AI-driven embodied intelligent adaptive learning method and system

Abstract

The application relates to the technical field of artificial intelligence, and an AI-driven embodied intelligent adaptive learning method and system, which comprises the following steps: confirming and receiving an adaptive learning instruction, obtaining a learning task type, searching a target benchmark template from a priori knowledge base by using external environment data and the learning task type, performing parameter initialization on the external environment data by using the target benchmark template to obtain an initialization parameter set, obtaining an optimized parameter node set by using an optimization algorithm and the initialization parameter set, performing parameter adjustment on the optimized parameter set by using the priori knowledge base to obtain a global parameter set, generating a control instruction set by using the global parameter set, driving the embodied intelligent agent by using the control instruction set, collecting a plurality of execution result data by using a perception driving unit, generating a detection report, and obtaining an optimized priori knowledge base. The application can improve the decision-making capability and task execution efficiency of the embodied intelligent agent in the adaptive learning process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-driven embodied intelligence adaptive learning method and system. Background Technology

[0002] AI-driven embodied intelligence adaptive learning is a technology that enables intelligent agents to learn autonomously and adapt to environmental changes through AI technology. This method utilizes the embodied agent's perception-driven, computation-driven, data-driven, and interaction-driven units to collect and process external environmental data in real time, and uses a pre-built prior knowledge base for parameter initialization and optimization. It can significantly improve the embodied agent's decision-making ability and task execution efficiency in complex and dynamic environments, enabling the embodied agent to continuously adjust and optimize its behavior based on real-time data, thereby better completing tasks.

[0003] Existing adaptive learning methods often suffer from latency and data loss when processing large amounts of real-time data, causing embodied agents to fail to respond promptly to environmental changes. Therefore, improving the decision-making ability and task execution efficiency of embodied agents in the adaptive learning process is an urgent technical problem to be solved. Summary of the Invention

[0004] This invention provides an AI-driven embodied intelligent adaptive learning method and system, the main purpose of which is to improve the decision-making ability and task execution efficiency of embodied intelligent agents in the adaptive learning process.

[0005] To achieve the above objectives, the present invention provides an AI-driven embodied intelligence adaptive learning method, comprising:

[0006] The adaptive learning instruction is received and parsed to obtain the learning task type. The embodied intelligent agent and 6G network communication link are identified. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit.

[0007] Visual image data and LiDAR point cloud data are acquired using a preset sampling time interval, a perception driving unit, and a 6G network communication link. External environment data, including multiple environmental parameters, is obtained using the visual image data and LiDAR point cloud data. A target benchmark template is retrieved from a pre-built prior knowledge base using the external environment data and the learning task type. Each environmental parameter in the multiple environmental parameters corresponding to the external environment data is initialized using the target benchmark template to obtain an initialization parameter set. The initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates.

[0008] Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0009] The optimized parameter node set is transmitted to the data-driven unit;

[0010] The optimization parameter set and confidence evaluation parameter set are extracted from the optimization parameter node set of the data-driven unit. The optimization parameter set is adjusted using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set, and the global parameter set is transmitted to the interaction-driven unit.

[0011] A control instruction set is generated using a global parameter set, the control instruction set is used to drive the embodied intelligent agent, and multiple execution result data of the embodied intelligent agent are collected using a perception driving unit. The multiple execution result data are transmitted to the data driving unit based on a 6G network communication link, and a detection report is generated based on the multiple execution result data in the data driving unit.

[0012] The prior knowledge base is updated based on the test report to obtain an optimized prior knowledge base.

[0013] Optionally, the step of acquiring external environment data including multiple environmental parameters using visual image data and LiDAR point cloud data includes:

[0014] The perception-driving unit is adjusted based on the learning task type. The adjusted perception-driving unit is used to perform real-time enhancement processing on visual image data to obtain a preprocessed image set. Then, noise reduction and motion compensation processing are performed on LiDAR point cloud data to obtain optimized point cloud data.

[0015] The preprocessed image set and optimized point cloud data are fused from multiple sources to obtain external environment data.

[0016] Optionally, the step of retrieving a target benchmark template from a pre-built prior knowledge base using external environment data and learning task type, and using the target benchmark template to initialize each of the multiple environmental parameters corresponding to the external environment data, yields an initialization parameter set, including:

[0017] By analyzing environmental parameters in external environment data using a pre-verified edge computing module, the current task encoding vector, current environment feature fingerprint, and device information corresponding to the environmental parameters are obtained. The following operations are performed on each of the multiple benchmark templates:

[0018] Extract the task encoding vector of the baseline template, and calculate the task type similarity using the current task encoding vector, the task encoding vector of the baseline template, and the pre-constructed task type similarity calculation formula;

[0019] Extract the environmental feature fingerprint of the benchmark template, and calculate the environmental similarity using the current environmental feature fingerprint, the environmental feature fingerprint of the benchmark template, and the pre-constructed environmental similarity calculation formula;

[0020] The hardware compatibility is obtained by comparing the device information with a pre-built device compatibility comparison table.

[0021] The total matching score is calculated using task type similarity, environment similarity, and hardware compatibility. The calculation formula is shown below:

[0022] ;

[0023] in, The total match score. For task type similarity, Indicates environmental similarity. For hardware compatibility, All are weighting coefficients;

[0024] The total matching degree is summarized to obtain a total matching degree set. The total matching degree in the total matching degree set is sorted in descending order to obtain a total matching degree sequence. Five candidate benchmark templates are identified in the total matching degree sequence. The five candidate benchmark templates are the benchmark templates corresponding to the first five total matching degrees in the total matching degree sequence. Each of the five candidate benchmark templates is input into a pre-built digital twin platform for verification to obtain five verification parameters. The candidate benchmark template corresponding to the verification parameter with the largest value among the five verification parameters is taken as the target benchmark template.

[0025] The environmental parameters are initialized using the verification parameters corresponding to the target benchmark template to obtain the initialization parameters. The initialization parameters are then summarized to obtain the initialization parameter set.

[0026] Optionally, the optimization algorithm for obtaining initialization parameters based on the initialization parameter set, which optimizes the initialization parameters to obtain optimized parameter nodes, includes:

[0027] The optimization algorithm is obtained by using the initialization parameters in the initialization parameter set and the pre-built algorithm mapping table;

[0028] The optimization algorithm is used to perform hierarchical optimization calculations on the initialization parameters to obtain optimized parameters for multiple edge nodes;

[0029] A fixed number of gradient descent steps are performed on the optimization parameters of the multiple edge nodes to obtain multiple preliminary optimization parameters;

[0030] The digital twin platform was used to verify and adjust several preliminary optimization parameters, resulting in several optimized parameters.

[0031] The confidence evaluation parameters are obtained by using a pre-built confidence evaluation method to evaluate each of the multiple optimization parameters.

[0032] The optimization parameters and their corresponding confidence evaluation parameters are integrated to generate optimization parameter nodes.

[0033] Optionally, the step of adjusting the optimization parameter set using a prior knowledge base and a confidence evaluation parameter set to obtain a global parameter set includes:

[0034] Based on the confidence assessment parameters in the confidence assessment parameter set, the optimization parameters in the optimization parameter set are divided into three types of partitioning parameters, namely high confidence parameters, medium confidence parameters, and low confidence parameters. The low confidence parameters are reinitialized using the prior knowledge base to obtain the corrected parameters.

[0035] The confidence parameter is fine-tuned using a prior knowledge base to obtain the fine-tuning parameter;

[0036] A global optimization strategy is obtained based on the aforementioned correction parameters, fine-tuning parameters, high-confidence parameters, and prior knowledge base.

[0037] The global optimization strategy is used to integrate the correction parameters, fine-tuning parameters, and high-confidence parameters to obtain a global parameter set.

[0038] Optionally, generating a control instruction set using a global parameter set includes:

[0039] A preliminary control instruction set is generated based on the global parameter set;

[0040] The interactive driving unit is used to perform syntax and semantic verification on each preliminary control instruction in the preliminary control instruction set.

[0041] If the syntax verification of the initial control instruction fails, the initial control instruction is revised according to the pre-built syntax rules to obtain a syntax revision instruction;

[0042] If the syntax verification of the preliminary control instruction passes, then the preliminary control instruction is considered a syntax-qualified instruction.

[0043] If the semantic verification of the initial control instruction is unclear, the initial control instruction is semantically optimized according to the pre-built task execution rules to obtain the semantically optimized instruction;

[0044] If the semantic verification of the preliminary control instruction is clear, then the preliminary control instruction shall be deemed a semantically qualified instruction.

[0045] The syntax-qualified instructions, semantic-qualified instructions, syntax-correction instructions, and semantic-optimization instructions are integrated to obtain the verified instruction set;

[0046] Similarity comparison is performed using the verified instruction set, the pre-confirmed historical successful instruction set, and a pre-built structural similarity algorithm to obtain a similarity evaluation value;

[0047] If the similarity evaluation value is less than the preset evaluation threshold, the verified instruction set is adjusted based on the learning task type to obtain the adjusted instruction set.

[0048] The adjusted instruction set is used as the verified instruction set, and the step of performing similarity comparison using the verified instruction set, the pre-confirmed historical successful instruction set, and the pre-built structural similarity algorithm to obtain a similarity evaluation value is repeated until the similarity evaluation value is greater than or equal to the evaluation threshold.

[0049] If the similarity evaluation value is greater than or equal to the evaluation threshold, the verified instruction set will be used as the candidate control instruction set.

[0050] The candidate control instruction set is pre-executed using a pre-built simulation execution environment. Once the effect of the pre-execution is confirmed to be the preset expected effect, the candidate control instruction set is determined as the control instruction set.

[0051] Optionally, the step of driving the embodied intelligent agent using a control instruction set and collecting multiple execution result data of the embodied intelligent agent using a perception-driven unit includes:

[0052] Data is collected when driving the embodied intelligent agent using a perception driving unit and a preset acquisition frequency to obtain multiple raw execution data.

[0053] Each of the multiple raw execution data is preprocessed to obtain multiple preprocessed execution data.

[0054] Data cleaning is performed on each of the multiple preprocessed execution data sets to obtain multiple accurate execution data sets.

[0055] Feature extraction is performed on each of the multiple accurate execution data sets to obtain multiple execution result data sets.

[0056] Optionally, the step of transmitting multiple execution result data to the data-driven unit based on the 6G network communication link, and generating a detection report based on the multiple execution result data in the data-driven unit, includes:

[0057] The multiple execution result data are transmitted to the data driving unit using a 6G network communication link to obtain the multiple execution result data after transmission.

[0058] The accuracy of the data is verified by performing a data accuracy check on the multiple execution result data after transmission, and the check results are obtained. The check results include check feasibility and check infeasibility.

[0059] After confirming that the inspection results are feasible, the multiple execution result data are compared and analyzed with multiple preset qualified performance data to obtain an analysis result report;

[0060] Anomalies are diagnosed in the analysis result report using preset anomaly diagnosis rules to obtain a detection report.

[0061] Optionally, updating the prior knowledge base based on the detection report to obtain an optimized prior knowledge base includes:

[0062] Extract feedback data from the test report;

[0063] The baseline template is updated using the feedback data to obtain the updated baseline template;

[0064] The updated baseline template is integrated into the prior knowledge base to obtain the optimized prior knowledge base.

[0065] To achieve the above objectives, the present invention also provides an AI-driven embodied intelligent adaptive learning system, comprising:

[0066] The learning task parsing module is used to confirm the receipt of adaptive learning instructions, parse the adaptive learning instructions, obtain the learning task type, and identify the embodied intelligent agent and the 6G network communication link. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit.

[0067] The parameter optimization module is used to collect visual image data and LiDAR point cloud data using a preset sampling time interval, a perception driving unit, and a 6G network communication link; to obtain external environment data including multiple environmental parameters using the visual image data and LiDAR point cloud data; to retrieve a target benchmark template from a pre-built prior knowledge base using the external environment data and the learning task type; to initialize each of the multiple environmental parameters corresponding to the external environment data using the target benchmark template, thereby obtaining an initialization parameter set; and to pass the initialization parameter set to the computation driving unit. The prior knowledge base includes multiple benchmark templates.

[0068] Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0069] The parameter adjustment module is used to transmit the set of optimized parameter nodes to the data-driven unit, extract the set of optimized parameters and the set of confidence evaluation parameters from the set of optimized parameter nodes in the data-driven unit, adjust the parameters of the set of optimized parameters using the prior knowledge base and the set of confidence evaluation parameters to obtain the global parameter set, and transmit the global parameter set to the interaction-driven unit.

[0070] The interactive feedback module is used to generate a control instruction set using a global parameter set, drive the embodied intelligent agent using the control instruction set, collect multiple execution result data of the embodied intelligent agent using a perception driving unit, transmit multiple execution result data to a data driving unit based on a 6G network communication link, generate a detection report based on multiple execution result data in the data driving unit, update the prior knowledge base based on the detection report, and obtain an optimized prior knowledge base.

[0071] To address the above problems, the present invention also provides an electronic device, the electronic device comprising:

[0072] Memory, storing at least one instruction; and

[0073] The processor executes the instructions stored in the memory to implement the AI-driven embodied intelligence adaptive learning method described above.

[0074] To address the aforementioned issues, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the AI-driven embodied intelligent adaptive learning method described above.

[0075] To address the problems described in the background art, this invention confirms the receipt of adaptive learning instructions, parses the adaptive learning instructions to obtain the learning task type, and identifies the embodied intelligent agent and the 6G network communication link. The embodied intelligent agent includes a perception-driven unit, a computing-driven unit, a data-driven unit, and an interaction-driven unit. It is evident that this invention, through the high-speed and low-latency characteristics of the 6G network communication link, enables real-time and efficient data interaction between the embodied intelligent agent and the management platform, ensuring the rapid transmission and execution of task instructions. Based on this, the present invention utilizes a preset sampling time interval, a perception driving unit, and a 6G network communication link to collect visual image data and LiDAR point cloud data. It then uses the visual image data and LiDAR point cloud data to obtain external environment data including multiple environmental parameters. Using the external environment data and the learning task type, it retrieves a target benchmark template from a pre-constructed prior knowledge base. The target benchmark template is then used to initialize each of the multiple environmental parameters corresponding to the external environment data, resulting in an initialization parameter set. This initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates. Therefore, the embodiments of the present invention can ensure the timeliness and accuracy of external environment data, providing a reliable data foundation for subsequent processing. Furthermore, based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the present invention uses the optimization algorithm to optimize the initialization parameters, obtaining optimized parameter nodes. Each optimized parameter node includes optimized parameters and confidence evaluation parameters. The optimized parameter nodes are then summarized to obtain an optimized parameter node set. Thus, the embodiments of the present invention efficiently optimize the initialization parameters through optimization algorithms, generating an optimized parameter node set containing optimized parameters and confidence evaluation parameters, significantly improving the efficiency and accuracy of parameter optimization. Next, the present invention transmits the optimized parameter node set to the data-driven unit. The optimized parameter set and confidence evaluation parameter set are extracted from the optimized parameter node set in the data-driven unit. The optimized parameter set is then adjusted using the prior knowledge base and the confidence evaluation parameter set to obtain a global parameter set. This global parameter set is then transmitted to the interaction-driven unit. It is evident that this embodiment of the present invention significantly improves the accuracy and reliability of the parameters by adjusting the optimized parameters, integrating the prior knowledge base and the confidence evaluation parameter set, and generating a global parameter set. Furthermore, the present invention uses the global parameter set to generate a control instruction set, drives the embodied intelligent agent using the control instruction set, and uses the perception-driven unit to collect multiple execution result data from the embodied intelligent agent. These multiple execution result data are transmitted to the data-driven unit via a 6G network communication link. A detection report is generated based on the multiple execution result data in the data-driven unit, and the prior knowledge base is updated based on the detection report to obtain an optimized prior knowledge base. It is evident that this embodiment of the present invention achieves adaptive optimization through a feedback mechanism, updating the prior knowledge base based on the execution result data, significantly improving the accuracy and efficiency of task execution.Therefore, the present invention can improve the decision-making ability and task execution efficiency of embodied intelligent agents in the adaptive learning process. Attached Figure Description

[0076] Figure 1 A flowchart illustrating an AI-driven embodied intelligence adaptive learning method provided in an embodiment of the present invention;

[0077] Figure 2 This is a functional block diagram of an AI-driven embodied intelligent adaptive learning system provided in an embodiment of the present invention;

[0078] Figure 3 This is a schematic diagram of the structure of an electronic device that implements the AI-driven embodied intelligent adaptive learning method according to an embodiment of the present invention.

[0079] Explanation of reference numerals in the attached figures:

[0080] 10. Electronic device; 11. Processor; 12. Memory; 13. Bus.

[0081] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0082] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0083] This application provides an AI-driven embodied intelligent adaptive learning method. The executing entity of the AI-driven embodied intelligent adaptive learning method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the AI-driven embodied intelligent adaptive learning method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0084] Reference Figure 1 The diagram shown is a flowchart illustrating an AI-driven embodied intelligence adaptive learning method according to an embodiment of the present invention. In this embodiment, the AI-driven embodied intelligence adaptive learning method includes:

[0085] S1. Confirm receipt of adaptive learning instruction, parse the adaptive learning instruction to obtain the learning task type, and confirm the embodied intelligent agent and 6G network communication link. The embodied intelligent agent includes: perception driving unit, computing driving unit, data driving unit and interaction driving unit.

[0086] It should be explained that adaptive learning instructions are issued by the embodied agent manager to initiate and configure the adaptive learning process of the embodied agent to adapt to different task requirements and environmental conditions. Learning task type refers to a specific type of intelligent task, such as image recognition, path planning, or data analysis. The 6G network communication link refers to a high-speed, low-latency communication link based on sixth-generation mobile communication technology, i.e., a 6G network, used to transmit data and instructions between the embodied agent and its management platform or other intelligent agents, ensuring efficient and real-time information interaction. An embodied agent is an intelligent system with physical form and perception capabilities, capable of performing learning tasks in its environment. It achieves adaptive learning and intelligent decision-making through the collaborative work of its various units. The embodied agent includes a perception-driven unit, a computation-driven unit, a data-driven unit, and an interaction-driven unit. For specific applications of these units, please refer to subsequent embodiments.

[0087] S2. Visual image data and LiDAR point cloud data are collected using a preset sampling time interval, a sensing drive unit, and a 6G network communication link. External environment data, including multiple environmental parameters, is obtained using the visual image data and LiDAR point cloud data.

[0088] It should be explained that the process of obtaining external environment data including multiple environmental parameters using visual image data and LiDAR point cloud data includes: adjusting the perception driving unit based on the learning task type, using the adjusted perception driving unit to perform real-time enhancement processing on the visual image data to obtain a preprocessed image set, and performing noise reduction and motion compensation processing on the LiDAR point cloud data to obtain optimized point cloud data.

[0089] The preprocessed image set and optimized point cloud data are fused from multiple sources to obtain external environment data.

[0090] It should be explained that the distributed sensing network is composed of multiple visual sensors, LiDAR, millimeter-wave radar, and ultrasonic sensors connected via 6G network communication links within the sensing drive unit. This enables multi-point, synchronous data acquisition and transmission. The sensing drive unit includes multiple environmental detection units, which in turn include visual sensors, LiDAR, millimeter-wave radar, and ultrasonic sensors. Adjusting the sensing drive unit based on the learning task type refers to adjusting the parameters of the sensing drive unit using the current learning task type. For example, adjusting the exposure time and frame rate of the visual sensor, the scanning frequency of the LiDAR, and the detection range of the millimeter-wave radar, allows the sensing drive unit to adapt to the requirements of the current learning task type, improving the accuracy and efficiency of external environmental data acquisition.

[0091] Furthermore, the preset sampling time interval is a fixed time interval set according to the learning task type requirements and data processing capabilities, such as sampling 10 times per second.

[0092] Understandably, visual image data is environmental information collected by visual sensors and represented in two-dimensional or three-dimensional image form, containing scene details and object appearance features. Real-time enhancement processing refers to the operation of instantly optimizing and improving the visual image data acquired by the perception-driven unit during the external environment data acquisition process. This includes adjusting the brightness, contrast, and sharpness of the image, and removing noise, to obtain clearer, more accurate, and easier-to-process and analyze image data. The preprocessed image set refers to the collection formed by summarizing the visual image data after real-time enhancement processing.

[0093] Furthermore, LiDAR point cloud data is data representing environmental distance and shape information in the form of a three-dimensional point cloud, collected by a LiDAR sensor. Optimized point cloud data refers to LiDAR point cloud data after noise reduction and motion compensation processing, which improves the accuracy and reliability of the LiDAR point cloud data. Multi-source data fusion utilizes convolutional neural networks to extract image features from a preprocessed image set, uses point cloud processing algorithms to extract LiDAR features from the optimized point cloud data, and fuses the image features and LiDAR features at the feature level to generate more comprehensive and accurate environmental representation data. Optionally, the VoxelNet algorithm is used as the point cloud processing algorithm. External environment data is data collected by the perception-driven unit to describe and analyze the environment surrounding the embodied intelligent agent, including visual image data and LiDAR point cloud data. Optionally, high-pass filtering technology is used to denoise the LiDAR point cloud data, and a motion compensation algorithm improved by a Gaussian mixture model is used to perform motion compensation processing on the LiDAR point cloud data.

[0094] For example, in intelligent driving scenarios, multi-source data fusion typically involves combining visual image data and LiDAR point cloud data. Algorithms are used to align, correct, and fuse these data to improve the perception of the surrounding environment. The purpose of multi-source data fusion is to compensate for the shortcomings of a single data source, such as the disadvantages of visual data in low light conditions or the limitations of LiDAR data in object recognition, thereby providing a more reliable information foundation for subsequent decision-making and control.

[0095] S3. Using external environment data and learning task type, retrieve target benchmark templates from the pre-built prior knowledge base, use the target benchmark templates to initialize each of the multiple environmental parameters corresponding to the external environment data, obtain an initialization parameter set, and pass the initialization parameter set to the computing drive unit. The prior knowledge base includes multiple benchmark templates.

[0096] Furthermore, the step involves retrieving a target benchmark template from a pre-built prior knowledge base using external environment data and learning task type, and then using the target benchmark template to initialize each of the multiple environmental parameters corresponding to the external environment data, resulting in an initialization parameter set, including:

[0097] By analyzing environmental parameters in external environment data using a pre-verified edge computing module, the current task encoding vector, current environment feature fingerprint, and device information corresponding to the environmental parameters are obtained. The following operations are performed on each of the multiple benchmark templates:

[0098] Extract the task encoding vector of the baseline template, and calculate the task type similarity using the current task encoding vector, the task encoding vector of the baseline template, and the pre-constructed task type similarity calculation formula;

[0099] Extract the environmental feature fingerprint of the benchmark template, and calculate the environmental similarity using the current environmental feature fingerprint, the environmental feature fingerprint of the benchmark template, and the pre-constructed environmental similarity calculation formula;

[0100] The hardware compatibility is obtained by comparing the device information with a pre-built device compatibility comparison table.

[0101] The total matching score is calculated using task type similarity, environment similarity, and hardware compatibility. The calculation formula is shown below:

[0102] ;

[0103] in, The total match score. For task type similarity, Indicates environmental similarity. For hardware compatibility, All are weighting coefficients;

[0104] The total matching degree is summarized to obtain a total matching degree set. The total matching degree in the total matching degree set is sorted in descending order to obtain a total matching degree sequence. Five candidate benchmark templates are identified in the total matching degree sequence. The five candidate benchmark templates are the benchmark templates corresponding to the first five total matching degrees in the total matching degree sequence. Each of the five candidate benchmark templates is input into a pre-built digital twin platform for verification to obtain five verification parameters. The candidate benchmark template corresponding to the verification parameter with the largest value among the five verification parameters is taken as the target benchmark template.

[0105] The environmental parameters are initialized using the verification parameters corresponding to the target benchmark template to obtain the initialization parameters. The initialization parameters are then summarized to obtain the initialization parameter set.

[0106] It should be explained that the pre-confirmed edge computing module refers to a computing module deployed near the perception-driving unit. It can perform preliminary processing and analysis of data at the data source, reducing data transmission latency and improving the real-time performance and response speed of the embodied agent's adaptive learning. Optionally, the edge computing module is the NVIDIA Jetson AGXXavier; other technologies can achieve the same effect, which will not be elaborated further here. Environmental parameters refer to various information used to describe and analyze the environment surrounding the embodied agent, constructing the embodied agent's perception of the surrounding environment. These parameters include temperature, humidity, light intensity, object position, velocity, acceleration, etc., and are mainly derived from visual image data and LiDAR point cloud data.

[0107] The current task encoding vector is a high-dimensional vector representation generated by the edge computing module after analyzing environmental parameters based on the current learning task type. The edge computing module first extracts key features related to the learning task from the environmental parameters, such as object position and velocity. These key features are the input conditions or influencing factors for completing the learning task. Furthermore, it defines target information based on the learning task's objective, such as object classification or path planning. The key features and target information are integrated to form a vector describing the current learning task type, i.e., the current task encoding vector. Optionally, a convolutional neural network algorithm is used to form this vector.

[0108] For example, when the learning task is image classification, key features include image size (e.g., 224×224 pixels), color space (e.g., RGB), texture features, and edge information. Target information includes classification categories (e.g., animals, plants, vehicles) or specific labels (e.g., cats, dogs, cars, airplanes). When the learning task is autonomous driving, key features include vehicle speed, acceleration, distance to the vehicle in front, lane line positions, and traffic sign types. Target information includes safe navigation (e.g., collision avoidance, staying within lanes), path planning (e.g., the optimal path from start to finish), and traffic signal recognition (e.g., red lights, green lights, yellow lights).

[0109] It should be understood that the current environmental feature fingerprint is a vector representation of the environment in which the embodied agent exists, based on environmental parameters, generated by extracting and fusing task features from multiple environmental parameters. It reflects the unique attributes and state of the environment. Optionally, a convolutional neural network algorithm can be used to form the vector. Task features refer to environmental attributes or conditions related to a specific learning task, reflecting the specific requirements of the learning task on the environment. For example, in autonomous driving tasks, road curvature, lane line clarity, and traffic sign visibility are examples. Device information refers to the detailed specifications and status information of various hardware devices involved in data acquisition and processing, including device model, sensor parameters, computing power, storage capacity, etc., reflecting the device's capabilities and limitations. The pre-built prior knowledge base is a knowledge base that is pre-established and stored, containing multiple benchmark templates. A benchmark template is a knowledge unit template in the pre-built prior knowledge base, representing the benchmark parameter configuration under different tasks and environments, including parameter configurations and behavioral patterns under specific task types, environmental characteristics, and device requirements. Each benchmark template has corresponding task encoding vectors, environmental feature fingerprints, and other attributes. The purpose of the benchmark templates is to provide a standard for parameter initialization of the embodied agent under different tasks and environments. By matching the current task and environment, the most suitable baseline template is selected. Its parameter configuration is then used to initialize the behavior of the embodied agent, enabling it to quickly adapt to new tasks and environments. Task type similarity is the similarity between the current learning task type and the corresponding learning task type of the baseline template. By quantifying the similarity between tasks, the template that best matches the current task is selected from multiple baseline templates. Combining environment similarity and hardware compatibility ensures that the selected baseline template meets the current requirements in terms of task type, environment adaptability, and device compatibility, thus providing the agent with the optimal parameter initialization scheme. Environment similarity is an indicator used to quantify the degree of similarity between the current environment and the baseline template environment in terms of features or conditions. Hardware compatibility is an indicator used to evaluate the degree of hardware-level adaptation between the current device and the baseline template. High hardware compatibility indicates that the device's hardware specifications (such as computing power, sensor type, and storage capacity) can meet the requirements of the learning task, ensuring smooth task execution. For example, if the task requires high-performance computing, and the current device's computing power is insufficient, low hardware compatibility will suggest that the user may need to upgrade the device or adjust the task parameters.

[0110] Furthermore, the formula for calculating the task type similarity is as follows:

[0111] ;

[0112] in, For task type similarity, Encode the vector for the current task. Encode the task vector. Indicates a modulo operation;

[0113] The formula for calculating environmental similarity is as follows:

[0114] ;

[0115] in, Indicates environmental similarity. Fingerprint representing the characteristics of the current environment Fingerprints represent environmental characteristics. Fingerprint length This represents the Hamming distance calculation function that takes the current environmental feature fingerprint and the environmental feature fingerprint as calculation parameters;

[0116] It should be explained that the task encoding vector is a high-dimensional vector representation used to uniquely identify and describe the learning task corresponding to the benchmark template. The environmental feature fingerprint of the benchmark template is a vector representation of the environment in which the benchmark template is applicable. Optionally, the vector is formed using a convolutional neural network algorithm. The pre-built device compatibility lookup table is a pre-established and stored table that lists the compatibility information between different devices and each benchmark template. It records in detail the degree of adaptation between each device model, sensor parameters, computing power, storage capacity, and other hardware specifications and each benchmark template. The device compatibility lookup table is shown below:

[0117] ;

[0118] It should be explained that "Complete device model match" means that the hardware model of the current device is exactly the same as the hardware model specified in the benchmark template. For example, if the benchmark template requires the use of NVIDIA Jetson AGX Xavier, and the current device is also an NVIDIA Jetson AGX Xavier, then the hardware compatibility is 1. "Same model, different version" means that the hardware model of the current device belongs to the same series as the hardware model specified in the benchmark template, but the version is different. For example, if the benchmark template requires the use of NVIDIA Jetson AGX Xavier, and the current device is an NVIDIA Jetson TX2 (same series, different version), then the hardware compatibility is 0.8. "Cross-platform compatibility mode" means that the hardware model of the current device does not belong to the same series as the hardware model specified in the benchmark template, but the device has cross-platform compatibility capabilities. In this case, the hardware architecture and performance characteristics of the device differ significantly from the requirements of the benchmark template, but the task can still be performed through software optimization and adaptation. For example, if the benchmark template requires the use of NVIDIA Jetson AGX Xavier, and the current device is a Raspberry Pi 4B (a completely different platform), then the hardware compatibility is 0.3.

[0119] Furthermore, the fingerprint length is determined during the task feature generation process of extracting and fusing multiple environmental parameters; it is equal to the dimension of the environmental feature fingerprint vector. The dimension of the environmental feature fingerprint vector refers to the number of features contained in the vector, each feature representing a quantified representation of the environment. The dimension of the environmental feature fingerprint vector is determined by the selected feature set, which is pre-defined based on the requirements of the learning task and the environmental data.

[0120] For example, in an autonomous driving task, the following five features were selected to construct an environmental feature fingerprint vector: road curvature, lane line clarity, traffic sign visibility, distance to the vehicle ahead, and vehicle speed. These five features are integrated into a 5-dimensional vector, meaning the environmental feature fingerprint vector has a dimension of 5, or the fingerprint length is 5.

[0121] Furthermore, the overall matching score is an indicator used to comprehensively evaluate the degree of matching between the current task and the benchmark template in terms of task type, environmental characteristics, and hardware compatibility. It is calculated by weighted summation, combining task type similarity, environmental similarity, and hardware compatibility. The overall matching score dataset is a collection of the overall matching scores of all benchmark templates. It contains the overall matching score value of each benchmark template for subsequent sorting and filtering operations. The overall matching score sequence is a sequence formed by arranging the overall matching scores in the overall matching score dataset in descending order. Candidate benchmark templates are selected from the overall matching score sequence, choosing the benchmark templates with the highest overall matching scores. The pre-built digital twin platform is a virtual simulation environment used to simulate and verify the performance of the benchmark templates in a real-world environment. Optionally, NVIDIA Omniverse is used as the digital twin platform; other technologies can achieve the same effect, which will not be elaborated upon here.

[0122] Furthermore, the step of inputting each of the five candidate benchmark templates into a pre-built digital twin platform for verification refers to loading each candidate benchmark template into the pre-built digital twin platform, setting up a test environment similar to the actual application scenario within the digital twin platform, including but not limited to environmental parameters, task types, and hardware configurations, simulating the operation of each candidate benchmark template, recording its performance in the simulated environment, including task completion status and performance indicators, and calculating the verification parameters for each candidate benchmark template based on the simulation results. Optionally, a performance evaluation model and data analysis algorithm are used to calculate the verification parameters. The verification parameters are the evaluation parameters obtained by verifying the candidate benchmark templates in the digital twin platform, and are used for the initial configuration of environmental parameters during the parameter initialization phase. The target benchmark template is the template that best matches the current task and environment, retrieved from the prior knowledge base based on the current learning task type and external environment data. Parameter initialization is the process of initially configuring various parameters in the external environment data processing process using the parameter values ​​in the target benchmark template. Initialization parameters are the parameters obtained through the initialization process. The parameter values ​​are set according to the configuration information in the target benchmark template. The initialization parameter set is a collection of all initialization parameters.

[0123] For example, suppose the parameters defined in the target benchmark template are in the order of {temperature, humidity, speed}, and the corresponding parameter values ​​are {25}. During parameter initialization, based on the configuration information of the target benchmark template, the parameters in the external environment data processing process are initially configured to obtain the initialization parameter set. The parameter order and values ​​in the initialization parameter set are consistent with the target benchmark template, i.e., the initialization parameter set is {temperature: 25...}. Humidity: 60%, Speed: 50km / h. The purpose of parameter initialization is to ensure that the external environment data processing can be initialized and configured according to the requirements of the target baseline template, providing a foundation for subsequent processing and optimization.

[0124] S4. Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0125] It should be explained that the optimization algorithm for obtaining initialization parameters based on the initialization parameter set uses the optimization algorithm to optimize the initialization parameters and obtain optimized parameter nodes, including:

[0126] The optimization algorithm is obtained by using the initialization parameters in the initialization parameter set and the pre-built algorithm mapping table;

[0127] The optimization algorithm is used to perform hierarchical optimization calculations on the initialization parameters to obtain optimized parameters for multiple edge nodes;

[0128] A fixed number of gradient descent steps are performed on the optimization parameters of the multiple edge nodes to obtain multiple preliminary optimization parameters;

[0129] The digital twin platform was used to verify and adjust several preliminary optimization parameters, resulting in several optimized parameters.

[0130] The confidence evaluation parameters are obtained by using a pre-built confidence evaluation method to evaluate each of the multiple optimization parameters.

[0131] The optimization parameters and their corresponding confidence evaluation parameters are integrated to generate optimization parameter nodes.

[0132] Understandably, the pre-built algorithm mapping table is a lookup table of algorithms that is pre-established and stored, used to determine the most suitable optimization algorithm based on different initialization parameter characteristics. The algorithm mapping table is built based on historical data and covers the correspondence between various parameter patterns and optimization algorithms. Optionally, a decision tree algorithm is used to identify the optimization algorithm from the algorithm mapping table. Optimization computation is a computational method that uses mathematical models to adjust initialization parameters, aiming to improve parameter performance and make them more suitable for the needs of a specific task. It typically involves defining an objective function and constraints, and finding the parameter values ​​that optimize the objective function through iterative computation. Optimization algorithms are specific methods used for optimization computation, such as gradient descent, genetic algorithms, and particle swarm optimization. Hierarchical optimization computation decomposes complex optimization tasks into multiple levels or stages, with each level focusing on a specific subset of parameters or optimization objectives, and optimizing step by step. For example, first optimize key parameters to achieve the main objective, and then optimize secondary parameters to refine and improve the overall effect. The purpose of using hierarchical optimization computation is to help simplify complex optimization tasks and improve the efficiency and effectiveness of initialization parameter optimization. Edge node optimization parameters are parameters obtained by running optimization algorithms on edge nodes in a distributed computing environment. Edge nodes typically refer to computing units located close to the data source or terminal device, capable of performing data processing and optimization locally, reducing data transmission latency. Optionally, NVIDIA Jetson series modules can be used as the edge nodes.

[0133] It should be understood that fixed-step gradient descent optimization is an iterative optimization method that updates parameters in the opposite direction of the gradient of the objective function over a fixed number of iterations to minimize the objective function. This method is simple and straightforward, and applicable to many machine learning and engineering optimization problems. For example, in machine learning, stochastic gradient descent (SGD) is used as the iterative optimization method, updating parameters by calculating the gradient using a single training sample or a mini-batch of samples in each iteration. A fixed number of iterations (e.g., 1000 steps) can be set to perform the optimization. The initial optimized parameters are those obtained after fixed-step gradient descent optimization.

[0134] Understandably, the verification and adjustment of multiple preliminary optimization parameters using a pre-built digital twin platform refers to simulating the actual application effects of these parameters in a virtual simulation environment. The effectiveness of the parameters is evaluated through simulation results, and adjustments are made as needed to ensure they perform well in a real environment. Verification involves evaluating the effects of the parameters by running them in a simulation environment, while adjustment involves optimizing the parameters using a genetic algorithm within the digital twin platform based on the verification results to improve actual performance. The optimized parameters are those obtained after verification and adjustment.

[0135] For example, suppose that in an autonomous driving scenario, the initial optimized parameters perform well on straight roads, but the vehicle frequently deviates from its lane on curves. Verification using a digital twin platform reveals that the vehicle's recognition and reaction time on curves is insufficient. Based on this result, relevant parameters are adjusted, increasing the weight given to curve recognition and optimizing the vehicle's steering control parameters to improve the vehicle's performance on curves.

[0136] It should be understood that pre-constructed confidence assessment methods are used to quantify the credibility and reliability of optimization parameters. These methods, based on statistics, machine learning, or other evaluation techniques, can provide a measure of the uncertainty of parameter performance. Optionally, Bayesian analysis can be used as the confidence assessment method. Optimization evaluation refers to applying confidence assessment methods (such as Bayesian analysis) to calculate the confidence assessment parameters of the optimization parameters, quantifying their uncertainty. The confidence assessment parameter is a metric obtained by calculating the confidence assessment of the optimization parameters, used to measure the credibility of the optimization parameters in practical applications, and the metric value is a value between 0 and 1. Quantifying the credibility of optimization parameters through confidence assessment parameters clarifies the direction of subsequent optimization, improving optimization efficiency and effectiveness.

[0137] It should be explained that integrating the optimization parameters and their corresponding confidence evaluation parameters to generate optimization parameter nodes refers to creating a data structure containing multiple optimization parameters and their confidence evaluation parameters. This structure provides a unified representation of the optimization parameters and their corresponding confidence evaluation parameters, facilitating subsequent processing and analysis. The optimization parameter node set is the collection of all optimization parameter nodes.

[0138] S5. Extract the optimization parameter set and confidence evaluation parameter set from the optimization parameter node set of the data-driven unit, and adjust the parameters of the optimization parameter set using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set.

[0139] It should be explained that the process of adjusting the optimization parameter set using a prior knowledge base and a confidence evaluation parameter set to obtain a global parameter set includes:

[0140] Based on the confidence assessment parameters in the confidence assessment parameter set, the optimization parameters in the optimization parameter set are divided into three types of partitioning parameters, namely high confidence parameters, medium confidence parameters, and low confidence parameters. The low confidence parameters are reinitialized using the prior knowledge base to obtain the corrected parameters.

[0141] The confidence parameter is fine-tuned using a prior knowledge base to obtain the fine-tuning parameter;

[0142] A global optimization strategy is obtained based on the aforementioned correction parameters, fine-tuning parameters, high-confidence parameters, and prior knowledge base.

[0143] The global optimization strategy is used to integrate the correction parameters, fine-tuning parameters, and high-confidence parameters to obtain a global parameter set.

[0144] It should be explained that SQL queries are used to extract the optimization parameter set and confidence assessment parameter set from the optimization parameter node set of the data-driven unit. The confidence assessment parameter set is a collection of multiple confidence assessment parameters extracted from the optimization parameter node set. The optimization parameter set is a collection of multiple optimization parameters extracted from the optimization parameter node set.

[0145] Understandably, parameter partitioning involves classifying optimization parameters into high-confidence, medium-confidence, or low-confidence parameters based on their specific confidence assessment values. High-confidence parameters are those with a confidence assessment value greater than or equal to 0.9; medium-confidence parameters are those with a confidence assessment value between 0.7 and 0.9; and low-confidence parameters are those with a confidence assessment value less than or equal to 0.7. Re-initializing low-confidence parameters using a prior knowledge base to obtain corrected parameters involves calling upon a large amount of historical data and experience models stored in the prior knowledge base. By matching the features of the low-confidence parameters, successful parameter configurations in similar scenarios are extracted from the prior knowledge base, and these are used as the basis for re-initializing the low-confidence parameters. Re-initialization refers to the operation of re-initializing the low-confidence parameters based on the obtained successful parameter configurations in similar scenarios. Fine-tuning of the mid-confidence parameters utilizes a prior knowledge base. The fine-tuned parameters refer to the direction and scope of fine-tuning provided by the prior knowledge base. Based on parameter adjustment experience in similar tasks and environments, subtle adjustments are made to the mid-confidence parameters. The corrected parameters are obtained by re-initializing the low-confidence parameters. The purpose of re-initializing the low-confidence parameters is to improve their adaptability to the current task and environment by incorporating successful experiences from the prior knowledge base, providing more reasonable initial values ​​for the low-confidence parameters. The purpose of fine-tuning the mid-confidence parameters is to further optimize them based on their existing performance, making them more accurately suited to the current task requirements and improving performance.

[0146] For example, in an autonomous driving task, the optimization parameters include speed, steering angle, and acceleration. After confidence evaluation, the confidence evaluation parameter corresponding to speed is 0.65, the confidence evaluation parameter corresponding to steering angle is 0.75, and the confidence evaluation parameter corresponding to throttle opening is 0.92. Based on the confidence evaluation parameters, the optimization parameters are divided into three categories: high-confidence parameter: throttle opening (0.92), medium-confidence parameter: steering angle (0.75), and low-confidence parameter: speed (0.65). Assume that successful parameter configurations in similar scenarios (in an urban road environment) stored in the prior knowledge base show that the speed is typically between 30-50 km / h. The current low-confidence speed parameter is 45 km / h (confidence 0.65). According to the prior knowledge base, the speed is reinitialized to 40 km / h. In similar scenarios, the steering angle usually needs to be fine-tuned according to the road curvature. The current steering angle is 30 degrees (confidence 0.75). Based on the prior knowledge base, the steering angle is fine-tuned to 32 degrees. The throttle opening is a high-confidence parameter (confidence 0.92), so it is directly included in the global parameter set without adjustment.

[0147] Furthermore, obtaining a global optimization strategy based on the aforementioned correction parameters, fine-tuning parameters, high-confidence parameters, and prior knowledge base includes utilizing historical data from the prior knowledge base and a pre-built learning model. Optionally, a convolutional neural network learning model is used as the learning model. Parameters from historical data are used as input to the learning model, and their corresponding system performance indicators are used as output. The relationship between the learning parameters and the optimization effect is used to train the learning model. The correction parameters, fine-tuning parameters, and high-confidence parameters are input into the trained learning model, and the optimal parameter combination and its corresponding system performance indicator are output, forming a global optimization strategy based on this. The system performance indicator is a quantitative indicator obtained by evaluating the performance of the parameters input into the learning model. Historical data refers to data collected in past related learning tasks. The parameter combination refers to the combination obtained by weighting the correction parameters, fine-tuning parameters, and high-confidence parameters according to different weights when inputting them into the trained learning model. The global optimization strategy refers to the optimization strategy formed by determining the weight of each type of input parameter using each type of input parameter and its corresponding system performance indicator. The input parameters are the correction parameters, fine-tuning parameters, and high-confidence parameters input into the learning model. By making reasonable use of parameters with different confidence levels, the global optimization strategy can optimize resource allocation, avoid resource waste, and improve resource utilization efficiency.

[0148] Furthermore, integrating the correction parameters, fine-tuning parameters, and high-confidence parameters using the global optimization strategy to obtain the global parameter set refers to weighting the correction parameters, fine-tuning parameters, and high-confidence parameters according to the weights determined in the global optimization strategy, and then summing the weighted correction parameters, fine-tuning parameters, and high-confidence parameters to obtain the global parameter set. The global parameter set is a collection of multiple parameters obtained by summing the weighted correction parameters, fine-tuning parameters, and high-confidence parameters. By integrating parameters with different confidence levels, the embodied agent can better cope with various complex environments and task requirements, enhancing the robustness and adaptability of the embodied agent.

[0149] S6. Generate a control instruction set using the global parameter set.

[0150] It should be explained that the generation of control instruction sets using the global parameter set includes:

[0151] A preliminary control instruction set is generated based on the global parameter set;

[0152] The interactive driving unit is used to perform syntax and semantic verification on each preliminary control instruction in the preliminary control instruction set.

[0153] If the syntax verification of the initial control instruction fails, the initial control instruction is revised according to the pre-built syntax rules to obtain a syntax revision instruction;

[0154] If the syntax verification of the preliminary control instruction passes, then the preliminary control instruction is considered a syntax-qualified instruction.

[0155] If the semantic verification of the initial control instruction is unclear, the initial control instruction is semantically optimized according to the pre-built task execution rules to obtain the semantically optimized instruction;

[0156] If the semantic verification of the preliminary control instruction is clear, then the preliminary control instruction shall be deemed a semantically qualified instruction.

[0157] The syntax-qualified instructions, semantic-qualified instructions, syntax-correction instructions, and semantic-optimization instructions are integrated to obtain the verified instruction set;

[0158] Similarity comparison is performed using the verified instruction set, the pre-confirmed historical successful instruction set, and a pre-built structural similarity algorithm to obtain a similarity evaluation value;

[0159] If the similarity evaluation value is less than the preset evaluation threshold, the verified instruction set is adjusted based on the learning task type to obtain the adjusted instruction set.

[0160] The adjusted instruction set is used as the verified instruction set, and the step of performing similarity comparison using the verified instruction set, the pre-confirmed historical successful instruction set, and the pre-built structural similarity algorithm to obtain a similarity evaluation value is repeated until the similarity evaluation value is greater than or equal to the evaluation threshold.

[0161] If the similarity evaluation value is greater than or equal to the evaluation threshold, the verified instruction set will be used as the candidate control instruction set.

[0162] The candidate control instruction set is pre-executed using a pre-built simulation execution environment. Once the effect of the pre-execution is confirmed to be the preset expected effect, the candidate control instruction set is determined as the control instruction set.

[0163] It should be explained that generating a preliminary control instruction set based on a global parameter set refers to generating a preliminary control instruction set based on a global parameter set using a parameter-instruction mapping model pre-built in the interactive driving unit. Optionally, an LSTM sequence generation model can be used as the parameter-instruction mapping model to convert the global parameter set into a preliminary control instruction set executable by the embodied agent.

[0164] Furthermore, syntax verification checks the compliance of preliminary control instructions based on pre-built syntax rules, optionally using instruction structure rules defined in the BNF paradigm. If a syntax error is found, such as a missing operand or action parameter, the syntax reviser is invoked to automatically complete the error and generate a syntax revision instruction. If the syntax is compliant, it is marked as a grammatically correct instruction. The syntax reviser is an automated tool used to correct syntax errors in preliminary control instructions. It analyzes the preliminary control instructions based on pre-built syntax rules (such as instruction structure rules defined in the BNF paradigm). If a syntax error is found (such as a missing operand or action parameter), the syntax reviser automatically completes the missing part or corrects the erroneous structure, generating a syntax revision instruction. A syntax revision instruction is an instruction corrected by the syntax reviser, while a grammatically correct instruction is an instruction that passes syntax verification directly without revision.

[0165] Understandably, semantic verification checks the executability of instructions based on pre-built task execution rules. If the semantics are ambiguous, semantic optimization is performed to generate semantically optimized instructions; if the semantics are clear, they are marked as semantically qualified instructions. Semantic optimization, when the semantics of an instruction are ambiguous, uses a perception-driven unit to refine the instruction. For example, when the initial control instruction does not specify the execution target, the perception-driven unit clarifies the target. Semantically optimized instructions are the initial control instructions that were semantically ambiguous after semantic verification, and semantically qualified instructions are those that have been confirmed as qualified after semantic verification. The verified instruction set is a collection of multiple instructions, summarizing grammatically qualified instructions, semantically qualified instructions, grammatically revised instructions, and semantically optimized instructions. Syntax and semantic verification effectively reduce errors in initial control instructions, improve their accuracy and reliability, and ensure that initial control instructions correctly guide the behavior of the embodied agent.

[0166] Furthermore, task execution rules are a predefined set of conditions and constraints based on the specific learning task type, used to determine whether the instructions are sufficiently explicit for execution. Task execution rules include the explicitness of the operation object, the completeness of the operation conditions, the logicality of the operation sequence, and environmental adaptability.

[0167] It should be explained that the pre-validated historical successful instruction set is a set of instructions extracted from the prior knowledge base that have been validated in similar learning tasks. Optionally, a cosine similarity algorithm is used as the structural similarity algorithm, which is used to quantify the similarity between the validated instruction set and the historical successful instruction set. Similarity comparison is a method used to evaluate the similarity assessment value between the validated instruction set and the historical successful instruction set. The similarity assessment value is a numerical value between 0 and 1, representing the degree of similarity between the two; the higher the value, the stronger the similarity. The preset assessment threshold is a pre-set similarity assessment standard; optionally, the assessment threshold is set to 0.8 here, used to determine whether the validated instruction set is sufficiently similar to the historical successful instruction set. Targeted adjustment involves revising the validated instruction set according to the learning task type, such as adjusting the instruction order, to improve the similarity assessment value. The adjusted instruction set refers to the set of validated instruction sets with similarity assessment values ​​less than the assessment threshold after targeted adjustment. The candidate control instruction set refers to the validated instruction set with similarity assessment values ​​greater than the assessment threshold.

[0168] Furthermore, the ROS simulation platform can be used as a simulated execution environment to simulate the execution environment of the embodied agent, thereby verifying the effectiveness of the candidate control instruction set. Pre-execution refers to simulating the candidate control instruction set in the model execution environment. The preset expected performance is based on the learning task type and the expected efficiency, such as image recognition accuracy ≥95% and robotic arm grasping success rate ≥90%. If the pre-execution performance meets the expected performance, the candidate control instruction set is determined as the control instruction set. The control instruction set is a collection of multiple control instructions that drive the embodied agent. Pre-executing the candidate control instruction set in the simulated execution environment can identify potential problems and errors in advance, avoiding the execution of incorrect instructions in the actual environment and reducing the risk of system crashes or task failures.

[0169] S7. Drive the embodied intelligent agent using a control instruction set, and collect multiple execution result data of the embodied intelligent agent using a perception driving unit.

[0170] It should be explained that the process of using a control instruction set to drive the embodied intelligent agent and using a perception-driven unit to collect multiple execution result data of the embodied intelligent agent includes:

[0171] Data is collected when driving the embodied intelligent agent using a perception driving unit and a preset acquisition frequency to obtain multiple raw execution data.

[0172] Each of the multiple raw execution data is preprocessed to obtain multiple preprocessed execution data.

[0173] Data cleaning is performed on each of the multiple preprocessed execution data sets to obtain multiple accurate execution data sets.

[0174] Feature extraction is performed on each of the multiple accurate execution data sets to obtain multiple execution result data sets.

[0175] It should be explained that the preset acquisition frequency is a fixed frequency set according to the type of learning task. For example, in an image recognition task, it can be set to acquire 30 frames per second, and in a robotic arm control task, it can be set to acquire 100 frames per second. Multiple raw execution data are the raw data generated when the embodied intelligent agent executes the control command set, including sensor data and actuator feedback data. Sensor data includes visual sensor data, distance sensor data, and environmental sensor data. Visual sensor data comes from visual sensors such as cameras and can be two-dimensional or three-dimensional image data, containing environmental scene details and object appearance features, for example, for environmental visual information in image recognition or visual navigation tasks. Distance sensor data comes from LiDAR, millimeter-wave radar, or ultrasonic sensors, presented in the form of point clouds or distance values, providing distance and shape information of environmental objects, used for tasks such as ranging, obstacle avoidance, and spatial perception. Environmental sensor data comes from temperature and humidity sensors, barometric pressure sensors, and light sensors, providing physical state information of the environment, such as temperature, humidity, barometric pressure, and light intensity, used for environmental perception. The preset acquisition frequency ensures the continuity and integrity of the data, avoiding data loss or incomplete acquisition.

[0176] Furthermore, the actuator feedback data includes position feedback data and velocity feedback data. Position feedback data, derived from position sensors, provides the embodied agent's current position information, such as the joint angles of the robotic arm or the displacement of the moving platform, for precise control and trajectory planning. Velocity feedback data, calculated from the position sensors, reflects the embodied agent's current speed and is used for speed control and motion coordination.

[0177] It should be explained that data preprocessing refers to the initial processing of the raw execution data, such as removing invalid data and filling in missing data. Preprocessed execution data is the data after this initial processing, improving data quality and usability. Optionally, Trifacta can be used to preprocess the raw execution data to identify and filter out raw execution data that does not conform to the format or has obvious errors, resulting in preprocessed execution data. Data cleaning further removes noise and outliers from the preprocessed execution data. Accurate execution data is the high-quality data after data cleaning. Optionally, OpenRefine can be used for data cleaning to improve the accuracy and reliability of the data by removing or correcting erroneous data, missing values, and outliers, resulting in accurate execution data. Feature extraction refers to extracting core features, such as image features or motion features, from the accurate execution data. Optionally, a convolutional neural network can be used as the feature extraction method. Core features refer to attributes or patterns in the data that have significant distinguishability and representativeness, effectively representing the essence and main information of the data. For example, in image data, core features can be edges, textures, shapes, colors, etc.; in motion data, core features can be position, velocity, acceleration, angular velocity, etc. Execution result data is data obtained by extracting features from a single accurate execution dataset, used for subsequent analysis and decision-making. Extracting features from multiple accurate execution datasets will yield multiple execution result datasets. By extracting key features, the dimensionality and complexity of the data are reduced, and processing speed is improved.

[0178] S8. Based on the 6G network communication link, multiple execution result data are transmitted to the data driving unit, and a detection report is generated based on the multiple execution result data in the data driving unit.

[0179] It should be explained that the process of transmitting multiple execution result data to the data-driven unit based on the 6G network communication link, and generating a detection report based on the multiple execution result data in the data-driven unit, includes:

[0180] The multiple execution result data are transmitted to the data driving unit using a 6G network communication link to obtain the multiple execution result data after transmission.

[0181] The accuracy of the data is verified by performing a data accuracy check on the multiple execution result data after transmission, and the check results are obtained. The check results include check feasibility and check infeasibility.

[0182] After confirming that the inspection results are feasible, the multiple execution result data are compared and analyzed with multiple preset qualified performance data to obtain an analysis result report;

[0183] Anomalies are diagnosed in the analysis result report using preset anomaly diagnosis rules to obtain a detection report.

[0184] Furthermore, the data accuracy verification check verifies the transmitted execution result data to ensure that the data has not been damaged or lost during transmission. Optionally, CRC check technology can be used for verification. The check result is divided into two types: if the data is complete and accurate, the check is feasible; if the data is damaged or lost, the check is not feasible. Through the data accuracy verification check, it can be confirmed whether the transmitted data is complete and intact. During data transmission, data loss or damage may occur due to network problems or storage media failures. The data accuracy verification check can promptly detect these problems and ensure data integrity. After confirming that the data accuracy verification check passes, multiple execution result data are compared and analyzed with multiple preset qualified performance data. Optionally, statistical analysis methods can be used to compare and analyze multiple execution result data with multiple preset qualified performance data.

[0185] It should be explained that the preset qualified performance data are performance standards set according to the learning task, such as image recognition accuracy ≥95% and robotic arm grasping success rate ≥90%. The comparison results form an analysis report. The analysis report includes a comparison between the actual performance indicators and the qualified performance data.

[0186] Understandably, the preset anomaly diagnosis rules are logical rules used to identify anomalies in the analysis result report. Based on these logical rules, the analysis result report is evaluated to determine if the embodied intelligent agent exhibits performance abnormalities or malfunctions. For example, if the image recognition accuracy is below 95%, an anomaly alarm is triggered; if the robotic arm's grasping success rate is repeatedly below 90%, it is determined that the robotic arm may be malfunctioning. The detection report is obtained by using the anomaly diagnosis rules to diagnose anomalies in the analysis result report.

[0187] S9. Update the prior knowledge base according to the test report to obtain the optimized prior knowledge base.

[0188] It should be explained that updating the prior knowledge base based on the detection report to obtain the optimized prior knowledge base includes:

[0189] Extract feedback data from the test report;

[0190] The baseline template is updated using the feedback data to obtain the updated baseline template;

[0191] The updated baseline template is integrated into the prior knowledge base to obtain the optimized prior knowledge base.

[0192] Furthermore, optionally, an SQL query method can be used as the method for extracting the feedback data.

[0193] Feedback data is key information extracted from the test report. This key information includes the accuracy or success rate of the embodied intelligent agent in executing control commands, execution errors, and targeted improvement suggestions.

[0194] It should be explained that updating the baseline template using feedback data refers to updating the parameter configuration and behavior patterns in the baseline template using feedback data to obtain an updated baseline template that is more in line with the actual task requirements and can provide more accurate initial parameter configurations for subsequent tasks.

[0195] Furthermore, integrating the updated benchmark template into the prior knowledge base means adding the updated benchmark template to the prior knowledge base to obtain an optimized prior knowledge base, making the knowledge system of the prior knowledge base richer and more complete, and providing stronger support for the continuous optimization and performance improvement of embodied intelligent agents.

[0196] To address the problems described in the background art, this invention confirms the receipt of adaptive learning instructions, parses the adaptive learning instructions to obtain the learning task type, and identifies the embodied intelligent agent and the 6G network communication link. The embodied intelligent agent includes a perception-driven unit, a computing-driven unit, a data-driven unit, and an interaction-driven unit. It is evident that this invention, through the high-speed and low-latency characteristics of the 6G network communication link, enables real-time and efficient data interaction between the embodied intelligent agent and the management platform, ensuring the rapid transmission and execution of task instructions. Based on this, the present invention utilizes a preset sampling time interval, a perception driving unit, and a 6G network communication link to collect visual image data and LiDAR point cloud data. It then uses the visual image data and LiDAR point cloud data to obtain external environment data including multiple environmental parameters. Using the external environment data and the learning task type, it retrieves a target benchmark template from a pre-constructed prior knowledge base. The target benchmark template is then used to initialize each of the multiple environmental parameters corresponding to the external environment data, resulting in an initialization parameter set. This initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates. Therefore, the embodiments of the present invention can ensure the timeliness and accuracy of external environment data, providing a reliable data foundation for subsequent processing. Furthermore, based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the present invention uses the optimization algorithm to optimize the initialization parameters, obtaining optimized parameter nodes. Each optimized parameter node includes optimized parameters and confidence evaluation parameters. The optimized parameter nodes are then summarized to obtain an optimized parameter node set. Thus, the embodiments of the present invention efficiently optimize the initialization parameters through optimization algorithms, generating an optimized parameter node set containing optimized parameters and confidence evaluation parameters, significantly improving the efficiency and accuracy of parameter optimization. Next, the present invention transmits the optimized parameter node set to the data-driven unit. The optimized parameter set and confidence evaluation parameter set are extracted from the optimized parameter node set in the data-driven unit. The optimized parameter set is then adjusted using the prior knowledge base and the confidence evaluation parameter set to obtain a global parameter set. This global parameter set is then transmitted to the interaction-driven unit. It is evident that this embodiment of the present invention significantly improves the accuracy and reliability of the parameters by adjusting the optimized parameters, integrating the prior knowledge base and the confidence evaluation parameter set, and generating a global parameter set. Furthermore, the present invention uses the global parameter set to generate a control instruction set, drives the embodied intelligent agent using the control instruction set, and uses the perception-driven unit to collect multiple execution result data from the embodied intelligent agent. These multiple execution result data are transmitted to the data-driven unit via a 6G network communication link. A detection report is generated based on the multiple execution result data in the data-driven unit, and the prior knowledge base is updated based on the detection report to obtain an optimized prior knowledge base. It is evident that this embodiment of the present invention achieves adaptive optimization through a feedback mechanism, updating the prior knowledge base based on the execution result data, significantly improving the accuracy and efficiency of task execution.Therefore, the present invention can improve the decision-making ability and task execution efficiency of embodied intelligent agents in the adaptive learning process.

[0197] like Figure 2 The diagram shown is a functional block diagram of an AI-driven embodied intelligent adaptive learning system provided in an embodiment of the present invention.

[0198] The AI-driven embodied intelligent adaptive learning system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the AI-driven embodied intelligent adaptive learning system 100 may include a learning task parsing module 101, a parameter optimization module 102, a parameter adjustment module 103, and an interactive feedback module 104. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0199] The learning task parsing module 101 is used to confirm the receipt of the adaptive learning instruction, parse the adaptive learning instruction, obtain the learning task type, and confirm the embodied intelligent agent and the 6G network communication link. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit.

[0200] The parameter optimization module 102 is used to collect visual image data and LiDAR point cloud data using a preset sampling time interval, a perception driving unit and a 6G network communication link, obtain external environment data including multiple environmental parameters using the visual image data and LiDAR point cloud data, retrieve a target benchmark template from a pre-built prior knowledge base using the external environment data and the learning task type, initialize each of the multiple environmental parameters corresponding to the external environment data using the target benchmark template to obtain an initialization parameter set, and pass the initialization parameter set to the calculation driving unit. The prior knowledge base includes multiple benchmark templates.

[0201] Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0202] The parameter adjustment module 103 is used to transmit the optimized parameter node set to the data driving unit, extract the optimized parameter set and confidence evaluation parameter set from the optimized parameter node set in the data driving unit, adjust the optimized parameter set using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set, and transmit the global parameter set to the interaction driving unit.

[0203] The interactive feedback module 104 is used to generate a control instruction set using a global parameter set, drive the embodied intelligent agent using the control instruction set, collect multiple execution result data of the embodied intelligent agent using a perception driving unit, transmit the multiple execution result data to the data driving unit based on a 6G network communication link, generate a detection report based on the multiple execution result data in the data driving unit, update the prior knowledge base based on the detection report, and obtain an optimized prior knowledge base.

[0204] In detail, the modules in the AI-driven embodied intelligent adaptive learning system 100 described in this embodiment of the invention employ the same methods as described above. Figure 1 The techniques used are the same as those described in the article, which are AI-driven embodied intelligence adaptive learning methods and can produce the same technical effects. Therefore, they will not be elaborated on here.

[0205] like Figure 3 The diagram shown is a structural schematic of an electronic device that implements an AI-driven embodied intelligent adaptive learning method according to an embodiment of the present invention.

[0206] The electronic device 1 may include a processor 10, a memory 11 and a bus 12, and may also include a computer program stored in the memory 11 and executable on the processor 10, such as an AI-driven embodied intelligent adaptive learning method program.

[0207] The memory 11 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device 1, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 1. Furthermore, the memory 11 includes both internal storage units and external storage devices of the electronic device 1. The memory 11 can be used not only to store application software and various types of data installed on the electronic device 1, such as the code of an AI-driven embodied intelligent adaptive learning method program, but also to temporarily store data that has been output or will be output.

[0208] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (such as AI-driven embodied intelligent adaptive learning methods) and calls data stored in the memory 11 to perform various functions of the electronic device 1 and process data.

[0209] The bus 12 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 12 can be divided into an address bus, a data bus, a control bus, etc. The bus 12 is configured to realize the connection and communication between the memory 11 and at least one processor 10, etc.

[0210] Figure 3 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 3 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0211] For example, although not shown, the electronic device 1 may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0212] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the electronic device 1 and other electronic devices.

[0213] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0214] The AI-driven embodied intelligent adaptive learning method program stored in the memory 11 of the electronic device 1 is a combination of multiple instructions, which, when run in the processor 10, can achieve the following:

[0215] The adaptive learning instruction is received and parsed to obtain the learning task type. The embodied intelligent agent and 6G network communication link are identified. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit.

[0216] Visual image data and LiDAR point cloud data are acquired using a preset sampling time interval, a perception driving unit, and a 6G network communication link. External environment data, including multiple environmental parameters, is obtained using the visual image data and LiDAR point cloud data. A target benchmark template is retrieved from a pre-built prior knowledge base using the external environment data and the learning task type. Each environmental parameter in the multiple environmental parameters corresponding to the external environment data is initialized using the target benchmark template to obtain an initialization parameter set. The initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates.

[0217] Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0218] The optimized parameter node set is transmitted to the data-driven unit;

[0219] The optimization parameter set and confidence evaluation parameter set are extracted from the optimization parameter node set of the data-driven unit. The optimization parameter set is adjusted using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set, and the global parameter set is transmitted to the interaction-driven unit.

[0220] A control instruction set is generated using a global parameter set, the control instruction set is used to drive the embodied intelligent agent, and multiple execution result data of the embodied intelligent agent are collected using a perception driving unit. The multiple execution result data are transmitted to the data driving unit based on a 6G network communication link, and a detection report is generated based on the multiple execution result data in the data driving unit.

[0221] The prior knowledge base is updated based on the test report to obtain an optimized prior knowledge base.

[0222] Specifically, the processor 10's implementation method for the above instructions can be found in [reference needed]. Figures 1 to 3 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0223] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0224] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following:

[0225] The adaptive learning instruction is received and parsed to obtain the learning task type. The embodied intelligent agent and 6G network communication link are identified. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit.

[0226] Visual image data and LiDAR point cloud data are acquired using a preset sampling time interval, a perception driving unit, and a 6G network communication link. External environment data, including multiple environmental parameters, is obtained using the visual image data and LiDAR point cloud data. A target benchmark template is retrieved from a pre-built prior knowledge base using the external environment data and the learning task type. Each environmental parameter in the multiple environmental parameters corresponding to the external environment data is initialized using the target benchmark template to obtain an initialization parameter set. The initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates.

[0227] Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set.

[0228] The optimized parameter node set is transmitted to the data-driven unit;

[0229] The optimization parameter set and confidence evaluation parameter set are extracted from the optimization parameter node set of the data-driven unit. The optimization parameter set is adjusted using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set, and the global parameter set is transmitted to the interaction-driven unit.

[0230] A control instruction set is generated using a global parameter set, the control instruction set is used to drive the embodied intelligent agent, and multiple execution result data of the embodied intelligent agent are collected using a perception driving unit. The multiple execution result data are transmitted to the data driving unit based on a 6G network communication link, and a detection report is generated based on the multiple execution result data in the data driving unit.

[0231] The prior knowledge base is updated based on the test report to obtain an optimized prior knowledge base.

[0232] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and actual implementations may have other classification methods.

[0233] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0234] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0235] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0236] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI-driven embodied intelligence adaptive learning method, characterized in that, The method includes: The adaptive learning instruction is received and parsed to obtain the learning task type. The embodied intelligent agent and 6G network communication link are identified. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit. Visual image data and LiDAR point cloud data are acquired using a preset sampling time interval, a perception driving unit, and a 6G network communication link. External environment data, including multiple environmental parameters, is obtained using the visual image data and LiDAR point cloud data. A target benchmark template is retrieved from a pre-built prior knowledge base using the external environment data and the learning task type. Each environmental parameter in the multiple environmental parameters corresponding to the external environment data is initialized using the target benchmark template to obtain an initialization parameter set. The initialization parameter set is then passed to the computation driving unit. The prior knowledge base includes multiple benchmark templates. Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set. The optimized parameter node set is transmitted to the data-driven unit; The optimization parameter set and confidence evaluation parameter set are extracted from the optimization parameter node set of the data-driven unit. The optimization parameter set is adjusted using the prior knowledge base and confidence evaluation parameter set to obtain the global parameter set, and the global parameter set is transmitted to the interaction-driven unit. A control instruction set is generated using a global parameter set, the control instruction set is used to drive the embodied intelligent agent, and multiple execution result data of the embodied intelligent agent are collected using a perception driving unit. The multiple execution result data are transmitted to the data driving unit based on a 6G network communication link, and a detection report is generated based on the multiple execution result data in the data driving unit. The prior knowledge base is updated based on the test report to obtain an optimized prior knowledge base.

2. The AI-driven embodied intelligence adaptive learning method as described in claim 1, characterized in that, The method of acquiring external environmental data, including multiple environmental parameters, using visual image data and LiDAR point cloud data includes: The perception-driving unit is adjusted based on the learning task type. The adjusted perception-driving unit is used to perform real-time enhancement processing on visual image data to obtain a preprocessed image set. Then, noise reduction and motion compensation processing are performed on LiDAR point cloud data to obtain optimized point cloud data. The preprocessed image set and optimized point cloud data are fused from multiple sources to obtain external environment data.

3. The AI-driven embodied intelligence adaptive learning method as described in claim 2, characterized in that, The target benchmark template is retrieved from a pre-built prior knowledge base using external environment data and learning task type. The target benchmark template is then used to initialize each of the multiple environmental parameters corresponding to the external environment data, resulting in an initialization parameter set, including: By analyzing environmental parameters in external environment data using a pre-verified edge computing module, the current task encoding vector, current environment feature fingerprint, and device information corresponding to the environmental parameters are obtained. The following operations are performed on each of the multiple benchmark templates: Extract the task encoding vector of the baseline template, and calculate the task type similarity using the current task encoding vector, the task encoding vector of the baseline template, and the pre-constructed task type similarity calculation formula; Extract the environmental feature fingerprint of the benchmark template, and calculate the environmental similarity using the current environmental feature fingerprint, the environmental feature fingerprint of the benchmark template, and the pre-constructed environmental similarity calculation formula; The hardware compatibility is obtained by comparing the device information with a pre-built device compatibility comparison table. The total matching score is calculated using task type similarity, environment similarity, and hardware compatibility. The calculation formula is shown below: ； in, The total match score. For task type similarity, Indicates environmental similarity. For hardware compatibility, All are weighting coefficients; The total matching degree is summarized to obtain a total matching degree set. The total matching degree in the total matching degree set is sorted in descending order to obtain a total matching degree sequence. Five candidate benchmark templates are identified in the total matching degree sequence. The five candidate benchmark templates are the benchmark templates corresponding to the first five total matching degrees in the total matching degree sequence. Each of the five candidate benchmark templates is input into a pre-built digital twin platform for verification to obtain five verification parameters. The candidate benchmark template corresponding to the verification parameter with the largest value among the five verification parameters is taken as the target benchmark template. The environmental parameters are initialized using the verification parameters corresponding to the target benchmark template to obtain the initialization parameters. The initialization parameters are then summarized to obtain the initialization parameter set.

4. The AI-driven embodied intelligence adaptive learning method as described in claim 3, characterized in that, The optimization algorithm for obtaining initialization parameters based on the initialization parameter set uses the optimization algorithm to optimize and calculate the initialization parameters to obtain optimized parameter nodes, including: The optimization algorithm is obtained by using the initialization parameters in the initialization parameter set and the pre-built algorithm mapping table; The optimization algorithm is used to perform hierarchical optimization calculations on the initialization parameters to obtain optimized parameters for multiple edge nodes; A fixed number of gradient descent steps are performed on the optimization parameters of the multiple edge nodes to obtain multiple preliminary optimization parameters; The digital twin platform was used to verify and adjust several preliminary optimization parameters, resulting in several optimized parameters. The confidence evaluation parameters are obtained by using a pre-built confidence evaluation method to evaluate each of the multiple optimization parameters. The optimization parameters and their corresponding confidence evaluation parameters are integrated to generate optimization parameter nodes.

5. The AI-driven embodied intelligence adaptive learning method as described in claim 4, characterized in that, The process of adjusting the optimization parameter set using a prior knowledge base and a confidence evaluation parameter set to obtain a global parameter set includes: Based on the confidence assessment parameters in the confidence assessment parameter set, the optimization parameters in the optimization parameter set are divided into three types of partitioning parameters, namely high confidence parameters, medium confidence parameters, and low confidence parameters. The low confidence parameters are reinitialized using the prior knowledge base to obtain the corrected parameters. The confidence parameter is fine-tuned using a prior knowledge base to obtain the fine-tuning parameter; A global optimization strategy is obtained based on the aforementioned correction parameters, fine-tuning parameters, high-confidence parameters, and prior knowledge base. The global optimization strategy is used to integrate the correction parameters, fine-tuning parameters, and high-confidence parameters to obtain a global parameter set.

6. The AI-driven embodied intelligence adaptive learning method as described in claim 5, characterized in that, The generation of control instruction sets using a global parameter set includes: A preliminary control instruction set is generated based on the global parameter set; The interactive driving unit is used to perform syntax and semantic verification on each preliminary control instruction in the preliminary control instruction set. If the syntax verification of the initial control instruction fails, the initial control instruction is revised according to the pre-built syntax rules to obtain a syntax revision instruction; If the syntax verification of the preliminary control instruction passes, then the preliminary control instruction is considered a syntax-qualified instruction. If the semantic verification of the initial control instruction is unclear, the initial control instruction is semantically optimized according to the pre-built task execution rules to obtain the semantically optimized instruction; If the semantic verification of the preliminary control instruction is clear, then the preliminary control instruction shall be deemed a semantically qualified instruction. The syntax-qualified instructions, semantic-qualified instructions, syntax-correction instructions, and semantic-optimization instructions are integrated to obtain the verified instruction set; Similarity comparison is performed using the verified instruction set, the pre-confirmed historical successful instruction set, and a pre-built structural similarity algorithm to obtain a similarity evaluation value; If the similarity evaluation value is less than the preset evaluation threshold, the verified instruction set is adjusted based on the learning task type to obtain the adjusted instruction set. The adjusted instruction set is used as the verified instruction set, and the step of performing similarity comparison using the verified instruction set, the pre-confirmed historical successful instruction set, and the pre-built structural similarity algorithm to obtain a similarity evaluation value is repeated until the similarity evaluation value is greater than or equal to the evaluation threshold. If the similarity evaluation value is greater than or equal to the evaluation threshold, the verified instruction set will be used as the candidate control instruction set. The candidate control instruction set is pre-executed using a pre-built simulation execution environment. Once the effect of the pre-execution is confirmed to be the preset expected effect, the candidate control instruction set is determined as the control instruction set.

7. The AI-driven embodied intelligence adaptive learning method as described in claim 6, characterized in that, The process of driving the embodied intelligent agent using a control instruction set and collecting multiple execution result data of the embodied intelligent agent using a perception-driven unit includes: Data is collected when driving the embodied intelligent agent using a perception driving unit and a preset acquisition frequency to obtain multiple raw execution data. Each of the multiple raw execution data is preprocessed to obtain multiple preprocessed execution data. Data cleaning is performed on each of the multiple preprocessed execution data sets to obtain multiple accurate execution data sets. Feature extraction is performed on each of the multiple accurate execution data sets to obtain multiple execution result data sets.

8. The AI-driven embodied intelligence adaptive learning method as described in claim 7, characterized in that, The process of transmitting multiple execution result data to the data-driven unit via the 6G network communication link, and generating a detection report based on the multiple execution result data in the data-driven unit, includes: The multiple execution result data are transmitted to the data driving unit using a 6G network communication link to obtain the multiple execution result data after transmission. The accuracy of the data is verified by performing a data accuracy check on the multiple execution result data after transmission, and the check results are obtained. The check results include check feasibility and check infeasibility. After confirming that the inspection results are feasible, the multiple execution result data are compared and analyzed with multiple preset qualified performance data to obtain an analysis result report; Anomalies are diagnosed in the analysis result report using preset anomaly diagnosis rules to obtain a detection report.

9. The AI-driven embodied intelligence adaptive learning method as described in claim 8, characterized in that, The step of updating the prior knowledge base based on the test report to obtain an optimized prior knowledge base includes: Extract feedback data from the test report; The baseline template is updated using the feedback data to obtain the updated baseline template; The updated baseline template is integrated into the prior knowledge base to obtain the optimized prior knowledge base.

10. An AI-driven embodied intelligent adaptive learning system, characterized in that, The system includes: The learning task parsing module is used to confirm the receipt of adaptive learning instructions, parse the adaptive learning instructions, obtain the learning task type, and identify the embodied intelligent agent and the 6G network communication link. The embodied intelligent agent includes: a perception driving unit, a computing driving unit, a data driving unit, and an interaction driving unit. The parameter optimization module is used to collect visual image data and LiDAR point cloud data using a preset sampling time interval, a perception driving unit, and a 6G network communication link; to obtain external environment data including multiple environmental parameters using the visual image data and LiDAR point cloud data; to retrieve a target benchmark template from a pre-built prior knowledge base using the external environment data and the learning task type; to initialize each of the multiple environmental parameters corresponding to the external environment data using the target benchmark template, thereby obtaining an initialization parameter set; and to pass the initialization parameter set to the computation driving unit. The prior knowledge base includes multiple benchmark templates. Based on the initialization parameter acquisition optimization algorithm in the initialization parameter set, the initialization parameters are optimized and calculated using the optimization algorithm to obtain optimization parameter nodes. The optimization parameter nodes include optimization parameters and confidence evaluation parameters. The optimization parameter nodes are summarized to obtain an optimization parameter node set. The parameter adjustment module is used to transmit the set of optimized parameter nodes to the data-driven unit, extract the set of optimized parameters and the set of confidence evaluation parameters from the set of optimized parameter nodes in the data-driven unit, adjust the parameters of the set of optimized parameters using the prior knowledge base and the set of confidence evaluation parameters to obtain the global parameter set, and transmit the global parameter set to the interaction-driven unit. The interactive feedback module is used to generate a control instruction set using a global parameter set, drive the embodied intelligent agent using the control instruction set, collect multiple execution result data of the embodied intelligent agent using a perception driving unit, transmit multiple execution result data to a data driving unit based on a 6G network communication link, generate a detection report based on multiple execution result data in the data driving unit, update the prior knowledge base based on the detection report, and obtain an optimized prior knowledge base.

Images