Photovoltaic paving robot following control method based on adaptive pose estimation

Abstract

The invention discloses a photovoltaic paving robot following control method based on self-adaptive pose estimation, and aims to solve the problem of collision between a mechanical arm and a photovoltaic bracket during non-paving ground operation. According to the method, a solid-state laser radar of a multiplexing robot is used for collecting point cloud data, a target coordinate system is constructed after preprocessing, plane extraction and edge and corner point detection, the attitude difference between the target coordinate system and a template coordinate system is calculated, and the coordinates of transition points of the mechanical arm are corrected by combining filtering optimization and a smooth interpolation algorithm. The method is simple in operation, does not need to additionally install a sensor, is low in operation requirement, can be directly deployed in an existing controller, can accurately avoid collision risks, is adaptive to complex terrains such as deserts and plateaus, and improves the installation safety and consistency of the photovoltaic module and the environmental adaptability of the robot.

Description

Technical Field

[0001] This invention relates to the field of industrial robot control technology, and more specifically to a photovoltaic paving robot following control method based on adaptive pose estimation. Background Technology

[0002] Centralized photovoltaic (PV) power plants are the mainstream form of PV power generation, and PV module installation is the core construction process that directly affects the construction cycle and operational efficiency of the power plant. Currently, PV module installation relies heavily on manual labor, resulting in high labor costs, long construction cycles, high safety risks associated with working at heights, and the risk of microcracks in modules due to improper operation. Therefore, the gradual promotion and application of PV module installation robots can significantly improve efficiency and reduce labor intensity and the risk of module damage.

[0003] With increasing demands for ecological protection and cost control, centralized photovoltaic power stations are mostly built in deserts, plateaus, and other areas, eliminating the need for ground leveling. Photovoltaic installation robots must operate on undulating, unpaved surfaces. Currently, the robotic arms of existing robots are fixed to a chassis, and their movement trajectories rely on preset fixed transition points. When the chassis changes position with the terrain, the robotic arm's preset trajectory is prone to interference and collisions with the already installed photovoltaic supports, affecting operational safety and continuity.

[0004] The problem of obstacle avoidance for robotic arms has been explored in related fields: in the field of autonomous driving, multiple sensors are used in conjunction with real-time detection to avoid obstacles, which requires a large amount of data processing and high computing power; in the field of robotic arm control, either similar real-time trajectory planning schemes are used, which are complex and costly, or collision detection sensors are used, which are "post-collision protection" and are not suitable for heavy industrial robotic arms. Summary of the Invention

[0005] Based on the problems described above, existing technologies cannot meet the obstacle avoidance requirements of the robotic arm when photovoltaic installation robots are operating on unpaved ground. There is an urgent need for an adaptive following control method that does not require additional sensors, has low computing power requirements, and is compatible with heavy-duty robotic arms, in order to solve the collision risk between the robotic arm and the photovoltaic support and ensure the safety and efficiency of the installation operation. To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A photovoltaic paving robot following control method based on adaptive pose estimation, the control method includes the following steps: S1. After the photovoltaic installation robot moves to the preset work point, it stops moving and starts its built-in solid-state lidar to scan data, collect raw point cloud data and preprocess it to form an effective point cloud covering the photovoltaic module that has been installed in the previous area. S2. The effective point cloud is segmented into planes using a random sampling consensus algorithm to extract the target plane where the photovoltaic module is located; S3. Perform edge detection and contour optimization on the target plane to obtain the continuous edges of the photovoltaic module; S4. Extract multiple straight lines from the continuous edge using a straight line detection algorithm and obtain their parameters; S5. Filter out straight line pairs with an included angle between 80° and 100° and a length that meets the preset threshold. S6. Calculate the intersection point of the line pair and extract the corner coordinates of the photovoltaic module; S7. Based on the normal to the target plane and the edge lines where the corner points are located, construct the target coordinate system; S8. Calculate the attitude difference between the target coordinate system and the preset template coordinate system. The attitude difference includes translation offset and rotation angle. The translation offset is the Euclidean distance between the origin of the target coordinate system and the origin of the template coordinate system. The rotation angle is the arithmetic mean of the angles between the corresponding coordinate axes of the target coordinate system and the template coordinate system. S9. The attitude differences are filtered and fused to obtain the optimized pose estimation result; the filtering process uses the Kalman filter algorithm. The core state prediction equation of this algorithm is: ; The core state update equation is: ; in, The subscripts k and k-1 represent the current time and the previous time, respectively. A represents the state transition matrix, B represents the control input matrix, u represents the control input, w represents the process noise, K represents the Kalman gain, Z represents the observation value, and H represents the observation matrix. S10. Based on the optimized pose estimation results, adjust the transition point coordinates of the robotic arm to achieve adaptive obstacle avoidance and following control.

[0006] In the preferred embodiment, in step S1, before starting the solid-state lidar scanning, the robot chassis is calibrated using the robot's built-in inertial measurement unit to ensure that the lidar scanning reference is consistent with the robot's body coordinate system; the point cloud preprocessing includes using a statistical filtering algorithm to remove isolated noise points and using a voxel grid downsampling algorithm to adjust the point cloud density to 500 to 1000 points per square centimeter. The statistical filtering algorithm is implemented as follows: for each point in the original point cloud... Find its K nearest neighbors in the point cloud and calculate the value of that point. Calculate the arithmetic mean of the Euclidean distances to the K nearest neighbors. and standard deviation Set the distance threshold to If point If the actual distance to any of its nearest neighbors is greater than the threshold, then the point will be... Points identified as noise are removed; where K ranges from 15 to 25 and α ranges from 1 to 2.

[0007] In the preferred embodiment, the scanning frequency of the solid-state lidar in step S1 is 10Hz to 20Hz, the horizontal angular resolution × vertical angular resolution is not higher than 0.1 degrees × 0.1 degrees, and the effective measurement distance range is not less than 0.5 meters and not more than 50 meters.

[0008] In a preferred embodiment, the preset number of iterations for the random sampling consensus algorithm in step S2 is 1000 to 5000, and the interior point distance threshold is set to 0.01 meters to 0.05 meters. The execution process of the algorithm includes: in each iteration, three non-collinear points are randomly selected from the effective point cloud, and an initial planar model is fitted based on the three points; The standard equation of the initial planar model is: Calculate the vertical distance from each of the remaining points in the effective point cloud to the initial plane, and mark the points whose distance is less than the inner point distance threshold as the inner points of this iteration; count the number of inner points in this iteration, and if the number of inner points accounts for more than 70% of the total number of points in the effective point cloud, then the initial plane model is determined to be valid, and the final plane fitted by the inner point set corresponding to the model is determined as the target plane.

[0009] In the preferred embodiment, after extracting corner points in step S6, the deviation between the distance between adjacent corner points and the side length of the standard photovoltaic module is calculated for verification. The formula for calculating the deviation δ is: ; in, The extracted Euclidean distance between adjacent corner points. The standard photovoltaic module's design side length is used; if the deviation δ exceeds 5%, the corner point extraction is deemed abnormal, and steps S2 to S6 need to be repeated.

[0010] In the preferred embodiment, if multiple corner points are extracted in step S7, multiple candidate target coordinate systems are constructed based on the different corner points; by calculating the overlap between each candidate target coordinate system, the three coordinate systems with the highest overlap are selected for subsequent average calculation of attitude differences; The formula for calculating the degree of overlap C is: ; Where d is the Euclidean distance between the origins of the two compared coordinate systems. The maximum allowable distance between origin points is set to 0.03 meters. This is the arithmetic mean of the angles between three pairs of corresponding coordinate axes in two coordinate systems. The maximum allowable angle between the coordinate axes is set to 2°; and Let be the weighting coefficient, satisfying , and All values ​​are 0.5.

[0011] In the preferred embodiment, the template coordinate system described in step S8 supports dynamic updating; The update rule is as follows: During continuous installation, record the attitude differences of the N most recently installed photovoltaic modules, calculate the arithmetic mean of these N translation offsets as the new translation reference, and calculate the arithmetic mean of these N rotation angles as the new rotation reference; when the deviations of the translation offsets and rotation angles in all N attitude differences relative to their arithmetic mean are less than the preset tolerance, generate and replace the original template coordinate system based on the new translation reference and rotation reference; wherein, the preset tolerance is that the translation offset does not exceed 0.01 meters and the rotation angle does not exceed 1°, and N is an integer greater than or equal to 3.

[0012] In the preferred embodiment, the correction of the transition point of the robotic arm in step S10 adopts a smooth transition algorithm, which realizes the motion trajectory planning from the original transition point to the compensated target transition point through cubic polynomial interpolation. The expression for the cubic polynomial interpolation is: Where τ is a normalized time variable, ranging from 0 to 1, corresponding to the start and end times of the transition phase, respectively. Let τ be the coordinates of the robotic arm's transition point in the task space at time τ; a, b, c, and d are the interpolation coefficients to be solved. The interpolation coefficients are obtained by solving for the set boundary conditions, which include: at the start of the transition, the position, velocity, and acceleration are known; at the end of the transition, the position, velocity, and acceleration are known; specifically: , , ; , , ; in, , , These represent the position, velocity, and acceleration of the original transition point, respectively. , , These are the position, velocity, and acceleration of the target transition point after compensation; By solving the above system of equations to determine the coefficients a, b, c, and d, the generated trajectory must satisfy the condition that the absolute value of its jerk (the derivative of acceleration) does not exceed 5 m / s³ throughout the entire transition process.

[0013] In the preferred embodiment, step S10 further includes a real-time monitoring step: Local point cloud data of the robotic arm's workspace is acquired in real time using LiDAR. A plane fitting algorithm is then used to identify and fit the plane equation of the supporting plane where the photovoltaic bracket is located. The standard form of the plane equation is as follows: ; Arbitrary sampling point on the planned trajectory of the computational robotic arm Vertical distance to the fitted plane The calculation formula is: ; If there exists a sampling point Q such that the distance is... Less than the preset safe distance threshold If the current movement of the robotic arm is immediately paused and an audible and visual alarm signal is triggered; wherein, the safe distance threshold... The value ranges from 0.1 meters to 0.2 meters.

[0014] In a preferred embodiment, the method further includes a fault-tolerant step: The formula for calculating the effective point cloud data volume ratio η is: ; in, This refers to the number of points contained in the effective point cloud obtained after preprocessing in step S1. This represents the total number of points in the original point cloud collected by the solid-state lidar in S1. When a fault occurs, the control system interrupts the current adaptive following control process, automatically calls and executes the preset safe transition trajectory, and controls the robotic arm to return to the safe initial position or standby position along the safe trajectory. At the same time, a clear fault warning signal is issued through the robot's alarm module, and all paving operations are suspended until the operator troubleshoots the fault and confirms that the system has returned to normal.

[0015] A photovoltaic paving robot following control method based on adaptive pose estimation has the following beneficial effects, including but not limited to the following: 1. Data is collected directly using the positioning solid-state lidar built into the photovoltaic paving robot, without the need for additional sensors, avoiding the calibration difficulties of multi-sensor fusion, adapting to existing hardware architecture, and easy to promote; 2. Point cloud processing is completed by utilizing the stationary gap before the robotic arm starts working. No real-time trajectory planning is required. It can be deployed directly on the existing controller without upgrading the hardware, thus lowering the threshold for implementation. 3. By accurately estimating the pose relationship through multi-step algorithms, combined with filtering optimization and real-time monitoring, the collision risk between the robotic arm and the photovoltaic support is avoided from the source, protecting the integrity of the components and equipment; 4. Supports dynamic updates of the template coordinate system to ensure consistency in batch installation; the robotic arm trajectory transitions smoothly, reducing mechanical impact and extending equipment lifespan; 5. It achieves adaptive trajectory correction to address terrain undulations, eliminating the need for flat ground and significantly improving the environmental adaptability of photovoltaic paving robots in desert, plateau, and other similar environments. Attached Figure Description

[0016] Figure 1 The solid-state lidar of this invention requires several seconds of scanning time to form a dense point cloud map; Figure 2 This invention extracts a plan view of a previously installed photovoltaic module; Figure 3 This invention obtains a planar edge map of a photovoltaic module through edge detection; Figure 4 This invention compares pairs of lines detected by the present invention and extracts pairs of lines with an included angle between 80 and 100°. Figure 5 It involves finding the intersection point using the shortest distance method to extract the corner point map of the photovoltaic module; Figure 6 The attitude difference diagram is calculated using this coordinate system and the template coordinate system. Detailed Implementation

[0017] Example 1: Photovoltaic module installation on slightly undulating, unpaved ground in desert areas This embodiment is applied to a centralized photovoltaic power station in the Northwest desert. The power station uses standard photovoltaic modules measuring 6.5 feet by 3.25 feet, weighing 60 kg. The ground undulation height of the working area does not exceed 0.3 meters, and the average daily solar irradiance is 80 to 100 kilolux. The photovoltaic installation robot is equipped with a tracked mobile chassis, a six-degree-of-freedom industrial robotic arm, and a solid-state LiDAR. The LiDAR scanning frequency is 15 Hz, and the horizontal angular resolution × vertical angular resolution is 0.1 degrees × 0.1 degrees. The controller supports the deployment of the PCL point cloud processing algorithm library. S1, such as Figure 1As shown, the photovoltaic installation robot moves to the preset work point via the automatic navigation system and then stops its chassis movement. The inertial measurement unit is then activated to calibrate the chassis attitude, with a calibration error not exceeding 0.5 degrees, ensuring that the LiDAR scanning reference is consistent with the robot's coordinate system. The solid-state LiDAR is controlled to continuously scan for 5 seconds, acquiring the original point cloud data covering the previously installed photovoltaic module. A statistical filtering algorithm is used to remove isolated noise points, with K set to 20 and α set to 1.5. A voxel grid downsampling algorithm is then used to adjust the point cloud density to 800 points per square centimeter, forming an effective point cloud. Figure 1 This is a schematic diagram of the dense point cloud formed in this step, clearly covering the photovoltaic module and the surrounding area; S2, such as Figure 2 As shown, the random sample consensus algorithm is used to segment the effective point cloud into a plane, with 3000 iterations and an interior point distance threshold of 0.03 meters. In each iteration, three non-collinear points are randomly selected to fit the initial planar model. The standard equation of the planar model is... Calculate the vertical distance from the remaining points to the plane, and mark points with a distance less than a threshold as inliers. When the number of inliers accounts for more than 70% of the total number of valid point clouds, the model is considered valid, and the final target plane is fitted with the corresponding set of inliers. Figure 2 This is a schematic diagram of the extracted photovoltaic module target plane, clearly separated from the background point cloud; S3, such as Figure 3 As shown, the Canny edge detection algorithm is used to obtain the initial edge contour of the target plane, and the interference of discrete noise points is eliminated by morphological dilation-erosion operation to obtain the continuous and complete edge of the photovoltaic module plane. Figure 3 This is an optimized schematic diagram of the photovoltaic module's edge, accurately representing the rectangular outline of the module; S4, such as Figure 4 As shown, the Hough line detection algorithm is used to extract 12 straight lines from a continuous edge, and the slope, intercept, and extension range parameters of each line are obtained. All lines are compared pairwise, the dot product of the direction vectors between the lines is calculated, and four pairs of lines are selected with an included angle between 85° and 95° and a length not less than 80% of the corresponding side length of the standard component. Figure 4 This is a schematic diagram of the filtered line pairs, retaining only the key lines corresponding to the four edges of the component; S5, such as Figure 5 As shown, the intersection points of four pairs of straight lines were calculated using the shortest distance criterion, resulting in four candidate corner points. After verifying the spacing between adjacent corner points, the standard component side length was 3.25 feet, with an allowable deviation of no more than 5%. One abnormal intersection point was eliminated, and finally, the coordinates of four valid corner points were extracted, with an error of no more than two millimeters. Figure 5 This is a schematic diagram of the extracted photovoltaic module corner points, precisely corresponding to the four corner positions of the module; S6, such as Figure 6As shown, the plane normal vector is calculated based on the target plane. The effective corner point at the lower left corner is taken as the origin, and the two edge lines where the corner point is located and the plane normal are taken as the X-axis, Y-axis and Z-axis respectively to construct the target coordinate system. The preset template coordinate system is retrieved, and the initial attitude difference with a translation offset of 0.02 meters and a rotation angle of 0.8 degrees is calculated through coordinate system transformation matrix operations. Figure 6 This diagram illustrates the attitude differences between the two coordinate systems, clearly showing the relative positions and angular relationships. S7. A Kalman filter algorithm is used to suppress noise from the initial attitude difference. Inertial measurement unit calibration data is then fused to obtain optimized pose estimation results, with a translational offset of 0.018 meters and a rotation angle of 0.7 degrees. This difference is converted into adjustment amounts for the joint angles of the robotic arm, and the transition point coordinates of the robotic arm are smoothly corrected using a cubic polynomial interpolation algorithm. The interpolation expression is: , The transition time variable τ is set to range from 0 to 1, and the boundary conditions are as follows: , , , , , ; After obtaining the interpolation coefficients, the motion trajectory is generated, ensuring that the absolute value of the jerk does not exceed 5 / second. 3 ; S8. When the robotic arm moves along the corrected trajectory, the LiDAR collects local point clouds in real time, fits the photovoltaic support plane, and calculates the distance from the trajectory sampling points to that plane, setting a safety distance threshold of 0.15 meters. No distances below the threshold occurred during the operation, and the robotic arm successfully completed the component installation. When the effective point cloud coverage of the LiDAR was only 28% during a particular operation, the system immediately activated a safe transition trajectory, controlled the robotic arm to return to its initial position, and issued a fault warning. After inspection and cleaning of the LiDAR lens, operation resumed.

[0018] Example 2: Photovoltaic module installation suitable for high-altitude, high-sunlight, undulating, unpaved ground. This embodiment is applied to a photovoltaic power station at an altitude of 4558 meters in Tibet. The power station uses lightweight photovoltaic modules weighing 39 kg, with dimensions of 1.6 m * 1 m. The ground undulation height of the working area does not exceed 0.5 m, the average daily solar intensity is 100 to 120 kilolux, and the temperature is between -10°C and 15°C. The photovoltaic installation robot is equipped with a low-temperature resistant tracked chassis, a six-degree-of-freedom industrial robotic arm, and a high-interference-resistant solid-state lidar. The lidar scanning frequency is 20 Hz, and the horizontal angular resolution × vertical angular resolution is 0.08 degrees × 0.08 degrees, making it suitable for high-altitude, low-oxygen environments without human intervention. S1, such as Figure 1As shown, the photovoltaic installation robot stops moving after reaching the preset work point and activates the inertial measurement unit to calibrate the chassis attitude. The calibration error does not exceed 0.3 degrees, compensating for the reference deviation caused by the tilt of the plateau terrain. Considering strong light interference, the solid-state lidar is controlled to continuously scan for 8 seconds to collect the original point cloud data of the previously installed photovoltaic module. A statistical filtering algorithm is used to remove isolated noise points, with K set to 25 and α set to 2.0 to enhance the anti-light interference capability. The point cloud density is adjusted to 1000 points per square centimeter using a voxel grid downsampling algorithm to form an effective point cloud. Figure 1 This is a schematic diagram of the high-density point cloud formed in this step, effectively filtering out false points caused by strong light; S2, such as Figure 2 As shown, the random sampling consensus algorithm is used to segment the effective point cloud into planes. The number of iterations is set to 5,000 to improve the accuracy of plane extraction in complex terrain, with an interior point distance threshold of 0.02 meters. In each iteration, three non-collinear points are randomly selected to fit the initial plane model. The standard equation of the plane model is... Calculate the perpendicular distances of the remaining points to the plane and label the interior points. When the proportion of interior points exceeds 75%, the model is considered valid, and the final target plane is fitted using the corresponding set of interior points. Figure 2 This is a schematic diagram of the target plane of the extracted photovoltaic module, effectively eliminating background interference caused by terrain undulations; S3, such as Figure 3 As shown, the target plane is first subjected to grayscale threshold filtering to suppress strong light reflection interference, and then the Canny edge detection algorithm is used to obtain the initial edge contour. After optimization through two morphological dilation-erosion operations, a continuous and complete photovoltaic module plane edge is obtained. Figure 3 This is an optimized schematic diagram of the photovoltaic module's edge, clearly showing the module's outline details; S4, such as Figure 4 As shown, 15 straight lines in a continuous edge are extracted using the Hough line detection algorithm. After obtaining the parameters of each line, they are compared pairwise to calculate the angle between the lines. Four pairs of lines are selected with an angle between 80° and 100° and a length not less than 90% of the corresponding side length of the standard component. Figure 4 This is a schematic diagram of the filtered straight line pairs, precisely preserving the four edge lines of the component; S5, such as Figure 5 As shown, the intersection points of four pairs of straight lines are calculated using the shortest distance criterion, resulting in four candidate corner points. Through verification of the spacing between adjacent corner points, the standard component has a side length of one meter, with an allowable deviation of no more than 5%. All intersection point deviations meet the requirements and are directly determined as valid corner point coordinates, with an error not exceeding 1.5 mm. Figure 5 This is a schematic diagram of the extracted photovoltaic module corner points, with corner point positioning accuracy adapted to the high-precision installation requirements in high-altitude areas; S6, such as Figure 6As shown, based on the target plane normal vector, a target coordinate system is constructed with the top-left effective corner point as the origin, and the two edge lines containing this corner point and the plane normal as the coordinate axes. The preset template coordinate system is retrieved, and the initial attitude difference with a translation offset of 0.04m and a rotation angle of 1.2° is calculated. Considering batch installation consistency, the template coordinate system dynamic update function is enabled, with N set to 5, and the attitude differences of the last five components are recorded and the average value is calculated. Figure 6 This is a schematic diagram illustrating the attitude differences between the two coordinate systems, clearly showing the offsets and angles that need to be adjusted; S7. The Kalman filter algorithm is used to fuse inertial measurement unit data to optimize attitude differences, resulting in a final translational offset of 0.038m and a rotation angle of 1.1°. This result is then converted into joint adjustment amounts for the robotic arm, and the transition point coordinates are corrected using a cubic polynomial interpolation algorithm. The boundary conditions are set as follows: , , , , , The absolute value of the acceleration of the generated motion trajectory does not exceed 5 / second. 3 To avoid impact on the robotic arm joints in low-temperature environments; S8. During the robotic arm's movement, the lidar collects local point clouds in real time and fits them to the photovoltaic support plane, calculating the distance from the trajectory sampling points to this plane. The safety distance threshold is set at 0.12m. If a sampling point is 0.1m away during operation, the system immediately pauses the robotic arm's movement and triggers an audible and visual alarm. Operation resumes after adjusting the transition point coordinates. The entire process requires no human intervention, avoiding safety risks associated with operation in high-altitude, hypoxic environments.

[0019] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention; no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A photovoltaic paving robot following control method based on adaptive pose estimation, characterized in that, The control method includes the following steps: S1. After the photovoltaic installation robot moves to the preset work point, it stops moving and starts its built-in solid-state lidar to scan data, collect raw point cloud data and preprocess it to form an effective point cloud covering the photovoltaic module that has been installed in the previous area. S2. The effective point cloud is segmented into planes using a random sampling consensus algorithm to extract the target plane where the photovoltaic module is located; S3. Perform edge detection and contour optimization on the target plane to obtain the continuous edges of the photovoltaic module; S4. Extract multiple straight lines from the continuous edge using a straight line detection algorithm and obtain their parameters; S5. Filter out straight line pairs with an included angle between 80° and 100° and a length that meets the preset threshold. S6. Calculate the intersection point of the line pair and extract the corner coordinates of the photovoltaic module; S7. Based on the normal to the target plane and the edge lines where the corner points are located, construct the target coordinate system; S8. Calculate the attitude difference between the target coordinate system and the preset template coordinate system. The attitude difference includes translation offset and rotation angle. The translation offset is the Euclidean distance between the origin of the target coordinate system and the origin of the template coordinate system. The rotation angle is the arithmetic mean of the angles between the corresponding coordinate axes of the target coordinate system and the template coordinate system. S9. The attitude differences are filtered and fused to obtain the optimized pose estimation result; the filtering process uses the Kalman filter algorithm. The core state prediction equation of this algorithm is: ; The core state update equation is: ; in, The subscripts k and k-1 represent the current time and the previous time, respectively. A represents the state transition matrix, B represents the control input matrix, u represents the control input, w represents the process noise, K represents the Kalman gain, Z represents the observation value, and H represents the observation matrix. S10. Based on the optimized pose estimation results, adjust the transition point coordinates of the robotic arm to achieve adaptive obstacle avoidance and following control.

2. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, In step S1, before starting the solid-state lidar scanning, the robot chassis is calibrated using the robot's built-in inertial measurement unit to ensure that the lidar scanning reference is consistent with the robot's body coordinate system; the point cloud preprocessing includes using a statistical filtering algorithm to remove isolated noise points and using a voxel grid downsampling algorithm to adjust the point cloud density to 500 to 1000 points per square centimeter. The statistical filtering algorithm is implemented as follows: for each point in the original point cloud... Find its K nearest neighbors in the point cloud and calculate the value of that point. Calculate the arithmetic mean of the Euclidean distances to the K nearest neighbors. and standard deviation Set the distance threshold to If point If the actual distance to any of its nearest neighbors is greater than the threshold, then the point will be... Points identified as noise are removed; where K ranges from 15 to 25 and α ranges from 1 to 2.

3. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, The scanning frequency of the solid-state lidar mentioned in step S1 is 10Hz to 20Hz, the horizontal angular resolution × vertical angular resolution is no higher than 0.1 degrees × 0.1 degrees, and the effective measurement distance range is no less than 0.5 meters and no more than 50 meters.

4. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, The preset number of iterations for the random sampling consensus algorithm in step S2 is 1000 to 5000, and the interior point distance threshold is set to 0.01 meters to 0.05 meters. The execution process of the algorithm includes: in each iteration, three non-collinear points are randomly selected from the effective point cloud, and an initial planar model is fitted based on the three points; The standard equation of the initial planar model is: Calculate the vertical distance from each of the remaining points in the effective point cloud to the initial plane, and mark the points whose distance is less than the inner point distance threshold as the inner points of this iteration; count the number of inner points in this iteration, and if the number of inner points accounts for more than 70% of the total number of points in the effective point cloud, then the initial plane model is determined to be valid, and the final plane fitted by the inner point set corresponding to the model is determined as the target plane.

5. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, After extracting corner points in step S6, the deviation between the distance between adjacent corner points and the side length of the standard photovoltaic module is calculated for verification. The formula for calculating the deviation δ is: ; in, The extracted Euclidean distance between adjacent corner points. The standard photovoltaic module's design side length is used; if the deviation δ exceeds 5%, the corner point extraction is deemed abnormal, and steps S2 to S6 need to be repeated.

6. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, If multiple corner points are extracted in step S7, multiple candidate target coordinate systems are constructed based on different corner points; by calculating the overlap between each candidate target coordinate system, the three coordinate systems with the highest overlap are selected for subsequent average calculation of attitude differences; The formula for calculating the degree of overlap C is: ; Where d is the Euclidean distance between the origins of the two compared coordinate systems. The maximum allowable distance between origin points is set to 0.03 meters. This is the arithmetic mean of the angles between three pairs of corresponding coordinate axes in two coordinate systems. The maximum allowable angle between the coordinate axes is set to 2°; and Let be the weighting coefficient, satisfying , and All values ​​are 0.

5.

7. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, The template coordinate system described in step S8 supports dynamic updates; The update rule is as follows: During continuous installation, record the attitude differences of the N most recently installed photovoltaic modules, calculate the arithmetic mean of these N translation offsets as the new translation reference, and calculate the arithmetic mean of these N rotation angles as the new rotation reference; when the deviations of the translation offsets and rotation angles in all N attitude differences relative to their arithmetic mean are less than the preset tolerance, generate and replace the original template coordinate system based on the new translation reference and rotation reference; wherein, the preset tolerance is that the translation offset does not exceed 0.01 meters and the rotation angle does not exceed 1°, and N is an integer greater than or equal to 3.

8. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, In step S10, the correction of the transition point of the robotic arm adopts a smooth transition algorithm, which realizes the motion trajectory planning from the original transition point to the compensated target transition point through cubic polynomial interpolation. The expression for the cubic polynomial interpolation is: Where τ is a normalized time variable, ranging from 0 to 1, corresponding to the start and end times of the transition phase, respectively. Let τ be the coordinates of the robotic arm's transition point in the task space at time τ; a, b, c, and d are the interpolation coefficients to be solved. The interpolation coefficients are obtained by solving for the set boundary conditions, which include: at the start of the transition, the position, velocity, and acceleration are known; at the end of the transition, the position, velocity, and acceleration are known; specifically: , , ; 、 、 ； in, , , These represent the position, velocity, and acceleration of the original transition point, respectively. , , These are the position, velocity, and acceleration of the target transition point after compensation; By solving the above system of equations to determine the coefficients a, b, c, and d, the generated trajectory must satisfy the condition that the absolute value of its jerk (the derivative of acceleration) does not exceed 5 m / s³ throughout the entire transition process.

9. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, The execution of step S10 also includes a real-time monitoring step: Local point cloud data of the robotic arm's workspace is acquired in real time using LiDAR. A plane fitting algorithm is then used to identify and fit the plane equation of the supporting plane where the photovoltaic bracket is located. The standard form of the plane equation is as follows: ; Arbitrary sampling point on the planned trajectory of the computational robotic arm Vertical distance to the fitted plane The calculation formula is: ； If there exists a sampling point Q such that the distance is... Less than the preset safe distance threshold If the current movement of the robotic arm is immediately paused and an audible and visual alarm signal is triggered; wherein, the safe distance threshold... The value ranges from 0.1 meters to 0.2 meters.

10. The photovoltaic paving robot following control method based on adaptive pose estimation according to claim 1, characterized in that, The method also includes fault-tolerant steps: The formula for calculating the effective point cloud data volume ratio η is: ； in, This refers to the number of points contained in the effective point cloud obtained after preprocessing in step S1. This represents the total number of points in the original point cloud collected by the solid-state lidar in S1. When a fault occurs, the control system interrupts the current adaptive following control process, automatically calls and executes the preset safe transition trajectory, and controls the robotic arm to return to the safe initial position or standby position along the safe trajectory. At the same time, a clear fault warning signal is issued through the robot's alarm module, and all paving operations are suspended until the operator troubleshoots the fault and confirms that the system has returned to normal.

Images