Multi-machine collaborative navigation optimization method based on graph neural network

Abstract

The invention relates to the technical field of robot navigation, in particular to a multi-machine collaborative navigation optimization method based on a graph neural network, which comprises the following steps: normalizing and splicing the number of pedestrians and the stenosis degree corresponding to each node to obtain the initial feature of each node; using a graph neural network model to obtain fusion features of each node, inputting a graph attention network model based on a self-defined distance attenuation mechanism, and after outputting target features of each node in single-machine navigation operation of the current robot, performing path planning on the current robot to generate a heuristic path of the current robot; performing path coincidence detection and adjustment operation on each robot to obtain a target heuristic path of each robot; and performing global updating on the node map according to a preset interval time, and regenerating a new heuristic path of each robot in combination with the updated current position of each robot. According to the method, the path planning efficiency is improved, navigation congestion caused by path overlapping is avoided, and the timeliness and accuracy of navigation decision are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of robot navigation, and in particular to a multi-robot collaborative navigation optimization method based on graph neural network. Background Art

[0002] In recent years, as a key technology for solving path planning and task allocation in multi-agent systems, multi-robot collaborative navigation technology has received extensive attention and rapid development. Traditional collaborative navigation methods usually rely on rule-making, mathematical optimization, or heuristic algorithms. These methods perform well in dealing with simple scenarios, but when facing complex dynamic environments or high-dimensional state spaces, they are often restricted by computational complexity and real-time constraints, and the specific manifestations are as follows:

[0003] (1) When generating initial feature points, existing path planning methods often use simple random or grid uniform sampling. In the case of complex obstacle layouts, the feature points will be unevenly distributed, making it difficult to quickly generate high-quality sampling points with reasonable distribution, resulting in an increase in the amount of calculation, reducing the accuracy and efficiency of path planning, and making it difficult to achieve ideal obstacle avoidance effects in the follow-up.

[0004] (2) During the path planning process, it is necessary to consider both global environmental information to avoid sub-optimal overall paths and local environmental features to cope with detailed changes. However, existing path planning methods mostly rely on local extraction methods based on adjacency relationships in node feature processing, lacking effective aggregation and comprehensive analysis of global graph structure information. As a result, in narrow areas or complex topological graphs, the optimal path cannot be recognized, and sub-optimal or even infeasible path planning schemes are easily generated. In addition, in the processing of the importance of node features in the graph structure, existing path planning methods mostly adopt fixed or heuristic allocation strategies. For example, simple weight calculations based on node degrees or distances ignore the importance changes of node features in dynamic environments and cannot dynamically adjust weight allocation, making it difficult to accurately identify the role of key nodes, and thus making it difficult to achieve an effective balance between global information aggregation and local weight allocation, affecting the effects and efficiency of obstacle avoidance and overall planning.

[0005] (3) Existing path planning methods mainly take path length or obstacle avoidance as optimization goals, ignoring dynamic information in the environment, such as pedestrian distribution and real-time pedestrian flow. They cannot make full use of complex features in the environment, resulting in the inability to avoid high pedestrian flow areas or complex obstacles in complex and dynamic environments, and thus unable to generate intelligent and safe paths, affecting the safety and efficiency of multi-robot collaborative navigation.

[0006] (4) In the multi-robot collaborative navigation task, when the target points of multiple robots are similar or close, using traditional single-robot path planning methods is likely to cause multiple robots to generate similar or overlapping paths. This path overlap not only increases the risk of collision or mutual blockage between robots but also reduces the overall navigation efficiency. Secondly, in terms of node scheduling in existing multi-robot collaborative navigation technologies, most follow static or simple rules and do not introduce priority scheduling. As a result, in dynamic obstacle avoidance scenarios or narrow areas, when the dynamic characteristics of key nodes (such as passing time, congestion situation) change, the path strategy cannot be adjusted in a timely manner. At the same time, due to insufficient prediction of the time consumption between nodes, in time-sensitive scenarios such as logistics distribution, the planned path is prone to failure, lacking reliability and practicality. In addition, during the multi-robot collaborative navigation process, existing technologies lack an effective real-time conflict detection and resolution mechanism, resulting in low navigation efficiency or conflicts in the multi-robot environment.

[0007] (5) In practical applications, information such as the distribution of pedestrians and the location of obstacles in the environment is dynamically changing, and node characteristics (such as the number of pedestrians, narrowness, etc.) also change over time. However, there is a lack of a real-time update mechanism for node characteristics in existing technologies, resulting in path planning that may be based on outdated information and unable to avoid new obstacles brought by crowded people or environmental changes in a timely manner. At the same time, there are deficiencies in the synchronization of the global information update frequency and path planning in existing technologies, resulting in information lag or inconsistency, thus affecting the navigation effect. Summary of the Invention

[0008] Therefore, the technical problems to be solved by the present invention are to overcome the problems in the prior art that the generation efficiency of feature points in the graph structure is low, resulting in reduced accuracy and efficiency of path planning; the limited ability to process node characteristics, making it difficult to balance the global perspective and local weight allocation, affecting the effect and efficiency of obstacle avoidance and overall planning; not considering the dynamically changing information in the environment, unable to make full use of environmental characteristics, resulting in insufficient fusion of global information and local information in path planning, affecting the safety and efficiency of multi-robot collaborative navigation; not introducing priority scheduling and insufficient prediction of node time, lacking a real-time multi-robot collaboration and conflict resolution mechanism, and the problems of path overlap and blockage in multi-robot path planning; the timeliness and dynamic update of node characteristics are insufficient, and there are defects in the synchronization of the global information update frequency and path planning, resulting in information lag or inconsistency, thus affecting the navigation effect.

[0009] To solve the above technical problems, the present invention provides a multi-robot collaborative navigation optimization method based on a graph neural network, including:

[0010] Based on the original map, obtain a node map; wherein, all nodes in the node map are evenly distributed in the non-obstacle area;

[0011] Based on the coordinates of each node, the current positions of each robot, and the current positions of each pedestrian, if the current position of any pedestrian is within a radius centered at the current node coordinate meters and within the lidar range of any robot in the node map, it is determined that there is the current pedestrian within a radius centered at the current node coordinate meters. Repeat this step to obtain the number of pedestrians corresponding to each node, and store the number of pedestrians corresponding to each node in the historical pedestrian flow database;

[0012] According to the obstacle distribution in the node map and the coordinates of each node, the proportion of obstacles within a radius centered at each node coordinate meters is used as the narrowness degree corresponding to each node;

[0013] Perform a normalization splicing operation on the number of pedestrians and the narrowness degree corresponding to each node to obtain the initial feature of each node;

[0014] Use a graph neural network model to update the initial feature of each node to obtain the fused feature of each node;

[0015] Perform a single - robot navigation operation on each robot in the node map in turn to obtain the heuristic path of each robot, including:

[0016] Input the fused feature of each node into a graph attention network model based on a custom distance attenuation mechanism, and output the target feature of each node in the current single - robot navigation operation of the current robot; wherein, the custom distance attenuation mechanism is to control the weight of nodes whose distance from the current robot exceeds the distance threshold according to the attenuation intensity control parameter;

[0017] Take the node closest to the current position of the current robot as the starting point of the current robot, and take the node closest to the target position of the current robot as the end point of the current robot; based on the target features of each node in the current single - robot navigation operation of the current robot, perform path planning on the current robot to generate the heuristic path of the current robot.

[0018] Preferably, before performing the single - robot navigation operation, number all the robots in the node map, and perform the single - robot navigation operation on each robot in turn according to the numbering order;

[0019] Perform a single - robot navigation operation on the first robot to generate the heuristic path of the first robot; store each node in the heuristic path of the first robot in the path node database;

[0020] Perform a single - robot navigation operation on the second robot to generate the heuristic path of the second robot;

[0021] Perform path overlap detection and adjustment operations on the second robot to obtain a target heuristic path for the second robot, including:

[0022] Compare each node in the heuristic path of the second robot with the nodes in the path node database. If there are overlapping nodes, use a priority-based node scheduling framework to perform overlap detection on the heuristic paths of the first robot and the second robot, and extract the overlapping node closest to the current position of the second robot as the current target overlapping node;

[0023] Based on the current positions of the first robot and the second robot, respectively obtain the distance between the current target coincidence node and the first robot, and the distance between the current target coincidence node and the second robot;

[0024] The first ellipse is drawn in the node map with the position of the current target coincident node and the position of the first robot as the long axis vertices of the first ellipse; the second ellipse is drawn in the node map with the position of the current target coincident node and the position of the second robot as the long axis vertices of the second ellipse;

[0025] Respectively obtain all nodes included in the first ellipse and all nodes included in the second ellipse, and respectively construct a node set of the first robot and a node set of the second robot;

[0026] After obtaining the sum of the historical pedestrian flows corresponding to the node set of the first robot and the sum of the historical pedestrian flows corresponding to the node set of the second robot from the historical pedestrian flow database, the time when the first robot reaches the current target coincident node is calculated as the first time point, and the time when the second robot reaches the current target coincident node is calculated as the second time point according to the relationship expression between the pedestrian flow and the robot speed and the time speed expression;

[0027] If the absolute value of the difference between the first time point and the second time point is greater than the time threshold, the heuristic path of the second robot is used as its target heuristic path, and each node in the target heuristic path of the second robot is stored in the path node database;

[0028] If the absolute value of the difference between the first time point and the second time point is less than or equal to the time threshold, all nodes in the heuristic path of the first robot in the node map are removed to obtain a new node map; based on the new node map, the second robot re-executes the single-machine navigation operation to obtain the target heuristic path of the second robot, and stores each node in the target heuristic path of the second robot in the path node database;

[0029] Perform single-robot navigation operation on the third robot to generate a heuristic path for the third robot; perform path coincidence detection operation on the third robot, compare each node in the heuristic path of the third robot with the nodes in the path node database. If there are coincident nodes, then adopt a node scheduling framework based on priority, and determine the heuristic path of coincidence detection and the target coincident node according to the situation of the coincident nodes.

[0030] Perform path coincidence detection and adjustment operations on each robot in the node map in sequence according to the number, and obtain the target heuristic path of each robot.

[0031] Preferably, the step of adopting a node scheduling framework based on priority and determining the heuristic path of coincidence detection and the target coincident node according to the situation of the coincident nodes includes:

[0032] If the coincident node only belongs to the nodes of the heuristic path of the first robot in the path node database, then adopt a node scheduling framework based on priority, perform coincidence detection on the heuristic path of the third robot and the heuristic path of the first robot, and extract the coincident node closest to the current position of the third robot as the current target coincident node.

[0033] If the coincident node only belongs to the nodes of the heuristic path of the second robot in the path node database, then adopt a node scheduling framework based on priority, perform coincidence detection on the heuristic path of the third robot and the heuristic path of the second robot, and extract the coincident node closest to the current position of the third robot as the current target coincident node.

[0034] If the coincident node belongs not only to the nodes of the heuristic path of the first robot in the path node database but also to the nodes of the heuristic path of the second robot in the path node database, then adopt a node scheduling framework based on priority. First, perform coincidence detection on the heuristic path of the third robot and the heuristic path of the first robot, and extract the coincident node closest to the current position of the third robot as the current target coincident node; then perform coincidence detection on the heuristic path of the third robot and the heuristic path of the second robot, and extract the coincident node closest to the current position of the third robot as the current target coincident node.

[0035] Preferably, the relational expression between the robot speed and the pedestrian flow is:

[0036] ;

[0037] Wherein, represents the speed of the th robot; represents the average speed of the robot; represents the The sum of the historical pedestrian flows corresponding to the set of feature points of the robot; Represents the historical pedestrian flow threshold; Represents a coefficient.

[0038] Preferably, the time - speed expression is:

[0039] ;

[0040] Wherein, Represents the time when the th robot reaches the current coincidence node; Represents the th robot's speed; Represents the th robot's distance from the current coincidence node.

[0041] Preferably, the expression of the custom distance attenuation mechanism is:

[0042] ;

[0043] Wherein, Represents the weight of the th node in the th robot's single - machine navigation operation; Represents the attenuation intensity control parameter; Represents the th robot's distance from the th node.

[0044] Preferably, based on the target features of each node in the current robot's single - machine navigation operation, using the A* algorithm, by connecting multiple nodes, path planning is performed for the current robot to generate the heuristic path of the current robot; wherein, the heuristic function of the A* algorithm is a node feature comparison function.

[0045] Preferably, the expression of the node feature comparison function is:

[0046] ;

[0047] Wherein, Represents the feature comparison value between the th node and the th node; Represents the target feature of the th node; Represents the target feature of the th node.

[0048] Preferably, after sequentially performing single - machine navigation operations on each robot in the node map and obtaining the heuristic path of each robot, it further includes:

[0049] Globally update the node map at a preset interval, and based on the current positions of each robot after the update, re-obtain the new number of pedestrians and narrowness of each node to obtain the new initial features of each feature point.

[0050] Based on the current position of each robot, re-perform the single-robot navigation operation to generate a new heuristic path for each robot.

[0051] Take each node in the heuristic path of each robot generated during different global update periods of the node map as the sub-goal of each robot during different global update periods of the planner in the Risk-RRT algorithm, which is used to guide each robot to navigate during different global update periods of the node map.

[0052] Preferably, use the Poisson disk sampling method to obtain the node map based on the original map; where the original map is in PGM format.

[0053] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0054] (1) For the multi-robot collaborative navigation optimization method based on graph neural network of the present invention, by using the Poisson disk sampling method to generate uniformly distributed feature points, it can more accurately represent the structural characteristics of the map, and at the same time avoid the feature points falling into the obstacle area, laying a high-quality foundation for subsequent path planning.

[0055] (2) For the multi-robot collaborative navigation optimization method based on graph neural network of the present invention, by combining the graph neural network model and the graph attention network model based on the custom distance attenuation mechanism to process the feature points, it fully integrates the global and local information, can extract more expressive node features, enhances the robustness and diversity of node feature expression, makes the input of path planning more accurate, and thus improves the overall effect of path planning; in addition, this method can obtain node features according to the dynamic environment information near the feature points (such as pedestrian distribution and regional narrowness), aggregate global information through the graph neural network model, and then realize the refined allocation of local weights through the graph attention network model based on the custom distance attenuation mechanism, so as to generate a heuristic path more in line with the current environment. This dynamic feature update mechanism effectively improves the adaptability of the robot in complex dynamic environments.

[0056] (3) The multi-robot collaborative navigation optimization method based on graph neural network according to the present invention fully considers dynamic environment information (such as pedestrian distribution and regional narrowness) and the cooperation problem among multiple robots, overcomes the deficiencies of blocked navigation paths, inflexible planning, and low cooperation efficiency in the prior art, and provides an innovative technical means for optimizing the robot navigation system. At the same time, a comparison mechanism of node features is introduced into the A* algorithm, enhancing the consideration of factors other than path length in path planning. This method can optimize the heuristic function by combining crowd distribution density and regional narrowness, preferentially select safer and smoother paths, and effectively improve the safety of robot navigation and the rationality of path selection.

[0057] (4) The multi-robot collaborative navigation optimization method based on graph neural network according to the present invention proposes a node scheduling framework based on priority and a path coincidence point detection and optimization mechanism, which can effectively avoid the congestion problem caused by the coincidence of multiple robot navigation paths. By adjusting the cost of the coincidence path points in the subsequent robot path planning and regenerating the path to avoid the coincidence points, the efficiency and stability of multi-robot collaborative navigation are improved, and resource conflicts are avoided.

[0058] (5) The multi-robot collaborative navigation optimization method based on graph neural network according to the present invention updates the global information at a set fixed time interval, and further realizes the purpose of real-time updating of node features, can maintain the timeliness of node features, ensure that the robot can accurately reflect the real-time state of the global environment when planning the path, thus avoiding crowded areas and improving the dynamic response ability of path planning. Brief Description of the Drawings

[0059] In order to make the content of the present invention easier to be clearly understood, the following further details the present invention according to specific embodiments of the present invention in conjunction with the drawings, wherein:

[0060] Figure 1 is a flowchart of a multi-robot collaborative navigation optimization method based on graph neural network provided by the present invention;

[0061] Figure 2 is the node map of Embodiment 1;

[0062] Figure 3 is a comparison graph of the historical pedestrian volume of nodes in each elliptical area of Embodiment 1;

[0063] Figure 4 is the node map of Embodiment 2;

[0064] Figure 5 is a schematic diagram of the generation of a single-robot heuristic path in Embodiment 2;

[0065] Figure 6It is a schematic diagram of the initial heuristic paths of the three robots in the second embodiment;

[0066] Figure 7 It is a schematic diagram of the heuristic paths before the global information update in the second embodiment;

[0067] Figure 8 It is a schematic diagram of the heuristic paths after the global information update in the second embodiment. Detailed implementation manners

[0068] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited are not intended to limit the present invention.

[0069] Embodiment 1

[0070] Refer to Figure 1 As shown, it is a flowchart of a multi-robot collaborative navigation optimization method based on a graph neural network provided by the present invention; specifically including:

[0071] S1: Using the Poisson disk sampling method, based on the original map, obtain a node map, as Figure 2 shown; wherein, all nodes in the node map are evenly distributed in the non-obstacle area; the original map is in PGM format;

[0072] S2: Based on the coordinates of each node, the current positions of each robot, and the current positions of each pedestrian, if the current position of any pedestrian is within a radius of meters centered on the current node coordinate and within the lidar range of any robot in the node map, it is determined that there is the current pedestrian within a radius of meters centered on the current node coordinate. Repeat this step to obtain the number of pedestrians corresponding to each node, and store the number of pedestrians corresponding to each node in the historical pedestrian flow database; wherein, the lidar range of any robot represents a lidar range with a radius of meters centered on the current position of the robot;

[0073] S3: According to the obstacle distribution in the node map and the coordinates of each node, take the proportion of obstacles within a radius of meters centered on each node coordinate as the narrowness degree corresponding to each node;

[0074] S4: Perform a normalization splicing operation on the number of pedestrians and the narrowness degree corresponding to each node to obtain the initial feature of each node;

[0075] S5: Update the initial features of each node using the graph neural network model to obtain the fused features of each node. When the model updates the node features, it fuses various node feature information and combines the adjacency relationships in the graph structure to extract features with a more global perspective. This fusion will enhance the expressive power of the features and provide more accurate inputs for subsequent tasks (such as path planning).

[0076] Among them, the Graph Convolutional Network (GCN) is a neural network model for graph-structured data, aiming to extract the feature representations of nodes by aggregating neighborhood node information. Its core idea is to generalize the convolutional operation from the traditional Euclidean space to the non-Euclidean space (such as the graph structure), enabling each node in the graph to update its own state by learning the features of its neighborhood. The process of the graph neural network mainly includes:

[0077] (1) Input representation of the graph: The input of the GCN model includes a node feature matrix and an adjacency matrix of the graph , where represents the number of nodes, represents the node feature dimension; the adjacency matrix represents the connectivity between nodes.

[0078] (2) Feature aggregation and update: The core operation of GCN is the aggregation of neighborhood features and the update of node features.

[0079] (3) Multi-layer stacking: By stacking multiple layers of GCN, feature aggregation for a larger neighborhood can be achieved, which is suitable for global feature extraction.

[0080] In a specific embodiment of the present invention, an unsupervised learning training method is adopted to train the graph neural network model using the initial features of each node to obtain a trained graph neural network for updating the features of each node.

[0081] S6: Perform single-robot navigation operations on each robot in the node map in turn to obtain the heuristic paths of each robot, including:

[0082] S61: Input the fused features of each node into the graph attention network model based on the custom distance attenuation mechanism, and output the target features of each node in the current single-robot navigation operation. Among them, the custom distance attenuation mechanism is to control the parameter according to the attenuation intensity and reduce the weights of the nodes whose distance from the current robot exceeds the distance threshold.

[0083] Among them, the expression of the custom distance attenuation mechanism is:

[0084] ;

[0085] Among them, represents the weight of the th node in the th single-robot navigation operation; represents the attenuation intensity control parameter; represents the distance between the th robot and the th node;

[0086] Among them, the Graph Attention Network (GAT) improves the neighborhood feature aggregation method of GCN by introducing an attention mechanism, enabling the influence of different neighborhood nodes on the target node to be adaptively adjusted through learnable attention weights; the main process of the graph attention network includes:

[0087] (1) Attention weight calculation: Before aggregating node features, the GAT model calculates the weight of each neighbor node through a self-attention mechanism;

[0088] (2) Normalization and feature aggregation: Normalize the attention weights of neighborhood nodes through the Softmax operation and use them for feature aggregation;

[0089] (3) Multi-head attention: To enhance the expression ability of the model, the GAT model usually adopts a multi-head attention mechanism, that is, calculates multiple independent attention weights simultaneously and merges the results;

[0090] In S61, the node features processed by the GCN model are transmitted to the GAT model with a custom distance attenuation mechanism added. For feature points farther from the current robot position, this model makes their features larger to prevent subsequent path planning from connecting nodes that are too far from the robot. That is, when calculating the weights of each node in the graph attention network model, through the custom distance attenuation mechanism and according to the attenuation intensity control parameter, the weights of nodes whose distance from the current robot exceeds the distance threshold are reduced; after the data processed by the GCN model passes through the GAT model again, the distribution of local weights is increased, making the feature values of each feature point consider both global aggregation and local weight distribution;

[0091] S62: Take the node closest to the current position of the current robot as the starting point of the current robot, and take the node closest to the target position of the current robot as the end point of the current robot; Based on the target features of each node in the current single-robot navigation operation, perform path planning for the current robot to generate a heuristic path for the current robot;

[0092] Among them, based on the target features of each node in the current single-robot navigation operation, using the A* algorithm, by connecting multiple nodes, path planning is performed on the current robot to generate a heuristic path for the current robot; among them, the heuristic function of the A* algorithm is a node feature comparison function; the present invention modifies the original heuristic function in the A* algorithm to a node feature comparison function, so that the A* algorithm not only considers the path length factor, but also adds the factor of node features (i.e., the information comparison of the global crowd);

[0093] Among them, the expression of the node feature comparison function is:

[0094] ;

[0095] Among them, represents the feature comparison value between the th node and the th node; represents the target feature of the th node; represents the target feature of the th node;

[0096] S63: During the process of performing single-robot navigation operations on each robot, it is also necessary to perform path coincidence detection and adjustment operations on each robot, including:

[0097] S631: Before performing the single-robot navigation operation, number all the robots in the node map, and perform single-robot navigation operations on each robot in sequence according to the numbering order;

[0098] S632: Perform a single-robot navigation operation on the first robot to generate a heuristic path for the first robot; store each node in the heuristic path of the first robot in the path node database;

[0099] S633: Perform a single-robot navigation operation on the second robot to generate a heuristic path for the second robot;

[0100] S634: Perform path coincidence detection and adjustment operations on the second robot to obtain the target heuristic path of the second robot, including:

[0101] Compare each node in the heuristic path of the second robot with the nodes in the path node database. If there are overlapping nodes, then use a node scheduling framework based on priority to perform coincidence detection on the heuristic paths of the first robot and the second robot, and extract the overlapping node closest to the current position of the second robot as the current target overlapping node;

[0102] Based on the current positions of the first robot and the second robot, respectively obtain the distances between the current target coincidence node and the first robot, and between the current target coincidence node and the second robot;

[0103] Taking the positions of the current target coincidence node and the first robot as the major axis vertices of the first ellipse, draw the first ellipse in the node map; taking the positions of the current target coincidence node and the second robot as the major axis vertices of the second ellipse, draw the second ellipse in the node map;

[0104] Respectively obtain all the nodes included in the first ellipse and all the nodes included in the second ellipse, and respectively construct the node set of the first robot and the node set of the second robot, as Figure 3 shown;

[0105] After obtaining the sum of the historical pedestrian flows corresponding to the node set of the first robot and the sum of the historical pedestrian flows corresponding to the node set of the second robot from the historical pedestrian flow library, according to the relationship expression between pedestrian flow and robot speed and the time-speed expression, calculate the time for the first robot to reach the current target coincidence node as the first time point, and the time for the second robot to reach the current target coincidence node as the second time point;

[0106] Among them, the relationship expression between the robot speed and the pedestrian flow is:

[0107] ;

[0108] Among them, represents the speed of the th robot; represents the average speed of the robot; represents the th sum of the historical pedestrian flows corresponding to the set of characteristic points of the robot; represents the historical pedestrian flow threshold; represents the coefficient;

[0109] The time-speed expression is:

[0110] ;

[0111] Among them, represents the time for the th robot to reach the current coincidence node; represents the th robot's speed; represents the th distance between the robot and the current coincidence node;

[0112] If the absolute value of the difference between the first time point and the second time point is greater than the time threshold, then use the heuristic path of the second robot as its target heuristic path, and store each node in the target heuristic path of the second robot into the path node database;

[0113] If the absolute value of the difference between the first time point and the second time point is less than or equal to the time threshold, then remove all the nodes in the heuristic path of the first robot from the node map to obtain a new node map; based on the new node map, the second robot re - executes the single - robot navigation operation to obtain the target heuristic path of the second robot, and store each node in the target heuristic path of the second robot into the path node database; among them, when the second robot re - executes the single - robot navigation operation, set the cost of the nodes where the second robot coincides with the first robot to be very large, so that the heuristic path regenerated by the second robot does not include these coincident nodes, that is, the two paths have no overlapping path points;

[0114] In summary, for the path node database, it is necessary to first initialize a path node database, and store the path list in the heuristic path of the first robot as the value into this database; among them, the path list in the heuristic path of the first robot includes the numbers of each node. For example, store the nodes numbered 1, 3, 5, 7 in the path list of the heuristic path of the first robot as the value into this database, and set the key of these values to "robot1"; then compare the nodes numbered 2, 5, 6, 8 in the heuristic path of the second robot with all the values in the path list in this database, and find that the node numbered 5 is a coincident node, and set the key of the node numbered 5 to "robot12", indicating the coincident node of "robot1" and "robot2";

[0115] S635: Perform a single - robot navigation operation on the third robot to generate the heuristic path of the third robot; perform a path coincidence detection operation on the third robot, compare each node in the heuristic path of the third robot with the nodes in the path node database. If there are coincident nodes, then use the node scheduling framework based on priority to determine the heuristic path for coincidence detection and the target coincident nodes according to the situation of the coincident nodes;

[0116] Among them, if the coincident nodes only belong to the nodes in the heuristic path of the first robot in the path node database, then use the node scheduling framework based on priority to perform coincidence detection between the heuristic path of the third robot and the heuristic path of the first robot, and extract the coincident node closest to the current position of the third robot as the current target coincident node;

[0117] If the overlapping node only belongs to the nodes of the heuristic path of the second robot in the path node database, a node scheduling framework based on priority is adopted to perform overlapping detection on the heuristic path of the third robot and the heuristic path of the second robot, and the overlapping node closest to the current position of the third robot is extracted as the current target overlapping node;

[0118] If the overlapping node belongs to not only the nodes of the heuristic path of the first robot in the path node database but also the nodes of the heuristic path of the second robot in the path node database, a node scheduling framework based on priority is adopted. First, perform overlapping detection on the heuristic path of the third robot and the heuristic path of the first robot, and extract the overlapping node closest to the current position of the third robot as the current target overlapping node; then perform overlapping detection on the heuristic path of the third robot and the heuristic path of the second robot, and extract the overlapping node closest to the current position of the third robot as the current target overlapping node;

[0119] S636: According to the numbers, perform path overlapping detection and adjustment operations on each robot in the node map in sequence to obtain the target heuristic path of each robot;

[0120] S7: Perform global update on the node map at a preset interval time. Combine the current positions of each robot after the update, and re-obtain the new number of pedestrians and narrowness of each node to obtain the new initial features of each feature point; among them, the present invention designs to perform global information update after seconds, aiming to make the features of the nodes have timeliness and can reflect the pedestrian flow distribution of the global map in real time, so as to avoid places with large pedestrian flow for the navigation path of the robot;

[0121] According to the current positions of each robot, re-execute the single-robot navigation operation to generate the new heuristic path of each robot;

[0122] S8: Use each node in the heuristic path of each robot generated during different global update periods of the node map as the sub-goals of each robot in the planner inside the Risk-RRT algorithm (Risk-based Rapid Random Tree algorithm) during different global update periods of the node map to guide each robot to navigate during different global update periods of the node map.

[0123] Embodiment 2

[0124] To verify the effectiveness of a multi-robot collaborative navigation optimization method based on graph neural network provided by the present invention, a demonstration is given in the scenario of three robots.

[0125] Execute the above S1 step to obtain the node map in Embodiment 2, as Figure 4 shown;

[0126] Perform single-robot navigation operations on a robot in the node map based on S2 - S62 to generate a heuristic path for the current robot, as Figure 5 shown;

[0127] Based on S63, there are three robots in the node map, namely robot , robot and robot . Now, robot , robot and robot perform single-robot navigation operations in sequence, as Figure 6 shown. It can be seen that the target points of robot and robot are the same, but their heuristic paths are very different, indicating that the "priority-based node scheduling framework" of the present invention is effective;

[0128] Based on S7, as Figure 7 shown, this is the heuristic path diagram of the three robots at the previous moment before global information update. It can be seen that there are several pedestrians in front of robot . Global information update is about to be performed, and single-robot navigation operations, path coincidence detection, and adjustment operations are re-performed on robot , robot and robot ;

[0129] After global information update, as Figure 8 shown, robot selects a heuristic path under the crowd to avoid the crowd; since no pedestrians are detected near robot and robot , their heuristic paths basically remain unchanged;

[0130] Based on S8, use the planner in the Risk-RRT algorithm to guide robot , robot and robot to navigate during the global update period of the current node map.

[0131] In summary, the present invention combines the robot lidar data, calculates the pedestrian flow (pedestrian_count) and narrowness around the node, and generates the initial features of the node through normalized splicing; uses the graph convolutional network (GCN) to aggregate and update the global information of the node features, introduces a distance attenuation mechanism to optimize the graph attention network (GAT), takes into account the global feature aggregation and local weight allocation, increases the attention to distant feature points, and improves the local path planning accuracy; introduces the comparison of node features into the heuristic function of the A* algorithm, optimizes the path planning with factors such as pedestrian flow and narrowness, avoids the irrationality of local paths caused by the single optimal length, comprehensively considers the path length, pedestrian flow and narrowness, and improves the rationality of the path planning result; for single-robot navigation, optimizes the path planning process and outputs a heuristic path that meets the multi-factor evaluation; in multi-robot collaborative navigation, performs coincidence detection and adjustment on the single-robot paths, detects path coincidence through the priority node scheduling framework, combines the historical pedestrian flow database to assist in path coincidence detection and adjustment, further improves the intelligence and dynamic adaptability of the path planning, and re-plans the path for the subsequent robot to avoid navigation jams caused by path overlap; designs a real-time update mechanism, inputs the latest pedestrian distribution into the GCN and GAT models every fixed time, and updates the node features using the trained weights to ensure that the navigation path can dynamically avoid crowded areas and ensure the timeliness and accuracy of navigation decisions; decomposes the heuristic path into several sub-goal points and gradually realizes the global navigation goal of the robot in combination with the Risk-RRT algorithm.

[0132] Obviously, the above embodiments are merely examples for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A multi-machine collaborative navigation optimization method based on graph neural network, characterized in that: include: Based on the original map, a node map is obtained; wherein all nodes in the node map are evenly distributed in a non-obstruction area; Based on the coordinates of each node, the current position of each robot and the current position of each pedestrian, if the current position of any pedestrian is within the radius of the current node coordinates meters and is within the laser radar range of any robot in the node map, the current node coordinates are taken as the center radius If there are current pedestrians within meters, repeat this step to obtain the number of pedestrians corresponding to each node, and store the number of pedestrians corresponding to each node in the historical pedestrian flow database; According to the obstacle distribution in the node map and the coordinates of each node, the radius is calculated based on the coordinates of each node. The ratio of obstacles within meters is used as the narrowness corresponding to each node; Perform normalized concatenation on the number of pedestrians and the degree of narrowness corresponding to each node to obtain the initial features of each node; Using the graph neural network model, the initial features of each node are updated to obtain the fused features of each node; Perform stand-alone navigation operations on each robot in the node map in turn to obtain the heuristic path of each robot, including: The fused features of each node are input into the graph attention network model based on the custom distance decay mechanism, and the target features of each node in the current robot stand-alone navigation operation are output; wherein the custom distance decay mechanism is to reduce the weight of the node whose distance from the current robot exceeds the distance threshold according to the decay strength control parameter; The node closest to the current position of the current robot is taken as the starting point of the current robot, and the node closest to the target position of the current robot is taken as the end point of the current robot; based on the target characteristics of each node in the stand-alone navigation operation of the current robot, the path of the current robot is planned to generate the heuristic path of the current robot.

2. According to the multi-machine collaborative navigation optimization method based on graph neural network in claim 1, it is characterized in that: Before performing the stand-alone navigation operation, all robots in the node map are numbered, and the stand-alone navigation operation is performed on each robot in sequence according to the numbering sequence; Performing a stand-alone navigation operation on the first robot to generate a heuristic path of the first robot; storing each node in the heuristic path of the first robot in a path node database; performing a single-machine navigation operation on the second robot to generate a heuristic path for the second robot; Perform path overlap detection and adjustment operations on the second robot to obtain a target heuristic path for the second robot, including: Compare each node in the heuristic path of the second robot with the nodes in the path node database. If there are overlapping nodes, use a priority-based node scheduling framework to perform overlap detection on the heuristic paths of the first robot and the second robot, and extract the overlapping node closest to the current position of the second robot as the current target overlapping node; Based on the current positions of the first robot and the second robot, respectively obtain the distance between the current target coincidence node and the first robot, and the distance between the current target coincidence node and the second robot; The first ellipse is drawn in the node map with the position of the current target coincident node and the position of the first robot as the long axis vertices of the first ellipse; the second ellipse is drawn in the node map with the position of the current target coincident node and the position of the second robot as the long axis vertices of the second ellipse; Respectively obtain all nodes included in the first ellipse and all nodes included in the second ellipse, and respectively construct a node set of the first robot and a node set of the second robot; After obtaining the sum of the historical pedestrian flows corresponding to the node set of the first robot and the sum of the historical pedestrian flows corresponding to the node set of the second robot from the historical pedestrian flow database, the time when the first robot reaches the current target coincident node is calculated as the first time point, and the time when the second robot reaches the current target coincident node is calculated as the second time point according to the relationship expression between the pedestrian flow and the robot speed and the time speed expression; If the absolute value of the difference between the first time point and the second time point is greater than the time threshold, the heuristic path of the second robot is used as its target heuristic path, and each node in the target heuristic path of the second robot is stored in the path node database; If the absolute value of the difference between the first time point and the second time point is less than or equal to the time threshold, all nodes in the heuristic path of the first robot in the node map are removed to obtain a new node map; based on the new node map, the second robot re-executes the single-machine navigation operation to obtain the target heuristic path of the second robot, and stores each node in the target heuristic path of the second robot in the path node database; Perform a stand-alone navigation operation on the third robot to generate a heuristic path for the third robot; perform a path coincidence detection operation on the third robot to compare each node in the heuristic path of the third robot with the nodes in the path node database; if there are coincident nodes, a priority-based node scheduling framework is used to determine the heuristic path for coincidence detection and the target coincident node according to the situation of the coincident nodes; According to the number, the path overlap detection and adjustment operations are performed on each robot in the node map in turn to obtain the target heuristic path of each robot.

3. According to claim 2, a multi-machine collaborative navigation optimization method based on graph neural network is characterized in that: The priority-based node scheduling framework is used to determine the heuristic path and target coincident node for coincidence detection according to the situation of coincident nodes, including: If the coincident nodes only belong to the nodes of the heuristic path of the first robot in the path node database, the priority-based node scheduling framework is used to perform coincidence detection on the heuristic path of the third robot and the heuristic path of the first robot, and the coincident node closest to the current position of the third robot is extracted as the current target coincident node; If the coincident node only belongs to the node of the heuristic path of the second robot in the path node database, the priority-based node scheduling framework is used to perform coincidence detection on the heuristic path of the third robot and the heuristic path of the second robot, and the coincident node closest to the current position of the third robot is extracted as the current target coincident node; If the overlapping node belongs not only to the node of the heuristic path of the first robot in the path node database, but also to the node of the heuristic path of the second robot in the path node database, a priority-based node scheduling framework is adopted to first perform an overlap detection on the heuristic path of the third robot and the heuristic path of the first robot, and extract the overlapping node closest to the current position of the third robot as the current target overlapping node; then perform an overlap detection on the heuristic path of the third robot and the heuristic path of the second robot, and extract the overlapping node closest to the current position of the third robot as the current target overlapping node.

4. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 2 is characterized in that: The relationship between the robot speed and pedestrian flow is expressed as: ； in, Indicates The speed of the robot; represents the average speed of the robot; Indicates The sum of historical pedestrian flows corresponding to the robot’s feature point set; represents the historical pedestrian flow threshold; Represents the coefficient.

5. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 2 is characterized in that: The time speed expression is: ； in, Indicates The time when the robot reaches the current coincident node; Indicates The speed of the robot; Indicates The distance between the robot and the current coincident node.

6. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 1 is characterized in that: The expression of the custom distance attenuation mechanism is: ； in, Indicates The robot single-machine navigation operation The weight of each node; represents the attenuation strength control parameter; Indicates The robot and The distance between nodes.

7. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 1 is characterized in that: Based on the target features of each node in the current stand-alone navigation operation of the robot, the A* algorithm is used to connect multiple nodes to plan the path of the current robot and generate a heuristic path for the current robot; among which, the heuristic function of the A* algorithm is a node feature comparison function.

8. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 7 is characterized in that: The expression of the node feature comparison function is: ； in, Indicates The node and The feature comparison value of each node; Indicates The target features of each node; Indicates The target feature of each node.

9. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 1 is characterized in that: Perform the single-machine navigation operation on each robot in the node map in turn, and after obtaining the heuristic path of each robot, it also includes: At preset intervals, the node map is globally updated, and the new number of pedestrians and narrowness of each node are re-acquired based on the current position of each robot after the update, so as to obtain the new initial features of each feature point. Based on the current position of each robot, the stand-alone navigation operation is re-executed to generate a new heuristic path for each robot; Each node in the heuristic path of each robot generated during different node map global update periods is used as the sub-goal of each robot in different node map global update periods inside the planner in the Risk-RRT algorithm to guide each robot to navigate during different node map global update periods.

10. The multi-machine collaborative navigation optimization method based on graph neural network according to claim 1 is characterized in that: A Poisson disk sampling method is used to obtain a node map based on an original map; wherein the original map is in a PGM format.

Images