A robot process skill generation system and method based on double helix coupling architecture

Abstract

The application provides a robot process skill generation system and method based on a double helix coupling architecture, applied to the technical field of intelligent manufacturing, robot autonomous decision and adaptive control, by designing "knowledge generation" and "performance evaluation" as two spiral chains that can evolve independently but are tightly coupled, and realizing real-time bidirectional feedback through a bidirectional coupling and evolution control center, forming a real closed-loop evolution system. The knowledge spiral layer is responsible for storing and dynamically mutating the process knowledge graph to generate candidate skills, and the performance spiral layer executes and quantitatively evaluates in a simulation or real environment, and the evaluation results directly drive the knowledge spiral layer to update the graph in a targeted manner, fundamentally breaking the barrier between knowledge and execution, forming a closed-loop co-evolution mechanism of "knowledge generation-performance verification-feedback update", so that the robot process skill generation system has the ability of continuous self-evolution, high sample efficiency, explainability and safety controllability.

Description

Technical Field

[0001] This application relates to the fields of intelligent manufacturing, robot autonomous decision-making and adaptive control technology, and specifically to a robot process skill generation system and method based on a double-helix coupling architecture. Background Technology

[0002] With the widespread application of industrial robots in unstructured, multi-variety, and small-batch production environments, higher demands are placed on the robots' ability to autonomously generate process skills. Traditional robot skill generation methods mainly rely on manual teaching, offline programming, or single reinforcement learning trial and error. However, these methods generally suffer from the following technical shortcomings in practical applications: First, there is a disconnect between knowledge and execution: In traditional methods, the construction of process knowledge (such as action sequences and constraints) and the evaluation of execution performance (such as success rate, efficiency, and safety) are usually completed in different stages. This leads to the common phenomenon of "knowing how to do it but not doing it well" or "executing it well but not knowing why it is good." The knowledge base cannot be automatically optimized based on the actual execution results. Second, the skill generation efficiency is low: existing methods rely on a lot of trial and error or human experience, the exploration space is huge, the sample cost is high, and it is difficult to guarantee the interpretability of the generated skills. Third, there is a lack of continuous evolution capability: most existing systems are open-loop or single-loop structures, the knowledge base is updated late, and the system cannot achieve self-improvement during task execution. Fourth, the knowledge representation structure is rigid: traditional knowledge representation is mostly static rules or fixed strategies, which are difficult to adjust dynamically once constructed. When the task or environment changes, the system cannot flexibly reorganize or expand while maintaining semantic and physical constraints, which seriously limits its applicability in flexible production.

[0003] Therefore, a new robotic process skill generation scheme is needed. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a robot process skill generation system and method based on a double-helix coupling architecture. By tightly coupling "knowledge generation" and "performance evaluation" in a closed loop, continuously feeding back and co-evolving, the system possesses autonomous innovation capabilities, high sample efficiency, interpretability, and security controllability. It also enables automatic identification of task requirements in complex dynamic environments, online evolution of knowledge graphs, closed-loop evaluation of execution performance, and strategy optimization, which can effectively improve the adaptability of embodied intelligent robots to scenarios and their generalization ability to perform tasks.

[0005] The embodiments in this specification provide the following technical solutions: This specification provides a robot process skill generation system based on a double-helix coupling architecture, including: a bidirectional coupling and evolution control layer module, a knowledge helix layer module, and a performance helix layer module; The bidirectional coupling and evolution control layer module is bidirectionally connected to the knowledge spiral layer module and the performance spiral layer module, respectively. The bidirectional coupling and evolution control layer module is used to determine the control strategy and send it to the knowledge spiral layer module based on the externally input task instructions and / or environmental state, as well as the performance evaluation results returned by the performance spiral layer module. Receive candidate skills generated by the knowledge spiral layer module, compile the candidate skills into executable instructions and send them to the performance spiral layer module; The system receives the performance evaluation results returned by the performance spiral layer module, generates a knowledge update instruction based on the performance evaluation results, and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform targeted mutation and update on the process knowledge graph. The knowledge spiral layer module is used to store and maintain the process knowledge graph, wherein the process knowledge graph organizes structured knowledge in the process domain in the form of a graph structure, such as nodes and edges; wherein nodes include tasks, objects, actions and constraints, and edges represent the associations and weights between nodes; Based on the received control strategy, the process knowledge graph is mutated to generate candidate skills; Update the process knowledge graph according to the received knowledge update instruction; The performance spiral layer module is used to receive and execute the executable instructions to run the candidate skills in a simulation environment and / or a real execution environment; Execution trajectory data is collected, and the execution trajectory data is quantitatively evaluated based on preset evaluation indicators to obtain performance evaluation results; The performance evaluation results are returned to the bidirectional coupling and evolution control layer module.

[0006] This specification also provides an embodiment of a robot process skill generation method based on a double-helix coupling architecture, the robot process skill generation method based on the double-helix coupling architecture includes: S1. The bidirectional coupling and evolution control layer module determines the control strategy and sends it to the knowledge spiral layer module based on the externally input task instructions and / or environmental status, as well as the performance evaluation results returned by the performance spiral layer module. S2. The knowledge spiral layer module performs mutation on the process knowledge graph according to the received control strategy to generate candidate skills; wherein, the process knowledge graph organizes the structured knowledge of the process domain in the form of a graph structure, the nodes include tasks, objects, actions and constraints, and the edges represent the association and weight between nodes; S3. The bidirectional coupling and evolution control layer module receives the candidate skills generated by the knowledge spiral layer module, compiles the candidate skills into executable instructions, and sends them to the performance spiral layer module. S4. The performance spiral layer module receives and executes the executable instructions, runs the candidate skills in a simulation environment and / or a real execution environment; collects execution trajectory data, and performs quantitative evaluation on the execution trajectory data based on preset evaluation indicators to obtain performance evaluation results; S5. Return the performance evaluation results to the bidirectional coupling and evolution control layer module; S6. The bidirectional coupling and evolution control layer module generates a knowledge update instruction based on the received performance evaluation result and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform directional mutation and update on the process knowledge graph. S7. Repeat steps S1 to S6 to form a closed-loop evolutionary process of knowledge generation, performance verification, and feedback update.

[0007] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: This application designs "knowledge generation" and "performance evaluation" as two independently evolving but tightly coupled spiral chains, achieving real-time bidirectional feedback through a central hub to form a true closed-loop evolutionary system. On one hand, the knowledge spiral layer generates candidate skills based on the dynamic variation of the process knowledge graph according to the control strategy, while the performance spiral layer executes and quantifies the evaluation in a simulation or real environment. The evaluation results directly drive the knowledge spiral layer to update the graph in a targeted manner, breaking down the barrier between knowledge and execution in traditional methods and enabling the system to have continuous self-evolution capabilities. On the other hand, the knowledge graph organizes nodes and relationships such as tasks, objects, actions, and constraints in a graph structure, supporting structural and parametric variations. This avoids the rigidity of static rules or fixed strategies, significantly improving the system's adaptability, skill generation efficiency, and interpretability in dynamic tasks and flexible production environments, while reducing the high sample costs caused by blind trial and error. Attached Figure Description

[0008] 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.

[0009] Figure 1 This is a schematic diagram of the overall structure of the double-helix coupling architecture in this application; Figure 2 This is a flowchart of the interaction between modules in the double-helix coupling architecture of this application. Detailed Implementation

[0010] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0011] 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.

[0012] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. 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.

[0013] 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 shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0014] 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.

[0015] Through in-depth research and improvement exploration of robotic process skill generation systems, the applicant discovered that existing methods for generating robotic process skills mainly rely on manual instruction, offline programming, or single reinforcement learning trial and error. These methods generally suffer from a serious deficiency in practical applications: a disconnect between knowledge and execution. Specifically, the construction of process knowledge (such as action sequences and constraints) and the evaluation of execution performance (such as success rate, efficiency, and safety) are usually completed at different stages or by different subsystems, lacking a real-time, two-way feedback path. This results in the knowledge base being unable to automatically optimize based on actual execution results, often leading to phenomena such as "knowing how to do it but not doing it well" or "performing well but not knowing why," severely limiting the robot's autonomous operation capabilities in unstructured, multi-variety, small-batch production environments.

[0016] Furthermore, the applicant found that traditional methods also suffer from low efficiency in skill generation. Due to their reliance on extensive trial and error or human experience, the systems have a huge exploration space, high sample costs, and poor interpretability and security of generated skills. Simultaneously, existing systems are mostly open-loop or single-loop structures, resulting in severely lagging knowledge base updates. They cannot achieve self-improvement during task execution and are difficult to transfer skills learned in one task to other related tasks. Moreover, traditional knowledge representations are mostly static rules or fixed strategies, which are difficult to dynamically adjust once constructed. When tasks or environments change, the system cannot flexibly reorganize or expand knowledge while maintaining semantic and physical constraints, severely limiting its applicability in flexible production.

[0017] Based on this, this specification proposes a robotic process skill generation system based on a double-helix coupling architecture. The overall concept is as follows: "Knowledge generation" and "performance evaluation" are designed as two independently evolving but tightly coupled helical chains. Real-time bidirectional feedback is achieved through bidirectional coupling and an evolutionary control center, forming a true closed-loop evolutionary system. The knowledge helix layer is responsible for storing and dynamically mutating the process knowledge graph to generate candidate skills. The performance helix layer executes and quantifies the evaluation in a simulation or real environment. The evaluation results directly drive the knowledge helix layer to update the graph in a targeted manner. This fundamentally breaks down the barrier between knowledge and execution, forming a closed-loop co-evolutionary mechanism of "knowledge generation - performance verification - feedback update". This enables the robotic process skill generation system to have continuous self-evolution capability, high sample efficiency, interpretability, and safety and controllability.

[0018] The technical solutions provided by the various embodiments of this application are described below with reference to the accompanying drawings.

[0019] like Figure 1As shown in the embodiments of this specification, a robot process skill generation system based on a double-helix coupling architecture is provided, including: Bidirectional coupling and evolution control layer module, knowledge spiral layer module, and performance spiral layer module; The bidirectional coupling and evolution control layer module is bidirectionally connected to the knowledge spiral layer module and the performance spiral layer module, respectively. The bidirectional coupling and evolution control layer module is used to determine the control strategy and send it to the knowledge spiral layer module based on the externally input task instructions and / or environmental state, as well as the performance evaluation results returned by the performance spiral layer module. Receive candidate skills generated by the knowledge spiral layer module, compile the candidate skills into executable instructions and send them to the performance spiral layer module; The system receives the performance evaluation results returned by the performance spiral layer module, generates a knowledge update instruction based on the performance evaluation results, and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform targeted mutation and update on the process knowledge graph. The knowledge spiral layer module is used to store and maintain the process knowledge graph, wherein the process knowledge graph organizes the structured knowledge of the process domain in the form of a graph structure, the nodes include tasks, objects, actions and constraints, and the edges represent the associations and weights between nodes; Based on the received control strategy, the process knowledge graph is mutated to generate candidate skills; Update the process knowledge graph according to the received knowledge update instruction; The performance spiral layer module is used to receive and execute the executable instructions to run the candidate skills in a simulation environment and / or a real execution environment; Execution trajectory data is collected, and the execution trajectory data is quantitatively evaluated based on preset evaluation indicators to obtain performance evaluation results; The performance evaluation results are returned to the bidirectional coupling and evolution control layer module.

[0020] During implementation, such as Figure 1 As shown, the bidirectional coupling and evolution control layer module (i.e., the bidirectional coupling and evolution control center) serves as the control hub of the entire system. Based on the current input (such as external task instructions and / or environmental state) and the performance evaluation results of the previous round, it dynamically determines the control strategy, which specifically includes: updating the mutation strategy; compiling the candidate skills output by the knowledge spiral layer module to generate executable instructions for the robot and sending them to the performance spiral layer module; and driving the knowledge spiral layer module to update based on the evaluation results returned by the performance spiral layer module. After receiving the mutation strategy issued by the bidirectional coupling and evolution control center, the knowledge spiral layer module performs mutation based on the current knowledge graph (including tasks, objects, actions, and constraint nodes), generates candidate skills (such as action sequences, parameter configurations, and subgraph recombination schemes), and sends the candidate skills back to the bidirectional coupling and evolution control center. The bidirectional coupling and evolutionary control center distributes the compiled executable instructions to the performance spiral layer module. This module executes the instructions in a simulation or real-world environment, simultaneously collecting data such as trajectory, status, and results. It then quantitatively evaluates the execution performance through multi-dimensional constraint verification (e.g., semantic / process, logical consistency, physical feasibility, safety), with evaluation metrics covering success rate, efficiency, energy consumption, and risk, resulting in a performance assessment. During execution, new task requirements or environmental observations may arise, such as workpiece deformation, sensor anomalies, or environmental interference. These changes can be extracted and fed back to the control center to trigger the next round of decision-making, while also being sent back to the external input layer, forming a cross-cycle macro-environmental closed loop. Through this mechanism, the system achieves a continuous evolutionary closed loop of "knowledge generation - execution verification - feedback evolution - knowledge regeneration," enabling the system to continuously adjust and adapt to new scenarios based on actual execution results.

[0021] This application designs "process knowledge" and "execution performance" as two independently evolving but tightly coupled "spirals," achieving a two-way feedback loop through a control center. This allows performance evaluation to directly drive knowledge mutation, thus realizing "learning by doing and doing by learning." Specifically: the knowledge spiral is responsible for storing and mutating the process knowledge graph (including node relationships such as tasks, objects, actions, and constraints) and generating candidate skills; the performance spiral executes candidate skills in simulation / real environments, collects trajectories and results, and performs multi-dimensional evaluations (success rate, efficiency, safety, etc.); the control center, as the "brain," dynamically determines mutation strategies, compiles and distributes tasks, drives knowledge updates, and combines external environment / task input and new task / environment observations to form a continuously evolving double-helix ecosystem.

[0022] In some embodiments, the bidirectional coupling and evolution control layer module includes a strategy decision module and a knowledge mutation scheduling module; Specifically, the strategy decision module (selecting mutation / compilation / update strategies) is responsible for dynamically selecting the control strategy that the knowledge spiral should adopt in the current evolution cycle based on the task instructions of the input layer, environmental observations, and the evaluation results of the performance spiral feedback. This enables adaptive and self-evolving strategy decisions in complex dynamic environments, ensuring that the knowledge spiral always evolves in the direction of improving the overall performance of the system. The knowledge mutation scheduling module (which determines mutation type / parameters) serves as the execution layer of the strategy decision module. It is responsible for refining abstract strategy instructions into specific knowledge mutation operation instructions, including: determining the mutation type (such as parameter adjustment, node / edge addition / deletion, combination, splitting, and adding a new subgraph), the mutation magnitude (such as the range of parameter perturbation, the proportion of node deletion), the variable length and iteration count, and coordinating the collaborative work of various modules within the knowledge spiral. Based on this, it simulates the "mutation" in biological evolution. choose The "retention" mechanism introduces diversity through controllable randomness while filtering out invalid or dangerous mutations using constraints. Specifically, it can employ crossover, mutation, and selection operators from genetic algorithms (such as single-point crossover and Gaussian perturbation mutation), Bayesian optimization for efficient searching of the optimal mutation parameter space, domain-driven parameterized mutation (such as random sampling of process parameters according to tolerance bands), and subgraph generation and optimization based on graph neural networks (such as GNNs). Based on graph mutation), thereby stimulating the exploration ability of the knowledge space while ensuring the rationality of the knowledge graph structure and semantic consistency, and providing diverse candidate skill solutions for the performance spiral.

[0023] In some embodiments, the bidirectional coupling and evolution control layer module further includes: The constraint verification module is used to perform feasibility screening on the candidate skills before compiling them into executable instructions, filter out candidate skills that violate preset semantic constraints, process logic or security boundaries, and send the candidate skills that pass the verification to the performance spiral layer module.

[0024] In practice, the feasibility of candidate skills generated by the knowledge spiral layer module is screened before they are compiled into executable instructions.

[0025] Specifically, only candidate skills that pass the constraint verification module's checks (such as violations of semantics, process logic, or security boundaries) will be issued to the performance spiral layer module. This filters out obviously infeasible or dangerous solutions before actual operation, significantly reducing the number of costly real-world trials and errors, and improving system security and execution efficiency.

[0026] In some embodiments, the bidirectional coupling and evolution control layer module is further used for: The knowledge subgraphs corresponding to candidate skills whose scores exceed a preset threshold in the performance evaluation results are marked as elite individuals and directly retained in the next generation of process knowledge graphs, while mutation operations on the elite individuals are suppressed.

[0027] Specifically, the bidirectional coupling and evolutionary control layer module marks the knowledge subgraphs corresponding to candidate skills whose scores exceed a preset threshold in the performance evaluation results as elite individuals. These elite individuals are directly retained in the next generation's process knowledge graph, and mutation operations on them are suppressed. This ensures that the evolutionary process steadily converges towards performance improvement, avoiding the loss of high-quality solutions already obtained due to overexploration. This embodiment ensures the safety, efficiency, interpretability, and sustainability of the double-helix evolutionary process through multi-level constraint verification, fitness pre-evaluation, elite retention, and intelligent exploration mechanisms.

[0028] In some embodiments, the control strategy includes at least one of the following: mutation strategy, knowledge compilation method, and knowledge graph update strategy.

[0029] Specifically, mutation strategies include structural mutation, parameter mutation, or hybrid mutation. Knowledge compilation methods: for example, rules for generating state-action-reward triples S / A / R; Knowledge graph update strategies include incremental updates, local reconstruction, or global retraining. To maximize long-term performance gains (such as success rate, efficiency, and security), the system can construct a "state-action-reward" mapping based on reinforcement learning policy gradient methods or rule-based expert systems. Furthermore, it can rapidly adapt to new tasks using deep reinforcement learning (such as DQN, PPO, and SAC algorithms), meta-learning (MAML, Reptile), similar task retrieval and policy transfer based on knowledge graph embedding, and policy parameter tuning based on Bayesian optimization. This enables adaptive and self-evolving policy decisions in complex dynamic environments, ensuring that the knowledge spiral always evolves in the direction of improving the overall performance of the system.

[0030] In some embodiments, the knowledge spiral layer module includes: The knowledge retrieval module is used to retrieve knowledge fragments that match the current scenario from the process knowledge graph based on the current task instructions, environmental state, and control strategy, as mutation seeds; wherein, the knowledge fragments include historical execution schemes of similar tasks, adapted action sequences, or associated constraints; The knowledge mutation module is used to perform structural and / or parametric mutation operations on the mutation seed according to the mutation strategy. The mutation operations include: node addition and deletion, edge reconnection, parameter perturbation, or subgraph recombination. The task compilation module is used to compile the knowledge structure formed after mutation into a state-action-reward triplet that can be executed by the performance spiral layer module.

[0031] Specifically, such as Figure 2As shown, the knowledge retrieval module (matching relevant knowledge based on input) is responsible for retrieving the most relevant knowledge fragments (such as historical execution plans of similar tasks, adapted action sequences, and associated constraints) from the process knowledge graph based on the task instructions, environmental status, and control layer strategies of the current input layer.

[0032] The basic principle is to use similarity retrieval and semantic matching based on graph databases (such as Neo4j) or embedded knowledge graphs (such as TransE and RotatE) to map input features to nodes and edges in the knowledge graph and extract the most relevant subgraphs or paths. This involves employing graph traversal algorithms (such as DFS and BFS), semantic similarity calculations (such as cosine similarity, Jaccard coefficient, and sentence similarity based on pre-trained language models), knowledge graph embedding retrieval (such as embedding input vectors with knowledge nodes for nearest neighbor matching), and context-aware retrieval based on attention mechanisms (such as multi-head attention weighting of the knowledge graph in Transformer) to quickly locate knowledge resources matching the current task, providing "knowledge raw materials" for subsequent mutation and compilation, and significantly reducing the cost of blind searching.

[0033] The Knowledge Mutation Module (Execution Mutation: Parameter Adjustment / Combination / Splitting / Addition) is the "engine" of the Knowledge Spiral. It is responsible for performing controllable structural and parametric mutation operations based on the knowledge provided by the Knowledge Retrieval Module. This includes adjusting action parameters (such as speed, acceleration, welding torch height), combining multiple actions to form new sequences, splitting complex tasks into sub-tasks, and adding knowledge nodes (such as new action types, new constraint rules) or edges (such as adding task-action associations).

[0034] The basic principle is to introduce appropriate randomness and domain constraints to break the local optima of the existing knowledge structure and explore better or more suitable process solutions for the new environment. The technical methods used include mutation operators in genetic algorithms (such as adding Gaussian noise to parameter vectors), graph structure editing operations (such as adding / deleting nodes and reconnecting edges), combinatorial optimization algorithms (such as dynamic programming and greedy strategies), and knowledge graph generation based on deep generative models (such as GraphVAE and Diffusion Model for Graphs). At the same time, constraint checking mechanisms (such as semantic consistency, physical feasibility, and safety boundaries) are embedded in the mutation process to ensure that the mutation results maintain semantic coherence and engineering feasibility in the knowledge graph.

[0035] The task compilation module (which generates S / A / R: State / Action / Reward) is responsible for compiling the knowledge structure (such as action sequences, parameter configurations, and constraint relationships) formed after mutation into the executable "State-Action-Reward" (S / A / R) triples of the performance spiral, so that the performance spiral can be executed and evaluated in simulation or real environments.

[0036] The principle is to transform abstract knowledge graph nodes and edges into specific executable instruction sequences (such as robot joint angle trajectories, velocity curves, and switch signals), and define state space (such as robot pose, workpiece state, and environmental obstacles), motion space (such as joint control variables and end effector instructions), and reward function (such as positive reward for successful task completion and negative reward for collision or timeout).

[0037] The technical approaches employed include state machine modeling (such as finite state automata), action space discretization or continuous encoding (such as PID parameter mapping and trajectory interpolation), reward function design (such as sparse reward + dense penalty and weighted scoring based on process standards), and policy compilation based on reinforcement learning (such as converting knowledge graphs into policy network inputs), thereby transforming the "possible solutions" at the knowledge level into "executable instances" at the performance level, providing a standard interface for subsequent evaluation and feedback.

[0038] It should be noted that the process knowledge graph is the "knowledge base" of the knowledge spiral, organizing structured knowledge in the process domain in the form of a graph structure. Nodes include tasks (such as "welding" and "handling"), objects (such as "type A workpiece" and "welding gun"), actions (such as "move to coordinates" and "trigger welding"), and constraints (such as "speed ≤ 50 mm / s" and "safety distance ≥ 10 cm"). Edges represent the relationships between nodes (such as task → action, action → object, constraint → action) and weights (such as execution frequency, reliability, and priority). This knowledge graph supports dynamic updates.

[0039] The principle is to express the semantic associations and dependencies between knowledge through graph structures, supporting efficient retrieval, reasoning, and evolution; it employs database storage (such as Neo4j, JanusGraph), knowledge graph embedding (such as TransE, RotatE, ComplEx) to realize semantic similarity calculation, graph neural networks (GNN) for knowledge reasoning and completion, and dynamic graph update mechanisms (such as incremental knowledge injection and version control) to build a scalable, evolvable, and reasonable foundation for process knowledge, supporting the continuous learning and optimization of the knowledge spiral.

[0040] In some embodiments, the performance spiral layer module includes: An execution environment module is used to execute the executable instructions in a simulation environment and / or a real robot environment; The trajectory acquisition module is used to periodically or event-triggeredly acquire the execution trajectory data, which includes at least: state variables, action variables, time information, and energy consumption data; The performance evaluation module is used to quantify the execution trajectory data into comparable performance scores by setting thresholds, formula calculations, or machine learning models, and obtain performance evaluation results. The results feedback module is used to structure and organize the performance evaluation results and transmit them to the bidirectional coupling and evolution control layer module.

[0041] Specifically, such as Figure 2 As shown, the execution environment module is responsible for executing the S / A / R instructions generated by the task compilation module in a controlled environment. This can be a high-precision simulation environment (such as ROS+Gazebo, Unity, NVIDIA Omniverse) or a real industrial robot system (such as UR, KUKA, FANUC). The principle is to simulate or actually drive the robot to execute action sequences and observe its behavior and results under specific environments and tasks. The technical methods used include physics engine modeling (such as Bullet, ODE, PhysX for simulation), robot kinematics / dynamics solving (such as inverse kinematics IK, trajectory planning RRT*), real-time control systems (such as ROS Control, PLC drive), and virtual-real synchronization technology (such as digital twin alignment), thereby testing the feasibility, stability, and performance indicators of candidate skills under safe, repeatable, and low-cost conditions. The trajectory acquisition module (recording state / action / time / energy consumption) is responsible for recording the complete execution trajectory data of the robot and its environment in real time during execution. This includes state variables (such as robot end-effector pose, joint angles, velocity, and acceleration), action variables (such as executed control commands, switch signals, and welding torch current), timestamps (accurately recording data at each moment), and energy consumption data (such as motor power consumption, battery consumption, and heat dissipation). Its principle is to periodically or event-triggeredly collect key indicators during execution through sensor and control system interfaces to construct a complete "execution log." The technical methods employed include data bus acquisition (such as CAN, EtherCAT, and ROS Topic subscription), time-series database storage (such as InfluxDB and TimescaleDB), multi-channel synchronous sampling (ensuring temporal alignment of state and action), and edge computing preprocessing (such as downsampling and feature extraction), thereby providing high-precision, high-dimensional raw data support for performance evaluation. The performance evaluation module (calculating success rate / efficiency / safety / energy consumption) is responsible for quantitatively evaluating the execution results based on the raw data recorded by the trajectory acquisition module, combined with predefined evaluation indicators (such as success rate, execution efficiency, safety margin, energy consumption level, etc., each indicator is assigned different weights). Its principle is to transform the raw trajectory data into comparable performance scores by setting thresholds, formula calculations, or machine learning models. The technical methods employed include indicator calculations (such as success rate = number of successful attempts / total number of attempts, efficiency = task completion time / theoretical minimum time, safety margin = (safe distance - actual distance) / safe distance), multi-dimensional evaluation matrices (such as weighted comprehensive scores), performance fluctuation analysis based on statistical process control, and machine learning-driven anomaly detection and performance prediction (such as LSTM predicting future energy consumption, Isolation Forest identifying abnormal trajectories), thereby objectively and comprehensively measuring the actual performance of candidate skills and providing data support for feedback analysis. The results feedback module (results organization → control center) is responsible for structuring the evaluation results (such as scores, rankings, failure reasons, and constraint violations) generated by the performance evaluation module and transmitting them to the "two-way coupling and evolution control center" to form a closed-loop feedback. Its principle is to transform the evaluation results into update instructions or strategy adjustment signals that the control center can understand, driving the next round of mutation and optimization in the knowledge spiral. The technical methods employed include result serialization (such as JSON, Protobuf), data compression and transmission protocols (such as TCP / IP, MQTT), structured feedback templates (such as {"score": 0.85, "failure_reason": "collision", "constraint_violation": "insufficient safe distance"}), and event-based triggering mechanisms (such as automatically triggering feedback upon task completion), thereby ensuring that the performance evaluation results are promptly and accurately transmitted back to the control center, driving continuous system evolution.

[0042] In some embodiments, the robotic process skill generation system further includes an input layer module, which includes: The environmental observation module is used to collect environmental information in real time through multimodal sensors. The environmental information includes at least: workpiece pose, obstacle distribution, and robot body state. The task instruction module is used to convert the task instruction into a semantic model and match it with the task node to generate a task description vector. The input integration module is used to preprocess multi-source heterogeneous data, generate standardized system input, and send it to the bidirectional coupling and evolution control layer module.

[0043] Specifically, the environmental observation module (sensor data / scene state), as the system's perception entry point, collects dynamic state information of the external physical environment and work scene in real time through multimodal sensors (such as industrial cameras, LiDAR, force / torque sensors, encoders, inertial measurement units, etc.). This includes workpiece pose, surface texture, obstacle distribution, robot joint angles and speeds, ambient light and temperature, etc., providing high-fidelity, low-latency raw data support for subsequent decision-making and execution. It mainly employs multi-sensor fusion technology (such as Kalman filtering, particle filtering, or deep learning-driven multimodal fusion networks), scene semantic segmentation and instance segmentation (based on CNN or Transformer architectures such as Mask R-CNN, SegmentAnything model), spatial localization and mapping (SLAM algorithms such as ORB-SLAM, LOAM, or vision-based dense reconstruction), and state estimation and anomaly detection (such as anomaly data recognition based on statistical process control or autoencoders), providing environmental perception data for the control layer's strategy decision-making and the mutation generation of the knowledge spiral. The task instruction module (goal / constraint / priority) carries the output integration module, clearly defining the operational goals the system needs to complete (e.g., "transfer to designated workstation"), execution constraints (e.g., maximum speed limit, safe distance, energy consumption limit, process quality requirements), and task execution priorities (e.g., priority for urgent tasks, priority for high-value tasks). This module transforms natural language or structured instructions into machine-understandable semantic models and matches them with task nodes in the knowledge graph to form executable task description vectors. It employs semantic parsing from Natural Language Processing (NLP) (e.g., intent recognition and slot filling based on BERT or GPT), task modeling languages ​​(e.g., PDDL planning domain definition), constraint satisfaction problem (CSP) modeling and solving, and priority scheduling algorithms (e.g., weighted sorting based on task urgency, resource utilization, or historical success rate) to transform high-level task requirements into specific goals and boundary conditions that are schedulable at the control layer, mutable in the knowledge spiral, and executable in the performance spiral. The input integration module (parsing / standardization / conflict detection) is responsible for uniformly processing multi-source heterogeneous data from environmental observations, mission instructions, and new environmental observations to eliminate differences in data format, inconsistencies in units, or semantic conflicts, thus constructing a structured and standardized system input. The basic principle is to perform data preprocessing: first, parsing the raw data (e.g., extracting key fields from JSON, ROS messages, or binary streams); then, performing standardization operations (e.g., normalization, unit conversion, coordinate system alignment); and finally, conflict detection (e.g., triggering alarms when the deviation of multiple sensors' estimations of the same object's position exceeds a threshold, or using a weighted fusion mechanism). This ensures that the input received by the system remains consistent in both semantics and numerical values. It employs various strategies, including regular expressions and structured parsers (e.g., JSON Schema validation), data standardization algorithms (e.g., Min-Max scaling, Z-Score standardization), and conflict detection and resolution strategies (e.g., fusion based on DS evidence theory, voting mechanisms, or conflict adjudication based on expert rules). This ensures that subsequent modules perform reasoning and decision-making on a unified data foundation, avoiding system failure or performance degradation due to input ambiguity.

[0044] In some embodiments, the robotic process skill generation system further includes an output layer module, which comprises: The new task identification module is used to extract failure patterns or user intervention signals from the execution logs of the performance spiral layer module and generate new task requirements. The new environment observation module is used to detect scene changes through multimodal sensors, identify new environmental features, and generate new environmental inputs; The output integration module is used to integrate the new task requirements and the new environment inputs to generate the next round of external inputs, which are then re-injected into the bidirectional coupling and evolution control layer module to form a cross-cycle closed loop.

[0045] Specifically, the new task identification module (discovering new requirements from results) is responsible for mining potential new task requirements or optimization directions (such as "need to add obstacle avoidance actions", "need to support new workpiece types", "need to improve welding speed") from the execution results of the performance spiral (such as task failure, inefficiency, user intervention). Its principle is based on data mining, pattern recognition or natural language generation technology to extract implicit requirements from result logs, user feedback or abnormal behavior, thereby turning "execution failure" into "opportunity for knowledge evolution" and driving the system to actively adapt to new tasks. The new environment observation module (sensors capture new scenes) is responsible for continuously monitoring environmental changes through sensors, identifying new scene features (such as new workpiece types, new layouts, new obstacles, and changes in lighting), and converting these changes into "new environment inputs" that the system can recognize. Its principle is multimodal perception and scene understanding, combined with historical knowledge graphs for scene matching and difference detection. The technical methods used include visual saliency detection (such as scene change detection based on deep learning), environmental feature extraction (such as point cloud registration and image feature matching), scene classification and recognition (such as ResNet and YOLO for workpiece / obstacle recognition), and knowledge graph-based scene matching (such as comparing new scene features with known scene nodes), thereby enabling the system to have "environmental self-adaptation" capabilities and respond to external changes in a timely manner. The output integration module (generating the next round of input) is responsible for integrating the "new needs / new scenarios" discovered by the new task identification module and the new environment observation module to generate the next round of "external environment / task input," which is then re-injected into the system input layer, forming a closed loop of "perception → decision → execution → evaluation → feedback → re-perception." Its principle is to transform scattered new needs and scenarios into a unified input structure (such as task instructions + environmental state), ensuring that the system can continuously receive new stimuli and initiate a new round of evolution. The technical methods employed include data structured encapsulation (such as converting new tasks to PDDL format and new scenarios to knowledge graph nodes), input standardization (maintaining consistency with the input integration module), conflict resolution and priority ranking (such as scheduling strategies when new tasks conflict with old tasks), and output caching and batch injection (avoiding frequent system overload), thereby ensuring that the double helix structure can operate continuously, stably, and adaptively, constantly evolving to cope with complex dynamic environments.

[0046] In some embodiments, the new task identification module generates the new task requirements using at least one of the following methods: Cluster analysis, association rule mining, natural language generation, or implicit demand inference based on reinforcement learning.

[0047] Specifically, high-frequency failure patterns can be identified through cluster analysis (such as K-means), association rules can be mined (such as the Apriori algorithm) to discover the correlation between tasks and failures, natural language generation (such as generating user-understandable requirement descriptions from failure logs), and implicit requirement inference based on reinforcement learning (such as inferring requirement gaps from reward changes) can be used to transform "execution failure" into "opportunities for knowledge evolution" and drive the system to proactively adapt to new tasks.

[0048] This invention establishes a closed-loop self-evolving system by bidirectionally coupling a knowledge generation and evolution mechanism (knowledge spiral layer) and an execution performance evaluation and feedback mechanism (performance spiral layer) with a control center at its core. This system forms a cycle of "knowledge-driven execution—execution feedback performance—performance-driven knowledge mutation—knowledge update and re-execution." The knowledge spiral layer not only includes knowledge retrieval, mutation, and compilation functions but also introduces an intelligent decision-making mechanism that dynamically adjusts mutation strategies based on task requirements and performance indicators. The performance spiral layer collects multi-dimensional performance data (such as success rate, efficiency, security, and energy consumption) from simulated or real execution environments and, combined with weight evaluation and result feedback, drives the online updating and optimization of the knowledge graph. The architecture employs a five-layer collaborative structure: "input layer—control layer—knowledge spiral layer—performance spiral layer—output layer." The data flow and feedback paths between layers are clear and explicit, achieving full-link closed-loop control from environmental perception, task parsing, knowledge evolution to performance verification.

[0049] The double helix structure in this application is essentially a closed-loop system of "genotype (knowledge) - phenotype (performance) - evolutionary control", which enables robot process skills to actively innovate, improve themselves and achieve cross-task transfer during task execution. It breaks through the limitations of traditional open-loop or single-loop skill generation modes, and completes the leap of robot skills from "passive execution" to "active evolution", providing core technical support for unstructured manufacturing and flexible production of multiple varieties.

[0050] Based on the same inventive concept, this application also provides a method for generating robot process skills based on a double-helix coupling architecture, the method comprising: S1. The bidirectional coupling and evolution control layer module determines the control strategy and sends it to the knowledge spiral layer module based on the externally input task instructions and / or environmental status, as well as the performance evaluation results returned by the performance spiral layer module. S2. The knowledge spiral layer module performs mutation on the process knowledge graph according to the received control strategy to generate candidate skills; wherein, the process knowledge graph organizes the structured knowledge of the process domain in the form of a graph structure, the nodes include tasks, objects, actions and constraints, and the edges represent the association and weight between nodes; S3. The bidirectional coupling and evolution control layer module receives the candidate skills generated by the knowledge spiral layer module, compiles the candidate skills into executable instructions, and sends them to the performance spiral layer module. S4. The performance spiral layer module receives and executes the executable instructions, runs the candidate skills in a simulation environment and / or a real execution environment; collects execution trajectory data, and performs quantitative evaluation on the execution trajectory data based on preset evaluation indicators to obtain performance evaluation results; S5. Return the performance evaluation results to the bidirectional coupling and evolution control layer module; S6. The bidirectional coupling and evolution control layer module generates a knowledge update instruction based on the received performance evaluation result and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform directional mutation and update on the process knowledge graph. S7. Repeat steps S1 to S6 to form a closed-loop evolutionary process of knowledge generation, performance verification, and feedback update.

[0051] In implementation, the core process of this application revolves around the bidirectional coupling and iterative evolution of the "knowledge spiral" and the "performance spiral," forming a closed loop of "input stimulus → knowledge mutation generation → performance evaluation → feedback-driven evolution → new input triggering." Specifically, as follows... Figure 1 As shown: External environment, task input (starting point): Inputs include: new task instructions (such as "welding type A workpiece"), environmental changes (such as workpiece position shift, lighting changes), and user interventions.

[0052] Output: Flows to the control center as the initial trigger signal for this round of evolution.

[0053] Two-way coupling and evolutionary control center (central hub): Input: Feedback from "External Environment / Task Input" and "New Task / New Environment Observation".

[0054] Output: Downward: Control command → Knowledge Spiral (trigger mutation / generation).

[0055] Functions: Based on the current input and the performance evaluation results of the previous round, update the mutation strategy (such as structural mutation intensity and parameter perturbation range); compile the candidate skills output by the "Knowledge Spiral" to generate executable instructions; receive the evaluation results of the "Performance Spiral" and drive the Knowledge Spiral to update.

[0056] Knowledge Spiral (Process Knowledge + Strategy): Input: "Mutation strategy" and "Update instructions" from the control center.

[0057] Output: Candidate Skills → Control Center.

[0058] Function: Based on the current knowledge graph (including tasks, objects, actions, and constraint nodes), perform structural mutation or parameter mutation; generate candidate process skills (such as action sequences, parameter configurations, and subgraph reorganization schemes).

[0059] Performance spiral (execution results + evaluation): Input: "Executable instructions" (i.e. compiled candidate skills) from the control center.

[0060] Output: Evaluation results → Control Center (driving evolution); simultaneously, "new observations" during execution → New task / new environment observation module.

[0061] Functions: Execute candidate skills in simulation or real environments, collect trajectories, states, and results; evaluate performance (success rate, efficiency, energy consumption, risk, etc.) through multi-dimensional constraint verification (semantic / technical, logical consistency, physical feasibility, safety, etc.).

[0062] New task / new environment observation (feedback node): Input: "Execution process observations" from Performance Spiral (such as workpiece deformation, sensor anomalies, environmental interference).

[0063] Output: Feedback is sent to the "External Environment, Task Input" module (forming an environmental closed loop) and the "Control Center" (triggering a new round of control decisions).

[0064] Function: Extract "new task features" or "environmental change signals"; Function: To enable the system to have environmental awareness and adaptive capabilities.

[0065] This application combines high knowledge generation efficiency (directed mutation reduces invalid exploration), strong execution reliability (multi-dimensional constraint verification ensures security), continuous evolution capability (closed-loop feedback enables knowledge spiral and performance spiral to iterate in synergy), and high interpretability (explicit storage of process logic in knowledge graph), thereby breaking through the technical bottlenecks of "static solidification of knowledge" and "isolated feedback of execution" in traditional intelligent manufacturing systems.

[0066] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robot process skill generation system based on a double-helix coupling architecture, characterized in that, include: Bidirectional coupling and evolution control layer module, knowledge spiral layer module, and performance spiral layer module; The bidirectional coupling and evolution control layer module is bidirectionally connected to the knowledge spiral layer module and the performance spiral layer module, respectively. The bidirectional coupling and evolution control layer module is used to determine the control strategy and send it to the knowledge spiral layer module based on the externally input task instructions and / or environmental state, as well as the performance evaluation results returned by the performance spiral layer module. Receive candidate skills generated by the knowledge spiral layer module, compile the candidate skills into executable instructions and send them to the performance spiral layer module; The system receives the performance evaluation results returned by the performance spiral layer module, generates a knowledge update instruction based on the performance evaluation results, and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform targeted mutation and update on the process knowledge graph. The knowledge spiral layer module is used to store and maintain the process knowledge graph, wherein the process knowledge graph organizes the structured knowledge of the process domain in the form of a graph structure, the nodes include tasks, objects, actions and constraints, and the edges represent the associations and weights between nodes; Based on the received control strategy, the process knowledge graph is mutated to generate candidate skills; Update the process knowledge graph according to the received knowledge update instruction; The performance spiral layer module is used to receive and execute the executable instructions to run the candidate skills in a simulation environment and / or a real execution environment; Execution trajectory data is collected, and the execution trajectory data is quantitatively evaluated based on preset evaluation indicators to obtain performance evaluation results; The performance evaluation results are returned to the bidirectional coupling and evolution control layer module.

2. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The bidirectional coupling and evolution control layer module also includes: The constraint verification module is used to perform feasibility screening on the candidate skills before compiling them into executable instructions, filter out candidate skills that violate preset semantic constraints, process logic or security boundaries, and send the candidate skills that pass the verification to the performance spiral layer module.

3. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The bidirectional coupling and evolution control layer module is also used for: The knowledge subgraphs corresponding to candidate skills whose scores exceed a preset threshold in the performance evaluation results are marked as elite individuals and directly retained in the next generation of process knowledge graphs, while mutation operations on the elite individuals are suppressed.

4. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The control strategy includes at least one of the following: mutation strategy, knowledge compilation method, and knowledge graph update strategy.

5. The robot process skill generation system based on a double-helix coupling architecture according to claim 4, characterized in that, The knowledge spiral layer module includes: The knowledge retrieval module is used to retrieve knowledge fragments that match the current scenario from the process knowledge graph based on the current task instructions, environmental state, and control strategy, as mutation seeds; wherein, the knowledge fragments include historical execution schemes of similar tasks, adapted action sequences, or associated constraints; The knowledge mutation module is used to perform structural and / or parametric mutation operations on the mutation seed according to the mutation strategy. The mutation operations include: node addition and deletion, edge reconnection, parameter perturbation, or subgraph recombination. The task compilation module is used to compile the knowledge structure formed after mutation into a state-action-reward triplet that can be executed by the performance spiral layer module.

6. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The performance spiral layer module includes: An execution environment module is used to execute the executable instructions in a simulation environment and / or a real robot environment; The trajectory acquisition module is used to periodically or event-triggeredly acquire the execution trajectory data, which includes at least: state variables, action variables, time information, and energy consumption data; The performance evaluation module is used to quantify the execution trajectory data into comparable performance scores by setting thresholds, formula calculations, or machine learning models, and obtain performance evaluation results. The results feedback module is used to structure and organize the performance evaluation results and transmit them to the bidirectional coupling and evolution control layer module.

7. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The robot process skill generation system further includes an input layer module, which comprises: The environmental observation module is used to collect environmental information in real time through multimodal sensors. The environmental information includes at least: workpiece pose, obstacle distribution, and robot body state. The task instruction module is used to convert the task instruction into a semantic model and match it with the task node to generate a task description vector. The input integration module is used to preprocess multi-source heterogeneous data, generate standardized system input, and send it to the bidirectional coupling and evolution control layer module.

8. The robot process skill generation system based on a double-helix coupling architecture according to claim 1, characterized in that, The robot process skill generation system further includes an output layer module, which comprises: The new task identification module is used to extract failure patterns or user intervention signals from the execution logs of the performance spiral layer module and generate new task requirements. The new environment observation module is used to detect scene changes through multimodal sensors, identify new environmental features, and generate new environmental inputs; The output integration module is used to integrate the new task requirements and the new environment inputs to generate the next round of external inputs, which are then re-injected into the bidirectional coupling and evolution control layer module to form a cross-cycle closed loop.

9. The robot process skill generation system based on a double-helix coupling architecture according to claim 8, characterized in that, The new task identification module generates the new task requirements using at least one of the following methods: Cluster analysis, association rule mining, natural language generation, or implicit demand inference based on reinforcement learning.

10. A method for generating robot process skills based on a double-helix coupling architecture, characterized in that, The robot process skill generation method based on the double-helix coupling architecture includes: S1. The bidirectional coupling and evolution control layer module determines the control strategy and sends it to the knowledge spiral layer module based on the externally input task instructions and / or environmental status, as well as the performance evaluation results returned by the performance spiral layer module. S2. The knowledge spiral layer module performs mutation on the process knowledge graph according to the received control strategy to generate candidate skills; wherein, the process knowledge graph organizes the structured knowledge of the process domain in the form of a graph structure, the nodes include tasks, objects, actions and constraints, and the edges represent the association and weight between nodes; S3. The bidirectional coupling and evolution control layer module receives the candidate skills generated by the knowledge spiral layer module, compiles the candidate skills into executable instructions, and sends them to the performance spiral layer module. S4. The performance spiral layer module receives and executes the executable instructions, runs the candidate skills in a simulation environment and / or a real execution environment; collects execution trajectory data, and performs quantitative evaluation on the execution trajectory data based on preset evaluation indicators to obtain performance evaluation results; S5. Return the performance evaluation results to the bidirectional coupling and evolution control layer module; S6. The bidirectional coupling and evolution control layer module generates a knowledge update instruction based on the received performance evaluation result and sends it to the knowledge spiral layer module to drive the knowledge spiral layer module to perform directional mutation and update on the process knowledge graph. S7. Repeat steps S1 to S6 to form a closed-loop evolutionary process of knowledge generation, performance verification, and feedback update.

Images