Graph neural network enhanced large model PDDL modeling method and system
By constructing a heterogeneous graph and combining a GNN checker with an LLM for PDDL modeling, the problems of inaccurate logical judgment and separation of verification in existing technologies are solved, and efficient and accurate automated PDDL modeling is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-10
AI Technical Summary
Existing PDDL automated modeling methods lack in-depth logical judgment, have inaccurate localization, and cannot detect predicate definition defects. Furthermore, the separation of validation between domain files and issue files results in limited interpretability and generalization ability.
By constructing a heterogeneous graph and performing semantic embedding, logical relationship judgment is performed using GNN verifier and MLP, and PDDL file is generated by combining LLM and feedback closed-loop correction to achieve node-level positioning and logical consistency assurance.
It improves the verification capability, logical consistency and positioning accuracy of PDDL automated modeling, realizes efficient automated correction and reduces manual intervention.
Smart Images
Figure CN121835830A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of automatic planning, in particular to a graph neural network enhanced large model PDDL modeling method and system. BACKGROUND
[0002] PDDL (Planning Domain Definition Language) is an internationally recognized standard description language for symbolic planning, which can realize fast and reliable task planning through corresponding planners, but manual creation and maintenance of complex PDDL domain (Domain) files and problem (Problem) files are time-consuming and error-prone.
[0003] In existing methods, the SPAR framework proposes a technology for generating PDDL for unmanned aerial vehicle tasks using LLM and performing preliminary verification, first constructs a dataset containing detailed description of professional PDDL files, designs specific thinking chain prompt words to guide LLM to generate PDDL files from natural language input, and checks the syntax through a traditional PDDL parser and judges whether the problem has a solution through an external planner; Jeon et al. used GNN and reinforcement learning methods for planning and solving tasks based on PDDL description, represented the PDDL file as a graph structure for input to the reinforcement learning model, and learned a general planning strategy.
[0004] However, PDDL automatic modeling verification lacks deep logical judgment and is not accurate, and lacks intelligent verification capability for domain abstraction rules and problem instances, which is specifically manifested in: The existing method of generating PDDL by LLM relies on external planners for verification, which can only determine whether the file has syntax errors or the problem has a solution, and cannot find logical defects such as predicates defining requirements but lacking production actions and the completeness of each part. When feedback errors, it is difficult to accurately locate the error position, resulting in low efficiency of LLM correction relying on blind trial and error.
[0005] The existing technology verifies the domain file and the problem file separately, judges the syntax of the domain file through the syntax verifier, and judges whether the problem file has a solution through the planner. This method has limited explainability and generalization ability when facing different problems. SUMMARY
[0006] In view of the deficiencies of the prior art, the present application provides a graph neural network enhanced large model PDDL modeling method and system, which solves the problems of weak verification capability, insufficient logical consistency guarantee, and inaccurate positioning of PDDL automatic modeling.
[0007] To achieve the above purpose, the present application is realized by the following technical scheme: A graph neural network enhanced large model PDDL modeling method, the modeling method comprising: S1. Heterogeneous graph construction and feature embedding The PDDL domain file and the problem file in text format are parsed by using the Python open source library pddlpy, a heterogeneous graph is constructed according to the parsing result of the PDDL document, and the semantic information in the PDDL file is integrated into the graph through embedding; S2. GNN checker construction and parameter training The GNN checker is constructed and trained based on the PDDL public data, and the GNN checker comprises a GNN network and an MLP; the GNN network sets an independent, trainable relationship-specific weight matrix for each PDDL logical relationship corresponding edge, realizing effective propagation of node semantic features; the MLP is used to accurately locate the defects and defect types existing in the PDDL file to specific nodes in the graph; S3. LLM generates PDDL and correction closed loop based on GNN checking feedback The numerical and localized defect prediction output by the GNN checker is converted into accurate and structured natural language feedback, guiding the LLM to automatically and high-quality correct the generated PDDL file.
[0008] Preferably, in S1, the parsing of the PDDL domain file and the problem file comprises extracting from the two: a type declaration set, a predicate declaration set, an action template set, an object instance set, an initial state and a target state.
[0009] Preferably, in S1, the heterogeneous graph comprises a domain subgraph and a problem subgraph; The domain subgraph comprises predicate nodes , action nodes and type nodes ; corresponding to the predicate declaration, action template and type declaration in the PDDL file respectively; and connected according to the premise condition, parameter constraint and action effect of each action, specifically comprising edges corresponding to the premise condition , edges corresponding to the added effect , edges corresponding to the deleted effect and edges corresponding to the parameter type constraint ; The problem subgraph comprises specific object nodes , initial state nodes and target state nodes , and the nodes are connected according to the adjacency relationship defined by the problem.
[0010] Preferably, in the semantic embedding process, for the nodes, the initial features are composed of semantic vectors and structure vectors. in, These are node semantic vectors derived by a pre-trained BERT model based on action templates, predicate declarations, or node state-related descriptions in a PDDL file. A structure vector derived from a pre-trained BERT model, containing node type, number of parameters, and parameter type; This is a vector concatenation function; The node set is ; For edges, only the edge type needs to be labeled, assigning a number to each type of edge, and the set of edge types is: .
[0011] Preferably, S2 specifically includes: Based on the collected PDDL modeling examples, data cleaning and enhancement were performed to obtain high-quality PDDL positive examples without duplicates or errors. PDDL negative examples were constructed by artificially introducing action structure errors, parameter slot errors, predicate errors, and start / endpoint errors, and adding error type annotations. The defect set is as follows: ; Both positive samples labeled 1 and negative samples labeled 0 are transformed into graph representations using S1; Build In a multi-layered GNN network, each layer aggregates feature information between each node and its neighbors through message passing. in, This represents the current GNN network layer. This is a learnable self-looping weight matrix used to preserve the semantic information of the nodes themselves. For each edge type A learnable relation-specific weight matrix used to represent semantic transformation rules of different logical relations; For type Edges and nodes A set of connected nodes; This is a normalization constant; This is the activation function for this layer; After After layer iterations, each node in the graph incorporates a large amount of local information, resulting in the final embedding matrix of the entire graph. This contains the logical information of the entire graph, which can be used to verify whether the PDDL modeling is complete and logically correct; for each defect... Construct a two-layer MLP to achieve support for any node. Final verification: in, , These are the weights and biases of the first layer of the MLP, respectively. , These are the weights and biases of the second layer of the MLP, respectively. ReLU is the activation function corresponding to the first layer of the MLP; Sigmoid is the activation function corresponding to the second layer of the MLP; From the range of Sigmoid values, we can know ; In GNN networks and MLP, , , All parameters are learnable; a joint loss function is constructed using cross-entropy loss: in, For the first The weight hyperparameters for this type of defect; The set of nodes in the sample that are associated with this type of defect; This is the actual label of the node; Based on this, it can be minimized using the Adam optimizer. This yields the GNN network and MLP used to judge the effectiveness of PDDL modeling, and sets validation results for each defect. Is the threshold acceptable? .
[0012] Preferably, S3 specifically includes: Typical cases from the PDDL positive examples are injected into the LLM as prompt words. The LLM then generates the initial PDDL domain file based on the user's task description context. and problem files At this point, the prompt words used by LLM are represented as follows: ; The next Step, via S1 and Both are converted into graph representations. The GNN verifier obtained from S2 is used to verify whether the two PDDL files are complete and logically sound, and the verification results for each node and each possible error are obtained. ;like This indicates that the node has a modeling logic error, and the node is removed. With error Integrate into a natural language description using a certain template. All errors are grouped into the feedback prompt word "Error" for this round. If "Error" is not an empty set, the constructed PDDL file is considered to have a defect. Let: Using new prompts, the LLM algorithm corrects the generated PDDL file, resulting in new modeling outcomes. and And proceed to the next round of verification; otherwise, the current... and This is the final PDDL modeling result file for the natural language description task.
[0013] A graph neural network-enhanced large model PDDL modeling system, the modeling system comprising: a heterogeneous graph construction and feature embedding module, a GNN validator construction and parameter training module, and a large model modeling and correction module; The heterogeneous graph construction and feature embedding module is used to parse PDDL domain files and problem files in text format, construct heterogeneous graphs based on the parsing results of PDDL documents, and perform refined semantic embedding based on the heterogeneous graphs. The GNN validator construction and parameter training module constructs and trains a GNN validator based on the publicly available PDDL data. This validator is used to determine whether the logical relationships in the PDDL file are reasonable, and serves as a validator in the LLM PDDL generation process. The large model modeling and correction module is used to transform the numerical and localized defect predictions output by the GNN verifier into accurate and structured natural language feedback, guiding the LLM to perform automated and high-quality correction of the generated PDDL file.
[0014] This invention provides a method and system for modeling large-scale PDDL models enhanced by graph neural networks. Compared with existing technologies, it has the following advantages: In this invention, the modeling method transforms PDDL domain files and problem files into graph representations and performs semantic embedding, enabling the GNN validator and LLM to understand the common-sense semantics and physical attributes behind these graph symbols, ensuring logical consistency. The feedback loop correction method using the GNN validator can more effectively detect problems such as missing content, logical errors, or predicates that are consumed but not generated, helping the LLM to adjust the modeling results more accurately. The MLP in the GNN validator can accurately locate the defects and defect types in the PDDL file at the node level, making the PDDL automated modeling have strong verification capabilities, ensure logical consistency, and provide accurate location. Attached Figure Description
[0015] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart of PDDL modeling as described in an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. 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.
[0018] This application provides a graph neural network-enhanced large-scale model (PDDL) modeling method and system, which solves the problems of weak verification capability, insufficient logical consistency guarantee, and inaccurate positioning in PDDL automated modeling.
[0019] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0020] Example: like Figure 1 As shown, this invention provides a large-scale PDDL modeling method enhanced by graph neural networks, the modeling method comprising: S1. Heterogeneous graph construction and feature embedding based on PDDL This step aims to transform PDDL text code into a numerical heterogeneous graph structure that can be processed by GNN. By fusing semantic and symbolic features, it lays the foundation for subsequent logical reasoning. This step is applied in the construction of GNN validators and in the PDDL modeling part using LLM.
[0021] The Python open-source library pddlpy is used to parse PDDL domain files and problem files in text format, and extract the following from them: type declaration set, predicate declaration set, action template set, object instance set, initial state, and target state.
[0022] Based on the parsing results of the PDDL document, a joint heterogeneous graph is constructed, consisting of a domain subgraph and a problem subgraph; the domain subgraph includes: predicate nodes. Action nodes and type nodes These correspond to the predicate declarations, action templates (e.g., `action move`), and type declarations (e.g., `robot`, `location`) in the PDDL file, respectively. Edges are then connected based on the preconditions, parameter constraints, and action effects corresponding to each action, specifically including edges for preconditions such as corresponding state constraints. Edges with added effects, such as corresponding insertion positions. Edges corresponding to deletion effects such as removal positions And the edges of parameter type constraints such as the type of the corresponding object. ; The problem subgraph includes: specific object nodes. Initial state node and target state node Each node is connected according to the adjacency relationship defined in the problem.
[0023] Semantic information from PDDL files is embedded into the graph, allowing the GNN to better understand the physical meaning behind PDDL symbols. For nodes, the initial features are composed of semantic vectors and structural vectors. in, These are node semantic vectors derived by the pre-trained BERT model based on action descriptions, predicate meanings, or node state-related descriptions in the PDDL file. A structure vector derived from a pre-trained BERT model, containing node type, number of parameters, and parameter type; This is a vector concatenation function; The node set is ; For edges, only the edge type needs to be labeled, assigning a number to each type of edge, and the set of edge types is: .
[0024] S2. Construction and Parameter Training of a GNN Validator for PDDL Document Validation This step aims to build and train a GNN deep learning model based on a publicly available PDDL database to determine whether the logical relationships in the PDDL file are reasonable, serving as a validator in the LLM PDDL generation process.
[0025] Based on the collected PDDL modeling examples, data cleaning and enhancement were performed to obtain high-quality PDDL positive examples without duplicates or errors. Then, by artificially introducing error messages such as action structure errors (e.g., missing conditional predicates or action effects), parameter slot errors (e.g., reversing the start and end points of a move action), predicate errors (e.g., a predicate is consumed but not generated), and start / endpoint errors (e.g., missing end points or inaccessible start / endpoints), and adding error type labels, PDDL negative examples were constructed. The defect set is as follows: .
[0026] Both positive samples labeled 1 and negative samples labeled 0 are transformed into graph representations using S1, which are then used for training the GNN validator.
[0027] Build In a multi-layered GNN network, each layer aggregates feature information between each node and its neighbors through message passing. in, This represents the current GNN network layer. This is a learnable self-looping weight matrix used to preserve the semantic information of the nodes themselves. For each edge type A learnable relation-specific weight matrix used to represent semantic transformation rules of different logical relations; For type Edges and nodes A set of connected nodes; This is a normalization constant; This is the activation function for this layer; After After layer iterations, each node in the graph incorporates a large amount of local information, resulting in the final embedding matrix of the entire graph. This contains the logical information of the entire graph, which can be used to verify whether the PDDL modeling is complete and logically correct; for each defect... Constructing a two-layer multilayer perceptron (MLP) enables the processing of any node. Final verification: in, , These are the weights and biases of the first layer of the MLP, respectively. , These are the weights and biases of the second layer of the MLP, respectively. ReLU is the activation function corresponding to the first layer of the MLP; Sigmoid is the activation function corresponding to the second layer of the MLP; From the range of Sigmoid values, we can know ; In GNN networks and MLP, , , All parameters are learnable; a joint loss function is constructed using cross-entropy loss: in, For the first The weight hyperparameters for this type of defect; The set of nodes in the sample that are associated with this type of defect; This is the actual label of the node; Based on this, optimizers such as Adam can be used to minimize... This yields the GNN network and MLP used to judge the effectiveness of PDDL modeling, and sets validation results for each defect. Is the threshold acceptable? .
[0028] S3.LLM generates PDDL and a corrected closed loop based on GNN verification feedback. This step is the online application phase of the system. It mainly transforms the numerical and localized defect predictions output by the GNN discriminator obtained in S2 into accurate and structured natural language feedback, guiding the LLM to automatically and effectively correct the generated PDDL file.
[0029] Typical cases from the PDDL positive examples are injected into the LLM as prompt words. The LLM then generates the initial PDDL domain file based on the user's task description context. and problem files At this point, the prompt words used by LLM are represented as follows: .
[0030] The next Step, via S1 and All are converted into graph representations. A verifier composed of the GNN and MLP obtained from S2 is used to verify whether the two PDDL files are complete and logically sound, and the verification results for each node and each possible error are obtained. ;like This indicates that the node has a modeling logic error, and the node is removed. With error Integrate into a natural language description using a certain template. (e.g., "The currently generated domain file (:actionmove) lacks action effects," etc.), and all errors are grouped into the feedback prompt word "Error" for this round. If "Error" is not an empty set, then the constructed PDDL file is considered to have a defect. Let: Using new prompts, the LLM algorithm corrects the generated PDDL file, resulting in new modeling outcomes. and And proceed to the next round of verification; otherwise, the current... and This is the final PDDL modeling result file for the natural language description task.
[0031] This invention provides a large-scale model PDDL modeling system enhanced by graph neural networks. The modeling system includes: a heterogeneous graph construction and feature embedding module, a GNN validator construction and parameter training module, and a large-scale model modeling and correction module. The heterogeneous graph construction and feature embedding module is used to parse PDDL domain files and problem files in text format, construct heterogeneous graphs based on the parsing results of PDDL documents, and perform refined semantic embedding based on the heterogeneous graphs. The GNN validator construction and parameter training module constructs and trains a GNN validator based on the publicly available PDDL data. This validator is used to determine whether the logical relationships in the PDDL file are reasonable, and serves as a validator in the LLM PDDL generation process. The large model modeling and correction module is used to transform the numerical and localized defect predictions output by the GNN verifier into accurate and structured natural language feedback, guiding the LLM to perform automated and high-quality correction of the generated PDDL file.
[0032] In summary, compared with the prior art, the present invention has the following beneficial effects: 1. In this embodiment of the invention, the modeling method transforms PDDL domain files and problem files into graph representations and performs semantic embedding, enabling the GNN validator and LLM to understand the common-sense semantics and physical attributes behind these graph symbols, ensuring logical consistency. The method of using the GNN validator for feedback loop correction can more effectively detect problems such as missing content, logical errors, or predicates that are consumed but not generated, helping the LLM to adjust the modeling results more accurately. The MLP in the GNN validator can accurately locate the defects and defect types in the PDDL file at the node level, making the PDDL automated modeling have strong verification capabilities, ensure logical consistency, and provide accurate location.
[0033] 2. In this embodiment of the invention, the modeling method can provide highly interpretable node-level error diagnosis for PDDL files by defining MLPs for different error types, and clearly locate the error location; by using graph construction, BERT semantic embedding and GNN to extract structural information, the model is endowed with common sense reasoning ability. The generalized GNN message passing mechanism and the node-based MLP diagnosis method do not have too many requirements on the structure of the input graph, have strong generalization ability, can meet the modeling needs of various PDDL tasks, and have strong interpretability and generalization ability.
[0034] 3. In this embodiment of the invention, the modeling method utilizes cue word engineering and GNN methods to realize a fully automated closed-loop PDDL modeling process from error detection to correction suggestions to LLM execution of corrections. It has low dependence on expert intervention, high automation efficiency, and low labor costs.
[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0036] 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 modeling large-scale PDDL models enhanced by graph neural networks, characterized in that, The modeling method includes: S1. Heterogeneous Graph Construction and Feature Embedding The Python open-source library pddlpy is used to parse PDDL domain files and issue files in text format. A heterogeneous graph is constructed based on the parsing results of the PDDL documents, and the semantic information in the PDDL files is integrated into the graph through embedding. S2. GNN Verifier Construction and Parameter Training A GNN validator is built and trained based on publicly available PDDL data. The GNN validator consists of a GNN network and an MLP. The GNN network sets an independent and trainable relation-specific weight matrix for each edge corresponding to a PDDL logical relation, enabling effective propagation of node semantic features. The MLP is used to accurately locate defects and defect types in the PDDL file to specific nodes in the graph. S3.LLM generates PDDL and a corrected closed loop based on GNN verification feedback. The numerical and localized defect predictions output by the GNN validator are transformed into accurate and structured natural language feedback, guiding the LLM to perform automated and high-quality correction of the generated PDDL files.
2. The graph neural network-enhanced large-scale PDDL modeling method as described in claim 1, characterized in that, In S1, the parsing of the PDDL domain file and the problem file includes extracting from both: a set of type declarations, a set of predicate declarations, a set of action templates, a set of object instances, an initial state, and a target state.
3. The graph neural network-enhanced large-scale PDDL modeling method as described in claim 2, characterized in that, In S1, the heterogeneous graph includes: a domain subgraph and a problem subgraph; The domain subgraph includes: predicate nodes Action nodes and type nodes These correspond to the predicate declarations, action templates, and type declarations in the PDDL file, respectively; and are connected based on the preconditions, parameter constraints, and action effects corresponding to each action, specifically including edges corresponding to the preconditions. , corresponding to the added effect edge Edges corresponding to deletion effects Edges with corresponding parameter type constraints ; The problem subgraph includes: specific object nodes. Initial state node and target state node Each node is connected according to the adjacency relationship defined in the problem.
4. The graph neural network-enhanced large-scale PDDL modeling method as described in claim 3, characterized in that, In the semantic embedding process, for a node, the initial features are formed by concatenating semantic vectors and structural vectors: in, These are node semantic vectors derived by a pre-trained BERT model based on action templates, predicate declarations, or node state-related descriptions in a PDDL file. A structure vector derived from a pre-trained BERT model, containing node type, number of parameters, and parameter type; This is a vector concatenation function; The node set is ; For edges, only the edge type needs to be labeled, assigning a number to each type of edge, and the set of edge types is: .
5. The graph neural network-enhanced large-scale PDDL modeling method as described in claim 1, characterized in that, S2 specifically includes: Based on the collected PDDL modeling examples, data cleaning and enhancement were performed to obtain high-quality PDDL positive examples without duplicates or errors. PDDL negative examples were constructed by artificially introducing action structure errors, parameter slot errors, predicate errors, and start / endpoint errors, and adding error type annotations. The defect set is as follows: ; Both positive samples labeled 1 and negative samples labeled 0 are transformed into graph representations using S1; Build In a multi-layered GNN network, each layer aggregates feature information between each node and its neighbors through message passing. in, This represents the current GNN network layer. This is a learnable self-looping weight matrix used to preserve the semantic information of the nodes themselves. For each edge type A learnable relation-specific weight matrix used to represent semantic transformation rules of different logical relations; For type Edges and nodes A set of connected nodes; This is a normalization constant; This is the activation function for this layer; After After layer iterations, each node in the graph incorporates a large amount of local information, resulting in the final embedding matrix of the entire graph. This contains the logical information of the entire graph, which can be used to verify whether the PDDL modeling is complete and logically correct; for each defect... Construct a two-layer MLP to achieve support for any node. Final verification: in, , These are the weights and biases of the first layer of the MLP, respectively. , These are the weights and biases of the second layer of the MLP, respectively. ReLU is the activation function corresponding to the first layer of the MLP; Sigmoid is the activation function corresponding to the second layer of the MLP; From the range of Sigmoid values, we can know ; In GNN networks and MLP, , , All parameters are learnable; a joint loss function is constructed using cross-entropy loss: in, For the first The weight hyperparameters for this type of defect; The set of nodes in the sample that are associated with this type of defect; This is the actual label of the node; Based on this, it can be minimized using the Adam optimizer. This yields the GNN network and MLP used to judge the effectiveness of PDDL modeling, and sets validation results for each defect. Is the threshold acceptable? .
6. The graph neural network-enhanced large-scale PDDL modeling method as described in claim 1, characterized in that, S3 specifically includes: Typical cases from the PDDL positive examples are injected into the LLM as prompt words. The LLM then generates the initial PDDL domain file based on the user's task description context. and problem files At this point, the prompt words used by LLM are represented as follows: ; The next Step, via S1 and Both are converted into graph representations. The GNN verifier obtained from S2 is used to verify whether the two PDDL files are complete and logically sound, and the verification results for each node and each possible error are obtained. ;like This indicates that the node has a modeling logic error, and the node is removed. With error Integrate into a natural language description using a certain template. All errors are grouped into the feedback prompt word "Error" for this round. If "Error" is not an empty set, the constructed PDDL file is considered to have a defect. Let: Using new prompts, the LLM algorithm corrects the generated PDDL file, resulting in new modeling outcomes. and And proceed to the next round of verification; otherwise, the current... and This is the final PDDL modeling result file for the natural language description task.
7. A large-scale PDDL modeling system enhanced with graph neural networks, characterized in that, The modeling system includes: a heterogeneous graph construction and feature embedding module, a GNN validator construction and parameter training module, and a large model modeling and correction module; The heterogeneous graph construction and feature embedding module is used to parse PDDL domain files and problem files in text format, construct heterogeneous graphs based on the parsing results of PDDL documents, and perform refined semantic embedding based on the heterogeneous graphs. The GNN validator construction and parameter training module constructs and trains a GNN validator based on the publicly available PDDL data. This validator is used to determine whether the logical relationships in the PDDL file are reasonable, and serves as a validator in the LLM PDDL generation process. The large model modeling and correction module is used to transform the numerical and localized defect predictions output by the GNN verifier into accurate and structured natural language feedback, guiding the LLM to perform automated and high-quality correction of the generated PDDL file.