Extensible multi-robot autonomous exploration method for intelligent information aggregation

Abstract

The invention discloses an extensible multi-robot autonomous exploration method for intelligent information aggregation in order to improve the cooperation efficiency of multi-robot exploration of an unknown environment and enhance the expandability of an exploration strategy. According to the method, graph representation is performed on an exploration environment, local information sharing is realized by utilizing edges in a graph, high-quality local neighborhood information is extracted by adopting a graph neural network and mutual information, observation information of a robot and high-quality domain aggregation information are aggregated, and effective exploration decisions can be made by the robots in different scales. Through actual verification, the method can be used as an important way for solving the problems of poor collaboration, poor expandability and the like of a multi-machine collaborative exploration strategy.

Description

Technical Field

[0001] The present invention relates to a multi-robot autonomous exploration field of an unknown environment, and in particular to an expandable multi-robot autonomous exploration method oriented to intelligent information aggregation. Background Art

[0002] With the rapid development of sensor measurement and wireless communication technology, multi-robot systems have shown great advantages in robustness, timeliness, economy and functionality. Therefore, multi-robot systems have been widely used in search and rescue, environmental monitoring, unknown environment mapping and planetary exploration. However, multi-robot collaborative exploration strategy is the core technology of the above applications and is not limited to the above applications. This strategy can maximize the collaborative advantages of multi-robot systems and plays an important role in expanding the coverage of robots in the environment and speeding up the creation of maps. It has great application potential in industry, military, surveying and other fields, and has important research value and significance.

[0003] At present, the research on multi-robot autonomous exploration of unknown environments can be roughly divided into non-learning-based methods and learning-based methods. Non-learning-based methods are less applicable when dealing with some challenging scenarios. Compared with non-learning methods, learning-based methods can be trained and learned in large amounts of data, and then discover potential patterns and connections in the data, which can better adapt to different application scenarios and problem requirements.

[0004] In recent years, deep learning can efficiently learn and extract features from complex data through deep neural networks, improving the feature representation and learning capabilities of intelligent systems; reinforcement learning can autonomously learn behaviors by setting reward functions and strategies to achieve adaptive exploration of the environment. Therefore, autonomous exploration algorithms based on deep reinforcement learning enable robots to have stronger sequential decision-making capabilities in unknown environments. A typical method is that the multi-robot collaborative exploration framework MAANS uses RGB images and posture data to build a centralized global map for each robot, applies CNN technology to extract spatial features, and determines the long-term goals of each robot through a hierarchical self-attention architecture, merges the maps of all robots, and creates a unified global map to guide the robot to complete path planning. However, this method does not take into account the limitations of the robot's sensor range and capabilities in real environments, and they can often only obtain local environmental information. In addition, the communication between robots is affected by factors such as bandwidth and communication protocols, and effective information sharing cannot be achieved. Finally, the number of robots may change dynamically depending on task requirements, environmental changes, or the availability of system resources (Yu C, Yang X, Gao J, et al. Learning efficient multi-agent cooperative visual exploration [C] / / European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 497-515.).

[0005] Therefore, to improve the efficiency of efficient collaborative exploration of multi-robot systems, two key scientific problems need to be solved urgently: one is how to achieve local information sharing between robots and extract the most valuable information from the local observations of the robots; the other is how to maintain the effectiveness of multi-robot collaborative exploration strategies when the number of robots is different.

[0006] To this end, representing the exploration environment as a graph, using the edges in the graph to share local information, and using graph neural networks and mutual information to extract high-quality local neighborhood information are important ways to solve problems such as poor collaboration and poor scalability of multi-machine collaborative exploration strategies. Summary of the invention

[0007] The purpose of the present invention is to propose a scalable multi-robot autonomous exploration method for intelligent information aggregation to address the problems that robots can only partially observe environmental information and cannot share effective information during multi-machine collaborative exploration, and the exploration strategy only considers a fixed number of robots.

[0008] The present invention proposes a scalable multi-robot autonomous exploration method for intelligent information aggregation with the goal of improving the collaborative efficiency of multi-machine exploration of unknown environments and enhancing the scalability of exploration strategies. By establishing a distance-based robot-robot graph, local information sharing is achieved by utilizing the edges in the graph, and high-quality local neighborhood information is extracted using graph neural networks and mutual information. The robots aggregate their own observation information with the high-quality local neighborhood information, which serves as a technical approach for robots of different scales to make effective exploration decisions.

[0009] The multi-robots include but are not limited to drones and unmanned vehicles.

[0010] The purpose of the present invention is achieved by at least one of the following technical solutions.

[0011] A scalable multi-robot autonomous exploration method for intelligent information aggregation includes the following steps:

[0012] S1. For each robot, a local map centered on a single robot is constructed, and the current posture information of the robot is estimated;

[0013] S2, converting the local map of each robot centered on a single robot into a global map centered on a single robot, and then obtaining a fused global map including all robots;

[0014] S3. Based on the global map centered on a single robot, a global goal, i.e., a long-term navigation goal, is generated for each robot through a scalable multi-robot information aggregation planning method;

[0015] S4, plan a path for each robot to achieve the long-term navigation goal in the fused global map and generate a sequence of short-term sub-goals;

[0016] S5. For each short-term sub-goal in the short-term sub-goal sequence of the robot, convert it into the relative distance and relative angle between the position of the short-term sub-goal and the current position of the robot, input it into the local strategy, and the local strategy outputs the navigation action of the robot.

[0017] The relative distance refers to the Euclidean distance between two points.

[0018] Further, step S3 includes the following steps:

[0019] S3.1, build Actor network (scalable multi-robot information aggregation planner) and Critic network;

[0020] S3.2, train the Actor network and the Critic network;

[0021] S3.3. Use the trained Actor network to generate a global goal for each robot.

[0022] Further, in step S3.1, the Actor network includes a first CNN feature extractor, a first information aggregation module and an action generator;

[0023] In the first CNN feature extractor, a weight-sharing convolutional neural network is applied as a feature extractor to extract the spatial feature map of each robot from the global map corresponding to each robot centered on a single robot;

[0024] In the first information aggregation module, the exploration environment of each robot is represented by a graph, and a distance-based robot-robot graph is established for each robot to obtain the local neighborhood information of each robot;

[0025] A graph neural network is used to aggregate the local neighborhood information of each robot, and the attention mechanism is used to prioritize the information of neighboring robots. At the same time, the graph mutual information is used to maximize the correlation of the input and output feature information of the robots, and a fixed-size vector is obtained, that is, the neighborhood aggregation information of each robot.

[0026] Connect the neighborhood aggregation information of each robot with the spatial feature map to obtain the fused spatial feature map of each robot [x i ,x′ i ], in the action generator, the Softmax function is used to first select an area from the fusion spatial feature map of each robot, and then the Sigmoid function is used to output a coordinate to obtain the relative position of the global target in the selected area.

[0027] Furthermore, the global map corresponding to each robot, centered on a single robot, includes the robot's current position, motion trajectory, previous goals, goal history, occupancy map, and obstacle map;

[0028] The global map centered on a single robot corresponding to each robot is transformed into the same global coordinate system, and a convolutional neural network is used to generate an H×H×B feature map; H corresponds to the discretization level of the scene, and B is the number of channels.

[0029] Furthermore, the exploration environment is represented by a graph, and a robot-robot graph G(i) of robot i is constructed, G(i)∈G(V,E), where node v∈V is a neighboring robot of robot i;

[0030] If robot i can sense other robots within a limited radius ρ, then the robot sensed by robot i is the neighboring robot of robot i; then there is an edge e∈E between robot i and the sensed robot, and messages can be transmitted bidirectionally between robot i and the sensed robot, and the edge feature is the Euclidean distance between robot i and the sensed robot;

[0031] All robot features are X M×D =(x1,......,x M ), assuming an empirical probability distribution P, where M represents the number of robots, i = 1 ~ M, D represents the length of the feature vector, x i represents the characteristics of robot i, that is, the global map centered on robot i; the characteristics of the neighboring robots of robot i are defined as X i , of size (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; create a matrix C of shape N(i)+1×N(i)+1 for robot i i , matrix C i The element c ij represents the distance between robot i and neighboring robot j; according to the definition of G(i), the neighborhood aggregation information of robot i is represented as x′ i =f(G i )=f(X i ,C i ).

[0032] Furthermore, in the first information aggregation module, a graph neural network with a message passing framework is used to infer the neighborhood aggregation information of robot i; the network adopts a unified message passing model, where each layer update is defined as;

[0033]

[0034] Among them, x′ i is the neighborhood aggregation information of robot i, x i and x j are the features of robot i and the jth neighboring robot of robot i, N(i) is the number of neighboring robots connected to robot i, W1 and W2 are learnable weight matrices;

[0035] The attention mechanism is used to prioritize the information of neighboring robots as follows:

[0036] α i,j is the attention coefficient between node i and node j, W2, W3 and W4 are learnable weight matrices;

[0037]

[0038] Among them, e ij is the feature of the edge connecting robot i and the jth neighboring robot of robot i, and c is the output dimension of this layer;

[0039] The size of the attention coefficient represents the importance of the neighboring robots, and the neighboring robots of robot i are prioritized according to the attention coefficient.

[0040] Furthermore, the graph mutual information is the feature X of the neighboring robots of robot i i The empirical probability distribution of i ), the neighborhood aggregate information x′ of robot i i The probability distribution of i ), the joint distribution is expressed as P(x′ i |X i ); if the conditional probability P(x′ i |X i ) are multiplied, then the global mutual information I(x′) i |X i ) can be decomposed into a weighted sum of local mutual information as follows;

[0041]

[0042] Among them, x j represents the characteristics of the jth neighboring robot of robot i, N(i) represents the number of neighboring robots of robot i, and the weight z of the jth neighboring robot of robot i ij Satisfies 1 / N(i)≤z ij ≤1;

[0043] Inspired by the decomposition of global mutual information into local mutual information, trainable weights can be constructed from the topological view of the graph; w ij The weights can be used to capture the internal relationships of the robots; thus, the neighborhood aggregate information x′ i With Figure G i =(X i ,C i ) is represented by the graph mutual information between:

[0044]

[0045] in, c ij is the adjacency matrix C i The elements in , σ(·) is the Sigmod function; the first term of this formula is used to measure the local mutual information I(x′ i |x j ) for the global mutual information I(x′) i |G i )’s contribution; I(x′i |x j ) is contributed by the robot feature representation x′ i and x j The second term represents z ij The mutual information between the robot and the robot is obtained by maximizing I(z ij ;c ij ), weight z ij It is used to constrain the topological relationship between robots; that is, when the jth neighboring robot of robot i is "closer" to robot i, z ij Bigger; This item makes up for the disadvantage of focusing only on robot features;

[0046] Finally, the neighborhood aggregate information of each robot i is x′ i .

[0047] Further, in step S3.1, the Critic network includes a second CNN feature extractor and a second information aggregation module and a linear layer;

[0048] The second CNN feature extractor has the same structure as the first CNN feature extractor; the second information aggregation module has the same structure as the first information aggregation module, but the output of the second information aggregation module is different from that of the first information aggregation module;

[0049] In the second information aggregation module, the neighborhood aggregation information of all robots is globally averaged and pooled to obtain the global information vector X agg , that is, the graph aggregation information of the robot, This vector is independent of the number of robots and is input into the linear layer in the Critic network to obtain the value of the state-action pair.

[0050] Furthermore, in step S3.2, Multi-Agent Proximal Policy Optimization (MAPPO) is used to train the Actor network and the Critic network.

[0051] The training data for the Actor and Critic networks is usually obtained by the robot interacting with the environment, that is, by performing actions in the environment and observing the results. This data is usually stored in the form of experience replay and used to train the network.

[0052] Furthermore, in step S5, the local strategy module in the multi-robot collaborative exploration framework MAANS is used to output the navigation action of the robot according to the relative distance and relative angle between the position of the input short-term sub-goal and the current robot position.

[0053] The beneficial effects of the technical solution provided by the present invention are:

[0054] By establishing a distance-based robot-robot graph, using the edges in the graph to share local information, and using graph neural networks and mutual information to extract high-quality local neighborhood information, the robot aggregates its own observation information with high-quality local neighborhood information to make effective exploration decisions. This will enhance the robot's understanding and cognition of the entire exploration area, improve the performance of collaborative exploration strategies, and enhance the flexibility, adaptability, and scalability of exploration strategies.

[0055] Improve the collaborative efficiency of multiple machines exploring unknown environments: Using the attention mechanism to prioritize the information of neighboring robots can distinguish the importance of different neighborhood information, help robots perceive unknown environments more carefully, and maximize the correlation of the input and output feature information of robots through graph mutual information, which is beneficial for robots to autonomously learn collaborative strategies and avoid repeated exploration.

[0056] Enhance the scalability of the exploration strategy: After the exploration environment is represented by a graph, the exploration strategy can adapt well to situations with different numbers of robots. At the same time, it only aggregates the local neighborhood information of the robots from the graph and extracts the information of the best value, reducing the increase in the complexity of training samples to improve the exploration efficiency and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of a flow chart of an expandable multi-robot autonomous exploration method for intelligent information aggregation in an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the structure of the Actor network in an embodiment of the present invention;

[0059] Figure 3 Schematic diagram of convolutional neural network hyperparameters in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present invention clearer, the specific implementation of the present invention will be described in detail below in conjunction with the accompanying drawings and embodiments. The embodiments described with reference to the accompanying drawings are exemplary and are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0061] The following describes a scalable multi-robot autonomous exploration method for information aggregation according to an embodiment of the present invention with reference to the accompanying drawings.

[0062] A scalable multi-robot autonomous exploration method for intelligent information aggregation, such as Figure 1 As shown, the following steps are included:

[0063] S1. For each robot, a local map centered on a single robot is constructed, and the current posture information of the robot is estimated;

[0064] S2, converting the local map of each robot centered on a single robot into a global map centered on a single robot, and then obtaining a fused global map including all robots;

[0065] S3. Based on the global map centered on a single robot, a global goal, i.e., a long-term navigation goal, is generated for each robot through a scalable multi-robot information aggregation planning method, including the following steps:

[0066] S3.1, build Actor network (scalable multi-robot information aggregation planner) and Critic network;

[0067] The Actor network includes a first CNN feature extractor, a first information aggregation module and an action generator;

[0068] In the first CNN feature extractor, a weight-sharing convolutional neural network is applied as a feature extractor to extract the spatial feature map of each robot from the global map corresponding to each robot centered on a single robot;

[0069] In the first information aggregation module, the exploration environment of each robot is represented by a graph, and a distance-based robot-robot graph is established for each robot to obtain the local neighborhood information of each robot;

[0070] A graph neural network is used to aggregate the local neighborhood information of each robot, and the attention mechanism is used to prioritize the information of neighboring robots. At the same time, the graph mutual information is used to maximize the correlation of the input and output feature information of the robots, and a fixed-size vector is obtained, that is, the neighborhood aggregation information of each robot.

[0071] Connect the neighborhood aggregation information of each robot with the spatial feature map to obtain the fused spatial feature map of each robot [x i ,x′ i ], in the action generator, the Softmax function is used to first select an area from the fusion spatial feature map of each robot, and then the Sigmoid function is used to output a coordinate to obtain the relative position of the global target in the selected area.

[0072] The global map centered on a single robot corresponding to each robot includes the robot's current position, motion trajectory, previous goals, goal history, occupancy map, and obstacle map;

[0073] In one embodiment, the size of the global map centered on a single robot is 240x240x6; 6 refers to 6 channels, each channel is the robot's current position, motion trajectory, previous goals, goal history, occupancy map and obstacle map.

[0074] The global map centered on a single robot corresponding to each robot is transformed into the same global coordinate system, and a convolutional neural network is used to generate an H×H×B feature map; H corresponds to the discretization level of the scene, and B is the number of channels.

[0075] In one embodiment, the convolutional neural network used includes 5 convolutional layers and a maximum pooling layer, and the parameters of each network layer are as follows: Figure 3 shown.

[0076] The exploration environment is represented by a graph, and a robot-robot graph G(i) of robot i is constructed, G(i)∈G(V,E), where node v∈V is a neighboring robot of robot i;

[0077] If robot i can sense other robots within a limited radius ρ, then the robot sensed by robot i is the neighboring robot of robot i; then there is an edge e∈E between robot i and the sensed robot, and messages can be transmitted bidirectionally between robot i and the sensed robot, and the edge feature is the Euclidean distance between robot i and the sensed robot;

[0078] All robot features are X M×D =(x1,......,x M ), assuming an empirical probability distribution P, where M represents the number of robots, i = 1 ~ M, D represents the length of the feature vector, x i represents the characteristics of robot i, that is, the global map centered on robot i; the characteristics of the neighboring robots of robot i are defined as X i , of size (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; create a matrix C of shape N(i)+1×N(i)+1 for robot i i , matrix C i The element c ij represents the distance between robot i and neighboring robot j; according to the definition of G(i), the neighborhood aggregation information of robot i is represented as x′ i =f(G i )=f(X i ,C i ).

[0079] In the first information aggregation module, a graph neural network with a message passing framework is used to infer the neighborhood aggregation information of robot i; the network adopts a unified message passing model, where each layer update is defined as;

[0080]

[0081] Among them, x′ i is the neighborhood aggregation information of robot i, x i and x j are the features of robot i and the jth neighboring robot of robot i, N(i) is the number of neighboring robots connected to robot i, W1 and W2 are learnable weight matrices;

[0082] The attention mechanism is used to prioritize the information of neighboring robots as follows:

[0083] α i,j is the attention coefficient between node i and node j, W2, W3 and W4 are learnable weight matrices;

[0084]

[0085] Among them, e ij is the feature of the edge connecting robot i and the jth neighboring robot of robot i, and c is the output dimension of this layer;

[0086] The size of the attention coefficient represents the importance of the neighboring robots, and the neighboring robots of robot i are prioritized according to the attention coefficient.

[0087] The graph mutual information is the feature X of the neighboring robots of robot i. i The empirical probability distribution of i ), the neighborhood aggregate information x′ of robot i i The probability distribution of i ), the joint distribution is expressed as P(x′ i |X i ); if the conditional probability P(x′ i |X i ) are multiplied, then the global mutual information I(x′) i |X i ) can be decomposed into a weighted sum of local mutual information as follows;

[0088]

[0089] Among them, x j represents the characteristics of the jth neighboring robot of robot i, N(i) represents the number of neighboring robots of robot i, and the weight z of the jth neighboring robot of robot i ij Satisfies 1 / N(i)≤z ij ≤1;

[0090] Inspired by the decomposition of global mutual information into local mutual information, trainable weights can be constructed from the topological view of the graph; w ij The weights can be used to capture the internal relationships of the robots; thus, the neighborhood aggregate information x′ i With Figure G i =(X i ,C i ) is represented by the graph mutual information between:

[0091]

[0092] in, c ij is the adjacency matrix C i The elements in , σ(·) is the Sigmod function; the first term of this formula is used to measure the local mutual information I(x′ i |x j ) for the global mutual information I(x′) i |G i )’s contribution; I(x′ i |x j ) is contributed by the robot feature representation x′ i and x j The second term represents z ij The mutual information between the robot and the robot is obtained by maximizing I(z ij ;c ij ), weight z ij It is used to constrain the topological relationship between robots; that is, when the jth neighboring robot of robot i is "closer" to robot i, z ij Bigger; This item makes up for the disadvantage of focusing only on robot features;

[0093] Finally, the neighborhood aggregate information of each robot i is x′ i .

[0094] The Critic network includes a second CNN feature extractor, a second information aggregation module and a linear layer;

[0095] The second CNN feature extractor has the same structure as the first CNN feature extractor; the second information aggregation module has the same structure as the first information aggregation module, but the output of the second information aggregation module is different from that of the first information aggregation module;

[0096] In the second information aggregation module, the neighborhood aggregation information of all robots is globally averaged and pooled to obtain the global information vector X agg , that is, the graph aggregation information of the robot, This vector is independent of the number of robots and is input into the linear layer in the Critic network to obtain the value of the state-action pair.

[0097] S3.2, train the Actor network and the Critic network;

[0098] S3.3. Use the trained Actor network to generate a global goal for each robot.

[0099] In one embodiment, Multi-Agent Proximal Policy Optimization (MAPPO) is used to train the Actor network and the Critic network.

[0100] The training data for the Actor and Critic networks is usually obtained by the robot interacting with the environment, that is, by performing actions in the environment and observing the results. This data is usually stored in the form of experience replay and used to train the network.

[0101] S4. In one embodiment, a fast marching method (FMM) is used to plan a path for each robot to achieve a long-term navigation goal in the fused global map, and a short-term sub-goal sequence is generated;

[0102] S5. In one embodiment, for a short-term sub-goal in the short-term sub-goal sequence of each robot, it is converted into the relative distance and relative angle between the position of the short-term sub-goal and the current position of the robot, and the local strategy module in the multi-robot collaborative exploration framework MAANS. (Yu C, Yang X, Gao J, et al. Learning efficient multi-agent cooperative visual exploration [C] / / European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 497-515.) is used to output the navigation action of the robot according to the relative distance and relative angle between the input position of the short-term sub-goal and the current position of the robot.

[0103] Matters not covered by the present invention are known technologies.

[0104] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A scalable multi-robot autonomous exploration method for intelligent information aggregation, characterized in that: The steps include: S1. For each robot, a local map centered on a single robot is constructed, and the current posture information of the robot is estimated; S2, converting the local map of each robot centered on a single robot into a global map centered on a single robot, and then obtaining a fused global map including all robots; S3. Based on the global map centered on a single robot, a global goal, i.e., a long-term navigation goal, is generated for each robot through a scalable multi-robot information aggregation planning method; S4, plan a path for each robot to achieve the long-term navigation goal in the fused global map and generate a sequence of short-term sub-goals; S5. For each short-term sub-goal in the short-term sub-goal sequence of the robot, convert it into the relative distance and relative angle between the position of the short-term sub-goal and the current position of the robot, input it into the local strategy, and the local strategy outputs the navigation action of the robot.

2. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 1, characterized in that: Step S3 includes the following steps: S3.

1. Build Actor network and Critic network; S3.2, train the Actor network and the Critic network; S3.

3. Use the trained Actor network to generate a global goal for each robot.

3. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 2 is characterized in that: In step S3.1, the Actor network includes a first CNN feature extractor, a first information aggregation module and an action generator; In the first CNN feature extractor, a weight-sharing convolutional neural network is applied as a feature extractor to extract the spatial feature map of each robot from the global map corresponding to each robot centered on a single robot; In the first information aggregation module, the exploration environment of each robot is represented by a graph, and a distance-based robot-robot graph is established for each robot to obtain the local neighborhood information of each robot; A graph neural network is used to aggregate the local neighborhood information of each robot, and the attention mechanism is used to prioritize the information of neighboring robots. At the same time, the graph mutual information is used to maximize the correlation of the input and output feature information of the robots, and a fixed-size vector is obtained, that is, the neighborhood aggregation information of each robot. Connect the neighborhood aggregation information of each robot with the spatial feature map to obtain the fused spatial feature map of each robot [x i ,x i ′], in the action generator, the Softmax function is used to first select an area from the fusion spatial feature map of each robot, and then the Sigmoid function is used to output a coordinate to obtain the relative position of the global target in the selected area.

4. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 3 is characterized in that: The global map centered on a single robot corresponding to each robot includes the robot's current position, motion trajectory, previous goals, goal history, occupancy map, and obstacle map; The global map centered on a single robot corresponding to each robot is transformed into the same global coordinate system, and a convolutional neural network is used to generate an H×H×B feature map; H corresponds to the discretization level of the scene, and B is the number of channels.

5. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 3, characterized in that: The exploration environment is represented by a graph, and a robot-robot graph G(i) of robot i is constructed, G(i)∈G(V,E), where node v∈V is a neighboring robot of robot i; If robot i can sense other robots within a limited radius ρ, then the robot sensed by robot i is the neighboring robot of robot i; then there is an edge e∈E between robot i and the sensed robot, and messages can be transmitted bidirectionally between robot i and the sensed robot, and the edge feature is the Euclidean distance between robot i and the sensed robot; All robot features are X M×D =(x1,......,x M ), assuming an empirical probability distribution P, where M represents the number of robots, i = 1 ~ M, D represents the length of the feature vector, x i represents the characteristics of robot i, that is, the global map centered on robot i; the characteristics of the neighboring robots of robot i are defined as X i , with a size of (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; create a matrix C of shape N(i)+1×N(i)+1 for robot i i , matrix C i The element c ij represents the distance between robot i and neighboring robot j; according to the definition of G(i), the neighborhood aggregation information of robot i is expressed as x i ′=f(G i )=f(X i ,C i ).

6. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 5, characterized in that: In the first information aggregation module, a graph neural network with a message passing framework is used to infer the neighborhood aggregation information of robot i; the network adopts a unified message passing model, where each layer update is defined as; Among them, x′ i is the neighborhood aggregation information of robot i, x i and x j are the features of robot i and the jth neighboring robot of robot i, N(i) is the number of neighboring robots connected to robot i, W1 and W2 are learnable weight matrices; The attention mechanism is used to prioritize the information of neighboring robots as follows: α i,j is the attention coefficient between node i and node j, W2, W3 and W4 are learnable weight matrices; Among them, e ij is the feature of the edge connecting robot i and the jth neighboring robot of robot i, and c is the output dimension of this layer; The size of the attention coefficient represents the importance of the neighboring robots, and the neighboring robots of robot i are prioritized according to the attention coefficient.

7. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 6, characterized in that: The graph mutual information is the feature X of the neighboring robots of robot i. i The empirical probability distribution of i ), the neighborhood aggregation information x of robot i i The probability distribution of ′ is expressed as P(x i ′), the joint distribution is expressed as P(x i ′|X i ); if the conditional probability P(x i ′|X i ) are multiplied, then the global mutual information I(x i ′|X i ) can be decomposed into a weighted sum of local mutual information as follows; Among them, x j represents the characteristics of the jth neighboring robot of robot i, N(i) represents the number of neighboring robots of robot i, and the weight z of the jth neighboring robot of robot i ij Satisfies 1 / N(i)≤z ij ≤1; Therefore, the neighborhood aggregation information x i ′ and graph G i =(X i ,C i ) is represented as follows: in, c ij is the adjacency matrix C i The first term of this formula is used to measure the local mutual information I(x i ′|x j ) for the global mutual information I(x i ′|G i )’s contribution; I(x i ′|x j ) is contributed by the robot feature representation x i ′ and x j The second term represents z ij The mutual information between the robot and the robot is obtained by maximizing I(z ij ;c ij ), weight z ij It is used to constrain the topological relationship between robots; that is, when the jth neighboring robot of robot i is "closer" to robot i, z ij Bigger; This item makes up for the disadvantage of focusing only on robot features; Finally, the neighborhood aggregate information of each robot i is x i ′.

8. A scalable multi-robot autonomous exploration method for intelligent information aggregation according to any one of claims 3 to 7, characterized in that: In step S3.1, the Critic network includes a second CNN feature extractor and a second information aggregation module and a linear layer; The second CNN feature extractor has the same structure as the first CNN feature extractor; the second information aggregation module has the same structure as the first information aggregation module, but the output of the second information aggregation module is different from that of the first information aggregation module; In the second information aggregation module, the neighborhood aggregation information of all robots is globally averaged and pooled to obtain the global information vector X agg , that is, the graph aggregation information of the robot, This vector is independent of the number of robots and is input into the linear layer in the Critic network to obtain the value of the state-action pair.

9. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 2, characterized in that: In step S3.2, the multi-agent proximal strategy is used to optimize the training of the Actor network and the Critic network.

10. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 1, characterized in that: In step S5, the local strategy module in the multi-robot collaborative exploration framework MAANS is used to output the navigation action of the robot according to the relative distance and relative angle between the input short-term sub-goal position and the current robot position.

Images