Multi-agent priority obstacle avoidance path planning method for dynamic obstacles

Abstract

The invention discloses a multi-agent priority obstacle avoidance path planning method for a dynamic obstacle, and relates to the technical field of mobile agent path planning, the method comprises the following steps: a global planning stage: optimizing an A * algorithm by introducing a heuristic scaling factor, an obstacle and direction penalty term and a neighborhood adaptive extension mechanism, a path is smoothed by combining broken line reconstruction and a Bezier curve; in the local planning stage, the weight of a DWA evaluation function is dynamically adjusted through fuzzy control, and the real-time obstacle avoidance robustness is improved. Meanwhile, a priority obstacle avoidance strategy and a dynamic look-ahead mechanism are designed, multi-agent conflicts are coordinated, and collective stagnation is avoided. Simulation results show that the method shows better path length, search efficiency and obstacle avoidance performance in static, unknown static and dynamic obstacle environments, effectively improves the multi-agent coordination stability, and is suitable for actual scenes such as warehouse logistics and the like.

Description

Technical Field

[0001] This invention relates to the field of mobile intelligent agent path planning technology, and in particular to a multi-agent priority obstacle avoidance path planning method for dynamic obstacles. Background Technology

[0002] With the continuous development of automation technology, mobile intelligent agents have been widely used in various fields such as warehousing and handling, service interaction, environmental disinfection, and express delivery, becoming core equipment for improving production efficiency and service quality. Path planning, as a key supporting technology in the field of mobile intelligent agents, has the core objective of planning a collision-free optimal or suboptimal path from the starting position to the target position for the intelligent agent, directly affecting the agent's operating efficiency, safety, and collaborative performance.

[0003] Currently, path planning algorithms are mainly divided into two categories: global path planning algorithms and local path planning algorithms. In the field of global path planning, the A* algorithm has become one of the most widely used heuristic search algorithms due to its advantages such as low computational cost, high search efficiency, and near-optimal planned paths. However, the traditional A* algorithm has obvious shortcomings in complex obstacle environments: redundant traversal of nodes leads to large memory consumption, too many path inflection points cause uneven trajectory, and it does not fully consider obstacle distribution information, resulting in paths that tend to be close to obstacle edges, posing a high risk of collision and failing to meet the requirements of multi-agent systems for path safety and controllability. To compensate for these shortcomings, related research has proposed a number of improvement schemes, such as expanding the neighborhood search direction to improve path smoothness, but sacrificing some search efficiency; using a bidirectional search strategy to reduce redundant nodes, but easily generating unreasonable paths that cross obstacles in complex environments; and optimizing the neighborhood expansion method to avoid path crossing obstacles, but without solving the path smoothness problem. None of these solutions have achieved a comprehensive optimization of search efficiency, path quality, and safety.

[0004] In the field of local path planning, the Dynamic Window Approach (DWA) is often used in conjunction with global path planning algorithms due to its good real-time performance and robustness to address the dynamic obstacle avoidance requirements in complex environments. However, the evaluation function weights of traditional DWA algorithms are fixed values, which can easily lead to local optima and path planning failure in complex scenarios such as dense obstacles and changing target distances. Although some studies have improved the effectiveness of dynamic obstacle avoidance by introducing fuzzy logic systems or integrating other intelligent algorithms to optimize the weight adjustment mechanism, they still do not fully consider the collaborative needs in multi-agent interaction scenarios.

[0005] Furthermore, with the widespread application of multi-agent systems, collaborative path planning among agents has become a technical challenge. Most existing path planning methods focus on single-agent scenarios, lacking conflict coordination mechanisms for multi-agent systems. In areas where agents intersect or where there are dense obstacles, problems such as mutual interference, path conflicts, and even collective stagnation easily arise. Meanwhile, while the fusion of global and local path planning algorithms can improve environmental adaptability, it still suffers from insufficient dynamic adjustment capabilities and poor robustness in complex environments where unknown static and dynamic obstacles coexist, making it difficult to meet the comprehensive performance requirements of multi-agent systems in practical applications.

[0006] In summary, the traditional A* algorithm suffers from problems such as high computational cost, uneven path smoothness, and insufficient security. The DWA algorithm is prone to getting trapped in local optima. Existing improved schemes fail to consider multi-dimensional performance indicators and lack effective multi-agent cooperation mechanisms, making them unsuitable for applications in complex and dynamic environments. Therefore, there is an urgent need for a path planning method that integrates the advantages of global planning and local obstacle avoidance, and possesses multi-agent cooperation capabilities, to address the shortcomings of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent priority obstacle avoidance path planning method for dynamic obstacles, which solves the problems of non-smooth paths, easy getting trapped in local optima, and multi-agent conflict stagnation in traditional algorithms, and is applicable to multi-agent path planning in complex dynamic environments.

[0008] To achieve the above objectives, this invention provides a multi-agent priority obstacle avoidance path planning method for dynamic obstacles, the steps of which are as follows: S1. Global Path Planning: An improved A* algorithm is used to generate an initial global path. The node expansion function of the improved A* algorithm introduces a variety of cost adjustment mechanisms, including a heuristic scaling factor to improve the problem of decreased search accuracy near the endpoint of the traditional A* algorithm, an obstacle penalty term to improve path safety, a direction penalty term to enhance the coherence of path direction, and an endpoint neighborhood adaptive expansion mechanism to optimize path convergence stability. S2, Path Smoothing Optimization: The initial global path generated by S1 is sequentially subjected to three-fold polyline reconstruction optimization and second-order Bézier curve smoothing to obtain a globally smooth path that satisfies the agent's kinematic constraints; S3, Local Path Planning: Guided by the globally smooth path optimized in S2, the DWA algorithm, which integrates fuzzy control and dynamic weight adjustment mechanism, adaptively adjusts the evaluation function weights according to dynamic obstacles and environmental conditions, and achieves local real-time obstacle avoidance within the global path framework, ensuring that path deviation is controllable. S4. Multi-agent cooperative control: Based on the global smooth path of S2, a priority obstacle avoidance strategy is designed by combining the relative distance between agents, attitude angle and target position. It is combined with a dynamic look-ahead-driven local target update mechanism to coordinate the conflicts when multiple agents meet along the global path, avoid deviation from the global optimal path and eliminate collective stagnation.

[0009] Preferably, the heuristic scaling factor mentioned in S1 is integrated into the total path cost function, the expression of which is:

[0010] in, The cumulative cost from the starting point to the current node. Let R be the heuristic cost from the current node to the target node, r be the straight-line distance from the starting point to the target node, and W be the heuristic cost from the current node to the target node. The weights are denoted by , where 'a' is the scaling factor, and its value is 1. The expression for the obstacle penalty term is:

[0011] in, This represents the local risk assessment value of the current node in the obstacle grid. This is the weighting coefficient, with a value of 0.5; The expression for the direction penalty term is:

[0012] in, The direction vector from the starting point to the current node. The direction vector from the current node to the expanding node. This is the directional penalty coefficient, with a value of 0.5.

[0013] Preferably, the neighborhood adaptive expansion mechanism described in S1 is as follows: in the early stage of the search, an 8-neighborhood expansion is used, and when the current node is less than 10 grids away from the target point, it switches to a 4-neighborhood expansion.

[0014] Preferably, the cubic polyline reconstruction optimization in S2 includes: First-order polyline optimization: Starting from the starting point, merge subsequent path points that can form a straight line to obtain a primary simplified path; Secondary polyline optimization: Simplify adjacent node pairs with small angles in the primary simplified path to further reduce inflection points; Triple-line optimization: Fitting with a three-segment neighborhood sliding window to replace or delete locally jittery nodes.

[0015] Preferably, the second-order Bézier curve smoothing process in S2 is as follows: the second-order Bézier curve defines the curve shape through n control points, and optimizes the path smoothness by adjusting the spatial position and weight distribution of the control points. The expression for the generated points is:

[0016] in Let be the i-th control point, n=3, and t be a parameter with a value range of . .

[0017] Preferably, the input variables of the fuzzy control dynamic weight adjustment mechanism described in S3 include the distance between the current agent and the nearest obstacle. Distance between the current agent and the target point The distance between the current agent and other agents Each input variable contains 4 fuzzy subsets, and dynamic adjustment is achieved through 64 fuzzy rules; the output variables are the target orientation term weight α, obstacle avoidance distance term weight β, and speed term weight γ of the evaluation function; The formula for normalizing input variables is:

[0018] in It means very close. It indicates that it is very far away.

[0019] Preferably, the rules for the priority obstacle avoidance strategy described in S4 are as follows: (1) Set agent a as the highest obstacle avoidance priority, and its obstacle avoidance behavior takes precedence over agents b and c; when the distance between agent a and the target point is greater than 0.2 meters, obstacle avoidance control is started; if the distance is greater than 2 meters, all obstacles are included in the obstacle avoidance calculation and obstacle avoidance trajectory is generated by DWA; if the distance is ≤2 meters, only the closest obstacle is avoided, and the dynamic window is re-evaluated and the obstacle information is updated. (2) The obstacle avoidance mechanism of agent b is based on the relative distance and angle with agent a: when the distance between agent b and other agents is less than 4 meters and the relative angle between them is less than the set threshold, the obstacle avoidance action stops; in other scenarios, if the distance to the target point is far, all obstacles are considered, and if the distance to the target point is close, only the nearest obstacle is considered. Path planning and obstacle avoidance are completed through DWA. (3) The obstacle avoidance of agent c requires simultaneous judgment of the relative distance and angle with agents a and b: when the distance between agent c and agent a or b is less than 4 meters and the corresponding relative angle is less than the set threshold, the obstacle avoidance action stops; in other scenarios, obstacle avoidance control is performed by DWA based on the principle of considering all obstacles for distant targets and only considering the nearest obstacle for near targets.

[0020] Preferably, the formula for calculating the look-ahead step size of the dynamic look-ahead-driven local target update mechanism described in S4 is:

[0021] Where L represents the look-forward step size, Let be the linear velocity of agent i at the current moment, k be the look-ahead coefficient, k is optimized to a value of 1.2, and l be the basic step size threshold.

[0022] Preferably, the parameters of the intelligent agent are set as follows: maximum speed 10m / s, maximum rotation speed 20rad / s, acceleration 0.4m / s², rotational acceleration 50rad / s², speed resolution 0.02m / s, and rotational speed resolution 1rad / s.

[0023] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) Through multi-dimensional improvements to the A* algorithm, the overall performance of global path planning is significantly enhanced. The introduction of heuristic scaling factors, obstacle penalty terms, and direction penalty terms effectively guides path search towards safe and efficient directions, avoiding the problems of paths being too close to obstacles and having too many redundant nodes in traditional algorithms; the neighborhood adaptive expansion mechanism balances the search range and convergence accuracy, enabling the algorithm to quickly generate feasible paths in complex environments. Combining cubic polyline reconstruction and second-order Bézier curve smoothing further optimizes the continuity and controllability of the path, eliminates the polyline redundancy of the gridded path, makes the trajectory more in line with the kinematic constraints of the intelligent agent, and reduces the control difficulty of subsequent trajectory tracking.

[0024] (2) At the local path planning level, the dynamic weight adjustment mechanism integrating fuzzy control gives the DWA algorithm stronger environmental adaptability. This mechanism can adaptively adjust the weight distribution of the evaluation function according to the real-time environmental state such as obstacle distance, target point distance and agent spacing. It prioritizes obstacle avoidance safety in dense obstacle areas and focuses on improving movement efficiency in open environments. It effectively overcomes the defect of traditional DWA algorithm being prone to getting trapped in local optima and ensures that the agent can achieve real-time and stable obstacle avoidance in unknown static and dynamic obstacle scenarios.

[0025] (3) Regarding multi-agent collaboration, the combination of priority obstacle avoidance strategy and dynamic look-ahead mechanism successfully solves the conflict coordination problem when multiple agents meet. By clarifying the priority levels and obstacle avoidance judgment criteria of different agents, orderly avoidance among agents is achieved, avoiding path planning failure caused by mutual interference; the dynamic look-ahead mechanism effectively eliminates the phenomenon of multiple agents simultaneously stagnating in areas with dense obstacles by adaptively adjusting local target points, ensuring the continuity and collaborative stability of the overall system movement. Overall, this invention achieves the organic unity of global path optimality, local obstacle avoidance robustness and multi-agent collaboration, greatly improving the engineering application capability of multi-agent systems in complex dynamic environments.

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating an embodiment of a multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to the present invention. Figure 2 This is a diagram illustrating the motion state transition process of a mobile intelligent agent according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the algorithm fusion process according to an embodiment of the present invention; Figure 4 This is a comparison chart of the traditional A* algorithm, the improved A* algorithm, and the path planning after curve smoothing in this embodiment of the invention. Figure 5 These are partial turning diagrams of three path planning methods according to embodiments of the present invention; Figure 6 The following diagrams are comparisons of the next path planning algorithms for different sizes of raster maps in this embodiment of the invention; where (a) is a comparison diagram of a 20×20 raster map, (b) is a comparison diagram of a 30×30 raster map, and (c) is a comparison diagram of a 60×60 raster map. Figure 7 This diagram illustrates path planning and local motion under a multi-agent priority obstacle avoidance mechanism according to an embodiment of the present invention; where (a) is the global path generation result of the multi-agent system, (b) is the local motion state under priority obstacle avoidance, (c) is the implementation process of the priority obstacle avoidance mechanism for agent a, and (d) is the implementation process of the priority obstacle avoidance mechanism for agents b and c. Figure 8 The following is a graph showing the kinematic parameter variation curves under the multi-agent priority obstacle avoidance mechanism in this embodiment of the invention; where (a) is the curve of agent attitude angle variation, (b) is the curve of agent linear velocity variation, and (c) is the curve of agent angular velocity variation. Figure 9 This is a diagram of the path planning and obstacle avoidance process of a multi-agent system in an environment with unknown static obstacles, according to an embodiment of the present invention; where (a) is the global path generation result of the multi-agent system, and (b) is the local motion state when there are unknown static obstacles. Figure 10 The diagram shows the kinematic parameter variation curves of a multi-agent system in an unknown static obstacle environment according to an embodiment of the present invention; where (a) is the curve of the agent's attitude angle variation, (b) is the curve of the agent's linear velocity variation, and (c) is the curve of the agent's angular velocity variation. Figure 11 This is a schematic diagram of the path planning and obstacle avoidance process of a multi-agent system in an environment with unknown dynamic obstacles, according to an embodiment of the present invention; where (a) is the global path generation result of the multi-agent system, and (b) is the local motion state when there are unknown dynamic obstacles. Figure 12 The following are the kinematic parameter variation curves of a multi-agent system in an unknown dynamic obstacle environment according to an embodiment of the present invention; wherein (a) is the curve of the agent's attitude angle variation, (b) is the curve of the agent's linear velocity variation, and (c) is the curve of the agent's angular velocity variation. Figure 13 The diagram shows a comparison of obstacle avoidance paths for multiple agents before and after the introduction of the dynamic look-ahead mechanism in this embodiment of the invention; where (a) is the path generation result without the introduction of the dynamic look-ahead mechanism, showing the phenomenon of collective stagnation of multiple agents, and (b) is the path generation result with the introduction of the dynamic look-ahead mechanism. Figure 14 The following are the kinematic parameter variation curves of multiple agents under the dynamic look-ahead mechanism of this invention: (a) is the agent's attitude angle variation curve, (b) is the agent's linear velocity variation curve, and (c) is the agent's angular velocity variation curve. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Example A multi-agent priority obstacle avoidance path planning method for dynamic obstacles, such as Figure 1 As shown, the steps are as follows: I. Specific Implementation Steps S1: Global Path Planning. The A* algorithm, a classic heuristic search method, is widely used in path planning tasks under static environments. However, traditional A* algorithms are prone to problems such as path backtracking, approaching obstacles, path discontinuity, and search redundancy in complex obstacle scenarios, making it difficult to meet the comprehensive requirements of multi-agent systems for path smoothness, safety, and real-time performance. Therefore, this paper designs an improved A* algorithm, retaining its global optimality characteristics, incorporating obstacle penalty, direction penalty, heuristic scaling, and path polyline optimization to enhance its adaptability in real-world environments. Specifically, it includes: 1. Initialize the grid map by inputting the starting and ending coordinates of the agent and the distribution information of obstacles, creating an obstacle grid map and marking impassable areas; 2. Construct an improved node expansion function for the A* algorithm. Based on the traditional A* neighborhood expansion strategy, the node expansion function proposed in this application introduces a variety of cost adjustment mechanisms, mainly including heuristic scaling factor, obstacle penalty term, direction penalty term, and endpoint neighborhood adaptive switching mechanism, which guides the path search to a safe, smooth and fast-converging solution from multiple dimensions.

[0032] (1) To improve the problem of decreased search accuracy near the endpoint in the traditional A* algorithm, a heuristic scaling factor is added to the total path cost function. The expression for the total path cost function is:

[0033] in, The cumulative cost from the starting point to the current node. Let R be the heuristic cost from the current node to the target node, r be the straight-line distance from the starting point to the target node, and W be the heuristic cost from the current node to the target node. The weight is used to reduce Impact on cost calculation; 'a' is the scaling factor, with a value of 1; this factor can maintain global exploratory nature in the early stages of the search, while enhancing target attractiveness in the later stages, thereby improving search efficiency and endpoint accuracy.

[0034] (2) The traditional A* algorithm does not fully consider obstacle distribution information during cost evaluation, which may result in the generated path being close to the obstacle edge, posing a high risk of collision. To improve the path's environmental adaptability, a penalty term based on obstacle density is introduced:

[0035] in, This represents the local risk assessment value of the current node in the obstacle grid. This is a weighting coefficient with a value of 0.5; this item can effectively guide the search to avoid areas with high-density obstacles, enhancing the feasibility and safety of the path.

[0036] (3) The directional consistency of the path has a direct impact on subsequent trajectory tracking and dynamic obstacle avoidance performance. To suppress unnecessary sharp turns and backswings in the path, a directional penalty term is introduced:

[0037] in, The direction vector from the starting point to the current node. The direction vector from the current node to the expanding node. This is the direction penalty coefficient, with a value of 0.5. This factor limits the deviation angle between the extension direction and the main path direction, thereby making the path more coherent and contributing to the smoothing of subsequent paths.

[0038] (4) Perform adaptive neighborhood expansion: In the early stage of the search, 8-neighbor expansion is used, including 8 directions: up, down, left, right and diagonal. When the distance between the current node and the target point is less than 10 grids, switch to 4-neighbor expansion, and only allow expansion in the up, down and left and right directions; avoid unnecessary offset near the endpoint due to diagonal expansion, and improve the path convergence stability.

[0039] 3. Iterate through the search until the target point is reached, generating an initial global path.

[0040] S2: Path Smoothing Optimization 1. Cubic Polyline Reconstruction: The improved A* algorithm can generate a better path, but due to the inherent discreteness of the raster map, the final path still suffers from problems such as dense polylines and frequent corners, which are detrimental to subsequent trajectory tracking and control. Therefore, this paper designs a cubic polyline reconstruction method based on geometric features after path generation to improve the overall continuity and controllability of the path. This method consists of the following three steps: (1) First-order polyline optimization: Starting from the starting point, check step by step whether the subsequent path points can form a straight line with the current starting point. If it is feasible, merge them into a polyline segment to obtain the primary simplified path.

[0041] (2) Secondary polyline optimization: Further examine the path after the first polyline optimization to find if there are adjacent node pairs with small angles that can be simplified into a polyline by connecting them with straight lines. This process can further reduce unnecessary vertices in the path and make the path more concise.

[0042] (3) Triple-line optimization: Three-segment neighborhood sliding window fitting is performed on the remaining corner points to evaluate the angle change and path smoothness index, and local jitter nodes are replaced or deleted to achieve simplified reconstruction.

[0043] 2. Second-order Bézier curve smoothing: In the A* algorithm, after finding the optimal path, there are often non-smooth turning points in the path. Curve optimization can be introduced to smooth the path. The shape of the Bézier curve can be changed by adjusting the position and weight of the control points, making the path smoother.

[0044] The vector representation of a Bézier curve defines its shape using a series of control points. This vector form of a Bézier curve uses a linear combination of control points. Assuming n control points are needed for definition, the generating points on the Bézier curve are... This can be expressed as:

[0045] in Let be the i-th control point, and t be a parameter with a value range of . .

[0046] The smoothness of a Bézier curve is achieved by the combination of control points and the parameter t in its vector representation. By adjusting the position and weight of the control points, the shape of the curve can be changed. When the distance between the control points and the weight distribution are reasonable, the Bézier curve will exhibit smooth characteristics.

[0047] The three-dimensional polyline reconstruction prioritizes addressing the redundancy of inflection points in the rasterized path (reducing computational load), while the second-order Bézier curve further optimizes the continuity of remaining corners (satisfying kinematic constraints). Polyline reconstruction alone cannot eliminate sharp corners, and Bézier curves alone can cause the path to deviate excessively from the global optimum. The combination of the two achieves the dual goals of simplification and smoothing.

[0048] S3: Local Path Planning The Dynamic Window Approach (DWA) is a method based on the agent's motion model to calculate an effective velocity search space under certain constraints. Based on the agent's parameters and kinematic model, the maximum linear velocity and maximum angular velocity achievable during motion can be obtained, thus providing a velocity space for the agent's movement. Multiple sets of velocities are sampled within this velocity search space, and a trajectory is simulated over a period of time. All sampled velocities are subject to safety constraints to ensure the agent's normal movement. Then, the optimal velocity command is selected based on an evaluation function, driving the mobile agent towards the optimal trajectory. During the movement, DWA continuously calculates and updates, ultimately reaching the target with the optimal plan.

[0049] 1. Establish a motion model The dynamic window method is a classic local path planning algorithm based on a kinematic model. Its core logic involves sampling the linear and angular velocities of the agent and simulating the agent's trajectory within each sampling period. The complete trajectory of the agent is composed of multiple smaller trajectories corresponding to multiple sampling periods. Due to the short sampling period, it can be assumed that the agent moves in a straight line within a single sampling period. The change in pose of the agent within this period can be expressed as:

[0050] The state of a mobile agent at the next time step can be represented as:

[0051] Where x and y represent displacements in two directions of the coordinate system, respectively; Let be the heading angle of the mobile agent at time t; , These represent the displacements of the mobile agent in the x and y directions, respectively. Indicates the sampling period; This represents the angular velocity of the mobile agent. The motion state transition process of the mobile agent is as follows: Figure 2 As shown.

[0052] 2. Velocity sampling The achievable speed of a mobile intelligent agent is limited by its own hardware performance and the environment, including limitations on speed, acceleration, angular velocity, angular acceleration, and braking distance. , and The intersection of these values ​​serves as the effective speed range.

[0053] (1) The agent's own maximum and minimum velocity and angular velocity constraints are expressed as follows:

[0054] (2) Motor acceleration and deceleration capability constraint, the expression is:

[0055] in , These represent the agent's current linear velocity and angular velocity, respectively. , These represent the maximum linear acceleration and the maximum angular acceleration, respectively. , These represent the maximum linear deceleration and the maximum angular deceleration, respectively.

[0056] (3) To ensure that the mobile intelligent agent can safely avoid obstacles, the agent needs to reduce its speed to 0 before colliding with the obstacle. Therefore, the sampling speed constraint under the maximum deceleration condition is expressed as:

[0057] in This represents the shortest distance between the trajectory and the obstacle.

[0058] 3. Optimization of the dynamic window method In DWA (Dynamic Impact Assessment), to evaluate the performance of multiple velocity trajectories, an evaluation function needs to be constructed to measure the trajectory's performance in terms of target guidance, obstacle avoidance safety, and operational efficiency. A typical evaluation function is shown below:

[0059] in: : Indicates the angular deviation between the end of the trajectory and the target point, reflecting the target guidance; : Indicates the minimum distance between the trajectory and the nearest obstacle, reflecting obstacle avoidance safety; : Indicates the linear velocity used in the trajectory, reflecting the movement efficiency; , , : This is the weight coefficient for the corresponding evaluation item, which is usually set to a fixed constant in the initial stage of the algorithm.

[0060] Traditional evaluation functions have the following problems: when obstacles are dense or nearby, the fixed... The value may not be sufficient to prompt the agent to slow down or evade in time; when approaching the target area, a fixed value... The evaluation value may not enhance the focusing behavior on the target; in multi-agent cooperation, it fails to consider the dynamic distance with other agents and lacks priority obstacle avoidance adjustment capabilities. To overcome these problems, a dynamic weight adjustment mechanism based on a fuzzy controller is introduced. This fuzzy system takes the environmental state as input and adaptively outputs three weight parameters to achieve real-time adjustment of the evaluation function. The controller structure is shown in the table below: Table 1. Structure of the fuzzy controller

[0061] Using 64 fuzzy rules (each input contains four fuzzy subsets), the controller can dynamically adjust the three weight coefficients based on the current local environment state. The controller's operation and adjustment logic is as follows: (1) Operating mechanism: Whenever the DWA control cycle executes fuzzy inference once, the output is ( , , The first three weight parameters of the evaluation function are replaced in real time.

[0062] (2) Input-output mapping principle: To ensure uniform processing, input variables are normalized to... Interval:

[0063] in It means very close. It indicates that it is very far away.

[0064] When the obstacle approaches ( Small, obstacle-heavy areas): , ,

[0065] Significantly increase obstacle avoidance weight, suppress speed weight, and moderately reduce goal orientation weight to prioritize safety.

[0066] When the distance to the obstacle is large ( Large, open area): , ,

[0067] At this point, obstacle avoidance weights are kept at a low level, while speed and target orientation weights are increased to ensure efficient robot movement.

[0068] When the target approaches ( Small):

[0069] The goal-oriented weight is significantly increased; if obstacles approach simultaneously, a high obstacle avoidance weight is maintained to achieve a balance between convergence and safety.

[0070] When the target is far away ( big):

[0071] The goal-oriented weight is kept at a high level to maintain the agent's continuous approach to the goal; the trade-off between speed and obstacle avoidance is dynamically determined by the distance to the obstacle.

[0072] When near other intelligent agents ( Small): ,

[0073] The speed weight is reduced and the obstacle avoidance weight is increased, thereby including nearby intelligent agents in the obstacle avoidance range and ensuring the safety of group operation.

[0074] When the distance to other agents is large ( big):

[0075] Speed ​​weight is kept at a high level to ensure overall efficiency; the weights of target orientation and obstacle avoidance are dynamically determined by the target distance and obstacle distance.

[0076] (3) Reasoning method: Gaussian membership function is used for input fuzzification, rule calculation is based on min-max logic, the maximum value is taken after aggregation, and the maximum membership mean method is used for defuzzification.

[0077] This dynamic weight adjustment strategy based on environmental state significantly improves the adaptability and safety of the DWA algorithm in complex, variable, and multi-agent scenarios, effectively avoiding the problem of poor performance of traditional fixed-coefficient methods in specific environments.

[0078] S4, Multi-agent Cooperative Control 1. In multi-agent systems, mutual interference between agents in complex dynamic environments can lead to a significant degrade in system performance. The priority obstacle avoidance method in this application makes a comprehensive judgment based on the relative distance, attitude angle, and target position between agents to achieve obstacle avoidance coordination among agents of different priorities, thereby avoiding path planning failure due to mutual interference when multiple agents are performing tasks together.

[0079] The priority obstacle avoidance mechanism is designed based on the following principles: First, agent A has the highest obstacle avoidance priority, and its obstacle avoidance behavior is executed first among all agents. Second, agents B and C determine whether obstacle avoidance is needed based on their distance from other agents and their relative angle, respectively. Specifically, when the distance between an agent and other agents exceeds a set threshold, its obstacle avoidance control mechanism is activated, and its obstacle avoidance strategy is adjusted according to changes in distance and angle. If the agent is far from the target point, it will consider all obstacles and perform path planning and obstacle avoidance using DWA (Distance-Based Avoidance). When the agent is close to the target point, it will only consider the nearest obstacle to ensure the immediacy and accuracy of the obstacle avoidance action.

[0080] For agent A, the core of its obstacle avoidance control lies in determining the relative distance to agents B and C. In the current mechanism, if the distance between agent A and the target point is greater than 0.2 meters, the obstacle avoidance process begins. In this case, if the distance to the target point is greater than 2 meters, agent A will include all obstacles (including those of other agents) in its obstacle avoidance calculation, thereby generating an obstacle avoidance trajectory. If the distance to the target point is relatively short, agent A only considers the closest obstacle to avoid. At this point, the system will re-evaluate the agent's dynamic window and execute obstacle avoidance actions based on the updated obstacle information.

[0081] The obstacle avoidance control mechanism of agent b is similar to that of agent a, but it also considers the relative angle with agent a. If the distance between agent b and other agents (such as agent a) is less than 4 meters and their relative angle is small (i.e., the angle between them is less than a certain threshold), agent b will stop its obstacle avoidance action to avoid conflict with other agents. In addition, when the relative distance and angle between agent b and other agents meet certain conditions, agent b will calculate its path and use a dynamic window method for obstacle avoidance control.

[0082] Agent C's obstacle avoidance strategy is similar to that of agents A and B, but its judgment criteria are more complex. During obstacle avoidance, agent C simultaneously considers the distance and angle information between itself and agents A and B to ensure effective avoidance when approaching other agents. If the distance to other agents is less than 4 meters and the relative angle is small, agent C will stop obstacle avoidance. Otherwise, agent C will perform dynamic window path planning based on the current obstacle information.

[0083] 2. Dynamic Forward-Looking Mechanism In complex, obstacle-dense environments, when multiple agents are simultaneously constrained by priority limitations in a narrow area, a collective deadlock phenomenon can easily occur, causing the overall system motion to be interrupted. To address this issue, this application introduces a dynamic lookahead-driven local target adaptive update strategy based on the priority obstacle avoidance mechanism.

[0084] (1) Principle of dynamic look-ahead mechanism This mechanism dynamically determines the look-ahead step size based on the agent's current velocity and pose, calculated using the following formula:

[0085] Where L represents the look-ahead step size (i.e., the number of nodes searched forward along the planned path). Let be the linear velocity of agent i at the current moment, and k be the look-ahead coefficient used to adjust the path look-ahead depth and sensitivity. The system updates the local target point in real time within each control cycle using this formula, enabling the agent to adaptively adjust its look-ahead range according to the velocity: it has a greater path perception capability when the velocity is higher, while maintaining strong local reaction accuracy when the velocity is lower.

[0086] During execution, the algorithm sequentially scans local nodes along the currently planned path, automatically searching for safe local target points located in non-obstacle areas. When a node on the path ahead is detected to be within an obstacle area, the algorithm continuously increments the step size L until a node in a safe area is found and set as the new target point, thus achieving dynamic path transitions and local replanning.

[0087] (2) Parameter selection and optimization During parameter tuning, it was found that when the look-ahead coefficient k=2, the look-ahead step size was too large. This led to problems such as local target skipping, trajectory prediction failure, and synchronized yielding among multiple agents in obstacle-dense areas, ultimately causing the system to stagnate. To address this, multiple simulations were conducted to verify the parameters. The look-ahead coefficient was then optimized to k=1.2. At this point, the look-ahead range was moderate, local target updates were smooth, and the agent achieved a better balance between local obstacle avoidance and global path tracking. The results show that appropriately reducing the look-ahead step size can effectively reduce the spatial overlap of local targets among multiple agents, lower the risk of synchronized stagnation in intersection areas, and thus achieve stable system operation and path continuity optimization.

[0088] The dynamic look-ahead mechanism does not change the core logic of the original priority rules: low-priority agents will still stop and yield when they should. The difference is that the introduction of the mechanism avoids all agents simultaneously entering a waiting state in the same area. After the improvement, only low-priority agents stop briefly when necessary, while other agents can continue to move forward at a low speed and adjust their paths, thus ensuring the continuous movement and global stability of the overall system.

[0089] 3. This application constructs a path planning fusion framework for multi-agent collaboration, organically combining the improved global optimal pathfinding capability of the A* algorithm with the real-time local obstacle avoidance performance of the dynamic window method. By introducing a dynamic weight adjustment mechanism based on fuzzy control, the evaluation function is adaptively optimized under different environmental states, effectively improving the robustness and flexibility of the algorithm in complex dynamic scenarios. Simultaneously, a priority obstacle avoidance strategy is designed to coordinate interaction conflicts among multiple agents, and a global-local collaborative update mechanism is constructed to cope with unknown obstacles and sudden environmental changes. The fusion process is as follows: Figure 3 As shown.

[0090] II. Simulation Verification To verify the effectiveness of the improved A* algorithm, simulation tests were conducted in multiple environments. The software used for the simulation tests was Matlab R2024a. Table 2 shows the set agent parameters, including maximum speed, maximum rotational speed, acceleration, rotational acceleration, velocity resolution, and rotational speed resolution.

[0091] Table 2 Agent Parameters

[0092] 1. To verify the effectiveness of path smoothing in optimizing path length and improving turning smoothness, this paper compares three path planning methods under the same start, end, and obstacle configurations: the traditional A* algorithm, the improved A* algorithm, and a method that incorporates Bézier curve smoothing based on the improved A* algorithm. Experimental results are as follows: Figure 4 As shown, see the partial turning situation. Figure 5 The corresponding path length data is listed in Table 3.

[0093] Table 3 Comparison of the three conditions

[0094] From an overall path perspective, all three methods successfully achieve obstacle avoidance navigation from the starting point to the destination, but their geometric features differ significantly. The path generated by the traditional A* algorithm exhibits obvious grid-like characteristics, with numerous polylines, especially redundant segments at corners, resulting in a longer path length and the risk of diagonally traversing obstacles. The improved A* algorithm introduces directional constraints and adaptive mechanisms during neighborhood expansion, making the path fit the boundary better when bypassing obstacles and reducing unnecessary turns. Furthermore, by employing second-order Bézier curve smoothing, the path corners are effectively rounded, resulting in a more natural and continuous overall trajectory.

[0095] The quantitative results show that the traditional A* algorithm has a path length of 46.28 m, which is reduced to 44.93 m by the improved A* algorithm, and further reduced to 44.88 m after smoothing. This demonstrates that the improved A* algorithm can optimize path length while maintaining feasibility, while curve smoothing further improves the continuity and controllability of the path.

[0096] In summary, path smoothing not only reduces redundant inflection points but also significantly improves the geometric characteristics at turns, making the trajectory more consistent with the kinematic constraints of the agent. This verifies the effectiveness of the improved combination of A* and Bézier curves in enhancing path quality.

[0097] 2. Comparative Analysis of Traditional A* Algorithm and Improved A* Algorithm To verify the effectiveness of the proposed improved algorithm, this paper compares the traditional A* algorithm, the extended 16-neighborhood algorithm, the bidirectional search algorithm, and the improved bidirectional A* algorithm under different scale grid maps. The experimental environment was the same, with the starting and ending points remaining consistent. The experimental results are as follows: Figure 6 As shown, (a) is a comparison of 20*20 maps, (b) is a comparison of 30*30 maps, and (c) is a comparison of 60*60 maps; relevant statistical data are shown in Table 4.

[0098] Table 4 Comparison of data under different maps

[0099] In terms of pathfinding efficiency, the improved A* algorithm achieved the best performance in all three map types. Taking a 60×60 map as an example, the traditional A* pathfinding time is 17.14 s, the extended 16-neighbor method is 11.77 s, the bidirectional search is 6.32 s, the 5-neighbor method is 3.16 s, the improved bidirectional method is 2.74 s, while the improved A* algorithm only takes 2.43 s. Compared to the traditional A* algorithm, the time is reduced by 85.8%, and compared to the extended 16-neighbor method and the bidirectional search method, the time is reduced by 79.3% and 61.5% respectively, even outperforming the 5-neighbor method and the improved bidirectional method. This shows that the improved algorithm can still maintain high search efficiency on large-scale maps.

[0100] In terms of the number of nodes expanded, the improved algorithm also demonstrates a significant advantage. In a 30×30 map, the traditional A* algorithm expands to 40 nodes, the 16-neighbor expansion to 32, and the bidirectional search to 41, while the improved algorithm, the 5-neighbor expansion, and the improved bidirectional search algorithm only expand to 21, representing a reduction of up to 47.5% compared to the traditional A* algorithm. In a 60×60 map, the improved algorithm expands to 35 nodes, a reduction of 58.3% compared to the traditional A* algorithm, and is also significantly better than the 16-neighbor expansion and bidirectional search, demonstrating its effective suppression of redundant expansion in high-dimensional search spaces.

[0101] Regarding path quality, the total number of steps and the number of vertices are not significantly different among the algorithms. However, the improved algorithm avoids path redundancy caused by the excessive expansion range of the 16-neighborhood algorithm while maintaining path rationality. It also effectively overcomes the defect of bidirectional search in generating unreasonable paths in complex environments. The 5-neighborhood and improved bidirectional methods can achieve better step counts in some cases, but the overall number of vertices is basically the same as that of the improved A*, and it does not show stronger smoothness.

[0102] In summary, the improved A* algorithm, by introducing direction penalties, obstacle penalties, and a dynamic neighborhood mechanism, significantly improves pathfinding efficiency and reduces the number of expanded nodes while ensuring path rationality. Under comprehensive comparison of multiple metrics, the proposed algorithm not only outperforms the traditional A* and extended 16-neighborhood algorithms, but also surpasses bidirectional search, 5-neighborhood, and improved bidirectional methods in most scenarios, verifying its practicality and robustness in path planning tasks in complex environments.

[0103] 3. Multi-agent obstacle avoidance performance test (1) Priority risk avoidance To verify the effectiveness of the proposed priority obstacle avoidance strategy in multi-agent cooperation, a three-agent simulation experiment was conducted in an environment with known global obstacles. In the experiment, agent a's starting and ending points were (2.5, 16.5) and (29.5, 16.5), respectively; agent b's starting and ending points were (29.5, 16.5) and (2.5, 16.5), respectively; and agent c's starting and ending points were (16.5, 2.5) and (16.5, 29.5), respectively. The global path planning results are as follows: Figure 7 As shown, the local motion state and the changes in control parameters are as follows: Figure 8 As shown.

[0104] The experimental results show that agent c began to decelerate after detecting a potential collision risk with agent b at the 180th control node, and came to a complete stop at the 202nd node. Subsequently, agent b detected a conflict risk with agent a at the 178th control node and stopped moving at the 213th node. At this point, agent a, with the highest priority, treated the stationary b and c as static obstacles and replanned its trajectory according to DWA to complete obstacle avoidance. As the conflict between a and b was resolved, agent b resumed movement at the 351st control node, while agent c, with the lowest priority, did not resume movement until the 389th node.

[0105] like Figure 8 As shown, agent a exhibits relatively large fluctuations in its attitude angle, which is a result of performing additional obstacle avoidance maneuvers, but its overall velocity change remains stable. In contrast, agents b and c remain stationary for extended periods during obstacle avoidance, with their velocity curves showing a distinct zero-velocity range. The total experimental time was 150.01 s.

[0106] (2) Environments with unknown static obstacles To verify the adaptability of the proposed algorithm in scenarios with unknown static obstacles, randomly distributed unknown obstacles were added to the global planning algorithm, and experiments were conducted in a three-agent cooperative environment. In the experiment, the starting and target points of agents a, b, and c were set as (3.5, 3.5) → (29.5, 29.5), (3.5, 29.5) → (29.5, 3.5), and (16.5, 2.5) → (16.5, 29.5), respectively. The global path planning results are shown below. Figure 9 As shown, the changes in speed and attitude during the motion are as follows: Figure 10 As shown.

[0107] Experimental results show that the multi-agent system can effectively avoid randomly generated unknown static obstacles while moving along a global path. Taking agent c as an example, it decelerates after detecting a potential collision risk with a high-priority agent at the 293rd control node, and its speed drops to 0 at the 348th node. After the high-priority agent leaves the danger zone, c resumes movement at the 511th node, ultimately successfully bypassing the obstacle and reaching the target point.

[0108] From the perspective of motion parameters, the three agents showed significant differences in their speed and attitude curves. Agents a and b exhibited relatively smooth motion trajectories, while agent c, with the lowest priority, had the longest stopping time, and its speed curve showed a prolonged period of zero speed. Final statistical data showed that the travel distances of a, b, and c were 38.19 m, 38.23 m, and 28.60 m, respectively, with a total experimental time of 240.27 s.

[0109] In summary, the improved priority obstacle avoidance mechanism can maintain the safety and cooperation of multi-agent systems in complex environments with unknown static obstacles, verifying the robustness and adaptability of the proposed method in scenarios with incomplete information.

[0110] (3) Environments with unknown dynamic obstacles Building upon experiments with unknown static obstacles, multiple unknown dynamic obstacles were further introduced to verify the adaptability of the proposed algorithm in complex dynamic environments. In the experiments, the starting and ending points of agent a were (3.5, 3.5) and (29.5, 29.5), respectively; those of agent b were (3.5, 29.5) and (29.5, 3.5), respectively; and those of agent c were (16.5, 2.5) and (16.5, 29.5), respectively. The experimental scenario is as follows: Figure 11 As shown, yellow boxes represent unknown dynamic obstacles, and red boxes represent unknown static obstacles. The moving speed of the dynamic obstacles is set to 0.46 m / s.

[0111] Experimental results show that agent a detected a dynamic obstacle at the 46th control node and began to decelerate, successfully avoiding it at the 104th node and returning to the global path at the 132nd node; agent b detected a dynamic obstacle at the 41st node, completed obstacle avoidance at the 70th node, and then returned to the global path at the 114th node; agent c detected a dynamic obstacle at the 43rd node, completed obstacle avoidance at the 83rd node, and returned to the global path at the 133rd node. Subsequently, agents b and c detected potential collision risks with agent a at the 271st and 260th control nodes, respectively, and decelerated and stopped according to a priority strategy until the conflict was resolved.

[0112] like Figure 12As shown, during obstacle avoidance, all three agents exhibited significant attitude angle fluctuations, especially with increased angle deviations under dynamic obstacle interference. However, the overall velocity curves remained stable without oscillations or irreversible stalls. Final statistical results indicate that the actual movement distances of agents a, b, and c were 37.32 m, 37.03 m, and 27.31 m, respectively, with a total experimental time of 177.93 s.

[0113] In summary, the proposed algorithm can ensure the path feasibility and safety of multi-agent systems in complex environments containing unknown static and dynamic obstacles, and maintains high robustness and cooperation under dynamic disturbances, further verifying the effectiveness of the proposed priority obstacle avoidance mechanism.

[0114] (4) Impact of dynamic look-ahead mechanism on obstacle avoidance of multi-agent systems and simulation analysis To further verify the effectiveness of the proposed dynamic look-ahead-driven local target update mechanism in multi-agent obstacle avoidance, comparative experiments were conducted in areas where multiple agents intersect and where obstacles are dense. The experimental scenarios are as follows: Figure 13 As shown, (a) is the simulation result without the introduction of the dynamic look-ahead mechanism, and (b) is the operation result after the mechanism is added only to agent b.

[0115] Depend on Figure 13 As shown in (a), without a dynamic look-ahead mechanism, when three agents simultaneously approach the central obstacle area, although agent a has the highest priority and can continue moving forward, agents b and c are both constrained by their priority during obstacle avoidance. When b and c simultaneously determine that there is a potential conflict ahead and execute stop-and-go, a "collective deadlock" phenomenon occurs, where multiple agents simultaneously stop and wait, resulting in overall motion being hindered. At this point, the mutual waiting and path restriction among low-priority agents create a local deadlock.

[0116] It should be noted that the dynamic look-ahead mechanism only applies to agent b in this experiment. This is because agent a has the highest priority and its movement is unconstrained; while agent c is at the end of the path and is mainly affected by b's behavior. In the multi-agent intersection area, agent b is located in the middle of the system and often faces both the priority pressure from a and the path constraints from c, making it the node most likely to trigger a bidirectional yielding deadlock. Therefore, applying the dynamic look-ahead mechanism to agent b allows it to proactively adjust its local target point and maintain a low-speed dynamic movement before detecting a potential conflict, thereby prompting c to make speed corrections and preventing all three agents from stopping simultaneously. This single-point optimization method not only simplifies the system control logic but also reflects the algorithm's characteristics of local driving and global stability improvement.

[0117] like Figure 13As shown in (b), after introducing the dynamic look-ahead mechanism, the system adaptively adjusts the look-ahead step size according to the real-time speed of agent b, and dynamically jumps to safe nodes in the local path to achieve continuous updating of local target points. When a node on the path ahead is detected to be in an obstacle area or a potential conflict zone, agent b can actively jump to the subsequent node as a new target point, thereby achieving dynamic reconstruction and continuous advancement of the path. Under this mechanism, agent a maintains the highest priority and moves smoothly, b stops briefly or adjusts its posture at low speed when necessary, and c dynamically corrects its own movement according to the state changes of b, thus effectively avoiding synchronous stillness of multiple agents.

[0118] like Figure 14 The figures show the corresponding attitude angle, linear velocity, and angular velocity variation curves. It can be seen that the velocity and angle changes of agent a are stable, while agent b experiences a brief deceleration and stops during the obstacle avoidance phase, and agent c makes coordinated adjustments according to the state changes of agent b. The linear velocity curves do not show any simultaneous drops to zero, and the angular velocity changes smoothly without oscillations or repeated turning, indicating that the system achieves a significant improvement in path continuity and control stability under the dynamic look-ahead mechanism.

[0119] In summary, although the dynamic look-ahead mechanism is only applied to intermediate agent b, the system can still achieve global motion coordination through its local proactive adjustments, successfully eliminating the collective stagnation phenomenon. This mechanism effectively improves the continuity and cooperation of the system in obstacle-dense scenarios, further verifying the superior performance of the proposed fusion improved A* and DWA algorithms in multi-agent dynamic environments.

[0120] III. Conclusion This application addresses the problems of traditional A* algorithm in complex environments, such as high computational cost, uneven path smoothness, difficulty in adapting to dynamic environments, and lack of multi-agent cooperation mechanism. It proposes a multi-agent path planning method that integrates improved A* and dynamic window method (DWA).

[0121] In the global planning phase, the algorithm effectively improves search efficiency and reduces the risk of paths approaching obstacles by introducing heuristic scaling factors, obstacle penalty terms, and direction penalty terms, combined with a neighborhood adaptive expansion strategy. Simultaneously, the use of piecewise linear optimization and second-order Bézier curve smoothing significantly improves path continuity and controllability. In the local planning phase, combining the DWA algorithm with a fuzzy control-based weight adaptive mechanism enables dynamic adjustment of the evaluation function weights under different environmental conditions, thereby enhancing the flexibility and robustness of local obstacle avoidance. Furthermore, the designed multi-agent priority obstacle avoidance strategy, combined with dynamic look-ahead, can coordinate conflicts in intersection areas, ensuring the overall system's coordination and safety.

[0122] Simulation results demonstrate that the proposed method exhibits strong comprehensive performance across various scenarios: it achieves optimal control over path length and node count in static environments; it maintains good obstacle avoidance success rate and motion stability in complex environments with both unknown static and dynamic obstacles; and it effectively resolves conflict and stagnation issues in multi-agent cooperative scenarios. Compared with traditional A* and extended neighborhood methods, the proposed method shows significant advantages in pathfinding efficiency, path smoothness, and multi-agent cooperative capabilities.

[0123] In summary, the improved multi-agent path planning method integrating A* and DWA proposed in this paper has high practicality and promotion potential in complex dynamic environments, providing effective technical support for the application of multi-agent systems in warehousing and logistics, inspection operations and other fields.

[0124] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0125] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multi-agent priority obstacle avoidance path planning method for dynamic obstacles, characterized in that, The steps are as follows: S1. Global Path Planning: An improved A* algorithm is used to generate an initial global path. The node expansion function of the improved A* algorithm introduces a variety of cost adjustment mechanisms, including a heuristic scaling factor to improve the problem of decreased search accuracy near the endpoint of the traditional A* algorithm, an obstacle penalty term to improve path safety, a direction penalty term to enhance the coherence of path direction, and an endpoint neighborhood adaptive expansion mechanism to optimize path convergence stability. S2, Path Smoothing Optimization: The initial global path generated in S1 is sequentially subjected to three-fold polyline reconstruction optimization and second-order Bézier curve smoothing to obtain a globally smooth path that satisfies the agent's kinematic constraints. S3, Local Path Planning: Guided by the globally smooth path optimized in S2, the DWA algorithm, which integrates fuzzy control and dynamic weight adjustment mechanism, adaptively adjusts the evaluation function weights according to dynamic obstacles and environmental conditions, and achieves local real-time obstacle avoidance within the global path framework, ensuring that path deviation is controllable. S4. Multi-agent cooperative control: Based on the global smooth path of S2, a priority obstacle avoidance strategy is designed by combining the relative distance between agents, attitude angle and target position. It is combined with a dynamic look-ahead-driven local target update mechanism to coordinate the conflicts when multiple agents meet along the global path, avoid deviation from the global optimal path and eliminate collective stagnation.

2. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The heuristic scaling factor mentioned in S1 is integrated into the total path cost function, the expression of which is: Where n refers to the current node in the search process, corresponding to a feasible state of the robot in the path planning scenario, including position, pose, or task stage. This represents the cumulative cost from the starting point to the current node. Let R be the heuristic cost from the current node to the target node, r be the straight-line distance from the starting point to the target node, and W be the heuristic cost from the current node to the target node. The weights, where a is the scaling factor, with a value of 1; The expression for the obstacle penalty term is: in, This represents the local risk assessment value of the current node in the obstacle grid. This is the weighting coefficient, with a value of 0.5; The expression for the direction penalty term is: in, The direction vector from the starting point to the current node. The direction vector from the current node to the expanding node. This is the directional penalty coefficient, with a value of 0.

5.

3. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The neighborhood adaptive expansion mechanism described in S1 is as follows: in the early stages of the search, an 8-neighborhood expansion is used, and when the current node is less than 10 grids away from the target point, it switches to a 4-neighborhood expansion.

4. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The cubic polyline reconstruction optimization described in S2 includes: First-order polyline optimization: Starting from the starting point, merge subsequent path points that can form a straight line to obtain a primary simplified path; Secondary polyline optimization: Simplify adjacent node pairs with small angles in the primary simplified path to further reduce inflection points; Triple-line optimization: Fitting with a three-segment neighborhood sliding window, replacing or deleting locally jittery nodes.

5. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The second-order Bézier curve smoothing process described in S2 is as follows: The second-order Bézier curve defines the curve shape through n control points, and optimizes the path smoothness by adjusting the spatial position and weight distribution of the control points. The expression for the generated points is: in, Let be the i-th control point, and t be a parameter with a value range of . .

6. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The input variables for the fuzzy control dynamic weight adjustment mechanism described in S3 include the distance between the current agent and the nearest obstacle. Distance between the current agent and the target point The distance between the current agent and other agents Each input variable contains 4 fuzzy subsets, and dynamic adjustment is achieved through 64 fuzzy rules; the output variables are the target orientation term weight α, obstacle avoidance distance term weight β, and speed term weight γ of the evaluation function; The formula for normalizing input variables is: in It means very close. It indicates that it is very far away.

7. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The rules for the priority obstacle avoidance strategy described in S4 are as follows: (1) Set agent a as the highest obstacle avoidance priority, and its obstacle avoidance behavior takes precedence over agents b and c; when the distance between agent a and the target point is greater than 0.2 meters, obstacle avoidance control is started; if the distance is greater than 2 meters, all obstacles are included in the obstacle avoidance calculation and obstacle avoidance trajectory is generated by DWA; if the distance is ≤2 meters, only the closest obstacle is avoided, and the dynamic window is re-evaluated and the obstacle information is updated. (2) The obstacle avoidance mechanism of agent b is based on the relative distance and angle with agent a: when the distance between agent b and other agents is less than 4 meters and the relative angle between them is less than the set threshold, the obstacle avoidance action stops; in other scenarios, if the distance to the target point is far, all obstacles are considered, and if the distance to the target point is close, only the nearest obstacle is considered. Path planning and obstacle avoidance are completed through DWA. (3) The obstacle avoidance of agent c requires simultaneous judgment of the relative distance and angle with agents a and b: when the distance between agent c and agent a or b is less than 4 meters and the corresponding relative angle is less than the set threshold, the obstacle avoidance action stops; in other scenarios, obstacle avoidance control is performed by DWA based on the principle of considering all obstacles for distant targets and only considering the nearest obstacle for near targets.

8. The multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The formula for calculating the look-ahead step size of the dynamic look-ahead-driven local target update mechanism described in S4 is as follows: Where L represents the look-forward step size, Let be the linear velocity of agent i at the current moment, k be the look-ahead coefficient, k is optimized to a value of 1.2, and l be the basic step size threshold.

9. A multi-agent priority obstacle avoidance path planning method for dynamic obstacles according to claim 1, characterized in that: The parameters of the intelligent agent are set as follows: maximum speed 10m / s, maximum rotation speed 20rad / s, acceleration 0.4m / s², rotational acceleration 50rad / s², speed resolution 0.02m / s, and rotational speed resolution 1rad / s.

Images