Body-equipped intelligent robot task decomposition system and method based on physical state deduction

Abstract

The invention discloses a task decomposition system and method for an intelligent robot with a body based on physical state deduction, and relates to the technical field of artificial intelligence robots. The system comprises a task instruction receiving and initial decomposition module, a state sensing and formatting input module, a task feasibility deduction module, a feedback and optimization closed loop module and a decision and output module. According to the method, a technical framework of sensing, deducing and optimizing a closed loop is constructed, so that the robot can perform virtual rehearsal on a task sequence before executing actual physical actions, thereby identifying environment state conflicts and execution path risks in advance, and fundamentally improving the task execution mode from passive response depending on trial and error, and improving the task execution efficiency. The method is converted into active reliable planning based on physical state deduction, and the adaptability, the safety and the deployment efficiency of the robot for executing complex tasks in an open environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence robot technology, and more specifically to an embodied intelligent robot task decomposition system and method based on physical state deduction. Background Technology

[0002] With the development of embodied intelligence and large language model (LLM) technologies, using LLM to understand and decompose high-level natural language instructions to generate task sequences that robots can execute has become a cutting-edge direction in robot autonomy. However, reliably applying the powerful semantic understanding capabilities of LLM to robots in the physical world still faces fundamental challenges.

[0003] The existing technical solutions mainly have the following problems and shortcomings: 1. The "Physical Illusion" Problem of LLM: Existing methods generally rely directly on LLM for task decomposition. LLM is essentially a language model, and its training data and inference process severely lack an embedded understanding of fundamental laws of the physical world, such as spatial geometry, object dynamics, and resource exclusivity. This leads to a list of sub-tasks that is logically sound but physically infeasible. For example, LLM might plan steps that violate common sense, such as "unscrewing a bolt without tools" or "opening a refrigerator door while holding an object with both arms." This "physical illusion" makes the decomposition results unusable for driving real robots.

[0004] 2. Lack of verification mechanism in open-loop decomposition: Current mainstream methods mostly adopt an open-loop model of "decompose once, execute directly." The system lacks an automated, physically-based feasibility verification step for the decomposition results output by LLM. Its correctness highly depends on the randomness of the prompt word engineering and LLM itself, and its reliability cannot be guaranteed. Once a physical contradiction occurs during decomposition, it will only be exposed during the actual robot execution stage, which may lead to task failure, equipment damage, or safety accidents, resulting in high risks and high debugging costs.

[0005] 3. Lack of Structured Representation and Reasoning Methods for Physical States: To enable machines to verify the physical feasibility of tasks, it is essential to provide a structured description of the robot's and environment's states and to model how actions change these states. Existing methods lack a reasonable physical state representation and reasoning method aligned with the robot's underlying control model. Current methods typically rely on unstructured natural language generated by large models to describe states, such as "the robot is near the table." This semantically ambiguous description cannot be linked with precise geometric and physical models such as robot kinematics models and environmental maps for computation. Consequently, the system cannot automatically and rigorously deduce and verify the geometric feasibility (e.g., whether it is reachable, whether there will be collisions) and physical feasibility (e.g., whether it is overweight, whether it is stable) of actions such as "moving to a certain place" or "grabbing something."

[0006] Therefore, how to construct a task representation that can be understood by LLM and support underlying physical automatic reasoning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide a physical state deduction-based embodied intelligent robot task decomposition system and method that overcomes or at least partially solves the above problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a task decomposition system for embodied intelligent robots based on physical state deduction, comprising: The task instruction receiving and initial decomposition module is used to establish an interaction channel between the user and the system, and to convert the user's natural language task requirements into a list of subtasks that the system can process. The state perception and formatted input module is used to acquire information about the robot's own state and the external environment state, and generate a current structured state description; The task feasibility simulation module is used to perform step-by-step physical state simulation based on the received sub-task list and the current structured state description to verify its physical feasibility. The feedback and optimization closed-loop module generates feedback information containing conflict reasons and modification suggestions when the task feasibility deduction module fails, based on the task feasibility deduction module, and drives the task instruction receiving and initial decomposition module to re-optimize the sub-task list. The decision and output module is used to format, manually confirm, and reuse the subtask list, while also connecting to the robot execution end.

[0009] Furthermore, the task instruction receiving and initial decomposition module includes: The human-computer interaction unit is used to provide a natural language or graphical interface and receive task instructions input by the user; The task decomposition unit is used to fill the task instructions and the current structured state description into the corresponding prompt word template, and call LLM to generate a list of subtasks that conforms to a preset format.

[0010] Furthermore, the state-aware and formatted input module includes: The robot state perception unit is used to read data from the robot's internal sensors in real time. The environmental perception unit collects images of the current environment through the camera device mounted on the robot body and calls the Visual Language Model (VLM) to generate a preliminary description.

[0011] Furthermore, the task feasibility simulation module includes: The input unit is used to receive the current structured state description and the subtask list, wherein each subtask in the subtask list includes the preconditions, task description and subsequent state changes; The deduction processing unit is used to traverse the subtask list sequentially and perform the following operations on each subtask: Prerequisite check: Extract the prerequisites for this subtask, call LLM and use the preset prerequisite check prompt template to determine whether the prerequisites are consistent with the current deduction state; State Update: If the preconditions are met, call LLM and update the current deduction state according to the task description of the subtask and subsequent state changes through the preset state update prompt template; Failure handling: If the preconditions are not met, the deduction is judged to have failed, and the identifier of the failed subtask and the unmet preconditions are recorded. The output unit is used to output the deduction results and the final deduction state sequence.

[0012] Furthermore, the feedback and optimization closed-loop module includes: The receiving unit is used to receive information about failed subtasks, reasons for failure, and the status context at the time of failure. The analysis and processing unit is used to invoke the LLM, analyze the reasons for failure, and generate natural language feedback to guide replanning. The feedback unit is used to output natural language feedback text that guides replanning.

[0013] Furthermore, the decision-making and output module includes: The feasibility task output unit is used to format the list of subtasks derived from the deduction to make it readable. The manual editing and confirmation unit is used to display the formatted list of subtasks and supports manual editing, review, and confirmation.

[0014] Furthermore, the decision-making and output module also includes: The task storage unit stores a list of valid subtasks that have been manually confirmed, which can be called by subsequent similar tasks to improve planning efficiency.

[0015] Secondly, embodiments of the present invention provide a task decomposition method for embodied intelligent robots based on physical state deduction, applied to the system described above. The method includes the following steps: S100: System initialization and knowledge preparation, predefine and store all prompt word templates, and configure robot kinematics model and sensor interface; S200: Real-time state perception and construction, based on sensor interface, synchronously collects environmental images and robot internal sensor data, calls VLM and fuses the environmental images and robot internal sensor data to generate the current structured state description; S300: Run the task instruction receiving and initial decomposition module, fill the received user task instructions and the current structured state description into the corresponding prompt word template, call LLM, and obtain the initial subtask list; S400: Feasibility closed-loop simulation and optimization. The current structured state description is used as the initial simulation state. The output sub-task list is traversed and verified one by one. Prerequisite checks and state updates are performed. If the prerequisites are met, the current structured state description is updated synchronously. If the prerequisites are not met, precise feedback is generated and LLM is driven to re-optimize the sub-task list until all sub-task lists pass the simulation. S500: Result output and confirmation. Outputs a list of subtasks that have passed verification, which is then sent to the robot execution engine after manual confirmation.

[0016] Furthermore, the prompt word template includes environment description, task breakdown, prerequisite checks, status updates, and feedback generation.

[0017] Furthermore, step S400 includes the following specific processes: Step S401: Set the current deduced state Current_State to be equal to the current structured state description State_Structured; Step S402: Iterate through each subtask S_i in the subtask list Subtask_List in sequence; Step S403: Prerequisite check: Fill the preconditions of the current deduction state Current_State and S_i into the check prompt word template, and call LLM to make a judgment; Step S404: If LLM returns "yes", then update the state: fill the update prompt word template with the current deduction state Current_State and the postconditions of S_i, call LLM, replace the current deduction state Current_State with its output, and jump to S402 to process the next subtask. Step S405: If the LLM returns "No" and the premise is not met, the deduction fails, triggering the feedback and optimization closed-loop module to generate feedback text. Step S406: Input the task instruction Task_Cmd, the current structured state description State_Structured, and the feedback text Feedback into the task decomposition unit to request the regeneration or partial adjustment of the subtask list Subtask_List, thus forming a closed-loop optimization.

[0018] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention constructs a closed-loop technical framework of perception, deduction, and optimization, enabling robots to virtually rehearse task sequences before executing actual physical actions. This allows for early identification of environmental state conflicts and execution path risks, fundamentally transforming the task execution mode from a passive response relying on trial and error to proactive and reliable planning based on physical state deduction. When deduction fails, the system can autonomously diagnose the cause and generate optimization suggestions, driving the task sequence to be regenerated, endowing the robot with the intelligence to learn from mistakes and self-correct. Simultaneously, the system transforms natural language instructions and physical world states into a unified representation that is understandable and reasonable by the machine, effectively bridging the semantic gap between high-level instructions and low-level control. The resulting validated task solutions can be stored and reused, forming a continuously accumulating robot task knowledge base, significantly improving the robot's adaptability, safety, and deployment efficiency in performing complex tasks in open environments.

[0019] 2. This invention improves the physical feasibility and reliability of task decomposition. Through the core task feasibility deduction module, the system can automatically identify and eliminate subtasks or sequences in the LLM output that violate physical laws before execution, eliminating "physical illusions" at the planning stage. This reduces the risk of task failure or danger caused by robot misplanning, resulting in a qualitative leap in the credibility and direct executability of the output results.

[0020] 3. This invention achieves automated verification and closed-loop optimization of the task decomposition process, upgrading the traditional open-loop decomposition to an intelligent closed loop with self-checking and correction capabilities. The system can automatically diagnose the physical nature of planning defects and drive iterative optimization, reducing reliance on manual debugging and expert experience. This makes the automated and reliable decomposition of complex tasks possible, significantly improving the system's autonomy and practicality.

[0021] 4. This invention provides interpretable and traceable decomposition results. The output of this invention is not only a list of subtasks, but also includes a complete sequence of physical state evolutions generated during the deduction process. This provides a clear physical explanation of "why the decomposition is performed in this way" and "what physical state changes each step will lead to," greatly enhancing the system's transparency and debuggability, and facilitating collaborative decision-making with human operators. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a flowchart of the task decomposition system for embodied intelligent robots based on physical state deduction provided in an embodiment of the present invention.

[0024] Figure 2 This is a flowchart illustrating the specific system workflow for task decomposition of an embodied intelligent robot based on physical state deduction, as provided in this embodiment of the invention.

[0025] Figure 3 This is a flowchart illustrating the task decomposition method for embodied intelligent robots based on physical state deduction provided in this embodiment of the invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: according to Figure 1 As shown, this embodiment of the invention provides a task decomposition system for an embodied intelligent robot based on physical state deduction, comprising five parts: a task instruction receiving and initial decomposition module, a state perception and formatted input module, a task feasibility deduction module, a feedback and optimization closed-loop module, and a decision and output module. in: The task instruction receiving and initial decomposition module, as the core interaction hub between the user and the robot system, bears the crucial responsibility of transforming and initially decomposing requirements, and is the starting point for task execution. For example, it can accurately identify the core requirements, target parameters, time constraints, and priority order in the user's verbal or written instructions, while filtering redundant information, correcting ambiguities, and establishing a standardized interaction channel.

[0028] The State Awareness and Formatted Input Module is the core component enabling the robot to perceive itself and its environment, providing a basis for decision-making. It is responsible for converting dynamic state information into a structured description of the current state that the system can recognize. The module collects two types of core information in real time through various sensors mounted on the robot, such as vision sensors, force sensors, position sensors, and environmental sensors: first, the robot's own state, including joint angles, battery level, actuator load, equipment malfunction status, and current pose; second, the external environmental state, encompassing obstacle distribution in the work area, target object position and orientation, ambient temperature and humidity, and ground flatness. The collected raw data is processed and structured according to preset data specifications, generating a structured description of the current state that includes dimensions such as state type, parameter values, update time, and reliability. This provides accurate data support for task feasibility simulation, avoids simulation deviations caused by state ambiguity, and ensures the safety and stability of task execution.

[0029] The task feasibility simulation module is a crucial step in verifying the possibility of task execution and avoiding physical conflicts. Based on a list of subtasks and a structured state description, it conducts step-by-step physical state simulations. The simulation process follows physical laws and robot kinematic constraints, simulating the changes in the robot's own state and the external environment state after each step, verifying the executability of each subtask under the current conditions. For example, when simulating the "grabbing a part" subtask, it verifies whether the robot's end effector can reach the target position, whether the grasping force is suitable for the weight of the part, and whether there are obstacles obstructing the grasping path. For subtasks with dependencies, the module prioritizes verifying whether the state after the completion of the preceding task meets the start conditions of the subsequent task, while also predicting potential physical conflicts, such as joint movement exceeding limits, equipment overload, and collisions with environmental objects. During the simulation, the module records the state change trajectory and verification results for each step. If a subtask simulation fails, subsequent simulations are immediately terminated, and the conflict node is locked, providing accurate conflict location data for the feedback optimization module, ensuring that physical feasibility risks are eliminated before task execution.

[0030] The feedback and optimization closed-loop module is responsible for handling problems and iteratively optimizing tasks after a feasibility simulation fails, ensuring the final implementation of task requirements. When the feasibility simulation module fails, this module immediately retrieves detailed data from the conflict nodes, deeply analyzes the core reasons for the conflict, such as insufficient robot movement range, inability to avoid obstacles, exceeding equipment limits, or errors in subtask dependency logic. It then combines this with the system's pre-set solution library to generate targeted modification suggestions, including adjusting the subtask execution order, optimizing movement paths, changing execution tools, and breaking down complex subtasks. Feedback information is presented to users or operators in an intuitive format, simultaneously driving the task instruction receiving and initial decomposition module to initiate secondary optimization, adjusting the subtask list and optimizing the task decomposition logic based on the modification suggestions.

[0031] The decision-making and output module, as the final stage of the task processing flow, is responsible for the final organization, confirmation, and execution integration of the sub-task list, while also enabling the reuse and accumulation of task data. The module first standardizes and formats the optimized sub-task list, supplementing details such as execution parameters, safety thresholds, and contingency plans, generating a current structured state description that conforms to the robot's execution interface specifications. After confirmation, the module connects the current structured state description to the robot's execution end, synchronizing task progress in real time and receiving execution feedback to ensure accurate instruction execution. Simultaneously, it stores the current structured state description in the task storage unit, marking task characteristics, applicable scenarios, and execution effects, forming a reusable task template. When similar needs arise subsequently, the module can quickly retrieve the template and fine-tune parameters, significantly shortening the task processing cycle.

[0032] This invention constructs a closed-loop technical framework of perception, deduction, and optimization, enabling robots to virtually rehearse task sequences before executing actual physical actions. This allows for the early identification of environmental state conflicts and execution path risks, fundamentally transforming the task execution mode from a passive response relying on trial and error to proactive and reliable planning based on physical state deduction. When deduction fails, the system can autonomously diagnose the cause and generate optimization suggestions, driving the task sequence to be regenerated, endowing the robot with the intelligence to learn from mistakes and self-correct. Simultaneously, the system transforms natural language instructions and physical world states into a unified representation that is understandable and reasonable by the machine, effectively bridging the semantic gap between high-level instructions and low-level control. The resulting validated task solutions can be stored and reused, forming a continuously accumulating robot task knowledge base, significantly improving the robot's adaptability, safety, and deployment efficiency in performing complex tasks in open environments.

[0033] according to Figure 2 As shown, the above-mentioned embodied intelligent robot task decomposition system based on physical state deduction of the present invention will be described in detail, mainly consisting of the following five hierarchical functional modules: 1. The task instruction receiving and initial decomposition module receives the task instruction input by the user through the human-computer interaction unit. On the one hand, it triggers the robot state perception and environment perception unit to execute. On the other hand, after the robot state perception and environment perception are completed, according to the instruction, the module incorporates the structured description of the environment and state into the task decomposition prompt word template to perform initial task decomposition, form a preliminary sub-task list, and provide it to the task feasibility simulation to perform the feasibility simulation of task decomposition.

[0034] The task instruction receiving and initial decomposition module consists of a human-computer interaction unit and a task decomposition unit. The two units communicate and interact in real time through an internal data interface to ensure the smoothness and stability of the overall module operation.

[0035] (1) Human-computer interaction unit: It serves as the front-end interaction platform for the task instruction receiving and initial decomposition module, and is also the core window for the robot to connect with users regarding their needs, ensuring the accuracy and completeness of instruction reception. This unit breaks through the traditional robot's single instruction input mode, providing both natural language interaction and graphical interface interaction as two mainstream operation methods, which can flexibly adapt to the operating habits, professional backgrounds, and usage scenarios of different users.

[0036] (2) Task decomposition unit: The task decomposition unit is the core computing hub of the task instruction receiving and initial decomposition module, and also the starting core module of the entire robot task planning system. It undertakes the key tasks of sensory data fusion, decomposition rule matching, and intelligent sub-task generation. Its operation effect directly determines the initial direction of task planning and the underlying logic of subsequent execution.

[0037] The specific workflow of this module is implemented through the following steps: Input: Task instruction Task_Cmd, current structured state description State_Structured.

[0038] Processing: Invoke the large language model and, based on the current state, recursively decompose the high-level task into a series of atomic or composite subtasks. Each subtask must adhere to a specified output format.

[0039] Output: A list of subtasks that conforms to the specification, which is a Subtask_List.

[0040] Example of key prompts (task breakdown): "You are a robot task planner. Current state: {Insert current structured state description State_Structured}. Your goal is: {Insert task instruction Task_Cmd}."

[0041] Please break down the objective into a series of subtasks. Each subtask must be described strictly according to the following JSON format: { "id": serial number, "description": "Natural language description of the subtask" "preconditions": ["precondition1", "precondition2", ...], / / Must be based on the state after the current or previous task. "postconditions":["postcondition 1", "postcondition 2", ...] / / Describes how the state changes after execution. } Please ensure logical consistency between subtasks. Only output JSON arrays. 2. The State Perception and Formatted Input Module is the core data support module of the embodied intelligent robot task decomposition system based on physical state deduction in this invention. It undertakes the key functions of physical world information collection, data standardization processing and machine-understandable format conversion, and is an important data bridge connecting the robot's physical perception and subsequent intelligent deduction.

[0042] This module integrates multi-dimensional perception data of the robot's own state and the external environment. Through a series of operations such as standardized collection, structured processing, and semantic conversion, it transforms the originally discrete and non-uniform physical perception information into structured state data that conforms to the input specifications of large language models and can directly drive logical deduction.

[0043] The state perception and formatted input module consists of two core units: the robot state perception unit and the environment perception unit. The two units achieve perception trigger synchronization, real-time data interaction, and result collaborative integration through an internal high-speed data interface, ensuring the overall perception and processing efficiency of the module.

[0044] (1) Robot state perception unit: This unit establishes real-time communication connections with various sensors inside the robot, enabling it to continuously and frequently read sensor data from the robot's core components. The data collection range covers key dimensions such as robot motion, execution, and positioning. Specifically, it includes joint angle, motion speed, and rotational torque data collected by the joint encoder; gripper opening / closing stroke data collected by the end effector opening / closing degree sensor; gripping force and contact pressure data collected by the pressure sensor; and robot pose, movement distance, and motion trajectory data collected by the odometer.

[0045] The robot state perception unit can perform real-time cleaning, noise filtering, and standardized quantization processing on the various sensor data collected, transforming the original analog signals or discrete digital signals into structured data with unified units and formats, and refreshing them in real time according to a preset update frequency, ensuring that the output robot state data can truly and accurately reflect the robot's current hardware operating status and execution capabilities.

[0046] To address the core requirements of task decomposition and feasibility analysis in this invention, robot state perception does not need to cover too many redundant dimensions; it only needs to focus on the two core physical states of "idle" and "holding an object." These two key states can be directly and reliably determined by core information such as motor operating parameters and end effector pressure sensor data collected by this unit, without the need for additional sensing devices or complex algorithms. This approach not only leverages the advantages of standardized data collection and processing mentioned earlier but also further simplifies the state judgment logic and improves overall perception efficiency.

[0047] For example, the following rules can be used to determine this: A. "Idle" status judgment Gripper: For electric grippers, directly read their opening angle or width sensor. When the opening degree reaches the preset "fully open" threshold, the status is "idle".

[0048] Vacuum suction cup: Reads the value from the vacuum pressure sensor. When the pressure is below the adsorption threshold, the status is "idle".

[0049] Electromagnet: Reads current or magnetic sensors; when power is off, it is "idle".

[0050] Simple rule: if (gripper_opening > equal to MAX_OPEN_THRESHOLD) then state equals "idle". B. Judgment of "Holding an Object" Status When the aforementioned sensor indicates that the gripper is not in the "fully open" state, it can logically be determined that the gripper is "holding an object".

[0051] (2) Environmental perception unit: The environmental perception unit is specifically designed to enable robots to visually perceive, semantically understand, and structurally describe their external working environment. Using a high-definition camera mounted on the robot as its core perception carrier, this unit first receives a perception trigger signal from the task instruction receiving and initial decomposition module, immediately activating the camera to capture images and accurately obtain real-time images of the working environment, ensuring the timeliness of environmental information. Second, the acquired environmental images are transmitted in real-time to a pre-integrated Visual-Language Model (VLM). This model performs visual feature extraction, object recognition, and scene analysis based on the image content, automatically generating a preliminary natural language description of the categories, locations, shapes, and spatial relationships of visible objects in the image. Finally, this unit deeply integrates the generated preliminary environmental description with the standardized robot state data output by the robot state perception unit. Following a preset structured cue word framework, it integrates the data to form a structured cue word, which is then further logically refined, formatted, and supplemented by the LLM, ultimately generating a unified, standardized, and machine-understandable current structured state description.

[0052] For example, the following keywords can be used for environmental perception: "You are a robot perception module. Based on the image, list the key objects in the field of view and describe the relative position (e.g., front, back, left, right, distance) of each object to the 'robot'. Output in the following JSON format, do not interpret:" { "objects":[ {"name":"Red Cup","relation":"On the table about 1 meter in front of the robot"}, {"name":"Kitchen door","relation":"Approximately 2 meters to the left of the robot, status: closed"} ] }” 3. The task feasibility deduction module, which also serves as the physical state deduction engine of this invention, undertakes the key function of executing the original physical state deduction algorithm of this invention. This module breaks through the technical bottlenecks of traditional robot task feasibility analysis, which relies solely on manually preset rules, cannot adapt to complex dynamic physical scenarios, and has a disconnect between deduction results and actual execution. By constructing a step-by-step execution-based deduction engine that simulates the actual operation process of the robot, it deeply integrates the abstract sub-task execution logic with specific physical state constraints, realizing intelligent, refined, and physical deduction of the entire task execution process. This fundamentally ensures the actual executability of the robot task decomposition results and avoids task execution failure due to physical constraint conflicts or unmet execution conditions.

[0053] The task feasibility simulation module integrates three core functional units: an input unit, a simulation processing unit, and an output unit. Its core components and specific workflow are as follows: (1) Input unit: The core function of this unit is to receive two main types of input data: the current structured state description and the subtask list. It then performs data format verification and preprocessing to ensure that the data conforms to the input specifications of the inference algorithm. After receiving the two types of data, it automatically performs data format verification and integrity checks. If any issues such as missing data or non-standard format are detected, the data will be fed back to the preceding modules in real time to ensure that the data input to the inference processing unit is complete, standardized, and usable, thus laying a data foundation for the accuracy of subsequent inferences.

[0054] (2) Deduction and processing unit: This unit is responsible for the crucial tasks of sequentially traversing the subtask list, performing prerequisite checks, updating the state, and handling failures. Its core function is to build a simulation engine that simulates the step-by-step execution. Using the current structured state description transmitted by the input unit as the initial simulation state, it sequentially performs prerequisite checks, state updates, and failure handling operations for each subtask S_i according to the preset execution order of the subtask list. This achieves a full-process, detailed simulation of the entire task execution process. The specific operation flow is as follows: Step A: Prerequisite Check The deduction processing unit first extracts the preset preconditions of the current subtask S_i in sequence, then calls the LLM and constructs a precise verification instruction containing "current deduction state" and "subtask preconditions" through a pre-configured standardized precondition check prompt template. This instruction is then input into the LLM, which intelligently determines whether each statement in the preconditions is satisfied in the current deduction state based on the existing structured factual evidence.

[0055] Typical examples of prompt words are: "Known facts: {Current deduction state}."

[0056] Question: Are all statements in {S_i.preconditions} true? Answer only 'yes' or 'no'.

[0057] If 'No,' please list the statements that are not true. Step B: Status Update If all the preconditions for LLM feedback are met, the deduction processing unit will immediately enter the state update phase. First, it extracts the task description and postconditions of the current subtask S_i. Then, it calls LLM and constructs a state update instruction containing "old deduction state", "subtask description", and "postconditions" using a pre-configured standardized state update prompt template. This instruction is then input into LLM, which intelligently generates an updated complete fact list based on the old state facts and the subtask execution effects, thus dynamically refreshing, supplementing, and improving the current deduction state.

[0058] Typical examples of prompt words are: "Known old facts: {Current deduction state}."

[0059] Now, {S_i.description} has occurred, and its effect is {S_i.postconditions}.

[0060] Please generate a new, complete list of facts.

[0061] Output only the new list. Step C: If the premise is not met, the deduction fails, and the failure point S_i and the reason are recorded.

[0062] (3) Output unit: After the simulation processing unit completes the full-process simulation, it will integrate, encapsulate, and output the simulation process data and results, ultimately outputting two core deliverables: the simulation results and the final simulation state sequence. The specific contents are as follows: The simulation results are presented in a clear and standardized text format, showing the final feasibility conclusion of this task simulation. The conclusion is divided into two types: if all subtasks in the list are successfully simulated, the output is "fully feasible", indicating that the list of subtasks has the conditions for actual execution under the current physical state; if the simulation fails at a certain subtask node, the output is "failed at subtask X", where X is the identifier of the specific subtask, and the unmet preconditions from the LLM feedback are attached as a description of the reason for failure.

[0063] Final deduction state sequence: Generates a structured state change sequence containing the entire deduction process. This sequence is organized according to the execution order of subtasks and records in detail the deduction state before and after the execution of each subtask, the premise verification results, the state update content, failure nodes and other key information, which fully restores the working process and state change logic of the deduction engine.

[0064] 4. Feedback and Optimization Closed-Loop Module: This module receives output from the physical state deduction engine, analyzes problems, generates optimization guidance, and drives system correction. It overcomes the limitations of traditional one-time system decomposition, transforming deduction failure information into precise instructions to drive targeted adjustments in task decomposition units, achieving dynamic optimization of task decomposition results.

[0065] The feedback and optimization closed-loop module is automatically activated when the simulation fails. It contains three main units: receiving, analysis and processing, and feedback, forming a standardized feedback process, as detailed below: (1) Receiving unit: As an information entry point, it synchronously receives three types of core information transmitted by the simulation module through a high-speed interface: full attributes of failed subtasks, unmet prerequisites fed back by LLM, and structured state context at the moment of failure, thus fully reconstructing the failure scenario.

[0066] (2) Analysis and processing unit: Leveraging the logical reasoning capabilities of LLM, the integrated failure information is input into customized prompts to guide the generation of accurate and feasible optimization suggestions.

[0067] Typical prompt words example: "In robot task planning, when the state is {failure state}, the subtask {failed subtask description} cannot be executed because its premise {unsatisfied premise} is not met."

[0068] Please analyze the reasons and provide a one-sentence suggestion for improvement (e.g., adjust the order, add a prerequisite task, modify the parameters).

[0069] Only provide suggestions. This design requires no preset rules, adapts to various scenarios, and ensures that the suggestions are operable and formatted consistently.

[0070] (3) Feedback unit: The suggestions generated by LLM are encapsulated into standardized feedback text, which clearly identifies failed subtasks, optimization directions, and brief reasons for failure. The text is in a machine-readable format, which supports automatic reading and adjustment by task decomposition units and can also be synchronized to the human-machine interface for manual reference. This ensures smooth data interaction between modules and takes into account both automatic optimization and manual intervention needs.

[0071] 5. The Decision and Output module is the core module of the system terminal. It is responsible for the formatted output of task planning results, manual review, and storage of high-quality templates. It serves as the hub connecting the planning process, robot execution, and knowledge accumulation. The module comprises three main units: feasible task output, manual editing and confirmation, and task storage. Its responsibilities are as follows: (1) Feasibility task output unit: Receive the subtask list Subtask_List that has passed the simulation, and convert it into a standardized list that can be intuitively displayed on the interface and directly read by the robot according to standardized rules. Add annotations such as execution priority to take into account human and machine needs, and lay the foundation for subsequent verification and execution.

[0072] (2) Manual editing confirmation unit: It provides a visual interface that simultaneously displays the task planning results and the complete simulation process, allowing experts to view, modify, approve, or reject them.

[0073] (3) Task storage unit: High-quality task planning templates, approved by humans, are categorized and stored, archived by task type and scenario, and indexed for quick retrieval by subsequent modules. This enables experience reuse, shortens the planning cycle, and fosters a virtuous cycle of knowledge accumulation.

[0074] Example 2: Based on the same inventive concept, embodiments of the present invention also provide a method for decomposing embodied intelligent robot tasks based on physical state deduction, referring to... Figure 3 As shown, the process includes the following steps S100~S500: Step S100: System initialization and knowledge preparation, predefine and store all prompt word templates, configure robot kinematics model and sensor interface; complete the basic groundwork before system operation, unify prompt word template standards, and ensure consistency of module calls; pre-configure hardware interfaces and models to avoid compatibility issues during operation and ensure efficient operation after system startup.

[0075] Specifically, step S100 includes: Step S101: Predefine and store all prompt word templates, which include environment description, task decomposition, prerequisite checks, status updates, and feedback generation.

[0076] Step S102: Configure the robot kinematics model and sensor interface.

[0077] Step S200: Real-time state perception and construction. Based on the sensor interface, environmental images and internal sensor data of the robot are collected synchronously. The VLM is called and the environmental images and internal sensor data of the robot are fused to generate the current structured state description.

[0078] Specifically, step S200 includes: Step S201: Synchronously acquire environmental images and data from the robot's internal sensors.

[0079] Step S202: The environment and state structured description generation module runs, calls VLM and merges data to generate the current structured state description State_Structured at the current moment.

[0080] Step S300: Run the task instruction receiving and initial decomposition module, fill the received user task instructions and the current structured state description into the corresponding prompt word template, call LLM to obtain the initial subtask list; decompose the task instructions and the current state to ensure that the subtask list is adapted to the actual scenario and avoids deviating from physical constraints; templated LLM calls improve decomposition efficiency, and the generated subtask list provides clear objects for subsequent deduction.

[0081] Specifically, step S300 includes: Step S301: Receive the task instruction Task_Cmd through the human-computer interaction unit.

[0082] Step S302: The task decomposition unit runs, filling the task instruction Task_Cmd and the current structured state description State_Structured into the corresponding prompt word template, calling LLM to obtain the initial subtask list Subtask_List.

[0083] Step S400: Feasibility closed-loop simulation and optimization. Using the current structured state description as the initial simulation state, the output subtask list is traversed and verified one by one. Prerequisite checks and state updates are performed. If the prerequisites are met, the current structured state description is updated synchronously. If the prerequisites are not met, precise feedback is generated and the LLM is driven to re-optimize the subtask list until all subtask lists pass the simulation. A closed-loop optimization mechanism is built to automatically correct simulation failures without manual intervention. The feasibility of each step is ensured by verifying each subtask.

[0084] Specifically, step S400 includes: Step S401: Set the current deduced state Current_State to be equal to the current structured state description State_Structured.

[0085] Step S402: Iterate through each subtask S_i in the Subtask_List sequentially.

[0086] Step S403: Prerequisite check: Fill the check prompt word template with the preconditions of the current deduction state Current_State and S_i, and call LLM to make a judgment.

[0087] Step S404: If LLM returns "Yes", then update the state: fill the update prompt word template with the current deduced state Current_State and the postconditions of S_i, call LLM, and replace the current deduced state Current_State with its output. Jump to S402 to process the next subtask.

[0088] Step S405: If the LLM returns "No" and the premise is not met, the deduction fails. The precise feedback generation module is triggered to generate feedback text.

[0089] Step S406: Input the task instruction Task_Cmd, the current structured state description State_Structured, and the feedback text Feedback into the task decomposition unit to request the regeneration or partial adjustment of the subtask list Subtask_List (i.e., return to step S302) to form a closed-loop optimization.

[0090] Step S500: Result output and confirmation. Output the list of verified subtasks, which are sent to the robot execution engine after manual confirmation. Through final manual review, potential vulnerabilities in extreme scenarios of the automated process are avoided, and task reliability is enhanced. Output a structured task sequence to ensure that the robot can execute it directly, thus connecting the final link between planning and execution.

[0091] Specifically, step S500 includes: Step S501: When all subtasks in the Subtask_List have passed the deduction (completed in a loop from S402 to S404), the planning is successful.

[0092] Step S502: Send the list of successful subtasks, Subtask_List, into the feasible task output unit for formatting.

[0093] Step S503: The formatted planning results are displayed in the manual editing and confirmation unit, awaiting final confirmation, editing, or storage in the task storage unit.

[0094] Step S504: The confirmed final plan is sent to the robot motion execution engine to drive the robot body to execute.

[0095] This invention constructs a complete task planning closed loop through five core steps. Each step is progressive and collaborative, forming a standardized, automated, and intelligent task decomposition system for embodied intelligent robots. The system is supported by prompt word templates and LLM and VLM technologies, integrating robot physical state and external environment perception. It automatically corrects planning problems through a closed-loop optimization mechanism, while introducing a manual confirmation step to strengthen risk control. This ensures both the efficiency and adaptability of task planning and improves execution reliability. Furthermore, the storage and reuse mechanism for high-quality solutions enables knowledge accumulation. As application scenarios accumulate, the system's planning efficiency and accuracy will continue to improve, making it widely adaptable to various complex work scenarios and providing core technological support for the autonomous task execution of embodied intelligent robots.

[0096] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0097] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A task decomposition system for embodied intelligent robots based on physical state deduction, characterized in that, include: The task instruction receiving and initial decomposition module is used to establish an interaction channel between the user and the system, and to convert the user's natural language task requirements into a list of subtasks that the system can process. The state perception and formatted input module is used to acquire information about the robot's own state and the external environment state, and generate a current structured state description; The task feasibility simulation module is used to perform step-by-step physical state simulation based on the received sub-task list and the current structured state description to verify its physical feasibility. The feedback and optimization closed-loop module generates feedback information containing conflict reasons and modification suggestions when the task feasibility deduction module fails, based on the task feasibility deduction module, and drives the task instruction receiving and initial decomposition module to re-optimize the sub-task list. The decision and output module is used to format, manually confirm, and reuse the subtask list, while also connecting to the robot execution end.

2. The system as described in claim 1, characterized in that, The task instruction receiving and initial decomposition module includes: The human-computer interaction unit is used to provide a natural language or graphical interface and receive task instructions input by the user; The task decomposition unit is used to fill the task instructions and the current structured state description into the corresponding prompt word template, and call LLM to generate a list of subtasks that conforms to a preset format.

3. The system as described in claim 1, characterized in that, The state awareness and formatted input module includes: The robot state perception unit is used to read data from the robot's internal sensors in real time. The environmental perception unit collects images of the current environment through the camera device mounted on the robot body and calls the VLM to generate a preliminary description.

4. The system as described in claim 1, characterized in that, The task feasibility simulation module includes: The input unit is used to receive the current structured state description and the subtask list, wherein each subtask in the subtask list includes the preconditions, task description and subsequent state changes; The deduction processing unit is used to traverse the subtask list sequentially and perform the following operations on each subtask: Prerequisite check: Extract the prerequisites for this subtask, call LLM and use the preset prerequisite check prompt template to determine whether the prerequisites are consistent with the current deduction state; State Update: If the preconditions are met, call LLM and update the current deduction state according to the task description of the subtask and subsequent state changes through the preset state update prompt template; Failure handling: If the preconditions are not met, the deduction is judged to have failed, and the identifier of the failed subtask and the unmet preconditions are recorded. The output unit is used to output the deduction results and the final deduction state sequence.

5. The system as described in claim 1, characterized in that, The feedback and optimization closed-loop module includes: The receiving unit is used to receive information about failed subtasks, reasons for failure, and the status context at the time of failure. The analysis and processing unit is used to invoke the LLM, analyze the reasons for failure, and generate natural language feedback to guide replanning. The feedback unit is used to output natural language feedback text that guides replanning.

6. The system as described in claim 1, characterized in that, The decision-making and output module includes: The feasibility task output unit is used to format the list of subtasks derived from the deduction to make it readable. The manual editing and confirmation unit is used to display the formatted list of subtasks and supports manual editing, review, and confirmation.

7. The system as described in claim 6, characterized in that, The decision and output module also includes: The task storage unit stores a list of valid subtasks that have been manually confirmed, which can be called by subsequent similar tasks to improve planning efficiency.

8. A method for task decomposition of embodied intelligent robots based on physical state deduction, characterized in that, The method of using the system as described in any one of claims 1-7 includes the following steps: S100: System initialization and knowledge preparation, predefine and store all prompt word templates, and configure robot kinematics model and sensor interface; S200: Real-time state perception and construction, based on sensor interface, synchronously collects environmental images and robot internal sensor data, calls VLM and fuses the environmental images and robot internal sensor data to generate the current structured state description; S300: Run the task instruction receiving and initial decomposition module, fill the received user task instructions and the current structured state description into the corresponding prompt word template, call LLM, and obtain the initial subtask list; S400: Feasibility closed-loop simulation and optimization. The current structured state description is used as the initial simulation state. The output sub-task list is traversed and verified one by one. Prerequisite checks and state updates are performed. If the prerequisites are met, the current structured state description is updated synchronously. If the prerequisites are not met, precise feedback is generated and LLM is driven to re-optimize the sub-task list until all sub-task lists pass the simulation. S500: Result output and confirmation. Outputs a list of subtasks that have passed verification, which is then sent to the robot execution engine after manual confirmation.

9. The method as described in claim 8, characterized in that, The prompt template includes environment description, task breakdown, prerequisite checks, status updates, and feedback generation.

10. The method as described in claim 8, characterized in that, Step S400 includes the following specific processes: Step S401: Set the current deduced state Current_State to be equal to the current structured state description State_Structured; Step S402: Iterate through each subtask S_i in the subtask list Subtask_List in sequence; Step S403: Prerequisite check: Fill the preconditions of the current deduction state Current_State and S_i into the check prompt word template, and call LLM to make a judgment; Step S404: If LLM returns "Yes", then update the state: fill the update prompt word template with the current deduction state Current_State and the postconditions of S_i, call LLM, replace the current deduction state Current_State with its output, and jump to S402 to process the next subtask. Step S405: If the LLM returns "No" and the premise is not met, the deduction fails, triggering the feedback and optimization closed-loop module to generate feedback text. Step S406: Input the task instruction Task_Cmd, the current structured state description State_Structured, and the feedback text Feedback into the task decomposition unit to request the regeneration or partial adjustment of the subtask list Subtask_List, thus forming a closed-loop optimization.

Images