A multi-agent path planning method for complex environment

Abstract

This invention belongs to the field of multi-agent path planning technology, specifically providing a multi-agent path planning method for complex environments. It treats all agents as a whole for movement, gradually expanding to generate collision-free paths. The method includes the following steps: initializing the structure; limiting the sampling range during the sampling phase and setting a dynamic threshold, gradually expanding the sampling area with increasing iterations, and employing an accept-reject sampling method to enhance the sampling frequency around the target point; using a multi-step expansion approach combined with a priority inheritance mechanism to allocate agent priorities and avoid path conflicts; applying rewiring operations to optimize path quality; resetting priorities based on the distance from the agent to the target point, and using a priority inheritance backtracking algorithm as a local connector to ensure all agents reach the target location efficiently and without conflict. This method demonstrates high-efficiency path planning capabilities in complex dynamic environments and is particularly suitable for industrial and medical fields.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent path planning technology, specifically relating to a multi-agent path planning method for complex environments. Specifically, in a shared environment, it plans paths for multiple agents (such as robots, unmanned vehicles, drones, etc.) so that each agent moves from its starting position to its target position while avoiding collisions with obstacles and other agents. Background Technology

[0002] Over the past few decades, mobile robot systems have demonstrated significant value across numerous industries. Multi-agent path planning (MAPF) systems are commonly used to perform a series of transportation tasks. However, deploying these MAPF systems in compact environments such as factories and hospitals presents challenges in effective scheduling and precise coordination. Especially in scenarios involving dozens to hundreds of agents operating simultaneously, ensuring that each agent can complete its task safely and efficiently without conflict becomes crucial. As mobile robot technology matures in industries such as manufacturing, warehousing, and healthcare, the demand for MAPF in real-world scenarios is growing. MAPF typically requires planning safe paths from a starting point to a destination for multiple robots or other mobile agents, avoiding collisions with environmental obstacles and other agents. In relatively simple environments, existing algorithms can adequately meet application requirements; however, in large-scale deployments or space-constrained scenarios, traditional algorithms are susceptible to the increased environmental complexity and the number of agents, making it difficult to maintain efficient and stable planning performance.

[0003] On the one hand, complex scenarios such as factories and hospitals often involve narrow passages, temporary obstacles, and dynamically changing work areas, which significantly increases the difficulty of coordination between agents. On the other hand, when there are many agents, even if the individual planning of each agent seems feasible, potential path conflicts and resource contention can arise during collaborative work. To address this, researchers have proposed various MAPF solutions, including search-based, metaheuristic-based, or rule-based conflict resolution methods. However, these methods often fall short in dealing with scenarios with frequent conflicts and fail to meet the demand for efficient coordination in complex scenarios. Most existing multi-agent path planning methods consider relatively idealized situations and do not fully account for the complexity and uncertainty of real-world environments. For example, the patent document CN 113778090A, which discloses a "mobile robot path planning method based on ant colony optimization and PRM algorithm," first requires pre-construction of a global PRM map in the environment and is highly dependent on the environment. Furthermore, its modeling and maintenance complexity and time overhead are considerable, and its adaptability and stability in multi-agent path planning conflict coordination in large-scale or compact space applications still need improvement. Summary of the Invention

[0004] To address the aforementioned issues in the background technology, this invention provides a multi-agent path planning method for complex environments. Based on the existing Discrete Fast Random Tree (DRT) algorithm, it focuses on improvements for real-world scenarios, enabling more effective exploration of the entire state space and playing a role in complex scenarios requiring close agent coordination, thereby enhancing the planning capabilities for multiple agents in complex environments. By optimizing the algorithm's sampling and expansion strategies and introducing mechanisms for inter-agent collaboration, it aims to solve applications in large-scale or compact spaces, especially for transport robots in scenarios such as remote factories, hospitals, and emergency rescue. Since these robots generally lack expensive detection equipment, they urgently need an efficient planning algorithm capable of quickly calculating conflict-free paths for multiple robots to ensure successful task completion under resource-constrained and complex environmental conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-agent path planning method for complex environments, comprising the following steps:

[0006] S1. Read the predetermined workspace map and obtain a set of task instances, which includes the number of multiple agents and the start and end positions of each agent.

[0007] S2. Based on the Discrete Fast Random Tree (dRRT) framework, a tree structure is constructed for path planning to facilitate subsequent path storage configuration and fast querying, treating the movement of all agents as a whole. Multiple collision-free paths from the starting point to the target point are generated through iterative expansion.

[0008] S3. Limit the sampling range of all agents to the shortest path between the current position and the target point, set a dynamic threshold, and gradually expand the sampling area as the number of iterations increases. Use the accept-reject sampling method, adjust the rejection rate according to the Gaussian distribution characteristics, enhance the sampling frequency of the area around the target point, and accelerate the convergence of agents to the target point.

[0009] S4. During path expansion, a priority inheritance mechanism is used to continuously generate effective configurations for multiple agents. Each agent is assigned a priority, and agents move and adjust according to their priority order to ensure that the paths of high-priority agents are not interfered with by low-priority agents.

[0010] S5. After each expansion, the application performs a rewiring operation on the paths of multiple agents to optimize the found paths.

[0011] S6. After completing the rewiring operation, reset the priority of each agent based on its distance from the target point, with agents closer to the target point having higher priority. Then, use the Priority Inheritance Backtracking (PIBT) algorithm (with a fixed number of steps) as a local connector to attempt to move all agents to their target positions efficiently and without conflict.

[0012] Furthermore, in step S1, the user needs to manually set a threshold. , used to determine the boundary values ​​of the initial sampling region.

[0013] Further, step S2 includes:

[0014] S21. Unlike traditional path planning methods, this method does not plan paths for individual agents and then adjust and resolve conflicts in the upper layer. Instead, it treats all agents as a high-dimensional whole, thus avoiding the occurrence of conflicts during movement.

[0015] S22. In a graph consisting of points and edges... Among them It represents the set of all points in the graph. It is the set of all edges in the graph.

[0016] Assumption Represents intelligent agents exist The position of time, within any discrete time interval. In this context, the movement of all agents must adhere to the following two constraints:

[0017] 1. For any two intelligent agents ,if Then there must be .

[0018] 2. For any two agents ,if and Then there must be .

[0019] Further, step S3 includes:

[0020] S31. An acceptance-rejection sampling method is used to dynamically adjust the sampling probability, ensuring that samples are concentrated more in the target area, thereby improving the efficiency of path planning. This method is highly adaptable and can achieve efficient multi-agent coordination in dynamic or highly constrained environments.

[0021] S32, Use a threshold The sampling area is dynamically controlled. When no path can be found within a certain area, the sampling area is dynamically adjusted to search for a feasible path in a larger area.

[0022] S33. Using the distance to the target point as a parameter, dynamically control the change in the acceptance probability of the sampling point. The closer the selected sampling point is to the target point, the greater the probability that the point will be accepted.

[0023] Further, step S4 includes:

[0024] S41. During the expansion phase, the priorities of all agents are fixed. When any two adjacent agents... and (Assuming an intelligent agent) Priority is higher than that of the agent When, if the intelligent agent The next moving goal is And intelligent agents The next moving goal is But intelligent agents The agent is stuck and unable to move. It will temporarily inherit the intelligent agent Prioritize movement and move the position first. Make room for intelligent agents .

[0025] S42. During the expansion phase, a multi-step expansion strategy was employed. Before expansion began, the priorities of all agents were determined based on their distances from their respective target points. During the expansion process, agents farther from the target point would move with higher priority, thus reaching the target point more quickly.

[0026] Further, step S5 includes:

[0027] S51. Due to the random search process, path redundancy is unavoidable. This invention uses a rewiring operation to generate high-quality paths, unlike dRRT which is suitable for a small number of agents. This invention uses A The algorithm evaluates the total cost for rerouting to improve path quality. Among the new placement points generated through multi-step expansion, A is utilized. The path is optimized and its quality improved by replacing the original parent node with the nearest neighbor node that has the lowest evaluation cost.

[0028] S52, A series of new configuration points are generated during the expansion process. arrive and for the first new configuration point Perform a nearest neighbor search. Among these neighboring points, select the one accessible via A. The point with the lowest algorithm evaluation cost is taken as The new parent node replaces the original parent node. This reconnection operation eliminates unnecessary branches, optimizes the path structure, and further improves path quality.

[0029] Further, step S6 includes:

[0030] S61. Local connected components play a crucial role in this algorithm. During the path search process, they determine whether multiple agents can move to their respective target points without collision from their current positions, thereby reducing redundant expansion operations and improving overall efficiency. Therefore, this invention employs a fixed-step PIBT algorithm as a local connected component, enabling PIBT to be combined with the expansion process, significantly reducing computational complexity.

[0031] S62. This communicator implementation does not require pre-calculation of the path before expansion. Instead, it fully utilizes the distance information already calculated during the expansion process to quickly complete subsequent calculations. When conflicts occur between agents, collisions can be avoided through a priority inheritance inversion mechanism. This mechanism, using path data from the expansion process, can quickly determine whether an agent can reach the target point, thereby effectively reducing unnecessary calculations. It is particularly suitable for multi-agent path planning in large-scale complex environments.

[0032] The present invention also includes other components that enable its normal use, all of which are conventional means in the art. In addition, any devices or components not limited in the present invention adopt the prior art in the art.

[0033] The beneficial effects of this invention are as follows:

[0034] 1. This invention provides a multi-agent path planning method for complex environments, which for the first time applies a sampling-based algorithm to the large-scale multi-agent path planning (MAPF) problem and achieves significant results. This algorithm can effectively solve path planning problems in all complex instances, such as Tunnle and T-shape, thus overcoming the current shortcomings of MAPF in solving complex instances.

[0035] 2. In response to the need for narrow spaces and high-density deployment in real-world environments, this invention can ensure the continuous operation of multiple agents while avoiding common conflicts and deadlocks, providing a new and feasible technical solution for large-scale multi-agent collaborative work.

[0036] 3. The core mechanism of the algorithm has been optimized to adapt to large-scale or space-constrained real-world environments. The improved algorithm exhibits stronger robustness and adaptability, ensuring the efficient operation of the MAPF system even in the face of certain uncertainties. Through a more intelligent sampling strategy and path optimization, the solution of this invention significantly improves the efficiency and reliability of path planning while reducing computation time and resource consumption.

[0037] 4. Compared with the traditional MAPF method, this scheme exhibits better adaptability and stability in complex environments, and reduces the computational cost of the algorithm to some extent. This research lays the technical foundation for the further application of multi-agent systems in warehousing and logistics, medical transportation, and other fields, and also provides a feasible reference for future expansion into more complex scenarios. Attached Figure Description

[0038] Figure 1 (a) and (b) are the specific range map of the sampling area and the distance-based acceptance probability map, respectively;

[0039] Figure 2 (a) and (b) are schematic diagrams of the multi-step expansion strategy in the path expansion process of the present invention;

[0040] Figure 3 (a), (b), (c), and (d) are schematic diagrams of the rewiring operation of the present invention;

[0041] Figure 4 (a), (b), and (c) are comparison results of three different sampling methods used in the examples;

[0042] Figure 5 The diagram illustrates six typical small-scale complex examples that can be solved by this invention.

[0043] Figure 6 This is a schematic diagram of the operation of the present invention on the FlexSim simulation platform;

[0044] Figure 7 This is a schematic diagram of the physical device used in the actual deployment of the present invention;

[0045] Figure 8 This is a flowchart illustrating the overall process of the path planning method of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0047] Example

[0048] like Figure 8 As shown, the present invention provides a multi-agent path planning method for complex environments, comprising the following steps:

[0049] S1. Read the predetermined workspace map and obtain a set of task instances, which includes the number of multiple agents and the start and end positions of each agent.

[0050] S2. Based on the Discrete Fast Random Tree (dRRT) framework, a tree structure is constructed for path planning to facilitate the storage configuration and fast query of subsequent paths. The movement of all agents is treated as a whole, and multiple collision-free paths from the starting point to the target point are generated through iterative expansion.

[0051] S3. Limit the sampling range of all agents to the shortest path between the current position and the target point, set a dynamic threshold, and gradually expand the sampling area as the number of iterations increases. Use the accept-reject sampling method, adjust the rejection rate according to the Gaussian distribution characteristics, enhance the sampling frequency of the area around the target point, and accelerate the convergence of agents to the target point.

[0052] S4. During path expansion, a priority inheritance mechanism is used to continuously generate effective configurations for multiple agents. A priority is assigned to each agent, and agents move and adjust according to their priority order to ensure that the paths of high-priority agents are not interfered with by low-priority agents.

[0053] S5. After each expansion, the application performs a rewiring operation on the paths of multiple agents to optimize the found paths.

[0054] S6. After completing the rewiring operation, reset the priority of each agent based on its distance from the target point, with agents closer to the target point having higher priority. Then, use the Priority Inheritance Backtracking (PIBT) algorithm (with a fixed number of steps) as a local connector to attempt to move all agents to their target positions efficiently and without conflict.

[0055] Specifically, in step S1, the user needs to manually set a threshold. , used to determine the boundary values ​​of the initial sampling region.

[0056] Specifically, step S2 includes:

[0057] S21. Unlike traditional path planning methods, this method does not plan paths for individual agents and then adjust and resolve conflicts in the upper layer. Instead, it treats all agents as a high-dimensional whole, thus avoiding the occurrence of conflicts during movement.

[0058] S22. In a graph consisting of points and edges... Among them It represents the set of all points in the graph. It is the set of all edges in the graph.

[0059] Assumption Represents intelligent agents exist The position of time, within any discrete time interval. In this context, the movement of all agents must adhere to the following two constraints:

[0060] 1. For any two intelligent agents ,if Then there must be .

[0061] 2. For any two agents ,if and Then there must be .

[0062] Specifically, step S3 includes:

[0063] S31. An acceptance-rejection sampling method is used to dynamically adjust the sampling probability, ensuring that samples are concentrated more in the target area, thereby improving the efficiency of path planning. This method is highly adaptable and can achieve efficient multi-agent coordination in dynamic or highly constrained environments.

[0064] S32, Use a threshold The sampling area is dynamically controlled. When no path can be found within a certain area, the sampling area is dynamically adjusted to search for a feasible path in a larger area.

[0065] S33. Using the distance to the target point as a parameter, dynamically control the change in the acceptance probability of the sampling point. The closer the selected sampling point is to the target point, the greater the probability that the point will be accepted.

[0066] The basic steps of the specific algorithm are shown in Algorithm 1 below:

[0067]

[0068] In the process of Algorithm 1 above, in each iteration, an accept-rejection sampling technique is used, and the sampling domain is dynamically updated based on the distance from the last point of the tree to the target point. Then, a multi-step expansion strategy is applied, continuously moving towards the target point. Finally, local connectors are used to determine whether all agents can reach the target point from their current positions. The algorithm terminates when a preset number of iterations or running time is reached.

[0069] This process is described in more detail as follows: Figure 1As shown in (a) and (b), this invention presents a schematic diagram of dynamic accept-reject sampling. In sampling-based algorithms, path quality and execution efficiency largely depend on the sampling method. Random sampling has been considered in existing dRRT algorithms. However, random sampling follows a uniform distribution and lacks strong guidance. Although Gaussian sampling is effective in continuous spaces, it becomes impractical in discrete graphs due to the time-consuming pre-computation required, and it also lacks sufficient target guidance. To improve target guidance, this invention introduces the concept of accept-reject sampling, which is highly suitable for complex environments. Accept-reject sampling allows for rapid processing of random discrete points without pre-computation. It provides a higher sampling probability toward the target, thereby improving the efficiency of the algorithm, especially in multi-agent scenarios.

[0070] Specifically, it includes:

[0071] S11. Consider a dynamic sampling region to maintain path feasibility by concentrating the sampling domain within the effective range;

[0072] For accept-rejection sampling, this invention considers a dynamic sampling region to maintain path feasibility while concentrating the sampling domain around the target point. The sampling region is described below:

[0073]

[0074] in Represents intelligent agents A possible sampling point within the sampling area. This is a threshold coefficient that is not fixed but increases with each iteration. This dynamic characteristic allows for the search for feasible paths in a larger region when a suitable path cannot be found in a smaller region. For example... Figure 1 As shown in (a), when When the value is 1, it shows the specific range of the sampling area.

[0075] S12. Use an accept-reject sampling method to guide the agent to the target point as quickly as possible;

[0076] definition The acceptance rates are as follows:

[0077]

[0078] in, The standard deviation represents the distance between the sampling point and the target point. Set directly to an incremental threshold . The gradual increase helps to adaptively adjust the rejection rate within the sampling area, with an emphasis on initially prioritizing the exploration of areas closer to the target point.

[0079] like Figure 1 As shown in (b), the sampling region of an agent and its distance-based acceptance probability, when When the value is 1, the sampling region of the agent is displayed. It can be observed that as the distance between the sampling point and the target point increases... The increase in probability The rate will decrease, meaning that sampling points farther from the target point are more likely to be rejected. When When = 0, Maximize to ensure that sampling at the target point is not rejected.

[0080] Driven by the aforementioned accept-rejection sampling model, a candidate sampling point is first generated, and its acceptance probability is calculated using the formula described above. This probability is then compared with a random number in the interval [0,1]. Compare them. If the current candidate point satisfies... If the condition is met, then accept the point; otherwise, select the next candidate point and repeat this process until a candidate point is accepted.

[0081] Through the aforementioned accept-rejection sampling, the search area is gradually expanded to a wider space, thereby ensuring efficiency while enhancing the comprehensive exploration of the solution space.

[0082] Specifically, step S4 includes:

[0083] S41. During the expansion phase, the priorities of all agents are fixed. When any two adjacent agents... and (Assuming an intelligent agent) Priority is higher than that of the agent When, if the intelligent agent The next moving goal is And intelligent agents The next moving goal is But intelligent agents The agent is stuck and unable to move. It will temporarily inherit the intelligent agent Prioritize movement and move the position first. Make room for intelligent agents .

[0084] S42. During the expansion phase, a multi-step expansion strategy was employed. Before expansion began, the priorities of all agents were determined based on their distances from their respective target points. During the expansion process, agents farther from the target point would move with higher priority, thus reaching the target point more quickly.

[0085] like Figure 2 (a), (b) and Figure 3 As shown in (a), (b), (c), and (d), the algorithm expands and rewires, introducing the proposed multi-step expansion method for generating collision-free paths from the starting point to the expanded point. The multi-step expansion process in each iteration is described in detail. The main steps of this expansion are as follows: First, random points are generated through dynamic accept-reject sampling. and find the distance The most recent tree node Then, from Departure direction The process involves multi-step expansion and collision resolution based on PIBT. Then, after the multi-step expansion is complete, the path is optimized by rewriting operations. Finally, the optimized expansion is completed, and the expanded nodes are added to the tree for use in the next iteration.

[0086] The detailed techniques for this step are as follows:

[0087] Figure 2 (a) and (b) illustrate the idea of ​​multi-step expansion. In each iteration, multi-step expansion generates sampling points based on adaptive features. For each intelligent agent, starting from the base point Starting from the beginning, the algorithm expands incrementally using multiple step sizes. The step size is chosen as a random integer between [1, L], considering that an excessively large step size may lead to path redundancy, while an excessively small step size may slow down the convergence speed of the overall path.

[0088] In each expansion step, this invention employs the priority inheritance and backtracking principle from PIBT, allowing multiple agents to sequentially choose their next position without interruption due to conflicts. Unlike PIBT's dynamic priority principle, this invention determines agent priorities through a grouping strategy in each expansion. After each multi-step expansion, this invention evaluates whether agents have reached the target position. Agents that have not yet reached the target, along with their adjacent or indirectly adjacent agents, are categorized into a high-priority group, whose priority is significantly increased to ensure smooth movement in subsequent expansion phases. Other agents are assigned to a low-priority group, maintaining their original priority. This strategy also uses breadth-first search (BFS) to explore the neighbors of agents that have not yet reached the target and identify which agents are adjacent or indirectly adjacent. This strategy effectively reduces path quality redundancy that may result from random agent movements.

[0089] Figure 3(a), (b), (c), and (d) illustrate the rewiring operation, which further optimizes path quality. Due to the random search process, path redundancy is not unavoidable in the proposed adRRT algorithm. However, this redundancy is not considered in the original dRRT algorithm. This invention employs a rewiring operation to ensure the generation of high-quality paths, and the rewiring strategy of this invention differs from the strategy in dRRT that targets a small number of agents, traversing every node of the tree. This invention uses A... The algorithm evaluates the total cost and prunes the path to improve quality. Specifically, after multi-step expansion, a series of new configurations are generated. Next, a query is performed to find the n nearest neighbors around the first new configuration. Among these neighbors, the one directly adjacent to the first new configuration and with the lowest cost (via A) is selected. The node selected for heuristic evaluation is chosen as the new parent node, replacing... This reconnection operation eliminates unnecessary branches and improves path quality.

[0090] The complete expansion process is as follows:

[0091]

[0092]

[0093] Specifically, step S5 includes:

[0094] S51. Due to the random search process, path redundancy is unavoidable. This invention uses a rewiring operation to generate high-quality paths, unlike dRRT which is suitable for a small number of agents. This invention uses A The algorithm evaluates the total cost for rerouting to improve path quality. Among the new placement points generated through multi-step expansion, A is utilized. The path is optimized and its quality improved by replacing the original parent node with the nearest neighbor node that has the lowest evaluation cost.

[0095] S52, A series of new configuration points are generated during the expansion process. arrive and for the first new configuration point Perform a nearest neighbor search. Among these neighboring points, select the one accessible via A. The point with the lowest algorithm evaluation cost is taken as The new parent node replaces the original parent node. This reconnection operation eliminates unnecessary branches, optimizes the path structure, and further improves path quality.

[0096] Specifically, step S6 includes:

[0097] S61. Local connected components play a crucial role in this algorithm. During the path search process, they determine whether multiple agents can move to their respective target points without collision from their current positions, thereby reducing redundant expansion operations and improving overall efficiency. Therefore, this invention employs a fixed-step PIBT algorithm as a local connected component, enabling PIBT to be combined with the expansion process, significantly reducing computational complexity.

[0098] S62. This communicator implementation does not require pre-calculation of the path before expansion. Instead, it fully utilizes the distance information already calculated during the expansion process to quickly complete subsequent calculations. When conflicts occur between agents, collisions can be avoided through a priority inheritance inversion mechanism. This mechanism, using path data from the expansion process, can quickly determine whether an agent can reach the target point, thereby effectively reducing unnecessary calculations. It is particularly suitable for multi-agent path planning in large-scale complex environments.

[0099] like Figure 4 (a), (b), and (c) illustrate three typical examples comparing the performance of different sampling methods: deploying 300 agents on the maze-32-32-2 map, 300 agents on the room-32-32-4 map, and 400 agents on the random-32-32-20 map. Each example underwent 20 different experiments. The acceptance-rejection sampling technique introduced in this invention was compared with random sampling and Gaussian sampling. Since Gaussian distributions in the discrete domain require pre-computation steps, their runtime starts from the fifth second, while the method of this invention effectively solved all instances within 100 seconds. Acceptance-rejection sampling consistently outperforms random sampling and Gaussian sampling, with a significantly higher success rate. Random sampling shows a slight improvement, but despite some minor enhancements, Gaussian sampling's performance remains unsatisfactory.

[0100] Figure 5 The paper showcases different algorithms in six small but challenging examples. Compared to traditional metaheuristic methods and graph search-based algorithms, the method of this invention demonstrates stronger stability and adaptability in small-scale but highly complex scenarios. Common algorithms often struggle to maintain good planning performance in complex environments with frequent conflicts, while the method of this invention can successfully handle all such complex problems.

[0101] Figure 6The algorithm proposed in this invention was validated on the industrial simulation platform FlexSim, considering a narrow-aisle warehouse scenario. FlexSim is a Windows-based object-oriented simulation environment, utilizing existing technology; details are not elaborated here. It can be used to establish discrete event flow processes. The scenario includes a 26×10 route map (136 feasible points), a high-density storage area with 30 AGVs deployed. In the simulation, 10 tasks were randomly generated and pre-assigned to each AGV. During the simulation, the algorithm continuously calculated paths for all AGVs, and the AGVs executed tasks immediately upon receiving them, thus achieving coordination and continuous task execution among all AGVs.

[0102] Figure 7 The algorithm of this invention was further tested using Sony's Toio robot in a real-world experiment. Toio is a small cubic mobile robot, approximately 32mm × 32mm × 19.2mm in size. These robots support a Python interface and Bluetooth transmission, and utilize optical sensors on their bottoms to acquire positional information for movement. A snapshot of path planning on a special playpad, involving a T-shape instance, is shown. All robots move on a grid map, with small black cylinders representing obstacles. In this instance, three robots move from the bottom right corner to the bottom left corner, changing their relative positions during the movement. The paths for all robots are calculated using adRRT within approximately 2 milliseconds, and the calculated path information is transmitted to the robots via Bluetooth.

[0103] In summary, this invention proposes a multi-agent pathfinding method based on accept-rejection sampling, specifically designed to solve the complex Multi-Agent Path Planning (MAPF) problem. Compared with existing state-of-the-art algorithms, this method significantly improves the success rate. In simulation environments, the proposed method demonstrates high efficiency and reliability in extremely complex scenarios, with all agents able to coordinate their movements continuously during task execution without any collisions. Furthermore, the method has been tested on a real-world physical robot platform, validating its effectiveness and stability in real-world environments. These validations confirm that the proposed method has significant advantages in path planning performance, particularly excelling in multi-agent coordination and complex environment handling, demonstrating broad application prospects.

[0104] The technical solutions of the present invention are not limited to the specific embodiments described above. Without departing from the scope and spirit of the described embodiments, many modifications and changes will be obvious to those skilled in the art. Any technical modifications made within the spirit and principles of the present invention will fall within the protection scope of the present invention.

Claims

1. A multi-agent path planning method for complex environments, characterized in that, Includes the following steps: S1. Read the predefined workspace map and obtain a set of task instances; This set of task instances includes the number of agents, as well as the starting and destination points for each agent; S2. Based on the discrete fast exploration random tree framework, construct a tree structure for path planning to facilitate the storage, configuration, and fast query of subsequent paths; treat the movement of all agents as a whole, and generate multiple collision-free paths from the starting point to the target point through iterative expansion; S3. Limit the sampling range of all agents to the shortest path between the current position and the target point, set a dynamic threshold, and gradually expand the sampling area as the number of iterations increases. Use the accept-reject sampling method, adjust the rejection rate according to the Gaussian distribution characteristics, enhance the sampling frequency of the area around the target point, and accelerate the convergence of agents from the sampling point to the target point. S4. During the multi-agent path expansion process, a priority inheritance mechanism is adopted to continuously generate effective configurations for multiple agents and assign a priority to each agent. Each agent moves and adjusts according to the priority order to ensure that the path of a high-priority agent is not interfered with by a low-priority agent. S5. After each path expansion, optimize the found path by rewiring multiple agent paths. S6. After the rewiring operation is completed, the priority of each agent is reconfigured according to the distance between each agent and the target point, so that the closer the agent is, the higher the priority. Then, a priority inheritance backtracking algorithm with a fixed number of steps is used as a local connector to try to move all agents to their target positions.

2. The multi-agent path planning method for complex environments according to claim 1, characterized in that: In step S1, the user needs to manually set a threshold. , used to determine the boundary values ​​of the initial sampling region.

3. The multi-agent path planning method for complex environments according to claim 1, characterized in that: Step S2, multi-agent path planning, is achieved by guiding all agents to move together towards the target point through random sampling points. Specifically, it includes: S21. Treat all agents as a high-dimensional whole and avoid conflicts during movement; S22. In a map consisting of points and edges. Among them It represents the set of all points on the map. It is the set of all edges in the map; Assumption Represents intelligent agents exist At any given moment, the movement of all agents must adhere to the following two constraints within any discrete time interval: For any two intelligent agents ,if ,but ; For any two intelligent agents ,if and ,but .

4. The multi-agent path planning method for complex environments according to claim 1, characterized in that: The acceptance-rejection sampling method used in step S3 can efficiently guide all agents to move towards the target point, specifically including: S31. The acceptance-rejection sampling method is used to dynamically adjust the sampling probability, ensuring that the samples are more concentrated in the target area, thereby improving the efficiency of path planning and enabling efficient multi-agent coordination in dynamic or highly constrained environments. S32, Use a threshold By dynamically controlling the sampling area, when no feasible solution can be found within a certain area, the sampling area can be dynamically adjusted to find feasible paths over a larger range. S33. Using the distance from the sampling point to the target point as a parameter, dynamically control the change in the acceptance probability of the sampling point. The closer the selected sampling point is to the target point, the greater the probability that the sampling point will be accepted.

5. The multi-agent path planning method for complex environments according to claim 1, characterized in that: Step S4, the multi-agent path expansion process, employs a multi-step expansion approach, specifically including: S41. During the path expansion phase, the priorities of all agents are fixed; when any two adjacent agents... and exist The positions at the time are respectively and Assuming an intelligent agent Priority is higher than that of the agent At that time, if the intelligent agent The next moving goal is And intelligent agents The next moving goal is But intelligent agents The agent is stuck and unable to move. It will temporarily inherit the intelligent agent Prioritize movement and move the position first. Make room for intelligent agents ; S42. During the path expansion phase, a multi-step expansion strategy is adopted. Before the expansion begins, the priority of all agents is determined based on their distance from their respective target points. During the expansion process, agents that are farther from the target point will move with higher priority, thus enabling them to reach the target point faster.

6. The multi-agent path planning method for complex environments according to claim 1, characterized in that: The multi-agent path rewiring operation in step S5 is used to optimize the obtained path, specifically including: S51. Due to the random search process, path redundancy is unavoidable. Therefore, it is necessary to use rerouting operations to generate high-quality paths. Rerouting is performed by evaluating the total cost to improve path quality. Among the new configuration points generated by multi-step expansion, the neighboring point with the lowest evaluated cost is used to replace the original parent node, thereby optimizing the path and improving path quality. S52. During the expansion process, a series of new configuration points are generated, and the nearest neighbor query is performed on the first new configuration point. The point with the lowest evaluation cost is selected as the new parent node to replace the original parent node. Then, the reconnection operation is performed to eliminate unnecessary branches and further optimize the path structure.

7. The multi-agent path planning method for complex environments according to claim 1, characterized in that: The local connected components proposed in step S6 employ a priority inheritance backtracking algorithm with a fixed number of steps, specifically including: S61. During the path search process, the local communicator determines whether multiple agents can move to their respective target points without collision from their current positions, thereby reducing redundant expansion operations. S62. The implementation of this local connected device does not require pre-calculation of the path before expansion, but makes full use of the distance information already calculated during the expansion process to further and quickly complete subsequent calculations; when conflicts occur between agents, collisions are avoided through the priority inheritance inversion mechanism; by using the path data during expansion, it can be quickly determined whether the agent can reach the target point, thereby reducing unnecessary calculations.

Images