A multi-robot formation path planning method

Abstract

The present invention discloses a multi-robot formation path planning method. The method comprises the following steps: 1) selecting a plurality of maps and generating a training path for each map according to a set conflict protection strategy to obtain a training data set; wherein each map is provided with a plurality of robots and target locations; 2) building a multi-robot path planning model, including a convolutional neural network, a Transformer sequence generation neural network, a GNN graph neural network and a Vote voting module; 3) using the training data set to train the multi-robot path planning model; 4) inputting a map provided with the initial positions and target positions of a plurality of robots into the optimized model to generate a path for each robot on the map. The present invention sets a dynamic confidence mechanism to vote on the next action results independently deduced by the two paths of the Transformer sequence generation neural network and the graph neural network, thereby improving the credibility of the path obtained by the robot.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a multi-robot formation path planning method based on a Transformer sequence generation neural network. Background Art

[0002] Multi-Agent Path Finding (MAPF) is a very challenging problem in the field of robotics research. Its main task is to find an optimal path from the starting point to the end point, and to avoid robots and obstacles along the way to ensure that no collisions occur. Therefore, how to realize how the robot reaches the destination in the shortest time and the safest route is the focus of research on this problem. Although the traditional single-robot solution that plans a path for each robot separately increases the success rate of the robot reaching the destination, this method works better in simple systems. As the number or density of robots increases, the efficiency of the entire system will decrease. The multi-robot solution plans the paths of all robots together, taking into account the possibility of various collisions, and greatly improves efficiency.

[0003] Under the same circumstances, the multi-robot solution can effectively schedule the paths of all robots so that they can reach their destination smoothly. Unlike the traditional method of calculating a complete path for each robot based on global information, the learning-based MAPF method uses local observations to determine the path decision on the length of the wire by learning the planning strategy of each robot. However, the existing learning-based MAPF algorithm cannot better avoid collisions. The key lies in the lack of effective features learned and shared between robots, which leads to the locking of some robots. At the same time, the aggregation of robots will also cause the accuracy of the training model to fluctuate. For each robot, all or part of the information of other robots is usually used to achieve collision-free. In multi-robot path planning, because the observation and communication distance is limited, the transmission of shared information between robots is affected, resulting in the robot being locked and unable to reach the target position ideally, causing the success rate trend to fluctuate during training. Summary of the Invention

[0004] To address the challenges of the existing technology, the present invention aims to provide a multi-robot formation path planning method based on a Transformer sequence-generated neural network. This method addresses how multiple robots within a region can reach their destination in the shortest possible time and with the safest route. It also addresses the issue of some robots being locked due to insufficient learning and sharing of effective features between robots, and addresses the accuracy and robustness of path planning models during autonomous navigation caused by clustered robots.

[0005] The technical solution of the present invention adopts the attention mechanism of the Transformer sequence generation neural network to perform global unified information fusion and scheduling, so that a locked robot can obtain effective data communication; a dynamic confidence mechanism is set to vote on the next action results of the two paths of the Transformer sequence generation neural network and the graph neural network (GNN) to independently deduce the next action results, so as to improve the credibility of the path obtained by the robot; the present invention proposes success rate and differential data variance evaluation indicators to verify the problems of robot locking and model accuracy oscillation.

[0006] The technical solution of the present invention is:

[0007] A multi-robot formation path planning method, the steps of which include:

[0008] 1) Select multiple maps and generate a training path for each map according to a set collision protection strategy to obtain a training data set; each map is set with multiple robots and target locations;

[0009] 2) Building a multi-robot path planning model; the multi-robot path planning model includes a convolutional neural network, a Transformer sequence generation neural network, a GNN graph neural network, and a Vote voting module;

[0010] 3) Using the training data set to train the multi-robot path planning model; the convolutional neural network extracts the information of each robot's field of view at time t in the input map as the field of view feature at time t and sends it to the Transformer sequence generation neural network and the GNN graph neural network respectively. The field of view feature of the i-th robot at time t is The Transformer sequence generation neural network fuses the vision features of all robots in the map and then uses the attention mechanism to transform them. The transformed features of each robot are obtained and sent to the Vote voting module. The vision feature of the i-th robot at time t after transformation is The GNN graph neural network sends the vision features of each robot at time t-1 to other robots within the robot's field of view at time t. Each robot predicts its own features at time t based on its own vision features at time t and the vision features of other robots at time t-1 received, and sends them to the Vote voting module. The i-th robot predicts its own features at time t as The Vote module predicts the movement path of each robot at time t based on all received features; i = 1 to N, where N is the total number of robots; for each robot, a predicted movement path is generated from the robot's initial position to the target position based on the robot's movement path at each moment; and the multi-robot path planning model is then optimized based on the deviation between the predicted movement path of each robot and the training path.

[0011] 4) A map with initial positions and target positions of multiple robots is input into the multi-robot path planning model optimized in step 3), and a path from the initial position to the target position is generated for each robot on the map.

[0012] The method according to claim 1 is characterized in that the collision protection strategy includes: if the inferred action of a robot will cause a collision with another robot or an obstacle, the action is replaced by an idle action; if the inferred actions of two robots will cause an edge collision, the corresponding two actions will be replaced by idle actions.

[0013] The method according to claim 1, wherein the multi-robot path planning model is a trained multi-robot path planning model; and the method for generating the data set used for training is:

[0014] 21) Place the same number of robots on maps of different sizes, set the starting and ending positions of each robot, and use the ECBS algorithm to generate a training path dataset;

[0015] 22) Change the map size and number of robots, set the starting and ending positions of each robot on the map, and use the ECBS algorithm to generate a training path dataset.

[0016] Furthermore, the convolutional neural network is composed of three layers of Conv convolution layer + BN normalization layer + ReLU nonlinear layer + Maxpooling maximum pooling layer and one layer of FC fully connected neural network connected in sequence. After the convolutional neural network feature extraction, the features of N robots at time t are obtained. is the visual field feature of the Nth robot at time t.

[0017] Furthermore, the Transformer sequence generation neural network includes a transformer branch model and an attention mechanism unit; first, the communication matrix W in the transformer branch model is initialized and trained according to the robot's field of view characteristics. Q 、W K and W VThe transformer branch model fuses and normalizes the robot’s field of view features and then inputs them into the attention mechanism unit. The attention mechanism unit processes the input data to obtain the features after the transformation space.

[0018] Furthermore, the GNN graph neural network generates a global connection graph G according to the visual field features received by each robot at time t t ; Then determine whether there are other robots in the robot's field of view at time t. If there are other robots in the robot's field of view, G t The corresponding position element in is updated to the inverse of the square of the Euclidean distance between the two robots; then according to the global connection graph G at time t t Calculate the graph translation operator S at time t t ; Then predict the characteristics of each robot

[0019] Furthermore, the Vote voting module cross-joins the features of the same robot output by the Transformer sequence generation neural network and the GNN graph neural network at time t, and then encodes the cross-joined features of each robot to obtain a coded feature Then, the dynamic confidence mechanism based on the Transformer structure is used to encode the features Voting is performed to obtain the moving path corresponding to each robot at time t.

[0020] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the above method.

[0021] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0022] This paper proposes a multi-robot continuous path planning method based on AI algorithm (Message-Aware Graph Transformer Networks, MSG Transformers). The specific technical solution is as follows:

[0023] Step 1: Set up a conflict prevention strategy and formulate the path planning problem of multiple robots as a sequential decision problem;

[0024] Step 2: Use ECBS (Enhanced Conflict-Based Search) to generate a training dataset. This dataset can reflect robot lockup and training accuracy oscillation.

[0025] Step 3: Define and train a network model to solve the robot locking and model accuracy oscillation problems, improve the robot's success rate in reaching the target position, and reduce model accuracy oscillation.

[0026] Step 4: Select evaluation indicators. Propose the evaluation indicator for evaluating the robot's successful arrival at the target position - success rate, and the evaluation indicator for evaluating the model's accuracy fluctuation - differential data variance;

[0027] Step 5: Compare the experimental results. Train the network model on the training set and check the training effect on the validation set. Finally, check the robot's arrival at the target position and the oscillation of the training model on the test set.

[0028] Furthermore, the conflict protection strategy in step 1 is specifically to implement a collision shielding protection mechanism, that is, the conflict of the planned path of the Agent in the ECBS algorithm is 0.

[0029] Furthermore, the data set of step 2 is specifically:

[0030] To ensure the optimization of locked robots and reduce model accuracy fluctuations, the dataset is mainly modified by increasing the map size to reduce the probability of robots meeting each other, forcing the robots to weaken their ability to obtain local information and thus form robot locks. The specific steps include:

[0031] Step 2.1: The same number of robots (20 robots) were trained on maps of different sizes (28×28, 40×40, and 50×50). After setting the starting and ending positions of the robots, the ECBS algorithm was used to generate the training path dataset.

[0032] Step 2.2: Change the map size (20×20 map, 28×28 map, 40×40 map, and 50×50 map) and the number of robots (10 robots and 20 robots). After setting the starting and ending positions of the robots, use the ECBS algorithm to generate a training path dataset.

[0033] Furthermore, the network model of step 3 is specifically as follows:

[0034] like Figure 2 As shown in the figure, the network model mainly consists of four parts: convolutional neural network, Transformer sequence generation neural network, GNN graph neural network and Vote voting module. The specific steps are as follows:

[0035] Step 3.1: Convolutional neural network performs feature extraction. The robot in each map extracts visual field features from the information within its field of view through a convolutional neural network.

[0036] Step 3.2: The Transformer sequence generation neural network layer introduces an attention mechanism. This uses relevance weighting to aggregate all robots' local map vision information. This allows for the integration of all vision resources for subsequent calculations, allowing the transformed spatial information to be more focused on learning.

[0037] Step 3.3: The GNN graph neural network uses a communication mechanism to transmit feature information within the robot's field of view to other robots within the field of view;

[0038] Step 3.4: A dynamic confidence mechanism based on the Transformer structure votes on the next action results of the Transformer sequence generation neural network and the GNN graph neural network for two independent paths generated by the same robot, thereby improving the credibility of the path obtained by the robot;

[0039] Step 3.5: Train the model, set the learning rate to 0.001 and the batch size to 640.

[0040] During the training process, the predicted value at each step is compared with the actual label to calculate the cross entropy loss for parameter update.

[0041] Furthermore, the experimental results of step 5 are more specific as follows:

[0042] When the number of robots remains unchanged, as the map size increases, the robot locking and model oscillation problems will become more and more serious.

[0043] Step 5.1: Compare and analyze the success rate trends of the GNN model alone, the Transformer model alone, the GNN+Transformer model with voting results biased towards the GNN, the GNN+Transformer model with voting results biased towards the Transformer, and this model. Determine the solution of the robot path search by the epoch when the success rate first reaches 0.9, the best success rate, and the epoch corresponding to the best success rate;

[0044] Step 5.2: Compare and analyze the success rate trends of each model, and determine the solution to the model oscillation problem through the difference mean, difference variance, and difference data variance of each model.

[0045] The advantages of the present invention are as follows:

[0046] (1) Compared with other multi-robot path planning methods, this method can extract important features and make decisions from local observation data, improve the success rate of robots reaching the target position and solve the problem of model accuracy oscillation.

[0047] (2) The attention mechanism of the Transformer sequence generation neural network is used to perform global unified information fusion and scheduling, so that robots in a locked state can obtain effective data communication.

[0048] (3) Voting on the next robot operation result on two independent paths, the Transformer sequence generation neural network and the GNN graph neural network, improves the credibility of the robot's obtained path. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the multi-robot continuous path planning method of the present invention.

[0050] Figure 2 This is a model structure diagram of the present invention.

[0051] Figure 3 This is a comparison chart of success rate results on different map configurations;

[0052] (a) 20×20 map with 10 robots, (b) 40×40 map with 20 robots, (c) 40×40 map with 20 robots, and (d) 50×50 map with 20 robots. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] To achieve the above objectives, the technical solution of the present invention is to use the attention mechanism of the Transformer sequence generation neural network to perform global unified information fusion and scheduling, so that a locked robot can obtain effective data communication; set up a dynamic confidence mechanism to vote on the next action results of the two paths of transformer and graph neural network (GNN) independently to improve the credibility of the path obtained by the robot; propose success rate and differential data variance evaluation indicators to verify the problems of robot locking and model accuracy oscillation.

[0055] This paper proposes a multi-robot continuous path planning method based on AI algorithms (Message-Aware Graph Transformer Networks, MSG Transformers). Compared with the solution of using only GNN for data communication, which results in the robot being unable to obtain valid data information and unable to reach the target location or increasing the time to reach the target location, we use the attention mechanism of the Transformer sequence generation neural network to replace the adjacency matrix with edge values ​​in part of the GNN graph neural network for global unified information fusion and scheduling, so that the locked robot can obtain effective data communication. In addition, we also use a dynamic confidence mechanism to vote on the next action results of the two paths of the Transformer sequence generation neural network and the GNN graph neural network independently to improve the credibility of the robot's path. The specific plan is as follows:

[0056] Step 1: Set up a collision prevention strategy. Using the ECBS algorithm, each robot needs to reach its destination as quickly as possible within the specified time range T while avoiding collisions with other robots and obstacles.

[0057] Step 2: Generate an experimental dataset in an environment with obstacles (the environment map is a grid area with a width of W steps and a height of H steps, with an area of ​​W×H) and a robot navigation scene dataset. The instantiated maps here include four categories: 20×20 maps for 10 robots, 28×28 maps for 20 robots, 40×40 maps for 20 robots, and 50×50 maps for 20 robots.

[0058] Step 3: Define the training network model and propose Message-Aware Graph Transformer Networks (MSGTransformers) to address robot locking and model accuracy fluctuations. This improves the robot's success rate in reaching the target location and reduces model accuracy fluctuations.

[0059] Step 4: Select evaluation indicators. The success rate is used to evaluate whether the robot reaches the target position, and the differential data variance is used to evaluate the model accuracy oscillation problem.

[0060] Step 5: Compare the experimental results. Compare and analyze the success rate and differential data variance of the only trans, only gnn, vote prefer trans, vote prefer gnn, and our model to determine the solution to the robot locking problem and model oscillation problem.

[0061] Furthermore, the conflict prevention strategy of step 1 is specifically as follows:

[0062] Step 1.1: If the inferred action would result in a collision with another robot or an obstacle, the action is replaced by an idle action;

[0063] Step 1.2: If the inferred actions of the two robots would result in an edge collision (making them swap positions), then those actions are replaced by idle actions.

[0064] Furthermore, the data set of step 2 is specifically:

[0065] Four types of maps are instantiated, and the obstacle density β is 10%. 600 maps of each type are used to optimize the robot locking and training accuracy oscillation problems, of which 480 are used for training, 60 for verification, and the remaining 60 for testing. The obstacle density β is calculated as shown in formula (1), where n robots is the number of robots, n obs is the number of obstacles.

[0066]

[0067] Furthermore, the network model of step 3 is specifically as follows:

[0068] Step 3.1: The convolutional neural network consists of 3 layers of Conv convolution layer + BN normalization layer + ReLU nonlinear layer + Maxpooling maximum pooling layer and one layer of FC fully connected neural network, which can obtain the features of each robot i within the field of view at time t After the convolutional neural network feature extraction, N robots can obtain new features at time t The formula is shown in (2);

[0069]

[0070] Step 3.2: Initialize and train the communication matrix W in the transformer branch model based on the robot's field of view map information Q 、W K and W V The transformer first transforms the robot’s field of view feature X t Fusion, d is the matrix W V The dimension size is then normalized, as shown in formula (3-1). Finally, the attention mechanism is introduced to centrally allocate resources through correlation weighting, so that the feature information after the conversion space More focused learning, such as formula (4-1);

[0071]

[0072]

[0073] At the same time, according to the robot's field of view map information, initialize and train the parameter matrix A in the GNN branch model k ,k represents the number of graph convolutions.

[0074] A global connection graph G can be obtained based on whether there are other robots in the field of view of each robot. t , represents the global robot connection status at time t. If there are other robots in the robot's field of view, the matrix G t The corresponding position element in is the reciprocal of the square of the Euclidean distance between the two robots. According to the global connection graph G t Graph-shift operator S t . At this point we can build the training structure GNN:

[0075]

[0076]

[0077]

[0078] The above model is a standard graph convolution and graph neural network structure. Every time the next motion path needs to be updated, the graph translation operator S is used. t The temporal relationship (4-1) before and after the robot's movement is described (3-2) and used in GNN to facilitate subsequent movement prediction (4-2); σ is a standard nonlinear function.

[0079] Step 3.3: The Vote layer cross-concatenates the features output by the Transformer layer and the GNN layer at time t Similarly, you can get all the robot splicing coding data in the map As shown in formula (5).

[0080]

[0081] Although attention will adaptively adjust the weight matrix a according to other data, its main attention still remains on itself, which makes the calculation of the weight coefficients of transformer and GNN more biased towards its own branch data, and the actual path planning effect is not good, as shown in formula (6).

[0082]

[0083] Therefore, according to the maximum value in the result, the influence of the two branch data biases with their own characteristics is eliminated and the next movement strategy of the robot is determined, W trans to pp 、W gun to ppIt is a fully connected neural network with only one layer as shown in formula (7).

[0084]

[0085] Step 3.4: Train the model, set the learning rate to 0.001 and the batch size to 640.

[0086] Furthermore, the evaluation indicators of step 4 are specifically:

[0087] Step 4.1: The success rate acc is the ratio of the number of robots that successfully reach the target position to the total number of test robots n. Calculate the success rate r Formula (8) is as follows. When all robots reach the goal before the timeout, the case is considered successful (completed);

[0088]

[0089] Step 4.2: For the problem of model accuracy oscillation, this paper proposes an evaluation criterion S based on the variance of differential data, as shown in formula (9). n ] is a set of data describing the success rate of the model at different epochs, and its differential data X diff =[x diff 2-1 ,x diff 3-2 ,…,x diff n-(n-1) ], The variance of the difference data is obtained μ is a minimum value used to prevent the variance from being equal to 0. When S>0: the larger S is, the more stable the precision oscillation curve rises; when S<0: the smaller S is, the more stable the precision oscillation curve falls.

[0090]

[0091] Furthermore, the experimental results of step 5 are more specific as follows:

[0092] The effectiveness of the model proposed in this paper is demonstrated through the following model results: only trans means only using transformer to obtain data communication between robots, and there is no voting mechanism; only gnn means only using GNN to obtain data communication between robots, and there is no voting mechanism; vote prefer trans means using transformer to perform self-attention on GNN results; vote prefer gnn means using GNN to perform self-attention on transformer results; our model uses a voting mechanism to vote on the next action results deduced by two independent paths, transformer and GNN.

[0093] Step 5.1: In order to make the indicators of the data distinguishable, we set the model’s success rate in subsequent epochs to be recorded when the accuracy of the model reaches or exceeds 90% for the first time, such as Figure 3 The success rate (acc) results of each model on different map configurations.

[0094] from Figure 3 As the map size increases, the success rate of each model decreases. Furthermore, as robot lock-up increases, the GNN-only model, which uses only GNN to communicate inter-robot data, significantly underperforms our model, which incorporates an attention mechanism. During the model ramp-up phase, our model remains trainable, achieving the largest best acc appear epoch. During the model oscillation phase, our model reaches the oscillation phase faster, achieving the smallest best acc appear epoch. Overall, although our model's time to 90% accuracy is slightly lower than that of the Trans-only model, it is faster than the GNN-only, Vote Prefer Trans, and Vote Prefer GNN models.

[0095] Step 5.2: The epoch when the first success rate of each model (only trans, only gnn, vote prefer trans, vote prefergnn, and our model) on maps of different sizes is 0.9, the best success rate, and the epoch corresponding to the best success rate are shown in Table 1.

[0096] Table 1: Success rate of each model

[0097]

[0098]

[0099] Table 2 shows the success rate data for each model. On a 20×20 map with 10 robots, each model achieved the highest success rate and the lowest oscillation. Our model achieved the highest accuracy and the latest arrival time, while Transformer achieved the fastest time to 90% accuracy. This indicates that models may not converge even when reaching or exceeding 90% accuracy, and that the Transformer can effectively accelerate model convergence. Furthermore, our model leverages the Transformer's rapid arrival time, effectively resolving the robot lockup issue.

[0100] Step 5.3: The difference mean, difference variance and difference data variance of each model (only trans, only gnn, vote prefer trans, vote prefergnn and our model) on maps of different sizes are shown in Table 2.

[0101] Table 2: Difference mean, difference variance and difference data variance of each model

[0102]

[0103] From the experimental data of the differential mean, differential variance, and differential data variance for each model in Table 2, it can be seen that in each group of models, the method proposed in this invention is far ahead of all other models in the differential data variance indicator, and the model can rise both quickly and steadily. In addition, the differential mean, which determines the ability to rise data, is the highest, and the differential variance, which determines the degree of offline data, is the lowest. This shows that the method proposed in this invention not only learns the ability of the Transformer sequence generation neural network to increase accuracy, but also avoids the problem of model accuracy fluctuation caused by robot aggregation.

[0104] While specific embodiments of the present invention have been disclosed for illustrative purposes, intended to facilitate understanding and implementation of the present invention, those skilled in the art will appreciate that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments, and the scope of protection claimed in the present invention shall be determined by the scope of the claims.

Claims

1. A multi-robot formation path planning method, comprising the following steps: 1) Select multiple maps and generate a training path for each map according to the set conflict protection strategy to obtain a training data set; Each map is equipped with multiple robots and target locations; 2) Build a multi-robot path planning model; the multi-robot path planning model includes a convolutional neural network, a Transformer sequence generation neural network, a GNN graph neural network, and a vote voting module; 3) Using the training data set to train the multi-robot path planning model; the convolutional neural network extracts the information of each robot's field of view at time t in the input map as the field of view feature at time t and sends it to the Transformer sequence generation neural network and the GNN graph neural network respectively. The field of view feature of the i-th robot at time t is The Transformer sequence generation neural network fuses the vision features of all robots in the map and then uses the attention mechanism to transform them. The transformed features of each robot are obtained and sent to the Vote voting module. The vision feature of the i-th robot at time t after transformation is The GNN graph neural network sends the vision features of each robot at time t-1 to other robots within the robot's field of view at time t. Each robot predicts its own features at time t based on its own vision features at time t and the vision features of other robots at time t-1 received, and sends them to the Vote voting module. The i-th robot predicts its own features at time t as The Vote module predicts the movement path of each robot at time t based on all received features; i = 1 to N, where N is the total number of robots; for each robot, a predicted movement path is generated from the robot's initial position to the target position based on the robot's movement path at each moment; and the multi-robot path planning model is then optimized based on the deviation between the predicted movement path of each robot and the training path. 4) A map with initial positions and target positions of multiple robots is input into the multi-robot path planning model optimized in step 3), and a path from the initial position to the target position is generated for each robot on the map.

2. The method according to claim 1, characterized in that The collision prevention strategy includes: if the inferred action of a robot will cause a collision with another robot or an obstacle, the action will be replaced by an idle action; if the inferred actions of two robots will cause an edge collision, the corresponding two actions will be replaced by idle actions.

3. The method according to claim 1, characterized in that The multi-robot path planning model is a trained multi-robot path planning model; the method for generating the data set used for training is: 21) Place the same number of robots on maps of different sizes, set the starting and ending positions of each robot, and use the ECBS algorithm to generate a training path dataset; 22) Change the map size and number of robots, set the starting and ending positions of each robot on the map, and use the ECBS algorithm to generate a training path dataset.

4. The method according to claim 1, 2 or 3, characterized in that: The convolutional neural network is composed of three layers of Conv convolution layer + BN normalization layer + ReLU nonlinear layer + Max pooling layer and one layer of FC fully connected neural network connected in sequence. After the convolutional neural network feature extraction, the features of N robots at time t are obtained. is the visual field feature of the Nth robot at time t.

5. The method according to claim 1, 2 or 3, characterized in that: The Transformer sequence generation neural network includes a transformer branch model and an attention mechanism unit. First, the communication matrix W in the transformer branch model is initialized and trained according to the robot's field of view characteristics. Q 、W K and W V ; The transformer branch model fuses and normalizes the robot's field of view features and then inputs them into the attention mechanism unit. The attention mechanism unit processes the input data to obtain the features after the transformation space.

6. The method according to claim 1, 2 or 3, characterized in that: The GNN graph neural network generates a global connection graph G according to the visual field features received by each robot at time t t ; Then determine whether there are other robots in the robot's field of view at time t. If there are other robots in the robot's field of view, G t The corresponding position element in is updated to the inverse of the square of the Euclidean distance between the two robots; then according to the global connection graph G at time t t Calculate the graph translation operator S at time t t ; Then predict the characteristics of each robot 7. The method according to claim 1, 2 or 3, characterized in that: The Vote voting module cross-joins the features of the same robot output by the Transformer sequence generation neural network and the GNN graph neural network at time t, and then encodes the cross-joined features of each robot to obtain a coded feature Then, the dynamic confidence mechanism based on the Transformer structure is used to encode the features Voting is performed to obtain the moving path corresponding to each robot at time t.

8. A server, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images