A knowledge graph-based task map construction and task planning method

Abstract

The application discloses a kind of based on knowledge graph's task map construction and task planning method and equipment, method includes: constructing based on knowledge graph's task definition template;Acquire scene knowledge graph and extract scene state by scene information extraction algorithm;With initial scene state as starting node, find suitable task and simulate execution to generate new scene state, construct with scene state as node, task instance as edge's task map;According to target state in task map search path to obtain subtask sequence, execute by real-time comparison actual and predicted scene state trigger re-planning.This method can independently construct task map, reduce planning complexity, with dynamic environment adaptive capacity, applicable to robot task planning field.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and robotics task planning technology, and more specifically, to a method and medium for task map construction and task planning based on knowledge graphs. Background Technology

[0002] Robot task planning has always been a core research direction in the field of artificial intelligence, especially in on-orbit operations in space environments, characterized by high structure and uncertain task requirements. Traditional task planning methods, such as the classic PDDLStream planner, can couple action sequence planning with continuous variable solving to achieve task motion planning that considers action feasibility. However, these methods typically fall under the category of classical planning, often requiring the manual definition of complex planning files, and suffer from high computational complexity when dealing with large-scale state spaces. Furthermore, existing technologies include knowledge graph-based robot task-level planning methods, which build task maps by abstracting environmental semantic understanding and robot actions into nodes and edges in a knowledge graph. However, these existing methods usually require predefining object attributes in the scene and using combinations of object attributes to represent environmental states. This limits their applicability to task scenarios where the possible states of objects are completely known and task requirements are fixed, lacking the ability to autonomously adapt to dynamically changing environments and unexpected tasks. Summary of the Invention

[0003] The purpose of this invention is to provide a method and medium for task map construction and task planning based on knowledge graphs, aiming to solve the problems of high dependence on predefined definitions, difficulty in coping with dynamic environmental changes, and high planning complexity in the existing technology.

[0004] To achieve the above objectives, the first aspect of this invention provides a method for constructing a task map and planning tasks based on a knowledge graph, comprising: constructing a task definition template based on a knowledge graph; acquiring a scene knowledge graph, extracting the scene state at the current moment according to a scene information extraction algorithm, wherein the scene information extraction algorithm queries all relations in the scene knowledge graph, converts each relation record into a string, and represents the scene state at the current moment based on a list determined by all strings; and constructing a task map, wherein the task map construction process includes: taking the initial scene state as the starting node, searching for applicable tasks in the current scene according to the task definition template, simulating the execution of the applicable tasks, and utilizing the scene... The scene information extraction algorithm obtains the new scene state after execution. If the new scene state does not exist in the task map, a new state node is created and added to the waiting queue. The current state node and the new state node are connected by edges with the applicable task. The waiting queue is processed in a loop until all reachable states are traversed, thus forming a graph structure with scene states as nodes and task instances as edges. Path search is performed in the task map according to the target state to obtain a sub-task sequence. When executing the sub-task sequence, the scene information extraction algorithm is used to obtain the actual scene state in real time, and the actual scene state is compared with the expected state recorded during the planning process. If they are inconsistent, replanning is triggered.

[0005] Furthermore, the task definition template represents the task as T=[P,O,C,E]; where P represents the task executor, referring to the object node in the scene knowledge graph that has the ability to perform the task; O represents the task object, referring to all participating object nodes in the task instance except for the executor; C represents the task execution condition, which is determined by querying the relationships and attributes of nodes in the knowledge graph; and E represents the task execution operation, which represents the impact of a specific task on the task scene by adding, deleting, modifying, and querying the relationships and attributes in the knowledge graph.

[0006] Furthermore, converting each relationship record into a string includes: obtaining the relationship type, start node name, and end node name from the relationship record; and concatenating the start node name, relationship type, and end node name to form a string format.

[0007] Furthermore, the simulated execution of applicable tasks includes: performing task execution operations according to the task definition template, and performing node relationship addition, deletion or modification operations in the scene knowledge graph.

[0008] Furthermore, the step of performing path search in the task map based on the target state to obtain a sub-task sequence includes: extracting the current initial scene state using a scene information extraction algorithm and matching the corresponding starting state node in the task map; searching in the task map for the optimal path from the starting state node to the ending state node representing the target state; and parsing the edge information on the optimal path to obtain an ordered sub-task sequence.

[0009] Furthermore, after obtaining the ordered sequence of subtasks, the method further includes: for each subtask in the subtask sequence, searching the pre-built experience sequence library to see if there is planning experience knowledge for that subtask; if it exists, the corresponding action sequence is directly reused; if it does not exist, a pre-built intelligent planning algorithm is called to solve the action sequence, and the solution is stored in the pre-built experience sequence library.

[0010] Furthermore, the triggering of replanning includes: before executing a subtask, using a scene information extraction algorithm to obtain the current actual scene state and comparing it with the expected pre-execution state corresponding to the subtask; after executing a subtask, using a scene information extraction algorithm to obtain the current actual scene state and comparing it with the expected post-execution state corresponding to the subtask; if any of the above comparison results are inconsistent, execution is stopped, and the task map is called again for planning with the current actual scene state as a new starting point.

[0011] Furthermore, the method is applied to the on-orbit operation scenario of a space multi-arm robot, and the scenario knowledge graph contains semantic information about the robot, target spacecraft, functional modules, batteries, and end effectors and their interrelationships.

[0012] Furthermore, the task types defined by the task definition template include at least functional module disassembly, functional module assembly, battery disassembly, battery assembly, end effector assembly, and end effector disassembly.

[0013] To achieve the above objectives, a second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0014] Because existing task planning methods rely heavily on predefined definitions, struggle to adapt to dynamic environmental changes, and have high planning complexity, this invention provides a task map construction and task planning method and medium based on knowledge graphs. The method improves the robustness and autonomy of the robot system by autonomously constructing task maps, reducing planning complexity, possessing dynamic replanning capabilities, and having an experience reuse mechanism.

[0015] Compared with existing technologies, the advantages of this invention are as follows: Autonomous construction of task maps. By designing a task definition template based on a knowledge graph and a scene information extraction algorithm, this invention can start from the initial scene state, autonomously explore and simulate the execution of applicable tasks, and gradually construct a complete task map with scene states as nodes and task instances as edges. This construction method does not rely on predefined combinations of object attributes, but is automatically generated through graph evolution, ensuring the completeness of the task map. Reduced planning complexity: After the task map is constructed, the originally complex task planning problem is transformed into a path search problem between nodes in the task map, greatly improving planning efficiency and enabling rapid task planning. Dynamic replanning capability: By comparing the actual scene state with the expected state in real time, this invention can quickly capture sudden changes in the environment (such as equipment failure) and immediately trigger replanning, effectively dealing with the uncertainty of task requirements and improving the robustness and autonomy of the robot system. Experience reuse mechanism: Introducing an experience sequence library allows for the reuse of already solved sub-task action sequences, avoiding repeated calls to complex intelligent planning algorithms and further improving planning speed. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an on-orbit replacement scenario provided by an embodiment of the knowledge graph-based task map construction and task planning method of the present invention; Figure 2 This is a schematic diagram of a knowledge graph representing an on-orbit replacement scenario provided in an embodiment of the knowledge graph-based task map construction and task planning method of the present invention; Figure 3 This is a schematic diagram of a task map construction example provided in an embodiment of the task map construction and task planning method based on knowledge graphs of the present invention; Figure 4 This is a flowchart of a method provided in an embodiment of the knowledge graph-based task map construction and task planning method of the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] Example 1: A method for task map construction and task planning based on knowledge graphs This embodiment provides a knowledge graph-based method for task map construction and task planning. This method can be applied to various electronic devices with computing capabilities, especially to space robot control systems and autonomous mobile robot planning systems. Figure 4 An optional flowchart of the method in this embodiment is shown, such as... Figure 4 As shown, the method specifically includes the following steps: Step S1: Construct a task definition template based on knowledge graph.

[0020] In this embodiment, to enable the robot system to understand and execute tasks, a standardized task definition template first needs to be constructed. The task definition template is designed based on a knowledge graph framework and is used to describe the execution logic of a task in a specific scenario.

[0021] Specifically, the task definition template represents a task as a quadruple. Where: P (Performer): Refers to the object node in the scenario knowledge graph that has the ability to perform tasks. For example, in the space on-orbit service scenario, it can be "Manipulator1" or "Manipulator2".

[0022] O (Task Object): Refers to all participating object nodes in the task instance except for the executor. For example, "Functional Module A-1" to be replaced, "Battery_3" as a spare battery, etc.

[0023] C (Task Execution Condition): Refers to the relational or attribute conditions that relevant object nodes must satisfy before executing a specific task. This condition is determined by querying the relationships and attributes of nodes in the knowledge graph. For example, for the "disassemble module" task, the conditions might include "the robotic arm has a matching end effector installed" and "the target module is located at the designated workstation".

[0024] E (Task Execution Operation, Effect): Refers to the impact on the scene knowledge graph under the premise that the execution conditions are met. This impact is represented by adding, deleting, modifying, and querying (CRUD) operations on the relationships and attributes in the knowledge graph. For example, after executing the "grabbing" task, a "holding" relationship will be established between the "robotic arm" and the "module" in the knowledge graph, while the "located_at" relationship between the module and the original workstation will be deleted.

[0025] Step S2: Obtain the scene knowledge graph and extract the scene state at the current moment according to the scene information extraction algorithm.

[0026] When constructing a task map, it is first necessary to extract key information from the scene knowledge graph at a specified time. This paper defines the list of all relationships in the graph as the scene state at that time and stores it in the knowledge graph as a key attribute in the form of state nodes. This process realizes the representation of the scene state at a specific time.

[0027] This step aims to transform complex graph-structured data into textual state data that is easy to compare and retrieve, thereby achieving a dimensionality reduction representation of the scene.

[0028] Specifically, the workflow of the scene information extraction algorithm is as follows: Input the knowledge graph (KG) of the current scene.

[0029] Query all triplet relationship records in the scenario knowledge graph.

[0030] Iterate through each relationship record and extract its relationship type (relationshipType), start node name (startName), and end node name (endName).

[0031] According to the preset concatenation rules, the above three pieces of information are concatenated into a string format. For example, the concatenation format can be set to "starting node name - relation type - ending node name".

[0032] All concatenated strings are combined to generate a list of key states. This list serves as the unique representation of the scene state at that moment.

[0033] Step S3: Construct the task map. The task map construction process includes: starting with the initial scene state as the starting node, searching for applicable tasks in the current scene according to the task definition template, simulating the execution of the applicable tasks, and using the scene information extraction algorithm to obtain the new scene state after execution; if the new scene state does not exist in the task map, creating a new state node and adding it to the waiting queue, connecting the current state node and the new state node with the applicable task as an edge; cyclically processing the waiting queue until all reachable states have been traversed, thus forming a graph structure with scene states as nodes and task instances as edges.

[0034] This step is one of the core steps of the present invention. By simulating and deducing, it automatically explores all possible scenario evolution paths starting from the initial state, thereby constructing a complete task map.

[0035] Specifically, the process of constructing a task map includes the following sub-steps: Initialization: Use the initial scene state as the starting node of the task map and create a pending queue (ASL) to put the index of the starting node into the queue.

[0036] Loop processing: When the queue to be processed is not empty, perform the following operations: Retrieve the first status index (status_idx) from the queue and reproduce the corresponding scene knowledge graph environment (EK_curr) based on the index.

[0037] The current scene knowledge is obtained by using a scene information extraction algorithm, and its corresponding starting state node in the task map is determined.

[0038] Based on the task definition template in step S1, search for all applicable tasks in the currently reproduced scene knowledge graph. The criterion for determining an applicable task is that the current graph meets the execution conditions in the task definition.

[0039] For each applicable task found, simulated execution is performed: based on the execution operations in the task definition, add, delete, or modify node relationships in the simulated graph.

[0040] The new scene knowledge after the simulation is executed is obtained by using a scene information extraction algorithm, and a new scene state is generated.

[0041] The knowledge graph in the scene information extraction algorithm obtains basic information about objects from the simulation scene according to preset rules and updates the knowledge graph. Its function is to highly summarize and convert the knowledge graph before and after the update into a string set.

[0042] Node and edge updates: Determine whether the new scene state already exists in the mission map.

[0043] If it does not exist, a new state node is created, and the index of the new node is added to the tail of the pending queue.

[0044] If it exists, then directly obtain the existing state node.

[0045] Finally, create an edge between the current state node and the new (or existing) state node. The edge's attribute is the applicable task (i.e., the task instance) that was just executed.

[0046] End: When the queue to be processed is empty, it means that all reachable states have been traversed and recorded, and the task map construction is complete.

[0047] Through the above process, the constructed task map forms a special graph structure: nodes represent scene states, and edges represent task instances. This structure clearly describes the evolution of the scene under different task-driven conditions, transforming the complex task planning problem into a path search problem on a graph.

[0048] Step S4: Perform path search on the task map based on the target state to obtain a sub-task sequence; when executing the sub-task sequence, use the scene information extraction algorithm to obtain the actual scene state in real time, and compare the actual scene state with the expected state recorded during the planning process. If they are inconsistent, trigger replanning.

[0049] Once the task map is built, the planning process becomes efficient and intuitive.

[0050] Path search: The current actual scene state is extracted using a scene information extraction algorithm, and the corresponding starting state node is matched in the task map. Simultaneously, the ending state node corresponding to the target state is determined. The optimal path from the starting node to the ending node is searched in the task map using a graph search algorithm (such as Dijkstra's algorithm, A* algorithm, etc.). The edge information on this path is parsed to obtain an ordered sequence of subtasks.

[0051] Execution and Monitoring: When actually executing the sub-task sequence, this invention introduces a closed-loop feedback mechanism.

[0052] Before executing each subtask, the current actual scene state is obtained using a scene information extraction algorithm and compared with the expected pre-execution state of the corresponding subtask recorded during the planning process.

[0053] After executing each subtask, the actual scene state is retrieved again and compared with the expected state after execution.

[0054] Replanning trigger mechanism: If any of the above comparison results are inconsistent (for example, a sudden event causes a module to be damaged, or an unexpected fault relationship occurs in the actual state), the current execution process is immediately stopped, and the current actual scenario state is used as the new starting state to re-call the task map for planning (i.e., re-execute step S4), thereby achieving a rapid response to sudden situations.

[0055] Example 2: Action sequence planning optimization based on experience base Based on the above embodiment one, this embodiment further optimizes the processing procedure after obtaining the sub-task sequence in step S4. Since the sub-task sequence only represents the task-level planning (such as "removing the battery"), the robot still needs specific action sequences (such as "moving joint 1", "closing the gripper", etc.) to execute.

[0056] Specifically, after obtaining the ordered sequence of subtasks, the method also includes: Build an experience sequence library: Pre-build a database (experience sequence library) to store experience knowledge for subtask planning.

[0057] Experience retrieval: For each subtask in the subtask sequence, the system first searches the experience sequence library to see if there is any planning experience knowledge for that subtask.

[0058] Reuse or solve: If an action sequence for "using robotic arm 1 to disassemble type A battery" exists (for example, the system has previously successfully planned such an action sequence), the corresponding action sequence can be reused directly from the library without recalculation, significantly reducing planning time.

[0059] If it does not exist, a pre-built intelligent planning algorithm (such as PDDLStream, state space search, etc.) is called to solve the action sequence, and the solution is stored in the experience sequence library to enrich the knowledge reserve for future reuse.

[0060] Example 3: Application of Space Multi-Arm Robot in On-Orbit Operation Scenarios The following is combined with Figures 1 to 3 Taking a specific scenario of on-orbit replacement of a space multi-arm robot as an example, the application of the method of the present invention will be described in detail.

[0061] Employing simulation scenarios such as Figure 1 Assume that the target spacecraft has 2 operable objects, and the space multi-arm robot has 1 spare object. In this mission, modules require batteries with the same serial number to operate, and modules with different serial numbers must be operated by different end effectors. It is essential to ensure that all functional modules on the target spacecraft function properly. A scenario knowledge graph should be established for this scenario, as follows: Figure 2 As shown.

[0062] By analyzing the mission, the on-orbit replacement mission is decomposed into six basic sub-tasks: functional module disassembly, functional module assembly, battery disassembly, battery assembly, end effector assembly, and end effector disassembly. Furthermore, to ensure that the constructed mission map covers all possible states within the capabilities of the space multi-arm robot, two special missions—functional module damage and battery damage—are introduced. The eight missions are defined using a mission definition method. Taking the functional module disassembly sub-task as an example, this mission can be described as follows: When a functional module on the target spacecraft is in a failed state, and a robotic arm equipped with a corresponding type of end effector exists, the robotic arm can transfer this functional module from its working position to a suitable, vacant storage location on the space multi-arm robot.

[0063] After completing all task definitions, and combining the scenario knowledge graph, Algorithm 2 was used to construct the task map for the on-orbit maneuvering task of the space multi-arm robot. Some task maps are shown below. Figure 3As shown, each node represents a unique scene state, and each edge represents a task instance, demonstrating how scene states transition and evolve between state nodes through the actions of different task instances. Specifically, by... Figure 3 It can be seen that state 525 can transition to state 1965, the target state in which the target spacecraft is operating normally, under the action of the two subtask instances "function_module_on" and "battery_on".

[0064] The design replacement requirement is that module A-1 is damaged. The task planning algorithm is called to solve the problem, and the task sequence is shown in Table 1.

[0065] Table 1. Subtask sequence for changing scenario 1

[0066] The design replacement requirements are that module A-1, module B-2, battery A-1, and battery B-1 are damaged. The task planning algorithm is called to solve the problem, and the task sequence is shown in Table 2.

[0067] Table 2. Scene Change Task Sequence

[0068] The design involves a sudden damage event. After the 14th subtask is executed, the functional module Module_1_1, which was originally working normally, is damaged. For this scenario, the result obtained during the initial planning is completely consistent with the previous one. Before the 15th subtask is executed, the system detects the sudden damage event that causes the expected state to be inconsistent with the actual state through state comparison. Therefore, the subsequent tasks are no longer executed, and the system enters replanning. The replanning task sequence is shown in Table 3.

[0069] Table 3 Replanning Service Sequence

[0070] The analysis of the simulation results shows that, under the combined effect of the proposed planning method and the replanning method, the space multi-arm robot can autonomously plan and execute tasks for target spacecraft with replacement needs within its own capabilities, thus verifying the effectiveness of the task map construction and task planning method proposed in this invention.

[0071] Based on the above embodiments, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods provided above.

Claims

1. A method for task map construction and task planning based on knowledge graphs, characterized in that, include: Construct a task definition template based on knowledge graphs; Obtain the scene knowledge graph, extract the scene state at the current moment according to the scene information extraction algorithm. The scene information extraction algorithm queries all the relationships in the scene knowledge graph, converts each relationship record into a string, and represents the scene state at the current moment based on the list determined by all the strings. The task map is constructed as follows: starting with the initial scene state as the starting node, the applicable task is found in the current scene according to the task definition template, the applicable task is simulated and executed, and the new scene state after execution is obtained using the scene information extraction algorithm; if the new scene state does not exist in the task map, a new state node is created and added to the waiting queue, and the current state node and the new state node are connected by the applicable task as an edge; the waiting queue is processed in a loop until all reachable states are traversed, thus forming a graph structure with scene states as nodes and task instances as edges; Based on the target state, a path search is performed on the task map to obtain a sequence of subtasks. When executing the sequence of subtasks, the actual scene state is obtained in real time using a scene information extraction algorithm, and the actual scene state is compared with the expected state recorded during the planning process. If they are inconsistent, replanning is triggered.

2. The method for task map construction and task planning based on knowledge graphs according to claim 1, characterized in that, The task definition template represents the task as T=[P,O,C,E]; Where P represents the task executor, which refers to the object node in the scene knowledge graph that has the ability to perform tasks; O represents the task object, referring to all participating object nodes in the task instance except for the executor; C represents the task execution conditions, which are determined by querying the relationships and attributes of nodes in the knowledge graph; E represents the task execution operation, which represents the impact of a specific task on the task scenario by adding, deleting, modifying, and querying the relationships and attributes in the knowledge graph.

3. The method for constructing a task map and planning a task based on a knowledge graph according to claim 1, characterized in that, The process of converting each relation record into a string includes: Retrieve the relation type, start node name, and end node name from the relation record; The starting node name, relationship type, and ending node name are concatenated to form a string.

4. The method for task map construction and task planning based on knowledge graphs according to claim 1, characterized in that, The simulated execution of applicable tasks includes: Based on the task execution operations in the task definition template, perform operations to add, delete, or modify node relationships in the scene knowledge graph.

5. The method for task map construction and task planning based on knowledge graphs according to claim 1, characterized in that, The step of performing path search on the task map based on the target state to obtain a sequence of sub-tasks includes: The current initial scene state is extracted using a scene information extraction algorithm, and the corresponding starting state node is matched in the task map. Search the task map for the optimal path from the starting state node to the ending state node representing the target state. The edge information on the optimal path is analyzed to obtain an ordered sequence of subtasks.

6. The method for constructing a task map and planning a task based on a knowledge graph according to claim 5, characterized in that, After obtaining the ordered sequence of subtasks, the method further includes: For each subtask in the subtask sequence, search the pre-built experience sequence library to see if there is any planning experience knowledge for that subtask. If it exists, the corresponding action sequence is reused directly; If it does not exist, a pre-built intelligent planning algorithm is invoked to solve for the action sequence, and the solution is stored in a pre-built experience sequence library.

7. The method for constructing a task map and planning a task based on a knowledge graph according to claim 1, characterized in that, The triggering of replanning includes: Before executing a subtask, the current actual scene state is obtained using a scene information extraction algorithm and compared with the expected pre-execution state corresponding to the subtask. After executing a subtask, the current actual scene state is obtained using a scene information extraction algorithm and compared with the expected state after the execution of the subtask. If any of the above comparison results are inconsistent, execution will be suspended, and the task map will be called again for planning, starting from the current actual scenario state.

8. A method for constructing task maps and planning tasks based on knowledge graphs according to any one of claims 1 to 7, characterized in that, The method is applied to the on-orbit operation scenario of a space multi-arm robot. The scenario knowledge graph contains semantic information about the robot, target spacecraft, functional modules, batteries, and end effectors, as well as their interrelationships.

9. The method for constructing a task map and planning a task based on a knowledge graph according to claim 8, characterized in that, The task types defined in the task definition template include at least functional module disassembly, functional module assembly, battery disassembly, battery assembly, end effector assembly, and end effector disassembly.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 9.

Images