An agent collaborative navigation method based on communication opinions

Abstract

The application relates to artificial intelligence and discloses an intelligent agent cooperative navigation method based on communication opinions, which comprises the following steps: each intelligent agent acquires a current local observation and receives a communication input at a previous time t-1; a communication graph of an inter-agent communication connection state is generated based on a neighborhood relationship at a previous time, the current local observation and an opinion set; a joint observation feature is generated by adding movement; the joint observation feature is input into a strategy network to output a current time action and an action opinion; the intelligent agent executes the action to promote system state evolution, receives an environment feedback reward to acquire a new local observation and a communication input, and enters a next decision cycle. By introducing the communication opinions and the action opinions, the application realizes adaptive selection of an inter-agent communication relationship and dynamic cooperation of behavior decision-making, so that the information transmission efficiency is ensured, and the overall coordination and safety of the group are improved.

Description

Technical Field

[0001] This application pertains to artificial intelligence and relates to a collaborative navigation method for intelligent agents based on communication opinions. Background Technology

[0002] With the rapid development of autonomous driving, drone swarming, and intelligent robot systems, multi-agent cooperative decision-making technology has become a core means of achieving swarm intelligence. Typical multi-agent tasks include area coverage, path planning, dynamic formation, cooperative obstacle avoidance, and resource scheduling. In these tasks, the perception range of a single agent is limited, and communication mechanisms are essential for sharing local observation information to achieve global cooperation. However, in real-world environments, communication channels are often limited, and information transmission is subject to latency, packet loss, or interference. This makes the communication mechanism a key factor restricting the performance of multi-agent systems.

[0003] Existing research has mainly followed two paths. One type of method employs a centralized architecture, where a central controller receives state information from all agents and performs global planning. This type of method can obtain relatively good global solutions, but as the number of agents increases, the computational complexity rises sharply, and the central node easily becomes a system bottleneck. The other type of method employs a distributed decision-making architecture, where each agent makes independent decisions based on local information and shares information through communication. Representative algorithms include CommNet, DIAL, and TarMAC, which encode and aggregate communication messages using learnable neural networks. However, these methods generally suffer from high communication overhead, severe information redundancy, and a lack of semantic interpretation of messages. Especially in some observable scenarios, agents cannot accurately determine which information is most valuable for the current decision, leading to wasted communication bandwidth and decision confusion.

[0004] To reduce communication load, some studies have introduced sparse communication mechanisms, such as automatically filtering important neighbor information through attention networks. However, these methods still separate communication from action policy optimization, failing to achieve true joint decision-making. Furthermore, existing algorithms often neglect system security issues and lack constraints such as speed, distance, or energy, leading to collisions or unstable behavior in practical implementation. In summary, while the closest current technical solutions have made some progress in communication selection, they still have significant shortcomings in communication efficiency, policy coupling, security constraints, and system stability. Summary of the Invention

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0006] To address the problems existing in related technologies, this disclosure provides an intelligent agent cooperative navigation method based on communication opinions, which solves the problem that the communication structure in the multi-agent communication and decision-making of the prior art lacks adaptability and cannot be dynamically adjusted in different task stages or different spatial regions, resulting in a large amount of redundant communication and bandwidth waste.

[0007] In some embodiments, an agent cooperative navigation method based on communication opinions is provided, the method comprising:

[0008] S10, at time Each intelligent agent Obtain current local observations and receive the previous moment -1 communication input The A set of opinions conveyed to other intelligent agents;

[0009] S20, based on the neighborhood relationship of the previous time step Current local observations and opinion collection Generate a communication graph showing the communication connection states between intelligent agents. ;

[0010] S30, based on communication diagram Current local observations And the additional terms of motion z i t Generate joint observation features i t ;

[0011] S40 will combine observation features Input Policy Network Output the action 'a' at the current moment. i t and opinions on actions The action opinions are semantic descriptions of the agent's future behavioral trends;

[0012] S50, the agent executes action a i t Drive system state evolution and receive rewards from environmental feedback. Acquire new local observations at time t+1 and communication input Re-execute S10.

[0013] Preferably, in step S20, based on the neighborhood relationship of the previous time step... Current local observations and opinion collection Generate a communication graph showing the communication connection states between intelligent agents. ,include:

[0014] Based on the neighborhood relationship at the previous time step Construct a communication input graph structure, and [o i t , As input features, the communication selection probability p is generated through the communication policy network. ij t The sparse communication graph C is obtained through the sparse constraint sampling algorithm. t The C t c represents the communication connection state between intelligent agents. ij t constitute.

[0015] Preferably, S30, based on the communication diagram Current local observations And the additional terms of motion z i t Generate joint observation features i t ,include:

[0016] Based on communication graph C t Through message aggregation functions By aggregating the features of neighboring agents and their current states, a message summary item M is obtained. i t , will o i t M i t Additional sports items z i t splicing to generate joint observation features i t .

[0017] Preferably, a centralized training and distributed execution CTDE mechanism is adopted; during the centralized training phase, the states, actions, communication graphs, and reward information of all agents are centrally input into the global optimizer to calculate a multi-objective joint loss function, wherein the multi-objective joint optimization function takes the form of:

[0018] ,

[0019] To reinforce learning objectives, For communication constraints, For consistency constraints, For safety constraints, α, β, and γ are weighting coefficients.

[0020] Preferably, the safety constraint item Used to punish collisions between intelligent agents, boundary violations, and excessive acceleration operations.

[0021] Preferably, the communication constraint item Used to limit communication redundancy and encourage the construction of sparse and efficient communication topologies.

[0022] Preferably, the consistency constraint term This is used to ensure the consistency of behavior between the proposed action and the actual action performed.

[0023] Preferably, the specific form of the opinion includes desired speed, heading direction, or strategy embedding features.

[0024] Preferably, the message aggregation function A graph neural network aggregation mechanism is adopted to fuse the local observation features and action opinion features of neighboring agents.

[0025] Preferably, the sparse constraint sampling algorithm is a top-k selection algorithm or a threshold filtering algorithm.

[0026] Compared with the prior art, the intelligent agent cooperative navigation method based on communication opinions provided in this disclosure can achieve the following technical effects:

[0027] This disclosure, through the introduction of communication opinions and action opinions, enables adaptive selection of communication relationships and dynamic coordination of behavioral decisions among intelligent agents, thereby improving the overall coordination and security of the group while ensuring information transmission efficiency.

[0028] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0029] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0030] Figure 1 This is a flowchart of an intelligent agent cooperative navigation method based on communication opinions provided in an embodiment of this disclosure;

[0031] Figure 2 This is a schematic diagram of a dynamic communication graph generation logic provided by an embodiment of the present disclosure, which allows multiple agents to select communication objects on demand based on task objectives and environmental constraints.

[0032] Figure 3 This is a schematic diagram of the core execution and training logic of a multi-agent cooperative navigation method based on communication opinions provided in this embodiment of the disclosure;

[0033] Figure label:

[0034] 1-Agent, 2-Obstacle, 3-Destination. Detailed Implementation

[0035] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0036] The following description and accompanying drawings fully illustrate specific embodiments of the invention to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Throughout this document, each embodiment may be referred to individually or collectively with the term "invention," which is merely for convenience and is not intended to automatically limit the scope of the application to any single invention or inventive concept if more than one invention is disclosed. Relational terms such as "first" and "second" are used herein only to distinguish one entity or operation from another entity or operation without requiring or implying any actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or vehicle that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or vehicle. Unless otherwise specified, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or vehicle that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the methods, products, etc., disclosed in the embodiments, since they correspond to the method sections disclosed in the embodiments, the descriptions are relatively simple; relevant details can be found in the method section description.

[0037] Existing multi-agent communication and decision-making communication structures lack adaptability, often employing fixed or fully connected topologies that cannot dynamically adjust to different task stages or spatial regions, resulting in significant redundant communication and wasted bandwidth. Communication and action decision-making processes are independent, with the communication network only responsible for information transmission rather than decision-driven processes. The lack of a mapping mechanism from communication semantics to task opinions leads to unstable collaborative behavior. The system lacks security and consistency constraints, potentially causing speed fluctuations, path conflicts, and collaborative delays when the environment changes dynamically or the number of agents increases. Existing training frameworks struggle to balance communication efficiency with global performance optimization; centralized solutions suffer from heavy computational burdens, while distributed solutions exhibit slow convergence and poor generalization.

[0038] To address the problems existing in related technologies, embodiments of this disclosure construct an intelligent agent cooperative navigation system that jointly optimizes communication and actions. By introducing a communication opinion mechanism, the agent can dynamically evaluate the communication value while performing tasks, autonomously select communication objects, and establish a sparse and efficient topology. Through a reinforcement learning framework, the system achieves joint optimization of policy, communication, and security, significantly reducing communication costs while ensuring cooperative effectiveness. Embodiments of this disclosure also design a multi-objective loss function that optimizes task rewards while considering communication costs, opinion consistency, and security constraints, thereby achieving efficient, stable, and scalable multi-agent cooperative control.

[0039] See Figure 1 This is an embodiment of an intelligent agent cooperative navigation method based on communication opinions. The method includes:

[0040] S10, at time Each intelligent agent Obtain current local observations and receive the previous moment -1 communication input The A set of opinions passed on to other intelligent agents.

[0041] It should be understood, specifically, at time Each intelligent agent Each holds its own current local observations And the communication input from the previous step. Here This represents the set of opinions transmitted to this agent by other agents via communication at the previous moment. Each opinion can be viewed as an expression of the agent's movement trend at a specific moment, represented as a high-level semantic description of the agent's next action trend or behavioral direction in the form of a latent vector, such as expected speed, heading, or policy embedded features.

[0042] Current local observation Data is collected by the agent's sensors, such as cameras, lidar, and GPS.

[0043] S20, based on the neighborhood relationship of the previous time step Current local observations and opinion collection Generate a communication graph showing the communication connection states between intelligent agents. .

[0044] Specifically, based on the neighborhood relationships of the previous time step Construct a communication input graph structure, and [o i t , As input features, the probability p of agent i communicating with all other agents j is generated through a communication policy network. ij t Used to depict time Does each agent need to establish a communication connection with the others? Where 0 ≤ p ij t ≤1, p ij t The larger the value of j, the more valuable the information j is for i's decision-making.

[0045] Sparse opinion-based communication graph structures can be obtained through sampling algorithms with sparsity constraints, such as top-k selection or threshold filtering. ,in Represents intelligent agents Is it related to intelligent agents? An information channel is established at this moment. The design of the communication mechanism enables dynamic, on-demand generation of communication structures, allowing agents to interact only with task-related or high-value agents within limited communication bandwidth, significantly reducing the communication burden. The sparse constraint sampling algorithm uses a differentiable Top-K selection mechanism, specifically:

[0046] The communication policy network outputs the communication probability of each neighbor (j). Differentiable sampling is performed using Gumbel-Softmax to obtain continuous approximations:

[0047]

[0048] in, Temperature parameters The initial value is 1.0, which is reduced to 0.5 every 1000 training steps.

[0049] In the inference phase, directly take If and only if (j) is in It is among the first (K=3) largest values, otherwise it is 0. See also Figure 2 , Figure 2 The diagram shows agent 1, obstacle 2, and destination 3. The dashed lines and different styles of arrows illustrate selective communication mechanisms: agents highly relevant to the task, such as o... i t o j t o l t Dense communication connections are established between agents, while agents with low relevance to the task, such as Okt, maintain only sparse communication. This intuitively presents the dynamic communication graph generation logic of multiple agents selecting communication objects on demand based on task objectives and environmental constraints.

[0050] S30, based on communication diagram Current local observations And the additional terms of motion z i t Generate joint observation features i t .

[0051] Specifically, upon receiving communication opinions Subsequently, each agent aggregates cross-agent information from its neighborhood based on this graph structure. The system employs a message aggregation function. By fusing the features of neighboring agents with the current state, a message summary item is obtained. , representing intelligent agents The characteristics of group information perceived at the current moment. Subsequently, the corresponding message items and local observations, motion-related additional items. Concatenation extended to joint input This allows communication information and local states to be integrated in a unified representation space.

[0052] The message aggregation function An aggregation mechanism based on graph neural networks is used to achieve the following:

[0053] The set of neighbors for each intelligent robot (i) is denoted as . , including all that meet The robot (j) is defined as follows: For its neighbor (j), its feature vector is:

[0054]

[0055] The Encoder is a three-layer Multilayer Perceptron (MLP), with the input dimension being the length of the concatenated observation and opinion vectors, and the output dimension being... Attention weights are calculated using a Graph Attention Network (GAT).

[0056]

[0057] in, and For learnable parameters, (|) denotes vector concatenation;

[0058] Aggregated messages: The output dimension is .

[0059] S40 will combine observation features Input Policy Network Output the action 'a' at the current moment. i t and opinions on actions The action opinions are semantic descriptions of the agent's future behavioral trends.

[0060] Specifically, joint observation features Input to policy network The network outputs two types of results: one is the specific action at the current moment. Secondly, opinions on actions. This refers to a high-level semantic expression of behavioral trends at future moments. Action / Opinion The introduction of this technology enables intelligent agents to make feedforward predictions of the next communication and motion trends when making decisions. For example, it can express high-level instructions such as driving, avoiding, and cooperating in the form of expected speed, heading direction, or embedded semantic vectors, thereby achieving continuous coordination of communication and action.

[0061] Policy Network The Actor-Critic architecture is adopted, and its structure is as follows:

[0062] Encoding layer: Input is After passing through two layers of MLP with dimensions of 256 and 128 respectively, the hidden state is output. ;

[0063] Action Header: Outputs actions mean and variance The final action is obtained by sampling using a Gaussian distribution;

[0064] Opinion Header: Output Action Opinion It is obtained through a linear layer mapping, with dimensions... ;

[0065] The PPO algorithm was used during training, with a clip parameter of 0.2 and a value function coefficient of 0.5.

[0066] S50, the agent executes action a i t Drive system state evolution and receive rewards from environmental feedback. Acquire new local observations at time t+1 and communication input Re-execute S10.

[0067] It should be understood that when the environment returns a reward based on the action performed. Afterwards, each intelligent entity will be in time. Acquire new local observations and receive the communication input generated in the previous step. This process re-enters the decision-making cycle. The entire process forms a strict causal chain in time, that is... This forms a stable reinforcement learning closed-loop structure. For the initial moment, the action opinion can be... Alternatively, a learnable initial token can be used to ensure the system's feasibility during the cold start phase.

[0068] At the training level, this invention employs a mechanism of centralized training and distributed execution of CTDE. During the centralized training phase, the states, actions, communication graphs, and reward information of all agents are centrally input into the global optimizer to calculate multiple joint loss functions. The optimization objectives include: reinforcement learning objectives. Used to update the action decision-making capabilities of the policy network; communication constraints. This is used to limit the redundancy of communication and encourage efficient sparse communication structures; consistency constraints. Ensure consistency between suggested actions and actual execution; safety constraints. This is used to penalize potentially risky actions such as collisions, exceeding limits, or excessive acceleration. The multi-objective joint optimization function takes the form of...

[0069]

[0070] The specific implementation of the multi-objective joint optimization function is as follows:

[0071] The total loss function is:

[0072] .

[0073] The PPO algorithm is used to optimize the action decision-making strategy and state value estimation accuracy of the agent.

[0074] ,in,

[0075] For strategic losses, This results in a loss of value.

[0076] ;

[0077] , where Proj is a linear mapping layer from action to opinion space;

[0078] ,in , Weighting coefficients: , , .

[0079] Through joint optimization, this invention can simultaneously balance communication efficiency, action accuracy, and system security during training, ultimately achieving the learning of the globally optimal strategy.

[0080] This embodiment uses a collaborative transport task involving four autonomous mobile robots as an example. Each robot is equipped with a LiDAR, an RGB-D camera, an IMU, and a wireless communication module. The task requires the four robots to work together to transport a cube of goods from the starting area A to the target area B, while avoiding dynamic obstacles and maintaining a stable formation.

[0081] At time t, each robot first acquires local observations through its sensors. This includes the robot's own pose, velocity, relative position to the cargo, distance to the nearest obstacle, and load status. Simultaneously, each robot receives action feedback broadcast by other robots from the previous moment. This opinion is a 32-dimensional semantic vector, encoding the expected movement direction and cooperative intent of each robot. The system bases its decisions on the neighborhood relationships from the previous time step. Current observations and opinion collection Communication selection probabilities are generated through a communication policy network. The communication graph at the current time is constructed using a Top-2 sparse sampling mechanism. This enables dynamic, on-demand communication.

[0082] Based on communication graph Each robot aggregates the observations and opinions of its neighboring robots through a graph attention network to form a message summary item. .Will , Additional motion items in the representation task phase splicing together joint observation features Input to policy network The policy network outputs two parts: the current action containing linear velocity and angular velocity commands. And opinions on the next move. The action suggestions are then broadcast to all robots for communication decisions in the next moment.

[0083] Robot performs actions Afterwards, the environment returns a reward. This includes rewards for approaching the target, rewards for maintaining formation, and penalties for collisions. System entry time. Obtain new observations With communication input The above process is repeated cyclically. The training phase adopts a centralized training and distributed execution architecture, and optimizes task reward, communication sparsity, consensus and motion safety through multi-objective loss functions, ultimately achieving adaptive multi-robot cooperative handling.

[0084] See Figure 3 The core execution and training logic of the multi-agent cooperative navigation method based on communication opinions is shown below. The functions and relationships of each step are as follows:

[0085] Local observation + historical communication input: the starting point of the process, the agent acquires its own local environmental perception, and at the same time receives the action opinions transmitted by other agents in the previous moment, providing initial information for decision-making.

[0086] Intent broadcasting / receiving opinions from others: The agent broadcasts its own action opinions to the outside world, while collecting opinions from neighboring agents to achieve intent sharing, providing a basis for subsequent communication decisions and information aggregation.

[0087] Opinion convergence and observation fusion form communication features: By fusing local observations with collected opinions from others, communication features for communication decision-making are generated, which contain comprehensive information on individual status and group intentions.

[0088] Communication Network: Selective Edge Connection Based on Others' Opinions to Generate a Sparse Communication Graph: Based on communication characteristics, the probability of choosing to communicate with other agents is calculated through a communication policy network, and then a sparse communication graph is constructed through sparse constraints, retaining only communication connections with key neighbors to avoid redundant transmission.

[0089] Message aggregation: Collect key neighbor information by graph: Based on sparse communication graph, message aggregation function is used to collect effective information of key neighbors and further refine the decision-making basis at the group level.

[0090] Policy Network: Outputs current action + next action suggestion: By fusing local observations and aggregated neighbor information, the policy network outputs two parts: one is the current action used to drive the agent to execute; the other is the next action suggestion used for communication in the next time step.

[0091] Environmental feedback: rewards and new communication inputs: After an action is applied to the environment, the environment returns two parts of feedback: one is a reward; the other is new communication inputs, which enable the process to enter the next closed loop.

[0092] CTDE Training and Joint Optimization: Using a centralized training and distributed execution CTDE mechanism, observations, actions, opinions, and rewards of all agents are collected as data during the training phase. The model is then globally optimized through a multi-objective joint optimization function. Meanwhile, communication statistics and regularization terms constrain the sparsity and effectiveness of communication behavior, thereby improving the overall performance of the system.

[0093] This invention utilizes a learnable communication opinion mechanism to achieve efficient multi-agent collaboration. Therefore, several feasible alternatives exist at the implementation level. For example, the communication opinion generation part can be replaced by a probabilistic side-sampling model based on a graph attention network (GAT) instead of reinforcement learning, further reducing training difficulty; the opinion consensus module can employ contrastive learning or self-supervised prediction methods to improve the robustness of semantic embedding; and the security constraint term can be implemented through a control barrier function (CBF) or dynamic security planning based on reachability sets. Furthermore, during the action execution phase, model predictive control (MPC) or distributed optimization algorithms can be combined to correct the policy output in real time, further improving execution accuracy and security. Although the above alternatives differ in their implementation, they all adhere to the core principles of adaptive communication, opinion consensus, and joint optimization, and therefore can be considered technically equivalent to this invention.

[0094] This invention solves the problem of cooperative navigation of multiple agents in partially observable dynamic environments by means of communication opinion mechanism and multi-objective joint optimization, and can be applied to fields such as autonomous driving, drone formation, and intelligent robots.

[0095] The foregoing description and accompanying drawings fully illustrate embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included or replace parts and features of other embodiments. As used herein, the term “and / or” means including one or more of the associated listed elements and all possible combinations thereof. Additionally, when used in this application, the terms “comprise” and its variations “comprises” and / or “comprising” refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase “comprising a…” does not exclude the presence of additional identical elements in the process, method, or vehicle that includes said element. In this document, each embodiment focuses on describing the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the above-described apparatuses, devices, and units can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based device that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for cooperative navigation of intelligent agents based on communication opinions, characterized in that, include: S10, at time Each intelligent agent Obtain current local observations and receive the previous moment -1 communication input The A set of opinions passed on to other intelligent agents; S20, based on the neighborhood relationship of the previous time step Current local observations and opinion collection Generate a communication graph showing the communication connection states between intelligent agents. ; S30, based on communication diagram Current local observations and sports additional items z i t Generate joint observation features i t ; S40 will combine observation features Input Policy Network Output the action 'a' at the current moment. i t and opinions on actions The action opinions are semantic descriptions of the agent's future behavioral trends; S50, the agent executes action a i t Drive system state evolution and receive rewards from environmental feedback. Acquire new local observations at time t+1 and communication input Re-execute S10; S20, based on the neighborhood relationship of the previous time step Current local observations and opinion collection Generate a communication graph showing the communication connection states between intelligent agents. ,include: Based on the neighborhood relationship at the previous time step Construct a communication input graph structure, and [o i t , As input features, the communication selection probability p is generated through the communication policy network. ij t The sparse communication graph C is obtained through the sparse constraint sampling algorithm. t The C t c represents the communication connection state between intelligent agents. ij t constitute; S30, based on communication diagram Current local observations and additional terms of motion z i t Generate joint observation features i t ,include: Based on communication graph C t Through message aggregation functions By aggregating the features of neighboring agents and their current states, a message summary item M is obtained. i t , will o i t M i t Additional sports items z i t splicing to generate joint observation features i t ; A centralized training and distributed execution CTDE mechanism is adopted. During the centralized training phase, the states, actions, communication graphs, and reward information of all agents are centrally input into the global optimizer to calculate a multi-objective joint loss function. The multi-objective joint optimization function takes the form of… ， To reinforce learning objectives, For communication constraints, For consistency constraints, For safety constraints, α, β, and γ are weighting coefficients.

2. The agent cooperative navigation method based on communication opinions according to claim 1, characterized in that, The sparse constraint sampling algorithm is either the top-k selection algorithm or the threshold filtering algorithm.

3. The agent cooperative navigation method based on communication opinions according to claim 1, characterized in that, The message aggregation function A graph neural network aggregation mechanism is adopted to fuse the local observation features and action opinion features of neighboring agents.

4. The agent cooperative navigation method based on communication opinions according to claim 1, characterized in that, The security constraints Used to punish collisions between intelligent agents, boundary violations, and excessive acceleration operations.

5. The agent cooperative navigation method based on communication opinions according to claim 4, characterized in that, The communication constraints Used to limit communication redundancy and encourage the construction of sparse and efficient communication topologies.

6. The agent cooperative navigation method based on communication opinions according to claim 4, characterized in that, The consistency constraint This is used to ensure the consistency of behavior between the proposed action and the actual action performed.

7. The agent cooperative navigation method based on communication opinions according to claim 1, characterized in that, The specific forms of the opinions include desired speed, heading direction, or strategy embedding features.

Images