Space manipulator task instruction analysis and execution plan generation method and system

Abstract

The invention discloses a space manipulator task instruction analysis and execution plan generation method and system. The method comprises the steps that a structured task event is extracted from a flight task document in a natural language form through combination of a large language model and a rule system; the task event obtained through analysis comprises a target configuration, the shortest path of the mechanical arm from the current configuration to the target configuration is calculated, and path planning is completed; according to each event of path planning, time scheduling and observation station screening of each event are completed in combination with track circle numbers, measurement and control resource availability and illumination and shielding conditions; the events are disassembled into atomic-scale actions according to a time sequence, and the atomic-scale actions are mapped into a bottom-layer control instruction sequence; wherein in the whole process, the large language model cooperatively works with an external path planner, an event scheduler and a configuration database through a function call mechanism or a knowledge base query interface; the method has the advantages that the integration ability and generalization ability in the aspects of processing space task time arrangement, resource screening, multi-modal information mapping and the like are high.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to a method and system for parsing mission instructions and generating execution plans for a space robotic arm. Background Technology

[0002] Space mission planning and command execution are fundamental research topics in the field of aerospace flight control. Their goal is to enable spacecraft systems to accurately and efficiently execute control commands issued by ground personnel, such as operating the space station's robotic arm, orbital roving, equipment maintenance, and attitude adjustment. These tasks are widely used in critical aerospace scenarios such as space station maintenance, on-orbit servicing, and space science experiments, placing increasingly higher demands on the intelligence and automation of mission execution.

[0003] In actual missions, mission instructions are typically recorded in flight procedure documents in the form of natural language or semi-structured tables, exhibiting strong unstructured characteristics. Manually extracting mission elements is time-consuming, labor-intensive, and prone to errors. Furthermore, mission execution requires comprehensive consideration of multiple spatiotemporal constraints, including telemetry and control resources, communication visibility windows, illumination obstruction, and execution sequence. Traditional methods primarily rely on manual analysis and rule-based systems for mission parsing and scheduling planning, resulting in low efficiency, poor scalability, and difficulty in meeting the real-time and accuracy requirements of complex missions. Especially in spacecraft on-orbit management, tasks such as remote control issuance, measurement and tracking, and attitude control are highly dependent on the coupling and coordination of orbital and resource conditions, essentially constrained optimization scheduling problems. Traditional manual scheduling methods struggle to comprehensively and in real-time respond to these complex constraints, easily leading to resource conflicts and scheduling failures.

[0004] Therefore, the rapid development of Large Language Models (LLMs) in recent years in natural language understanding and knowledge reasoning has brought new opportunities for intelligent parsing and automatic planning of space missions. Utilizing LLMs to parse natural language instructions and combining them with graph structure modeling to achieve path search and state planning has become a promising research direction. For example, Chinese Patent Publication No. CN120542462A discloses a hybrid multi-robot collaborative method and system driven by a large language model. Meanwhile, some open-source or private domain models (such as the "Aerospace Private Domain Large Model") have been initially applied to instruction parsing and orbit calculation tasks, and graph search methods have also achieved good results in robot operation tasks. However, existing methods still suffer from insufficient integration capabilities and limited generalization capabilities in handling task scheduling, resource selection, and multimodal information mapping. Summary of the Invention

[0005] The technical problem to be solved by this invention is that existing methods still have insufficient integration capabilities and limited generalization capabilities in handling space mission scheduling, resource selection, and multimodal information mapping.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a method for parsing and generating execution plans for space robotic arms, comprising:

[0007] S1. Flight mission documents in natural language form are used to extract structured mission events through a large language model combined with a rule system;

[0008] S2. The parsed task events contain the target configuration. Calculate the shortest path from the current configuration to the target configuration for the robotic arm, and complete the path planning.

[0009] S3. Based on the events of the path planning, combined with the track circle number, the availability of telemetry and control resources, and the lighting and occlusion conditions, complete the time scheduling and station selection for each event;

[0010] S4. Events are broken down into atomic actions in chronological order and mapped to a sequence of underlying control instructions.

[0011] Throughout the process, the large language model works in collaboration with the external path planner, event scheduler, and configuration database through function call mechanisms or knowledge base query interfaces.

[0012] This invention leverages the powerful natural language understanding capabilities of a large language model to extract structured mission events from semantically complex flight mission documents. The large language model acts as the parsing kernel, outputting these structured events under specified prompt templates. This fully utilizes the language understanding capabilities of the pre-trained large language model, enhancing the system's robustness and generalization in multimodal information mapping across various scenarios. By combining path planning and timing scheduling, and considering the constraints of telemetry and control resources, it automatically generates orbital mission execution plans that meet these constraints. Specifically, it combines orbit circle numbers, telemetry and control resource availability, and lighting and occlusion conditions to complete the timing scheduling and station selection for each event, improving the accuracy and real-time performance of path planning. Events are decomposed into atomic-level actions in chronological order and mapped to a sequence of underlying control commands, further optimizing the efficiency of mission timing and resource selection. Therefore, the overall solution demonstrates strong integration and generalization capabilities in handling space mission timing, resource selection, and multimodal information mapping.

[0013] Furthermore, S1 includes:

[0014] Assuming the original flight control mission command document is a natural language sequence, the goal is to parse the set of mission events from it:

[0015]

[0016] Where i is the i-th event, N is the total number of events, and t i e is the event start time expression. id is the event name. i For event instruction code, a i Special notes are added for events; the task of the large language model is to map the input flight control mission instruction document into a structured event list ε, ultimately obtaining an event structured table:

[0017]

[0018] Furthermore, S2 includes:

[0019] Pathfinding is achieved using breadth-first search or heuristic search, with the goal of finding the path within a given initial state s. init and target state s goal Finding the shortest path between the given conditions minimizes the total cost; this process is encapsulated in the path lookup function PlanPath(s). init ,s goal The output is the intermediate configuration path and its action sequence.

[0020] Furthermore, S3 includes:

[0021] S31, e obtained from each parsing j Contains a time expression t relative to the task reference time T0 j The system first standardizes it into an offset δ j =Parse(t j The offset represents the expected execution time of the event in the task flow; Parse is the time offset conversion function; j is the j-th event; as mentioned above, e i The name of the event is the name of the i-th event. Therefore, for convenience, the j-th event will be directly described as event e in the following text. j .

[0022] S32, Given each ground tracking and control resource r i′ Its set of visible windows is This represents the start time of the k-th visible window, which belongs to the i′-th station. This represents the end time of the k-th visible window, which belongs to the i′-th station, where k represents the count of the visible windows. i′ This represents the total number of visible windows, where i′ represents the i′-th station, i.e., the ground control and measurement resource; the system sequentially judges event e. j Does each resource window meet the executable requirements?

[0023] S33. For resource windows that meet all conditions, the system records event e. j The candidate execution records are:

[0024]

[0025] in: For the event at station r i′ The executable window on; p is the offset time of the event execution relative to the start point of the window. j These are the event parameters; the final result constitutes the event scheduling table:

[0026]

[0027] Furthermore, the system sequentially judges event e. j Whether the executable conditions are met in each resource window, including:

[0028] Event e j For an operation to be executed in each resource window, the following three conditions must be met simultaneously.

[0029] 1) Track visibility constraint: The offset execution time point should be within a certain resource window, i.e.:

[0030]

[0031] 2) Geometric visibility constraint: There is no Earth obstruction between the target and the telemetry and control station, and the antenna field of view meets the communication requirements, i.e.:

[0032] Vis(e j ,r i′ ,T0+δ j ) = True

[0033] Vis represents a function that calculates whether there is Earth obstruction between the tracking station and the target, i.e. whether the target is on the other side of the Earth from the location of the tracking station. True indicates that the condition is true.

[0034] 3) Lighting and attitude constraints: If the event requires lighting conditions, then the following must be satisfied simultaneously:

[0035] θ sun (T0+δ j )∈[θ min ,θ max ]

[0036]

[0037] E sun =I0·cos(θ) sun )≥E min

[0038] Where, θ sun () represents the angle between the direction of solar incidence and the normal to the mission-sensitive surface, [θ min ,θ max] represents the permissible range of incident angles, POS(e j ,T0+δ j ) indicates that at time T0+δ i At that time, the task event e is executed. j The on-orbit target position, Ω shadow () indicates the Earth's shadow area; E sun Represents the effective incident energy, I0 is the solar constant, and E min This is the minimum effective incident energy;

[0039] Thus event e j The feasibility of execution is determined as follows:

[0040]

[0041] τ j Indicates event e j The start time of the allocation.

[0042] Furthermore, S4 includes:

[0043] Abstract events into atomic events in, This represents the h-th atomic event selected from the atomic event library. T is a collection of encapsulated atomic events. h P is the start time of the atomic event execution. h Let A be a parameter vector, and K be the total number of atomic events; each A n Corresponding to a group of multiple events The task blocks constituted; the scheduling process of atomic events on the time axis is T. h+1 =T h +ΔT h +d h , where d h Let ΔT be the execution duration of the h-th atomic event. h To allow for the interval between events, the atomic event scheduling table is finally obtained:

[0044]

[0045] Furthermore, the large language model includes a function tool invocation mechanism, a retrieval enhancement generation mechanism, and a response structure fusion mechanism. The function tool invocation mechanism refers to defining a set of external functions. During the parsing of user task instructions, the large language model determines whether to trigger a certain external function based on the embedded prompt structure or model reasoning. The function return value is then used as context input to the large language model for the next step of reasoning. The retrieval enhancement generation mechanism refers to automatically generating retrieval query instructions when the large language model needs to reason about a certain task but lacks relevant background knowledge. It calls the knowledge base interface to perform similarity retrieval, and the large language model concatenates the returned content into the current context to achieve dynamic injection of external knowledge. The response structure fusion mechanism refers to concatenating the explanatory text output by the large language model, the structured results returned by the function module, and the standardized output template structure. The standardized output template structure includes data format organization, field identification, and semantic constraints.

[0046] This invention also provides a system for parsing and generating execution plans for space robotic arms, comprising:

[0047] The instruction document parsing module is used to extract structured mission events from flight mission documents in natural language form through a large language model combined with a rule system.

[0048] The state transition diagram and path planning module is used to parse the target configuration contained in the obtained task events, calculate the shortest path from the current configuration to the target configuration of the robotic arm, and complete the path planning.

[0049] The time planning and telemetry and control resource scheduling module is used to complete the time scheduling and station selection for each event based on the path planning, combined with the track circle number, telemetry and control resource availability, lighting and occlusion conditions;

[0050] The atomic event abstraction and control instruction mapping module is used to decompose events into atomic-level actions in chronological order and map them into a sequence of underlying control instructions.

[0051] The large model and tool / knowledge base collaboration mechanism is used throughout the process. The large language model works in conjunction with the external path planner, event scheduler and configuration database through function call mechanism or knowledge base query interface.

[0052] Furthermore, the instruction document parsing module is also used for:

[0053] Assuming the original flight control mission command document is a natural language sequence, the goal is to parse the set of mission events from it:

[0054]

[0055] Where i is the i-th event, N is the total number of events, and t ie is the event start time expression. i d is the event name. i For event instruction code, a i Special notes are added for events; the task of the large language model is to map the input flight control mission instruction document into a structured event list ε, ultimately obtaining an event structured table:

[0056]

[0057] Furthermore, the state transition diagram and path planning module is also used for:

[0058] Pathfinding is achieved using breadth-first search or heuristic search, with the goal of finding the path within a given initial state s. init and target state s goal Finding the shortest path between the given conditions minimizes the total cost; this process is encapsulated in the path lookup function PlanPath(s). init ,s goal The output is the intermediate configuration path and its action sequence.

[0059] Furthermore, the time planning and telemetry resource scheduling module is also used for:

[0060] S31, e obtained from each parsing j Contains a time expression t relative to the task reference time T0 j The system first standardizes it into an offset δ j =Parse(t j The offset represents the expected execution time of the event in the task flow; Parse is the time offset conversion function; j is the j-th event;

[0061] S32, Given each ground tracking and control resource r i′ Its set of visible windows is This represents the start time of the k-th visible window, which belongs to the i′-th station. This represents the end time of the k-th visible window, which belongs to the i′-th station, where k represents the count of the visible windows. i′ This represents the total number of visible windows, where i′ represents the i′-th station, i.e., the ground control and measurement resource; the system sequentially judges event e. j Does each resource window meet the executable requirements?

[0062] S33. For resource windows that meet all conditions, the system records event e. j The candidate execution records are:

[0063]

[0064] in: For the event at station r i′ The executable window on; p is the offset time of the event execution relative to the start point of the window. j These are the event parameters; the final result constitutes the event scheduling table:

[0065]

[0066] Furthermore, the system sequentially judges event e. j Whether the executable conditions are met in each resource window, including:

[0067] Event e j For an operation to be executed in each resource window, the following three conditions must be met simultaneously.

[0068] 1) Track visibility constraint: The offset execution time point should be within a certain resource window, i.e.:

[0069]

[0070] 2) Geometric visibility constraint: There is no Earth obstruction between the target and the telemetry and control station, and the antenna field of view meets the communication requirements, i.e.:

[0071] Vis(e j ,r i′ ,T0+δ j ) = True

[0072] Vis represents a function that calculates whether there is Earth obstruction between the tracking station and the target, i.e. whether the target is on the other side of the Earth from the location of the tracking station. True indicates that the condition is true.

[0073] 3) Lighting and attitude constraints: If the event requires lighting conditions, then the following must be satisfied simultaneously:

[0074] θ sun (T0+δ j )∈[θ min ,θ max ]

[0075]

[0076] E sun =I0·cos(θ) sun )≥E min

[0077] Where, θ sun () represents the angle between the direction of solar incidence and the normal to the mission-sensitive surface, [θ min ,θ max ] represents the permissible range of incident angles, POS(e j ,T0+δj ) indicates that at time T0+δ j At that time, the task event e is executed. j The on-orbit target position, Ω shadow () indicates the Earth's shadow area; E sun Represents the effective incident energy, I0 is the solar constant, and E min This is the minimum effective incident energy;

[0078] Thus event e j The feasibility of execution is determined as follows:

[0079]

[0080] τ j Indicates event e j The start time of the allocation.

[0081] Furthermore, the atomic event abstraction and control instruction mapping module is also used for:

[0082] Abstract events into atomic events in, This represents the h-th atomic event selected from the atomic event library. T is a collection of encapsulated atomic events. h P is the start time of the atomic event execution. h Let A be a parameter vector, and K be the total number of atomic events; each A h Corresponding to a group of multiple events The task blocks constituted; the scheduling process of atomic events on the time axis is T. h+1 =T h +ΔT h +d h , where d h Let ΔT be the execution duration of the h-th atomic event. h To allow for the interval between events, the atomic event scheduling table is finally obtained:

[0083]

[0084] Furthermore, the large language model includes a function tool invocation mechanism, a retrieval enhancement generation mechanism, and a response structure fusion mechanism. The function tool invocation mechanism refers to defining a set of external functions. During the parsing of user task instructions, the large language model determines whether to trigger a certain external function based on the embedded prompt structure or model reasoning. The function return value is then used as context input to the large language model for the next step of reasoning. The retrieval enhancement generation mechanism refers to automatically generating retrieval query instructions when the large language model needs to reason about a certain task but lacks relevant background knowledge. It calls the knowledge base interface to perform similarity retrieval, and the large language model concatenates the returned content into the current context to achieve dynamic injection of external knowledge. The response structure fusion mechanism refers to concatenating the explanatory text output by the large language model, the structured results returned by the function module, and the standardized output template structure. The standardized output template structure includes data format organization, field identification, and semantic constraints.

[0085] The advantages of this invention are:

[0086] (1) This invention utilizes the powerful natural language understanding capabilities of a large language model to extract structured mission events from semantically complex flight mission documents. The large language model serves as the parsing kernel, outputting structured events under specified prompt templates. This fully leverages the language understanding capabilities of the pre-trained large language model, enhancing the robustness and generalization of the system's multimodal information mapping across multiple scenarios. By combining path planning and timing scheduling, and considering the constraints of telemetry and control resources, the system automatically generates orbital mission execution plans that meet the constraints. In other words, by combining orbit circle numbers, telemetry and control resource availability, and lighting and occlusion conditions, the system completes the timing scheduling and station selection for each event, improving the accuracy and real-time performance of path planning. Events are decomposed into atomic-level actions in chronological order and mapped to a sequence of underlying control commands, further optimizing the efficiency of mission timing and resource selection. Thus, the overall solution demonstrates strong integration and generalization capabilities in handling space mission timing, resource selection, and multimodal information mapping.

[0087] (2) This invention utilizes a large language model to directly parse human natural language instructions, achieving a fully automated process from instruction to execution plan, significantly reducing manual intervention and rule design workload, and exhibiting a high degree of automation. The method based on a pre-trained large model has good generalization ability for different task scenarios, adapting to varying task descriptions and environmental conditions. For example, a robot system driven by natural language instructions can still complete long-term tasks in complex environments; the use of graph structure and state space simplification strategies ensures minimal performance degradation in new scenarios. The overall solution has strong generalization ability and robustness. This invention combines graph search algorithms and time scheduling optimization, enabling accurate calculation of optimal paths and time arrangements while meeting complex constraints (such as resource visibility windows, event dependencies, etc.). This invention comprehensively considers orbital observation geometric constraints and telemetry and control resource limitations during modeling, integrating factors such as visibility windows, task priorities, and time offsets into the planning framework. Thus, even in complex environments with multiple satellites and ground stations, feasible scheduling windows that satisfy all constraints can be selected, enabling the overall solution to handle complex constraints.

[0088] (3) Regarding task scheduling, this invention generates candidate execution records for each parsed event. Record items include station identifier, available window start and end, event relative window offset, and parameter vector. The system uses the candidate set as an intermediate representation of the scheduling, and performs feasibility checks on the candidate records by calling time standardization and window matching functions item by item, ultimately forming an event scheduling table that satisfies physical constraints. This intermediate representation decouples the scheduling logic from the input representation, thereby improving the system's reusability and generalization ability for different task descriptions and different telemetry and control resource tables. Regarding resource selection, the resource selection module encapsulates track visibility, geometric occlusion, and illumination / attitude determination into independent callable functions, and filters each candidate execution record as input. The judgment process can be combined with a retrieval-enhanced generation (RAG) mechanism, injecting station capability tables and historical records as auxiliary evidence in real time. The return value of the function call is used for immediate feasibility judgment; candidates that do not meet the conditions will be eliminated or marked as requiring manual confirmation, thereby ensuring the engineering feasibility of the scheduling results. In terms of multi-source information mapping, this invention uniformly converts various inputs (natural language instructions, configuration descriptions, resource tables, etc.) into predefined text templates. A Large Language Model (LLM) acts as the parsing kernel, outputting structured events under specified prompt templates. When the parsing results require precise numerical support, the LLM triggers a functionalization tool to obtain the numerical calculation results and incorporates them into the final structured output. This unified textualization strategy, combined with the functionalization tool, leverages the language understanding capabilities of the pre-trained LLM while ensuring the accuracy of numerical judgments, thus enhancing the system's robustness and generalization capabilities across multiple scenarios.

[0089] (4) To address the problems of fuzzy natural language descriptions and difficulty in accurately extracting domain constraints in traditional flight mission instruction parsing, this invention designs an intelligent parsing framework based on a unified text modality. By converting natural language instructions, configuration descriptions, resource tables, and path information into structured text input, the Large Language Model (LLM) completes the semantic extraction and accurate mapping of mission events. This design simplifies the complexity of multimodal fusion, enhances the system's adaptability to different mission descriptions and input formats, and improves the accuracy and generalization of mission parsing. To address the limitations of traditional path planning, such as its inability to dynamically cope with complex aerospace mission environments and limited resource constraints, this invention proposes an integrated constraint optimization path planning method. By introducing a heuristic search algorithm based on mission requirements, this method can find the optimal path among complex orbital constraints, resource conditions, and mission objectives, and considers the transformation of robotic arm configurations and the calculation of the shortest path, improving the accuracy and real-time performance of path planning. To address the problems of high availability judgment and low efficiency of cross-mission resource sharing in traditional mission scheduling, this invention designs a resource scheduling framework based on a retrieval-enhanced generation (RAG) mechanism. By introducing an intelligent agent with knowledge of spacecraft control, relevant telemetry, tracking, and command (TT&C) resources and observation window forecasts are dynamically retrieved during mission analysis and injected into the large language model as auxiliary evidence, further optimizing the efficiency of mission scheduling and resource selection. This method not only enhances the generalization ability of missions under new TT&C resources but also improves the engineering reliability of the scheduling scheme. Addressing the problems of low efficiency in fine-grained instruction generation and difficulty in parallel scheduling in traditional mission planning systems, this invention proposes a control instruction generation method based on atomic event abstraction. By decomposing mission events into atomic-level actions in chronological order and mapping them to standardized low-level control instruction sequences, redundancy and complexity in the instruction mapping process are reduced, improving system execution efficiency and instruction reusability. This method supports parallel scheduling and combinatorial optimization of mission events, significantly improving the system's response speed and operational stability. Attached Figure Description

[0090] Figure 1 This is a system framework diagram of a method for parsing and generating execution plans for a space robotic arm according to an embodiment of the present invention;

[0091] Figure 2 This is a schematic diagram of instruction parsing in a space robotic arm task instruction parsing and execution plan generation method disclosed in an embodiment of the present invention;

[0092] Figure 3 This is a schematic diagram of atomic event abstraction and control instruction mapping in a space robotic arm task instruction parsing and execution plan generation method disclosed in an embodiment of the present invention. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0094] Example 1

[0095] During space mission execution, flight controllers typically face a series of challenges: First, they must accurately extract mission information (such as event names, execution conditions, time offsets, etc.) from complex, unstructured flight procedure documents; second, some missions involve configuration switching of robotic arms or spacecraft, requiring the planning of optimal paths in state transition diagrams; third, all events must be scheduled for execution within available orbital windows and ground telemetry and control resources, with scheduling meeting physical constraints such as lighting, occlusion, and communication; furthermore, mission execution units must possess controllable and combinable operational granularity, translating complex language instructions into low-level action sequences; finally, throughout the entire process, an intelligent engine is needed to understand human language, connect to knowledge bases, and invoke computational functions to achieve automatic mission closure.

[0096] Based on the above five typical requirements, this invention proposes a method for parsing mission instructions and generating execution plans for a space robotic arm, such as... Figure 1 As shown, the system consists of the following five modules: (1) instruction document parsing module; (2) state transition diagram and path planning module; (3) time planning and measurement and control resource scheduling module; (4) atomic event abstraction and control instruction mapping module; (5) large model and tool / knowledge base collaboration mechanism. The flight mission document in natural language format is first input into the instruction parsing module, which extracts structured mission events through a large language model combined with a rule system. The parsed mission events, containing motion requirements such as target configuration, are sent to the state transition diagram path planning module, which automatically calculates the shortest path for the robotic arm from the current configuration to the target configuration. The path planning results are then input into the time planning and telemetry resource scheduling module, which, considering factors such as orbit circle number, telemetry resource availability, lighting and occlusion conditions, completes the time scheduling and station selection for each event. Subsequently, the events are broken down into atomic-level actions in chronological order, and the atomic event abstraction and control instruction mapping module maps them into specific low-level control instruction sequences. Throughout this process, the large language model, through function call mechanisms or knowledge base query interfaces, collaborates with the external path planner, event scheduler, and configuration database to form an intelligent closed-loop system for mission parsing and scheduling. The overall system structure is as follows: Figure 1 As shown, the details are as follows:

[0097] (1) Command document parsing module

[0098] Suppose the original flight control mission command document is a natural language sequence D = w1, w1, ..., w T The goal is to extract the set of task events from it:

[0099]

[0100] Where t i For the event start time expression (e.g., "T0+1m30s"), e i For the event name (e.g., "Release Grippers"), d i For event instruction code, a i Special notes are added for events. The task of the large language model is to map the input instruction document D into a structured list of events ε.

[0101] The large language model is defined as a conditional probabilistic generative model P(Y|X), where X is a combination of instructions and several example prompts, and Y is a structured event element. The parsing objective of the large language model is to maximize the following log-likelihood:

[0102]

[0103] To improve the stability and controllability of the analysis, this module divides the task into sub-objectives, such as... Figure 2 As shown:

[0104] Time offset extraction: Normalizing the time expression to a relative number of seconds t i =ToSeconds(T j +Δt)

[0105] / R command identification: using a rule base to determine e i Does it belong to the category of instructions that require a response / jump?

[0106] Instruction code extraction: Extract the instruction code string d based on format rules (e.g., "TPXXXXG"). i

[0107] The final structured event table is obtained as follows:

[0108]

[0109] (2) State transition diagram and path planning module

[0110] Let the set of reachable states of the space robotic arm be . State transitions can be achieved using a directed graph. It means that the side Indicates from state s i Can be transferred to s jThe transfer cost is c(s) i ,s j )≥0.

[0111] The goal is to achieve the desired result in a given initial state s. init and target state s goal Find the shortest path π = {s0, s1, ..., s} L}, which minimizes the total cost:

[0112]

[0113] This module employs breadth-first search (BFS) or heuristic search (such as A...). * Path finding is performed. To adapt to the large model inference structure, the state graph G is converted into the following two structures:

[0114] Adjacency list

[0115] Edge mapping table

[0116] This module is encapsulated as a path lookup function PlanPath(s init ,s goal The output is the intermediate configuration path and its action sequence, which can be called by the atomic event planning module.

[0117] (3) Time Planning and Measurement & Control Resource Scheduling Module

[0118] Task events must be executed under physical constraints such as the availability of telemetry and control resources and sufficient lighting. Let the set of task events be ε={e1,…,e N The orbital circle number is c∈Z + Ground-based telemetry and control resources are set up as The available windows for each resource are:

[0119]

[0120] The goal of time scheduling is to schedule each event e j Assign start time τ j The following multiple constraints must be satisfied:

[0121] 1. The time scheduling meets the availability window of telemetry and control resources, i.e., there are ground stations available. Make:

[0122]

[0123] 2. The lighting conditions meet the mission requirements, specifically including the following:

[0124] The star target at that time τ j Located in the sunlit area, that is:

[0125] θ sun (τ j )∈[θ min ,θ max ]

[0126] Where θ sun (τ j ) represents the angle between the direction of solar incidence and the normal to the mission-sensitive surface, [θ] min ,θ max [ ] represents the permissible range of incident angles.

[0127] No Earth obstruction, meaning the solar vector and the spacecraft's line of sight relative to Earth are not blocked by the Earth itself. (Using Ω) shadow (τ j If ) represents the Earth's shadow region, then:

[0128]

[0129] If the mission involves requirements such as solar power supply or optical imaging, the effective incident energy E under attitude control must be met simultaneously. sun ≥E min ,in:

[0130] E sun =I0·cos(θ) sun ), where I0 is the solar constant.

[0131] Therefore, event e j The feasibility of execution is determined as follows:

[0132]

[0133] This condition is used by the task scheduling module to filter feasible time windows that meet the physical environment conditions of the task.

[0134] The specific process is as follows:

[0135] 1. Event Time Offset Analysis

[0136] Each parsed task event e j Contains a time expression t relative to the task reference time T0 j (e.g., “T0+5m”), the system first normalizes it into an offset δ. j ,Right now:

[0137] δ j =Parse(t j (Unit: seconds)

[0138] This offset represents the expected execution time of the event within the task flow. Parse represents the time offset transformation function.

[0139] 2. Resource window matching and feasibility screening

[0140] Given each ground tracking and control resource Its set of visible windows is:

[0141]

[0142] This represents the start time of the k-th visible window, which belongs to the i′-th station. This represents the end time of the k-th visible window, which belongs to the i′-th station, where k represents the count of the visible windows. i′ This represents the total number of visible windows, where i′ represents the i′-th station, i.e., the ground control and measurement resource; the system sequentially judges event e. j To determine whether an item is executable within a resource window, it must simultaneously meet the following three conditions:

[0143] Track visibility constraint: The offset execution time point should be within a certain resource window, i.e.:

[0144]

[0145] Geometric visibility constraint: There is no Earth obstruction between the target and the telemetry and control station, and the antenna field of view meets the communication requirements, i.e.:

[0146] Vis(e j ,r i′ ,T0+δ j ) = True

[0147] Vis represents a function that calculates whether there is Earth obstruction between the tracking station and the target, i.e. whether the target is on the other side of the Earth from the location of the tracking station. True indicates that the condition is true.

[0148] Illumination and attitude constraints: If the event involves imaging, power supply, or other conditions requiring illumination, then the following conditions must be met simultaneously:

[0149] θ sun (T0+δ j )∈[θ min ,θ max ]

[0150]

[0151] E sun =I0·cos(θ) sun )≥E min

[0152] 3. Event scheduling table generation

[0153] For a resource window that meets all the above conditions, the system logs event e.j The candidate execution records are:

[0154]

[0155] in:

[0156] For this event at station r i′ The executable window on;

[0157] This is the offset of the event execution time relative to the start point of the window;

[0158] p j These are event parameters (such as configuration, port number, execution mode, etc.).

[0159] The final results constitute the event scheduling table:

[0160]

[0161] This event scheduling table is a key intermediate representation layer in the system, recording the candidate execution opportunities for the parsed flight control mission semantic actions within each resource availability window. The execution time for each event is represented as an offset relative to the available window; its global absolute time or unified scheduling order has not yet been determined.

[0162] The table already meets physical constraints such as track visibility, telemetry and control resource accessibility, and light occlusion, and possesses structured representation and temporal sequence information, but it is still in a "candidate state" within the task organization process. Subsequent modules will build upon this foundation, encapsulating and abstracting events based on an atomic event library and semantic aggregation rules, and completing global time scheduling and control command mapping. Therefore, this event scheduling table plays a crucial bridging role in the entire task planning process, serving as a vital semantic layer interface connecting natural language parsing and control execution generation.

[0163] (4) Atomic event abstraction and control instruction mapping module

[0164] To achieve the automatic conversion of flight control tasks from semantic events to physical execution commands, this invention proposes an atomic event modeling and control mapping method based on task semantic abstraction, constructing an intermediate semantic bridge between the "event scheduling table" and the "control command table." This method, by introducing an atomic event library, realizes a hierarchical mapping control chain from events to actions and then to commands, such as... Figure 3 As shown:

[0165] First, the event scheduling table output by the system in the previous module is in the following form:

[0166]

[0167] These low-granularity events exhibit semantic redundancy and structural repetition at the scheduling level. Therefore, this invention proposes a task semantic encapsulation mechanism that, based on task context and semantic aggregation rules, abstracts continuous event blocks into more semantically consistent atomic events.

[0168]

[0169] in, This indicates the h-th atomic event selected from the atomic event library (such as "grab component", "release target", "switch configuration", etc.). T is a collection of encapsulated atomic events. h For its start execution time, P h This is a parameter vector. Each A h Corresponding to a group of multiple events The resulting task blocks have a mapping relationship:

[0170]

[0171] This process can leverage rule matching or large language models for clustering and classification, achieving semantic compression and unification of task concepts. Atomic events constitute the second layer of representation structure for task scheduling, and their scheduling process on the timeline is as follows:

[0172] T h+1 =T h +ΔT h +d h

[0173] Where d h Let ΔT be the execution duration of the h-th atomic event. h This allows for a time interval between events. The final atomic event scheduling table is obtained as follows:

[0174]

[0175] This scheduling table simplifies the planning process, improves the semantic consistency of task organization, and has strong human-machine consensus and engineering interpretability.

[0176] Subsequently, each atomic event A h The system call mapping function Φ decodes the data into a corresponding set of underlying operation instructions:

[0177]

[0178] in This represents the actual control commands from the operation command library, such as motor rotation commands, gripper closure commands, configuration switching commands, etc. This represents the relative time of the sub-operation. Finally, a unified format control instruction scheduling table is constructed:

[0179]

[0180] This table is the final output of the system, which can be directly accessed by the execution system (such as a flight control command scheduler or a robotic arm control system).

[0181] In summary, this module, by introducing an intermediate atomic event layer, not only enhances the abstraction capabilities of task modeling but also enables efficient and secure transformation from natural language task understanding to device-level control commands. Structurally, it aligns with the three-layer mapping chain of "semantics-event-command" in flight control tasks, providing key technical support for the integration of large models and mission plans.

[0182] (5) Mechanism for integrating large models with tools / knowledge bases

[0183] To achieve closed-loop automation among multiple stages of tasks, such as task parsing, path planning, and scheduling execution, this invention designs a hybrid interaction mechanism that integrates a Large Language Model (LLM), a Function Tooling interface, and a Retrieval-Augmented Generation (RAG) system. This mechanism enables the large model to possess both open-ended natural language understanding capabilities and controllable computational and precise knowledge support capabilities.

[0184] This mechanism mainly consists of the following three components:

[0185] 1) Function tool calling mechanism

[0186] Language models are essentially probabilistic generative models, generating content that is flexible but lacks structural precision. To address this issue, this invention employs a "function call" approach to outsource precise computation tasks to external modules. The set of external functions is defined as follows:

[0187]

[0188] Each f i This represents a callable structured computation interface, for example:

[0189] *f1=PlanPath(s start ,s goal State diagram path planning

[0190] *f2=ComputeTime(e j Task time calculation

[0191] *f3 = FindVisibleStation(t, target): Selecting control stations

[0192] Language models in parsing user task instructions During the process, the system uses embedded prompts or model reasoning to determine whether a function f should be triggered. i Then, fill in the necessary parameters. The function's return value is then used as context input to continue the model's inference process.

[0193] y = LLM(x,f i (x))=LLM(x′) where x′=x∪Result[f i ]

[0194] This mechanism ensures that even when the model itself is not good at numerical reasoning, constraint calculation, graph search, etc., it can perform accurately and return the results to the dialogue context.

[0195] 2) Search Enhancement Generation Mechanism

[0196] While language models possess powerful semantic understanding capabilities, their built-in knowledge is fixed and limited, failing to cover real-time tasks or private domain knowledge. This invention separates knowledge storage from model parameters through the RAG mechanism, constructing an external task knowledge base. Includes: configuration status diagram data, control command specifications, telemetry and control plan and ground station attributes, and aerospace operation procedures.

[0197] When a model needs to reason about a certain task but lacks relevant background knowledge, it will automatically generate a search query. Call the knowledge base interface to perform similarity retrieval:

[0198]

[0199] Where sim(·,·) is the semantic similarity calculation function, and θ is the relevance threshold. The model concatenates the returned content into the current context, achieving dynamic injection of external knowledge.

[0200] x′=x∪Retrieve(q)

[0201] Enhance the accuracy of understanding task instructions by combining them with context.

[0202] 3) Response structure fusion mechanism

[0203] To achieve a complete conversion from natural language task instructions to machine-executable control plans, this module designs a ternary response structure fusion strategy of "language generation + function result + structure template". This strategy not only retains the language description capabilities generated by the large model, but also combines external function return values ​​with a unified data structure template to ensure that the output is both readable and callable.

[0204] The overall response output is represented as follows:

[0205]

[0206] in, This indicates a structural splicing operation. y text Explanatory text output by a large language model; y func For structured results returned by function modules (such as path sequences, scheduling times); y schema This standardizes the output template structure, used to organize data formats, field identifiers, semantic constraints, etc.

[0207] This structure fusion mechanism ensures that the output is both readable and executable, and supports the direct generation of intermediate task representations (such as JSON / tables) for scheduling or control systems.

[0208] By combining the above three mechanisms, this invention realizes a closed-loop task control framework of natural language-structured reasoning-precise computation-knowledge injection, enabling large models to flexibly understand, accurately execute and continuously enhance in real space tasks, and greatly improving the system's generalization, robustness and engineering adaptability.

[0209] Through the above technical solutions, this invention utilizes a large language model to directly parse human natural language commands, achieving a fully automated process from command to execution plan, significantly reducing manual intervention and rule design workload. The method based on a pre-trained large model exhibits excellent generalization ability across different task scenarios, adapting to varying task descriptions and environmental conditions. For example, a robot system driven by natural language commands can still complete long-sequence tasks in complex environments; the use of graph structures and state-space simplification strategies ensures minimal performance degradation in new scenarios. Combining graph search algorithms and time scheduling optimization, this method can accurately calculate the optimal path and timing while meeting complex constraints (such as resource visibility windows and event dependencies). The modeling process comprehensively considers orbital observation geometric constraints and telemetry and control resource limitations, integrating factors such as visibility windows, task priorities, and time offsets into the planning framework. Thus, even in complex environments with multiple satellites and ground stations, feasible scheduling windows that satisfy all constraints can be selected. This invention can be widely applied in space mission planning and control systems, enabling intelligent parsing and automatic scheduling of flight mission commands in real-world mission scenarios. In terms of implementation, it can be embedded as modular software into spacecraft mission management systems, space station ground control platforms, or space mission auxiliary decision-making systems to provide real-time mission analysis and scheduling services; it can also be deployed on backend servers to support batch planning of multiple missions and collaborative scheduling of orbital resources.

[0210] Example 2

[0211] Based on Embodiment 1, Embodiment 2 of the present invention also provides a space robotic arm task instruction parsing and execution plan generation system, including:

[0212] The instruction document parsing module is used to extract structured mission events from flight mission documents in natural language form through a large language model combined with a rule system.

[0213] The state transition diagram and path planning module is used to parse the target configuration contained in the obtained task events, calculate the shortest path from the current configuration to the target configuration of the robotic arm, and complete the path planning.

[0214] The time planning and telemetry and control resource scheduling module is used to complete the time scheduling and station selection for each event based on the path planning, combined with the track circle number, telemetry and control resource availability, lighting and occlusion conditions;

[0215] The atomic event abstraction and control instruction mapping module is used to decompose events into atomic-level actions in chronological order and map them into a sequence of underlying control instructions.

[0216] The large model and tool / knowledge base collaboration mechanism is used throughout the process. The large language model works in conjunction with the external path planner, event scheduler and configuration database through function call mechanism or knowledge base query interface.

[0217] Specifically, the instruction document parsing module is also used for:

[0218] Assuming the original flight control mission command document is a natural language sequence, the goal is to parse the set of mission events from it:

[0219]

[0220] Where i is the i-th event, N is the total number of events, and t i e is the event start time expression. i d is the event name. i For event instruction code, a i Special notes are added for events; the task of the large language model is to map the input flight control mission instruction document into a structured event list ε, ultimately obtaining an event structured table:

[0221]

[0222] Specifically, the state transition diagram and path planning module is also used for:

[0223] Pathfinding is achieved using breadth-first search or heuristic search, with the goal of finding the path within a given initial state s. init and target state s goal Finding the shortest path between the given conditions minimizes the total cost; this process is encapsulated in the path lookup function PlanPath(s). init ,s goalThe output is the intermediate configuration path and its action sequence.

[0224] Specifically, the time planning and measurement and control resource scheduling module is also used for:

[0225] S31, e obtained from each parsing j Contains a time expression t relative to the task reference time T0 j The system first standardizes it into an offset δ j =Parse(t j The offset represents the expected execution time of the event in the task flow; Parse is the time offset conversion function; j is the j-th event;

[0226] S32, Given each ground tracking and control resource r i′ Its set of visible windows is This represents the start time of the k-th visible window, which belongs to the i′-th station. This represents the end time of the k-th visible window, which belongs to the i′-th station, where k represents the count of the visible windows. i′ This represents the total number of visible windows, where i′ represents the i′-th station, i.e., the ground control and measurement resource; the system sequentially judges event e. j Does each resource window meet the executable requirements?

[0227] S33. For resource windows that meet all conditions, the system records event e. j The candidate execution records are:

[0228]

[0229] in: For the event at station r i′ The executable window on; p is the offset time of the event execution relative to the start point of the window. j These are the event parameters; the final result constitutes the event scheduling table:

[0230]

[0231] More specifically, the system sequentially judges event e j Whether the executable conditions are met in each resource window, including:

[0232] Event e j For an operation to be executed in each resource window, the following three conditions must be met simultaneously.

[0233] 1) Track visibility constraint: The offset execution time point should be within a certain resource window, i.e.:

[0234]

[0235] 2) Geometric visibility constraint: There is no Earth obstruction between the target and the telemetry and control station, and the antenna field of view meets the communication requirements, i.e.:

[0236] Vis(e j ,r i′ ,T0+δ j ) = True

[0237] Vis represents a function that calculates whether there is Earth obstruction between the tracking station and the target, i.e. whether the target is on the other side of the Earth from the location of the tracking station. True indicates that the condition is true.

[0238] 3) Lighting and attitude constraints: If the event requires lighting conditions, then the following must be satisfied simultaneously:

[0239] θ sun (T0+δ j )∈[θ min ,θ max ]

[0240]

[0241] E sun =I0·cos(θ) sun )≥E min

[0242] Where, θ sun () represents the angle between the direction of solar incidence and the normal to the mission-sensitive surface, [θ min ,θ max ] represents the permissible range of incident angles, POS(e j ,T0+δ j ) indicates that at time T0+δ j At that time, the task event e is executed. j The on-orbit target position, Ω shadow () indicates the Earth's shadow area; E sun Represents the effective incident energy, I0 is the solar constant, and E min This is the minimum effective incident energy;

[0243] Thus event e j The feasibility of execution is determined as follows:

[0244]

[0245] τ j Indicates event e j The start time of the allocation.

[0246] More specifically, the atomic event abstraction and control instruction mapping module is also used for:

[0247] Abstract events into atomic events in, This represents the h-th atomic event selected from the atomic event library. T is a collection of encapsulated atomic events. h P is the start time of the atomic event execution. h Let A be a parameter vector, and K be the total number of atomic events; each A h Corresponding to a group of multiple events The task blocks constituted; the scheduling process of atomic events on the time axis is T. h+1 =T h +ΔT h +d h , where d h Let ΔT be the execution duration of the h-th atomic event. h To allow for the interval between events, the atomic event scheduling table is finally obtained:

[0248]

[0249] Specifically, the large language model includes a function tool invocation mechanism, a retrieval enhancement generation mechanism, and a response structure fusion mechanism. The function tool invocation mechanism refers to defining a set of external functions. During the parsing of user task instructions, the large language model determines whether to trigger a certain external function based on the embedded prompt structure or model reasoning. The function return value is then used as context input to the large language model for the next step of reasoning. The retrieval enhancement generation mechanism refers to automatically generating retrieval query instructions when the large language model needs to reason about a certain task but lacks relevant background knowledge. It calls the knowledge base interface to perform similarity retrieval, and the large language model concatenates the returned content into the current context to achieve dynamic injection of external knowledge. The response structure fusion mechanism refers to concatenating the explanatory text output by the large language model, the structured results returned by the function module, and the standardized output template structure. The standardized output template structure includes data format organization, field identification, and semantic constraints.

[0250] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for parsing mission instructions and generating execution plans for a space robotic arm, characterized in that, include: S1. Flight mission documents in natural language form are used to extract structured mission events through a large language model combined with a rule system; S2. The parsed task events contain the target configuration. Calculate the shortest path from the current configuration to the target configuration for the robotic arm, and complete the path planning. S3. Based on the events of the path planning, combined with the track circle number, the availability of telemetry and control resources, and the lighting and occlusion conditions, complete the time scheduling and station selection for each event; S4. Events are broken down into atomic actions in chronological order and mapped to a sequence of underlying control instructions. Throughout the process, the large language model works in collaboration with the external path planner, event scheduler, and configuration database through function call mechanisms or knowledge base query interfaces.

2. The method for parsing and generating execution plans for a space robotic arm according to claim 1, characterized in that, S1 includes: Assuming the original flight control mission command document is a natural language sequence, the goal is to parse the set of mission events from it: Where i is the i-th event, N is the total number of events, and t i The event start time expression, e i d is the event name. i For event instruction code, a i Special notes are added for events; the task of the large language model is to map the input flight control mission instruction document into a structured event list ε, ultimately obtaining an event structured table:

3. The method for parsing and generating execution plans for a space robotic arm according to claim 1, characterized in that, S2 include: Pathfinding is achieved using breadth-first search or heuristic search, with the goal of finding the path within a given initial state s. init and target state s goal Finding the shortest path between the given conditions minimizes the total cost; this process is encapsulated in the path lookup function PlanPath(s). init ,s goal The output is the intermediate configuration path and its action sequence.

4. The method for parsing and generating execution plans for a space robotic arm according to claim 1, characterized in that, S3 include: S31, e obtained from each parsing j Contains a time expression t relative to the task reference time T0 j The system first standardizes it into an offset δ j =Parse(t j The offset represents the expected execution time of the event in the task flow; Parse is the time offset conversion function; j is the j-th event; S32, Given each ground tracking and control resource r i′ Its set of visible windows is This represents the start time of the k-th visible window. K represents the end time of the k-th visible window. i′ This represents the total number of visible windows, where i′ represents the i′-th station, i.e., the ground control and measurement resource; the system sequentially judges event e. j Does each resource window meet the executable requirements? S33. For resource windows that meet all conditions, the system records event e. j The candidate execution records are: in: For the event at station r i′ The executable window on; p is the offset time of the event execution relative to the start point of the window. j These are the event parameters; the final result constitutes the event scheduling table:

5. The method for parsing and generating execution plans for a space robotic arm according to claim 4, characterized in that, The system sequentially judges event e j Whether the executable conditions are met in each resource window, including: Event e j For an operation to be executed in each resource window, the following three conditions must be met simultaneously. 1) Track visibility constraint: The offset execution time point should be within a certain resource window, i.e.: 2) Geometric visibility constraint: There is no Earth obstruction between the target and the telemetry and control station, and the antenna field of view meets the communication requirements, i.e.: Vis(e j ,r i′ ,T0+δ j )=True Vis represents a function that calculates whether there is Earth obstruction between the telemetry and control station and the target; True indicates that the condition is true. 3) Lighting and attitude constraints: If the event requires lighting conditions, then the following must be satisfied simultaneously: i sun (T0+δ j )∈[θ min ,i max ] E sun =I0·cos(θ sun )≥E min Where, θ sun () represents the angle between the direction of solar incidence and the normal to the mission-sensitive surface, [θ min ,θ max ] represents the permissible range of incident angles, POS(e j ,T0+δ j ) represents time T0+δ j At that time, the task event e is executed. j The on-orbit target position, Ω shadow () indicates the Earth's shadow area; E sun Represents the effective incident energy, I0 is the solar constant, and E min This is the minimum effective incident energy; Thus event e j The executability is determined as follows: τ j Indicates event e j The start time of the allocation.

6. The method for parsing and generating execution plans for a space robotic arm according to claim 4, characterized in that, S4 include: Abstract events into atomic events in, This represents the h-th atomic event selected from the atomic event library. T is a collection of encapsulated atomic events. h P is the start time of the atomic event execution. h Let A be a parameter vector, and K be the total number of atomic events; each A h Corresponding to a group of multiple events The task blocks constituted; the scheduling process of atomic events on the time axis is T. h+1 =T h +ΔT h +d h , where d h Let ΔT be the execution duration of the h-th atomic event. h To allow for the interval between events, the atomic event scheduling table is finally obtained:

7. The method for parsing and generating execution plans for a space robotic arm according to claim 1, characterized in that, The large language model includes a function tool invocation mechanism, a retrieval enhancement generation mechanism, and a response structure fusion mechanism. The function tool invocation mechanism defines a set of external functions. During the parsing of user task instructions, the large language model determines whether to trigger a specific external function based on embedded prompts or model reasoning. The function's return value is then used as context input for further reasoning. The retrieval enhancement generation mechanism automatically generates retrieval query instructions when the large language model lacks relevant background knowledge for reasoning about a task. It calls a knowledge base interface for similarity retrieval, and the large language model appends the returned content to the current context, achieving dynamic injection of external knowledge. The response structure fusion mechanism combines the explanatory text output by the large language model, the structured results returned by function modules, and the standardized output template structure. The standardized output template structure includes data format organization, field identifiers, and semantic constraints.

8. A system for parsing mission instructions and generating execution plans for a space robotic arm, characterized in that, include: The instruction document parsing module is used to extract structured mission events from flight mission documents in natural language form through a large language model combined with a rule system. The state transition diagram and path planning module is used to parse the target configuration contained in the obtained task events, calculate the shortest path from the current configuration to the target configuration of the robotic arm, and complete the path planning. The time planning and telemetry and control resource scheduling module is used to complete the time scheduling and station selection for each event based on the path planning, combined with the track circle number, telemetry and control resource availability, lighting and occlusion conditions; The atomic event abstraction and control instruction mapping module is used to decompose events into atomic-level actions in chronological order and map them into a sequence of underlying control instructions. The large model and tool / knowledge base collaboration mechanism is used throughout the process. The large language model works in conjunction with the external path planner, event scheduler and configuration database through function call mechanism or knowledge base query interface.

9. The space robotic arm task instruction parsing and execution plan generation system according to claim 8, characterized in that, The instruction document parsing module is also used for: Assuming the original flight control mission command document is a natural language sequence, the goal is to parse the set of mission events from it: Where i is the i-th event, N is the total number of events, and t i The event start time expression, e i d is the event name. i For event instruction code, a i Special notes are added for events; the task of the large language model is to map the input flight control mission instruction document into a structured event list ε, ultimately obtaining an event structured table:

10. A space robotic arm task instruction parsing and execution plan generation system according to claim 8, characterized in that, The state transition diagram and path planning module is also used for: Pathfinding is achieved using breadth-first search or heuristic search, with the goal of finding the path within a given initial state s. init and target state s goal Finding the shortest path between the given conditions minimizes the total cost; this process is encapsulated in the path lookup function PlanPath(s). init ,s goal The output is the intermediate configuration path and its action sequence.

Images