Multi-robot cooperative navigation method based on anonymity enhancement and adaptive decision-making

Abstract

A multi-robot cooperative navigation method based on anonymity-enhanced perception and adaptive decision-making includes the following steps: S1: Local environmental feature extraction based on residual networks; S2: Anonymity-enhanced perception mechanism fusing relative position information; S3: Probability distribution prediction and priority prediction of decision actions; S4: Adaptive coordinated decision-making based on priority-learnable PIBT. This invention designs a message fusion mechanism based on position encoding in the feature latent space to solve the "message anonymity" problem, and introduces a priority-learnable PIBT strategy at the decision layer to solve the problems of unreasonable priority allocation and priority inversion, thereby improving the generalization and flexibility of multi-robot cooperative navigation and enabling efficient cooperative navigation in a distributed, communicable environment.

Description

Technical Field

[0001] This invention relates to a robot navigation method, specifically a multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making, belonging to the field of cooperative navigation technology in multi-robot systems. Background Technology

[0002] With the widespread application of multi-robot systems in warehousing and logistics, intelligent manufacturing, and urban delivery, efficient and conflict-free cooperative navigation has become crucial for ensuring the safety and operational efficiency of multi-robot collaborative operation. Multi-robot cooperative navigation refers to the coordinated operation of multiple robots through perception, communication, and decision-making mechanisms. Without centralized control, each robot designs an independent and continuous motion trajectory from its starting point to its destination based on environmental dynamics and task objectives. Through strict spatiotemporal constraint management, collisions are effectively avoided, achieving optimal scheduling and collaborative operation of global resources.

[0003] Based on the way global information is used, multi-robot cooperative navigation methods can be divided into two main categories: centralized and distributed. Centralized methods, through centralized management, plan all robots uniformly and can obtain globally optimal or near-optimal solutions. However, due to the high concentration of computing resources and decision-making power, they typically incur huge communication overhead and lack scalability. If the central processing unit fails, the entire system may be paralyzed. Distributed methods, on the other hand, rely on each robot's local view to independently generate decisions and exchange information with neighboring robots through a distributed coordination mechanism, thereby achieving globally conflict-free cooperative navigation. This design significantly improves the system's robustness and reliability. However, due to the lack of a global perspective, this method is prone to conflicts, which may lead to conflicts or deadlocks in severe cases. To address this issue, current research introduces communication mechanisms, mainly focusing on two key collaborations: 1) Feature-level collaboration: aggregating the observation features of neighbors to enhance the robot's understanding of the dynamic environment; 2) Decision-level collaboration: employing a conflict shielding mechanism to resolve action conflicts and ensure path safety.

[0004] In distributed approaches, graph neural networks (GNNs) have emerged as a cutting-edge tool for handling local communication and collaboration among robots. GNNs capture the relational structure of a system by modeling each robot as a node in a graph and communication links as edges, fusing neighbor information through graph convolution or message passing. This approach is naturally suited to the dynamic communication graphs of multi-robot systems, effectively learning information exchange and coordination strategies among robots. However, most existing methods implicitly extract positional information from local grid observations, such as the layout of surrounding obstacles, the positions of other robots, and the projection of the target, without using explicit node identity cues. This anonymity of message sources undermines the model's ability to perceive relative relationships between robots in dynamic environments, increasing the risk of conflict. Position encoding is a technique widely studied in deep learning in recent years, especially relative position encoding, which has been used in sequence models such as Transformer and BERT to enhance the model's perception of spatial relationships between elements. Recent work has used the relative positions of robots as edge features in the graph to encode distance and orientation information between robots. For example, in graph-based multi-robot autonomous collaboration, researchers have set the features of each edge as a relative vector of the robot's position and encoded it through linear transformation before passing it to a graph neural network. This approach allows the model to utilize spatial topological information more directly, improving its robustness to changes in environmental geometry. However, research combining relative position encoding and graph neural networks is still limited, failing to fully leverage the advantages of relative position encoding in message fusion.

[0005] Furthermore, conflict avoidance mechanisms at the decision-making level are crucial for achieving conflict-free planning. Since each robot makes independent decisions based on its own local strategy, spatial conflicts between neighboring robots often arise. Existing conflict avoidance strategies mainly fall into two categories: one is to "freeze" robots that may cause conflict, i.e., make them wait; the other is to assign priorities to each robot and plan them sequentially according to priority. While the former can quickly eliminate conflicts, it can lead to deadlock and cannot fully utilize the complete probability distribution of actions already learned by the robots; the latter, while possessing a certain degree of scheduling flexibility, often causes priority inversion problems due to improper priority allocation, leading to a decline in overall system performance.

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

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

[0008] Improper priority allocation can lead to priority reversal in decision-making: While current conflict avoidance algorithms for robots that experience conflicts due to freezing can ensure that no conflict occurs, they are also prone to blocking and deadlock. Other priority-based methods can alleviate the above problems to some extent, but they are quite sensitive to priority allocation. If the priority allocation is improper, it can cause priority reversal, leading to a decrease in the overall system performance.

[0009] Insufficient collaborative capabilities lead to poor system scalability: Many learning-based cooperative navigation methods are trained on small-scale scenarios, and their performance often drops significantly when the number of robots or the scale of the scenario increases substantially. Existing research shows that large-scale multi-robot cooperative navigation remains an open challenge. Even with the use of techniques such as graph neural networks, current methods still suffer from insufficient success rate and completeness in larger scenarios or environments with high-density obstacles. Summary of the Invention

[0010] The purpose of this invention is to provide a multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making. This method designs a message fusion mechanism based on position encoding in the feature latent space to solve the "message anonymity" problem, and introduces a priority-learnable PIBT strategy at the decision layer to solve the problems of unreasonable priority allocation and priority inversion, thereby improving the generalization and flexibility of multi-robot cooperative navigation and enabling efficient cooperative navigation in a distributed and communicable environment.

[0011] To achieve the above objectives, this invention provides a multi-robot cooperative navigation method based on anonymity perception enhancement and adaptive decision-making, comprising the following steps:

[0012] 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.

[0013] S2: Anonymous perception enhancement mechanism that integrates relative position information; During message fusion, a relative position encoding mechanism is introduced to robots within the communication range to solve the anonymity problem. Specifically, the spatial position information of neighboring robots relative to themselves is encoded into a structured vector and combined with the local observation features transmitted by neighboring robots 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.

[0014] 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 (expert data is generated using the ECBS algorithm in this patent). 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.

[0015] 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.

[0016] The anonymous perception enhancement mechanism that integrates relative position information proposed in step S2 of this invention specifically includes the following two sub-steps:

[0017] 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;

[0018] S22: Neighbor message fusion based on relative position encoding; when fusing neighbor messages using 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.

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

[0020] 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.

[0021] 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.

[0022] The priority-learnable PIBT adaptive coordination decision-making action proposed in step S4 of this invention 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 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.

[0024] 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.

[0025] S43: The backtracking coordination strategy based on priority inheritance addresses the priority inversion problem. During sequential execution, if a priority inversion occurs, 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.

[0026] Compared with existing technologies, this invention introduces a relative position encoding mechanism, enabling the model to solve the message anonymity problem and thus significantly enhance its ability to perceive environmental topology. This encoding method embeds the geometric relationships between nodes in the graph into the attention calculation process, which not only improves the accuracy of feature fusion but also enhances the model's transferability and robustness when faced with unseen topology structures.

[0027] This invention introduces a priority-learnable mechanism, enabling robots to dynamically adjust their own priorities based on local observations. This overcomes the congestion and priority reversal 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 conflict-free action paths, effectively reducing the conflict rate between robots, avoiding robot deadlock, and improving overall navigation efficiency.

[0028] This invention is entirely based on a distributed architecture design, where robots perform local perception and decision-making coordination only within their communication radius, without the need for a central scheduler. While maintaining a high success rate and high coordination efficiency, it significantly reduces reliance on communication bandwidth and real-time performance. Even when the number of robots expands to tens or hundreds, the system still maintains good scalability and task completion efficiency. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of multi-robot path planning in a distributed, communicable environment according to the present invention;

[0030] Figure 2 This refers to the process of extracting the corresponding relative position code from the designed position coding matrix in this invention;

[0031] Figure 3 This is the message fusion calculation process with location encoding in this invention;

[0032] Figure 4 This is a flowchart illustrating the overall framework and workflow of multi-robot cooperative navigation planning in this invention.

[0033] Figure 5 The figures show the comparison between the algorithm proposed in this invention and three distributed algorithms on four evaluation metrics. Figures (a) to (d) show the comparison results of the success rate (SR), Makespan growth rate (MIR), Flowtime growth rate (FIR), and prediction conflict rate (PCR) on datasets with the same density, respectively. Figures (e) to (h) show the changing trends and performance of these four metrics for each algorithm on datasets 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 achieve cooperative navigation of multiple robots in a distributed, communicable environment, such as Figure 2 and Figure 3 As shown, this invention proposes a multi-robot cooperative navigation method based on anonymity-enhanced perception and adaptive decision-making, integrating latent space (feature level) and decision-level coordination. The specific implementation steps are as follows:

[0040] S1: Local Environment Feature Extraction Based on Residual Networks; First, the observation information collected by the robot is input into a convolutional neural network (CNN) to increase the channel dimension and enhance the expressive power of the low-level features. Then, the features are fed into a residual network (ResNet) composed of multiple residual modules to extract local observation features with a certain spatial structure perception capability. The extracted features are flattened and mapped to a unified latent space representation through a multilayer perceptron (MLP), providing a structurally consistent feature foundation for subsequent neighbor message fusion and adaptive decision-making.

[0041] like Figure 4 The observation information collected by the robot shown It is a three-channel input matrix of shape (2l+1)×(2l+1), where l is the observation radius and the first channel is the first channel. Encode a portion of the observation environment of the i-th robot, where 1 represents an obstacle and 0 represents a passable area; the second passage The third channel represents the target position of the i-th robot or its projection onto the boundary of the local view; The self-robot i located at the center and the other robots within its field of vision are encoded.

[0042] S2: Anonymous perception enhancement mechanism integrating relative position information; message anonymity can be addressed by assigning a unique ID to each node. However, in distributed robot navigation scenarios, robots constantly enter and leave limited communication ranges, making it impractical to assign a globally unique ID to each robot. Therefore, by assigning unique IDs to grid cells within the communication range, any robot occupying that cell can use its corresponding grid ID as its own ID. The position encoding of this invention is a mature and effective method for generating such IDs.

[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 2This demonstrates the process of extracting the corresponding relative position code from the designed position coding matrix P based on the discretized number of the neighboring robots within a local grid space.

[0049] Inspired by the position encoding strategy in the Transformer model, this invention introduces a fixed, non-learnable sine-cosine position encoding method. This method generates a unique vector representation of position using periodic functions of different frequencies, requiring no training, resulting in high computational efficiency. It is suitable for multi-robot systems with relatively limited resources or requiring rapid deployment. The calculation formula for the fixed position encoding is as follows:

[0050]

[0051] Where: pos ij Let d represent the position of robot j relative to robot i, and d be the dimension of the position vector.

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

[0053] Learnable Relative Position Encoding: Inspired by the learnable position encoding mechanism in the BERT model, this invention introduces a learnable relative position encoding method, enabling the model to automatically capture more complex and flexible spatial relationships. This invention pre-sets a trainable embedding matrix for different relative positions. Each relative position corresponds to an embedding vector, which is continuously optimized during training through gradient backpropagation. Compared with fixed sine and cosine coding, this method has the following advantages: it can automatically learn task-related positional features without requiring manual coding rules; it has stronger expressive power and can adapt to more complex environments.

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

[0055]

[0056] S22: Neighbor message fusion based on relative position encoding; such as Figure 3As shown, when using the graph attention mechanism to fuse neighbor messages, the relative position encoding obtained in S21 is embedded into the local observation features of the neighboring robots. The attention score is calculated using formula (1), and then a weighted sum is used to generate an anonymous perception enhancement feature x that simultaneously contains its own local observation features, the local observation features of the neighbors, and their relative position information. i ′;

[0057]

[0058] Where: σ is the activation function; It is the neighbor within robot i's communication range (including robot i itself); W (l) These are the parameters of the graph attention convolution in the l-th layer; α i,j It is a normalized attention score; | is the number of communicable neighbors of robot i; e (i,j) It is derived from formula (1).

[0059] 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.

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

[0061] S31: Learning the probability distribution of robot actions based on latent space aggregation features; treating the actions chosen by experts under given observations as "basic facts", expert data D={O t ,a *} t＝1,2，…，T Generated by ECBS, where O t It is the set of local observations of all robots at time t, a * It is the set of all robot actions planned by ECBS based on global observations, where T is the total number of steps. It has learned a policy π. θ This strategy maps the aggregated local observations *m* at the current time step to expert operations. The strategy parameter θ is optimized as follows:

[0062]

[0063] Where: a * It is the set of actions for all robots planned by ECBS based on global observations, m iIt is the feature of robot i after aggregating neighbor messages, a i ~π θ (m i ) is π θ Based on m i The probability distribution of actions learned.

[0064] S32: Robot Priority Learning Based on Latent Space Aggregation Features; Priority learning aims to capture the implicit priority strategy in classic centralized planners, i.e., learning whether the robot should approach the target or avoid conflict based on its local observations. This invention employs Enhanced Conflict-Based Search (ECBS) as the classic centralized planner, considered an efficient and bounded suboptimal multi-robot pathfinding algorithm. The priority learning model and action decision model share the same underlying architecture, differing only in the parameters of the final fully connected layer. This architectural coupling promotes stronger synergy between priority learning and imitation learning, thereby improving the overall effectiveness of the model.

[0065] The priority dataset is constructed as follows: Based on expert data provided by ECBS, if a robot's operation brings it closer to or keeps it at the target position, a priority value of 1 is assigned. If the operation does not cause a change in the distance to the target, 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 s represents the current position of robot i. i ′ is the next position of robot i, g i It is its target location or its projection on the boundary of the local view, d(x,y) represents the Manhattan distance from x to y, p i This indicates the priority of robot i at the current moment.

[0068] S4: Priority-based learnable PIBT adaptive coordination decision-making action; this strategy utilizes the dynamic priority learned from local observations in S32 to effectively enhance the rationality of priority allocation. When the system detects a priority reversal situation that may cause passage obstruction, it will automatically trigger the PIBT mechanism, allowing the current robot to temporarily inherit the permissions of the higher-priority robot. Simultaneously, combined with the predicted action probability distribution in S31, the passage order and actions among the robots are dynamically adjusted, thereby achieving adaptive decision-making among the robots.

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

[0070] S41: Action sampling and preliminary conflict detection based on action probability distribution; the robot samples actions according to the predicted action probability distribution 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.

[0071] 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. This helps reduce the occurrence of priority inversion and lowers the probability of conflict.

[0072] S43: The backtracking coordination strategy based on priority inheritance addresses the priority inversion problem. During sequential execution, if a priority inversion occurs, 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.

[0073] This invention constructs a novel distributed multi-robot navigation method with communication capabilities, namely PE-PLPIBT, based on an imitation learning framework combined with a proposed message fusion mechanism with PE and a conflict shielding strategy based on PLPIBT. 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-PLPIBT, and LBPE-PLPIBT, respectively.

[0074] Example

[0075] Taking multi-robot path planning as an example: Dataset setup: 10 robots were randomly distributed in a 20×20 grid scene, and 30,000 cases were collected. In each case, the environment layout and the robot's start and target positions were randomly generated. In these cases, 70% were used as the training set, 15% as the validation set, and the remaining 15% as the test set. In addition, we conducted generalization tests in two different settings: (1) changing the map size while keeping the robot density constant (same density dataset); (2) increasing the robot density while keeping the map size fixed (density increase dataset), with an obstacle density of 10% in all maps. Algorithm comparison: In order to evaluate the effectiveness of the three proposed algorithms, three distributed methods based on imitation learning were introduced as baselines. The distributed methods are MAGAT, GAT, and GL. MAGAT and GAT are classic distributed communicable multi-robot navigation methods. GL extracts global position information through an MLP layer on the basis of MAGAT and connects it with local observation features for message fusion.

[0076] Figure 5 The results of the three algorithms proposed in this invention and the comparison algorithm on different datasets are shown, along with four metrics: success rate (SR), Makespan growth rate (MIR), Flowtime growth rate (FIR), and prediction conflict rate (PCR). The smaller the values ​​of MIR, FIR, and PCR, the better the algorithm's performance, while SR shows the opposite. Figure 5 The `ad` and `eh` tables show the results of all algorithms on datasets with the same density and datasets with increasing density, respectively. Figure 5 As can be seen, the method of this invention has excellent search capabilities and strong generalization ability. Specifically, FPE-PIBT achieved the best overall ranking. This invention also demonstrates that simply concatenating features extracted from global coordinates does not produce a significant performance improvement. Figure 6 Convergence curves for all algorithms on four metrics during training were plotted, visually illustrating that the method of this invention has excellent search efficiency and convergence capability.

[0077] This invention introduces relative position encoding to explicitly capture distance and orientation information between robots and integrates it into a graph attention mechanism, significantly enhancing the representation capability of spatial structures and enabling the model to seamlessly migrate from small-scale training to large-scale deployment. Simultaneously, it employs a learnable priority strategy, combining action probability distribution with the PIBT mechanism, to dynamically adjust robot priorities, effectively alleviating congestion and priority inversion, rapidly generating conflict-free decisions, and preventing deadlocks. More importantly, in a fully distributed architecture, robots only need to communicate with their neighbors to achieve efficient collaboration, significantly reducing communication bandwidth and latency, and maintaining high success rates and scheduling efficiency in large-scale scenarios with dozens to hundreds of robots.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for multi-robot cooperative navigation based on anonymous perception enhancement and adaptive decision making, characterized in that, Comprise the following steps: S1: local environment feature extraction based on residual network; each robot encodes its local environment using a perception network based on ResNet structure, from which local observation features with spatial representation ability are extracted, providing a unified hidden space basis for subsequent message fusion and decision action; S2: anonymous perception enhancement mechanism of message fusion; in the message fusion process, a relative position coding mechanism is introduced for robots within the communication range. Specifically, the spatial position information of neighbor robots relative to themselves is encoded as a structured vector, and combined with the local observation features transmitted by neighbor robots to generate anonymous perception enhancement features; Message fusion is carried out synchronously in the process of message transmission and reception, thereby enhancing the perception of the local environment while maintaining the consistency of the hidden space; S3: action probability distribution prediction and priority prediction of decision action; each robot learns high-quality action strategy and implicit scheduling priority from expert demonstration data provided by traditional centralized methods with global vision through imitation learning. Specifically, each robot uses two multilayer perceptrons to predict the action probability distribution that may be taken by itself based on the anonymous perception enhancement features obtained in step S2, as well as the priority of action execution by each robot within the current communication range; 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, making the current robot temporarily inherit the high priority of the robot, while combining the action probability distribution predicted in step S3 to dynamically adjust the traffic order and execution action between robots, thereby realizing adaptive decision between robots; The anonymous perception enhancement mechanism of message fusion proposed in step S2 specifically includes the following two sub-steps: S21: position coding based on discrete positions within the communication range; the positions of neighbor robots are discretely numbered in the local grid space, and relative position coding with different spatial perception capabilities is constructed to provide structural prior for subsequent message fusion; S22: neighbor message fusion based on relative position coding; when using graph attention mechanism to fuse neighbor messages, the relative position coding obtained in step S21 is embedded into the local observation features of neighbor robots, and finally anonymous perception enhancement features containing local observation features of themselves, neighbor robots and their relative position information are generated.

2. The method of claim 1, wherein, The action probability distribution prediction and priority prediction proposed in step S3 specifically include the following two sub-steps: S31: Robot action probability distribution learning based on anonymous perception enhanced features; learn and fit the decision-making action strategy in the expert demonstration data through a neural network. The expert strategy used comes from a centralized planner with a global view and 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. The neural network aligns with the expert demonstration through a cross-entropy loss function, enabling the robot to approximately reproduce high-quality decisions of centralized planning in a distributed environment. S32: Priority learning of robots based on anonymous perception enhanced features; through a neural network, capture the scheduling logic implied in the centralized planner to achieve dynamic learning and prediction of priorities; The neural network input is the anonymous perception enhanced features generated in step S2, and the output is the action execution priority score of the robot within the current communication range. The learned priority reflects both task urgency and potential conflict risk in the neighborhood to guide the robot to make autonomous trade-offs between advancing and yielding.

3. The method of claim 2, wherein, The PIBT adaptive coordination decision action based on priority learning 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 based on the action probability distribution predicted in S31, and detects conflicts for the sampled actions within the communication range. If no conflict is detected, each robot can directly execute the sampled action without further coordination; S42: Conflict resolution based on the order execution mechanism of learnable priority; if a conflict is detected, the system will refer to the priority score of the robot learned in S32, and execute the candidate actions in order of priority to achieve effective conflict avoidance; S43: Backtracking coordination strategy based on priority inheritance to deal with the problem of priority reversal; In the sequential execution process, if priority reversal occurs, i.e., a low-priority robot blocks a high-priority robot, the priority inheritance mechanism in PIBT is triggered to temporarily raise the priority of the key robot in the blocked path to collaboratively resolve local conflicts and clear the path, preventing the system from entering a deadlock state.

Images