A robot autonomous behavior driving method based on state knowledge graph

Abstract

The present invention belongs to the field of artificial intelligence robot task behavior planning and knowledge graph technology. It is a method for driving robot autonomous behavior based on a state knowledge graph, comprising the following steps: Step 1: Constructing a task group knowledge graph; Step 2: Constructing a triple knowledge graph; Step 3: Constructing a state knowledge graph; Step 4: Designing a decision and driving method; Step 5: Designing a task decision closed-loop control system. The method of the present invention uses a knowledge graph and a dynamic perception system to generate an automatic robot behavior driving method. Using this solution, any robot behavior can be combined with a dynamic perception system. This combination process does not require excessive human editing and intervention. As a user of the technology, one only needs to set the state threshold and the desired attachment state to complete the autonomous control of the robot's behavior.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence robot task behavior planning and knowledge graph technology, and in particular relates to a robot autonomous behavior driving method based on a state knowledge graph. Background Art

[0002] Currently, robot behavior control relies heavily on rule-based constraints. Without receiving instructions, robots are unable to spontaneously generate behavioral commands. They are often forced to perform repetitive tasks based on pre-defined automated instructions by the robot designer, and are unable to autonomously plan their behavior based on environmental changes. To address this, we need to design a computational method based on task knowledge triples and state knowledge graphs. This approach allows the robot to determine the appropriate behavior based on changes in perceived parameters within the scene, informed by the knowledge graph. Summary of the Invention

[0003] The purpose of the present invention is to provide a robot autonomous behavior driving method based on state knowledge graph to solve the above technical problems.

[0004] To solve the above technical problems, the specific technical solution of the robot autonomous behavior driving method based on the state knowledge graph of the present invention is as follows:

[0005] A robot autonomous behavior driving method based on a state knowledge graph includes the following steps:

[0006] Step 1: Build a knowledge graph of the task group;

[0007] Step 2: Construct triple knowledge graph;

[0008] Step 3: Build a state knowledge graph;

[0009] Step 4: Design Decisions and Driving Methodology:

[0010] Step 4.1: Convert the perceptual information into a spatial surface distribution;

[0011] Step 4.2: Design a method to transform perception information into a state change algorithm based on the description of the knowledge graph.

[0012] Step 5: Design the task decision closed-loop control system.

[0013] Furthermore, the step 1 includes the following specific steps:

[0014] Classify task knowledge nodes into two categories: basic task action nodes and task nodes;

[0015] For the action node, the parameters include name, motionId, dsp, weight, vars, and collection, which respectively represent the task name, action execution primitive ID, action scenario description information, action weight, parameter variable table, and collection judgment;

[0016] For task nodes, the parameters include name, collection, and weight, which represent the node name, collection judgment, and weight respectively.

[0017] Furthermore, the step 2 includes the following specific steps:

[0018] Construct a node, which is composed of nouns and contains a list of noun features; the relationship between nodes is a transitive verb node vt, which means that the source node source acts on the target node through the action vt; loops can appear; again, construct real-time tracking information of people and objects in the scene based on environmental changes.

[0019] Furthermore, the step 3 includes the following specific steps:

[0020] The nodes of the state knowledge graph represent two relative states, and the edges are directed edges, indicating the transfer actions between states. The transfer actions between states must correspond to the nodes in the task knowledge. The state knowledge must provide a method for obtaining perception information and an algorithm for changing the state from perception. Through such calculations, real-time environmental information changes are fed back to scene changes.

[0021] Furthermore, the state knowledge in step 3 is defined as follows: a knowledge node of state knowledge is denoted as S, and the relationship between states is calculated as R; where S records a state symbol, and R records an action or task. Two S nodes and a directed R form a triple, denoted as S1–R->S2. By recording such triples, the set {S1–R->S2} is constructed, denoted as the state knowledge graph. Furthermore, step 4.1 includes the following specific steps:

[0022] Step 4.1.1: Extract sampling points from the knowledge base to obtain the specific value of the surface distribution, and then use the basic quadratic function to interpolate to form a specific surface. The surface must be a convex surface and is considered to be a conditional decision space.

[0023] Step 4.1.2: Get the position of the instantiated target in the scene, and then place its surface center (0,0) point at the position of the object in the scene;

[0024] Step 4.1.3: A distribution surface appears in the scene, which is considered to be the influence of the state on the overall environment.

[0025] Furthermore, the step 4.2 includes the following specific steps:

[0026] Step 4.2.1: For any perception, classify it at the perception level, rely on different state recognition algorithms, classify it into the corresponding state description in the state knowledge graph, and then attach it to the instance of the corresponding scene;

[0027] Step 4.2.2: The strength of the instance is given by perception, denoted as f. The state distribution formed by the surrounding is calculated in real time by state knowledge and perception. The surface will give the state of each spatial distribution point, which is expressed as follows:

[0028] f=F(x,y,z)

[0029] Step 4.2.3: Surfaces with the same state can be superimposed in space. This way, when a decision is made relative to any point, a scalar state is obtained. This scalar is used as another scalar from the robot's position, which can be considered as a vector indicating the influence of the state on the robot and the strength of the state's effect on the robot:

[0030] f sum =∑F(x,y,z).

[0031] Furthermore, the step 5 includes the following specific steps:

[0032] Step 5.1: For a state s in the prefab set, search for the expression of the state in the state surface in the scene. If the state does not exist, no processing is performed;

[0033] Step 5.2: If there is an instance of the state s, then calculate the impact of the state s on the robot. If the state s is not in an extreme state, then the robot will make such a decision that the current state s changes towards the prefabricated state;

[0034] Step 5.3: At the same time, all events that can change state s towards the prefabricated state can be queried in the knowledge base. This event will serve as the main task to trigger the autonomous decision;

[0035] Step 5.4: When making autonomous decisions and a task selection occurs, the knowledge base is used to search for the state to determine whether there is an instance of the state in the scene. If there is an instance of the state in the scene, the distribution of the state surface in space will affect the robot's current decision.

[0036] Furthermore, the steps affecting decision making in step 5 include type 1: robot selection of concurrent tasks and type 2: state changes triggering robot behavior;

[0037] Type 1: The robot's selection of concurrent tasks includes:

[0038] When a robot is performing a task, it receives multiple tasks at the same time. Each task is recorded as Task. Then the concurrent tasks at this time constitute a set {Task}.

[0039] The robot's task decomposition process: The task is decomposed and expanded to form a task tree, recorded as Tree-task. For the state set in the scene, if a state forms an association with the task, then the state will affect the execution of the current task. If the parent node of the task depends on the selection of the condition, then the state strength of the current robot at that position is used as the selection weight of the task, recorded as

[0040] W=F(x1,y1,z1)

[0041] Then, there will be more tasks in this state on the task tree. These tasks form a set. All tasks in the set can calculate a weight based on F(x1, y1, z1). The weights are added together to obtain the overall weight W of the task tree for this state at the position. s ,The overall weight is carried out when performing task selection;

[0042] Type 2: State changes triggering robot behaviors include:

[0043] When the robot is in a static state, for one or more changes in the surrounding state, the robot can plan the state in the scene to the desired expression mode based on the unstable factors in the state. This process will produce multiple behaviors, which are combined into a complete task. Driven by task planning, the robot's operating behavior makes the state in the scene return to the desired state. It can be described as a mathematical process: for an environment, the original state set is {s1, s2, s3}, which has an opposite state set, denoted as {s4, s5, s6}, where the states correspond one to one. The initial state perception is {s1, s5, s6} and the desired state {s1, s5, s3}. According to the description of state knowledge, it is concluded that its state s6 is not in the desired state. Therefore, the environment has a trend of transferring from s6->s3. This trend is obtained in the state knowledge to promote the action of this trend.

[0044] Furthermore, in step 5, for the layers formed by superimposing different surfaces, the following method is used to influence the robot task-level decision:

[0045] Type 1: The robot selects concurrent tasks in the following steps:

[0046] A1: Get all task types in a task tree and get tasks in the selected binding state;

[0047] A2: Specify the states that can be monitored and the states in the expected scenarios for a task and scenario;

[0048] A3: Select all states in the scenario and index the task types that constitute state transitions in the state knowledge base;

[0049] A4: Compare the selected task with the tasks found in the knowledge graph to obtain the intersection task of the two.

[0050] A5: Select any task in Task in turn. The set of states that can affect the task is recorded as {State}. According to State, W is calculated based on the robot task target and robot position. s ;

[0051] A6: According to W s Select the robot task and finally select the task that can make the State W s The maximum descent is achieved in the direction of the desired state;

[0052] A7: Follow steps A5 and A6 to select and decide on the content of all tasks, and finally complete the selection of all tasks;

[0053] Type 2: State changes trigger robot behavior:

[0054] B1: Set the initial state set and formulate the desired state set;

[0055] B2: Use real-time perception to monitor the status of each object in the scene;

[0056] B3: Take each state in the state set in turn and check whether the state is close to the expected state;

[0057] B4: If the state is not close to the desired state, search the state knowledge base for a task that can move the state toward the desired state.

[0058] B5: Check all states in turn and obtain all tasks that can make the state move toward the goal. These tasks form a set.

[0059] B6: Query the task knowledge base to see whether the combination of tasks can be classified into a large task;

[0060] B7: If they cannot be classified into one large task, all tasks are executed according to the queue;

[0061] Tasks are decomposed and judged one by one according to type one. After forming a task tree, they are expanded to the smallest executable primitives in sequence, and then drive the robot to perform specific functions.

[0062] The present invention's method for autonomously driving robot behavior based on a state knowledge graph offers the following advantages: It utilizes a knowledge graph and a dynamic perception system to generate an automated method for driving robot behavior. This approach allows for the integration of arbitrary robot behaviors with the dynamic perception system without requiring extensive human editing or intervention. Users of this technology only need to set state thresholds and desired attachment states to achieve autonomous control of the robot's behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flowchart of the method of the present invention;

[0064] Figure 2 This is a schematic diagram of the task knowledge graph structure of the present invention;

[0065] Figure 3 Schematic diagram of the state knowledge graph structure of the present invention;

[0066] Figure 4 This is a schematic diagram of the state layer overlay of the present invention. DETAILED DESCRIPTION

[0067] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a robot autonomous behavior driving method based on a state knowledge graph of the present invention in conjunction with the accompanying drawings.

[0068] like Figure 1 As shown, a robot autonomous behavior driving method based on a state knowledge graph of the present invention includes the following steps:

[0069] Step 1: Build a knowledge graph for the task group.

[0070] like Figure 2 As shown in the figure, the structure of the task knowledge graph is as follows: any node is composed of a task primitive, which represents a basic action for the robot. Any robot behavior is composed of a series of these actions. Since actions are simply combined to complete the primary task combination, these tasks do not have specific actions and serve as the root node of the action node. Therefore, we classify task knowledge nodes into two types: basic task action nodes (leaf nodes) and task nodes (non-leaf nodes).

[0071] For the action node, the parameters include name, motionId, dsp, weight, vars, and collection, which respectively represent the task name, action execution primitive ID, action scenario description information, action weight, parameter variable table, and collection judgment.

[0072] For task nodes, the parameters include name, collection, and weight, which represent the node name, collection judgment, and weight respectively.

[0073] Step 2: Build a triple knowledge graph.

[0074] Construct a node, which consists of a noun and contains a list of noun features. The relationship between nodes is a transitive verb node vt, which means that the source node source acts on the target node through the action vt. Note: loops can occur (set the maximum number of hops during traversal to avoid infinite loops). Again, construct real-time tracking information for people and objects in the scene based on environmental changes.

[0075] Step 3: Build a state knowledge graph.

[0076] The state knowledge graph represents the relationship between states and actions. It is obtained through semantic networks and generally consists of two relative states and an action, indicating that one state becomes another state through the execution of an action. The most basic actions can be classified into tasks. For the state knowledge graph, we design it as follows. Figure 3 As shown, the nodes in the state knowledge graph represent two opposing states, and the edges are directed edges, representing the transition actions between states. Transition actions between states must correspond to nodes in the task knowledge. Furthermore, state knowledge must provide methods for acquiring sensory information and algorithms for converting sensory information into state changes. Through these calculations, real-time environmental information changes can be fed back into scenario changes.

[0077] State knowledge is defined as follows: a knowledge node in state knowledge is denoted as S, and the relationship between states is calculated as R; S records a state symbol (e.g., clean, warm), and R records an action or task (e.g., cleaning, heating). Two S nodes and a directed R form a triple, denoted S1–R->S2. By recording these triples, we form the set {S1–R->S2}, denoted as the state knowledge graph. This constructs a number of structural frameworks for state knowledge. Furthermore, state knowledge nodes must define how states are distributed in the surrounding space to facilitate subsequent decision-making.

[0078] Step 4: Design decisions and driving methods.

[0079] Step 4.1: First, the perception information needs to be converted into a spatial surface distribution. The steps are as follows:

[0080] 1. Take out the sampling points from the knowledge base to obtain the special value of the surface distribution, and then use the basic quadratic function to interpolate to form a specific surface (the surface must be a convex surface), which can be considered as a conditional decision space.

[0081] 2. Get the position of the instantiated target in the scene, and then place its surface center (0,0) point at the position of the object in the scene.

[0082] 3. This will create a distribution surface in the scene, which can be considered as the influence of the state on the overall environment.

[0083] Step 4.2: Design a method to transform perception information into a state change algorithm based on the description of the knowledge graph. The method is as follows:

[0084] 1. For any perception, classification is performed at the perceptual level, relying on different state recognition algorithms (many algorithms are already mature, currently taking facial expressions as an example). For example, by recognizing a person's facial expression, it is classified into the corresponding state description in the state knowledge graph and then attached to the instance of the corresponding scene.

[0085] 2. The strength of the instance is given by perception, denoted as f, and the state distribution formed by the surrounding is calculated in real time based on state knowledge and perception. The surface will give the state of each spatial distribution point, expressed as follows:

[0086] f=F(x,y,z)

[0087] 3. Surfaces of the same state can be superimposed in space, so that when a decision is made relative to any point, a scalar state is obtained. This scalar is used as another scalar from the robot's position, and it can be considered that there is a vector that indicates the impact of the state on the robot and the intensity of the state's effect on the robot.

[0088] f sum =∑F(x,y,z)

[0089] Step 5: Design the task decision closed-loop control system.

[0090] State transition can be applied to robot decision-making, involving state changes. This approach assumes that the robot is given a pre-defined set of states, and all of its actions in the scene are designed to achieve this set of states. Thus, in static scenes, the robot naturally makes decisions about the selected parts of the task based on state trends. Similarly, in a dynamic scene, as objects and people change, state changes will also occur, leading to task decisions and forming a closed-loop control system. This process can be described as follows:

[0091] 1. For a state s in the prefab set, search for the expression of the state in the state surface in the scene. If the state does not exist, no processing is performed;

[0092] 2. If there is an instance of the state s, then calculate the impact of the state s on the robot. If the state s is not in an extreme state, then the robot will make such a decision that the current state s changes towards the prefabricated state.

[0093] 3. At the same time, all events that can change state s towards the prefabricated state can be queried in the knowledge base. This event will serve as the main task to trigger autonomous decision-making

[0094] 4. When making autonomous decisions and a task selection occurs, the knowledge base is used to search for the state to determine whether there is an instance of the state in the scene. If there is an instance of the state in the scene, the distribution of the state surface in space will affect the robot's current decision.

[0095] 5. Such as Figure 4 As shown in Figure 1, for layers formed by superimposing different surfaces, the following method is used to influence the robot's task-level decision-making. The steps that influence the decision-making are divided into two different types, which are broken down and expanded in the following.

[0096] Type 1: Robot’s selection of concurrent tasks

[0097] When a robot is performing a task, it receives multiple tasks at the same time. Each task is recorded as Task. Then the concurrent tasks at this time constitute a set {Task}.

[0098] The robot's task decomposition process: The task is decomposed and expanded to form a task tree, recorded as Tree-task. For the state set in the scene, if a state forms an association with a task, then the state will affect the execution of the current task. If the parent node of the task depends on the selection of the condition, then we use the state strength of the current robot at that position as the selection weight of the task, recorded as

[0099] W=F(x1,y1,z1)

[0100] Then, there will be more tasks in this state on the task tree. These tasks can form a set. All tasks in the set can be weighted according to F(x1, y1, z1). Here, the weights are added together to obtain the overall weight Ws of the task tree for this state. The overall weight is used when selecting tasks.

[0101] Type 2: State changes trigger robot behavior

[0102] When the robot is in a static (idle) state, for one (or more) state changes in the surroundings, the robot can plan the state in the scene to the desired expression mode based on the unstable factors in the state of the scene. This process will produce multiple behaviors. We need to combine these behaviors into a complete task and realize that under the drive of task planning, the robot's operating behavior makes the state in the scene return to the desired state. It can be described as a mathematical process: for an environment, the original state set is {s1, s2, s3}, which has an opposite state set, denoted as {s4, s5, s6}, and the states correspond one to one. For the initial state perception,

[0103] {s1, s5, s6} and the desired state {s1, s5, s3}. According to the description of state knowledge, we can conclude that its state s6 is not in the desired state. Therefore, the environment has a trend of transitioning from s6->s3, and this trend can be obtained in the state knowledge to promote the action of this trend.

[0104] Below we deal with these two different types separately:

[0105] Type 1: The robot selects concurrent tasks in the following steps:

[0106] 1. Get all task types on a task tree and get tasks in the selected binding state

[0107] 2. Specify the states that can be monitored and the states in the expected scenario for a task and scenario.

[0108] 3. Select all states in the scene and index the task types that constitute state transitions in the state knowledge base

[0109] 4. Then, select the task and compare it with the tasks found in the knowledge graph to obtain the intersection of the two tasks.

[0110] 5. Select any task in Task in turn, and the set of states that can affect the task is recorded as {State}. According to State, W will be calculated based on the robot task target and robot position. s

[0111] 6. Follow W s Select the robot task and finally select the task that can make the State W s The maximum descent is achieved in the direction of the desired state.

[0112] 7. Follow steps 5 and 6 to select and decide on the content of all tasks. Finally, complete the selection of all tasks.

[0113] 8. Type 2: State changes trigger robot behavior

[0114] 9. Set the initial state set and formulate the expected state set

[0115] 10. Use real-time perception to monitor the status of each object in the scene (including the robot itself)

[0116] 11. Take each state in the state set in turn and check whether the state is close to the expected state

[0117] 12. If the state is not close to the expected state, search the state knowledge base for tasks that can move the state toward the expected state.

[0118] 13. Check all states in turn and obtain all tasks that can make the state move towards the goal. These tasks constitute a set

[0119] 14. Query the task knowledge base to see if the combination of tasks can be classified into a large task

[0120] 15. If the tasks cannot be classified into one large task, all tasks will be executed according to the queue.

[0121] 16. Tasks are decomposed and judged one by one according to type 1. After forming a task tree, they are expanded to the smallest executable primitives in sequence, and then drive the robot to perform specific functions.

[0122] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. A robot autonomous behavior driving method based on state knowledge graph, characterized in that: The steps include: Step 1: Build a knowledge graph of the task group; Classify task knowledge nodes into two categories: basic task action nodes and task nodes; For the action node, the parameters include name, motionId, dsp, weight, vars, and collection, which respectively represent the task name, action execution primitive ID, action scenario description information, action weight, parameter variable table, and collection judgment; For task nodes, the parameters are name, collection, and weight, which represent the node name, collection judgment, and weight respectively; Step 2: Construct triple knowledge graph; Construct a node, which is composed of nouns and contains a list of noun features; The relationship between nodes is a transitive verb node vt, which means that the source node source acts on the target node through the action vt; loops can occur; secondly, real-time tracking information of people and objects in the scene is constructed based on environmental changes; Step 3: Build a state knowledge graph; The nodes of the state knowledge graph represent two relative states, and the edges are directed edges, representing the transition actions between states. The transition actions between states must correspond to the nodes in the task knowledge. The state knowledge must provide a method for acquiring perception information and an algorithm for changing the perception state. Through such calculations, real-time environmental information changes are fed back to scene changes. The state knowledge of step 3 is defined as follows: a knowledge node of state knowledge is denoted as S, and the relationship between states is calculated as R; where S records a state symbol, and R records an action or task. Two S nodes and a directed R form a triple, denoted as S1–R->S2. By recording such triples, a set {S1–R->S2} is formed, which is denoted as the state knowledge graph; Step 4: Design Decisions and Driving Methodology: Step 4.1: Convert the perceptual information into a spatial surface distribution; Step 4.1.1: Extract sampling points from the knowledge base to obtain the specific value of the surface distribution, and then use the basic quadratic function to interpolate to form a specific surface. The surface must be a convex surface and is considered to be a conditional decision space. Step 4.1.2: Get the position of the instantiated target in the scene, and then place its surface center (0,0) point at the position of the target in the scene; Step 4.1.3: A distribution surface appears in the scene, which is considered to be the influence of the state on the overall environment; Step 4.2: Design a method to transform perception information into a state change algorithm based on the description of the knowledge graph. Step 4.2.1: For any perception, classify it at the perception level, rely on different state recognition algorithms, classify it into the corresponding state description in the state knowledge graph, and then attach it to the instance of the corresponding scene; Step 4.2.2: The strength of the instance is given by perception, denoted as f. The state distribution formed by the surrounding is calculated in real time by state knowledge and perception. The surface will give the state of each spatial distribution point, which is expressed as follows: f=F(x,y,z) Step 4.2.3: Surfaces with the same state can be superimposed in space. This way, when a decision is made relative to any point, a scalar state is obtained. This scalar is used as another scalar from the robot's position, which can be considered as a vector indicating the influence of the state on the robot and the strength of the state's effect on the robot: f sum =∑F(x,y,z); Step 5: Design a task decision closed-loop control system; Step 5.1: For a state s in the prefab set, search for the expression of the state in the state surface in the scene. If the state does not exist, no processing is performed; Step 5.2: If there is an instance of the state s, then calculate the impact of the state s on the robot. If the state s is not in an extreme state, then the robot will make such a decision that the current state s changes towards the prefabricated state; Step 5.3: At the same time, all events that can change state s towards the prefabricated state can be queried in the knowledge base. This event will serve as the main task to trigger the autonomous decision; Step 5.4: When making autonomous decisions and a task selection occurs, the knowledge base is used to search for the state to determine whether there is an instance of the state in the scene. If there is an instance of the state in the scene, the distribution of the state surface in space will affect the robot's current decision.

2. The robot autonomous behavior driving method based on state knowledge graph according to claim 1 is characterized in that: The steps affecting decision making in step 5 include type 1: the robot’s selection of concurrent tasks and type 2: state changes triggering robot behavior; Type 1: The robot's selection of concurrent tasks includes: When a robot is performing a task, it receives multiple tasks at the same time. Each task is recorded as Task. Then the concurrent tasks at this time constitute a set {Task}. That is, the robot's task decomposition process: the task is decomposed and expanded to form a task tree, recorded as Tree-task. For the state set in the scene, if a state forms an association with the task, then the state will affect the execution of the current task. If the parent node of the task depends on the selection of the condition, then the state strength of the current robot at the current position is used as the selection weight of the task, recorded as W=F(x1,y1,z1) Then, there will be more tasks in this state on the task tree. These tasks form a set. All tasks in the set can calculate a weight based on F(x1, y1, z1). The weights are added together to obtain the overall weight W of the task tree for this state at the position. s ,The overall weight is carried out when performing task selection; Type 2: State changes triggering robot behaviors include: When the robot is in a static state, for one or more changes in the surrounding state, the robot can plan the state in the scene to the desired expression mode based on the unstable factors in the state. This process will produce multiple behaviors, which are combined into a complete task. Driven by task planning, the robot's operating behavior makes the state in the scene return to the desired state. It can be described as a mathematical process: for an environment, the original state set is {s1, s2, s3}, which has an opposite state set, denoted as {s4, s5, s6}, where the states correspond one to one. The initial state perception is {s1, s5, s6} and the desired state {s1, s5, s3}. According to the description of state knowledge, it is concluded that its state s6 is not in the desired state. Therefore, the environment has a trend of transferring from s6->s3. This trend is obtained in the state knowledge to promote the action of this trend.

3. The robot autonomous behavior driving method based on state knowledge graph according to claim 2 is characterized in that: In step 5, for the layers formed by superimposing different surfaces, the following method is used to influence the robot's task-level decision: Type 1: The robot selects concurrent tasks in the following steps: A1: Get all task types in a task tree and get tasks in the selected binding state; A2: Specify the states that can be monitored and the states in the expected scenarios for a task and scenario; A3: Select all states in the scenario and index the task types that constitute state transitions in the state knowledge base; A4: Compare the selected task with the tasks found in the knowledge graph to obtain the intersection task of the two. A5: Select any task in Task in turn. The set of states that can affect the task is recorded as {State}. According to State, W is calculated based on the robot task target and robot position. s ; A6: According to W s Select the robot task and finally select the task that can make the State W s The maximum descent is achieved in the direction of the desired state; A7: Follow steps A5 and A6 to select and decide on the content of all tasks, and finally complete the selection of all tasks; Type 2: State changes trigger robot behavior: B1: Set the initial state set and formulate the desired state set; B2: Use real-time perception to monitor the status of each object in the scene; B3: Take each state in the state set in turn and check whether the state is close to the expected state; B4: If the state is not close to the desired state, search the state knowledge base for a task that can move the state toward the desired state. B5: Check all states in turn and obtain all tasks that can make the state move toward the goal. These tasks form a set. B6: Query the task knowledge base to see whether the combination of tasks can be classified into a large task; B7: If they cannot be classified into one large task, all tasks are executed according to the queue; Tasks are decomposed and judged one by one according to type one. After forming a task tree, they are expanded to the smallest executable primitives in sequence, and then drive the robot to perform specific functions.

Images