Robot task planning method and system based on graph representation learning and electronic equipment

Abstract

The invention relates to the technical field of robot control, and provides a robot task planning method and system based on graph representation learning and electronic equipment, and the method comprises the following steps: constructing a priori knowledge reasoning model; according to the obtained natural language information of the robot execution service, based on the constructed priori knowledge reasoning model, robot planning elements to be processed for planning are extracted to serve as graph nodes, the relation between the robot planning elements serves as the edges of a graph, and a directed graph is constructed; vectorizing information of the constructed directed graph and natural language information of robot execution service to obtain vectorized representation of task planning; and based on the deep learning model, a progressive method of template generation and template filling is adopted to generate staged task planning of robot actions. Effective graphic representation is carried out on robot task planning by considering service execution requirements, and a service-related natural language is autonomously converted into selective task planning under the guidance of reinforcement learning.

Description

Technical Field

[0001] The present disclosure relates to the technical field of robot control, and more particularly, to a robot task planning method, system, and electronic device based on graph representation learning. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the development of artificial intelligence technology, natural language-based knowledge learning has become an important way for humans to improve their own skill levels. In the field of home service, there is a large amount of natural language information related to home service execution on the network. Enabling home service robots to have natural language learning and understanding capabilities and be able to autonomously generate task plans based on natural language will greatly improve the service skills of robots and enhance the intelligence level of their task execution. Robots are widely used in home services. Among them, robot actions can be divided into low-level actions and high-level actions. Low-level actions mainly target the basic operations of the robot and involve specific motion parameters, such as motion distance, joint angle, grasping direction, etc. High-level actions, on the other hand, are more semantic and abstract action descriptions, such as "grasp", "put down", "deliver", etc., which can more intuitively express the intentions and goals of robot tasks. In home service robot task planning, it mainly involves the recognition and execution of high-level actions to improve the autonomous decision-making ability of the robot and make it better adapt to service requirements in different home environments.

[0004] Currently, most home service robot task planning methods adopt rule-based design methods. These methods mainly focus on the structural representation of natural language, design corresponding planning templates, and achieve task planning through template matching. For example, Misra et al. used natural language processing (NLP) technology to construct information extraction templates through semantic tags (POS, Part of Speech), extract information elements from natural language sentences, and generate corresponding task plans based on the templates. However, this method has great limitations on the structure of natural language, is difficult to handle variable and complex language expression forms, and lacks flexibility.

[0005] With the development of deep learning technology, researchers have gradually attempted to endow robots with learning abilities, enabling them to autonomously convert natural language into task planning, improving the robot's ability to understand natural language and the intelligence level of task planning. However, the inventors found in their research that there are still many problems with existing task planning methods based on deep learning. First of all, these methods usually do not consider the actual execution environment of the robot in home services, and the generated task planning often disconnects from the actual environment, lacking environmental adaptability, resulting in greater uncertainty that the robot may face when performing specific tasks. Secondly, the task planning representation of existing methods is relatively single, usually only generating a unique execution plan and unable to provide multiple coping strategies, making it difficult for the robot to adjust the task planning in abnormal situations and affecting the successful execution of the task. In addition, most existing task planning methods based on deep learning limit the scale and vocabulary of natural language in order to reduce computational complexity and training difficulty, restricting the robot's ability to understand complex and diverse natural language. The existence of these problems makes the current task planning methods difficult to meet the execution needs of actual home service robots, restricting the robot's adaptability and autonomous decision-making ability in complex environments. Summary of the Invention

[0006] To solve the above problems, the present disclosure proposes a robot task planning method, system and electronic device based on graph representation learning, effectively graphically represents the robot task planning considering service execution requirements, and autonomously converts service-related natural language into a selective task planning under the guidance of reinforcement learning, guiding the robot to have the ability to handle abnormal situations and complete unfamiliar service tasks, ultimately improving the robot's intelligence level and service skills.

[0007] To achieve the above object, the present disclosure adopts the following technical solutions:

[0008] One or more embodiments provide a robot task planning method based on graph representation learning, including the following steps:

[0009] Based on item characteristic reasoning and scene co-occurrence relationship modeling, construct a prior knowledge reasoning model;

[0010] According to the obtained natural language information of the robot's execution service, based on the constructed prior knowledge reasoning model, extract the robot planning elements to be processed in the planning as graph nodes, and use the relationships between the robot planning elements as the edges of the graph to construct a directed graph;

[0011] Extract the adjacency matrix of the directed graph, and perform vectorization representation on the vocabulary and vocabulary positions in the natural language information related to the robot service, and obtain the vectorization representation of the task planning after fusion;

[0012] Taking natural language information related to robot services as input and the vectorized representation of task planning as output, train the constructed task planning model;

[0013] Obtain the natural language information instruction to be executed by the service robot. Based on the trained task planning model, use the progressive method of template generation and template filling to generate the task planning result of the robot's actions.

[0014] One or more embodiments provide a robot task planning system based on graph representation learning, including:

[0015] A prior construction module, configured to construct a prior knowledge reasoning model based on item characteristic reasoning and scene co-occurrence relationship modeling;

[0016] A task planning graphical representation module, configured to, according to the obtained natural language information of the robot performing services, based on the constructed prior knowledge reasoning model, extract the robot planning elements to be processed in the planning as graphical nodes, and use the relationship between the robot planning elements as the edges of the graph to construct a directed graph;

[0017] A vectorized representation module, configured to extract the adjacency matrix of the directed graph, and perform vectorized representation on the vocabulary and vocabulary positions in the natural language information related to robot services, and fuse them to obtain the vectorized representation of task planning;

[0018] A training module, configured to take natural language information related to robot services as input and the vectorized representation of task planning as output, and train the constructed task planning model;

[0019] A task planning generation module, configured to obtain the natural language information instruction to be executed by the service robot. Based on the trained task planning model, use the progressive method of template generation and template filling to generate the task planning result of the robot's actions.

[0020] An electronic device, including a memory and a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps in the above-mentioned robot task planning method based on graph representation learning are completed.

[0021] Compared with the prior art, the beneficial effects of the present disclosure are:

[0022] (1) This disclosure is based on graph learning. It constructs a directed graph using items, actions, similar items, and existing scenarios, and uses a deep model for task planning generation. The task planning is represented using a graph structure (Graph Representation), making the task planning not just a linear sequence but a graph model that includes items, actions, similar items, and scenarios. The task planning is optimized based on the relational edges of the graph, having stronger task expression capabilities and adaptability.

[0023] (2) This disclosure proposes a progressive task planning generation strategy of template generation and template filling. That is, first generate a task planning template composed of part-of-speech tags (POS), and then fill the template to form a task planning for the robot. This reduces the complexity of autonomous task planning generation, can meet the system requirements of task planning without restricting the scale of natural language, and obtains an accurate task planning based on reinforcement learning to ensure the successful execution of the robot service.

[0024] The advantages of this disclosure and the advantages of additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of this disclosure. The schematic embodiments and descriptions thereof are used to explain this disclosure and do not constitute a limitation to this disclosure.

[0026] Figure 1 is the overall flowchart of the robot task planning method based on graph representation learning in Embodiment 1 of this disclosure;

[0027] Figure 2 is the prior knowledge structure block diagram of Embodiment 1 of this disclosure;

[0028] Figure 3 is the schematic diagram of the graphical representation of the task planning of the robot in Embodiment 1 of this disclosure;

[0029] Figure 4 is the flowchart of the vectorization representation method in the robot task planning process of Embodiment 1 of this disclosure;

[0030] Figure 5 is the task planning template generation flowchart of Embodiment 1 of this disclosure;

[0031] Figure 6 is the task planning template filling flowchart of Embodiment 1 of this disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following further describes this disclosure in conjunction with the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features in the present disclosure can be combined with each other. The embodiments will be described in detail below with reference to the drawings.

[0035] Embodiment 1

[0036] In the technical solutions disclosed in one or more embodiments, as Figures 1 to 6 shown, a robot task planning method based on graph representation learning includes the following steps:

[0037] Step 1: Based on item characteristic reasoning and scene co-occurrence relationship modeling, construct a prior knowledge reasoning model;

[0038] Step 2: According to the obtained natural language information of the robot performing services, based on the constructed prior knowledge reasoning model, extract the robot planning elements to be processed in the plan as graph nodes, and use the relationships between the robot planning elements as the edges of the graph to construct a directed graph;

[0039] Step 3: Extract the adjacency matrix of the directed graph, and perform vectorized representation on the vocabulary and vocabulary positions in the natural language information related to the robot service. After fusion, obtain the vectorized representation of the task planning;

[0040] Step 4: Use the natural language information related to the robot service as the input and the vectorized representation of the task planning as the output to train the constructed task planning model;

[0041] Step 5: Obtain the natural language information instruction to be executed by the service robot, and based on the trained task planning model, use a progressive method of template generation and template filling to generate the task planning result of the robot action.

[0042] This embodiment is based on Graph Learning. A directed graph is constructed with items, actions, similar items, and existence scenarios, and a deep model is used to generate task planning. The task planning is represented by a graph structure (Graph Representation), making the task planning not just a linear sequence but a graph model that includes items, actions, similar items, and scenarios. The task planning is optimized based on the relational edges of the graph (sequence, function, similarity, location), having stronger task expression ability and adaptability. A progressive task planning generation strategy of template generation and template filling is proposed, that is, first generating a task planning template composed of part-of-speech tags (POS), and then filling the template to form a task planning for the robot, reducing the complexity of autonomous generation of task planning, meeting the system requirements of task planning without restricting the scale of natural language, and obtaining an accurate task planning based on reinforcement learning to ensure the successful execution of the robot service.

[0043] In this embodiment, the task planning model autonomously converts service-related natural language into a task planning for robot service execution. The overall block diagram is as Figure 1 shown. The service-related natural language information for the robot is the service instruction sent by the user to the robot; the service-related natural language is the text information used to describe the execution of home services, including service execution actions, household items, and other elements required for performing this service.

[0044] As Figure 2 shown, in step 1, constructing the prior knowledge reasoning model includes item prior and scenario prior. The item prior is the similar item reasoning model, and the scenario prior is the scenario reasoning model. The construction method is as follows:

[0045] Step 11: For the item prior, structurally represent the characteristics of the item, and establish an inference mechanism based on the structured features to infer substitutable items;

[0046] Step 11.1: Structural representation of item characteristics: Use ontology-based representation technology to represent the prior knowledge of the item; select the shape, size, and function of the item as reasoning characteristics, and organize the reasoning characteristics into a knowledge base with a tree structure, that is, obtain the item prior information base:

[0047] An item can have multiple reasoning characteristics, such as it can include reasoning characteristic 1, reasoning characteristic 2, and reasoning characteristic 1, etc.

[0048] A specific example:

[0049] Parent node: Item name, such as "cup";

[0050] Subnodes: inference characteristics of items, such as "cylindrical", "capacity", "300ml", "drinking function", etc.

[0051] Step 11.2: Establish a reasoning mechanism based on the reasoning characteristics of the item to identify the replaceable items, including:

[0052] An end-to-end reasoning model is constructed using a long short-term memory network (LSTM), which takes the feature data in the tree-structured knowledge base of different items as input to identify the similarity of reasoning characteristics of different items.

[0053] Setting a similarity threshold of the item reasoning characteristics, and screening similar items of the item with a similarity higher than the similarity threshold as replaceable items;

[0054] The end-to-end reasoning model constructed in this embodiment can infer alternative items with similar functions or characteristics from the item prior information library based on the characteristics of the newly input item; this reasoning method enables the robot to infer suitable alternative items based on the item characteristics when the target item is not available, thereby enhancing the flexibility of task execution.

[0055] Step 12: Based on the scene prior, the co-occurrence relationship between items and scenes is mined, and reasonable items are recommended in the set scene: natural language processing is used to extract robot planning element vocabulary such as items and scenes contained in the service execution requirements of the service robot, and statistical learning is used to perform statistical modeling on the extracted robot planning elements to form a probabilistic item-scene association model as the scene prior;

[0056] Step 12.1: Extraction of scene semantic elements, including:

[0057] Use natural language processing (NLP) technology to analyze the service robot service-related language information and extract key elements such as objects and scenes involved;

[0058] Use image recognition to label and classify objects and scenes in pictures or video data;

[0059] Step 12.2, Item-Scene Association Model:

[0060] Using statistical learning methods, we model the item-scene co-occurrence relationship and form a probabilistic item-scene association model to predict the most likely items to appear in different scenes.

[0061] For example, in the “office” scenario, the co-occurrence probability of documents, computers, and printers is high;

[0062] In the "kitchen" scenario, the co-occurrence probability of pots, ingredients, and seasonings is high;

[0063] The item-scene association model of this embodiment can output the occurrence probabilities of items in different scenarios. For example, when a robot is looking for a remote control, it should search the living room first, then the bedroom, rather than the kitchen. This reasoning method enables the robot to use environmental information to efficiently locate the target item, optimize the search strategy, and improve the task execution efficiency.

[0064] In this embodiment, prior knowledge is constructed respectively around the item prior and the scene prior. According to the item reasoning characteristics and the item-scene association relationship, the substitute items and the proposed existence scenarios of the items can be inferred, enabling the robot to effectively handle abnormal situations such as item absence and item non-location, increasing the selectivity of the robot's task planning and execution, and improving the environmental adaptability.

[0065] In step 2, the robot planning elements cover the core information that the robot needs to understand and process during the task planning process, and can provide the necessary context support for task generation, and can include items, actions, similar items, and existence scenarios;

[0066] Optionally, the relationships between the robot planning elements can include sequential relationships, functional relationships, similarity relationships, and positional relationships, etc.;

[0067] Such as Figure 3 and Figure 4 As shown, in step 2, the constructed directed graph is used as a graphical representation of the task planning. Specifically: according to the service-related natural language of the robot service execution requirements, four types of elements, namely actions, items, similar items (substitute items), and existence scenarios, are used as nodes, and the sequential relationship R 顺序 , functional relationship R 作用 , similarity relationship R 相似 , and positional relationship R 位置 are used as the edges between the nodes to construct a directed graph to represent the associations between actions, items, and the environment during the service execution process;

[0068] In this embodiment, the sequential relationship, functional relationship, similarity relationship, and positional relationship are used as the edges of the graph to represent the mutual connections between the above nodes, forming a directed graph;

[0069] In step 3, the language text information is converted into digital data, the adjacency matrix vector of the directed graph is extracted, and the natural language information related to the robot service is vectorized, and the vectorized representation of the task planning is obtained through fusion. As Figure 4 shown, the specific steps are as follows;

[0070] Step 31: Extract the adjacency matrix in the directed graph to obtain the vocabulary in the natural language information related to the robot service;

[0071] Step 32: Perform relationship representation on the adjacency matrix, perform lexical representation on the word vectors, and perform positional encoding on the lexical positions to obtain the vectorized digital representation;

[0072] In this embodiment, the vectorized representation of task planning includes three parts: relationship representation, lexical representation, and positional representation. The task planning is expressed in the form of a vector. Specifically:

[0073] Relationship representation: For the adjacency matrix of the directed graph, the graphic nodes and edges that make up the task planning are expressed in matrix form as the relationship representation;

[0074] Lexical representation: Using the word2vector method, the words in the service-related natural language information of the robot's execution service are represented in the form of word vectors as the lexical vectors. Word2Vec is a neural network-based word embedding technology used to convert the words in the text into continuous vectors of a fixed dimension.

[0075] Positional representation: Using one-hot encoding to perform positional encoding on the words in the natural language information of the robot's execution service to obtain the positional representation;

[0076] The robot task planning vector includes: the adjacency matrix vector of the directed graph, the word vectors corresponding to the semantic words in the task planning, and the one-hot vectors representing the positions of the semantic words in the task planning. These three categories together constitute the task planning vector.

[0077] Step 32: Vector fusion: Perform feature fusion on the digital vectors corresponding to the relationship representation, lexical representation, and positional representation through the attention mechanism to obtain the vectorized representation of the task planning;

[0078] Specifically, fuse the relationship vector, lexical vector, and positional vector in a concatenated manner and perform selective feature extraction under the guidance of the attention mechanism as the task planning representation result; Denote the obtained task planning sequence:

[0079] [Action 1, Item 1 (Similar Item 1, Scene 1)];

[0080] [Action 1, Item 1 (Similar Item 1, Scene 1)];

[0081] ……

[0082] [Action n, Item n (Similar Item n, Scene n)];

[0083] In the above steps 1 to 3, a task planning sequence is obtained from the natural language information (training data) of the robot performing services. First, prior knowledge reasoning is performed, and similar items and multiple possible scenarios can be obtained. Then, based on the fusion of the directed graph and the vectorized natural language data, multiple task plans are formed into a sequence. When a task plan executed by the robot fails, the next alternative task plan can be selected for execution. The above process involves prior knowledge reasoning (model), directed graph construction, and information fusion, with a complex processing process. Moreover, the results obtained at each step may not be accurate, requiring manual correction during the training and construction of the dataset, and the process is cumbersome and not conducive to the real-time generation of task planning. In this embodiment, the multiple task plans formed into a sequence obtained in step 3 are used as the model output, a task planning model is constructed and trained, and directly using the natural language information related to the robot service as the input, a task planning sequence containing multiple task plans can be obtained; this improves the real-time performance and accuracy of the robot task planning, reduces the failure rate of the robot executing commands, and improves the customer experience.

[0084] In step 4, a task planning model is constructed, including a template generation module and a template filling module. Two identical models can be built by combining the attention mechanism and the Transformer model, serving as the template generation model and the template filling model respectively, and generating task plans in a staged manner in the order from template generation to template filling.

[0085] The above task planning model includes two modules, namely the template generation module and the template filling module. Each module is designed based on the Transformer model, and both modules have corresponding inputs and outputs. The input of the template generation module is the natural language information related to the service, and the output is the task planning template. The input of the template filling model is the template vector and the vector corresponding to the natural language related to the service, and the output is the task plan with selectivity.

[0086] In step 4, a progressive method of template generation and template filling is adopted based on the deep learning model. First, a task planning template is generated based on the part-of-speech (POS) attributes, and then template filling is optimized using reinforcement learning to progressively generate the task plan for the robot's actions. The method for training the constructed task planning model includes the following steps:

[0087] Step 41: Construct a template generation model based on the Transformer model, and under the guidance of the attention mechanism, generate a task planning template characterized by part-of-speech (POS, Part of Speech) attributes.

[0088] Step 41-1: Obtain the input vector: Perform vector representation on the natural language information of the robot performing services, and fuse the word vector, sentence vector, and paragraph vector as the input vector of the template generation model.

[0089] Characterize the task planning template by part-of-speech attributes, and complete the generation of the task planning template with the assistance of the attention mechanism. The template generation process is as follows Figure 5 shown. Perform word vector representation on the service-related natural language information, and respectively obtain sentence vectors and segment vectors based on the word vectors. Then introduce weight coefficients to fuse the three types of vectors, and use the fused vector as the input vector of the template generation model.

[0090] Step 41-2: For the obtained input vector, based on the template generation model constructed by the attention mechanism and Transformer, generate a task planning template characterized by part-of-speech attributes (POS, Part of Speech);

[0091] In this embodiment, the template generation module is based on the Transformer network model. By taking the fused vector jointly composed of word vectors, sentence vectors and segment vectors as the input, under the guidance of the attention mechanism, generate a task planning template characterized by part-of-speech attributes (POS, Part of Speech). The form of the task planning template is as follows:

[0092] VB([NN,…],rel:[<VB,NN,lv-n>])

[0093] Among them, VB represents the verb part of speech, referring to the action contained in the sentence, which corresponds to the service execution action here; NN represents the noun part of speech, referring to the items contained in the sentence, which can correspond to household items; rel represents the association relationship between the item and the action, corresponding to the action relationship, sequence relationship, similarity relationship and position relationship.

[0094] In this embodiment, the template generation and template filling methods are realized through Transformer. First, generate a task planning template based on the grammatical structure, and then optimize the filling through reinforcement learning to improve the adaptability and generalization ability of the task planning. The process of template filling is described below;

[0095] Step 42: Build a template filling model based on the Transformer model. Use the template vector output by the template generation model and the input vector composed of word vectors, sentence vectors and segment vectors extracted from the natural language information of the robot's execution of the service as the input. Based on the reinforcement learning algorithm, fill the generated task planning template, and adjust the model parameters based on the set incentives to obtain the trained task planning model. The template filling process is as follows Figure 6 shown, including the following steps:

[0096] Step 42-1: Use the template vector and the input vector composed of word vectors, sentence vectors, and paragraph vectors as inputs, and transmit them to the constructed template filling model. Perform a dot product on the template features and the input vector to fill the planning template and form a task planning result.

[0097] Step 42-2: Based on the task planning result obtained in Step 42-1, calculate the semantic similarity incentive and the action logic incentive. Use the stochastic gradient descent method to iteratively adjust the generation strategy of the task planning through the incentive values corresponding to the semantic similarity incentive and the action logic incentive, and obtain the trained task planning model.

[0098] In this embodiment, the incentives in the reinforcement learning process mainly include the semantic similarity incentive and the action logic incentive.

[0099] (1) Semantic similarity incentive: For the similarity comparison of the input information and the output action sequence at the sentence level and the document level respectively, use the similarity as the semantic similarity incentive.

[0100] Specifically, the determination method of the semantic similarity incentive at the sentence level is as follows: At the sentence level, use the semantic role labeling (SRL) method to extract the comparison elements of the input sentence (i.e., the natural language information of the service robot's service execution requirements) and the output action sequence. Based on the elements extracted by SRL, perform a semantic similarity comparison with the corresponding action sequence at the sentence level, and the comparison result is the semantic similarity incentive at the sentence level.

[0101] Optionally, the extracted comparison elements can be expressed as:

[0102] (f A0 ,f A0 ,f A0 ,V);

[0103] Among them, f A0 represents the doer of the action; f A1 represents the direct recipient of the action, which can be a tool; f A2 refers to the indirect recipient of the action, which can be an item affected by the tool.

[0104] For example, when the robot wipes the table with a rag, here f A0 represents the robot, f A1 represents the rag, and f A2 represents the table. Finally, based on the elements extracted by SRL, perform a semantic similarity comparison with the corresponding action sequence at the sentence level, and the comparison result is the semantic similarity incentive at the sentence level.

[0105] Specifically, the method for determining the semantic similarity incentive at the document level is as follows: Compare the document vector composed of paragraph vectors with the average vector composed of each subtask planning vector respectively, and the result is used as the semantic similarity incentive at the document level;

[0106] Specifically, the task planning is generated based on service-related natural language information, and the corresponding natural language information can be converted into word vectors, sentence vectors, and paragraph vectors; the average vector corresponding to each paragraph vector is used as the document vector;

[0107] The subtask planning vector represents the task planning subtask vector, and the task planning is an overall composed of a combination of a series of actions and items. Therefore, each group of actions and items is a subtask, and they can be converted into subtask planning vectors

[0108] (2) Action logic incentive: In this embodiment, taking the document as a unit, extract the actions in the natural language information of the robot's execution service and the actions in the task planning respectively to form action lists L1 and L2; use the dynamic time warping (DTW, Dynamic Time Wrapping) method to compare the sequence similarity of L1 and L2, and the comparison value is used as the action logic incentive;

[0109] In this embodiment, different input vectors (word vectors, sentence vectors, paragraph vectors) are fused together when generating the task planning, and the combination of reinforcement learning and attention mechanism helps to generate a selective and accurate task planning.

[0110] In this embodiment, the random gradient descent method is used to guide the reinforcement learning process, and finally realize the conversion of service-related natural language into task planning represented by element labels.

[0111] The expression form of the task planning is: [action, item (similar item, existing scene)]. Among them, the similar item can represent a single item name or a list of item names, and the existing scene corresponds to the home scene where the item may appear in the action sequence.

[0112] Step 5: Obtain the natural language information instruction to be executed by the service robot, and based on the trained task planning model, use a progressive method of template generation and template filling to generate the task planning result of the robot's action; specifically,

[0113] Step 51: Obtain the input vector: Perform vector expression on the natural language information instruction to be executed, and fuse the obtained word vectors, sentence vectors, and paragraph vectors as the input vector of the trained template generation model; the implementation process of this step is the same as that of step 41;

[0114] Step 52: For the obtained input vector, generate a task planning template characterized by part-of-speech (POS) attributes through a template generation model; the implementation process of this step is the same as that of step 41;

[0115] Step 53: Use the template vector output by the template generation model and the input vector composed of word vectors, sentence vectors, and paragraph vectors extracted from the natural language information instruction to be executed as inputs, and transmit them to the trained template filling model. Perform a dot product on the template features and the input vector to fill the planning template and form a task planning result. The implementation process of this step is the same as that of step 42;

[0116] In this embodiment, the task planning representation of the existing method is relatively single, usually only generating a unique execution plan and unable to provide multiple coping strategies, resulting in the problem that it is difficult for the robot to adjust the task planning in abnormal situations and affecting the successful execution of the task. A progressive task planning generation strategy of template generation and template filling is proposed, that is, first generate a task planning template composed of part-of-speech tags (POS), and then fill the template to form a task planning for the robot, reducing the complexity of autonomous generation of task planning, meeting the system requirements of task planning without restricting the scale of natural language, and obtaining an accurate task planning based on reinforcement learning to ensure the successful execution of the robot service.

[0117] Embodiment 2

[0118] Based on Embodiment 1, a robot task planning system based on graph representation learning is provided in this embodiment, including:

[0119] A prior construction module, configured to construct a prior knowledge reasoning model based on item feature reasoning and scene co-occurrence relationship modeling;

[0120] A task planning graph representation module, configured to, according to the obtained natural language information of the robot's execution service, based on the constructed prior knowledge reasoning model, extract the robot planning elements to be processed in the planning as graph nodes, and use the relationships between the robot planning elements as the edges of the graph to construct a directed graph;

[0121] A vectorized representation module, configured to extract the adjacency matrix of the directed graph, and perform vectorized representation on the vocabulary and vocabulary positions in the natural language information related to the robot service, and obtain a vectorized representation of the task planning after fusion;

[0122] A training module, configured to use the natural language information related to the robot service as input and the vectorized representation of the task planning as output to train the constructed task planning model;

[0123] The task planning generation module is configured to obtain the natural language information instruction to be executed by the service robot, and based on the trained task planning model, adopt a progressive method of template generation and template filling to generate the task planning result of the robot's actions.

[0124] It should be noted here that each module in this embodiment corresponds to each step in Embodiment 1, and the specific implementation process is the same, so it will not be repeated here.

[0125] Embodiment 3

[0126] Based on Embodiment 1, this embodiment provides an electronic device, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps in the robot task planning method based on graph representation learning described in Embodiment 1 are completed.

[0127] The foregoing is only the preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

[0128] Although the specific implementation manners of the present disclosure have been described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solution of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A robot task planning method based on graph representation learning, characterized in that: The steps include: Construct a priori knowledge reasoning model based on item feature reasoning and scene co-occurrence relationship modeling; According to the natural language information of the robot execution service obtained, based on the constructed prior knowledge reasoning model, the robot planning elements to be processed by the planning are extracted as graph nodes, and the relationship between the robot planning elements is used as the edge of the graph to construct a directed graph; Extract the neighbor matrix of the directed graph, and vectorize the words and word positions in the natural language information related to the robot service, and then get the vectorized representation of the task planning after fusion; The constructed task planning model is trained with natural language information related to the robot service as input and vectorized representation of the task planning as output; The natural language information instructions to be executed by the service robot are obtained, and based on the trained task planning model, a progressive method of template generation and template filling is adopted to generate the task planning results of the robot action.

2. The robot task planning method based on graph representation learning as claimed in claim 1, characterized in that: Constructing a priori knowledge reasoning model including item prior and scene prior; According to the prior of items, the characteristics of items are structured, and a reasoning mechanism is established based on the structured characteristics to infer the substitutable items; For scene priors, natural language processing is used to extract the vocabulary of robot planning elements such as items and scenes included in the service execution requirements of the service robot. The extracted robot planning elements are statistically modeled using statistical learning methods to form a probabilistic item-scene association model as scene prior.

3. The robot task planning method based on graph representation learning as claimed in claim 1, characterized in that: Robot planning elements, including objects, actions, similar objects, and existing scenarios; The relationships between robot planning elements include sequential relationships, action relationships, similarity relationships, and position relationships.

4. The robot task planning method based on graph representation learning according to claim 1, characterized in that: According to the service-related natural language of the robot service execution requirements, a directed graph is constructed with four types of elements, namely actions, objects, similar objects, and existence scenarios, as nodes, and sequential relationships, action relationships, similarity relationships, and position relationships as edges between nodes, to represent the relationship between actions, objects, and environment during the service execution process.

5. The robot task planning method based on graph representation learning according to claim 1, characterized in that: Extract the neighbor matrix vector of the directed graph, and vectorize the natural language information related to the robot service, and integrate them to obtain the vectorized representation of the task planning. Specifically: Extract the neighbor matrix in the directed graph to obtain the vocabulary in the natural language information related to the robot service; Represent the relationship of the neighbor matrix, represent the word vector, encode the word position, and obtain the vectorized digital expression; The attention mechanism is used to fuse the digital vectors corresponding to the relationship representation, vocabulary representation, and position representation to obtain a vectorized representation of the task planning.

6. The robot task planning method based on graph representation learning as claimed in claim 5, characterized in that: For the neighbor matrix of the directed graph, the graph nodes and graph edges that constitute the task planning are expressed in a matrix as a relational representation; Using the word2vector method, the vocabulary in the service-related natural language information of the robot's execution of services is represented in the form of word vectors as vocabulary vectors; One-hot encoding is used to positionally encode the words in the natural language information of the robot's execution service to obtain positional representation.

7. The robot task planning method based on graph representation learning according to claim 1, characterized in that: The method for training the constructed task planning model includes the following steps: A template generation model is built based on the Transformer model. Under the guidance of the attention mechanism, a task planning template represented by word class attributes is generated. A template filling model is constructed based on the Transformer model. The template vector output by the template generation model and the input vector consisting of word vectors, sentence vectors and segment vectors extracted from the natural language information of the robot execution service are used as input. The generated task planning template is filled based on the reinforcement learning algorithm, and the model parameters are adjusted based on the set incentives to obtain the trained task planning model.

8. The robot task planning method based on graph representation learning as claimed in claim 7, characterized in that: A method for filling a generated task planning template based on a reinforcement learning algorithm, adjusting model parameters based on a set incentive, and obtaining a trained task planning model, including: The template vector and the input vector composed of word vector, sentence vector and segment vector are used as input and transmitted to the constructed template filling model. The template features and the input vector are multiplied together to fill the planning template and form the task planning result. Based on the obtained task planning results, the semantic similarity incentive and action logic incentive are calculated. The stochastic gradient descent method is used to iteratively adjust the task planning generation strategy through the incentive values ​​corresponding to the semantic similarity incentive and action logic incentive to obtain the trained task planning model.

9. A robot task planning system based on graph representation learning, characterized in that: include: A priori building module, configured to build a priori knowledge reasoning model based on item feature reasoning and scene co-occurrence relationship modeling; The task planning graph representation module is configured to extract the robot planning elements to be processed by the planning as graph nodes based on the acquired natural language information of the robot execution service and the constructed prior knowledge reasoning model, and to construct a directed graph with the relationship between the robot planning elements as the edges of the graph; A vectorization representation module is configured to extract a neighbor matrix of a directed graph and perform vectorization representation on words and word positions in natural language information related to the robot service, and obtain a vectorization representation of the task planning after fusion; A training module, configured to take natural language information related to the robot service as input and a vectorized representation of the task plan as output, to train the constructed task planning model; The task planning generation module is configured to obtain natural language information instructions to be executed by the service robot, and based on the trained task planning model, uses a progressive method of template generation and template filling to generate task planning results for robot actions.

10. An electronic device, characterized in that: The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps in the robot task planning method based on graph representation learning as described in any one of claims 1 to 8 are completed.

Images