Intelligent agent off-line and on-line logistics path planning method and system

Abstract

The invention discloses an agent off-line and on-line logistics path planning method and system, and relates to the technical field of path planning, and the core idea of the method is that through the extraction and clustering analysis of obstacle boundary points, the environment complexity is defined, and then the step length and the bias guide growth probability are self-adapted; in the expansion tree growth stage, a state space is divided, a value function is set to provide sampling area reference for non-target-oriented growth, and an off-line path is simplified through pruning to serve as an on-line execution benchmark. Virtual target points are introduced to increase escape force, and possible zero resultant force balance points are damaged; repulsive force is regulated and controlled in a multi-level mode, global and local obstacle avoidance is dynamically balanced, and the emergency obstacle avoidance capacity is improved while the safe distance is ensured. According to the method, off-line path planning and on-line obstacle avoidance real-time planning are fused, the high-value pre-calculated off-line path is used as guidance, on-line obstacle avoidance planning is dynamically and efficiently executed, and it is ensured that efficient and effective path planning is achieved when an intelligent agent executes a task.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agent path planning, and in particular to an intelligent agent offline online logistics path planning method and system. Background Art

[0002] As a crucial innovative tool in modern science and technology, intelligent agents offer high flexibility and low cost. With the rapid iteration and innovation of intelligent agent technology, their exceptional mobility, affordability, and versatility are demonstrating broad application prospects across numerous fields. Whether in civilian or military applications, the scope of intelligent agents is constantly expanding, making them a crucial tool for improving efficiency and reducing costs.

[0003] Path planning, as the foundation of intelligent agent motion control, is a crucial step in the agent's task execution process, impacting its operational efficiency. When performing tasks, intelligent agents often face complex and changing environments. Path planning helps them find the optimal path between their starting point and destination. Especially in resource-constrained environments, proper path planning can maximize battery utilization and improve task execution efficiency.

[0004] Agent path planning is typically divided into offline and online strategies. These two strategies each focus on specific tasks, complementing each other to ensure efficient task execution. Offline path planning relies on prior information about the entire environment to calculate and generate an optimal path from the starting point to the destination. It is typically used to construct a global path. However, due to its high reliance on environmental information, it struggles to cope with unexpected situations in dynamic or uncertain environments, making it more suitable for tasks in static environments.

[0005] Compared to offline path planning, online path planning places greater emphasis on the agent's ability to adjust instantly. It relies on the agent's sensors to perceive the surrounding environment in real time. Through dynamic obstacle avoidance strategies, it adjusts its route in real time to flexibly respond to environmental changes. This makes online planning more advantageous in complex and dynamic scenarios. However, online planning also faces the challenge of balancing real-time performance and safety. If these two factors are not properly addressed, they will seriously restrict the agent's operational efficiency. Therefore, it is crucial for the agent to develop an efficient offline-online combined path planning method that can combine the pre-planning capabilities of offline planning with the real-time adjustment advantages of online planning, thereby ensuring the efficient and stable operation of the agent in complex environments.

[0006] RRT can efficiently search for feasible paths in complex and unstructured environments, and has strong adaptability. However, when facing complex environments, the search path may be slow, and is greatly affected by the density of obstacles. Online path planning based on APF can adjust the path according to the real-time information of dynamic obstacles to avoid collision. However, for complex-shaped obstacles or narrow channels, APF may be trapped in a certain local area and cannot proceed.

[0007] In order to overcome the limitations of RRT and APF, the two are combined to realize the complementary advantages of offline and online path planning. However, there are many challenges in this process. For example, how to effectively integrate the adaptability of RRT and the real-time performance of APF to ensure that the path planning is both fast and accurate; how to maintain the stability of the APF potential field in a dynamic environment to reduce the error of path planning; and how to reduce the computational complexity of RRT in complex environments to improve the efficiency of path planning. SUMMARY

[0008] The method aims to provide an intelligent agent offline and online logistics path planning method and system to solve the contradiction between real-time performance, globality and computational complexity of the path planning algorithm in the prior art.

[0009] To achieve the above purpose, the following technical solutions are adopted: an intelligent agent offline and online logistics path planning method, which comprises: defining a region value function by dividing a state space to distinguish sampling value, realizing sampling bias when there is no target orientation; clustering the density of obstacle boundary points to analyze the complexity of the environment, and dynamically adjusting the step length and target bias orientation probability of the fast search random tree algorithm according to different complexity regions; generating an initial path based on static environment information, and removing redundant nodes through pruning optimization; using the artificial potential field method to realize real-time perception of the state of surrounding static and dynamic obstacles, combining a virtual target point to construct a strategy, and maintaining the navigation trend to the target point; for different types of obstacles, different avoidance priorities are set through multi-level regulation of repulsive force to realize the obstacle avoidance effect and stable planning path of the intelligent agent in a dynamic environment.

[0010] As a preferred scheme of the intelligent agent offline and online logistics path planning method, the state space is divided based on the spatial dynamic 16 division method of the position coordinates of the new growth point, the random tree growth point is taken as the center point, the space is divided into 4 regions through cross division, and the 4 spaces are further evenly divided to obtain 16 sub-regions.

[0011] The region value function is constructed by combining the distance of the target point, the ideal optimal path and the obstacle distribution factors, and the evaluation function value is used to guide the biased sampling through the dynamic region division method.

[0012] The value function based on the distance of the target point, the value function based on the distance between the regional center point and the ideal optimal path, and the value function based on the obstacle density are constructed, and the comprehensive value is calculated by comprehensively considering the value functions.

[0013] As a preferred scheme of the intelligent agent offline-online logistics path planning method, the analysis of the environment complexity comprises clustering the obstacle boundary points by using the DBSCAN algorithm, extracting the shape features and spatial distribution of the obstacles, and quantifying the environment complexity based on the clustering analysis and the minimum circumscribed rectangle of the obstacles.

[0014] As a preferred scheme of the intelligent agent offline-online logistics path planning method, the step length of the dynamic adjustment of the rapid search random tree algorithm comprises defining the step length by uniformizing the environment complexity.

[0015] The target bias guiding probability is used to guide the expansion tree to quickly approach the target by setting the target guiding probability, and the target bias probability is defined by uniformizing the environment complexity.

[0016] As a preferred scheme of the intelligent agent offline-online logistics path planning method, the pruning optimization comprises creating a pruning node set, connecting the newly added points with the subsequent path points one by one from the starting point, and checking whether there is a collision in the connection process.

[0017] When a collision is detected, the node before the collision occurs is added to the pruning node set, and the subsequent nodes are continuously detected for collision, and all nodes in the pruning node set are connected in order to obtain an offline path that is simplified and does not pass through the collision points.

[0018] As a preferred scheme of the intelligent agent offline-online logistics path planning method, the virtual target point construction strategy comprises dynamically adjusting the introduction of the virtual target point p vend A partial escape guide is added to the node that falls into the target unreachable or local optimal, which moves away from the obstacle while continuing to move towards the target point.

[0019] The virtual target point is initialized as the actual target point, and all obstacles within a certain range considering the current position of the intelligent agent are set, when the intelligent agent moves towards the obstacle, the virtual target point will be offset based on the actual target point, the size of the offset depends on the distance between the intelligent agent and the obstacle, and the offset direction always points to the direction away from the obstacle, and the virtual target attraction F vatt is:

[0020] F vatt =k att ·d(pnow ,p vend )

[0021] wherein, k att is a repulsive force coefficient, p now represents the current position of the agent, p vend represents the position of the virtual target point.

[0022] As a preferred scheme of the agent offline-online logistics path planning method, the multi-level regulation of the repulsive force comprises introducing a near-distance emergency obstacle avoidance repulsive force term based on a traditional repulsive force formula, enhancing the sensitivity of the agent to the obstacle through an exponential form, and setting the repulsive force as follows,

[0023]

[0024] wherein, k req is a repulsive force coefficient, p now represents the current position of the agent, p obs represents the position of the center point of the obstacle, p end represents the actual target point position, and d safe represents a specified range of the repulsive force influence.

[0025] Another object of the present application is to provide an agent offline-online logistics path planning system, which can realize efficient path planning of the agent in a complex environment.

[0026] To solve the above technical problems, the present application provides the following technical scheme: an agent offline-online logistics path planning system, comprising an environment complexity analysis module (100), an offline path planning module (200), an offline path pruning module (300) and an online path adjustment module (400);

[0027] The environment complexity analysis module (100) is used for an initial static state space, evaluates the environment complexity by using a density clustering method, and dynamically changes the parameters of the RRT algorithm and the sampling bias.

[0028] The offline path planning module (200) generates an initial path based on an improved RRT algorithm in combination with the environment complexity analysis result.

[0029] The offline path pruning module (300) prunes and simplifies the offline path, so that the path is more suitable for online following.

[0030] The online path adjustment module (400) is based on the artificial potential field method, perceives the surrounding static and dynamic obstacles in real time, dynamically adjusts the path of the agent, ensures the obstacle avoidance ability and planning safety, and maintains the overall navigation trend to the target point.

[0031] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the off-line and on-line logistics path planning method of the agent as described above when executing the computer program.

[0032] A computer readable storage medium stores a computer program, and the computer program implements the steps of the off-line and on-line logistics path planning method of the agent as described above when executed by a processor.

[0033] The present application has the beneficial effects that: the present application adopts the environment complexity analysis mechanism, accurately evaluates the environment characteristics through the density clustering method, and dynamically adjusts the parameters of the RRT algorithm, so that the path sampling process can adapt to different environments, improve the adaptability and calculation efficiency of the path planning, and thus solve the problem that the fixed parameters are difficult to adapt to complex environments.

[0034] The present application optimizes the initial planning path in the offline stage through the path pruning optimization strategy, removes the redundant nodes, reduces the path complexity, makes the on-line path adjustment more smooth, and improves the following performance and execution efficiency of the agent.

[0035] The present application adopts the artificial potential field method to realize real-time perception of static and dynamic obstacles, sets different avoidance priorities for different types of obstacles, and combines the virtual target point adjustment strategy, so as to maintain the navigation trend to the target point while ensuring the obstacle avoidance ability, thereby realizing the balance between the obstacle avoidance effect and the path stability, and improving the safety and reliability of the agent in the dynamic environment. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0037] Figure 1 A flow chart of an off-line and on-line logistics path planning method of an agent is provided for an embodiment of the present application.

[0038] Figure 2 A sampling bias region division example diagram of an off-line and on-line logistics path planning method of an agent is provided for an embodiment of the present application.

[0039] Figure 3 An obstacle clustering partition example diagram of an off-line and on-line logistics path planning method of an agent is provided for an embodiment of the present application.

[0040] Figure 4A simple environment online planning example diagram of an agent offline-online logistics path planning method provided for an embodiment of the present application.

[0041] Figure 5 A complex environment online planning example diagram of an agent offline-online logistics path planning method provided for an embodiment of the present application.

[0042] Figure 6 A module structure diagram of an agent offline-online logistics path planning system provided for an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0044] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0045] Embodiment 1, refer to Figures 1-3 , for the first embodiment of the present application, the embodiment provides an anti-interference type indoor positioning method, including, as shown in Figure 1

[0046] S1: Define the region value function by dividing the state space to distinguish the sampling value, realize the sampling bias when there is no target orientation

[0047] Refer to Figure 2 . First, take p new as the center point, divide the space into 4 regions by cross division, and then divide the 4 spaces again, finally get 16 sub-regions. Define the value function to evaluate the sampling priority of each region according to the factors such as the distance between the center point of the cutting region and the target point, the distance from the ideal optimal path and the distribution of obstacles in the region.

[0048] The value function based on the distance from the target point is calculated as follows:

[0049]

[0050] Where diste X(i) represents the distance between the center point of the region X(i) and p end ​distance between the current node and the goal node, Line is the length of the diagonal of the state space, representing the maximum possible distance. The sampling process is guided by quantifying the distance between the extended tree nodes and the goal node, with regions closer to the goal node having higher values R1, avoiding wasting computational resources in regions far from the goal.

[0051] The value function based on the ideal optimal path is calculated as follows:

[0052]

[0053] where distz X(i) represents the straight-line distance between the center point of region X(i) and q ideal , distz max is the farthest distance between all region center points and q ideal . It can be seen that regions closer to q ideal have higher values R2, and these regions are more likely to contain nodes of low-cost paths.

[0054] The value function based on the obstacle distribution is calculated as follows:

[0055]

[0056] where Points X(i) represents the number of boundary points of obstacles in region X(i), and the larger this value, the more obstacles or the larger area occupied by obstacles in this region, Area X(i) represents the area of region X(i). Regions with higher values R3 have fewer obstacles per unit area and more open space, making them suitable for sampling and path expansion.

[0057] Based on the above value functions, the comprehensive value can be calculated as follows:

[0058] NR sum (i) = w1NR1(i) + w2NR2(i) + w3NR3(i)

[0059] where w1, w2, and w3 represent the weights of the normalized value functions NR1, NR2, and NR3, respectively, and w1 + w2 + w3 = 1. The calculation of the normalized value functions is as follows:

[0060]

[0061] Further, to determine the corresponding weights of each value function, the analytic hierarchy process is used for calculation, and after comprehensive consideration, it is believed that the value function R1 has the highest importance, while R2 and R3 have comparable and lower importance. According to the Saaty scale method (1-9 scale), the calculation can obtain the optimal weight distribution as follows: w1 = 0.714, w2 = 0.143, and w3 = 0.143.

[0062] In one feasible embodiment, the determination of the respective weights of the value functions can be achieved by dynamic weight distribution based on entropy weight method. Specifically, in the initialization phase, historical sampling data of each region (such as the distance from the target point, the obstacle density, etc.) is collected, and the information entropy of the data of each value function (distance, path, obstacle) is calculated. The lower the entropy value, the higher the degree of differentiation of the index. According to the entropy value, the weight is dynamically adjusted, and the index with high degree of differentiation is given a higher weight (such as automatically increasing the weight of the distance from the target). As the path search progresses, the entropy value is periodically recalculated, and the weight distribution is dynamically optimized.

[0063] In another feasible embodiment, the determination of the respective weights of the value functions can be achieved by fuzzy comprehensive evaluation. Specifically, the value function values (such as distance, path, obstacle) of each region are divided into "high, medium, low" fuzzy levels, and rules are defined based on expert experience (for example, "if the distance from the target is short and the obstacle is few, then the weight is the highest"). Through the fuzzy rule base, the comprehensive value of each region is inferred, and the weight distribution result is output. The fuzzy output is converted into precise weight values, which are used to calculate the sampling probability.

[0064] The sampling bias probability P(i) of each region can be calculated as follows:

[0065]

[0066] R sum is normalized, and the normalized value is further converted into a cumulative probability distribution value using the cumulative distribution function. During the random tree expansion process, a random number rand ∈ [0, 1] is generated as the basis for selecting the sampling bias region.

[0067] S2: Cluster the environment complexity according to the density of obstacle boundary points, and dynamically adjust the step size and target bias orientation probability of the fast search random tree algorithm according to different complexity regions.

[0068] Initialize the state space, the starting point and target point positions of the agent, and other parameters. Define the environment complexity according to the obstacle space distribution state, obtain the complexity value EC of each region, and calculate it as follows:

[0069]

[0070] where S Area(i) represents the minimum circumscribed rectangle area of the cluster region Area(i), and Points(i) represents the number of obstacle boundary points contained in the cluster region Area(i). Referring to Figure 3 It can be seen that the more obstacle points per unit area in a region, the more dense the distribution of obstacles, and the higher the corresponding environment complexity.

[0071] In one feasible embodiment, the environment complexity is obtained by environment feature extraction through a convolutional neural network. Specifically, the state space is divided into a grid map, the static obstacle area is marked as a binary image, and a pre-trained lightweight CNN model (such as MobileNet) is used to extract the local complexity features of the grid map. According to the CNN output feature values, the area is divided into three levels of "simple, medium, and complex", the high complexity area reduces the step size, and the low complexity area increases the target bias probability.

[0072] In another feasible embodiment, the environment complexity can also be obtained by complexity evaluation based on region connectivity. Specifically, the flood fill algorithm is used to detect the connectivity of each region to the target point, and the skeleton extraction algorithm is used to calculate the width of the channel in the region. Narrow areas are marked as high complexity. Narrow areas shorten the step size and reduce the target orientation probability, and wide areas are the opposite.

[0073] Adaptive RRT bias sampling area based on environment complexity. To achieve adaptive bias orientation probability in different complex areas, the environment complexity EC needs to be normalized. The calculation is as follows:

[0074]

[0075] Where EC(i) is the environment complexity of region i.

[0076] The next random tree expansion direction is determined by generating a random number rand ∈ [0, 1] as follows:

[0077]

[0078] Where p end is the path planning target point, NEC(i) is the normalized environment complexity of region i, and rand is the generated random number. The more complex the environment, the lower the target orientation growth probability; the more open the environment, the higher the target orientation growth probability. When the tree is expanded, the sampling point is obtained based on the sampling bias area X(i) selected in the above steps.

[0079] Adaptive RRT step length based on environment complexity. For open areas, i.e. NEC = 0, the step size is set to the basic step size step basic , which ensures that the expansion tree can grow quickly and reduce the search time; for areas with certain environment complexity, the step size is appropriately shortened, so that the expansion tree can more accurately adapt to the distribution of obstacles and avoid being trapped in local optimum due to too large step size. The specific adjustment formula of the step size is as follows:

[0080] step = step basic · [1-NEC(i)] 1.5

[0081] S3: Generate initial path based on static environment information, and remove redundant nodes through pruning optimization.

[0082] Offline path pruning, initialize pruning node set, connect new added points with subsequent path points in turn from the starting point, and check for collision. If collision occurs, add the node before the collision to the pruning node set, and continue to detect subsequent nodes. Finally, connect all nodes in the pruning node set in turn to obtain the simplified offline path.

[0083] S4: Real-time perception of surrounding static and dynamic obstacle states using artificial potential field method, combined with virtual target point construction strategy to maintain navigation trend to target point.

[0084] Dynamic adjustment of virtual target point p based on obstacle position vend Add partial escape leads to nodes trapped in target unreachable or local optimal to achieve moving away from obstacles while continuing to move towards the target point. Virtual target attraction calculation is shown in Table 1.

[0085] Table 1 Virtual target attraction construction algorithm

[0086]

[0087]

[0088] The total attraction of online planning is as follows:

[0089] F att_total = F att + F vatt

[0090] Further, the repulsion formula is defined as follows: On the basis of traditional repulsion, introduce the short-distance emergency obstacle avoidance repulsion term F req2 , adjust the nonlinear decay characteristics of repulsion through exponential form, strengthen the sensitivity of short-distance obstacle avoidance, and enhance the obstacle avoidance effect.

[0091]

[0092] where k req is the repulsion coefficient, p now represents the current position of the agent, p obs represents the position of the obstacle center point, p end represents the actual target point position, and d safe represents the specified range of repulsion influence.

[0093] S5: For different types of obstacles, set different avoidance priorities through multi-level regulation of repulsion to achieve obstacle avoidance effect and stable planning path of the agent in dynamic environment.

[0094] During the online planning process, the agent has different priorities for avoiding static and dynamic obstacles, and prioritizes the correlation coefficients between the two. For static obstacles, a smaller repulsion coefficient is set, k req =k req_static , weaken its repulsive force on the intelligent body and reduce the adjustment range; set a larger safety distance range, that is, d safe =d safe_static , ensuring that there is always enough buffer space between the agent and the static obstacle. In contrast, the position and speed of dynamic obstacles are uncertain, and the collision risk is higher. The agent needs to perceive the state of dynamic obstacles in real time and quickly adjust the path to ensure safety. To this end, a larger repulsion coefficient is set. At this time, k req =k req_dynamic , enhance the impact of dynamic obstacles on the agent's obstacle avoidance. Set a smaller safety distance range, i.e. d safe =d safe_dynamic , allowing the agent to quickly adjust its path when approaching dynamic obstacles. req_static <k req_dynamic , d safe_static >d safe_dynamic .

[0095] The total repulsive force of online planning is as follows:

[0096] F req_total =F req_s +F req_d

[0097] The net force on the agent's online planning is as follows:

[0098] F total =F att_total +F req_total

[0099] Under the combined force, the agent's position is updated as follows, where p now Indicates the current position of the agent, and step is the current step length of the agent:

[0100]

[0101] Example 2, reference Figure 4 and Figure 5, the second embodiment of the present application is demonstrated. Different from the first embodiment, the path planning method of the agent off-line / on-line combination further comprises the step of verifying the technical effectiveness of the method. Specifically, to fully explain the application effect of the method, the embodiment adopts the comparative test method to compare the traditional technical solution with the method proposed in the present application. The experimental results are compared and analyzed by scientific demonstration method, aiming to verify the actual performance of the method of the present application.

[0102] To further verify the performance of the method of the present application in complex and variable environment, two typical scenes are selected for comparative test: simple environment with sparse obstacle distribution EC=0, and complex environment with dense obstacle distribution and uneven distribution EC increasing. In the two environments, off-line and on-line path planning verification is carried out. The purpose is to evaluate the planning efficiency and obstacle avoidance ability of the method of the present application in the face of high complexity state space environment by scientific method.

[0103] To verify the planning performance of the agent off-line and on-line logistics path planning method (HOOPP) proposed in the present application, RRT*, RRT-Connect and GB-RRT algorithms are selected as off-line planning comparison algorithms, and the traditional APF algorithm is selected as on-line planning comparison algorithm. The planning path length and planning time are taken as the main measurement index, and the effectiveness of the HOOPP method of the embodiment is analyzed. To eliminate the potential influence of equipment difference on algorithm performance evaluation, the off-line and on-line planning processes are carried out on the same equipment, and Matlab2021b is used as the experimental simulation platform to ensure the consistency and comparability of data processing.

[0104] Considering the system computing power, the experimental parameters are set to common configuration. To ensure the fairness of the comparative experiment, the parameter configuration of RRT* algorithm, GB-RRT* algorithm and Bi-RRT* algorithm is consistent with the algorithm of the present application.

[0105] Through experimental simulation observation, it is known that the above algorithms successfully achieve the off-line path planning task in a simple environment. The sampling points of the RRT* algorithm, the RRT-Connect algorithm and the GB-RRT algorithm are uniformly distributed in the state space in a random state, and the paths planned by the RRT* algorithm and the RRT-Connect algorithm are more tortuous and farther from the ideal optimal path. Since the GB-RRT algorithm has a certain target bias probability, compared with the other two algorithms, only a small number of sampling times are required, and the planned path has an obvious growth trend towards the target. The HOOPP algorithm in the simple environment with EC = 0 can not only maintain a bias growth probability of 0.5, but also make more sampling points concentrate in the area close to the target point and the ideal optimal path through the division and selection of the biased sampling area, so as to quickly find the target point and reduce the number of required sampling points.

[0106] Table 2 shows the off-line path planning performance of the four algorithms in a simple environment, wherein the average deviation distance is calculated as the average of the straight line distances between all path nodes and the ideal optimal path. The double-target-domain bias strategy of HOOPP enables the off-line path to be quickly planned in the shortest time, while avoiding invalid sampling and improving sampling efficiency. In addition, the algorithm has a low average deviation distance, fully embodying the role of the value function R2 in the selection of the sampling bias area. The GB-RRT algorithm has a high target growth probability, and its performance is close to that of the HOOPP algorithm, and it can also quickly plan a low-cost path in a simple environment with sparse obstacles, and approach the ideal optimal path. In comparison, the random sampling strategy of RRT* results in a lower path quality, a longer planning time, and a larger deviation from the ideal path, with the average deviation distance being nearly 1.6 times that of HOOPP. The bidirectional expansion strategy of RRT-Connect improves the path quality and reduces the planning time to some extent, but there is still a big difference between it and the HOOPP algorithm, with the planning time difference being about 2.3 times and the iteration number difference being nearly 2.8 times.

[0107] Table 2 Comparison of performances of four algorithms in a simple environment

[0108]

[0109]

[0110] The four algorithms are used for offline path planning in complex environment. Since the sampling height of RRT* algorithm and RRT-Connect algorithm is random, the sampling points are uniformly distributed in the state space, so the planned path is relatively tortuous and deviates far from the ideal optimal path. The GB-RRT algorithm has the target bias characteristic. Compared with the simple environment, the dense obstacles in the complex environment increase the collision times caused by target bias, and more sampling points are needed to achieve effective path planning. The HOOPP proposed in the application can realize path planning with fewer sampling points by adaptive target-oriented growth probability and biased sampling region selection based on the complexity of the environment, and the sampling points are mostly distributed near the target point.

[0111] Table 3 shows the offline path planning performance of the four algorithms in the complex environment. The target bias strategy of the GB-RRT algorithm is prone to cause more sampling failures in the complex obstacle environment, which increases the number of iterations of the algorithm, about 2.6 times of HOOPP, so the overall planning time is also longer. In contrast, HOOPP can more accurately approach the ideal optimal path while constantly approaching the target region through the definition of the value function, thereby completing the planning in fewer iterations. GB-RRT algorithm and HOOPP have similar path length, but GB-RRT algorithm has obvious waste of computing resources and is less efficient. RRT-Connect uses a bidirectional tree strategy, and the overall planning performance is slightly better than RRT* algorithm. However, since both of them rely on random sampling strategy, their planning performance is still significantly inferior to HOOPP, and the number of iterations required is more than 1.8 times of HOOPP.

[0112] Table 3 Comparison of performance of four algorithms in complex environment

[0113]

[0114]

[0115] Reference Figure 4 In which, the yellow circle represents an unknown dynamic obstacle, which is marked as No. 1 dynamic obstacle and No. 2 dynamic obstacle respectively, and performs horizontal uniform circular motion in the state space. The blue rectangle represents an unknown static obstacle, the green dashed line represents the pruning path, the red solid line represents the online path adjusted in real time according to the environment, and the pink dot represents the agent. It can be seen from Figure 4 (b) and Figure 4 (d) that the agent can perceive dynamic obstacles in time and effectively avoid them during online navigation. For Figure 4The static obstacles in (c) are avoided by the HOOPP through a multi-level repulsive force regulation mechanism, which ensures the stability of the path and effectively maintains a safe distance. For unknown static obstacles appearing on the offline path, the HOOPP also effectively avoids them, as shown in Figure 4 (f). After the agent gradually moves away from the influence area of the obstacle, it constantly approaches the pruned path and finally successfully reaches the target point.

[0116] Referring to Figure 5 , the agent can effectively avoid the dynamic obstacles in the complex environment in real time according to Figure 5 (b) and Figure 5 (d). For unknown static obstacles in Figure 5 (c), the agent successfully perceives and effectively avoids them, ensuring the stability of the progress. Figure 5 (e) shows that in the obstacle-dense area, the pruned path is tangent to multiple obstacle boundaries. To ensure the safety of the path, the agent effectively maintains a safe distance through repulsive force regulation to avoid collision and successfully reaches the target point. During online planning, the agent always follows the pruned path as a guide and adjusts the path according to the real-time environment. After gradually moving away from the influence area of the obstacle, the path gradually approaches the pruned path, ensuring the efficiency and safety of the planning.

[0117] Embodiment 3 is the third embodiment of the present application, which is different from the first two embodiments:

[0118] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the parts that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of software products, which are stored in a storage medium and include instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0119] The logic and / or steps represented in the flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be for example but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example by optically scanning the paper or other suitable medium, then

[0120] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example by optically scanning the paper or other suitable medium, then

[0121] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), and / or the like.

[0122] Example 4, with reference to Figure 6For an embodiment of the application, an intelligent agent offline-online logistics path planning system is provided, characterized by comprising an environment complexity analysis module 100, an offline path planning module 200, an offline path pruning module 300, and an online path adjustment module 400; the environment complexity analysis module 100 is used for an initial static state space, uses a density clustering method to evaluate the environment complexity, and dynamically changes the parameters of the RRT algorithm and the sampling bias to improve the adaptability of path planning; the offline path planning module 200 generates an initial path based on an improved RRT algorithm combined with the environment complexity analysis result, ensures the feasibility and global optimization effect of the path; the offline path pruning module 300 prunes and simplifies the offline path to reduce redundant nodes, improve the path feasibility, and make it more suitable for online following; the online path adjustment module 400 is based on the artificial potential field method, perceives the surrounding static and dynamic obstacles in real time, and dynamically adjusts the path of the intelligent agent to ensure the obstacle avoidance ability and planning safety, while maintaining the overall navigation trend to the target point.

Claims

1. An intelligent offline online logistics path planning method, characterized by: include, By dividing the state space and defining the regional value function to distinguish the sampling value, the sampling bias can be achieved when there is no goal guidance. The environment complexity is analyzed by clustering the density of obstacle boundary points, and the step size and target bias guidance probability of the fast search random tree algorithm are dynamically adjusted according to different complexity areas; Generate an initial path based on static environment information and remove redundant nodes through pruning optimization; The artificial potential field method is used to perceive the state of surrounding static and dynamic obstacles in real time, and combined with the virtual target point construction strategy, the navigation trend towards the target point is maintained; For different types of obstacles, different avoidance priorities are set through multi-level regulation of repulsion, so as to achieve obstacle avoidance effect and stable path planning of the intelligent agent in a dynamic environment.

2. The method for offline online logistics path planning of an intelligent agent according to claim 1, characterized in that: The state space is divided into four regions by a cross division based on the spatial dynamic 16-division method of the new growth point position coordinates, with the random tree growth point as the center point, and then the four spaces are evenly divided to obtain 16 sub-regions; The regional value function comprehensively considers the distance to the target point, the ideal optimal path and the obstacle distribution factors to construct an evaluation function, and combines the dynamic region division method to guide the bias sampling through the evaluation function value; This includes constructing a value function based on the distance to the target point, constructing a value function based on the distance between the center point of the area and the ideal optimal path, and constructing a value function based on the density of obstacles. The above value functions are combined to calculate the comprehensive value.

3. The method for offline online logistics path planning of an intelligent agent according to claim 2, characterized in that: The analysis of environmental complexity includes clustering obstacle boundary points using the DBSCAN algorithm, extracting shape features and spatial distribution of obstacles, and using an environmental complexity quantification method based on the minimum circumscribed rectangle of obstacles based on cluster analysis.

4. The method for offline online logistics path planning of an intelligent agent according to claim 3, characterized in that: The dynamically adjusting the step size of the fast search random tree algorithm includes defining the step size by normalizing the complexity of the environment; The target bias-guided probability and the target-guided growth strategy guide the expansion tree to quickly approach the target by setting the target-guided probability, and define the target bias probability by homogenizing the environment complexity.

5. The method for offline online logistics path planning of an intelligent agent according to claim 4, characterized in that: The pruning optimization includes creating a set of pruned nodes, connecting the newly added points with subsequent path points one by one starting from the starting point, and checking whether there is a collision during the connection process; When a collision is detected, the node before the collision is added to the pruned node set, and collision detection is continued for subsequent nodes. All nodes in the pruned node set are connected in sequence to obtain a simplified offline path that does not pass through the collision point.

6. The method for offline online logistics path planning of an intelligent agent according to claim 5, characterized in that: The virtual target point construction strategy includes dynamically adjusting and introducing the virtual target point p based on the obstacle position. vend , for nodes that are trapped in unreachable targets or local optimality, some escape guides are added to keep them away from obstacles while continuing to move towards the target point; Initialize the virtual target point to the actual target point, considering all obstacles within the set range of the agent's current position. When the agent moves towards the obstacle, the virtual target point will be offset based on the actual target point. The size of the offset depends on the distance between the agent and the obstacle, and the offset direction always points away from the obstacle. The virtual target gravity F vatt for: F vatt =k att ·d(p now ,p vend ) Among them, k att is the gravitational coefficient, p now Indicates the current position of the agent, p vend Indicates the location of the virtual target point.

7. The method for offline online logistics path planning of an intelligent agent according to claim 6, characterized in that: The multi-level control of the repulsive force includes introducing a close-range emergency obstacle avoidance repulsive term based on the traditional repulsive force formula, and enhancing the agent's sensitivity to obstacles in an exponential form. The repulsive force is set as follows: Among them, k req is the repulsion coefficient, p now Indicates the current position of the agent, p obs represents the center point of the obstacle, p end Indicates the actual target point position, d safe Indicates the specified range of influence of repulsive force.

8. An intelligent offline online logistics path planning system according to any one of claims 1 to 7, characterized in that: It includes an environment complexity analysis module (100), an offline path planning module (200), an offline path pruning module (300) and an online path adjustment module (400); The environment complexity analysis module (100) is used for the initial static state space, uses a density clustering method to evaluate the environment complexity, and adapts the dynamic changes of parameters and sampling bias of the RRT algorithm; The offline path planning module (200) generates an initial path based on an improved RRT algorithm and in combination with an environmental complexity analysis result; The offline path pruning module (300) prunes and simplifies the offline path to make the path more suitable for online following; The online path adjustment module (400) is based on an artificial potential field method, and perceives surrounding static and dynamic obstacles in real time, and dynamically adjusts the path of the intelligent body to ensure obstacle avoidance capability and planning safety, while maintaining the overall navigation trend towards the target point.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of an intelligent offline-online logistics path planning method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of an intelligent offline-online logistics path planning method described in any one of claims 1 to 7 are implemented.

Images