Intelligent path planning system for complex grassland unmanned vehicle

Abstract

The application belongs to the technical field of computer data processing, and relates to path planning, model training and intelligent control technology, in particular to an intelligent path planning system for unmanned work vehicles in complex grasslands. The system obtains historical operation data of the unmanned work vehicle, such as a planned path, an actual trajectory, speed attenuation, steering correction, traction output change and work state, divides the grassland work area into planning units with spatial and time indexes, and trains a grassland passage difficulty model according to the passage correlation between adjacent planning units. Further, a path planning cost map containing a basic distance cost, a passage difficulty cost, a steering continuity cost and a historical risk attenuation cost is constructed according to the model output, and a work path of the unmanned work vehicle is output. The system can convert the operation feedback in the complex grassland into an updateable path cost, reduce the trajectory deviation and frequent re-planning caused by low adhesion and high resistance areas, and improve the path execution stability.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing technology, and relates to path planning, model training and intelligent control technology, specifically an intelligent path planning system for unmanned vehicles operating in complex grasslands. Background Technology

[0002] Path planning for unmanned vehicles operating in complex grasslands typically uses the boundaries of the work area, starting point, ending point, work width, and restricted areas as basic data. A computer system discretizes the grassland area into grids, nodes, or flight strips, and then performs path search within this discrete space. Conventional inventions often use fixed-resolution grid maps to represent the grassland surface, designating areas outside the boundaries as impassable and areas inside the boundaries as uniformly passable areas, generating work paths based on factors such as path length, number of turns, and coverage sequence. For continuous work tasks, the system usually first determines the main work direction based on the grassland outline, then generates reciprocating or zoned coverage lines, and connects adjacent coverage lines using graph search methods. The key to this type of invention lies in converting irregular grassland boundaries into a computable path map, enabling the unmanned vehicle to complete travel and work according to a preset path. Because surface resistance, grass cover, low-lying water accumulation, historical ruts, and adhesion conditions at different locations within the grassland are difficult to fully represent in the initial map, existing systems often treat areas within the same boundary with the same or approximately the same passage cost, failing to fully reflect the spatial differences in drivability within complex grasslands.

[0003] In existing path planning systems, some inventions utilize the historical trajectories of unmanned vehicles to correct paths. The conventional approach involves recording the vehicle's actual driving trajectory, speed changes, yaw rate, or control corrections, and adding locations where deviations, stops, or detours occurred as risk locations to the map during the next operation. This method typically converts historical anomalies into fixed obstacles, fixed high-cost grids, or manually marked areas, which are then avoided by the path search algorithm. While this approach can memorize some recurring anomalies, it is insufficient for representing the directional, temporal, and continuously propagating difficulties in grassland environments. For example, when entering the same planned location from different directions, the vehicle's lateral deviation, steering correction, and traction output may differ; the same area may also exhibit different traffic characteristics at different operating times due to ground moisture, repeated compaction, or grass obstruction. If historical experience is only represented by fixed anomalies or static high-cost grids, the path planning system struggles to distinguish between localized, occasional deviations and stable traffic difficulties, and also finds it difficult to make reasonable cost estimates for adjacent areas that have not been fully traversed.

[0004] Existing similar technologies include path cost estimation for autonomous vehicles based on learning models. These inventions typically input environmental images, terrain classification, historical driving states, or trajectory errors into the model, outputting the probability of road or region accessibility, and then combining this with traditional path search algorithms to generate routes. In grassland unmanned operation scenarios, to reduce mapping complexity, the system usually only overlays accessibility probabilities or risk scores onto existing raster maps and updates path costs at fixed intervals. This approach is applicable to areas with structured roads or clearly defined obstacle boundaries, but complex grasslands are unstructured operation areas where there is a lag and spatial correlation between vehicle operating status and surface conditions. A single trajectory point or single passage record cannot directly represent the true accessibility difficulty of the corresponding area. If the model training does not consider the continuous access relationships between adjacent planning units, differences in entry directions, differences in passage time periods, and sources of abnormal samples, the generated cost distribution is prone to abrupt changes, drift, or excessive smoothing, causing the path search results to deviate from the actual executable paths of the vehicle.

[0005] The main technical problem with existing intelligent path planning systems for unmanned vehicles operating in complex grasslands is that the travel costs relied upon for path planning lack dynamic learning and spatial correlation representation based on historical vehicle operation data. This makes it difficult for the system to transform operational data such as deviations between planned and actual trajectories, speed decay, steering corrections, traction output changes, and operational status into continuously updated grassland travel difficulty information. The technical reason for this is that existing inventions often use static maps, fixed outliers, or single risk scores as the basis for path search. They fail to divide grasslands into planning units with spatial and temporal indices, fail to perform credibility stratification and directional time-period correlation on multiple travel records within the same planning unit, and fail to infer the travel difficulty of insufficiently traversed areas through the graph structure relationships between adjacent planning units. Consequently, in complex grasslands, the system easily generates paths that enter areas with low adhesion, high resistance, or accumulated offset, resulting in increased path execution deviations and frequent replanning. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent path planning system for unmanned vehicles operating in complex grasslands, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical invention adopted by this invention is as follows: The intelligent path planning system for unmanned vehicles operating in complex grasslands includes a processor and a memory. The memory stores program instructions executable by the processor, which, when executed, cause the processor to: Acquire historical operational data of unmanned vehicles operating in complex grasslands, including planned path, actual trajectory, speed decay, steering correction, traction output changes, and operational status. The grassland operation area is divided into planning units with spatial and temporal indexes; Traffic difficulty labels are generated based on historical operational data within the same planning unit. A grassland access difficulty model is trained based on the access associations between adjacent planning units; The system constructs a path planning cost map based on the grassland access difficulty model and outputs the operation path of the unmanned vehicle.

[0008] Preferably, when the processor divides the planning unit, it performs coordinate unification and effective operation range closure processing on the boundary of the grassland operation area, and maps the planned path and the actual trajectory to the same spatial reference system; The size of the planning unit is determined based on the operating width of the unmanned vehicle, the density of historical trajectories, and the curvature of the boundary. An index record containing the entry direction, exit direction, passage time period, and adjacent relationships is established for each planning unit. Trajectory segments spanning multiple planning units are segmented and assigned to different units, so that the historical operation data and the corresponding planning units form updatable data pairs.

[0009] Preferably, when generating the accessibility label, the processor determines the trajectory deviation sequence based on the lateral offset, heading offset, and return time between the planned path and the actual trajectory; The motion response sequence is determined based on the speed decay, steering correction, and traction output changes. The segments in the operational status that represent temporary parking, manual takeover, non-operational movement, or task switching are marked as samples to be excluded; Within the same planning unit, the trajectory deviation sequence and the motion response sequence are time-aligned to generate passage difficulty labels associated with the entry direction and passage time period.

[0010] Preferably, when the processor trains the grassland access difficulty model, it constructs each planning unit as a graph node and constructs adjacent planning units with continuous access relationships as directed edges. The node features and edge features are formed using the aforementioned passage difficulty labels, boundary distances, historical passage counts, differences in entry directions, and differences in adjacent node labels; The implicit representation of nodes is updated using a graph structure learning approach, and the estimated passage difficulty is output for planning units that have not formed effective passage samples. Different sample weights were set for manually taken-over samples and non-continuous passage samples during the training process.

[0011] Preferably, when constructing the path planning cost map, the processor configures a basic distance cost, a passage difficulty cost, a turning continuity cost, and a historical risk attenuation cost for each planning unit; The passage difficulty cost is output by the grassland passage difficulty model, the turning continuity cost is determined by the entry and exit directions of adjacent path segments, and the historical risk attenuation cost is determined by the most recent abnormal passage time and the number of abnormal repetitions. A set of candidate paths is generated in the global search, and the operation path of the unmanned vehicle is selected according to the total cost and path continuity of each candidate path.

[0012] Preferably, the processor further performs credibility stratification on multiple historical operation data within the same planning unit. First, it determines the record completeness based on timestamp continuity, positioning sampling interval, trajectory segment completeness, and operation status consistency. Then, it determines the sample consistency based on the synchronous changes between the trajectory deviation sequence and the motion response sequence. When the same planning unit has opposite passage difficulty labels at different passage times, sub-label groups are established according to the passage time, entry direction, exit direction and adjacent planning unit status, and the sub-label groups are associated with and stored with the corresponding planning unit, passage sample source and update time.

[0013] Preferably, the processor also sets direction-sensitive propagation constraints in the directed edges, and the direction-sensitive propagation constraints are jointly determined by the departure direction of the previous planning unit, the entry direction of the next planning unit, the boundary curvature change, and the historical trajectory offset direction. When the difference in accessibility labels between adjacent planning units exceeds the preset hierarchical conditions, direct label propagation is stopped and a transition node representation is established; When a planning unit that has not formed a valid passage sample is located between multiple transition nodes, the passage difficulty estimate is generated based on the distance attenuation and directional consistency represented by the transition nodes.

[0014] Preferably, when generating the candidate path set, the processor first forms basic search constraints based on the start point, end point, and required coverage area of ​​the grassland operation task, and then retains multiple paths with similar costs but different spatial distributions in the path planning cost map. For each candidate path, generate a path signature consisting of continuous planning units, and record the continuous length, occurrence position, and entry direction of the planning units with high traversability difficulty in the path signature; During the operation of the unmanned vehicle, the corresponding path signatures are locally sorted and updated based on the newly generated passage difficulty labels.

[0015] Preferably, the processor further establishes an incremental update version chain for the grassland access difficulty model, the incremental update version chain including planning units, sub-label groups, node implicit representations, transition node representations, and associated version numbers of access difficulty estimates; When the newly added historical running data only affects some planning units, local training is only performed on the affected planning units, adjacent planning units with directed edge connections, and the corresponding sub-label groups. After local training is completed, the difference in the estimated travel difficulty before and after the update is written into the version difference record, and a stable version is retained for use in the path planning cost graph.

[0016] Preferably, the processor continuously receives newly generated trajectory segments and matches them with the current path signature when the unmanned vehicle is executing the unmanned vehicle's work path; When the planning unit corresponding to the trajectory segment forms a level inconsistency record with the estimated travel difficulty value in the stable version, the unaffected path signature segment is locked, and the local path planning cost map is reconstructed only for the planning unit with inconsistent level and the connecting segments before and after it. The reconstructed candidate local paths are subjected to signature splicing verification. If the splicing verification passes, the corresponding segment in the current path signature is replaced, and the replacement result is written into the version difference record.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By dividing complex grassland operation areas into planning units with spatial and temporal indices, and generating accessibility labels based on planned paths, actual trajectories, speed decay, steering corrections, traction output changes, and operational status, the path planning system can transform the operational feedback of unmanned vehicles during actual operations into calculable and updatable accessibility costs. Compared to path searching based solely on static maps or fixed anomalies, this invention can distinguish the accessibility of different planning units and incorporate entry direction, passage time, and adjacency relationships into the data organization process, preserving the differences in the same grassland area under different access conditions. Accessibility relationships between adjacent planning units are established through graph structure learning; areas without sufficient accessibility samples can also obtain accessibility estimates based on neighboring areas, boundary distances, historical passage counts, and label differences. The path planning cost map simultaneously includes basic distance costs, accessibility costs, steering continuity costs, and historical risk decay costs, ensuring that candidate path generation is not solely dependent on geometric distance but is also constrained by the actual vehicle operating status. As a result, the system can reduce the number of times the operation path is planned to areas with low adhesion, high resistance, or cumulative deviation, thereby reducing the probability of trajectory execution deviation and repeated replanning, and enabling unmanned vehicles to obtain more stable executable paths in complex grasslands.

[0018] 2. By performing credibility stratification on multiple historical operation data within the same planning unit, the system can eliminate interference from abnormal samples caused by temporary parking, manual takeover, non-operational movement, or task switching, reducing the impact of erroneous samples on the accessibility label. By establishing sub-label groups according to the time period, entry direction, exit direction, and status of adjacent planning units, the system can retain the directional and temporal differences in grassland accessibility, avoiding the merging of operational deviations with different causes into a single risk value. By setting direction-sensitive propagation constraints in directed edges and establishing transition node representations between adjacent planning units with significant label differences, the accessibility estimate can maintain continuity during spatial propagation, reducing unreasonable abrupt changes in the cost graph. By incrementally updating the version chain to record the version relationship of planning units, sub-label groups, node implicit representations, and accessibility estimates, adding new historical operation data only triggers local training in the affected area and its adjacent areas. Through path signature and local path planning cost graph reconstruction, when local level inconsistencies occur during path execution, unaffected path segments can be retained, and only affected segments and their connecting segments can be replaced, reducing the computational burden of full-field recalculation and maintaining consistency before and after path updates. Attached Figure Description

[0019] Figure 1 This provides an overview of the system's overall processing flow from historical data to job path output; Figure 2 A process diagram is created for grassland area discretization, trajectory mapping, and planning unit indexing; Figure 3 This describes the logic for generating accessibility labels and the process for processing sample credibility. Figure 4 The process of training a grassland accessibility model and constructing a path planning cost map; Figure 5 The execution-time update process triggers local training, version difference recording, and local path replacement for newly added running data. Detailed Implementation

[0020] refer to Figure 1In one embodiment, the intelligent path planning system for unmanned vehicles operating in complex grasslands includes a processor and a memory. The memory stores program instructions that can be executed by the processor. When the program instructions are executed, the processor completes the following tasks: receiving historical grassland operation data, discretizing the grassland operation area, generating accessibility labels, training the grassland accessibility model, constructing a path planning cost map, and outputting the unmanned vehicle's operation path. The system does not rely on modifications to the vehicle's structure; instead, it utilizes data records generated by the unmanned vehicle during previous operations to learn path costs. Historical operation data includes the planned path, actual trajectory, speed decay, steering correction, traction output changes, and operational status. The planned path represents the target trajectory that the unmanned vehicle should have traversed in the previous path planning round. The actual trajectory represents the real position sequence formed during vehicle execution. Speed ​​decay represents the degree of reduction in actual speed relative to the planned speed. Steering correction represents the directional adjustments made by the vehicle to return to the planned path. Traction output changes represent the changes in the vehicle's power response when passing through different positions in the grassland. The operational status distinguishes data sources such as normal operation, temporary parking, manual takeover, non-operational movement, and task switching. The above data is uniformly written into the historical operation data cache area, along with timestamps, trajectory segment identifiers, and job task identifiers, so that the subsequent label generation process can eliminate interference caused by non-operational factors.

[0021] refer to Figure 2 In this embodiment, the system divides the grassland operation area into planning units with spatial and temporal indices. The spatial index is used to determine the location of the planning unit within the grassland operation area, while the temporal index is used to distinguish the passage performance of the same location at different times. Since complex grassland boundaries are usually not regular rectangles, the system first performs a closure check on the operation area boundaries to confirm that the boundary point sequence can form a closed area. For boundary records where the first and last points do not coincide, the system completes the closed connection segment according to the directional extension relationship of adjacent boundary points and marks the area within the boundary as a candidate operation range. The planned path and the actual trajectory are mapped to the same spatial reference system before entering the planning unit division to avoid coordinate deviations caused by different data sources. Coordinate unification can be achieved using the following transformation formula: ; ; in, and Indicates the first The horizontal and vertical coordinates of the original coordinate points and This indicates the coordinates of the reference origin for the grassland operation area. This represents the rotation angle of the original coordinate system relative to the unified coordinate system. and Represents the transformed, unified coordinates. Operators and These represent the cosine and sine operations, respectively. For example, when the original coordinate point is... The reference origin is Rotation angle When, the conversion result is , The point is mapped to a unified plane coordinate system with the grassland reference origin as the reference. If the rotation angle is not 0, the same calculation formula simultaneously performs translation and rotation, so that the trajectory point and the grassland boundary point are in the same computational space.

[0022] In this embodiment, the planning unit size is not a fixed value, but is determined jointly based on the unmanned vehicle's operating width, historical trajectory density, and boundary curvature. The operating width reflects the coverage scale of the operating path, the historical trajectory density reflects the sufficiency of existing operational data for a certain grassland area, and the boundary curvature reflects the irregularity of the grassland edge. For areas with high boundary curvature or high historical trajectory density, the planning unit size is reduced accordingly to preserve local traffic differences; for areas with relatively gentle boundaries and sparse historical trajectories, the planning unit size is kept within a larger range to avoid forming an overly fragmented path map. The planning unit size can be determined according to the following formula: ; in, Indicates the side length of the planning unit. Indicates the working width. Represents the normalized boundary curvature. Represents the normalized historical trajectory density. , and These represent the calculation coefficients used to balance the effects of operating width, boundary curvature, and trajectory density. and These represent the lower and upper limits of the side length of the planning unit, respectively. This indicates a truncation operation that restricts intermediate calculated values ​​to between the lower and upper bounds. For example, when , , , , , , , At that time, the intermediate calculated value is Since it lies between the lower and upper limits, the side length of the planning unit is taken as... The advantage of this embodiment is that the planning unit can be adjusted according to the complexity of the grassland boundary and the distribution of historical data, so that the learning of accessibility difficulty has a clear data carrying unit, and avoids the inadequacy of a single static map in expressing the differences within the grassland.

[0023] In one embodiment, the system performs field processing on historical operation data and assigns each trajectory segment to a corresponding planning unit. The start and end positions of a trajectory segment may span multiple planning units. The processor divides the trajectory segment into several sub-segments according to the intersection points of the trajectory points and the boundaries of the planning units, with each sub-segment corresponding to only one planning unit. For points that cross the boundaries of planning units, the system determines whether the point belongs to the departure point of the previous planning unit or the entry point of the next planning unit based on the trajectory direction and time sequence. An index record is established for each planning unit, including a spatial index, a time index, an entry direction, a departure direction, a passage period, and adjacent planning unit numbers. The entry direction is determined by the direction of the trajectory points before and after the trajectory segment enters the planning unit, the departure direction is determined by the direction of the trajectory points before and after the trajectory segment leaves the planning unit, the passage period is determined by the timestamp interval corresponding to the trajectory segment, and the adjacency relationship is determined by the planning unit boundary sharing relationship and the continuous traversability relationship. For areas outside the grassland boundaries and prohibited entry areas, no traversability index is established.

[0024] Table 1 shows the processing standards between historical operation data and planning units, which are used to illustrate the calculation purpose of different data fields in the process of generating accessibility difficulty.

[0025] Table 1. Processing methods between historical operational data and planning units

[0026] In this embodiment, the data fields in Table 1 are not stored in isolation, but rather form operation samples using trajectory segment identifiers as the primary key. Each operation sample includes a planned path segment, an actual trajectory segment, a speed sequence, a steering correction sequence, a traction output sequence, and an operational status sequence. Before generating accessibility labels, the processor checks whether the data timestamps within the same operation sample are continuous. If a field is missing but the trajectory segment remains complete, the system marks the operation sample as a low-completeness sample, rather than deleting it directly. For samples with manual intervention or temporary parking, the system marks them as samples to be excluded based on their operational status, ensuring that subsequent label generation processes do not misjudge speed reductions or trajectory deviations caused by human intervention as grassland accessibility difficulties. The advantage of this embodiment is that historical operation data can form traceable data pairs at the planning unit level, giving accessibility labels a clear data source and spatial location basis.

[0027] refer to Figure 3In one embodiment, the processor determines a trajectory deviation sequence based on the lateral offset, heading offset, and return time between the planned path and the actual trajectory. The lateral offset represents the vertical distance from the actual trajectory point to a segment of the planned path; the heading offset represents the angle between the actual trajectory direction and the planned path direction; and the return time represents the time it takes for the vehicle to return to the vicinity of the planned path after deviating from it. For multiple trajectory points within the same planning unit, the processor calculates the average lateral offset, average heading offset, and average return time, and combines them into a trajectory deviation. The trajectory deviation can be calculated using the following formula: ; in, Indicates the first The trajectory deviation of each running sample. Indicates the first The average lateral offset of each running sample within its corresponding planning unit. Indicates the average heading deviation. Indicates the return time. The baseline quantity representing the regression time. , and This represents the calculated weights of each item in the trajectory deviation, and remains fixed when used for the same model version. This represents the absolute value operation. This represents pi. For example, when... , , , , , , , At that time, the trajectory deviation was The larger this value, the more difficult it is for a vehicle to stably pass through the corresponding planning unit according to the planned path.

[0028] In this embodiment, the processor also determines the motion response sequence based on speed decay, steering correction, and traction output changes. Speed ​​decay reflects the vehicle's speed change after being hindered by grass or low ground adhesion within the corresponding planning unit; steering correction reflects the frequency of directional adjustments made by the vehicle to return to the planned path; and traction output changes reflect the changes in the vehicle's dynamic response as it passes through the area. The motion response quantities can be calculated using the following formula: ; in, Indicates the first Motion response of a running sample This represents the actual average speed within the corresponding planning unit. Indicates the planned speed. Indicates when Take when greater than 0 ,when When less than or equal to 0, the positive part of the operation is taken. Indicates the number of reversal corrections. This indicates the number of trajectory sampling points used for statistics within the planning unit. This represents the average traction output. This indicates the planned traction output reference quantity. This represents a stable term used to avoid a denominator of 0. , and This represents the weighting of each item in the motion response. For example, when... , , , , , , , , , At that time, the motion response quantity is The advantage of this embodiment is that the difficulty of passage is no longer determined solely by positional deviation, but is also calculated in conjunction with vehicle motion response, which can reduce the impact of a single trajectory error on the labeling results.

[0029] In one embodiment, before generating accessibility labels, the processor filters job states and stratifies sample credibility. Segments of job states indicating temporary stops, manual takeover, non-job movement, or task switching are marked as samples to be excluded. For samples containing brief missing data but still maintaining trajectory continuity, the system retains the data but reduces the credibility. Credibility is determined by record completeness, consistency between trajectory deviation and motion response, and valid status marking. Record completeness indicates whether fields such as planned path, actual trajectory, speed, steering correction, traction output, and job state are complete; consistency indicates whether trajectory deviation and motion response change in the same direction within the same planning unit; valid status marking excludes abnormal job states. Sample credibility can be calculated using the following formula: ; in, Indicates the first The credibility of each running sample Indicates record completeness, with a value ranging from 0 to 1. Indicates the consistency truncation value. This indicates a valid status flag; 1 indicates normal operation, and 0 indicates samples to be excluded. This indicates the operation of taking the smaller value. For example, when... , , , , At that time, the credibility was If the same sample is marked as manually taken over, then The confidence level is 0, and this sample does not participate in the effective accumulation of the difficulty label.

[0030] Table 2 shows the correspondence between job status and sample processing method, which is used to explain the filtering logic in the process of generating access difficulty labels.

[0031] Table 2. Correspondence between work status and sample processing method

[0032] In this embodiment, multiple passage records may exist within the same planning unit. The processor will group running samples with the same or similar time period, entry direction, and exit direction into the same sub-label group. The sub-label group is not a simple average of samples, but rather retains the directional and time period characteristics of the running samples. The passage difficulty label can be calculated using the following formula: ; in, Indicates spatial index as Grouped by time period Directional grouping The difficulty tag for passage, This represents the set of running samples belonging to this spatial index, grouped by time period and direction. and This indicates the weights of trajectory deviation and motion response in the label calculation. This indicates a stable term used to avoid a denominator of 0. For example, when a sub-label group contains two samples with confidence levels of 0.88 and 0.70, trajectory deviations of 0.199 and 0.250, and motion response values ​​of 0.210 and 0.280, , , At that time, the difficulty label was The advantage of this embodiment is that the label generation process utilizes trajectory deviation, motion response, sample reliability, passage time period, and entry direction simultaneously, which can distinguish between occasional anomalies and stable passage difficulties.

[0033] In one embodiment, the system employs a graph-based learning approach to train the grassland access difficulty model. The processor constructs each planning unit as a graph node and adjacent planning units with continuous access relationships as directed edges. The direction of the directed edges is determined by the trajectory direction of a vehicle entering another planning unit, with different directions corresponding to different edge features. Node features include access difficulty label, boundary distance, historical passage count, entry direction statistics, exit direction statistics, and most recent update time. Edge features include the exit direction of the previous planning unit, the entry direction of the next planning unit, differences in adjacent node labels, boundary curvature changes, and historical trajectory offset direction. During model training, planning units with valid access samples use the access difficulty label as a supervision signal, while planning units without valid access samples obtain implicit representations through the propagation of adjacent node features. For manually intervened samples and discontinuous access samples, the processor sets lower sample weights or excludes them directly to prevent the training data from being contaminated by abnormal operating conditions.

[0034] In this embodiment, graph structure learning can be accomplished through iterative updates of node representations. The implicit representation of a planning unit node is calculated jointly using its own features and the features of its incoming adjacent nodes, specifically according to the following formula: ; in, Indicates spatial index as The planning unit in the first Implicit representation of nodes in round iteration This represents the implicit representation of the node after the next iteration. This indicates that the spatial index is accessible. The set of adjacent nodes of the planning unit. Represents the spatial index of adjacent nodes. Indicates from adjacent nodes To the node Direction-sensitive propagation weights Indicates from node To the node Edge features, and This represents the parameter matrix obtained from model training. This represents the vector concatenation operation. This represents a nonlinear mapping function. For example, in a simplified one-dimensional calculation, if... Adjacent nodes are represented as 0.6, and edge features are represented as 0.2. , The calculation of the concatenated vector is as follows: , and will If processed using the identity mapping, the update result is: This update method allows insufficiently traversed planning units to obtain computational information related to direction, boundaries, and traversal history from neighboring planning units. The advantage of this embodiment is that the traversal difficulty of grasslands is not limited to locations where anomalies have occurred, but is inferred within continuously traversable areas through the graph structure relationships of adjacent planning units.

[0035] In one embodiment, the processor sets direction-sensitive propagation constraints in directed edges. These constraints are jointly determined by the departure direction of the preceding planning unit, the entry direction of the following planning unit, boundary curvature changes, and historical trajectory offset directions. If two adjacent planning units are geometrically adjacent, but a vehicle needs to drastically change direction when entering the following unit, the propagation weight is reduced. If the historical trajectory offset direction is consistent with the boundary curvature change direction, it indicates that the offset in this area may be caused by boundary constraints or continuous changes in grassland terrain, and the propagation weight is retained. If the difference in accessibility labels between adjacent planning units exceeds the hierarchical condition, the processor does not directly propagate the labels but instead establishes a transition node representation between the two planning units. The transition node representation does not correspond to the actual grassland area but is used to describe the connection relationship when there is a sudden change in accessibility between adjacent planning units. For planning units located between multiple transition nodes and lacking sufficient samples, the processor generates accessibility estimates based on the distance decay and direction consistency of the transition node representations.

[0036] In this embodiment, direction-sensitive propagation constraints are also used to handle narrow connected regions and regions near boundaries in grasslands. For adjacent planning units within a narrow region, if only a few directions are continuously passable, the system only establishes directed edges for passable directions, excluding geometrically adjacent but not continuously passable planning units from the propagation set. For planning units near boundaries, the system writes boundary distance and boundary curvature into node features, enabling the model to distinguish path deviations caused by boundary proximity from path deviations caused by low adhesion within the grassland. If the same planning unit generates opposite travel difficulty labels under different entry directions, the system retains the node additional states grouped by different directions and calls the label consistent with the current path direction during path search. In this way, different travel difficulty estimates can be obtained when a vehicle passes through the same region from east to west versus from south to north. The advantage of this embodiment is that the model training process preserves the directionality and continuity of complex grassland travel difficulty, reducing the cost of misjudgments caused by simple spatial smoothing.

[0037] refer to Figure 4In one embodiment, the processor constructs a path planning cost map based on a grassland accessibility model. Each traversable planning unit in the path planning cost map is configured with a base distance cost, accessibility cost, turning continuity cost, and historical risk attenuation cost. The base distance cost is determined by the geometric distance between planning units, the accessibility cost is output by the grassland accessibility model, the turning continuity cost is determined by the change between the entry and exit directions of adjacent path segments, and the historical risk attenuation cost is determined by the most recent abnormal passage time and the number of abnormal repetitions. The total path cost can be calculated using the following formula: ; in, This represents a candidate path composed of multiple planning units. This represents the total cost of the candidate paths. Indicates spatial index as The planning unit is located in the candidate path. This represents the basic distance cost of the planning unit. This represents the estimated accessibility value output by the grassland accessibility model. This indicates the cost of transitioning to continuity. This indicates the cost of historical risk attenuation. , , and This represents the weights used in calculating various costs. For example, for a planning unit in a candidate path, if... , , , and take , , , Then the local cost of this planning unit is The total cost of the candidate paths is the sum of the local costs of each planning unit.

[0038] In this embodiment, the historical risk attenuation cost is used to describe the sustained impact of historical anomalies on the current path planning, preventing an earlier anomaly from accumulating a high cost for an extended period, and also preventing recently occurring consecutive anomalies from being quickly ignored. The historical risk attenuation cost can be calculated using the following formula: ; in, Indicates spatial index as The historical risk attenuation cost of the planning unit This indicates the number of abnormal passages recorded for this planning unit within the stable version. This indicates the time interval since the most recent abnormal access record. Indicates the time scale of risk decay. This indicates exponential operations with the natural constant as the base. For example, when... , , At that time, the cost of historical risk decay is If no anomalies occur within the same planning unit for an extended period, As the number of abnormal passages increases, the cost of risk attenuation decreases; conversely, as the number of abnormal passages increases, the cost of risk attenuation increases. The processor generates multiple candidate paths in a global search and selects the unmanned vehicle's operating path based on the total cost and path continuity. The advantage of this embodiment is that the path search is simultaneously affected by geometric distance, passage difficulty, turning continuity, and historical anomalies, resulting in a path that better matches the vehicle's executable state in complex grasslands.

[0039] In one embodiment, when generating a candidate path set, the processor does not only retain the path with the lowest total cost, but also retains multiple paths with similar costs but different spatial distributions. The system uses the start and end points of the pasture operation task and the area to be covered as basic search constraints. After performing a graph search on the path planning cost graph, it filters candidate paths based on spatial differences. If two candidate paths have similar total costs but pass through different high-difficulty planning units, both candidate paths are retained; if the planning units passed through by two candidate paths highly overlap, only the path with the lower total cost or better turning continuity is retained. For each candidate path, the processor generates a path signature. The path signature consists of continuous planning unit numbers, entry direction sequences, exit direction sequences, continuous length of high-difficulty planning units, the location of high-difficulty occurrence, and path version number. The path signature is used to identify which segment of the given path the current trajectory belongs to during vehicle execution and is used for subsequent local sorting updates.

[0040] Table 3 shows the main fields of the candidate path signature and version record, which are used to explain how the candidate path is invoked during execution.

[0041] Table 3. Key Fields of Candidate Path Signatures and Version Records

[0042] refer to Figure 5In this embodiment, during the vehicle's execution of the operation path, the system continuously receives newly generated trajectory segments and updates the local ranking of relevant path signatures based on the planning units corresponding to the trajectory segments. If the difficulty label of a certain segment in a path signature increases due to the addition of new samples, the system does not immediately discard the entire path. Instead, it recalculates the contribution of that segment and its adjacent segments to the total path cost and performs a local comparison with the candidate path signatures. If the candidate path can avoid the newly added high-difficulty segments and is consistent with the connection direction of the currently executed path, then the local segments of the candidate path can enter the replacement candidate set. If the new difficulty change only affects the segments that the vehicle has already passed, the system writes the change record to the version difference record without changing the currently unexecuted segments. The advantage of this embodiment is that the candidate path set and path signatures enable the path planning to have execution-period traceability, and can complete local ranking and candidate segment selection without recalculating the entire path.

[0043] In one embodiment, the system establishes an incremental update version chain for the grassland accessibility model. The incremental update version chain includes the planning unit, sub-label group, node implicit representation, transition node representation, and the associated version number of the accessibility difficulty estimate. Whenever new historical running data enters the system, the processor first determines whether the planning unit corresponding to the new data already has a sub-label group; if it does, the system updates the sample set and accessibility label of the corresponding sub-label group; if it does not exist, the system establishes a new sub-label group based on the passage time period, entry direction, and exit direction. The processor searches for adjacent planning units with directed edges connected to the affected planning unit and determines these planning units as the local training range. Local training only updates the affected planning units, adjacent planning units, and related transition node representations, without retraining the entire model. After training, the system writes the difference in accessibility estimates before and after the update into the version difference record and retains a stable version that can be called by the path planning cost graph. The stable version is used for current path planning and vehicle execution; the temporary version during training does not directly participate in path output.

[0044] In this embodiment, the version difference record can be calculated using the following formula: ; in, Indicates spatial index as The amount of version differences in the planning unit. This represents the estimated difficulty of passage obtained after local training. This represents the estimated difficulty of passage in the stable version before local training. This indicates a discrepancy in difficulty level; the value is 1 if the difficulty level changes, and 0 if it does not. This indicates the amount of difference between the current execution trajectory and the stable version level. For example, when , , , At that time, the version difference was If the estimated difficulty value changes little and no level changes, the version difference is mainly determined by the difference in estimated values. The processor determines whether to rebuild the local path planning cost map based on the version difference. The advantage of this embodiment is that newly added running data can enter the difficulty model through the version chain, while the current execution path still calls the stable version, avoiding frequent path switching caused by intermediate training results.

[0045] In one embodiment, when the unmanned vehicle executes a work path, the processor matches the newly generated trajectory segment with the current path signature. The matching process is based on the sequence of planning units, combined with the entry direction, exit direction, and time sequence for judgment. If the planning unit corresponding to the actual trajectory segment is consistent with the planning unit in the current path signature, and the directional difference is within the allowable range, the system determines that the vehicle will still execute along the current path signature; if the actual trajectory segment deviates from the current path signature but is still within the adjacent passable planning unit, the system calculates the impact of the deviation on subsequent path connections; if the planning unit corresponding to the actual trajectory segment has a level inconsistency record with the estimated passability in the stable version, the system locks the unaffected path signature segment and reconstructs the local path planning cost map only for the level inconsistent planning unit and its preceding and following connecting segments. The preceding and following boundaries of the reconstruction range are jointly determined by the location of high difficulty occurrence, the change of entry direction, and the location of the area that must be covered, to avoid local replacement disrupting the work continuity of unexecuted segments.

[0046] In this embodiment, after the local path planning cost map is reconstructed, the processor generates multiple candidate local paths and performs splicing verification with the current path signature. The splicing verification includes start-point connection, end-point connection, direction connection, coverage area connection, and version connection. Start-point connection confirms that the candidate local path can enter from the current vehicle's planning unit; end-point connection confirms that the candidate local path can return to the unaffected segment of the original path signature; direction connection confirms that the candidate local path's entry and exit directions are consistent with the preceding and following segments; coverage area connection confirms that the candidate local path does not omit any necessary coverage areas; and version connection confirms that the estimated travel difficulty value of the candidate local path comes from the same stable version or a replaceable version. If the splicing verification passes, the processor replaces the corresponding segment in the current path signature with the candidate local path and writes the replacement result to the version difference record. If the splicing verification fails, the system retains the current path signature and continues to search for local candidates in the affected segments. The advantage of this embodiment is that local changes during path execution can be limited to the affected segments, while unaffected path segments remain stable, allowing vehicles to maintain continuous operation paths in complex grasslands.

[0047] In a preferred embodiment, the system manages cases where the same planning unit generates opposite passage difficulty labels in different passage periods by grouping them. The processor divides the passage period into several time period groups, and each time period group, together with the entry and exit directions, constitutes a sub-label group index. If the same planning unit exhibits low passage difficulty in one time period group and high passage difficulty in another time period group, the system does not directly average the two labels but retains two sub-label groups. When the path planning cost graph calls the passage difficulty estimate, it selects the corresponding sub-label group based on the current operation period and the candidate path direction; if there is no corresponding sample in the current operation period, the system calls the sub-label group with higher confidence in the adjacent time period group and assigns the cost value based on the estimate output from graph structure learning. For opposite labels in different direction groups, the system selects the corresponding label according to the entry direction of the candidate path and does not directly write passage records with opposite directions into the same label calculation process.

[0048] In this embodiment, the sub-label group is also stored in association with the sample source and update time. The sample source is used to track which task and trajectory segment the passage difficulty label comes from, and the update time is used to determine the position of the label in the version chain. When a new sample differs significantly from the labels in the existing sub-label group, the processor first checks the credibility of the new sample; if the credibility of the new sample is low, it is only written into the observation record and the stable label is not updated; if the credibility of the new sample is high, the processor writes it into the sub-label group and triggers local training. Through this processing method, multiple passage states within the same planning unit are preserved as a searchable data structure, and the corresponding cost can be invoked based on the task time and driving direction during path planning. The advantage of this embodiment is that the passage difficulty label will not be distorted due to the mixing of samples from different directions or time periods, so that the dynamically changing passage performance in complex grasslands can be continuously expressed.

[0049] In a preferred embodiment, the system estimates the passage difficulty for planning units that have not formed valid passage samples. Planning units without valid passage samples include those never traversed by vehicles, those with only manual intervention samples, those with only non-operational movement samples, and those with sample completeness below the usage condition. During graph structure learning, the processor does not directly generate supervision labels for these planning units. Instead, it generates estimates using the implicit node representations and directed edge features of adjacent planning units. If an unsampled planning unit is adjacent to multiple sampled planning units, the system calculates the contribution of each adjacent node to its implicit representation based on directed edge propagation weights and directional consistency. If the unsampled planning unit is located between regions with significant label differences, the system prioritizes using transition node representations to avoid directly inheriting high or low difficulty labels from one side. If the unsampled planning unit is close to the grassland boundary, boundary distance and boundary curvature are used as node features in the estimation, ensuring that the estimated value does not deviate from the grassland spatial morphology.

[0050] In this embodiment, the estimated value of unsampled areas does not directly replace the true label. The system records the data source of this estimated value as model estimation in the path planning cost map, and after subsequent vehicles pass through the area, write the actual running samples into the corresponding planning unit to generate a new passage difficulty label. If the difference between the true label and the estimated value is large, the processor marks the planning unit as a model correction area and performs local training on its adjacent edge propagation weights. In this way, unsampled areas can still obtain a computable cost during the initial planning, and can enter the normal label update process after generating real samples. The advantage of this embodiment is that the system can handle areas in grasslands that have not been fully traversed, reduce blank areas in the path cost map, and gradually correct the estimation results with real running data in subsequent operations.

[0051] Furthermore, in this embodiment, the system can output the grassland operation path as a data object composed of a planning unit sequence, a direction sequence, and execution constraints. The planning unit sequence represents the spatial locations that the unmanned vehicle needs to traverse, the direction sequence represents the direction of entering and leaving each planning unit, and the execution constraints include the areas that must be covered, the prohibited areas, and the path connection sections. When outputting the path, the system also outputs the version of the grassland accessibility model, the version of the path planning cost map, and the path signature number. New trajectory segments formed during vehicle execution are associated with the above version records, so that subsequent path deviations or local cost changes can be traced back to the corresponding model version. If the same path signature is partially replaced multiple times, the system saves the start and end planning units of each replaced section, the local path before replacement, the local path after replacement, and the version difference record that triggered the replacement. This data object can be reused by subsequent operation tasks and can also be used as a new source of historical operation data to participate in the next round of accessibility learning. The advantage of this embodiment is that the path output result not only includes the executable route but also the model version and path signature, making path planning, path execution, and subsequent learning form a closed data processing flow.

[0052] In a preferred embodiment, the processor performs time-series alignment on the trajectory deviation sequence and motion response sequence. Since the time intervals for actual trajectory sampling, speed recording, steering correction recording, and traction output recording may differ, the system uses the actual trajectory timestamp as a reference to map the speed, steering correction, and traction output data to adjacent trajectory point intervals. If multiple speed records exist within the same trajectory point interval, the system calculates their average as the speed for that interval; if multiple steering correction records exist within the same trajectory point interval, the system counts the number of corrections and retains the maximum correction magnitude; if a traction output record lacks a certain time value, the system uses linear interpolation with adjacent records within the same trajectory segment. After time-series alignment, the processor maps each trajectory point interval to a planning unit, forming the trajectory deviation sequence and motion response sequence within the planning unit. If the number of trajectory points in a planning unit is too small, the system marks that sample as a low-completeness sample and reduces its impact through confidence calculation.

[0053] In this embodiment, temporal alignment is also used to determine the synchronous changes between trajectory deviation and motion response. If a vehicle experiences an increase in lateral offset within a planning unit, and simultaneously experiences a synchronous increase in speed decay, steering correction frequency, or traction output changes, the consistency of that sample is high. If only lateral offset occurs while the motion response remains stable, the system considers the sample potentially affected by positioning errors or short-term path tracking errors, reducing its reliability. If only motion response changes while the trajectory still conforms to the planned path, the system retains the sample but writes it into the motion response dominant label. Through the above processing, the passage difficulty label does not rely on a single data field but is formed based on the temporal correlation of multi-source operational data. The advantage of this embodiment is that the label generation process can adapt to different data sampling intervals and reduce the interference caused by single record anomalies in the passage difficulty judgment.

[0054] In a preferred embodiment, the processor calculates the steering continuity cost in the path planning cost graph. The steering continuity cost is determined by the change in the angle between the entry and exit directions of adjacent path segments. The steering continuity cost is low when the candidate path maintains a smooth direction change within a continuous planning unit; it increases when the candidate path needs to frequently change direction between adjacent planning units. The system does not use vehicle structural parameters as the basis for judgment, but rather calculates based on the path direction sequence and the adjacency relationship of planning units. For paths near the boundary, if a direction change is required to conform to the grassland boundary, the system also considers the boundary curvature characteristics; if the path direction change is consistent with the boundary curvature change, the steering continuity cost is corrected according to the boundary constraints; if the path direction change is unrelated to the boundary curvature and causes high-frequency steering, the system increases the corresponding local cost. In this way, path planning can both adapt to irregular boundaries and avoid forming unnecessary broken paths within the grassland.

[0055] In this embodiment, the processor writes the basic distance cost, accessibility cost, turning continuity cost, and historical risk decay cost into the same cost map, instead of performing independent screening separately. When generating candidate paths, the system calculates the total cost for each candidate path and simultaneously records the continuous length and location of high-accessibility planning units. If a candidate path, although short in distance, passes through multiple high-accessibility planning units, its total cost will increase with both accessibility cost and historical risk decay cost; conversely, if another candidate path is slightly longer but has a lower accessibility distribution and better turning continuity, it may be selected as the output path. This selection process is completed by the cost map calculation and does not rely on manual annotation. The advantage of this embodiment is that path planning can avoid path execution deviations caused by simply finding the shortest distance in complex grasslands and maintain good directional continuity in the output path.

[0056] In a preferred embodiment, the processor performs boundary control on local path replacement. If a planning unit with an inconsistent level is located ahead of the current vehicle, the system searches for the affected segment forward from the planning unit where the current vehicle is located. If the inconsistent planning unit has already been passed by a vehicle, the system only writes a version difference record and does not replace the current path signature. For segments that need to be replaced, the system extends forward to the nearest directionally stable planning unit and backward to a planning unit that can reconnect with the original path. A directionally stable planning unit is a planning unit whose entry and exit directions change little among multiple adjacent planning units, and whose corresponding passage difficulty level has not changed. Through this boundary control method, local replacement will not arbitrarily spread from the affected point to the entire path range, and unaffected path signature segments can continue to use the stable version.

[0057] In this embodiment, after the local candidate path is generated, the processor also checks its relationship with the area that must be covered. If the local replacement path bypasses the high-traffic-difficulty area but causes the necessary coverage area to be missed, the system does not accept the local path; if the local replacement path passes through the necessary coverage area but causes discontinuity with the original path direction, the system keeps it as a candidate and continues to search for a path with better directional connection; if multiple candidate local paths all meet the requirements of coverage area connection and directional connection, the system selects the path with the lower local total cost for replacement. After the replacement is completed, the new path signature retains the original path signature number and adds a replacement segment version record, so that the execution data can correspond to the path status before and after the replacement. The advantage of this embodiment is that online path adjustment can take into account local avoidance, coverage continuity, and version traceability, avoiding the destruction of the overall operation path due to local replanning.

[0058] Furthermore, in a preferred embodiment, the system stores historical running data, difficulty labels, model versions, path planning cost maps, and path signatures in a unified data structure. Each piece of historical running data points to one or more planning units, each planning unit points to multiple sub-label groups, each sub-label group points to the set of running samples participating in the calculation, each model version points to the corresponding node implicit representation and difficulty estimate, and each path signature points to the model version and cost map version called when generating the path. After receiving new data, the processor traces the affected planning units and path signatures according to the data structure and decides whether to trigger label updates, local training, cost map reconstruction, or path segment replacement. For data that has not triggered path segment replacement, the system still writes it into the historical running data for subsequent model version training. For data that has triggered replacement, the system saves the replacement result along with the triggering conditions, enabling the next round of path planning to identify that the region has undergone execution-time path adjustments.

[0059] In this embodiment, historical operational data is entered into the planning unit index after coordinate unification. The processor generates sample calculation values ​​based on trajectory deviation and motion response, and then filters or downweights abnormal samples through confidence stratification. Samples within the same planning unit are grouped into sub-labels according to time period and direction, and these sub-labels generate accessibility labels. The grassland accessibility model is trained using planning units as graph nodes and continuous accessibility relationships as directed edges, generating learning results for sampled areas and estimated values ​​for unsampled areas. The path planning cost graph calls the accessibility estimate and combines distance, steering continuity, and historical risk decay to generate candidate paths. New data generated during vehicle execution enters the incremental update version chain, and affected sections are partially replaced through path signatures. Through the above data processing chain, the intelligent path planning system for unmanned vehicles in complex grasslands can continuously use real operational feedback to correct path costs. The advantage of this embodiment is that the system has an implementable computational process from data acquisition, label generation, model training, path planning to execution-period updates, and can fully cover the technical implementation of dynamic accessibility learning and path planning in complex grasslands.

Claims

1. An intelligent path planning system for unmanned vehicles operating in complex grasslands, characterized in that: It includes a processor and a memory, the memory storing program instructions executable by the processor, which, when executed, cause the processor to: Acquire historical operational data of unmanned vehicles operating in complex grasslands, including planned path, actual trajectory, speed decay, steering correction, traction output changes, and operational status. The grassland operation area is divided into planning units with spatial and temporal indexes; Traffic difficulty labels are generated based on historical operational data within the same planning unit. A grassland access difficulty model is trained based on the access associations between adjacent planning units; The system constructs a path planning cost map based on the grassland access difficulty model and outputs the operation path of the unmanned vehicle. When dividing the planning unit, the processor performs coordinate unification and effective operation range closure processing on the boundary of the grassland operation area, and maps the planned path and the actual trajectory to the same spatial reference system. The size of the planning unit is determined based on the operating width of the unmanned vehicle, the density of historical trajectories, and the curvature of the boundary. An index record containing the entry direction, exit direction, passage time period, and adjacent relationships is established for each planning unit. The trajectory segments spanning multiple planning units are segmented and assigned to different units, so that the historical operation data and the corresponding planning units form updatable data pairs. When generating the accessibility difficulty label, the processor determines the trajectory deviation sequence based on the lateral offset, heading offset, and return time between the planned path and the actual trajectory. The motion response sequence is determined based on the speed decay, steering correction, and traction output changes. The segments in the operational status that represent temporary parking, manual takeover, non-operational movement, or task switching are marked as samples to be excluded; Within the same planning unit, the trajectory deviation sequence and the motion response sequence are time-aligned to generate passage difficulty labels associated with the entry direction and passage time period.

2. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 1, characterized in that, When training the grassland access difficulty model, the processor constructs each planning unit as a graph node and constructs adjacent planning units with continuous access relationships as directed edges. The node features and edge features are formed using the aforementioned passage difficulty labels, boundary distances, historical passage counts, differences in entry directions, and differences in adjacent node labels; The implicit representation of nodes is updated using a graph structure learning approach, and the estimated passage difficulty is output for planning units that have not formed effective passage samples. Different sample weights were set for manually taken-over samples and non-continuous passage samples during the training process.

3. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 2, characterized in that, When constructing the path planning cost map, the processor configures a basic distance cost, a passage difficulty cost, a turning continuity cost, and a historical risk attenuation cost for each planning unit. The passage difficulty cost is output by the grassland passage difficulty model, the turning continuity cost is determined by the entry and exit directions of adjacent path segments, and the historical risk attenuation cost is determined by the most recent abnormal passage time and the number of abnormal repetitions. A set of candidate paths is generated in the global search, and the operation path of the unmanned vehicle is selected according to the total cost and path continuity of each candidate path.

4. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 3, characterized in that, The processor also performs credibility stratification on multiple historical operation data within the same planning unit. First, it determines the record completeness based on timestamp continuity, positioning sampling interval, trajectory segment completeness, and operation status consistency. Then, it determines the sample consistency based on the synchronous changes between the trajectory deviation sequence and the motion response sequence. When the same planning unit has opposite passage difficulty labels at different passage times, sub-label groups are established according to the passage time, entry direction, exit direction and adjacent planning unit status, and the sub-label groups are associated with and stored with the corresponding planning unit, passage sample source and update time.

5. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 4, characterized in that, The processor also sets direction-sensitive propagation constraints in the directed edges. The direction-sensitive propagation constraints are jointly determined by the departure direction of the previous planning unit, the entry direction of the next planning unit, the boundary curvature change, and the historical trajectory offset direction. When the difference in accessibility labels between adjacent planning units exceeds the preset hierarchical conditions, direct label propagation is stopped and a transition node representation is established; When a planning unit that has not formed a valid passage sample is located between multiple transition nodes, the passage difficulty estimate is generated based on the distance attenuation and directional consistency represented by the transition nodes.

6. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 5, characterized in that, When generating a set of candidate paths, the processor first uses the starting point, ending point, and required coverage area of ​​the grassland operation task to form basic search constraints, and then retains multiple paths with similar costs but different spatial distributions in the path planning cost graph. For each candidate path, generate a path signature consisting of continuous planning units, and record the continuous length, occurrence position, and entry direction of the planning units with high traversability difficulty in the path signature; During the operation of the unmanned vehicle, the corresponding path signatures are locally sorted and updated based on the newly generated passage difficulty labels.

7. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 6, characterized in that, The processor also establishes an incremental update version chain for the grassland access difficulty model, the incremental update version chain including planning units, sub-label groups, node implicit representations, transition node representations and the associated version number of the access difficulty estimate; When the newly added historical running data only affects some planning units, local training is only performed on the affected planning units, adjacent planning units with directed edge connections, and the corresponding sub-label groups. After local training is completed, the difference in the estimated travel difficulty before and after the update is written into the version difference record, and a stable version is retained for use in the path planning cost graph.

8. The intelligent path planning system for unmanned vehicles operating in complex grasslands according to claim 7, characterized in that, The processor continuously receives newly generated trajectory segments and matches them with the current path signature when the unmanned vehicle is executing the unmanned vehicle's work path; When the planning unit corresponding to the trajectory segment forms a level inconsistency record with the estimated travel difficulty value in the stable version, the unaffected path signature segment is locked, and the local path planning cost map is reconstructed only for the planning unit with inconsistent level and the connecting segments before and after it. The reconstructed candidate local paths are subjected to signature splicing verification. If the splicing verification passes, the corresponding segment in the current path signature is replaced, and the replacement result is written into the version difference record.

Images