Intelligent task planning method based on large model

Abstract

The invention provides an intelligent task planning method based on a large model, and relates to the technical field of unmanned systems. Input information is generated by obtaining user requirements and system perception data; generating original task content according to the input information in combination with the large model; performing task decomposition on the original task content to generate a basic task set; acquiring basic information of the current execution unit, and generating a task allocation strategy in combination with the basic task set; and calling a target interface corresponding to the execution unit according to the task allocation strategy. By combining task network thinking and large model reasoning ability, an automatic closed-loop process of finding and solving the problems of the unmanned system is realized, and the technical effect of improving the rationality of intelligent planning of the unmanned system is realized.

Description

Technical Field

[0001] The present application relates to the technical field of unmanned systems, and in particular to an intelligent task planning method based on a large model. Background Art

[0002] With the development of generative artificial intelligence, large language models (LLM) and related technologies are becoming the key force in promoting the advancement of general artificial intelligence technology. At present, a series of large models represented by GPT-4o, GLM4, etc. have played an important role in various industries. In this context, the intelligent task planning technology of unmanned systems will also usher in new changes. Large models can expand the autonomous exploration, planning and action capabilities of unmanned systems in complex open environments, and improve generalization and transferability.

[0003] Unmanned systems and their intelligent task planning technologies have been widely used in modern military, aerospace, logistics warehouses and other fields, but their intelligence and automation still have the following two major problems. The first is the problem of generalization and transferability. The design process of unmanned systems and their planning subsystems usually considers individual scenarios and a small number of fixed tasks, with customized models, algorithms and processes, resulting in weak generalization. The second is the problem of broken links in the four processes of problem discovery, task creation, task planning and task execution of unmanned systems. Problem discovery and task creation are often completed by experts before and during the construction of unmanned systems. It is a manual process, while task allocation, scheduling, path planning or action planning in task planning are solved through planning models and intelligent algorithms. Specifically, engineers need to preset a large number of subtask execution modules and program and compile them to achieve task execution. From problem discovery to problem solving (task completion), it has experienced repeated alternations of "manual → computer → man → computer", resulting in low fault tolerance for process incoherence.

[0004] Mission planning for unmanned systems is a complex and multi-layered process that requires consideration of not only the capabilities and limitations of individual devices, but also the coordination between different devices and the optimal allocation of overall tasks. Lack of generalization means that a large amount of reconstruction is required for different scenarios, which consumes a lot of manpower and material resources. Incoherent task loops will reduce the usability of unmanned systems and fail to effectively promote unmanned and intelligent systems.

[0005] Therefore, how to improve the rationality of intelligent planning of unmanned systems has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] In order to achieve the technical effect of improving the rationality of intelligent planning of unmanned systems, the present application provides an intelligent task planning method based on a large model.

[0007] In the first aspect, the present application provides an intelligent task planning method based on a large model using the following technical solutions:

[0008] An intelligent task planning method based on a large model, comprising:

[0009] Obtain user needs and system perception data to generate input information;

[0010] Generate original task content based on the input information and the large model;

[0011] Decomposing the original task content to generate a basic task set;

[0012] Obtaining basic information of the current execution unit, and combining the basic task set to generate a task allocation strategy;

[0013] The target interface corresponding to the execution unit is called according to the task allocation strategy.

[0014] Optionally, the step of obtaining user requirements and system perception data to generate input information includes:

[0015] Obtain user input information and generate user intent recognition results through intent recognition;

[0016] Acquiring system perception data, wherein the system perception data includes external environment data and internal state data;

[0017] Generate a potential demand identification result according to the external environment data and the internal state data;

[0018] Input information is generated according to the user intention identification result and the potential demand identification result.

[0019] Optionally, the step of generating original task content according to the input information in combination with the large model includes:

[0020] Get the task understanding template in the big model;

[0021] The input information is sent to the task understanding template so that the task understanding template creates the original task content T Raw

[0022] T Raw =M Comprehension (Situation,Instruction)

[0023] Where M represents the large model parameter, M Comprehension Indicates that the task understanding template was adopted.

[0024] Optionally, the step of performing task decomposition on the original task content to generate a basic task set includes:

[0025] Get the target prompt tool in the big model to build M Decompose ;

[0026] By constructing M Decompose For the original task content T Raw Perform decomposition to generate composite tasks;

[0027] Decompose the load task to generate a basic task set

[0028]

[0029] in, express Can be E i Execution, the basic task is to be executed by an execution unit E i ∈E performs the task, and the task decomposition is performed at the lth level.

[0030] Optionally, the step of obtaining basic information of the current execution unit and combining the basic task set to generate a task allocation strategy includes:

[0031] Invoke the retrieval enhancement generation process in the large model;

[0032] Obtaining basic information of the current execution unit, wherein the basic information includes execution unit capability information and execution unit status information;

[0033] Determining capability constraints according to the execution unit capability information;

[0034] Determining a state constraint according to the execution unit state information;

[0035] generating an execution platform according to the capability constraint and the state constraint;

[0036] The basic task set and the execution platform are passed as input to the search enhancement generation process to obtain a task allocation strategy.

[0037] Optionally, the step of calling the target interface corresponding to the execution unit according to the task allocation strategy includes:

[0038] The function of the target interface corresponding to the execution unit is denoted as M Drive ;

[0039] Input the task allocation strategy A into the function of the target interface to obtain the output as the code set C and execution logic L of the driving execution unit

[0040] M Drive (A) = (C, L);

[0041] Among them, code set C is the interface calling function generated by the large model, that is,

[0042]

[0043] in It is the execution unit E i Execute B p The code exists only if Capability(B p ,E i )=1;

[0044] The target interface corresponding to the execution unit is called according to the execution logic L.

[0045] Optionally, after the step of calling the target interface corresponding to the execution unit according to the task allocation strategy, the method further includes:

[0046] Acquire execution environment information and execution result information through preset sensor devices;

[0047] Generate an evaluation report based on the execution result information, the execution environment information and the original task content;

[0048] Determine whether the task allocation strategy needs to be optimized based on the evaluation report.

[0049] In a second aspect, the present application provides an intelligent task planning system based on a large model, and the intelligent task planning system based on a large model includes:

[0050] An input information module, used to obtain user requirements and system perception data to generate input information;

[0051] An original task content module, used to generate original task content according to the input information combined with the large model;

[0052] A basic task set module, used for performing task decomposition on the original task content to generate a basic task set;

[0053] A task allocation strategy module is used to obtain basic information of the current execution unit and generate a task allocation strategy in combination with the basic task set;

[0054] The target interface module is used to call the target interface corresponding to the execution unit according to the task allocation strategy.

[0055] In a third aspect, the present application provides a large-model-based intelligent task planning device, the device comprising: a memory and a processor, the processor executing the method described above when running the large-model-based intelligent task planning program stored in the memory.

[0056] In a fourth aspect, the present application provides a computer-readable storage medium, characterized in that it includes instructions, which, when executed on a computer, enable the computer to execute the method described above.

[0057] In summary, the present application includes the following beneficial technical effects:

[0058] This application generates input information by acquiring user needs and system perception data; generates original task content based on the input information combined with a large model; decomposes the original task content to generate a basic task set; acquires basic information of the current execution unit, combines it with the basic task set to generate a task allocation strategy; and calls the target interface corresponding to the execution unit according to the task allocation strategy. By combining task network thinking and large model reasoning capabilities, an automated closed-loop process from problem discovery to problem solving for unmanned systems is realized, achieving the technical effect of improving the rationality of intelligent planning for unmanned systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the structure of an intelligent task planning device based on a large model in the hardware operating environment involved in the embodiment of the present application;

[0060] Figure 2 It is a flowchart of the first embodiment of the intelligent task planning method based on a large model of the present application;

[0061] Figure 3 This is a schematic diagram of the comparison effect between the traditional planning technology and GTN in the first embodiment of the intelligent task planning method based on a large model of the present application;

[0062] Figure 4 It is a schematic diagram of the actual operation process of the first embodiment of the intelligent task planning method based on a large model of the present application;

[0063] Figure 5 This is a schematic diagram of the application architecture of the execution entity GTN of the first embodiment of the intelligent task planning method based on a large model of the present application;

[0064] Figure 6 This is a schematic diagram of the implementation path of the execution subject GTN of the first embodiment of the intelligent task planning method based on a large model of the present application;

[0065] Figure 7 It is a structural block diagram of the first embodiment of the intelligent task planning device based on a large model of the present application. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below through the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0067] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a large-model-based intelligent task planning device in the hardware operating environment involved in the embodiment of the present application.

[0068] like Figure 1 As shown, the intelligent task planning device based on the large model may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0069] Those skilled in the art will understand that Figure 1 The structure shown in does not constitute a limitation on the large model-based intelligent task planning device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0070] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and an intelligent task planning program based on a large model.

[0071] exist Figure 1In the big model-based intelligent task planning device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the present application can be set in the big model-based intelligent task planning device, and the big model-based intelligent task planning device calls the big model-based intelligent task planning program stored in the memory 1005 through the processor 1001, and executes the big model-based intelligent task planning method provided in the embodiment of the present application.

[0072] The present application embodiment provides an intelligent task planning method based on a large model, referring to Figure 2 , Figure 2 This is a flow chart of the first embodiment of the large model-based intelligent task planning method of this application.

[0073] In this embodiment, the intelligent task planning method based on the large model includes the following steps:

[0074] Step S10: Obtain user requirements and system perception data to generate input information.

[0075] In the specific implementation, in order to execute the intelligent task planning method based on the large model, the execution subject of this embodiment is the Generative Task Network (GTN).

[0076] It can be understood that in this embodiment, GTN is a new unmanned system driving paradigm that integrates intelligent task planning and automatic execution built for unmanned systems. Figure 3The comparison between traditional planning technology and GTN shown in the figure summarizes the complete process of intelligent planning and execution of unmanned systems, and compares and analyzes the differences between the traditional intelligent planning technology stack and the GTN technology stack in each link. The traditional method mainly relies on people to analyze the situation, raise questions, and model tasks, and use computers and algorithms to plan, solve, and schedule tasks, and then people program to drive the unmanned system to perform tasks. The GTN method is mainly based on large models, which can autonomously create, automatically decompose, and allocate tasks based on situation data, and automatically program and compile to execute tasks. Therefore, compared with traditional intelligent planning technology, GTN has the advantage of "integration of task generation, planning, and execution". The control of unmanned systems is an integrated and coherent process, not just a pre-defined task sequence. At the same time, in the GTN process, people and large models can achieve decoupled collaboration, which will be reflected in the GTN process design below. For example, the user input can be empty, and the large model only needs to receive the current situation data to execute the complete GTN process. If the user has special needs, he can also intervene in the GTN process to input the needs to the large model. GTN can still analyze its needs and make task planning and execution processes. In addition, the running results of each process in GTN depend on the prompt words of the prompt project. The application scenario migration and generalization of the unmanned system will become very quick and easy, and only a small modification of the prompt template is required.

[0077] It should be noted that the steps of obtaining user needs and system perception data to generate input information include: obtaining user input information and generating user intention recognition results through intention recognition; obtaining system perception data, which includes external environment data and internal state data; generating potential demand recognition results based on external environment data and internal state data; and generating input information based on user intention recognition results and potential demand recognition results.

[0078] GTN draws on the structure of hierarchical task networks and uses large models to effectively decompose and transform complex initial tasks into executable interface operation tasks. The core of GTN lies in its ability to intelligently parse user instructions and situation data results perceived by the system, and then build a dynamic task planning and execution network. In this network, tasks are broken down into subtasks of different levels and types, and are completed by various execution units.

[0079] It should be noted that the execution entity GTN of this embodiment includes the following key elements:

[0080] Definition 1: Execution Platform

[0081] The execution platform refers to the hardware and software infrastructure that provides the execution unit operating environment. It can support the flexible deployment, monitoring, task planning and scheduling of multiple heterogeneous execution units, denoted as E = {E1 ,E 2 ,E 3 ,…,E m}, where m is the total number of execution units, E i Represents the i-th execution unit of the execution platform.

[0082] Definition 2: Executant

[0083] An execution unit is a functional unit used to complete basic tasks, which can be a drone, an unmanned vehicle, or a ground computer. Each execution unit has a set of common attributes. To facilitate understanding, viewing, and calling, these attributes are summarized into basic attributes, functional attributes, and state attributes. It can be expressed as E i ={A i,base ,A i,func ,A i,state}, where A i,base = {b i1 ,b i2 ,b i3 ,…} represents execution unit E i The basic attribute set of A, similarly, its functional attribute set can be further expressed as A i,func ={f i1 ,f i2 ,f i3 ,…}, and its state attribute set is represented by A i,state ={s i1 ,s i2 ,s i3 ,…}. b in these three attribute sets ij 、f ij and ij Represents execution unit E i The j-th basic attribute, functional attribute and state attribute of the execution unit. The basic attributes are some immutable descriptive parameters of the execution unit. Each functional attribute has its corresponding calling interface. The state attributes have been connected to the current situation information for real-time acquisition and update.

[0084] Definition 3: Raw Task

[0085] The initial task refers to the task created by the large model based on the user input instructions and the situation data results perceived by the system. The initial task content is recorded as T Raw ={D 1 ,D 2 ,D 3 ,D 4}. The task background D 1 Provide background information of the task, including the reason and purpose of the task; task objectives 2Clarify the ultimate goal of the task and the criteria for successful completion; resource requirements D 3 Indicates the knowledge, execution units and their interfaces that may be used to complete the task; Risk analysis D 4 Indicates the degree of danger to people and objects caused by completing the task. Tasks with too high risk will be stopped. You can add more task descriptions as needed.

[0086] Define 4 sub-tasks

[0087] A subtask is a smaller task decomposed from an initial task. It may be included in the initial task or a more detailed task. A subtask is an intermediate result in the task decomposition process. It can be a basic task or a composite task.

[0088] Definition 5 Compound Task

[0089] A composite task is a task that can be further decomposed into multiple subtasks. It can be divided into a basic task or a composite task of the next level. As the intermediate result of the task network, the composite task undertakes the upper-level task while also providing more detailed information for the next-level task decomposition. The task decomposition process of the first level is recorded as

[0090]

[0091] in refers to the qth subtask in the decomposition result of the lth layer, and Q is the total number of subtasks decomposed by the current composite task decomposition layer l. Note that when When it is a basic task, Q = 1. For the remaining complex tasks after decomposition at this level, the decomposition continues to the next level, that is,

[0092]

[0093] Definition 6 Basic Task

[0094] A basic task refers to a task that has been decomposed to the finest granularity in GTN. It can be directly executed by an execution unit or interface tool without further decomposition.

[0095] remember Indicates that the initial task is decomposed into only basic tasks, where B p is the pth basic task, P is the completion of the initial task T Raw The total number of basic tasks required. Each basic task B p It can be associated with an execution unit, indicating that the execution unit is used to execute the basic task. This association can be represented by a mapping relationship:

[0096]

[0097] The basic task ensures that the final result of task decomposition can be understood and executed by the specific execution unit, so that the task can be completed effectively.

[0098] In the specific implementation, the actual operation process of the execution subject of this embodiment is as follows Figure 4 As shown in the figure, the operation process of GTN mainly takes system perception and user requirements as input, and uses the big model to perform autonomous task creation, task decomposition, task allocation and task execution. At the same time, the big model is also responsible for checking and updating the output content, perceiving and evaluating the execution results, etc.

[0099] Step S20: Generate original task content based on input information and the large model.

[0100] It should be noted that the step of generating the original task content according to the input information combined with the large model includes: obtaining the task understanding template in the large model; sending the input information to the task understanding template so that the task understanding template creates the original task content T Raw

[0101] T Raw =M Comprehension (Situation,Instruction)

[0102] Where M represents the large model parameter, M Comprehension Indicates that the task understanding template was adopted.

[0103] It should be noted that the GTN system in this embodiment can not only better respond to the explicit needs of users, but also predict and meet the implicit needs of users according to changes in the system environment, thereby realizing personalized and intelligent task management.

[0104] Step S30: Decompose the original task content into tasks to generate a basic task set.

[0105] It should be noted that the steps of decomposing the original task content to generate a basic task set include: obtaining a target prompt tool in the large model to construct an M Decompose ; By constructing M Decompose For the original task content T Raw Decompose to generate composite tasks; decompose load tasks to generate a set of basic tasks

[0106]

[0107] in, express Can be E iExecution, the basic task is to be executed by an execution unit E i ∈E performs the task, and the task decomposition is performed at the lth level.

[0108] It should be noted that the task decomposition process of GTN draws on the technical ideas of traditional hierarchical task networks to divide and conquer complex problems. Raw After the self-creation is completed, the M built using the big model and its CoT, ToT and other prompting technologies Decompose T Raw Reasonably decompose into multiple levels of subtasks, and decompose them multiple times until all are decomposed into basic tasks

[0109] Step S40: Obtain basic information of the current execution unit, and combine it with the basic task set to generate a task allocation strategy.

[0110] It should be noted that the steps of obtaining the basic information of the current execution unit and combining it with the basic task set to generate a task allocation strategy include: calling the retrieval enhancement generation process in the large model; obtaining the basic information of the current execution unit, the basic information including the execution unit capability information and the execution unit status information; determining the capability constraints based on the execution unit capability information; determining the status constraints based on the execution unit status information; generating an execution platform based on the capability constraints and the status constraints; and passing the basic task set and the execution platform as input to the retrieval enhancement generation process to obtain the task allocation strategy.

[0111] It is understandable that a successful interface call not only requires the correct calling of interface functions and parameter passing, but also requires consideration of how to arrange the calling logic and order of different interfaces to ensure the correct execution of tasks. This can be achieved simultaneously in the process of generating code for the large model, so the arrangement and call are included as a process node of GTN. Interface arrangement and call refers to generating arranged interface call code according to the task allocation plan, and driving each execution unit to complete the assigned basic tasks according to the correct logic and order.

[0112] Step S50: calling the target interface corresponding to the execution unit according to the task allocation strategy.

[0113] It should be noted that the step of calling the target interface corresponding to the execution unit according to the task allocation strategy includes: recording the function of the target interface corresponding to the execution unit as M Drive ; Input the task allocation strategy A into the function of the target interface to obtain the output as the code set C and execution logic L of the driving execution unit

[0114] M Drive (A) = (C, L);

[0115] Among them, code set C is the interface calling function generated by the large model, that is,

[0116]

[0117] in It is the execution unit E i Execute B p The code exists only if Capability(B p ,E i )=1; call the target interface corresponding to the execution unit according to the execution logic L.

[0118] In the specific implementation, the execution logic L defines the logical order and dependencies of the execution of each basic task to ensure that the tasks are performed as expected.

[0119] L = {l 1 ,l 2 ,l 3 ,…}

[0120] Among them l j The essence of this is a reorganization of the basic tasks. j Each of them consists of one or more basic tasks' execution code In the same j Indicates that these tasks are executed in parallel, and l j+1 Yes j The next serial interface arrangement and calling code set.

[0121] It should be noted that after the step of calling the target interface corresponding to the execution unit according to the task allocation strategy, it also includes: obtaining execution environment information and execution result information through a preset sensor device; generating an evaluation report based on the execution result information, execution environment information and original task content; and judging whether the task allocation strategy needs to be optimized based on the evaluation report.

[0122] In specific implementation, perception and evaluation refers to the system using sensors and other information collection tools to monitor environmental changes and task execution during the execution process, and evaluate the completion and quality of the task. This step is crucial for the closed-loop task management system, which provides feedback information for the continuous improvement and optimization of the system.

[0123] The perception and evaluation function can be recorded as M Evaluate , it collects the current environment data and system status information Situation, the execution result Situation' and the initial task T Raw As input, the output is an evaluation report Report on the task execution.

[0124] M Evaluate(Situation,Situation',T Raw )→Report

[0125] The evaluation report includes evaluations of multiple dimensions, including whether the task is completed, the quality of completion, resource utilization efficiency, environmental adaptability, etc. This information will be used to guide the creation, decomposition, allocation and driving process of subsequent tasks to achieve continuous optimization of the system.

[0126] It can be understood that the application architecture of the execution entity GTN of this embodiment is as follows Figure 5 As shown, including:

[0127] The perception layer is the front end of GTN, responsible for collecting and understanding sensor data from the external environment and internal hardware. The modules of the perception layer include:

[0128] (1) Internal hardware devices and sensors: The hardware devices and their sensors collect situation data from the external environment and the status information of the system, ensuring that the system has real-time understanding of the external and internal status. The hardware devices are also the execution units used by the system to perform tasks.

[0129] (2) Situation processing model: Converts the collected data into situation information and provides input to the knowledge layer for further mission planning and decision-making.

[0130] The knowledge layer is the core of the architecture that enables GTN to operate efficiently. It is responsible for storing, managing, and updating various professional knowledge. The modules of the knowledge layer include:

[0131] (1) Expert domain knowledge: Contains expert knowledge in a specific task domain, which is used to guide the generation and driving of domain tasks.

[0132] (2) Hardware control interface knowledge: Instructions and input and output parameters related to hardware control and driving, ensuring that basic tasks can be correctly executed by the hardware.

[0133] (3) Algorithm interface knowledge: Contains the interface information required for algorithm execution, which is used to select the appropriate algorithm and configure its parameters.

[0134] (4) Model interface knowledge: This refers to the interface information related to model calling, which is used to select and configure the model required for task execution.

[0135] (5) Situation knowledge: It comes from the perception layer and is processed by the situation processing model. It is used to understand, store, and predict possible situations during task planning and execution.

[0136] The retrieval layer is the core module of RAG technology. It is designed for situations where the amount of data is huge and it is not convenient to input all of it into a large model for question-answering reasoning. It can reduce information redundancy and reduce large model hallucinations.

[0137] The retrieval layer uses text segmentation, word embedding and vector retrieval technology to quickly find knowledge related to the current task to enhance the accuracy and success rate of task planning and driving. The modules of the retrieval layer include:

[0138] (1) Integrated knowledge base: stores and manages all relevant knowledge, including domain knowledge, hardware control interface knowledge, etc.

[0139] (2) Knowledge segmentation module: divides a large knowledge base into smaller pieces that are easier to manage and retrieve.

[0140] (3) Knowledge embedding module: responsible for converting knowledge into Embeddings and storing it in the vector database.

[0141] (4) Vector database: stores and manages knowledge in the form of embeddings and supports fast retrieval.

[0142] (5) Ranking model: Sort the knowledge according to the relevance of the retrieval results and select the most relevant knowledge for task generation.

[0143] The generation layer is the output layer of this application architecture. It is responsible for generating specific tasks and execution plans based on the retrieved knowledge. It mainly uses large language models and their prompt engineering technology to generate task autonomous creation, task decomposition, task allocation, interface arrangement and calling according to task requirements. The modules of the generation layer include:

[0144] (1) Large model: A large language model is used to generate and drive tasks. If the large model is a closed-source model, the API is directly called. If it is an open-source model, in order to further enhance its task planning and execution capabilities for unmanned systems, it needs to be optimized using methods such as fine-tuning or preference alignment.

[0145] (2) Prompt technology: Design and optimize the prompt words of the large model, such as using CoT and ToT to guide the large model to complete the task planning and execution process of the unmanned system. In addition, to enhance the effect of GTN, a chain messaging (CM) method is also used. This method optimizes the historical content of the dialogue with the large model. In the GTN process, only the necessary upper-level process text is retained to the next level, rather than retaining all of it, to improve the accuracy of question and answer and the response speed.

[0146] (3)RAG: The relevant knowledge obtained from the retrieval layer is provided to the designed GTN process nodes and combined with the prompt words in the nodes, thus avoiding the input of a large amount of useless information (relative to the current process node) into the large model and enhancing the generation effect.

[0147] The execution entity GTN of this embodiment realizes the full process automation from environment perception to task generation, task planning, interface arrangement and calling through the close cooperation of these four levels. Each level plays a key role to ensure that the task network can respond to complex and dynamic needs efficiently and accurately.

[0148] It should be noted that if Figure 6 In the implementation path shown in the figure, GTN organizes the unmanned system interface and existing external knowledge and model tools through prompt engineering to jointly complete the entire GTN process, including autonomous task creation, task decomposition, task allocation, interface arrangement and call, verification and execution, perception and evaluation. The multi-agent architecture is adopted in the prompt engineering. Each process node is a new dialogue independently constructed with flexible input, and the chain message passing method is used to selectively retain the upstream node dialogue information and pass it to the next process node, rather than passing all historical dialogues. This chain message passing design method helps to improve the performance of large models, reduce the waste of dialogue tokens, and the portability and generalization performance of application scenarios. The current situation data, API tools, and Cases in the figure are all from the knowledge integration database used for the RAG process. After the RAG process, the search results are delivered to the prompt of the GTN process node. Among them, due to safety and ethical considerations, after the interface arrangement and call are completed, human users need to intervene in the verification and execution steps to manually verify its safety, compliance, and feasibility. Only after passing the verification can the task code be compiled and executed.

[0149] The prompt template of the GTN process node in the implementation path has "{}" content, which represents the flexible input part in the prompt template, which can be adjusted according to the application scenario and task. This is the specific implementation form of CM.

[0150] This embodiment generates input information by acquiring user needs and system perception data; generates original task content based on the input information combined with the large model; decomposes the original task content to generate a basic task set; acquires the basic information of the current execution unit, combines the basic task set to generate a task allocation strategy; and calls the target interface corresponding to the execution unit according to the task allocation strategy. By combining the task network thinking and the large model reasoning ability, an automated closed-loop process from problem discovery to problem solving of the unmanned system is realized, achieving the technical effect of improving the rationality of the intelligent planning of the unmanned system.

[0151] In addition, an embodiment of the present application also proposes a computer-readable storage medium, on which a program for intelligent task planning based on a large model is stored. When the program for intelligent task planning based on a large model is executed by a processor, the steps of the method for intelligent task planning based on a large model as described above are implemented.

[0152] Reference Figure 7 , Figure 7 This is a structural block diagram of the first embodiment of the intelligent task planning system based on a large model of this application.

[0153] like Figure 7 As shown, the intelligent task planning system based on a large model proposed in the embodiment of the present application includes:

[0154] An input information module 10, used to obtain user requirements and system perception data to generate input information;

[0155] The original task content module 20 is used to generate the original task content according to the input information combined with the large model;

[0156] A basic task set module 30 is used to perform task decomposition on the original task content to generate a basic task set;

[0157] The task allocation strategy module 40 is used to obtain the basic information of the current execution unit and generate a task allocation strategy in combination with the basic task set;

[0158] The target interface module 50 is used to call the target interface corresponding to the execution unit according to the task allocation strategy.

[0159] It should be understood that the above is only an example and does not constitute any limitation on the technical solution of the present application. In specific applications, technicians in this field can make settings as needed, and the present application does not impose any limitation on this.

[0160] This embodiment generates input information by acquiring user needs and system perception data; generates original task content based on the input information combined with the large model; decomposes the original task content to generate a basic task set; acquires the basic information of the current execution unit, combines the basic task set to generate a task allocation strategy; and calls the target interface corresponding to the execution unit according to the task allocation strategy. By combining the task network thinking and the large model reasoning ability, an automated closed-loop process from problem discovery to problem solving of the unmanned system is realized, achieving the technical effect of improving the rationality of the intelligent planning of the unmanned system.

[0161] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present application. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0162] In addition, for technical details not described in detail in this embodiment, please refer to the large model-based intelligent task planning method provided in any embodiment of the present application, which will not be repeated here.

[0163] In addition, it should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0164] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0165] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory (ROM) / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0166] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An intelligent task planning method based on a large model, characterized in that: include: Obtain user needs and system perception data to generate input information; Generate original task content based on the input information and the large model; Decomposing the original task content to generate a basic task set; Obtaining basic information of the current execution unit, and combining the basic task set to generate a task allocation strategy; The target interface corresponding to the execution unit is called according to the task allocation strategy.

2. The large model-based intelligent task planning method according to claim 1, characterized in that: The step of obtaining user requirements and system perception data to generate input information includes: Obtain user input information and generate user intent recognition results through intent recognition; Acquiring system perception data, wherein the system perception data includes external environment data and internal state data; Generate a potential demand identification result according to the external environment data and the internal state data; Input information is generated according to the user intention identification result and the potential demand identification result.

3. The large model-based intelligent task planning method according to claim 1, characterized in that: The step of generating the original task content according to the input information combined with the large model includes: Get the task understanding template in the big model; The input information is sent to the task understanding template so that the task understanding template creates the original task content T Raw T Raw =M Comprehension (Situation,Instruction) Where M represents the large model parameter, M Comprehension Indicates that the task understanding template was adopted.

4. The large model-based intelligent task planning method according to claim 1, characterized in that: The step of performing task decomposition on the original task content to generate a basic task set includes: Get the target prompt tool in the big model to build M Decompose ; By constructing M Decompose For the original task content T Raw Perform decomposition to generate composite tasks; Decompose the load task to generate a basic task set in, express Can be E i Execution, the basic task is to be executed by an execution unit E i ∈E performs the task, and the task decomposition is performed at the lth level.

5. The large model-based intelligent task planning method according to claim 1, characterized in that: The step of obtaining basic information of the current execution unit and combining the basic task set to generate a task allocation strategy includes: Invoke the retrieval enhancement generation process in the large model; Obtaining basic information of the current execution unit, wherein the basic information includes execution unit capability information and execution unit status information; Determining capability constraints according to the execution unit capability information; Determining a state constraint according to the execution unit state information; generating an execution platform according to the capability constraint and the state constraint; The basic task set and the execution platform are passed as input to the search enhancement generation process to obtain a task allocation strategy.

6. The large model-based intelligent task planning method according to claim 1 is characterized in that: The step of calling the target interface corresponding to the execution unit according to the task allocation strategy includes: The function of the target interface corresponding to the execution unit is denoted as M Drive ; Input the task allocation strategy A into the function of the target interface to obtain the output as the code set C and execution logic L of the driving execution unit M Drive (A)=(C,L); Among them, code set C is the interface calling function generated by the large model, that is, in It is the execution unit E i Execute B p The code exists only if Capability(B p ,E i )=1; The target interface corresponding to the execution unit is called according to the execution logic L.

7. The large model-based intelligent task planning method according to claim 1 is characterized in that: After the step of calling the target interface corresponding to the execution unit according to the task allocation strategy, the method further includes: Acquire execution environment information and execution result information through preset sensor devices; Generate an evaluation report based on the execution result information, the execution environment information and the original task content; Determine whether the task allocation strategy needs to be optimized based on the evaluation report.

8. An intelligent task planning system based on a large model, characterized in that: The large model-based intelligent task planning system includes: An input information module, used to obtain user requirements and system perception data to generate input information; An original task content module, used to generate original task content according to the input information combined with the large model; A basic task set module, used for performing task decomposition on the original task content to generate a basic task set; A task allocation strategy module is used to obtain basic information of the current execution unit and generate a task allocation strategy in combination with the basic task set; The target interface module is used to call the target interface corresponding to the execution unit according to the task allocation strategy.

9. An intelligent task planning device based on a large model, characterized in that: The device comprises: a memory and a processor, and the processor executes the method according to any one of claims 1 to 7 when running the large model-based intelligent task planning program stored in the memory.

10. A computer-readable storage medium, characterized in that: The method comprises instructions, which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 7.

Images