A mobile robot control method and system supporting edge-side intelligent reasoning

Abstract

This invention discloses a mobile robot control method and system supporting edge-side intelligent reasoning, belonging to the field of robot control technology. The method includes: determining an initial ground state matrix and the heading angle of the centerline of the cultivation rack channel based on the current scene; obtaining a local ground state matrix based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and characteristic parameters, wherein the characteristic parameters include ground normal reaction force, slip rate, and driving torque; obtaining a global ground state map based on the local ground state matrix and interactive data, wherein the interactive data is obtained from adjacent mobile robots; and obtaining motion control parameters based on the global ground state map and ground state parameters. This invention accurately stitches together the local ground state data of multiple robots through communication between robots, and avoids the risk of vehicle shaking caused by sudden changes in ground state by using dynamic parameter adaptation driven by the global ground state and pre-aiming control at the end of the cultivation rack channel.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically to a mobile robot control method and system that supports edge-side intelligent reasoning. Background Technology

[0002] With the large-scale development of facility agriculture globally, the cultivation area and operational complexity of mushroom cultivation bases, such as button mushrooms, continue to increase. Traditional manual spraying and inspection methods can no longer meet the demands for efficient and precise production. Therefore, multi-robot collaborative operation has become the core development direction of smart mushroom cultivation. This involves multiple robots working in parallel along the cultivation racks of button mushrooms to perform tasks such as inoculation, watering, and harvesting. However, the mushroom house is a closed, high-humidity environment. Intelligent micro-mist humidification and condensate settling in the greenhouse can cause slippery areas on the floor of the aisles. High humidity can also easily cause attenuation of communication and positioning signals. Furthermore, the operation aisles in the mushroom house have strong linear constraints. Traditional robot perception, communication, and control solutions are difficult to adapt to these scenario requirements, thus existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a mobile robot control method that supports end-side intelligent reasoning. By coordinating multiple robots and calibrating the coordinates of the heading angle constraint of the mushroom cultivation rack channel, control of the mobile robot can be achieved from the supporting end-side.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A mobile robot control method supporting edge-side intelligent reasoning, wherein the mobile robot control method is executed by each mobile robot, and the mobile robot control method includes:

[0006] Determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario;

[0007] Based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and the characteristic parameters, a local ground state matrix is ​​obtained. The characteristic parameters include the ground normal reaction force, slippage rate, and driving torque.

[0008] A global ground state map is obtained based on the local ground state matrix and the interaction data, wherein the interaction data is obtained from neighboring mobile robots;

[0009] Motion control parameters are obtained based on the global ground state map and ground state parameters.

[0010] As a further improvement of the present invention, the step of determining the initial ground state matrix based on the current scenario includes:

[0011] The grid resolution is obtained based on the spacing between button mushroom cultivation units in the current scene;

[0012] The dimensions of the initial ground state matrix are determined based on the grid resolution and its corresponding effective coverage area.

[0013] The initial adhesion coefficient and initial flatness are determined based on historical data to obtain default state parameters, and each element in the initial ground state matrix is ​​assigned the default state parameter.

[0014] As a further improvement of the present invention, the step of obtaining the local ground state matrix based on the initial ground state matrix, the heading angle of the cultivation rack channel centerline, and characteristic parameters includes:

[0015] The current ground adhesion coefficient is obtained based on the aforementioned characteristic parameters;

[0016] Obtain the current lateral tilt angle and longitudinal pitch angle to determine the current ground flatness;

[0017] Obtain the current obstacle type label from lidar point cloud data;

[0018] The local ground state matrix is ​​obtained based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, the current ground adhesion coefficient, the current ground flatness, and the current obstacle type label.

[0019] As a further improvement of the present invention, the interaction data is obtained based on neighboring mobile robots and includes:

[0020] The adjacent mobile robots obtain the core network based on their corresponding local ground state matrix;

[0021] Extract the state change data of the core network to obtain the differential data to be broadcast;

[0022] Data confidence is obtained based on sensor health and time decay coefficient;

[0023] The interaction data is obtained based on the differential data to be broadcast and the data confidence level.

[0024] As a further improvement of the present invention, the step of obtaining a global ground state map based on the local ground state matrix and the interaction data includes:

[0025] A target function is constructed based on the local ground state matrix and interactive data;

[0026] Based on the objective function and dynamic matching rules, an overlapping grid matching set is obtained;

[0027] The confidence-weighted matching error is obtained based on the overlapping grid matching set and the data confidence level.

[0028] The local matrix coordinate offset is obtained by minimizing the confidence-weighted matching error.

[0029] The global ground state map is obtained based on the local matrix coordinate offset.

[0030] As a further improvement of the present invention, obtaining the global ground state map based on the local matrix coordinate offset includes:

[0031] Based on the local matrix coordinate offset and the interaction data, the local matrix data is obtained;

[0032] Based on the dynamic Hube threshold and the residual direction weights, the direction-weighted Hube loss function is obtained;

[0033] The robust fusion weight coefficients are obtained based on the direction-weighted Hubey loss function and the regularized weights.

[0034] The global ground state map is obtained based on the robust fusion weight coefficients and the local matrix data.

[0035] As a further improvement of the present invention, the step of obtaining motion control parameters based on the global ground state map and ground state parameters includes:

[0036] The robot's positioning grid coordinates are obtained based on the global ground state map.

[0037] The positioning fusion parameters are obtained based on the robot's positioning-related grid coordinates;

[0038] The robot's global coordinates are obtained based on the localization fusion parameters;

[0039] Motion control parameters are obtained based on the robot's global coordinates and the ground state parameters.

[0040] As a further improvement of the present invention, the step of obtaining motion control parameters based on the robot's global coordinates and the ground state parameters includes:

[0041] Based on the robot's global coordinates and the global ground state map, the basic differential ratio and basic drive torque are obtained;

[0042] The dynamic motion parameters are obtained based on the basic differential ratio, the basic drive torque, and the first-order low-pass filter formula.

[0043] The motion control parameters are obtained based on the ground state parameters and the dynamic motion parameters.

[0044] As a further improvement of the present invention, the step of obtaining the motion control parameters based on the ground state parameters and the dynamic motion parameters includes:

[0045] Based on the robot's global coordinates and the global ground state map, a trigger command is generated;

[0046] Based on the trigger command and the ground state parameters, the pre-aiming steering angle and deceleration speed are obtained;

[0047] The motion control parameters are obtained based on the dynamic motion parameters, the pre-aiming steering angle, and the deceleration speed.

[0048] This invention provides a mobile robot control system that supports edge-side intelligent reasoning, comprising:

[0049] The initialization module is used to determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario.

[0050] The calculation module is used to obtain a local ground state matrix based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and characteristic parameters, wherein the characteristic parameters include ground normal reaction force, slippage rate, and driving torque.

[0051] An interaction module is used to obtain a global ground state map based on the local ground state matrix and the interaction data, wherein the interaction data is obtained from neighboring mobile robots;

[0052] The control module is used to obtain motion control parameters based on the global ground state map and ground state parameters.

[0053] This invention constructs a perception fusion mechanism based on the structural features of the cultivation rack passage on the support end. Utilizing a scene-adaptive robust weighted fusion algorithm, it accurately stitches together the local ground state data of multiple robots, fundamentally overcoming the limitations of traditional single-source perception. This results in a fused global map that better conforms to the continuous structural features of agricultural ground, effectively suppressing abnormal data interference. Furthermore, by combining coordinate calibration logic with the heading angle constraint of the cultivation rack passage, the accuracy of map alignment is improved, adapting to signal attenuation and straight-line constraint requirements. At the motion control level, through dynamic parameter adaptation driven by the global ground state and pre-aiming control at the end of the cultivation rack passage, the control logic achieves smooth transition of motion parameters and advance prediction of steering actions, avoiding vehicle vibration caused by sudden changes in ground state and ensuring the stability of robot operation. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the steps of a mobile robot control method supporting edge-side intelligent reasoning according to the present invention.

[0055] Figure 2A schematic diagram illustrating the steps to obtain the local ground state matrix;

[0056] Figure 3 This is a diagram illustrating the interaction between robots. Detailed Implementation

[0057] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0058] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0059] like Figure 1 As shown, this embodiment provides a mobile robot control method that supports edge-side intelligent reasoning. The mobile robot control method is executed by each mobile robot and includes:

[0060] Determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario;

[0061] Based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and characteristic parameters, the local ground state matrix is ​​obtained. The characteristic parameters include the ground normal reaction force, slippage rate, and driving torque.

[0062] A global ground state map is obtained based on the local ground state matrix and interaction data, with the interaction data obtained from neighboring mobile robots.

[0063] Motion control parameters are obtained based on the global ground state map and ground state parameters.

[0064] The method provided in this embodiment can be applied to large-scale button mushroom cultivation bases. The button mushrooms adopt a multi-layer cultivation rack cultivation mode, and parallel operation channels are formed between the cultivation racks. The channels are concrete floors with drainage slopes. The water mist generated by the humidification operation in the mushroom house settles due to gravity and the condensate drips from the greenhouse, which will cause humid areas in the channels. At the same time, multiple mobile robots (such as inoculation robots, harvesting robots, and environmental inspection robots) may be operating in parallel along the channels. This embodiment does not limit the specific types of robots.

[0065] This embodiment uses two robots as an example to introduce the subsequent steps. Each mobile robot currently has only one adjacent mobile robot. After the two mobile robots start, all onboard sensing and communication hardware is first activated and calibrated, including vision, lidar, pressure sensors, wheel speed sensors, inertial measurement units, positioning modules, and short-range communication modules. This ensures that each hardware component can stably collect and transmit data. To avoid signal conflicts when multiple robots communicate simultaneously, dedicated communication time slices can be allocated to robots with different priorities. For example, a high-priority drug delivery robot uses a longer time slice to ensure that its core operational data can be transmitted preferentially and stably, without packet loss or delay due to signal congestion. This embodiment does not limit the priority setting method or the specific time slice length. This embodiment also does not limit the communication method between mobile robots. For example, edge-side ZigBee short-range communication can be used, and each mobile robot performs directional communication, meaning each mobile robot only sends interactive data to its corresponding adjacent mobile robot. Directional communication can be achieved by carrying the ID of the adjacent mobile robot in the data packet header. Furthermore, the adjacent mobile robots corresponding to each mobile robot can be pre-assigned. For example, mobile robots that perform parallel tasks at the same time can be designated as adjacent mobile robots according to the work schedule, and the ID of the adjacent mobile robot corresponding to each mobile robot can be stored in the cache area inside each mobile robot to facilitate the subsequent sending of interactive data.

[0066] Specifically, such as Figure 2 As shown, the basic rules and initial state for multi-robot collaborative operation are first established, including configuring communication time slots according to task priority, initial ground state matrix, and preloading the heading angle of the cultivation rack channel centerline, providing prerequisites for subsequent perception and collaboration. Then, based on the initial ground state matrix and the heading angle of the cultivation rack channel centerline, the ground state (adhesion coefficient, flatness, obstacle type) is calculated by collecting data from multiple sensors, and smoothing preprocessing is performed in combination with the direction constraints of the cultivation rack channel to obtain a local ground state matrix that better fits the scene. Next, based on the interaction data sent by adjacent mobile robots, the local ground state matrix corresponding to the adjacent mobile robots is obtained. The heading angle of the cultivation rack channel centerline is used to complete the precise calibration of multiple matrices, and finally, the global ground state map is obtained by fusion. Then, based on the global ground state map and ground state parameters, motion control parameters are obtained, and control commands for cultivation rack channel operation and end-point steering are generated to achieve precise motion control of the end-point autonomous operation.

[0067] This embodiment constructs a perception fusion mechanism based on the structural features of the cultivation rack channel on the support end. Utilizing a scene-adaptive robust weighted fusion algorithm, it accurately stitches together the local ground state data of multiple robots, fundamentally overcoming the limitations of traditional single-source perception. This makes the fused global map more closely match the continuous ground structure features of the mushroom cultivation base, effectively suppressing abnormal data interference. Furthermore, combined with coordinate calibration logic constrained by the heading angle of the cultivation rack channel, it improves the accuracy of map alignment and adapts to signal attenuation and linear constraint requirements. At the motion control level, through dynamic parameter adaptation driven by the global ground state and pre-aiming control at the end of the cultivation rack channel, it achieves smooth transition of motion parameters and advance prediction of steering actions from a control logic perspective, avoiding vehicle shaking caused by sudden changes in ground state and ensuring the stability of robot operation.

[0068] Furthermore, this embodiment provides a step for determining an initial ground state matrix based on the current scenario, including:

[0069] The grid resolution is obtained based on the spacing between button mushroom cultivation units in the current scene;

[0070] The dimensions of the initial ground state matrix are determined based on the grid resolution and its corresponding effective coverage area.

[0071] The initial adhesion coefficient and initial flatness are determined based on historical data, and the default state parameters are obtained. Each element in the initial ground state matrix is ​​then assigned the default state parameter.

[0072] The initial ground state matrix is ​​a data matrix stored in grid form, which is the ground state of the surrounding local area perceived by the mobile robot through sensors. In this embodiment, the size of the surrounding local area is not limited, for example, it is 1m×1m. Specifically, the grid resolution is the actual ground area corresponding to a single grid in the initial ground state matrix, which needs to be adapted to the spacing between the button mushroom cultivation units and the perception range of the mobile robot. However, this embodiment does not limit its specific value. For example, when the spacing between the button mushroom cultivation racks is about 0.2 meters and the effective perception range of the mobile robot is about 1m×1m, in order to accurately match the plant distribution and the perception range, it can be determined that a single grid corresponds to an actual ground area of ​​0.2m×0.2m. Therefore, the 1m×1m perception range needs to be divided into a 5×5 grid matrix. The 5×5 specification is the dimension corresponding to the initial ground state matrix. The effective perception range of the mobile robot refers to the physical space range in which the various sensors carried by the mobile robot can stably and reliably collect effective environmental perception data in a specific working environment. It is the perception boundary determined by the sensor hardware performance and scene adaptability.

[0073] Next, default state parameters are set, including the initial adhesion coefficient, initial flatness, and obstacle type. For example, historical data can be statistically analyzed. If the ground is dry, the adhesion coefficient for most plots is between 0.42 and 0.58, with 72% of the samples concentrated in the 0.48-0.52 range, and the arithmetic mean is 0.502. To balance the stability of the initial baseline with scene adaptability, an approximation of the statistical average of 0.5 can be used as the initial adhesion coefficient. This not only reflects the actual adhesion capability of the ground but also avoids misjudgments in power output during the startup phase due to deviations from default values. The plots are areas defined based on the mobile robot's operating area, and their range is larger than the mobile robot's effective perception range. Concrete surfaces naturally exhibit flatness fluctuations due to a 1‰-3‰ drainage slope, construction expansion joints, and long-term micro-deformation. If the lateral tilt angle and longitudinal pitch angle of the corresponding area are statistically concentrated in the 0.5°-1.5° range, multiple flatness values ​​are obtained based on the relationship between flatness and tilt angle. Finally, the median (e.g., 80) is selected as the initial flatness value. The obstacle type can be determined based on historical obstacle detection data from the mobile robot's startup phase, and a unified numerical coding rule can be set for the system, such as 0 representing no obstacles, 1 representing rigid obstacles, and 2 representing flexible obstacles. Historical data can be obtained through sampling; for example, those skilled in the art can select multiple plots within the mobile robot's operating area at preset intervals, obtain the adhesion coefficient and flatness of each plot, and store them. The above values ​​are merely examples, and this embodiment does not impose limitations. The method for obtaining the adhesion coefficient and flatness is the same as the method for each subsequent mobile robot to obtain the current ground adhesion coefficient and current ground flatness, and will not be elaborated upon here.

[0074] Furthermore, this embodiment provides a step for obtaining a local ground state matrix based on an initial ground state matrix, the heading angle of the cultivation rack channel centerline, and characteristic parameters, including:

[0075] The current ground adhesion coefficient is obtained based on the characteristic parameters;

[0076] Obtain the current lateral tilt angle and longitudinal pitch angle to determine the current ground flatness;

[0077] Obtain the current obstacle type label from lidar point cloud data;

[0078] Based on the initial ground state matrix, the heading angle of the cultivation rack channel centerline, the current ground adhesion coefficient, the current ground flatness, and the current obstacle type label, the local ground state matrix is ​​obtained.

[0079] The characteristic parameters include ground normal reaction force, slip ratio, and driving torque. Specifically, the ground adhesion coefficient is a parameter reflecting the gripping ability between the wheel and the ground. The higher the value, the stronger the ground grip, which can be used to judge the risk of wheel slippage. The ground normal reaction force is the vertical support force of the ground on the wheel, which is directly collected by the pressure sensor. The slip ratio reflects the degree of wheel slippage. It can be obtained by collecting the actual linear velocity of the wheel by the wheel speed sensor and obtaining the theoretical linear velocity when the wheel has no slippage. Then, the actual linear velocity is subtracted from the theoretical linear velocity, and the ratio of the difference to the theoretical linear velocity is recorded as the slip ratio. The driving torque is the torque that drives the robot to rotate the wheel, which is directly output by the control module of the power system. The current method for calculating the coefficient of adhesion is to multiply the normal reaction force of the ground by (1 minus the slip rate) and then divide by the driving torque. The purpose is to comprehensively evaluate the ground's grip capability by combining ground support force, slippage degree, and power output. Specifically, when the wheels slip, the actual friction force decreases significantly. Therefore, this embodiment uses (1 minus the slip rate) to correct the normal force. The driving torque is the torque output by the mobile robot's power system, which needs to be converted into the driving force of the wheels. The greater the driving torque, the stronger the grip required by the mobile robot. Using it as the denominator is to balance the grip force provided by the ground with the actual grip force required by the mobile robot, thereby assessing in real time whether the current ground grip capability meets the driving requirements.

[0080] Ground smoothness is a score used to reflect the degree of ground bumpiness. The lateral tilt angle / longitudinal pitch angle are the tilt angles of the ground in the lateral and longitudinal directions, which are collected by the inertial measurement unit. Then, the square root of the sum of the squares of the lateral tilt angle and the longitudinal pitch angle is used to obtain the comprehensive bumpiness. The comprehensive bumpiness is then mapped back to a score to obtain the current ground smoothness. For example, the current ground smoothness = 100 - 10 × comprehensive bumpiness, where 100 represents a completely smooth ground and 10 is a coefficient used to keep the score change consistent with the perception of the actual bumpiness. This embodiment does not limit the specific value of the coefficient, and those skilled in the art can determine it according to the actual situation.

[0081] Next, point cloud data of the surrounding environment collected by the lidar is acquired. By identifying the shape and density of the point cloud, it distinguishes between unobstructed objects, rigid obstacles (such as ground expansion joints), and flexible obstacles (such as mushroom debris knocked down during harvesting), and assigns corresponding digital coding labels. For example, firstly, the point cloud data is denoised, and then the RANSAC algorithm is used to fit the ground plane, dividing the point cloud into ground points and non-ground points. Non-ground points are potential obstacles. The RANSAC algorithm is existing technology, and this embodiment will not elaborate on it. Then, for non-ground points, their height relative to the ground is calculated. The aspect ratio and point cloud density are used to determine the specific obstacle type. This embodiment does not limit the criteria for classifying obstacle types. For example, if the height of the point cloud is <0.01m and the density is <5 points / cm³, it is recorded as no obstacle. If the height of the point cloud is between 0.01-0.03m, the aspect ratio is >5, and the density is >10 points / cm³, it corresponds to a linear micro-protrusion ground expansion joint and is recorded as a rigid obstacle. If the height of the point cloud is between 0.01-0.05m, the aspect ratio is >2, and the density is between 3-8 points / cm³, it corresponds to slender mushroom debris and is recorded as a flexible obstacle.

[0082] Next, the directional constraint weights of the cultivation rack channels are set to conform to the linear structural features of the cultivation rack channels. Different weight values ​​need to be set for the smoothness of the initial ground state matrix in different directions. In this embodiment, the specific values ​​of the weights are not limited. For example, since the cultivation rack channels are distributed in a continuous straight line, the changes in ground state (such as dryness and wetness, flatness) along the direction of the cultivation rack channels are smaller and more continuous, so a higher weight such as 0.3 is set; the changes in ground state in the direction perpendicular to the cultivation rack channels are greater, so a lower weight such as 0.1 is set. Finally, the current ground adhesion coefficient, the current ground flatness, and the current obstacle type label are combined with the heading angle of the center line of the cultivation rack channels according to the weight ratio along the cultivation rack channels and perpendicular to the cultivation rack channels to smooth the initial ground state matrix and obtain the local ground state matrix.

[0083] Specifically, the heading angle of the cultivation rack aisle centerline refers to the angle between the physical centerline of the cultivation rack aisle and the reference direction of the coordinate system in the coordinate system of the mobile robot operation. This embodiment does not restrict the method of obtaining the heading angle of the cultivation rack aisle centerline. For example, the cultivation rack aisle can be surveyed on-site, and multiple key coordinate points of the aisle centerline can be collected. By fitting these coordinate points, the straight line equation of the aisle can be obtained, and then the heading angle of the centerline can be calculated. Then, the heading angle of the cultivation rack aisle is... Mapped to a unit direction vector along the channel. and the unit direction vector of the vertical channel Next, for each grid in the initial ground state matrix, its neighborhood is determined. This embodiment does not limit the size of the neighborhood; for example, it can be set to a 3×3 neighborhood. The weighted smoothing weight of each grid within the neighborhood is then calculated. For example, for each grid in the initial ground state matrix located at the... Line number Column grid The grid within its neighborhood Weighted smoothing weights ,in yes Compared to The displacement vector, This indicates the weight corresponding to the direction along the cultivation rack aisle. This indicates the weight corresponding to the direction of the vertical cultivation rack channel. and These represent the mesh within the neighborhood compared to the mesh. The offsets in the row and column directions are then used to weight the ground adhesion coefficients of each grid in the neighborhood according to the weighted smoothing weights of each grid in the neighborhood, so as to obtain the smoothed ground adhesion coefficients of each grid. Similarly, the smoothed ground flatness of each grid can be obtained. For obstacle types, a weighted voting method can be used, that is, the weighted votes of each obstacle type in the neighborhood are counted, and the type with the highest number of votes is the final obstacle type of the grid. Then, based on the smoothed ground adhesion coefficients, ground flatness, and obstacle types, the local ground state matrix is ​​obtained. In this embodiment, there are no restrictions on the value of each of the above weights.

[0084] This embodiment accurately quantifies the ground adhesion coefficient based on the fusion calculation of ground support force, slip rate, and power output. Combined with the flatness score and obstacle type label obtained from tilt angle and LiDAR point cloud data, it significantly improves the perception dimension and accuracy of the ground state at the end, providing richer decision-making basis for subsequent motion control. At the same time, by smoothing the differential weights along the cultivation rack channel and the vertical cultivation rack channel, the local ground state matrix is ​​made to better fit the linear structural characteristics of the tomato cultivation rack channel, effectively suppressing noise interference from single sensor data and reducing local matrix state fluctuations. This lays a solid perception foundation for the accurate alignment and global fusion of subsequent multi-robot maps, ultimately achieving deep adaptation between the end-side perception results and the agricultural scenario, improving the stability and accuracy of multi-robot collaborative operations from the source.

[0085] Furthermore, this embodiment provides a step of obtaining interactive data based on neighboring mobile robots, including:

[0086] Neighboring mobile robots obtain the core network based on their corresponding local ground state matrix;

[0087] Extract the state change data of the core network to obtain the differential data to be broadcast;

[0088] Data confidence is obtained based on sensor health and time decay coefficient;

[0089] Interactive data is obtained based on the differential data to be broadcast and the data confidence level.

[0090] Specifically, the steps for adjacent mobile robots to determine their corresponding local ground state matrix are the same as described above, and will not be repeated in this embodiment. The core grid refers to the critical operating area directly in front of and below the mobile robot, and changes in the core grid state have a significant impact on motion control. The edge grid refers to the peripheral secondary area, and changes in its state have a smaller impact on operation. This embodiment does not limit the specific size of the core grid and the edge grid. For example, the local ground state matrix can be divided into a 3×3 core grid in the middle and a 2×2 edge grid on the periphery. Then, the ground state is judged to have changed significantly based on the state change threshold. Data broadcasting is triggered only when the change exceeds the threshold. This embodiment does not limit the threshold. For example, if the change in the adhesion coefficient exceeds 0.05, or the change in the flatness exceeds 5, the state of the core grid is considered to have changed significantly, and a core grid broadcast trigger flag is generated.

[0091] Next, the data corresponding to the grids whose states have changed in the core grid are extracted as the differential data to be broadcast, and the data confidence level is calculated. For example, the current health score of the sensor can be multiplied by a coefficient that decays over time to obtain the final confidence value, which is used to evaluate the reliability of the data. The health score takes a value of 0-1, and can be determined by weighting noise rate, packet loss rate, and drift rate. The noise rate is the proportion of high-frequency noise in the sensor data, the packet loss rate is the proportion of packet loss in the transmission of sensor data, and the drift rate is the cumulative drift of the sensor data. This embodiment does not limit the weight values, and the time decay coefficient is used. It can be determined by the decay rate and the time interval of data collection Determine, for example, set the decay rate. , At this time, the time decay coefficient Finally, the differential data to be broadcast and the data confidence scores are packaged to obtain the interactive data, such as... Figure 3 As shown, each mobile robot needs to send and receive interactive data to its neighboring mobile robots. For example, mobile robot A and its corresponding neighboring mobile robot B both need to send interactive data to each other.

[0092] Furthermore, this embodiment provides a step for obtaining a global ground state map based on a local ground state matrix and interactive data, including:

[0093] Construct an objective function based on the local ground state matrix and interactive data;

[0094] Based on the objective function and dynamic matching rules, an overlapping grid matching set is obtained;

[0095] Based on the overlapping grid matching set and the data confidence, the confidence-weighted matching error is obtained;

[0096] The local matrix coordinate offset is obtained by minimizing the confidence-weighted matching error.

[0097] The global ground state map is obtained based on the local matrix coordinate offset.

[0098] Specifically, upon receiving the interactive data, it can parse the other party's initial ground state matrix, current coordinates, data confidence level, and the other party's own heading angle, etc., and then use the heading angle of the centerline of the cultivation rack channel as the reference. Using the global benchmark, compare the heading angle of the other party with that of the benchmark. If the other party's heading angle deviates from the benchmark, construct an objective function for calibration. The objective function is:

[0099] ;

[0100] in, This represents the coordinate offset of the local matrix obtained from the final solution, including the lateral position offset. Longitudinal position offset and heading angle rotation offset , This indicates the search for the objective function that will achieve its minimum value. , Represents the set of overlapping meshes. Each grid in Summation, This represents the spatial coordinates of the grid. This represents the data confidence level corresponding to itself; the data confidence level is the same for each grid. This represents the grid in the local ground state matrix corresponding to itself. The corresponding value, This represents the initial ground state matrix corresponding to adjacent mobile robots, after... and Translated network The corresponding value, Denotes the square of the L2 norm. This represents the weighting coefficient of the heading angle constraint term, used to balance the importance of mesh matching error and heading angle deviation. This embodiment does not impose restrictions on its specific value; for example, it can be set to 0.1. This represents the heading angle of the adjacent mobile robots. In this embodiment, the grid state of the two matrices is matched by the summation term, and the heading angle constraint term ensures that the direction of the matrix is ​​consistent with the global reference of the cultivation rack channel, avoiding the direction deviation caused by local noise. Finally, the matrices corresponding to multiple mobile robots can be accurately stitched into a global map.

[0101] For example, the dynamic matching rule can be set to match along the 4-neighborhood of the cultivation rack channel. That is, in the physically overlapping area of ​​the two matrices (the local ground state matrix corresponding to itself and the initial ground state matrix corresponding to the adjacent mobile robot), with the heading angle of the center line of the cultivation rack channel as the axis, matching points are only found in the four adjacent grids in front, behind, left, and right along the direction of the cultivation rack channel. For example, based on the objective function, in the overlapping area corresponding to the two matrices, state similarity matching points with an adhesion coefficient difference < 0.05 and a flatness difference < 5 can be selected to obtain the overlapping grid matching set. 0.05 and 5 are only examples and are not limited in this embodiment. For each matching point, the Euclidean distance between its corresponding grid values ​​in the two matrices is calculated as the original error, which is then multiplied by the data confidence level. If the data confidence level of a point is high, the error weight of that point is larger; if the confidence level is low, the error weight is smaller. The final weighted result of the original errors of all matching points is the confidence-weighted matching error. This embodiment does not limit the specific value of the weight. However, during the iterative solution of the objective function, the confidence-weighted matching error is constantly changing because the local matrix coordinate offset obtained in each iteration is different.

[0102] The purpose of obtaining the offset through confidence-weighted matching error is to correct the position and orientation deviations of the initial ground state matrices corresponding to adjacent mobile robots. This embodiment does not limit the specific method for solving the objective function; for example, iterative least squares can be used. The iteration termination condition is that the difference between two consecutive confidence-weighted matching errors is less than a preset difference. This embodiment does not limit the specific value of the preset difference. The final obtained local matrix coordinate offset is used to correct the opponent's initial ground state matrix, including lateral and longitudinal positional movement and directional rotation, so that the opponent's initial ground state matrix is ​​completely aligned with the reference direction of the cultivation rack channel centerline. After calibration, the heading angle deviation of the opponent's initial ground state matrix is ​​controlled within a small range, improving the mesh matching accuracy and ensuring that the local matrices of multiple robots can be accurately stitched into a global map without misalignment or deviation. The initial ground state matrix retains the finest-grained ground details collected by the sensors, while the smoothed matrix is ​​the result of processing based on the orientation constraints of the cultivation rack channel, which has already filtered out some high-frequency noise and local features. Directly correcting the initial ground state matrix can completely preserve these local real features while aligning with the global heading angle reference, avoiding the loss of key information due to premature smoothing. Therefore, this embodiment is corrected based on the initial ground state matrix corresponding to the adjacent mobile robot, while using the smoothed local ground state matrix itself to provide a globally stable reference benchmark and ensure the overall stability of the global ground state map. If both mobile robots use the smoothed local ground state matrix, it is easy to lose the real small-scale features of the ground.

[0103] Furthermore, this embodiment provides a step for obtaining a global ground state map based on local matrix coordinate offsets, including:

[0104] The local matrix data is obtained based on the local matrix coordinate offset and the interactive data;

[0105] Based on the dynamic Hube threshold and the residual direction weights, the direction-weighted Hube loss function is obtained;

[0106] The robust fusion weight coefficients are obtained based on the direction-weighted Hubey loss function and regularized weights;

[0107] A global ground state map is obtained based on the robust fusion weight coefficients and local matrix data.

[0108] The initial ground state matrix corresponding to adjacent mobile robots is aligned by coordinate offset calibration and recorded as local matrix data. Then, the local ground state matrix and local matrix data corresponding to the robot need to be fused to obtain the global ground state map. Since the elements corresponding to each coordinate position in the matrix are three features: ground adhesion coefficient, ground flatness, and obstacle type, the three features need to be fused separately. The global ground state map is then obtained based on the fusion result of the three features.

[0109] Taking the ground adhesion coefficient as an example, firstly, the ground adhesion coefficient corresponding to each grid in the overlapping area is extracted from the local matrix data, and the standard deviation is calculated. Then, the standard deviation is multiplied by the threshold corresponding to the Huber loss function to obtain the dynamic Huber threshold. This embodiment does not restrict the threshold; for example, a commonly used threshold of 1.345 in the Huber loss function can be selected. The dynamic Huber threshold is used as the outlier judgment standard. Specifically, for each grid in the overlapping area, its ground adhesion coefficient in the local matrix data and its ground adhesion coefficient in its corresponding local ground state matrix are obtained. The absolute value of the difference between these two ground adhesion coefficients is calculated. If the absolute value is less than or equal to the dynamic Huber threshold, the loss value corresponding to that grid can be calculated based on the quadratic loss. The quadratic loss is more sensitive to smaller errors and can better fit subtle fluctuations in the ground condition. If the absolute value is greater than the dynamic Hube threshold, the loss value corresponding to the grid can be calculated based on the linear loss. Linear loss increases the error penalty for larger errors more slowly, which can avoid outliers excessively interfering with the fusion result. This represents the absolute value of the difference. Indicates the dynamic Hubble threshold. This represents the direction weight, which is based on... Once the corresponding direction is determined, for example, obtain the ground adhesion coefficient corresponding to the grid in the local matrix data and its corresponding coordinates in its local ground state matrix. Calculate the position difference vector between the two coordinates, and then calculate the dot product of the position difference vector with the unit direction vector along the channel and the unit direction vector perpendicular to the channel, respectively. If the absolute value of the dot product calculated with the unit direction vector along the channel is greater than the absolute value of the dot product calculated with the unit direction vector perpendicular to the channel, then... The corresponding direction is along the channel direction, and the opposite is perpendicular to the channel direction. If it is along the channel direction, a higher directional weight is set, such as 1.2; if it is perpendicular to the channel direction, a lower directional weight is set, such as 0.8. The reason is that the cultivation rack channels extend in a straight line, and the ground conditions (such as ground adhesion coefficient and flatness) are more continuous. Therefore, the error along the channel direction has a greater impact on the stability of the global map. Assigning a higher weight can make the fusion result more closely fit the continuous structure of the channel. Then, iterative operations are performed based on the calculated loss value to obtain the fusion result of the grid for the ground adhesion coefficient. For example, for the grid... The fusion result for the ground adhesion coefficient in the first iteration for:

[0110] ;

[0111] in, Indicates the mesh after the iteration operation is completed. After adding subscripts to the fused results of the ground adhesion coefficient, The index 1 in the table corresponds to the first iteration, and since this embodiment uses two mobile robots as an example, the index... The value can be 1 or 2. When taking 1, This represents the grid in the local ground state matrix corresponding to itself. The corresponding ground adhesion coefficient, This indicates the confidence level of the data it corresponds to. When taking 2, Represents the grid in local matrix data The corresponding ground adhesion coefficient, This indicates the data confidence level corresponding to adjacent mobile robots. Indicates standard deviation, Represents a grid Regarding the neighborhood, this embodiment does not limit the size of the neighborhood; for example, the neighborhood can be set as a grid. The four grids in the four directions of up, down, left, and right. Represents the coordinates of the grid within the neighborhood. This represents the directional weight of the corresponding grid within the neighborhood. This represents the fused result of the ground adhesion coefficient obtained by this grid in the previous adjacent iteration within the neighborhood. Since this is the first iteration, therefore... The ground adhesion coefficient can be represented by the local ground state matrix corresponding to the grid, by adding a weighted summation term of the neighborhood input. This can ensure that the fusion result of the current grid is consistent with the result of the neighboring grid, enhancing the continuity of the global map. This represents the regularization weight, used to control the constraint strength of the channel direction. The larger the value, the more it conforms to the linear structure of the channel. In this embodiment, its value is not limited; for example, it can be set to 0.1. Represents a grid For the robust fusion weighting coefficient of the ground adhesion coefficient, this coefficient will assign a larger fusion ratio to the matrix with more stable state and higher confidence, and a smaller ratio to the matrix with large fluctuations and low confidence. This represents the grid in the local ground state matrix corresponding to itself. The corresponding ground adhesion coefficient and the grid in the local matrix data The average value of the corresponding ground adhesion coefficient.

[0112] After obtaining the result of the first iteration, repeat the above steps based on the result of the first iteration to obtain the result of the second iteration. Then, the difference between the two iteration results is calculated. If the difference is less than a preset termination threshold, the iteration is terminated. This embodiment does not limit the preset termination threshold; for example, it can be set to... Conversely, repeat the above steps until the difference between two adjacent iterations is less than a preset termination threshold, then output the final iteration result as the grid. Similarly, for the fusion result of the ground adhesion coefficient, the fusion result of the ground adhesion coefficient for each grid can be obtained. Similarly, for ground flatness or obstacle type, the above method can be used to obtain the fusion result of ground flatness or obstacle type for each grid. It is only necessary to replace the value corresponding to the ground adhesion coefficient with the value corresponding to the ground flatness or obstacle type. This embodiment will not elaborate on this. Finally, based on the fusion result of the three features and the element corresponding to each grid in the non-overlapping area, a global ground state map is obtained. The element corresponding to each grid in the global ground state map is composed of the fusion result of the ground adhesion coefficient, ground flatness and obstacle type corresponding to that grid.

[0113] This embodiment achieves robust generation of a global ground state map through dynamic threshold adaptation, directional weighted fusion, and iterative optimization: dynamic Hubey threshold accurately distinguishes between normal fluctuations and outliers, directional weighting fits the structural features of the cultivation rack channel, and iterative fusion improves the stability and accuracy of the global map. Ultimately, this significantly enhances the state continuity and outlier detection capability of the global map. In this map, state fluctuations along the cultivation rack channel direction are significantly reduced, and the accuracy of outlier detection is greatly improved, enabling a more accurate reflection of the actual ground state of the agricultural factory and providing a solid perception foundation for multi-robot collaborative operations.

[0114] Furthermore, in addition to the steps described above, this embodiment provides another step for obtaining a global ground state map based on a local ground state matrix and interactive data. This can achieve global optimization through a distributed architecture, ensuring both the accuracy of positioning and mapping, and reducing the communication overhead of multi-robot collaboration.

[0115] Specifically, a distributed factor graph is first constructed. The global factor graph contains the core data nodes and constraint factors of all mobile robots. The node part consists of two parts: first, the real-time position and attitude data of each mobile robot, with attitude data being the pose information output by SLAM, including planar position and tilt angles in three directions; second, the ground state data perceived by each mobile robot, including the adhesion coefficient in the form of a 3D mesh, flatness, and geometric feature information such as the cultivation rack columns extracted by SLAM. The constraint factor part contains five types of core constraints: the first type is the local SLAM pose constraint of the mobile robot, ensuring smooth and consistent changes in the robot's position and attitude at different times; the second type is the correlation constraint between local pose and ground state, such as the smaller fluctuation of the mobile robot's tilt angle corresponding to a flat ground; the third type is the feature correlation constraint of the local 3D mesh, used to bind the geometric features extracted by SLAM with ground state parameters, so that each mesh has both numerical parameters and environmental feature labels; the fourth type is the interaction constraint between multiple mobile robots, ensuring that the data of adjacent mobile robots remains consistent in overlapping work areas; and the fifth type is the SLAM closed-loop constraint, providing a basis for global consistency calibration when a mobile robot revisits a previously worked area.

[0116] For example, taking a mobile robot as an example, its corresponding data nodes include pose nodes and ground state nodes. The pose node specifically includes position coordinates and three attitude (roll / pitch / yaw) angles. The ground state node has a ground adhesion coefficient, ground flatness and obstacle type corresponding to each grid in its corresponding grid matrix. Furthermore, based on the point cloud data collected by SLAM and the preset feature template of the cultivation rack column, it can be identified whether there is a cultivation rack column in the surrounding area. If there is, the ground state node also needs to include the feature vector corresponding to the cultivation rack column. The feature vector can be obtained from the average curvature and aspect ratio of the point cloud data corresponding to the cultivation rack column. This embodiment does not limit this.Each core constraint can be represented as a corresponding error value. For the first type of constraint, it is necessary to obtain the current position coordinates of the mobile robot and the position coordinates of the adjacent previous moment. Based on the offset between the coordinates and the channel direction, it is determined whether its movement direction deviates from the channel direction. If so, the offset between the coordinates is recorded as the error value; otherwise, the error value is recorded as 0. The determination of whether its movement direction deviates from the channel direction can be combined with a preset offset threshold, which is not limited in this embodiment. For the second type of constraint, firstly, the flatness level corresponding to the current ground is determined based on the average ground flatness. For example, the flatness level includes excellent, good, and poor. When the flatness level is excellent, it means that the ground is flat. Then, based on the corresponding flatness level, the current... The maximum allowable tilt angle corresponding to the previous level is specified in this embodiment. The method of classifying the flatness levels, the specific standards for each level, and the maximum allowable tilt angle for each level are not limited. Those skilled in the art can set these according to actual conditions. Then, the pitch angle corresponding to the mobile robot is obtained, and it is determined whether it is greater than the corresponding maximum allowable tilt angle. If so, the error value is recorded as the difference between the pitch angle and the maximum allowable tilt angle; otherwise, the error value is recorded as 0. For the third type of constraint, the corresponding obstacle type needs to be obtained. If the obstacle type is a rigid obstacle and the ground state node includes the feature vector corresponding to the cultivation rack column, the error value is recorded as 0; for all other cases, the error value is recorded as 1. For the fourth type of constraint… First, it is necessary to determine whether there is an overlapping area between the current location and its adjacent mobile robot. If not, the error value is recorded as 0. If there is, for both mobile robots, the average ground adhesion coefficient of the overlapping area is calculated, and it is determined whether the difference between the two average ground adhesion coefficients is greater than a preset difference in ground adhesion coefficients. If so, the error value is recorded as the difference between the two average ground adhesion coefficients; otherwise, the error value is recorded as 0. For the fifth type of constraint, which corresponds to two error values, it is necessary to determine whether the current position is a position visited by an adjacent mobile robot through interaction with the mobile robot. If so, the average ground adhesion coefficient and the average ground flatness value corresponding to when the adjacent mobile robot reached that position are obtained, and the error value is calculated. The difference between the current ground adhesion coefficient and the current ground flatness is calculated. If the difference between the current ground adhesion coefficient and the current ground flatness is greater than the preset ground adhesion coefficient difference, the first error value is recorded as the difference between the current ground adhesion coefficient and the current ground flatness. Otherwise, the first error value is recorded as 0. If the difference between the current ground flatness and the current ground flatness is greater than the preset ground flatness difference, the second error value is recorded as the difference between the current ground flatness and the current ground flatness. Otherwise, the second error value is recorded as 0. It can be determined whether the current position is a position that the adjacent mobile robot has reached by comparing the similarity between the feature vector corresponding to the cultivation rack column obtained now and the feature vector corresponding to the cultivation rack column obtained by the adjacent robot before.

[0117] Each mobile robot only needs to maintain its own local data and constraints, without storing the global data of all mobile robots. Local data includes its own pose node and ground state node, and local constraints include its own SLAM pose constraints, pose and ground state association constraints, and 3D mesh feature association constraints. That is, only the values ​​corresponding to the first three types of constraints are maintained.

[0118] When multiple robots need to interact, they will not transmit the complete ground state matrix or SLAM point cloud data, but only the key information summary of the overlapping area. The key information summary refers to the values ​​corresponding to the first three types of constraints and the ground adhesion coefficient, ground flatness and obstacle type corresponding to each grid in the overlapping area.

[0119] Then, a distributed optimization algorithm was adopted to break down the global optimization task into local computation for each mobile robot and simple interactions between robots. Global data consistency was achieved through multiple iterations. Specifically, the global objective was to minimize the deviation of all constraint factors to ensure accurate positioning and reliable ground state data. After transformation by the distributed algorithm, the local optimization objective of each robot included two parts: first, minimizing the deviation of its own local constraints; and second, maintaining consistency with the data in the overlapping areas with adjacent mobile robots. At the same time, adjustment parameters were introduced to coordinate the differences between adjacent mobile robots.

[0120] During iterative solving, each robot first calculates its own pose and ground state optimal values ​​based on its own sensor data (SLAM pose information, ground perception adhesion coefficient, etc.) to ensure the rationality of local data. Then, adjacent mobile robots exchange the optimal values ​​and adjustment parameters of overlapping areas, update their mutual interaction constraint deviations, and confirm the data differences between them. Next, based on the difference data fed back by adjacent mobile robots, each robot adjusts its own adjustment parameters. These parameters are continuously optimized with iteration, gradually reducing the data deviation between adjacent mobile robots. Finally, it checks whether the position error of the overlapping area of ​​all adjacent mobile robots is less than a set threshold. This embodiment does not impose restrictions on this. If the requirement is met, the iteration stops, and at this point, the pose and ground state data of all robots have achieved global consistency. If not, it returns to the first step and recalculates until the convergence requirement is met.

[0121] Furthermore, when a robot detects a closed loop, it will add this constraint to its local optimization objective and perform an additional iterative calculation to force the global data to meet the closed loop consistency requirement, thus avoiding the problem of reasonable local data but global misalignment.

[0122] Since each robot's 3D mesh is bound to its own SLAM pose, the mesh coordinates of adjacent mobile robots will automatically align through the aforementioned distributed iterative process, eliminating the need for additional complex stitching algorithms. Ground state parameters in overlapping areas (such as adhesion coefficient and flatness) will be automatically weighted and fused using a confidence-based process that minimizes deviations through interactive constraints, making it more efficient and accurate than the original individual stitching steps. In the event of a communication interruption, the robot can continue optimizing based on its local data, maintaining operational continuity. Once communication is restored, simply exchanging adjustment parameters and iteration counts allows for rapid synchronization of global data, without needing to re-stitch the entire map.

[0123] For example, assuming there are two mobile robots A and B, we will focus on mobile robot A to introduce the above iteration and global ground state map generation steps. First, based on the optimization objective, the objective function formula is determined as follows:

[0124] ;

[0125] in, The local data corresponding to A in the overlapping area of ​​A and B. To optimize the objective, namely, to find the corresponding value for A. To minimize the value of the function that follows. This represents the set of error values ​​corresponding to each constraint factor of A. express An error value corresponding to a constraint factor. This represents the sum of squares of all error values ​​corresponding to A. and These represent the vectors corresponding to the overlapping regions in the local data corresponding to A and B, respectively. They can be obtained by normalizing each parameter included in the local data corresponding to the overlapping regions and then concatenating them. The square norm of a vector. This represents the adjustment parameters between A and B. The adjustment parameters are initially a vector of 0 and need to be updated in subsequent iterations. This indicates the transpose, and then the objective function is solved. This embodiment does not restrict the method for solving the objective function; for example, the Gauss-Newton method can be used. For ease of distinction, the vector corresponding to the overlapping region obtained by the solution is denoted as... Similarly, for mobile robot B, the vector corresponding to the overlapping region can be obtained by solving for it, denoted as... ,judge and If the difference between the values ​​is less than a preset threshold, the iteration stops. This embodiment does not impose a limit on the preset threshold; for example, it can be set to 0.01. If not, it is necessary to determine whether the current position of A is a position that B has visited based on the closed-loop constraints. If not, according to... and Updated adjustment parameters, updated adjustment parameters ,in This represents the penalty coefficient. This embodiment does not limit its specific value; for example, it can be set to 0.1. Similarly, for B, its corresponding adjustment parameter needs to be updated. The update method is the same, and will not be described in detail in this embodiment. Next, based on the updated... , and Repeat the above steps until the condition corresponding to the preset threshold is met, then stop the iteration; if the current position of A is a position that B has reached, first update the adjustment parameters according to the above steps, and then obtain the vector corresponding to when B reached that position. And add constraint terms to the original objective function. Subsequent iterations are all based on the objective function after adding constraints, until the conditions corresponding to the preset threshold are met, at which point the iteration stops. The determination of whether the current position of A is a position that B has reached is performed only once and does not need to be repeated in subsequent iterations.

[0126] The final output is obtained from the last iteration. and ,right and The values ​​in the data are weighted, and the corresponding weights can be set to be the same to obtain the value of each grid in the overlapping area. Then, the remaining grids corresponding to A and B are stitched together with the grids in the overlapping area to obtain the global ground state map.

[0127] This embodiment overcomes the limitations of computing power and communication in traditional centralized optimization by deeply integrating distributed factor graphs with SLAM, while also solving the problem of insufficient accuracy in simple distributed collaboration: the distributed architecture allows each robot to undertake only local computation, greatly reducing hardware pressure and communication overhead; factor graph optimization ensures global consistency of localization and ground state data through multiple constraints, and the binding of 3D meshes with SLAM features makes map stitching more accurate.

[0128] Furthermore, this embodiment provides a step for obtaining motion control parameters based on a global ground state map and ground state parameters, including:

[0129] The robot's localization grid coordinates are obtained from the global ground state map;

[0130] The positioning fusion parameters are obtained based on the robot's positioning association grid coordinates;

[0131] The robot's global coordinates are obtained based on the localization fusion parameters;

[0132] Motion control parameters are obtained based on the robot's global coordinates and ground state parameters.

[0133] In this context, the robot localization associated grid coordinates are coordinates that bind the robot's physical position to the grid structure of the global ground state map. Their core function is to ensure a one-to-one correspondence between the localization result and the ground state data. To guarantee localization accuracy, a multi-source fusion localization algorithm can be used, which integrates multiple localization algorithms. For example, absolute localization (such as UWB), relative localization (RSSI), and inertial pre-integration can be used. Absolute localization refers to obtaining the absolute coordinates of the mobile robot in the global coordinate system by measuring the time difference between the base station and the tag. Relative localization refers to estimating the relative distance between the mobile robot and known anchor points by the received signal strength, thereby determining the coordinates. Inertial pre-integration refers to calculating the mobile robot's trajectory by integrating the acceleration and angular velocity of the inertial sensor, thereby obtaining the coordinates. All of the above methods for obtaining coordinates are existing technologies, and this embodiment will not use them. Without going into details, the positioning fusion parameters are the core control parameters of the multi-source fusion positioning algorithm. They are used to coordinate the contribution ratio of different positioning data sources and noise processing rules. They mainly include data source weights and filtering smoothing coefficients. The data source weights are used to control the contribution ratio of absolute positioning, relative ranging, and inertial pre-integration during fusion. The filtering smoothing coefficients are used to process the noise of each data source to make the positioning results more stable. This embodiment does not limit the specific values. For example, in the initial state, the data source weights can be set to 0.5 for absolute positioning, 0.3 for relative ranging, and 0.2 for inertial pre-integration, and the filtering smoothing coefficient can be set to 0.1 to balance noise and response speed.

[0134] A signal strength threshold is then set. When the relative positioning signal strength is detected to be lower than this threshold, it is determined that the current situation is a signal attenuation scenario. During signal attenuation, the inertial pre-integration weight in the positioning fusion parameters is automatically adjusted from low to high weight, so that the inertial data can play a greater role in the fusion. When the signal returns to normal, the weight is adjusted back to the initial value, and the system re-relies on the high-precision data of absolute positioning and relative ranging. For example, the signal strength threshold is set to -70dBm. When the RSSI signal strength is detected to be -75dBm, it is determined to be an attenuation scenario. The inertial pre-integration weight is increased from 0.2 to 0.8, while the absolute positioning and relative ranging weights are reduced to 0.1. When the signal recovers to -65dBm, the weight is adjusted back to the initial value. This value is only an example to illustrate the logic of weight adjustment.

[0135] Specifically, since high humidity is a continuous production condition in this embodiment, when the UWB / RSSI signal is attenuated by water vapor, the noise of its observation data will increase significantly and the accuracy will drop sharply. However, the IMU is not affected by the humidity environment. Therefore, dynamically adjusting the weights through the signal strength threshold is to ensure positioning continuity when the signal is attenuated and to ensure positioning accuracy when it is normal. Essentially, it uses the complementarity of multiple sensors to improve the robustness of the system. The data source weights will be used in subsequent iteration steps. The higher the weight, the smaller the observation covariance of the corresponding data source, which in turn affects the calculation of the Kalman gain. Finally, the contribution ratio of the data source to the fusion result is determined in the observation update step.

[0136] Next, the dynamically adjusted positioning fusion parameters are substituted into the multi-source fusion filtering logic, and the positioning results are continuously optimized through iterative calculations until the error stabilizes at the lowest level. For example, an extended Kalman filter can be used for iterative coordinate adjustment. First, the predicted value is obtained based on the coordinates corresponding to the previous iteration and the offset obtained from inertial pre-integration. Since this is the first iteration, the coordinates corresponding to the previous iteration can be represented by the coordinates corresponding to UWB. Then, the observation values ​​are obtained based on the UWB coordinates and the RSSI ranging results. For example, the UWB coordinates and the relative distance output by RSSI can be concatenated to form the observation vector. Then, based on the Kalman gain and observation matrix Obtain the observation values The Kalman gain is calculated using the prediction covariance and the observation covariance. The observation covariance is obtained based on the aforementioned weights. For example, since the observation vector is obtained from the x-coordinate, y-coordinate, and relative distance output by the UWB, the observation covariance is a 3×3 diagonal matrix. The three diagonal elements represent the measurement noise variance in the x-direction, the measurement noise variance in the y-direction, and the noise variance of the relative distance, respectively. A larger weight corresponds to a smaller noise variance. This embodiment does not impose specific numerical limitations; for example, the noise variance can be set to [value missing] under normal conditions. During attenuation, the noise variance is set to... The observation matrix is ​​used to convert the predicted robot coordinates into predicted observation values, which are the coordinates corresponding to the current iteration number. The above steps are repeated until the coordinate change between two iterations is less than the error threshold, at which point the iteration stops. In this embodiment, the error threshold is not limited, for example, it can be set to 1 cm. Finally, the output coordinates are set as global coordinates. The principle of extended Kalman filtering is existing technology, and this embodiment will not elaborate on it.

[0137] Ultimately, the precise global coordinates of the mobile robot are obtained, the positioning error is stabilized within a very small range, and the error accumulation rate during long-term operation is significantly reduced, preventing positioning drift due to prolonged operation. Furthermore, these coordinates correspond one-to-one with the robot's positioning grid coordinates, allowing direct access to the ground state parameters of the corresponding grid, providing accurate position and environmental information for subsequent motion control.

[0138] This embodiment binds the robot's physical position to the grid structure of the global ground state map as robot positioning-related grid coordinates, enabling the positioning results to directly interface with ground state data. This achieves deep linkage between perception and control, and subsequent motion control can dynamically adjust parameters based on the current grid's wetness, flatness, and other states. On the other hand, by dynamically adjusting the positioning fusion parameters, the weight of inertial pre-integration is automatically increased in signal attenuation scenarios, ensuring positioning continuity. Ultimately, this stabilizes the robot's positioning error within a small range, reducing the cumulative positioning error rate over long-term operations. This provides a reliable position reference for precise motion control of subsequent cultivation rack channel operations and end-effector steering, fundamentally improving the stability and accuracy of multi-robot collaborative operations.

[0139] Furthermore, this embodiment provides a step for obtaining motion control parameters based on the robot's global coordinates and ground state parameters, including:

[0140] Based on the robot's global coordinates and the global ground state map, the basic differential ratio and basic drive torque are obtained;

[0141] The dynamic motion parameters are obtained based on the basic differential ratio, basic drive torque, and first-order low-pass filter formula.

[0142] Motion control parameters are obtained based on ground condition parameters and dynamic motion parameters.

[0143] Specifically, after obtaining the global coordinates, the corresponding grid is located in the global ground state map. The ground adhesion coefficient and ground flatness data of that grid are extracted. Combining the extracted ground adhesion coefficient and ground flatness, the basic differential ratio and basic driving torque are calculated. The basic differential ratio is the ratio of the rotational speeds of the left and right wheels of the mobile robot, used to adjust steering flexibility and straight-line driving stability. A larger value results in more flexible steering, while a smaller value results in more stable straight-line driving. The basic driving torque is the power output intensity that drives the wheels to rotate, determined by the ground adhesion capability. The higher the ground adhesion coefficient, the greater the torque that can be output, preventing the wheels from rubbing against each other. For example, the ground adhesion coefficient can be multiplied by the scaling factor to obtain the basic differential ratio. The scaling factor is in the range of 0-1. This embodiment does not limit its specific value. Its specific value can be determined according to the actual operation requirements. If more flexible steering is required, the scaling factor can be set to 0.6. Then, the ideal ground flatness (100) is subtracted from the current ground flatness. The difference is divided by 100 to map the ground flatness to the reasonable range corresponding to the torque. Then, the mapping result is multiplied by the upper limit of the maximum torque that the mobile robot can output to obtain the basic driving torque.

[0144] Then, based on the first-order low-pass filter formula, the new dynamic motion parameters are fused with the dynamic motion parameters obtained at the previous moment in a fixed proportion. The dynamic motion parameters corresponding to the previous moment have a higher proportion, while the current basic parameters have a lower proportion, making the parameter changes smoother and preventing sudden torque increases or differential ratio jumps. For example, the fixed proportion is set to 0.2, so the weight corresponding to the new dynamic motion parameters is 0.2, and the weight corresponding to the dynamic motion parameters obtained at the previous moment is 0.8. Next, the smoothed dynamic motion parameters are corrected in conjunction with ground condition adaptation rules. For example, when the ground adhesion coefficient corresponding to the grid is detected to be extremely low (such as a slippery area), the driving torque is further reduced; when the flatness is extremely poor (such as being in a construction expansion joint), the differential ratio is adjusted to improve passability. The ground condition adaptation rule is a parameter correction rule preset for different ground features, such as reducing torque in slippery areas, increasing the differential ratio on roads with low flatness, and decelerating in advance before obstacles. The final generated dynamic motion control parameters will be directly sent to the robot's power and steering system to control the wheel speed and torque output. For example, if the ground adhesion coefficient is less than 0.2, it means that it may be in a slippery area. At this time, the obtained basic driving torque is multiplied by the coefficient 0.6 to reduce the torque. If the ground flatness is less than 80, it means that it may be in a construction expansion joint. The basic differential ratio can be multiplied by the coefficient 1.2 to increase the differential ratio. This value is only an example to illustrate the logic of reducing torque and increasing differential ratio. This embodiment does not limit its specific values.

[0145] This embodiment first binds the robot's precise global coordinates to the grid of the global ground state map, allowing the basic motion parameters to directly match the adhesion coefficient and flatness of the current grid, avoiding insufficient adaptation of general parameters in scenarios such as alternating wet and dry conditions and bumpy roads. Second, smooth transition processing suppresses vehicle vibration caused by parameter abrupt changes, improving driving stability and equipment lifespan. Finally, combined with the correction of ground state adaptation rules, the scenario adaptability of parameters is further optimized, ultimately significantly reducing wheel slippage rate and vehicle vibration amplitude. This ensures the continuity and accuracy of operations in the cultivation rack channel while reducing impact losses in the power system, improving the overall efficiency and reliability of multi-robot collaborative operations from the motion control perspective.

[0146] Furthermore, this embodiment provides a step for obtaining motion control parameters based on ground state parameters and dynamic motion parameters, including:

[0147] Generate trigger commands based on the robot's global coordinates and the global ground state map;

[0148] Based on the trigger command and ground status parameters, the pre-aiming steering angle and deceleration speed are obtained;

[0149] Motion control parameters are obtained based on dynamic motion parameters, pre-aiming steering angle, and deceleration speed.

[0150] Specifically, based on the robot's precise global coordinates and the global ground state map, the system calculates the straight-line distance from the robot's current position to the end of the current cultivation rack aisle in real time. When this distance is detected to be less than a preset distance threshold, the system immediately generates a pre-aiming control trigger command for the end of the cultivation rack aisle, notifying subsequent modules to enter the turning preparation state. This avoids emergency turning only when approaching the end. For example, if the preset distance threshold is 2m, based on the pre-aiming control trigger command, it is necessary to collect in real time the offset between the robot's current position and the center line of the cultivation rack aisle, as well as the remaining end distance (i.e., the straight-line distance from the previous position to the end of the aisle). Then, combining the offset and the end distance, the system calculates the pre-aiming turning angle and the travel speed after deceleration. If the offset is large, the steering angle will be adjusted accordingly to correct the path; if the remaining distance is short, the deceleration will be greater to ensure smooth steering. For example, first, the ratio of the offset to the end distance is calculated, which represents the slope that needs to be corrected. Then, the arctangent function is used to convert the slope into a pre-aiming steering angle, ensuring that the robot turns in advance and corrects to the centerline just when it reaches the end. The travel speed after deceleration can be obtained based on a linear deceleration strategy, such as calculating the ratio of the end distance to a preset distance threshold and multiplying this ratio by the base speed to obtain the travel speed after deceleration. The base speed is the speed at which the robot operates normally in the channel and can be obtained according to the actual situation. Finally, the dynamic motion parameters, the pre-aiming steering angle, and the travel speed after deceleration are integrated into a complete motion control parameter set, which then generates corresponding motion control commands that include straight-line operation in the cultivation rack channel and smooth steering at the end. This allows the mobile robot to execute subsequent motion operations according to the motion control parameter set, achieving smoothness and accuracy throughout the operation.

[0151] This application provides a mobile robot control system that supports edge-side intelligent reasoning, including:

[0152] The initialization module is used to determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario.

[0153] The calculation module is used to obtain the local ground state matrix based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and characteristic parameters. The characteristic parameters include the ground normal reaction force, slippage rate, and driving torque.

[0154] The interaction module is used to obtain a global ground state map based on the local ground state matrix and interaction data. The interaction data is obtained from neighboring mobile robots.

[0155] The control module is used to obtain motion control parameters based on the global ground state map and ground state parameters.

[0156] This application embodiment constructs a perception fusion mechanism based on the structural features of the cultivation rack channel on the support end. Utilizing a scene-adaptive robust weighted fusion algorithm, it accurately stitches together the local ground state data of multiple robots, fundamentally overcoming the limitations of traditional single-source perception. This makes the fused global map more closely match the continuous structural features of agricultural ground, effectively suppressing abnormal data interference. Furthermore, combined with coordinate calibration logic constrained by the heading angle of the cultivation rack channel, it improves the accuracy of map alignment and adapts to signal attenuation and straight-line constraint requirements. At the motion control level, through dynamic parameter adaptation driven by the global ground state and pre-aiming control at the end of the cultivation rack channel, it achieves smooth transition of motion parameters and advance prediction of steering actions from a control logic perspective, avoiding vehicle vibration caused by sudden changes in ground state.

[0157] Furthermore, the control method provided in this embodiment can be applied not only to agricultural scenarios such as the cultivation of mushrooms like button mushrooms, but also to industrial scenarios such as intelligent warehousing and collaborative production line operations. Specifically, when applied to industrial scenarios, the parameters of the specific scenario should be adjusted. For example, the cultivation rack aisle in the button mushroom cultivation scenario can be replaced with the shelf aisle / production line aisle in the industrial scenario, and the spacing between cultivation units can be replaced with the shelf spacing / production line unit spacing. The default state parameters also need to be adjusted to adapt to the industrial scenario. In motion control, the upper limit of torque can be increased according to the load requirements of the industrial robot, and the pre-aiming turning threshold is set to 2.0 meters to adapt to the longer working distance of the industrial aisle. The ground state adaptation rules are adjusted to be specific to the industrial scenario. For example, the torque is reduced in oily areas to prevent slippage, and the differential ratio is increased to quickly avoid obstacles left on the pallet to prevent material tipping.

[0158] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0160] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0161] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A mobile robot control method supporting end-side intelligent inference, characterized by, The mobile robot control method is executed by each mobile robot, and the mobile robot control method includes: Determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario; Based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and the characteristic parameters, a local ground state matrix is ​​obtained. The characteristic parameters include the ground normal reaction force, slippage rate, and driving torque. A global ground state map is obtained based on the local ground state matrix and the interaction data, wherein the interaction data is obtained from neighboring mobile robots; Motion control parameters are obtained based on the global ground state map and ground state parameters.

2. The mobile robot control method of claim 1, wherein, The step of determining the initial ground state matrix based on the current scenario includes: The grid resolution is obtained based on the spacing between button mushroom cultivation units in the current scene; The dimensions of the initial ground state matrix are determined based on the grid resolution and its corresponding effective coverage area. The initial adhesion coefficient and initial flatness are determined based on historical data to obtain default state parameters, and each element in the initial ground state matrix is ​​assigned the default state parameter.

3. The mobile robot control method of claim 1, wherein, The process of obtaining the local ground state matrix based on the initial ground state matrix, the heading angle of the cultivation rack channel centerline, and characteristic parameters includes: The current ground adhesion coefficient is obtained based on the aforementioned characteristic parameters; Obtain the current lateral tilt angle and longitudinal pitch angle to determine the current ground flatness; Obtain the current obstacle type label from lidar point cloud data; The local ground state matrix is ​​obtained based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, the current ground adhesion coefficient, the current ground flatness, and the current obstacle type label.

4. The mobile robot control method of claim 1, wherein, The interaction data is obtained from neighboring mobile robots and includes: The adjacent mobile robots obtain the core network based on their corresponding local ground state matrix; Extract the state change data of the core network to obtain the differential data to be broadcast; Data confidence is obtained based on sensor health and time decay coefficient; The interaction data is obtained based on the differential data to be broadcast and the data confidence level.

5. The mobile robot control method of claim 4, wherein, The step of obtaining a global ground state map based on the local ground state matrix and interaction data includes: A target function is constructed based on the local ground state matrix and interactive data; Based on the objective function and dynamic matching rules, an overlapping grid matching set is obtained; The confidence-weighted matching error is obtained based on the overlapping grid matching set and the data confidence level. The local matrix coordinate offset is obtained by minimizing the confidence-weighted matching error. The global ground state map is obtained based on the local matrix coordinate offset.

6. The mobile robot control method of claim 5, wherein, The step of obtaining the global ground state map based on the local matrix coordinate offset includes: Based on the local matrix coordinate offset and the interaction data, the local matrix data is obtained; Based on the dynamic Hube threshold and the residual direction weights, the direction-weighted Hube loss function is obtained; The robust fusion weight coefficients are obtained based on the direction-weighted Hubey loss function and the regularized weights. The global ground state map is obtained based on the robust fusion weight coefficients and the local matrix data.

7. The mobile robot control method of claim 6, wherein, The process of obtaining motion control parameters based on the global ground state map and ground state parameters includes: The robot's positioning grid coordinates are obtained based on the global ground state map. The positioning fusion parameters are obtained based on the robot's positioning-related grid coordinates; The robot's global coordinates are obtained based on the localization fusion parameters; Motion control parameters are obtained based on the robot's global coordinates and the ground state parameters.

8. A mobile robot control method supporting edge-side intelligent reasoning according to claim 7, characterized in that, The process of obtaining motion control parameters based on the robot's global coordinates and the ground state parameters includes: Based on the robot's global coordinates and the global ground state map, the basic differential ratio and basic drive torque are obtained; The dynamic motion parameters are obtained based on the basic differential ratio, the basic drive torque, and the first-order low-pass filter formula. The motion control parameters are obtained based on the ground state parameters and the dynamic motion parameters.

9. A mobile robot control method supporting edge-side intelligent reasoning according to claim 8, characterized in that, The step of obtaining the motion control parameters based on the ground state parameters and the dynamic motion parameters includes: Based on the robot's global coordinates and the global ground state map, a trigger command is generated; Based on the trigger command and the ground state parameters, the pre-aiming steering angle and deceleration speed are obtained; The motion control parameters are obtained based on the dynamic motion parameters, the pre-aiming steering angle, and the deceleration speed.

10. A mobile robot control system supporting edge-side intelligent reasoning, characterized in that, include: The initialization module is used to determine the initial ground state matrix and the heading angle of the cultivation rack channel centerline based on the current scenario. The calculation module is used to obtain a local ground state matrix based on the initial ground state matrix, the heading angle of the centerline of the cultivation rack channel, and characteristic parameters, wherein the characteristic parameters include ground normal reaction force, slippage rate, and driving torque. An interaction module is used to obtain a global ground state map based on the local ground state matrix and the interaction data, wherein the interaction data is obtained from neighboring mobile robots; The control module is used to obtain motion control parameters based on the global ground state map and ground state parameters.

Images