Task planning method, control device and storage medium

Abstract

The invention provides a task planning method, a control device and a storage medium, and belongs to the technical field of long-range planning. The method comprises the following steps: constructing a corresponding rich semantic learning graph according to a target planning problem of a PDDL problem format; and inputting the rich semantic learning graph into the trained hybrid planning network, calculating a heuristic value corresponding to each state, guiding to search a transfer solution from an initial state to a target state, obtaining a planning action sequence of a target planning problem, and constructing the hybrid planning network based on node attention and a relational graph neural network. According to the method, semantic information is extracted by adopting a large language model, and a rich semantic learning graph and a hybrid planning network which are used for fully capturing PDDL problem information to assist task planning are constructed; the success rate of the planning problem can be improved by utilizing the rich semantic learning graph and the hybrid network architecture, and the reliability and the high efficiency of long-range planning are effectively guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of long-range planning, and particularly relates to a task planning method, a control device and a storage medium. BACKGROUND

[0002] Long-range planning, as the core capability of intelligent systems to execute tasks in complex environments, has decisive significance in the fields of robot operation, autonomous driving and intelligent logistics. However, with the increase of problem complexity, reliable and efficient planning methods face severe challenges. Although large language models have made significant progress in natural language understanding, directly applying large language models to planning tasks has the problem of insufficient reliability, and it is difficult to generate high-quality planning sequences. At the same time, symbolic planning methods rely on domain-specific heuristic functions and are difficult to generalize across domains. To solve the above problems, heuristic learning methods based on neural networks have emerged.

[0003] In the related art, heuristic learning methods based on neural networks still have significant limitations in expressiveness, generalization and computational efficiency. SUMMARY

[0004] The purpose of the present application is to provide a task planning method, a control device and a storage medium, aiming to solve the problem that related planning methods cannot fully utilize the prior information of the planning problem, resulting in a low planning success rate.

[0005] According to a first aspect of the present application, a task planning method is provided, which comprises: acquiring a target planning problem in PDDL (Planning Domain Definition Language) problem format; constructing a corresponding rich semantic learning graph according to the target planning problem in PDDL problem format; inputting the rich semantic learning graph into a trained hybrid planning network to calculate the heuristic value corresponding to each state, guide the search for a transition solution from the initial state to the target state, and obtain a planning action sequence of the target planning problem. Wherein, the hybrid planning network is constructed based on node attention and relation graph neural network.

[0006] In an optional implementation, the target planning problem in PDDL problem format comprises a domain file and a problem file. Wherein, the domain file comprises object class, predicate mode and action mode definition, and the problem file comprises instance object, instance state and target state.

[0007] In an optional implementation, the constructing a rich semantic learning graph corresponding to the goal planning problem in the PDDL problem format comprises: constructing a multi-relation and multi-type node directed learning graph according to the goal planning problem in the PDDL problem format; generating a feature vector including semantic prior information corresponding to each node of the directed learning graph; and obtaining the rich semantic learning graph based on the feature vector including semantic prior information corresponding to each node.

[0008] In an optional implementation, the directed learning graph comprises a schema subgraph and an instance subgraph. The schema subgraph describes domain general knowledge, and comprises action nodes, predicate nodes and formal parameter nodes. The semantic relations of the nodes are connected according to the preconditions, add effects and delete effects of actions. The instance subgraph describes states and goals of the goal planning problem, and comprises instance object nodes, state nodes and goal nodes. The connections of the nodes are defined according to the ownership relations of states and objects. The schema subgraph and the instance subgraph are connected through predicate nodes.

[0009] In an optional implementation, the directed learning graph further comprises: a semantic shortcut constructed by adding a connection edge between the object node and the corresponding state node; and a configurable expert node connected to the state node and the goal node.

[0010] In an optional implementation, the feature vector including semantic prior information comprises a category feature vector, a parameter index encoding feature vector, a schema action and predicate feature vector, a large model encoding semantic feature vector and an expert prior heuristic feature vector.

[0011] In an optional implementation, after the rich semantic learning graph is input into the trained hybrid planning network, the task planning method further comprises: performing an attention mechanism to enable the trained hybrid planning network to generate a feature fusion representation before performing neighborhood aggregation. For each feature vector including semantic prior information corresponding to a node, each channel feature is padded to a uniform dimension and split to form a corresponding feature sequence, and the feature fusion representation is generated based on the feature sequence corresponding to each node.

[0012] In an optional implementation, the constructing a hybrid planning network based on node attention and a relational graph neural network comprises: using node attention to fuse heterogeneous features within a node and generate a relation-conditioned node embedding; and using a relational graph neural network to perform information propagation on a graph structure to capture relation dependencies and structure dependencies for effective heuristic estimation.

[0013] According to a second aspect of the present application, a control device is provided, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the task planning method described above.

[0014] According to a third aspect of the present application, a machine-readable storage medium is provided, which stores instructions for causing a machine to execute the task planning method described above.

[0015] The technical solution described above is to solve the problem that related planning methods cannot fully utilize prior information of a planning problem, resulting in a low planning success rate. The embodiments of the present application provide a task planning method, which comprises: obtaining a target planning problem in a PDDL problem format; constructing a corresponding rich semantic learning graph according to the target planning problem in the PDDL problem format; inputting the rich semantic learning graph into a trained hybrid planning network, calculating a heuristic value corresponding to each state, guiding the search of a transition solution from an initial state to a target state, and obtaining a planning action sequence of the target planning problem, and constructing a hybrid planning network based on node attention and a relational graph neural network. The embodiments of the present application use a large language model to extract semantic information and construct a rich semantic learning graph and a hybrid planning network for fully capturing PDDL problem information to assist task planning. The rich semantic learning graph can provide structured and semantically rich input representations for neural networks by constructing structured data and multi-channel features for neural network learning; the hybrid planning network realizes decoupling of node internal semantic fusion and cross-node relationship propagation through node attention and a relational graph convolutional neural network, the former focuses on multi-source feature integration, and the latter focuses on structure information transmission, so that the hybrid planning network can efficiently learn heuristic functions. The embodiments of the present application use the rich semantic learning graph and the hybrid network architecture to improve the success rate of planning problems and effectively ensure the reliability and efficiency of long-range planning.

[0016] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and achieved by the structures and processes as described in the specification and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0018] Figure 1is a flowchart of a task planning method provided by an example embodiment of the present application.

[0019] Figure 2 is a schematic diagram of a PDDL planning problem provided by an example embodiment of the present application.

[0020] Figure 3 is a structural schematic diagram of an undirected learning graph provided by an example embodiment of the present application.

[0021] Figure 4 is a schematic diagram of an output large model encoding semantic feature vector provided by an example embodiment of the present application.

[0022] Figure 5 is a schematic diagram of a feature vector provided by an example embodiment of the present application.

[0023] Figure 6 is a schematic diagram of a hybrid planning network provided by an example embodiment of the present application.

[0024] Figure 7 is a comparison diagram of success rates of task planning based on the present application and other methods provided by an example embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0026] In related technologies, for a planning method based on a large language model, the large language model shows great potential in natural language understanding, but since the large language model is not designed for planning tasks, there are limitations in directly applying the large language model to planning tasks. The large language model (for example, ChatGPT) is essentially a sequence generation model trained on a large amount of text data, which is trained to predict the next most likely word, rather than to understand the structure and logical relationship of the problem. For example, when asking the large language model to solve a simple Tower of Hanoi planning problem, it may generate a seemingly reasonable but actually unfeasible step sequence. This ability is different from the ability to generate a correct sequence based on an understanding of the nature of the problem.

[0027] The classical symbolic planning method models the planning problem using an explicit symbolic model, for example, the planning domain definition language (PDDL) defines related objects , state representation and actions For the current state To the target state The transition solution can be obtained by searching the action sequence. We obtained, among which, Refers to the first There are several actions involved. Therefore, classical symbolic programming provides a theoretically guaranteed solution that ensures correctness and interpretability. Despite these advantages, the performance of classical planners largely depends on heuristic search, and designing efficient heuristic functions to estimate state quality typically requires extensive domain knowledge and experimental verification. For example, for the PDDL programming problem, It is the mainstream heuristic function for solving this type of planning problem. State heuristics are estimated by calculating the minimum action costs to reach the target state. While these methods perform well in specific domains, they have limitations. These heuristic functions have poor generalization ability across different planning problems, and the search space grows exponentially with the size of the planning problem, resulting in low computational efficiency.

[0028] For heuristic learning methods based on neural networks, related techniques enhance classical planning through learning-based heuristics. Neural network-based methods attempt to improve planning efficiency by learning structured representations of planning problems and then computing approximate heuristic functions. For example, hypergraph network-based frameworks are used to learn domain-specific heuristic functions; or graph convolutional networks are used to process structured representations of PDDL problems. However, these methods often struggle to generalize effectively across different domains and fail to effectively utilize the structural information in symbolic representations and the semantic knowledge related to the planning problem.

[0029] To address the issue that related task planning methods cannot fully utilize prior information about the planning problem, resulting in a low success rate, this application provides a task planning method based on a large language model and neural networks.

[0030] Please refer to Figure 1 The task planning method provided in this application embodiment may include the following steps: Step S110: Obtain the goal planning problem in PDDL problem format.

[0031] The preferred PDDL problem format in this application embodiment may include a domain file and a problem file. The domain file may include object categories, predicate patterns, and action pattern definitions, while the problem file may include instance objects, instance states, and target states.

[0032] In the embodiments of the present application, the goal planning problem in PDDL problem format (hereinafter can be referred to as PDDL problem) can include a domain file (domain.pddl) and a problem file (problem.pddl). In the embodiments of the present application, the domain file can include object class, predicate mode (i.e., state representation, hereinafter referred to as predicate) and action mode (hereinafter referred to as action) definition; the problem file can include instance object, instance state and goal state. Please refer to Figure 2 For example, the object class in the domain file of the goal planning problem fragment is types, including block class. The predicate is predicates, including on, on-table and clear, etc., wherein on includes two parameters x and y, indicating that the object corresponding to the parameter x is on the object corresponding to y, and the other predicate parameters and meanings are similar. The action is action, including stack action, indicating the stacking block action. The action can include the following key elements: parameters, precondition and effect; and can be divided into addition effect and deletion effect according to whether the effect is negative. The objects in the problem file are instance objects, init and goal can represent the initial instance state and the target instance state respectively, and the instance state can be composed of instantiated predicates, for example, (on a b).

[0033] Step S120: constructing a corresponding rich semantic learning graph according to the goal planning problem in PDDL problem format.

[0034] In the preferred embodiments of the present application, step S120 can include: step S121, constructing a multi-relation and multi-type node undirected learning graph according to the goal planning problem in PDDL problem format; step S122, generating a corresponding feature vector including semantic prior information for each node of the undirected learning graph; and step S123, obtaining a rich semantic learning graph based on the feature vector including semantic prior information corresponding to each node.

[0035] The preferred undirected learning graph of the embodiments of the present application can include a mode subgraph and an instance subgraph. The preferred mode subgraph of the embodiments of the present application describes general knowledge in the field, and the mode subgraph can include action nodes, predicate nodes and formal parameter nodes, and the semantic relationship of each node is connected according to the premise condition, addition and deletion effect of the action. The preferred instance subgraph of the embodiments of the present application describes the state and goal of the goal planning problem, and the instance subgraph can include instance object nodes, state nodes and goal nodes, and the connection of each node is defined according to the ownership relationship of the state and the object. The mode subgraph and the instance subgraph are connected through the predicate nodes.

[0036] In the embodiments of the present application, in order to solve the problem that it is difficult for the neural network to directly extract structured information from the PDDL planning problem, resulting in difficulty in learning the heuristic function, a multi-relation and multi-type node undirected learning graph can be constructed according to the goal planning problem in the PDDL problem format as shown in Figure 2 Please refer to Figure 3 For example, the schema subgraph describes the domain general knowledge of the goal planning problem, which can include action nodes (stack), predicate nodes (on, hand-empty, clear and holding), and formal parameter nodes (which can also be referred to as formal schema parameter nodes, such as the three middle nodes in Figure 3 The semantic relationship of each node is connected according to the precondition, addition and deletion effect of the action, for example, corresponding to the red, black and blue edges in Figure 3 The example subgraph can describe the state and goal of the goal planning problem, which can include example object nodes (a, b, c), current state nodes (on (a, b)), goal nodes (on (b, c)), formal example parameter nodes and predicate nodes (on). In the embodiments of the present application, the schema subgraph and the example subgraph can be connected through the predicate nodes for information propagation. The connection relationship of the example subgraph can be defined by the relationship between the state and the corresponding object. For example, the state node (on (a, b)) is connected to the object nodes a and b through the formal example parameter nodes 1 and 2.

[0037] The preferred undirected learning graph in the embodiments of the present application can also include: a semantic shortcut constructed by adding a connection edge between the object node and the corresponding state node; and a configurable expert node connected to the state node and the goal node.

[0038] In the embodiments of the present application, in order to further improve the propagation efficiency of information in the undirected learning graph and enhance the utilization ability of the domain knowledge of the hybrid planning network below, the embodiments of the present application are extended on the basis of the above undirected learning graph structure, which can include: 1) object state direct connection relationship. A direct connection edge is added between the object node and the corresponding state node to construct a semantic shortcut, so as to significantly reduce the number of message passing required for the object information to affect the state representation and accelerate the information flow. For example, as shown in the boxed part of Figure 3 The state node on (a, b) and the object nodes a and b are directly connected. 2) Configurable expert node for encoding domain-specific heuristic prior, corresponding to the expert feature below, each expert node can be connected to the state node and the goal node, so that the heuristic information can be injected into the undirected learning graph in a structured manner and participate in subsequent reasoning. For example, as shown in the lower boxed part of Figure 3 The expert node hconnected with the state node on (a, b) and the goal node on (b, c).

[0039] The feature vector of the preferred embodiment of the present application including semantic prior information can include a category feature vector, a parameter index encoding feature vector, a mode action and predicate feature vector, a large model encoding semantic feature vector, and an expert prior heuristic feature vector.

[0040] In the embodiment of the present application, to solve the problem that the related method cannot provide the related prior information of the PDDL planning problem, on the basis of the above-mentioned undirected learning graph, the embodiment of the present application further provides multi-class learning graph features. Each feature has a feature vector with related and heterogeneous information, so for each node, a corresponding rich information representation (i.e., a feature vector including semantic prior information) can be constructed. In the embodiment of the present application, the feature vector of each node can include: 1) a category feature vector (also called a one-hot vector), used to identify the semantic category of the node, which can include predicates, actions, instance objects, instance states, instance positive goals, and instance negative goals; for other nodes, such as parameter nodes, a zero vector is assigned, for example, for the action stack, the category feature can be [0, 1, 0, 0, 0, 0], indicating that the node is a semantic category, i.e., an action.

[0041] 2) a parameter index encoding feature vector, used to represent the index embedding assigned to the formal parameter node in the action or predicate, the parameter node can be assigned to a fixed embedding vector according to its order index in the mode. For example, the action stack has two parameters, then the first parameter node can be assigned to the feature embedding corresponding to index 1, for example, [-0.0061, 0.4370, -0.6705, -0.5995].

[0042] 3) a mode action and predicate feature vector, used to distinguish different predicates and actions in the same domain, which can be assigned another set of fixed embedding vectors according to its symbolic identification. For example, the action stack can be assigned an embedding vector, for example, [0.6080, -0.5208, -0.2288, 0.5538].

[0043] The parameter index encoding feature vector and the mode action and predicate feature vector are respectively for parameter nodes and mode nodes, and the feature part of other nodes can also be a zero vector, as follows.

[0044] 4) Large model encoding semantic feature vector, using large language model to introduce high-level semantic priori for PDDL problem. In the embodiment of the application, the semantic embedding of action and predicate can be obtained by inputting the goal planning problem (problem description) in PDDL problem format and related prompt words into the large language model, and the feature vector including semantic priori information is generated for each mode node. For example, GPT-4 (thousand questions and other large language models can also be used) can be used to output the formatted semantic feature vector by inputting the prompt word example and input format as shown in the following table. Figure 4 For example, the semantic feature vector of the action stack is [0.5, 0.3, -0.2, 0.4], and the predicate clear is [0.6, 0.4, -0.1, 0.2].

[0045] 5) Expert prior heuristic feature vector is a vector including domain-specific heuristic priori combined with related expert information. For example, for the n-puzzle embodiment problem, the Manhattan distance between the current position of each block and the target position can be considered as the prior heuristic value. For example, the final feature of the action stack is [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0.6080, -0.5208, -0.2288, 0.5538, 0.5, 0.3, -0.2, 0.4, 0], which has 19 dimensions. The meaning and dimension of each type of feature correspond to Figure 5 .

[0046] In the embodiment of the application, the multi-channel representation constructed by the category feature vector of each node, the parameter index encoding feature vector, the mode action and predicate feature vector, the large model encoding semantic feature vector, and the expert prior heuristic feature vector can form a structured graph representation that can be learned by a neural network, that is, a rich semantic learning graph.

[0047] Step S130: inputting the rich semantic learning graph into the trained hybrid planning network, calculating the heuristic value corresponding to each state, guiding the transition solution from the initial state to the target state, and obtaining the planning action sequence of the target planning problem.

[0048] Among them, the hybrid planning network is constructed based on node attention and relational graph neural network.

[0049] In the preferred embodiment of the application, after inputting the rich semantic learning graph into the trained hybrid planning network, the task planning method can further include: performing an attention mechanism to make the trained hybrid planning network generate a feature fusion representation before performing neighborhood aggregation. Among them, for the feature vector including semantic priori information corresponding to each node, the channel features are filled to a unified dimension and split to form a corresponding feature sequence, and the feature fusion representation is generated based on the feature sequence corresponding to each node.

[0050] In the embodiments of the present application, each node in the rich semantic learning graph can be associated with a set of heterogeneous feature channels, which have significant differences in scale and semantics, as shown in Figure 5 The relevant graph neural network usually processes them through a single linear transformation. However, this linear mixing method is difficult to effectively capture the meaningful interaction between different feature types. To overcome this limitation, the embodiments of the present application can perform attention mechanisms within the node in the mixed planning network, so that the mixed planning network can learn the structured fusion method between the feature channels before performing neighborhood aggregation, to generate a feature fusion representation. For example, for each node feature , the channel features can be padded to a uniform dimension and split to form a feature sequence , so as to regard the node as a small sequence composed of multiple feature tokens. For example, the action stack feature in the above embodiments can be split into , wherein the zero-padded category feature vector , the zero-padded parameter index encoding feature vector , and so on. For the node sequence , the following scaled dot-product attention can be used:

[0051] , wherein is obtained by projection transformation from the sequence , respectively denoting the query, key and value sequences, is the vector dimension for scaling; the attention output can be restored to a fixed-dimensional relation-conditioned node embedding by flattening or projection.

[0052] The preferred mixed planning network constructed based on the node attention and the relation graph neural network in the embodiments of the present application can include: using the node attention to fuse the heterogeneous features within the node and generate a relation-conditioned node embedding; and using the relation graph neural network to perform information propagation on the graph structure to capture the relation dependence and structural dependence for effective heuristic estimation.

[0053] In alternative embodiments of the present application, the node attention can be replaced by a lightweight linear attention or a gating mechanism to improve the inference speed.

[0054] In the embodiments of the present application, based on the above node attention, to efficiently learn the heuristic function from the rich semantic learning graph and further calculate the accurate heuristic value, the embodiments of the present application provide a mixed planning network, which combines the node attention within the node with the relation graph neural network, as shown in Figure 6In the embodiments of the present application, node attention is utilized to fuse heterogeneous features within nodes and generate relation-conditioned node embeddings. Relation graph neural networks are utilized to propagate information on graph structure to capture relation dependencies and structure dependencies that are crucial for effective heuristic estimation. In the embodiments of the present application, the relation graph neural networks can encode relation inductive bias by assigning independent transformation matrices for each relation type, thereby modeling how different semantic relations affect the interaction between nodes; and on this basis, replace the standard linear transformation in each relation with a richer, relation-specific node representation generated by the node attention module. The embodiments of the present application can decouple the semantic fusion within nodes and the relation propagation between nodes, with the former focusing on integrating multi-source heterogeneous features and the latter focusing on propagating these fused information on graph structure. Based on this, the hybrid planning network can derive heuristic values of PDDL problems from node features through global pooling and a simple feedforward network, for calculating the pros and cons of the current state and guiding efficient search for planning. Further, the label heuristic values can be calculated according to existing planning problem datasets and using other classical planning methods to build training and test sets, train the hybrid planning network; and input the above rich semantic learning graph into the trained hybrid planning network to calculate the heuristic values corresponding to each state, guide the transition solution from the initial state to the target state, and obtain the planning action sequence of the target planning problem.

[0055] The embodiments of the present application are explained by taking the shopping guide robot scenario as an example. The domain file and the problem file of the shopping guide scenario can be defined according to the shopping guide scenario. The domain file can include robot executable actions (such as moving, grabbing, etc.), shopping guide scenario item categories, and shopping guide scenario state predicates; the problem file can include instance objects (such as a water bottle, a refrigerator, a starting area, a refrigerator area, etc.), an initial state (such as the robot being in the starting area and the water bottle being in the refrigerator), and a target state to be set. If the robot needs to complete the task planning of taking out the water bottle, the target state can be set as the water bottle being in the robot's hand. For example, the domain file and the problem file can be converted into a rich semantic learning graph; the heuristic value of the current state is calculated by using the trained hybrid planning network, the transition solution from the initial state to the target state is guided, and a reliable planning action sequence is obtained. For example, moving to the refrigerator area, opening the refrigerator, taking out the water bottle, and closing the refrigerator. If the actions defined in the domain file completely correspond to the robot executable actions, the planning action sequence can be directly mapped to the robot action sequence, thereby ensuring the reliability of the robot actions.

[0056] According to this, in order to solve the problem that the related planning method cannot fully utilize the prior information of the planning problem, resulting in a low success rate of planning, the embodiments of the present application provide a task planning method, which comprises: obtaining a target planning problem in PDDL problem format; constructing a corresponding rich semantic learning graph according to the target planning problem in PDDL problem format; inputting the rich semantic learning graph into the trained hybrid planning network, calculating the heuristic value corresponding to each state, guiding the transition solution from the initial state to the target state, obtaining the planning action sequence of the target planning problem, and constructing a hybrid planning network based on node attention and relational graph neural network. The embodiments of the present application use a large language model to extract semantic information, and construct a rich semantic learning graph and a hybrid planning network for fully capturing PDDL problem information to assist task planning. The rich semantic learning graph can provide structured and semantically rich input representations for neural networks by constructing structured data and multi-channel features for neural network learning; the hybrid planning network realizes the decoupling of node internal semantic fusion and cross-node relationship propagation through node attention and relational graph convolutional neural network, the former focuses on multi-source feature integration, and the latter focuses on structure information transmission, so that the hybrid planning network can efficiently learn heuristic functions. The embodiments of the present application can improve the success rate of planning problems by using rich semantic learning graphs and hybrid network architectures, effectively ensuring the reliability and efficiency of long-range planning.

[0057] Further, in order to overcome the problem that neural networks are difficult to directly extract structured information from PDDL planning problems to learn heuristic functions, the embodiments of the present application construct an undirected learning graph composed of multiple relationships and multiple types of nodes based on PDDL problems. In order to make up for the lack of PDDL prior information in related methods, the embodiments of the present application construct node representations of rich semantic learning graphs based on five types of heterogeneous node features. Since each feature channel has significant differences in scale and semantics, a single transformation used by related graph neural networks cannot effectively fuse them. In this regard, the embodiments of the present application use node attention to dynamically learn the interaction weights between each feature before neighborhood aggregation through a self-attention mechanism, and generate the fused node representation. In order to efficiently learn heuristic functions, the embodiments of the present application construct a hybrid planning network, which is different from the related relational graph convolutional network using a fixed linear transformation for each type of relationship. The embodiments of the present application use the rich node representation output by the node attention to replace the original input, fuse heterogeneous features through node attention, and generate relationship-conditioned node embeddings; input them into the relational graph convolutional network for graph propagation to capture the relationship and structural dependencies that are critical to heuristic estimation.

[0058] Further, compared with the method of directly using a large language model for planning, the embodiments of the present application avoid the unreliability problem caused by the lack of explicit modeling of action logic and state transition constraints in the large language model; by using the large language model only for extracting high-level semantic priors and integrating them into a structured graph representation, the semantic understanding ability of the large language model is retained, and the logical rigor of the planning process is ensured. Compared with the symbolic planning method, the embodiments of the present application do not need to rely on artificially designed domain-specific heuristic functions, but automatically learn more general and more accurate state evaluation functions in a data-driven manner, thereby achieving stronger generalization ability and higher solution success rate in multiple different PDDL domains. Compared with related planning methods based on graph neural networks, the embodiments of the present application not only construct a more rich graph structure to accommodate related prior knowledge, but also fuse semantic embeddings and expert heuristic information from a large language model; at the same time, by effectively integrating multiple source heterogeneous features through intra-node attention mechanism and relational graph neural network, the learned heuristic function has obvious improvement in expression ability, generalization performance and computational efficiency. Figure 7 The task planning method provided by the embodiments of the present application is shown, and the planning problem success rate is compared with other methods. The values in the table indicate the planning success rate, the horizontal axis is eight classical planning problems, and the vertical axis is each planning method. Blind is a blind search method, FF is a classical heuristic algorithm, HGN and GOOSE are heuristic planning algorithms based on graph neural networks. It can be obtained that the task planning method achieves a very high success rate on four planning tasks, and the overall planning success rate reaches 78.0%, which is significantly better than other methods, and at the same time, the efficiency of the embodiments of the present application is verified.

[0059] The embodiments of the present application also provide a control device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to realize the task planning method described above.

[0060] The embodiments of the present application also provide a machine-readable storage medium, which stores instructions for causing a machine to execute the task planning method described above.

[0061] It should be noted that the control device and the machine-readable storage medium described above can realize the task planning method provided by the embodiments described above, and the specific implementation manner can be referred to the description of the task planning method in the above embodiments, which will not be described here.

[0062] It can be understood that the circuit structure, name and parameter described in the above embodiments are only examples. Those skilled in the art can also easily combine and adjust the structural characteristics of the above multiple embodiments according to the use needs, and the concept of the present application should not be limited to the specific details of the above examples.

[0063] Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood that modifications can be made to the foregoing embodiments, or additional implementations can be implemented, without departing from the spirit and scope of the embodiments.

Claims

1. A method of task planning, characterized by, The task planning method comprises: acquiring a target planning problem in a PDDL problem format; constructing a corresponding rich semantic learning graph according to the target planning problem in the PDDL problem format; inputting the rich semantic learning graph into a trained hybrid planning network, calculating a heuristic value corresponding to each state, guiding a transition solution from an initial state to a target state, and obtaining a planning action sequence of the target planning problem, wherein the hybrid planning network is constructed based on node attention and a relational graph neural network.

2. The task planning method of claim 1, wherein, The target planning problem in the PDDL problem format comprises a domain file and a problem file, wherein the domain file comprises object categories, predicate patterns and action pattern definitions, and the problem file comprises instance objects, instance states and a target state.

3. The task planning method of claim 1, wherein, The construction of the corresponding rich semantic learning graph according to the target planning problem in the PDDL problem format comprises: constructing a multi-relation and multi-type node undirected learning graph according to the target planning problem in the PDDL problem format; generating a corresponding feature vector comprising semantic prior information for each node of the undirected learning graph; and obtaining the rich semantic learning graph based on the feature vector comprising semantic prior information corresponding to each node.

4. The task planning method of claim 3, wherein, The undirected learning graph comprises a pattern subgraph and an instance subgraph, wherein the pattern subgraph describes domain general knowledge, the pattern subgraph comprises action nodes, predicate nodes and formal parameter nodes, and the semantic relationship of each node is connected according to the premise conditions, addition and deletion effects of actions, the instance subgraph describes the states and target of the target planning problem, the instance subgraph comprises instance object nodes, state nodes and target nodes, and the connection of each node is defined according to the ownership relationship between states and objects, the pattern subgraph and the instance subgraph are connected through predicate nodes.

5. The task planning method of claim 4, wherein, The undirected learning graph further comprises: a semantic shortcut constructed by adding a connection edge between the object node and the corresponding state node; and a configurable expert node connected to the state node and the target node.

6. The task planning method of claim 4, wherein, The feature vector comprising semantic prior information comprises a category feature vector, a parameter index encoding feature vector, a pattern action and predicate feature vector, a large model encoding semantic feature vector and an expert prior heuristic feature vector.

7. The task planning method of claim 3, wherein, After the rich semantic learning graph is input into the trained hybrid planning network, the task planning method further comprises: performing an attention mechanism to enable the trained hybrid planning network to generate a feature fusion representation before performing neighborhood aggregation, wherein for the feature vector comprising semantic prior information corresponding to each node, each channel feature is filled to a uniform dimension and split to form a corresponding feature sequence, and the feature fusion representation is generated based on the feature sequence corresponding to each node.

8. The task planning method of claim 1, wherein, The construction of the hybrid planning network based on node attention and a relational graph neural network comprises: utilizing node attention to fuse heterogeneous features within a node and generate a relation-conditioned node embedding; and utilizing a relational graph neural network to perform information propagation on a graph structure to capture relation dependence and structure dependence for effective heuristic estimation.

9. A control device characterized by comprising: The control device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the task planning method according to any one of claims 1-8.

10. A machine-readable storage medium, characterized in that, The machine readable storage medium stores instructions, which cause the machine to execute the task planning method according to any one of claims 1-8.

Images