WO2025102453A1 - 基于语义网络和知识库控制机器人任务决策的系统 - Google Patents

基于语义网络和知识库控制机器人任务决策的系统 Download PDF

Info

Publication number
WO2025102453A1
WO2025102453A1 PCT/CN2023/136721 CN2023136721W WO2025102453A1 WO 2025102453 A1 WO2025102453 A1 WO 2025102453A1 CN 2023136721 W CN2023136721 W CN 2023136721W WO 2025102453 A1 WO2025102453 A1 WO 2025102453A1
Authority
WO
WIPO (PCT)
Prior art keywords
node
semantic
nodes
knowledge base
action
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2023/136721
Other languages
English (en)
French (fr)
Inventor
宋伟
周元海
朱世强
孙尧
任杰
穆宗昊
方伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Lab
Original Assignee
Zhejiang Lab
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Lab filed Critical Zhejiang Lab
Publication of WO2025102453A1 publication Critical patent/WO2025102453A1/zh
Anticipated expiration legal-status Critical
Pending legal-status Critical Current

Links

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
    • B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
    • B25J9/00—Program-controlled manipulators
    • B25J9/16—Program controls
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Definitions

  • the present invention belongs to the technical field of robots and artificial intelligence, and in particular to a system for controlling robot task decision-making based on semantic networks and knowledge bases.
  • the robot task planning output of the prefabricated model is a sequence of behaviors, which results in a large number of redundant actions and behaviors when building complex robot tasks, and these behaviors and tasks will not be fully covered during runtime.
  • the patent document with publication number CN114153943A discloses a method based on knowledge graph construction.
  • a system and method for building a robot behavior tree the system includes: a knowledge graph and a behavior tree, the knowledge graph is constructed according to an AOG format through a group of directed connected action nodes, the action nodes include an action category and a corresponding group of action features and action action objects, and the directed connection connects the action nodes according to the compared action features;
  • the behavior tree takes the subject and the object as the action action objects according to the task of the subject-predicate-object structure, classifies the predicate into a group of action features corresponding to the action category, constructs the corresponding behavior according to the action features, constructs the behavior node according to the behavior, and swaps the order of each behavior node of the behavior tree according to the directed connection of the sub-action node;
  • the method includes: S1, constructing an object knowledge base; S2, constructing an action knowledge base; S3, constructing a knowledge graph; S4, constructing a
  • the above technical solution constructs a behavior tree based on the knowledge graph, but many behaviors do not exist in the knowledge graph, which will cause the failure or inaccuracy of the behavior tree construction.
  • the purpose of the present invention is to provide a system for controlling robot task decision-making based on semantic network and knowledge base, combining semantic network, large language model and knowledge base to generate decision behavior tree for robot task.
  • the present invention provides a system for controlling robot task decision-making based on semantic network and knowledge base, including a semantic network module, a large language model module, a knowledge base module, a task decision module and a variable storage module;
  • the semantic network module is used to provide a semantic network for constructing a task decision behavior tree, and the semantic network is supplemented and expanded by the question-answering results and knowledge base of the large language model;
  • the large language model module is used to use the large language model to perform knowledge question answering and output question answering results
  • the knowledge base module is used to provide knowledge for constructing a task decision behavior tree
  • the task decision module is used to determine the task based on the semantic network module, the large language model module, the knowledge base
  • the module builds a task decision behavior tree. Specifically, it builds nodes of the decision tree by querying the semantic network, fills in node parameters through the question and answer results of the large language model, and fills in node parameters through the knowledge base.
  • variable storage module is used to store temporary variables when constructing a task decision behavior tree.
  • the semantic network is a directed connected graph structure, which sequentially connects associated semantic nodes, uses in-degree and directed connections to represent the weights and possibilities between semantic nodes, and each semantic node includes an identifier and a parameter list, wherein the directed connection relationship includes a relationship type and relationship attributes, and the relationship attributes include routing possibilities, execution order, and condition descriptions.
  • the semantic nodes include semantic nodes with single semantics, semantic nodes with attached decorative information, semantic nodes with attached conditions, and semantic nodes that allow targets to be sets, so as to adapt to the construction of task decision behavior trees.
  • the large language model pre-stores knowledge, which can perform question-and-answer search based on input questions and output question-and-answer results when expanding the semantic network and constructing the task decision behavior tree. At the same time, it can search based on the task decomposition problem and provide multiple possible results for task planning.
  • the question-and-answer results and possible results are connected to the semantic network as expanded knowledge, or added to the nodes of the task decision behavior tree as supplementary knowledge.
  • the knowledge base is refined knowledge, including an action knowledge base and an object knowledge base, wherein the action knowledge base includes actions, action types to which actions belong, action features, and action objects; the object knowledge base includes objects and object attributes.
  • the step of constructing a task decision behavior tree includes:
  • an object instance corresponding to the target object is searched in the object knowledge base as the target object instance of the action;
  • search the semantic node set corresponding to the behavior set in the semantic network then build a task decision behavior tree according to the semantic node set, and add node parameters to the action nodes in the task decision behavior tree, where the node parameters come from at least one of the target object instance searched from the object knowledge base, the dynamically perceived scene instance, and the question and answer result obtained through multiple rounds of dialogue based on the large language model.
  • searching for a semantic node set corresponding to the behavior set in the semantic network, and then constructing a task decision behavior tree according to the semantic node set includes:
  • the semantic node set corresponding to the behavior includes a single semantic node, a semantic node with decorative information, a semantic node with conditions, and a semantic node with a set as the target;
  • the action nodes, sequence node selection nodes, condition nodes, and decoration nodes generated according to the topological structure corresponding to the semantic node set constitute the task decision behavior tree.
  • the rule knowledge is queried, a selection node is constructed based on the rule knowledge, and a condition node is generated under the selection node based on the rule knowledge.
  • the execution order of each sub-behavior in the task decision behavior tree is swapped according to the execution order of the sub-action nodes.
  • a task decision behavior tree verification process is also included, and the verification conditions are:
  • the sub-behaviors under each behavior must correspond to all sequential nodes or selection nodes.
  • Parallel nodes can coexist with sequential nodes.
  • Selection nodes must contain conditional nodes and the conditional node judgment Boolean parameters contained in the conditional nodes.
  • Any behavior must contain an action identification parameter to complete the action.
  • the present invention has the following beneficial effects:
  • the present invention introduces the design thinking of complex semantic networks, so that there is a dependency relationship between action knowledge, and the robot's behavior can be constructed only by relying on semantic networks; the task decision behavior tree is finally generated.
  • behavior trees are superior to behavior sequences in terms of control logic structure. Behavior trees can represent both the order and the hierarchy between action behaviors, and the same behavior plan will not appear repeatedly.
  • FIG. 1 is a schematic diagram of the structure of a system for controlling robot task decision-making based on a semantic network and a knowledge base provided in an embodiment
  • FIG2 is a schematic diagram of generating an action node from a terminal semantic node provided by an embodiment
  • FIG3 is a schematic diagram of a semantic node generating a behavior tree for implementing the provided terminal attached condition
  • FIG4 is a schematic diagram of a behavior tree generated by a semantic node with decoration information provided by an embodiment
  • FIG5 is a schematic diagram of generating a behavior tree for semantic nodes of a target set provided by an embodiment
  • FIG6 is an overall flow chart of generating a task decision behavior tree provided by an embodiment.
  • the system 100 for controlling robot task decision-making based on semantic network and knowledge base includes a semantic network module 110 , a large language model module 120 , a knowledge base module 130 , a task decision module and a variable storage module 140 .
  • the semantic network model 110 is used to provide a semantic network for constructing a task decision behavior tree, and the semantic network is supplemented and expanded by the question-answering results and knowledge base of the large language model.
  • the semantic network corresponds to the decomposition structure of the task, and the action node M is established with the minimum controllable action of the robot.
  • a task is decomposed into multiple required subtasks, some of which have conditional judgments and selection values.
  • the semantic network is a directed connected graph structure, which sequentially connects and associates semantic nodes.
  • the in-and-out degrees and directed connection relationships are used to represent the weights and possibilities between semantic nodes.
  • each semantic node is an action node, it includes the action unique identifier and the parameter list vars of its target object.
  • the directed connection relationship includes the relationship type and relationship attributes.
  • the relationship attributes include the routing possibility prop, the execution order seq, and the condition description condition. For the relationship attribute of the routing possibility prop, if the subordinate child nodes of a semantic node lack the execution order parameter, then prop will be used as the sorting basis.
  • the relationship type is the association attribute between semantic nodes, including subordination (is), action (action), state transfer (transform), spatial relationship (space), etc. Regardless of the type of relationship connection, it follows the constraints of the execution order seq from small to large.
  • the condition description condition indicates that there are conditions for the decomposition relationship. It is necessary to establish decoration nodes and conditional nodes on the corresponding decision behavior tree to solve this problem.
  • Directed links include two different states: feasibility connection and necessity connection.
  • the routing possibility prop indicates the selection strength of the action node.
  • Semantic nodes include semantic nodes with single semantics, semantic nodes with decorative information, semantic nodes with conditions, and semantic nodes that allow targets to be sets, in order to adapt to the construction of task decision behavior trees.
  • Single semantics contains various action semantics.
  • the large language model module 120 is used to use the large language model to perform knowledge question-answering and output question-answering results.
  • the large language model pre-stores knowledge, and can perform question-answering search based on input questions and output question-answering results when expanding the semantic network and constructing the task decision behavior tree. At the same time, it can search based on the task decomposition problem and give multiple possible results of task planning.
  • the question-answering results and multiple possible results are connected to the semantic network as expanded knowledge, or added to the nodes of the task decision behavior tree as supplementary knowledge.
  • the knowledge base module 130 is used to provide knowledge for building a task decision behavior tree. It is refined knowledge used to store knowledge of real-time dynamic scenes, including an action knowledge base and an object knowledge base.
  • the object knowledge base includes objects and object attributes in the scene.
  • the object includes character information ⁇ People ⁇ and environmental object information ⁇ Env ⁇ , and the attribute information contained in both includes character name, character title, character coordinates, character features, object name, object coordinates, object features, historical trajectory and social posture.
  • the motion knowledge base includes actions, the motion type motion to which the action belongs, motion features, and the objects acted upon by the action.
  • Motion features refer to the robot hardware components required to perform the action and the necessary parameter table for performing the action.
  • a type identifier is constructed for each action category ID, a semantic association is constructed for the action feature, and a target object is associated with the action object.
  • the target object is an object instance created in the computer memory according to its data format in the object knowledge base, has a unique object symbol, and uses a unique action ID to act on the target object and the object parameter list vars.
  • the object The parameter list vars includes location, target object, target person, language content, grasping space posture information, interaction target, preprocessing function, etc., as shown in Table 1.
  • the task decision module 140 is used to construct a task decision behavior tree based on the semantic network module, the large language model module, and the knowledge base module. Specifically, the nodes of the decision tree are constructed by querying the semantic network, the node parameters are filled in through the question and answer results of the large language model, and the node parameters are filled in through the knowledge base.
  • conditional rules For a decision behavior tree, the calculation of conditional rules is related to the robot's perception information. It is calculated by real-time conditional rules. When it is determined that the current conditional rule can be executed, a multi-way selection branch behavior is constructed to obtain the real-time conditional judgment result of Boolean type through the synchronous condition-rule-execution feature.
  • targets for the decision behavior tree the target object is queried according to the subject corresponding to the behavior, and the location information, action category motion, description information, language information, job type, and target list are injected into the object parameter list of the action node.
  • the action category motion is injected into the behavior as content, and the task type and a random number are used as the unique identification ID of the decision behavior tree and cached in the database. For actions where the parameter object target cannot be determined at all, it is determined that the behavior tree construction has failed.
  • constructing a task decision behavior tree includes: determining the predicate and target object corresponding to the task according to the task understanding command and the task classification result; searching the object instance corresponding to the target object in the object knowledge base according to the target object as the target object instance for the action; classifying the predicate into the action category in the action knowledge base, and constructing a behavior set according to a set of task semantics corresponding to the action category; searching the semantic node set corresponding to the behavior set in the semantic network, and then constructing the task decision behavior tree according to the semantic node set, and adding node parameters to the action nodes in the task decision behavior tree, wherein the node parameters come from at least one of the target object instance searched from the object knowledge base, the dynamically perceived scene instance, and the question and answer result obtained through multiple rounds of dialogue based on the large language model.
  • a semantic node set corresponding to a behavior set is searched in the semantic network, and the semantic node set represents the formation of an action node and a sub-behavior set, wherein the sub-behavior is composed of an action node.
  • the action is split into action nodes of different behaviors according to whether the relationship type of the semantic node has attached conditions, and a selection node is added above the action node.
  • a decoration node is constructed and a decoration attribute is added.
  • a selection node is added to the decoration node, and the conditional nodes are placed below the selection node in sequence.
  • the sequence nodes with conditions are also expanded according to this mode.
  • the execution order of each sub-behavior in the task decision behavior tree is swapped according to the execution order of the sub-action node.
  • the node type is out of range, or an empty node appears when generating sequence nodes and selection nodes, it is recorded as a failure in the behavior tree construction process.
  • the process of constructing a behavior tree based on the semantic network is as follows: find robot data in the robot database according to the task trigger data, find the object entity as the action object in the object or dynamic scene library, and fail to build when the object entity does not exist in the scene; the task is classified into a specific action by the classifier, and the action corresponds to the label list features of the action features in the action knowledge base, that is, the action features in the action node M; fill the parameter list vars of the action object, and the specific parameters are obtained from the dest information in the acquired scene information and additional supplementary information, and the language content is obtained from the corpus of the corresponding target in the knowledge base; construct the corresponding behavior according to the action features, and choose to use movement, behavior or language dialogue according to the behavior features.
  • the completed behavior is recorded as an action node of the behavior tree.
  • the subtask is taken as the core in the semantic network to find the subgraph corresponding to the subtask.
  • the subgraph is represented as all the next-order nodes containing the action node.
  • the edges in the semantic network description have conditional information, decoration information, and prop variables, they are corresponded to three types of behavior tree logic nodes, namely selection nodes, sequence nodes, and parallel nodes (when the semantic network relationship has conditional information, it is a selection node, and in other cases it is a sequence node.
  • the nodes with the same sequence number in the sequence node are summarized as parallel nodes).
  • a sequence node is constructed.
  • the first node under the sequence node is recorded as the corresponding conditional node, followed by the corresponding action node. Nodes with the same sequence are planned to be parallel nodes.
  • a selection behavior node is constructed. When constructing, the conditional judgment cond(env) is performed to obtain the real-time conditional judgment result of Boolean type.
  • a sequential behavior node is constructed. When constructing, it is sorted. According to the seq sorting method, the order of each child node of the behavior tree is swapped from large to small.
  • the subject When injecting parameters into the decision behavior tree, the following method is used for any node: the subject always uses the subject s in the task understanding, and performs environment and character target queries based on the subject s to obtain the location information, the id of the target object (i.e., the action category motion), and the description information to inject into the target parameter list of the action node; the action id is injected into the behavior as the content, and the predicate v in the task understanding information is used as the name of the behavior tree.
  • the knowledge base For behaviors whose parameters cannot be determined, the knowledge base is searched to find actions with similar features for processing; for actions whose object parameters cannot be determined at all, the behavior cannot be constructed, and the entire behavior tree construction is judged to have failed.
  • a task decision behavior tree verification process is also performed, and the verification conditions are:
  • the sub-behaviors under each behavior must correspond to all sequential nodes or selection nodes.
  • Parallel nodes can coexist with sequential nodes.
  • Selection nodes must contain conditional nodes and the conditional node judgment Boolean parameters contained in the conditional nodes.
  • Any behavior must contain an action identification parameter to complete the action.
  • the variable storage module 150 is used to store temporary variables when constructing the task decision behavior tree.
  • the temporary variables are the scene information read during the dynamic perception process, the question-answering results of multiple rounds of dialogue solved by the large language model, and the knowledge read from the knowledge base.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Mechanical Engineering (AREA)
  • Robotics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于语义网络和知识库控制机器人任务决策的系统，包括语义网络模块用于提供构建任务决策行为树的语义网络，该语义网络通过大语言模型的问答结果和知识库进行补充拓展；大语言模型模块用于利用大语言模型进行知识问答输出问答结果；知识库模块用于提供构建任务决策行为树的知识；任务决策模块用于依据语义网络模块、大语言模型模块、知识库模块构建任务决策行为树，具体通过查询语义网络构建决策树的节点，通过大语言模型的问答结果进行节点参数的填充，通过知识库进行节点参数的填充；变量存储模块用于存储在构建任务决策行为树时的临时变量。基于该系统构建决策行为树更加准确。

Description

基于语义网络和知识库控制机器人任务决策的系统 技术领域
本发明属于机器人及人工智能技术领域，尤其是涉及一种基于语义网络和知识库控制机器人任务决策的系统。
背景技术
现有技术中，机器人的控制更多依赖于固定程序化的脚本、知识符号或者新出的大模型，从技术应用层面来说程序化脚本和知识符号无法自适应环境改变，大语言模型能够处理动态场景下的机器人任务控制，但是无法快速高效的进行决策，且训练成本巨大。
传统知识库只存放物体、人物、常识、动作等知识，然后使用预制的机器人规划方法进行规划，在此过程中的知识只能形成各个孤立的节点，并且参数属性需要在机器人执行任务的时候才能同步并获得注入。再者，知识库的规模限制了机器人的自主决策控制能力和多样性，使得机器人的作业行为过于程序化和僵化。
与此同时，通过预制模型的机器人任务规划输出为一个个的行为序列，这样构建复杂机器人任务时出现大量冗余动作和行为，这些行为和任务在运行时并不会被完全覆盖到。
对此，迫切需要在知识库、图谱计算、大语言模型的基础上进行改进和创新，依托于机器人作业的大数据语义网络基础，设计一种新的方案，将语义网络应用到机器人行为决策上来。
公开号为的CN114153943A的专利文献公开了一种基于知识图谱构 建机器人行为树的系统及方法，系统包括：知识图谱及行为树，知识图谱通过一组有向连接的动作节点，根据AOG格式构建，动作节点包括动作类别及其对应的一组动作特征、动作作用对象，有向连接根据比对动作特征连接动作节点；行为树根据主谓宾结构的任务，将主语和宾语作为动作作用对象，将谓语分类到动作类别对应的一组动作特征，根据动作特征构建对应的行为，根据行为构建行为节点，并根据子动作节点有向连接调换行为树各行为节点的顺序；方法包括：S1，构建对象知识库；S2，构建动作知识库；S3，构建知识图谱；S4，构建行为树。
上述技术方案基于知识图谱构建行为树，但是很多行为在知识图谱中是不存在的，会造成行为树构建的失败或者不准确。
发明内容
鉴于上述，本发明的目的是提供一种基于语义网络和知识库控制机器人任务决策的系统，结合语义网络、大语言模型以及知识库为机器人任务生成决策行为树。
为实现上述发明目的，本发明提供的一种基于语义网络和知识库控制机器人任务决策的系统，包括语义网络模块、大语言模型模块、知识库模块、任务决策模块以及变量存储模块；
所述语义网络模块用于提供构建任务决策行为树的语义网络，该语义网络通过大语言模型的问答结果和知识库进行补充拓展；
所述大语言模型模块用于利用大语言模型进行知识问答输出问答结果；
所述知识库模块用于提供构建任务决策行为树的知识；
所述任务决策模块用于依据语义网络模块、大语言模型模块、知识库 模块构建任务决策行为树，具体通过查询语义网络构建决策树的节点，通过大语言模型的问答结果进行节点参数的填充，通过知识库进行节点参数的填充；
所述变量存储模块用于存储在构建任务决策行为树时的临时变量。
优选地，所述语义网络为有向连接的图结构，顺序连接关联语义节点，使用出入度和有向连接表示语义节点之间的权重和可能性，每个语义节点包括标志符和参数列表，其中，有向连接关系包括关系类型和关系属性，关系属性包括路由可能性、执行顺序、以及条件描述，语义节点包括单个语义的语义节点、附带装饰信息的语义节点、附带条件的语义节点以及允许目标为集合的语义节点，以适配任务决策行为树的构建。
优选地，所述大语言模型预存知识，能够在拓展语义网络和任务决策行为树构建时，基于输入问题进行问答搜索，输出问答结果，同时基于任务分解问题进行搜索，给出任务规划的多种可能性结果，问答结果和可能性结果作为拓展知识连接到语义网络上，或者作为补充知识补充到任务决策行为树的节点中。
优选地，所述知识库为精细化知识，包括动作知识库和对象知识库，其中，动作知识库包括动作、动作所属动作类型、动作特征、以及动作作用对象；对象知识库包括对象、对象属性。
优选地，所述构建任务决策行为树包括：
根据任务理解命令和任务分类结果，确定任务对应的谓语以及目标对象；
依据目标对象在对象知识库中搜索目标对象对应的对象实例作为动作作用的目标对象实例；
将谓语分类到动作知识库中动作类别，并依据动作类别对应的一组任 务语义构建行为集合；
在语义网络中搜索行为集合对应的语义节点集合，然后按照语义节点集合中构建任务决策行为树，并为任务决策行为树中的动作节点添加节点参数，其中节点参数来自于从对象知识库中搜索的目标对象实例、动态感知的场景实例以及基于大语言模型经过多轮对话得到的问答结果中的至少一个。
优选地，所述在语义网络中搜索行为集合对应的语义节点集合，然后按照语义节点集合构建任务决策行为树，包括：
行为对应的语义节点集合包括单个语义节点、附带装饰信息的语义节点、附带条件的语义节点以及目标为集合的语义节点；
依据语义节点集合对应的拓扑结构构建决策行为树时，当语义节点为单个语义节点时，则对应生成动作节点并顺序执行；当语义节点为附带条件的语义节点时，为附带条件的语义节点生成分支，具体生成一个选择节点，并在选择节点下连接一个条件节点和分支的动作节点；当语义节点为当附带装饰信息的语义节点时，为附带装饰信息的语义节点生成装饰节点，并在装饰节点下生成顺序节点和动作节点；当语义节点为允许目标为集合的语义节点时，为目标为集合的语义节点生成多个装饰节点，针对每个装饰节点按照上述方式进行展开生成；
依据语义节点集合对应的拓扑结构生成的动作节点、顺序节点选择节点、条件节点、装饰节点构成任务决策行为树。
优选地，具体构建时，当语义节点附带条件时，查询规则知识，依据规则知识构建选择节点，并在选择节点下依据规则知识生成条件节点。
优选地，在构建任务决策行为树时，根据子动作节点的执行顺序调换任务决策行为树中各子行为的执行顺序。
优选地，在构建任务决策行为树时，还包括任务决策行为树校验过程，校验条件为：
每个行为下的子行为必须对应全是顺序节点或者选择节点，并行节点能够与顺序节点共存，选择节点必须包含条件节点及条件节点包含的条件节点判断布尔参数，任何行为必须包含完成动作的动作标识参数。
与现有技术相比，本发明具有的有益效果至少包括：
本发明在动作知识的基础上，将复杂语义网络的设计思维引入，使得动作知识间存在依存关系，仅仅依靠语义网络就可以构建出机器人的行为；最终生成的任务决策行为树。相较于语言、脚本控制序列，行为树在控制逻辑结构上优于行为序列，行为树既能够表示顺序又能够表示动作行为间的层次，而且相同的行为规划不会重复出现。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图做简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动前提下，还可以根据这些附图获得其他附图。
图1是实施例提供的基于语义网络和知识库控制机器人任务决策的系统的结构示意图；
图2是实施例提供的末端语义节点生成动作节点示意图；
图3是实施提供的末端附带条件的语义节点生成行为树的示意图；
图4是实施例提供的附带装饰信息的语义节点生成行为树的示意图；
图5是实施例提供的目标为集合的语义节点生成行为树的示意图；
图6是实施例提供的生成任务决策行为树的总体流程图。
具体实施方式
为使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例对本发明进行进一步的详细说明。应当理解，此处所描述的具体实施方式仅仅用以解释本发明，并不限定本发明的保护范围。
如图1所示，实施例提供的基于语义网络和知识库控制机器人任务决策的系统100，包括语义网络模块110、大语言模型模块120、知识库模块130、任务决策模块以及变量存储模块140。
语义网络模型110用于提供构建任务决策行为树的语义网络，该语义网络通过大语言模型的问答结果和知识库进行补充拓展。其中，语义网络对应任务的分解结构，以机器人具有的可被控制的最小动作建立动作节点M，使用任务分解的方法，将一个任务分解为多个需要的子任务，其中一些子任务存在条件判定和选择值。
语义网络为有向连接的图结构，顺序连接关联语义节点，使用出入度和有向连接关系表示语义节点之间的权重和可能性，每个语义节点当为动作节点时，包括动作唯一标志符及其作用目标对象的参数列表vars，其中，有向连接关系包括关系类型和关系属性，关系属性包括路由可能性prop、执行顺序seq、以及条件描述condition，针对路由可能性prop这一关系属性，如果一个语义节点的下属子节点缺乏执行顺序参数，那么prop将作为排序依据。当作为条件关系时，prop将决定该节点被放置到行为树中的可能性，是一个随机数落入一定范围域的方法。其中，关系类型是语义节点之间的关联属性，有从属(is)、动作(action)、状态转移(transform)、空间关系(space)等。无论是何种关系的连接，都遵循按照执行顺序seq从小到大进行约束，条件描述condition表示分解关系存在条件，需要在对应决策行为树上建立装饰节点、条件节点用来解决这个问题。
有向链接包括可行性连接和必然性连接两种不同的状态，当出现可能性连接状态时，路由可能性prop表示动作节点的选择强度。语义节点包括单个语义的语义节点、附带装饰信息的语义节点、附带条件的语义节点以及允许目标为集合的语义节点，以适配任务决策行为树的构建。单个语义包含各种动作语义。
大语言模型模块120用于利用大语言模型进行知识问答输出问答结果。其中，大语言模型预存知识，能够在拓展语义网络和任务决策行为树构建时，基于输入问题进行问答搜索，输出问答结果，同时基于任务分解问题进行搜索，给出任务规划的多种可能性结果，问答结果和多种可能性结果作为拓展知识连接到语义网络上，或者作为补充知识补充到任务决策行为树的节点中。
知识库模块130用于提供构建任务决策行为树的知识，其为精细化知识，用于存放实时动态场景的知识，包括动作知识库和对象知识库。其中，对象知识库包括场景中的对象及对象属性。该对象包括人物信息{People}、环境物体信息{Env}，两者均包含的属性信息包括人物名称、人物称呼、人物坐标、人物特征、物体名称、物体坐标、物体特征、历史轨迹和社交姿态。在构建决策行为树时，根据机器人信息和任务分类结果，在对象知识库中检索对应的目标对象作为动作作用对象参数vars填充的依据。
动作知识库包括动作、动作所属动作类型motion、动作特征、以及动作作用对象。动作特征是指执行动作时需要的机器人硬件组件、执行动作的必要参数表。为每个动作类别ID构建类型标识，为动作特征构建语义关联，为动作作用对象关联目标对象，目标对象是在对象知识库中根据其数据格式在计算机存储器中创建的一个对象实例，具有唯一的对象标志符号，使用唯一的动作ID作用目标对象和对象参数列表vars。其中，对象 参数列表vars包括地点、目标物体、目标人物、语言内容、抓取空间位姿信息、交互目标、预处理功能等，具体如表1所示。
表1
任务决策模块140用于依据语义网络模块、大语言模型模块、知识库模块构建任务决策行为树，具体通过查询语义网络构建决策树的节点，通过大语言模型的问答结果进行节点参数的填充，通过知识库进行节点参数的填充。
具体地，结合任务并使用语义网络，通过选择的条件规则构建决策行 为树，条件规则计算与机器人感知信息相关，由实时条件规则计算，判决当前条件规则可以执行时，构建多路选择分支行为，获得通过同步条件-规则-执行特征获取布尔类型的实时条件判断结果。对决策行为树进行目标生成时，针对行为根据行为对应的主语进行目标对象查询，获取位置信息、动作类别motion、描述信息、语言信息、作业类型、目标列表注入动作节点的对象参数列表中，动作类别motion作为内容注入行为中，将任务类型和一个随机数作为决策行为树的唯一标志ID，缓存在数据库中。对于完全无法确定参数对象target的动作，判定行为树构建失败。
更具体地，如图6所示，构建任务决策行为树包括：根据任务理解命令和任务分类结果，确定任务对应的谓语以及目标对象；依据目标对象在对象知识库中搜索目标对象对应的对象实例作为动作作用的目标对象实例；将谓语分类到动作知识库中动作类别，并依据动作类别对应的一组任务语义构建行为集合；在语义网络中搜索行为集合对应的语义节点集合，然后按照语义节点集合中构建任务决策行为树，并为任务决策行为树中的动作节点添加节点参数，其中节点参数来自于从对象知识库中搜索的目标对象实例、动态感知的场景实例以及基于大语言模型经过多轮对话得到的问答结果中的至少一个。
依据语义节点集合对应的拓扑结构构建决策行为树时，如图2所示，当语义节点为单个语义节点时，则对应生成动作节点并顺序执行；如图3所示，当语义节点为附带条件的语义节点时，为附带条件的语义节点生成分支，具体生成一个选择节点，并在选择节点下连接一个条件节点和分支的动作节点；如图4所示，当语义节点为当附带装饰信息的语义节点时，为附带装饰信息的语义节点生成装饰节点，并在装饰节点下生成顺序节点和动作节点；如图5所示，当语义节点为允许目标为集合的语义节点时， 为目标为集合的语义节点生成多个装饰节点，针对每个装饰节点按照上述方式进行展开生成；依据语义节点集合对应的拓扑结构生成的动作节点、顺序节点选择节点、条件节点、装饰节点构成任务决策行为树。
具体地，在语义网络中寻找与行为集合对应的语义节点集合，该语义节点集合表示形成动作节点和子行为集合，其中子行为由动作节点组成。针对语义节点集合，依据语义节点的关系类型是否存在附带条件，将动作拆分不同行为的动作节点，并在动作节点上方加入选择节点，当一个任务存在装饰信息时，构建装饰节点并添加装饰属性，装饰节点中加入一个选择节点，依次将条件节点放置在选择节点下方，对于具有条件的顺序节点也按照这个模式进行展开，无论是顺序节点还是条件节点，都必须要按照这个模式进行处理，直至一个子行为树构建完成，构建时，根据子动作节点的执行顺序调换任务决策行为树中各子行为的执行顺序。在构建机器人行为树的过程中，出现获取场景中特征数据失败、节点类型不在范围内情况，或者再产生顺序节点和选择节点时，出现空节点，那么记作行为树构建处理失败。
实例中，依据语义网络构建行为树的过程为：根据任务触发数据在对机器人数据库中找到机器人数据，在对象或者动态场景库中找到对象实体作为动作作用对象，当在场景中不存在对象实体时构建失败；任务经过分类器分类到一个具体动作，该动作与动作知识库中的动作特征的标签列表features对应，即动作节点M中的动作特征；对动作作用对象的参数列表vars进行填充具体参数从获取到的场景信息和额外补充信息中的dest信息中获取，语言内容从知识库中对应目标的语料库中获得；依照动作特征构建对应行为，根据行为特征选择使用运动、行径还是语言对话，完成后的行为记作行为树的一个动作节点。
对于已经构建的动作节点，在语义网络中以子任务为核心，找到子任务对应的子图，该子图表示为包含该动作节点的所有下一阶节点，根据语义网络描述中边是否存在条件信息、装饰信息、prop变量，将其对应成为三种行为树逻辑节点，分别是选择节点、顺序节点、并行节点(语义网路关系具有条件信息时为选择节点，其他情况为顺序节点。对于顺序节点中存在的相同序号节点则一并归纳为并行节点)，对于已经是叶子节点的情形，则构建一个顺序节点，顺序节点下的第一个节点记为对应的条件节点，之后为对应的动作节点，将顺序相同的节点拟定为平行节点，对于所有的{Behavior}sel构建选择行为节点，构建时进行条件判断cond(env)，获得布尔类型的实时条件判断结果，对于所有的{Behavior}seq构建顺序行为节点，构建时进行排序，按照seq排序方法由大到小调换行为树各个子节点的顺序。
在决策行为树进行参数注入时，针对任意一个节点，使用如下方法进行参数注入：主体始终使用任务理解中的主语s，根据主语s进行环境、人物目标查询，获取位置信息、目标对象的id(即动作类别motion)、描述信息注入动作节点的目标参数列表中；动作id作为内容注入行为中，使用任务理解信息中的谓语v作为行为树名称，对于无法确定参数的行为，进行知识库检索，寻找feature相似的动作进行处理；对于完全无法确定对象参数的动作，将导致无法构建行为，此时判定整个行为树构建失败。
实施例中，还进行任务决策行为树校验过程，校验条件为：
每个行为下的子行为必须对应全是顺序节点或者选择节点，并行节点能够与顺序节点共存，选择节点必须包含条件节点及条件节点包含的条件节点判断布尔参数，任何行为必须包含完成动作的动作标识参数。
变量存储模块150用于存储在构建任务决策行为树时的临时变量。其 中临时变量为动态感知过程中读取的场景信息、大语言模型进行多轮对话求解的问答结果以及从知识库中读取的知识。
以上所述的具体实施方式对本发明的技术方案和有益效果进行了详细说明，应理解的是以上所述仅为本发明的最优选实施例，并不用于限制本发明，凡在本发明的原则范围内所做的任何修改、补充和等同替换等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种基于语义网络和知识库控制机器人任务决策的系统，其特征在于，包括语义网络模块、大语言模型模块、知识库模块、任务决策模块以及变量存储模块；
    所述语义网络模块用于提供构建任务决策行为树的语义网络，该语义网络通过大语言模型的问答结果和知识库进行补充拓展；
    所述大语言模型模块用于利用大语言模型进行知识问答输出问答结果；
    所述知识库模块用于提供构建任务决策行为树的知识；
    所述任务决策模块用于依据语义网络模块、大语言模型模块、知识库模块构建任务决策行为树，具体通过查询语义网络构建决策树的节点，通过大语言模型的问答结果进行节点参数的填充，通过知识库进行节点参数的填充；
    所述变量存储模块用于存储在构建任务决策行为树时的临时变量。
  2. 根据权利要求1所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，所述语义网络为有向连接的图结构，顺序连接关联语义节点，使用出入度和有向连接表示语义节点之间的权重和可能性，每个语义节点包括标志符和参数列表，其中，有向连接关系包括关系类型和关系属性，关系属性包括路由可能性、执行顺序、以及条件描述，语义节点包括单个语义的语义节点、附带装饰信息的语义节点、附带条件的语义节点以及允许目标为集合的语义节点，以适配任务决策行为树的构建。
  3. 根据权利要求2所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，有向链接包括可行性连接和必然性连接两种不同 的状态，当出现可能性连接状态时，路由可能性prop表示动作节点的选择强度。
  4. 根据权利要求1所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，所述大语言模型预存知识，能够在拓展语义网络和任务决策行为树构建时，基于输入问题进行问答搜索，输出问答结果，同时基于任务分解问题进行搜索，给出任务规划的多种可能性结果，问答结果和多种可能性结果作为拓展知识连接到语义网络上，或者作为补充知识补充到任务决策行为树的节点中。
  5. 根据权利要求1所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，所述知识库为精细化知识，包括动作知识库和对象知识库，其中，动作知识库包括动作、动作所属动作类型、动作特征、以及动作作用对象；对象知识库包括对象、对象属性。
  6. 根据权利要求1所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，所述构建任务决策行为树包括：
    根据任务理解命令和任务分类结果，确定任务对应的谓语以及目标对象；
    依据目标对象在对象知识库中搜索目标对象对应的对象实例作为动作作用的目标对象实例；
    将谓语分类到动作知识库中动作类别，并依据动作类别对应的一组任务语义构建行为集合；
    在语义网络中搜索行为集合对应的语义节点集合，然后按照语义节点集合中构建任务决策行为树，并为任务决策行为树中的动作节点添加节点参数，其中节点参数来自于从对象知识库中搜索的目标对象实例、动态感知的场景实例以及基于大语言模型经过多轮对话得到的问答结果中的至 少一个。
  7. 根据权利要求6所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，所述在语义网络中搜索行为集合对应的语义节点集合，然后按照语义节点集合构建任务决策行为树，包括：
    行为对应的语义节点集合包括单个语义节点、附带装饰信息的语义节点、附带条件的语义节点以及目标为集合的语义节点；
    依据语义节点集合对应的拓扑结构构建决策行为树时，当语义节点为单个语义节点时，则对应生成动作节点并顺序执行；当语义节点为附带条件的语义节点时，为附带条件的语义节点生成分支，具体生成一个选择节点，并在选择节点下连接一个条件节点和分支的动作节点；当语义节点为当附带装饰信息的语义节点时，为附带装饰信息的语义节点生成装饰节点，并在装饰节点下生成顺序节点和动作节点；当语义节点为允许目标为集合的语义节点时，为目标为集合的语义节点生成多个装饰节点，针对每个装饰节点按照上述方式进行展开生成；
    依据语义节点集合对应的拓扑结构生成的动作节点、顺序节点选择节点、条件节点、装饰节点构成任务决策行为树。
  8. 根据权利要求7所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，具体构建时，当语义节点附带条件时，查询规则知识，依据规则知识构建选择节点，并在选择节点下依据规则知识生成条件节点。
  9. 根据权利要求7所述的基于语义网络和知识库控制机器人任务决策的系统，其特征在于，在构建任务决策行为树时，根据子动作节点的执行顺序调换任务决策行为树中各子行为的执行顺序。
  10. 根据权利要求7所述的基于语义网络和知识库控制机器人任务决 策的系统，其特征在于，还包括任务决策行为树校验过程，校验条件为：
    每个行为下的子行为必须对应全是顺序节点或者选择节点，并行节点能够与顺序节点共存，选择节点必须包含条件节点及条件节点包含的条件节点判断布尔参数，任何行为必须包含完成动作的动作标识参数。
PCT/CN2023/136721 2023-11-13 2023-12-06 基于语义网络和知识库控制机器人任务决策的系统 Pending WO2025102453A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202311505193.1A CN117556894A (zh) 2023-11-13 2023-11-13 基于语义网络和知识库控制机器人任务决策的系统
CN202311505193.1 2023-11-13

Publications (1)

Publication Number Publication Date
WO2025102453A1 true WO2025102453A1 (zh) 2025-05-22

Family

ID=89822628

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/136721 Pending WO2025102453A1 (zh) 2023-11-13 2023-12-06 基于语义网络和知识库控制机器人任务决策的系统

Country Status (2)

Country Link
CN (1) CN117556894A (zh)
WO (1) WO2025102453A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120258747A (zh) * 2025-06-06 2025-07-04 北京致远互联软件股份有限公司 基于机器人流程自动化的人机协同方法及系统
CN120307305A (zh) * 2025-06-17 2025-07-15 北京理工大学长三角研究院(嘉兴) 一种装配机械臂任务规划方法及装置
CN120370756A (zh) * 2025-06-25 2025-07-25 深圳墨影科技有限公司 基于大模型的机器人低代码可编辑任务序列生成系统
CN120363218A (zh) * 2025-06-26 2025-07-25 中亿丰数字科技集团股份有限公司 一种建筑机器人任务规划与技能学习方法及系统
CN121722088A (zh) * 2026-02-25 2026-03-24 北京国联视讯信息技术股份有限公司 基于AI Agent的人机协作控制方法及系统

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119904865B (zh) * 2024-12-25 2025-09-30 西安电子科技大学广州研究院 基于场景模糊感知与推理的问答方法、装置、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190392820A1 (en) * 2019-08-05 2019-12-26 Lg Electronics Inc. Artificial intelligence server for setting language of robot and method for the same
US20200298401A1 (en) * 2019-01-03 2020-09-24 Lucomm technologies, Inc Semantic Robotic System
CN114153943A (zh) * 2021-11-22 2022-03-08 之江实验室 一种基于知识图谱构建机器人行为树的系统及方法
CN114310883A (zh) * 2021-12-24 2022-04-12 深圳巨隆基科技有限公司 一种基于多知识库的机械臂自主装配方法
CN117009490A (zh) * 2023-08-31 2023-11-07 之江实验室 基于知识库反馈的生成式大语言模型的训练方法和装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3214510B1 (en) * 2016-03-03 2021-06-30 Magazino GmbH Controlling process of robots having a behavior tree architecture
CN116663416A (zh) * 2023-06-01 2023-08-29 西安空天仿真科技有限公司 一种基于行为树的cgf决策行为仿真方法
CN116737908B (zh) * 2023-07-19 2025-07-04 北京百度网讯科技有限公司 知识问答方法、装置、设备和存储介质
CN117033608B (zh) * 2023-09-28 2023-12-22 中国电子科技集团公司第十研究所 一种基于大语言模型的知识图谱生成式问答方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200298401A1 (en) * 2019-01-03 2020-09-24 Lucomm technologies, Inc Semantic Robotic System
US20190392820A1 (en) * 2019-08-05 2019-12-26 Lg Electronics Inc. Artificial intelligence server for setting language of robot and method for the same
CN114153943A (zh) * 2021-11-22 2022-03-08 之江实验室 一种基于知识图谱构建机器人行为树的系统及方法
CN114310883A (zh) * 2021-12-24 2022-04-12 深圳巨隆基科技有限公司 一种基于多知识库的机械臂自主装配方法
CN117009490A (zh) * 2023-08-31 2023-11-07 之江实验室 基于知识库反馈的生成式大语言模型的训练方法和装置

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120258747A (zh) * 2025-06-06 2025-07-04 北京致远互联软件股份有限公司 基于机器人流程自动化的人机协同方法及系统
CN120307305A (zh) * 2025-06-17 2025-07-15 北京理工大学长三角研究院(嘉兴) 一种装配机械臂任务规划方法及装置
CN120370756A (zh) * 2025-06-25 2025-07-25 深圳墨影科技有限公司 基于大模型的机器人低代码可编辑任务序列生成系统
CN120370756B (zh) * 2025-06-25 2025-09-02 深圳墨影科技有限公司 基于大模型的机器人低代码可编辑任务序列生成系统
CN120363218A (zh) * 2025-06-26 2025-07-25 中亿丰数字科技集团股份有限公司 一种建筑机器人任务规划与技能学习方法及系统
CN121722088A (zh) * 2026-02-25 2026-03-24 北京国联视讯信息技术股份有限公司 基于AI Agent的人机协作控制方法及系统

Also Published As

Publication number Publication date
CN117556894A (zh) 2024-02-13

Similar Documents

Publication Publication Date Title
Zhou et al. Llm-bt: Performing robotic adaptive tasks based on large language models and behavior trees
CN117556894A (zh) 基于语义网络和知识库控制机器人任务决策的系统
CN114153943B (zh) 一种基于知识图谱构建机器人行为树的系统及方法
CN110309289A (zh) 一种句子生成方法、句子生成装置及智能设备
CN114201885B (zh) 基于改进行为树的兵力实体行为仿真元建模方法及系统
Ma et al. Explorllm: Guiding exploration in reinforcement learning with large language models
CN109189469A (zh) 基于反射的安卓应用微服务化方法及系统
CN118095349B (zh) 一种大模型记忆机制的实现方法
CN109409592A (zh) 动态环境下移动机器人的最优策略解决方法
CN118332120A (zh) 结合大语言模型和Agent思想的任务动态规划方法
CN118544343A (zh) 一种多机器人系统的任务规划方法、装置、介质及产品
CN114328188B (zh) 一种基于模型的航天器控制器软件测试用例自动生成方法
Mokhtari et al. Experience-based robot task learning and planning with goal inference
CN121210074A (zh) 一种任务路径规划方法与装置
Newell Learning, Generality and Problem-Solving
CN115328129A (zh) 一种基于状态知识图谱的机器人自主行为驱动方法
CN117829242B (zh) 模型处理方法及相关设备
Kotsenko et al. Automated search for several alternative logical inferences in a mivar knowledge base
CN117435749B (zh) 生成知识图谱的方法、装置和存储介质
CN117993518A (zh) 一种基于生成式ai的智能策略生成系统及方法
CN119761514A (zh) 基于知识图谱和大模型的机器人操作任务规划方法及系统
Wang et al. Two fundamental questions concerning BIM data representation for machine learning
CN115796288A (zh) 一种基于知识库的动态场景下推理任务的方法和装置
Patil et al. A Review on Multi-Agent Data Mining Systems
Chetoui et al. Course recommendation model based on Knowledge Graph Embedding

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23958722

Country of ref document: EP

Kind code of ref document: A1