An extensible multi-robot autonomous exploration method for intelligent information aggregation

Abstract

In order to improve the cooperation efficiency of multi-robot exploration in unknown environment and enhance the scalability of exploration strategy, the application discloses an intelligent information aggregation oriented scalable multi-robot autonomous exploration method. The method realizes local information sharing by using the edges in the graph, extracts high-quality local neighborhood information by using the graph neural network and mutual information, aggregates the observation information of the robot and the high-quality field aggregation information, and realizes that the robot can make effective exploration decisions under different scales. Through actual verification, the application can be used as an important way to solve the problems of poor cooperation and scalability of multi-robot collaborative exploration strategy.

Description

Technical Field

[0001] This invention relates to the field of autonomous multi-robot exploration of unknown environments, and specifically to a scalable autonomous multi-robot exploration method oriented towards intelligent information aggregation. Background Technology

[0002] With the rapid development of sensing and wireless communication technologies, multi-robot systems have demonstrated significant advantages in robustness, timeliness, economy, and functionality. Therefore, multi-robot systems have been widely applied in fields such as search and rescue, environmental monitoring, mapping of unknown environments, and planetary exploration. However, multi-robot collaborative exploration strategies are a core technology for these applications, but not limited to them. This strategy can maximize the collaborative advantages of multi-robot systems, playing a crucial role in expanding the robot's coverage area in the environment and accelerating map creation. It has enormous application potential in industry, military, and surveying fields, and possesses significant research value and significance.

[0003] Current research on multi-robot autonomous exploration of unknown environments can be broadly categorized into non-learning-based methods and learning-based methods. Non-learning-based methods have limited applicability in challenging scenarios. In contrast, learning-based methods can be trained and learned from large amounts of data, uncovering underlying patterns and connections, and thus better adapting to different application scenarios and problem requirements.

[0004] In recent years, deep learning has enabled robots to efficiently learn and extract features from complex data through deep neural networks, improving the feature representation and learning capabilities of intelligent systems. Reinforcement learning, on the other hand, allows robots to autonomously learn behaviors by setting reward functions and policies, achieving 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 approach is the multi-robot collaborative exploration framework MAANS, which uses RGB images and pose data to construct 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. Finally, it merges the maps of all robots to create a unified global map, guiding the robots in path planning. However, this method does not consider the limitations of the robot's sensor range and capabilities in real-world environments, where they often only acquire local environmental information. Furthermore, communication between robots is affected by factors such as bandwidth and communication protocols, hindering effective information sharing. Finally, the number of robots may vary 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, improving the efficiency of multi-robot collaborative exploration urgently requires solving two key scientific problems: first, how to achieve local information sharing among robots and extract the most valuable information from the robots' local observations; and second, how to maintain the effectiveness of multi-robot collaborative exploration strategies when the number of robots is different.

[0006] Therefore, representing the exploration environment as a graph, using the edges in the graph to achieve local information sharing, and employing graph neural networks and mutual information to extract high-quality local neighborhood information are important ways to solve the problems of poor cooperation and poor scalability of multi-machine collaborative exploration strategies. Summary of the Invention

[0007] The purpose of this invention is to address the problems that robots can only partially observe environmental information and cannot effectively share information during multi-robot collaborative exploration, and that exploration strategies only consider a fixed number of robots. This invention proposes a scalable multi-robot autonomous exploration method oriented towards intelligent information aggregation.

[0008] This invention proposes a scalable multi-robot autonomous exploration method oriented towards intelligent information aggregation, aiming to improve the collaborative efficiency of multi-robot exploration of unknown environments and enhance the scalability of exploration strategies. By establishing a distance-based robot-robot graph, local information sharing is achieved using the edges in the graph. High-quality local neighborhood information is extracted using graph neural networks and mutual information. The robot aggregates its own observation information with the high-quality local neighborhood information, serving as a technical approach for robots of different scales to make effective exploration decisions.

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

[0010] The objective of this 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, construct a local map centered on the single robot, and estimate the robot's current posture information;

[0013] S2. Convert the local map of each robot centered on a single robot into a global map centered on a single robot, thereby obtaining a fused global map that includes all robots;

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

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

[0016] S5. For each robot's short-term sub-objective sequence, convert the short-term sub-objective's position into the relative distance and relative angle between the short-term sub-objective's position and the current robot position, input the relative distance and relative angle into the local policy, and the local policy outputs the robot's navigation action.

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

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

[0019] S3.1 Construct the Actor network (an extensible multi-robot information aggregation planner) and the Critic network;

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

[0021] S3.3. Use the trained Actor network to generate a global target 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-shared convolutional neural network is used as a feature extractor to extract the spatial feature map of each robot from the global map 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 built 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 an attention mechanism is used to prioritize the information of neighboring robots. At the same time, graph mutual information is used to maximize the correlation of the input and output feature information of the robots, resulting in a fixed-size vector, which is the neighborhood aggregated information of each robot.

[0026] By linking the neighborhood aggregation information of each robot with the spatial feature map, a fused spatial feature map of each robot is obtained [x]. i ,x′ i In the motion generator, the Softmax function is used to select a region from the fused 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 within the selected region.

[0027] Furthermore, the global map corresponding to each robot, centered on the single robot, includes the robot's current position, movement trajectory, previous targets, target history, occupied map, and obstacle map;

[0028] The global map corresponding to each robot, centered on a single robot, is transformed to the same global coordinate system, and an H×H×B feature map is generated using a convolutional neural network; 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, where G(i)∈G(V,E), and node v∈V is a neighboring robot of robot i.

[0030] If robot i can perceive other robots within a finite radius ρ, then the robots perceived by robot i are its neighboring robots; then there is an edge e∈E between robot i and the perceived robots, and messages can be transmitted bidirectionally between robot i and the perceived robots, with the edge feature being the Euclidean distance between robot i and the perceived robots.

[0031] All robot characteristics are X M×D =(x1,......,x M Assume an empirical probability distribution P, where M represents the number of robots, i = 1 to M, D represents the length of the eigenvector, and x... i The characteristics of robot i are represented by the global map centered on robot i; the characteristics of neighboring robots of robot i are defined as X. i The matrix C is of size (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; a matrix C of shape N(i)+1×N(i)+1 is created for robot i. i Matrix C i element c ij Let x' represent the distance between robot i and its 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] Where, x′ i For the neighborhood aggregation information of robot i, x i and x j Let N(i) be the features of robot i and the j-th neighboring robot of robot i, where N(i) is the number of neighboring robots connected to robot i, and W1 and W2 are learnable weight matrices.

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

[0036] α i,j W1 is the attention coefficient between node i and node j, and W2, W3, and W4 are learnable weight matrices.

[0037]

[0038] Among them, e ij Let c be the feature of the edge connecting robot i and the j-th neighboring robot of robot i, and let c be the output dimension of this layer;

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

[0040] Furthermore, the graphical mutual information will include the features X of robot i's neighboring robots. i The empirical probability distribution is represented as P(X) i ), the neighborhood aggregation information x′ of robot i i The probability distribution is represented as P(x′) i The joint distribution is represented as P(x′). i |X i If the conditional probability P(x′) i |X i If x' and x'' are multiplied, then the global mutual information I(x'') is... i |X i It can be decomposed into a weighted sum of local mutual information, as follows;

[0041]

[0042] Where, x j Let z represent the characteristics of the j-th neighboring robot of robot i, N(i) represent the number of neighboring robots of robot i, and z represent the weight of the j-th neighboring robot of robot i. ij Satisfying 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 Weights can be used to capture the robot's intrinsic relationships; therefore, neighborhood aggregation information x′ i With Figure G i =(X i C i The graphical mutual information between ) is represented as:

[0044]

[0045] in, c ij Let C be the adjacency matrix. i The elements in the expression, σ(·), are the Sigmoid functions; the first term of the expression measures the local mutual information I(x′). i |x j ) for global mutual information I(x′) i |G i The contribution of I(x′);i |x j The contribution of ) is through the robot feature representation x′ i and x j The similarity between them is used for calculation; the second term represents z. ij Mutual information in the relationship between the robot and the robot; by maximizing I(z) ij c ij ), weight z ij This constraint is used to define the topological relationships between robots; that is, when the j-th neighboring robot of robot i is "closer" to robot i, z... ij Larger; this overcomes the shortcoming of focusing only on robot features;

[0046] Finally, the neighborhood aggregation information for each robot i is x′. i .

[0047] Further, in step S3.1, the Critic network includes a second CNN feature extractor, 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 outputs of the second information aggregation module and the first information aggregation module are different.

[0049] In the second information aggregation module, the neighborhood aggregation information of all robots is subjected to a global average pooling operation to obtain the global information vector X. agg That is, the robot's graph aggregation information. This vector is independent of the number of robots and is fed into a linear layer in the Critic network to obtain the value of the state-action pair.

[0050] Furthermore, in step S3.2, the Actor network and Critic network are trained using Multi-Agent Proximal Policy Optimization (MAPPO).

[0051] Training data for Actor and Critic networks is typically obtained through robot interactions with their environment, i.e., by performing actions in the environment and observing the results. This data is usually stored in the form of experience replays and used to train the network.

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

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

[0054] By constructing a distance-based robot-robot graph and utilizing the edges within the graph to achieve local information sharing, high-quality local neighborhood information is extracted using graph neural networks and mutual information. The robot then aggregates its own observation information with this high-quality local neighborhood information to make effective exploration decisions. This enhances the robot's understanding and cognition of the entire exploration area, improves the performance of collaborative exploration strategies, and strengthens the flexibility, adaptability, and scalability of the exploration strategies.

[0055] Improving the collaborative efficiency of multi-robot exploration of unknown environments: Prioritizing information from neighboring robots using attention mechanisms can distinguish the importance of different neighboring information, helping robots to perceive the unknown environment more carefully. Maximizing the correlation of input and output feature information of robots through graph mutual information is beneficial for robots to learn collaborative strategies autonomously and avoid repeated exploration.

[0056] Enhancing the scalability of exploration strategies: After the exploration environment is represented graphically, the exploration strategy can adapt well to different numbers of robots. At the same time, it only aggregates the local neighborhood information of robots in the graph and extracts the information with the best value, reducing the increase in training sample complexity and thus improving exploration efficiency and scalability. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a scalable 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 Actor network structure in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the hyperparameters of the convolutional neural network in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. The embodiments described with reference to the accompanying drawings are exemplary and used to illustrate the present invention, but are not intended to limit the scope of the present invention.

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

[0062] A scalable multi-robot autonomous exploration method oriented towards intelligent information aggregation, such as Figure 1 As shown, it includes the following steps:

[0063] S1. For each robot, construct a local map centered on the single robot, and estimate the robot's current posture information;

[0064] S2. Convert the local map of each robot centered on a single robot into a global map centered on a single robot, thereby obtaining a fused global map that includes 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 using a scalable multi-robot information aggregation planning method, including the following steps:

[0066] S3.1 Construct the Actor network (an extensible multi-robot information aggregation planner) and the 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-shared convolutional neural network is used as a feature extractor to extract the spatial feature map of each robot from the global map 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 built 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 an attention mechanism is used to prioritize the information of neighboring robots. At the same time, graph mutual information is used to maximize the correlation of the input and output feature information of the robots, resulting in a fixed-size vector, which is the neighborhood aggregated information of each robot.

[0071] By linking the neighborhood aggregation information of each robot with the spatial feature map, a fused spatial feature map of each robot is obtained [x]. i ,x′ i In the motion generator, the Softmax function is used to select a region from the fused 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 within the selected region.

[0072] Each robot's corresponding global map, centered on the single robot, includes the robot's current position, movement trajectory, previous targets, target history, occupied map, and obstacle map;

[0073] In one embodiment, the global map size centered on a single robot is 240x240x6; the 6 refers to 6 channels, each channel showing the robot's current position, movement trajectory, previous targets, target history, occupied map, and obstacle map.

[0074] The global map corresponding to each robot, centered on a single robot, is transformed to the same global coordinate system, and an H×H×B feature map is generated using a convolutional neural network; 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 max-pooling layer, and the parameters of each network layer are as follows: Figure 3 As shown.

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

[0077] If robot i can perceive other robots within a finite radius ρ, then the robots perceived by robot i are its neighboring robots; then there is an edge e∈E between robot i and the perceived robots, and messages can be transmitted bidirectionally between robot i and the perceived robots, with the edge feature being the Euclidean distance between robot i and the perceived robots.

[0078] All robot characteristics are X M×D =(x1,......,x M Assume an empirical probability distribution P, where M represents the number of robots, i = 1 to M, D represents the length of the eigenvector, and x... i The characteristics of robot i are represented by the global map centered on robot i; the characteristics of neighboring robots of robot i are defined as X. i The matrix C is of size (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; a matrix C of shape N(i)+1×N(i)+1 is created for robot i. i Matrix C i element c ij Let x' represent the distance between robot i and its 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] Where, x′ i For the neighborhood aggregation information of robot i, x i and x j Let N(i) be the features of robot i and the j-th neighboring robot of robot i, where N(i) is the number of neighboring robots connected to robot i, and W1 and W2 are learnable weight matrices.

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

[0083] α i,j W1 is the attention coefficient between node i and node j, and W2, W3, and W4 are learnable weight matrices.

[0084]

[0085] Among them, e ij Let c be the feature of the edge connecting robot i and the j-th neighboring robot of robot i, and let c be the output dimension of this layer;

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

[0087] The graphical mutual information will include the features X of robot i's neighboring robots. i The empirical probability distribution is represented as P(X) i ), the neighborhood aggregation information x′ of robot i i The probability distribution is represented as P(x′) i The joint distribution is represented as P(x′). i |X i If the conditional probability P(x′) i |X i If x' and x'' are multiplied, then the global mutual information I(x'') is... i |X i It can be decomposed into a weighted sum of local mutual information, as follows;

[0088]

[0089] Where, x j Let z represent the characteristics of the j-th neighboring robot of robot i, N(i) represent the number of neighboring robots of robot i, and z represent the weight of the j-th neighboring robot of robot i. ij Satisfying 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 Weights can be used to capture the robot's intrinsic relationships; therefore, neighborhood aggregation information x′ i With Figure G i =(X i C i The graphical mutual information between ) is represented as:

[0091]

[0092] in, c ij Let C be the adjacency matrix. i The elements in the expression, σ(·), are the Sigmoid functions; the first term of the expression measures the local mutual information I(x′). i |x j ) for global mutual information I(x′) i |G i The contribution of I(x′); i |x j The contribution of ) is through the robot feature representation x′ i and x j The similarity between them is used for calculation; the second term represents z. ij Mutual information in the relationship between the robot and the robot; by maximizing I(z) ij c ij ), weight z ij This constraint is used to define the topological relationships between robots; that is, when the j-th neighboring robot of robot i is "closer" to robot i, z... ij Larger; this overcomes the shortcoming of focusing only on robot features;

[0093] Finally, the neighborhood aggregation information for 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 outputs of the second information aggregation module and the first information aggregation module are different.

[0096] In the second information aggregation module, the neighborhood aggregation information of all robots is subjected to a global average pooling operation to obtain the global information vector X. agg That is, the robot's graph aggregation information. This vector is independent of the number of robots and is fed into a linear layer in the Critic network to obtain the value of the state-action pair.

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

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

[0099] In one embodiment, the Actor network and Critic network are trained using Multi-Agent Proximal Policy Optimization (MAPPO).

[0100] Training data for Actor and Critic networks is typically obtained through robot interactions with their environment, i.e., by performing actions in the environment and observing the results. This data is usually stored in the form of experience replays and used to train the network.

[0101] S4. In one embodiment, the 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-target in the short-term sub-target sequence of each robot, the position of the short-term sub-target is converted into the relative distance and relative angle between the position of the short-term sub-target and the current position of the robot. The local policy module in the multi-robot cooperative 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 based on the input relative distance and relative angle between the position of the short-term sub-target and the current position of the robot.

[0103] Matters not covered in this invention are common knowledge.

[0104] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A scalable multi-robot autonomous exploration method for intelligent information aggregation, characterized in that, Includes the following steps: S1. For each robot, construct a local map centered on the single robot, and estimate the robot's current posture information; S2. Convert the local map of each robot centered on a single robot into a global map centered on a single robot, thereby obtaining a fused global map that includes 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 using a scalable multi-robot information aggregation planning method; this includes the following steps: S3.1 Construct an Actor network and a Critic network; 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-shared convolutional neural network is used as a feature extractor to extract the spatial feature map of each robot from the global map 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 built 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 an attention mechanism is used to prioritize the information of neighboring robots. At the same time, graph mutual information is used to maximize the correlation of the input and output feature information of the robots, resulting in a fixed-size vector, which is the neighborhood aggregated information of each robot. The neighborhood aggregation information of each robot is connected with the spatial feature map to obtain the fused spatial feature map of each robot. In the action generator, the Softmax function is used to select a region from the fused 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 region. S3.2 Training the Actor network and Critic network; S3.

3. Use the trained Actor network to generate a global target for each robot; S4. In the fused global map, plan a path for each robot to achieve the long-term navigation goal and generate a short-term sub-goal sequence; S5. For each robot's short-term sub-objective sequence, convert the short-term sub-objective's position into the relative distance and relative angle between the short-term sub-objective's position and the current robot position, input the relative distance and relative angle into the local policy, and the local policy outputs the robot's navigation action.

2. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 1, characterized in that, Each robot's corresponding global map, centered on the single robot, includes the robot's current position, movement trajectory, previous targets, target history, occupied map, and obstacle map; The global map corresponding to each robot, centered on a single robot, is transformed to the same global coordinate system, and an H×H×B feature map is generated using a convolutional neural network; H corresponds to the discretization level of the scene, and B is the number of channels.

3. The scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 1, characterized in that, The exploration environment is represented by a graph, and a robot-robot graph G(i) of robot i is constructed, where G(i)∈G(V,E), and node v∈V is a neighboring robot of robot i. If robot i can perceive other robots within a finite radius ρ, then the robots perceived by robot i are its neighboring robots; then there is an edge e∈E between robot i and the perceived robots, and messages can be transmitted bidirectionally between robot i and the perceived robots, with the edge feature being the Euclidean distance between robot i and the perceived robots. All robot characteristics are X M×D =(x1,......,x M Assume an empirical probability distribution P, where M represents the number of robots, i = 1 to M, D represents the length of the eigenvector, and x... i The characteristics of robot i are represented by the global map centered on robot i; the characteristics of neighboring robots of robot i are defined as X. i The matrix C is of size (N(i)+1)×D, where N(i) represents the number of neighboring robots of robot i; a matrix C of shape N(i)+1×N(i)+1 is created for robot i. i Matrix C i element c ij Let x' represent the distance between robot i and its 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 ).

4. A scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 3, 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: Where, x′ i For the neighborhood aggregation information of robot i, x i and x j Let W1 and W2 be the features of robot i and the features of the j-th neighboring robot of robot i, respectively. N(i) is the number of neighboring robots connected to robot i, and W1 and W2 are learnable weight matrices. The attention mechanism is used to prioritize information from neighboring robots, as follows: α i,j W1 is the attention coefficient between node i and node j, and W3, W4 and W5 are learnable weight matrices. Among them, e ij Let c be the feature of the edge connecting robot i and the j-th neighboring robot of robot i, and let c be the output dimension of this layer; The magnitude of the attention coefficient represents the importance of neighboring robots, and the neighboring robots of robot i are prioritized according to the attention coefficient.

5. A scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 4, characterized in that, The graphical mutual information will include the features X of robot i's neighboring robots. i The empirical probability distribution is represented as P(X) i ), the neighborhood aggregation information x′ of robot i i The probability distribution is represented as P(x′) i The joint distribution is represented as P(x′). i |X i If the conditional probability P(x′) i |X i If x' and x'' are multiplied, then the global mutual information I(x'') is... i |X i It can be decomposed into a weighted sum of local mutual information, as follows; Where, x j Let z represent the characteristics of the j-th neighboring robot of robot i, N(i) represent the number of neighboring robots of robot i, and z represent the weight of the j-th neighboring robot of robot i. ij Satisfying 1 / N(i)≤z ij ≤1; Therefore, the neighborhood aggregation information x′ i With Figure G i =(X i C i The graphical mutual information between ) is represented as: in, c ij Let C be the adjacency matrix. i The elements in the expression, σ(·), are the Sigmoid functions; the first term of the expression measures the local mutual information I(x′). i |x j ) for global mutual information I(x′) i |G i The contribution of I(x′); i |x j The contribution of ) is through the robot feature representation x′ i and x j The similarity between them is used for calculation; the second term represents z. ij Mutual information in the relationship between the robot and the robot; by maximizing I(z) ij c ij ), weight z ij This constraint is used to define the topological relationships between robots; that is, when the j-th neighboring robot of robot i is "closer" to robot i, z... ij Larger; this overcomes the shortcoming of focusing only on robot features; Finally, the neighborhood aggregation information for each robot i is x′. i .

6. A scalable multi-robot autonomous exploration method for intelligent information aggregation according to any one of claims 1 to 5, characterized in that, In step S3.1, the Critic network includes a second CNN feature extractor, 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 outputs of the second information aggregation module and the first information aggregation module are different. In the second information aggregation module, the neighborhood aggregation information of all robots is subjected to a global average pooling operation to obtain the global information vector X. agg That is, the robot's graph aggregation information. This vector is independent of the number of robots and is fed into a linear layer in the Critic network to obtain the value of the state-action pair.

7. A scalable multi-robot autonomous exploration method for intelligent information aggregation according to claim 1, characterized in that, In step S3.2, a multi-agent proximal strategy is used to optimize the training of the Actor network and the Critic network.

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

Images