Multi-robot cooperative exploration method and system based on relationship-gated graph neural network

Abstract

The application provides a multi-robot cooperative exploration method and system based on a relational gate graph neural network, relates to the fields of mobile robots and artificial intelligence technology, and comprises the following steps: acquiring observation information of an environment to be explored, constructing an environment map, and extracting a candidate exploration target point set; constructing a topological graph containing robot nodes, target point nodes, and multiple types of relationship edges for representing different semantic relationships according to the robot set and the candidate exploration target point set, and performing feature coding on the nodes and the relationship edges to obtain a graph structure representation; inputting the graph structure representation into a relational gate graph neural network, weighting and fusing different relationship information through a gating mechanism to obtain a fused feature representation of the nodes; generating a target allocation result between the robots and the target points based on the fused feature representation, and controlling the robots to perform an exploration task on the allocated target points according to the result; and the application realizes more efficient and coordinated multi-robot cooperative exploration.

Description

Technical Field

[0001] This invention relates to the fields of mobile robots and artificial intelligence technology, specifically to a multi-robot collaborative exploration method and system based on relation-gated graph neural networks. Background Technology

[0002] In the exploration of unknown environments, multi-robot systems can significantly improve the efficiency of environmental mapping through collaborative perception and task allocation. Existing methods mainly include heuristic rule-based methods, optimization modeling methods, and information theory-based methods.

[0003] However, existing technologies typically rely on manually designed utility functions or simple distance metrics, which are difficult to effectively model the complex relationships between multiple robots and candidate targets. At the same time, in the process of multi-robot collaboration, problems such as target conflict, repeated exploration, and insufficient information utilization are likely to occur. Moreover, traditional methods are difficult to uniformly represent the interaction structure between multiple agents and lack the ability to adaptively model different relationship types, thus limiting the collaborative decision-making performance of the system in complex environments.

[0004] Therefore, existing multi-robot collaborative exploration processes suffer from problems such as insufficient relationship modeling capabilities, unreasonable target allocation, and low exploration efficiency. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a multi-robot collaborative exploration method and system based on relation-gated graph neural networks, enabling more efficient and coordinated multi-robot collaborative exploration.

[0006] According to some embodiments, the present invention adopts the following technical solution: Multi-robot cooperative exploration methods based on relation-gated graph neural networks include: Acquire observational information about the environment to be explored, construct an environmental map, and extract a set of candidate exploration target points; Based on the robot set and the candidate exploration target point set, a topological graph is constructed, which includes robot nodes, target point nodes, and multi-type relation edges used to represent different semantic relationships. The nodes and relation edges are then feature-encoded to obtain a graph structure representation. The graph structure is used to represent the input relation-gated graph neural network. Different relation information is weighted and fused through a gating mechanism to obtain the fused feature representation of the nodes. Based on the fused feature representation, a target allocation result between the robot and the target point is generated, and each robot is controlled to perform the exploration task of the allocated target point according to the result.

[0007] According to some embodiments, the present invention adopts the following technical solution: A multi-robot collaborative exploration system based on relation-gated graph neural networks includes: The information acquisition module is configured to: acquire observation information of the environment to be explored, construct an environment map, and extract a set of candidate exploration target points; The feature encoding module is configured to: construct a topological graph containing robot nodes, target point nodes, and multi-type relation edges representing different semantic relationships based on the robot set and the candidate exploration target point set; and perform feature encoding on the nodes and relation edges to obtain a graph structure representation. The feature fusion module is configured to: input the graph structure representation into a relation-gated graph neural network, and perform weighted fusion of different relation information through a gating mechanism to obtain the fused feature representation of the nodes; The target allocation module is configured to: generate a target allocation result between the robot and the target point based on the fused feature representation, and control each robot to perform an exploration task on the allocated target point according to the result.

[0008] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the multi-robot cooperative exploration method based on a relation-gated graph neural network.

[0009] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the multi-robot cooperative exploration method based on a relation-gated graph neural network.

[0010] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the multi-robot cooperative exploration method based on relation-gated graph neural networks.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires environmental observation information and extracts candidate exploration target points, constructs a topological graph containing robot nodes, target point nodes, and multiple types of relationship edges, and performs feature encoding. Then, it uses a relation-gated graph neural network to perform gated weighted fusion of different relationship information, and finally generates target allocation results to control the robot to perform exploration tasks. It can make full use of various semantic relationship information between robots and between robots and target points. Through the gating mechanism, it adaptively fuses the importance of different relationships, thereby improving the accuracy of target allocation and the efficiency of collaborative exploration. It avoids allocation conflicts or exploration redundancy caused by single relationship modeling. At the same time, the structured representation of the graph neural network enables the model to flexibly adapt to different numbers of robots and target points, enhances the generalization ability and robustness of the method in dynamic unknown environments, and ultimately achieves more efficient and coordinated multi-robot collaborative exploration. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0013] Figure 1 This is a flowchart of the method in Example 1. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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 invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0017] Example 1 One embodiment of the present invention provides a multi-robot cooperative exploration method based on a relation-gated graph neural network, such as... Figure 1 As shown, it includes: Step S1: Obtain observation information of the environment to be explored, construct an environment map, and extract a set of candidate exploration target points; Step S2: Based on the robot set and the candidate exploration target point set, construct a topology graph containing robot nodes, target point nodes, and multi-type relation edges used to represent different semantic relationships, and perform feature encoding on the nodes and relation edges to obtain the graph structure representation; Step S3: Input the graph structure representation into the relation-gated graph neural network, and use the gating mechanism to perform weighted fusion of different relation information to obtain the fused feature representation of the nodes; Step S4: Based on the fused feature representation, generate the target allocation result between the robot and the target point, and control each robot to perform the exploration task of the allocated target point according to the result.

[0018] As one embodiment, the multi-robot collaborative exploration method based on relation-gated graph neural networks of the present invention achieves more efficient and coordinated multi-robot collaborative exploration. The specific implementation process is described below using a two-dimensional unknown environment as an example: I. Modeling the problem of multi-robot cooperative exploration in a two-dimensional unknown environment Let a multi-robot system contain N mobile robots, denoted as set. Unknown environments are represented using occupancy grid maps, a widely used representation method for robot environmental perception and path planning. It divides the environment into many small grids, each corresponding to a discrete spatial region, where each grid unit... Corresponding to an occupancy probability Each grid cell typically has three possible states: occupied, free, and unknown. These states can be described by binary tags or probability values.

[0019] robot The state at time t is represented as:

[0020] in, Indicates spatial location, Indicates the orientation angle. This indicates speed information.

[0021] II. Multi-robot collaborative exploration based on modeling The goal of multi-robot cooperative exploration methods is to maximize the exploration efficiency of unknown environments through cooperative strategies while ensuring safe obstacle avoidance. Specifically, this means maximizing the known area within a limited time. Step 1: Acquire observation information of the current unknown environment through the sensors of the multi-robot system, and construct a local grid map based on each robot; identify the boundary between the known free area and the unknown area in the local grid map, thereby extracting a set of candidate exploration target points.

[0022] Each robot perceives its surroundings using LiDAR or vision sensors and updates its local occupancy grid map based on sensor data, that is, updates the state of the perceived grid cells, such as when a grid cell is detected. If there are obstacles at the location, then the grid cell will be... The state is set to "occupied", thus obtaining a local occupancy grid map represented by a matrix. Each item in the matrix is ​​the state of the corresponding grid cell, and the matrix is ​​used as the robot's current observation information.

[0023] In a raster map, rasters are divided into three categories: (1) The free region is known and consists of grid cells in a free state; (2) The known obstacle area is composed of grid cells in an occupied state; (3) Unknown region, which is composed of grid cells in an unknown state.

[0024] The leading edge point is defined as the boundary position between the known free region and the unknown region, i.e., a raster cell that satisfies the following conditions:

[0025] To avoid computational complexity issues caused by an excessive number of frontier points, clustering methods (such as distance-based clustering algorithms) can be used to group the frontier points, thereby obtaining a set of candidate exploration target points. ,in, The j-th candidate exploration target point can be a frontier point or a region composed of adjacent frontier points obtained through clustering.

[0026] To evaluate the exploration value of different target points, an information gain function is introduced:

[0027] in, Indicates the target point The local area centered on The raster information entropy is expressed by the formula:

[0028] in, For grid cells The probability of occupancy.

[0029] Furthermore, a distance attenuation factor can be introduced to correct the information gain:

[0030] in, To get the robot to the target point distance, This is the attenuation coefficient.

[0031] Step 2: Based on the robot set and the candidate exploration target point set, construct a multi-robot-target point relationship topology graph; The topology diagram includes: - Node set: includes robot nodes and front edge nodes; - Edge set: includes robot-front-point relationship edges, robot-robot relationship edges, and front-point-front-point relationship edges; Different types of edges are used to represent different semantic relationships, including spatial distance relationships, reachability relationships, and collaborative constraint relationships.

[0032] The set of multiple robots and target points is represented as a graph structure. Node set middle, For a set of robot nodes, The set of target point nodes.

[0033] edge set include: (1) Robot-target point relationship edge This is used to describe the reachability and cost relationship between the robot and the target point; (2) Robot-robot cooperative relationship edge This is used to describe the cooperation and collision avoidance constraints between robots; (3) Spatial adjacency relationship edge between target points It is used to describe the spatial distribution relationship between target points.

[0034] For any pair of nodes (i,j), establish corresponding edges in the graph when they satisfy the distance threshold or reachability constraint.

[0035] Step 3: Perform feature encoding on the nodes and edges in the topological graph to obtain a unified graph structure representation.

[0036] Specifically, the characteristics of a robot node include its spatial position, orientation angle, velocity information, current observation information, and historical path trajectory information, defined as:

[0037] in, Represents robots Spatial location, Represents robots The orientation angle, Represents robots Speed ​​information, For robots Current observation information; robot Historical path trajectory recorded This includes the coordinates of its previous motion sequence.

[0038] Target point node features include spatial location, information gain, and neighborhood structure features, defined as:

[0039] in, Indicates the target point Spatial location, Indicates the target point Corrected information gain; for the target point Its frontier point density is denoted as This index reflects the spatial neighborhood structure characteristics of a target point, that is, the distribution of other target points within a certain range around the target point. It measures the spatial concentration of target points and is expressed by the formula:

[0040] in, Indicates Let r be the set of target points within a neighborhood centered at a radius of r. The number of target points in the neighborhood. A target point in the set of target points.

[0041] To improve the stability of model training, the above node features are normalized:

[0042] Where μ and σ are the characteristic mean and standard deviation, respectively.

[0043] Edge features include the relative distance between nodes, orientation information, and reachability constraints, and are defined as follows:

[0044]

[0045]

[0046] in, For nodes With nodes The relative distance between them For directional information, The path reachability index is used to represent reachability constraint information. If a node... With nodes The value is 1 if a path exists, otherwise it is 0.

[0047] Through the above encoding, the state information and environmental structure information of multiple robots are uniformly mapped to the graph structure representation space.

[0048] Step 4: Input the above graph structure representation into the relation-gated graph neural network for feature learning and information fusion; during the information transmission process, a relation gating mechanism is introduced for different types of relation edges to adaptively weight the information contributions of different relations, thereby obtaining a high-dimensional representation of nodes that integrates multi-relation information, including: (1) Perform feature transformations on different relation types respectively; In relation-gated graph neural networks, the graph structure contains various types of relations, such as cooperative relations between robots, matching relations between robots and target points, and continuity relations between frontier points. Each relation has its unique semantics and function, so it is necessary to perform separate feature transformations on different types of relations.

[0049] Specifically, for each edge under the same relation type, a multilayer perceptron is used to encode information by combining the node features at both ends of the edge and the corresponding edge features, thus obtaining relation edge-level features. This can be expressed as a formula:

[0050] in, and Representing nodes respectively With nodes initial characteristics, Representing relations The edge features below, It is a relationship The corresponding multilayer perceptron, Represents a node With nodes In relationship The lower-level features.

[0051] In the feature encoding stage, different relation types use independent parameter mapping networks to transform the node interaction information under the corresponding relation, so that the network can learn the specific semantic representation corresponding to different relations. It should be noted that this process is for each specific edge under the same relation type, rather than generating a unified feature for a certain type of relation as a whole.

[0052] (2) Calculate the importance weights of each relation using a gating function; In the feature fusion process of graph neural networks, different types of relations have different effects on node updates. In order to enable the network to adjust the degree of attention to each relation according to the task requirements at different exploration stages, a gating function is used to calculate the weight of each relation. The relation gating mechanism dynamically adjusts the importance of each relation through a learned gating network.

[0053] Specifically, after completing the edge-level feature transformation under different relation types, the node features in the current graph are first processed. Relationship-level features Perform global aggregation to obtain graph state information that can represent the current exploration state. ,in, This represents the layer number of the graph neural network. The graph state information comprehensively reflects the spatial distribution, target association, and environmental exploration structure of the current multi-robot system, and serves as the input to the gating network. Expressed as a formula:

[0054] in, and Let these represent the set of nodes and the set of edges in the graph, respectively. The global aggregation function can summarize the features of all nodes in the graph into a global representation vector. In this embodiment, the global aggregation function adopts a weighted pooling method based on the attention mechanism. During the aggregation process, each node is assigned a different weight according to its contribution to the current exploration task.

[0055] Graph state information based on the current exploration state Calculate the relationship gating coefficient The formula is as follows:

[0056] in, It is the weight matrix of the gated network. It is a bias term. It is an activation function used to map the output to the interval [0, 1], representing a relation. The importance of each relationship is determined. In this way, the network can adaptively adjust the information contribution of various relationships based on the current state of exploration.

[0057] (3) Weighted aggregation of adjacent node information based on weights; In the message passing phase of a relation-gated graph neural network, each node not only aggregates information from its neighboring nodes but also performs weighted aggregation based on the weights of each relation; for each layer, nodes... The feature update depends on the information of its neighboring nodes, and the update formula is:

[0058] in, This represents the set of all relationship types in the graph, including robot collaboration relationships, robot-target point matching relationships, and frontier point continuity relationships, etc. Represents a node In the The updated feature representation of the layer. This represents the layer number in a graph neural network. It is a node In relationship The set of neighboring nodes below, Indicates the first Layer nodes To the node Transmission in relation The relationship between the edge features is as follows: Based on relationships The aggregation operation of neighbor node information is usually a weighted average, and the aggregation result of each relation is multiplied by the corresponding gating coefficient. This ensures that the influence of different relationships in information transmission is dynamically adjustable; this weighted aggregation operation enables the features of each node to integrate information from different types of relationships, and adapts to the task requirements at different stages according to the adjustment of the gating mechanism.

[0059] Through the above steps, the fused feature representation of each node is finally obtained, and the robot... and target point The fusion features are represented by respectively using and express.

[0060] Step 5: Generate target assignment results based on fusion features The output of the relation-gated graph neural network is a high-level feature representation of each robot and the target point, reflecting the relationship between each robot and the target point in the current exploration state. Based on these feature representations, target assignment decisions are made to ensure that the multi-robot system can efficiently allocate tasks and coordinate its exploration behavior.

[0061] In a multi-robot collaborative exploration task, each robot needs to select a target to explore. Let the current time be... The robot collection is The set of target points is Each robot A target point needs to be selected. The target assignment problem can be formalized as a graph matching problem, where the matching relationship between robot nodes and target node is described by an affinity matrix.

[0062] In the feature learning stage of the graph neural network, high-level feature representations of each robot and target point have been obtained. By calculating the similarity or fit between the robot and the target point, an affinity matrix is ​​constructed. ,in Represents robots and target point The affinity between them is calculated using the following formula:

[0063] in, It is a mapping function used to calculate the matching degree between the robot and the target point. In this embodiment, the mapping function adopts the cosine similarity function.

[0064] The affinity matrix reflects the compatibility between each robot and each target point. The higher the affinity, the more suitable the robot is for performing the exploration task at that target point.

[0065] The graph matching problem essentially requires an affinity matrix... The goal is to find a one-to-one matching relationship such that each robot is matched with a target point, and each target point can only be selected by one robot. To achieve this, the Sinkhorn algorithm is used for approximate double-random matrix normalization. The Sinkhorn algorithm can approximate the constraint structure of the linear assignment problem in continuous space while maintaining end-to-end differentiability, making it suitable for joint optimization with graph neural networks.

[0066] Specifically, the Sinkhorn algorithm uses alternating row and column normalization operations to normalize the affinity matrix. Transform into an approximate double random matrix Its elements Represents robots With the target point The probability of a match between them is expressed by the formula:

[0067] in, It is a probability matrix that satisfies the following conditions:

[0068] The Sinkhorn algorithm guarantees matrix integrity. The system satisfies the double random constraint, meaning that the sum of each row and each column is 1. This implies that each robot is assigned a target point, and each target point will only be assigned to one robot.

[0069] During the execution phase, the matrix This translates into a final target allocation decision; specifically, it allows for the selection of the highest probability match (i.e., selecting...). The largest element), thus assigning a target point to each robot.

[0070] Step 6: Perform path planning and task execution based on the target allocation results. Based on the target allocation results in step 5, a target point has been assigned to each robot. These target points represent the exploration locations that each robot needs to go to in the current round. In step 6, the motion path of each robot needs to be generated based on these target points to ensure that the robot can move smoothly from its current position to the target point.

[0071] The goal of short-term path planning is to generate a safe and feasible path for each robot under given long-term objective conditions, and then drive the robot to execute that path through motion control commands. Here, the short-term path planning module does not directly participate in the optimization of long-term objective selection; instead, it stably completes objective tracking and path execution based on the results of higher-level decisions (i.e., objective allocation).

[0072] Step 7: Collaborative Updates and Optimization After completing the task, the overall exploration performance of the multi-robot system needs to be evaluated. The evaluation aims to determine whether the system has achieved its predetermined goals based on the actual performance in the current round. To guide the multi-robot system to achieve efficient collaborative exploration, a comprehensive reward function is designed:

[0073] The parameters are explained in detail below: (1) Information gain reward , indicating the selection of the target point The resulting information benefits can be expressed by the formula:

[0074] in, These represent the number of robots and the number of target points, respectively. Represents robots With the target point The probability of a match between them. Indicates the target point The corresponding information gain is usually calculated from the area of ​​the unknown region near the target point.

[0075] (2) Exploration efficiency reward , representing the increase in the area of ​​the explored region per unit time, is expressed by the formula:

[0076] in, Indicates the time interval The incremental area of ​​the explored region; It is the time step, representing each round or control cycle.

[0077] (3) Path cost penalty :

[0078] in, These represent the number of robots and the number of target points, respectively. Represents robots With the target point The probability of a match between them. Represents robots To the target point The path length, i.e., the relative distance.

[0079] By independently updating the strategy of each robot and using the aforementioned comprehensive reward function for target evaluation, the collaborative decision-making among multiple robots is gradually optimized. After each update, the robots can adjust target allocation and path planning according to the new strategy, further improving overall collaborative efficiency.

[0080] Example 2 One embodiment of the present invention provides a multi-robot cooperative exploration system based on a relation-gated graph neural network, comprising: The information acquisition module is configured to: acquire observation information of the environment to be explored, construct an environment map, and extract a set of candidate exploration target points; The feature encoding module is configured to: construct a topological graph containing robot nodes, target point nodes, and multi-type relation edges representing different semantic relationships based on the robot set and the candidate exploration target point set; and perform feature encoding on the nodes and relation edges to obtain a graph structure representation. The feature fusion module is configured to: input the graph structure representation into a relation-gated graph neural network, and perform weighted fusion of different relation information through a gating mechanism to obtain the fused feature representation of the nodes; The target allocation module is configured to: generate a target allocation result between the robot and the target point based on the fused feature representation, and control each robot to perform an exploration task on the allocated target point according to the result.

[0081] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the multi-robot collaborative exploration method based on a relation-gated graph neural network.

[0082] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions. When the computer instructions are executed by a processor, the multi-robot cooperative exploration method based on relation-gated graph neural networks is implemented.

[0083] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the multi-robot cooperative exploration method based on relation-gated graph neural network.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-robot cooperative exploration method based on relation-gated graph neural networks, characterized in that, include: Acquire observational information about the environment to be explored, construct an environmental map, and extract a set of candidate exploration target points; Based on the robot set and the candidate exploration target point set, a topological graph is constructed, which includes robot nodes, target point nodes, and multi-type relation edges used to represent different semantic relationships. The nodes and relation edges are then feature-encoded to obtain a graph structure representation. The graph structure is used to represent the input relation-gated graph neural network. Different relation information is weighted and fused through a gating mechanism to obtain the fused feature representation of the nodes. Based on the fused feature representation, a target allocation result between the robot and the target point is generated, and each robot is controlled to perform the exploration task of the allocated target point according to the result; The multi-type relationship edges include: the association relationship edge between the robot and the target point, the cooperation relationship edge between robots, and the spatial adjacency relationship edge between target points; After completing the edge-level feature transformation under different relation types, the node features in the current graph are first processed. Relationship-level features Perform global aggregation to obtain graph state information that can represent the current exploration state. ,in, This represents the layer number of the graph neural network; the graph state information comprehensively reflects the spatial distribution, target association, and environmental exploration structure of the current multi-robot system, and serves as the input to the gating network. Expressed as a formula: ； in, and Let these represent the set of nodes and the set of edges in the graph, respectively. This represents a global aggregation function that summarizes the features of all nodes in the graph into a single global representation vector. i For nodes; Representing relations The edge features below; Graph state information based on the current exploration state Calculate the relationship gating coefficient The formula is as follows: ； in, It is the weight matrix of the gated network. It is a bias term. It is an activation function that represents a relation. The degree of importance.

2. The multi-robot cooperative exploration method based on relation-gated graph neural networks as described in claim 1, characterized in that, The candidate exploration target point is the frontier point, which is the boundary position between the known free region and the unknown region.

3. The multi-robot cooperative exploration method based on relation-gated graph neural networks as described in claim 1, characterized in that, The gating mechanism calculates the importance weights of different relationship types based on node features and relationship edge features using a gating function, and then performs weighted aggregation of adjacent node information from different relationships.

4. The multi-robot cooperative exploration method based on relation-gated graph neural networks as described in claim 1, characterized in that, The target assignment result is obtained by constructing the robot-target point affinity matrix and using the Sinkhorn algorithm to generate an approximate double-random matching matrix.

5. The multi-robot cooperative exploration method based on relation-gated graph neural networks as described in claim 1, characterized in that, The process of generating the target allocation result includes: Based on the robot node fusion features and target point node fusion features output by the relation-gated graph neural network, the affinity between each robot and each target point is calculated, and an affinity matrix is ​​constructed. The affinity matrix is ​​approximated by double random normalization using the Sinkhorn algorithm to obtain the probability matching matrix; Based on the probability matching matrix, a target point is assigned to each robot, forming a target assignment result.

6. The multi-robot cooperative exploration method based on relation-gated graph neural networks as described in claim 1, characterized in that, Also includes: After the robot performs the exploration task, the collaborative exploration effect is evaluated according to the comprehensive reward function, which includes information gain reward, exploration efficiency reward and path cost penalty. Among them, the information gain reward is calculated based on the weighted sum of the probability of the robot being assigned to the target point and the information gain of the target point; the exploration efficiency reward is calculated based on the increase in the area of ​​the explored region per unit time; and the path cost penalty is calculated based on the weighted sum of the probability of the robot being assigned to the target point and the path length from the robot to the target point.

7. A multi-robot cooperative exploration system based on relation-gated graph neural networks, characterized in that, include: The information acquisition module is configured to: acquire observation information of the environment to be explored, construct an environment map, and extract a set of candidate exploration target points; The feature encoding module is configured to: construct a topological graph containing robot nodes, target point nodes, and multi-type relation edges representing different semantic relationships based on the robot set and the candidate exploration target point set; and perform feature encoding on the nodes and relation edges to obtain a graph structure representation. The feature fusion module is configured to: input the graph structure representation into a relation-gated graph neural network, and perform weighted fusion of different relation information through a gating mechanism to obtain the fused feature representation of the nodes; The target allocation module is configured to: generate a target allocation result between the robot and the target point based on the fused feature representation, and control each robot to perform an exploration task on the allocated target point according to the result; The multi-type relationship edges include: the association relationship edge between the robot and the target point, the cooperation relationship edge between robots, and the spatial adjacency relationship edge between target points; After completing the edge-level feature transformation under different relation types, the node features in the current graph are first processed. Relationship-level features Perform global aggregation to obtain graph state information that can represent the current exploration state. ,in, This represents the layer number of the graph neural network; the graph state information comprehensively reflects the spatial distribution, target association, and environmental exploration structure of the current multi-robot system, and serves as the input to the gating network. Expressed as a formula: ； in, and Let these represent the set of nodes and the set of edges in the graph, respectively. This represents a global aggregation function that summarizes the features of all nodes in the graph into a global representation vector; i is a node. Representing relations The edge features below; Graph state information based on the current exploration state Calculate the relationship gating coefficient The formula is as follows: ； in, It is the weight matrix of the gated network. It is a bias term. It is an activation function that represents a relation. The degree of importance.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the multi-robot cooperative exploration method based on a relation-gated graph neural network as described in any one of claims 1-6.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the multi-robot cooperative exploration method based on relation-gated graph neural networks as described in any one of claims 1-6.

Images