Limited field angle multi-robot cooperative patrol path planning method and device

Abstract

The invention provides a limited field angle multi-robot cooperative patrol path planning method and device, and relates to the technical field of multi-robot cooperative patrol path planning. The method comprises the following steps: planning task data based on a multi-robot cooperative patrol path of a limited field angle, and constructing a graph structure code and current state data; constructing a multi-robot cooperative patrol path based on the current state data and the graph structure codes; and sending the multi-robot cooperative patrol path to each of the plurality of robots, and driving each of the plurality of robots to execute a corresponding decision action. The method is used for solving the problems that the overall task processing efficiency is low and the collaborative planning effect is limited due to the fact that a multi-robot collaborative patrol path planning method in the prior art cannot fully utilize the sensing capacity of the robot in the environment with the limited field angle, and the execution efficiency of the multi-robot collaborative patrol path is improved.

Description

Technical Field

[0001] This invention relates to the field of multi-robot cooperative patrol path planning technology, and in particular to a method and apparatus for multi-robot cooperative patrol path planning with a limited field of view. Background Technology

[0002] Multi-robot systems are widely used in intelligent logistics, automated inspection, security monitoring, and unmanned swarm collaboration. How to achieve efficient collaboration, rational division of labor, and intelligent path and motion planning among multiple robots in a shared environment is one of the core research problems in the field of robotics. With the development of robot perception capabilities, more and more robots are equipped with cameras, depth sensors, or other directional observation devices with field-of-view characteristics. Limited field-of-view perception methods can not only provide robots with high-precision, directional local environmental information, but also achieve focused observation of key areas and dynamic coverage of the task area by reasonably adjusting the orientation. In applications such as target search, inspection monitoring, and visual acquisition, utilizing the directional characteristics of the field of view can significantly improve task execution efficiency.

[0003] In multi-robot systems with limited field of view, robot planning involves not only selecting spatial positions but also deciding on orientation angles. Joint planning of position and orientation helps achieve optimal coverage of information acquisition and complementarity of observation capabilities among robots, thereby improving the overall collaborative performance of the system. However, traditional multi-robot patrol path planning methods, such as graph search-based, sampling planning, or model predictive control algorithms, suffer from high computational complexity and low planning efficiency when dealing with high-dimensional action spaces coupled with position and orientation. Furthermore, these methods often fail to fully utilize the perceptual advantages of the field of view, hindering effective coordination among multiple robots to achieve dynamic field of view coverage and task optimization.

[0004] Deep reinforcement learning has significant advantages in handling high-dimensional state spaces and complex decision-making tasks, providing a new approach for multi-robot cooperative planning. However, existing multi-robot cooperative patrol path planning methods cannot fully utilize the robots' perception capabilities in environments with limited field of view, resulting in low overall task processing efficiency and limited cooperative planning effectiveness. Summary of the Invention

[0005] This invention provides a method and apparatus for multi-robot cooperative patrol path planning with a limited field of view. It addresses the problem that existing multi-robot cooperative patrol path planning methods cannot fully utilize the robots' perception capabilities in environments with limited field of view, resulting in low overall task processing efficiency and limited cooperative planning effectiveness. This invention improves the execution efficiency of multi-robot cooperative patrol paths. The technical solution proposed by this invention is as follows: In a first aspect, the present invention provides a method for planning cooperative patrol paths for multiple robots with a limited field of view, comprising: Based on multi-robot cooperative patrol path planning task data with limited field of view, construct graph structure encoding and current state data; Based on the current state data and the graph structure encoding, a multi-robot collaborative patrol path is constructed; The multi-robot collaborative patrol path is sent to each of the multiple robots, driving each of the multiple robots to execute the corresponding decision-making action.

[0006] In one embodiment, the multi-robot cooperative patrol path planning task data with a limited field of view is constructed as a graph structure model; wherein, the task data contains all target points that need to be visited by multiple robots; In the graph structure model, each target point is defined as a node of the graph structure, and the connectivity between any two target points and the corresponding distance parameters are modeled as edges and attributes of the graph structure. The nodes of the graph structure are used to record the spatial location information of each target point, including two-dimensional coordinate data used to characterize the location of the target point; The edges of the graph structure are used to describe the relationships between target points, including connection relationship data indicating the connectivity between any two target points and edge weight data characterizing the connectivity cost between two target points; wherein, the edge weight data is a distance metric of the Euclidean distance between the corresponding target points.

[0007] In one embodiment, the graph structure encoding includes fixed keys, fixed values, fixed view keys, and fixed graph embeddings; the graph structure encoding is constructed through an encoder network, which includes a first linear layer, a second linear layer, a graph encoder module, and a gating network module; the graph structure encoding is constructed based on the multi-robot cooperative patrol path planning task data with a limited field of view, including: The multi-robot cooperative patrol path planning task data with limited field of view is input into the first linear layer to obtain the initial node feature embedding and initial edge feature embedding output by the first linear layer. The initial node feature embedding and the initial edge feature embedding are input into the graph encoder module to obtain the node feature embedding representation and edge feature embedding representation output by the graph encoder module; The node features are embedded into the second linear layer to obtain the fixed key, fixed value, and fixed view key output by the second linear layer; The node feature embedding representation and the edge feature embedding representation are input into the gated network module to obtain the fixed graph embedding output by the gated network module.

[0008] In one embodiment, constructing a multi-robot cooperative patrol path based on the current state data and the graph structure encoding includes: The current state data and the graph structure encoding are input into the decoder network to obtain the individual node value estimation vector and the individual perspective value estimation vector output by the decoder network. The individual node value estimation vector includes a set of individual node values, and the node corresponding to the largest individual node value in the set of individual node values ​​is determined as the individual node decision action. The individual perspective value estimation vector includes a set of individual perspective values, and the perspective corresponding to the largest individual perspective value in the set of individual perspective values ​​is determined as the individual perspective decision action. Individual robots update their current state data based on individual node decision actions and individual perspective decision actions. The decision actions corresponding to individual robots are individual decision actions, and the individual robot is any one of the plurality of robots. The individual decision actions include individual node decision actions and individual perspective decision actions. Individual robots move to the corresponding target node based on their individual node decision actions. Individual robots also move to the corresponding target node based on their individual perspective decision actions and set a corresponding orientation. Once an individual robot reaches the target node with the specified orientation, it inputs the updated current state data and graph structure encoding into the decoder network. This iteratively generates the individual node value estimation vector and individual perspective value estimation vector for the next time step, and accordingly instructs the individual robot to proceed to the corresponding node with the corresponding orientation until the task allocation for all robot cooperative patrol paths is completed.

[0009] In one embodiment, the decoder network includes a third linear layer, a masked multi-head attention mechanism module, a node decoder module, and a direction decoder module; the step of inputting the current state data and the graph structure encoding into the decoder network to obtain the individual node value estimation vector and the individual viewpoint value estimation vector output by the decoder network includes: The current state data and the graph structure encoding are input into the third linear layer to obtain the current state feature embedding output by the third linear layer; The current state feature embedding and the fixed graph embedding are summed and used as query variables, the fixed value and the fixed key are input into the masked multi-head attention mechanism module to obtain the encoded node features in the current state output by the masked multi-head attention mechanism module. The encoded node features in the current state and the fixed viewpoint key are input into the node decoder module to obtain the individual node value estimation vector output by the node decoder module; The encoded node features in the current state, the individual node value estimation vector, and the fixed graph embedding are input into the orientation decoder module to obtain the individual viewpoint value estimation vector output by the orientation decoder module.

[0010] In one embodiment, before constructing graph structure encoding and current state data from the multi-robot cooperative patrol path planning task data based on a limited field of view, the method further includes: Acquire training data, wherein the training data includes individual node value estimation vectors and individual perspective value estimation vectors acquired by the multiple robots in a historical time period, as well as individual node decision actions and individual perspective decision actions executed. Initialize the parameters of the encoder network and decoder network, and iteratively train the encoder network and decoder network based on the training data until the training termination condition is met, so as to obtain the trained encoder network and decoder network.

[0011] Secondly, the present invention also provides a multi-robot cooperative patrol path planning device with a limited field of view, comprising the following modules: The encoding module is used to construct graph structure encoding and current state data for multi-robot cooperative patrol path planning task data based on a limited field of view; The decoding module is used to construct a multi-robot collaborative patrol path based on the current state data and the graph structure encoding; The sending module is used to send the multi-robot collaborative patrol path to each of the multiple robots, driving each of the multiple robots to execute the corresponding decision-making action.

[0012] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the limited field-of-view multi-robot cooperative patrol path planning method as described in the first aspect above.

[0013] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the finite field-of-view multi-robot cooperative patrol path planning method as described in the first aspect above.

[0014] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the limited field-of-view multi-robot cooperative patrol path planning method as described in the first aspect above.

[0015] Based on the above technical solution, the beneficial effects of the present invention compared with the prior art are as follows: The method and apparatus for multi-robot cooperative patrol path planning with a limited field of view provided by this invention are based on multi-robot cooperative patrol path planning task data with a limited field of view when constructing the multi-robot cooperative patrol path planning scheme. This means that the limited field of view characteristic is fully considered in the process of constructing graph structure encoding and current state data, as well as in the further construction of the cooperative patrol path. Through this targeted modeling, robots can better utilize their perception capabilities in a limited field of view environment, thereby enabling the paths between robots to cooperate with each other. This collaborative path planning method can fully leverage the advantages of multiple robots, improve the ability to cope with complex environments, and thus enhance the effect of cooperative planning. Accurately sending the constructed multi-robot cooperative patrol path to each robot and driving them to execute the corresponding decision actions ensures that all robots act according to a unified plan, avoiding confusion and conflict caused by inconsistent actions between robots, and improving the overall task processing efficiency.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the multi-robot cooperative patrol path planning method with limited field of view provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the encoder network provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the decoder network provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of the multi-robot cooperative patrol path planning device with limited field of view provided by the present invention.

[0023] Figure 5This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] Limited field-of-view multi-robot cooperative patrol path planning refers to designing reasonable control strategies and algorithms to ensure that multiple robots maintain coordination and consistency when performing cooperative patrol tasks, thereby achieving high overall efficiency of the multi-robot system. Deep Reinforcement Learning (DRL) is a technique that combines deep learning and reinforcement learning. In limited field-of-view multi-robot cooperative patrol path planning, DRL can be applied to various aspects such as path planning, action decision-making, and task allocation.

[0026] In related technologies, multi-robot cooperative patrol path planning methods based on deep reinforcement learning with limited field of view mostly focus on optimizing the robot's position nodes, without effectively considering the robot's orientation parameters. This results in the inability to fully utilize the robot's perception capabilities in environments with limited field of view, leading to low overall task processing efficiency and limited cooperative planning effects.

[0027] In summary, how to construct a limited field-of-view multi-robot cooperative patrol path planning method that can fully utilize the advantages of visual directionality while effectively reducing the complexity of multi-robot joint planning under the perception conditions of limited field of view, thereby further improving the task execution efficiency, group cooperation ability and environmental adaptability of multi-robot systems, has become an urgent technical problem to be solved in the field of reinforcement learning.

[0028] Based on this, embodiments of the present invention provide a method and apparatus for planning cooperative patrol paths for multiple robots with limited field of view.

[0029] Reference Figure 1 As shown, the multi-robot cooperative patrol path planning method with a limited field of view includes the following: S110. Based on the multi-robot cooperative patrol path planning task data with limited field of view, construct graph structure encoding and current state data.

[0030] The limited field-of-view multi-robot cooperative patrol path planning method provided in this embodiment can be executed by a multi-robot system. A multi-robot system is a complex system encompassing multiple robots, characterized by the requirement that the multiple robots can collaboratively execute tasks. Within this task framework, each robot possesses independent cognitive and execution capabilities, and through interaction with other robots, they jointly solve the current limited field-of-view multi-robot cooperative patrol path planning task. Within this task framework, the limited field-of-view multi-robot cooperative patrol path planning is applicable to various environmental scenarios, including but not limited to street grids, building maps, and other map environments.

[0031] A limited field of view (FLA) refers to the limitation of a robot's sensors, such as cameras, depth cameras, and LiDAR, on the physical space it can perceive at any given moment, which is restricted by their orientation and the inherent sensing angle of the sensor. Typically, such sensors are directional, only acquiring environmental information within a certain angular range directly in front of them, and cannot achieve 360° omnidirectional perception. In this invention, the limited FLA specifically means that the robot can only observe target or environmental features within a fan-shaped or cone-shaped area of ​​its current orientation. The robot's perception capability is directly related to its orientation; adjusting the orientation changes its observation area. A single robot cannot simultaneously observe all directions; therefore, multi-robot collaboration and dynamic orientation adjustments are necessary to achieve comprehensive coverage of the entire task area. Thus, in path planning, in addition to planning where the robot will move to, it is also necessary to plan the orientation in which the robot arrives at that location to maximize its perception efficiency and task coverage.

[0032] The multi-robot cooperative patrol path planning task data mentioned above refers to a set of structured data used to describe robot states and target point information. It is modeled in graph form to facilitate algorithm processing and path planning. The multi-robot cooperative patrol path planning task data includes a target point set and a robot set. The target point set represents all key locations that need to be visited or observed by the robots. Each target point contains its two-dimensional coordinate information, and the number of target points is denoted as N. The robot set represents all robots participating in the cooperative patrol. Each robot has a unique robot ID and records its current position coordinates; the number of robots is denoted as M. The task data is modeled as a graph. The node set V contains all target point nodes and the robot's initial position node. The edge set E describes the connectivity between nodes. An edge exists between any two nodes if they are directly connected. Each edge contains two attributes: edge weight and edge angle feature. The edge weight is the Euclidean distance between the two nodes, representing the movement cost. The edge angle feature describes the direction angle from one node to another, facilitating subsequent orientation planning.

[0033] After establishing the graph structure model, the graph undergoes encoding processing. Specifically, this involves multi-level abstraction and integration of node features, edge features, and global topological relationships in the graph to generate a fixed-dimensional graph structure encoding representation. Simultaneously, current state data is constructed and maintained. Current state data includes the current robot ID. and the node number to which it belongs Remaining robot ratio characteristics Remaining target point ratio characteristics Distance features, including the distance the robot has already traveled. The farthest distance from all target points to the robot's return point The furthest distance among the remaining target points from the robot's return point ,in, This indicates the number of robots currently in use. This indicates the number of target points currently used.

[0034] S120. Based on the current state data and the graph structure encoding, construct a multi-robot collaborative patrol path.

[0035] Current state data reflects the current task execution progress, including the proportion of robots that have completed tasks, the proportion of target points that have been visited, the current position of each robot, and the cumulative distance traveled. Graph structure encoding contains a topological representation of the environment after feature extraction. A two-step decision-making mechanism is used to plan patrol paths for each robot, a process that fully integrates current state data and graph structure encoding. When making location decisions, current state data and graph structure encoding are jointly analyzed. Current state data provides real-time dynamic information about the task, including real-time features such as the proportion of target points to be visited, the proportion of available robots, the cumulative distance traveled by each robot, and the current position of the robot. Graph structure encoding provides a topological representation of the environment, including static features such as node features, edge connections, and global structural information. By fusing these two types of features, based on the current global state and environmental structural features, the expected benefits of the robot moving to each target point to be visited are evaluated, forming a location value assessment result. The target point with the highest benefit is selected as the robot's next destination.

[0036] In the orientation decision-making phase, the perception effectiveness of the robot when arriving at the target location is evaluated by combining structural feature information about the surrounding environment from the graph structure encoding with information about task completion and collaboration requirements from the current state data. Specifically, the feature representation of node neighborhood relationships from the graph structure encoding is used to analyze the area that different orientations can cover; at the same time, information about covered and uncovered areas from the current state data is combined to calculate the value assessment of each possible orientation, and the orientation that maximizes perception coverage efficiency or task completion quality is selected as the arrival posture.

[0037] This planning process is conducted sequentially in an iterative manner. Each time, only the joint decision action of "target position-target orientation" is determined for the current robot. After the robot's state is updated (including updating the robot's position, marking visited target points, etc.), new current state data is generated. This data is then combined with a fixed graph structure encoding to plan for the next robot or the next decision moment of the same robot. This process is repeated until all robots have obtained a complete patrol path sequence, forming a collaborative patrol scheme covering all target points.

[0038] S130. Send the multi-robot collaborative patrol path to each of the multiple robots, and drive each of the multiple robots to execute the corresponding decision action.

[0039] The decision-making actions corresponding to individual robots are individual decision-making actions. An individual robot is any one of multiple robots, and individual decision-making actions include individual node decision-making actions and individual perspective decision-making actions. After receiving its corresponding patrol path, each individual robot autonomously executes the patrol instructions in the path according to the individual decision-making actions. In this way, multiple robots independently complete the selection of task actions and motion control based on their own planned paths and corresponding orientations, thereby achieving efficient and orderly execution of multi-robot collaborative patrol tasks.

[0040] Existing methods typically fail to specifically model the field of view characteristics, leaving robots unable to fully utilize their perception capabilities in environments with limited field of view. This invention, when constructing a multi-robot cooperative patrol path planning scheme, is based on multi-robot cooperative patrol path planning task data with limited field of view. This means that the limited field of view characteristic is fully considered during the construction of graph structure encoding, current state data, and further cooperative patrol path construction. This targeted modeling allows robots to better utilize their perception capabilities in environments with limited field of view, enabling their paths to coordinate with each other. This collaborative path planning method fully leverages the advantages of multiple robots, improves the ability to cope with complex environments, and thus enhances the effectiveness of cooperative planning. Accurately sending the constructed multi-robot cooperative patrol path to each robot and driving them to execute corresponding decision actions ensures that all robots act according to a unified plan, avoiding confusion and conflict caused by inconsistent robot actions, and improving overall task processing efficiency.

[0041] Furthermore, this invention constructs graph structure encoding and current state data based on multi-robot cooperative patrol path planning task data with a limited field of view, and then constructs multi-robot cooperative patrol paths based on this data. This construction method does not directly adopt parallel interaction and joint training strategies, but rather avoids excessive expansion of the state and action space caused by the increase in the number of robots to a certain extent through data encoding and state utilization, thereby reducing the training difficulty.

[0042] In this embodiment, the multi-robot cooperative patrol path planning task data with a limited field of view is constructed as a graph structure model; wherein, the task data includes all target points that need to be visited by multiple robots; In the graph structure model, each target point is defined as a node of the graph structure, and the connectivity between any two target points and the corresponding distance parameters are modeled as edges and attributes of the graph structure. The nodes of the graph structure are used to record the spatial location information of each target point, including two-dimensional coordinate data representing the location of the target point. The edges of the graph structure are used to describe the association between target points, including connection data indicating the connectivity between any two target points and edge weight data representing the connection cost between two target points. The edge weight data is a distance metric of the Euclidean distance between the corresponding target points.

[0043] Specifically, during the task data construction process, all target points that need to be accessed by the robot and the robot itself are abstracted as nodes in a graph, and connecting edges are constructed based on the spatial positional relationships between the nodes to form a graph structure corresponding to the multi-robot cooperative patrol path planning task data with a limited field of view. .in, Used to represent a set of nodes Used to represent edge sets. Specifically, graph structures. It can be represented equally as three parts: the target point coordinate matrix, the robot position matrix, and the edge attributes. Target point coordinate matrix , here This indicates the number of target points, and each row represents the two-dimensional coordinates of a target point. Robot position matrix. , here This indicates the number of robots, with each row representing the robot's two-dimensional coordinates. (Edge attributes) It contains feature information in three dimensions, namely edge weights and by and The edge angle features formed, among which Representing vectors The direction angle.

[0044] This invention achieves a structured representation of complex physical environments by abstracting target points in patrol missions as nodes and the connectivity between target points as edges. This allows the algorithm to process spatial information and topological relationships within a unified framework. Based on this, precise two-dimensional coordinate data in nodes clearly characterize the spatial location of target points, while connectivity identifiers and Euclidean distance weights in edge attributes quantify the reachability and movement costs between locations. This provides a precise data foundation for subsequent path cost calculation and optimization. In particular, this structured modeling approach is naturally suited to the collaborative planning needs of scenarios with limited field of view. The edge weights directly reflect the physical cost of robot movement. Combined with node position information, this allows the algorithm to optimize perception orientation while planning movement paths, thereby maximizing coverage efficiency within a limited perception range. Furthermore, the graph structure has good algorithm compatibility and scalability, enabling seamless integration with deep learning models such as graph neural networks, facilitating the construction of end-to-end reinforcement learning decision systems. The model can also dynamically update the states of nodes and edges according to task progress, flexibly adapting to environmental changes and large-scale scene expansion. Moreover, graph structure modeling effectively compresses the path search space and avoids invalid paths by explicitly expressing connectivity relationships, significantly reducing the complexity of multi-robot collaborative planning. Simultaneously, by uniformly encoding position, orientation, and connectivity relationships, the model achieves integrated optimization of movement paths and observation orientation, thereby improving task execution efficiency while enhancing the collaborative perception capability and overall robustness of the multi-robot system.

[0045] This invention achieves intelligent path planning for multi-robot cooperative patrols under limited field of view by deeply integrating a deep neural network architecture with a reinforcement learning optimization mechanism. The deep neural network employs an encoder-decoder structure to realize an end-to-end mapping from multi-robot cooperative patrol path planning task data to the robot position-orientation joint action sequence. The reinforcement learning optimization mechanism optimizes network parameters through a policy gradient algorithm, aiming to minimize the maximum distance traveled by the robot, and improves training efficiency and generalization ability by leveraging symmetric transformation and shared baseline. The planning process of multi-robot cooperative patrols under limited field of view is described below.

[0046] In this embodiment, the graph structure encoding described above is constructed through an encoder network, such as... Figure 2 As shown, the encoder network includes a first linear layer, a second linear layer, a graph encoder module, and a gating network module. In this embodiment, the linear layer is a basic feedforward neural network structure unit that performs feature mapping by linearly transforming the input features. The linear layer typically consists of a weight matrix and bias terms, which can map the input vector to an output space of a specified dimension, used to achieve functions such as feature compression, dimensionality expansion, or feature reorganization, thereby providing a more expressive feature representation for subsequent network structures. Graph structure encoding includes fixed keys, fixed values, fixed viewpoint keys, and fixed graph embedding. The graph structure encoding of the multi-robot cooperative patrol path planning task data based on a limited field of view, as described in S110 above, includes: S1101. Input the multi-robot cooperative patrol path planning task data with limited field of view into the first linear layer to obtain the initial node feature embedding and initial edge feature embedding output by the first linear layer.

[0047] The data for multi-robot cooperative patrol path planning tasks with limited field of view (corresponding to) Figure 2 The target point coordinate matrix, robot position matrix, and edge attributes are input into the first linear layer, yielding the initial node feature embedding and initial edge feature embedding output by the first linear layer. Here, in a multi-robot cooperative patrol path planning task with a limited field of view, after the data is input into the first linear layer, the first linear layer is used to extract initial features and uniformly embed and encode the target point positions, robot positions, and edge attributes. The processed data specifically includes: target point position embedding. Robot position embedding Initial edge feature embedding After completing the linear mapping described above, the processed target point positions are embedded. With robot position embedding Stack them to form the initial node feature embedding. .

[0048] S1102. Input the initial node feature embedding and the initial edge feature embedding into the graph encoder module to obtain the node feature embedding representation and edge feature embedding representation output by the graph encoder module.

[0049] Embed the initial node features and initial edge feature embedding Input the graph encoder module to obtain the node feature embedding representation output by the graph encoder module. Edge feature embedding representation The graph encoder is... L A graph neural network is constructed by stacking layers, and each layer of the graph neural network can be specifically represented as: in, The learnable parameter matrix; SiLU is the activation function; GN represents the graph normalization operation; Indicates a node The features of neighboring nodes are aggregated by mean. For the Sigmoid function; This represents digit-by-digit Hadamard multiplication; Indicates the first layer The encoding corresponding to the node. Indicates the first The encoding corresponding to node j in layer j. The updated node feature embedding for layer l+1 represents the new node feature vector obtained by node i after passing through the graph neural network message passing and feature transformation from layer l to layer l+1. The updated edge feature embedding at layer l+1 represents the new edge feature vector obtained after computation by the graph neural network at layer l+1, representing the edge from node i to node j. The superscript l+1 indicates the next layer after the current layer, i.e., the result after being updated by the current layer of the graph neural network. The subscript i represents the i-th node in the graph (target point or robot position). The above update mechanism allows node features to be influenced not only by their own historical state, but also to adaptively aggregate effective features from neighboring nodes through the edge gating mechanism; at the same time, the edge features comprehensively consider the feature states of the two connected nodes during the update, thereby achieving a strengthened expression of the directionality and correlation between nodes in the graph structure.

[0050] S1103. Embed the node features into the second linear layer to obtain the fixed key, fixed value and fixed view key output by the second linear layer.

[0051] The node features are embedded into the representation and input to the second linear layer, yielding the fixed key, fixed value, and fixed viewpoint key output by the second linear layer. Specifically, the node features are embedded into the representation. Input to the second linear layer; the output of the second linear layer is a fixed key. Fixed value and fixed view button : S1104. Input the node feature embedding representation and the edge feature embedding representation into the gated network module to obtain the fixed graph embedding output by the gated network module.

[0052] Embedding node features into representation Edge feature embedding representation Input the gated network module and obtain the fixed graph embedding output by the gated network module. Specifically, the gated network module first embeds a representation of the node features output by the graph encoder module. Edge feature embedding representation Aggregation is performed to obtain the global representation of the nodes. Global representation of edges : in, Represents a node The node feature embedding representation, Indicates a node The features of neighboring nodes are aggregated by mean. Then, by introducing a gating mechanism, the global features of nodes and edges are fused to obtain a fixed graph embedding. : in, and represents the learnable parameters, and GELU represents the activation function.

[0053] In summary, the task data for multi-robot cooperative patrol path planning with a limited field of view is input into the encoder network, and the encoder network outputs a fixed key... Fixed value and fixed view button and fixed image embedding Among them, fixed key Fixed value and fixed view button and fixed image embedding Together they form a graph structure encoding.

[0054] The encoder network of this invention performs preliminary feature embedding on the original graph structure data through a first linear layer, mapping heterogeneous information such as target point coordinates, robot position, and edge attributes to a unified vector space. This achieves a standardized representation of the task data and lays a compatible foundation for subsequent deep feature extraction. The graph encoder module, as the core processing unit, achieves deep fusion of local structural information and global topological relationships through multi-round message passing and neighborhood aggregation of node and edge features. The node feature embedding representation output by this module not only includes the node's own attribute information but also incorporates the structural features of its neighborhood, particularly strengthening the association information related to direction perception. The edge feature embedding representation further encodes the directionality and weight characteristics of connectivity relationships. This deep encoding mechanism enables the model to fully understand the spatial structure and perceptual constraints of the environment under a limited field of view.

[0055] The fixed keys, fixed values, and fixed viewpoint keys generated by the second linear layer provide a stable query and comparison benchmark for the attention mechanism in the subsequent decoding process. The fixed keys and fixed values ​​retain feature summaries of each node in the graph, enabling the decoder to quickly retrieve environmental information during planning. The fixed viewpoint key specifically encodes directional features related to orientation decisions, providing an orientation reference for the robot's viewpoint selection. The design of these three fixed components ensures both the efficiency of information acquisition and enhances the stability and interpretability of decoding decisions.

[0056] The gated network module generates a fixed graph embedding that incorporates the overall semantics of the environment by adaptively fusing global features of nodes and edges. This embedding, as a high-level abstract representation of the environment, provides the decoder with task-level contextual information, enabling path planning to be performed under the guidance of a global optimization objective. The fixed graph embedding, together with the aforementioned three fixed components, forms a graph structure encoding system that creates a multi-granularity environment representation system from local details to global semantics.

[0057] The hierarchical coding architecture provided by this invention transforms raw graph data into a multi-component coded representation rich in semantic information through a collaborative feature extraction mechanism. This coding not only fully preserves the topological structure and directional characteristics of the environment, but also enhances the computational efficiency and decision consistency of the subsequent decoding process through fixed processing, thereby significantly improving the perceptual adaptability and optimization effect of multi-robot cooperative patrol path planning under limited field of view.

[0058] In this embodiment, the process of constructing a multi-robot cooperative patrol path is completed through a decoder network. The construction of the multi-robot cooperative patrol path based on the current state data and the graph structure encoding described in S120 includes: The current state data and the graph structure encoding are input into the decoder network to obtain the individual node value estimation vector and the individual perspective value estimation vector output by the decoder network. The individual node value estimation vector includes a set of individual node values, and the node corresponding to the largest individual node value in the set of individual node values ​​is determined as the individual node decision action. The individual perspective value estimation vector includes a set of individual perspective values, and the perspective corresponding to the largest individual perspective value in the set of individual perspective values ​​is determined as the individual perspective decision action. Individual robots update their current state data based on individual node decision actions and individual perspective decision actions. The decision actions corresponding to individual robots are individual decision actions, and the individual robot is any one of the plurality of robots. The individual decision actions include individual node decision actions and individual perspective decision actions. Individual robots move to the corresponding target node based on their individual node decision actions. Individual robots also move to the corresponding target node based on their individual perspective decision actions and set a corresponding orientation. Once an individual robot reaches the target node with the specified orientation, it inputs the updated current state data and graph structure encoding into the decoder network. This iteratively generates the individual node value estimation vector and individual perspective value estimation vector for the next time step, and accordingly instructs the individual robot to proceed to the corresponding node with the corresponding orientation until the task allocation for all robot cooperative patrol paths is completed.

[0059] This invention first inputs the current state data reflecting the real-time progress of the task and the graph structure encoding the topology of the encoded environment into a decoder network. Through joint inference, it outputs individual node value estimation vectors and individual viewpoint value estimation vectors for the robot currently being planned, achieving a collaborative value assessment of the moving target and the observation orientation. In terms of the decision-making mechanism, the target position is determined by selecting the maximum value from the node value estimation vectors, and the optimal observation orientation is determined by selecting the maximum value from the viewpoint value estimation vectors, forming a joint "position-orientation" decision action. This design enables the robot to not only choose an efficient movement path under limited perception constraints but also optimize the observation direction when reaching the target, thereby improving the perception gain and task coverage of a single step. After executing this decision, the robot synchronously updates its current state data, such as marking visited nodes, updating the robot's position and task progress, and then calls the decoder network again based on the updated state to make the next decision, forming a closed-loop iterative process of "decision-execution-update". This sequential decision-making mechanism, through gradual state updates and replanning, can dynamically adapt to task progress and environmental changes, achieving adaptive task allocation. By planning only one action for each robot at a time, the high-dimensional multi-robot joint decision-making problem is decomposed into a series of single-robot single-step decision-making problems, significantly reducing decision complexity and computational burden. Through iterative accumulation of single-step optimized decisions, a coordinated patrol path is ultimately formed, ensuring global coordination while avoiding conflicts and redundant coverage. More importantly, this method deeply integrates the characteristics of a limited field of view into the decision loop, explicitly optimizing orientation selection through viewpoint value estimation. This allows robots to autonomously adjust their observation direction based on environmental structure and task status, maximizing their directional perception capabilities. Finally, by iteratively assigning complete patrol paths to all robots, the collaborative goals of full target coverage, optimized orientation, and balanced resource utilization are achieved, significantly improving the patrol efficiency and perception quality of multiple robots under limited field-of-view constraints.

[0060] In this embodiment, as Figure 3 As shown, the decoder network includes a third linear layer, a masked multi-head attention mechanism module, a node decoder module, and a direction decoder module; the step of inputting the current state data and the encoded graph structure into the decoder network to obtain the individual node value estimation vector and the individual viewpoint value estimation vector output by the decoder network includes: S210. Input the current state data and the graph structure encoding into the third linear layer to obtain the current state feature embedding output by the third linear layer.

[0061] The current state data (corresponding to) Figure 3 Current robot number in and the node number to which it belongs Remaining robot ratio characteristics Remaining target point ratio characteristics Distance already traveled The farthest distance from all target points to the return point The furthest distance from the remaining target point to the return point Input the third linear layer to obtain the current robot state code output by the third linear layer. Current robot task progress code Distance encoding of the current robot ,in, Encode the current robot number. Encode the node where the robot is currently located. Encode the current robot state. Current robot task progress code Distance encoding of the current robot Summing yields the final output of the third linear layer, which is the current state feature embedding. .

[0062] S220. The current state feature embedding and the fixed graph embedding are summed as query variables, the fixed value and the fixed key are input into the masked multi-head attention mechanism module to obtain the encoded node features in the current state output by the masked multi-head attention mechanism module.

[0063] Embed the current state features Embedded with fixed graph The sum is used as the query variable and compared with a fixed value. and fixed key The inputs are combined with the masked multi-head attention mechanism module to obtain the encoded node features in the current state, as output by the masked multi-head attention mechanism module. Encoding node features in the current state. The specific calculation method is as follows: The MASK mechanism is used to mask visited nodes to prevent attention from being allocated to invalid locations. Here, we will explain the Masked Multi-Head Attention (MMHA) mechanism. This mechanism employs a masked multi-head attention module consisting of eight attention heads. Multi-head attention constructs multiple independent attention heads in parallel, enabling the model to learn multiple relevance patterns simultaneously in different feature subspaces, thereby improving representational capabilities. Building upon this, masked multi-head attention further incorporates a masking operation to block attention allocation to visited nodes, ensuring that attention is applied only to the currently selectable node.

[0064] S230. Input the encoded node features in the current state and the fixed viewpoint key into the node decoder module to obtain the individual node value estimation vector output by the node decoder module.

[0065] Encode node features in the current state and fixed view button The input node decoder module is used to obtain the individual node value estimation vector output by the node decoder module. The node decoder encodes node features based on the current state. And combined with a fixed viewpoint key To generate individual node value estimation vectors Specifically, the individual node value estimation vector The calculation is defined as follows: in, For learnable parameters, constants To explore the amplification factor, which can be used to facilitate exploration during the training process. In this invention, the latent space dimensions of the encoder and decoder networks are set as follows: tanh is the hyperbolic tangent function. S240. Input the encoded node features in the current state, the individual node value estimation vector, and the fixed graph embedding into the orientation decoder module to obtain the individual viewpoint value estimation vector output by the orientation decoder module.

[0066] Encode node features in the current state Individual node value estimation vector and fixed image embedding Input the direction decoder module to obtain the individual viewpoint value estimation vector output by the direction decoder module. The directional decoder module first estimates the value vector of each individual node. Select the corresponding node Then, combined with fixed image embedding and nodes coding Obtain the value estimation vector from the individual perspective The calculation process is as follows: in, These are learnable parameters. Let represent a d×d real matrix.

[0067] The process of selecting nodes based on the individual node value estimation vector is as follows: Assuming there are 10 nodes, the individual node value estimation vector is a vector of length 10, where each element represents the selection probability of each node. During training, nodes are selected by directly sampling from this vector, with each element representing the probability of selecting that node. During evaluation, the id corresponding to the element with the largest value in this vector is directly selected, indicating the next node to be selected.

[0068] It should be noted that the decision-making action corresponding to an individual robot is an individual decision-making action, and the individual robot is any one of the plurality of robots. The individual decision-making action includes individual node decision-making actions and individual perspective decision-making actions. The individual node decision-making action is used to select one of multiple candidate target points as the robot's next target position, and the individual robot moves to the corresponding target node according to its individual node decision-making action. The individual perspective decision-making action is used to discretize the robot's selectable orientation, for example, dividing the 360° orientation range into several equal parts (such as dividing it into 6 equal parts), and selecting one from this discrete orientation set as the corresponding orientation decision. The individual robot moves to the corresponding target node and sets the corresponding orientation according to its individual perspective decision-making action.

[0069] The decoder network iteratively generates individual node value estimation vectors and individual viewpoint value estimation vectors for the current individual robot at the next time step, and instructs the individual robot to head towards the corresponding node in the specified direction. It should be noted that the decoder network generates individual node value estimation vectors and individual viewpoint value estimation vectors only for one individual robot in each iteration. When the individual robot returns to the return point, its patrol path allocation is considered complete. The decoder network then begins allocating patrol paths for the next individual robot, and this process continues until all individual robots have completed the task allocation for cooperative patrol paths.

[0070] The decoder network of this invention employs a hierarchical modular structure to process current state data and graph structure encoding. It achieves accurate node value and viewpoint value estimation through multi-stage feature fusion and an attention mechanism. First, the third linear layer performs linear transformation and feature compression on the input state data and graph encoding, generating embedding vectors representing the current task progress and robot state, providing a suitable query representation for subsequent attention calculations. The masked multi-head attention mechanism module fuses the current state feature embedding with fixed graph embeddings into query variables and performs attention calculations using fixed keys and values, achieving dynamic focusing on environmental node features. This module uses a masking mechanism to shield visited nodes, ensuring attention is only distributed on valid candidate nodes, avoiding repeated visits and resource waste. The multi-head design supports parallel capture of various association patterns between nodes, enhancing the model's ability to perceive complex spatial relationships and task dependencies. Its output encoded node features fuse current state information and global environmental semantics, providing context-rich node representations for subsequent value estimation. The node decoder module estimates node value based on encoded node features and fixed viewpoint keys, generating individual node value estimation vectors by calculating the matching degree between each candidate node and the current state and orientation priors. This module utilizes orientation-aware key vectors to enhance the assessment of node observability, making value estimation more closely aligned with the perception characteristics under limited field of view. This ensures that the selected target node not only has low path cost but also high observation potential. The value maximization selection mechanism further guarantees the local optimality of the decision, providing clear goal guidance for robot movement.

[0071] After identifying the target node, the orientation decoder module performs viewpoint value estimation by integrating encoded node features, node value estimation, and fixed graph embedding. This module evaluates the perception coverage efficiency and environmental cooperation benefits when the robot reaches the target with different orientations by fusing local node features with global graph semantics, generating individual viewpoint value estimation vectors. Viewpoint decision-making not only considers the observation optimization of the current node but also incorporates the global task state and multi-robot cooperation requirements, enabling orientation selection to both enhance immediate perception gains and promote balanced progress in the overall task.

[0072] This invention's decoder network achieves end-to-end mapping from state and environmental features to node and viewpoint values ​​through hierarchical processing of linear transformation, masked attention, node decoding, and orientation decoding. This design balances spatial efficiency in path planning with perceptual effectiveness in orientation optimization. By employing attention masking and orientation perception mechanisms, it significantly enhances the targetedness and synergy of decision-making, thereby effectively improving the task execution efficiency and perception coverage quality of multi-robot collaborative patrols under limited field-of-view constraints.

[0073] In this embodiment, before constructing the graph structure encoding and current state data based on the multi-robot cooperative patrol path planning task data with a limited field of view as described in S110 above, the method further includes: S1001. Acquire training data, wherein the training data includes individual node value estimation vectors and individual perspective value estimation vectors acquired by the multiple robots in a historical time period, as well as individual node decision actions and individual perspective decision actions executed. S1002. Initialize the parameters of the encoder network and decoder network, and iteratively train the encoder network and decoder network based on the training data until the training termination condition is met, so as to obtain the trained encoder network and decoder network.

[0074] The parameters of the encoder and decoder networks are initialized, and the encoder and decoder networks are iteratively trained using the REINFORCE loss function based on the training data until the training termination condition is met, resulting in the trained encoder and decoder networks. Specifically, the REINFORCE training objective based on a symmetric shared baseline can be expressed as: in, Describe the policy objective function J For network parameters θ gradient estimation, a (i,j) Indicates the first i Training samples (Figure) G i ) in the j Action sequences generated by a policy network (i.e., encoder and decoder networks) under a certain symmetric transformation. This indicates that for the i-th graph G i The result after applying the j-th symmetric transformation.

[0075] Indicates the first One image, This represents the encoder and decoder networks. B is the batch size, with each batch containing B distinct images. . This represents the set of symmetric transformations applied to each graph, and the number of symmetric transformations. ,sequence Indicates that for the first i The batch of instances and their first The action sequence generated by a symmetric transformation, with a shared baseline defined as . Indicates a given sequence of actions a (i,j) and picture The corresponding reward size is determined by the input action sequence, which represents the node access paths of all robots. The reward size is the negative of the longest path length among all robots, and the training objective is to minimize the maximum walking distance among all robots. Indicates the target points visited by the robot in sequence. and the corresponding observation orientation .

[0076] This invention, before constructing the graph structure encoding and current state data for multi-robot cooperative patrol path planning tasks with limited field of view, introduces a training mechanism based on historical data, achieving a leap from experience-based learning to autonomous decision-making. This training process uses the robot's value estimation vector and decision actions from historical tasks as training data, constructing a mapping learning foundation from environmental state to decision behavior. This allows the encoder and decoder networks to gradually master the inherent laws of cooperative patrol under limited field of view through iterative optimization. First, the construction of the training data fully reflects the historical experience and decision-making logic of task execution. The value estimation vectors of individual nodes and individual perspectives reflect the benefit assessments of different positions and orientations in historical tasks, while the corresponding decision actions reflect the actual behavioral choices. These data together constitute a complete triplet of state-value-action, providing the network with a learning signal that combines supervision and reinforcement. This enables the model not only to learn specific decisions but also to understand the value assessment basis behind those decisions, thus achieving more accurate generalization reasoning in subsequent tasks. Second, by initializing network parameters and iteratively optimizing them based on training data, an end-to-end joint training mechanism is achieved. The encoder network learns an efficient mapping from raw graph data to compact structure encoding, gradually mastering how to extract the most relevant feature information for perception within a limited field of view in the environment. The decoder network, on the other hand, learns the transformation relationship from state encoding to value estimation, gradually optimizing the collaborative strategy for node selection and orientation decision-making. This joint training enables the two networks to adapt to each other and co-evolve during parameter updates, ultimately forming a highly efficient path decision-making model, namely the combination of the encoder and decoder networks described above.

[0077] The following describes the finite field-of-view multi-robot cooperative patrol path planning device provided by the present invention. The finite field-of-view multi-robot cooperative patrol path planning device described below and the finite field-of-view multi-robot cooperative patrol path planning method described above can be referred to in correspondence with each other.

[0078] The finite field-of-view multi-robot cooperative patrol path planning device provided by this invention refers to... Figure 4 As shown, it includes: Encoding module 310 is used to construct graph structure encoding and current state data for multi-robot cooperative patrol path planning task data based on a limited field of view. Decoding module 320 is used to construct a multi-robot collaborative patrol path based on the current state data and the graph structure encoding; The sending module 330 is used to send the multi-robot collaborative patrol path to each of the multiple robots, driving each of the multiple robots to execute the corresponding decision action.

[0079] Specifically, according to embodiments of the present invention, any multiple modules among the encoding module, decoding module, and transmitting module can be merged into one module, or any one of these modules can be split into multiple sub-modules. Alternatively, at least some functions of one or more of these modules can be combined with at least some functions of other modules and implemented in the same module. According to embodiments of this application, at least one of the above modules can be at least partially implemented as a hardware circuit, such as a field-programmable gate array, a programmable logic array, a system-on-a-chip, an application-specific integrated circuit, etc., or implemented in any one or a combination of software, hardware, and firmware. Alternatively, at least one of the above modules can be at least partially implemented as a computer program module, which, when executed by a processor, can implement corresponding functional steps. The limited field-of-view multi-robot cooperative patrol path planning device provided in the embodiments of this application, by introducing a node decoder module and a direction decoder module, simultaneously optimizes the nodes and the viewpoint, thereby solving the problem that related technologies for multi-robot cooperative patrol path planning do not effectively consider the robot's orientation parameters, resulting in the robot's inability to fully utilize its perception capabilities in a limited field-of-view environment, leading to low overall task processing efficiency and limited cooperative planning effect, and improving the execution efficiency of multi-robot cooperative patrol paths.

[0080] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions from the memory 430 to execute a multi-robot cooperative patrol path planning method with a limited field of view.

[0081] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the limited field-of-view multi-robot cooperative patrol path planning method provided by the above methods.

[0083] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the finite field-of-view multi-robot cooperative patrol path planning method provided by the methods described above.

[0084] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for planning cooperative patrol paths for multiple robots with a limited field of view, characterized in that, include: Based on multi-robot cooperative patrol path planning task data with limited field of view, construct graph structure encoding and current state data; Based on the current state data and the graph structure encoding, a multi-robot collaborative patrol path is constructed; The multi-robot collaborative patrol path is sent to each of the multiple robots, driving each of the multiple robots to execute the corresponding decision-making action.

2. The multi-robot cooperative patrol path planning method with limited field of view according to claim 1, characterized in that, The multi-robot cooperative patrol path planning task data with a limited field of view is constructed as a graph structure model; wherein, the task data contains all target points that need to be visited by multiple robots; In the graph structure model, each target point is defined as a node of the graph structure, and the connectivity between any two target points and the corresponding distance parameters are modeled as edges and attributes of the graph structure. The nodes of the graph structure are used to record the spatial location information of each target point, including two-dimensional coordinate data used to characterize the location of the target point; The edges of the graph structure are used to describe the relationships between target points, including connection relationship data indicating the connectivity between any two target points and edge weight data characterizing the connectivity cost between two target points; wherein, the edge weight data is a distance metric of the Euclidean distance between the corresponding target points.

3. The multi-robot cooperative patrol path planning method with limited field of view according to claim 1, characterized in that, The graph structure encoding includes fixed keys, fixed values, fixed view keys, and fixed graph embeddings; the graph structure encoding is constructed through an encoder network, which includes a first linear layer, a second linear layer, a graph encoder module, and a gating network module. The multi-robot cooperative patrol path planning task data based on a limited field of view is used to construct a graph structure encoding, including: The multi-robot cooperative patrol path planning task data with limited field of view is input into the first linear layer to obtain the initial node feature embedding and initial edge feature embedding output by the first linear layer. The initial node feature embedding and the initial edge feature embedding are input into the graph encoder module to obtain the node feature embedding representation and edge feature embedding representation output by the graph encoder module; The node features are embedded into the second linear layer to obtain the fixed key, fixed value, and fixed view key output by the second linear layer; The node feature embedding representation and the edge feature embedding representation are input into the gated network module to obtain the fixed graph embedding output by the gated network module.

4. The multi-robot cooperative patrol path planning method with limited field of view according to claim 3, characterized in that, The construction of a multi-robot cooperative patrol path based on the current state data and the graph structure encoding includes: The current state data and the graph structure encoding are input into the decoder network to obtain the individual node value estimation vector and the individual perspective value estimation vector output by the decoder network. The individual node value estimation vector includes a set of individual node values, and the node corresponding to the largest individual node value in the set of individual node values ​​is determined as the individual node decision action. The individual perspective value estimation vector includes a set of individual perspective values, and the perspective corresponding to the largest individual perspective value in the set of individual perspective values ​​is determined as the individual perspective decision action. Individual robots update their current state data based on individual node decision actions and individual perspective decision actions. The decision actions corresponding to individual robots are individual decision actions, and the individual robot is any one of the plurality of robots. The individual decision actions include individual node decision actions and individual perspective decision actions. Individual robots move to the corresponding target node based on their individual node decision actions. Individual robots also move to the corresponding target node based on their individual perspective decision actions and set a corresponding orientation. Once an individual robot reaches the target node with the specified orientation, it inputs the updated current state data and graph structure encoding into the decoder network. This iteratively generates the individual node value estimation vector and individual perspective value estimation vector for the next time step, and accordingly instructs the individual robot to proceed to the corresponding node with the corresponding orientation until the task allocation for all robot cooperative patrol paths is completed.

5. The multi-robot cooperative patrol path planning method with limited field of view according to claim 4, characterized in that, The decoder network includes a third linear layer, a masked multi-head attention mechanism module, a node decoder module, and a direction decoder module; the step of inputting the current state data and the encoded graph structure into the decoder network to obtain the individual node value estimation vector and the individual viewpoint value estimation vector output by the decoder network includes: The current state data and the graph structure encoding are input into the third linear layer to obtain the current state feature embedding output by the third linear layer; The current state feature embedding and the fixed graph embedding are summed and used as query variables, the fixed value and the fixed key are input into the masked multi-head attention mechanism module to obtain the encoded node features in the current state output by the masked multi-head attention mechanism module. The encoded node features in the current state and the fixed viewpoint key are input into the node decoder module to obtain the individual node value estimation vector output by the node decoder module; The encoded node features in the current state, the individual node value estimation vector, and the fixed graph embedding are input into the orientation decoder module to obtain the individual viewpoint value estimation vector output by the orientation decoder module.

6. The multi-robot cooperative patrol path planning method with limited field of view according to claim 4, characterized in that, Before constructing graph structure encoding and current state data from the multi-robot cooperative patrol path planning task data based on a limited field of view, the method further includes: Acquire training data, wherein the training data includes individual node value estimation vectors and individual perspective value estimation vectors acquired by the multiple robots in a historical time period, as well as individual node decision actions and individual perspective decision actions executed. Initialize the parameters of the encoder network and decoder network, and iteratively train the encoder network and decoder network based on the training data until the training termination condition is met, so as to obtain the trained encoder network and decoder network.

7. A multi-robot cooperative patrol path planning device with a limited field of view, characterized in that, include: The encoding module is used to construct graph structure encoding and current state data for multi-robot cooperative patrol path planning task data based on a limited field of view; The decoding module is used to construct a multi-robot collaborative patrol path based on the current state data and the graph structure encoding; The sending module is used to send the multi-robot collaborative patrol path to each of the multiple robots, driving each of the multiple robots to execute the corresponding decision-making action.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the limited field-of-view multi-robot cooperative patrol path planning method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the finite field-of-view multi-robot cooperative patrol path planning method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the finite field-of-view multi-robot cooperative patrol path planning method as described in any one of claims 1 to 6.

Images