A tree-plate and inter-row oriented orchard weeding robot

Abstract

The application discloses an orchard weeding robot facing tree disks and inter-rows, comprising a self-driving chassis, a positioning and navigation system, a visual identification system, a spraying execution system, a variable spraying system and a control system. The visual identification system adopts super green feature extraction, OTSU threshold segmentation and SegFormer semantic segmentation combined with maximum inscribed circle radius geometry screening to accurately distinguish between benign and malignant weeds. The spraying execution system includes a mechanical arm end nozzle and a front end nozzle group, which are respectively used for targeted spraying in the tree disk and the inter-row. The control system takes the trunk detected by the laser radar as the center of the circle to define the tree disk, and automatically switches the execution mechanism: the mechanical arm sprays in the tree disk, and the nozzle group sprays according to the position in the inter-row. The variable spraying system fuses coverage, density, wind speed, leaf inclination and temperature and humidity, optimizes the flow through a physical drift model and a lightweight neural network, and accurately applies pesticides through adaptive PID closed-loop control. The application realizes collaborative weeding in the tree disk and the inter-row, improves the efficiency and reduces the amount of pesticide.

Description

Technical Field

[0001] This invention belongs to the fields of smart agricultural equipment technology, computer vision technology and artificial intelligence application, and specifically relates to an orchard weeding robot for tree basins and rows. Background Technology

[0002] In today's era of rapid advancements in agricultural information technology, orchard weed control is considered the scientific foundation and core technology of "smart orchards," as well as a bridge to orchard management informatization, and has become a hot research area both domestically and internationally. Orchard weeds compete with fruit trees for water, fertilizer, and light resources, and provide breeding grounds for pests and diseases, seriously affecting fruit yield and quality. Precision weeding refers to the use of intelligent identification and targeted application technology to distinguish between benign and noxious weeds in complex orchard environments, and to remove noxious weeds at specific locations. This helps reduce the use of chemical pesticides, protect the ecological balance of orchards, lower production costs, and guide orchard production and management that are limited by time and labor conditions.

[0003] Publication number CN211558608U proposes a multi-functional spraying and intelligent weeding integrated machine. This method uses a robotic arm to grab and chop weeds, while simultaneously using a water tank and rotating nozzles for spraying. This method uses visual recognition to identify the location of weeds, with the robotic arm performing the grabbing action and the nozzles spraying the herbicide, thus combining mechanical and chemical weeding. However, this method only uses a single robotic arm actuator, resulting in low efficiency in grabbing and spraying weeds one by one, making it difficult to cover large areas between rows. Furthermore, it does not consider the need for zoned operations between tree basins and rows, and is prone to collisions with fruit trees when operating near tree basins. Publication number CN118160492B proposes a multi-area automatic weeding robot, employing a dual-mechanism design of a row-to-row weeding arm and inter-plant weeding blades, with adjustable posture via lifting outriggers to adapt to different terrains. This method achieves zoned operations between rows and plants, but it uses mechanical cutting, which cannot selectively distinguish between benign and malignant weeds, and mechanical cutting can easily damage the root system of fruit trees, making it unsuitable for fine operations in tree basin areas. However, neither of the above methods can simultaneously achieve high-precision obstacle avoidance and weeding in the tree basin area and efficient continuous coverage in the inter-row area within the same work cycle.

[0004] A comprehensive analysis of existing technologies reveals that traditional orchard weeding robots mostly employ a single actuator. They may only be equipped with a robotic arm, offering high precision but low efficiency, or a spray boom, which, while efficient, cannot avoid the tree basin. This makes it impossible to simultaneously meet the diverse operational needs of the tree basin and rows within the same work cycle. With the continuous advancements in LiDAR, depth cameras, and embedded AI computing platforms, especially the development of deep learning-based visual recognition technology, robots now possess real-time perception and decision-making capabilities, which is of great significance for precise weeding in orchards. However, existing deep learning-based orchard weeding robots primarily use a single deep learning model for weed identification. Under conditions of blurred boundaries and changing lighting, they are prone to misclassifying benign weeds as noxious ones. Furthermore, they lack real-time perception and fusion of environmental parameters (wind speed, leaf tilt angle, temperature, and humidity), making it difficult to achieve environmentally adaptive variable spraying, leading to pesticide waste and environmental pollution. Summary of the Invention

[0005] (a) Technical problems to be solved This invention aims to address the inherent contradiction in existing orchard weeding robots, which, due to their single actuator, single recognition model, and static spraying decisions, cannot simultaneously achieve high-precision obstacle avoidance weeding around the tree base and high-efficiency continuous coverage between rows within the same work cycle. Specifically: existing robotic arm solutions offer high precision but low efficiency, making it difficult to cover large areas between rows; existing boom spraying solutions are efficient but cannot avoid obstacles around the tree base; existing weed identification methods cannot distinguish between benign and malignant weeds; and existing variable-rate spraying systems ignore the influence of environmental parameters such as wind speed and leaf tilt angle on droplet drift, leading to pesticide waste and environmental pollution.

[0006] (II) Technical Solution To address the aforementioned problems, this invention provides the following technical solution: a weeding robot for orchards that is designed for use between tree basins and rows, as detailed below.

[0007] An orchard weeding robot for tree basins and rows includes the following components: The self-propelled drive-by-wire chassis adopts a four-wheel independent drive and four-wheel differential control structure to carry various functional modules and autonomously navigate and move within the orchard. The positioning and navigation system includes a differential positioning RTK module and a lidar navigation module, which are used to provide centimeter-level positioning accuracy and environmental map construction; A visual recognition system, including a depth camera and an image processing unit connected to it, is used to acquire weed images in real time and identify noxious weeds; The spraying execution system includes a multi-degree-of-freedom robotic arm and an atomizing nozzle installed at the end of the robotic arm, as well as a nozzle assembly installed at the front of the robot, which are used for precise targeted spraying of noxious weeds in the tree basin area and between fruit tree rows, respectively. The variable spray system includes a flow sensor, pressure sensor, wind speed sensor, temperature and humidity sensor, diaphragm pump and solenoid valve, which is used to dynamically adjust the spray flow rate based on visual and environmental parameters; The control system includes a host computer and a slave computer; the host computer is used to run weed recognition algorithms and area determination algorithms and generate control commands, and the slave computer is used to receive commands and drive the chassis, robotic arm and nozzle assembly to move in coordination.

[0008] Preferably, the orchard weeding robot identifies benign and noxious weeds in the following way: Step S1: Extract the green weed region from the RGB image acquired by the depth camera through super green feature extraction and OTSU threshold segmentation, and perform morphological opening and closing operations on the region to remove noise, and output the weed mask region; Step S2: Input the weed mask region into the pre-trained SegFormer semantic segmentation model, identify and segment the benign weed region through forward inference of the model, and output the benign weed mask map; Step S3: Considering the morphological differences between benign weed leaves and noxious weeds, perform distance transformation on each connected component, calculate the maximum Euclidean distance from all foreground pixels to the nearest background pixel in the connected component, and use it as the maximum inscribed circle radius of the connected component. Use the inscribed circle of the calculated region to perform geometric feature filtering to further distinguish benign weeds. Step S4: Perform a pixel-by-pixel subtraction operation between the weed mask region and the benign weed mask image to extract the pixel regions not covered by the benign weed mask image in the weed mask region and output the candidate regions for malignant weeds.

[0009] Preferably, the control system switches modes in the following manner: When the noxious weeds are determined to be located in the tree basin area, the chassis brakes and the robotic arm spraying unit is activated. After receiving the target, the robotic arm controller plans an obstacle avoidance trajectory from the side standby position through inverse kinematics and moves to a position 0.2m directly above the target position. The end atomizing nozzle is then activated for targeted spraying. After spraying is completed, the robotic arm returns to the standby position, and the chassis releases the brakes and resumes movement after the robotic arm returns to its position. When a noxious weed is determined to be located in the inter-row area, the host computer sends a deceleration command to the slave computer. The chassis decelerates but does not stop. The host computer calculates the geometric offset distance based on the pixel horizontal coordinate and depth value of the weed in the image, combined with the camera's intrinsic parameters. Then, it obtains the actual offset distance based on the real-time wind speed. The host computer sends the offset distance to the slave computer via the CAN bus. The slave computer selects the corresponding nozzle in the front nozzle group based on the weed's location and the offset distance's belonging range. When both the tree basin area and the inter-row area have noxious weeds, the control system prioritizes processing the weeds in the tree basin area. The weeds in the inter-row area are marked and processed on the next pass, or they are processed in parallel through time-sequence scheduling.

[0010] Preferably, the region determination is performed as follows: The lidar scans the environment at a frequency of 10Hz, performs a DBSCAN clustering algorithm based on Euclidean distance on the acquired 3D point cloud, identifies and extracts the global coordinates of the center of each fruit tree trunk, and presets the tree basin radius R, which can be adjusted in the system parameters according to the fruit tree variety, age and crown width; the host computer receives the world coordinates of the malignant weeds output by the visual recognition system, traverses all tree trunk coordinates, calculates the Euclidean distance, takes the minimum value and records the corresponding tree trunk number, and obtains the distance d between the weed and the tree trunk; if d ≤ R, it is determined that the malignant weed is located in the tree basin area, and the "tree basin semantics" is output; if d > R, it is determined that the malignant weed is located in the inter-row area, and the "inter-row semantics" is output; the region determination results are published through the ROS2 topic / weed_semantic for the execution mechanism scheduling module to subscribe to.

[0011] Preferably, when the robotic arm spraying unit operates within the tree basin area, the control system, based on the real-time tree trunk position detected by the lidar, and the host computer, using the MoveIt2 framework of ROS2, inputs the current joint angle of the robotic arm, the target spraying pose, and tree trunk obstacles into the motion planner. The motion planner uses the Rapid Random Tree Exploration (RRT) algorithm to search for a collision-free trajectory from the current standby position to the target spraying position in the joint space of the robotic arm, with a search step size of 0.05 rad and a maximum number of iterations of 5000. If no feasible trajectory is found within a specified time, the PRM algorithm is switched to retry. After successful planning, the host computer sends the trajectory to the robotic arm controller via the CAN bus, and the robotic arm moves along the trajectory. During the movement, the lidar continuously detects the tree trunk position. If the tree trunk position changes by more than 0.05 m, the host computer immediately stops the current trajectory and re-executes the above steps to achieve dynamic obstacle avoidance. After the robotic arm reaches the target spraying position, it performs the spraying action. After spraying is completed, it returns to the standby position along the original path or the re-planned path.

[0012] Preferably, the control system is built on the ROS2 architecture; the host computer and the slave computer are connected via USB serial port, and a custom CAN-over-UART protocol is used to encapsulate ROS2 topic data into CAN frames for transmission; the host computer and the robotic arm controller are directly connected via CAN bus, and the robotic arm controller receives trajectory commands as a CAN slave station; the three-dimensional spatial coordinates of the noxious weeds output by the vision recognition system are published through ROS2 topics, and the area determination node subscribes to the topic and outputs / weed_semantic, and the actuator scheduling node subscribes to / weed_semantic and publishes topics such as chassis speed and steering commands, robotic arm target pose, and nozzle gating commands according to semantics; the publication time of all control commands is synchronized to the system clock through the NTP protocol to ensure that all nodes work together within a 10ms control cycle.

[0013] Preferably, the variable spraying system achieves precise spraying through the following steps: Step B1: Obtain the coverage rate c and density ρ of the current noxious weed area through a visual recognition system, and determine the basic flow rate Q using a pre-set two-dimensional lookup table. base , Step B2: Using the depth image acquired by the depth camera, for each connected component, take its local neighborhood of point cloud, and use the least squares method to perform local plane fitting on the point cloud of the noxious weed area to estimate the average leaf tilt angle of each weed connected component. Take the average leaf tilt angle of all pixels in the connected component as the representative tilt angle θ of the weed area. leaf , Step B3: Collect real-time wind speed v using a wind speed sensor wind The ambient humidity H is collected by a temperature and humidity sensor, and Q is converted into humidity Q. base v wind θ leaf The combination of H represents the current decision state vector; Step B4: Construct a physical drift compensation model and calculate the drift compensation coefficients. Where k1 is an empirical coefficient obtained from calibration experiments; construct the humidity correction coefficient K. humidity When H > 70%, K humidity = 0.95, otherwise K humidity = 1.00; Q base With K drift K humidity Multiply to output the corrected flow rate Q. adjusted ; Step B5: Place Q adjusted v wind θ leafThe four parameters H are input to a pre-trained lightweight neural network. This network has a single hidden layer structure with 4 input nodes, 8 hidden nodes, and 1 output node. It outputs optimized flow instructions after forward inference. Step B6: Based on real-time wind speed v wind The geometric offset distance is dynamically corrected, and the actual offset distance Δx is calculated, where Δx is the compensation amount for fog droplet drift caused by wind speed, Δx=k2*v wind *sin φ, k2 is the drift offset coefficient, and φ is the angle between the wind direction and the chassis forward direction; Step B7: Send the optimized flow command and offset distance to the lower-level machine via the CAN bus. The lower-level machine uses the optimized flow as the target value and the real-time feedback from the flow sensor as the observed value.

[0014] Preferably, the lower-level machine performs adaptive PID closed-loop control: the lower-level machine receives the optimized flow command issued by the upper-level machine through the CAN bus, collects the actual flow in real time through the flow sensor, calculates the flow deviation and the deviation change rate, adjusts the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd online, calculates the control quantity according to the adjusted PID parameters, and finally outputs the PWM duty cycle to the diaphragm pump drive pin, and opens the solenoid valve of the corresponding nozzle according to the actual offset distance.

[0015] (III) Beneficial Effects Compared with the prior art, the present invention has at least the following positive technical effects.

[0016] In existing technologies, weed identification often employs a single deep learning model or a single color threshold segmentation. Under conditions of blurred boundaries or shading, these methods cannot effectively distinguish between beneficial and harmful weeds, often resulting in the indiscriminate spraying of all green vegetation and the mistaken killing of beneficial weeds, thus disrupting orchard biodiversity and soil ecological balance. This invention extracts green weed regions through supergreen feature extraction and OTSU threshold segmentation, then identifies beneficial weeds using SegFormer semantic segmentation. Finally, a distance transformation is performed on candidate regions for harmful weeds, utilizing the geometric feature of the "maximum inscribed circle radius" to eliminate misjudged areas at the edges of beneficial weeds. This forms a two-level collaboration between deep learning coarse segmentation and geometric feature fine screening. This solution solves the problem of misjudgment at boundaries caused by a single model, effectively protecting beneficial weeds and maintaining orchard biodiversity.

[0017] (2) In the prior art, orchard weeding robots are usually equipped with only a single actuator: the pure robotic arm solution has high precision, but spraying one by one is inefficient and difficult to cover a large area between rows; the pure spray boom solution has high efficiency, but cannot flexibly avoid obstacles around the tree basin. This invention proposes an "area determination module" and "mode switching logic": the tree basin area is defined with the center of the tree trunk detected by the lidar as the center and a preset radius R. When the noxious weeds are located in the tree basin area, the chassis brakes and the robotic arm avoids obstacles and sprays at a fixed point; when they are located in the area between rows, the chassis decelerates but does not stop, and the front nozzle group selects and sprays according to the lateral position of the weeds. This solution simultaneously meets the needs of high-precision obstacle avoidance operation in the tree basin and continuous and efficient coverage between rows within the same work cycle.

[0018] (3) Existing variable spraying systems typically adjust the flow rate based on a single parameter, such as weed density or coverage, ignoring the influence of environmental factors like wind speed, leaf tilt angle, temperature, and humidity on droplet drift and deposition. This results in wasted pesticides on windy days and excessive deposition on humid days. This invention designs a variable spraying process: a base flow rate is determined by visually recognized weed coverage and density; leaf tilt angle is estimated in real-time using a depth camera; wind speed and humidity are integrated to construct a physical drift compensation model; the optimized flow rate is then output via a lightweight neural network nonlinear correction; the offset distance of the nozzle selection is dynamically corrected based on wind speed; and finally, the system is precisely executed by a lower-level machine through adaptive PID closed-loop control. This scheme achieves precise, on-demand pesticide application by integrating multiple visual and environmental parameters, saving over 90% of pesticides compared to traditional uniform spraying, significantly reducing pesticide waste and environmental pollution. Attached Figure Description

[0019] Figure 1 This is the front view of the present invention.

[0020] Figure 2 This is an overall structural diagram of the present invention.

[0021] Figure 3 This is a diagram of the internal structure of the upper part of the present invention.

[0022] Figure 4 This is a schematic diagram of the flow closed-loop control principle of the present invention.

[0023] Figure 5 This is a schematic diagram of the traffic optimization model based on shallow neural networks of the present invention.

[0024] In the attached diagram: 1. Depth camera, 2. Robotic arm, 3. Depth camera, 4. Support frame, 5. Hub servo motor, 6. Front nozzle, 7. Walking wheel, 8. Robot chassis, 9. Solenoid valve, 10. LiDAR, 11. Depth camera support, 12. Diaphragm pump, 13. Atomizing nozzle, 14. Drug delivery tube, 15. RTK antenna, 16. Medicine tank, 17. Display screen, 18. Robotic arm base, 19. Robotic arm power supply, 20. Jetson upper computer control board, 21. STM32 lower computer control board. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and implementation examples.

[0026] like Figure 1 , Figure 2 and Figure 3 As shown, one embodiment of the present invention provides an orchard weeding robot for tree basins and between rows. The robot comprises six main modules: a self-propelled wire-controlled chassis, a positioning and navigation system, a vision recognition system, a spraying execution system, a variable spraying system, and a control system. The self-propelled wire-controlled chassis consists of a robot chassis 8 and wheels 7. The wheels 7 are fixedly mounted on the bottom of the chassis 8 and employ a four-wheel independent drive and four-wheel differential steering structure to achieve autonomous navigation and movement within the orchard. The positioning and navigation system includes an RTK antenna 15 and a lidar 10. The RTK antenna 15 is a differential positioning RTK module, and the lidar 10 is used for environmental perception and tree trunk detection. The vision recognition system includes a depth camera 1 and a depth camera 3. The depth camera 3 is fixedly mounted on the front of the chassis for real-time acquisition of weed images. The spraying execution system includes a multi-degree-of-freedom robotic arm 2 and an atomizing nozzle 15 mounted at the end of the robotic arm, as well as a nozzle assembly (front nozzle 13) mounted at the front of the robot, for precise targeted spraying of noxious weeds within the tree basin area and between fruit tree rows. The variable-rate spraying system includes a medicine tank 16, a diaphragm pump 12, a flow sensor, a pressure sensor, a wind speed sensor, a temperature and humidity sensor, and a solenoid valve 9, used to dynamically adjust the spray flow rate based on visual and environmental parameters. The control system includes a host computer 20 and a slave computer 21. The host computer 20 uses a Jetson Orin Nano module, and the slave computer 21 uses an STM32 controller. The two are connected via a USB serial port. The host computer also communicates with the robotic arm controller via a CAN bus. A display screen 17 is mounted on the upper part of the chassis to display the system's operating status.

[0027] In this embodiment, the self-propelled wire-controlled chassis adopts an aluminum alloy frame, and the four walking wheels 7 are driven by hub servo motors 5 respectively. The speed and rotation angle commands are received from the lower-level machine 21 via the CAN bus. A bracket 4 is installed on the chassis, and a robotic arm base 18 is installed on the bracket, located at the center of the chassis, to fix the robotic arm 2 and balance the overturning moment generated during operation.

[0028] In this embodiment, the pesticide tank 16 of the variable spraying system is installed at the rear of the bracket 4 to provide pesticides for weeding the tree basin and between rows. The diaphragm pump 12 is installed on the left side of the bracket 4 and can be steplessly speed-regulated via PWM signals. The front nozzle 6 has four independently controlled atomizing nozzles installed at equal intervals on the front crossbeam of the chassis. The robotic arm 2 of the spraying execution system is a six-degree-of-freedom collaborative robotic arm, powered by the robotic arm power supply 19. The atomizing nozzle 13 is installed at the end of the robotic arm. The atomizing nozzle 13 and the front nozzle 6 respectively realize tree basin weeding and between-row weeding.

[0029] In its implementation, this invention first involves mapping and initialization: the robot is placed at the orchard entrance, the host computer 20 and slave computer 21 are activated, the LiDAR 10 begins scanning, and the RTK antenna 15 acquires initial latitude and longitude coordinates. The robot travels around the orchard boundary, and the host computer 20 runs the Cartographer SLAM algorithm to construct a two-dimensional grid map. Simultaneously, it identifies and records the global coordinates of the centers of all fruit tree trunks through LiDAR point cloud clustering. Then, based on the grid map and the fruit tree locations, the host computer 20 uses the coverage path planner in the Nav2 framework to generate inter-row round-trip paths. During the robot's movement, the depth camera 1, mounted on the front of the chassis, continuously captures field images and transmits them to the host computer 20 in real time. The host computer 20 processes and recognizes the image: First, it extracts green weed regions through super-green feature extraction and OTSU threshold segmentation to obtain a weed mask image. Then, it inputs the weed mask image into a pre-trained SegFormer semantic segmentation model to identify and segment benign weed regions. Distance transformation is performed on the benign weed candidate regions, and the maximum inscribed circle radius of each connected component is calculated. Regions with radii greater than 57 pixels are removed to obtain accurate benign weed regions. The weed mask image is then subtracted from the benign weed mask image to obtain the malignant weed regions. Finally, combined with the depth image, the host computer 20 calculates the 3D world coordinates of each malignant weed relative to the robot. The host computer 20 automatically determines the region based on the distance d between the weed coordinates and the center of the nearest fruit tree trunk: a preset tree basin radius R. If d ≤ R, it is determined to be a tree basin region, and tree basin weeding mode is executed; if d > R, it is determined to be an inter-row region, and inter-row weeding mode is executed.

[0030] In specific implementation of this invention, the tree basin weeding mode follows these steps: The robotic arm 2 remains on the side, and the depth camera 1 acquires real-time image data of the weeds on the side. When a malignant weed is detected, the host computer 20 sends a braking command to the slave computer 21, and the chassis 8 stops moving. The lidar 10 scans the tree basin area in real-time to acquire tree trunk point cloud data. Based on the ROS2 MoveIt2 framework, the host computer 20 plans a collision-free trajectory for the robotic arm 2 from the standby position to the target spraying position, setting the tree trunk point cloud as an obstacle. The robotic arm 2 moves along the planned trajectory to the target position, and the host computer 20 triggers variable spraying decisions, generating optimized flow commands. The slave computer 21 adjusts the PWM duty cycle of the diaphragm pump 12 and opens the solenoid valve 9 of the atomizing nozzle 13. After spraying is completed, the robotic arm 2 returns to the standby position, and the chassis 8 resumes movement.

[0031] In specific implementation of this invention, the steps of the inter-row weeding mode are as follows: When the depth camera 3 detects a persistent weed, the host computer 20 sends a deceleration command to the slave computer 21, causing the chassis 8 to decelerate but not stop. The host computer 20 calculates the lateral offset relative to the chassis centerline based on the lateral pixel coordinates of the weed in the image, and selects the corresponding nozzle among the front nozzles 6 to activate accordingly. The host computer 20 triggers a variable spraying decision, the slave computer 21 adjusts the diaphragm pump and activates the solenoid valve of the selected nozzle, and after spraying is completed, the nozzle is closed, and the chassis 8 returns to its original speed and continues to move.

[0032] In specific implementation, the variable spraying decision is based on... Figure 4 The specific steps are as follows: The host computer 20 extracts the coverage and density of noxious weeds from the visual recognition results and determines the basic flow rate using a preset lookup table. A wind speed sensor collects real-time wind speed, a depth camera estimates blade tilt angle, and a temperature and humidity sensor collects humidity. The drift compensation coefficient and humidity correction coefficient are calculated to obtain the corrected flow rate. All data are then input... Figure 5 The shallow neural network shown outputs the final optimized flow command. The host computer sends the command to the slave computer via the CAN bus. The slave computer uses a fuzzy adaptive PID algorithm to dynamically adjust the PWM duty cycle of the diaphragm pump, stabilizing the actual flow rate within ±5%, and simultaneously opening the solenoid valve of the corresponding nozzle.

[0033] In practical implementation of this invention, when multiple weeds are detected simultaneously in the tree basin and between rows, the host computer 20 processes them according to priority: first, weeds in the tree basin area are processed; second, weeds closest to the base in the between-row area are processed; the coordinates of the remaining weeds in the between-row area are temporarily stored in a cache queue. After the tree basin operation is completed, the base resumes movement and processes the weeds in the cache queue sequentially; if the weed coordinates are outside the coverage range of the front nozzle, the target is discarded. If new weeds are detected in the tree basin during the between-row operation, the current between-row spraying is immediately terminated, and tree basin weeding is performed first.

[0034] After the robot completes the weeding operation in the entire orchard area according to the planned path, the host computer 20 detects that the distance between the current position and the starting point is less than 0.5m, and determines that the operation is complete. The host computer 20 sends a stop command to the slave computer 21, the chassis 8 brakes, the robotic arm 2 returns to the home position, the diaphragm pump 12 stops, all solenoid valves close, the display screen 17 displays the operation statistics, and the system enters standby mode, completing a complete weeding operation.

[0035] The specific examples described in this application are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the specific examples described herein, or substitute them by similar means, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An orchard weeding robot for tree basins and rows, characterized in that... Includes the following parts: The self-propelled drive-by-wire chassis adopts a four-wheel independent drive and four-wheel differential control structure to carry various functional modules and autonomously navigate and move within the orchard. The positioning and navigation system includes a differential positioning RTK module and a lidar navigation module, which are used to provide centimeter-level positioning accuracy and environmental map construction; A visual recognition system, including a depth camera and an image processing unit connected to it, is used to acquire weed images in real time and identify noxious weeds; The spraying execution system includes a multi-degree-of-freedom robotic arm and an atomizing nozzle installed at the end of the robotic arm, as well as a nozzle assembly installed at the front of the robot, which are used for precise targeted spraying of noxious weeds in the tree basin area and between fruit tree rows, respectively. The variable spray system includes a flow sensor, pressure sensor, wind speed sensor, temperature and humidity sensor, diaphragm pump and solenoid valve, which is used to dynamically adjust the spray flow rate based on visual and environmental parameters; The control system includes a host computer and a slave computer; the host computer is used to run weed recognition algorithms and area determination algorithms and generate control commands, and the slave computer is used to receive commands and drive the chassis, robotic arm and nozzle assembly to move in coordination.

2. The orchard weeding robot for tree basins and rows according to claim 1, characterized in that, This orchard weeding robot identifies benign and noxious weeds using the following methods: Step S1: Extract the green weed region from the RGB image acquired by the depth camera through super green feature extraction and OTSU threshold segmentation, and perform morphological opening and closing operations on the region to remove noise, and output the weed mask region; Step S2: Input the weed mask region into the pre-trained SegFormer semantic segmentation model, identify and segment the benign weed region through forward inference of the model, and output the benign weed mask map; Step S3: Considering the morphological differences between benign weed leaves and noxious weeds, perform distance transformation on each connected component, calculate the maximum Euclidean distance from all foreground pixels to the nearest background pixel in the connected component, and use it as the maximum inscribed circle radius of the connected component. Use the inscribed circle of the calculated region to perform geometric feature filtering to further distinguish benign weeds. Step S4: Perform a pixel-by-pixel subtraction operation between the weed mask region and the benign weed mask image to extract the pixel regions not covered by the benign weed mask image in the weed mask region and output the candidate regions for malignant weeds.

3. The orchard weeding robot for tree basins and rows according to claim 1, characterized in that, The control system switches modes in the following manner: When the noxious weeds are determined to be located in the tree basin area, the chassis brakes and the robotic arm spraying unit is activated. After receiving the target, the robotic arm controller plans an obstacle avoidance trajectory from the side standby position through inverse kinematics and moves to a position 0.2m directly above the target position. The end atomizing nozzle is then activated for targeted spraying. After spraying is completed, the robotic arm returns to the standby position, and the chassis releases the brakes and resumes movement after the robotic arm returns to its position. When a noxious weed is determined to be located in the inter-row area, the host computer sends a deceleration command to the slave computer. The chassis decelerates but does not stop. The host computer calculates the geometric offset distance based on the pixel horizontal coordinate and depth value of the weed in the image, combined with the camera's intrinsic parameters. Then, it obtains the actual offset distance based on the real-time wind speed. The host computer sends the offset distance to the slave computer via the CAN bus. The slave computer selects the corresponding nozzle in the front nozzle group based on the weed's location and the offset distance's belonging range. When both the tree basin area and the inter-row area have noxious weeds, the control system prioritizes processing the weeds in the tree basin area. The weeds in the inter-row area are marked and processed on the next pass, or they are processed in parallel through time-sequence scheduling.

4. The orchard weeding robot for tree basins and rows according to claim 3, characterized in that, The region determination is performed as follows: The LiDAR scans the environment at a frequency of 10Hz, and performs a DBSCAN clustering algorithm based on Euclidean distance on the acquired 3D point cloud to identify and extract the global coordinates of the center of each fruit tree trunk. The tree basin radius R is preset and can be adjusted in the system parameters according to the fruit tree variety, age, and crown width. The host computer receives the world coordinates of the malignant weeds output by the visual recognition system, traverses all tree trunk coordinates, calculates the Euclidean distance, takes the minimum value and records the corresponding tree trunk number, and obtains the distance d between the weed and the tree trunk. If d ≤ R, the malignant weed is determined to be located in the tree basin area, and "tree basin semantics" is output. If d > R, the malignant weed is determined to be located in the inter-row area, and "inter-row semantics" is output. The region determination results are published through the ROS2 topic / weed_semantic for the execution mechanism scheduling module to subscribe to.

5. The orchard weeding robot for tree basins and rows according to claim 1 or 3, characterized in that, When the robotic arm spraying unit operates within the tree basin area, the control system, based on the real-time tree trunk position detected by the lidar, and the host computer using the ROS2 MoveIt2 framework, inputs the current joint angle of the robotic arm, the target spraying pose, and tree trunk obstacles into the motion planner. The motion planner uses the Rapid Random Tree Exploration (RRT) algorithm to search for a collision-free trajectory from the current standby position to the target spraying position in the joint space of the robotic arm, with a search step size of 0.05 rad and a maximum number of iterations of 5000. If no feasible trajectory is found within the specified time, it switches to the PRM algorithm for retry. After successful planning, the host computer sends the trajectory to the robotic arm controller via the CAN bus, and the robotic arm moves along the trajectory. During the movement, the lidar continuously detects the position of the tree trunk. If the position of the tree trunk changes by more than 0.05m, the host computer immediately stops the current trajectory and re-executes the above steps to achieve dynamic obstacle avoidance. After the robotic arm reaches the target spraying position, it performs the spraying action. After the spraying is completed, it returns to the standby position along the original path or the re-planned path.

6. The orchard weeding robot for tree basins and rows according to claim 1, characterized in that, The control system is built on the ROS2 architecture. The host computer and the slave computer are connected via USB serial port and adopt a custom CAN-over-UART protocol to encapsulate ROS2 topic data into CAN frames for transmission. The host computer and the robotic arm controller are directly connected via CAN bus, with the robotic arm controller acting as a CAN slave to receive trajectory commands. The three-dimensional spatial coordinates of the noxious weeds output by the vision recognition system are published through ROS2 topics. After the area determination node subscribes to the topic, it outputs / weed_semantic. The actuator scheduling node subscribes to / weed_semantic and publishes topics such as chassis speed and steering commands, robotic arm target pose, and nozzle gating commands according to semantics. The publication time of all control commands is synchronized to the system clock through the NTP protocol to ensure that all nodes work together within a 10ms control cycle.

7. The orchard weeding robot for tree basins and rows according to claim 1, characterized in that, The variable spraying system achieves precise spraying through the following steps: Step B1: Obtain the coverage rate c and density ρ of the current noxious weed area through a visual recognition system, and determine the basic flow rate Q using a pre-set two-dimensional lookup table. base ; Step B2: Using the depth image acquired by the depth camera, for each connected component, take its local neighborhood of point cloud, and use the least squares method to perform local plane fitting on the point cloud of the noxious weed area to estimate the average leaf tilt angle of each weed connected component. Take the average leaf tilt angle of all pixels in the connected component as the representative tilt angle θ of the weed area. leaf ; Step B3: Collect real-time wind speed v using a wind speed sensor wind The ambient humidity H is collected by a temperature and humidity sensor, and Q is converted into humidity Q. base θ leaf v wind The combination of H represents the current decision state vector; Step B4: Construct a physical drift compensation model and calculate the drift compensation coefficients. Where k1 is an empirical coefficient obtained from calibration experiments; a humidity correction coefficient k is constructed. humidity When H > 70%, k humidity = 0.95, otherwise k humidity = 1.00; Q base With K drift K humidity Multiply to output the corrected flow rate Q. adjusted ; Step B5: Place Q adjusted v wind θ leaf The four parameters H are input to a pre-trained lightweight neural network. This network has a single hidden layer structure with 4 input nodes, 8 hidden nodes, and 1 output node. It outputs optimized flow instructions after forward inference. Step B6: Based on real-time wind speed v wind Dynamic correction is applied to the geometric offset distance, and the actual offset distance Δx is calculated, where Δx is the compensation amount for fog droplet drift caused by wind speed, Δx = k2 × v wind ×sinφ, k2 is the drift offset coefficient, and φ is the angle between the wind direction and the chassis forward direction; Step B7: Send the optimized flow command and offset distance to the lower-level machine via the CAN bus. The lower-level machine uses the optimized flow as the target value and the real-time feedback from the flow sensor as the observed value.

8. The orchard weeding robot for tree basins and rows according to claim 7, characterized in that, The lower-level machine performs adaptive PID closed-loop control: the lower-level machine receives the optimized flow command issued by the upper-level machine through the CAN bus, collects the actual flow in real time through the flow sensor, calculates the flow deviation and the rate of change of deviation, adjusts the proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd online, calculates the control quantity based on the adjusted PID parameters, and finally outputs the PWM duty cycle to the diaphragm pump drive pin, and opens the solenoid valve of the corresponding nozzle according to the actual offset distance.

Images