Multi-robot collaborative navigation method based on anonymous perception enhancement and adaptive decision

Abstract

The invention discloses a multi-robot collaborative navigation method based on anonymous perception enhancement and adaptive decision. The method comprises the following steps: S1, extracting local environment features based on a residual network; s2, fusing an anonymous perception enhancement mechanism of relative position information; s3, decision action probability distribution prediction and priority prediction; s4, adaptively coordinating decision actions based on priority learnable PIBT; according to the method, a message fusion mechanism based on position coding is designed in the feature hidden space to solve the problem of'message anonymity ', and a priority learnable PIBT strategy is introduced into a decision layer to solve the problems of unreasonable priority distribution and priority inversion, so that generalization and flexibility of multi-robot collaborative navigation are improved; efficient collaborative navigation in a distributed communicable environment can be realized.

Description

TECHNICAL FIELD

[0001] The application relates to a robot navigation method, in particular to a multi-robot cooperative navigation method based on anonymous perception enhancement and adaptive decision-making, and belongs to the technical field of cooperative navigation in a multi-robot system. BACKGROUND

[0002] With the wide application of multi-robot systems in warehouse logistics, intelligent manufacturing and urban distribution, efficient and conflict-free cooperative navigation has become the key to ensuring the safety and efficiency of multi-robot cooperative operation. Multi-robot cooperative navigation refers to the linkage of perception, communication and decision-making mechanism among multiple robots, in the absence of central control, to design an independent and continuous motion trajectory for each robot from the starting point to the ending point according to the environmental dynamics and task objectives, and to effectively avoid collisions through strict spatiotemporal constraints management, so as to achieve optimal scheduling and cooperative operation of global resources.

[0003] According to the use mode of global information, multi-robot cooperative navigation methods can be divided into centralized and distributed methods. The centralized method can obtain a globally optimal or near-optimal solution through unified planning of all robots by central management, but due to the high concentration of computing resources and decision-making power, it usually brings huge communication overhead and poor scalability, and once the central processing unit fails, the entire system may be paralyzed. The distributed method relies on the local field of view of each robot to generate independent decisions and exchange information with neighboring robots through a distributed coordination mechanism, thereby realizing globally conflict-free cooperative navigation. This design significantly improves the robustness and reliability of the system. However, due to the lack of global perspective, this method is prone to conflicts and may even lead to conflicts or deadlocks in severe cases. In order to solve this problem, current research introduces a communication mechanism, mainly focusing on two key collaborations: 1) feature-level collaboration: aggregating the observed features of neighbors to enhance the understanding of dynamic environments by robots. 2) decision-level collaboration: using a conflict shielding mechanism to solve action conflicts and ensure path safety.

[0004] In the distributed approach, graph neural networks have become a frontier tool for handling local communication and coordination among robots. Graph neural networks model each robot as a node in the graph and communication links as edges in the graph, and fuse neighbor information in a graph convolution or message passing manner, thereby capturing the relational structure of the system. This approach is naturally suitable for dynamic communication graphs of multi-robot systems and can effectively learn information exchange and coordination strategies among robots. However, most existing methods only implicitly extract location information from local grid observations, such as the surrounding obstacle layout, the positions of other robots, and the projection of the target, without using explicit node identity cues. This anonymity of the message source undermines the model's ability to perceive the relative relationships between robots in dynamic environments, increasing the risk of conflicts. Position encoding is a technique that has been extensively studied in the field of deep learning in recent years, and relative position encoding in particular has been used in sequence models such as Transformers and BERT to enhance the model's perception of spatial relationships between elements. In recent years, there have been efforts to encode the relative positions of robots as edge features in graphs to encode distance and direction information between robots. For example, in graph-based multi-robot autonomous collaboration, researchers set the feature of each edge to be the relative vector of the robot positions and pass it to the graph neural network after encoding through a linear transformation. This allows the model to more directly utilize spatial topology information and improves robustness to changes in environmental geometry. However, there have been few studies combining relative position encoding with graph neural networks, and the advantages of relative position encoding in message fusion have not been fully utilized.

[0005] In addition, conflict avoidance mechanisms at the decision-making layer are crucial in achieving collision-free planning. Since each robot makes independent decisions based on its own local strategy, it often leads to spatial conflicts between neighboring robots. Existing conflict avoidance strategies mainly fall into two categories: one is to perform a "freeze" operation on robots that may cause conflicts, i.e., to let them stay and wait; the other is to assign priorities to each robot and plan in order of priority. The former can quickly eliminate conflicts, but can fall into a deadlock state and cannot fully utilize the complete motion probability distribution learned by the robots. The latter has some scheduling flexibility, but often causes priority inversion problems due to improper priority assignment, leading to a decline in overall system performance.

[0006] In summary, existing multi-robot collaborative navigation methods have the following main defects:

[0007] Message anonymity leads to insufficient spatial relationship modeling: Many distributed planning or learning methods rely only on local observation features to implicitly perceive the location relationships of surrounding robots, making it difficult to accurately know the source of the message (anonymity) and flexibly capture the constantly changing relative spatial relationships between robots. The lack of effective relative position encoding methods results in insufficient generalization ability of the model to environmental changes.

[0008] The unreasonable priority allocation leads to the priority inversion of decision: although the conflict avoidance algorithm of the current frozen conflict robot can guarantee that no conflict movement occurs, it is easy to fall into the situation of congestion deadlock. Although other priority-based methods can alleviate the above problems to a certain extent, they are sensitive to priority allocation, and if the priority allocation is improper, the priority inversion problem will occur, resulting in the decline of the overall performance of the system.

[0009] The insufficient collaboration capability leads to poor system scalability: many learning collaborative navigation methods are trained in small-scale scenarios, and when the number of robots or the size of the scene increases significantly, the system performance often decreases significantly. Existing research shows that large-scale multi-robot collaborative navigation problem is still an open challenge, and even if the graph neural network technology is used, the success rate and integrity of the current method in a larger scene or high-density obstacle environment are obviously insufficient. SUMMARY

[0010] The purpose of the present application is to provide a multi-robot collaborative navigation method based on anonymous perception enhancement and adaptive decision, which designs a message fusion mechanism based on position coding in the feature latent space to solve the "message anonymity" problem, and introduces a priority learnable PIBT strategy in the decision layer to solve the unreasonable priority allocation and priority inversion problem, improve the generalization and flexibility of multi-robot collaborative navigation, and realize efficient collaborative navigation in a distributed communicable environment.

[0011] In order to achieve the above purpose, the present application provides a multi-robot collaborative navigation method based on anonymous perception enhancement and adaptive decision, comprising the following steps:

[0012] S1: local environment feature extraction based on residual network; each robot encodes its local observed environment by using a perception network based on ResNet structure, and extracts local observation features with spatial representation capability from it, providing a unified latent space basis for subsequent message fusion and action decision;

[0013] S2: anonymous perception enhancement mechanism for fusing relative position information; in the message fusion process, a relative position coding mechanism is introduced for the robots within the communication range to solve the anonymity problem. Specifically, the spatial position information of the neighbor robot relative to itself is coded into a structured vector, and is combined with the local observation features transmitted by the neighbor robot to generate "anonymous perception enhancement features". The fusion is carried out simultaneously in the process of message transmission and reception, thereby enhancing the perception of the surrounding environment while maintaining the consistency of the latent space;

[0014] S3: Decision action probability distribution prediction and priority prediction; each robot learns high-quality action strategy and implicit scheduling priority from expert demonstration data provided by traditional centralized method with global view through imitation learning (expert data is generated by using ECBS algorithm in this patent), specifically, each robot uses two multilayer perceptrons respectively to predict the action probability distribution that can be taken by itself next based on the anonymous perception enhanced features obtained by using step S2, and the priority of action execution of each robot in the current local communication range;

[0015] S4: PIBT adaptive coordination decision action based on learnable priority; an improved priority inheritance and backtracking strategy PLPIBT is designed, which effectively enhances the rationality of system priority allocation by using the priority learned in step S3, and when the system detects a priority inversion situation that may cause traffic congestion, the PIBT mechanism is automatically triggered, the current robot temporarily inherits the high priority robot's permission, and the traffic order and execution action between robots are dynamically adjusted by combining the action probability distribution predicted in step S3, so as to realize adaptive decision between robots.

[0016] The anonymous perception enhancement mechanism based on relative position information proposed in step S2 of the application specifically includes the following two sub-steps:

[0017] S21: Constructing position code based on discrete position in communication range; the position of the neighbor robot is discretized and numbered in the local grid space, and a relative position code with different spatial perception ability is constructed to provide structural prior for subsequent message fusion;

[0018] S22: Neighbor message fusion based on relative position code; when the graph attention mechanism is used to fuse the neighbor message, the relative position code obtained in step S21 is embedded into the local observation feature of the neighbor robot, and finally the anonymous perception enhanced feature containing the local observation feature of itself, the local observation feature of the neighbor and the relative position information is generated.

[0019] The action decision probability distribution prediction and priority prediction proposed in step S3 of the application specifically includes the following two sub-steps:

[0020] S31: Robot action probability distribution learning based on anonymous perception enhanced features; learn and fit the action decision-making strategy in the expert demonstration data through a neural network, the expert strategy used is derived from a centralized planner with a global view, which has strong decision-making ability; in the training stage, the system takes the anonymous perception enhanced features fused in step S2 as input to predict the action probability distribution that the robot may take next, and the strategy network is aligned with the expert demonstration through the cross-entropy loss function, so that the robot can approximately reproduce the high-quality decisions of centralized planning in a distributed environment;

[0021] S32: Robot priority learning based on anonymous perception enhanced features; this step also uses a neural network to capture the scheduling logic implied in the expert centralized planner, and realizes dynamic learning and prediction of priority; the network input is the anonymous perception enhanced features generated in step S2, and the output is the action execution priority score of the robot in the current local communication range, the learned priority reflects both task urgency and potential conflict risk in the neighborhood, to guide the robot to make autonomous trade-off between advancing and yielding.

[0022] The PIBT adaptive coordination decision action based on learnable priority proposed in step S4 of the application specifically includes the following three sub-steps:

[0023] S41: Action sampling and preliminary conflict detection based on action probability distribution; the robot samples actions according to the action probability distribution predicted in S31, and detects conflicts of the sampled actions within the local communication range, if no conflict is detected, the robot can directly execute the sampled action without further coordination;

[0024] S42: Conflict resolution based on the order execution mechanism of learnable priority; if a conflict is detected, the system will refer to the robot priority score learned in S32, and execute the candidate actions in order of priority, thereby achieving effective conflict avoidance;

[0025] S43: Backtracking coordination strategy based on priority inheritance to deal with priority inversion problem; during the sequential execution, if priority inversion occurs, i.e. a low-priority robot blocks a high-priority robot, the priority inheritance mechanism in PLPIBT is triggered to temporarily raise the priority of the key robot in the blocked path, to cooperatively complete local conflict resolution and path unblocking, and prevent the system from entering a deadlock state.

[0026] Compared with the prior art, the model can solve the message anonymity problem by introducing a relative position coding mechanism, thereby significantly enhancing the perception ability of the environmental topology. The coding method embeds the geometric relationship between nodes in the graph into the attention calculation process, which not only improves the accuracy of feature fusion, but also enhances the migration ability and robustness of the model when facing unseen topological structures.

[0027] The priority learning mechanism introduced in the application enables the robot to dynamically adjust its priority based on local observations, overcoming the congestion and priority inversion problems caused by fixed or regularized priorities in traditional methods. The learned action probability distribution combined with the improved PIBT coordination strategy enables the system to quickly generate a conflict-free actionable path, effectively reducing the conflict occurrence rate between robots, avoiding robot deadlock, and improving overall navigation efficiency.

[0028] The application is completely based on a distributed architecture design, and the robot only performs local perception and decision-making coordination within the communication radius, without the need for a central scheduler to coordinate. While maintaining high success rate and high coordination efficiency of the system, the dependence on communication bandwidth and real-time performance is greatly reduced. Even when the number of robots expands to tens or hundreds, the system still has good scalability and task completion efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A schematic diagram of multi-robot path planning in a distributed communicable environment in the application;

[0030] Figure 2 A process of extracting corresponding relative position codes from the designed position coding matrix in the application;

[0031] Figure 3 A message fusion calculation process with position coding in the application;

[0032] Figure 4 A whole framework workflow diagram of multi-robot cooperative navigation planning in the application;

[0033] Figure 5 Comparison results of the algorithm proposed in the application and three distributed algorithms in four evaluation indicators, wherein figures (a) to (d) correspond to the comparison results of success rate (SR), makespan growth rate (MIR), flowtime growth rate (FIR), and predicted conflict rate (PCR) four indicators on the same density dataset; figures (e) to (h) show the change trend and performance of the four indicators on the dataset with increasing density;

[0034] Figure 6Figure 1 shows the convergence process of the proposed algorithm and three distributed algorithms on the training set for various metrics. Figure 2 shows the convergence curve of SR, Figure 3 shows the convergence curve of MIR, Figure 4 shows the convergence curve of FIR, and Figure 5 shows the convergence curve of PCR. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0036] like Figure 1 As shown, in the communication-based distributed multi-robot navigation method, each robot utilizes its own local observations while sharing observation or decision-making information with neighboring robots through communication mechanisms, thereby promoting environmental perception or decision coordination. All robots in this invention reside in a W×H grid environment and have a fixed w×h local field of view (…). Figure 1 The field of view is represented by a box of the same color as the robot, which helps to extend the strategy to any world scale. Each robot can communicate with other robots within its radius r (for example, robots that communicate with robot 3 are connected by red lines).

[0037] For a system consisting of n robots, its communication topology at time t can be represented by graph G. t =(V, E) t W t ) represents, where V = (v1, v2, ..., v n () is a collection of robots. Let V be the set of edges at time t; if (v i v j )∈E t This indicates that robots i and j can communicate, meaning the distance between them is less than r. E t Using the adjacency matrix S t This means that if (v i v j )∈E t but otherwise W t The weight of the edge represents the relative importance (attention) of the message received from the neighboring robot, and is calculated by the position encoding and class key query attention mechanism proposed in step S1 (the thickness of the red line in the figure represents the importance of the message).

[0038] In a discrete environment, to avoid conflict, the robot can move to an adjacent grid cell or remain in place (wait). Therefore, this invention divides the robot's actions into five discrete actions: up, left, down, right, and wait, and represents them using one-hot encoding for easy use in model learning.

[0039] To realize the cooperative navigation of multiple robots in a distributed communicable environment, as shown in Figure 2 and Figure 3 The present application proposes a multi-robot cooperative navigation method based on anonymous perception enhancement and adaptive decision-making, which integrates hidden space (feature level) and decision level coordination, and the specific implementation steps are as follows:

[0040] S1: Local environment feature extraction based on residual network; first, the observation information collected by the robot is input into the convolutional neural network (CNN) to improve the channel dimension and enhance the expression ability of the bottom layer features. Then, the features are sent to the residual network (ResNet) composed of multiple residual modules to extract local observation features with certain spatial structure perception ability. The extracted features are flattened and mapped to a unified hidden space representation through multi-layer perception (MLP) to provide a consistent feature basis for subsequent neighbor message fusion and adaptive decision-making.

[0041] As shown in Figure 4 the observation information collected by the robot is a three-channel input matrix with a shape of (2l+1)×(2l+1), where l is the observation radius, the first channel encodes part of the observed environment of the i-th robot, where 1 represents an obstacle and 0 represents a passable area; the second channel represents the target position of the i-th robot or its projection on the boundary of the local view; the third channel encodes the self robot i located at the center and other robots within its field of view.

[0042] S2: Anonymous perception enhancement mechanism for fusing relative position information; message anonymity can be solved by assigning a unique ID to each node. However, in the distributed robot navigation scenario, robots constantly enter and leave the limited communication range, so it is not practical to assign a globally unique ID to each robot, so by assigning a unique ID to the grid cell within the communication range, any robot occupying the cell can use its corresponding grid ID as its own ID. The position encoding of the present application is a mature and effective method to generate such ID.

[0043] By introducing a relative position encoding mechanism, the spatial position information of other robots relative to themselves is encoded into structured vectors, and feature fusion is performed by combining these vectors with local observation features transmitted from neighbors. This fusion process occurs during message propagation and reception, effectively alleviating the problem of insufficient spatial modeling caused by robot anonymity, thereby improving the robot's ability to perceive the state of its companions in the surrounding environment. Specifically, in the interaction between robot i and robot j within its communication range, the attention score not only depends on the feature representations of both robots but also explicitly introduces their relative position encoding. The attention score can be formally expressed as:

[0044]

[0045] Where: W is a parameter matrix; x i The features extracted from a graph neural network layer on robot i, x i T It is x i The transpose of x, if it is the 0th level, then x i =v i ;x j It is the feature extracted by a graph neural network layer on the neighboring robot j within the communication range; pos ij This represents the position of robot j relative to robot i; if pos ij =k, which means that robot j is located in the (2r+1)×(2r+1) local grid map of robot i. Row (k%r) column, r is the communication radius; P is the designed relative position coding matrix, P(pos ij ) represents the position code of robot j relative to robot i, so it can only be a relative position code. The calculation method of part a of the formula is the same as that of the original attention score. Part b of the formula contains the core of the proposed relative position code, which is the solution to the anonymity problem.

[0046] The anonymous perception enhancement mechanism that integrates relative position information proposed in step S2 of this invention specifically includes the following two sub-steps:

[0047] S21: Construct position codes based on discretized positions within the communication range; Discretize and number the positions of neighboring robots in the local grid space, and construct relative position codes with different spatial perception capabilities to provide structural priors for subsequent message fusion;

[0048] This invention explicitly encodes the relative position information of other robots within the communication range and embeds it into a graph attention mechanism, thereby enhancing the model's ability to perceive spatial structure. Inspired by position encoding strategies in sequence models, this invention proposes three forms of relative position encoding mechanisms. Figure 2The process of extracting the corresponding relative position encoding from the designed position encoding matrix P according to the discrete number of the neighbor robot in the local grid space is shown.

[0049] Fixed non-learnable relative position encoding; inspired by the position encoding strategy in the Transformer model, the present application introduces a fixed, non-learnable sine-cosine position encoding method. This method generates a unique vector representation for the position through periodic functions of different frequencies, without participating in training, with high computational efficiency, and is suitable for multi-robot systems with relatively limited resources or fast deployment. The calculation formula of the fixed position encoding is as follows:

[0050]

[0051] where: pos ij represents the position of robot j relative to robot i, and d is the dimension of the position vector.

[0052] If pos ij =m, then P(pos ij ,2k) and P(pos ij ,2k+1) represent the kth scalar and the k+1th scalar of the mth position vector of robot i, respectively.

[0053] Learnable relative position encoding; inspired by the learnable position encoding mechanism in the BERT model, the present application introduces a learnable relative position encoding method, which enables the model to automatically capture more complex and flexible spatial relationships. The present application presets a trainable embedding matrix for different relative positions, each relative position corresponds to an embedding vector, which is continuously optimized through gradient backpropagation during training. Compared with the fixed sine-cosine encoding, this method has the following advantages: it can automatically learn task-related position features without manually setting encoding rules; it has stronger expression ability and can adapt to more complex environments.

[0054] Learnable bias relative position encoding; inspired by the T5 model, a learnable bias-based position encoding is proposed. Unlike the previous two methods, it does not directly generate vectors for each node, but introduces a learnable scalar bias in the message weight calculation process. Compared with the learnable position encoding, it has fewer parameters and is easier to train. It replaces the b term in formula (1) with a learnable bias term β, the specific formula is as follows:

[0055]

[0056] S22: neighbor message fusion based on relative position encoding; as Figure 3As shown, when using the graph attention mechanism to fuse the neighbor message, the relative position code obtained by S21 is embedded into the local observation feature of the neighbor robot, the attention score is calculated by formula (1), and then weighted summation is generated to generate an anonymous perception enhanced feature x containing the local observation feature of itself, the local observation feature of the neighbor and the relative position information thereof i ′;

[0057]

[0058] Wherein: sigma is an activation function; is the neighbor (including the robot i itself) in the communication range of the robot i; W (l) is the parameter of the lth layer graph attention convolution; alpha i,j is the normalized attention score; |is the number of communicable neighbors of the robot i; e (i,j) derived from formula (1).

[0059] S3: decision action probability distribution prediction and priority prediction; each robot learns a high-quality action strategy and implicit scheduling priority from the expert demonstration data provided by the traditional centralized method with global vision through imitation learning, specifically, each robot uses two multilayer perceptrons respectively to predict the action probability distribution that may be taken by itself next step based on the anonymous perception enhanced feature obtained by step S2, and the priority of the action execution of each robot in the current local communication range;

[0060] The decision action probability distribution prediction and priority prediction proposed in step S3 of the application specifically includes the following two sub-steps:

[0061] S31: robot action probability distribution learning based on latent space aggregation feature; the action selected by the expert under a given observation is regarded as a "basic fact", and the expert data D={O t ,a *} t＝1,2，…，T generated by ECBS, wherein O t is the set of local observation values of all robots at time t, a * is the set of actions of all robots planned by ECBS according to the global observation value, and T is the total number of steps. It learns a strategy pi θ that maps the aggregated local observation value m at the current time step to the expert operation. The strategy parameter theta is optimized as follows:

[0062]

[0063] Wherein: a * is the set of actions of all robots planned by ECBS according to the global observation value, and m iis the feature of robot i after aggregating neighbor messages, a i ~ π θ (m i ) is π θ based on m i learned action probability distribution.

[0064] S32: Robot priority learning based on latent space aggregated features; priority learning aims to capture the priority strategy implied in classical centralized planners, i.e., according to the local observation of the robot, learn whether it should approach the target or avoid conflicts. In this invention, enhanced conflict-based search (ECBS) is adopted as a classical centralized planner, which is considered to be an efficient and bounded suboptimal multi-robot pathfinding algorithm. The priority learning model in this invention shares the same underlying architecture with the action decision model, only differing in the parameters of the final fully connected layer. This architecture coupling promotes stronger synergy between priority learning and imitation learning, thereby improving the overall effectiveness of the model.

[0065] The construction of the priority dataset is as follows: based on the expert data provided by ECBS, if the operation of the robot makes it closer to the target or keeps it at the target position, it is assigned a priority value of 1. If the operation does not cause the distance to the target to change, the priority is set to 0.5. Otherwise, if the operation increases the distance to the target, the priority is set to 0. The mathematical formula is as follows:

[0066]

[0067] where: s i represents the current position of robot i, s i ' is the next position of robot i, g i is its target position or its projection on the boundary of the local view, d(x,y) represents the Manhattan distance from x to y, p i represents the priority of robot i at the current time.

[0068] S4: PIBT adaptive coordination decision action based on priority learnability; this strategy effectively enhances the rationality of the system in assigning priorities by utilizing the dynamic priorities learned based on local observations in S32. When the system detects a priority inversion situation that may trigger a traffic jam, it will automatically trigger the PIBT mechanism, causing the current robot to temporarily inherit the authority of the high-priority robot. At the same time, combined with the predicted action probability distribution in S31, the passing order and execution action between robots are dynamically adjusted, thereby realizing adaptive decision-making between robots.

[0069] The PIBT adaptive coordination decision action based on priority learnability proposed in step S4 of the invention specifically includes the following three sub-steps:

[0070] S41: Action sampling and preliminary conflict detection based on action probability distribution; robots sample actions according to the predicted action probability distribution in S31, and detect conflicts of the sampled actions within the local communication range. If no conflict is detected, each robot can directly execute the sampled action without further coordination.

[0071] S42: Conflict resolution based on the order execution mechanism of learnable priority; if a conflict is detected, the system will refer to the robot priority score learned in S32, and execute each candidate action in turn according to the priority, thereby achieving effective conflict avoidance. This helps to reduce the occurrence of priority inversion and reduce the possibility of conflict.

[0072] S43: Backtracking coordination strategy based on priority inheritance to deal with priority inversion problem; in the process of sequential execution, if priority inversion occurs, i.e. a low-priority robot blocks a high-priority robot, the priority inheritance mechanism in PLPIBT is triggered to temporarily improve the priority of key robots in the blocked path, thereby achieving local conflict resolution and path unblocking, and preventing the system from entering a deadlock state.

[0073] Based on the proposed message fusion mechanism with PE and conflict shielding strategy based on PLPIBT in the framework of imitation learning, a new distributed multi-robot navigation method with communication ability, PE-PLPIBT, is constructed. Since the proposed PE includes three variants: fixed, learnable and learnable bias, the algorithms using these three different PE methods are named FPE-PLPIBT, LPE-PLPIPT and LBPE-PLPIBT, respectively.

[0074] Embodiments

[0075] Take multi-robot path planning as an example: dataset setting: 10 robots are randomly distributed in a 20x20 grid environment, and 30000 cases are collected. In each case, the environment layout and the starting and target positions of the robots are randomly generated. Among these cases, 70% are used as the training set, 15% are used as the validation set, and the remaining 15% are used as the test set. In addition, we conduct generalization tests under two different settings: (1) change the map size while keeping the robot density constant (same density dataset); (2) increase the robot density while keeping the map size fixed (density increase dataset), and the obstacle density of all maps is 10%. Comparative algorithm: In order to evaluate the effectiveness of the three algorithms proposed, three distributed methods based on imitation learning are introduced as baselines. The distributed methods are MAGAT, GAT and GL. MAGAT and GAT are classic distributed communicable multi-robot navigation methods, and GL extracts global position information through an MLP layer and combines it with local observation features for message fusion based on MAGAT.

[0076] Figure 5 The success rate (SR), Makespan increase rate (MIR), Flowtime increase rate (FIR) and predicted conflict rate (PCR) of the three algorithms proposed by the application and the comparative algorithms on different datasets are shown, where the smaller the values of MIR, FIR and PCR, the better the performance of the algorithm, and SR is the opposite. In Figure 5 a-d and e-h in the table show the results of all algorithms on the same density dataset and the density increase dataset, respectively, from Figure 5 It can be seen from the table that the methods of the application have good search ability and strong generalization ability. Among them, FPE-PIBT obtains the best overall ranking, and the application also shows that simply connecting the features extracted from the global coordinates does not produce significant performance improvement. Figure 6 The convergence curves of all algorithms on the four indicators during training are plotted, which intuitively illustrates that the methods of the application have good search efficiency and convergence ability.

[0077] By introducing relative position encoding, the application explicitly captures the distance and direction information between robots and integrates it into the graph attention mechanism, greatly enhancing the representation ability of the spatial structure, enabling the model to seamlessly migrate from small-scale training to large-scale deployment; at the same time, by using a learnable priority strategy, combining action probability distribution with PIBT mechanism, dynamically adjusting the priority of robots, effectively alleviating congestion and priority inversion, quickly generating conflict-free decisions and preventing deadlocks; more importantly, under the fully distributed architecture, robots only need to communicate with their neighbors, which not only significantly reduces the communication bandwidth and delay, but also maintains high success rate and scheduling efficiency in large-scale scenarios with dozens to hundreds of robots.

[0078] The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and practice described. Accordingly, all such variations and modifications are intended to be included within the scope of the application as defined in the following claims.

Claims

1. A multi-robot cooperative navigation method based on anonymity-enhanced perception and adaptive decision-making, characterized in that, Includes the following steps: S1: Local environment feature extraction based on residual network; Each robot uses a ResNet-based perception network to encode its local observation environment and extract local observation features with spatial representation capabilities, providing a unified latent space basis for subsequent message fusion and action decision-making. S2: Anonymous perception enhancement mechanism that integrates relative position information; During message fusion, a relative position encoding mechanism is introduced to the robot within the communication range. Specifically, the spatial position information of the neighboring robot relative to itself is encoded into a structured vector and combined with the local observation features transmitted by the neighboring robot to generate "anonymous perception enhancement features". This fusion is carried out simultaneously during message propagation and reception, thereby enhancing the perception of the surrounding environment while maintaining the consistency of the latent space. S3: Prediction of decision action probability distribution and priority prediction; Each robot learns high-quality action strategies and implicit scheduling priorities from expert demonstration data provided by traditional centralized methods with a global perspective through imitation learning. Specifically, each robot uses two multilayer perceptrons to predict the probability distribution of its next possible action based on the anonymized perception enhancement features obtained in step S2, as well as the priority of each robot's action execution within the current local communication range. S4: Priority-based learnable PIBT adaptive coordination decision-making action; An improved priority inheritance and backtracking strategy PLPIBT, which integrates learnable priorities, is designed. This strategy utilizes the priorities learned in step S3 to effectively enhance the rationality of priority allocation in the system. When the system detects a priority reversal situation that may cause passage blockage, it will automatically trigger the PIBT mechanism, allowing the current robot to temporarily inherit the permissions of the higher-priority robot. At the same time, combined with the action probability distribution predicted in step S3, the passage order and execution actions between robots are dynamically adjusted, thereby achieving adaptive decision-making between robots.

2. The multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making according to claim 1, characterized in that, The anonymized perception enhancement mechanism that integrates relative position information proposed in step S2 specifically includes the following two sub-steps: S21: Construct position codes based on discretized positions within the communication range; Discretize and number the positions of neighboring robots in the local grid space, and construct relative position codes with different spatial perception capabilities to provide structural priors for subsequent message fusion; S22: Neighbor message fusion based on relative position encoding; when fusing neighbor messages using the graph attention mechanism, the relative position encoding obtained in step S21 is embedded into the local observation features of the neighboring robot, and finally an anonymous perception enhancement feature is generated that simultaneously contains its own local observation features, the local observation features of the neighbor and their relative position information.

3. The multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making according to claim 2, characterized in that, Step S3, which proposes prediction of the probability distribution of action decisions and priority prediction, specifically includes the following two sub-steps: S31: Learning the probability distribution of robot actions based on anonymized perception-enhanced features; The system learns and fits the action decision-making strategies in expert demonstration data through a neural network. The expert strategies used are derived from a centralized planner with a global perspective and strong decision-making capabilities. During the training phase, the system uses the anonymized perception-enhanced features obtained in step S2 as input to predict the probability distribution of the robot's next possible action. The policy network is aligned with the expert demonstration through a cross-entropy loss function, enabling the robot to approximately reproduce the high-quality decisions of centralized planning in a distributed environment. S32: Robot priority learning based on anonymized perception enhancement features; This step also uses a neural network to capture the scheduling logic implicit in the expert centralized planner and realize dynamic learning and prediction of priorities; The network input is the anonymized perception enhancement features generated in step S2, and the output is the robot's action execution priority score within the current local communication range. The learned priority reflects both the urgency of the task and incorporates the potential conflict risk in the neighborhood, so as to guide the robot to make autonomous trade-offs between moving forward and yielding.

4. The multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making according to claim 1, characterized in that, The priority-learnable PIBT adaptive coordination decision-making action proposed in step S4 specifically includes the following three sub-steps: S41: Action sampling and preliminary conflict detection based on action probability distribution; The robot samples actions according to the action probability distribution predicted in S31, and performs conflict detection on the sampled actions within the local communication range. If no conflict is detected, each robot can directly execute the sampled actions without further coordination. S42: Conflict resolution is achieved through a sequential execution mechanism based on learnable priorities. If an action conflict is detected, the system will refer to the robot priority score learned in S32 and execute the candidate actions in order of priority, thereby achieving effective conflict avoidance. S43: A backtracking coordination strategy based on priority inheritance is used to address the priority inversion problem; If a priority inversion occurs during sequential execution, i.e. a low-priority robot obstructs the passage of a high-priority robot, the priority inheritance mechanism in PLPIBT is triggered. By temporarily increasing the priority of the key robot in the blocked path, the local conflict is resolved and the path is cleared in coordination, preventing the system from entering a deadlock state.

Images