusion layer, the semantic encoder introduces a multi-head attention mechanism to calculate the attention weights between the structural embedding vector, the second semantic vector, and the relation vector, generating a weighted sum. After linear transformation and normalization, the final structural-semantic joint embedding vector is obtained. This vector uniformly represents the topological structure of the assembly process and the semantic information it contains, providing a semantically aware graph representation for subsequent mutation operations and solving the problem that traditional static knowledge bases cannot evolve on their own.

[0071] Next, the mutation operator mutates the knowledge graph based on these structural semantic joint embedding vectors. For example, when a robot needs to assemble a new type of irregular component, the structural mutation module in the mutation operator may perform operations such as adding, deleting, merging, splitting, and recombining subgraphs along the structural dimension. It might add a new node for "adjusting component orientation" or merge the "pickup" and "pre-positioning" steps. After each structural operation, the system immediately invokes symbolic reasoning verification, for example, verifying that the "placement" operation must follow the "pickup" operation to ensure that the newly generated graph is logically valid. Simultaneously, the parameter mutation module performs bounded random perturbations on node attributes, for example, injecting Gaussian noise or sampling uniform intervals into the grasping torque parameter of the "pickup resistor" node to adapt to the physical characteristics of the new component. For edge attributes, it adjusts the associated parameters based on historical performance data combined with current fitness predictions, for example, adjusting the welding time of the "weld resistor" operation. After all parameters have been mutated, symbolic inference verification and lightweight physical simulation are invoked again to verify whether the parameters meet the predefined constraints. For example, it checks whether a new gripping torque will cause component damage or whether a new soldering time will cause overheating. This controlled mutation method, which undergoes multi-level constraint verification, avoids the inefficiency and uninterpretability of traditional random search and ensures the security of skill generation from the source.

[0072] Subsequently, the fitness evaluator performs a rapid, multi-dimensional pre-evaluation of the candidate knowledge graph set generated by the mutation operator. For example, for multiple candidate graphs with different assembly orders or parameters, the performance prediction submodule adopts a hybrid architecture combining a graph neural network regression model and a Bayesian performance predictor. It concatenates the candidate's structural semantic joint embedding vector with the current environmental state vector (such as obstacle positions and lighting conditions) as model input, quickly predicting the multi-objective performance of each candidate graph in the target execution environment, such as assembly success rate, time efficiency, and energy consumption. The semantic consistency evaluation submodule calculates the cosine similarity of the graph embeddings and the ontology relation distance of the graphs before and after mutation, normalizes the results, and then performs a weighted sum to obtain a semantic consistency score, ensuring that the mutated process flow remains consistent with the original task objective. The feasibility verification submodule performs symbolic logic verification through a solver and calls a lightweight physics engine to perform physical simulation verification, doubly checking the physical realizability of the candidate graphs, for example, verifying whether the robot arm's motion trajectory will cause collisions. Finally, the comprehensive scoring and decision submodule normalizes, weights, fuses, and thresholds these outputs to calculate a comprehensive fitness score. This multi-objective, pre-evaluation mechanism significantly improves exploration efficiency and further strengthens the safety guarantee for skill generation, solving the problem of the separation between knowledge and execution in traditional methods.

[0073] Finally, the selection mechanism dynamically guides the direction of knowledge evolution based on the comprehensive fitness score output by the fitness evaluator. The elite retention submodule employs a deterministic Top-K selection combined with a historical performance weighting strategy, selecting the elite set in descending order of the fitness vector's main score to ensure that high-fitness, high-quality knowledge (e.g., an efficient and safe assembly process) is inherited by the next generation. The Bayesian optimization exploration submodule uses a Gaussian process combined with an expected improvement strategy to build models for non-elite candidates, calculate expected improvement values, and select the candidate with the largest expected improvement value as the exploration set. This intelligently discovers and explores new regions with high potential but not yet fully validated, avoiding getting trapped in local optima. The cross-recombination submodule performs subgraph isomorphic matching based on the VF2 algorithm on the elite set and the exploration set. Within the matching region, it performs structural single-point cross-cross, parametric arithmetic cross-cross, or BLX-α cross-cross to generate offspring candidates that combine the advantages of the parent generation with novel exploration ideas. For example, combining an efficient subprocess from an elite graph with a newly discovered parameter setting from an exploration graph. After feasibility verification, legitimate offspring are added to the selected set. The evolutionary state management submodule establishes an evolutionary state table, records key data for each round of evolution, dynamically adjusts parameters in the evolutionary strategy configuration based on convergence and diversity indices, and feeds updates back to the Bayesian optimization model and mutation operator. Through this closed-loop knowledge evolution process, the system can continuously learn and optimize assembly process knowledge, achieving a balance between the inheritance of high-quality knowledge and the efficient exploration of new knowledge. This overcomes the problems of static knowledge bases being unable to evolve on their own and lacking cross-task generalization capabilities, enabling the robot to adapt to rapidly changing task requirements and environments.

[0074] In the scenario described above, where a multi-armed mobile robot performs a complex electronic component assembly task, this approach improves the task success rate and shortens the average execution time. Furthermore, all generated tasks are verified through symbolic logic and physical simulation, eliminating the possibility of safety incidents. Compared to purely random search or manual rule design, this technical solution significantly improves sample efficiency, generalization ability, security, and interpretability, validating its practicality and innovation.

[0075] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, used for the automatic evolution of the process knowledge graph of multi-armed mobile robots, characterized in that, include: A semantic encoder is configured to encode the topological structure and node and / or edge semantic information of a process knowledge graph into a unified structural semantic joint embedding vector by fusing graph neural networks with symbolic and / or language embeddings, thereby providing a semantically aware graph representation for variation. The mutation operator is configured to add, delete, merge, split, and reorganize nodes and / or edges in the structural dimension based on the joint embedding vector of the structural semantics, and perform bounded random perturbation and Bayesian optimization-guided adjustment in the parameter dimension. It also performs multi-level constraint verification through symbolic reasoning and lightweight physical simulation to generate a set of candidate knowledge graphs. The fitness evaluator is configured to predict multi-objective performance based on a graph neural network regression model and a Gaussian process, and to perform a multi-objective performance pre-evaluation on the candidate knowledge graph set generated by mutation, combined with semantic consistency scores and feasibility indicators, and to calculate the comprehensive fitness. The selection mechanism is configured to dynamically guide the direction of knowledge evolution based on the comprehensive fitness, employing strategies such as elite retention, Bayesian optimization exploration, graph structure cross-reorganization, and evolutionary state management. This achieves a balance between the inheritance of high-quality knowledge and the efficient exploration of new knowledge, forming a closed-loop knowledge evolution process.

2. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 1, characterized in that: The process knowledge graph is G_t=(V,E); where V is the set of nodes, E is the set of edges, each node v_i contains an attribute vector A_i, and each edge e_j contains a relation type R_j.

3. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 2, characterized in that: The semantic encoder employs a graph convolutional network or a graph attention network as a structural feature extractor, taking the adjacency matrix A_d and the node feature matrix X as input. Through multi-layer graph convolution operations H^(l+1)=σ(D^{-1 / 2}A_dD^{-1 / 2}H^(l)W^(l)), it aggregates neighbor structural information layer by layer to obtain the structural embedding vector S_i∈ ^d_s; where D is the degree matrix, W^(l) is the learnable weight matrix, and σ is the activation function.

4. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 3, characterized in that: The semantic encoder performs domain word embedding or pre-trained language model encoding on the node v_i and the relation type R_j, mapping each symbolized text t_k to a first semantic vector T_k∈ ^d_t, and combined with ontology embedding, the relation type R_j is vectorized as O_j∈ ^d_o.

5. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 4, characterized in that: The semantic encoder introduces a multi-head attention mechanism in the fusion layer, calculating the attention weights α{i,j}=softmax(f(S_i,T_i,O_j)) between the structural embedding vector S_i and the second semantic vector T_i and relation vector O_j, and generating a weighted sum F_i=∑jα_{i,j}·[S_i‖T_i‖O_j]; where ‖ denotes vector concatenation. After linear transformation and normalization, the final structural semantic joint embedding vector E_i∈ ^d_e.

6. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 1, characterized in that: The mutation operator includes two collaborative working units: a structural mutation module and a parameter mutation module. The structural mutation module adopts a dual-channel mechanism of graph structure editing operation and symbolic reasoning verification. The graph structure editing operation guides the node mutation operation, and the symbolic reasoning verification drives the edge mutation verification. Through the subgraph recombination operation of the VF2 subgraph isomorphism algorithm, the symbolic reasoning verification is called after each edge operation to verify the logical legality of the graph structure. The parameter mutation module employs a hybrid strategy of bounded random perturbation and Bayesian guided adjustment. It performs continuous parameter perturbation on node attributes by injecting Gaussian noise, sampling uniform intervals, or discretizing enumeration, and performs correlation parameter mutation on edge attributes guided by historical performance data combined with current fitness prediction. After all parameters are mutated, the symbolic inference verification is called again to verify whether the parameters meet the predefined constraints.

7. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 1, characterized in that: The fitness evaluator includes: a performance prediction submodule, a semantic consistency evaluation submodule, a feasibility verification submodule, and a comprehensive scoring and decision-making submodule; The performance prediction submodule adopts a hybrid architecture that combines a graph neural network regression model with a Bayesian performance predictor. It concatenates the candidate structural semantic joint embedding vector with the environment state vector as the model input and outputs multi-objective performance prediction values ​​as well as prediction mean and variance. The semantic consistency evaluation submodule calculates the graph embedding cosine similarity and ontology relation distance of the graphs before and after mutation, and then normalizes and weights the sum to obtain the semantic consistency score SC∈[0,1]. The feasibility verification submodule performs symbolic logic verification through the solver and physical simulation verification by calling the lightweight physics engine, thus double-checking the physical realizability of the candidate graph. The comprehensive scoring and decision-making submodule normalizes, weights, fuses, and filters the outputs of the performance prediction submodule, semantic consistency evaluation submodule, and feasibility verification submodule.

8. The knowledge mutation system for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion according to claim 1, characterized in that: The selection mechanism includes: an elite retention submodule, a Bayesian optimization exploration submodule, a crossover and recombination submodule, and an evolutionary state management submodule; The elite retention submodule employs a deterministic Top-K screening combined with a historical performance weighting strategy, selecting an elite set E_set in descending order of fitness vector principal scores, and calculating the historical performance index (HPI) of candidates in the elite set E_set. i =γ·Average success rate of all executions + (1 γ)·The inverse of the stability variance of each execution, marking candidates whose HPii are higher than the preset threshold as G_parent or G_i; where γ is an adjustable weight; The Bayesian optimization exploration submodule uses a Gaussian process combined with an expected improvement strategy to construct a model for non-elite candidates and calculate the expected improvement value EI(X)=∫_{y}^{∞}(y y)·p(y|X)dy, select the top K2 candidates with the largest expected improvement value as the exploration set Expl_set; where y is the current best fitness value, and p(y|X) is the probability density function predicted by the Gaussian process; The cross-recombination submodule performs subgraph isomorphic matching based on the VF2 algorithm on the elite set and the exploration set. Within the matching region, it performs structural single-point cross-cross, parametric arithmetic cross-cross, or BLX-α cross-cross to generate candidate offspring G_child. After feasibility verification, the legal offspring are added to the selected set. The evolutionary state management submodule establishes an evolutionary state table (EST) to record key data for each round of evolution. It dynamically adjusts the parameters in the evolutionary strategy configuration based on the convergence and diversity indices in the evolutionary state table (EST) and updates the evolutionary state table (EST) to feed back to the Bayesian optimization model and mutation operator.

9. A knowledge mutation algorithm for embodied intelligent mobile robots based on semantic guidance and multi-strategy fusion, characterized in that: Includes the following steps: First, the semantic encoder receives the process knowledge graph at the current moment and uses graph neural networks and symbolic and / or language embedding techniques to encode its topological structure and semantic information on nodes and / or edges into a unified semantic joint embedding vector. Next, the mutation operator uses semantic joint embedding vectors to perform controlled and interpretable mutations on the knowledge graph in both structural and parametric dimensions. In the structural dimension, it performs operations such as adding, deleting, merging, splitting, and reorganizing subgraphs of nodes and / or edges. In the parametric dimension, it applies bounded random perturbations to continuous parameters and guides the adjustment through Bayesian optimization. At the same time, the mutation operator calls the symbolic inference engine and lightweight physical simulation to perform multi-level constraint verification to ensure that the generated candidate knowledge graph is logically valid, physically feasible, and semantically consistent. Subsequently, the fitness evaluator performs a rapid, multi-dimensional pre-evaluation of the candidate knowledge graph set generated by the mutation operator. It uses a graph neural network regression model and a Gaussian process to predict the multi-objective performance of each candidate knowledge graph in the target execution environment. It also calculates a comprehensive fitness score by combining semantic consistency score and feasibility flag. Finally, the selection mechanism dynamically guides the direction of knowledge evolution based on the comprehensive fitness score output by the fitness evaluator. It ensures that high-fitness, high-quality knowledge is inherited to the next generation through an elite retention strategy; intelligently discovers and explores new regions with high potential but not yet fully validated through Bayesian optimization of the exploration strategy; generates offspring candidates that combine the advantages of the parent generation with the novelty of exploration through graph structure cross-recombination; and dynamically adjusts the evolution strategy through evolutionary state management to achieve rapid convergence and continuous maintenance of diversity in the knowledge evolution process.

Images