Task processing method and device applied to humanoid large model agent

Abstract

The invention provides a task processing method and device applied to a human-like large model agent. The method comprises the following steps: acquiring an instruction and preference information of a user; performing intention recognition and machine understanding according to the instruction and the preference information, reasoning the instruction based on long and short term memory, and decomposing to obtain a plurality of sub-tasks corresponding to the instruction; calling the plug-in tool set to execute the subtask, and performing exception handling and dynamic adjustment based on environment feedback; and aggregating the execution results of the sub-tasks, generating a final response for the instruction, and sending the final response to the user terminal. According to the task processing method applied to the humanoid large model agent, decomposition, logical reasoning and multi-step action plan making of a complex target can be achieved, and the universality and robustness of the agent in a complex scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a task processing method and apparatus for humanoid large-scale intelligent agents. Background Technology

[0002] With the rapid evolution of large-scale model technology, LLM-based agents are becoming an important path to achieving artificial general intelligence (AGI). Human-like large-scale model agents not only possess powerful language understanding and generation capabilities, but also can complete complex, multi-step autonomous decision-making and actions through mechanisms such as environmental perception, task planning, tool invocation, memory of experience, and self-reflection. The overall architecture of a human-like large-scale model agent typically includes core components such as perception, memory, planning, action, and reflection. These modules must work collaboratively to support reasoning and interaction in dynamic environments.

[0003] However, existing human-like large-scale intelligent agent systems still face significant challenges in areas such as module decoupling, state consistency, and long-term memory management. Specifically:

[0004] High module coupling: The boundaries between functional modules (such as perception, memory, planning, and action) are blurred, and they are highly interdependent, resulting in low adaptability in complex scenarios.

[0005] State consistency is difficult to guarantee: In complex environments, the internal state of an agent is prone to inconsistency with the external environment, affecting the accuracy of decision-making and the reliability of actions.

[0006] These challenges constrain the large-scale deployment and continuous evolution of human-like large-model intelligent agents in real-world, complex environments. Therefore, there is an urgent need for a large-model intelligent agent architecture that balances versatility and robustness. Summary of the Invention

[0007] This invention provides a task processing method and apparatus for humanoid large model intelligent agents, which can solve the defects of existing humanoid large model intelligent agent systems in terms of module decoupling and state consistency.

[0008] This invention provides a task processing method applied to a human-like large-scale intelligent agent, comprising:

[0009] Obtain user commands and preference information;

[0010] Intent recognition and machine understanding are performed based on the instructions and preference information. Reasoning is performed on the instructions based on long short-term memory, and multiple sub-tasks corresponding to the instructions are decomposed.

[0011] The subtasks are executed by calling a plug-in toolset, and exception handling and dynamic adjustments are performed based on environmental feedback.

[0012] The execution results of each of the subtasks are aggregated to generate a final response to the instruction, and the final response is sent to the user terminal.

[0013] According to the present invention, a task processing method applied to a human-like large-scale intelligent agent is provided, wherein the instruction is reasoned based on long short-term memory to decompose the instruction into multiple sub-tasks corresponding to the instruction, including:

[0014] Short-term memory is used to cache dialogue context information, and semantic retrieval technology is used to retrieve relevant knowledge, historical experience and preference information corresponding to the instruction from the long-term memory system.

[0015] Search results are generated based on the relevant knowledge, historical experience, and preference information.

[0016] The subtask corresponding to the instruction is obtained based on the search results.

[0017] According to the present invention, a task processing method applied to a humanoid large model intelligent agent, after decomposing the instruction into multiple sub-tasks, further includes:

[0018] Based on the execution results of the subtasks or changes in the environment, the execution order or execution strategy of subsequent subtasks is dynamically adjusted.

[0019] According to the present invention, a task processing method applied to a humanoid large model intelligent agent, wherein the subtask is executed by calling a plug-in toolset, including:

[0020] The subtask is executed by calling at least one of the following tools through a standardized application programming interface or unified communication protocol: vision / perception analysis tool, IoT / device control tool, database query tool, robot motion control tool, script and sandbox execution tool, and voice and multimedia interaction tool.

[0021] According to the present invention, a task processing method applied to a humanoid large model intelligent agent is provided, wherein the decomposition to obtain multiple sub-tasks corresponding to the instruction and the invocation of a pluggable toolkit to execute the sub-tasks are jointly completed by multiple intelligent agents through a division of labor and negotiation mechanism.

[0022] According to a task processing method for a humanoid large model intelligent agent provided by the present invention, after sending the final response to the user terminal, the method further includes:

[0023] Obtain user feedback on the final response, and update and / or optimize relevant model parameters in the long-term memory based on the feedback.

[0024] According to a task processing method for a humanoid large model intelligent agent provided by the present invention, the process of calling a plug-in toolset to execute the subtask further includes:

[0025] The execution process of the sub-task is subject to service quality monitoring and security and trust control; wherein the security and trust control includes at least one of operation compliance verification, dynamic permission review, and operation behavior audit and traceability.

[0026] The present invention also provides a task processing device for a humanoid large model intelligent agent, comprising:

[0027] The instruction receiving module is used to acquire user instructions;

[0028] The task decomposition module performs intent recognition and machine understanding based on the instruction and the preference information, and infers the instruction based on long short-term memory to decompose it into multiple sub-tasks corresponding to the instruction.

[0029] The task execution module is used to call the plug-in toolset to execute the sub-tasks and to perform exception handling and dynamic adjustments based on environmental feedback;

[0030] The response feedback module is used to aggregate the execution results of each of the subtasks, generate a final response to the instruction, and send the final response to the user terminal.

[0031] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the task processing method applied to a humanoid large model intelligent agent as described above.

[0032] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the task processing method applied to a humanoid large model intelligent agent as described above.

[0033] The present invention provides a task processing method and apparatus for humanoid large-scale intelligent agents. This method acquires user instructions and preference information; performs intent recognition and machine understanding based on the instructions and preference information; reasons about the instructions based on long short-term memory, decomposing them into multiple sub-tasks corresponding to the instructions; executes the sub-tasks by calling a plug-in toolset, and performs anomaly handling and dynamic adjustments based on environmental feedback; aggregates the execution results of each sub-task, generates a final response to the instructions, and sends the final response to the user terminal. The task processing method for humanoid large-scale intelligent agents provided by this invention can achieve the decomposition of complex goals, logical reasoning, and multi-step action plan formulation, significantly improving the versatility and robustness of the intelligent agent in complex scenarios. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the task processing method for humanoid large-scale intelligent agents provided by the present invention;

[0036] Figure 2 This is a schematic diagram of an optional human-like large model intelligent agent architecture in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the task processing device for humanoid large model intelligent agents provided by the present invention;

[0038] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0040] The following is combined Figure 1 and Figure 2 This invention describes a task processing method applied to a large humanoid intelligent agent.

[0041] The task processing method of the present invention applied to humanoid large-scale intelligent agents is mainly aimed at humanoid large-scale intelligent agent architectures, such as... Figure 2 This invention demonstrates an optional human-like large-scale intelligent agent architecture, which mainly includes a task planning layer, a memory reasoning layer, a tool usage layer, and a model management layer. The method provided by this invention can decompose complex goals, perform logical reasoning, and formulate multi-step action plans. It can also dynamically adjust strategies based on environmental feedback, significantly improving the agent's versatility and robustness in complex scenarios. Furthermore, it enhances the agent's scalability through a systematic, modular, and highly scalable architecture. Figure 1 As shown, the task processing method for humanoid large-scale intelligent agents provided by this invention includes the following steps:

[0042] S1. Obtain user instructions and preference information.

[0043] Specifically, this invention supports acquiring multimodal, multi-terminal user commands, including but not limited to commands from "human users" or "robot terminal devices." For example, a user can say to the smart home hub via a mobile app, "Adjust the living room temperature to a comfortable level and play some relaxing music." This is a composite voice command that includes both environmental control and entertainment needs.

[0044] S2. Based on instructions and preference information, perform intent recognition and machine understanding, reason about instructions based on long short-term memory, and decompose the instructions into multiple sub-tasks.

[0045] This step corresponds to Figure 2The task planning layer and memory reasoning layer in the system aim to decompose complex goals into a sequence of executable subtasks using "short-term and long-term memory" and "big-small model coordinated reasoning." In this step, after intent recognition and machine understanding (text / visual / multimodal semantic understanding) based on instructions and user preference information, a clear task objective (i.e., the overall task) is formulated. This objective is then further decomposed into numerous subtasks by integrating short-term and long-term memory, machine learning reasoning, and big-small model coordinated reasoning. For example, taking the user instruction "adjust the living room temperature to a comfortable level and play some relaxing music" as an example, the task planning layer understands that "comfortable" likely corresponds to 24 degrees Celsius (combined with user preference memory) and decomposes the instruction into: Subtask 1 - Query the current living room temperature; Subtask 2 - Adjust the air conditioner to 24 degrees Celsius; Subtask 3 - Retrieve the user's favorite "relaxing music" playlist; Subtask 4 - Control the speakers to play the playlist. These subtasks call the task execution planning module to perform multi-step planning, including reasoning, learning, multi-step execution, fault tolerance, memory, and feedback learning. This primarily involves breaking down complex tasks into single, simplified subtasks, enabling the decomposition of complex goals, logical reasoning, and multi-step action plan formulation, as well as dynamic strategy adjustment based on environmental feedback. It transforms ambiguous user instructions into explicit, structured, and executable operational plans. This is key to achieving "autonomous decision-making" and handling complex tasks, solving the agent's "what to do" and "what to do first and what to do later" problems. Its ability to "dynamically adjust strategies" directly enhances the system's robustness and adaptability.

[0046] This involves reasoning about instructions based on long short-term memory, decomposing the instructions into multiple sub-tasks, including:

[0047] Short-term memory is used to cache dialogue context information, and semantic retrieval technology is used to retrieve relevant knowledge, historical experience and preference information corresponding to the instructions from the long-term memory system.

[0048] Retrieval results are generated based on relevant knowledge, historical experience, and preference information;

[0049] The subtasks corresponding to the instructions are obtained based on the search results.

[0050] Specifically, the method provided by this invention forms short-term memory through the historical context of the current session or task, performs semantic retrieval by combining it with a vector database in long-term memory, recalls relevant knowledge, historical experience or user preference information to form retrieval results, and injects them into prompts and reasoning basis to enhance decision-making accuracy and meet the personalized requirements of the current user.

[0051] In an optional embodiment of the present invention, after decomposing the instructions into multiple subtasks, it is also necessary to dynamically adjust the execution order or execution strategy of subsequent subtasks based on the execution results of the subtasks or changes in the environment. Taking the above-mentioned subtask sequence of "Subtask 1 - Query the current temperature of the living room; Subtask 2 - Adjust the air conditioner to 24 degrees Celsius; Subtask 3 - Retrieve the user's favorite 'relaxing music' playlist; Subtask 4 - Control the speakers to play the playlist" as an example, assuming that when executing Subtask 3 - retrieving the user's favorite "relaxing music" playlist, the user's favorite "relaxing music" playlist cannot be found, then when executing sub-step 4, the execution strategy is dynamically adjusted to "Subtask 4 - Control the speakers to randomly play an existing pure music playlist".

[0052] S3. Call the plug-in toolset to execute subtasks and perform exception handling and dynamic adjustments based on environmental feedback.

[0053] The corresponding step Figure 2 The tool usage layer in the system implements a "perception-execution" closed loop. For planned sub-tasks, it performs multi-step execution based on the task content, including but not limited to calling short-term memory data, long-term memory data, internal and external tools, and large and small models for inference. Short-term memory refers to the context cache maintained by the agent during a single task execution, containing the input, output, and intermediate states of the current session. It directly serves as the prompt input for large model inference and is limited by the context window length. Long-term memory refers to a retrievable knowledge base stored in an external vector database. It is used to persistently store experience, facts, and user data across sessions and injects them into the inference process on demand through semantic retrieval, enabling knowledge reuse and personalization.

[0054] Specifically, through standardized application programming interfaces (APIs / SDKs, etc.) or unified communication protocols, the agent autonomously selects and invokes suitable tools to execute subtasks. These tools include, but are not limited to, at least one of the following: vision / perception analysis tools, IoT / device control tools, database query tools, robot motion control tools, script and sandbox execution tools, and voice and multimedia interaction tools. Transforming planned subtasks into actual actions is the "executor" of the agent's capabilities. The plug-in design of the tool layer solves the problems of "module decoupling" and "scalability," enabling the agent to easily integrate new capabilities.

[0055] Using the same subtask sequence as an example: "Subtask 1 - Query the current living room temperature; Subtask 2 - Adjust the air conditioner to 24 degrees Celsius; Subtask 3 - Retrieve the user's favorite 'relaxing music' playlist; Subtask 4 - Control the speakers to play the playlist," let's assume that during the execution of Subtask 2, the intelligent agent calls the smart air conditioner's API to set the temperature through the "IoT / Device Control" module in the tool layer. If the air conditioner is offline, causing the call to fail, the fault tolerance module will capture the exception, trigger a retry, or check the device status through the "Environmental Awareness" module and report "Task execution failed" feedback.

[0056] In an optional embodiment of the present invention, the steps of "decomposing the instructions into multiple sub-tasks" in step S2 and "calling the plug-in toolset to execute the sub-tasks" in step S3 are jointly completed by multiple agents through a division of labor and negotiation mechanism. By adopting a multi-agent cooperation mechanism, the agent architecture can flexibly adapt to various complex task scenarios from the digital space to the physical world.

[0057] In an optional embodiment of the present invention, the process of invoking the plug-in toolset to execute subtasks further includes:

[0058] The execution process of subtasks is subject to service quality monitoring and security and trust control. Security and trust control includes at least one of the following: operational compliance verification, dynamic permission review, and operational behavior auditing and traceability. When a subtask fails, the humanoid big data model intelligent agent architecture provides environmental awareness, feedback learning modules, and fault tolerance modules to resolve anomalies and ensure successful task re-execution. Service quality monitoring and security and trust control are crucial for the "scalable deployment" of intelligent agents in enterprise applications. Service quality monitoring encompasses service quality assurance throughout the entire execution lifecycle of the intelligent agent, including performance, reliability, response speed, resource utilization (CPU / GPU / memory / disk I / O, etc.), throughput, stability, emergency scheduling, fault tolerance, and recovery, ensuring that the intelligent agent "not only does it right, but also does it well and quickly." Security and trustworthiness control is a core capability system that ensures compliant, reliable, explainable, and auditable behavior. It encompasses compliance control (such as a built-in compliance rule engine, approval requirements for sensitive operations, and dynamic verification of regional / permission policies), access control (dynamic permission allocation based on RBAC / ABAC, and manual confirmation for temporary permissions), and auditable and traceable operations (such as when an agent ID calls which tools / data / interfaces, the environmental state at the time, and log records). When high-risk behaviors are identified, timely alerts and interventions are provided. Simultaneously, automatic anonymization and encrypted data transmission and storage are required for data privacy and security. Security and trustworthiness control addresses enterprises' core requirements for "compliant, reliable, explainable, and auditable behavior," ensuring the system's robustness and controllability. For example, when an agent attempts to perform a high-risk operation such as "deleting database logs," the "compliance rule engine" of the security and trustworthiness management module will intercept the request, trigger an "approval required" process, and record the operation attempt in the audit log.

[0059] S4. Aggregate the execution results of each subtask, generate the final response to the instruction, and send the final response to the user terminal.

[0060] In this step, the results of all subtasks are aggregated into the final task implementation result, i.e., the final response to the instruction, which is fed back to the terminal device and presented to the user. In an optional embodiment of the present invention, various integration methods can be adopted, including but not limited to RESTful API interfaces, gRPC, multi-language SDKs (python / C++ / Java / node.js, etc.), and unified control protocol MCP, reflecting the "compatibility" and "advanced nature" of the architecture, ensuring that the capabilities of the intelligent agent can be seamlessly integrated into existing enterprise IT systems and terminal devices. Taking the above subtask sequence of "Subtask 1 - Query the current temperature of the living room; Subtask 2 - Adjust the air conditioner to 24 degrees Celsius; Subtask 3 - Retrieve the user's favorite 'relaxing music' playlist; Subtask 4 - Control the speakers to play the playlist" as an example, the execution results of all subtasks ("Temperature has been set to 24℃", "Playing the 'Afternoon Relaxation' playlist") are aggregated into a natural language response or structured data, which is returned to the mobile app via API and presented to the user by the app.

[0061] In an optional embodiment of the present invention, after sending the final response to the user terminal in step S4, the method further includes:

[0062] S5. Obtain user feedback on the final response, and update the long-term memory and / or optimize the relevant model parameters in the agent based on the feedback.

[0063] This step corresponds to Figure 2 The model management layer contains a secure and controllable local model of a specific size, enabling version control, canary releases, hot updates, and A / B testing. This facilitates refined management of the model iteration process, ensuring the stability and traceability of the production environment and fully supporting enterprises in building a deployable, operational, and evolving private intelligent agent application system. This step, based on online learning and continuous optimization mechanisms, utilizes user feedback (explicit ratings or implicit behaviors) and machine learning to call semi-automatic / fully automatic modules to update long-term memory or dynamically fine-tune the model, optimizing the model's inference process. This makes the intelligent agent more intelligent, achieving "continuous learning and iteration," transforming it from a static system into an organism capable of continuous evolution through interaction. For example, regarding the execution results of the subtasks mentioned above ("Temperature set to 24℃", "Playing 'Afternoon Relaxation' playlist"), assuming the user feedback is "Too cold," the system will capture this feedback and associate it with the current context (instruction "Adjust to Comfort", execution result 24℃). Through machine learning modules, it may be possible to explore the unknown mechanism by which "comfort" is closer to 26°C for this user, and store this knowledge in long-term memory, or directly fine-tune the relevant parameters of the local model.

[0064] The following example, using a specific smart home control scenario, illustrates the task processing method provided by this invention for humanoid large-scale intelligent agents:

[0065] Suppose a user issues command A through a smart speaker: "Make the house more comfortable," accompanied by user preference information B (historical data shows that "comfortable" as defined by users at night typically includes "temperature 24℃" and "playing soft music"). The processing flow for command A and preference information B is as follows:

[0066] 1. Command accepted:

[0067] The system receives instruction A, "Make your home more comfortable," along with associated user preference information B.

[0068] 2. Task Planning:

[0069] Intent Recognition and Understanding: The core of the task planning module's understanding instruction A is "to improve home comfort".

[0070] Memory reasoning: Short-term working memory caches the current conversation context. The long-term memory system retrieves user preference information B ("comfortable" at night = temperature 24℃ + soft music) and environmental information such as the current season and time through semantic retrieval.

[0071] Task breakdown: Based on the above reasoning, the task is broken down into ordered subtasks: ① Check the current temperature and humidity in the living room; ② If the temperature is uncomfortable, adjust the air conditioner to 24℃; ③ Check and play the "light music" playlist.

[0072] 3. Task Execution:

[0073] The task execution module calls plug-in tools via a standardized API:

[0074] Call the IoT device query tool to execute subtask ① and obtain the current temperature as 28℃.

[0075] Call the air conditioning control tool to execute subtask ② and set the temperature to 24℃.

[0076] The multimedia playback tool is invoked to execute subtask ③, which starts playing music.

[0077] Exception handling: If the air conditioner is found to be offline while executing subtask ②, the fault tolerance module will capture the exception and attempt to reconnect or notify the user via speaker that "the air conditioner seems to be offline, please check".

[0078] 4. Results Feedback: The results feedback module aggregates the results of each subtask and generates a natural language response: "The living room temperature has been adjusted from 28℃ to 24℃ and light music has started playing," which is then broadcast through the speaker.

[0079] 5. Learning Optimization:

[0080] User A reports: "It's a bit cold." The learning optimization module collects this negative feedback and triggers analysis: the user feels cold with the current temperature set to 24℃. This module may use machine learning to analyze the new knowledge that "User A's comfortable temperature under the current specific conditions (such as a rainy day) may be higher than 24℃" and store it as a new experience record in the long-term memory system. Alternatively, it may fine-tune the local small model so that its understanding of "comfortable" in similar environments in the future is more inclined to 26℃.

[0081] In summary, the task processing method for humanoid large-scale intelligent agents provided by this invention has the following beneficial effects:

[0082] Task planning and dynamic decision-making capabilities: It can decompose complex goals, perform logical reasoning and formulate multi-step action plans, and dynamically adjust strategies based on environmental feedback, significantly improving the autonomy and adaptability of intelligent agents in complex scenarios.

[0083] Modularization and decoupling: By clearly dividing the layers of task planning, memory reasoning, tool use, and learning optimization, and defining standardized interfaces, the system achieves high cohesion and low coupling, improving scalability and maintainability.

[0084] State consistency and decision-making ability: Through the synergy of short-term working memory and long-term memory systems, the contextual coherence and knowledge continuity of the reasoning process are ensured.

[0085] Efficient long-term memory and continuous learning: Efficient management of long-term memory is achieved by utilizing vector databases and semantic retrieval technology, and the agent is endowed with the ability to continuously learn and evolve through a user feedback-driven optimization mechanism.

[0086] Secure and controllable with enterprise-level reliability: Built-in service quality monitoring and security and trust control mechanisms throughout the entire task execution process ensure that the intelligent agent's behavior is compliant, reliable, and auditable, meeting the high requirements of enterprise-level applications for stability, security, and traceability.

[0087] Flexibility and versatility: The plug-in tool integration approach and multi-agent collaboration mechanism support agents seamlessly calling various external tools, translating language understanding capabilities into actual business operations. This allows the agent architecture to flexibly adapt to a variety of complex task scenarios, from the digital space to the physical world. It supports multiple role-based agents working collaboratively through division of labor, negotiation, and communication, and can flexibly combine professional capabilities (such as analysis and visualization) to efficiently complete comprehensive tasks that are difficult for a single agent to handle.

[0088] Based on the same inventive concept, the present invention also provides a task processing device for humanoid large model intelligent agents. The task processing device for humanoid large model intelligent agents provided by the present invention is described below. The task processing device for humanoid large model intelligent agents described below and the task processing method for humanoid large model intelligent agents described above can be referred to in correspondence.

[0089] like Figure 3 As shown, the task processing device for humanoid large model intelligent agents provided by the present invention includes an instruction receiving module 31, a task decomposition module 32, a task execution module 33, and a response feedback module 34.

[0090] The instruction receiving module 31 is used to obtain user instructions.

[0091] The task decomposition module 32 performs intent recognition and machine understanding based on the instruction and the preference information, and infers the instruction based on long short-term memory to decompose it into multiple sub-tasks corresponding to the instruction.

[0092] The task execution module 33 is used to call the plug-in toolset to execute the sub-tasks and to perform exception handling and dynamic adjustments based on environmental feedback.

[0093] The response feedback module 34 is used to aggregate the execution results of each of the subtasks, generate a final response to the instruction, and send the final response to the user terminal.

[0094] In an optional embodiment of the present invention, the task processing device applied to the humanoid large model intelligent agent further includes an iterative optimization module, which is used to obtain user feedback on the final response and update the long-term memory and / or optimize the relevant model parameters in the intelligent agent based on the feedback.

[0095] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute the task processing methods provided by the above methods for humanoid large-scale intelligent agents, which include:

[0096] Obtain user commands and preference information;

[0097] Intent recognition and machine understanding are performed based on the instructions and preference information. Reasoning is performed on the instructions based on long short-term memory, and multiple sub-tasks corresponding to the instructions are decomposed.

[0098] The subtasks are executed by calling a plug-in toolset, and exception handling and dynamic adjustments are performed based on environmental feedback.

[0099] The execution results of each of the subtasks are aggregated to generate a final response to the instruction, and the final response is sent to the user terminal.

[0100] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the task processing method for a humanoid large model intelligent agent provided by the above methods, the method comprising:

[0102] Obtain user commands and preference information;

[0103] Intent recognition and machine understanding are performed based on the instructions and preference information. Reasoning is performed on the instructions based on long short-term memory, and multiple sub-tasks corresponding to the instructions are decomposed.

[0104] The subtasks are executed by calling a plug-in toolset, and exception handling and dynamic adjustments are performed based on environmental feedback.

[0105] The execution results of each of the subtasks are aggregated to generate a final response to the instruction, and the final response is sent to the user terminal.

[0106] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the task processing methods for humanoid large model intelligent agents provided by the above methods, the method comprising:

[0107] Obtain user commands and preference information;

[0108] Intent recognition and machine understanding are performed based on the instructions and preference information. Reasoning is performed on the instructions based on long short-term memory, and multiple sub-tasks corresponding to the instructions are decomposed.

[0109] The subtasks are executed by calling a plug-in toolset, and exception handling and dynamic adjustments are performed based on environmental feedback.

[0110] The execution results of each of the subtasks are aggregated to generate a final response to the instruction, and the final response is sent to the user terminal.

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units 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. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A task processing method applied to a human-like large-scale intelligent agent, characterized in that, include: Obtain user commands and preference information; Intent recognition and machine understanding are performed based on the instructions and preference information. Reasoning is performed on the instructions based on long short-term memory, and multiple sub-tasks corresponding to the instructions are decomposed. The subtasks are executed by calling a plug-in toolset, and exception handling and dynamic adjustments are performed based on environmental feedback. The execution results of each of the subtasks are aggregated to generate a final response to the instruction, and the final response is sent to the user terminal.

2. The task processing method applied to a humanoid large-scale intelligent agent according to claim 1, characterized in that, The reasoning based on long short-term memory decomposes the instruction into multiple sub-tasks corresponding to the instruction, including: Short-term memory is used to cache dialogue context information, and semantic retrieval technology is used to retrieve relevant knowledge, historical experience and preference information corresponding to the instruction from the long-term memory system. Search results are generated based on the relevant knowledge, historical experience, and preference information. The subtask corresponding to the instruction is obtained based on the search results.

3. The task processing method applied to a humanoid large-scale intelligent agent according to claim 1, characterized in that, After decomposing the instruction into multiple subtasks, the process also includes: Based on the execution results of the subtasks or changes in the environment, the execution order or execution strategy of subsequent subtasks is dynamically adjusted.

4. The task processing method applied to a humanoid large-scale intelligent agent according to claim 1, characterized in that, Invoking the pluggable toolset to execute the subtasks includes: The subtask is executed by calling at least one of the following tools through a standardized application programming interface or unified communication protocol: vision / perception analysis tool, IoT / device control tool, database query tool, robot motion control tool, script and sandbox execution tool, and voice and multimedia interaction tool.

5. The task processing method applied to a humanoid large-scale intelligent agent according to claim 1, characterized in that, The decomposition yields multiple subtasks corresponding to the instruction, and the invocation of a pluggable toolset to execute the subtasks is completed jointly by multiple intelligent agents through a division of labor and negotiation mechanism.

6. The task processing method applied to a humanoid large-scale intelligent agent according to claim 1, characterized in that, After sending the final response to the user terminal, the process also includes: Obtain user feedback on the final response, and update and / or optimize relevant model parameters in the long-term memory based on the feedback.

7. The task processing method applied to a humanoid large-scale intelligent agent according to any one of claims 1 to 6, characterized in that, The process of invoking the pluggable toolset to execute the subtask also includes: The execution process of the sub-task is subject to service quality monitoring and security and trust control; wherein the security and trust control includes at least one of operation compliance verification, dynamic permission review, and operation behavior audit and traceability.

8. A task processing device applied to a humanoid large-scale intelligent agent, characterized in that, include: The instruction receiving module is used to acquire user instructions; The task decomposition module performs intent recognition and machine understanding based on the instruction and the preference information, and infers the instruction based on long short-term memory to decompose it into multiple sub-tasks corresponding to the instruction. The task execution module is used to call the plug-in toolset to execute the sub-tasks and to perform exception handling and dynamic adjustments based on environmental feedback; The response feedback module is used to aggregate the execution results of each of the subtasks, generate a final response to the instruction, and send the final response to the user terminal.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the task processing method for humanoid large model intelligent agents as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the task processing method for humanoid large model intelligent agents as described in any one of claims 1 to 7.

Images