A Mars rover mission planning method, apparatus, and system for HTN

Abstract

This invention discloses a method, apparatus, and system for Mars rover mission planning oriented towards HTN (Human Threat Network), belonging to the field of Mars exploration technology. The method includes: inputting known domain knowledge data corresponding to Mars rover mission execution data into a constructed domain knowledge learning model; constructing a decomposition relationship between composite tasks and operator sequences based on a hierarchical task network structure learning model by decomposing relational factors; obtaining the hierarchical task network structure of composite tasks using an iterative optimization method; outputting domain knowledge elements using the operator and decomposition method learning model; and fusing the outputs of the two models to obtain a fused domain knowledge file. This provides a powerful reference for intelligent planners to compile Mars rover mission domain knowledge, thereby realizing Mars rover mission planning. This application overcomes the technical problems of low efficiency, high labor costs, and low modification efficiency in the construction of domain knowledge for Mars rover mission planning.

Description

Technical Field

[0001] This invention belongs to the field of Mars exploration technology, and more specifically, relates to a Mars rover mission planning method, apparatus and system for HTN. Background Technology

[0002] Mars rover missions are lengthy, typically lasting several months or even years. Furthermore, the complex and ever-changing deep-space environment poses numerous challenges to mission safety. Additionally, significant communication delays between the Mars rover and Earth limit data transmission capabilities. Given these complex constraints, the dynamic and unpredictable extraterrestrial environment, and limited communication, research into intelligent mission planning for Mars rovers is ongoing to provide effective decision support. HTN planning, as a widely used intelligent mission planning method, is one such approach. Figure 1 The conceptual model of the planner is presented, showing that its input mainly consists of three parts: domain knowledge, system state, and task network. Domain knowledge is a "prescription" and tool for planning problems within a specific domain, typically containing specific decomposition and execution rules, represented as a set of decomposition methods and operators. The system state and task network together constitute the planning domain problem, describing the real-time state information of the current environment and the task objectives to be achieved.

[0003] Existing research on intelligent planning for Mars rover missions assumes the existence of a complete domain knowledge model. Domain knowledge is the foundation for applying intelligent planning methods; reasonable and complete domain knowledge can effectively improve planning efficiency and increase the accuracy of planning results; conversely, inadequate domain knowledge will severely hinder intelligent planning research. However, domain knowledge is often difficult to acquire, and its construction is frequently a bottleneck in knowledge engineering. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a Mars rover mission planning method, apparatus, and system oriented towards HTN (Hyper-Hyper-Number). Its purpose is to input known domain knowledge data corresponding to Mars rover mission execution data into a constructed domain knowledge learning model. Based on a hierarchical task network structure learning model and an operator and decomposition method learning model, it outputs decomposition methods and domain knowledge elements oriented towards HTN. The resulting domain knowledge file, obtained by fusing these two elements, provides a powerful reference for intelligent planners to compile Mars rover mission domain knowledge, thereby achieving Mars rover mission planning. This overcomes the technical problems of low efficiency, high labor costs, and low modification efficiency in Mars rover mission planning domain knowledge construction.

[0005] To achieve the above objectives, according to one aspect of the present invention, a Mars rover mission planning method for HTN is provided, comprising: S1: Construct a domain knowledge learning model for the Mars rover, including a hierarchical task network structure learning model and an operator and decomposition method learning model; S2: Input the domain knowledge known data corresponding to the task execution data into the domain knowledge learning model, so that the hierarchical task network structure learning model can learn the domain knowledge known data and construct the decomposition relationship between the composite task and the operator sequence by decomposing relation factors. The hierarchical task network structure of the composite task is obtained by using the iterative optimization method. Then, the operator and decomposition method learning model is used to learn the domain knowledge known data, the decomposition method and the hierarchical task network structure and output the domain knowledge elements. Finally, the domain knowledge file obtained by fusing the HTN-oriented decomposition method and the domain knowledge elements is output. The domain knowledge elements include: the preconditions of the decomposition method, the preconditions of the operators, and the addition and deletion states of the operators. S3: Input the domain knowledge file into the HTN planner to obtain an action plan for task execution.

[0006] In one embodiment, the hierarchical task network structure learning model includes: The decomposition structure optimization module is used to input the known domain knowledge data, take the execution task name as the composite task name, construct the decomposition relationship between the composite task and the operator sequence using heuristic decomposition relation factors, and finally obtain the composite task decomposition method structure. The hierarchical task network learning module is used to obtain the hierarchical task network structure based on the composite task decomposition method structure and using a recursive learning method.

[0007] In one embodiment, the decomposition structure optimization module is configured to: Randomly select an execution task name np and its corresponding operator sequence ax. If the matching success rate is 1, it means that the currently set operator sequence has a good matching degree. At this time, only the length of the operator sequence is considered. Randomly select an action execution sequence with a matching instance from the action execution sequence set, and add or delete operator sequences according to the action execution sequence. Calculate the decomposition structure matching rate Q of all execution task names in the execution task name set after the change. If the Q of all execution task names does not decrease, the optimization is retained; otherwise, it is not retained, and the next round of optimization is performed. Repeat the above steps to obtain an ideal task decomposition structure, where Q is 1 for all task names.

[0008] In one embodiment, the hierarchical task network learning module is used for: Using a bottom-up approach, starting with the operator sequence obtained from the final decomposition of the task, the most frequently occurring operator subsequence is selected as a hierarchical task network for decomposition methods, and new method names are automatically generated to represent the corresponding subsequences. The corresponding subsequences in the operator sequence are then replaced with method names, and the operator sequence probability matrix for the task is obtained again. By iterating the same method repeatedly through the operator sequence, a hierarchical task network for executing task decomposition methods is finally obtained.

[0009] In one embodiment, the operator and decomposition method learning model includes: The domain knowledge construction module is used to encode the action execution trajectory into a relational database, encode domain knowledge information into logical formulas, and obtain the degree of satisfaction of domain knowledge information by learning the weights of logical formulas, thus initially constructing domain knowledge. The conflict detection and repair module is used to substitute the initially constructed domain knowledge into the original action execution trajectory data, detect conflicts, obtain the specific location and data of the conflicts, and store them in the conflict set; randomly select a conflict in the conflict set, find a way to repair the conflict and repair it, and then obtain the domain knowledge elements.

[0010] In one embodiment, the domain knowledge elements further include resource constraints; the operator and decomposition method learning model further includes: The resource constraint acquisition module is used to obtain corresponding resource constraints by taking the domain knowledge elements and the historical resource data as inputs. The resource constraints include resource change time and resource change amount.

[0011] In one embodiment, the domain knowledge known data includes: the action execution sequence of the patrol vehicle, the action execution trajectory, the name of the execution task, the name of the operator, the state description, and the parameter type; The sources of the patrol device's action execution trajectory include: action execution data, historical status data, historical resource data, and historical status data of the external environment.

[0012] In one embodiment, before step S2, the method further includes: uniformly converting the format of the task execution data into a preset format, and then extracting features from it to obtain the domain knowledge known data.

[0013] According to another aspect of the present invention, a Mars rover mission planning device for HTN is provided, comprising: The module constructs a domain knowledge learning model for the Mars rover, including a hierarchical task network structure learning model and an operator and decomposition method learning model. The acquisition module is used to input the domain knowledge known data corresponding to the task execution data into the domain knowledge learning model, so that the hierarchical task network structure learning model can learn the domain knowledge known data and construct the decomposition relationship between the composite task and the operator sequence by decomposing relation factors. The hierarchical task network structure of the composite task is obtained by using an iterative optimization method. Then, the operator and decomposition method learning model is used to learn the domain knowledge known data, the decomposition method and the hierarchical task network structure and output the domain knowledge elements. Finally, the domain knowledge file obtained by fusing the HTN-oriented decomposition method and the domain knowledge elements is output. The domain knowledge elements include: the preconditions of the decomposition method, the preconditions of the operators, and the addition and deletion states of the operators. The planning module is used to input the domain knowledge file into the HTN planner to obtain action plans for task execution.

[0014] According to another aspect of the present invention, a Mars rover mission planning system for HTN is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0015] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This application constructs a domain knowledge learning model for Mars rovers. Known domain knowledge data corresponding to the execution data of the Mars rovers' mission is input into the domain knowledge learning model. Based on a hierarchical task network structure learning model and an operator and decomposition method learning model, it outputs decomposition methods and domain knowledge elements oriented towards HTN. The resulting domain knowledge file, obtained by fusing these two models, can provide a powerful reference for intelligent planners to compile domain knowledge for Mars rovers' missions, overcoming the low efficiency, high labor costs, and low modification efficiency faced in constructing domain knowledge for Mars rovers' mission planning. This promotes the application of HTN intelligent planning methods in the field of Mars rovers' mission planning, providing domain knowledge model support for the subsequent autonomous planning and execution of Mars rovers.

[0016] (2) The hierarchical task network structure learning model described in this scheme includes a decomposition structure optimization module and a hierarchical task network learning module. The decomposition structure optimization module constructs decomposition relationships by decomposing relation factors, and the hierarchical task network learning module obtains the hierarchical task network by using an iterative optimization method. The algorithm model has low computational complexity and high accuracy.

[0017] (3) In this scheme, the decomposition structure optimization module randomly selects an action execution sequence that matches the instance from the action execution sequence set, and adds or deletes operator sequences according to the action execution sequence. It then calculates the decomposition structure matching rate Q of all execution task names in the set after the change. If the Q of all execution task names does not decrease, the optimization is retained; otherwise, it is not retained, and the next round of optimization is performed until an ideal execution task decomposition structure is finally obtained. This scheme ensures that the execution task name np and the operator sequence ax are scientifically matched to obtain an ideal execution task decomposition structure, providing a foundation for subsequent algorithms.

[0018] (4) The hierarchical task network learning module of this scheme uses a bottom-up approach to select the most frequently occurring operator subsequence as a hierarchical task network for decomposing methods, and automatically generates new method names to represent the corresponding subsequences. It then replaces the corresponding subsequences in the operator sequence with the method names and obtains the operator sequence probability matrix for executing the task again. Through iteration, a scientific hierarchical task network structure is obtained to ensure the effectiveness of domain knowledge.

[0019] (5) The operator and decomposition method learning model described in this scheme includes: a domain knowledge construction module and a conflict detection and repair module. Based on a relational database, through weight learning, conflict detection and repair, more reliable information such as preconditions, addition status, and deletion status is obtained, thereby improving the completeness and effectiveness of domain knowledge and providing a reliable information foundation for task planning.

[0020] (6) The operator and decomposition method learning model described in this scheme also includes: a resource constraint acquisition module, which improves the resource constraint information of domain knowledge and provides convenience for task planning.

[0021] (7) This solution converts the format of task execution data into a preset format and then extracts its features to obtain the known data of the domain knowledge. It can obtain the input data required by the method, realize the unification of the format of task execution data, the extraction of effective information and the deletion of error information. Attached Figure Description

[0022] Figure 1 A schematic diagram of the conceptual model for HTN planning.

[0023] Figure 2 This is a schematic diagram of a Mars rover mission planning method for HTN provided in an embodiment of the present invention.

[0024] Figure 3a A schematic diagram of the algorithm framework corresponding to the domain knowledge learning model of a Mars rover provided in an embodiment of the present invention.

[0025] Figure 3b The main flowchart of task execution data processing provided in an embodiment of the present invention is shown.

[0026] Figure 4 This is a schematic diagram of an action execution trajectory conversion method provided in an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram illustrating a parameter type acquisition method provided in an embodiment of the present invention.

[0028] Figure 6 The diagram shows the algorithm framework for a hierarchical task network structure learning model provided in one embodiment of the present invention.

[0029] Figure 7 The diagram shows the algorithm framework for the decomposition structure optimization module provided in one embodiment of the present invention.

[0030] Figure 8 The diagram shows the algorithm framework for the hierarchical task network learning module provided in one embodiment of the present invention.

[0031] Figure 9 This is an algorithmic framework diagram of an operator and decomposition method learning model provided in an embodiment of the present invention.

[0032] Figure 10 The flowchart shows the algorithm for the conflict detection and repair module provided in one embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] like Figure 2 As shown, a Mars rover mission planning method for HTN is provided, including: S1: Construct a domain knowledge learning model for the Mars rover, including a hierarchical task network structure learning model and an operator and decomposition method learning model; S2: Input the domain knowledge known data corresponding to the task execution data into the domain knowledge learning model, so that the hierarchical task network structure learning model can learn the domain knowledge known data and construct the decomposition relationship between the composite task and the operator sequence by decomposing relation factors. The hierarchical task network structure of the composite task is obtained by using the iterative optimization method. Then, the operator and decomposition method learning model is used to learn the domain knowledge known data, decomposition method and hierarchical task network structure and output domain knowledge elements. Finally, the domain knowledge file obtained by fusing the decomposition method and domain knowledge elements for HTN is output. The domain knowledge elements include: the preconditions of the decomposition method, the preconditions of the operator, and the addition and deletion states of the operator. S3: Input the domain knowledge file into the HTN planner to obtain action plans for task execution.

[0035] Figure 3a This paper illustrates the algorithm framework for the domain knowledge learning model of a Mars rover. Mars rover mission execution data, generated and stored during the rover's historical exploration cycles as it executes corresponding action commands, serves as a crucial information source and basis for the domain knowledge learning algorithm. However, because this data is stored in parallel by multiple subsystems of the Mars rover, format incompatibility issues exist; furthermore, many subsystems automatically add storage records according to time nodes, resulting in a large amount of redundant data unrelated to domain knowledge. Additionally, due to the uncertainty of the Mars rover's external environment, the stored mission execution data may contain significant errors or inconsistencies. Therefore, this invention first processes the Mars rover's mission execution data to extract domain knowledge-related information as known domain knowledge data.

[0036] In one embodiment, the known domain knowledge data may include a rover action execution sequence, a rover action execution trajectory, a rover task name, a rover operator name, a rover state description, and a rover parameter type. Specifically, the rover action execution sequence describes the action commands successfully executed by the Mars rover; the rover action execution trajectory describes the successful action commands of the Mars rover and the state information before and after execution; the rover task name describes the names and parameter types of all execution tasks involved in the Mars rover task execution data; similarly, the rover operator name describes the names and parameter types of all operators involved in the Mars rover task execution data; and the parameter type describes all types of parameters designed in the Mars rover task execution data and the specific instances involved for each type of parameter.

[0037] Figure 3bThis demonstrates the main workflow of mission execution data processing. The ground planning system and the Mars rover are two entity objects and the sources of mission execution data. Mission execution data is recorded by the ground planning system during mission planning, while action execution data, historical state data, historical resource data, and historical external environment state data are recorded by the Mars rover during action command execution. There are six types of domain knowledge known data that need to be obtained: mission name, state type, operator name, parameter type, action execution sequence, and action execution trajectory. The action execution trajectory is derived from action execution data, historical state data, historical resource data, and historical external environment state data. Based on the domain knowledge characteristics provided by the domain knowledge known data, it can be divided into instance-based domain knowledge known data and abstract domain knowledge known data. Instance-based domain knowledge known data describes the instantiated planning results, including action execution sequences and action execution trajectories, and serves as an important reference for domain knowledge learning.

[0038] Figure 4 This demonstrates the method for transforming action execution trajectories. Action execution trajectories can be derived from information in action execution data, historical state data, and historical state data of the external environment. (Action) The information comes from action execution data. When an action is successfully executed, the data from a single action execution data point is... and Converted into action execution trajectory External environmental conditions Actions are recorded based on historical data of the external environment. After receiving the information, from Obtain the action start time from the corresponding action execution data. Use it as historical data of the external environment Record all external environmental state data at the current time point as the external environmental state. Mars rover status Data derived from the historical status of the Mars rover is also used to record actions. After obtaining the information, get the action start time. Historical status data of Mars rovers At this point, a consistency check is performed to retrieve the previous action. End time And determine the time interval. Same inside Has the status changed? If the status remains unchanged, record it in the Mars rover status log. If an inconsistent state exists, it indicates that the consistency check has failed and the state information is incorrect. In this case, the state will not be recorded in the Mars rover's state. .

[0039] Figure 5 This demonstrates how parameter types are obtained. Parameter types describe the kinds and instances of all parameters involved in the task execution data, and can be represented as follows: ,in Indicates the parameter type. ( () represents a parameter instance of that type in the task execution data. The parameter type retrieval method reads information about the parameters from the task execution data, determines the parameter type, and stores it in a parameter set of the corresponding type. For example, it records all... The parameter of type can be used to obtain a subset of parameters of type location. .

[0040] In one embodiment, the task execution data can be processed first to obtain domain knowledge known data. Using this domain knowledge known data as input, a hierarchical task network structure learning model is used to obtain the decomposition method and its hierarchical task network structure. An operator and decomposition method learning model is used to obtain the preconditions of the decomposition method, the preconditions of the operators, and information on addition and deletion effects. Combining these domain knowledge elements with the domain knowledge known data yields a complete domain knowledge file. The acquisition of the domain knowledge known data can involve data format conversion operations, changing the format of the Mars rover's task execution data and organizing it into a new data format based on the information contained within the data. Abstract domain knowledge known data describes the specific learning scope of the domain knowledge, including the execution task name, state type, operator name, and parameter type, and is the information source for some elements of the domain knowledge. Since the Mars rover's task execution data is instantiated, the acquisition of abstract domain knowledge known data requires abstraction operations to summarize and generalize the information contained in the task execution data, extracting new data types from it.

[0041] Figure 6 This demonstrates the algorithmic framework for the hierarchical task network structure learning model. The algorithm's input is a set A of action execution sequences, and the elements in the set... , where n is the sequence of actions to be executed. The length of the action sequence may vary. It is a successfully executed action; a set of tasks NT, the elements in the set. , and action execution sequence Correspondingly, there is a set of task names NP and an operator name set O. The task execution data is one type of task execution data; the other data belongs to domain knowledge and is obtained by processing the task execution data. The algorithm output is a set of HTN decomposition methods M, where each decomposition method m includes its name. hierarchical task network The hierarchical task network learning algorithm mainly consists of two stages. In the task decomposition method structure learning stage, the algorithm uses the name of the task to be executed as the name of the composite task, and uses heuristic decomposition relation factors to construct the decomposition relationship between the composite task and the operator sequence, ultimately obtaining the composite task decomposition method structure. In the hierarchical task network learning stage, based on the composite task decomposition method structure, the algorithm uses a recursive learning method to obtain the complete hierarchical task network structure, which can then be transmitted to planners for adjustment.

[0042] Figure 7 The algorithm framework corresponding to the decomposition structure optimization module is demonstrated. The algorithm first randomly selects an execution task name np and its corresponding operator sequence ax. If a match is found, the success rate is... If the operator sequence is equal to 1, it indicates that the current operator sequence has a good matching degree. In this case, only lengthening the operator sequence is considered. The algorithm randomly selects an action execution sequence with a matching instance from the action execution sequence set, and adds or deletes operator sequences according to the action execution sequence. It then calculates the Q of all execution task names in the execution task name set after the change. If the Q of all execution task names does not decrease, the optimization is retained; otherwise, it is not retained, and the next round of optimization is performed. This process is repeated until an ideal execution task decomposition structure is obtained, where the Q of all execution task names is 1. Decomposition structure matching rate Represented as: in Execute the task name in the list of all executed tasks. Quantity, It is the sequence of operators obtained by decomposing the name of the currently executing task. It is an instance in the action execution sequence data that can match the task decomposition structure.

[0043] Figure 8This paper demonstrates the algorithmic framework for the hierarchical task network learning module. Using a bottom-up approach, it starts with the operator sequence obtained from the final decomposition of the task. The most frequently occurring operator subsequence is selected as a hierarchical task network for a decomposition method, and a new method name is automatically generated to represent this subsequence. The corresponding subsequences in the operator sequence are then replaced with these method names. After obtaining the operator sequence with the replaced method names, the probability matrix of the operator sequence for the task execution is obtained again. This same method is iterated repeatedly in the operator sequence until the hierarchical task network for decomposing the task execution method is finally obtained.

[0044] Figure 9 This paper presents the algorithmic framework for an operator and decomposition method learning model. Based on knowledge of Machine Learning Networks (MLN), a learning algorithm for operators and decomposition methods is proposed. The algorithm transforms action execution trajectories into a relational database of Markov logic networks (MLNs), and domain knowledge into logical formulas for MLNs. By learning the weights of these logical formulas, the algorithm obtains the degree of satisfaction of the formulas, thereby acquiring relevant information about operators and decomposition methods. The algorithm's input consists of a set of action execution trajectories T from a Mars rover mission, a finite set of state types ST, a set of operator names O, and a set of decomposition methods M (with known method names). hierarchical task network The algorithm's output is a prerequisite for the decomposition method. (This is followed by a list of parameters and their types, which is not translated here.) Preconditions of operators The addition status of operators and deletion status The algorithm comprises a domain knowledge construction module and a conflict detection and repair module. In the domain knowledge construction module, the algorithm encodes action execution trajectories into a relational database and domain knowledge information into logical formulas. By learning the weights of these logical formulas, the algorithm obtains the degree of satisfaction of the domain knowledge information, thus initially constructing the domain knowledge. Based on this constructed domain knowledge, the algorithm performs conflict detection, repairs the domain knowledge, and ensures its usability. Building upon this, and using historical resource data as input, a resource constraint acquisition module is designed based on the acquired domain knowledge elements, providing a method for acquiring resource constraints.

[0045] Figure 10 The algorithm flow for the conflict detection and repair module is demonstrated. First, the initially constructed domain knowledge is substituted into the existing action execution trajectory data to detect conflicts, obtaining their specific locations and data, which are stored in a conflict set. After detection, a conflict is randomly selected from the set, and a method to repair the conflict is searched.

[0046] There are several ways to resolve conflicts. For conflicts where operators simultaneously add and delete a state, one possible solution is to remove the conflicting state during the addition or deletion process. To resolve state conflicts, one possible solution is to remove the operator. Delete state in state Alternatively, select an operator for stage j from stage 1 to stage i. Add a state to its added state. The modified domain knowledge is re-applied and conflicts are re-detected. The number of conflicts at this point is used as an indicator to evaluate the repair methods. All conflict repair methods are compared, and the optimal one is selected for application. If the evaluation indicator data is the same, the conflict repair method with the higher logical formula weight is prioritized. After completing the conflict repair, other conflicts are selected, and the above steps are repeated iteratively until the number of conflicts is reduced to 0. At this point, the loop ends, and the current domain knowledge is used as the output, which is the domain knowledge file.

[0047] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Mars rover mission planning method for HTN, characterized in that, include: S1: Construct a domain knowledge learning model for the Mars rover, including a hierarchical task network structure learning model and an operator and decomposition method learning model; S2: Input the domain knowledge known data corresponding to the task execution data into the domain knowledge learning model, so that the hierarchical task network structure learning model can learn the domain knowledge known data and construct the decomposition relationship between the composite task and the operator sequence by decomposing relation factors. The hierarchical task network structure of the composite task is obtained by using the iterative optimization method. Then, the operator and decomposition method learning model is used to learn the domain knowledge known data, the decomposition method and the hierarchical task network structure and output the domain knowledge elements. Finally, the domain knowledge file obtained by fusing the HTN-oriented decomposition method and the domain knowledge elements is output. The domain knowledge elements include: the preconditions of the decomposition method, the preconditions of the operators, and the addition and deletion states of the operators. S3: Input the domain knowledge file into the HTN planner to obtain an action plan for task execution; The hierarchical task network structure learning model includes: The decomposition structure optimization module is used to input the known domain knowledge data, take the execution task name as the composite task name, construct the decomposition relationship between the composite task and the operator sequence using heuristic decomposition relation factors, and finally obtain the composite task decomposition method structure. The hierarchical task network learning module is used to obtain the hierarchical task network structure based on the structure of the composite task decomposition method using a recursive learning method. The decomposition structure optimization module is used to: randomly select an execution task name np and its corresponding operator sequence ax. If the matching success rate is 1, it indicates that the currently set operator sequence has a good matching degree; at this time, only the lengthening of the operator sequence is considered; randomly select an action execution sequence with a matching instance from the action execution sequence set, and add or delete operator sequences according to the action execution sequence, and calculate the decomposition structure matching rate Q of all execution task names in the execution task name set after the change. If the Q of all execution task names does not decrease, the optimization is retained; otherwise, it is not retained, and the next round of optimization is performed; repeat the above steps to finally obtain an ideal execution task decomposition structure, so that the Q of all execution task names is 1; The operator and decomposition method learning model includes: The domain knowledge construction module is used to encode the action execution trajectory into a relational database, encode domain knowledge information into logical formulas, and obtain the degree of satisfaction of domain knowledge information by learning the weights of logical formulas, thus initially constructing domain knowledge. The conflict detection and repair module is used to substitute the initially constructed domain knowledge into the original action execution trajectory data, detect conflicts, obtain the specific location and data of the conflicts, and store them in the conflict set; randomly select a conflict in the conflict set, find a way to repair the conflict and repair it, and then obtain the domain knowledge elements. The known domain knowledge data includes: the rover's action execution sequence, action execution trajectory, execution task name, operator name, state description, and parameter type; the sources of the rover's action execution trajectory include: action execution data, historical state data, historical resource data, and historical state data of the external environment.

2. The Mars rover mission planning method for HTN as described in claim 1, characterized in that, The hierarchical task network learning module is used for: Using a bottom-up approach, starting with the operator sequence obtained from the final decomposition of the task, the most frequently occurring operator subsequence is selected as a hierarchical task network for decomposition methods, and new method names are automatically generated to represent the corresponding subsequences. The corresponding subsequences in the operator sequence are then replaced with method names, and the operator sequence probability matrix for the task is obtained again. By iterating the same method repeatedly through the operator sequence, a hierarchical task network for executing task decomposition methods is finally obtained.

3. The Mars rover mission planning method for HTN as described in claim 1, characterized in that, The domain knowledge elements also include resource constraints; the operator and decomposition method learning model also includes: The resource constraint acquisition module is used to obtain corresponding resource constraints by taking the domain knowledge elements and historical resource data as input. The resource constraints include resource change time and resource change amount.

4. The Mars rover mission planning method for HTN as described in any one of claims 1-3, characterized in that, Before step S2, the method further includes: converting the format of the task execution data into a preset format, and then extracting features from it to obtain the known domain knowledge data.

5. A Mars rover mission planning device for HTN, characterized in that, The method for executing the HTN-oriented Mars rover mission planning method according to any one of claims 1-4 includes: The module constructs a domain knowledge learning model for the Mars rover, including a hierarchical task network structure learning model and an operator and decomposition method learning model. The acquisition module is used to input the domain knowledge known data corresponding to the task execution data into the domain knowledge learning model, so that the hierarchical task network structure learning model can learn the domain knowledge known data and construct the decomposition relationship between the composite task and the operator sequence by decomposing relation factors. The hierarchical task network structure of the composite task is obtained by using an iterative optimization method. Then, the operator and decomposition method learning model is used to learn the domain knowledge known data, the decomposition method and the hierarchical task network structure and output the domain knowledge elements. Finally, the domain knowledge file obtained by fusing the HTN-oriented decomposition method and the domain knowledge elements is output. The domain knowledge elements include: the preconditions of the decomposition method, the preconditions of the operators, and the addition and deletion states of the operators. The planning module is used to input the domain knowledge file into the HTN planner to obtain action plans for task execution.

6. A Mars rover mission planning system for HTN, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

Images