Lawn robot dynamic task scheduling method based on artificial intelligence

Abstract

The invention belongs to the field of lawn robot dynamic task scheduling, and particularly relates to a lawn robot dynamic task scheduling method based on artificial intelligence, which comprises the steps of deploying sensors to acquire data and accessing external data to generate an input set; calculating the area of a regional shadow zone, dividing the regional shadow zone into hybrid task slices according to blocks and time, and generating a hybrid task slice set; constructing an evaluation algorithm to calculate the workable probability of the lawn robot, and generating a working schedule of each task slice; calculating priority scores of the task pieces, and training by adopting a priority scheduling algorithm to generate a candidate task sequence; setting a passing token and a charging token for the lawn robot, and generating an execution list and a token queue; splicing the executable shadow zones, performing path re-planning on the lawn robot, and generating a forward re-planning path; according to the invention, the completion time can be shortened and invalid driving can be reduced, so that the lawn area which is not treated and shielded is reduced, and the harvesting energy consumption is reduced.

Description

Technical Field

[0001] This invention belongs to the field of dynamic task scheduling for lawn robots, and specifically relates to a dynamic task scheduling method for lawn robots based on artificial intelligence. Background Technology

[0002] The Chinese patent application CN202510231890.5 discloses a visual SLAM method for intelligent robots applicable to dynamic lawn scenes. This invention extracts global feature points from images; matches feature points in adjacent frames; combines object detection and epipolar geometry for dynamic feature point detection, filtering out potential dynamic feature points contained within the global feature points; and designs the visual odometry part of the SLAM system as a dual-threaded parallel method, enabling information exchange between the two threads on the same host through sockets.

[0003] While this invention improves the robot's navigation performance in complex and dynamic environments and ensures its ability to accurately understand and respond to changes in the surrounding environment, it still presents the following problems in mountainous or plain photovoltaic power plant sites: Existing lawn robots mostly perform mowing according to fixed cycles and fixed areas; however, weather conditions, such as rain causing muddy roads and water accumulation, insufficient power, and maintenance shutdowns, often lead to work delays and plan mismatches. Rainfall reduces the accessibility of harvesting paths and slows down the operation, resulting in more rework and increased energy consumption, negatively extending completion time. Furthermore, mountainous and plain areas differ in sunlight and water / heat conditions, leading to significant spatial unevenness in grass growth: high energy consumption for mowing in areas of vigorous growth and low energy consumption in areas of sparse growth; existing robots that allocate resources to fixed areas suffer from resource allocation issues. The imbalance between task load and workload leads to the following problems: robots responsible for sparse areas can occupy charging spots to return to charging when they still have sufficient power, while robots responsible for lush areas are hindered by occupied spots when they run out of power or need to charge after completing their work, resulting in charging resource conflicts and unreasonable scheduling; in addition, in photovoltaic power station scenarios with sufficient sunlight, robot operations are prone to overheating and increased energy consumption; when grass grows too vigorously and is close to the edge of the module, there is a risk of touching the module and causing safety hazards; in summary, the existing technology lacks adaptive prioritization based on environmental and load differences and cross-regional resource coordination capabilities, and cannot dynamically adjust the task order and charging spot occupation according to multiple factors such as weather, growth momentum, accessibility, energy, and safety, making it difficult to achieve priority treatment of key areas and reasonable global resource scheduling, such as priority harvesting of shading sensitive areas, the area around the module, and firebreaks.

[0004] To address the aforementioned issues, this invention proposes a dynamic task scheduling method for lawn robots based on artificial intelligence. This method involves modeling shadow zones driven by the sun's trajectory—dividing the workable timetable into hybrid task slices based on blocks and time—generating candidate sequences through priority scoring with historical regression weights and learning-based scheduling—jointly issuing corridor access tokens and charging tokens—and using a closed loop of incremental rearrangement and shortest path replanning by stitching shadow zones in the same direction under token constraints. This approach minimizes shading loss and total completion time, eliminates oncoming traffic congestion and charging contention, and reduces ineffective driving energy consumption while ensuring maintenance window priority and safety constraints. Summary of the Invention

[0005] This invention provides an AI-based dynamic task scheduling method for lawnmowers, aiming to solve the following problems: Existing robot scheduling methods mostly mow surrounding lawns according to fixed cycles, but lawnmower mowing can be delayed due to weather conditions, insufficient robot power, and machine maintenance; In addition, due to differences in sunlight, the growth of grass in mountainous and plain areas is different, and the energy consumption of mowing is closely related to the growth of grass. The more vigorous the grass growth, the higher the energy consumption, and the sparser the grass growth, the lower the energy consumption. Existing lawnmowers mow lawns according to fixed areas, which leads to uneven resource allocation; Furthermore, in mountainous and plain areas with abundant sunlight, lawnmowers may overheat during operation, as excessive grass growth in abundant sunlight may touch the photovoltaic panels and cause a fire. Existing technologies lack the adaptive ability to prioritize the treatment of critical areas.

[0006] The technical solution adopted by the present invention to solve the above problems is: a dynamic task scheduling method for lawn robots based on artificial intelligence, including: step S1, deploying sensors to collect data and accessing external data, and generating an input set from the collected data and external data;

[0007] Step S2: Based on the input set, calculate the area of ​​the region shaded zone, divide the region shaded zone into mixed task slices according to blocks and time, and generate a set of mixed task slices;

[0008] Step S3: Based on the mixed task slice set, calculate the probability of the lawn robot being able to work by constructing an evaluation algorithm, and generate the work schedule for each task slice;

[0009] Step S4: Based on the work schedule of each task slice and the mixed task slices, calculate the priority score of the task slices, train the priority scheduling algorithm, and generate candidate task sequences.

[0010] Step S5: Based on the candidate task sequence, set access tokens and charging tokens for the lawn robot, and generate an execution list and token queue;

[0011] Step S6: Based on the execution list and token queue, the executable shadow bands are spliced ​​together, the path of the lawn robot is replanned, and a forward replanning path is generated.

[0012] Step S7: Based on steps S1-S6, perform backhaul and closed-loop to generate a globally optimized dynamic adjustment strategy.

[0013] As a preferred embodiment, the specific steps for calculating the area of ​​the shaded region are as follows:

[0014] Based on the solar incident vector SUN, azimuth angle ANGLE, photovoltaic panel array arrangement matrix JU(x,y) in the input set of step S2, and the set parameters of each photovoltaic module in the field, each photovoltaic module is projected parallel to the ground axis z=0 along the unit vector -SUN, which is opposite to the direction of sunlight propagation, to obtain the shadow domain Shdow. An area exponential function is then constructed to calculate the shadow area of ​​the projected shadow domain, used to quantitatively characterize the shadow coverage of the block × time unit. The formula for the area exponential function is:

[0015] ,

[0016] Let SIZE be the area of ​​the shadow region where the ground point (x, y) falls at time t, representing the shadow exponential function. Let x be the area of ​​the shadow region where x points east and y points north. For indicating function symbols, The dot product symbol. Let t be the shaded region at time step t, where t is the time step and Shdow is the shaded region.

[0017] As a preferred embodiment, the specific steps for dividing the region shadow band into hybrid task slices according to blocks and time, and generating a set of hybrid task slices are as follows:

[0018] Based on the obtained area shaded region, the site is spatially discretized within the planning time window according to the operation corridor or management grid. Specifically, spatial discretization according to the operation corridor involves dividing the site into spatial sub-segments SP along the center line of the corridor using a fixed step size. Then, temporal discretization is performed: equally spaced time slices TP are generated at the fixed step size. On the Cartesian product of the above spatial sub-segments and time slices, a hybrid task unit MD is constructed. For each hybrid task slice, the following are calculated and recorded: area shaded region SIZE, area shaded region Shdow, area accessibility and geographical attributes, maintenance restricted access time window and charging accessibility, and workable marker. The hybrid task slice MD and its attributes are stored in the hybrid task slice set GET in a block-first, time-ordered manner and provided externally as standard fields, forming an entry with a clear start and end time and continuous operation length.

[0019] As a preferred implementation, the specific steps for generating the work schedule for each task segment by constructing an evaluation algorithm to calculate the operational probability of the lawn robot based on the hybrid task segment set are as follows:

[0020] Based on the hybrid task slice set GET, an evaluation algorithm is constructed to assess the operational probability of each task slice in the hybrid task slice set and generate a work schedule for each task slice.

[0021] The steps for defining the evaluation algorithm are as follows:

[0022] R1 is the driving sequence ALL of the grey model, composed of surface humidity W, wind speed F, cloud cover WIN, and ground temperature TEP from the input set. The calculation formula for the driving sequence is as follows:

[0023] ,

[0024] ALL is the driving sequence. This is the surface humidity weighting coefficient, where W is the surface humidity. It is a weighting factor for wind speed. The maximum wind speed for operation is F, where F is the wind speed. WIN is the weighting factor for cloud cover. It is the weighting coefficient for ground temperature, TEP is the ground temperature, and TEP1 is the calibrated ground temperature;

[0025] By driving sequences, environmental factors such as humidity, wind speed, cloud cover, and ground temperature are mapped into monotonic signals of workability, providing quantitative input for the evaluation algorithm;

[0026] R2, based on the driving sequence, obtains the predicted values ​​of the driving sequence through grey model fitting. The algorithm is used to evaluate the probability that the lawn robot can operate, and the calculation formula for the evaluation algorithm is as follows:

[0027] ,

[0028] Let be the workability probability function, representing the probability that block index i and time slice index j are workable. This represents a workable function, where i represents the block index and j represents the time slice index. This represents the activation function. Indicates the calibration coefficient. Represents the predicted value of the driving sequence;

[0029] The workability probability of each task slice is obtained by constructing an evaluation algorithm, and a workability threshold is set based on the input set. 1. Determine whether the area is a restricted access window for maintenance, the corridor width threshold Hallmin, and the slope threshold SLmax. If all conditions are met, mark it as a workable area to obtain the workable area.

[0030] R3, based on the workable area, records the time slice index of the expected start and end times of workable time. By merging adjacent time slices within the same corridor Hall that are all workable according to time continuity, the work schedule item is obtained. All work items are written into a set to obtain the work schedule WrokTab.

[0031] As a preferred embodiment, the specific steps for calculating the priority score of a task segment based on the work schedule of each task segment and the mixed task segments are as follows:

[0032] Based on the work schedule WrokTab and the mixed task slice set GET, features are constructed for each work time item in WrokTab and normalized to the [0,1] interval. A priority scoring algorithm is constructed to calculate the priority score of each work time item. This score is used to convert the task items in the work schedule WrokTab and the mixed task slice set GET into comparable scores. Candidate task sequences are formed in descending order of benefits and constraints, providing a direct basis for token issuance and execution list distribution in step S5.

[0033] As a preferred embodiment, the specific steps for calculating the priority score of a task segment based on the work schedule of each task segment and the mixed task segments further include:

[0034] The definition steps of the priority scoring algorithm are as follows:

[0035] T1, based on each work time item, extract and construct scoring elements from the item: shadow bandwidth intensity, available window, reachability and rainfall penalty, and normalize the extracted elements to obtain the construction elements of the priority scoring algorithm;

[0036] T2, Based on the constructive elements of the priority scoring algorithm, a priority scoring algorithm is constructed to obtain the priority score of each task objective. The formula for the priority scoring algorithm is:

[0037] ,

[0038] Where Mark represents the priority score. Here, EShdow represents the shadow bandwidth intensity weighting factor. WW represents the available window weight coefficient. Here, ARR is the accessibility weighting coefficient, and ARR is the accessibility rating. Here, is the rainfall penalty weighting coefficient, and Prain is the rainfall penalty. Here, WA represents the cross-region cost weighting coefficient. PLwork represents the workability probability weighting coefficient, and PLwork is the workability probability function.

[0039] As a preferred implementation, the specific steps for training using a priority scheduling algorithm to generate a candidate task sequence are as follows:

[0040] By reading data from the work schedule WrokTab and the mixed task slice set GET, an original element set is obtained, which serves as the basis for issuing tokens and distributing execution lists in step S5. A priority scheduling algorithm is used for training, and historical work records are collected to form a training set. A benefit label BE is constructed from historical samples, and a scoring function is obtained by training priority scoring weights and scoring thresholds using ridge regression. Hard and soft constraints are set for the scoring function to obtain the feasible region. A sliding window is set to recalculate Mark, and incremental reordering is triggered when the rainfall probability RAIN, surface humidity W, and robot battery power change, resulting in an online adaptive candidate task queue.

[0041] As a preferred implementation, the specific steps for setting access tokens and charging tokens for the lawn robot based on the candidate task sequence, and generating an execution list and token queue are as follows:

[0042] Based on the candidate task sequence, a time slot table is established by traversing the candidate task sequence. Access tokens and charging tokens are sent to the corridor and charging station corresponding to each item, respectively, generating a token queue. Specifically: For the access queue established at the corridor entrance, access tokens are issued to lawn robots, and the order of robots waiting to enter the corridor and their access tokens is recorded for one-way passage and to avoid passing vehicles; for the charging queue established at the exit, charging tokens are issued to lawn robots that have completed their tasks, access tokens are collected, and the order of robots waiting to occupy the charging station and their charging tokens is recorded for orderly queuing and on-arrival service, resulting in one-way passage and orderly occupancy; by reading the token queue and the candidate task queue, an ordered list of execution items is obtained.

[0043] As a preferred implementation, the specific steps for splicing executable shadow bands based on the execution list and token queue to replan the path of the lawn robot and generate a forward replanning path are as follows:

[0044] By reading the execution list and token queue, the available task pieces from the acquired token entries are selected and spliced ​​into a strip channel according to the direction consistency rule. D* Lite is used to perform short-circuit search in the channel graph, where the cost is set as: distance + reverse penalty + departure penalty, and yield point and occupation conflict constraints are included to obtain the forward replanning path.

[0045] As a preferred implementation, the specific steps for generating a globally optimized dynamic adjustment strategy based on steps S1-S6, which involve performing backhaul and loop closure, are as follows:

[0046] By collecting and transmitting the remaining battery power, token queue status, and execution list status of the lawn robot, and synchronizing the timestamps according to the time of the input set in step S1, a unified data stream that can be used for online judgment is obtained as the basis for triggering judgment. By setting triggers {RAIN≥0.3, W≥surface humidity threshold Wmin, lawn robot operating body temperature>65°C, detection of passage obstacle, change in token queue, and update of maintenance prohibition time window}, an event queue is obtained for incremental rearrangement. By recalculating Mark in step S4, scanning all items in descending order of Mark according to the calculated Mark, and applying a greedy algorithm + taboo list to adjust the queue and tokens, and partially updating the forward replanning path and execution list in step S6, a dynamic adjustment strategy is obtained.

[0047] The beneficial effects of this invention are as follows:

[0048] 1. By integrating weather, power, and maintenance window data to generate a dynamic input set and rearranging it in real time, the system can achieve immediate avoidance and task replacement scheduling for tasks that are slowed down by mud and water accumulation, thereby reducing delays, rework, and energy consumption.

[0049] 2. By incorporating grass growth intensity and energy consumption per unit area into priority scoring, and combining access tokens and charging tokens to coordinate resource occupation control, a task and charging sequence is obtained that prioritizes lush areas and gives way to sparse areas. This can alleviate the uneven distribution of resources between areas and prevent high-load robots from running out of power and charging piles from being occupied by low-load tasks for a long time.

[0050] 3. By constructing key area weights based on shadow zone, light intensity, and temperature rise threshold, and triggering path replanning and operation window adjustments during periods of overheating or strong light, priority is given to the treatment of areas around photovoltaic panels and high-risk corridors, and temperature rise avoidance is achieved. This can reduce energy consumption and equipment damage, and reduce the risk of fire caused by touching photovoltaic panels. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention.

[0052] Figure 2These are technical effect diagrams of the present invention, wherein the black bar chart represents the present invention, and the gray bar chart represents the prior art. Detailed Implementation

[0053] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0054] Example 1 This application proposes a dynamic task scheduling method for lawn robots based on artificial intelligence, combined with Figure 1 As shown, a dynamic task scheduling method for lawn robots based on artificial intelligence includes:

[0055] Step S1: Deploy sensors to collect data and access external data, and generate an input set from the collected data and external data;

[0056] Specifically, data is collected by deploying meteorological stations, surface humidity sensors, anemometers, ground temperature sensors, light meters, corridor width laser rangefinders, and slope sensors in the site area, and external data is also accessed, including hourly weather forecasts, maintenance work orders, and power-on time periods. The data sampling period is set to 5 minutes for sensor data collection and external data retrieval. All data collected by the sensors and externally accessed data are preprocessed, including noise reduction and missing measurement linear interpolation. After unifying the timestamps and aligning them, the units are normalized and written to the cloud database in a fixed format to obtain a structured input set, providing a consistent data source for steps S2-S7.

[0057] The fixed format is based on the following parameters: solar incident vector (SUN), azimuth angle (ANGLE), cloud cover (WIN), rainfall probability (RAIN), wind speed (F), surface humidity (W), ground temperature (TEP), corridor width (Hall), slope (SL), access level (DE), maintenance restricted access time window (FIX), charging position status (CE), and robot status (ST).

[0058] Step S2: Based on the input set, calculate the area of ​​the region shaded zone, divide the region shaded zone into mixed task slices according to blocks and time, and generate a set of mixed task slices;

[0059] Specifically, based on the solar incident vector SUN, azimuth angle ANGLE, photovoltaic panel array arrangement matrix JU(x,y) in the input set of step S2, and the set parameters of each photovoltaic module in the field, each photovoltaic module is projected parallel to the ground axis z=0 along the unit vector -SUN, which is opposite to the direction of sunlight propagation, to obtain the shadow domain Shdow. Here, x points east, y points north, and z is the vertical coordinate pointing towards the sky to represent the height of the module and to calculate the projection onto the ground. The origin is set at the field measurement reference point. An area exponent function is constructed to calculate the shadow area of ​​the projected shadow domain, which is used to quantitatively characterize the shadow coverage of the block × time unit. The formula for the area exponent function is:

[0060] ,

[0061] Let SIZE be the area of ​​the shadow region where the ground point (x, y) falls at time t, representing the shadow exponential function. Let x be the area of ​​the shadow region where x points east and y points north. For indicating function symbols, The dot product symbol. The shaded region is defined at time step t, where t is the time step and Shdow is the shaded region.

[0062] Based on the obtained area shaded region, the site is spatially discretized within the planning time window according to the operation corridor or management grid. Specifically, spatial discretization according to the operation corridor involves dividing the site into spatial sub-segments SP along the center line of the corridor using a fixed step size. Then, temporal discretization is performed: equally spaced time slices TP are generated at a fixed step size. Hybrid task units MD are constructed on the Cartesian product of the above spatial sub-segments and time slices. For each hybrid task slice, the following are calculated and recorded: area shaded region SIZE, area shaded region Shdow, area accessibility and regional attributes, maintenance restricted access time window and charging accessibility, and workable marker. The hybrid task slice MD and its attributes are stored in the hybrid task slice set GET in a block-priority and time-ordered manner and provided as standard fields. These fields are used to merge adjacent and workable segments within the same block according to time continuity, forming entries with clear start and end times and continuous operation lengths, which serve as input for generating work schedules and calculating priority scores.

[0063] The origin is set at the site measurement benchmark point. By selecting a permanent marker point in the site and conducting a total station traverse closure measurement, and then connecting it with the national coordinate network, a local ENU coordinate system is established at the origin. After least square adjustment, the plane and elevation closure differences are ≤4mm and ≤7mm, respectively. The coordinates of this point are determined as the measurement benchmark point POINT and fixed as the origin for record.

[0064] Step S3: Based on the mixed task slice set, calculate the probability of the lawn robot being able to work by constructing an evaluation algorithm, and generate the work schedule for each task slice;

[0065] Specifically, based on the hybrid task slice set GET, an evaluation algorithm is constructed to evaluate the operability probability of each task slice in the hybrid task slice set and generate a work schedule for each task slice.

[0066] The steps for defining the evaluation algorithm are as follows:

[0067] R1 is the driving sequence ALL of the grey model, composed of surface humidity W, wind speed F, cloud cover WIN, and ground temperature TEP from the input set. The calculation formula for the driving sequence is as follows:

[0068] ,

[0069] ALL is the driving sequence. This is the surface humidity weighting coefficient, where W is the surface humidity. It is a weighting factor for wind speed. The maximum wind speed for operation is F, where F is the wind speed. WIN is the weighting factor for cloud cover. It is the weighting coefficient for ground temperature, TEP is the ground temperature, and TEP1 is the calibrated ground temperature;

[0070] By driving sequences, environmental factors such as humidity, wind speed, cloud cover, and ground temperature are mapped into monotonic signals of workability, providing quantitative input for the evaluation algorithm;

[0071] R2, based on the driving sequence, obtains the predicted values ​​of the driving sequence through grey model fitting. The algorithm is used to evaluate the probability that the lawn robot can operate, and the calculation formula for the evaluation algorithm is as follows:

[0072] ,

[0073] Let be the workability probability function, representing the probability that block index i and time slice index j are workable. This represents a workable function, where i represents the block index and j represents the time slice index. This represents the activation function. Indicates the calibration coefficient. Represents the predicted value of the driving sequence;

[0074] An evaluation algorithm is constructed to obtain the workability probability of each task slice. To determine whether a task slice can execute a job, a job rate threshold is set based on the input set. The system determines whether the area is within the restricted access window for maintenance, the corridor width threshold (Hallmin), and the slope threshold (SLmax). The area is marked as workable only after all of the following conditions are met, thus obtaining the workable area: The block does not fall into the maintenance access window. ,and ;

[0075] R3, based on the workable area, records the time slice index of the expected start and end times of workable time. By merging adjacent time slices that are all workable within the same corridor Hall according to time continuity, a work schedule item is obtained. All work items are written into a set to obtain the work schedule WrokTab, which provides clear time boundaries, continuous work length and work confidence for priority scoring and scheduling training in step S4.

[0076] The grey model described above uses a time series prediction model based on a single-variable first-order differential equation. By accumulating the original sequence and estimating the parameters using least squares, a smoothed prediction value is obtained to calculate the probability of workability.

[0077] Step S4: Based on the work schedule of each task slice and the mixed task slices, calculate the priority score of the task slices, train the priority scheduling algorithm, and generate candidate task sequences.

[0078] Specifically, based on the work schedule WrokTab and the mixed task slice set GET, features are constructed for each work time item in WrokTab and normalized to the [0,1] interval. A priority scoring algorithm is constructed to calculate the priority score of each work time item. This score is used to convert the task items in the work schedule WrokTab and the mixed task slice set MD into comparable scores. Candidate task sequences are formed in descending order of benefits and constraints, providing a direct basis for token issuance and execution list distribution in step S5.

[0079] The definition steps of the priority scoring algorithm are as follows:

[0080] T1, based on the following contained in each work time item: Halli, tstrat, tend, con, PLwork, and Where Halli represents the block identifier at corridor field i, tstrat is the start time indicating the start time of the continuous workable segment, tend is the end time indicating the end time of the continuous workable segment, con is the continuous work length indicating the spatial continuous distance covered by the entry, and PLwork is the intra-segment representation of the workability probability. The sum of the shadow areas of the segments represents the sum of the shadow areas of all mixed task slices within this item, where i represents the block index, j represents the time slice index, and Size is the shadow exponential function; the following scoring elements are extracted and constructed from the item: shadow bandwidth intensity, available window, reachability, and rainfall penalty, and the extracted elements are normalized to obtain the construction elements of the priority scoring algorithm;

[0081] T2, Based on the constructive elements of the priority scoring algorithm, a priority scoring algorithm is constructed to obtain the priority score of each task objective. The formula for the priority scoring algorithm is:

[0082] ,

[0083] Where Mark represents the priority score. Here, EShdow represents the shadow bandwidth intensity weighting factor. WW represents the available window weight coefficient. Here, ARR is the accessibility weighting coefficient, and ARR is the accessibility rating. Here, is the rainfall penalty weighting coefficient, and Prain is the rainfall penalty. Here, WA represents the cross-region cost weighting coefficient. PLwork represents the workability probability weighting coefficient, and PLwork is the workability probability function.

[0084] The priority scoring algorithm can obtain the priority score of each task, which can be used for the dynamic scheduling of lawn robots. It can prioritize the execution of high-priority tasks, thereby avoiding fires and optimizing task scheduling.

[0085] By reading data from the Work Schedule (WrokTab) and the Mixed Task Piece Set (GET), a raw element set is obtained, which serves as the basis for issuing tokens and distributing execution lists in step S5. A priority scheduling algorithm is used for training, and historical job records are aggregated to form a training set. A benefit label (BE) is constructed from historical samples, and a scoring function is obtained by training priority scoring weights and scoring thresholds using ridge regression. Hard and soft constraints are set for the scoring function. The hard constraints are: the planned start and end times must not conflict with the maintenance restricted access time window of the block; access tokens and charging tokens must be available at the planned time; the lawn robot's battery level must be greater than the battery threshold; and only one task is allowed to be active in the same block at any given time. The soft constraints are: the number of lawn robot cross-area switching times ≤ the cross-area switching threshold; and the total field-changing travel distance ≤ the total field-changing travel threshold. This yields the feasible region. All items are then scanned in descending order by Mark. The algorithm employs a greedy algorithm and a tabu list to select tasks with higher marginal returns that do not violate constraints within the feasible region. It simultaneously updates the tabu list and feasible region until no feasible gains are available, resulting in a conflict-free candidate task sequence. This sequence is used to mark and temporarily prohibit candidates that conflict with selected items in terms of time, space, or resources during task selection. By assigning start and end times and required token types to selected items, a candidate task queue, corresponding token queue, and execution list are generated. A sliding window is set to recalculate the Mark every 30 seconds, and incremental reordering is triggered when the rainfall probability (RAIN), surface humidity (W), and robot battery level change, resulting in an online adaptive candidate task queue. For example, if the rainfall probability (RAIN) changes from 0% to 100%, or the surface humidity (W) exceeds the surface humidity threshold, muddy ground may cause the lawn robot to slip, or the robot's battery may be insufficient to complete the task, requiring task reordering.

[0086] The greedy algorithm described above obtains a fast and feasible sequence of candidate tasks by iteratively adding the task with the highest marginal benefit that does not violate the constraints within the feasible region each time.

[0087] The taboo list obtains a dynamic shielding and rearrangement guidance mechanism for conflicts by registering candidate items that conflict with selected items in terms of time, space or resources as temporarily prohibited and setting a taboo period.

[0088] The priority scoring weights and scoring thresholds trained by ridge regression are obtained by constructing feature vectors and benefit labels from historical samples, minimizing them using ridge regression, and then searching for the scoring thresholds on the validation set based on accuracy. These are then used for feasible region screening and candidate task sequence construction.

[0089] Step S5: Based on the candidate task sequence, set access tokens and charging tokens for the lawn robot, and generate an execution list and token queue;

[0090] Specifically, based on the candidate task sequence, a time slot table is established for the candidate task sequence by traversing the candidate task sequence, and access tokens and charging tokens are sent to the corridor and charging position corresponding to the item, respectively, to generate a token queue. Specifically: for the access queue established according to the corridor entrance, access tokens are issued to the lawn robot, and the order of the robots waiting to enter the corridor and their access tokens is recorded for one-way passage and to avoid meeting each other; for the charging queue established according to the exit, charging tokens are issued to the lawn robot that has completed its task, access tokens are collected, and the order of the robots waiting to occupy the charging position and their charging tokens is recorded for orderly queuing and on-arrival service, thereby achieving one-way passage and orderly occupancy, and reducing meeting each other and waiting; by reading the token queue and the candidate task queue, an ordered list of execution items is obtained;

[0091] Step S6: Based on the execution list and token queue, the executable shadow bands are spliced ​​together, the path of the lawn robot is replanned, and a forward replanning path is generated.

[0092] Specifically, by reading the execution list and token queue, the operable task segments (MDs) from the acquired token entries are selected, and the task segments are spliced ​​into strip-shaped channels according to the direction consistency rule. For example, if the centerline heading of task segment MD is approximately 90° eastward, satisfying the direction consistency rule and being time-continuous, then the strip-shaped channel is obtained by performing operations on adjacent segments of the same corridor. If there is a single time gap and the time gap is less than the gap threshold, it is filled according to the operable flags and splicing continues. For cross-corridor transitions, continuous channels are spliced ​​only when the heading difference of the transition segment is ≤30° and there is a yield point. D* Lite is used to perform short-circuit search on the channel map, where the cost is set as: distance + reverse penalty + departure penalty, and yield point and occupancy conflict constraints are included to obtain the forward replanning path.

[0093] The short-circuit search using D* Lite involves running D* Lite on the strip channel graph to incrementally update the edge cost consisting of distance + reverse penalty + exit from the shaded zone penalty, thereby obtaining the shortest feasible path that converges quickly under obstacle and token changes, which can reduce replanning overhead.

[0094] The yielding point is determined by using the nearest avoidable location within the corridor, either preset or dynamically selected, such as a location with sufficient width, no restricted access, and where temporary parking is permitted. This allows for an orderly yielding and de-congestion strategy when vehicles are occupying the road, thereby avoiding traffic congestion.

[0095] Step S7: Based on steps S1-S6, perform backhaul and closed-loop to generate a globally optimized dynamic adjustment strategy;

[0096] Specifically, by collecting and transmitting the remaining battery power, token queue status, and execution list status of the lawn robot, and synchronizing the timestamps according to the time of the input set in step S1, a unified data stream that can be used for online judgment is obtained as the basis for triggering judgment. By setting triggers {RAIN≥0.3, W≥surface humidity threshold Wmin, lawn robot operating body temperature>65°C, detection of passage obstacles, changes in the token queue, and updates to the maintenance prohibition time window}, an event queue is obtained for incremental rearrangement. By recalculating Mark in step S4, scanning all items in descending order of Mark according to the calculated Mark, and applying a greedy algorithm + tabu list to adjust the queue and tokens, and partially updating the forward replanning path and execution list in step S6, a dynamic adjustment strategy is obtained that can perform parameter updates, queue updates, execution list updates, and forward replanning path updates, thereby enabling dynamic adjustment without interrupting the operation.

[0097] Figure 2 This is a technical effect diagram of a dynamic task scheduling method for lawn robots based on artificial intelligence. The black bars represent the present invention, and the gray bars represent the prior art. The lower the value, the better. As can be seen from the figure, the technical effect of the present invention is better than that of the prior art.

[0098] Example 2 is an AI-based dynamic task scheduling method for lawn robots, based on Example 1. This Example 2 is an example used in conjunction with Example 1. The specific steps are as follows:

[0099] Step 1: Initialize the token and queue;

[0100] Specifically, by setting access tokens at corridor entrances and charging tokens at charging stations, access queues and charging queues are established, enabling one-way passage and orderly station occupancy, thus avoiding passing vehicles and charging competition.

[0101] Step 2: Read the execution list and schedule the path;

[0102] Specifically, by reading the candidate task sequence to generate an execution list for each robot, and splicing the workable task pieces into a strip channel according to the consistent direction rule, running D* Lite to obtain the forward replanning path and issuing it, it can ensure forward movement and reduce detours.

[0103] Step 3: Perform online monitoring and set trigger conditions;

[0104] Specifically, by collecting and transmitting the remaining battery power, token queue status, and execution list status of the lawn robot, and synchronizing the timestamps according to the time of the input set in step S1, a unified data stream that can be used for online judgment is obtained as the basis for triggering judgment; by setting triggers {RAIN≥0.3, W≥surface humidity threshold Wmin, lawn robot operating body temperature>65°C, detection of passage obstacle, change in token queue, update of maintenance prohibition time window}, an event queue is obtained for incremental rearrangement.

[0105] Step four: Perform incremental rearrangement and closed-loop update;

[0106] Specifically, by recalculating Mark and adjusting the queue and tokens using a greedy algorithm and tabu list, the execution list and Mark are locally updated and recorded, achieving the same technical effect as in Example 1. This can shorten the completion time, reduce invalid driving and traffic congestion, and decrease the unprocessed shading area and the energy consumption of the lawn robot.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic task scheduling method for lawn robots based on artificial intelligence, characterized in that: Step S1: Deploy sensors to collect data and access external data, and generate an input set from the collected data and external data; Step S2: Based on the input set, calculate the area of ​​the region shaded zone, divide the region shaded zone into mixed task slices according to blocks and time, and generate a set of mixed task slices; Step S3: Based on the mixed task slice set, calculate the probability of the lawn robot being able to work by constructing an evaluation algorithm, and generate the work schedule for each task slice; Step S4: Based on the work schedule of each task slice and the mixed task slices, calculate the priority score of the task slices, train the priority scheduling algorithm, and generate candidate task sequences. Step S5: Based on the candidate task sequence, set access tokens and charging tokens for the lawn robot, and generate an execution list and token queue; Step S6: Based on the execution list and token queue, the executable shadow bands are spliced ​​together, the path of the lawn robot is replanned, and a forward replanning path is generated. Step S7: Based on steps S1-S6, perform backhaul and closed-loop to generate a globally optimized dynamic adjustment strategy.

2. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for calculating the area of ​​the shaded region are as follows: Based on the solar incident vector SUN, azimuth angle ANGLE, photovoltaic array arrangement matrix JU(x,y) in the field, and set parameters of each photovoltaic module in the field, each photovoltaic module is projected parallel to the ground axis z=0 along the unit vector -SUN opposite to the direction of sunlight propagation in step S2 to obtain the shadow domain Shdow. An area exponential function is constructed to calculate the area of ​​the shadow region projected into the shadow domain, which is used to quantitatively represent the shadow coverage of the block × time unit. The formula for the area exponential function is: ， Let SIZE be the area of ​​the shadow region where the ground point (x, y) falls at time t, representing the shadow exponential function. Let x be the area of ​​the shadow region where x points east and y points north. For indicating function symbols, The dot product symbol. Let t be the shaded region at time step t, where t is the time step and Shdow is the shaded region.

3. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for dividing the area shadow band into hybrid task slices according to blocks and time, and generating a set of hybrid task slices are as follows: Based on the obtained area of ​​shadow, the site is spatially discretized according to the operation corridor or management grid within the planning window. Specifically, the spatial discretization of the site according to the operation corridor is as follows: the corridor is denoted as Hall, and the spatial sub-segment SP is obtained by dividing along the center line of the corridor with a fixed step size. Subsequently, time discretization is performed: equally spaced time slices (TP) are generated at fixed step sizes; on the Cartesian product of the aforementioned spatial sub-segments and time slices, hybrid task units (MD) are constructed, and for each hybrid task slice, the following are calculated and recorded: region shaded area (SIZE), region shaded domain (Shdow), region accessibility and geographical attributes, maintenance restricted access time window and charging accessibility, and workable marker. The hybrid task slice MD and its attributes are stored in the hybrid task slice set (GET) in a block-first, time-ordered manner and provided externally as standard fields, forming entries with clear start and end times and continuous operation lengths.

4. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for calculating the operational probability of the lawn robot based on the hybrid task slice set and generating the work schedule for each task slice by constructing an evaluation algorithm are as follows: Based on the hybrid task slice set GET, an evaluation algorithm is constructed to assess the operational probability of each task slice in the hybrid task slice set and generate a work schedule for each task slice. The steps for defining the evaluation algorithm are as follows: R1 is the driving sequence ALL of the grey model, composed of surface humidity W, wind speed F, cloud cover WIN, and ground temperature TEP from the input set. The calculation formula for the driving sequence is as follows: ， ALL is the driving sequence. This is the surface humidity weighting coefficient, where W is the surface humidity. It is a weighting factor for wind speed. The maximum wind speed for operation is F, where F is the wind speed. WIN is the weighting factor for cloud cover. It is the weighting coefficient for ground temperature, TEP is the ground temperature, and TEP1 is the calibrated ground temperature; By driving sequences, environmental factors such as humidity, wind speed, cloud cover, and ground temperature are mapped into monotonic signals of workability, providing quantitative input for the evaluation algorithm; R2, based on the driving sequence, obtains the predicted values ​​of the driving sequence through grey model fitting. The algorithm is used to evaluate the probability that the lawn robot can operate, and the calculation formula for the evaluation algorithm is as follows: ， Let be the workability probability function, representing the probability that block index i and time slice index j are workable. This represents a workable function, where i represents the block index and j represents the time slice index. This represents the activation function. Indicates the calibration coefficient. Represents the predicted value of the driving sequence; The workability probability of each task slice is obtained by constructing an evaluation algorithm, and a workability threshold is set based on the input set.

1. Determine whether the area is a restricted access window for maintenance, the corridor width threshold Hallmin, and the slope threshold SLmax. If all conditions are met, mark it as a workable area to obtain the workable area. R3, based on the workable area, records the time slice index of the expected start and end times of workable time. By merging adjacent time slices within the same corridor Hall that are all workable according to time continuity, the work schedule item is obtained. All work items are written into a set to obtain the work schedule WrokTab.

5. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for calculating the priority score of a task slice based on the work schedule of each task slice and the mixed task slices are as follows: Based on the work schedule WrokTab and the mixed task slice set GET, features are constructed for each work time item in WrokTab and normalized to the [0,1] interval. A priority scoring algorithm is constructed to calculate the priority score of each work time item. This score is used to convert the task items in the work schedule WrokTab and the mixed task slice set GET into comparable scores. Candidate task sequences are formed in descending order of benefits and constraints, providing a direct basis for token issuance and execution list distribution in step S5.

6. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 5, characterized in that: The specific steps for calculating the priority score of a task segment based on the work schedule of each task segment and the mixed task segments also include: The definition steps of the priority scoring algorithm are as follows: T1, based on each work time item, extract and construct scoring elements from the item: shadow bandwidth intensity, available window, reachability and rainfall penalty, and normalize the extracted elements to obtain the construction elements of the priority scoring algorithm; T2, Based on the constructive elements of the priority scoring algorithm, a priority scoring algorithm is constructed to obtain the priority score of each task objective. The formula for the priority scoring algorithm is: ， Where Mark represents the priority score. Here, EShdow represents the shadow bandwidth intensity weighting factor. WW represents the available window weight coefficient. Here, ARR is the accessibility weighting coefficient, and ARR is the accessibility rating. Here, is the rainfall penalty weighting coefficient, and Prain is the rainfall penalty. Here, WA represents the cross-region cost weighting coefficient. PLwork represents the workability probability weighting coefficient, and PLwork is the workability probability function.

7. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for generating candidate task sequences using a priority scheduling algorithm are as follows: By reading data from the work schedule WrokTab and the mixed task slice set GET, the original element set is obtained, which serves as the basis for issuing tokens and distributing execution lists in step S5. The priority scheduling algorithm is used for training, and historical job records are collected to form a training set. The benefit label BE is constructed using historical samples, and the priority scoring weights and scoring thresholds are trained using ridge regression to obtain the scoring function. Hard and soft constraints are set for the scoring function to obtain the feasible region. A sliding window is set to recalculate Mark, and incremental reordering is triggered when the rainfall probability RAIN, surface humidity W, and robot battery power change, resulting in an online adaptive candidate task queue.

8. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for setting access tokens and charging tokens for the lawn robot based on the candidate task sequence, and generating an execution list and token queue are as follows: Based on the candidate task sequence, a time slot table is established by traversing the candidate task sequence. Access tokens and charging tokens are sent to the corridor and charging station corresponding to each item, respectively, generating a token queue. Specifically: For the access queue established at the corridor entrance, access tokens are issued to lawn robots, and the order of robots waiting to enter the corridor and their access tokens is recorded for one-way passage and to avoid passing vehicles; for the charging queue established at the exit, charging tokens are issued to lawn robots that have completed their tasks, access tokens are collected, and the order of robots waiting to occupy the charging station and their charging tokens is recorded for orderly queuing and on-arrival service, resulting in one-way passage and orderly occupancy; by reading the token queue and the candidate task queue, an ordered list of execution items is obtained.

9. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for splicing executable shadow bands based on the execution list and token queue to replan the path of the lawn robot and generate a forward replanning path are as follows: By reading the execution list and token queue, the available task pieces from the acquired token entries are selected and spliced ​​into a strip channel according to the direction consistency rule. D* Lite is used to perform short-circuit search in the channel graph, where the cost is set as: distance + reverse penalty + departure penalty, and yield point and occupation conflict constraints are included to obtain the forward replanning path.

10. The method for dynamic task scheduling of a lawn robot based on artificial intelligence according to claim 1, characterized in that: The specific steps for generating a globally optimized dynamic adjustment strategy based on the backhaul and loop closure performed in steps S1-S6 are as follows: By collecting and transmitting the remaining battery power, token queue status, and execution list status of the lawn robot, and synchronizing the timestamps according to the time of the input set in step S1, a unified data stream that can be used for online judgment is obtained as the basis for triggering judgment. By setting triggers {RAIN≥0.3, W≥surface humidity threshold Wmin, lawn robot operating body temperature>65°C, detection of passage obstacle, change in token queue, and update of maintenance prohibition time window}, an event queue is obtained for incremental rearrangement. By recalculating Mark in step S4, scanning all items in descending order of Mark according to the calculated Mark, and applying a greedy algorithm + taboo list to adjust the queue and tokens, and partially updating the forward replanning path and execution list in step S6, a dynamic adjustment strategy is obtained.

Images