A path planning method and device for multiple agents

Abstract

The present application relates to a kind of path planning method and device for multiple agents, it is related to multiple agent cooperative control technical field, it is solved that the signal strength between agent is not included in multiple agent cooperative planning system in prior art, leading to multiple agent system in actual operating environment Information link interruption, decision out of sync and other problems prone to occur.The method includes: according to the obstacle point cloud collected by each agent to establish three-dimensional probability occupancy map;The position of each agent and three-dimensional probability occupancy map are input into the pre-trained communication signal prediction deep neural network model, and the communication signal strength prediction three-dimensional map is obtained;Communication signal strength prediction three-dimensional map and three-dimensional probability occupancy map are fused, and two-dimensional grid map is established;Based on the income function optimization of greedy strategy, the highest income target point is selected for each agent;Based on the position of each agent and its corresponding highest income target point, trajectory is planned for it in two-dimensional grid map.

Description

Technical Field

[0001] This invention relates to the field of multi-agent cooperative control technology, and in particular to a path planning method and apparatus for multi-agent systems. Background Technology

[0002] With the rapid development of unmanned systems technology, multi-agent cooperative search has demonstrated irreplaceable value in critical scenarios such as post-disaster rubble rescue and resource exploration in complex terrain. The core requirement of these scenarios is to achieve rapid discovery, location, and confirmation of targets within a designated area through the division of labor and cooperation of multiple agents, while ensuring the continuity and reliability of the mission in complex environments. However, these scenarios often involve complex environments with severe electromagnetic interference, unstable or interrupted communication, and complex terrain, posing significant bottlenecks to existing multi-agent cooperative search technologies in addressing such environments.

[0003] To address the aforementioned issues, traditional multi-agent cooperative search and control strategies (such as those for autonomous vehicles and drones) largely treat communication resources as ideal conditions, focusing primarily on planning optimization and task allocation, while neglecting the dynamic changes in communication signals in the actual environment. Existing technologies lack a correlation between communication signal strength and robot motion states, and do not incorporate signal strength between agents into the multi-agent cooperative planning system. This leads to problems such as information link interruptions and asynchronous decision-making in real-world operating environments. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to solve the problems of information link interruption and decision-making asynchrony that occur in multi-agent systems in actual operating environments due to the fact that the existing technology does not incorporate the signal strength between agents into the multi-agent cooperative planning system.

[0005] In a first aspect, embodiments of the present invention provide a path planning method for multiple agents, including: A three-dimensional probabilistic occupancy map is built based on the obstacle point cloud collected by each intelligent agent; The positions of each agent and the three-dimensional probability occupancy map are input into a pre-trained deep neural network model for predicting communication signals to obtain a three-dimensional map for predicting communication signal strength. The three-dimensional map of predicted communication signal strength is fused with the three-dimensional probability occupancy map, and a two-dimensional grid map is established based on the fusion result. The area in the three-dimensional map of predicted communication signal strength where the communication signal strength is less than a preset signal strength threshold is set as the occupancy area of ​​the obstacle. The reward function optimization based on a greedy strategy selects the target point with the highest reward for each agent from multiple target points in the two-dimensional grid map, wherein the reward function includes communication cost based on communication signal strength; and Based on the location of each agent and its corresponding target point with the highest profit, a trajectory is planned for it in the two-dimensional grid map.

[0006] Further improvements to the path planning method described above include building a 3D probabilistic occupancy map based on the obstacle point cloud collected by each agent, including: Based on the field of view of each agent, the obstacle point cloud is transformed into an obstacle point cloud within the field of view of the agent. Based on the map resolution, the point cloud of obstacles in the agent's field of vision is converted into a three-dimensional occupied map; Based on the positions of each agent and the line-of-sight connections between the point clouds of obstacles in the agents' field of vision, the 3D occupancy map is transformed into a 3D probabilistic occupancy map; and The current state of each pixel in the 3D probabilistic map is updated based on the obstacle point cloud, wherein the state of each pixel in the 3D probabilistic map includes unknown, occupied, and unoccupied.

[0007] Based on further improvements to the above path planning method, the pre-trained deep neural network model for predicting communication signals includes: Path network prediction model and attenuation network prediction model; among which, The path network prediction model takes into account the potential location sequence of the communication signal receiver agent, the location of the communication signal transmitter agent, and the three-dimensional probability occupancy map as input, and outputs features of the communication signal propagation path. The input to the attenuation network prediction model includes the characteristics of the communication signal propagation path and the three-dimensional probability occupancy map, and the output is a three-dimensional map of communication signal strength prediction.

[0008] Based on further improvements to the above path planning method, the potential location sequence of the communication signal receiver agent... Determined by the following method:

[0009] in, The threshold for the driving distance of the intelligent agent receiving the communication signal. The potential location of the intelligent agent receiving the communication signal. The current position of the intelligent agent receiving the communication signal. For calculation and distance, The pixels in the three-dimensional probabilistic occupied map that are in an unoccupied state.

[0010] Further improvements to the above path planning method include fusing the communication signal strength prediction 3D map with the 3D probability occupancy map, including: The state of each pixel in the 3D probability-occupied map is updated according to the following formula:

[0011] in, To preset the signal strength threshold, The pixels in the map that occupy the three-dimensional probability The strength of the communication signal.

[0012] Based on a further improvement of the above path planning method, the revenue function is:

[0013]

[0014] in, Let i be the total cost for the i-th agent to reach the j-th target point from its current position. Let J be the size of the unknown region surrounding the j-th target point. Let be the path distance from the current position of the i-th agent to the j-th target point. The delay penalty is used to punish agents for changing their behavior at the target point. Communication rewards The communication signal strength in the region of target point j. Let J be the integral of the communication signal strength of the agent traveling to the target point j region. As a weighting factor, , , They are respectively , , The weighting coefficients.

[0015] Based on further improvements to the above path planning method, and based on the agent's position and the target point with the highest reward, the planned trajectory for it in the two-dimensional grid map includes: An improved A* algorithm is used to generate a global planning trajectory from the agent's position to the target point with the highest reward on the two-dimensional grid map, wherein the cost function used by the improved A* algorithm includes a signal cost based on the communication signal strength; The global planning trajectory is optimized using a trajectory optimization method based on B-spline curve parameterization, wherein the objective function used by the trajectory optimization method includes a signal strength constraint term based on the communication signal strength.

[0016] Based on a further improvement to the path planning method described above, the signal cost is determined by the following formula:

[0017] in, The signal cost for the child nodes. The signal cost of the parent node. These are the global signal weighting coefficients. For the agent to move from its starting position to its child node The length of the trajectory generated in time, K is the distance from the agent's starting position to the child node. The total number of sampling points of the trajectory generated at that time. The value cost for each of the sampling points; in, Determined by the following formula:

[0018] in, The two-dimensional raster map is defined as follows: α is the weighting factor, and β is the sensitivity factor.

[0019] Based on the further improvement of the above path planning method, the signal strength constraint term is determined by the following formula:

[0020] in, For signal strength constraints, Nc is the number of control points for the B-spline curve. This refers to the two-dimensional raster map.

[0021] Secondly, an embodiment of the present invention provides a path planning device for multiple agents, comprising: The unit is used to build a 3D probabilistic occupancy map based on the obstacle point cloud collected by each agent; The unit is used to input the position of each agent and the three-dimensional probability occupancy map into a pre-trained deep neural network model for predicting communication signals, so as to obtain a three-dimensional map for predicting communication signal strength. The fusion establishment unit is used to fuse the communication signal strength prediction three-dimensional map with the three-dimensional probability occupancy map, and establish a two-dimensional grid map based on the fusion result, wherein the area in the communication signal strength prediction three-dimensional map where the communication signal strength is less than a preset signal strength threshold is set as the occupancy area of ​​the obstacle. A filtering unit is used for optimizing the reward function based on a greedy strategy, selecting the target point with the highest reward for each agent from multiple target points in the two-dimensional grid map, wherein the reward function includes communication cost based on communication signal strength; and The trajectory planning unit is used to plan trajectories for each agent in the two-dimensional grid map based on the position of each agent and its corresponding target point with the highest benefit.

[0022] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. The present invention can plan and optimize the search trajectory of a multi-agent system in an unknown environment based on the prediction model of inter-agent communication signals. It effectively avoids the situation where the agent enters the signal blind zone and loses communication during the actual search process. It solves the problem in the existing multi-agent cooperative search method that ignores the strength of inter-agent communication signals, which easily leads to information link interruption and asynchronous decision-making.

[0023] 2. The present invention incorporates the consideration of communication signal strength into the planning and control system of the intelligent agent. It avoids the intelligent agent from entering the signal blind zone on a global scale while optimizing the local trajectory, so that the intelligent agent can still move along the high signal strength area during the collaborative search process, thus ensuring the stability of the communication link of the multi-robot system.

[0024] 3. Compared with traditional prediction methods, the neural network-based communication prediction module proposed in this invention has significantly improved prediction accuracy and can accurately predict the signal strength between machines under denial conditions.

[0025] 4. The introduction of communication signal strength in the present invention effectively helps multi-robot systems avoid entering signal coverage blind spots, ensuring the stability of the communication link of multi-robot systems.

[0026] 5. The present invention uses a "prediction-adjustment" closed loop, which enables the system to autonomously respond to changes in terrain and fluctuations in interference intensity, and achieve global optimal search in complex scenarios.

[0027] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0028] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 A flowchart illustrating a multi-agent path planning method according to an embodiment of the present invention is shown.

[0029] Figure 2 A schematic diagram of an example of a deep neural network model for predicting communication signals according to an embodiment of the present invention is shown.

[0030] Figure 3A structural block diagram of a multi-agent path planning device according to an embodiment of the present invention is shown. Detailed Implementation

[0031] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0032] According to one aspect of the present invention, a path planning method for multiple agents is provided. Figure 1 A flowchart illustrating a multi-agent path planning method according to an embodiment of the present invention is shown. Figure 1 As shown, this multi-agent path planning method includes: Step 101: Establish a 3D probabilistic occupancy map based on the obstacle point cloud collected by each agent.

[0033] In this embodiment, whenever the agent collects a frame of obstacle point cloud, it is represented as a set of points of the obstacle in three-dimensional space. dot set Each obstacle point The coordinates in the three-dimensional coordinate system are The visual sensors equipped on the agent have a limited field of view. Therefore, to ensure that the agent can perform a comprehensive and accurate scan of the search area, only points within the field of view of the agent's visual sensors are updated into the map, and the updated point cloud is then added to the map. It can be represented as:

[0034] in, It is the location of the intelligent agent. It refers to the field of view range of the intelligent agent. It refers to the field of vision range of the intelligent agent. Point cloud of obstacles in the field of view of the intelligent agent.

[0035] In this embodiment, in order to visualize the point cloud of obstacles in the agent's field of vision... Updating the map requires adjusting the point cloud of obstacles within the agent's field of vision according to the map's resolution. Perform adaptation adjustments using the following formula:

[0036] in, It's the map resolution. Indicates the starting element of the map Coordinates in a three-dimensional coordinate system It is a 3D occupancy map. These are the pixels of a 3D occupied map. A 3D occupied map refers to a 3D map that includes obstacles. ... In a 3D occupied map, the spatial points traversed by the line of sight between the agent's own position and obstacles are:

[0037] Where m is the number of points on the line of sight, and k is the k-th point representing the obstacle. It is a three-dimensional point map.

[0038] In this embodiment, a three-dimensional spatial point map is used. and 3D occupied map By merging the data, a three-dimensional probability occupancy map can be obtained.

[0039] In this embodiment, based on a certain frame of obstacle point cloud Obtain a three-dimensional point map and 3D occupied map Afterwards, at this time, Map elements in An occupied state was observed once. Map elements in An unoccupied state is observed once. The state value of each pixel in the 3D probabilistic occupancy map is then updated according to the following formula:

[0040]

[0041]

[0042]

[0043] in, This represents the state value of the pixel in the map that occupies a 3D probability in the previous observation. This represents the probability that a pixel in a 3D probabilistically occupied map will be observed to be in an occupied state once. This represents the probability that a pixel in a 3D probabilistically occupied map is observed to be in an occupied state once.

[0044] The state corresponding to the state value of each pixel in the 3D probability-occupied map can be determined by the following formula:

[0045] in, This represents the threshold between the unoccupied and occupied states. This represents the initial state value.

[0046] Step 102: Input the positions of each agent and the three-dimensional probability occupancy map into the pre-trained deep neural network model for predicting communication signals to obtain a three-dimensional map for predicting communication signal strength.

[0047] In this embodiment, a deep neural network model for predicting communication signals is used to predict the strength of communication signals between agents, which can transform a three-dimensional probability occupancy map into a three-dimensional map for predicting communication signal strength.

[0048] Figure 2 A schematic diagram illustrating the structure of an example deep neural network model for predicting communication signals according to an embodiment of the present invention is shown. Figure 2 As shown, the RSSI (Received Signal Strength Indication) signal prediction network model includes a path prediction network model and an attenuation prediction network model. In this embodiment, the path prediction network model and the attenuation prediction network model can be trained separately using a data-driven approach. The path prediction model can be designed as a fully connected network model. Its input layer includes the potential location sequence of the receiving agent (such as an autonomous vehicle), the location of the transmitting agent, and a probability occupancy map of the 3D obstacle space point cloud. The output layer includes the features of the RSSI signal propagation path. It should be noted that... Figure 2 The 3D obstacle spatial point cloud pair probabilistic occupancy map in the image is the same as the 3D probabilistic occupancy map in this paper. The attenuation network prediction model can be designed using a fully connected network model. Its input layer includes the characteristics of the RSSI signal propagation path and the 3D obstacle spatial point cloud pair probabilistic occupancy map, and the output layer is the 3D map of RSSI signal strength prediction. Figure 2 In this study, the training data for the RSSI signal prediction network model includes the location of the RSSI signal sender and receiver, as well as the amplitude of the received RSSI signal.

[0049] like Figure 2 As shown, the potential position sequence of the receiver agent can be obtained by estimating the agent's range of motion. Specifically, it can be obtained according to the following formula:

[0050] in, It is the threshold for the agent's driving distance. It is the potential location of the intelligent agent. It is the current position of the agent. It calculates the distance between the agent's potential location and its current location. It can be obtained using the A* algorithm or other trajectory planning algorithms. It is the potential position sequence of the agent.

[0051] exist Figure 2 In this context, the potential location sequence of the agent is... The location of the RSSI signal sender and the probability occupancy map of the three-dimensional obstacle space point cloud are input into the RSSI signal prediction network model, which can generate the signal strength of the potential location of the agent, i.e., the three-dimensional map of communication signal strength prediction in this embodiment.

[0052] Step 103: Fuse the predicted three-dimensional map of communication signal strength with the three-dimensional probability occupancy map, and establish a two-dimensional grid map based on the fusion result.

[0053] In this embodiment, to prevent the agent from entering a communication signal blind zone, the 3D map of predicted communication signal strength obtained in step 102 and the 3D probability occupancy map obtained in step 101 can be fused. During the fusion process, areas in the 3D map of predicted communication signal strength where the communication signal strength is less than a preset signal strength threshold can be designated as obstacle occupancy areas.

[0054] In some embodiments, fusing a communication signal strength prediction 3D map with a 3D probability occupancy map includes: Update the state of each pixel in the 3D probability-occupied map according to the following formula:

[0055] in, To preset the signal strength threshold, Pixels occupying the map in three dimensions The strength of the communication signal.

[0056] In this embodiment, the fused 3D map can be converted into a 2D raster map. Specifically, the communication signal strengths of pixels with the same x and y coordinates but different heights in the fused 3D map can be added together to obtain a communication signal strength value at that height.

[0057] Step 104: Optimize the reward function based on the greedy strategy, and select the target point with the highest reward for each agent from multiple target points in the two-dimensional grid map.

[0058] In this embodiment, the current state of each element in a two-dimensional raster map can be searched to find a set of boundary points used to distinguish between unknown and known areas of the map. A Rapid Random Tree (RRT) algorithm can be used for this search. After extracting the boundary point set, all boundary points can be filtered using a clustering algorithm to remove redundant boundary points.

[0059] If a multi-agent system has N agents and M target points to be searched in the search area, optimization can be achieved using a greedy strategy's reward function. This involves iterating through all combinations to ensure that each agent has a target point with the highest reward. The reward function includes the communication cost based on the communication signal strength.

[0060] In some embodiments, the payoff function is:

[0061]

[0062] in, Let i be the total cost for the i-th agent to reach the j-th target point from its current position. Let J be the size of the unknown region surrounding the j-th target point. Let be the path distance from the current position of the i-th agent to the j-th target point. The delay penalty is used to punish agents for changing their behavior at the target point. Communication rewards The communication signal strength in the region of target point j. Let J be the integral of the communication signal strength of the agent traveling to the target point j region. As a weighting factor, , , They are respectively , , The weighting coefficients.

[0063] Step 105: Based on the position of each agent and the target point with the highest profit, plan a trajectory for them in the two-dimensional grid map.

[0064] In this embodiment, after selecting the target point with the highest reward for each agent in the multi-agent system, the trajectory of each agent from its current position to its highest-reward target point can be planned on a two-dimensional grid map. The A* algorithm or other trajectory planning algorithms can be used for trajectory planning.

[0065] In some implementations, step 105 includes: Step 1050: Use the improved A* algorithm to generate a global planning trajectory from the agent's position to the target point with the highest reward on the two-dimensional grid map.

[0066] In this embodiment, the cost function used by the improved A* algorithm includes a signal cost based on the communication signal strength.

[0067] In this embodiment, the two-dimensional raster map can first be converted into a value-cost map according to the following formula:

[0068] in, The given two-dimensional grid map is defined by α, which is a weighting factor, and β, which is a sensitivity factor. The larger α is, the more likely the agent is to take a longer route to find strong signals. β is typically set to β≥1.

[0069] The cost function of the A* algorithm is defined as follows:

[0070] in, For heuristic functions, Starting from the target point with the highest return, run the 2D Dijkstra algorithm once on the full 2D raster map to calculate the shortest Euclidean distance from the current raster to the target raster. It consists of the following three parts:

[0071] It represents the distance cost, indicating the cumulative Euclidean distance traveled by the intelligent agent; It is a kinematic cost. and This is the cost term used in the traditional A* algorithm. In this embodiment, the improved A* algorithm refers to the addition of a signal cost. Signal cost Determined by the following formula:

[0072] in, The signal cost for the child nodes. The signal cost of the parent node. These are the global signal weighting coefficients. For the agent to move from its starting position to its child node The length of the trajectory generated in time, K is the distance from the agent's starting position to the child node. The total number of sampling points of the trajectory generated at that time. The value cost for each of the sampling points. In this embodiment, the parent node is the currently traversed grid node whose cost has been calculated; the child nodes are candidate path nodes whose costs are to be calculated, extending from the parent node to the surrounding adjacent passable grids.

[0073] Step 1052: Optimize the global planning trajectory using a trajectory optimization method based on B-spline curve parameterization.

[0074] The global planning trajectory generated in step 1050 can be parameterized using B-spline curves, which are uniquely determined by the degree pb, Nc control points {Q1,Q2,…,QNc}, and the node vector {t1,t2,…,tM}. The optimization objective function J is:

[0075] in, It is an obstacle avoidance item. It is a smoothing term. It is a feasible option. It is a multi-vehicle collision avoidance system. This is a signal strength constraint term. It should be noted that... , , , This is the cost term used in traditional trajectory optimization methods based on B-spline curve parameterization. This embodiment improves upon this by adding a signal strength constraint term. .

[0076] Signal strength constraint Determined by the following formula:

[0077] in, For signal strength constraints, Nc is the number of control points for the B-spline curve. The given 2D raster map. This is a gradient-based optimization term that generates a gradient force if a pathpoint is in a weak signal area, pushing the pathpoint towards a nearby strong signal raster.

[0078] According to another aspect of the present invention, a path planning apparatus for multiple agents is provided. For example... Figure 2As shown, the device 200 includes: a building unit 201, used to build a three-dimensional probabilistic occupancy map based on obstacle point clouds collected by each agent; a obtaining unit 202, used to input the positions of each agent and the three-dimensional probabilistic occupancy map into a pre-trained deep neural network model for predicting communication signals to obtain a three-dimensional map for predicting communication signal strength; a fusion building unit 203, used to fuse the three-dimensional map for predicting communication signal strength with the three-dimensional probabilistic occupancy map, and build a two-dimensional grid map based on the fusion result, wherein the area in the three-dimensional map for predicting communication signal strength where the communication signal strength is less than a preset signal strength threshold is set as the occupancy area of ​​the obstacle; a filtering unit 204, used to select the target point with the highest benefit for each agent from multiple target points in the two-dimensional grid map based on a greedy strategy's benefit function optimization, wherein the benefit function includes a communication cost based on the communication signal strength; and a trajectory planning unit 205, used to plan a trajectory for each agent in the two-dimensional grid map based on the position of each agent and its corresponding target point with the highest benefit.

[0079] It is understandable that unit 201 in device 200 Operation of Unit 205 Figure 1 Step 101 Step 105 is similar and will not be described in detail here.

[0080] Compared with the prior art, the embodiments of the present invention can achieve at least one of the following beneficial effects: 1. The present invention can plan and optimize the search trajectory of a multi-agent system in an unknown environment based on the prediction model of inter-agent communication signals. It effectively avoids the situation where the agent enters the signal blind zone and loses communication during the actual search process. It solves the problem in the existing multi-agent cooperative search method that ignores the strength of inter-agent communication signals, which easily leads to information link interruption and asynchronous decision-making.

[0081] 2. The present invention incorporates the consideration of communication signal strength into the planning and control system of the intelligent agent. It avoids the intelligent agent from entering the signal blind zone on a global scale while optimizing the local trajectory, so that the intelligent agent can still move along the high signal strength area during the collaborative search process, thus ensuring the stability of the communication link of the multi-robot system.

[0082] 3. Compared with traditional prediction methods, the neural network-based communication prediction module proposed in this invention has significantly improved prediction accuracy and can accurately predict the signal strength between machines under denial conditions.

[0083] 4. The introduction of communication signal strength in the present invention effectively helps multi-robot systems avoid entering signal coverage blind spots, ensuring the stability of the communication link of multi-robot systems.

[0084] 5. The present invention uses a "prediction-adjustment" closed loop, which enables the system to autonomously respond to changes in terrain and fluctuations in interference intensity, and achieve global optimal search in complex scenarios.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A path planning method for multi-agent systems, characterized in that, include: A three-dimensional probabilistic occupancy map is built based on the obstacle point cloud collected by each intelligent agent; The positions of each agent and the three-dimensional probability occupancy map are input into a pre-trained deep neural network model for predicting communication signals to obtain a three-dimensional map for predicting communication signal strength. The three-dimensional map of predicted communication signal strength is fused with the three-dimensional probability occupancy map, and a two-dimensional grid map is established based on the fusion result. The area in the three-dimensional map of predicted communication signal strength where the communication signal strength is less than a preset signal strength threshold is set as the occupancy area of ​​the obstacle. The reward function optimization based on a greedy strategy selects the target point with the highest reward for each agent from multiple target points in the two-dimensional grid map, wherein the reward function includes communication cost based on communication signal strength; and Based on the location of each agent and its corresponding target point with the highest profit, a trajectory is planned for it in the two-dimensional grid map.

2. The path planning method according to claim 1, characterized in that, A 3D probabilistic occupancy map is constructed based on the obstacle point cloud collected by each agent, including: Based on the field of view of each agent, the obstacle point cloud is transformed into an obstacle point cloud within the field of view of the agent. Based on the map resolution, the point cloud of obstacles in the agent's field of vision is converted into a three-dimensional occupied map; Based on the positions of each agent and the line-of-sight connections between the point clouds of obstacles in the agents' field of vision, the 3D occupancy map is transformed into a 3D probabilistic occupancy map; and The current state of each pixel in the 3D probabilistic map is updated based on the obstacle point cloud, wherein the state of each pixel in the 3D probabilistic map includes unknown, occupied, and unoccupied.

3. The path planning method according to claim 2, characterized in that, The pre-trained deep neural network model for predicting communication signals includes: Path network prediction model and attenuation network prediction model; among which, The path network prediction model takes into account the potential location sequence of the communication signal receiver agent, the location of the communication signal transmitter agent, and the three-dimensional probability occupancy map as input, and outputs features of the communication signal propagation path. The input to the attenuation network prediction model includes the characteristics of the communication signal propagation path and the three-dimensional probability occupancy map, and the output is a three-dimensional map of communication signal strength prediction.

4. The path planning method according to claim 3, characterized in that, The potential location sequence of the communication signal receiver agent Determined by the following method: in, The threshold for the driving distance of the intelligent agent receiving the communication signal. The potential location of the intelligent agent receiving the communication signal. The current position of the intelligent agent receiving the communication signal. For calculation and distance, The pixels in the three-dimensional probabilistic occupied map that are in an unoccupied state.

5. The path planning method according to claim 2, characterized in that, The fusion of the predicted 3D map of communication signal strength with the 3D probability occupancy map includes: The state of each pixel in the 3D probability-occupied map is updated according to the following formula: in, To preset the signal strength threshold, The pixels in the map that occupy the three-dimensional probability The strength of the communication signal.

6. The path planning method according to claim 2, characterized in that, The profit function is: in, Let i be the total cost for the i-th agent to reach the j-th target point from its current position. Let J be the size of the unknown region surrounding the j-th target point. Let be the path distance from the current position of the i-th agent to the j-th target point. The delay penalty is used to punish agents for changing their behavior at the target point. Communication rewards The communication signal strength in the region of target point j. Let J be the integral of the communication signal strength of the agent traveling to the target point j region. As a weighting factor, , , They are respectively , , The weighting coefficients.

7. The path planning method according to claim 2, characterized in that, Based on the location of each agent and its corresponding target point with the highest reward, the trajectory planned for them in the two-dimensional grid map includes: An improved A* algorithm is used to generate a global planning trajectory from the agent's position to the target point with the highest reward on the two-dimensional grid map, wherein the cost function used by the improved A* algorithm includes a signal cost based on the communication signal strength; The global planning trajectory is optimized using a trajectory optimization method based on B-spline curve parameterization, wherein the objective function used by the trajectory optimization method includes a signal strength constraint term based on the communication signal strength.

8. The path planning method according to claim 7, characterized in that, The signal cost is determined by the following formula: in, The signal cost for the child nodes. The signal cost of the parent node. These are the global signal weighting coefficients. For the agent to move from its starting position to its child node The length of the trajectory generated in time, K is the distance from the agent's starting position to the child node. The total number of sampling points of the trajectory generated at that time. The value cost for each of the sampling points; in, Determined by the following formula: in, The two-dimensional raster map is defined as follows: α is the weighting factor, and β is the sensitivity factor.

9. The path planning method according to claim 7, characterized in that, The signal strength constraint term is determined by the following formula: in, For signal strength constraints, Nc is the number of control points for the B-spline curve. This refers to the two-dimensional raster map.

10. A path planning device for multiple agents, characterized in that, include: The unit is used to build a 3D probabilistic occupancy map based on the obstacle point cloud collected by each agent; The unit is used to input the position of each agent and the three-dimensional probability occupancy map into a pre-trained deep neural network model for predicting communication signals, so as to obtain a three-dimensional map for predicting communication signal strength. The fusion establishment unit is used to fuse the communication signal strength prediction three-dimensional map with the three-dimensional probability occupancy map, and establish a two-dimensional grid map based on the fusion result, wherein the area in the communication signal strength prediction three-dimensional map where the communication signal strength is less than a preset signal strength threshold is set as the occupancy area of ​​the obstacle. A filtering unit is used for optimizing the reward function based on a greedy strategy, selecting the target point with the highest reward for each agent from multiple target points in the two-dimensional grid map, wherein the reward function includes the communication cost based on the communication signal strength. as well as The trajectory planning unit is used to plan trajectories for each agent in the two-dimensional grid map based on the position of each agent and its corresponding target point with the highest benefit.

Images