An unmanned autonomous body path planning method and system based on path diversity

Abstract

This invention belongs to the field of intelligent connected and autonomous driving technology, specifically a method and system for unmanned autonomous path planning based on path diversity. It obtains all passable neighboring nodes of the current node, calculates and normalizes the comprehensive suitability of each neighboring node to obtain the probability distribution of each neighboring node being selected, and constructs a dynamic probability wheel based on this distribution. According to the probability distribution of the dynamic probability wheel, the next node is randomly selected from the neighboring nodes, and the current node is updated as the next node. The above steps are repeated until the current node is the target destination, and the path from the starting point to the target destination is output. This transforms path selection from "deterministic optimal solution" to "probabilistic random selection," reducing the predictability of the path; it solves the core defects of existing technologies, such as highly convergent paths and strong predictability.

Description

Technical Field

[0001] This invention belongs to the field of unmanned autonomous vehicle motion control technology, specifically an unmanned autonomous vehicle path planning method and system based on path diversity. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] Path planning is one of the key technologies for autonomous navigation of unmanned autonomous vehicles (such as inspection robots, logistics robots, autonomous vehicles, and drones). Current mainstream path planning algorithms, such as the A* algorithm, Dijkstra's algorithm, artificial potential field method, and various learning-based planning methods, are essentially optimization solutions. That is, given a starting point and a destination, they search for an "optimal" or "near-optimal" path based on objective functions such as distance, time, or energy consumption.

[0004] However, these methods share a common characteristic: when the starting point, ending point, and environmental map are fixed, the path results output by the algorithm are highly similar. For the same task scenario, different unmanned autonomous vehicles or multiple tasks performed by the same vehicle tend to choose almost identical paths. This high degree of path repetition makes the trajectory of the unmanned autonomous vehicle highly predictable, allowing external observers to accurately predict its route through simple observation or data analysis. Summary of the Invention

[0005] This invention provides a path planning method and system for unmanned autonomous entities based on path diversity, aiming to solve the technical defects of existing path planning algorithms, which pursue a single optimal path, resulting in highly similar and predictable paths for multiple unmanned autonomous entities or multiple tasks in the same scenario.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention discloses an unmanned autonomous path planning method based on path diversity, comprising the following steps: Obtain all accessible neighboring nodes of the current node, calculate the comprehensive fitness degree of each neighboring node and normalize it to obtain the probability distribution of each neighboring node being selected, and construct a dynamic probability roulette based on this. Based on the probability distribution of the dynamic probability roulette, the next node is determined from the adjacent nodes in a random manner, and the current node is updated as the next node; Repeat the above steps until the current node is the target destination, and output the path from the starting point to the target destination.

[0007] Furthermore, the overall fitness degree is obtained by using at least two of the following: regional feasibility function, directional feasibility function, memory activation intensity function, and node availability function, and through intersection operation.

[0008] Furthermore, the regional feasibility function is constructed as follows: a polygonal soft-constraint feasible region is constructed based on the starting point and the target endpoint; for nodes located inside the feasible region, the regional feasibility remains high and flat; for nodes close to the boundary of the feasible region, the regional feasibility decays smoothly; for nodes outside the feasible region, the regional feasibility approaches zero.

[0009] Furthermore, the directional feasibility function is constructed as follows: taking the direction from the current node to the target endpoint and the neighboring nodes of the target endpoint as the reference, the closer the travel direction of the adjacent node is to the reference direction, the higher the directional feasibility; the directional feasibility is calculated using a smooth window function, so that the allowed directional range transitions continuously.

[0010] Furthermore, the memory activation intensity function includes a short-term memory activation intensity function and a long-term memory activation intensity function; the short-term memory activation intensity function records the number of times a node is visited in a single path planning, and the short-term memory activation intensity decreases as the number of visits increases; the long-term memory activation intensity function records the cumulative frequency of visits to nodes in historical path planning, and decays and forgets according to a preset coefficient; the memory activation intensity function decreases monotonically with the increase of the number of visits.

[0011] Furthermore, the node availability function is a binary function: when a node is a faulty or dangerous node, the node availability is 0, indicating that the node is unavailable; otherwise, the node availability is 1, indicating that the node is available.

[0012] Furthermore, after completing a path planning operation, the number of short-term visits is reset to zero, and the number of long-term visits is multiplied by a preset forgetting coefficient to reduce the impact of historical paths on subsequent path planning.

[0013] A second aspect of the present invention discloses an unmanned autonomous path planning system based on path diversity, comprising: The dynamic probability roulette module is configured to: obtain all passable neighboring nodes of the current node, calculate the comprehensive fitness degree of each neighboring node and normalize it to obtain the probability distribution of each neighboring node being selected, and construct a dynamic probability roulette based on this. The node update module is configured to: determine the next node from the adjacent nodes in a random manner according to the probability distribution of the dynamic probability roulette, and update the current node as the next node; The iterative output module is configured to repeat the above steps until the current node is the target endpoint, and output the path from the starting point to the target endpoint.

[0014] A third aspect of the present invention discloses a computer program product including computer-readable instructions, which, when executed on an electronic device, enable the electronic device to implement the aforementioned unmanned autonomous path planning method based on path diversity.

[0015] A fourth aspect of the present invention discloses an electronic device, including at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor being used to execute the computer program, enabling the electronic device to implement the aforementioned unmanned autonomous path planning method based on path diversity.

[0016] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: The path selection process is transformed from "deterministic optimal solution" to "probabilistic random selection." At each decision node, each candidate node is assigned a different probability of being selected based on its overall fitness, and the next node is randomly selected under this probability distribution. Because randomness is introduced in each selection, even if the starting point and the destination are the same, the paths generated by multiple planning processes exhibit a high degree of diversity globally, effectively reducing the predictability of the paths. However, by guiding the random walk through dimensions such as memory constraints in the overall fitness, it is ensured that the path always moves towards the target destination, avoiding problems such as excessive detours and deviations from the target caused by pure random walks. This improves the diversity of paths while ensuring the accessibility and reliability of the task, solving the core defects of existing technologies, such as highly similar paths and high predictability. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 A flowchart illustrating a path planning method provided in one or more embodiments of the present invention; Figure 2 A schematic diagram of the regional feasibility function provided for one or more embodiments of the present invention; Figure 3 A schematic diagram of the directional feasibility function provided for one or more embodiments of the present invention; Figure 4 A schematic diagram of the memory activation strength function provided in one or more embodiments of the present invention; Figure 5 A schematic diagram of a normal path provided for one or more embodiments of the present invention; Figure 6 The path and heatmap provided for when some nodes are unavailable in one or more embodiments of the present invention; Figure 7This is a schematic diagram of path entropy analysis provided for one or more embodiments of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The reason why existing path planning techniques suffer from highly convergent paths lies in the fact that existing methods model path planning as a single-objective optimization problem. Under fixed starting and ending points and environmental constraints, it inevitably converges to a unique or a few "optimal solutions." This singular output logic leads to the defects of path rigidity and high predictability at the ontology level, allowing malicious entities to observe and predict the route and carry out interference or information interception at specific nodes. Furthermore, in situations where paths are highly fixed, unmanned autonomous entities repeatedly pass through the same nodes and road segments. If applied to communication-sensitive environments (such as inspection tasks in classified areas), a large number of information packets repeatedly passing through the same nodes may lead to data interception by communication nodes on fixed routes, resulting in the leakage of critical communication data and thus reducing information security.

[0022] Therefore, this solution provides a path planning method and system for unmanned autonomous agents based on path diversity. It abandons the approach of finding a single optimal path and constructs a "controllable random walk under soft constraints" mechanism. By constructing a dynamic probability roulette, it integrates multi-dimensional constraints such as the feasible region, the direction of travel, and the historical visit memory of nodes into a probabilistic selection criterion. This allows the unmanned autonomous agent to naturally generate highly diverse path distributions while moving towards the destination, thus fundamentally breaking the path rigidity and solving the aforementioned problems.

[0023] The "controllable random walk under soft constraints" mechanism does not rely on strict geometric restrictions. Instead, it constructs a flexible feasible region through region constraints, direction constraints, and memory constraints. This ensures the subject moves towards the destination while providing sufficient path selection space, enabling controllable walks, increasing path diversity and unpredictability, and providing a foundation for counter-information theft. The feasible region exhibits smooth boundary transitions, without abrupt changes or path jumps, ensuring the stability of the unmanned autonomous subject's movement.

[0024] Dynamic Probability Roulette: Constructs a normalized probability distribution based on multi-dimensional activation levels, and randomly selects nodes based on cumulative probability to achieve a natural exploration mechanism of "high activation priority, low activation allowed". This fundamentally avoids path solidification, reduces path predictability, and alleviates path clustering problems.

[0025] In this scheme, the activation degree is the output value of each degree function, with a value range of [0, 1]. It is used to characterize the membership degree of a candidate node in a certain dimension. The closer the value is to 1, the higher the degree of conformity in that dimension.

[0026] This solution will be explained using a robot as an example of an unmanned autonomous entity.

[0027] An unmanned autonomous agent path planning method based on path diversity includes the following steps: S1. Construct a probabilistic roulette structure for unmanned autonomous path selection as the basis for random node selection. The core purpose is to reduce path predictability and improve anti-information theft capabilities.

[0028] The roulette wheel is not a fixed structure, but is dynamically generated by the combined activation degree of all neighboring nodes of the current node. Its construction steps are as follows: a. For all walkable adjacent nodes of the current node (excluding faulty and dangerous nodes), calculate their respective comprehensive activation degree; b. Normalize the combined activation values ​​of all adjacent nodes to form a probability distribution; c. Construct a cumulative probability sequence based on the probability distribution to form a complete dynamic probability roulette; d. The width of each segment in the roulette is proportional to the overall activation level of the corresponding node. The higher the activation level, the greater the probability of being selected.

[0029] This roulette wheel is not a preset roulette wheel, but is generated in real time by four degree functions. This ensures that the path is naturally and diversely distributed while maintaining progress toward the destination, reducing path predictability, enhancing anti-information theft capabilities, and incidentally alleviating path clustering problems.

[0030] S1-1. The roulette structure depends on four degree functions: regional feasibility function, directional feasibility function, memory activation strength function, and node availability function.

[0031] Among them, the memory activation intensity function is divided into long-term memory activation intensity function and short-term memory activation intensity function. The four types of functions work together to ensure that the path is controllable and the task is achievable, while maximizing the diversity and unpredictability of the path and adapting to the needs of various scenarios.

[0032] For a single candidate node, its overall node fit represents the overall recommendation level of the candidate node at the current moment, and is calculated by the following formula (parameters are in square brackets []): Candidate node overall fit = Regional feasibility function [candidate node] × Directional feasibility function [candidate node] × Long-term memory activation intensity of candidate node × Short-term memory activation intensity of candidate node × Node availability.

[0033] In this scheme, the overall fitness score refers to the comprehensive score by which a candidate node is selected under the current state. It is obtained by fuzzy intersection operation (multiplication is used in this embodiment) based on regional feasibility, directional feasibility, short-term memory activation strength, long-term memory activation strength, and node availability. The overall fitness score reflects the degree of superiority or inferiority of the candidate node relative to other candidate nodes; a higher score indicates that it is more in line with the current selection preference.

[0034] Suppose there are N candidate nodes, numbered i. For each candidate node, the sum of its probabilities is 1. Therefore, the probability of the i-th candidate node is as follows: The probability of the i-th candidate node = the overall fitness of the i-th candidate node ÷ the sum of the overall fitness of all nodes.

[0035] S1-1-1. Construction of the regional feasibility function: Based on the starting point, the ending point and their adjacent nodes (combined with the efficient regional division of different scenarios), a polygonal soft constraint feasible region is constructed to limit the reasonable range of movement of the unmanned autonomous body and avoid inefficient areas and areas with weak communication.

[0036] In this scheme, the feasible region refers to a polygonal (elliptical in this embodiment) area constructed based on the starting and ending points. It is used to define the spatial range within which the unmanned autonomous entity can reasonably roam, avoiding dangerous or faulty areas. The feasible region has the characteristic of smooth boundary transition; the closer a node is to the boundary, the smoother the decay of its regional feasibility becomes, without any abrupt changes.

[0037] The soft constraints in this scheme are implemented using a probability roulette mechanism. Specifically, each candidate node is assigned a different probability of being selected based on its overall fitness. Nodes with high fitness have a higher probability of being selected, while nodes with low fitness have a lower probability of being selected. However, nodes with low fitness are not absolutely unselectable. The selection mechanism of "high probability first, low probability allowed" formed in this way is the soft constraint, which is different from the absolute prohibition of hard constraints (such as obstacles).

[0038] For any node in the space, calculate its region-feasible activation degree: 1. Located within the feasible region (high-efficiency region), the activation level remains high and flat; 2. Activation decays smoothly near the feasible region boundary; 3. Outside the feasible region (inefficient region), the activation rate approaches 0.

[0039] This activation level ensures that movement is controllable, does not deviate from the mission direction, and does not enter inefficient areas, providing the roulette wheel with a basis for regional legitimacy and security, and indirectly enhancing the ability to prevent information theft.

[0040] This embodiment uses an ellipse (a polygon with infinite sides) as the region feasibility function, such as... Figure 2As shown. It is understood that this scheme does not restrict the specific shape, and any polygon including the start and end points, as well as the adjacent nodes of the start and end points, is acceptable.

[0041] The calculation steps are as follows: Step 1: Find the starting point, then find the set of neighboring nodes of the starting point, and calculate the set of the starting point and its neighbors = the set of neighboring nodes of the starting point ∪ {starting point}.

[0042] Step 2: Calculate the centroid of the starting point and all nodes in its neighbor set, denoted as C1. Calculate the distance from C1 to the farthest node in the starting point and its neighbor set, and use this distance as the radius R1.

[0043] Step 3: Find the endpoint, then find the set of neighboring nodes of the endpoint, and calculate the endpoint and its neighbor set = the set of neighboring nodes of the endpoint ∪ {the endpoint}.

[0044] Step 4: Calculate the centroid of the endpoint and all nodes in its neighbor set, denoted as C2. Calculate the distance from C2 to the endpoint and its neighbor set, and use it as the radius R2.

[0045] Step 5: Determine that the major axis of the ellipse = 0.5 × (distance between C1 and C2 + R1 + R2).

[0046] Step 6: Determine that the minor axis of the ellipse = MAX{√[major axis of the ellipse × MAX{R1,R2}], parameter 1}.

[0047] Step 7: Calculate the vector between C1 and C2 and standardize it to obtain V, which is a 2-dimensional column vector. V and its normal vector are placed side by side to obtain a 2×2 matrix Z. At the same time, matrix A = DIAG{1÷(ellipse major axis × ellipse major axis), 1÷(ellipse minor axis × ellipse minor axis)}.

[0048] Step 8: Calculate the transformation operator H = 0.5 × V × (distance between C1 and C2 - R1 + R2) + C1.

[0049] Step 9: The square of the ellipse [candidate node] = (candidate node - H) transpose × Z × A × Z transpose × (candidate node - H).

[0050] Step 10: Calculate the regional feasibility function for candidate nodes: The regional feasibility function [candidate node] = 0.5 × (TANH[parameter 3 × (parameter 2 - square of ellipse [candidate node])] + 1).

[0051] Among them, parameter one is the minimum value of the minor axis, which is used to prevent the feasible region from being too narrow due to the minor axis of the ellipse being too small. The specific value can be set according to the safety distance requirements of the actual scenario. Parameter 2 is the expansion rate, which controls the extent of expansion of the feasible region. The larger the value, the larger the range of the feasible region. Parameter 3 is the marginal gradient of the regional feasibility function, which controls the steepness of the feasibility decay near the boundary. The larger the value, the faster the decay and the steeper the boundary transition.

[0052] S1-1-2. Directional Feasibility Function: Based on the direction from the current node to the endpoint and the endpoint's neighborhood, construct the directional activation function: 1. The closer the direction of adjacent nodes is to the target direction, the higher the activation degree; 2. Employ a smooth window function to ensure a continuous transition within the allowed direction range; 3. Supporting a certain degree of directional tolerance allows for path diversity and reduces predictability. The directional feasibility function provides the roulette wheel with a goal-oriented soft constraint, avoiding ineffective detours and target exposure during random walks, thus improving anti-information theft capabilities while ensuring task efficiency.

[0053] Directional feasibility function such as Figure 3 As shown, where, Figure 3 (a) is the directional feasibility function. Figure 3 (b) Comparison of candidate node comprehensive fitness (○) and directional feasibility (×). The calculation steps of the directional feasibility function are as follows: Step 1: Calculate the target heading from the current node to the destination and all nodes in its neighbor set. Assume there are M nodes in the destination and its neighbor set, i.e., M target headings. The feasibility function for each target heading is calculated as follows: Feasibility function [target heading] = 0.5 × TANH(parameter 4 × (target heading - parameter 5)) - 0.5 × TANH(parameter 4 × (target heading + parameter 5)).

[0054] The second step is to perform a union operation on the feasibility functions of all routes (the candidate routes are the routes from the current node to the candidate node): The directional feasibility function [candidate heading] = MAX{feasibility function [j-th target heading], j = 1, 2, ..., M}.

[0055] Among them, parameter four is the marginal gradient of the directional feasibility function, which is used to control the steepness of the transition of directional feasibility from the allowed direction to the non-allowed direction; Parameter 5 is the acceptable angle deviation, used to set the half-width angle of the directional tolerance, that is, within this angle range, it is considered a reasonable direction toward the target endpoint.

[0056] S1-1-3. Memory activation strength function.

[0057] To avoid path solidification and reduce path predictability, short-term and long-term memory activation levels are constructed to enhance anti-information theft capabilities and incidentally alleviate path clustering. 1. Short-term memory activation strength function: records the number of times a node in a single path has been visited; the more times a node is visited, the lower its activation level. 2. Long-term memory activation strength function: records the frequency of node access in historical tasks and decays forgetting according to a coefficient; 3. The activation rate decreases monotonically with the number of visits, causing the system to tend to select less frequently used nodes, avoid path duplication, and reduce the risk of being predicted.

[0058] like Figure 4 As shown, Figure 4 (a) represents the long-term / short-term memory activation intensity of candidate nodes. Figure 4 (b) is the overall fit of the nodes after a planning is completed (the lighter the node color, the lower the fit).

[0059] The memory activation strength function provides the roulette wheel with diverse and unpredictable constraints, and is one of the core supports for achieving anti-information theft.

[0060] Short-term memory activation strength function [number of short-term nodes passed] = EXP[-parameter six × number of short-term nodes passed^4].

[0061] Long-term memory activation strength function [number of times long-term nodes are passed] = EXP[-parameter 7 × number of times long-term nodes are passed^4].

[0062] Parameters six and seven control the rate of decrease in short-term and long-term memory, respectively. The smaller the value, the slower the decrease in activation as the number of visits increases. To ensure that short-term memory has a stronger inhibitory effect on repeated visits to nodes within a single path than long-term memory, the value of parameter six should be greater than that of parameter seven.

[0063] In this scheme, the difference between the short-term memory activation strength function and the long-term memory activation strength function lies in their different reset mechanisms: short-term memory records the number of times a node is visited during a single path planning process. After each path planning is completed (i.e., when the target destination is reached), the short-term visit count of all nodes is reset to zero, so that the next path planning is no longer affected by the short-term memory of the previous path within the same task cycle; long-term memory records the cumulative visit frequency across tasks. It is not directly reset to zero, but is only multiplied by the forgetting coefficient (parameter eight) after each path planning is completed to reduce the weight, so that nodes with low historical visit frequency or those that have not been visited for a long time in long-term memory gradually recover a higher activation level, avoiding the path from becoming solidified again on a long-term scale.

[0064] Initially, all nodes have 0 short-term and long-term access counts. At this time, both short-term and long-term memory activation strengths are at their maximum values ​​(1), meaning that nodes have the highest selection tendency before they are visited. The starting point is included in the access count.

[0065] S1-1-4. Node availability function construction: It is only a binary value of 0-1. 0 indicates that the node is unavailable (faulty node, dangerous node, weak communication node that is easily stolen), and 1 indicates that the node is available. It avoids security risks from the source, improves path security and reliability, and adapts to high security requirements (e.g., police scenarios).

[0066] S2. Integrated Decision Making and Controlled Stochastic Execution: For the current planning location, the regional feasibility function, directional feasibility function, memory activation intensity function, and node availability function of its reachable subsequent locations are combined, and the integrated selection tendency of each subsequent location is obtained through intersection operation; Based on this tendency, the next position or path segment is selected by introducing controllable randomness. The core purpose is to reduce the predictability of the path and strengthen the ability to prevent information theft. Iteration and Output: Repeat step S2 until the endpoint is reached or the termination condition is met, and output a path; by executing this method multiple times or planning for multiple agents, a path set with high information entropy can be generated, further reducing the predictability of the overall path, reducing the risk of being stolen by batch information, and incidentally mitigating the security risks caused by path aggregation.

[0067] In this scheme, "intersection operation" refers to performing a fuzzy intersection operation on the output values ​​of multiple degree functions. In this embodiment, multiplication in the Zadeh fuzzy system is used as the fuzzy intersection operator, that is, the output values ​​of each degree function are directly multiplied to obtain the comprehensive fit degree. However, the scope of protection of this scheme is not limited to this; other reasonable fuzzy operation methods (such as the minimum operation) are also applicable.

[0068] In this embodiment, "multiple executions of this method or planning for multiple agents can generate a set of paths with high information entropy" is achieved as follows: When planning paths for multiple unmanned autonomous agents simultaneously, each agent needs to share the same long-term memory, that is, the path selection of all agents is affected by the unified historical access record, thereby achieving natural distribution of paths at the global level; when executing tasks multiple times for a single unmanned autonomous agent, the agent independently maintains its long-term memory, and subsequent path planning is still affected by previous historical tasks to continuously maintain path diversity.

[0069] This embodiment presents an unmanned autonomous path planning method based on path diversity. The overall process is as follows: Figure 1 As shown, it includes the following steps: Initialization: For all nodes in the local graph, calculate the region feasibility function and eliminate dangerous or weakly connected nodes.

[0070] Step 1: For the current node, find the neighbor set, which consists of N candidate nodes, with i as the candidate node number.

[0071] Step 2: Calculate the directional feasibility function, short-term memory activation strength function, and long-term memory activation strength function in sequence based on the node information.

[0072] Step 3: Calculate the overall fitness of candidate nodes; Candidate node overall fit = Regional feasibility function [candidate node] × Directional feasibility function [candidate node] × Long-term memory activation intensity of candidate node × Short-term memory activation intensity of candidate node × Node availability.

[0073] Step 4: Create roulette wheel F; The probability of the i-th candidate node = the overall fitness of the i-th candidate node ÷ the sum of the overall fitness of all nodes; F = {the i-th candidate node: the probability of the i-th candidate node, i = 1, 2, ..., N}.

[0074] Step 5: Make a decision based on the path decision-making process of the probability roulette: select the Jth candidate node.

[0075] Step 6: Current node ← Jth candidate node; the number of times the Jth candidate node passes the long-term node test increases by 1, and the number of times the Jth candidate node passes the short-term node test increases by 1.

[0076] Repeat steps one through six until the current node equals the endpoint (if the current node is not equal to the endpoint, then return to step one); if the endpoint itself is not in the neighbor list, then proceed to the next step.

[0077] Step 7: For all nodes: Long-term node pass count = Long-term node pass count × Parameter 8; Meanwhile, short-term node pass count = 0. Begin the next planning phase to avoid long-term path fixation and continuously improve anti-information theft capabilities.

[0078] Parameter 8 is the forgetting coefficient, with a value ranging from (0, 1), which controls the rate of forgetting in long-term memory. The smaller the value of parameter 8, the faster the forgetting rate; the closer the value is to 1, the slower the forgetting rate.

[0079] The path planning results based on this scheme are as follows: Figures 5-7 As shown, Figure 5 A schematic diagram of the path under normal circumstances is shown. Figure 6 The path and heatmap are shown when some nodes are unavailable. Figure 7 This is a schematic diagram of the path entropy analysis results, where, Figure 7 (a) is the path entropy obtained after 50,000 planning iterations. Figure 7 (b) is the conditional entropy obtained after 50,000 planning iterations. Figure 7 (c) represents the success rate and entropy percentage of 50,000 planning operations.

[0080] Overall, this scheme shifts from the existing "deterministic optimality" paradigm to a "probabilistic diversity" paradigm. Specifically, existing path planning algorithms (such as A* and Dijkstra's algorithm) essentially belong to the optimization paradigm, that is, under fixed start and end points and environmental constraints, they search for a deterministic path based on a single objective, a weighted multi-objective function, or a Pareto front. This deterministic solution logic, when the start, end, and environmental map are fixed, results in highly similar algorithm outputs and strong path predictability. This scheme, however, transforms path selection from "deterministic optimality selection" to "probabilistic random selection," that is, by constructing a dynamic probability roulette wheel, each candidate node is assigned a probability of being selected, and then a random decision is made under this probability distribution. The essence of this design is that even if a node has the "highest score" in the current state, it only has a higher probability of being selected, not a certainty; conversely, nodes with lower scores still have a possibility of being selected. This mechanism makes path planning no longer converge to a single solution, but naturally generates diverse path distributions, fundamentally breaking the inherent mechanism of path fixation.

[0081] To avoid the lack of constraints that can lead to problems such as excessively long detours, deviations from the target, or even failure to reach the destination due to random decision-making, this solution proposes a soft constraint mechanism. The overall suitability of each candidate node is jointly determined by multiple degree functions through fuzzy intersection operations. These functions guide the random walk from multiple dimensions, including spatial range (regional feasibility), direction of travel (directional feasibility), access history (memory activation strength), and node state (node ​​availability). Among them, regional feasibility restricts the autonomous agent's walking range to a feasible domain constructed based on the starting point and the destination, while directional feasibility ensures that the path generally moves towards the target destination. Together, they guarantee the controllability of the path, ensuring that the path does not deviate from the task scope and always moves towards the destination. Unlike the absolute prohibition of hard constraints, soft constraints only suppress the selection of nodes that do not meet the conditions by reducing the probability, rather than completely eliminating them. This achieves a balance between "controllability" and "diversity," realizing the effect of "controllable but not rigid."

[0082] The core obstacle to path diversity lies in the fact that even with the introduction of randomness, the lack of a memory mechanism for "already traversed paths" can lead to repeated selection of the same nodes, causing statistically repetitive paths. Therefore, this scheme employs a dual activation strength function combining short-term and long-term memory to address this issue. Short-term memory records the number of times a node is visited during a single path planning process, resetting it to zero after each destination is reached. Its function is to prevent repeated visits to the same nodes, thus avoiding loops or backtracking. Long-term memory records the cumulative frequency of visits across tasks, not resetting it directly but using a forgetting coefficient for attenuation and weighting. Its function is to continuously suppress the repeated selection of high-frequency nodes in multiple tasks or multi-agent collaborative planning, allowing low-frequency or long-unvisited nodes to gradually regain their selection probability. By utilizing short-term memory to ensure that single paths avoid loops and backtracking, and long-term memory to ensure continuous dispersion across paths over time, the synergistic effect of both makes the system naturally inclined to select "less frequently used" nodes, thus automatically generating a high-information-entropy path set without relying on external intervention or complex collaborative communication.

[0083] Meanwhile, in scenarios where multiple autonomous agents simultaneously execute tasks, if each agent plans independently without sharing information, multiple agents may still choose the same or similar paths, leading to path clustering. This solution addresses this issue by using a shared long-term memory mechanism: all autonomous agents share the same long-term access record. Whenever an agent passes through a node, the node's long-term access count increases by one, thus reducing the probability of that node being selected by other agents. This makes path dispersion global and spontaneous, rather than requiring real-time negotiation and dynamic coordination between agents. Each agent only needs to read the shared long-term memory to automatically choose a different path from others, greatly reducing the complexity of collaborative communication and improving the system's robustness.

[0084] Furthermore, the calculation of each degree function involves simple mathematical operations (product, exponentiation, hyperbolic tangent, etc.), without involving complex iterative optimization or large-scale matrix operations, and can run in real time on embedded platforms with limited computing power. Moreover, the comprehensive adaptability of each node is calculated by a degree function with four dimensions that can be explicitly interpreted. Each decision node can be backtracked to analyze why a certain node was chosen, improving the traceability of the decision. In addition, the parameters during path planning (such as feasible region size, directional tolerance, forgetting rate, etc.) all have clear physical meanings, and the parameters can be intuitively adjusted according to the actual needs of the scenario to adapt to different task requirements without the need for a complex parameter tuning process.

[0085] Accordingly, an unmanned autonomous path planning system based on path diversity includes: The dynamic probability roulette module is configured to: obtain all passable neighboring nodes of the current node, calculate the comprehensive fitness degree of each neighboring node and normalize it to obtain the probability distribution of each neighboring node being selected, and construct a dynamic probability roulette based on this. The node update module is configured to: determine the next node from the adjacent nodes in a random manner according to the probability distribution of the dynamic probability roulette, and update the current node as the next node; The iterative output module is configured to repeat the above steps until the current node is the target endpoint, and output the path from the starting point to the target endpoint.

[0086] This approach transforms path selection from a "deterministic optimal solution" to a "probabilistic random selection." At each decision node, candidate nodes are assigned different probabilities of selection based on their overall fitness, and the system randomly selects the next node under this probability distribution. This fundamentally breaks the inherent mechanism of path fixation: on the one hand, because randomness is introduced in each selection, even if the starting point and destination are the same, the paths generated by multiple planning processes exhibit a high degree of diversity globally, effectively reducing path predictability; on the other hand, by guiding the random walk through dimensions such as memory constraints in the overall fitness, it ensures that the path always moves towards the target destination, avoiding problems such as excessive detours and deviations from the target caused by pure random walks. This improves path diversity while ensuring task reachability and reliability. Therefore, it can achieve a balance between path diversity and controllability in a lightweight manner without relying on complex collaborative communication, solving the core defects of existing technologies such as highly convergent paths and strong predictability.

[0087] Correspondingly, a computer program product includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the aforementioned unmanned autonomous path planning method based on path diversity.

[0088] Correspondingly, an electronic device includes at least one processor and a memory connected to the processor, the memory being used to store computer programs; the processor is used to execute the computer programs, enabling the electronic device to implement the aforementioned unmanned autonomous path planning method based on path diversity.

[0089] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for unmanned autonomous path planning based on path diversity, characterized in that, Includes the following steps: Obtain all accessible neighboring nodes of the current node, calculate the comprehensive fitness degree of each neighboring node and normalize it to obtain the probability distribution of each neighboring node being selected, and construct a dynamic probability roulette based on this. Based on the probability distribution of the dynamic probability roulette, the next node is determined from the adjacent nodes in a random manner, and the current node is updated as the next node; Repeat the above steps until the current node is the target destination, and output the path from the starting point to the target destination.

2. The path diversity based path planning method for an unmanned autonomous body as claimed in claim 1 wherein, The overall fitness score is obtained by intersecting at least two of the following functions: regional feasibility function, directional feasibility function, memory activation intensity function, and node availability function.

3. The path planning method for path diversity based on the unmanned autonomous body according to claim 2, characterized in that, The regional feasibility function is constructed as follows: a polygonal soft-constraint feasible region is constructed based on the starting point and the target endpoint; for nodes located inside the feasible region, the regional feasibility remains high and flat; for nodes close to the boundary of the feasible region, the regional feasibility decays smoothly; for nodes outside the feasible region, the regional feasibility approaches zero.

4. The path planning method for path diversity based on the unmanned autonomous body according to claim 2, characterized in that, The directional feasibility function is constructed as follows: the direction from the current node to the target endpoint and the neighboring nodes of the target endpoint are used as the reference. The closer the travel direction of the adjacent node is to the reference direction, the higher the directional feasibility. The directional feasibility is calculated using a smooth window function, so that the allowed directional range transitions continuously.

5. The path diversity based path planning method for an unmanned autonomous agent according to claim 2, wherein, The memory activation strength function includes a short-term memory activation strength function and a long-term memory activation strength function. The short-term memory activation strength function records the number of times a node is visited in a single path planning session. As the number of visits increases, the short-term memory activation strength decreases. The long-term memory activation strength function records the cumulative frequency of visits to nodes in historical path planning sessions and decays the frequency of forgetting according to a preset coefficient. The memory activation strength function decreases monotonically with the increase of the number of visits.

6. The path diversity based path planning method for an unmanned autonomous body as claimed in claim 2, wherein, The node availability function is a binary function: when a node is a faulty or dangerous node, the node availability is 0, indicating that the node is unavailable; otherwise, the node availability is 1, indicating that the node is available.

7. The path diversity based path planning method for an unmanned autonomous body as claimed in claim 1 wherein, After completing a path planning operation, the number of short-term visits is reset to zero, and the number of long-term visits is multiplied by a preset forgetting coefficient to reduce the impact of historical paths on subsequent path planning.

8. A path planning system for an autonomous agent based on path diversity, the system comprising: include: The dynamic probability roulette module is configured to: obtain all passable neighboring nodes of the current node, calculate the comprehensive fitness degree of each neighboring node and normalize it to obtain the probability distribution of each neighboring node being selected, and construct a dynamic probability roulette based on this. The node update module is configured to: determine the next node from the adjacent nodes in a random manner according to the probability distribution of the dynamic probability roulette, and update the current node as the next node; The iterative output module is configured to repeat the above steps until the current node is the target endpoint, and output the path from the starting point to the target endpoint.

9. A computer program product, characterised in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the steps in the unmanned autonomous path planning method based on path diversity as described in any one of claims 1-7.

10. An electronic device, comprising: It includes at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor is used to execute the computer program, enabling the electronic device to perform the steps in the unmanned autonomous path planning method based on path diversity as described in any one of claims 1-7.

Images