Robot experiment operation planning and scheduling system and method based on graph neural network

Abstract

The application discloses a robot experiment operation planning and scheduling system and method based on a graph neural network, and the system comprises a pre-training module, a perception module, a decision module and a control module; the pre-training module constructs an initial strategy graph according to a simulation scene and in combination with a prior knowledge base, continuously iteratively updates a decision graph through an experiment flow decision graph learning network according to the initial strategy graph and the simulation scene, and obtains an optimal decision graph; the decision module generates an optimal experiment operation flow scheme according to the optimal decision graph and quantity and category information of a plurality of operation objects obtained by the perception module; and the control module generates a control instruction according to position information of the plurality of operation objects, end pose information of a robot and the optimal experiment operation flow scheme, and controls the robot to perform corresponding experiment operation actions. The application realizes intelligent planning of an overall experiment operation flow of a plurality of operation objects based on a deep learning framework and in combination with a graph neural network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot intelligent strategy planning and control, and particularly relates to a robot experimental operation planning and scheduling system and method based on a graph neural network. BACKGROUND

[0002] Flexible and intelligent planning of experimental processes is one of the most challenging problems in the field of robot intelligent strategy planning in recent years. Generally, experimental processes in chemical laboratories have characteristics such as a variety of operation objects, flexible and variable experimental processes, and unstructured overall task frameworks. The particularity of these tasks itself brings great challenges to laboratory intelligentization.

[0003] At present, laboratory intelligentization basically adopts a scheme in which a robot performs simple operations in a standard laboratory to replace manual operations. The robot does not have a management and planning and scheduling system, and can only complete each operation according to the experimental process sequence when performing a chemical experiment. The experimental process is basically fixed, and has disadvantages such as poor generalization, weak flexibility, and inability to be combined and embedded with other experimental processes. This scheme is only applicable to the case where the operation objects, experimental processes, and laboratory environment are fixed and unchangeable. For example, when the number of test tubes to be operated changes, the robot behavior will report an error, and the process planning cannot be intelligently completed. When the placement order of the test tubes changes, the experiment will fail, and even safety problems may be caused due to improper operation. In addition, the existing scheme has poor portability and adaptability, and also causes higher manufacturing and design costs and power consumption. SUMMARY

[0004] To solve the above problems of the prior art, the present application provides a robot experimental operation planning and scheduling system and method based on a graph neural network. Based on a deep learning framework and combined with a graph neural network model, the present application realizes intelligent planning of the overall experimental operation process of multiple operation objects, improves the degree of laboratory intelligentization, and makes the laboratory intelligent operation have higher flexibility and adaptability.

[0005] In a first aspect, the present disclosure provides a robot experimental operation planning and scheduling system based on a graph neural network.

[0006] A robot experimental operation planning and scheduling system based on a graph neural network comprises a pre-training module, a perception module, a decision module, and a control module.

[0007] The perception module is configured to identify and acquire the number, type, and position information of multiple operation objects in an actual experimental environment.

[0008] The pre-training module is configured to construct an initial strategy graph according to a simulation scenario and in combination with a prior knowledge base, to iteratively update a decision graph through an experimental procedure decision graph learning network according to the initial strategy graph and the simulation scenario, to obtain an optimal decision graph, and to transmit the optimal decision graph to the decision module.

[0009] The decision module is configured to generate an optimal experimental operation procedure scheme according to the optimal decision graph and quantity and category information of the plurality of operation objects.

[0010] The control module is configured to generate a control instruction according to position information of the plurality of operation objects, end pose information of the robot, and the optimal experimental operation procedure scheme, and to control the robot to perform a corresponding experimental operation action.

[0011] In a further technical solution, the perception module transmits the quantity and category information of the plurality of operation objects to the decision module, and transmits the position information of the plurality of operation objects to the control module.

[0012] In a further technical solution, the control module is electrically connected to the robot in an actual experimental environment, and is configured to acquire end pose information of the robot in real time and control actions of the robot.

[0013] In a further technical solution, obtaining the optimal decision graph includes the following steps:

[0014] constructing a random planning strategy graph according to a simulation scenario, and then constructing an initial strategy graph in combination with a prior knowledge base;

[0015] iteratively updating a decision graph through an experimental procedure decision graph learning network according to the initial strategy graph and the simulation scenario, and obtaining an optimal decision graph.

[0016] In a further technical solution, the strategy graph is a triple graph representation G=(u,V,E).

[0017] wherein, represents a node set with a cardinality of N v , v i represents an attribute of an i-th node; is a set of edges with a cardinality of N e , e k represents an attribute of a k-th edge, s k and r k respectively represent indexes of a sender node and a receiver node of the k-th edge; and u represents a global attribute.

[0018] In a further technical solution, the experimental procedure decision graph learning network includes a plurality of fully connected graph neural network blocks connected in sequence, each graph neural network block is configured to update an input decision graph, and each graph neural network block includes three update functions φ and three aggregation functions ρ.

[0019] Further technical solutions, the decision graph is input into the graph neural network block, and an updated decision graph is output, including the following steps:

[0020] Based on the node set, the edge set and the global attribute in the decision graph, the updated edge set E' is obtained through the aggregation function and the update function. e The update function is used to update the edge set, and the updated edge set E' is obtained.

[0021] The updated edge set E' is aggregated through the aggregation function. e→v The aggregation function is used to aggregate the updated edge set.

[0022] According to the node set, the global attribute and the aggregated updated edge set, the updated node set V' is obtained through the aggregation function and the update function. v The update function is used to update the node set, and the updated node set V' is obtained.

[0023] The updated node set V' is aggregated through the aggregation function. e→u The aggregation function is used to aggregate the updated node set.

[0024] Based on the global attribute, the aggregated updated edge set and the node set, the updated global attribute u' is obtained through the aggregation function and the update function. u The update function is used to update the global attribute, and the updated global attribute u' is obtained.

[0025] Further technical solutions, the updated decision graph is obtained, the mechanical arm is controlled to perform a corresponding operation process in a simulation environment based on the updated decision graph, and the decision graph is evaluated by using a reward function r.

[0026] According to the evaluation result, the decision graph is continuously updated until an optimal evaluation result is obtained, and the decision graph corresponding to the optimal evaluation result is output, which is the final decision graph.

[0027] In a second aspect, the present disclosure provides a robot experiment operation planning and scheduling method based on a graph neural network.

[0028] A robot experiment operation planning and scheduling method based on a graph neural network includes:

[0029] The number, type and position information of a plurality of operation objects in an actual experiment environment are obtained.

[0030] According to the simulation scene, the initial strategy graph is constructed in combination with the prior knowledge base, the decision graph is continuously updated through the experiment flow decision graph learning network based on the initial strategy graph and the simulation scene, and the optimal decision graph is obtained.

[0031] Based on the optimal decision graph and the number, kind information of multiple operation objects, an optimal experiment operation process scheme is generated;According to the optimal experiment operation process scheme and the position information of the operation object and the end pose information of the robot, a control instruction is generated to control the robot to perform corresponding experiment operation actions.

[0032] Further technical solutions, obtaining the optimal decision graph, comprising the following steps:

[0033] According to the simulation scene, a random planning strategy graph is constructed, and then an initial strategy graph is constructed in combination with a priori knowledge base;

[0034] According to the initial strategy graph and the simulation scene, the decision graph is continuously updated through the experiment flow decision graph learning network, and the optimal decision graph is obtained.

[0035] The above one or more technical solutions have the following beneficial effects:

[0036] 1、The application provides a robot experiment operation planning and scheduling system and method based on a graph neural network, which is based on a deep learning framework, adopts a learning idea of a graph neural network, realizes intelligent planning of the overall experiment operation process of multiple operation objects, improves the intelligent degree of the laboratory, makes the intelligent operation of the laboratory have higher flexibility and adaptability, and has better expansion performance and generalization ability.

[0037] 2、The application proposes a graph neural network module with a certain paradigm, supports flexible expression, internal configuration and multi-block architecture combination, can adapt to various task scenes, can perform task reconstruction in various scenes, and has good popularization in the field of complex skill learning and combined generalization. DETAILED DESCRIPTION

[0038] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their description serve to explain the application, and do not constitute an improper limitation on the application.

[0039] Figure 1 A schematic diagram of the robot experiment operation planning and scheduling system based on the graph neural network described in the embodiments of the application;

[0040] Figure 2 A flowchart for obtaining the experiment operation flow decision strategy in the embodiments of the application;

[0041] Figure 3 A schematic diagram for defining the graph representation in the embodiments of the application;

[0042] Figure 4 A structural schematic diagram of the experiment flow decision graph learning network in the embodiments of the application;

[0043] Figure 5A structural schematic diagram of a graph neural network in an embodiment of the present application. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0045] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a reference to the presence of a feature, step, operation, device, component, and / or combinations thereof.

[0046] Embodiment one

[0047] The present embodiment provides a robot experimental operation planning and scheduling system based on a graph neural network, as shown in Figure 1 The pre-training module is constructed and run in a simulation environment, and learns the experimental process design strategy of multiple operation objects with different characteristics, different quantities, and different experimental requirements based on the idea of a graph neural network. According to the simulated experimental environment, the pre-training module generates a decision graph, thereby providing the intelligent decision module with prior knowledge of experimental operations, achieving the purpose of flexibly designing experimental process strategies for a known plurality of operation objects. It should be noted that due to the flexibility of simulation, the decision prior knowledge has good adaptability in subsequent optimization and embedding work. The perception module collects the quantity, type, and position information of the experimental operation objects in the experimental environment, and provides the corresponding information to the decision module and the control module, respectively. The quantity and type information of the experimental operation objects is sent to the decision module, and the position information of the experimental operation objects is sent to the control module. The decision module generates a safe and effective optimal experimental operation process scheme under the current experimental environment according to the decision graph and the quantity and type of the experimental operation objects. The control module generates corresponding control instructions according to the optimal experimental operation process method, the position information of each operation object in the experimental environment, and the end pose information of the robot, controls the robot to perform corresponding actions, and completes the intelligent experimental operation of the robot.

[0048] Specifically, the robot experimental operation planning and scheduling system based on a graph neural network proposed in the present embodiment includes a pre-training module, a perception module, a decision module, and a control module.

[0049] The perception module includes a depth camera, an identification system, a calibration system, a pose estimation system, etc., for identifying and obtaining the number, type, and position information of a plurality of operating objects in an actual experimental environment, and transmitting the number and type information of the plurality of operating objects to the decision module and transmitting the position information of the plurality of operating objects to the control module.

[0050] The pre-training module is configured to construct an initial strategy graph according to a simulation scene and in combination with a prior knowledge base, to iteratively update a decision graph through an experimental process decision graph learning network according to the initial strategy graph and a simulation environment, to obtain an optimal decision graph, and to transmit the optimal decision graph to the decision module.

[0051] The decision module is configured to generate an optimal experimental operation process scheme according to the received optimal decision graph and the number and type information of the plurality of operating objects, and to transmit the optimal experimental operation process scheme to the control module.

[0052] The control module is electrically connected to a robot in an actual experimental environment, can obtain robot end pose information in real time, and controls the action execution of the robot. The control module generates a control instruction according to the received position information of the plurality of operating objects, the robot end pose information, and the optimal experimental operation process scheme, and controls the robot to perform a corresponding experimental operation action.

[0053] The above system can realize intelligent experimental operation in various task scenarios, can adaptively plan and schedule experimental operation processes according to the number and position of test tubes in a chemical experiment, controls the robot to perform corresponding experimental operations, realizes complex skill learning, and has better intelligent experimental operation portability and adaptability. Further, the implementation of the above system focuses on generating an optimal decision graph through a pre-training module. The optimal decision strategy for generating an experimental process includes: giving a basic graph structure according to known experimental process conditions, building a graph learning network to learn a decision graph, and then giving experimental operation process planning actions in combination with experimental environment information to control a robot in a simulation environment to complete an operation task. Then, quality evaluation feedback is performed according to process planning completion efficiency and other indicators / standards, and the above steps are repeatedly performed to continuously optimize the strategy. After multiple iterations, an optimal experimental process intelligent decision strategy is finally generated, i.e., an optimal decision graph is generated.

[0054] As shown in Figure 2 the step of obtaining the optimal decision graph includes:

[0055] Step S1: According to a simulation scene, a random planning strategy graph is constructed, and then an initial strategy graph is constructed in combination with a prior knowledge base.

[0056] Specifically, the pre-training module is constructed and run in a simulation environment. First, a random planning strategy graph is constructed according to the simulation scene. The experimental environment information is built in the simulation environment to obtain the simulation scene, which includes not only the robot / robotic arm and mobile platform, but also the laboratory environment, such as the desktop, various operation objects and other basic objects involved in the actual operation process. According to the simulation environment or simulation scene, the scene description matrix is defined, the robot arm end position is defined, the arm action representation is defined, and the graph representation is defined.

[0057] (1) Define the scene description matrix O m×n

[0058] The scene description matrix is a set of m operation objects, each of which is represented by an n-dimensional vector encoding the features of the operation object. In this embodiment, a 4-dimensional vector is used to describe the laboratory reagent operation scene using reagent index and three-dimensional spatial coordinates.

[0059] (2) Define the robot arm end position P = {x, y, z}

[0060] In the robot arm end position P = {x, y, z}, x, y, and z represent the respective position coordinates of the robot arm in the world coordinate system.

[0061] (3) Define the robot arm action representation A = {θ1, θ2, θ3, θ4, θ5, θ6, θ7}

[0062] In this embodiment, the robot arm is a seven-axis robot arm, and in the robot arm action representation A = {θ1, θ2, θ3, θ4, θ5, θ6, θ7}, θ i＝1：7 is the joint angle of the seven-axis robot arm.

[0063] (4) Define the graph representation G = (u, V, E)

[0064] As shown in Figure 3 , the triple graph representation G represents the overall graph information containing global attributes, represents a node set with a cardinality of N v , where v i represents the attributes of the i-th node, i.e., the reagent type and position of the i-th test tube; is a set of edges with a cardinality of N e , where e k represents the attributes of the k-th edge, s k and r k represent the indices of the sender node and the receiver node of the k-th edge, respectively, i.e., whether the pouring order between certain operation objects and other operation objects is feasible; u represents the global attribute, i.e., the fastest and optimal scheme planning under the premise of ensuring the feasibility of the experimental process.

[0065] In addition, in order to ensure the safety of the experimental operation, a priori knowledge base K is defined. K represents the basic chemical safety operation requirements, which has the highest priority of laboratory operation, for example, for the mixing of concentrated sulfuric acid and water, only water can be added first, otherwise it will cause safety hazards. That is, the basic safety requirements of the experiment are actually the priori knowledge base K, and the graph representation is defined according to the simulation scene, and the graph representation at this time is the strategy graph G of the random planning, and on this basis, the initial strategy graph G0 with safety operation behavior planning ability is obtained by combining the basic safety requirements of the experiment.

[0066] Step S2, according to the initial strategy graph and the simulation scene, the decision graph is continuously updated by the experimental process decision graph learning network, and the optimal decision graph is obtained.

[0067] Firstly, the experimental process decision graph learning network is constructed, which is actually a graph neural network, and the network structure is as shown in Figure 4 The network structure is as shown in

[0068] As shown in Figure 5 Each GNN block is used to update the input decision graph, and three update functions φ and three aggregation functions ρ are defined in the GNN block, which are respectively:

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] Wherein, φ e function is used to map to all edges to calculate each edge update, φ v function is used to map to all nodes to calculate each node update, φ u function is used to apply once as a global update; ρ function is used to take a set as input, and simplify it into a single element representing aggregated information.

[0076] In each GNN block, the input decision graph is first applied to each edge φ e function, and the input parameters of the function are Return the updated edge parameters e' k which corresponds to the connection relationship between two operation objects. The set of outputs of each edge connected to each node i is The set of edges involved by all nodes (i.e., traversing all nodes i) is The set is the updated edge set. That is, based on the node set, edge set and global attribute in the decision graph, the edge set is updated by the function e to obtain the updated edge set E'.

[0077] Secondly, the updated edge set is projected by the function e→v to update and aggregate the edges connected to the vertex i into , which is convenient for use in the next node update. The role of this step is to realize the aggregation of all connection relationships on the i-th operation object. That is, the updated edge set is aggregated by the function e→v

[0078] Thirdly, based on the updated edge set, the function v is applied to each node i to calculate the updated node attribute v' i . That is, based on the aggregated updated edge set, node set and global attribute, the node set is updated by the function v to obtain the updated node set V'. The function v is used to calculate the result similar to each operation object after the edge information, and the set of outputs of each node The set of nodes is the updated node set.

[0079] Again, the updated node set V' is aggregated by the function e→u to update all nodes into which is then used for the next global update; at the same time, the updated edge set E' is aggregated by the function e→u to update all edges into which is then used for the next global update.

[0080] Finally, based on the global attribute, the aggregated updated edge set and node set, the function u is applied to calculate the updated global attribute to obtain the updated global attribute u'.

[0081] Through the above steps, the decision graph is input into the GNN block to output the updated decision graph.

[0082] ​In the simulation environment, an updated decision graph is obtained, the robot arm is controlled to perform a corresponding operation process based on the decision graph, the decision graph is evaluated by using a reward function r, such as evaluating the quality of the process planning, feedback is performed according to the evaluation result, the updating of the decision graph is iterated continuously until an optimal evaluation result is obtained, and at this time the output decision graph is the final decision graph.

[0083] In this embodiment, a reward function r = f (G0, P) is defined, where P represents the end position of the robot arm, and the reward function r gives a strategy evaluation value according to evaluation indexes such as the time used for the robot arm operation and the smoothness of the operation process when the end position of the robot arm is different, the position where the operation target is placed is different, and there are multiple sequence schemes in the planning under the premise that the experimental operation is completely safe, and the higher the strategy evaluation value is, the better. Through continuous iterative updating, the highest strategy evaluation value r max The corresponding planning strategy is the optimal scheme under the current operation environment, that is, the optimal decision graph G perf .

[0084] On the basis of obtaining the optimal strategy graph through the above pre-training, an optimal experimental process is generated through environment interaction and strategy graph, and the robot arm is controlled by the control module to complete corresponding actions.

[0085] Embodiment two

[0086] The embodiment provides a robot experimental operation planning and scheduling method based on a graph neural network, and the method comprises the following steps:

[0087] The number, type and position information of a plurality of operation objects in an actual experimental environment are obtained.

[0088] According to the simulation scene, an initial strategy graph is constructed in combination with a priori knowledge base, and the decision graph is iteratively updated by using an experimental process decision graph learning network according to the initial strategy graph and the simulation scene, so as to obtain an optimal decision graph.

[0089] Based on the optimal decision graph and the number and type information of the plurality of operation objects, an optimal experimental operation process scheme is generated, and a control instruction is generated according to the optimal experimental operation process scheme, the position information of the plurality of operation objects and the end position information of the robot, so that the robot performs corresponding experimental operation actions.

[0090] Further, the above obtaining of the optimal decision graph specifically comprises the following steps:

[0091] According to the simulation scene, a random planning strategy graph is constructed, and then an initial strategy graph is constructed in combination with a priori knowledge base.

[0092] According to the initial strategy graph and the simulation scene, the decision graph is iteratively updated by using an experimental process decision graph learning network, so as to obtain an optimal decision graph.

[0093] The steps involved in the above embodiment two correspond to embodiment one, and the specific implementation can refer to the relevant description part of embodiment one.

[0094] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0095] The above describes the specific embodiments of the present application in combination with the accompanying drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A robot experiment operation planning and scheduling system based on a graph neural network, characterized in that, The pre-training module, the perception module, the decision module and the control module are included. The perception module is used for identifying and acquiring the quantity, category and position information of a plurality of operation objects in an actual experimental environment. The pre-training module is used for constructing an initial strategy graph according to a simulation scene and in combination with a priori knowledge base, continuously updating a decision graph through an experimental flow decision graph learning network according to the initial strategy graph and the simulation scene, obtaining an optimal decision graph, and transmitting the optimal decision graph to the decision module. The optimal decision graph is obtained by the following steps: According to the simulation scene, a stochastic planning strategy graph is constructed, and an initial strategy graph is constructed in combination with a prior knowledge base; according to the initial strategy graph and the simulation scene, a decision graph is iteratively updated through an experimental process decision graph learning network to obtain an optimal decision graph; the experimental process decision graph learning network comprises a plurality of fully connected graph neural network blocks connected in sequence, each graph neural network block is used for updating an input decision graph, and each graph neural network block comprises three update functions and three aggregation functions . ​ The decision graph is input into a graph neural network block, and an updated decision graph is output, including the following steps: Based on the node set, the edge set and the global attribute in the decision graph, by updating the function, updating the edge set, obtaining the updated edge set ; updated edge set by aggregating the updated edge set; According to the node set, the global attribute and the updated edge set, the node set is updated by an update function to obtain an updated node set ; updated node set after aggregate function operates on the updated node set; Based on the global attribute, the aggregated updated edge set and node set, the global attribute is updated by applying an update function to obtain an updated global attribute ;​ The decision module is used for generating an optimal experimental operation flow scheme according to the optimal decision graph and the quantity and category information of the plurality of operation objects. The control module is used for generating a control instruction according to the position information of the plurality of operation objects, the robot end pose information and the optimal experimental operation flow scheme, and controlling the robot to perform corresponding experimental operation actions.

2. The robot experiment operation planning and scheduling system based on a graph neural network according to claim 1, characterized in that, The perception module transmits the quantity and category information of the plurality of operation objects to the decision module, and transmits the position information of the plurality of operation objects to the control module.

3. The robot experiment operation planning and scheduling system based on a graph neural network according to claim 1, characterized in that, The control module is electrically connected with the robot in the actual experimental environment, and is used for acquiring the robot end pose information in real time and controlling the action execution of the robot.

4. The robot experiment operation planning and scheduling system based on a graph neural network according to claim 1, characterized in that, The policy graph is a triple graph representation ); wherein, represents a set of nodes with cardinality , represents an attribute of the i th node; is a set of edges with cardinality , represents an attribute of the k th edge, and represent the indices of the sender node and the receiver node of the k th edge, respectively; u represents a global attribute.

5. The robot experiment operation planning and scheduling system based on a graph neural network according to claim 1, characterized in that, Obtain the updated decision graph, control the robot to perform the corresponding operation process in the simulation environment based on the updated decision graph, and utilize the reward function r Evaluate the decision graph; According to the evaluation results, the decision graph is continuously updated until the optimal evaluation result is obtained, and the decision graph corresponding to the optimal evaluation result is output as the final decision graph.

6. A robot experiment operation planning and scheduling method based on a graph neural network, characterized in that, The quantity, category and position information of a plurality of operation objects in an actual experimental environment are acquired. An initial strategy graph is constructed according to a simulation scene and in combination with a priori knowledge base, a decision graph is continuously updated through an experimental flow decision graph learning network according to the initial strategy graph and the simulation scene, and an optimal decision graph is obtained. The decision graph is input into a graph neural network block, and an updated decision graph is output, including the following steps: According to the simulation scene, a random planning strategy graph is constructed, and an initial strategy graph is constructed in combination with a prior knowledge base; according to the initial strategy graph and the simulation scene, a decision graph is iteratively updated through an experimental process decision graph learning network to obtain an optimal decision graph; the experimental process decision graph learning network comprises a plurality of fully connected graph neural network blocks connected in sequence, each graph neural network block is used for updating an input decision graph, and each graph neural network block comprises three update functions and three aggregation functions . ​ An optimal experimental operation flow scheme is generated based on the optimal decision graph and the quantity and category information of the plurality of operation objects, and a control instruction is generated according to the optimal experimental operation flow scheme, the position information of the plurality of operation objects and the robot end pose information, so as to control the robot to perform corresponding experimental operation actions. Based on the node set, the edge set and the global attribute in the decision graph, by updating the function, updating the edge set, obtaining the updated edge set ; updated edge set by aggregating the updated edge set; According to the node set, the global attribute and the updated edge set, the node set is updated by an update function to obtain an updated node set ; updated node set after aggregate function operates on the updated node set; Based on the global attribute, the aggregated updated edge set and node set, the global attribute is updated by applying an update function to obtain an updated global attribute ;​ ​

Images