Multi-information domain multi-resolution humanoid dual-layer motion planning method

Abstract

The application discloses a two-layer motion planning method combining a human-like global direction guiding path and a local actual strain path based on a multi-information domain and a multi-resolution scene description model. First, referring to the actual process of human motion planning, a low-resolution rough scene description model is required in global guiding path generation, and the local actual path is generated by flexibly adjusting the current high-resolution actual scene information, based on the global guiding path and the local actual path, which are completely decoupled, so that the global guiding path can be completely followed or completely regenerated, thereby improving the flexibility of traditional motion planning algorithms. Secondly, in order to fully describe various constraint information, task requirements in different information domains are flexibly described at global and local levels, so that the diversity of motion planning results at global and local levels can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motion planning algorithm, and particularly to a multi-information domain multi-resolution humanoid double-layer motion planning method BACKGROUND

[0002] The so-called path planning refers to finding an optimal or approximate optimal path from a starting state to a target state in a state space and avoiding all obstacles according to a certain criterion. Path planning is one of the basic behavior capabilities of various agents in the physical world, and is a key to guaranteeing that the agents do not collide with other obstacles and individuals and better advance towards the target in real life. In the research of agent motion planning, this basic behavior characteristic should also be embodied, so that the agent can move according to a natural and reasonable path and smoothly reach the target position. Realistically and effectively simulating the path planning behavior of the agent can not only improve the intelligence of the path planning algorithm, but also provide an important foundation guarantee for higher-level multi-agent behavior simulation.

[0003] The existing path planning method can be divided into two sub-problems of environment modeling and path search, and the basic idea is to first model the environment model, and then use a search algorithm to find the optimal path between the starting point and the target point. However, in the process of environment modeling and path search, most of the existing motion planning algorithms only consider the spatial accessibility relationship, the environment model resolution is fixed, and less consideration is given to high-level task requirements and individual preference constraints and other information, resulting in a single motion planning result and insufficient flexibility and fidelity. Therefore, based on the probabilistic path graph algorithm, a multi-information domain and multi-resolution hierarchical description model is proposed, different constraint information is described at the global and local two levels, and different resolution probabilistic path graphs are constructed at the global and local levels, which can guarantee that the motion planning result meets the various task constraint requirements in different information domains, generates diversified motion paths, and improves the flexibility and solving efficiency of the traditional motion planning algorithm.

[0004] For the above two sub-problems, the motion planning algorithm can be divided into two key steps: first, a scene description model needs to be defined, and on the basis of properly describing various geometric relationship information in the motion scene, the connectivity graph between the free spaces in the motion scene is constructed; then, based on the connectivity graph, the generation of a specific motion path is completed. For the construction of the connectivity graph of the motion scene, the traditional method can be divided into two categories: one is the direct description idea, which uses various graph algorithms (such as unit grid, triangular facet, Voronoi diagram, etc.) to split the complex motion scene into basic free spaces (Cell Decomposition) that are connected to each other, and then directly describes the geometric connectivity relationship based on the geometric connectivity relationship, such as various navigation mesh algorithms. The second is the random sampling idea, which constructs a path graph (Roadmap) by random sampling of the motion scene to indirectly describe the geometric connectivity relationship of the space, such as the probabilistic roadmap PRM and the rapid random expansion tree RRT. However, for complex dynamic scenes, in addition to the static geometric connectivity relationship, it is also necessary to fully describe various dynamic change information. In recent years, people have proposed various improved algorithms to simultaneously describe information such as gap information with obstacles and space-time change information, but it is difficult to perfectly describe all information based on a certain specific description model. In order to improve the adaptability to the dynamic changes of the scene (such as dynamic obstacles and mutual cooperation and avoidance between multiple agents), the current motion planning algorithm is usually decoupled into two parts: global navigation and local navigation.

[0005] According to the above situation, the applicant has developed a multi-information domain multi-resolution human-like double-layer motion planning method in view of the fact that the existing motion planning algorithm only considers the spatial accessibility relationship in the environment modeling and path search process, less considers individual preference constraint information, and the coupling between local path and overall path is too tight, resulting in single motion planning result, insufficient flexibility and realism. SUMMARY

[0006] The purpose of the present application is to provide a multi-information domain multi-resolution human-like double-layer motion planning method to solve the problem of the current motion planning algorithm in the environment modeling and path search process, which only considers the spatial accessibility relationship, less considers individual preference constraint information, and the coupling between local path and overall path is too tight, resulting in single motion planning result, insufficient flexibility and realism.

[0007] To achieve the above purpose, the present application provides the following technical scheme: a multi-information domain multi-resolution human-like double-layer motion planning method, characterized in that it comprises the following steps:

[0008] Step one: problem mathematical modeling; assume the motion scene is W, which has m obstacle regions O obs m}, the free space excluding obstacles is represented by C free , i.e. W = C free + O obs , there are n task target regions R = {R1, …, R n} in the motion scene, which are not adjacent to each other, satisfying (i≠j), the agent's target task is to plan an optimal collision-free motion path according to certain high-level spatiotemporal task logic T R , the order of reaching target regions is represented as Π i = {π i1 … π in}, where π i represents the agent reaching target region R R , i1…in is the specific order arrangement between 1…n that satisfies the specific task constraints T i , C n (i = 1…k) is the preference constraint imposed on the agent according to the actual path conditions during path generation;

[0009] Step two: representation of high-level spatiotemporal task constraints; adopt the idea of describing temporal logic requirements and spatial constraints separately, describe temporal logic requirements using rule expressions, and describe spatial constraint requirements using a combination of global PRM and local PRM,

[0010] 1) sequentially reaching task regions, only requiring to reach regions R1, R2, …, R n , the temporal logic requirements are described using rule expressions as follows:

[0011] φ seq (π1, π2, … π n ) = π1π2… π n , where π i represents reaching target region R i ;

[0012] where perm(1, n) represents all possible permutations of {1, …, n};

[0013] 2) for the case of alternately reaching regions and , it can be represented as:

[0014] 3) more complex high-level task constraints can be described using the above basic rule expressions, such as reaching regions Then in any order to reach the area Can be expressed as:

[0015]

[0016] The above formula is further expanded to the case of dividing the task target into several groups and reaching different groups in sequence, which can be expressed as:

[0017] φ groups (π1,π2,…π n )=φ cov (G1)φ cov (G2)…φ cov (G k )

[0018] Where, G1,G2,…G k Indicates that the task {π1,π2…π n} is divided into k groups;

[0019] Then convert it into the form of DFA (Deterministic Finite Automata) to uniformly describe it; give all possible temporal logic paths;

[0020] Step three: generation of global probability path graph; the basic PRM algorithm is used to generate the global motion path graph;

[0021] Step four: generation of global diversified guided path;

[0022] After obtaining the global probability path graph, the global guided path generation needs to meet the temporal logic constraint graph DFA and the global probability path graph PRM at the same time.

[0023] First, based on the temporal logic constraint description graph DFA, all possible task completion paths Z l (l=1,…n) are obtained, and then according to the global probability path graph, each task completion path Z l The corresponding optimal path Q l On the global probability path graph PRM. The path cost can be simply expressed by the Euclidean distance between the corresponding nodes in the probability path graph PRM, and further introduce global constraints such as the blocking degree and danger degree of different areas, modify the path cost by describing the constraints as path weight, so as to generate the global guided path with preference.

[0024] Finally, based on various search algorithms, find the path Q with the minimum cost in all Q l (l=1,…n) Q = min(Q1,…,Q n) as the final global guidance path, the globally optimal guidance path Q = {q0, q1, … q k g}, wherein q0 is the start position of the global guidance path on the PRM, q g is the end point of the global guidance path on the PRM, {q0, q1, … q k g} is a path on the PRM that satisfies the temporal logic constraints and has the minimum cost;

[0025] Step five: generation of a local high-resolution probabilistic path graph;

[0026] After considering the local preference constraints, the accuracy of the scene description model is improved, and the local constraints are fully expressed; the tunnel concept is introduced, the guidance path is expanded from a one-dimensional straight line to a two-dimensional or three-dimensional safe tunnel, the original global guidance path is taken as the central axis, a certain width is expanded outward, and sampling points are randomly added within the tunnel range to generate a high-resolution local probabilistic path graph; by taking the global guidance path as a basic directional constraint of the movement path between stage sub-tasks, the agent can generate diversified movement trajectories according to its own preferences within the tunnel range, and can also ensure the consistency of the movement direction with the overall global guidance path;

[0027] Specifically, it includes:

[0028] 1) First, determine whether the nodes q i and q j at the two end points of the current global guidance path segment satisfy sufficient sampling, that is, whether l(q i , q j ) < r(q i ) + r(q j ) is satisfied, wherein l(q i , q j ) is the length of the straight line path (q i , q j ), and r(q i ) represents the distance from the global sampling point q i to the nearest obstacle. If the sampling is insufficient, increase the sampling points on the current path segment until the sampling is sufficient;

[0029] 2) Obtain the tunnel width according to the formula , wherein x represents a certain point on the straight line l(q i , q j );

[0030] 3) Within the tunnel width determined by the formula, continue to increase the sampling points by using the uniform random sampling method to generate a local high-resolution probabilistic path graph;

[0031] ​​Step six: local diversification of motion path generation; after obtaining the local high-resolution probability path graph, the similar method as the global constraint description can be used to change the weight of different paths by adding the definition of local preference constraints to obtain different local motion paths during local path generation;

[0032] Step seven: motion path re-planning under dynamic obstacle conditions; after the dynamic obstacle appears, the invalid edges on the probability path graph are first calculated by the possible impact area of the dynamic obstacle, and then removed. Then, the motion path is re-planned and generated at the local level. If an effective motion path cannot be obtained, the guided path is re-planned and generated at the global level, ensuring the adaptability and computational efficiency of the dynamic planning result of the motion path.

[0033] Specifically, the following steps are included:

[0034] 1) Motion path re-planning; first, according to the possible position range of the dynamic obstacle in the current and future period of time, the edges on the probability path graph that may collide with the dynamic obstacle are removed to generate a new probability path graph and re-search to generate a motion path. If the invalid edges on the probability path graph are removed and a new motion path from the current position to the target region of the sub-task cannot be found at the local level, a new global guided path needs to be generated by re-planning at the global level.

[0035] 2) Dynamic recovery and update of the probability path graph; the dynamic recovery check time interval of the probability path graph is appropriately determined, and according to the overall algorithm framework, the dynamic update of the probability path graph is performed after the agent reaches a certain sampling point position on the global probability path graph. If the probability path graph has changed, i.e., the original invalid edge has become valid, the guided path needs to be re-planned and generated at the global level. If a better global guided path can be found, the local motion path needs to be re-planned and generated according to the local constraint preference at the local level. Through the dynamic adjustment of the motion path at the local and global levels, the dynamic adaptability and flexibility of the algorithm framework are ensured.

[0036] Preferably, the task requirements in different information domains, such as the task completion time and logical relationship requirements on the motion path, the collision-free requirement, and the preference constraint information of the virtual individual, are flexibly described at the global and local levels, respectively, to achieve the diversity of the motion planning result at the global and local levels.

[0037] Preferably, during global path graph generation, sampling is first performed in each target region of the sub-task, and then sampling is performed in the free space outside the target region, thereby ensuring that a probability path graph connecting all target regions can be finally generated.

[0038] Preferably, in the global path map generation, in the global optimal path generation from the initial region to the target region, in the global guidance path generation, the temporal logic constraint graph and the global probability path graph need to be satisfied at the same time, and after obtaining all the feasible global motion paths, the path with the minimum cost is found. The path cost can be simply represented by the Euclidean distance between the corresponding nodes in the probability path graph, and further global constraints such as the blocking degree and the danger degree of different regions are introduced to generate diversified global guidance paths.

[0039] Preferably, according to different situations of global task constraints, although only the path with the lowest cost is selected as the final global guidance path, the entire algorithm framework generates and saves all other possible paths, which can ensure the rapidity and flexibility of the path dynamic replanning process in dynamic scene handling.

[0040] Preferably, in the local high-resolution PRM generation, the local resolution of the probability path graph is improved by increasing the number of local sampling points. When generating new sampling points, the invention makes full use of the directional constraint information provided by the previous global guidance path. The original global guidance path is taken as the central axis, and the sampling points are randomly increased within the tunnel range to generate a high-resolution local probability path graph. By taking the global guidance path as a basic directional constraint between stage sub-tasks, different agents can generate diversified motion trajectories according to their own preferences, while ensuring that the motion direction is basically consistent with the global guidance path.

[0041] Preferably, the dynamic recovery and update method of the probability path graph. The influence of various dynamic obstacles in the virtual scene is often temporary, and after a period of time, the invalid edges will become valid edges again as the dynamic obstacles move to other places. Based on the above consideration, during the process of avoiding dynamic obstacles, the removed invalid edges are saved, and after a period of time, it is checked whether the original invalid edges are valid again. In order to ensure the running efficiency of the algorithm, it is necessary to appropriately determine the dynamic recovery check time interval of the probability path graph. The invention performs dynamic update of the probability path graph after the agent reaches a certain sampling point position on the global probability path graph. If the probability path graph has changed, i.e. the original invalid edge is valid again, it is necessary to first perform the re-planning and generation of the guidance path at the global level. If a better global guidance path can be found, it is necessary to further re-plan and generate the local motion path according to the local constraint preferences at the local level. During the dynamic recovery and update process of the entire probability path graph, since the original invalid edges are preserved, and the dynamic recovery and update calculation of the probability path graph is only performed when the agent reaches a certain state point on the global low-resolution probability path graph, the running efficiency of the entire algorithm framework will not be significantly affected.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This humanized two-layer motion planning method includes steps such as mathematical modeling of the problem, representation of high-level spatiotemporal logical task constraints, generation of global probabilistic path graphs, generation of global diversified guided paths, generation of local high-resolution probabilistic path graphs, generation of local diversified motion paths, and replanning of motion paths under dynamic obstacle conditions. In contrast, traditional methods often require local paths to deviate as little as possible from the global path. The planning of local paths is largely dependent on the overall path planning. When encountering local obstacles, the local path has little autonomy for temporary avoidance, and the high coupling between global and local planning greatly reduces the flexibility of the overall planning and the realism of the final result. This method, however, only uses the global path as a directional guide, and then flexibly adjusts the local path according to the actual situation. It can either follow the global path completely or completely regenerate the path, adjusting flexibly according to the specific situation. Through the two-layer approach, the multi-resolution perception of information, the multi-information domain of perception content, and the cooperative relationship between the two layers, it effectively solves the core problem of the excessively tight coupling between local and global paths, and constructs a two-layer motion planning algorithm framework of global directional guidance and local dynamic processing. Drawing inspiration from the global-local decoupling approach in human motion planning, this method divides path planning into two layers: First, based on a low-resolution scene description model, a coarse, directional global path is generated. This global path serves only as a global directional guide; consistency between the actual local motion path and the global path is not required beyond the overall directional guidance. Second, based on a high-resolution local scene description model, the current actual motion path is generated. Its advantage lies in the fact that the actual planning result, in terms of the local motion path, can satisfy the directional requirements of the global task while also being flexibly adjusted based on the current dynamic scene information. This ensures the flexibility of the algorithm framework, enables rapid processing of dynamically changing scenes, and improves the flexibility and solution efficiency of traditional motion planning algorithms. Attached Figure Description

[0044] Figure 1 This is the overall structural frame of the present invention;

[0045] Figure 2 For the present invention The corresponding DFA;

[0046] Figure 3 For the present invention The corresponding DFA;

[0047] Figure 4 This is a schematic diagram of the global probability path of the present invention;

[0048] Figure 5 This is a schematic diagram of the tunnel width determination method of the present invention;

[0049] Figure 6 This is a schematic diagram illustrating the motion scene environment type constraints of the present invention;

[0050] Figure 7 This is the simulation scenario and global guidance path for this invention;

[0051] Figure 8 This is the global boot path after adding global constraints in this invention;

[0052] Figure 9 This invention generates localized personalized motion paths.

[0053] Figure 10 This is a possible result of the local motion path replanning in this invention;

[0054] Figure 11 This refers to the dynamic obstacle influence area and invalid edge of the present invention;

[0055] Figure 12 This describes the dynamic recovery and update process of the probability path graph in this invention. Detailed Implementation

[0056] 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.

[0057] Please see Figure 1 This invention provides a technical solution: a human-like two-layer motion planning method with multiple information domains and multiple resolutions, including the following steps: mathematical modeling of the problem, representation of high-level spatiotemporal logical task constraints, generation of global probability path graphs, generation of global diversified guidance paths, generation of local high-resolution probability path graphs, generation of local diversified motion paths, and replanning of motion paths under dynamic obstacle conditions.

[0058] Motion planning algorithms can be divided into two key steps: First, a scene description model needs to be defined. Based on the appropriate description of various geometric relationships in the motion scene, a connectivity graph between the free reachable spaces of the motion scene is constructed. Then, based on the connectivity graph, the specific motion path is generated.

[0059] Traditional methods for constructing connectivity graphs for motion scenes can be divided into two main categories. The first is a direct description approach, which utilizes various computer graphics algorithms (such as cell meshes, triangular patches, and Voronoi diagrams) to decompose complex motion scenes into interconnected basic free reachable spaces (cell decompositions), and then directly describes them based on geometric connectivity, such as various navigation mesh algorithms. The second is a random sampling approach, which constructs a roadmap by randomly sampling the motion scene to indirectly describe spatial geometric connectivity, such as probabilistic path graphs (PRM) and fast randomized tree expansion (RRT). However, for complex dynamic scenes, in addition to static geometric connectivity, it is necessary to fully describe various dynamic changes. Recently, various improved algorithms have been proposed to simultaneously describe information such as gaps between obstacles and spatiotemporal changes, but it is difficult to achieve a perfect description of all information based on a single specific description model. To improve the adaptability to dynamic changes in the scene (such as dynamic obstacles and cooperation and avoidance among multiple agents), current motion planning algorithms are usually decoupled into two parts: global navigation and local navigation.

[0060] Based on the scene description model's depiction of the connectivity relationships between different reachable regions in the motion space, the global planning algorithm directly generates a global planning path using intelligent search algorithms such as A* or Dijkstra's algorithm, based on the connectivity graph. Since global planning does not consider dynamic changes in the scene, local planning further adjusts the global path locally based on the current scene's changing information.

[0061] The purpose of this solution is not to improve the PRM algorithm, but to develop a human-like coarse-to-fine motion planning algorithm. The underlying scene description model is multi-resolution; for distant locations, only a low-resolution map and corresponding low-resolution PRM are needed, with PRM serving as a tool. During global path planning, high resolution is not required. When moving to a local area, the local motion path is dynamically adjusted based on the current situation and specific constraints. The key is to first generate a coarse global path using a low-resolution PRM, and then dynamically adjust local responses based on this global path during the movement to address dynamically changing scene problems.

[0062] Overall Approach

[0063] The overall framework of the hierarchical motion planning algorithm proposed in this scheme is as follows: Figure 1 As shown.

[0064] The algorithm is generally divided into two levels: global and local. At the global level, firstly, the high-level task constraints are described in a regularized manner to automatically obtain the arrival sequences of all target regions that meet the task requirements. Then, the motion scene is randomly sampled to generate a global low-resolution PRM. Finally, based on the arrival sequences of all target regions that meet the task requirements and the global PRM, an optimal collision-free path connecting the initial region and all target location regions is generated, i.e., the global guidance path. In addition to meeting the high-level spatiotemporal logic task requirements, the global guidance path can also incorporate various global preference constraints to ensure its diversity.

[0065] At the local level, sampling points are first added as needed between each stage of sub-tasks (between the currently reached sub-task area and the next sub-task target area) based on the specific situation (whether there are local preference constraints) to generate a local high-resolution probabilistic path map; then, based on the agent's own individual characteristics and motion preferences, a diversified local motion path connecting the currently reached area and the next sub-task target area is searched and generated.

[0066] The proposed algorithm exhibits a hierarchical nature in two aspects: firstly, it is hierarchical across different information domains, combining global high-level task constraints with local individual preference constraints. Different types of task constraints can be defined at both the global and local levels, ensuring the diversity of path results; secondly, it is hierarchical in resolution, combining low-resolution global levels with high-resolution local description models. This partially solves the problem of traditional probabilistic path graph algorithms lacking global guiding directional information due to random selection of sampling points, reducing the difficulty of solving the path generation problem and improving the algorithm's operating efficiency.

[0067] If dynamic obstacles arise during path planning, the process begins at the local level. Based on the obstacle's current and future possible locations, edges on the probabilistic path graph that might collide with the obstacle are removed. The path is then replanned based on the pruned probabilistic path graph. If, in some cases, removing certain edges from the probabilistic path graph fails to find a local path to the target region of the current subtask, the global guiding path needs to be replaced at the global level. This combination of dynamic adjustments to the motion paths at both the local and global levels ensures the algorithm's dynamic adaptability and flexibility.

[0068] Specifically, it includes the following steps:

[0069] Problem Description

[0070] Assume the motion scenario is W, which contains m obstacle regions O. obs ={O1…O m}, the free configuration space excluding obstacles is represented by C freeThis means that W = C free +O obs Within the motion scene, there are n non-adjacent task target regions R = {R1, ..., R2}. n},satisfy The intelligent agent is now required to perform tasks according to certain high-level spatiotemporal logic requirements T. R (For example, in addition to reaching the target position G, it is also necessary to pass through the task areas R1, R2, and R3 first, and further specific requirements can be made on the arrival order of the task areas R1, R2, and R3), and a collision-free optimal motion path is planned.

[0071] The arrival order of a target area that meets the task requirements can be represented as Π. i ={π i1 …π in}, where π i This indicates that the agent has reached the target region R. i i1…in are to satisfy the specific task constraints T R A specific sequential arrangement of 1…n. (Except for the high-level spatiotemporal logic constraint T) R In addition, various preference constraints can be imposed on the agent during the path generation process. i (i = 1…k) to generate a personalized motion path that differs from other intelligent agents. Therefore, the problem to be solved in this scheme can be qualitatively described as: generating an optimal collision-free path that completes a task objective with time-constrained logic and satisfies the requirements of multiple information domains, including scene constraints and its own personalized constraints.

[0072] Representation of high-level spatiotemporal logic task constraints

[0073] For high-level spatiotemporal logic task constraints, this solution adopts a separate description approach for temporal logic requirements and spatial constraints. Spatial constraints are described using a combination of global high-resolution PRM and local low-resolution PRM, while temporal logic requirements are first described using regular expressions and then converted into a Deterministic Finite Automata (DFA) for unified characterization. For example, tasks arrive sequentially in task regions R1, R2…, R… n The temporal logic requirements can be described using a regular expression as follows:

[0074] φ seq (π1,π2,…π n )=π1π2…π n

[0075] In the formula π i This indicates that the target area R has been reached. i Similarly, it is only required to reach regions R1, R2, ..., R nHowever, the order is not limited, and it can be represented as:

[0076]

[0077] Where perm(1,n) represents all possible permutations of {1,…,n}.

[0078] For alternating arrival areas and The situation can be represented as:

[0079]

[0080] More complex high-level task constraints can be described by combining the above basic rule expressions, such as arriving at the region in any order. Then arrive at the area in any order. It can be represented as:

[0081]

[0082] The formula can be further extended to the case where the task objective is divided into several groups, and different groups are reached sequentially, which can be expressed as:

[0083] φ groups (π1,π2,…π n )=φ cov (G1)φ cov (G2)…φ cov (G k )

[0084] Among them, G1, G2, ... G k This indicates that the task {π1, π2, ..., π} will be assigned to the task {π1, π2 n Divide into k groups.

[0085] Once the high-level temporal logic task requirements are described using basic rule expressions, standard algorithms from automata theory can automatically convert these rule expressions into a DFA model, providing all possible temporal logic paths, such as... and They can be described as follows: Figure 2 and Figure 3 As shown in the figure.

[0086] In the diagram, z i This indicates the completion status of a specific task during the movement process, such as Figure 5 In this context, z4 indicates that the target region R4 has only been reached, while z 14 This indicates that the target regions R2, R3, and R4 have been reached. Each task completion path Z from the initial state to the final state... l ={z0,…,z g} corresponds to a possible order of arrival at a target region that meets the temporal logic requirements, where l = 1, ..., n, n is the number of all possible paths, z0 is the task's initial state, z g The task is now complete.

[0087] Generation of global probabilistic path graph

[0088] This scheme directly uses the basic Probabilistic Path Recognition (PRM) algorithm to generate the global motion path graph. The probabilistic path graph is represented by an undirected graph (V, E, COST), with the node set V = (q1…q...). n node q in ) i According to a certain probability (such as a uniform random distribution), they are randomly selected from free space C. free Select one edge (q) from edge set E. i ,q j ) represents a connection configuration q generated by the local planner LP. i and q j To generate collision-free paths, this scheme uses straight-line connections, where COST represents the cost of a specific edge. During global path graph generation, sampling is first performed in the target regions of each subtask, and then sampling is performed in the free configuration space excluding the target regions. This ensures that a probabilistic path graph connecting all target regions is ultimately generated. The pseudocode for the global PRM generation process is as follows:

[0089]

[0090]

[0091] In the above procedure flow, the initial path graph G = (V, E) is empty. Steps 03 to 06 aim to generate sampling points in the target region of the sub-project. Steps 07 to 13 randomly select N from free space. node This process involves finding nodes, thus constructing the node set V. Then, for each node q belonging to V, step 15 aims to select the k node from V that is closest to q based on a distance function. neighbor_node There are nodes, this k neighbor_node The nodes constitute the set Nk of neighboring nodes of q. q The purpose of steps 16-20 is to utilize the local planner to solve the problem between q and Nk. q Edges are constructed between nodes to complete the path map. This scheme uses a simple and fast line planner to connect sampled points and complete the path map construction.

[0092] Generation of global diverse boot paths

[0093] After obtaining the global probabilistic path graph (PRM), as follows:Figure 4 As shown, if certain weights are defined for each edge, a globally optimal guided path Q = {q0, q1, ... q} can be generated based on various search algorithms. k ,q g}, where q0 is the starting position of the global boot path on PRM, q g The global boot path ends on PRM, {q0,q1,…q k ,q g} represents the path on the PRM that satisfies the temporal logic constraints and has the minimum cost.

[0094] This scheme uses Dijkstra's shortest path algorithm to generate the optimal path from the initial region to the target region. During optimization, the path cost can be simply represented by the Euclidean distance between corresponding nodes in the probabilistic path graph (PRM). Furthermore, global constraints are introduced, such as the degree of congestion and danger in different regions, to generate a preferred global guidance path. This scheme describes the constraints as path weights and calculates the path weights between two nodes (q...). i ,q j Path cost between:

[0095] COST(q i ,q j ) = w i ×l1(q i ,q k )+l2(q k ,q j )

[0096] In the formula, l(q) i ,q j )=l1(q i ,q k )+l2(q k ,q j ) represents node q i and q j The Euclidean distance between them, l1(q) i ,q k ) represents the straight line l(q) i ,q j Located at constraint C i The portion within the affected area, l2(q) k ,q j ) represents the straight line l(q) i ,q j Located at constraint C i The part outside the scope of influence, w i (w i >0) represents a global constraint C i The weighting factor represents the constraint C. iThe impact on path costs within its influence range. i When the value is greater than 1, the path cost increases, indicating that the constraint requires the virtual individual to move away from the current region; when the value is less than 1, the path cost increases. i When the value is less than 1, the path cost is reduced, indicating that the constraint requires the virtual individual to prefer the current region. For the simple case, in constraint C... i Within the scope of influence, w i The value of can be a constant, remaining unchanged, or it can be a variable, having different values ​​in different regions. If a path segment is affected by multiple constraints simultaneously, simply multiply the weight factors of the different constraints together, i.e. Then, the path cost is calculated using the weighting factor w, which represents the combined effect of different constraints. After obtaining the new path costs for all affected edges, the preferred path that satisfies different global constraints can be obtained by selecting the weighted shortest path during global path generation.

[0097] When generating the global bootstrap path, both the temporal logic constraint graph (DFA) and the global probabilistic path graph (PRM) must be satisfied. First, all possible task completion paths Z need to be obtained based on the DFA. l (l=1,…n), and then, based on the global probability path graph, find each task completion path Z. l The optimal path Q on the global probabilistic path graph PRM l Finally, in all Q l Find the path with minimum cost Q = min(Q1, ..., Qn) in the subset (l = 1, ..., n). n (This serves as the final global boot path.)

[0098] like Figure 3 As shown, if the requirement is to start from the initial position and reach the target area R... g Beforehand, you must reach sub-task areas R1, R2, R3, and R4. The order in which they are reached is not important. Therefore, you can start by... Figure 11 Generate all possible task completion paths, such as:

[0099] Z1 = {z0, z1, z5, z} 11 ,z g}={π1,π2,π3,π4}, Z2={z0,z2,z5,z 11 ,z g There are 24 paths in total, such as {π2, π1, π3, π4}. Then, based on each specific task completion path, a specific global motion path is generated in the global PRM, such as according to... Figure 3 It can be seen that the global motion path corresponding to Z1 is:

[0100] Q1 = {q0, q1, q7, q8, q 10 ,q 11 ,q 13 ,q 14 ,q 18 ,q 20 ,q 21 ,q 22 ,q 23}

[0101] The global motion path corresponding to Z2 is:

[0102] Q2={q0,q2,q8,q7,q 12 ,q 13 ,q 14 ,q 18 ,q 20 ,q 21 ,q 22 ,q 23}

[0103] After obtaining all feasible global motion paths Q l After (l=1,…24), find the path with the minimum cost Q=min(Q1,…,Q 24 This is the final determined global boot path.

[0104] In the above calculation process, depending on the different global task constraints, although only the path with the lowest cost is selected as the final global guiding path, the entire algorithm framework generates and saves all other possible paths (Q1,…,Q). n These candidate path results can ensure the speed and flexibility of the dynamic path replanning process in handling dynamic scenarios.

[0105] Diverse motion path generation at the local level

[0106] In practice, when generating specific motion paths, in addition to satisfying various high-level basic task constraints, agents often need to generate diverse local paths based on their own preferences. Therefore, to further improve the realism and personalization of different individual motion trajectories, between stage sub-tasks (any two sub-target regions arrived at sequentially), it is necessary to further generate diverse local motion paths based on the virtual individual's preference constraints and using global guidance paths.

[0107] Generation of local high-resolution probabilistic path maps

[0108] When considering local preference constraints, it is necessary to improve the accuracy of the scene description model to fully express the local constraints. Therefore, in local path generation, this scheme adopts a multi-resolution approach, increasing the local resolution of the probabilistic path graph by adding local sampling points. Since a global guiding path has already been generated, new sampling points cannot be generated simply by randomization; the directional constraint information provided by the previous global guiding path should be fully utilized. Based on this, this scheme introduces the concept of a tunnel, expanding the guiding path from a one-dimensional straight line into a two-dimensional or three-dimensional safe tunnel. Using the original global guiding path as the central axis, the tunnel expands outward with a certain width, randomly adding sampling points within the tunnel area to generate a high-resolution local probabilistic path graph. By using the global guiding path as a basic directional constraint for the movement paths between stage subtasks, different agents within the tunnel area can generate diverse movement trajectories according to their own preferences while ensuring that the movement direction is basically consistent with the global guiding path. Therefore, to ensure the real effectiveness of the local path and the efficiency of the algorithm, it is crucial to determine an appropriate tunnel width.

[0109] If the node q in the global boot path i q j satisfy

[0110] l(q i ,q j )<r(q i )+r(q j )

[0111] In the formula, l(q) i ,q j ) is a straight path (q) i ,q j The length of r(q) i ) represents the global sampling point q i The distance to the nearest obstacle can prove the connection q. i q j The straight line l(q) i ,q j It must be a collision-free path. For example... Figure 3 As shown, the line l(q2,q3) satisfies the formula, so it is definitely a collision-free path. In this case, we can define line l(q2,q3) as having sufficient sampling. However, the line l(q1,q2) does not satisfy the above formula, so we consider the path (q1,q2) to be undersampled. That is, we cannot directly determine whether l(q1,q2) is a collision-free path. If it does not satisfy the above formula, we can first... i q j Add sampling point q at the middle position ij And judge (q) in turn. i ,qij ), (q ij ,q j If the above formula is not satisfied, then continue adding sampling points between the two sampling points on the global path until the formula is satisfied. Figure 5 As shown, a sampling point q is added between (q1, q2). 12 Afterwards, the straight path l(q1,q2) becomes a fully sampled path, and the straight path (q1,q2) becomes a fully sampled path. 12 ) and (q2,q 12 The tunnel width can be determined using the formula.

[0112] Based on the above definition and judgment conditions for sufficient sampling, this scheme proposes an adaptive method for determining tunnel width. Two points q on the global path... i q j Once the sufficient sampling condition is met, the tunnel width can be approximated using the following formula:

[0113]

[0114] In the formula, x represents the straight line l(q) i ,q j A point on ). If the tunnel width is limited using the w(x) value obtained above, it can be guaranteed that the vast majority of new sampling points are located in free space C. free Inside.

[0115] After increasing the sampling points on the straight path to obtain an adaptive tunnel width, the sampling points can be further increased within the tunnel width range determined by the formula using a uniform random sampling method to generate a locally high-resolution probability path map.

[0116] Local Diversified Motion Path Generation

[0117] After obtaining a local high-resolution probabilistic path map, a method similar to the global constraint description can be used during local path generation. This involves adding locally defined preference constraints and changing the weights of different paths to obtain different local motion paths. During local path generation, this scheme adds environment type constraints to different paths, such as trails, grasslands, dirt roads, and forests. Figure 4 As shown, the path (q) traverses two different environment types. i ,q j The cost of ) can be expressed as:

[0118] COST(q i ,q j )=w1l1+w2l2

[0119] Where w1 and w2 are the agent's preference weights for different environment types, and l1 and l2 are the path (q) weights.i ,q j The length of the path in the corresponding environment type. If two different individuals have different weight preferences for the same environment type, different local motion paths can be generated.

[0120] After obtaining the weights of different edges in the high-resolution motion path graph between stage subtasks, an optimal motion path connecting the current position and the target region of the stage subtask can be generated based on Dijkstra's shortest path algorithm. Although different agents use the same probabilistic path graph, the weights of the same edges in the same probabilistic path graph are directly determined by the specific agent's individual requirements. The weights change as the agent's requirements change. Therefore, diverse local paths can be generated based on the individual requirements of different agents.

[0121] By introducing the concept of a tunnel, we can ensure that the local guidance path is basically along the direction of the global guidance path, which solves the problem of lack of global guidance direction information when randomly selecting sampling points in the generation of high-resolution probabilistic path maps. On the other hand, the global guidance path can be extended from a one-dimensional straight line into a two-dimensional or three-dimensional safe tunnel, which further improves the flexibility and diversity of local motion path generation.

[0122] Movement path replanning under dynamic obstacle conditions

[0123] As the agent moves forward along the planned path, unforeseen dynamic events may occur during the path generation process, such as the sudden appearance of moving obstacles or other agents on the planned path. In such cases, it is necessary to replan the current movement path to achieve intelligent avoidance of dynamic obstacles.

[0124] Replanning of movement paths

[0125] From the perspective of motion path planning and generation, the impact of dynamic obstacles lies in invalidating certain edges on the probabilistic path graph that could potentially collide with the dynamic obstacle. Therefore, it is first necessary to remove edges on the probabilistic path graph that may collide with the dynamic obstacle based on the current and future possible location range of the obstacle, generate a new probabilistic path graph, and then re-search and generate motion paths. For example, in... Figure 2 On the probability path graph shown, if the current agent is at q 12 The sub-task area to be reached is R3. If the dynamic obstacle is located on edge l(q) 12 ,q 13 At this point, remove it and, at the local level, re-plan and generate a path from q. 12 The local path to subtask target region R3. Based on... Figure 2 As shown, two candidate paths can be obtained: q 12 →q 11 →q13 and q 12 →q 15 →q 13 To further improve the environmental adaptability of the re-planned path generation, this scheme further predicts the future positions of dynamic obstacles, thus avoiding various dynamic obstacles in advance. Let the current state of the dynamic obstacle be S = (x, v, t), where x represents the current position of the dynamic obstacle, v represents the current velocity of the dynamic obstacle, and t represents the current time. After a time interval Δt, the position of the dynamic obstacle may be:

[0126] S'=(x+(v+Δv)×Δt,v+Δv,t+Δt)

[0127] In the formula, Δv = {±α} represents the range of possible velocity values ​​of the dynamic obstacle, and the possible velocity direction changes of the dynamic obstacle can be further defined as follows: Based on the above formula, a semi-circular potential area of ​​dynamic obstacles with the current velocity direction as the central axis can be obtained. Edges in the probability path graph within this area are then removed as invalid edges, and the local motion path is re-planned and generated. For example... Figure 4 As shown, if the dynamic obstacle is on edge l(q) 12 ,q 13 ) up, and towards q 11 Move in the direction of, except for removing the currently invalid edge l(q) 12 ,q 13 In addition, it is also necessary to calculate the possible range of motion of the dynamic obstacle in the future based on the above formula, since the dynamic obstacle is moving towards q at this time. 11 The direction of movement is so that edge l(q) moves. 12 ,q 11 The edge ) might become invalid, therefore the result of replanning the motion path should be q. 12 →q 15 →q 13 .

[0128] In some cases, if removing invalid edges from the probabilistic path graph fails to find a new path from the current position to the subtask target region at the local level, it is necessary to re-plan the path at the global level to generate a new global guidance path. Since all possible paths have been obtained from the DFA model graph during the global guidance path generation process, a new guidance path can be generated quickly. For example, in the example above, the agent is located at q... 12 At this point, according to the DFA model shown in the figure, the agent is in the task-completed state z5. In the current situation, there are only two global motion paths to choose from: the original path z5→z 11Since the effects of dynamic barriers cannot be achieved, the global boot path should be changed to z5→z 12 In other words, the agent needs to reach the subtask target region R4 first, instead of the originally planned R3.

[0129] In summary, when a dynamic obstacle appears, this scheme first calculates the invalid edges on the probability path graph by the area that the dynamic obstacle may affect and removes them. Then, it first re-plans and generates the motion path at the local level. If an effective motion path cannot be obtained, it then re-plans and generates the guiding path at the global level, thus ensuring the adaptability and computational efficiency of the motion path dynamic planning results.

[0130] Dynamic recovery and update of probabilistic path graph

[0131] In reality, the impact of various dynamic obstacles in virtual scenes is often temporary. That is, invalid edges removed in the previous stage may become valid again after a period of time as the obstacles move to other locations. Based on this consideration, during obstacle avoidance, removed invalid edges are saved, and after a period of time, their validity is re-checked. To ensure the algorithm's efficiency, the dynamic recovery check interval of the probabilistic path graph needs to be appropriately determined. If the interval is too short, not only will the system's efficiency decrease, but the changes in invalid edges may be minimal; if the interval is too long, the optimality of the agent's motion planning results may not be guaranteed. Therefore, according to the overall algorithm framework, the probabilistic path graph is dynamically updated after the agent reaches a sampling point on the global probabilistic path graph. If the probabilistic path graph changes, i.e., the original invalid edges become valid again, the guiding path needs to be re-planned and generated at the global level first. If a better global guiding path can be found, the local motion path needs to be re-planned and generated at the local level based on local constraints and preferences.

[0132] During the dynamic update and recovery process of the entire probabilistic path graph, since the original invalid edges are retained and the dynamic recovery and update calculation of the probabilistic path graph is only performed when the agent reaches a certain state point on the global low-resolution probabilistic path graph, the running efficiency of the entire algorithm framework is not significantly affected.

[0133] The above method will then be verified through simulation: See [link / reference] Figures 7 to 12 :

[0134] The simulation experiment scenario is shown in the figure. If the temporal logic requirement is... Based on temporal logic constraints and the global probability path graph, the globally optimal guiding path is:

[0135] Q1 = {q0, q2, q8, q7, q...}12 ,q 13 ,q 14 ,q 16 ,q 19 ,q 21 ,q 23 ,q 27}

[0136] If global blocking level constraints are added during the global path cost calculation process, as shown in the figure, the globally optimal boot path obtained is:

[0137] Q2={q0,q2,q8,q7,q 12 ,q 13 ,q 14 ,q 17 ,q 18 ,q 20 ,q 22 ,q 23 ,q 27}

[0138] Furthermore, if individual terrain preference constraints, such as grassland (w), are added between stage subtasks R3 and R4,... g ), desert (w d ), forest (w f Different terrains, such as those shown in the figure, require the first step of configuring the global guide path q. 14 →q 17 →q 18 →q 20 →q 22 →q 23 On both sides, sampling points are added within the tunnel area determined by the formula. If a certain agent selects w g =1.5, w d =2, w f =3, and the final local motion path is as follows Figure 7 As shown by the arrow in the image, the virtual individual prefers grassland and desert terrain, avoiding forest terrain. If other virtual individuals have different preferences, personalized movement paths that meet specific preference requirements can be easily generated based on the specific weights. When constructing a local high-resolution probabilistic path map, the original global movement path sampling points and their interconnections can still be utilized, such as... Figure 7 As shown, the long path q is directly used when generating the local motion path. 20 →q 22 and q 22 →q 23 This improves the path generation efficiency of the overall algorithm framework.

[0139] If along the local path (q) 14,q 17 ,q 18 ,q 20 ,q 22 ,q 23 During the movement, a dynamic obstacle occurs. Under the current circumstances, according to... Figure 5 The DFA model shown has the agent already reached task state z. 11 Since only the sub-task target region R4 has not been reached, the motion path can only be re-planned and generated at a local level. Figure 8 The dashed lines in the diagram represent all possible specific motion path directions. If the agent reaches state q... 12 At times, such as Figure 9 The dynamic obstacles shown in the diagram are represented by semicircles, and their potential impact areas need to be removed first. Figure 9 Invalid edges are represented by dashed lines, and then new motion paths are re-planned and generated at the local level (q). 12 ,q1,q4,q6,q 14 In the current situation, since a locally effective motion path can be obtained, there is no need for global-level replanning. However, if more dynamic obstacles appear, for example... Figure 9 The complex situation shown cannot currently be found at the local level from q 12 The feasible path to the subtask target region R3 requires further global path replanning and generation at the global level. According to... Figure 5 The DFA model shown, in the current situation, first reaching the subtask target region R4 and then returning to R3 becomes the only option. Clearly, this is not an ideal result compared to the original global guidance path, requiring more time. However, according to the probabilistic path graph dynamic recovery and update mechanism proposed in this chapter, when the agent reaches q4 along the new global guidance path, the previously invalid edges l(q4,q6) and l(q5,q6) will become valid again. Therefore, the replanning process will be restarted, and the agent will find that reaching the subtask region R3 first at the global level is a better choice. The agent will then further adjust the generated local path (q4,q6,q6) based on the generated local path (q4,q6,q6). 14 First, it reaches the subtask region R3, ensuring the dynamic adaptability of the entire algorithm framework.

[0140] Therefore, the superiority of this solution has been fully verified.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A human-like two-layer motion planning method with multiple information domains and multiple resolutions, characterized in that, Includes the following steps: Step 1: Mathematical modeling of the problem; assume the motion scenario is as follows: Among them are Obstacle areas Free configuration space without obstacles It means, that is There are non-adjacent task target areas within the motion scene. ,satisfy , ( The target task of the intelligent agent is: to meet certain high-level spatiotemporal task logic requirements. Plan a collision-free optimal motion path; The arrival order of a target area is represented as follows: This indicates that the agent has reached the target area. , To meet specific task constraints One The specific order in which they are arranged; To impose various preference constraints on the agent during the path generation process based on the actual path conditions; Step Two: Representation of High-Level Spatiotemporal Logic Task Constraints; This step employs a separate description approach for temporal logic requirements and spatial constraint requirements. Temporal logic requirements are described using regular expressions, while spatial constraint requirements are described using a combination of global and local PRM approaches. Specifically, this includes: 1) Arrive at the mission area in sequence; only arrival at the area is required. The temporal logic requires that it be described first using regular expressions: In the formula This indicates that the target area has been reached. ; ,in, express All possible permutations; 2) For alternating arrival areas and The situation is represented as follows: ； 3) More complex high-level task constraints can be described by combining the basic rule expressions from steps 1) and 2) above, such as arriving at the region in any order. Then arrive at the area in any order. , can be represented as: ； （1） Formula (1) can be further expanded to represent the case where the task objective is divided into several groups, and different groups are reached in sequence, as follows: ； （2） in, Indicates the task Divided into Groups; Then it is converted into the form of DFA (Deterministic Finite Automata) to uniformly characterize it; giving all possible temporal logic paths; Step 3: Generation of the global probabilistic path graph; The basic PRM algorithm is used to generate the global motion path graph; When generating the global path graph, it is first necessary to sample in the target regions of each subtask, and then sample in the free configuration space excluding the target regions, so as to ensure that a probabilistic path graph connecting all target regions can be generated in the end. Step 4: Generation of globally diverse boot paths; After obtaining the global probabilistic path graph, when generating the global guiding path, the temporal logic constraint graph must be satisfied simultaneously. and global probability path graph ; First, we need to describe the temporal logic constraint diagram. Get all possible task completion paths Then, based on the global probability path graph, find the completion path for each task. In the global probability path diagram The corresponding optimal path Path cost is obtained through the path cost availability probability path graph. The Euclidean distance between corresponding nodes is simply represented, and a global constraint is further introduced. By describing the constraint as a path weight, the path cost is modified, thereby generating a preferred globally guided path. Finally, based on various search algorithms, in all... Find the path with the minimum cost. As the final global boot path, the globally optimal boot path ,in For the global boot path in The starting position on For the global boot path in The end point on the way, for The path that satisfies the temporal logic constraints and has the lowest cost; Step 5: Generation of local high-resolution probabilistic path graph; After considering local preference constraints, improve the accuracy of the scene description model and achieve full expression of local constraints; Introduce the concept of a tunnel, expand the guidance path from a one-dimensional straight line into a two-dimensional or three-dimensional safe tunnel, take the original global guidance path as the central axis, expand outward with a certain width, randomly add sampling points within the tunnel range, and generate a high-resolution local probabilistic path graph. By using the global guidance path as a basic directional constraint for the movement path between stage sub-tasks, the agent can generate diverse movement trajectories according to its own preferences within the tunnel range, while ensuring the overall consistency of the movement direction with the global guidance path. Specifically, it includes: 1) First, determine the nodes at both endpoints of the current global bootstrap path segment. , Whether the sampling is sufficient, i.e., whether it satisfies... , It is a straight path Length, Represents global sampling points If the sampling is insufficient for the distance to the nearest obstacle, add sampling points on the current path segment until sufficient sampling is achieved; 2) According to the formula Obtain the tunnel width. Represents a straight line A point on; 3) Within the tunnel width range determined by the formula, the sampling points are further increased using a uniform random sampling method to generate a local high-resolution probability path map; Step Six: Generation of Local Diverse Motion Paths; After obtaining the local high-resolution probabilistic path map, the local path generation can be performed using the same method as the global constraint description. By adding locally defined preference constraints and changing the weights of different paths, different local motion paths can be obtained. Step 7: Replanning the motion path under dynamic obstacle conditions; After a dynamic obstacle appears, invalid edges on the probability path graph are first obtained by calculating the area that may be affected by the dynamic obstacle and then removed. Then, the motion path is replanned and generated at the local level. If an effective motion path cannot be obtained, the guiding path is replanned and generated at the global level, which ensures the adaptability and computational efficiency of the motion path dynamic planning results. Specifically, the following steps are included: 1) Replanning of the motion path; First, based on the possible location range of the dynamic obstacle in the current and future period, remove the edges on the probabilistic path graph that may collide with the dynamic obstacle, generate a new probabilistic path graph, and re-search and generate the motion path; If after removing invalid edges on the probabilistic path graph, if a new motion path from the current position to the sub-task target area cannot be found at the local level, then it is necessary to re-plan the path at the global level and generate a new global guidance path. 2) Dynamic recovery and update of the probabilistic path graph; To ensure operational efficiency, the probabilistic path graph is dynamically updated after the agent reaches a sampling point on the global probabilistic path graph. If the probabilistic path graph changes, i.e., the original invalid edges become valid again, the guiding path needs to be re-planned and generated at the global level first. If a better global guiding path can be found, the local motion path needs to be re-planned and generated at the local level according to the local constraint preferences.

2. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, The task requirements under different information domains are flexibly described at both the global and local levels, while simultaneously achieving diversity in motion planning results at both the global and local levels.

3. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, When generating the global path graph, it is first necessary to sample in the target regions of each subtask, and then sample in the free configuration space excluding the target regions to generate a probabilistic path graph connecting all target regions.

4. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, When generating the global path graph, when generating the globally optimal path from the initial region to the target region, and when generating the global guiding path, based on the temporal logic constraint graph and the global probability path graph, after obtaining all feasible global motion paths, the path with the minimum cost is found. The path cost can be simply represented by the Euclidean distance between corresponding nodes in the probability path graph, and further global constraints are introduced to generate a variety of preferred global guiding paths.

5. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, Depending on the different global task constraints, the path with the lowest cost is selected as the final global guiding path. However, the entire algorithm framework generates and saves all other possible paths, improving the speed and flexibility of the dynamic replanning process in handling dynamic scenarios.

6. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, During the generation of local high-resolution PRM, the local resolution of the probability path map is improved by adding local sampling points. When generating new sampling points, the directional constraint information provided by the previous global guidance path is used to expand the guidance path from a one-dimensional straight line to a two-dimensional or three-dimensional safe tunnel with a preset width, using the original global guidance path as the central axis. Sampling points are randomly added within the tunnel to generate a high-resolution local probability path map. By using the global guidance path as a basic directional constraint for the movement path between stage subtasks, the agent can ensure that the movement direction is consistent with the global guidance path within the tunnel, and at the same time, diverse movement trajectories can be generated.

7. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, This includes a flexible method for determining tunnel width. If the nodes in the global boot path , satisfy ； In the formula, It is a straight path Length, Represents global sampling points The distance to the nearest obstacle can prove the connection. , straight line It must be a collision-free path, in which case a straight line can be defined. The sampling is sufficient; but the straight line If the formula is not satisfied, then the path is considered a straight line. The sampling was insufficient, meaning it cannot be directly determined. Is it a collision-free path? If not, then we can first... , Add sampling points in the middle position and judge in turn. , If the formula is not satisfied, continue to add sampling points between the two sampling points on the global path until the formula is satisfied. Based on the above definition and judgment conditions, if two points on the global path... , If the sufficient sampling criterion is met, the tunnel width can be approximated using the following formula: ； In the formula, Represents a straight line If a point on the graph is obtained using the above formula... By limiting the tunnel width, it can be ensured that the vast majority of new sampling points are located in free space. Inside; After increasing the sampling points on the straight path to obtain an adaptive tunnel width, the sampling points can be further increased within the tunnel width range determined by the above formula using a uniform random sampling method to generate a local high-resolution probability path map.

8. The human-like two-layer motion planning method with multiple information domains and multiple resolutions according to claim 1, characterized in that, It also includes a dynamic recovery and update method for the probabilistic path graph. After a period of time, invalid edges will become valid edges again as dynamic obstacles move to other places. During the process of avoiding dynamic obstacles, the removed invalid edges are saved. The impact of various dynamic obstacles in the virtual scene is often temporary. After a period of time, the original invalid edges are re-checked to see if they have become valid again. After the agent reaches a certain sampling point position on the global probabilistic path graph, the probabilistic path graph is dynamically updated. If the probabilistic path graph changes, that is, the original invalid edges become valid again, the guidance path needs to be re-planned and generated at the global level first. If a better global guidance path is found, the local motion path needs to be re-planned and generated at the local level according to the local constraint preferences. By combining the dynamic adjustment of motion paths at the local and global levels, the dynamic adaptability and flexibility of the algorithm framework are ensured.

Images